You are on page 1of 22

Comput. Methods Appl. Mech. Engrg.

188 (2000) 505±526


www.elsevier.com/locate/cma

Factorization methods for the numerical approximation of


Navier±Stokes equations
Al®o Quarteroni a,b,*, Fausto Saleri a, Alessandro Veneziani c
a
Dipartimento di Matematica ``F. Brioschi'', Politecnico di Milano, Piazza Leonardo da Vinci 32, I-20133 Milano, Italy
b
D
epartement de Math 
ematiques, Ecole Polytechnique Fed
erale de Lausanne, CH-1015, Lausanne, Switzerland
c
Dipartimento Scienti®co e Tecnologico, Universit
a degli Studi di Verona Ca' Vignal 2, Strada Le Grazie, I-37134 Verona, Italy
Received 28 October 1998

Abstract

We investigate a general approach for the numerical approximation of incompressible Navier±Stokes equations based on splitting
the original problem into successive subproblems cheaper to solve. The splitting is obtained through an algebraic approximate fac-
torization of the matrix arising from space and time discretization of the original equations. Several schemes based on approximate
factorization are investigated. For some of these methods a formal analogy with well known time advancing schemes, such as the
projection Chorin±Temam's, can be pointed out. Features and limits of this analogy (that was earlier introduced in B. Perot, J. Comp.
Phys. 108 (1993) 51±58) are addressed. Other, new methods can also be formulated starting from this approach: in particular, we
introduce here the so called Yosida method, which can be investigated in the framework of quasi-compressibility schemes. Numerical
results illustrating the different performances of the different methods here addressed are presented for a couple of test cases. Ó 2000
Elsevier Science S.A. All rights reserved.

Keywords: Navier±Stokes equations; Algebraic splitting; Projection schemes; Quasi-compressibility methods

1. Introduction

The Navier±Stokes equations describing the motion of an incompressible Newtonian ¯uid are
ou
‡ …u  r†u ÿ mDu ‡ rp ˆ f;
ot
…1†
r  u ˆ 0;
for x 2 X; 0 < t 6 T . Here, X denotes a bounded domain of Rm …m ˆ 2 or 3†, u the velocity of the ¯uid, p
the ratio between the pressure and the constant ¯uid density, m the kinematic viscosity and f a possible
external body force. The ®rst equation in (1) represents the conservation of momentum, while the second
one is the incompressibility constraint forcing the conservation of mass. In order to have a well posed
problem, equations (1) must be provided with suitable initial and boundary conditions.
The algebraic system arising from the space-time discretization of this problem can be very large when
addressing real life applications. The ``saddle-point'' nature of the problem, with the pressure acting as a
lagrangian multiplier of the incompressibility constraint, yields a coupling between velocity and pressure
also at the discrete level (see [2]).

*
Corresponding author. Tel.: +39-2-2399-4525; fax: +39-2-2399-4588.
E-mail address: aq@mate.polimi.it (A. Quarteroni).

0045-7825/00/$ - see front matter Ó 2000 Elsevier Science S.A. All rights reserved.
PII: S 0 0 4 5 - 7 8 2 5 ( 9 9 ) 0 0 1 9 2 - 9
506 A. Quarteroni et al. / Comput. Methods Appl. Mech. Engrg. 188 (2000) 505±526

Di€erent methods have been devised for the solution of this problem, in order to reduce the compu-
tational complexity through a separate computation of velocity and pressure ®elds. The ancestor is the
Chorin±Teman projection method (see [7,31]) which is based on the Helmholtz decomposition principle.
Other methods move from the introduction of a ``small'' perturbation in the mass conservation law, re-
laxing therefore the incompressibility constraint. Since the new system resembles the one for compressible
¯ows, these approaches are known as quasi-compressibility methods (see e.g. [2,26]).
In the framework of ®nite volume discretization of the incompressible Navier±Stokes equations, Perot
(see [21]) proposed a fractional step method for the solution of the discrete linear system based on an in-
complete (or approximate) block LU factorization of the system matrix. The basic idea is to approximate the
original unsplit system with a sequence of block-triangular systems, whose solution is obviously cheaper.
Should the splitting error introduced have the same behavior than the error of the unsplit problem, then the
reduction of computational effort would be obtained without a meaningful loss of accuracy. Perot has
pointed out an analogy between block LU factorization and the Chorin±Temam projection method.
However, this analogy is more formal than substantial, and in fact the algebraic re-interpretation of the
classical projection schemes (the latter being fully driven by the physics) allows our new method to be more
accurate than the original ones. Indeed, a pure ``physical'' splitting as the one underlying the Chorin±
Temam method would demand spurious boundary conditions for the pressure, which eventually com-
promise the solution accuracy. On the other hand, the explicit prescription of such conditions is by no
means demanded by our methods. Moreover, the algebraic factorization approach is easy to implement and
generalize. We start showing that different projection schemes proposed in literature in recent years gen-
eralizing Chorin±Temam's, can be reformulated as being suitable block-factorization of the original system.
Since the analogy is not always a strict equivalence, it is worthwhile to investigate its feature and limits. In
particular, we prove that if the space discretization is based on either spectral methods or ®nite differences
on a staggered grid, then the Perot block factorization and the Chorin±Temam scheme do actually coincide.
We then propose several possible generalizations and modi®cations of Perot's idea, based on the use of
inexact LU factorization, through the replacement in the L and U factors of the inverse elliptic matrix by
suitable approximations that are cheaper to use. The nature of this approximation can be characterized as
either momentum preserving or mass preserving; in the former case it modi®es the continuity equation
(yielding a quasi-compressibility method), while in the latter the mass is conserved, but the momentum
equations are perturbed. In particular, we introduce a new quasi-compressibility method that is given the
name of Yosida method as it is based on the Yosida regularization of the Laplace operator.
The paper is organized as follows. In Section 2 we recall some basic issues about the space and time
discretization of Navier±Stokes equations. In Section 3 we introduce the linear system arising from the
discretization and linearization of the Navier±Stokes problem and the algebraic factorization which is the
basis for the methods we want to consider. In Section 4 we consider a general framework for inexact
factorizations of the linear systems at hand. In particular, in Section 5 it is shown that the Perot splitting
can be recovered as a particular case of this inexact factorization. The extent of the analogy between such
factorization and several variants of Chorin±Temam projection method is stated in this section.
In Section 6, we introduce the Yosida scheme, and investigate the links with other schemes proposed in
literature, in particular with respect to pseudo-compressibility methods.
In Section 7, the properties of some of the introduced schemes are numerically investigated on a couple
of test cases. Our emphasis is on two methods, the so called ``algebraic'' Chorin±Temam method, intro-
duced in [21] and the Yosida method. The obtained results show very good performances of these methods
with respect to the classical Chorin±Temam projection method and the quasi-compressibility method based
on a stabilization with the Laplacian of the pressure. Indeed, the results obtained with a ®nite element
P1 isoP2 ; P1 space discretization show better accuracy for the computed velocity and pressure ®leds and a
very good ful®llment of the divergence free constraint.

2. Numerical approximation: algebraic formulation

The analytical solution of problem (1), together with the associated initial and boundary conditions, is
seldom known in a closed form: in order to compute a numerical solution, discretization with respect to the
A. Quarteroni et al. / Comput. Methods Appl. Mech. Engrg. 188 (2000) 505±526 507

time and the spatial variables is carried out. Although the approaches can be very di€erent, typically the
combined time and space discretization yields as each time level tn‡1 in a system of equations of the form

Ayn‡1 ˆ bn‡1 ; …2†

where
     
C DT Wn‡1 b1n‡1
Aˆ ; yn‡1 ˆ ; bn‡1 ˆ : …3†
D 0 Pn‡1 b2n‡1
We have supposed that the time interval of interest …0; T Š has been divided into N subintervals ‰tk ; tk‡1 Š,
where tk‡1 ÿ tk ˆ Dt; k ˆ 0; 1; . . . ; N ÿ 1.
The ®rst row of (3) arises from the discretization of the momentum equation, the second from the mass
T
conservation. The unknowns ‰Wn‡1 ; Pn‡1  Šn‡1are
 related to the nodal values of the velocity and pressure
solution. Typically, it is W ˆ U ˆ Ui , where Uki is the value of the computed velocity vector at the
n‡1 n‡1

ith node and at time tk ; similarly Pn‡1 ˆ Pn‡1 ˆ ‰Pin‡1 Š, where Pik is the value of the computed pressure at the
T
ith node and at time tk . However, also different choices for the unknowns ‰Wn‡1 ; Pn‡1 Š are possible, as we
clarify in the next sections.
Due to the presence of the non-linear convective term of the momentum equation …1†1 , in fully implicit
time-stepping the submatrix C will depend on Wn‡1 , so (3) is actually a system of non-linear equations.
Suitable linearization strategies can be adopted in order to obtain a linear system; we will consider some of
these strategies in the next subsections. However, in the sequel, we will suppose that system (3) has already
been linearized in some way; moreover, we will assume that the matrix A is non-singular. It is well known
that if C is non-singular and (positive or negative) de®nite, A is invertible if and only if Ker…DT † ˆ 0, which
is the case if the Ladyzhenskaja±Babuska±Brezzi (LBB) condition between the discretization spaces for the
velocity and the pressure holds (see e.g. [2,25, ch. 9]).

