Solution of the steady Navier–Stokes equations

by a Newton FEM

J. Canton and M. Tugnoli

Dipartimento di Scienze e Tecnologie Aerospaziali, Politecnico di Milano, Via La Masa
34, 20156 Milano, Italy; jacopo.canton / matteo.tugnoli
@mail.polimi.it

Abstract

This work describes a Finite Element Newton Method for the solution of the stationary
Navier–Stokes equations for two-dimensional incompressible flows. We start from the weak
variational formulation of the problem and adopt an unequal order interpolation P1 -P2 for
pressure and velocity. Rather general boundary conditions are considered. The Newton
method for the nonlinear system of coupled equations is written in a particularly transparent
incremental form and the Jacobian linear system is solved by means of a direct algorithm
(MUMPS). The results of some numerical tests are provided to demonstrate the correctness
and capability of the method.

Key words: Stationary Navier–Stokes equations, Newton method for incompressible flows
PACS:

Preprint submitted to Elsevier Science 14 October 2015

.

2
Contents

1 Introduction 4

2 Governing equations and boundary conditions 5

3 Weak formulation and completion of the boundary conditions 6

3.1 From strong equations to weak formulation 6

3.2 The weak formulation 7

3.3 Completing the boundary conditions 8

3.4 Compatibility condition and desingularization 8

4 Newton method 9

4.1 Linearized Navier–Stokes equations 10

4.2 Incremental Newton method for quadratically nonlinear equations 11

4.3 Incremental Newton method for the Navier–Stokes equations 13

5 Continuation 15

5.1 Simple continuation 16

5.2 Continuation under varying boundary conditions 17

6 Mixed finite element discretization 19

7 Direct solution of the large sparse system 19

8 Numerical comparison 20

9 Conclusions 21

References 21

3
1 Introduction

We develop a classical Newton method for the approximate solution of the steady-
state Navier–Stokes equations with the aim of calculating very accurate solutions of
a base incompressible viscous flow. In fact, in dynamical system theory a base (i.e.,
time-independent) flow can represent a fixed point in the phase space, to be used for
investigating the linear and nonlinear stability of the steady current around its neigh-
bourhood.

The Newton iterative method is more suitable than a time marching technique because of
its efficiency: only a few iterations are sufficient to reach convergence to a very accurate
steady solution, provided the initial guess is chosen not too far from the solution, while
a time stepping technique may achieve stationarity only asymptotically at the end of a
long process, whose convergence is slower at higher Reynolds numbers and may become
impossible near a Hopf bifurcation.

Other authors have already obtained solutions to the steady Navier–Stokes equations
by means of the Newton method, see, for instance, [2]. In the present work, special
attention is devoted to the use of the incremental form of the Newton iteration, which
fully exploits the quadratic nature of the nonlinearity of the Navier–Stokes system. An
algorithm of the utmost simplicity is obtained which is expected to be optimal both in the
imposition of the boundary conditions and with respect to the cancellation of numerical
errors.

The content of the paper is organized as follows: in the next section, the Navier–Stokes
governing equations are presented together with their boundary conditions for the veloc-
ity; in the subsequent section, the mathematical problem is recast in a weak variational
form, discovering the complete set of boundary conditions for pressure and velocity.
Section 4 provides a description of the Newton method and gives the detailed proof of
its incremental form in the particular case of a quadratic nonlinearity. The adaptation of
the Newton solver to a restart option as well as to a continuation strategy is described
in section 5. In section 6 the spatial discretization of the equations is described and the
appropriate finite element spaces for approximating the flow fields u and p are intro-
duced. A few considerations on the sparsity of the linear algebraic system of equations
coupling velocity and pressure unknowns are given. Section 7 is dedicated to the direct
solution of the Jacobian linearized system by means of a direct method (MUMPS) and
includes the symbolic-numerical factorization, as well as an efficient way of evaluating
the Jacobian matrix within the Newton iteration to reduce the numerical cost of the
algorithm. The work ends in section 8 with a comparison of the computed numerical
solution with an experimentally determined flow field.

4
2 Governing equations and boundary conditions

This work is concerned with the determination of the solution to the nonlinear differ-
ential system of the stationary Navier–Stokes equations for a fluid of uniform density.
Assuming that the nonlinear term is expressed in the usual convective form, the equations
governing the steady state, written in a fully dimensionless form, are

(
Re (u ·∇)u − ∇ 2 u + ∇p = f ,
(2.1)
∇· u = 0,

where the pressure has been rescaled by multiplying it by the Reynolds number Re .
The right-hand side f = f (r) is a known force field. The problem is set on a finite
domain Ω whose boundary is divided into three parts, ∂Ω = Γ1 ∪ Γ2 ∪ Γ3 , to allow the
assignement of different boundary conditions on each Γi . The strong statement of the
problem is: Find (u(r), p(r)), such that



Re (u ·∇)u − ∇ 2 u + ∇p = f ,

∇· u = 0,



u|Γ1 = b1 , (2.2)

n̂ · u|Γ2 = n̂ · b2 ,




n̂×u = n̂×b ,
| Γ3 3

where n̂ is the outward pointing normal. The domain Ω can be two- or three-dimensional.
In the boundary values, b1 represents the prescribed velocity defined only on Γ1 , n̂ · b2
represents actually the scalar datum bn,2 defined only on Γ2 , and n̂×b3 is the tangential
velocity defined only on Γ3 . The first boundary condition assigns the velocity vector on
Γ1 and thus is a full Dirichlet condition for the vector unknown.

