You are on page 1of 185

High-resolution unstructured nite-volume

methods for conservation laws


A. Guardone
1
L. Quartapelle

1
Politecnico di Milano, Dipartimento di Ingegneria Aerospaziale, Via La Masa 34,
20158 Milano, Italy. e-mail: guardone@aero.polimi.it
.
Abstract
The node-pairs representation of Galerkin and Finite-Volume methods for solving
conservation laws over unstructured multidimensional grids is described. This ap-
proach allows for the factorization of the grid-dependent metric quantities in the
computation of the nite-element integrals, thus leading to very efcient high-
resolution methods for simulating compressible ows in domains of arbitrary
shape. A procedure for imposing the boundary conditions in nonlinear hyper-
bolic systems within the assumed discretization framework is also presented. The
node-pair representation of second-order spatial derivative terms required to im-
plement LaxWendroff and TaylorGalerkin schemes is also described. The high-
resolution Galerkin and Finite-Volume formulations are rst developed for a scalar
conservation law and then extended to nonlinear systems of hyperbolic equations.
Finally, an original TaylorGalerkin scheme for the mass conservation law over an
arbitrary velocity eld is derived in both standard and node-pair formulations.
.
Contents
1 Introduction 1
1.1 Scalar conservation law . . . . . . . . . . . . . . . . . . . . . . . 4
2 Finite element method 6
2.1 Scalar conservation law . . . . . . . . . . . . . . . . . . . . . . . 6
2.2 Boundary conditions . . . . . . . . . . . . . . . . . . . . . . . . 9
2.3 Spatially discrete form of the equations . . . . . . . . . . . . . . 13
2.4 Approximation of a diffusion term . . . . . . . . . . . . . . . . . 13
2.4.1 Basic property of diagonal stiffness elements . . . . . . . 15
3 Flux reinterpolation 17
3.1 Scalar conservation law . . . . . . . . . . . . . . . . . . . . . . . 17
3.2 Treatment of boundary conditions . . . . . . . . . . . . . . . . . 19
3.3 Spatially discrete form of the equations . . . . . . . . . . . . . . 21
4 Node-pair representation 23
4.1 Node-pairs and metric vectors of interaction . . . . . . . . . . . . 23
4.2 Proof of the split of domain and boundary contributions . . . . . . 25
4.2.1 Proof of the domain integral indentity . . . . . . . . . . . 25
4.2.2 Final transformation of the boundary term . . . . . . . . . 28
4.3 Node-pair form of the discrete equations . . . . . . . . . . . . . . 30
4.4 Treatment of boundary conditions . . . . . . . . . . . . . . . . . 31
4.4.1 Duplication and augmentation of boundary nodes . . . . . 31
4.4.2 Duplication of boundary edges in 3D problems . . . . . . 33
4.5 Diffusion term in node-pair form . . . . . . . . . . . . . . . . . . 35
5 Finite-Volume method on nonstructured meshes 38
5.1 Finite-Volume spatial discretization . . . . . . . . . . . . . . . . 38
5.2 Upwind Finite-Volume scheme . . . . . . . . . . . . . . . . . . . 40
5.3 The bridge between nite volumes and nite elements . . . . . . . 44
1
6 Taylor-series-based integration schemes 47
6.1 Nonlinear scalar conservation law . . . . . . . . . . . . . . . . . 47
6.2 The fully discrete form of the equations . . . . . . . . . . . . . . 49
6.3 TaylorGalerkin scheme in node-pair form . . . . . . . . . . . . . 49
6.3.1 The bulk TG term . . . . . . . . . . . . . . . . . . . . . . 50
6.3.2 The boundary TG term . . . . . . . . . . . . . . . . . . . 52
6.3.3 Node-pair TaylorGalerkin scheme . . . . . . . . . . . . 54
6.4 Linear ux function . . . . . . . . . . . . . . . . . . . . . . . . . 54
7 High-resolution scheme in node-pair form 59
7.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59
7.2 Extended node-pairs . . . . . . . . . . . . . . . . . . . . . . . . 61
7.3 High-resolution scheme for steady solutions . . . . . . . . . . . . 62
7.4 High-resolution scheme for unsteady solutions . . . . . . . . . . . 64
7.5 Limiters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66
8 Thermodynamics of gases 68
8.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
8.2 Denition of the gas properties . . . . . . . . . . . . . . . . . . . 69
8.3 Polytropic and nonpolytropic behaviour . . . . . . . . . . . . . . 74
8.4 Pressure function of the conservative variables . . . . . . . . . . . 78
8.5 Speed of sound . . . . . . . . . . . . . . . . . . . . . . . . . . . 80
8.6 Some particular gas models . . . . . . . . . . . . . . . . . . . . . 81
8.6.1 Polytropic ideal gas . . . . . . . . . . . . . . . . . . . . . 81
8.6.2 Nonpolytropic ideal gas . . . . . . . . . . . . . . . . . . 82
8.6.3 Polytropic van der Waals gas . . . . . . . . . . . . . . . . 82
8.6.4 Nonpolytropic van der Waals gas . . . . . . . . . . . . . 83
9 Euler equations of gasdynamics 85
9.1 Conservation laws of gasdynamics . . . . . . . . . . . . . . . . . 85
9.2 Euler equations in one dimension . . . . . . . . . . . . . . . . . . 86
9.2.1 Conservation variables . . . . . . . . . . . . . . . . . . . 86
9.2.2 Quasilinear form . . . . . . . . . . . . . . . . . . . . . . 87
9.2.3 Eigenstructure . . . . . . . . . . . . . . . . . . . . . . . 88
9.3 Euler equations in two dimensions . . . . . . . . . . . . . . . . . 89
10 Roe linearization of gasdynamic equations 92
10.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92
10.2 Principles of Roe linearization . . . . . . . . . . . . . . . . . . . 93
10.2.1 Denition of Roe linearization . . . . . . . . . . . . . . . 93
10.2.2 General solution of Roe linearization . . . . . . . . . . . 95
2
10.2.3 Solution in Jacobian form . . . . . . . . . . . . . . . . . 95
10.3 Linearization of Euler equations . . . . . . . . . . . . . . . . . . 96
10.4 Solution of Euler linearization . . . . . . . . . . . . . . . . . . . 97
10.5 Polytropic van der Waals uid . . . . . . . . . . . . . . . . . . . 102
10.6 Ideal gases . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 106
10.6.1 Flux functions homogeneous of degree one . . . . . . . . 107
10.6.2 Nonpolytropic ideal gas . . . . . . . . . . . . . . . . . . 108
10.6.3 Polytropic ideal gas . . . . . . . . . . . . . . . . . . . . . 110
11 High-resolution schemes for systems of conservation laws 113
11.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 113
11.2 Multidimensional upwind scheme for systems . . . . . . . . . . . 114
11.3 Boundary conditions for system of conservation laws . . . . . . . 115
11.4 High-resolution scheme for steady solutions . . . . . . . . . . . . 118
11.5 A general limiter function for hyperbolic systems . . . . . . . . . 121
11.6 High-resolution scheme for unsteady solutions . . . . . . . . . . . 123
12 Boundary conditions in nonlinear hyperbolic systems 127
12.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 127
12.2 Conservative, characteristic and
physical variables . . . . . . . . . . . . . . . . . . . . . . . . . . 128
12.3 The boundary values for a scalar unknown . . . . . . . . . . . . . 130
12.4 Steps of the boundary procedure for a system . . . . . . . . . . . 131
12.5 Alternative boundary procedure . . . . . . . . . . . . . . . . . . 134
A Discrete differential operators in node-pair weak form 140
B Nomenclature 142
C Algorithms 144
D Error analysis of the TaylorGalerkin method 152
D.1 Basic third-order TG scheme . . . . . . . . . . . . . . . . . . . . 152
D.2 Two-step third-order TG scheme . . . . . . . . . . . . . . . . . . 156
D.3 Two-step fourth-order TG schemes . . . . . . . . . . . . . . . . . 162
D.4 Vector advection equation . . . . . . . . . . . . . . . . . . . . . . 166
D.5 Mass conservation equation . . . . . . . . . . . . . . . . . . . . . 170
.
If you think we are moving pictures, he said,
You ought to pay you know. Moving pictures are
not made to be looked at for nothing. Nohow!
Robert Gilmore Alice in Quantumland
.
Chapter 1
Introduction
The TaylorGalerkin scheme of Donea [3] has been proven successful for the
computation of unsteady hyperbolic problems and it has been used in a large
number of applications, ranging from multidimensional advection problems [5]
to the solution of the multidimensional shallow water equation [1]. Different
schemes have been proposed depending on the considered set of equations and on
the aimed-at accuracy, but all rest on the same machinery for the time discretization
(by means of Taylor series) and for the space discretization (by means of nite
elements). As pointed out in [5], TaylorGalerkin schemes are to be considered
as the natural extension of LaxWendroff nite-difference schemes to the nite
element framework. In subsequent elaborations of the TaylorGalerkin approach,
Selmin [22] introduced a two-step procedure for time advancing. This lead to a
newclass of schemes, in which, differently fromthe original ones, no modication
to the standard mass matrix is needed.
Although many applications of TaylorGalerkin schemes can be found in the
literature, the authors are not aware of any review work on the subject, the work of
different authors being mainly focused on specializing the TaylorGalerkin scheme
to the problem under study. Moreover, the issue of the treatment of boundary
conditions for the hyperbolic problem in the TaylorGalerkin framework, in either
a strong or a weak sense, has received almost no attention in the literature.
A well known drawback of the TaylorGalerkin scheme lies in that it is built
upon nite element spaces of piecewise linear interpolations, so that this scheme
is not suitable for the computation of discontinuous or shocked solutions, as those
encountered, for instance, in the solution of the compressible Euler equations. To
overcome these difculties, in [6] a two-step procedure that includes an articial
dissipation operator was proposed. However, it is the authors opinion that this
issue is far from being claried and deserves further attention.
The rst part of the present work is intended as a reviewof the TaylorGalerkin
scheme for conservation laws; an error analysis of the scheme is also given in an
1
appendix. For completeness, the description of the nite element and of Taylor
Galerkin schemes includes also an original procedure for the imposition of the
proper (in ow or inlet) boundary conditions in a weak form.
TaylorGalerkin schemes are presented for both a multidimensional nonlinear
conservation law governing a scalar unknown and for a linear advection equation
(with variable coefcients) over a given advection eld, possibly nonsolenoidal and
time-dependent, as it is the case of the mass conservation equation for compressible
ows. An original TaylorGalerkin scheme is presented for the mass conservation
law. This particular scheme can be applied in the solution of the Navier-Stokes
equations in those formulations where the mass conservation equation is tackled
is a separate step from the momentumand energy equations, to determine the uid
density at a given time level. The new density so calculated is then employed to
advance the momentumand the energy variables, using an explicit time integration
scheme; such an approach is taylored to deal with the incomplete parabolic charac-
ter of the compressible NavierStokes equations. Moreover, the Taylor-Galerkin
scheme for linear advection can also be applied to the solution of the advection
problem arising from one or two-equation turbulent models.
In the second part of the report, we review the node-pair representation of
the Galerkin nite element method introduced by Selmin [23]. In particular a
node-pair formulation of the TaylorGalerkin schemes is derived as a step towards
their use for the solution of Euler and Navier-Stokes equations for compressible
ows. Thanks to the factorization of all the geometry-dependent quantities in the
computation of the nite element integrals, a great improvement in the overall
efciency of the code is achieved over the standard nite-element assembling
procedures. As a consequence, the proposed node-pair-based TaylorGalerkin
schemes opens the way to the use of standard nite volume stabilization techniques,
such as upwind schemes, high-resolution schemes and articial viscosity methods.
Interestingly enough, the node-pair representation of nite element schemes and
TaylorGalerkin schemes is discovered to provide a bridge between nite volume
and LaxWendroff schemes in the context of unstructured spatial discretizations.
This report is organized as follows. In Chapter 2 the nite element approxima-
tion of a nonlinear scalar conservation lawis presented, and the discrete formulation
of inlet boundary conditions for a hyperbolic problem is described. The nite ele-
ment approximation of a diffusion termand of a second-order directional derivative
is detailed to construct TaylorGalerkin schemes for nonlinear conservation law.
In Chapter 3, an approximate technique for evaluating the nonlinear terms based on
the reinterpolation of uxes, using the same local basis functions of the unknown
variable, is introduced. This step is preliminary to the derivation, in Chapter 4,
of the node-pair representation of the discrete equations. Chapter 5 introduces
the Finite-Volume method in the context of unstructured spatial discretizations.
The node-pair formulation of the upwind scheme is described together with the
2
relationship between nite volumes and elements, as established in Selmins basic
work [22].
Chapter 6 is devoted to the analysis of schemes based on the Taylor serie ex-
pansion in the time step, such as LaxWendroff and TaylorGalerkin schemes. In
particular the evaluation of the second-order directional derivative term is detailed
to build schemes of this kind for a nonlinear conservation law, both in the stan-
dard Galerkin form and in the node-pair representation. In Chapter 7 an original
implementation of high-resolution and TaylorGalerkin schemes for discontinu-
ous solutions over unstructured meshes is presented for a scalar conservation law.
Before describing the extension of the high resolution unstructured method to
the nonlinear hyperbolic system of the gasdynamic equations, some preparatory
material is introduced.
Chapter 8 presents the thermodynamical properties of gases and describes the
gas models that will considered in the subsequent applications. The Euler equa-
tions of gasdynamics governing the motion of a compressible but invisicd uid are
recalled in Chapter 9 for both one- and multidimensional ows; the eigenstructre
of these systems of nonlinear hyperbolic equations is derived. Chapter 10 presents
the idea of local linearization introduced by Roe to replace the Riemann problemat
each interface by an approximate linear problemthat is equivalent fromthe conser-
vation viewpoint to the original nonlinear problem. In particular, the linearization
is sought for by determining an intermediate state such that the Jacobian matrix
satises Roes conservation conditions. The solution of the linearization problem
is provided for Euler equations made complete by some ideal and nonideal gas
models.
All these physical and mathematical components are needed to develop the
high-resolution schemes for solving the multidimensional Euler equations which
are detailed in Chapter 11. The last Chapter provides a detailed account of the
original procedure for satisfying the boundary conditions in the simple case of
one-dimensional hyperbolic systems, whose multidimensional counterpart was
employed in the preceding chapter.
In the appendices, the node-pair representation of differential operators in weak
form is summarized (Appendix A) and a short table of the nomenclature is given
(Appendix B). In Appendix C, an algorithm for the construction of the divergence
operator in node-pair format is detailed and, in Appendix D, the two-step Taylor
Galerkin schemes are recalled and an error analysis of these schemes is provided.
3
1.1 Scalar conservation law
We start by considering a multidimensional conservation lawfor the scalar quantity
u, written in the following divergence form
u
t
+ f (u) = 0, (1.1)
where the ux f (u) R
d
is a given vector function of the unknown u(x, t ) R,
x R
d
, t [0, T] (see gure 1.1). Here, d = 2 or d = 3.

Figure 1.1: The domain and its boundary in two spatial dimensions.
The conservation law under consideration must be supplemented by proper initial
and inlet boundary conditions, so that the complete initial boundary value problem
(IBVP) reads (e.g. [20])
_

_
u
t
+ f (u) = 0,
u(x, 0) = u
0
(x),
u
|
in
= a(s, t ), s
in
,
(1.2)
where u
0
(x) and a(s, t ) are known functions, dened their respective domain
and
in
. The coordinate s over the boundary is indicated here as a scalar
variable with reference to problems in two dimensions, while it becomes a vector
(of dimension two) in the three-dimensional case. The inow or inlet portion

in
of the domain boundary is dened as (gure 1.2)

in
def
=
_
x

n(x) a(u(x, t )) < 0


_
4
a

out

in
Figure 1.2: The (local) direction of the advection eld a with respect to the outward
normal n identies the inow and outow portions of the boundary.
with n(x) denoting the outward normal unit vector and with the advection velocity
dened by
a(u)
def
=
d f (u)
du
.
The remaining part of the boundary will be referred to as the ouow portion of
and will be denoted by
out
, so that =
in

out
. Note that the
denition of
in
is a function of the trace of the solution u(x, t ) at a given time,
so the inowboundary is in general dependent on time and a more precise notation
for it would be
t
in
, and similarly for
t
out
For a complete analysis of the well-posedness of the IBVP (1.2) including the
compatibility conditions on the data we refer to the monograph of Godlewski and
Raviart [9], chapter 5.
5
Chapter 2
Finite element method
In this chapter, the scalar conservation law (1.2) is discretized in space by means
of the standard nite element method (FEM). For later convenience, the FEM
approximation of the boundary terms is detailed, together with a procedure to
impose inlet boundary conditions in this framework.

Figure 2.1: Triangulation of the domain .