2.1. Time discretization

We recall some approaches for time-discretization of the Navier±Stokes equations, together with some
linearization strategies, that amenate to a linear system of the form (3).
Euler±Newton method: Assume that the Navier±Stokes equations are advanced in time by the backward
Euler scheme, then the resulting non-linear boundary value problem at the new time level tn‡1 may be
linearized around un ; pn . This corresponds precisely to carry out one iteration of Newton method. The
resulting equations read:
un‡1 un
ÿ mDun‡1 ‡ rpn‡1 ‡ un  run‡1 ‡ un‡1  run ˆ ‡ un  run ‡ f n‡1 ; …4†
Dt Dt

r  un‡1 ˆ 0; …5†
yielding a ®rst order scheme with respect to Dt.
Euler semi-implicit method: It is derived from (4) by assuming that un‡1  run approximately equals
u  run . We have therefore
n

un‡1 un
ÿ mDun‡1 ‡ rpn‡1 ‡ un  run‡1 ˆ ‡ f n‡1 …6†
Dt Dt

together with (5); this is still a ®rst order scheme with respect to Dt.
Crank±Nicolson/Adams±Bashforth method: This time, we use the trapezoidal (Crank±Nicolson) rule to
advance the linear part of the momentum equations, while the convective term is linearly extrapolated from
the two previous time-levels. We obtain

2 n‡1 2
u ÿ mDun‡1 ‡ rpn‡1 ‡ 3un  run ÿ unÿ1  runÿ1 ˆ un ‡ f n ‡ f n‡1 ‡ mDun ÿ rpn …7†
Dt Dt
508 A. Quarteroni et al. / Comput. Methods Appl. Mech. Engrg. 188 (2000) 505±526

together with (5); this is a second order scheme with respect to Dt. The treatment of the convective term
corresponds to having used the Adams±Bashforth second order scheme (see e.g. [8]).

2.2. Space discretization

The algebraic system (2) is now obtained after discretizing the space variables in the previous equations.
With this aim, let us consider for instance a Galerkin approximation and, for the sake of simplicity, the case
of homogeneous Dirichlet conditions u ˆ 0 on oX. Then, the Galerkin approximation of the problem reads:
for each n P 0 ®nd whn‡1 2 Vh and phn‡1 2 Qh such that
8 ÿ  ÿ  ÿ n‡1  ÿ  ÿ n‡1 
< m whn‡1 ; vh ‡ a wn‡1 n
h ; vh ‡ b wh ; vh ‡ g vh ; ph
n‡1
ˆ h ; vh ;
…8†
: gÿwn‡1 ; q  ˆ 0;
h h

for all vh 2 Vh and qh 2 Qh , where fVh ; h > 0g and fQh ; h > 0g are families of ®nite dimensional subspaces
for the velocity and the pressure ®elds, respectively. In (8) we have set:
Z Z
a
m…w; v† ˆ w  v dx; a…w; v† ˆ m rw : rv dx;
Dt X X
…9†
Z
g…w; q† ˆ ÿ r  wq dx;
X

and the value of a, the actual expression of bn …; † and the one of hn‡1 depend on the time discretization
scheme as well as on the linearization adopted.
Let us consider some instances, starting from the ®nite element discretization. We introduce a triangu-
lation of the domain X, i.e. a ®nite decomposition of X  into triangles or quadrilaterals in 2D or tetrahe-
drons or parallelograms in 3D. As usual, denote with h the maximum of the inner diameters of the
elements. Then, assume that Vh is the space of piecewise polynomial functions on every element of the
decomposition of X, continuous in X and vanishing on the boundary oX. Similarly Qh is the space of
piecewise polynomial functions on every element of the decomposition. In order to satisfy the LBB com-
patibility condition, the choice of the polynomial degrees for the velocity and the pressure must undergo
some restriction. As an instance, in the case of triangular elements, we could assume that Vh is the space of
quadratic functions on every element of the mesh, while Qh is given by the continuous piecewise linear
functions on every element. Other suitable couples of compatible spaces can be found, e.g., in [2,13,25].
Denote with fui giˆ1;...;Nu a set of basis functions for the space Vh and with fwi giˆ1;...;Np the set of basis
functions for Qh . Finally, set:
R 
(i) M ˆ ‰mij Š ˆ X ui  uj dx (the mass matrix),
(ii) K ˆ ‰kij Š ˆ ‰a…ui ; uj †Š (the sti€ness matrix),
PNu R 
(iii) B1 …W† ˆ ‰b1;ij …W†Š ˆ kˆ1 Wk X …uk  ruj †  ui dx ,
PNu R 
(iv) B2 …W† ˆ ‰b2;ij …W†Š ˆ kˆ1 Wk X …uj  ruk †  ui dx ,

(v) D ˆ ‰dij Š ˆ ‰g…uj ; wi †Š,


(vi) F ˆ ‰fi Š ˆ ‰…f; ui †Š and U0 ˆ ‰U0i Š ˆ ‰…u0 ; ui †Š.
If we combine the Euler±Newton method and a ®nite element space discretization, we obtain the system
(2) upon setting:
Wn ˆ Un ; Pn ˆ Pn ; C ˆ Dt1 M ‡ K ‡ B1 …Wn † ‡ B2 …Wn †;
ÿ  …10†
bn‡1
1 ˆ Fn‡1 ‡ 1
Dt
M ‡ B1 …Wn † Wn ; b2n‡1 ˆ 0:

Similar relations can be easily obtained when using the Euler semi-implicit method or the Crank±
Nicolson/Adams±Basforth scheme, still in combination with ®nite elements.
A. Quarteroni et al. / Comput. Methods Appl. Mech. Engrg. 188 (2000) 505±526 509

Fig. 1. Marker±Cell staggered grid.

If we consider a spectral discretization, we assume that the space of shape functions of approximate both
m
velocity and pressure is given by polynomials of high degree. More precisely, we will set Vh  VN ˆ ‰Q0N Š
being QN the space of global algebraic polynomials of degree less than or equal to N in each variable
xi ; i ˆ 1; . . . ; m, and Q0N the subspace of those vanishing on oX. In order to satisfy the LBB condition a
suitable (although not exclusive) choice consists of setting Qh ˆ QN ÿ2 , where the star means that the av-
erage on X is null. For more details see [8]. Proceeding as in the ®nite element case, we can still recover a
system like (2). However, because of the choice of global polynomial spaces, the exact evaluation of the
integrals corresponding to the terms (i)±(vi) of Section 2.1 is computationally intensive. For this reason, the
integrals are approximated by suitable Gaussian numerical quadrature formulae (see [8]).
More generally, the computational domain can be preliminary partitioned into big quadrilateral ele-
ments (called spectral elements), then Gauss±Lobatto integration can be carried out within every element
(see e.g. [3]). The corresponding scheme is a generalized Galerkin method (i.e. a Galerkin method with
numerical integration), which can be interpreted as a spectral collocation method owing to the particular
choice of integration points.
Finally, a non-Galerkin approach, which however still gives a system in the form (2) is the one based on
the ®nite differences discretization. A typical choice for the stencil is the so called Marker±Cell staggered
grid, represented in Fig. 1 (see e.g. [34]), having Dx ˆ Dy ˆ h. The structure of the system obtained by
means of this spatial discretization is actually the same of (3).
Observe that, in this case, the mass matrix M is equal to the identity matrix, while in spectral collocation
or ®nite volume methods, the mass matrix M is a diagonal matrix. In all these cases, the matrix M ÿ1 is easily
computable. Conversely, in the ®nite element methods, the mass matrix M is not diagonal; however, a
common strategy to make effective the computations is the so-called mass-lumping, that is replace M with
ML , a suitable diagonalization of the original (velocity) mass matrix: for different strategies of lumping
(especially for polynomial spaces with degree greater than 1) see e.g. [1]. In the case of linear elements,
lumping amounts to perform the computation of the mass-matrix elements via a ®rst-order accurate
quadrature formula (see. e.g. [25, ch. 6]). In the sequel, M will denote either the original mass matrix or its
lumped approximation (if it is not diagonal).