For 3D problems, the boundary conditions on Γ2 and Γ3 are not sufficient to supplement
the system of equations. In fact, the second condition assigns only the normal compo-
nent of velocity, while the two tangential components remain unspecified, and the third
condition fixes the two tangential components, while the normal component remains
unspecified. Thus on Γ2 two scalar boundary conditions are laking and on Γ3 only one
scalar condition is laking.

With reference to 2D problems, the second and the third conditions in (2.2) fix only
a scalar quantity, and therefore these boundary conditions are similarly not sufficient
to supplement the equation system. In this case, a single scalar boundary conditions
is laking, since the tangential and normal component of velocity remains unspecified
respectively on Γ2 and Γ3 .

5
3 Weak formulation and completion of the boundary conditions

3.1 From strong equations to weak formulation

To obtain the weak variational formulation of problem (2.2) using the Galerkin method,
we multiply the first equation by a vector test function v and the second by a scalar
one q, whose functional spaces and smoothness properties will be discussed below. By
integrating over the domain Ω, we obtain:
 Z Z Z Z
2
Re
 v ·(u ·∇)u − v ·∇ u + v ·∇p = v· f,
Z Ω Ω Ω Ω (3.1)
 q∇· u = 0.

Ω

As usual, the purpose of the weak formulation is to lighten the smoothness required by
the solution. The integration by parts of the term with ∇p gives:
Z Z Z Z Z
v ·∇p = − (∇· v) p + ∇· (v p) = − p ∇· v + p n̂ · v (3.2)
Ω Ω Ω Ω ∂Ω

thanks to the derivative of the product of two functions and the divergence theorem. The
integration of the Laplacian term requires a little bit more attention as it involves the
product of two vector functions. The following vector identity

∇ 2 u = −∇×∇×u + ∇(∇· u) (3.3)

will be used to deal with the two contributions of the Laplacian of a vector field. The
integration by parts af these two terms is made possible by reminding the two following
identities for the derivative of a product

∇· ( A×B) = (∇× A) · B − A · ∇×B, (3.4a)

∇· ( A f ) = (∇· A) f + A · ∇f. (3.4b)

The application of the first identity with A → v and B → ∇×u and of the second
equation with A → v and f → ∇· u yields, respectively,

∇· (v×∇×u) = (∇×v) · (∇×u) − v ·∇×∇×u, (3.5a)

∇· (v∇· u) = (∇· v)(∇· u) + v · ∇(∇· u). (3.5b)

The Laplacian of equation (3.3), scalarly multiplied by v and integrated over Ω, be-
comes: Z Z Z
2
v·∇ u = − v · ∇×∇×u + v ·∇(∇· u).
Ω Ω Ω

6
Now, by applying equations (3.5b) and (3.5a) for the derivatives of products to the terms
on the RHS and by reordering the terms, we obtain:
Z Z Z
2
v · ∇ u = − (∇×v) · (∇×u) − (∇· v)(∇· u)
Ω Ω Ω
Z Z
+ ∇· (v×∇×u) + ∇· (v∇· u).
Ω Ω

The divergence theorem applied to the two last integrals gives:

Z Z Z
2
v · ∇ u = − (∇×v) · (∇×u) − (∇· v)(∇· u)
Ω Ω Ω
Z Z (3.6)
+ n̂×v · ∇×u + n̂ · v ∇· u,
∂Ω ∂Ω

where the commutative property of the mixed (scalar-vector) product a×b · c = a · b×c
has been exploited to transform n̂ · v×∇×u into n̂×v · ∇×u in the first boundary in-
tegral. The form of the boundary conditions in problem (2.2) indicates that the test
functions v must fulfill the following homogeneous boundary conditions:

 v|Γ = 0,
 1


n̂ · v|Γ2 = 0, (3.7)



n̂×v| = 0,
Γ3

These boundary conditions complete the necessary steps to deduce the weak formulation.

3.2 The weak formulation

The weak formulation can now be obtained by substituting equations (3.6) and (3.2)
into (3.1). Furthermore, the boundary integrals of (3.6) simplify since, by virtue of the
boundary conditions (3.7) on v, n̂×v vanishes on Γ1 ∪ Γ3 and n̂ · v vanishes on Γ1 ∪ Γ2 ,
to give
 Z Z Z


 Re v ·(u ·∇)u + [(∇×v) · (∇×u) + (∇· v)(∇· u)] − (∇· v) p


 Ω Z Ω Z Z Ω
= v·f + n̂×v · ∇×u − ( p − ∇· u) n̂ · v,

 Z Ω Γ2 Γ3


 q∇· u = 0.