2.1 Scalar conservation law
The scalar conservation law (1.1) is recast in a weak or variational form by
applying the classical Galerkin nite element method. The weak form of the
equation is obtained (e.g. [20]) by multiplying the differential equation by test
6
functions belonging to a suitable
1
space V H
1
(), and integrating over the
domain as follows
_
,
u
t
_
+
_
, f (u)
_
= 0,
where (, ) denotes the inner product in L
2
(). To simplify the presentation, the
discussion of how the boundary conditions are imposed is postponed to Chapter
2.2. An integration by parts gives
_
,
u
t
_

_
, f (u)
_
+
_

n f (u) = 0. (2.1)
Let us nowmove to the discrete formof the equations above; if a nite dimensional
space V
h
H
1
(), of weighting functions
h
of linear or bilinear Lagrangian
type is considered (gure 2.1) , we obtain
_

i
u
t

_

i
(
i
) f (u) +
_

i
n f (u) = 0, (2.2)
where the shorthand notation

i

i
has been introduced. In writing
the boundary integral, we have taken into account that the element basis function

i
V
h
of each node i vanishes on the boundary
i
of its support
i
, but for
nodes on the domain boundary. The support
i
is the union of the sets
e
,
e
denoting the e-th nite element, which contain the node i (cf. gure 2.2 and 2.3).

i
i

i
Figure 2.2: The shape function
i
(x) in one spatial dimension and its support
i
.
1
The introduction of the Sobolev space H
1
() in the context of hyperbolic problems cannot be
justied mathematically unless the solution has a regularity typical of elliptic or parabolic (second-
order) problems.
7
i
Figure 2.3: The support
i
of the shape function
i
(x) in two spatial dimensions.
Let us assume that the solution u(x, t ) is approximated by the function u
h
(x, t )
obtained by an expansion in the same space of the weighting functions
h
(x) as
follows,
u(x, t ) u
h
(x, t ) =

kN
u
k
(t )
k
(x), (2.3)
where u
k
(t ) is the value of the approximate solution at node k and at time t (gure
2.4) and N denotes the set of all nodes of the triangulation. The weak form of the
conservation law now reads
_

i
u
h
t

_

i
(
i
) f (u
h
) +
_

i
n f (u
h
) = 0.
u
u
h
Figure 2.4: The function u
h
(x, t ) versus u(x, t ) in one spatial dimension for linear La-
grangian nite elements.
8
Substituting the expansion of u
h
into the rst term and transferring the other two
terms into the right-hand side gives

kN
i
__

ik

i

k
_
du
k
dt
=
_

i
(
i
) f (u
h
)
_

i
n f (u
h
).
By recalling the denition of the elements of the consistent mass matrix, namely,
M
i k
=
_

ik

i

k
, (2.4)
the (Bubnov) Galerkin nite element approximation of the scalar conservation
law assumes the

kN
i
M
i k
du
k
dt
=
_

i
(
i
) f (u
h
)
_

i
n f (u
h
). (2.5)
2.2 Boundary conditions
Let us recall the nonlinear conservation-lawequation endowed with inlet boundary
condition as in (1.2), namely,
_
_
_
u
t
+ f (u) = 0,
u
|
in
= a(s, t ) s
in
.
Under this inlet condition the boundary integral of (2.1) can be split in the sum of
two integrals, as follows,
_

n f (u) =
_

out
n f (u) +
_

in
n f (a), (2.6)
where
out
def
= \
in
and where the integrand function over the inlet part

in
of the domain boundary has been evaluated by taking into account the pre-
scribed boundary condition u
|
in
= a(s, t ). We remind that the partition of the
boundary in
in
and
out
depends on the advection eld a and hence, in
the nonlinear case of interest here, the partition depends on the boundary value of
the unknown u, so that the inow and outow portions of may change with
time and ought to be denoted more appropriately as

t
in
and
t
out
.
The type of boundary condition changes along the boundary and the points of
transition, where the sign of n a changes, may fall, in general, inside a boundary
9
edge. This represents the most difcult occurrence, in which case the imposition
of the boundary condition for the hyperbolic equation is particularly critical and
requires an ad hoc special treatment, see below. By contrast, when the change of
the boundary condition type from inow to ouow occurs across two consecutive
edges, the account of the boundary condition less complicated.
The evaluation of the boundary integrals (2.6) of function n f (u) over the
solution-dependent domains
in
and
out
in the context of the nite element
method can be performed as follows. Let us introduce the nite element approxi-
mation of the unknown u and of the function a dened on the boundary,
u(x, t )
|

out
u
out
h
(s, t )
def
=

kN

out
u
k
(t )

k
(s), s
out
a(s, t ) a
h
(s, t )
def
=

kN

in
a
k
(t )

k
(s), s
in
where

k
is the trace of the weight function
k
on , namely,

k
(s)
def
=
k
(x(s)), x(s) ,
and N

out
and N

in
are the boundary nodes belonging to the outow and inow
portions of , respectively. The nite element approximation of the boundary
integral (2.6) now reads
_

out

i
n f (u) +
_

in

i
n f (a)
_

out

i
n f (u
out
h
) +
_

in

i
n f (a
h
).
For completeness, we describe the numerical computation of the above integral by
means of approximate quadrature formulas. If a Gaussian numerical integration
is considered, the distinction between
in
and
out
pertains only to the Gauss
points, where the integrand is evaluated. The two contributions to the boundary
integral will be computed as follows
_

out

i
n f (u) +
_

in

i
n f (a)

gG

out

i
(x
g
) n
g
f
_
k
u
k

k
(x
g
)
_
p
g
+

gG

in

i
(x
g
) n
g
f
_
k
a
k

k
(x
g
)
_
p
g
,
where the subscript g is used to denote Gauss integration points and p
g
is the
corresponding integration weight. The two sets of Gauss integration points G

out
10
Grid node
Gauss point
a
a

G
out
G
out
G
out
G
in
G
in
G
in
Figure 2.5: Gauss points (indicated with empty circles) belonging to G

out
and G

in
.
and G

in
of
out
and
in
, respectively, are dened as follows (gure 2.5)
G

out
def
=
_
g G

n
g
a
_
k
u
k

k
(x
g
)
_
0
_
G

in
def
=
_
g G

n
g
a
_
k
u
k

k
(x
g
)
_
< 0
_
In other words, at each Gauss integration point the (interpolated) value of the
unknown is used to evaluate the advection eld, and, consequently, to determine
whether an inlet boundary condition has to be imposed or not.
The approximate integration procedure outlined above stems directly from
equation (2.6), in that the resulting algorithm retrieves rst the partitions
out
and
in
and then performs the integration. This procedure is computationally
inefcient, since it can be either time consuming (the isoparametric transformation
of the whole boundary is to be performed twice) or memory consuming (the coor-
dinate of Gauss point and the value of the shape functions are to be stored for the
whole boundary at the same time). In practice, since in the usual implementationof
the nite element method the computation of the integrals is performed by splitting
them into a sum of elemental contributions, the above procedure can be applied
on every element independently, without loss of efciency. In the following, we
will write the boundary integrals in an alternative but equivalent form that is more
11
suitable for descriptive purposes and also to deal with different formulation of the
nite element method.
First of all, we want to encompass the different denitions of the integral
over
out
and
in
within a uniform procedure. This can be accomplished by
introducing the function
u(s)
def
=
_
_
_
u if n(s) a(u(s)) 0
a if n(s) a(u(s)) < 0
(2.7)
where the binary operator | indicates the possible choice between u and a, depend-
ing on the local direction of the advection eld a(u) with respect to the outward
normal n(s), s , as shown in gure 2.6. It follows that the boundary term
u
a
u(s)
n a
Figure 2.6: The function u(s) in two spatial dimensions.
(2.6) of the weak equation can be written in the compact form
_

out
n f (u) +
_

in
n f (a) =
_

n f (u).
Coming now to the discrete representation of the unknown and of the prescribed
boundary value we have
_

n f (u)
_

i
n f (u
h
(s)). (2.8)
Thus, in this formulation, the subdivision of the boundary into the sets
in
and
out
is obtained automatically through evaluating the integral of a function
of the unknown u
h
or of the datum a
h
over the whole boundary .
12
2.3 Spatially discrete form of the equations
For completeness, we summarize here the detailed form of the spatially discrete
counterpart of the conservation law (1.1) according to the standard nite element
approach. We have

kN
i
M
i k
du
k
dt
=
_

i
(
i
) f (u
h
)
_

i
n f (u
h
(s)), (2.9)
where we have introduced the following (semi-) discrete approximation of the
unknown
u(x, t ) u
h
(x, t ) =

kN
i
u
k
(t )
k
(x).
The function u is dened as
u
h
(s)
def
=
_
_
_
u
h
if n(s) a(u
h
) 0
a
h
if n(s) a(u
h
) < 0
in which the following expansion of the boundary datum has been considered
a(s, t ) a
h
(s, t ) =

kN

i
a
k
(t )

k
(s), s
in
.
As well known [7], the nite element method described above suffers from a
deterioration of accuracy for increasing t in hyperbolic and advection problems.
In chapter 6, we shall recall the TaylorGalerkin scheme of Donea [3] which has
been devised to overcome these difculties for the particular case of transient
solutions.
2.4 Approximation of a diffusion term
As a preliminary step in the derivation of the TGscheme for the scalar conservation
law, the spatial discretization of a diffusion term is presented. Let us consider the
problem of approximating a term of the type
(u) (u
h
)
by means of the nite element method, where u is a scalar unknown, for example
the temperature or a velocity component, and is a (possibly variable) coefcient,
for example a thermal conductivity or a viscosity coefcient. Once expressed in
13
weak formby projection onto a nite dimensional space V
h
H
1
() of weighting
functions
h
V
h
of Lagrangian type, the viscous term integrated by parts reads
_

i
(u
h
) =
_

i
(
i
) u
h
+
_

i
n u
h
. (2.10)
The evaluation of the boundary term in the expression above depends on the kind
of boundary conditions specied for u, as a consequence of the elliptic nature of
the term under examination: thus Dirichlet or Neumann boundary conditions are
usually prescribed on different parts of . Suppose that these conditions are
u = a on
D
and
(u/n) n u
h
= b on
N
,
(2.11)
where
D

N
= . Then, the weighting functions are to be chosen in the
space H
1
D
(), namely, such that
i
= 0 on
D
(see gure 2.7).
Dirichlet
boundary condition
Neumann
boundary condition
x x
0
x
1
1
0
Figure 2.7: Shape functions belonging to the Sobolev space H
1
D
([x
0
, x
1
]) in one spatial
dimension. A Dirichlet boundary condition is to be imposed at x = x
0
, while a Neumann
boundary condition is assumed at x = x
1
.
As a consequence, expression (2.10) becomes
_

i
(u
h
) =

kN
i
__

ik
(
i
)
k
_
u
k
+
_

i

N

i
b. (2.12)
Moreover, if the viscosity coefcient is constant, the diffusion termcan be written
in the form
_

i
(u
h
) =

kN
i
K
i k
u
k
+
_

i

N

i
b (2.13)
14
where the elements K
i k
of the stiffness matrix K are dened as follows
K
i k
def
=
_

i
,
k
_
=
_

ik
(
i
)
k
(i = k), (2.14)
which is symmetric. Here, the usual symbol K for the stiffness matrix has been
retained, which should not be confused with the node index k or the calligraphic
letter N denoting node sets.
In conclusion, if an expansion of the Neumann boundary term into the nite
dimensional space of weighting function is considered, i.e.,
b(s, t )

kN

N
b
k
(t )

k
(s), s
N
where

k
is the trace of the k-th shape function
k
on dened in (2.2), and
N

N
is the set of the Neumann boundary nodes, we have, for a constant dissipative
coefcient ,
_

i
(U) =

kN
i
K
i k
u
k
+

k(N

i
N

N
)
_
_

ik

i

k
_
b
k
,
where the shorthand notation

i k

i

k
has been introduced.
Instead, for nonconstant , the discrete representation of the Laplacian term under
consideration reads
_

i
(u
h
) =

kN
i
K

i k
u
k
+

k(N

i
N

N
)
_
_

ik

i

k
_
b
k
, (2.15)
where we have introduced the more general stiffness matrix
K

i k
def
=
_

i
,
k
_
=
_

ik
(
i
)
k
. (2.16)
2.4.1 Fundamental property of diagonal elements
of the stiffness matrix
The diagonal elements of the stiffness matrix have an important property that
will prove very useful in the explicit matrix multiplication within the edge-based
organization of the spatially discrete data that will be introduced in the next chapter.
The diagonal element of each row of the stiffness matrix is the opposite of the sum
15
of all off-diagonal elements of the considered row, namely, the following relation
holds
K
ii
=

kN
i,=
K
i k
. (2.17)
This identity is a simple consequence of the fact that the stiffness matrix is the
discrete representation of the (negative) Laplacian for the Neumann problem and
that the solution to this problem is dened up to an arbitrary additive constant.
This arbitrariness is possible in the discrete problem only provided that the sum of
all elements of each raw of the matrix vanishes.
The detailed proof the identity follows from the local support property of the
shape functions and the basic relation

kN
e
i

k
(x) = 1, x
e
, e E,
over any nite element
e
of the triangulation. This relation implies that

i
(x) = 1

kN
e
i
, k=i

k
(x) x
e
,
and hence

i
(x) =

kN
e
i
, k=i

k
(x) x
e
.
From the denition of the diagonal element we have
K
ii
def
=
_

i
|
i
|
2
=

eN
e
i
_

e
|
i
|
2
.
Using the relation for the gradient just established, we have
K
ii
=

eN
e
i
_

e
(
i
)

kN
e
i
, k=i

k
=

eN
e
i

kN
e
i
, k=i
_

e
(
i
)
k
=

kN
i
, k=i
_

ik
(
i
)
k
=

kN
i,=
K
i k
.
16
Chapter 3
Flux reinterpolation
In the following sections, we shall introduce an approximation of the ux function
which makes the evaluation of the corresponding discrete terms much simpler and
more efcient than in the classical Galerkin formulation. Such an approximation is
based on reinterpolating the ux function f (u
h
) using the ux values calculated at
the nodes and the same basis functions employed in the expansion of the unknown.
The ux reinterpolation has been frequently considered in the solution of problems
by means of the nite element method and it is known as group-representation
method or group approximation or under other names; see, for example, [18].
The interpolation of the ux is actually an exact representation in the particular
case of a linear ux function f (u) = a u, with a constant; on the contrary, for a
given arbitrary f (u) the reinterpolation introduces numerical errors whose analysis
is beyond the scope of this work. Nevertheless, it would be worth investigating the
effects of using such an approximation especially in the case of Lagrangian nite
elements of order higher than one.
3.1 Scalar conservation law
Let us rst examine the volume integral of (2.5) obtained by the integration by
parts of the divergence f (u) of the ux function, namely,
_

i
f (u
h
) =
_

i
f
_
kN
u
k

k
_
(3.1)
which constitutes the approximation of the corresponding integral under the con-
sidered Galerkin approach. We now reinterpolate the ux by means of the same
17
i
k

i

k
Figure 3.1: The intersection
ik
of the supports
i
and
k
of nodes i and k.
basis functions
h
, i.e., we make the approximation
f (u
h
(x, t )) = f
_
kN
u
k
(t )
k
(x)
_

kN
f (u
k
(t ))
k
(x) =

kN
f
k
(t )
k
(x)
where f
k
(t )
def
= f (u
k
(t )). Thus, the spatial integral (3.1) is approximated by
_

i
f (u
h
)
i

_

i
(x)

kN
f
k
(t )
k
(x) =

kN
i
f
k
(t )
_

k
(x)
i
(x)
and also, denoting more precisely the actual integration domain (cf. gure 3.1),
_

i
f (u
h
)

kN
i
f
k

ik

i
(3.2)
In a similar way, let us nowevaluate the boundary term provided by the integration
by parts. Resorting to the same reinterpolation of the uxes, we have
_

i
n f (u
h
)
_

kN
n f (u
k
)
k
=
_

kN
i
n f (u
k
)
k
=

kN
i
f
k

ik

i

k
n
where, again, f
k
= f (u
k
). The boundary integral will be different from zero only
provided that also the node k belongs to , so that, to indicate this, the range of
18
the summation index in the previous relation is restricted appropriately as follows
_

i
n f (u
h
)

kN

i
f
k

ik

i

k
n, (3.3)
where N

represents the set of all boundary nodes.


In conclusions, once the approximation based on the reinterpolation of the ux
has been introduced, the discrete Galerkin formulation of the conservation law
reads

kN
i
M
i k
du
k
dt
=

kN
i
f
k

ik

kN

i
f
k

ik

i

k
n. (3.4)
The introduction of the ux reinterpolation has allowed us to recast the integrals
in space involving the (time-dependent) ux function as the product of the time-
dependent nodal value f
k
(t ) and xed spatial integrals involving the shape func-
tions
i
and
k
. Therefore, the computation of the integrals dependent on the mesh
geometry appearing in the discrete equation can be performed once and for all at
the beginning of the computation, in a preprocessing phase. This leads to a strong
reduction in the CPU time requirements with respect to the standard Galerkin ap-
proach (cf. equation (2.5)). On the other hand, such an approximation introduces
an integration error whose analysis is beyond the scope of this work.
3.2 Treatment of boundary conditions
To simplify the derivation of expression (3.4) we did not consider the proper bound-
ary conditions for the initial boundary value problem(1.2). Moreover, in the deriva-
tion of (3.4), and in particular in (3.3), we assumed a constant ux function over
the intersection of the two support
i
and
k
of the nodes i and k. Wheneverthe
type of the boundary condition changes on a boundary element edge, as described
in Section 2.2, this approximation is no longer valid and the variable f cannot be
factorized anymore. In the following we exclude this occurrence and discuss the
theme of the imposition of the boundary conditions assuming that they can change
type only when passing across different boundary elements, as depicted in 3.2.
In this section, we address the problemof imposing the proper inlet boundary con-
ditions, and, more generally, of imposing boundary condition of different type in
this framework. Under the ux reinterpolation already introduced, we seek for
a suitable form of the boundary term which preserves the envisaged factorization
of the geometry-dependent part from the time-dependent one. The resulting for-
mulation is obtained without introducing any approximation other than the usual
19
a
a
a
a

out

in
i k m
j
n
n
n
Figure 3.2: Over the boundary
k
of the support
k
of node k boundary conditions of
different kind are to be imposed: on
k

j
an inlet boundary condition is required,
while
k

i
is an outow region, that is, no boundary conditions are to be imposed.
group representation of the uxes in the case of linear and bilinear Lagrangian
elements, and a rst order approximation of the normal direction in the case of
elements of order p > 1.
We perform the following reinterpolation of the ux function on the boundary
on which we impose
f (s, t )

kN

f
k
(t )

k
(s), (3.5)
where
f
k
(t )
def
= f
_
u(s
k
, t )
_
. (3.6)
In the last expression, function u(s, t ) is dened as in (2.7), namely:
u(s, t )
def
=
_
_
_
u(s, t ) if n(s) a(u(s, t )) 0
a(s, t ) if n(s) a(u(s, t )) < 0
(3.7)
where u(s, t ) = u(x, t )
|
, that is the trace of the solution u on the boundary at
time t . [Note that a is the boundary value prescribed on u and is not the intensity
of the advection velocity a.]
In the reinterpolation, the dependence of f (s, t ) on the boundary coordinate s
cannot be dropped, as done for instance for the coefcient f
k
(t ) of the reinterpo-
lation in the domain, for u here depends on s. By substituting (3.3) in (2.8), we
20
have
_

i
(s) n f (s, t )

kN

i
_

i
(s)

k
(s) n(s) f
k
(t ).
In the case of linear or bilinear elements, the normal n is constant over the boundary

e
belonging to the element e, so that we can dene the outward normal of
element e as
n
e
def
= n(s) = constant s
e
(3.8)
and the following restriction of f (u(s, t )) over the e-th element as (gure 3.3)
f
e
k
(t )
def
= f (u(s
k
, t )) s
e
k
, (3.9)
which is a function of the time t only and can be factorized in the computation of
the space integrals, as
_

i
n f (u)

eE

kN

i
f
e
k
(t )
_

ik

k
n. (3.10)

out

out

in
i k m j
n a n a
f (u
m
)
f (u
i
)
f (u
k
)
f (a
k
)
f (a
j
)
Figure 3.3: Flux reinterpolation at boundaries for the situation sketched in gure 3.2.
3.3 Spatially discrete form of the equations
For completeness, let us recall here the complete form of the spatially discretized
conservation law (1.1) under the hypothesis of ux reinterpolation. We have

kN
i
M
i k
du
k
dt
=

kN
i
f
k

ik

i
+

eE

kN

i
f
e
k

_

ik

k
n, (3.11)
21
where we have dened as usual
f
k
(t )
def
= f (u
k
(t )) (3.12)
and
f
e
k
(t )
def
= f
_
u(s
k
, t )
_
, s
i

k

e
(3.13)
22
Chapter 4
Node-pair representation
In this chapter we introduce a more convenient way of writing the ux terms
of the discrete equations which has been proposed by Vittorio Selmin [23]. In
his formulation, these contributions are expressed as a summation over pairs of
interacting nodes in such a way quantities dependent on time are factorized out
from quantities dependent on the assumed spatial discretization. As a results, all
the quantities associated with geometrical features of the mesh can be evaluated
in a preprocessing phase of the computation. Moreover, the ux contributions are
found to be expressed in a quasi one-dimensional formthat permits the exploitation
of all algorithmic tools developed for the solution of one-dimensional conservation
law equations and systems, as, for example, upwind schemes.
4.1 Node-pairs and metric vectors of interaction
Let us introduce the following metric vector quantities

i k
def
=
_

ik
(
i

k

k

i
), (i = k) (4.1)
which are different from zero only if
i

k
= 0, i.e., if the nodes i and k
interacts in the nite element sense. Each couple of such interacting nodes
is called node-pair (gure 4.1). In terms of these metric vectors, the discrete
counterpart of the considered scalar conservation law can be recast in the compact
form [23]

kN
i
M
i k
du
k
dt
=

kN
i,=
f
k
+ f
i
2

i k
+

kN

i,=
f
k
f
i
2

_

ik

i

k
n

kN

i,=
f
k

ik

i

k
n
(4.2)
23
i
k
Figure 4.1: The node-pair i -k.
where the last two summations must be noticed to involve only boundary nodes.
The metric vectors
i k
s occurring in equation (4.2) above satisfy the following
fundamental properties [25]:

i k
=
ki
, (antisymmetry) (4.3a)

ii
= 0 (4.3b)

kN
i

i k
= 0, for any internal point i (4.3c)

kN
i

i k
=
_

i
n =
_

i
n (4.3d)
Properties (4.3a) and (4.3b) are important for ensuring a conservative numerical
scheme, while properties (4.3c) and (4.3d) are a necessary consequence of the
fact that a constant ux must give a zero contribution to the change in u
h
at each
internal point. The properties listed above allow to identify the quantities
i k
as metric vectors which behave similarly to the integrated normal vectors of a
control volume in the nite volume approach. From this viewpoint, properties
(4.3c) and (4.3d) guarantees that the control volume is closed. In Selmin [23], an
equivalence theorem is given which relates the node-pair formulation of the nite
element method to the standard nite volume approach over suitable constructed
control volumes on triangular meshes. The equivalence holds for the discrete
equations associated with internal nodes, but it breaks down for those associated
with boundary nodes.
24
4.2 Proof of the split of domain and
boundary contributions
Equation (4.2) is proved starting formthe discrete equation (3.4), which is repeated
here for convenience:

kN
i
M
i k
du
k
dt
=

kN
i
f
k

ik

kN

i
f
k

ik

i

k
n, (4.4)
and using the following identity

kN
i
f
k

ik

i
=

kN
i,=
f
k
+ f
i
2

i k
+

kN

i,=
f
k
f
i
2

_

i

k
n
(4.5)
which splits the domain integral into domain contributions and boundary contribu-
tions. The direct substitution of (4.5) into equation (4.4) gives (4.2). The boundary
contributions in (4.2) are recast in a different more convenient form by means of a
second identity that will be discussed below.
4.2.1 Proof of the domain integral indentity
The standard method used in nite element schemes for computing the integral
on the left hand side of the relation above is to assemble the contributions coming
from each element e in the mesh, exploiting the local support property of the shape
functions, as follows

kN
i
f
k

ik

i
=

eE
i

kN
e
f
k

i
, (4.6)
where E
i
is the set of the elements having the node i in common (bubble around
node i ) and N
e
is the set of the nodes of element e. The rst summation on the
right-hand side is limited to the elements contained in the support
i
of node i ,
which are the only ones to give a nonzero contribution to integrals containing the
function
i
. In fact, if we indicate by
e
the subset of pertaining to the e-th
element ( =

eE

e
) we have, by denition
i
=

eE
i

e
. Let us consider
the identity (gradient theorem)
_

e
(
i

k
) =
_

i

k
n (4.7)
valid for any continuous function
i

k
, which gives the following relation
_

i
=
_

k
+
_

i

k
n (4.8)
25
i k
e
1
e
2

e
1
i k

e
2
i k

i k
Figure 4.2: Elemental contributions to the metric vector
ik
.
By virtue of this identity one can deduces
_

i
=
1
2
_

i
+
1
2
_

i
=
1
2
_

i
+
1
2
_

k
+
_

i

k
n
_
=

e
i k
2
+
1
2
_

i

k
n (4.9)
where in the last equality we have introduced the elemental contributions
e
i k
to
the metric vectors
i k
, according to the denition (gure 4.2)

e
i k
def
=
_

ik

e
(
i

k

k

i
)
In terms of these elemental contributions the metric vector associated with the pair
i -k can be expressed as

i k
=

e(E
i
E
k
)

e
i k
Using (4.9), the integral (4.6) becomes

kN
i
f
k

ik

i
=

eE
i

kN
e
f
k

e
i k
2

1
2
_

i

k
n
_
(4.10)
26
On the other hand, from equation (4.9) it follows that

e
i k
= 2
_

i
+
_

i

k
n
and also, recalling (4.8),

e
i k
= 2
_

k

_

i

k
n
Summing up this relation for all nodes k belonging to element
e
and using the
elementary property

kN
e
k
(x) = 0, x
e
, we obtain

kN
e
_

e
i k
+
_

i

k
n
_
= 0
Summing up this relation for all elements belonging to E
i
and multiplying the total
by the constant vector f
i
, we obtain the following relation

eE
i

kN
e
f
i

e
i k
+
_

i

k
n
_
= 0
Multiplying this relation by 1/2 and adding it to the right hand side of (4.10) we
obtain

kN
i
f
k

ik

i
=

eE
i

kN
e
f
k
+ f
i
2

e
i k
+

eE
i

kN
e
f
k
f
i
2

_

i

k
n.
By recalling that
e
i k
= 0 for e (E
i
E
k
) and
i k
=

e(E
i
E
k
)

e
i k
and that

ii
= 0, we can rearrange the right-hand side of the last relation so as to obtain

kN
i
f
k

ik

i
=

kN
i,=
f
k
+ f
i
2

i k
+

kN
i,=
f
k
f
i
2

_

i

k
n,
where N
i,=
denotes, as already specied, all mesh points belonging to
i
except
for the node i itself. Since
i
= 0 on
i
unless the node i belongs to , the
last term can be written as

kN
i,=
f
k
f
i
2

_

ik

i

k
n
The boundary integral will be different from zero only provided that also the node
k belongs to , so that, to indicate this explicitly, the index of the summation is
rewritten as follows

kN

i,=
f
k
f
i
2

_

ik

i

k
n
27
Therefore the nal result is

kN
i
f
k

ik

i
=

kN
i,=
f
k
+ f
i
2

i k
+

kN

i,=
f
k
f
i
2

_

ik

i

k
n,
that is, the identity (4.5).
Exploiting the identity (4.5) just established, the starting equation (4.4) reduces
to equation (4.2) that will be written compactly as

kN
i
M
i k
du
k
dt
=

kN
i,=
f
k
+ f
i
2

i k
+BT
i
, (4.11)
after having introduced the following notation for indicating together the two
boundary contribution terms
BT
i
=

kN

i,=
f
k
f
i
2

_

ik

i

k
n

kN

i,=
f
k

ik

i

k
n. (4.12)
4.2.2 Final transformation of the boundary term
The two integrals contributing to expression (4.12) can be recast in an alternative
manner that involves the use of metric vectors associated with node-pairs belonging
to the boundary as well as with the nodes of the boundary. The nal form gives
the boundary terms where the boundary conditions can and must be enforced in
the Galerkin FE approache, in a weak form.
Initially, let us rewrite the second term (apart from its sign) of BT
i
by isolating
the contribution of the node i , as follows

kN

i
f
k

ik

i

k
n =

kN

i,=
f
k

ik

i

k
n + f
i

ik

i
n
We now add and subtract the term

kN

i,=
f
i

ik

i

k
n
to the relation above to obtain

kN

i
f
k

ik

i

k
n
=

kN

i,=
( f
k
f
i
)
_

ik

i

k
n + f
i

kN

i
_

ik

i

k
n.
28
By exploiting the local support property of the shape functions andthe basic relation

kN
e

k
(x) = 1, x
e
, e E,
we nally obtain

kN

i
f
k

ik

i

k
n =

kN

i,=
( f
k
f
i
)
_

ik

i

k
n + f
i

i
n
By substituting this result into (4.12) we note that there are two equal terms but
with coefcients
1
2
and 1, so that the boundary contribution to the i -th discrete
equation assumes the following from:
BT
i
=

kN

i,=
f
k
f
i
2

_

ik

i

k
n f
i

i
n, (4.13)
where the rst integral is over the domain

i k
=
i k
, and therefore it
exists only for i, k , with however i = k since for i = k the vector coefcient
vanishes.
We are therefore led to introduce the following boundary metric vectors

i k
def
=
_

ik

i

k
n, i, k N

(i = k)

i
def
=
_

i
n, i N

in terms of which the boundary contribution is written in the nal form


BT
i
=

kN

i,=
f
k
f
i
2

i k
f
i

i
. (4.14)
We note that

i k
is dened only for i, k N

and under the condition


i

k
=
. Moreover,

i
is similarly dened only for i N

.
The symmetry property

i k
=

ki
can be contrasted with the antisymmetry
property
i k
=
ki
. However, the coefcients in front of the two metric vectors
are respectively symmetric and antisymmetric, so that the two corresponding scalar
products share one and the same property of antisymmetry with respect to the
indices i and k. Stated in other words, the contribution of all node-pairs, both the
domain terms and the boundary ones, repects conservation exactly.
29
4.3 Node-pair form of the discrete equations
Let us now recall the conservation law (1.1) we started from, i.e.,
u
t
+ f (u) = 0.
Once the metric vectors have been introduced, the complete nal form of the
spatially discrete conservation law in the node-pair representation reads

kN
i
M
i k
du
k
dt
=

kN
i,=

i k

f
i
+ f
k
2

kN

i,=

i k

f
k
f
i
2

i
f
i
, (4.15)
where f
i
def
= f (u
i
). The last, nodal, contribution

i
f
i
is present only when node
i lies on the boundary : in other words, when i correponds to an internal point,
the respective equation does not contain any term of this form. This circumstance
is reminded by the superscript index appended to boundary metric vector

i
.
To summarize all elements of the node-pair formulation, we collect here the
denition of the mass matrix M and of the metric vectors ,

and

, namely
M
i k
=
_

ik

i

k
,

i k
=
_

ik
(
i

k

k

i
), (i = k)

i k
=
_

ik

i

k
n, i, k N

(i = k)

i
=
_

i
n, i N

Remarkably, all the geometrical information pertaining to the triangulation are


contained in the metric quantities above.
Furthermore, the evaluation of the summations in the right-hand side of the
discrete equations can be performed quite easily by processing sequentially the
list of node-pairs and accumulating the left and right contributions with the
proper sign according to their overall antisymmetry. More precisely, the list of
all interacting node-pairs involved by the domain contribution term is stored in a
suitable array whose entries contain the order numbers of two nodes of each node-
pair. This array contains therefore the connectivity information of the entire mesh
seen from the viewpoint of knowing the interaction between the nodes as brought
about by the overlapping of the their respective support. By means of this nod-pair
30
connectivity, the summation of the terms on the right-had side can be performed
explicitly by accounting for the contribution of each node-pair termto the quantities
of two nodes. A similar array for node-pair-to-nodes connectivity is memorized
for the set of node-pairs involved by the summation associated with the boundary.
The corresponding boundary connectivity information allows to account for the
boundary integral expressed in the node-pair format, similarly to what has been
explained for the domain integral. Finally, the contribution of the third termis taken
into account by looping on the list of all boundary nodes. The latter must therefore
been memorized as another boundary information for the considered mesh. Thus,
the node-pair representation of the spatial triangulation comprises three differents
pieces of information for the mesh topology, one pertaining to the domain and
the other two associated with the boundary, while a FEM mesh is described by
only two connectivity element-to-nodes arrays, one for the domain elements and
the other for the boundary element. The resulting algorithm for performing the
explicit accumulation of the right-hand side in node-pair format is straight forward
and an example for a conservation law is given in appendix C.
4.4 Treatment of boundary conditions
For simplicity, the node-pair form of the spatially discrete equations has been de-
rived in the previous sections without considering the imposition of the boundary
conditions of the considered hyperbolic problem. In particular, the boundary inte-
gral has been rearranged by assuming that a constant reinterpolated ux function
f
i
over the support
i
of i . This assumption cannot be accepted when boundary
conditions change frominowto outowalong the same element, as pointed out in
section 3.2. As a consequence, the envisaged factorization of the ux function in
the boundary integrals is not allowed and the formulation must be modied slightly
so as to include the possibility of imposing boundary conditions of different type.
Let us recall the form of the boundary integral which has been obtained under
the hypothesis of ux reinterpolation. We have
_

i
n f (u)

eE

kN
i
f
e
k

ik

k
n (4.16)
where
f
e
k
(t )
def
= f
_
u(s
k
, t )
_
, s
e
(4.17)
4.4.1 Duplication and augmentation of boundary nodes
To simplify the analysis, let us assume that the partition of the boundary in its
inow and outow parts does not occurs inside any boundary elements. In
31
other words, we are assuming that the frontier between these two parts is precisely
located at points (in 2D) and along edges (in 3D) which are on the border of the
boundary elements.
Under this assumption each boundary element can be associated either to the
inow boundary or to the outow boundary, according to the sign of n f ,
where f is evaluated in the interior of the boundary element. This circumstance
allows tosplit the nodal term

i
f
i
inits contributions due tothe different boundary
elements. In 2D problems, there are always two distinct contributions, stemming
from the two boundary edges containing the boundary node i . Therefore one can
augment articially the number of boundary nodes, by duplicating all of them,
so as to compute the term of the purely nodal contribution still as a single cycle
running on a double number of articially augmented boundary node.
On the contrary, in 3D problems there a few of boundary element containing
the boundary node i as one of their vertices, the typical number being around 5 or
6. In this case, it is necessary to known the number N
i
of these boundary triangles
associated with the boundary node i . Then one can augment articially the number
of boundary nodes by replicating N
i
times the surface node i .
By summarizing, while the numbering of domain nodes is left unchonged,
for each boundary node i one denes the set

N
i
of the two duplicated boundary
nodes in two-dimensional problems or of the N
i
-folded boundary nodes in three-
dimensional problems.
Consideringnowthe surfacial node-pair term
1
2

i k
( f
k
f
i
) inthree-dimensional
problems, a difculty with the imposition of boundary conditions changing from
infow to outow type across the common edge of two boundary triangles is en-
countered, similar to that just discussed for the nodal metric vectors. This difculty
requires to split the two elemental contributions of each boundary metric vector

i k
(see gure 4.3). Accordingly, elemental boundary metric vectors could be
dened as follows:

,e
i k
=
_

ik

i

k
n, i, k N

, e E

(i = k)

,e
i
=
_

i

e

i
n, i N

, e E

In terms of these elemental contributions the complete boundary term could be


written as follows:
BT
i
=

kN

i,=

eE

ik

,e
i k

f
e
k
f
e
i
2

eE

,e
i
f
e
i
(4.18)
The double summation appearing in the rst (node-pair) term is retained here to
have an expression valid in general for both two and three-dimensional problems.
32

i
k
m
j
n
e
n
m

e
i

m
i

m
j

e
k

i k

i j
e
Figure 4.3: Boundary metric quantities for elements m and e in two spatial dimensions.
Thanks to the augmented set of boundary nodes just described, the elemental
contribution
,e
i
to the metric vector

i
will be regarded as an independent nodal
metric vector and will be indicated by

, to make explicit the fact that they belong


to the augmented set of boundary nodes. In this way, the term for the boundary
nodal contribution reduces to a summation over the articially extended set of the
duplicated (in 2D) or multiplicated (in 3D) boundary nodes, as follows:

eE

,e
i
f
e
i


N
i

,
where

are the boundary metric vectors duplicated (2D) or multiplicated (3D) in


conformity with the treatment of the boundary nodes. Note that with the assumed
notation the superscript e referring to the element becomes unnecessary.
4.4.2 Duplication of boundary edges in 3D problems
As anticipated, a similar treatment can be adopted also for the term associated
with the node-pair boundary contribution for three-dimensional equations. Each
boundary node-pair can be duplicated to account for the two boundary triangles
involved by the considered pair. The elemental contributions
,e
i k
are then consid-
ered as referring to an articially doubled set of node-pairs on the boundary, that
will be indicated by

N

. (For the two dimensional equations



N

consists simply
of the set of all the node pairs on the boundary, whose number is equal to the
33
number of edges along the boundary.) Thus, by interpreting the rst summation
over k N

i,=
as running over the duplicted set of boundary node-pairs, the second
summation over the elemental contribution is made to disappear as follows:

kN

i,=

eE

ik

,e
i k

f
e
k
f
e
i
2

k

N

i,=

k
f

2
where the number of the boundary metric vectors

k
is twice that of edges on the
domain boundary in three dimensions. Again we note that the assumed notation of
the extended set of metric vectors

k
eliminate the need to refer to the elements in
the ux evaluation: in fact the number of ux values on the boundary is equal to the
number of the augmented boundary nodes. The expression above is the general
form of the boundary contribution due to the surfacial node-pairs valid for the
three-dimensional equation. In the two-dimensional case there is no duplication of
the boundary edeges, which are parts of the discretized curve delimining the plane
computational domain. Therefore in two dimensions, the considered boundary
contribuition would reduce to the simpler standard summation

kN

i,=

eE

ik

,e
i k

f
e
k
f
e
i
2

kN

i,=

i k

f
k
f
i
2
2D only.
In the following, the previous more general expression of this term will be always
considered to obtain equations suitable for implementing the numerical schemes
also in the three-dimensional case.
Both treatments for the boundary terms allow a very convenient simplication
of the program, only at the expense of interpreting the boundary as if it were
composed of fully disconnected edges in 2D or fully disconnected triangles in 3D.
The boundary contribution to the discrete equations will be written as follows
BT
i
=

k

N

i,=

k
f


N
i

.
It is worth reminding that in 2D the set

N
i
has always two elements while in 3D
it contains a number of elements equal to the number of triangles over the surface
which contain the boundary node i . On the other hand, the set