3. Block factorizations of the algebraic system

For the sake of simplicity, in this section assume that Wn‡1 ˆ Un‡1 and Pn‡1 ˆ Pn‡1 . In the following
Section 5.2, we will examine the schemes arising from di€erent choices for W and P.
If Nu denotes the number of degree of freedom for each component of the velocity ®eld and Np the
number of degrees of freedom for the pressure ®eld, the dimension of the system (2) is m  Nu ‡ Np , which
can be pretty large number especially in 3D problems. A possible strategy to reduce the computational
510 A. Quarteroni et al. / Comput. Methods Appl. Mech. Engrg. 188 (2000) 505±526

e€ort can be to resort a suitable splitting of the problem, for instance generating subproblems for the
velocity and the pressure separately.
In this context, a typical approach consists of generating a linear system for the pressure, by formal
elimination of the velocity vector. As a matter of fact, from (2) and (3) we obtain
Un‡1 ˆ C ÿ1 …b1 ÿ DT P†;
…11†
DC ÿ1 DT P ˆ DC ÿ1 b1 ÿ b2 :
The solution of the system (2) and (3), therefore, can be obtained by solving the system …11†2 for the
pressure and recovering afterwards the velocity from the …11†1 . The computational e€ort of this approach,
however, can be unsatisfactory, since the algebraic solution of …11†2 is generally still fairly expensive (see
[25]), Section 9.6.1). In any case, we remark that this strategy can be reinterpreted in the framework of
suitable LU block factorizations of system (3). Indeed, we have
  
C 0 I C ÿ1 G
Aˆ ; …12†
D ÿ DC ÿ1 G 0 I

where G ˆ DT . This factorization introduces a splitting for the computation of velocity and pressure vari-
ables. As a matter of fact, we can compute the exact values Un‡1 and Pn‡1 by means of the sequence of steps:
(
CUe n‡1 ˆ bn‡1 ;
1
L ÿ step :
e n‡1
D U ÿ DC ÿ1 G P e n‡1 ˆ bn‡1 ;
2
( …13†
Un‡1 ‡ C ÿ1 GPn‡1 ˆ U e n‡1 ;
U ÿ step :
Pn‡1 ˆ P e n‡1

or, equivalently:
e n‡1 ˆ bn‡1
CU …intermediate velocity computation†;
1
ÿ1
ÿ DC GP n‡1 e n‡1 …pressure computation†;
ˆ bn‡1 ÿ D U …14†
2
n‡1 e
CU ˆ C U ÿ GPn‡1
n‡1
…end-of-step velocity computation†;
which amounts exactly to the solution of (11). However, in order to reduce the computational e€ort,
suitable inexact factorizations (i.e. approximations of the exact block factorizations (12)) can be intro-
duced. These approximations are acceptable if they induce an error not larger than the truncation error due
to the time discretization itself. An interesting feature is that, in some cases, it is possible to ®nd an analogy
between approximate factorizations and other classical schemes that obtain a splitting of the computation
of velocity and pressure starting from a completely di€erent approach, such as the projection schemes. This
analogy, already observed by Perot in the case of the Chorin±Temam method (see [21]), will be carefully
investigated in the following sections for several other schemes that have an algebraic counterpart in terms
of inexact factorizations.

4. Inexact factorizations

The computational complexity of the scheme (13) can be reduced provided C ÿ1 in (12) is replaced by a
simpler matrix. In an abstract form, we adopt the following viewpoint. Suppose that the matrix C ÿ1 is
approximated by a matrix H1 in the L-block and H2 in the U one. The following approximate matrix is
obtained
    
C 0 I H2 G C CH2 G
Aapprox ˆ ˆ : …15†
D ÿ DH1 G 0 I D D…H2 ÿ H1 †G

The ®rst conclusion that can be drawn from (15) is that the mass conservation is guaranteed if H1 ˆ H2 .
Indeed, setting H ˆ H1 ˆ H2 …6ˆ C ÿ1 †, we have
A. Quarteroni et al. / Comput. Methods Appl. Mech. Engrg. 188 (2000) 505±526 511

 
C CHG
Aapprox ˆ AH ˆ : …16†
D 0

Using AH instead of A amounts therefore to modify the pressure operator in the momentum equation
and the amount of this perturbation depends obviously on the extent at which H approximates C ÿ1 .
Noticeable instances of this kind of ``mass-preserving'' strategy are considered in the next subsection.
Another strategy consists of taking in (15) H2 ˆ C ÿ1 and H1 6ˆ H2 , yielding an approximate matrix of the
form
 
C G
Aqc ˆ with Q ˆ D…H1 ÿ C ÿ1 †G: …17†
D Q

Matrices like (17) arise in the framework of quasi-compressibility and pressure regularization methods,
which entail a regularization of the continuity equation by adding terms depending on the pressure or its
derivatives. Notice, however, that this time the strategy is ``momentum-preserving'', since the momentum
equation is unperturbed. We will further develop this issue in Section 6.
The third and most general strategy is to consider H1 6ˆ H2 and both di€erent than C ÿ1 . This amounts to
modify both momentum and continuity equations: although very interesting in principle, this strategy will
not be considered in the present work.

Remark 4.1. The block LU factorization (12) is not the only possible way to split the matrix A. Different
fractional step schemes can be formally obtained based, e.g., on a L1 D1 R1 block factorization, L1 being block-
lower-triangular, D1 block-diagonal and R1 block-upper-triangular. Also in this case, inexact factorizations
ought to be used in order to obtain schemes more effective than the solution based on (11).

5. Projection methods

The principle of projection methods is to compute the velocity and the pressure ®elds separately, through
the computation of an intermediate auxiliary velocity, which is then projected on the subspace of the so-
lenoidal vector functions. The theoretical framework is provided by the Helmholtz decomposition principle
or Ladhyzhenskaja theorem (for the proof, see e.g. [13]). Denoting, as usual, by L2 …X† the space of functions
whose square is integrable in X and by H 1 …X† the space of L2 …X† functions with ®rst derivative in L2 …X† (in
bold face are denoted the corresponding spaces for vector functions), the decomposition principle reads:

Theorem 5.1. Any vector function v 2 L2 …X† can be uniquely represented as v ˆ w ‡ rw with w 2 L2 …X†,
r  w ˆ 0; w  n ˆ 0 on oX and w 2 H 1 …X†.

In this way, the vector v is split into its solenoidal part w and its irrotational part rw.

5.1. Non-incremental schemes

The most classical projection method is the Chorin±Temam scheme (see [7,31]); its semi-discrete version,
for a fully homogeneous Dirichlet problem (i.e. u ˆ 0 on oX), reads as follows. Let u0 be given.
1. Find a preliminary (intermediate step) velocity ~un‡1 of the velocity ®eld as the solution of the following
semi-discrete problem:

1 n‡1
un‡1 ‡ …u  r†u ˆ f n‡1
f~u ÿ un g ÿ lD~ in X; …18†
Dt
with ~un‡1 ˆ 0 on oX; u and u are to be chosen suitably for the treatment of the non-linear term (e.g.
u ˆ u ˆ un for the explicit Euler, ˆ un‡1 for the implicit one and u ˆ un and u ˆ un‡1 for the semi-
implicit Euler).
512 A. Quarteroni et al. / Comput. Methods Appl. Mech. Engrg. 188 (2000) 505±526

2. Determine un‡1 and pn‡1 as being the solution of:


8
< 1 fun‡1 ÿ ~un‡1 g ‡ rpn‡1 ˆ 0
Dt in X; …19†
:
r  un‡1 ˆ 0;

with un‡1  n ˆ 0 on oX.