Ω

The boundary integrals involve derivatives of the velocity field and the pressure itself. It
is necessary to include additional boundary conditions of derivative type, which will be
found to be the vector generalization of the Neumann condition supplementing a scalar
elliptic equation.

7
3.3 Completing the boundary conditions

The two boundary integrals in the previous weak equation need to have an assigned value.
Therefore, we have to assume that two new quantities n̂×∇×u and ( p − ∇· u) on Γ2
and Γ3 , must be prescribed. We will denote the prescribed boundary values by n̂×c2
and q3 , respectively. This leads to the following completion of the boundary conditions
on Γ2 and Γ3 : 
 u| = b1 ,
 Γ1


n̂ · u|Γ2 = n̂ · b2 , n̂×∇×u|Γ2 = n̂×c2 , (3.8)



n̂×u = n̂×b , ( p − ∇· u)| = q .
| Γ3 3 Γ3 3

The weak formulation is now made complete and reads: Find u ∈ H 1 (Ω) satisfying
the boundary conditions

u|Γ1 = b1 , n̂ · u|Γ2 = n̂ · b2 , n̂×u|Γ3 = n̂×b3 , (3.9)

and p ∈ L 2 (Ω) such that, ∀v ∈ H 1 (Ω) satisfying the homogeneous boundary condi-
tions
v|Γ1 = 0, n̂ · v|Γ2 = 0, n̂×v|Γ3 = 0, (3.10)
and ∀q ∈ L 2 (Ω), the following sistem of weak equations
 Z Z Z


 Re v ·(u ·∇)u + [(∇×v) · (∇×u) + (∇· v)(∇· u)] − (∇· v) p


 Ω Ω Z Z Z Ω
= v·f + n̂×v · c2 − n̂ · v q3 , (3.11)

 Z Ω Γ2 Γ3


 q∇· u = 0,

Ω

is fulfilled. Note that nothing needs to be done to the nonlinear term thanks to Sobolev’s
immersion theorem that states that, as u, v ∈ H 1 (Ω), the term v · (u ·∇)u ∈ L 2 (Ω).
For additional details see [5].

3.4 Compatibility condition and desingularization

Depending on the type of boundary conditions, it may occur that the velocity boundary
values are required to satisfy a condition of compatibility in order that the problem can
be solved. In fact, when the normal component of velocity is prescribed on the entire
boundary, the total net flux of mass passing through the boundary Γ must be zero, not to
violate the incompressibility constraint. This circumstance may be met in two specific
different situations, which are both characterized by the fact that Γ3 = ∅.

The first possibility is when the vector velocity is prescribed on the entire boundary, so

8
that Γ1 = Γ and the only existing boundary condition would read most simply

u|Γ = b. (3.12)

In this case the compatibility condition on the datum b is

I
n̂ · b = 0. (3.13)
Γ

The second possibility is when Γ = Γ1 ∪ Γ2 and the velocity conditions on the two
portions Γ1 and Γ2 of the boundary could be written as

u|Γ1 = b1 and n̂ · u|Γ2 = bn,2 , (3.14)

to emphasize that the quantity specified on Γ2 is a scalar value bn,2 . In this case the
compatibility condition of the boundary data would read as
Z Z
n̂ · b1 + bn,2 = 0. (3.15)
Γ1 Γ2

We insist that either of the two compatibility conditions will be present only provided
that Γ3 = ∅.

Of course, in both the considered cases the pressure will be defined only up to an
undetermined additive constant and the operator associated with the problem will be
singular. It is therefore necessary to deal with such a singularity properly. One possible
method consists in fixing (arbitrarily) the value of pressure at an arbitray point in the
domani Ω. Thus we introduce the idea of desingularizing the problam by adding the
following pressure boundary condition

p(r ⋆ ) = p⋆ (3.16)

at a single arbitrary point r ⋆ ∈ Ω, with p⋆ denoting a completely arbitrary value, for

instance zero.

4 Newton method 1

To formulate a Newton method for solving the weak variational problem (3.10)-(3.11)
numerically it would be necessary to derive first a discretized version of the problem.
However, instead of introducing a finite element approximation for this purpose, we
prefer to analyze the Newton iterative method in the context of the differential problem

1
The first part of the present section has been borrowed from [1] and adapted to our need.

9
(2.1) in the original strong form. We draw the attention on a quite general nonlinear
problem expressed in the following form:

N (x) = 0, (4.1)

where N (x) is a vector of ν nonlinear equations, namely, a set of ν nonlinear functions

of the ν unknowns (x1 , . . . , xν ) = x. The Newton algorithm is derived by expanding
N (x) in a Taylor series in the neighbourhood of an approximate solution xn , at the n-th
step, and by truncating the series to first order:

N (x) ≈ N (xn ) + J(xn )(x − xn ), (4.2)

where J(x) represents the Jacobian matrix of the system computed in x, in indicial
notation Ji,i ′ (x) = ∂ Ni (x1 , . . . , xν )/∂ xi ′ , with 1 ≤ i, i ′ ≤ ν. By substituting the linear
approximation (4.2) in the original nonlinear problem (4.1), a linear system of equations
is obtained at each step whose solution xn+1 represents an updated solution of the iterative
procedure, namely
J(xn ) xn+1 = J(xn ) xn − N (xn ). (4.3)
The Newton iteration, after introducing the unknown variation δxn+1 ≡ xn+1 − xn , can
be written alternatively and equivalently in the so-called incremental form

J(xn )(xn+1 − xn ) = −N (xn ), (4.4)

to be followed by the update assignment xn+1 ← xn + δxn+1 .

4.1 Linearized Navier–Stokes equations

For the Navier–Stokes equations (2.1), the nonlinear system (4.1) assumes the form
   
2
Re (u ·∇)u − ∇ u + ∇p   f 
N (u, p) =   −   = 0. (4.5)
∇· u 0

The Jacobian matrix operator can be written as

 
u 2
∂N (u, p) Re L − ∇ ∇
J (u) = = , (4.6)
∂(u, p)
∇· 0

where the linear operator L u has been introduced, acting on the space of differentiable
vector functions and producing a vector field, according to the definition

L u w = (u ·∇)w + (w ·∇)u (4.7)

10
which depends linearly on the vector variable u, assumed to be a known velocity field u =
u(r). For the sake of the readability, the intuitive "slot" notation . . . can be introduced:

L u . . . = (u ·∇) . . . + (. . . ·∇)u (4.8)

which highlights the position held by the element that the operator is acting on.

For the Navier–Stokes equations, the linear (4.3) assumes the form
   
un+1 un
J (un )   = J (un )   − N (un , pn ). (4.9)
pn+1 pn

Thus, by exploiting the definition of N (u, p) in (4.5) and that of the linearizing operator
L u given by (4.8), the Navier–Stokes nonlinear problem (2.1) is approximated at each
iteration by the linearized system for the unknown pair (un+1 , pn+1 )
(
Re L un − ∇ 2 un+1 + ∇pn+1 = Re (un ·∇)un + f ,


(4.10)
∇· un+1 = 0.

Here, all the boundary conditions have been omitted, the complete discussion about
them being already established in 2 and in 3.3. The focus is now, in fact, more on the
development of the algorithm logical procedure whereas further specific applications of
the Newton algorithm will be introduced later.

The set of equations (4.10) defines a linear incompressible viscous problem which
contains the sum of the two linearized convection terms, (un ·∇)un+1 and (un+1 ·∇)un .

If the starting guess of the iterative procedure is sufficiently near to the solution, the
convergence rate of the Newton algorithm is quadratic, and thus very fast with respect
to other algorithms, such as the fixed point method.

4.2 Incremental Newton method for quadratically nonlinear equations

The incremental Newton method as expressed in (4.4) is now formulated for a particular
class of nonlinear equations containing only quadratic nonlinearities. The nonlinear
system of equations is written formally as follows:

N (x) = 0, (4.11)

where
N (x) ≡ Lx + Q(x, x) + F. (4.12)
Here L is a simple linear operator (namely a matrix of linear operators acting on the
components of the vector unknown x), Q(x, y) is a bilinear quadratic form, possibly not

11
symmetric, and F is a forcing term. The system of the Navier–Stokes equations falls in
this particular class. The bilinearity of the quadratic form Q(x, y) is expressed by the
relations:

Q(x1 + x2 , y) = Q(x1 , y) + Q(x2 , y), (4.13a)

Q(x, y1 + y2 ) = Q(x, y1 ) + Q(x, y2 ). (4.13b)

The operatorial Jacobian matrix, evaluated at x, is simply:

J (x) ≡ L + Q(x, ...) + Q(..., x). (4.14)

Applying the Newton method in standard (i.e. nonincremental) form

J (xn ) xn+1 = J (xn ) xn − N (xn ) (4.15)

to the special quadratically nonlinear operator (4.12) we have

J (xn ) xn+1 = Lxn + 2Q(xn , xn ) − [Lxn + Q(xn , xn ) + F], (4.16)

which simplifies to
J (xn ) xn+1 = Q(xn , xn ) − F. (4.17)
On the other side, according to the incremental form of the Newton method (4.4), the
system for the approximate solution at (n + 1)-th step assumes the form:

J (xn ) δxn+1 = −Lxn − Q(xn , xn ) − F, (4.18)

where δxn+1 = xn+1 −xn represents the solution increment at this step. We now consider
the non-incremental form (4.17) applied at the n-th step and use also the definition (4.14)
of the Jacobian J (x), to give