N

in 2D is
nothing but N

while in 3D it always contains twice the number of the boundary


edges of the mesh.
For subsequent reference, the discrete conservation law with the boundary
34
terms expressed in the new form is written here explicitly:

kN
i
M
i k
du
k
dt
=

kN
i,=

i k

f
i
+ f
k
2

k

N

i,=

k
f


N
i

.
(4.19)
Of course, the two terms associated with the boundary are present only when the
node i lies on the domain boundary: the absence of these terms for the discrete
equations associated to internal nodes is explicitly indicated by appending the
superscript to the boundary metric vectors

k
and

.
It can be noted that the three summations in the right-hand side of the last
equation can be easily coded in the program by means of do-loops. The three
loops run on the three sets of all the node-pairs, all the (augmented) boundary
node-pairs and all the (augmented) boundary nodes, respectively. The superscript
e referring to the element is present in the two boundary terms simply due to the
fact that the ux on the boundary depends on the boundary value of u, which in
turn depends on the inow or outow nature of each boundary element.
4.5 Diffusion term in node-pair form
In section 2.4, the nite element discretization of a diffusion termhas been derived;
in this section we address the problem of expressing the weak form of such a term
under the node-pair format. We anticipate that when the diffusion coefcient is
not constant, the resulting node-pair form can be achieved only at the expense of
introducing an additional error with respect to spatial discretization error of the
classical nite element procedure.
Let us now consider the weak form of the diffusion term. By integrating by
parts we have
_

i
(u
h
) =

kN
i
__

ik
(
i
)
k
_
u
k
+
_

i
b (4.20)
According to the standard nite element method the two integrals on the right-hand
side of this expression are computed by splitting them into a sum over the mesh
elements and a sum over the boundary sides, respectively. Let us examine how
these summations are evaluated by employing the same node-pair data structure
previously introduced.
35
By exploiting the property

kN
e
k
(x) = 0, x
e
, on each nite
element
e
of the mesh, we may calculate
(
i
) u
h
= (
i
)

kN
e
u
k

k
=

kN
e
(
i
)
k
u
k
=

kN
e
i,=
(
i
)
k
u
k
+(
i
)
i
u
i
, x
e
.
By subtracting the term u
i

kN
e
k
, which sum up to zero, we obtain
(
i
) u
h
=

kN
e
i,=
(
i
)
k
u
k
u
i

kN
e
i,=

k
=

kN
e
i,=
(
i
)
k
(u
k
u
i
), x
e
Therefore, the volume integral on the right-hand side of expression (4.20) for the
second-derivative can be rewritten as
_

i
(
i
) u
h
=

eE
i
_

e
(
i
) u
h
=

eE
i

kN
e
i,=
__

e
(
i
)
k
_
(u
k
u
i
)
=

kN
i,=
__

ik
(
i
)
k
_
(u
k
u
i
)
where in the last relation we have used the fact that the integral vanishes outside

i k
. Therefore, the Galerkin discretization of the diffusion term (4.20) is
_

i
(u
h
) =

kN
i,=
_

i
,
k
_

ik
(u
k
u
i
) +
_

i

N

i
b (4.21)
in which the node-pairs data structure has been put into evidence.
If we now approximate
|
ik
with a constant value
i k
for each node pair i -k,
i.e., for example, by taking

|
ik

i k
=
1
2
(
i
+
k
), (4.22)
36
we obtain the following approximation
_

i
(u
h
)

kN
i,=

i k
_

i
,
k
_

ik
(u
k
u
i
) +
_

i

N

i
b (4.23)
where the volume integral has been made independent from the function (x) and
therefore can be evaluated once and for all at the beginning of the computation as the
other metric vectors already introduced. Note that the considered approximation
preserves the symmetry of the stiffness matrix. Of course, for constant the
technique reduces to the exact nite element expression. Recalling the denition
(2.14) of the stiffness matrix element K
i k
, the proposed approximation yields
_

i
(u
h
)

kN
i,=

i k
K
i k
(u
k
u
i
) +

k(N

i
N

N
)
_
_

ik

i

k
_
b
k
(4.24)
where the boundary Neumann datumhas been replaced by its interpolant according
to the approximation
b(s, t )

k(N

i
N

N
)
b
k
(t )

k
(s), s
N
. (4.25)
37
Chapter 5
Finite-Volume method
on nonstructured meshes
In this chapter we study the nite volume method for unstructured meshes. We
derive the upwind scheme based on this spatial discretization technique. The semi-
discrete equations are rst expressed in the standard nite-volume format. Then,
thy are expressed in the node-pair format, using the data structure and metric quan-
tities already described in Chapter 4, by taking advantage of Selmins theorem of
equivalence between nite volumes and nite elements in node-pair representation
[23].
5.1 Finite-Volume spatial discretization
Let us write the standard nite volume scheme for the scalar conservation law
(1.1). As well know, the nite volume method moves from the conservation law
written in the following integral form
d
dt
_
C
u(x, t ) =
_
C
n f (u), C ,
where n indicates the outward normal vector of the region C . Let us now
consider the discrete counterpart of the above equation by considering a certain
number of nite volumes C
i
, with boundary C
i
, each of them surrounding a single
node i of the triangulation of (see gure 5.1), namely,
d
dt
_
C
i
u(x, t ) =
_
C
i
n f (u), i N. (5.1)
In the following, we assume that the nite volumes C
i
satisfy the constraints:
38

i
C
i
Figure 5.1: Finite volume discretization of in two spatial dimensions. The underlying
FEM discretization using triangular meshes is shown (dashed lines).

C
i


C
k
= , i, k N, i = k, (5.2a)
_
kN
C
k
= , i, k N, i = k, (5.2b)
i C
i
i / C
k
, i, k N, i = k. (5.2c)
Condition (5.2a) assures that the open sets

C
i
, which are the internal parts of the
nite volumes C
i
, are nonoverlapping, while condition (5.2c) implies that each
nite volume C
i
is associated with a single node i . Alternatively, nite volumes
can be chosen to be coincident with the elements of the triangulation, that is,
C
e
=
e
; the resulting scheme is said to be a cell-centred nite volume scheme,
to be contrasted with the above node- or vertex-centred approach.
Over each nite volume C
i
, the unknown u(x, t ) is now approximated by its
spatial average as
u(x, t ) u
i
(t )
def
=
1
|C
i
|
_
C
i
u(x, t ), x C
i
. (5.3)
In other words, we consider the following piecewise constant approximation of
the unknown over (cf. gure 5.2), i.e.,
u(x, t ) u
h
(x, t )
def
=

kN
u
k
(t ) I
k
(x), (5.4)
39
u
u
h
Figure 5.2: Piecewise constant approximation u
h
of u in one spatial dimension.
where, borrowing the nomenclature of nite elements, the shape functions are
chosen to be the characteristic functions I
k
(x)
I
k
(x) =
_
1 x C
k
0 x / C
k
Therefore, if the nite volume does not change in time, we can write
d
dt
_
C
i
u(x, t ) |C
i
|
du
i
dt
.
As a rst step toward the node-pair representation, the right hand side of (5.1) is
rearranged so as to put into evidence the node-pair structure of the data, namely,
_
C
i
n f (u) =

kN
i,=
_
C
ik
n
i
f (u) +
_
C

i
n f (u), (5.5)
where n
i
denotes the outward normal with respect to the volume C
i
. We notice
that n
i
= n
k
over C
i k
= C
i
C
k
. In the nite volume jargon, the set C
i k
is
often referred to as the cell interface between the volumes C
i
and C
k
(gure 5.3).
5.2 Upwind Finite-Volume scheme
We can now consider the issue of imposing inlet boundary conditions for the
hyperbolic problem. By means of the same procedure detailed in section 2.2, the
boundary condition is enforced in a weak sense throught the boundary integral by
substituting the unknown u with the function u dened in equation (2.7), which is
repeated here for convenience,
u(s, t )
def
=
_
_
_
u(s, t ) if n(s) a(u(s, t )) 0
a(s, t ) if n(s) a(u(s, t )) < 0
40
i
k
C
i
C
k
C
i k

i k
Figure 5.3: Cell interface C
ik
between the nite volumes C
i
and C
k
.
where a(u) = d f (u)/du and a(s, t ) denotes the prescribed boundary value on the
inlet portion
in
of the boundary. It should be reminded that a is the boundary
value prescribed on u and is not the intensity of the advection velocity a. This
boundary function is replaced by its piecewise constant approximation, as follows,
a(s, t ) a
h
(s, t ) =

kN
a
k
(t ) I

k
(x),
where
I

k
(s)
def
= I
k
(x(s)), x(s) ,
denotes the trace of the shape function I
k
over . The integral over the domain
boundary in (5.5) is therefore evaluated as usual by the expression
_
C

i
n f =
_
C

i
n f (u). (5.6)
Due to the piecewise constant approximation considered, over C
i
u
h
and
a
h
assume the constant values u
i
and a
i
, respectively, so that u
h
(s, t ) is constant
over C
i
as long as the outward normal n(s) is constant over C
i
. In
standard nite volume discretizations, a constant (mean) normal is considered over
C
i
, so that u
h
assumes a constant
1
value and can therefore be factorized in
the computation of the boundary integral, namely,
_
C

i
n f (u
h
(s)) f
i
(t )
_
C

i
n
with
f
i
(t )
def
= f (u
h
(s)) = constant s C
i
.
1
If a cell-centred volume is considered over a grid of linear elements, the normal n is indeed a
constant vector over C
e
.
41
Here, by taking advantage of the analysis performed in section 3.2, we choose to
consider a nonconstant normal over C
i
and split the boundary integral as
follows
_
C

i
n f (u
h
(s)) =

eE

i
f
i
(t )
_
C
i

n (5.7)
where
e
is the subregion of corresponding to element e of the nite element
triangulation and where we have introduced the following vector quantity
f
i
(t )
def
= f (u(s
i
)) = constant s C
i

e
. (5.8)
Let us consider now the nite volume approximation of the domain integral.
First, we notice that the integrals
_
C
ik
n
i
f (u), (5.9)
are undened at cell interfaces in the nite volume framework due to the piecewise
constant approximation u
h
chosen for u; with the discrete unknown u
h
discontin-
uous across C
i k
. Therefore we introduce a numerical ux f
i k
, representing an
approximation of f (u) at the cell interface C
i k
, so as to have
_
C
ik
n
i
f (u)
_
C
ik
n
i
f
i k
.
If we select a constant f
i k
over C
i k
, the numerical ux can be factorized in the
computation of the above integral to obtain
_
C
ik
n
i
f (u) f
i k

_
C
ik
n
i
= f
i k

i k
,
where

i k
def
=
_
C
ik
n
i
.
We notice in passing that
i k
=
ki
. As an example, a second order approxima-
tion of (5.9) can be obtained by choosing the following centred approximation of
f (u) over C
i k
f
i k
=
f (u
i
) + f (u
k
)
2
.
To allowfor the inclusion of a more general approximation of (5.9), it is prefer-
able to introduce an integrated numerical function q
i k
, which depends on u
i
and
u
k
as well as on the integrated normal
i k
and which provides the following ap-
proximation
_
C
ik
n
i
f (u) q
i k
= q(u
i
, u
k
,
i k
).
42
The integrated numerical ux q
i k
is assumed to satisfy the following conditions:
q(u
i
, u
k
,
i k
) = q(u
k
, u
i
,
ki
) (Conservation) (5.10a)
q(u, u,
i k
) = f (u)
i k
(Consistency) (5.10b)
Condition (5.10a) is required to obtain a conservative scheme since it implies that
the integral of the ux computed at the cell interface C
i k
is equal (up to the sign)
to the one computed at the cell interface C
k
C
i
. In terms of the numerical ux
f
i k
, the condition is
q
i k
f
i k

i k
= f
ki

ki
q
ki
,
and it can be satised by selecting a symmetric numerical ux, namely, f
i k
= f
ki
,
since
i k
=
ki
.
The fully discrete form of (1.1) under the nite volume approximation is there-
fore,
|C
i
|
du
i
dt
=

kN
i,=
q(u
i
, u
k
,
i k
)

eE

,e
i
f
i
(t ) (5.11)
where

,e
i
def
=
_
C
i

n,
the functional form of the integrated numerical ux still remaining to be selected.
Consider now the following denition for the numerical ux
q(u
i
, u
k
,
i k
)
def
=
i k

f
i
+ f
k
2

1
2
| a
i k
| (u
k
u
i
) (5.12)
where a linearization projected along the direction
i k
has been introduced, namely,
a
i k
def
=
_

i k

f
k
f
i
u
k
u
i
if u
k
= u
i
,

i k

d f (u)
du

u
i
u
k
if u
k
= u
i
.
(5.13)
The integrated numerical ux (5.12) leads to well know rst order (in space and
time) nite volume upwind scheme for the conservation law, namely,
|C
i
|
du
i
dt
=

kN
i,=
_

i k

f
i
+ f
k
2

1
2
| a
i k
| (u
k
u
i
)
_


N
i

, (5.14)
where the duplication or multiplication of the boundary nodes in 2D and 3D,
respectively, has been assumed to obtain the algorithmically more convenient rep-
resentation of the elemental contributions
,e
i
to the averaged normal vectors.
43
5.3 The bridge between nite volumes
and nite elements
Let us assume that a nite volume discretization of a two-dimensional domain is
derived from the underlying grid of triangular elements in the following way:
Each element e of the grid is divided into three subelements delimited by
the medians of the triangle itself, as in gure 5.4. The sets
i
,
k
and
j
associated with the vertices i , k and j of triangle e, respectively, are delimited
by the two element sides from each node and the two segments that join the
middle of each element side to the center of gravity x
g
of the element, where
x
g
def
=
1
|
e
|
_

e
x.
The nite volume C
i
associated with node i is given by the union of all the
subelements having i as a vertex, namely,
C
i
def
=
_
eE
i

i
.
An example of a nite volume C
i
assembled following the above prescriptions is
given in gure 5.4.
In [23], the following identities are given for metric quantities generated over
the nite volume dicretization described above:
|C
i
| =
_

i
=

kN
i
M
i k
= L
i
, (5.15a)

i k
=
i k
, (5.15b)

i
=

i
. (5.15c)
The proof of the important relation (5.15a) is direct. It can be obtained by con-
sidering a representative triangle of the mesh and focusing on that part of its area
contributing to the cell associated to one its vertex, the node i in relation (5.15a).
This portion of the triangle is dened by the medians and its area and can be cal-
culated geometricaly. On the other hand, the contribution to the mass elements
M
i k
coming from the integrals inside the considered triangle can be evaluated
analytically for linear interpolations.
For instance, let us consider the representative generic triangle with vertices
x
1
= (0, 0), x
2
= (a, 0) and x
3
= (b, c), and let us take the part inside this triangle
44
i
k
j
C
i

e
i

e
k

e
j
e
Figure 5.4: Construction of the nite volume C
i
satisfying the nite elements/nite vol-
umes equivalence requirements.
of cell around the node (0, 0) coincident with its rst vertex. The intersection of
the medians of the triangle is easily nd to be x
m
= ((a + b)/3, c/3). Thus the
area A of the cell portion inside the triangle is made of two components and is
given by
A =
1
2
a
2
y
m
+
1
2

b
2
+c
2
2
d
where d is the distance of x
m
to the triangle side oblique with respect to the
Cartesian axes. Since d =
1
3
ac/

b
2
+c
2
, we obtain
A =
ac
12
+
ac
12
=
ac
6
.
To evaluate (the contribution to) the mass elements M
1,1
, M
1,2
and M
1,3
the basis
functions over the considered triangle are needed. The three linear interpolation
functions are easily found to be
_

1
(x, y) = 1
x
a

_
1
b
a
_
y
c
,

2
(x, y) =
x
a

by
ac
,

3
(x, y) =
y
c
.
45
The integrals contributing to the mass elements are
_ _
triangle

k
dx dy =
_
c
0
dy
_
a+(b1)y/c
by/c

1
(x, y)
k
(x, y) dx.
By a change of variables y Y = y/c and x X = (x b)/a the three
integrals give the mass contributions
M
1,1
=
ac
12
, M
1,2
=
ac
24
, M
1,3
=
ac
24
,
which add to geometrical result A = ac/6, as required. This proof can be
extended in three dimensions to tetrahedral elemenst with linear interpolations.
As a consequence, the upwind nite volume scheme (5.14) can be written in
terms of the nite element metric quantities introduced so far as
L
i
du
i
dt
=

kN
i,=
_

i k

f
i
+ f
k
2

1
2
| a
i k
| (u
k
u
i
)
_


N
i

, (5.16)
where we recall the denition of the linearization projected along the direction
i k
a
i k
def
=
_

i k

f
k
f
i
u
k
u
i
if u
k
= u
i

i k

d f (u)
du

u
i
u
k
if u
k
= u
i
.
It can be noticed that the above upwind nite volume method can be written in the
usual conservation form, i.e.,
L
i
du
i
dt
=

kN
i,=
q
i k


N
i

,
by introducing the integrated numerical ux
q
i k
=
i k

f
i
+ f
k
2

1
2
| a
i k
| (u
k
u
i
).
This result is important because it shows that the unstructured nite volume upwind
method presented here extends to multidimensional equations the schemes written
in conservation form for one-dimensional problems.
46
Chapter 6
Taylor-series-based integration
schemes
In this chapter we describe how to compute the terms occurring in LaxWendroff
scheme as well as in TaylorGalerkin scheme of Donea [3]. First, we examine
the second-order correction term arising from the Taylor expansion in time of the
multidimensional conservation law (1.1) for a scalar quantity u. Then, the Taylor
Galerkin scheme for a conservation law endowed with a linear ux function is
derived. In Appendix D, the numerical properties of the original TaylorGalerkin
scheme and of some two-step versions of it are recalled.
6.1 Nonlinear scalar conservation law
The TaylorGalerkin scheme is based on the idea of performing the discretization
in time before performing the discretization in space; the former is accomplished by
means of a Taylor expansion at time level n, while the latter is performed by means
of the standard Galerkin nite element method. More precisely, the unknown at
the new time level t
n+1
is obtained by means of the following Taylor expansion at
time t
n
u
n+1
= u
n
+t
t
u
n
+
1
2
(t )
2

2
t
u
n
+
1
6
(t )
3

3
t
u
n
+ O
_
(t )
4
_
,
where, with standard notation, the subscript indicates derivation with respect to
time. The terms involving the time derivative are obtained from the governing
equation, i.e.,

t
u = f (u),
47
and by taking the time derivatives of the same equation to give,

2
t
u =
t
f =
_

t
f
_
=
_
d f
du

t
u
_
=
_
d f
du
( f )
_
= (a f ),
where we have substituted for the advection velocity a = a(u) = d f (u)/du. The
weak form of the term (a f ) is therefore easily found, using integration by
parts, in the form
_
, (a f )
_
=
_
, a f
_
+
_
n a f
=
_
a , a u
_
+
_
n a f ,
or, indicating explicitly the dependence on the unknown u,
_
, [a(u) f (u)]
_
=
_
a(u) , a(u) u
_
+
_
n a(u) f (u).
(6.1)
Thus, for the nonlinear conservation law
t
u + f (u) = 0, the second-order
term of the TaylorGalerkin method gives a symmetric correction. This term is
the weak variational expression of the second-order directional derivative, along
the (local) direction of the advection eld a(u).
To simplify the exposition, the description of TG schemes considered so far
has not examined how to impose the boundary condition. We now discuss this
point and we distinguish between the case of a nonlinear conservation law and that
of a conservation law with a linear ux.
For the nonlinear conservation law, there is only one boundary term, namely,
_
n a(u) a(u) u,
stemming from the integration by parts. Note that in this expression the vector u
most be evaluated on the boundary where, in general, has a nonzero normal
component. The determination of this component involves also values of u at
internal points of .
The semi-discrete equations for the TaylorGalerkin method for the scalar
conservation law reads
(, u
n+1
u
n
)
t
=
_
, f (u
n
)
_

n f (u
n
)

t
2
_
a
n
, a
n
u
n
_
+
t
2
_

n a
n
a
n
u
n
48
6.2 The fully discrete form of the equations
Let us consider rst the second-order TaylorGalerkin scheme (TG2 or Lax-
Wendroff Finite Element in [5]) for the nonlinear conservation law (1.1). This
scheme requires to solve only a mass matrix problem, and therefore represents the
basic component to build the two-step schemes of third- and fourth-order accuracy
to be described in appendix D,
If we now introduce the standard nite element approximation in space, the
second-order TG term considered here becomes, omitting the boundary term and
the functional dependence of a for simplicity,
_

i
, [a(u
h
) f (u
h
)]
_

i
=
_
a
i
, a

kN

k
u
k
_

i
=

kN
i
_
a
i
, a
k
_

ik
u
k
(6.2)
The spatially discrete counterpart of boundary integral contribution to the second-
order term of the TaylorGalerkin scheme is

kN

i
_

ik

i
n a(u
h
) a(u
h
) u
h
.
The complete equation of the fully discrete TG2 scheme for the conservation
law (1.1) reads

kN
i
M
i k
u
n+1
k
u
n
k
t
=
_

i
, f
_
kN
i
u
n
k

k
__

i

_

i
n f
_
u
n
h
(s)
_

t
2

kN
i
_
a
n

i
, a
n

k
_

ik
u
n
k
+
t
2

kN

i
_

ik

i
n a
n
a
n
u
n
h
with a
n
def
= a(u
n
h
).
For a description of the properties of different TG schemes we refer to [5] and
[6]. Two-step TG schemes are described in Appendix D, which closely follow
reference [19]. In this appendix, the error analysis for TG schemes is performed
to compare single-step to two-step TG schemes; it must be emphasized that all
these methods include the evaluation (either in an explicit or implicit way) of the
second-order directional derivative described so far.
6.3 TaylorGalerkin scheme in node-pair form
In the this section, by taking advantage of the (approximate) node-pair represen-
tation of a viscous term give in section 4.5 and by following the guidelines for
49
the construction of a TaylorGalerkin schemes given in section 6.1, the node-pair
representation of the bulk contribution to the TG terms is derived. This amounts
to introduce an approximation of the velocity eld in the computation of the TG
terms, thus allowing for a very compact and efcient representation of them.
6.3.1 The bulk TG term
Let us recall the expression of the second order termoccurring in the TGdiscretiza-
tion of the nonlinear scalar conservation law (1.1), namely,