This step can be reformulated in a way allowing the computation of pn‡1 and un‡1 separately. Indeed, by
formally applying the divergence operator on (19), we obtain the following Neumann problem for pn‡1 :

1
Dpn‡1 ˆ un‡1
r~ in X;
Dt
…20†
n‡1
op
ˆ0 on oX;
on

Observe that the Neumann conditions on the pressure are a by-product of the Ladyzhenskaja theorem,
and however they are not veri®ed by the solution of the unsplit original problem. This problem of the
pressure conditions, analyzed by Rannacher (see [26]), is particularly relevant for non-Dirichlet problems,
for which is not clear what has to be prescribed (see also [19]).
Finally, the end-of-step velocity ®eld is obtained by updating ~un‡1 as follows:

un‡1 ˆ ~un‡1 ÿ Dtrpn‡1 : …21†

With reference to the Ladyzhenskaja theorem, the intermediate velocity ~un‡1 plays the role of the vector
v, while the end-of-step velocity un‡1 corresponds to the solenoidal vector w and Dt pn‡1 to the scalar w (see
(21)).
On the basis of the above decomposition, the Chorin±Temam splitting is able to ensure the mass con-
servation of the solution (as the end-of-step velocity is divergence free). However, the tangential component
of the velocity ®eld cannot be controlled on the boundary, as from the Helmholtz decomposition principle
we can prescribe only a condition on the normal component of the velocity ®eld u. Another potential
drawback of the Chorin±Temam method arises from the spurious boundary conditions for the pressure,
introduced by the second of the (20).
A discrete counterpart of the Chorin±Temam scheme can be obtained in the framework of the abstract
approximate factorization (15) proceeding as follows (see [21,32]). Suppose for instance to adopt a semi-
implicit linearization in the step (18), which corresponds to choose u  un and u  un‡1 . Then, using the
same notation as in (10), it is easily veri®ed that the actual expression for the matrix C is

1
Cˆ M ‡ S; …22†
Dt
where S ˆ K ‡ B1 (the dependence of B1 on Wn is understood). Then
 ÿ1
ÿ1 1 ÿ1
C ˆ M ‡S ˆ Dt…I ‡ DtM ÿ1 S† M ÿ1 : …23†
Dt

Now, in the inexact factorization (15) we could use the following approximation of C ÿ1

H ˆ H1 ˆ H2 ˆ DtM ÿ1 : …24†

This is a ®rst order approximation of C ÿ1 with respect to Dt. As a matter of fact, suppose that
q…DtM ÿ1 S†, the spectral radius of DtM ÿ1 S, is less than one, which is obviously the case if Dt < 1=q …M ÿ1 S†.
Then, we can write
A. Quarteroni et al. / Comput. Methods Appl. Mech. Engrg. 188 (2000) 505±526 513

X
1
j j
C ÿ1 ˆ Dt …ÿ1† …DtM ÿ1 S† M ÿ1 ˆ Dt…I ÿ DtM ÿ1 S ‡   †M ÿ1 : …25†
jˆ0

With the approximation (23), (14) becomes


e n‡1 ˆ bn‡1 ;
CU …26†
1

e n‡1 ;
ÿDtDM ÿ1 GPn‡1 ˆ b2n‡1 ÿ D U …27†

e n‡1 ÿ DtM ÿ1 GPn‡1 :


Un‡1 ˆ U …28†
The matrix DM ÿ1 G has some analogy with the discretization of the Laplace operator ÿD ˆ ÿr  …r†,
especially in the case where M ˆ I, since D is related to the operator (ÿr) and G ˆ DT to the gradient
(indeed, sometimes it is referred to as discrete Laplacian (see e.g. [6,32])). On this ground, the analogy
between (27) and (20) can be easily inferred. On the other hand (26) is the space discretization of (18), i.e.
the fully discrete counterpart of the ®rst step in the Chorin±Temam scheme. Finally, (28) is clearly related
to the updating step (21). This analogy was already observed by Perot (see [21]) in the framework of a ®nite
volume discretization. As we will point out in Section 5.3, this identi®cation between (26)±(28) and the
Chorin±Temam scheme in general is only formal, being based on a formal analogy between DM ÿ1 G and the
Laplace operator (with homogeneous Neumann conditions), although, in some cases, this analogy is ac-
tually a rigorous equivalence, as we will point out later on.
In conclusion, we outline the fact that the algebraic subproblems (26)±(28) (as well as other subproblems
obtained by other approximate factorizations, see e.g., (59) in Section 6.1) are from one hand inspired by
the simpler subproblems underlying the classical projection method (namely, di€usion and transport step
separated from the propagation step). From the other hand, these subproblems are furtherly easined by the
fact that no additional boundary condition has to be provided.

Remark 5.1. Approximation (23) (as well as the splitting (26)±(28)) can be improved if we take instead of (24)
a second order approximation of C ÿ1 , that is
H1 ˆ H2 ˆ Dt…I ÿ DtM ÿ1 S†M ÿ1 ; …29†
or a third order one:
 ÿ 2 
H1 ˆ H2 ˆ Dt I ÿ DtM ÿ1 S ‡ Dt2 M ÿ1 S M ÿ1 : …30†

For more details about this strategy,in the framework of spectral element methods, see [9]. In particular, as
observed in [9], in the case of the Stokes problem, the matrix (29) is not definite positive for all values of the
Reynolds number, whereas the matrix (30) is always definite positive.

Remark 5.2. As already pointed out, the conditions imposed on the end-of-step velocity are u  n ˆ 0 instead
of u ˆ 0. In practice, the problem can be bypassed by forcing the null homogeneous boundary condition also on
the tangential velocity in an essential way (see [11]).
Another strategy consists of assuming the intermediate velocity (which satisfies the complete homogeneous
Dirichlet conditions) as the correct velocity field to be considered (see [12,16]).
In Section 6.1 we introduce an inexact factorization scheme, the Yosida method, for which the end-of-step
velocity, conversely, satisfies automatically the correct set of conditions.

Remark 5.3. When the matrix C ÿ1 is approximated according to (24), the inexact block LU-factorization (12)
becomes
    
C 0 I DtM ÿ1 G C G ‡ DtSM ÿ1 G
ACT ˆ ˆ : …31†
D ÿ DtDM ÿ1 G 0 I D 0
514 A. Quarteroni et al. / Comput. Methods Appl. Mech. Engrg. 188 (2000) 505±526

This is the algebraic formulation of the problem which is actually solved if the splitting (26) is performed.
Since
 
0 DtSM ÿ1 G
ECT  ACT ÿ A ˆ ; …32†
0 0

we conclude that the discrete projection method amounts to perturb the momentum equation of original
problem by a O…Dt† term acting on the pressure.

5.2. Incremental schemes

A modi®cation of the original Chorin±Temam method (18)±(20) is represented by the so-called incre-
mental projection scheme (see e.g. [16,18,29,30]). In its time-discrete/space-continuous formulation it reads
as follows. Let u0 be given and let us denote by p0 an (arbitrary) initial condition for the pressure.
1. Find a preliminary (intermediate-step) velocity ~un‡1 as the solution of the following semi-discrete
problem:
1 n n‡1 o
~u ÿ un ÿ mD~ un‡1 ˆ f n‡1 ÿ rpn in X;
un‡1 ‡ …un  r†~ …33†
Dt

with ~un‡1 ˆ 0 and oX;


2. Determine un‡1 and pn‡1 as the solution of
8 n o
< 1 un‡1 ÿ ~un‡1 ‡ rÿpn‡1 ÿ pn  ˆ 0
Dt in X …34†
: n‡1
Du ˆ0

with un‡1  n ˆ 0 on oX. Equivalently, (34) can be replaced by the two steps:

1
D…pn‡1 ÿ pn † ˆ un‡1
r~ in X
Dt
…35†
n‡1 0
op op
ˆ on oX
on on

and
ÿ 
un‡1 ˆ ~un‡1 ÿ Dtr pn‡1 ÿ pn : …36†
This scheme has been analyzed by Guermond and Guermond and Quartapelle (see [12,16]). Under
regularity hypothesis on the forcing term and the initial conditions, they proved that if the solution of the
unsplit discrete problem (3) is ®rst order accurate with respect to Dt and h and Dt is suf®ciently small, then,
also the incremental projection scheme is ®rst order accurate. More precisely, if en denotes the error on the
velocity ®eld and gn the error on the pressure at time tn , then
" #1=2 " #1=2
X
N
2
X
N
2
n
max ke kL2 …X† ‡ Dtken kH1 …X† ‡ Dtkgn kL2 …X† 6 c…Dt ‡ h†: …37†
0<n 6 N
nˆ0 nˆ0

A discrete counterpart of the incremental projection scheme is readily obtained as an inexact factor-
ization of the matrix A. Indeed, let us set Wn‡1 ˆ Un‡1 and
Pn‡1 ˆ Pn‡1 ÿ Pn : …38†
A. Quarteroni et al. / Comput. Methods Appl. Mech. Engrg. 188 (2000) 505±526 515

Then, the system (2) modi®es correspondingly as follows


 
n‡1 n‡1 GPn
Ay ˆ b ÿ  rn‡1 : …39†
0

Using the same matrix decomposition as in (15) with the approximation of C ÿ1 indicated in (24), the
fully discrete counterpart of the incremental projection scheme (33)±(36) can still be expressed by (26)±(28),
taking into account the di€erent right hand side rn‡1 .