J (xn−1 ) xn = Q(xn−1 , xn−1 ) − F = Lxn + Q(xn−1 , xn ) + Q(xn , xn−1 ). (4.19)

Isolating the quantity −Lxn − F from this equation and substituting it into the rhs of
the incremental equation (4.18), we find

J (xn ) δxn+1 = −Q(xn , xn ) + Q(xn−1 , xn ) + Q(xn , xn−1 ) − Q(xn−1 , xn−1 )

(4.20)
= −Q(δxn , xn ) + Q(δxn , xn−1 )

where the bilinearity (4.13a) of Q with respect to its first argument has been exploited
and the variation of the unknown δxn = xn − xn−1 at the previous iteration has been
introduced. Exploiting finally the bilinearity (4.13b) of Q with respect to its second
argument gives
J (xn ) δxn+1 = −Q(δxn , δxn ). (4.21)
This is the linear system at the (n + 1)-th iteration, for n = 1, 2, . . . , of the incremental
Newton method for the linear-quadratic nonlinear problem. Remarkably enough, the
right hand side of equation (4.21) depends only on the quadratic term Q, evaluated only

12
at the increment δxn : this makes it fully transparent the quadratic convergence rate of
the Newton iteration.

However, the iteration (4.21) is not defined at the first Newton iteration (i.e., n = 0) since
δx0 is not defined. Thus, the first Newton iteration of the incremental method requires
a special consideration. We will assume that the initial guess x0 is chosen to be the
solution to the linear problem, namely, that x0 satisfies Lx0 = −F (the Stokes problem
in the Navier–Stokes solver). In this particular case, we will adopt the incremental form
also at the first iteration, which reads

J (x0 ) δx1 = −N (x0 ) = −Lx0 − Q(x0 , x0 ) − F. (4.22)

By exploiting the definition Lx0 = −F of the initial guess, the first Newton iteration in
incremental form simplifies to

J (x0 ) δx1 = −Q(x0 , x0 ) (4.23)

and can be dealt with by the general equation

J (xn ) δxn+1 = −Q(δxn , δxn ), (4.24)

which becomes valid for any n = 0, 1, . . . , provided we introduce the simple definition
(
x0 n=0
δxn ≡ (4.25)
xn − xn−1 n≥1

This incremental formulation of the Newton method for the incompressible Navier–
Stokes equations has been suggested by Franco Auteri. The derivation of the iteration
formula (4.21) just obtained can be dubbed LQ-Theorem, since it is a consequence of
the Linear & Quadratic character of the Navier–Stokes equations. The iteration (4.21)
as well as its generalization (4.24)–(4.25) to include the first iterate in incremental form
is particularly elegant and compact, and its application to the steady Navier–Stokes
equations is straightforward.

4.3 Incremental Newton method for the Navier–Stokes equations

Now the compact formulation described in section 4.2 is applied to the system of the
steady Navier–Stokes equations. The operator of the linear part of the Navier–Stokes
system (2.1) is:
 
2
−∇ ∇
L= , (4.26)
∇· 0

13
while the quadratic operatorial function corresponding to the nonlinear part reads
 
Re (u ·∇)w
Q(u, w) =  . (4.27)
 
0

Finally, the forcing term for the system is simply:

 
 f
F =  . (4.28)
0

Let us now introduce the increment unknown vector δx for the system:
 
δu
δx =   . (4.29)
δp

Then, by substituting the explicit expressions of J (u) from definition (4.6) and of
Q(u, w) from (4.27) into the final iteration equation (4.24), the incremental form of the
Newton method for the Navier–Stokes problem is obtained. It is easily found to consist
in the following linear boundary value problem, for any n ≥ 1,
(
Re L un − ∇ 2 δun+1 + ∇δpn+1 = −Re (δun ·∇)δun ,


(4.30)
∇· (δun+1 ) = 0.

A simple stopping criteria for the iterations can be a norm of the difference between
two subsequent iterations, which will stop the iterations when falling under a prescribed
tolerance.

A particular attention must be paid to the initial guess of the iterative method, along
with the necessary boundary conditions. As the initial guess (u0 , p0 ), we assume the
solution to the Stokes problem, namely:

2
−∇ u0 + ∇p0 = f ,

∇· u0 = 0, (4.31)

NonH.B.C.


The nonhomogeneous boundary conditions to be imposed in the Stokes problem are

the same as in the Navier–Stokes one. Specifically, the imposition of the velocity com-
ponents (essential conditions) is requested by the set up of the problem and must be
imposed regardless of the equations used to model the fluid motion. On the other hand,
the additional conditions applied to the pressure and the vorticity normal component on
Γ2 and Γ3 (natural conditions) are generated from the integration by parts of the pres-
sure divergence and viscous Laplacian terms, which are equally present in the Stokes
equation.

14
Once the Stokes system has been solved, the solution satisfies all of the boundary condi-
tons. The Newton iteration is then recast for the increment of the solution, which implies
the imposition of fully homogeneous boundary conditions.