t
u + f (u) = 0
with a nonlinear ux f (u) = f (u(x, t )), i.e., (cf. equation (6.2))
_

i
, [a(u
h
) f (u
h
)]
_

i
=
_
a(u
h
)
i
, a(u
h
)

kN
u
k

k
_

i
=

kN
i
_
a(u
h
)
i
, a(u
h
)
k
_

ik
u
k
.
Then, by the same argument followed in the analysis for the calculation of the
diffusion term, we can write, for each element e,
(a
i
) a u
h
= (a
i
) a

kN
e
u
k

k
=

kN
e
(a
i
) (a
k
) u
k
=

kN
e
i,=
(a
i
) (a
k
) u
k
+(a
i
) (a
i
) u
i
.
By subtracting the term
u
i
(a
i
) a

kN
e

k
,
which is equal to zero over the element e since

kN
e
k
= 0 for any x
e
,
we obtain
(a
i
) a u
h
=

kN
e
i,=
(a
i
) (a
k
u
k
) u
i
(a
i
) a

kN
e
i,=

k
=

kN
e
i,=
(a
i
) (a
k
) (u
k
u
i
), x
e
.
50
Thus, the second-order term of the TaylorGalerkin method can be written in the
node-pair format as follows
_

i
, [a(u
h
) f (u
h
)]
_

i
=

kN
i,=
_
a(u
h
)
i
, a(u
h
)
k
_

ik
(u
k
u
i
).
(6.3)
Let us nowintroduce the approximation of consideringa locallyconstant advec-
tion eld over the intersection
i k
of the supports of the shape functions associated
with node i and k; for instance we may choose
a(u
h
)
|
ik
a
i k
= a
_
1
2
(u
i
+u
k
)
_
.
This approximation allows us to write the considered TG term in the following
compact form
_

i
, [a(u
h
) f (u
h
)]
_

kN
i,=
K
i k
: [a
i k
a
i k
] (u
k
u
i
), (6.4)
where, for each node-pair, we have introduced the (symmetric) metric tensor
K
i k
def
=
_

ik

i

k
(i = k), (6.5)
which can be regarded as a tensor generalization of the stiffness matrix K
i k
asso-
ciated with the couple i k.
When the tensor diffusion term above is to be evaluated implicitly (as is the
case, for example, when the TG3 scheme is employed), it is necessary to consider
also the tensor elements
K
ii
def
=
_

i

i
.
For two-dimensional equations, the meaning of each contracted term in the
summation (6.4) is
K
i k
: [a
i k
a
i k
] = [
x

h
]
i k
(a
x,i k
)
2
+[
x

h
]
i k
a
x,i k
a
y,i k
+[
y

h
]
i k
a
y,i k
a
x,i k
+[
y

h
]
i k
(a
y,i k
)
2
= [
x

h
]
i k
(a
x,i k
)
2
+2 [
x

h
]
i k
a
x,i k
a
y,i k
+[
y

h
]
i k
(a
y,i k
)
2
,
the subscript x and y denoting the vector component in the x and y direction,
respectively.
51
6.3.2 The boundary TG term
Let us consider now the boundary integral stemming from the second order term
of the TaylorGalerkin scheme, namely,
_
n a a u
_

i
n a a u
h
,
where the function

i
is dened only on the boundary, while u
h
is dened also in
the volume in order to be able to compute the gradient.
In spatially discrete form this surface term will be approximated as the corre-
sponding volume term. This means to introduce the approximation
_

i
n a a u
h

kN

i,=
_

ik

i
n a
i k
a
i k

_
kN
i
u
k

k
_
.
The evaluation of the right-hand side to compute the gradient of the variable u
on the surface of the computational domain. It is therefore necessary to know the
behaviour of the eld u in the layer of nite elements touching the boundary. More
precisely, the layer where the eld u is used consists of the nite element with (at
least) one edge in 2D or one triangular face in 3D lying on the boundary. As a
consequence, the evaluation of the considered boundary term involves also values
of degrees of freedom at points in the interior of the computational domain. We
denote by N

i
the subsect of nodes belonging to this layer of domain elements and
which belong to the support of the boundary node i . Exploiting the local support
property

kN
e
k
(x) = 1, we have

_

kN

i
u
k

k
_
=
_

kN

i,=
u
k

k
_
+(u
i

i
)
=
_

kN

i,=
u
k

k
_
+u
i

_
1

kN

i,=

k
_
=

kN

i,=
(u
k
u
i
)
k
.
Substituting this results in the integral we obtain
_

i
n a a u
h

kN

i,=
_

ik

i
n a
i k
a
i k

k
(u
k
u
i
).
52
We are therefore led to introduce the following boundary metric tensor
K

i k
def
=
_

ik

i
n
k
, with i N

and k N

i
.
According to this tensor notation, the boundary term associated with second-order
operator of the TaylorGalerkin scheme will be evaluated as follows
_

h
n a a u
h
=

kN

i,=
K

i k
: [a
i k
a
i k
] (u
k
u
i
).
Concerning this term, a fews remarks are in order. First, notice that the calculation
of the gradient on the surface require to know values of the gradient in the interior
of the elements belonging to the aforementioned layer and it is not possible to
reduce the TG boundary integral elds dened only on the surface.
Second the fact that in the boundary term of the TG method it is wrong to
introduce any information in order to impose the boundary conditions. The genesis
of this term is in fact quite different from the other boundary term resulting from
the integration by parts of the ux of the conservation law. The former comes from
increasing the time accuracy to second order and integrating by parts the second
order directional derivative, but the nature of the equation remains hyperbolic, so
that including boundary conditions in this surface integral as if the equation would
contain a true diffusion term is not correct.
The last remark is about the relevance of this term in actual computations. The
inclusion of this term assures second order time accuracy of the solution on the
boundary. Therefore its omission can be expected to reduce the global quality of
the discrete solution only marginally. Considering the complexity of the actual
implementation of the boundary integral involving the gradient, this observation
might be somewhat relaxing.
53
6.3.3 Node-pair TaylorGalerkin scheme
Combining all the terms of the TG scheme considered so far in the node-pair
format, we arrive at the following system of fully discrete equations

kN
i
M
i k
u
n+1
k
u
n
k
t
=

kN
i,=

i k

f
n
i
+ f
n
k
2

k

N

i,=

f
n

k
f
n


N
i

f
n

t
2

kN
i,=
K
i k
: [a
n
i k
a
n
i k
] (u
n
k
u
n
i
)
+
t
2

kN

i,=
K

i k
: [a
n
i k
a
n
i k
] (u
n
k
u
n
i
).
(6.6)
Although the difculties associated with implementing the boundary integral in-
volving the gradient should not be underestimated, its rather limited inuence on
the global accuracy of the solution must be reminded.
It is interesting to note that the above numerical scheme can be written in the
following form

kN
i
M
i k
u
n+1
k
u
n
k
t
=

kN
i,=
q
n
i k


N
i

f
n

,
by introducing the integrated numerical ux
q
n
i k
=
i k

f
n
i
+ f
n
k
2
+
t
2
K
i k
: [a
n
i k
a
n
i k
] (u
n
k
u
n
i
)
+

f
n

k
f
n

2

t
2

k

N

,=
K

k
: [a
n
k
a
n
k
] (u
n
k
u
n

).
A summation involving an internal node remains in the last term since the original
boundary integral involves the gradient of u
n
, whose evaluation needs a knowledge
of u
n
at some internal points near to the boundary.
6.4 Linear ux function
The expression of the TG scheme presented in the previous section is valid only
when the ux is indeed a function of the unknown alone, that is, only if f (x, t ) =
54
f (u(x, t )), which allows the substitution f (u) = a(u) u. We now want to
consider a ux function with an explicit spatial dependence, i.e., f = f (u, x, t ),
and will focus our attention on a linear ux of the form
f (x, t ) = a(x, t ) u(x, t ),
where a(x, t ) is a possibly nonuniform advection eld. The corresponding con-
servation law states that the quantity u is conserved within the velocity eld a, and
reduces to the advection equation provided that a is solenoidal. The obvious exam-
ple is the mass conservation lawfor a compressible uid, where the scalar unknown
is the mass density and the ux function has the form f (, x, t ) = v(x, t )(x, t ),
v being a known velocity eld.
For the linear ux f (x, t ) = a(x, t ) u(x, t ) considered here, the second-order
time derivative of the unknown reads

2
t
u =
t
_
(au)
_
=
_

t
(au)
_
=
_
a
t
u +u
t
a
_
.
By substituting
t
u = (au), we have

2
t
u =
_
a
_
(au)
_
+u
t
a
_
=
_
a (a u) + a u a u
t
a
_
.
The weak form of this TG term reads
_
,
2
t
u
_
=
_
,
_
a (a u) + a u a u
t
a
__
=
_
,
_
a (a u)
__
+
_
,
_
a u a u
t
a
__
.
The rst term in the relation above has been already dealt with in the previous
section, with here a = a(x, t ) instead of a = a(u). The second term is new and
contains two contributions. The rst contribution is different from zero only if the
advection eld is not solenoidal, i.e., if the scalar eld
(x, t )
def
= a(x, t ),
is different from zero everywhere in , while the second contribution is nonzero
only provided that the advection eld varies with time, i.e., if
t
a(x, t ) = 0. Using
integration by parts in the expression above we obtain
_
,
2
t
u
_
=
_
a , a u
_

_
, au u
t
a
_
+
_
n a a u +
_
n ( au u
t
a).
(6.7)
55
Considering the spatially discretized counterpart of this expression we have
_

i
,
2
t
u
h
_
=

kN
i
_
a
i
, a
k
_
u
k

kN
i
_

i
, ( a
t
a)
k
_
u
k
+
_

i
n a a u
h
+

kN

i
__

i

k
n ( a
t
a)
_
u
k
.
(6.8)
In the particular case of a solenoidal and steady advection eld, expression
(6.7) simplies to
_
,
2
t
u
_
=
_
a , a u
_
+
_
n a a u, a = 0,
t
a = 0,
which still represents the weak form of the second-order directional derivative,
along the direction of the solenoidal advection eld a(x), exactly as in the case of
a nonlinear conservation law, for which however the velocity eld a = a(u) is not
required to be solenoidal.
On the other hand, for the linear ux function the additional boundary term

kN

i
_
_

ik

i

k
n ( a
t
a)
_
u
k
.
must be included.
Let us nowassume that the advectioneldof the linear uxfunction f (u, x, t ) =
a(x, t ) u(x, t ) is solenoidal but that it can be time-dependent, so that the ux func-
tion assumes the form f (u, x, t ) = a(x, t ) u(x, t ). In this case the TG scheme for
the conservation law with a linear ux function reads

kN
i
M
i k
u
n+1
k
u
n
k
t
=

kN
i
_

i
, a
n

k
_

ik
u
n
k
+

kN

i
_

ik

i

k
n a
n
u
n
k

t
2

kN
i
_
a
n

i
, a
n

k
_

ik
u
n
k
+
t
2

kN

i
_

ik

i
n a
n
a
n
u
n
h
,
where a
n
def
= a(x, t
n
).
To conclude, we notice that whenever the expression of the divergence is
not known analytically, as, for example, when the velocity eld is know through a
discrete representation, the scalar eld can be retrieved as the weak divergence
of a, by solving a simple mass problem in the form:
_
,
_
=
_
, a
_
=
_
, a
_
+
_

n a.
56
In the standard nite element framework of interest here, this problem is approxi-
mated by

kN
i
M
i k

k
=
_

i
a
i

_

i
n a, (6.9)
where
(x, t )
h
(x, t )
def
=

kN

k
(t )
k
(x). (6.10)
In the case of the linear ux function under examination, the node-pair com-
putation of TG terms is performed as follows. The term
_
a , a u
_
gives a
contribution similar to (6.4), where here a = a(x, t ), and has already been dis-
cussed. Therefore, we focus on the term involving the divergence of the velocity
eld, i.e.,
_
, a u
_
, The quantity a u is then approximated by
a u

kN

k
(t )
k
(x), (6.11)
where

k
(t ) =
k
(t ) a(x
k
, t ) u
k
(t ) (6.12)
Whenever the expression of the divergence is not known analytically, as, for
example, when the velocity eld is know through a discrete representation, the
scalar eld can be retrieved as the weak divergence of a, by solving a simple
mass problem in the form:
_
,
_
=
_
, a
_
=
_
, a
_
+
_

n a
which, in the node-pair formulation considered here reads

kN
i
M
i k

k
=

kN

i,=

i k

a
i
+ a
k
2

k

N

i,=

i k

a
k
a
i
2


N
i

,
where (x, t )
h
(x, t ) =

kN

k
(t )
k
(x) once the values
k
have been de-
termined, the node-pair representation of the considered TG term is obtained by
recalling that
_

i
, ( a u)
_

kN
i,=

i k

i
+
k
2
+

k

N

i,=

i k

k

i
2
+


N
i

.
(6.13)
57
A similar treatment can be applied to the term
_
, u
t
a
_
, by introducing the
approximation
u
t
a

kN

k
(t )
k
(x), (6.14)
where

k
(t ) = u
k
(t )
t
a(x
k
, t ). (6.15)
58
Chapter 7
High-resolution scheme
in node-pair form
In this chapter, numerical schemes for the computation of solutions to a scalar
conservation laws in the presence of discontinuities are given, for both steady-
state and unsteady problems. Both schemes are based on the idea of blending
the previously introduced TaylorGalerkin scheme with a rst-order nite volume
upwind scheme by means of a high-resolution technique (e.g. LeVeque [16]).
The discretized equations of both schemes share the quasi-monodimensional
formcharacteristic of the node-pair formulation. This offers the possibility of using
standard limiters to switch to high order Galerkin or TaylorGalerkin schemes in
smooth regions.
Although a theory of TVD scheme for multidimensional problems is still lack-
ing, such an approach seems to us capable of merging the one-dimensional theory of
upwind schemes and the genuine multidimensional properties of TaylorGalerkin
schemes. In the one-dimensional case the proposed approach reduces to the stan-
dard Roe/LaxWendroff high resolution scheme detailed in [16].
7.1 Introduction
In the previous section 5.3, the rst order nite volume upwind scheme has been
rewritted according to the node-pair representation derived for Galerkin and Taylor-
Galerkin schemes. We now want to introduce a class of high-resolution scheme
which take full advantage of this common formulation. We refer to LeVeque [16]
for a description of high-resolution scheme and for the denition of the usual
concept of monotonicity preserving and TVD schemes.
As well known, in a nite volume setting, a high-resolution scheme can be
derived by blending a scheme of low-order of accuracy with a high-order scheme
59
by means of limiters: the low order scheme, which is monotonicity preserving, is
to be used where the solution is discontinuous; elsewhere, the high-order scheme is
considered. ATVD (Total Variation Diminishing) scheme is obtained by selecting
a proper limiter which acts as a sensor for discontinuities. The general form of the
high resolution integrated numerical ux reads
q
HR
i k
= q
L
i k
+
_
q
H
i k
q
L
i k
_
() (7.1)
where q
L
i k
and q
H
i k
are the low order and the high order integrated numerical uxes,
respectively, and () is the limiter.
1
The limiter is assumed to be a function of
the ratio of the slope in the upwind side with that at the node-pair center:

def
=
slope
upw
slope
center
. (7.2)
For one-dimensional problems and assuming a uniform grid with an ordered num-
bering of the grid points, the slope can be replaced by the upwind and centered
variations of the solution. In this case the variable of the limiter assumes the simpler
form
=
u
upw
i k
u
cen
i k
=
u
upw
i k
u
k
u
i
,
where u
upw
i k
stands for the jumps of the solution evaluated from the upwind
direction, that is, in the direction opposite to the local advection velocity, namely,
since i = j and k = j +1,
u
upw
i k
=
_
u
j
u
j 1
if a
i k
> 0,
u
j +2
u
j +1
if a
i k
< 0.
For multidimensional problems on unstructured grids, the upwind value of the
solution is more difcult to be dened, see below.
In all cases, care must be taken in evaluating the variable which is not dened
when u
cen
i k
= 0, that is, for a locally constant solution. The denition of the
limiter function is therefore to be slightly modied to allow for locally constant
solutions. To this purpose, we recall that the TVDproperty of the scheme is assured
provided that the limiter is a bounded function and, in particular
0 () 2.
Moreover, the following limits exist
lim
0
() = 0 and lim

() =

,
1
The standard notation for the limiter function () is used, which is not to be confused with
the FEM shape functions introduced before.
60
i
k
i

Figure 7.1: The extended node-pair data structure.


which, for a constant u
upw
i k
, correspond to u
cen
i k
and u
cen
i k
0, respec-
tively. The actual value of

ranges from 1 (Minmod limiter) to 2 (Superbee


and van Leer limiters) if the limiter function is contained whitin the so-called
second-order TVD region of Sweby [26]. Moreover, we recall that:
() = 0 if 0.
7.2 Extended node-pairs
For evaluating the limiter on unstructured multidimensional grids in the context of
the quasi one-dimensional node-pair representation of the discrete equations, it is
convenient to introduce an extended node-pair structure [10]. In this formulation
for each node-pair i k two additional nodes, i

and k

, are considered, such that


(x
i
x
k
) (x
i
x
k
)
|x
i
x
k
||x
i
x
k
|
= max
lN
i,=
,l=k
(x
i
x
k
) (x
l
x
k
)
|x
i
x
k
||x
l
x
k
|
, (7.3)
and
(x
k
x
i
) (x
k
x
k
)
|x
k
x
i
||x
k
x
k
|
= min
lN
k,=
,l=i
(x
k
x
i
) (x
k
x
l
)
|x
k
x
i
||x
k
x
l
|
. (7.4)
In other words, the node-pair i

-i is the one, among all the node-pairs connected


with i , excluding the node-pair i -k, which is best aligned with the direction i -k, as
sketched in gure 7.1. A similar denition holds for the node-pair k-k

.
Over the unstructured mesh, the upwind variation u
upw
i k
is dened by rescal-
ing the difference u
i
u
i
or u
k
u
k
so that it corresponds, at a xed slope,
to the same distance between the two central nodes of the considered node-pair.
61
Thus, depending on the the sign of the average velocity a
i k
, the upwind variation
u
upw
i k
will be dened by
u
upw
i k
def
=
_