5.2.1. Van Kan scheme


The Van Kan scheme in the time-discrete/space-continuous formulation reads as follows (see [33] and
also, e.g. [23]):
un‡1 of the velocity ®eld as the solution of the following semi-discrete problem
1. Find a preliminary guess ~
1 n n‡1 o 1 1 3 1
~u ÿ un ÿ mD~ un‡1 ‡ rpn ˆ f n‡1=2 ‡ mDun ÿ …un  r†un ‡ …unÿ1  r†unÿ1 in X; …40†
Dt 2 2 2 2
with ~un‡1 ˆ 0 on oX and f n‡1=2 ˆ …1=2†…f n‡1 ‡ f n †;
2. Determine un‡1 and pn‡1 as the solution of
1 n n‡1 o 1 ÿ 
u ÿ ~un‡1 ‡ r pn‡1 ÿ pn ˆ 0;
Dt 2 in X …41†
n‡1
ru ˆ0
with un‡1  n ˆ 0 on oX, equivalently:
ÿ  2
D pn‡1 ÿ pn ˆ r  ~ un‡1 in X;
Dt
rpn‡1  n ˆ rpn  n on oX …42†
Dt ÿ 
un‡1 ˆ ~un‡1 ÿ rpn‡1 ÿ rpn :
2
Now, suppose to adopt the Crank±Nicolson/Adams±Bashforth time discretization (7). The fully discrete
problem for a homogeneous Dirichlet problem becomes:
 
1 1 1
M ‡ K Un‡1 ‡ GPn‡1
Dt 2 2
1 ÿ n‡1  1 1 1 3 1 ÿ  …43†
ˆ F ‡ F ‡ MUn ÿ KUn ÿ GPn ÿ B1 …Un †Un ‡ B1 Unÿ1 Unÿ1
n
2 Dt 2 2 2 2
DUn‡1 ˆ 0;

where the meaning of the symbols is the same as in (3).


We can show that the discrete counterpart of the Van Kan scheme (40)±(42) can be regarded as an
inexact factorization of the matrix A corresponding to the system (43). Indeed, if we de®ne
1 1   1
Svk ˆ K; Cvk ˆ M I ‡ DtM ÿ1 Svk ; Gvk ˆ G; …44†
2 Dt 2
the system (43) can be formulated as (3), with:
   
Cvk Gvk n‡1 Un‡1
Aˆ ; y ˆ ;
D 0 Pn‡1 ÿ Pn …45†
1ÿ  1 3 1 ÿ 
bn‡1
1 ˆ Fn‡1 ‡ Fn ‡ MUn ÿ Svk Un ÿ Gvk Pn ÿ B1 …Un †Un ‡ B1 Unÿ1 Unÿ1 :
2 Dt 2 2
Now, performing on A the splitting (12), with Cvk instead of C and Gvk instead of G and considering the
ÿ1
inexact factorization arising from the replacement of Cvk by DtM ÿ1 , the fully discrete counterpart of the
Van Kan scheme is obtained. A different (not LU) factorization of the matrix A corresponding to the Van
516 A. Quarteroni et al. / Comput. Methods Appl. Mech. Engrg. 188 (2000) 505±526

Kan scheme can be found in [10]. Moreover, an extensive analysis of (a reformulation of) the Van Kan
scheme can be found in [22].

5.3. Some properties of the discrete Laplace operator

As we pointed out previously, the analogy between the matrix DM ÿ1 G and the approximation of the
Laplace operator ÿD is only formal as it is not (in general) a rigorous equivalence. Consider, for instance,
the case of ®nite element space discretization; the matrix M is not the identity (is not even a diagonal
matrix, if the mass lumping is not adopted). Furthermore, even though the matrix ÿD is the discretization
of the divergence operator and G that of the gradient operator, we could not expect ÿDM ÿ1 G to coincide
with the discretization of the Laplacian, even if we would assume M ˆ I.
However, there are cases in which the equivalence holds in a rigorous way, i.e. for particular space
discretizations, as stated in the following propositions.

Proposition 5.1. Assume that the space discretization is obtained through a finite difference scheme based on
the MAC grid of Fig. 1. The discretization of the operator ÿD associated to homogeneous Neumann conditions
corresponds exactly to the matrix DDT (equivalently, to DM ÿ1 G since M ˆ I and G ˆ DT in this case).

Proof. Just for the sake of simplicity suppose that the domain X is entirely represented in Fig. 1, i.e. the
dotted line of the picture actually correspond to the boundary oX. A second order (with respect to h) choice
for the discretization of Dp evaluated in the linear node i; j is given by
1 1
…Dh p†i;j ˆ …pi‡1;j ÿ 2pi;j ‡ piÿ1;j † ‡ 2 …pi;j‡1 ÿ 2pi;j ‡ pi;jÿ1 †: …46†
h2 h
In the same node, suppose that the discrete divergence operator …r†h applied to the vector v ˆ ‰v…1† ; v…2† Š
is de®ned by
1  …1† …1†
 1
…2† …2†

…rh  v†i;j ˆ vi‡1=2;j ÿ viÿ1=2;j ‡ vi;j‡1=2 ÿ vi;jÿ1=2 …47†
h h
and, correspondingly, the discrete gradient operator rh is de®ned componentwise by
   
oh q 1 oh q 1
ˆ …qi‡1;j ÿ qi;j †; ˆ …qi;j‡1 ÿ qi;j †: …48†
ox i‡1=2;j h oy i;j‡1=2 h

For what concerns the boundary node, let us consider for instance the evaluation of ÿDp in the node
i ‡ 1; j, next to the vertical edge. Since the value pi‡2;j is unavailable, it has to be provided by the homo-
T
geneous Neumann condition (where n ˆ ‰1; 0Š ):
…rh p†i‡3=2;j  n ˆ pi‡2;j ÿ pi‡1;j ˆ 0: …49†

We can therefore adopt de®nition (46) with the condition (49) and obtain
1 1
…Dh p†i‡1;j ˆ …ÿpi‡1;j ‡ pi;j † ‡ 2 …pi‡1;j‡1 ÿ 2pi‡1;j ‡ pi‡1;jÿ1 †: …50†
h 2 h
Similar de®nitions can be easily obtained on the other sides of oX.
For what concerns the application of the discrete divergence operator on the boundary nodes, since it is
obtained on vectors vanishing on the boundary, we simply get
1  …1† …1†
 1
…2† …2†

…rh  v†i‡1;j ˆ vi‡3=2;j ÿ vi‡1=2;j ‡ vi‡1;j‡1=2 ÿ vi‡1;jÿ1=2
h h
1 …1†
 1
…2† …2†

ˆ ÿ vi‡1=2;j ‡ vi‡1;j‡1=2 ÿ vi‡1;jÿ1=2 ; …51†
h h
together with similar relations on the other boundary nodes.
A. Quarteroni et al. / Comput. Methods Appl. Mech. Engrg. 188 (2000) 505±526 517

On the basis of (47) and (48), the successive application of the discrete gradient and discrete divergence
operators turns out to be exactly the discretization of the Laplace operator, given by (46). Indeed, we have
    !     !
1 oh p oh p 1 oh p oh p
…rh  …rh p††i;j ˆ ÿ ‡ ÿ
h ox i‡1=2;j ox iÿ1=2;j h oy i;j‡1=2 oy i;jÿ1=2
1 1
ˆ …pi‡1;j ÿ 2pi;j ‡ piÿ1;j † ‡ 2 …pi;j‡1 ÿ 2pi;j ‡ pi;jÿ1 † ˆ …Dh p†i;j : …52†
h2 h
Similarly, it is readily veri®ed that the successive application of (48) and (51) leads exactly to (50). This
means that the matrix corresponding to Dh is exactly given by DDT . 