We can now summarize the algorithm steps:

(1) Initially, the Stokes system is solved with the complete nonhomogeneous boundary
conditions:


 −∇ 2 u0 + ∇p0 = f ,

∇· u0 = 0,



u0 |Γ1 = b1 , (4.32)

n̂ · u0 |Γ2 = n̂ · b2 , n̂×∇×u0 |Γ2 = n̂×c2 ,





n̂×u | = n̂×b , ( p − ∇· u )| = q .
0 Γ3 3 0 0 Γ3 3

(2) The increment δun is defined as follows

(
u0 n=0
δun = (4.33)
un − un−1 n≥1

and similarly for the δpn .

(3) For any n ≥ 0, solve the linear incompressible problem, supplemented by fully
homogeneous boundary conditions:

Re L un − ∇ 2 δun+1 + ∇δpn+1 = −Re (δun ·∇)δun ,





∇· (δun+1 ) = 0,



δun+1 |Γ1 = 0, (4.34)

n̂ · δun+1 |Γ2 = 0, n̂×∇×(δun+1 )|Γ2 = 0,





n̂×δu | = 0, [δp
n+1 Γ3 n+1 − ∇· (δun+1 )]|Γ3 = 0.

(4) Update (
un+1 ← un + δun+1
(4.35)
pn+1 ← pn + δpn+1
The iterative method is stopped when the relative error of the incremental solution
is less than a prescribed small threshold value.

5 Continuation

For solutions at high Reynolds numbers, the Newton method may fail when the Stokes
flow is too far from the nonlinear solution. In these cases, it is necessary to resort to
a continuation method. This means to compute the solution corresponding to a given

15
Reynolds number Re in a step-wise manner, i.e., by computing different intermediate
solutions for smaller values of the Reynolds number.

5.1 Simple continuation

Suppose a solution for an intermediate Reynolds number Re∗ < Re has been found:
namley we have calculated the fields u∗ and p∗ , solutions to the steady Navier–Stokes
equations for the Reynolds number Re∗ . Then, one applies the Newton method a second
time, for the nonlinear Navier–Stokes equation corresponding to the final, originally
intended to, Reynolds number Re > Re∗ , starting from the previous solution u∗ and p∗
taken as a second initial guess.

The Newton algorithm used in the second stage of the process is analogous to that
considered before, with only a modification in the first step n = 0. In fact, while the first
system, for the (first) incremental unknows δu1 and δp1 , is still written in incremental
form, the new initial guess is taken as the solution of the nonlinear steady problem and
the latter is characterized by the previous, different, Reynolds number Re∗ .

The first incremental unknown (δu1 , δp1 ) is defined by the equation

 
δu1
J (u∗ )   = −N (u∗ , p∗ ), (5.1)
δp1

which implies, for the momentum equation,

∗
Re L u − ∇ 2 δu1 + ∇δp1 = − Re (u∗ ·∇)u∗ − ∇ 2 u∗ + ∇p ∗ − f .

 
(5.2)

But (u∗ , p∗ ) satisfies the nonlinear steady-state equation Re∗ (u∗ ·∇)u∗ −∇ 2 u∗ +∇p∗ −
f = 0 so that −∇ 2 u∗ + ∇p∗ − f = −Re∗ (u∗ ·∇)u∗ . Substituting this result into the
right-hand side of (5.2) yields
∗
Re L u − ∇ 2 δu1 + ∇δp1 = −(Re − Re∗ )(u∗ ·∇)u∗ ,


(5.3)

where the occurrence of the Reynolds number variation (Re − Re∗ ) can be noticed. As
a coonsequence, the complete problem for n = 0 reads
∗

Re L u − ∇ 2 δu1 + ∇δp1 = −(Re − Re∗ )(u∗ ·∇)u∗ ,





∇· (δu1 ) = 0,



δu1 |Γ1 = 0, (5.4)

n̂ · δu1 |Γ2 = 0, n̂×∇×(δu1 )|Γ2 = 0,





n̂×δu | = 0,
1 Γ3 [δp1 − ∇· (δu1 )]|Γ3 = 0.

This is the only modification of the Newton algorithm necessary to implement the
continuation strategy.

16
5.2 Continuation under varying boundary conditions

As stated in section 1, the purpose of the Newton algorithm described so far, with the
possibility of a restart option, is the computation of a base flow, whose stability fea-
tures are under investigation. Interesting behaviours may occur at quite high Reynolds’
numbers or for different values of other parameters. As an example, considering the
flow past a translating and rotating circular cylinder, the Reynolds number, defined by
Re = U D/ν, is accompanied by a second dimensionless number α which represents the
ratio of the rotational velocity of the cylinder surface Ω D/2 to the translational velocity
U , to give α = Ω D/(2U ).