_
(u
i
u
i
)
|x
k
x
i
|
|x
i
x
i
|
if a
i k
0,
(u
k
u
k
)
|x
k
x
i
|
|x
k
x
k
|
if a
i k
< 0.
(7.5)
In this way, exactly as on uniform 1D grids, the variable of the limiter over the
unstructured grid can be still expressed by the ratio of solution variations
=
u
upw
i k
u
cen
i k
=
u
upw
i k
u
k
u
i
. (7.6)
7.3 High-resolution scheme for steady solutions
In this section, a high-resolution scheme for steady problems is derived. As the
low order scheme, we will use the rst order (in space and time) nite volume
upwind scheme (5.16). The high order scheme is given by the standard Galerkin
scheme (4.15) in node-pair representation. As anticipated, for the latter we will
introduce the mass lumping approximation (7.18): as a consequence, the resulting
scheme is only rst order accurate in time and second order in space. The high
order scheme reads
L
i
du
i
dt
=

kN
i,=

i k

f
i
+ f
k
2

k

N

i,=

k
f


N
i

. (7.7)
We notice in passing that the above scheme is equivalent, on triangular meshes, to
the standard nite volume scheme with a centred integrated numerical ux, that is,
q(u
i
, u
k
,
i k
) =
i k

f
i
+ f
k
2
,
with the exception of the node-pair boundary terms.
The low order and high order schemes are now mixed using the denition
q
HR
= q
L
+
_
q
H
q
L
_
().
Combining the upwind scheme (5.16) with the second order scheme (7.7) we obtain
62
the high-resolution scheme for solving steady-state problems:
L
i
du
i
dt
=

kN
i,=
_

i k

f
i
+ f
k
2

1
2
| a
i k
| (u
k
u
i
)
_


N
i

kN
i,=
1
2
| a
i k
| (u
k
u
i
) (
i k
)

k

N

i,=

k
f

2
(

k
).
(7.8)
where
f
k
(t )
def
= f
_
u
k
(s, t )
_
, s .
Note that the above numerical scheme is coincident, on triangular meshes, to the
corresponding nite volume high-resolution scheme but for a node-pair contribu-
tion at boundaries, which stems from the Galerkin approximation, as shown in
[23].
We now introduce the following limiter in Rebays form
(u
cen
, u
upw
)
def
=
_
u
upw
u
cen
_
u
cen
, (7.9)
which deals with the situation of a locally constant solution automatically, since
when u
cen
= 0,
(0, u
upw
) =

0 = 0.
By means of the limiter in Rebays form the nal scheme reads
L
i
du
i
dt
=

kN
i,=
_

i k

f
i
+ f
k
2

1
2
| a
i k
| (u
k
u
i
)
_


N
i

kN
i,=
1
2
| a
i k
| (u
cen
i k
, u
upw
i k
)

1
2

k

N

i,=

k
f

k
u

_
u
cen

k
, u
upw

k
_
,
(7.10)
The high-resolution scheme for syeady solutions can be recast in conservation form
L
i
du
i
dt
=

kN
i,=
q
HR
i k


N
i

. (7.11)
by introducing the integrated numerical ux
q
HR
i k
=
i k

f
i
+ f
k
2

1
2
| a
i k
| (u
k
u
i
) +
1
2
| a
i k
| (u
cen
i k
, u
upw
i k
)
+
1
2

k
f

k
u

_
u
cen

k
, u
upw

k
_
.
(7.12)
63
with the provisio that, when in the last boundary term u

k
= u

, the corresponding
contribution is absent since it is also f

k
f

= f (u

k
) f (u

) = 0.
7.4 High-resolution scheme for unsteady solutions
For unsteady problems, the use of a scheme with a higher order accuracy in time
is preferable. Therefore, the high resolution scheme will be derived by using the
same rst order (in space and time) nite volume upwind scheme (5.16) as a low-
order scheme and the TaylorGalerkin scheme as the high-order one. Actually,
the time accuracy of the latter is reduced when the mass lumping approximation
(7.18) is introduced, to give the following (second order in time and space) scheme
L
i

t
u
i
t
=

kN
i,=

i k

f
i
+ f
k
2

k

N

i,=

k
f


N
i

t
2

kN
i,=
K
i k
: [a
i k
a
i k
] (u
k
u
i
)
+
t
2

kN

i,=
K

i k
: [a
i k
a
i k
] (u
k
u
i
),
(7.13)
where u
i
= u
n+1
i
u
n
i
and where all the quantities on the right hand side are
understood to be evaluated at time t = t
n
. The above scheme is equivalent, on
triangular meshes, to the standard LaxWendroff nite volume scheme but for
boundary terms.
The resulting high-resolution scheme is obtained by combining the second
order TG scheme with rst order upwind method, to read:
L
i

t
u
i
t
=

kN
i,=
_

i k

f
i
+ f
k
2

1
2
| a
i k
| (u
k
u
i
)
_


N
i

kN
i,=
_
1
2
| a
i k
| +
t
2
K
i k
: [a
i k
a
i k
]
_
(u
k
u
i
) (
i k
)
+
t
2

kN

i,=
K

i k
: [a
i k
a
i k
] (u
k
u
i
) (
i k
)

k

N

i,=

k
f

2
(

k
).
(7.14)
64
Note that the boundary integral of the second derivative contribution to the TG
scheme is split in a nodal part (not limited) and a node-pair part (limited).
Alternatively, by substituting the limiter in Rebays form (7.9) in the volumic
term, the high-resolution scheme can be written as
L
i

t
u
i
t
=

kN
i,=
_

i k

f
i
+ f
k
2

1
2
| a
i k
| (u
k
u
i
)
_


N
i

kN
i,=
_
1
2
| a
i k
| +
t
2
K
i k
: [a
i k
a
i k
]
_

_
u
cen
i k
, u
upw
i k
_
+
t
2

kN

i,=
K

i k
: [a
i k
a
i k
]
_
u
cen
i k
, u
upw
i k
_

1
2

k

N

i,=

k
f

k
u

_
u
cen

k
, u
upw

k
_
.
(7.15)
To evaluate the boundary contribution to the limited second order term requires
to introduce ctitious points external to the boundary to have an extended node-
pair also for the edges with one point on the boundary and one points inside
the computational region. For this boundary term, the value of the variable at
surface nodes can be taken in the form u
i
, which includes the treatment of the
boundary value (inowor outow) introduced for taking into account the boundary
condition. The values at the articially introduced external points could be dened
by extrapolation, e.g., a linear extrapolation, or eventually a quadratic one.
As a nal remark, we notice that the above numerical scheme can be written
in the usual conservation form, i.e.,
L
i

t
u
i
t
=

kN
i,=
q
HR
i k


N
i

, (7.16)
by introducing the high-resolution integrated numerical ux
q
HR
i k
=
i k

f
i
+ f
k
2

1
2
| a
i k
| (u
k
u
i
)
+
_
1
2
| a
i k
| +
t
2
K
i k
: [a
i k
a
i k
]
_

_
u
cen
i k
, u
upw
i k
_

t
2

k

N

,=
K

k
: [a
k
a
k
]
_
u
cen
k
, u
upw
k
_
+
1
2

k
f

k
u

_
u
cen

k
, u
upw

k
_
,
(7.17)
65
always with the provisio that, when in the last boundary term u

k
= u

, the corre-
sponding contribution is absent since it is also f

k
f

= f (u

k
) f (u

) = 0.
7.5 Limiters
Without entering into the details, we list here the most common limiters that are
currently employed in the numerical simulations. We give rst the form of the
limiter function that depends on only a single variable ()
() =
_

_
|| +
1 +||
van Leer
max(0, min(1, )) minmod
max(0, min(1, 2), min(, 2)) superbee
max(0, min((1 +)/2, 2, 2)) monotonized central
However, as already anticipated, these functions cannot be used directly in this
form, since the variable = u
upw
/u
cen
is undened when u
cen
= 0, which
occurs not rarely. It is therefore convenient to replace the standard limiters ()
by their counterpart (a, b) function of two variablesas suggested by Rebay.
We give the list of the limiters in this safely implementable form
(a, b) =
_

_
a |b| +|a| b
|a| +|b| +
van Leer
1
2
[sgn(a) +sgn(b)] min(|a|, |b|) minmod
1
2
[sgn(a) +sgn(b)]
max
_
min(|a|, 2|b|), min(2|a|, |b|)
_ superbee
max
_
0, min((a +b)/2, 2a, 2b)
_
+min
_
0, max((a +b)/2, 2a, 2b)
_ monot. cent.
max
_
0,
(a +b) ab
a
2
+b
2
+
_
Van Albada
with = 10
8
, for instance.
To conclude this introduction to the high-resolution schemes for scalar equa-
tions, it is worthwhile to point out an important difference existing between nite
volumes and nite elements as far as the discretization of the time-derivative is
concerned. In nite volume schemes, the discrete equation for the i -th degree
of freedom includes the time derivative of the unknown associated with node i
only. As a consequence, the discrete equations are all decoupled from each other
66
whenever an explicit time integration scheme is considered. On the other hand,
in the nite element discretization, the introduction of the consistent mass matrix
leads to a coupling of the equations, even if an explicit time-integration procedure
is considered. This intrinsic impliciteness of the scheme affords the high phase-
speed accuracy of Galerkin and TaylorGalerkin schemes with respect to nite
differences and nite volumes.
Under a different point of view, we can regard the mass matrix as the projecting
operator of the solution from time t
n
to time t
n+1
; the projection is taken in the
piecewise linear space of nite elements, thus it will lead to oscillations when
discontinuous solutions are considered (Gibbs phenomenon). However, if the
consistent mass matrix is substituted by its diagonally lumped counterpart, that
is, by the approximation

kN
i
M
i k
du
k
dt
L
i
du
i
dt
, (7.18)
this difculty is overcome, although at the price of a severe reduction of time accu-
racy. In the following, the high accuracy of the standard nite element approxima-
tion will be tradeoff with the possibility of computing discontinuous solutions, and
we will therefore introduce the mass lumping approximation (7.18) in the nite
element schemes.
67
Chapter 8
Thermodynamics of gases
8.1 Introduction
Before considering the Euler equations of gasdynamics for developing numerical
solution schemes, it is necessary to introduce the mathematical concept and rela-
tions that are needed to describe the thermodynamic properties of a gas. This is
the main purpose of the present chapter, where different physical models of gases
will be presented. A possible source of confusion in the denition of a gas model
in the context of the Euler equations is the circumstance that only one thermody-
namic equation of state is needed to close the hyperbolic system of gasdynamics
equations. This is to be contrasted with the well konwn fundamental thermody-
namic principle that two equations of state are necessary to completely specify the
thermodynamic properties of a uid, at least in the case of a simple (monocom-
ponent) uid system, see, e.g., Callen [2], that we are presently interested with.
In this connection, it is worth to remind that a second equation of state, although
not explicitely used in the solution of the Euler equations, plays actually a role in
selecting the physically relevant unique solution within the innite set of weak so-
lutions of the nonlinear hyperbolic problemthrough the intermediate of the entropy
condition.
In this chapter, following Callen [2] and Galgani and Scotti [8] the properties
of a thermodynamic system are dened focussing on the central role played by its
fundamental thermodynamic relation. This relation is the starting theoretical point
to introduce the intensive thermodynamic variables and to obtain the equations of
state of the system, both of which in this approach appear to be derived concepts,
as recalled here. Ideal and nonideal gas models will be introduced, allowing
for different behaviours of the specic heat, usually denoted as polytropic and
nonpolytropic.
The gas models are described initially adopting the extensive variable N that
68
represents the mole number, as typical in the treatment of thermodynamic systems.
Subsequently, the mass M of the system is adopted since this variable is more
suitable in situations where the variable mass density is the leading lady and specic
quantities (i.e., per unit mass) are present, as it happens in a uid dynamic context.
Anticipating its use in the Euler equations written for the conservative variables,
the equation of state for pressure expressed in terms of specic internal energy and
mass density is used to build a pressure function dependent on the conservative
variables, as required for closing the system of the gasdynamic equations. In
the same vein, expressions for the sound speed in the gas will be derived also
as a function of the conservative variables, to be exploited in Roe linearization
procedure for arbitrary gas models addressed in a later chapter. In this connection,
the pressure function of the conservative variables is also examined to identify the
conditions assuring its rst-order homogeneity. This mathematical condition is
necessary for the ux functions of Euler equations to satisfy the same homogeneity
property which is relevant for the linearization procedure of this hyperbolic system.
8.2 Denition of the gas properties
For a mono-component simple uid [2] in thermodynamic equilibrium, all the
thermodynamic properties of the system are dened by a fundamental relation
which species an extensive quantity of the system, the entropy S, in terms of the
extensive variables (internal) energy
1
E, volume V and number of moles N of the
system,
S = S(E, V, N). (8.1)
The fundamental laws of thermodynamic theory are summarized in the following
mathematical properties of the function S(E, V, N), stated by Galgani and Scotti
[8]:
Homogeneity of degree one (or rst-order) with respect to its variables,
Strictly monotonicity with respect to the energy,
Superadditivity in the sense specied in [8].
The equations of state are obtained by taking the partial derivatives of the fun-
damental relation and by assuming these functions as denition of the intensive
variables, for details see Callen. It follows that the equations of state are not in-
dependent from each other. In fact, the derivatives of the equations of state have
1
The meaning of symbol E in this section, but only in this section, is different from that in the
rest of the paper. Here, E is the internal energy of the system, while in the rest of the paper E is
the internal energy density (per unit volume) of the gas, cf. (9.6).
69
some relationships between them which are a direct consequence the theorem on
the equality of mixed partial derivatives (Schwartz theorem). On the other hand, for
the simple thermodynamic systemconsidered here, of the three (compatible) equa-
tions of state only two are necessary and sufcient to reconstruct the fundamental
relation by means of the Gibbs-Duhem equation [2].
With the aid of the basic mathematical tools of thermodynamics, we are in
place to dene the gases we are interested in for our uid dynamic purposes. Let
us rst introduce the standard denition of an ideal gas:
Denition 8.2.1 A gas is said to be ideal if its pressure P, temperature T and
volume V satisfy the following equation of state
P =
NRT
V
(8.2)
where R is the universal constant of gases.
2
2
A more general equation of state that corresponds to a gas model taking into
account interatomic or intermolecular long-range attraction and short-range repul-
sion is given by the so-called van der Waals uid model [27], for which we have
the following denition,
Denition 8.2.2 We call a gas a vander Waals gas if the thermodynamic variables
P, V and T satisfy
P =
NRT
V bN

aN
2
V
2
(8.3)
where a and b are constants dependent on the gas. 2
The van der Waals uid model is a good approximation of a gas with atomic or
molecular interaction always under the assumption of elastic collisions. It is im-
portant to emphasize that the attributes ideal and van der Waals, as specied in
the previous denitions, are not sufcient to characterize a gas completely insofar
as a second equation of state has not yet been given. However, as anticipated, this
lacking equation of state cannot be given in a completely independent manner from
the rst one, since both must be a consequence of one and the same fundamental
relation S(E, V, N). To discover the form of the second equation of state, namely
E = E(T, V, N), compatible with the pressure equation of state P = P(T, V, N)
we proceed as in Example 4 of section 12.8, page 693, in Calculus of Several Vari-
ables, by Robert Adams, Sixth Edition 2006, [or page 189 of the Italian translation
Calcolo differenziale 2, Funzioni di pi variabili, CEA, 2007].
2
The numerical value of the universal gas constant is R = 8.314J/mole K. This fundamental
constant is given by relation R = k
B
N
A
, where k
B
= 1.3807 10
23
J/K is Boltzmann constant
and N
A
= 6.022 10
23
mole
1
is the Avogadro number.
70
The internal energy E and the entropy S of the system are thermodinamic
quantities that depend on the state variables and hence may be expressed as function
of any two variables, of T, V and T, together with the third variable N. Let us
choose T, V and N for the independent variables and so write E = E(T, V, N) and
S = S(T, V, N). By denition of the thermodynamic quantities, the innitesimal
change of energy is expressed in terms of changes of the independent variables by
the differential equation
dE = T dS P dV +dN,
where is the electrochemical potential. To deduce the formof the energy equation
of state E = E(T, V, N) implied by the idel gas pressure equation P(T, V, N) =
NRT/V we calculate the differentials dE and dS and substitute them into the
differential relation to obtain
E
T
dT +
E
V
dV +
E
N
dN
= T
_
S
T
dT +
S
V
dV +
S
N
dN
_
P dV +dN.
Divide by T and substitute RN/V for P/T (fromthe equation of state) and collect
the coefcients dT, dV and dN on opposite sides of the equations to get
_
1
T
E
T

S
T
_
dT +
_
1
T
E
V

S
V
+
RN
V
_
dV
=
_

1
T
E
N
+
S
N
+
_
dN
Since T, V and N are independent variables, the three coefcients of the innites-
imals must vanish. Hence, in particular,
S
T
=
1
T
E
T
,
S
V
=
1
T
E
V
+
RN
V
.
Now differentiate the rst equation with respect to V and the second with respect
to T. Using the equality two mixed partial derivatives for both S(T, V, N) and
71
E(T, V, N), we obtain the desired result:

V
_
1
T
E
T
_
=

2
S
V T
=

2
S
T V
=

T
_
1
T
E
V
+
RN
V
_
1
T

2
E
V T
=
1
T
2
E
V
+
1
T

2
E
T V
1
T
2
E
V
= 0.
It follows that E/V = 0, so E is independent of V, namely E(T, V, N) =
E(T, N). Moreover, since E and N are extensive variables while T is intensive,
the energy equation of state of the ideal gas must be of the form
E(T, V, N) = f (T) N, (8.4)
where f (T) is an arbitrary function, but for restrictions implied by the thermody-
namic stability, see below.
By the same argument, the second compatible equation of state for pressure
function P(T, V, N) = NRT/(V bN) aN
2
/V
2
, of the van der Waals gas is
found to be in the form
E(V, T, N) = f (T) N
aN
2
V
, (8.5)
whith still f (T) an arbitrary function.
In both cases, the arbitrary function f (T) is related to the molar heat capacity
at constant volume
c
v
=
1
N
_
d

Q
dT
_
V,N
=
1
N
_
E
T
_
V,N
, (8.6)
by the obvious relation
f

(T) = c
v
(T) (8.7)
and the thermodynamic stability condition c
v
> 0, see [2], implies f

(T) > 0.
From the physical viewpoint, the form of function f (T) depends on the ro-
tational and vibrational internal degrees of freedom of the molecules of the gas.
The contribution of these physical effects to the internal energy of the gas can be
determined theoretically by means of arguments of statistical mechanics and will
be discussed later. Of course, once the function c
v
(T) of the molar heat capacity
at constant volume has been established either theoretically or by experimental
measurements, the function f (T) will be dened by direct integration:
f (T) =
_
T
T
0
c
v
(T) dT. (8.8)
72
The various possible forms for the function f (T) lead to a further specication
of the thermodynamic properties of the gas according to the following denition
Denition 8.2.3 A gas is said to be polytropic if
f

(T) = constant (8.9)
where the constant is by denition the molar specic heat at constant volume c
v
.
When the function f

(T) is not a constant, the gas is said to be nonpolytropic. 2
In a uid dynamic context the thermodynamic variables of interest are the
specic quantities, i.e, per unit mass, instead of those per unit mole. It is therefore
convenient torewrite all the previous equations of state bysubstitutingmole number
N of the system with its total mass M, the two quantities being proportional to
each other according to the relation N = M/
0
, where
0
is the molecular weight
of the gas. A simple calculation gives the new equations of state for the pressure,
by introducing the gas-dependent constant
R = R/
0
, (8.10)
and = M/Vthe mass density. The new equations of state of the ideal and van
der Waals gas are
P(T, ) = RT and P(T, ) =
RT
1 b
a
2
, (8.11)
respectively, where the two new van der Waals constants a and b are obtained from
the old ones a and b by the ratios a/
2
0
and b/
0
. Similarly, the second equation of
state compatible respectively with the ideal gas and the van der Waals gas becomes
e(T) = f (RT) and e(T, ) = f (RT) a, (8.12)
where e = E/M is the internal energy per unit mass and the new function f
is obtained from the previous one by the scaling it, f /
0
, while its argument is
written as RT for later convenience. Correspondingly, the quantity c
v
is redened
in the following per-unit-mass version
c
v
=
1
M
_
d