From Proposition 5.1 we can therefore conclude that in the case of a ®nite di€erence discretization on
the MAC staggered grid of Fig. 1, the analogy between the LU block factorization (15), (24) and the
Chorin±Temam (or Van Kan) scheme is not only formal, but rigorous.

Proposition 5.2. Assume that the space discretization is obtained through a pseudo-spectral collocation
method. The discretization of the operator ÿD associated to Neumann bounary conditions corresponds exactly
to the matrix DM ÿ1 G, where D is the space discretization of (the opposite of ) the divergence operator, op-
erating on vectors having null normal component on the boundary and G ˆ DT is the discretization of the
gradient operator.

Proof. Suppose that the discrete velocity components belong to PN0 …x† (i.e. polynomials with degree N and
vanishing on the boundary), while the discrete pressure belongs to PN ÿ2 …x†. This choice ensures the full-
®llment of the LBB inf±sup condition (see e.g. [3,25]). The basis functions for uh are denoted by fuj gjˆ1;...;Nu ;
while for the pressure by fwj gjˆ1;...;Np : Moreover, suppose that the scalar products are computed via
Legendre Gauss±Lobatto integration formulas, according to the generalized Galerkin approach: such
scalar products will be denoted by …; †N . The weights of the Gauss±Lobatto formula will be denoted by wj ,
and the nodes by Pj . The discrete Laplace operator for the pressure is given by the matrix A…sp† such that
A…sp†i;j ˆ …rwi ; rwj †N (with associated homogeneous Neumann boundary conditions). Now, denoting by dij
the Kronecker symbol, we have
ÿ  X
Mij ˆ ui ; uj N ˆ ui …Pk †uj …Pk †wk ˆ dik djk wk ˆ dij wj …53†
k

and
X ÿ  X X X X X
‰DM ÿ1 GŠi;j ˆ Dik M ÿ1 G kj ˆ Dik Mklÿ1 Glj ˆ Dik wÿ1
k dkl Gkj ˆ wÿ1
k Dik Gkj : …54†
k k l k l k

Since also in the generalized Galerkin approach DT ˆ G, if every component of ui is a polynomial with
degree N vanishing on oX and wj is a polynomial with degree N ÿ 2, we have
Z
Dki ˆ ÿ…r  ui ; wk †N ˆ ÿ…r  ui ; wk † ˆ …ui ; rwk † ÿ …ui  n†wk dc ˆ …ui ; rwk † ˆ …ui ; rwk †N
oX

ˆ rwk …Pi †wi ˆ Gik : …55†

(Notice, in fact, that the tangential components of the functions ui do not enter in our proof). Therefore,
(54) becomes
X X X
‰DM ÿ1 GŠi;j ˆ wÿ1
k Dik Gkj ˆ wÿ1
k Gki Gkj ˆ wÿ1 2
k rwi …Pk †rwj …Pk †wk
k k k
X
ˆ rwi …Pk †rwj …Pk †wk ˆ …rwi ; rwj †N ˆ ‰A…sp†i;j Š; …56†
k

which concludes the proof. 


518 A. Quarteroni et al. / Comput. Methods Appl. Mech. Engrg. 188 (2000) 505±526

6. Quasi-compressibility methods and inexact factorizations

As previously pointed out, projection methods pursue the reduction of the computational e€ort for the
numerical solution of the incompressible Navier±Stokes equations through a suitable splitting of the
original problem. Other methods aim at stabilizing the solution of system (2) through the introduction of a
``small'' perturbation in the mass conservation law. For a suitable choice of the perturbation term, the LBB
condition can be ``circumvented'' (see [17]). This means that, for instance, the same linear elements can be
adopted both for the velocity and pressure unknowns, without the onrise of spurious pressure modes (see
e.g. [25]).
Typically, the perturbed continuity equation may take one of the following forms (see [26] for the
associated denominations):
op
ru‡ ˆ 0; p…t ˆ 0† ˆ p0 …artificial compressibility method†;
ot
r  u ‡ p ˆ 0 …penalty method†; …57†
r  u ÿ Dp ˆ 0; rp  njC ˆ 0 …Petrov±Galerkin method†:

Having relaxed the incompressibility constraint, these methods are usually addressed as quasi-com-
pressibility (or pseudo-compressibility) methods. The perturbation parameter e must be suf®ciently large to
have a signi®cant regularizing effect, but as small as possible to minimizing the perturbations on the in-
compressibility equation. However, usually the solution of the Navier±Stokes equations via a quasi-com-
pressibility approach involves the solution of linear systems in the form (17), being Q associated to the
speci®c form of the perturbation term. Classical quasi-compressibility methods as the ones speci®ed by (57)
are solved in an unsplit way. In other words, the original system (2) is ®rstly perturbed in the form (17), and
then the perturbed system (17) is solved without reductions to smaller problems. Conversely, in Section 6.1,
we introduce a method which has been originally set up in the framework of inexact factorizations, so that
it is obtained as the successive solution of smaller problems, but it can be regarded also as a quasi-com-
pressibility method. Indeed, it really corresponds to solve a problem in the form (17), where Q has an
(apparently) new form.

6.1. The Yosida scheme

Let us consider the approximation (15) of the matrix A, assuming

H1 ˆ DtM ÿ1 ; H2 ˆ C ÿ1 : …58†

This choice corresponds to the following scheme:


e n‡1 ˆ bn‡1 …intermediate velocity computation†;
CU 1

ÿ DtDM ÿ1 GPn‡1 ˆ bn‡1 ÿ D Ue n‡1 …pressure computation†; …59†


2
e n‡1 ÿ GPn‡1 …end-of-step velocity computation†:
CUn‡1 ˆ C U

As previously pointed out, the matrix corresponding to this approximation turns out to be
    
C 0 I C ÿ1 G C G
AY ˆ ˆ ; …60†
D ÿ DtDM ÿ1 G 0 I D Q

being

Q ˆ ÿD…DtM ÿ1 ÿ C ÿ1 †G; …61†

which is actually a matrix in the form (17). However, di€erently from the other quasi-compressibility
schemes, this scheme should be able to couple the regularizing e€ect of the perturbing term with the
framework of inexact factorizations involving a sequence of smaller problems.
A. Quarteroni et al. / Comput. Methods Appl. Mech. Engrg. 188 (2000) 505±526 519

The perturbation term (61) has a new form in the context of pseudo-compressibility methods. However,
suppose to deal with a Stokes problem (with homogeneous Dirichlet boundary conditions), so that the
matrix C has the form
1
Cˆ M ‡K
Dt
being K the stiffness matrix obtained as discretization of the Laplace operator on the velocity vector.
Furthermore, suppose to adopt a space discretization involving the identity matrix as mass matrix M. In
this case, the perturbing term Q becomes
ÿ1
Q ˆ ÿDtD…I ÿ …I ‡ DtK† †G ˆ ÿDt2 DYG;
where
1  
Y ˆ I ÿ …I ‡ DtK†ÿ1 : …62†
Dt
In Section 5 we considered the matrix product ÿDG as a formal analog of the Laplace operator r  …r†.
In a similar way, we could investigate the formal analog of the matrix Q. In particular, denoting with K the
Laplace operator (with associated Dirichlet boundary conditions), matrix Y given by (62) can be formally
related to the operator
1 ÿ1

YDt ˆ I ÿ …I ‡ DtK† …63†
Dt
which is commonly known as Yosida regularization of the Laplace operator K (see [5, ch. 7]). Therefore,
the formal analog of the matrix Q is given by the pseudo-differential operator

Q ˆ Dt2 r  …YDt r†:


This circumstance explains the name given to the method.
In Section 7 we will illustrate some numerical results obtained by the Yosida method, comparing its
performances with the ones of other methods based both on the inexact factorizations and the quasi-
compressibility approaches. For the analysis of this method, see [24]. Here, we limit ourselves to remark
that it is possible to prove that the end-of-step velocity of the Yosida scheme in a Dirichlet problem satisfy
completely the boundary conditions (not only the normal component as in the case of the Chorin±Temam
scheme: see Remark 5.2).