In this case, the continuation strategy is to compute a solution to the nonlinear equations
for lower/different values of the parameters and to use it as the initial guess to start a
subsequent iteration in order to determine the solution for the final values of the parame-
ters. In general, the subsequent application of Newton method involves the change of the
value of both Re and α. For instance, after a nonlinear solution (u∗ , p∗ ) for the values
Re∗ and α ∗ has been determined, one aims at calculating the solution for the different
values Re > Re∗ and α 6 = α ∗ . Since the new value of the rotation velocity parameter
α enters the problem only at the level of the boundary conditions, the Newton method
described so far is unable to compute the final solution and needs to be modified for
accommodating different nonhomogeneous boundary values during the continutation
process.

First, we have to consider the boundary values bi∗ , i = 1, 2, 3, c∗2 and q3∗ , associated
with the intermediate values Re∗ and α ∗ of the continuation parameter. Then, for the
subsequent (final) values Re > Re∗ and α 6 = α ∗ , different boundary values bi , c2
and q3 are to be imposed. As far as the boundary values are concerned, for the flow
around a translating and rotating circular cylinder, only the rotational velocity of the
cylindrical surface may change during the continuation. More precisely, the Dirichlet
condition enforcing no-slip tangential velocity condition τ̂ · u∗ |Γcyl = α ∗ is replaced by
τ̂ · u|Γcyl = α, with α 6 = α ∗ .

The modified Newton algorithm for the continuation with the possibility of varying the
boundary values reads as follows.

(1) compute the solution (u∗0 , p0∗ ) to the Stokes problem (4.32) enforcing the (inter-
mediate) nonhomogeneous boundary values bi∗ , i = 1, 2, 3, c∗2 and q3∗ :


−∇ 2 u∗0 + ∇p0∗ = f ,

∗
∇· u0 = 0,



u∗0 |Γ1 = b∗1 , (5.5)
n̂ · u∗ |Γ = n̂ · b∗ , n̂×∇×u∗0 |Γ2 = n̂×c∗2 ,



 0 2 2
n̂×u∗ | = n̂×b∗ , ( p∗ − ∇· u∗ )| = q .

0 Γ3 3 0 0 Γ3 3

17
 This solution is indicated by (u∗0 , p0∗ ) to remind that it depends on the value α ∗ ,
which enters the problem through the tangential boundary condition n̂×u∗0 |Γ3 =
n̂×b∗3 , which for the two-dimensional cylinder problem reduces to τ̂ · u∗ |Γcyl = α ∗ .
(2) use (u∗0 , p0∗ ) as initial guess for Newton’s algorithm (steps 2, 3 and 4 in section 4.3)
with the intermediate values Re∗ (α ∗ is not involved due to the homogeneity of the
velocity boundary conditions) to compute the intermediate solution (u∗ , p∗ );
(3) modify the parameters to their final values (Re, α);
(4) Start a new Newton iteration with initial guess provided by (u∗ , p∗ ). The first
Newton iteration, with n = 0 from (u∗ , p∗ ) to (u1 , p1 ) is performed in the not
incremental form by imposing the final nonhomogeneous boundary values bi , c2
and q3 . The first-step unknown x1 = (u1 , p1 ) is solution to the equation J (x ∗ ) x1 =
J (x ∗ ) x ∗ − N (x ∗ ). In particular, the first vector equation reads
∗
Re L u − ∇ 2 u1 + ∇p1


∗
= Re L u − ∇ 2 u∗ + ∇p∗ − Re (u∗ ·∇)u∗ + ∇ 2 u∗ − ∇p∗ + f (5.6)

= 2Re (u∗ ·∇)u∗ − ∇ 2 u∗ − Re (u∗ ·∇)u∗ + ∇ 2 u∗ + f

which reduces to
∗
Re L u − ∇ 2 u1 + ∇p1 = Re (u∗ ·∇)u∗ + f .


(5.7)

As a consequence, the complete problem for the first iteration would read
∗

Re L u − ∇ 2 u1 + ∇p1 = Re (u∗ ·∇)u∗ + f ,





∇· u1 = 0,



u1 |Γ1 = b1 , (5.8)

n̂ · u1 |Γ2 = n̂ · b2 , n̂×∇×u1 |Γ2 = n̂×c2 ,





n̂×u | = n̂×b ,
1 Γ3 3 [ p1 − ∇· u1 ]|Γ3 = q3 .