Q
dT
_
V,M
=
1
M
_
E
T
_
V,M
, (8.13)
and therefore now denotes the specic heat at constant volume.
73
8.3 Polytropic and nonpolytropic behaviour
Agas model to be completely dened requires the specication of the formof func-
tion f (RT) or some equivalent substitute function as, for instance, the specic heat
at constant volume c
v
(T) of the gas. Different gas models are therefore obtained
starting from the ideal gas and van der Waals gas depending on the assumed func-
tion c
v
(T). As anticipated, the form of the dependence on temperature of the
specic heat is determined by the internal degrees of freedom of the molecules of
the gas and can also be different depending on temperature range.
The most simplest situation is that of atomic gases which have no internal
degrees of freedom so that their energy is due only to the translationary motion
of the atoms, possible along the three spatial directions. According to the energy
equipartition theorem of kinetic theory of gases, each degree of freedom gives a
constant contribution equal to R/2 to the specic heat at constant volume, so that
c
atomic
v
= 3R/2. All atomic gases are therefore represented correctlty as polytropic
ideal gases with internal specic energy given by e
atomic
(T) =
3
2
RT. This relation
is often written introducing the following important parameter for polytropic ideal
gas

def
=
c
P
c
v
, (8.14)
called specic heat ratio. For polytropic ideal gases also the specic heat at
constant pressure is constant and it can be shown that c
P
= c
v
+ R, a relation
attribute to Meyer. Substiting Meyer relation into the denition of constant
gives immediately
= 1 +
R
c
v
c
v
=
R
1
, (8.15)
so that the specic internal energy of any polytropic ideal gas can be written also
as
e(T) =
RT
1
. (8.16)
For atomic gases
atomic
= 5/3.
Let us nowconsider molecular gases. Beyondthe translatorymotioninthe three
spatial directions each molecule can rotate around some axes through its center and
the atoms of the molecule can performoscillatory motions around their equilibriun
positions. These internal degrees of freedom of the molecule can contribute to the
internal energy of the gas, but classical mechanics and quantum mechanics give a
very different description of these contributions.
From the classical viewpoint, any degree of freedom gives a xed contribution
to the specic heat at constant volume, which is constant, does not depend on
74
temperature and is equal to R/2. Therefore, according to classical mechanics the
molecular gases are described as polytropic ideal gases, similarly to the atomic
ones, but only with a different value of the constant . The latter depends on
the number of atoms in the molecule and also on whether its shape is linear or
non-linear since the number of possible rotations is different in the two cases:
linear molecules, including the diatomic ones, can rotate only around two axes
perpendicular the molecular axis while non-linear molecules can rotate around
three perpendicular axes.
This difference in the possible rotational motions of the molecule has also
an implication on the number of degrees of freedom available for the molecular
vibrations. In fact, assuming a molecule with #atoms, the number of vibrational
degress of freedom is obtained by subtracting form the total number 3#atoms of
degrees of freedoms of all atoms in the molecule the number of degrees of freedom
for the translations and rotations of the rigid molecule as a whole. The latter can
be 2 or 3 depending the linear or non-linear shape of the molecule. In short:
#vibrational d.o.f. =
_
3#atoms 5 linear molecule
3#atoms 6 non-linear molecule
In other terms, a linear molecule has one more vibrational mode than a non-linear
molecule with the same number of atoms.
Each vibrational mode is characterized by the presence of kinetic energy and
of potential (elastic) energy and both energies contributes to the internal energy of
the gas. Therefore, always according to the classical equipartition theorem each
vibrational mode gives a contribution to c
v
equal to R/2 + R/2 = R.
In conclusion, within the description of classical mechanics, the specic heat
of atomic and molecular gases is given by
c
classical
v
=
_

_
3
2
R atomic gas
1
2
(6#atoms 5) R gas with linear molecules
3(#atoms 1) R gas with non-linear molecules
The corresponding constant is

classical
=
_

_
5/3 atomic gas
3(2#atoms 1)
6#atoms 5
gas with linear molecules
3#atoms 2
3(#atoms 1)
gas with non-linear molecules
75
This is the description provided by the classical mechanics. The values c
classical
v
just established are in disagreement with the experimental results, except at very
high temperature. The measurements show that the specic heat of gases expe-
riences strong variations with temperature until it eventually saturates to the its
classical values c
classical
v
. The explanation of this behaviour is provided by quan-
tum mechanics and was actually one important experimental verication of its
validity.
Quantum mechanically, molecular rotations and atomic vibrations within a
molecule are discontinuous. This means that the states of possible rotations for the
molecule and of possible oscillations of its atoms are discrete, with a nite separa-
tion between them. In other words, molecular rotations and atomic vibrations can
exist at the microscopic level only for some discretely dened values of angular
velocity and oscillatory speed. The consequence of the quantum discretization is
that the equipartition theorem is no more valid and the rotation and vibration of
the molecule may not contribute to the internal energy of the gas, depending on
its temperature. In fact, when the temperature is such that k
B
T is much smaller
than the energy separation between the rst quantumrotational mode and the mode
without rotation, molecular rotations are impossible and do not contribute to the
gas energy. Only when k
B
T becomes comparable with the aforementioned energy
gap, the quantum rotations of the molecule can give a contribution to the energy
and hence to the specic heat. For higher temperatures, the molecular rotations
become fully accessible and their discrete nature has no more consequence so that,
in this temperature range, the contribution of rotation kinetic energy to the internal
energy is as predicted by classical statistical mechanics. Therefore, according to
this picture, the contribution to the specic heat of each molecular rotation is a
temperature dependent function that goes from zero to the limit value R/2 pre-
dicted by the classical equipartition theorem. The exact analytical expression of
this function is provided by quantum statistical mechanics, see, i.e., [13], but is
not reported here since it is not necessary for the following reason. The values of
temperature at which the transition occurs, called rotational temperatures, are
characteristic of the considered molecule since they are proportional to the mo-
ment of inertia of the molecule around the principal axes of rotation, and their
typical values for most gases are much lower than the ambient temperature. Thus,
in virtually all gasdynamic applications it can be assumed that the contribution of
molecular rotations to the specic heat is as described classically.
Coming now to the vibrations of the atoms within the molecule, they may not
contribute to the gas internal energy by quantum mechanical reasons similar to
those explained for the molecular rotations, since also the atomic oscillations are
discretized. In polyatomic molecules the possible oscillations can be described in
terms of elastic normal modes. Their dynamics is however governed by quantum
mechanics and also the energy levels of the quantum oscillators are discrete, simi-
76
larly to quantum rotation. Therefore, the contribution of each normal mode to the
energy and to the specic heat is similarly dependent on temperature with a tran-
sition from zero to the limit value R/2 + R/2 = R predicted for each oscillation
mode by the equipartition theorem. The transition temperatures for the quantum
activation of vibrational modes are called vibrational temperatures. They depend
on the elastic structure of the molecule and their values for many gases relevant
for applications are larger than 700 K. It follows that in the range of temperature
300 K < T < 700 K it is possible to consider the molecule as perfectly rigid with
no atomic oscillation permitted and with the rotation fully described by the clas-
sical mechanics. In this temperture interval the molecular gas is well represented
by a polytropic ideal gas with a value of that is easily deduced by the counting
of the allowed classical contributions to the specic heat. Such an approximation
can be called intermediate and the corresponding value of c
v
is
c
inter
v
=
_

_
3
2
R atomic gas
5
2
R gas with linear molecules
3R gas with non-linear molecules
while the value is

inter
=
_

_
5/3 atomic gas
7/5 gas with linear molecules
4/3 gas with non-linear molecules
To go beyond intermediate approximation, it is necessary to know the function of
the vibrational contribution to the specic heat provided by the quantum statistical
mechanics. Lukily, the exact analytical solution is found to be the relatively simple
expression
c
vib
v
(T) = R
_
T

T
_
2
e
T

/T
_
e
T

/T

_
2
(8.17)
where T

is the vibrational temperature of the considered normal mode of oscilla-


tions. Assuming now that the molecule has N

vibrational modes, the expression


of the specic heat of the gas including all possible contributions (translational,
rotational and vibrational) to the interna enery of the gas is
c
v
(T) = R +
N
vib

=1
_
T

T
_
2
e
T

/T
_
e
T

/T

_
2
. (8.18)
where the parameter can assume only the following two values
=
_
_
_
3
2
+
2
2
=
5
2
= 2.5 linear molecule
3
2
+
3
2
= 3 non-linear molecule
77
since the nonvibrational part of the internal energy includes always the same trans-
lational kinetic energy while the rotational kinet energy is different for a linear or
nonrectlinear molecule. Agas model with the specic heat (8.17) is nonpolitropic.
We notice in passing that when a vibrational mode is multiple (degeneracy), the
summation in expression (8.17) can be limited only to the terms associated to the
different eigenvalues by including a factor equal to their molteplicity.
8.4 Pressure function of the conservative variables
It is now useful, as a step toward the analysis of the the Euler equations of gasdy-
namics, to introduce the pressure as a function of the conservative variables of the
hyperbolic system. This function is obtained from the thermodynamic equation
of state P = P(e, ) by composition of function. As a consequence the partial
derivatives of the newpressure functions can be obtained formthose of the function
P = P(e, ) by means of the chain rule. On the other hand, since this equation
of state depends on only two variables the three partial derivatives of the pressure
function of the conservative variables are not independent and one of them can
be expressed in terms of the other two. Finally, for a subsequent application to
Roe linearization of the Euler equations, it is interesting to deduce the thermody-
namic condition under which the pressure function of the conservation variables
is a homogeneous function of degree one.
Let us consider a pressure equation of state in the form P = P(e, ) and
let u = (, m, E
t
)
T
be the set of conservative variables for the one-dimensional
Euler equations of gasdynamics. The pressure as a function of the these dynamical
variables will be dened by
(u)
def
= P
_
E
t

m
2
2
2
,
_
. (8.19)
The partial derivatives of (u) with respect to the conservative variables u can be
regrouped to dene the gradient:

u
(u) =
_

(u),
m
(u),
E
t (u)
_
T
.
78
By means of the chain rule, we have

(u) = P

_
E
t

m
2
2
2
,
_

_
E
t

2

m
2

3
_
P
e
_
E
t

m
2
2
2
,
_
, (8.20)

m
(u) =
m

2
P
e
_
E
t

m
2
2
2
,
_
, (8.21)

E
t (u) =
1

P
e
_
E
t

m
2
2
2
,
_
, (8.22)
where the subscripts to the function P(e, ) denote partial derivative with respect to
its variables e and . The last relation can be inverted to obtain the partial derivative
P
e
(e, ) in terms of
E
t(u). Then, substiting into the rst relation, also the partial
derivative P

(e, ) can be expressed in terms of

(u) and
E
t (u), to give
P
e
(u) =
E
t (u), (8.23)
P

(u) =

(u) +
_
E
t

m
2

2
_

E
t (u). (8.24)
These relations are useful to dene the speed of sound c as a function of the
conservative variables.
Form the fact that the three-variable function (u) is dened from a function
with only two independent variables it follows that the three partial derivatives
of (u) are not independent. In fact from equations (8.21) and (8.22) we obtain
immediately

m
(u) =
m

E
t (u). (8.25)
Coming nowto the rst-order homogeneity of the uxes of the Euler equations,
this property depends on the behaviour of the pressure function since all the terms
different from pressure contributing to the Euler uxes are already homogeneous
of degree one. As a consequence, the ux function is homogeneous of degree
one if and only if the scalar function (u) satises the condition of rst-order
homogeneity
u
u
(u) = (u). (8.26)
Using the expressions (8.20)(8.22) of the partial derivatives of (u) it is imme-
79
diate to see that the scalar product of the Euler condition above reads
u
u
(u) =

(u) +m
m
(u), +E
t

E
t (u)
= P


_
E
t

m
2

2
_
P
e

m
2

2
P
e
+
E
t

P
e
= P

+
_

E
t

+
m
2

2

m
2

2
+
E
t

_
P
e
= P

Thus the pressure function (u) is rst-order homogeneous only provided the
pressure equation of state P = P(e, ) satises the differential equation
P

(e, ) = P(e, ). (8.27)


Similar relations among the partial derivatives can be established starting from
the alternative equation of state for pressure P =

P(E, ). For instance, from the
denition

(u) =

P
_
E
t

m
2
2
,
_
, (8.28)
one obtains the following expressions of the partial derivatives

P
E
(E, ) and

(E, ) in terms of
E
t

and

P
E
(u) =
E
t

(u), (8.29)

(u) =


(u)
m
2
2
2

E
t

(u). (8.30)
8.5 Speed of sound
The relations obtained in the previous section allow us to express the local speed
of sound of the uid,
[c(s, )]
2
= P

(s, ), (8.31)
where s is the specic entropy (per unit mass), as a function of the thermody-
namic and uid dynamic quantities of interest. The speed of sound c is needed to
completely dene the eigenstructure of the Euler equations of gasdynamics.
80
By using the relations given in section 8.4 it is possible to show that the speed
of sound can be expressed in the three following equivalent forms
[c(e, )]
2
=
P(e, )

2
P
e
(e, ) + P

(e, ), (8.32)
[c(E, )]
2
=
E + P(E, )

P
E
(E, ) + P

(E, ), (8.33)
[c(u)]
2
=

(u) +
_
E
t
+(u)

m
2

2
_

E
t (u). (8.34)
8.6 Some particular gas models
In the present section, different gas models are presented. The gases introduced
here are characterizedfollowingthe ideal/vander Waals andpolytropic/nonpolytropic
specications established in section 8.2. For each gas model, the corresponding
equation of state P(e, ) is presented, together with the explicit expression of the
speed of sound c. The condition (8.27) for the function P(e, ) will be checked
to acertain whether the pressure function (u) of the conservation variables is
rst-order homogeneous for each particular gas model.
8.6.1 Polytropic ideal gas
The simplest model presented is the polytropic ideal gas, whose thermodynamic
properties are completely dened by the pair of equations of state:
P(T, ) = RT and e(T) =
RT
1
. (8.35)
By substituting RT from the second equation of state into the rst equation, the
pressure P is obtained as a function of the specic variables e and
P(e, ) = ( 1) e. (8.36)
Now, from equation (8.32), the sound speed of the polytropic ideal gas is found to
be
[c(e)]
2
= ( 1) e and [c(T)]
2
= RT. (8.37)
The condition (8.27) on the pressure function (8.36) associated with the consid-
ered model gives the relation ( 1)e = ( 1) e, which is identically satised.
Thus, the Euler uxes for a polytropic ideal gas are always homogeneous functions
of degree one.
81
8.6.2 Nonpolytropic ideal gas
When function f (RT) is assumed to be not linear so that the specic heat at con-
stant volume c
v
is not constant, one obtains the gas model denoted as nonpolytropic
ideal gas, which dened by the two equations of state
P(T, ) = RT and e(T) = f (RT). (8.38)
Since, for stability, f

> 0, the function f is invertible and so the pressure P(e, )
is given by
P(e, ) = f
1
(e) = g(e) , (8.39)
having introduced the inverse function g = f
1
. The speed of sound of the
nonpolytropic ideal gas reads
[c(e)]
2
= g(e)[1 +g

(e)]. (8.40)
By the differentiation rule of the inverse function we have g

(e) = [ f
1
(e)]

=
1/f

(RT). But f

(RT) = c
v
(T), so the sound speed can also be written as a
function of temperature by the alternative relation
[c(T)]
2
=
_
1 +
1
c
v
(T)
_
RT, (8.41)
involving directly the specic heat at constant volume c
v
(T) instead of the inverse
function g(e).
To check condition (8.27) assuring rst-order homogeneity of (u), substi-
tute P(e, ) = g(e) into this relation, which yields g(e) = g(e) , which
is identically satised. Hence the nonpolytropic ideal gas model leads to rst-
order homogeneous Euler uxes f(u). Therefore, the rst-order homogeneity is
a consequence of the ideal character of the gas, irrespective of the polytropic or
nonpolytropic property.
8.6.3 Polytropic van der Waals gas
The polytropic van der Waals gas model represents a gas with a constant specic
heat at constant volume c
v
but that accounts for interatomic or intermolecular
forces, which are instead fully neglected in the ideal gas. The two equations of
state characterizing the polytropic van der Waals gas are
P(T, ) =
RT
1 b
a
2
and e(T, ) =
RT
1
a, (8.42)
82
and the equation of state giving the pressure as a function of e and reads
P(e, ) = ( 1)
e +a
2
1 b
a
2
. (8.43)
The critical value of the density, pressure andtemperature are givenbythe following
relations

c
=
1
3b
, P
c
=
a
27b
2
, (RT)
c
=
8a
27b
. (8.44)
The speed of sound is obtained from equation (8.32), which gives
[c(e, )]
2
= ( 1)
e +a
(1 b)
2
2a. (8.45)
Thus, in the case of polytropic van der Waals the sound speed c depends both on e
and , while in the ideal gas models c is a function only of e. By eleminating the
specic energy e in favor of the temeprature T, the sound speed of the polytropic
van der Waals gas can also be written as
[c(T, )]
2
=
T R
(1 b)
2
2a. (8.46)
To verify whether the pressure function (u) of the conservative variables is
rst-order homogeneous by checking condition (8.27), we rst compute

P(e, )

= ( 1)
e +2a
2
ab
3
(1 b)
2
2a
2
.
This quantity is different from the pressure function (8.43) except when a = 0 and
b = 0, in which case the van der Waals gas model degenerates to the polytropic
ideal gas. Consequently, the pressure function (u) for a polytropic van der Waals
gas is not homogeneous.
8.6.4 Nonpolytropic van der Waals gas
The last model considered is the nonpolytropic van der Waals gas. The equations
of state characterizing the thermodynamic properties of this gas model are written
as
P(T, ) =
RT
1 b
a
2
and e(T, ) = f (RT) a, (8.47)
where f (RT) is the function related to the specic heat at constant volume c
v
(T)
in the known manner. Introducing again the inverse function g = f
1
, we have
P(e, ) =
g(e +a)
1 b
a
2
(8.48)
83
and
[c(e, )]
2
=
[1 + g

(e +a)]g(e +a)
(1 b)
2
2a. (8.49)
By eliminating the specic energy e in favor of the temperature T thanks to the
differentiation rule for the inverse funtion, we can obtain the alternative expression
for the sound speed of the nonpolytropic van der Waals gas
[c(T, )]
2
=
_
1 +
1
c
v
(T)
_
RT
(1 b)
2
2a (8.50)
that involves directly the specic heat at constant volume c
v
(T) in place of the
inverse function g(e).
To check the homogeneity condition (8.27), we compute

P(e, )