Remark 6.1. The incremental formulation of the Yosida scheme is immediately deduced as for the algebraic
projection method, by solving the steps:
e n‡1 ˆ bn‡1 ÿ GPn ;
CU 1
e n‡1 ;
ÿ DtDM ÿ1 GdPn‡1 ˆ b2n‡1 ÿ D U …64†
e n‡1 ÿ GdPn‡1 :
CUn‡1 ˆ C U

Remark 6.2 (Incompatible finite elements). The Yosida method that we have introduced is well defined
provided that the matrix DM ÿ1 G in (60) (or (64)) is non-singular. This property requires the finite elements
spaces for velocity and pressure to satisfy the compatibility LBB condition. If this would not be the case, then
the lower triangular matrix of the factorization (60) could be modified as follows
  
C 0 I C ÿ1 G
AY ˆ ; …65†
D ÿ Dt…DM ÿ1 G ‡ Mp † 0 I
R
where Mp is the pressure mass-matrix (i.e. Mp ˆ ‰mp;ij Š ˆ ‰ X wi wj dxŠ† and  > 0 is a convenient parameter.
The analysis of this stabilized scheme will make the object of a future note.
520 A. Quarteroni et al. / Comput. Methods Appl. Mech. Engrg. 188 (2000) 505±526

7. Numerical results

We consider two di€erent test cases: the Kim and Moin problem, for which an exact solution is available
and the problem of a ¯ow past a circular cylinder.

7.1. Kim and Moin problem

As is well known, in order to make well posed the Navier±Stokes problem, m scalar conditions have to
be prescribed on oX. As an instance, we could suppose that oX is split into two parts, CD (Dirichlet
boundary) and CN (Neumann boundary), where we require that
u ˆ g; x 2 CD ; s ˆ ÿpn ‡ mru  n ˆ h; x 2 CN ; …66†
where g and h are given for all t > 0, n is the outward unit normal vector and s is the normal component of
the stress tensor. We suppose that meas…CD † is always positive; in the case of a pure Dirichlet problem, we
have that CD coincides with the whole boundary oX. This is the case of the following Kim and Moin
problem (see [20]). Indeed, let us consider the square domain X ˆ …0:25; 1:25†  …0:5; 1:5† and the Navier±
Stokes problem with the Dirichlet conditions illustrated in Fig. 2, whose exact solution is:
2
u1 ˆ ÿ cos…2px† sin…2py† eÿ8p mt ;
2
u2 ˆ sin…2px† cos…2py† eÿ8p mt ; …67†
ÿ16p2 mt
p ˆ ÿ 0:25…cos…4px† ‡ cos…4py†† e :
There u1 and u2 denote the velocity components along x and y, respectively, p is the pressure and m
(which we set equal to 0.01) is the kinematic viscosity.
In this test case, we compare the performances of four time advancing methods:
1. The incremental Chorin±Temam projection method (33), (35), (36), which we will denote by CTD
(Chorin±Temam di€erential); in particular, we prescribe the ``spurious'' condition …35†2 for the pressure
problem, according to the Helmholtz decomposition principle.
2. The algebraic counterpart of the Chorin±Temam incremental method (see 39) denoted as CTA
(Chorin±Temam algebraic).
3. A quasi-compressibility method (QC) where the mass conservation equation is substituted by
r  u ˆ DtDp
(see (57) and [4]) and condition …20†2 has been associated to the Laplace operator of the pressure.
4. The incremental Yosida scheme (64), denoted by YOS.

Fig. 2. Boundary conditions for the Kim and Moin problem.


A. Quarteroni et al. / Comput. Methods Appl. Mech. Engrg. 188 (2000) 505±526 521

The spatial discretization is based on the ®nite element method, using P1 isoP2 elements for the velocity
unknowns and P1 (continuous) elements for the pressure (see e.g. [25]). The time discretization makes use of
the backward Euler formula with a semi-implicit linearization of the convective term (see (6)). The linear
systems associated to the solution of the problem have been solved by the BiCGStab method (see [28]) with
a tolerance on the residual set to 10ÿ8 . The problem has been solved on three grids, with decreasing step size
(h1 ˆ 1=16; h2 ˆ 1=32; h3 ˆ 1=64) and with di€erent time step size (from Dt ˆ 2ÿ3 up to Dt ˆ 2ÿ11 ). We
measured the error between the computed solution uh ; ph and the exact one u; p given by (67) according to
the following norms:

Z …1=2†
kek ˆ kuh ÿ ukL1 …L2 …X†† ˆ max j uh ÿ u j2 dx ;
0<t 6 T X

Z T Z …1=2†
2
kgk ˆ kph ÿ pkL2 …L2 …X†† ˆ j ph ÿ p j dx dt :
0 X

The above integrals are computed through a gaussian high order integration formula.
Fig. 3 shows the error behaviour when Dt decreases for the four adopted methods. In these simulations,
the space discretization error was small enough with respect to the time discretization error (h ˆ 1=64).
From these pictures it is possible to deduce that the error reduces linearly with Dt in both norms. Moreover,
splitting methods (CTD, CTA, YOS) perform better than QC. More precisely, CTD and CTA behave
similarly, while the YOS method halves the errors provided by CTA and CTD. The behaviour of YOS,
CTA and CTD methods when h is decreased is illustrated in Fig. 4 (Dt ˆ 1=1024). As expected, working on
the same grid the three methods give practically the same (space discretization) errors, and the error
reduction is quadratic with respect to h.
The results provided by CTD and CTA are considerably di€erent when considering a mixed Dirichlet/
Neumann problem. Working on the same grids, we have considered the same domain of Fig. 2, with the
solution (67), but this time Neumann conditions are prescribed on the vertical right boundary. The CTD
method for a mixed problem has been implemented according to the algorithm proposed by Gresho and
Sani in [15, p. 735]. Fig. 5 illustrates the velocity and pressure errors of CTD, CTA and YOS methods for
h ˆ 1=32 when Dt decreases. It is evident that the spurious pressure boundary conditions that must be
prescribed on the Neumann boundary for the CTD method a€ect signi®cantly the accuracy of the
solution.

Fig. 3. Velocity error (left) and pressure error (right) when the time step size decrease for the four adopted methods, for a ®xed spatial
grid (uniform, with h ˆ 1=64); the dotted line is the reference line for the error decrease rate.
522 A. Quarteroni et al. / Comput. Methods Appl. Mech. Engrg. 188 (2000) 505±526

Fig. 4. Velocity error (left) and pressure error (right) when the space step size decrease (Dt ˆ 1=512) for CTD, CTA and YOS methods.

Fig. 5. Velocity error (left) and pressure error (right) when the time step size decrease for CTD, CTA and YOS methods, for a ®xed
spatial grid (h ˆ 1=32) on the mixed Dirichlet/Neumann, Kim and Moin test case.

It is also interesting to observe the extent at which the divergence free constraint is satis®ed by the four
methods. Figs. 6±8 show the time evolution of maxx2X jr  uh j; for the particular choice h ˆ 1=64 and
Dt ˆ 1=1024. Observe however that similar conclusions can be drawn by observing the results obtained on
di€erent space and time step size. The CTA-solution satis®es exactly the mass conservation (bearing in
mind that the residual for the linear system solver is set equal to 10ÿ8 ± Fig. 7), while the CTD solution
satis®es that constraint only for the limit case h ˆ 0; however, the violation of the divergence-free constraint
is signi®cant only limitedly to the very ®rst time levels (Fig. 6, right).
Both QC and YOS methods violate the incompressibility equation, being pseudo-compressibility
methods (see Figs. 6 and 8). However, it is interesting to observe the (quite unexpected) circumstance that
also YOS-solution, like the CTD one, after a certain number of time levels (independently of Dt), shows a
strong reduction of the computed divergence.
All these results suggest that the methods based on inexact factorization, i.e. the methods that adopt a
splitting of the fully discrete problem, are more accurate than those based on the splitting of the di€erential
problem. In particular, the Yosida method (YOS) presented here performs better than the algebraic version
of the Chorin±Temam method (CTA), with an excellent ful®llment of the incompressibility constraint. The
theoretical analysis of this method, with a proof of the linear behaviour of the error, is carried out in [24].
A. Quarteroni et al. / Comput. Methods Appl. Mech. Engrg. 188 (2000) 505±526 523

Fig. 6. Evolution in time of the maximum of jr  uh j for the solution computed by the QC (left) and CTD (right) methods.

Fig. 7. Evolution in time of the maximum of jr  uh j for the solution computed by the CTA method.