The solution of the first iteration depends on the final values of the parameter Re
present in the momentum equation and of α which enters the boundary condition
for the tangential component of velocity.
(5) Define the increment of the first step as

δu1 ≡ u1 − u∗ and δp1 ≡ p1 − p ∗ . (5.9)

(6) For n ≥ 1, use the incremental Newton iteration with the final parameter val-
ues (Re, α) but imposing fully homogeneous boundary conditions, as in (4.34),
repeated here for the sake of completeness,

Re L un − ∇ 2 δun+1 + ∇δpn+1 = −Re (δun ·∇)δun ,





∇· (δun+1 ) = 0,



δun+1 |Γ1 = 0, (5.10)

n̂ · δun+1 |Γ = 0,

n̂×∇×(δun+1 )|Γ2 = 0,

 2

n̂×δu | = 0, [δp
n+1 Γ3 n+1 − ∇· (δun+1 )]|Γ3 = 0,

18
 with the standard update
(
un+1 ← un + δun+1
(5.11)
pn+1 ← pn + δpn+1

The iterative method is stopped when the relative error of the incremental variable
(δun+1 , δpn+1 ) is less than a prescribed small threshold value. Then, the solution
(un+1 , pn+1 ) provides a sufficiently accurate representation of the exact solution
(u, p) corresponding to the parameter values Re and α.
(7) Eventually, consider other new values (Renext , α next ) of the continuation parameters
and repeat the process from step 3.

This modified version of the Newton algorithm admits variable boundary conditions in
a continuation framework and can be used to increase the values of the parameters in
a step-wise manner, guaranteeing convergence, while the previous Newton algorithm,
with the initial guess based on the Stokes solution, is limited to enforce once and forall
a fixed set of boundary values.

6 Mixed finite element discretization

7 Direct solution of the large sparse system

19
8 Numerical comparison

In this section we present some numerical results obtained by means of a program that
implements the Newton method described in this report. We compare the numerical
results with the experimental visualizations presented in [3]. The program has been
developed by Luigi Quartapelle and Franco Auteri together with the first author using
the Fortran 90 language. It implements the Newton iterative method to solve the nonlinear
system of the stationary incompressible Navier–Stokes equations, using a finite elements
method on a given triangulation. The program solves the above mentioned nonlinear
system over a generic two-dimensional grid of triangular Lagrangian elements, allowing
for rather general boundary conditions. The only constraint on the boundary conditions
is that the borders where the user wants to impose natural boundary conditions, i.e. Γ2
and Γ3 , have to be parallel to the Cartesian axes to avoid mixing the velocity components
u x and u y . The program has been used to solve a classical experimental and numerical
setup: a cylinder in a stationary, uniform velocity field in the far field region. A series of
different grids, more and less refined, and a series of different Reynolds numbers have
been tested, all with excellent results.

For imposing the boundary conditions, the boundary is split in four different parts.

• On the cylinder surface we impose zero velocity, namely b = 0. Then, the external
boundary is split in three parts:
• an inflow portion Γ1 , where the uniform velocity b1 = U x̂ is imposed, namely
u|Γ1 = U x̂.
• a lateral portion Γ2 , which is parallel to the x axis, where only the normal veloc-
ity component is imposed, equal to zero, together with zero vorticity specification,
namely: u y |Γ2 = 0 (i.e. n̂ · b2 = 0), with the boundary integral over Γ2 being zero
(n̂×c2 = 0).
• an outflow portion Γ3 parallel to the y axis, where only the tangential velocity com-
ponent is imposed, equal to zero, together with a uniform (arbitrary, for instance zero)
constant value of pressure, namely: u y |Γ3 = 0 (i.e. n̂×b3 = 0), with the boundary
integral over Γ3 being zero (q3 = 0).

As an example, we presented here the results of a calculation on a grid made of ≈ 150 000
elements, ≈ 75 000 linear nodes and ≈ 300 000 parabolic nodes. The steady flow
at a Reynolds number Re = 41. The solution, thanks to the second order Newton
method converged after just 6 iterations with a relative tolerance of 10−12 defined by
relative error = kδxk/kxk. The streamlines obtained from the numerical solution are
depicted in fig. 8. Other calculations have been performed with success with up to
800 000 elements which corresponds to approximately 1 600 000 unknowns. The same
numerical results are presented over an experimental visualization made at the same
Reynolds number in fig. 8. The fair agreement with the experimental results is evident,
the only slight difference is caused by a small lack of symmetry in the visualization.
More comparison, with measures of the wake length and angle of attachment at different

20
 Fig. 8.1. Streamlines at Re = 41

Fig. 8.2. Streamlines superimposed on experimental flow visualization at Re = 41

Reynolds numbers, could have been made with [4] but are left, for the moment being,
for a subsequent investigation.

9 Conclusions

References

[1] F. AUTERI AND L. Q UARTAPELLE, Newton–Krylov method with Stokes preconditioning

for steady incompressible flow in a sphere. In prepapration, 2012.
[2] S YAMSUDHUHA AND D. J. S YLVESTER, Efficient solution of the steady-state Navier–
Stokes equations using a multigrid preconditioned Newton–Krylov method. Int. J. Numer.
Meths. Fluids, 43, 1407–1427, 2003.
[3] M. VAN DYKE, An Album of Fluid Motion. The Parabolic Press, 1982.
[4] S. W. C HURCHILL, Viscous Flows. The Practical Use of Theory. Butterworths Publishers,
1988.
[5] A. Q UARTERONI, Modellistica numerica per problemi differenziali, Springer, Seconda
edizione, 2003.

21