Fig. 8. Evolution in time of the maximum of jr  uh j for the solution computed by the YOS method. On the right a detail of the
diagram from t ˆ 0:0625 up to t ˆ 1.
524 A. Quarteroni et al. / Comput. Methods Appl. Mech. Engrg. 188 (2000) 505±526

7.2. Flow past a circular cylinder

In the present section we consider a mixed Dirichlet/Neumann problem, corresponding to having


meas…CN † > 0 in (66). In the case of mixed problems, the issue of which kind of condition should be
prescribed on the pressure problem (20) (or 35) is quite controversial (see however [14]). Instead, in the
algebraic splitting approach there is no need to give more boundary conditions than those originally
prescribed for the Navier±Stokes problem. In the sequel, we illustrate the results obtained with method
YOS on the mixed problem of a ¯ow past a cylinder, proposed by Simo and Armero in [27] and illustrated
in Fig. 9. We have focused our analysis on results obtained using the YOS methods as they are by far better
than those obtained with the other three methods CTD, QC and CTA.
The space discretization is again P1 isoP2 for velocity and P1 for the pressure on an unstructured mesh
(see Fig. 9). The time step used is Dt ˆ 0:05, on the time interval, (0, 150]. The initial conditions (not
compatible with the boundary conditions) prescribe a null velocity ®eld on the whole boundary.
For this problem, the lift coecient CL and drag coecient CD are de®ned respectively by
Z 2p Z 2p
CL ˆ s2 d/ and CD ˆ s1 d/;
0 0

where the integrals are carried out on the cylinder and s1 and s2 are the x- and y-components of the normal
stresses (see (66)); the Strouhal number is de®ned as
D
St ˆ ;
j u1 j T

being D ˆ 1 the diameter of the cylinder, ju1 j ˆ 1 the far ®eld velocity and T the period of oscillations,
which on the basis of experimental data (see [27]) is equal to 6, so that we have St ˆ 1=6.

Fig. 9. Flow past a cylinder problem (see [27]): geometry, boundary conditions, parameters and grid.
A. Quarteroni et al. / Comput. Methods Appl. Mech. Engrg. 188 (2000) 505±526 525

Our numerical results are in good agreement with this value of St and with the numerical results pro-
vided by Simo and Armero, as illustrated in Fig. 10, for what concerns the lift and drag coecients, and in
Figs. 11 and 12 for what concerns velocity ®eld, pressure contours and streamlines on the semiperiod from
t ˆ 120 to t ˆ 123.

Fig. 10. Flow past a circular cylinder. Evolution of the lift (left) and drag (right) coecients.

Fig. 11. Flow past a circular cylinder. Details of the periodic solution on a half time period. Velocity ®eld (above) and contour pressure
(below) lines at t ˆ 120 (left) and t ˆ 123 (right) are reported.

Fig. 12. Flow past a circular cylinder. Details of the periodic solution on a half time period. Streamlines at t ˆ 120 (left) and t ˆ 123
(right) are reported.
526 A. Quarteroni et al. / Comput. Methods Appl. Mech. Engrg. 188 (2000) 505±526

References

[1] J. Akin, Finite Elements for Analysis and Design, Academic Press, New York, 1986.
[2] F. Brezzi, M. Fortin, Mixed and Hybrid Finite Element Methods, Springer-Verlag Series in Computational Mathematics 15, 1991.
[3] C. Bernardi, Y. Maday, Approximations Spectrales de Problemes aux Limites Elliptiques, Springer, Paris, 1992.
[4] F. Brezzi, J. Pitkaranta, On the stabilization of ®nite element approximation of the Stokes equations, Ecient solutions of Elliptic
Systems, Vieweg, Braunschweig, 1984, pp. 11±19.
[5] H. Brezis, Analyse Fonctionnelle, Masson, Paris, 1983.
[6] R. Codina, J. Blasco, A ®nite element formulation for the stokes problem allowing equal velocity, Comp. Meth. Appl. Mech.
Engrg. 143 (1997) 373±391.
[7] A. Chorin, Numerical solution of the Navier±Stokes equations, Math. Comput. 22 (1968) 745±762.
[8] C. Canuto, M. Hussaini, A. Quarteroni, T. Zang, Spectral Methods in Fluid Dynamics, Springer, New York, 1988.
[9] W. Couzy, Spectral element discretization of the unsteady Navier±Stokes equations and its iterative solution on parallel
computers, Ph.D. Thesis, EPFL, Lausanne 1995.
[10] J.K. Dukowicz, A.S. Dvinsky, Approximate factorization as a high order splitting for the implicit incompressible ¯ow equations,
J. Comp. Phys. 102 (1992) 336±347.
[11] M. Fortin, Finite element methods and Navier±Stokes equations, Acta Numerica (1993) 239±284.
[12] J. Guermond, L. Quartapelle, Analyse de convergence d'une approximation par elements ®nis et methode de projection pour les
equations de Navier±Stokes instationnaires, LIMSI Reports 95-14, 1995.
[13] V. Girault, P. Raviart, Finite element methods for Navier±Stokes equations. Theory and Algorithms, Springer-Verlag Series in
Computational Mathematics 5, 1991.
[14] P. Gresho, Some current CFD issues relevant to the incompressible Navier±Stokes equations, Comp. Meth. Appl. Mech. Engrg.
87 (1991) 201±252.
[15] P.M. Gresho, R.L. Sani, Incompressible Flow and the Finite Element Method, Wiley, New York, 1998.
[16] J. Guermond, Some implementations of projection methods for Navier±Stokes equations, MZAN 30 (5) (1996) 637±667.
[17] T. Hughes, L.P. Franca, M. Balestra, A new ®nite element formulation for computational ¯uid dynamics: V. Circumventing the
Babuska Brezzi condition, Comp. Meth. Appl. Mech. Engrg. 59 (1986) 85±99.
[18] D. Hilbert, An ecient Navier±Stokes solver and its applications to ¯uid ¯ow in elastic tubes, Colloquia Societatis Janos Bolyai 50
(1987) 423±431.
[19] G. Karniadakis, M. Israeli, S. Orszag, High-order splitting methods for the incompressible Navier±Stokes equations, J. Comp.
Phys. 97 (1991) 414±443.
[20] J. Kim, P. Moin, Application of a fractional step method to incompressible Navier±Stokes equations, J. Comp. Phys. 59 (1985)
308±323.
[21] B. Perot, An analysis of the fractional step method, J. Comp. Phys. 108 (1993) 51±58.
[22] A. Prohl, Projection and Quasi-Compressibility Methods for Solving the Incompressible Navier±Stokes Equations, Advances in
Numerical Mathematics, B.G. Teubner, Stuttgartt, 1997.
[23] A. Pinelli, A. Vacca, A. Quarteroni, A spectral multi-domain method for the numerical simulation of turbulent ¯ows, J. Comp.
Phys. (1997).
[24] A. Quarteroni, F. Saleri, A. Veneziani, The Yosida method for the unsteady Navier±Stokes equations, J. Math. Pure Appl. 78
(1999) 473±503.
[25] A. Quarteroni, A. Valli, Numerical Approximation of Partial Di€erential Equations, Springer-Verlag Series in Computational
Mathematics 23, 1994.
[26] R. Rannacher, On Chorin's projection methods for the incompressible Navier±Stokes equations, Navier±Stokes Equations:
Theory and Numerical Analysis, Lecture Notes in Mathematics (1530), Springer, Berlin, 1991, pp. 167±183.
[27] J. Simo, F. Armero, Unconditional stability and long-term behaviour of transient algorithms for the incompressible Navier±
Stokes and Euler equations, Comp. Meth. Appl. Mech. Engrg. 111 (1994) 111±154.
[28] Y. Saad, Iterative Methods for Sparse Linear Systems, ITP, PWS, Boston, 1996.
[29] J. Shen, On error estimates of projection methods for Navier±Stokes equations: ®rst order schemes, SIAM J. Numer. Anal. 29 (1)
(1992) 57±77.
[30] J. Shen, On error estimates of projection methods for Navier±Stokes equations: second order schemes, Math. Comp. 65 (215)
(1997) 1039±1065.
[31] R. Temam, Sur l'approximation de la solution des equations de Navier±Stokes par la methode de pas fractionnaires (II), Arch.
Rat. Mech. Anal. 33 (1969) 377±385.
[32] S. Turek, On discrete projection methods for the incompressible Navier±Stokes equations: an algorithmical approach, Comp.
Meth. Appl. Mech. Engrg. 143 (1997) 271±288.
[33] J. Van-Kan, A second order accurate pressure-correction scheme for viscous incompressible ¯ow, SIAM J. Sci. Stat. Comp. 7
(1986) 870±891.
[34] F. Warlow, F. Welch, Numerical calculation of time-dependent viscous incompressible ¯ow of ¯uid with free surface, Phys. Fluids
8 (1965) 2182±2189.

You might also like