=
g(e +a)
(1 b)
2
+
ag

(e +a)
2
1 b
2a
2
.
This expression is equal to the pressure function (8.48) only provided that a = 0
and b = 0. Therefore, also the nonpolytropic van der Waals gas fails to give rst-
order homogeneous uxes, as its polytropic counterpart. In conclusion both van
der Waals gas models are characterized by a pressure function of the conservative
variables which is not rst-order homogeneous. This property will be found to be
useful in the linearization of Euler equations of gasdynamic for nonideal gases.
84
Chapter 9
Euler equations of gasdynamics
In the present chapter we introduce the equations governing the ow of a com-
pressible, but nonviscous, uid. This set of equations is denoted as the system
of Euler equations of gasdynamics. The fundamental equations of uid dynamics
are actually the NavierStokes equations for viscous compressible uids, which
include the description of viscous effects as well as of heat condution phenomena.
Our interest is however focussed on the study of numerical schemes for invisicd
ows so that the set of the full NavierStokes equations fall outside the scope of
our analysis. We will present the Euler equations for one-dimensional and mul-
tidimensional ows. Moreover, the equations will be stated initially in terms of
the physical variables, density, velocity and total energy density, and will be sub-
sequently expressed in terms of the so-called conservation variables, namely, the
density of mass, momentumand total energy. Finally, the hyperbolic systemof the
Euler equations for the conservation variables will be written in quasilinear form
and the corresponding eigenstructure of the system will be derived.
9.1 Conservation laws of gasdynamics
The hyperbolic problem of our main concern is that associated with the equations
of gasdynamics, namely the Euler equations for compressible nonviscous ows.
These equations are a mathematical statement of the conservation laws for the mass
of a uid and of balance laws for momentum and total energy of the uid. In the
one-dimensional case, these equations take the form
_

t
+
x
(u) = 0,

t
(u) +
x
(u
2
+ P) = 0,

t
E
t
+
x
[(E
t
+ P) u] = 0,
(9.1)
85
where (x, t ) is the mass density, u(x, t ) the velocity and E
t
(x, t ) the total energy
density. The pressure P is given as a known function of two thermodynamic
variables, typically the specic internal energy e and the mass density: P =
P(e, ). Such a function is called an equation of state and its precise formdepends
on the thermodynamic properties of the gas under examination, as discussed in the
previous section.
For multidimensional ows, the Euler equations of gasdynamics assume the
following form
_

t
+ (u) = 0,

t
(u) + (u u + P) = 0,

t
E
t
+ [(E
t
+ P) u] = 0,
(9.2)
where denotes tensor product. The meaning of the tensor product uu in any
orthogonal coordinate system, including Cartesian, cylindrical an spherical ones, is
provided by the relation uu = (u )u+u (u) = (u )(u)+u u.
The relationship between the total energy density E
t
and the internal specic energy
e is given by
E
t
= e +
1
2
|u|
2
. (9.3)
9.2 Euler equations in one dimension
9.2.1 Conservation variables
For developing numerical schemes for nonlinear hyperbolic systems stemming
from conservation laws it is necessary to use as unknonws the variables associated
with the quatities that are conserved. For the gasdynamic equations the conserva-
tion variables are density of mass , of momentumm = u and of the total energy
E
t
. In terms of these unknowns, the Euler equations for one-dimensional ows
becomes
_

t
+
x
m = 0,

t
m +
x
(m/
2
+ P) = 0,

t
E
t
+
x
[(E
t
+ P) m/] = 0,
(9.4)
where m = u is the density of momentum in direction of x axis. The relation
between the energies assumes the form
E
t
= E +
m
2
2
, (9.5)
86
and the pressure equation of state for closing the system will be used in the form
P = P(e, ) = P
_
E
t

m
2
2
2
,
_
. (9.6)
The nonlinear system of conservation laws (9.4) can be written in quasilinear form
by introducing following vector of the conservation variables:
u =
_
_

m
E
t
_
_
, (9.7)
belonging to the domain u [R
+
R R
+
], and the corresponding ux vector
f(u) =
_
_
_
_
_
m
m
2

+ P
(E
t
+ P)
m

.
_
_
_
_
_
(9.8)
With these denitions, the system of conservation laws can be written compactly
as

t
u +
x
f(u) = 0. (9.9)
9.2.2 Quasilinear form
By introducing the Jacobian matrix
A(u) =
f(u)
u
, (9.10)
the system of Euler equations in one dimension can be recast in the quasilinear
form:

t
u +A(u)
x
u = 0. (9.11)
The explicit expression of the Jacobian A(u) for the Euler equations requires to
know the dependence of the pressure function P on the unknown variables. As a
consequence, the complete denition of the problemrequires to specify the relation
between the pressure and the conservation variables. With the thermodynamic
equation of state assumed in in the form P = P(e, ), the closure of the hyperbolic
system will be provided by introducing the composed function of the conservation
variables:
(u) = (, m, E
t
) P
_
E
t

m
2
2
2
,
_
. (9.12)
87
Accordingly, the ux functions of the one-dimensional Euler equations reads
f(u) =
_
_
_
_
_
m
m
2

+(u)
_
E
t
+(u)
_
m

,
_
_
_
_
_
(9.13)
and the Jacobian matrix of the ux will be
A(u) =
_
_
_
_
_
_
_
_
0 1 0

m
2

2
+

(u)
2m

+
m
(u)
E
t (u)
m

E
t
+(u)

(u)
_
E
t
+(u)

+
m

m
(u)
m

_
1 +
E
t (u)
_
_
_
_
_
_
_
_
_
(9.14)
9.2.3 Eigenstructure
with the associated eigenvalues

1,3
(u) =
m

c(u),
2
(u) =
m

, (9.15)
where c(u) is given in (8.34). The corresponding right eigenvectors are
R(u) =
_
_
_
_
1 1 1
u c u u +c
h
t
uc u
2

E
t
h
t
+uc
_
_
_
_
, (9.16)
and the matrix of the left eigenrows is given by
L(u) =
1
2c
2
_
_
_
_

+uc
m
c
E
t
2(

c
2
) 2
m
2
E
t

uc
m
+c
E
t
_
_
_
_
, (9.17)
where u = m/ and c = c(u). The left eigenrows are normalized so that
L(u) R(u) = R(u) L(u) = I.
The representation of the jump u of the conservation variables in terms of
the characteristic variables is given by
v = L(u)u (9.18)
88
where P = P
r
P

= (u
r
)(u

). The systemof Euler equations is strictly


hyperbolic if c(u) = 0, i.e., from (8.34) if

(u) +
_
E
t
+(u)

m
2

2
_

E
t (u) > 0, (9.19)
that is
E + P(E, )

P(E, )
E
+
P(E, )

> 0. (9.20)
Conditions assuring strict hyperbolicity and genuine nonlinearity of the hyperbolic
system are given in [9]; here, we only want to remind that the characteristic eld
associated with the second eigenvalue is always linearly degenerate, since one has

2
(u) r
2
(u) =
_

2
,
1

, 0
_ _
1,
m

,
m
2

(u)

E
t (u)
_
T
=
m

2
+
m

2
= 0.
(9.21)
9.3 Euler equations in two dimensions
In two dimensions the Euler equations of gasdynamics are dened by the following
funtions of the tensor ux f = (f
x
, f
y
)
T
f
x
(u) =
_
_
_
_
_
_
_
_
_
m
m
2

+(u)
mn

_
E
t
+(u)
_
m

_
_
_
_
_
_
_
_
_
and f
y
(u) =
_
_
_
_
_
_
_
_
_
n
nm

n
2

+(u)
_
E
t
+(u)
_
n

_
_
_
_
_
_
_
_
_
, (9.22)
where u = (, m, E
t
)
T
[R
+
R
2
R
+
], and m = (m, n)
T
. After introducing
the two Jacobian matrices of the Cartesian components of the tensor ux
A
x
(u) =
f
x
(u)
u
and A
y
(u) =
f
y
(u)
u
, (9.23)
their explict expressions is found by direct differentiation
A
x
(u) =
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
0 1 0 0

m
2

2
+

(u)
2m

+
m
(u)
n
(u)
E
t (u)

mn

2
n

0
m

E
t
+(u)

(u)
_
E
t
+(u)

+
m

m
(u)
m

n
(u)
m

_
1 +
E
t (u)
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
(9.24)
89
and
A
y
(u) =
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
0 0 1 0

nm

2
n

n
2

2
+

(u)
m
(u)
2n

+
n
(u)
E
t (u)
n

E
t
+(u)

(u)
_
n

n
(u)
E
t
+(u)

+
n

m
(u)
n

_
1 +
E
t (u)
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
(9.25)
The two Jacobian matrices can be combined to dene the Jacobian hypermatrix of
the tensor ux
A(u) = (A
x
(u), A
y
(u))
T
, (9.26)
We now dene the p p matrix A(u, ) as
A(u, ) = A(u)
=
x
A
x
(u) +
y
A
y
(u)
=
x
A
x
(u(w)) +
y
A
y
(u(w)),
(9.27)
where in the last line the notations u(w) means that the conservative variables are
given as a function of the vector w = (, u, h
t
)
T
. This choice is made only to
simplify the notation. We have
A(u(w), ) =
_
_
_
_
_
_
_
_
_
0
x

y
0
u u +
x

u +u
x
(1
E
t ) u
y
v
x

E
t
x

E
t
v u +
y

v
x
u
y

E
t u +v
y
(1
E
t )
y

E
t
u (h
t

) h
t

x
u u
E
t h
t

y
v u
E
t u (1 +
E
t )
_
_
_
_
_
_
_
_
_
(9.28)
The eigenvalues of A(u, ) are

1,4
(u, ) =
m

c(u)||,
2,3
(u, ) =
m

, (9.29)
with c(u) given in (8.34) and the matrix of the associated right eigenvectors is
R(u, ) =
_
_
_
_
_
_
_
1 1 0 1
u c
x
u
y
u +c
x
v c
y
v
x
v +c
y
h
t
c u |u|
2

E
t
u

h
t
+c u
_
_
_
_
_
_
_
, (9.30)
where

= 0 and =

||
. (9.31)
90
The matrix of the left eigenrows is
L(u, ) =
1
2c
2
_
_
_
_
_
_
_

+c u
m
c
x

n
c
y

E
t
2(

c
2
) 2
m
2
n
2
E
t
2c
2
u

2c
2

y
2c
2

x
0

c u
m
+c
x

n
+c
y

E
t
_
_
_
_
_
_
_
. (9.32)
91
Chapter 10
Roe linearization of gasdynamic
equations
10.1 Introduction
Godunovmethodrequires the solutionof Riemannproblems for the Euler equations
at every cell interface and at each time step. Due to the nonlinear character of these
problems their solution is necessarily iterative and thus can be expensive in practice.
In general, the computational effort needed to nd the exact solution of the Riemann
problem can be high, due to its nonlinear character. However, the full knowledge
of the Riemann solution is redundant insofar as a cell average is performed by
Godunov method at the end of the time step. For this reason one can think that the
essential elements of the information can be provided also by a suitable reduction
or approximation of the Riemann problem. In other words, the exact Riemann
problem might be replaced by some simpler approximate counterpart which is
capable of respecting fundamental properties of the nonlinear governing equations
and, at the same time, of reducing the computational effort substantially.
This chapter introduces a method proposed by Roe for linearizing the Euler
equations of gasdynamics in order to replace the solution of nonlinear Riemann
problems at cell interfaces by some linear substitutive equivalent. The fundamen-
tal idea of Roe was to build the approximate linearized counterpart of the exact
Riemann problem in a manner that mantains conservation. This requirement is es-
sential for obtaining numerical solutions which are weak solutions to the nonlinear
hyperbolic system.
Aderivation of Roes matrix in Jacobian formand using conservative variables
is presented, following the analysis of [12] and [11]. In other words, the lineariz-
ing matrix is chosen to be coincident with the Jacobian matrix A(u) = f(u)/u,
with u denoting the conservative variables of Euler euqations and f the correspond-
92
ing uxes, so that the unknown of the linearization problem is assumed to be an
intermediate state.
The linearization problem is solved for different gas models, including the
original polytropic ideal gas, to outline the consequences of the thermodynamic
properties of each gas. The polytropic ideal gas will be found to represent a very
peculiar situation, brought about by the mathematically degenerate nature of its
pressure function as a function specic energy and density, so that this gas turns
out not to be the most convenient starting point for understanding Roes method.
For this reason, the analysis starts by considering gas models endowed with
more general thermodynamic properties than under the assumption of ideal gases.
The linearization for the polytropic van der Waals gas is described in detail from
[12] and [11]. For this rather general gas model it turns out that the average of
velocity and specic total enthalpy, using the square root of density as a weight,
proposed by Roe for the polytropic ideal gas are still useful. Subsequently, the
analysis of the linearization problem is specialized to the ideal gas where the Roe
averages are found to be possible only for the polytropic ideal gas but not in the
nonpolytropic case.
10.2 Principles of Roe linearization
10.2.1 Denition of Roe linearization
Within the framework of a Godunov-type solution scheme, it is necessary to intro-
duce the primitive idea of the Riemann problem for the hyperbolic system (9.9),
which means to consider the following initial condition for it:
u(x, 0) =
_
u

if x < 0
u
r
if x > 0
(10.1)
In a Finite Volume spatial discretization, for example, a Riemann problemis stated
at each interface separating e ach couple of neighboring discrete cells, taking the
nodal value of the solution as initial condition. In the Godunov method, the Rie-
mann problem is solved exactly (see [16] for a complete denition of the method)
at each cell interface and the solution obtained is then averaged in space to nd the
average cell values of the discrete solution at the end of the time step. Most of the
detailed information that is given by the exact solution of the Riemann problem is
lost in this averaging: therefore, to reduce the computational cost of the numerical
scheme, the Riemann problem can be replaced by a suitable approximation.
In the approximate Riemann solver method proposed by Roe, the Riemann
93
problem is replaced by a linearized problem as follows:
u
t
+
f(u)
x
= 0
u
t
+A

(u

, u
r
)
u
x
= 0, (10.2)
where A

(u

, u
r
) is a matrix which depends on the left and right states and which
must satisfy the constraints implied by the following denitions [21]:
Denition 10.2.1 Matrix A

(u

, u
r
) is called a Roe linearization of the hyper-
bolic system with ux f(u) and Jacobian matrix A(u) = f(u)/u if (u

, u
r
)
A

(u

, u
r
) is a mapping fromR
pp
into the set of pp matrices with the following
properties
i) A

(u

, u
r
)(u
r
u

) = f(u
r
) f(u

),
ii) A

(u

, u
r
) has real eigenvalues and a corresponding set of eigenvectors that
form a basis of R
p
,
iii) A

(u

, u
r
) A(u) as (u

, u
r
) u.
2
Condition i) assures that the resulting scheme is conservative and that it solves
the Riemann problemexactly when the right and left states are connected bya single
discontinuity satisfying the Rankine-Hugoniot jump condition [16]. In fact, if the
states u

and u connect a single propagating discontinuity, the RankineHugoniot


holds
(u
r
u

) = f(u
r
) f(u

),
where is the propagating velocity of the discontinuity. But the vector jump
u
r
u

satises condition i) which implies


(u
r
u

) = f(u
r
) f(u

) = A

(u

, u
r
)(u
r
u

)
so that u
r
u

is also an eigenvector of the linearizing matrix. Therefore a solution


to the nonlinear problem consisting of a single discontinuity is represented by the
approximate linearized problem by a (linear) discontinuity propagating with the
the same velocity of the exact solution.
Condition ii) requires that the approximating systemis solvable and hyperbolic,
while condition iii) guarantees that smooth ows are resolved correctly. Roes
matrix A

(u

, u
r
) is not uniquely dened and its actual determination is addressed
in the next two sections.
94
10.2.2 General solution of Roe linearization
In principle, the determination of a matrix A

satisfying the denition above for a


given pair (u

, u
r
) requires to evaluate all its p p elements, which are related by
the p quantitative conditions stated in i). Therefore, we have to solve a systemof p
(hopefully independent) equations in the p
2
unknowns a

i, j
(u

, u
r
), 1 i, j p,
assuming that the merely qualitative conditions ii) and iii) are satised. Such a
general solution for the linearization matrix is therefore a (p
2
p)-parameter
family of solutions. In particular, for the Euler equations in one spatial dimension,
where p = 3, we have a six-parameter family of solutions.
On the other hand, in order to fulll the qualitative condition ii) and iii) on the
functions a

i, j
(u

, u
r
), one has to select their form properly. For instance, some
rather general expressions for these functions have been proposed for the Euler
equations in [17]. In these works, the functional form of the coefcients a

i, j
is
partially xed by choosing a quasi-Jacobian form of Roes matrix.
10.2.3 Solution in Jacobian form
A very simple way of satisfying the quantitative condition i) together with the
qualitative condition ii) is to take

A coincident with the Jacobian matrix A(u) =
f(u)/u of the original hyperbolic system, evaluated at a suitable intermediate
state u, which will be determinate by the following condition
A( u(u

, u
r
))(u
r
u

) = f(u
r
) f(u

). (10.3)
where the notation u = u(u

, u
r
) emphasizes the dependence of the unknown
intermediate state on the states at the interface. If the linearizing matrix is chosen
in Jacobian form, the unknown becomes the state vector u and therefore one has
a system with a number or equations apparently equal to number unknowns, see
below. We notice in passing that the simple choice of an intermediate state in the
form
u =
u

+u
r
2
leads to a matrix that does not satisfy condition i) except in the trivial and irrelevant
case of a linear hyperbolic system, as discussed in [16].
For subsequent reference, it is convenient to state the Roe conditions for the
particular case of a solution sought for in Jacobian form:
i) A( u)(u
r
u

) = f(u
r
) f(u

) (conservation)
ii) A( u) (hyperbolicity, automatically satised)
iii) u(u

, u
r
) u as (u

, u
r
) u (consistency).
In the following only linearizations in Jacobian form will be considered.
95
10.3 Linearization of Euler equations
The Roe linearization problem for the Euler equations is obtained from condi-
tion (10.3) specialized to the ux function (9.13). The rst component of Roe
conservation condition gives the trivial equation
m = m (10.4)
which is identically satised. The second component of the condition leads to the
equation
_

m
2

2
+

( u)
_
+
_
2 m

+
m
( u)
_
m +
E
t ( u)E
t
=
_
m
2

+
_
which, introducing the gradient of the pressure function (u), can be written more
compactly as
()
_
m

_
2
2(m)
m

+
_
m
2

_
+
u
( u) u = 0, (10.5)
where writing or P is only a matter of taste since = (u
r
) (u

) =
P
r
P

= P. Finally, the third component of Roe conservation condition gives


_

m

2
_

E
t
+( u)
_
+
m

( u)
_
+
_
1

_

E
t
+( u)
_
+
m

m
( u)
_
m
+
m

_
1 +
E
t ( u)
_
E
t
=
_
m

(E
t
+)
_
.
Introducing the gradient, as in the previous equation, we obtain

m

2
_

E
t
+( u)
_
+
1

_

E
t
+( u)
_
m +
m

_
E
t
+
_
+
m

_

u
( u) u
_
=
_
m

(E
t
+)
_
.
(10.6)
The resulting system for the unknown intermediate state u = ( , m,

E
t
)
T
consists
of the two scalar equations
_

_
()
_
m

_
2
2(m)
m

+
_
m
2

_
+
u
( u) u = 0,

E
t
+( u)

_
m
m

_
+
m

_
E
t
+
_
+
m

_

u
( u) u
_
=
_
m

(E
t
+)
_
.
(10.7)
96
The set of intermediate states solution to the two conditions consists therefore in a
one-parameter family of solutions.
10.4 Solution of Euler linearization
Let us focus on the algebraic form of the system (10.7). Both equations contain
the term

u
( u) u ,
which depends on the unknowns , m and

E
t
, so that the equations are coupled
together. As a consequence, a substantial simplication in the solution of the
system resulting from the linearization problem might be obtained if such a term
would disappear. In this case, the rst equation would become a single equation
for the unknown m/ to be subsequently substituted in the second one to obtain
an explicit expression for the ratio [

E
t
+ ( u)]/ . We notice in passing that, if

u
( u) u = , then the rst equation of (10.7) becomes coincident with its
counterpart of the p-system.
It seems therefore convenient to impose the additional condition

u
( u) u = (10.8)
as a supplementary equation for dening a unique intermediate state u, simulta-
neously with the two linearization equations of (10.7). We are therefore led to
characterize a unique intermediate state u as the solution to the following system
of three equations
_

u
( u) u = ,
()
_
m

_
2
2(m)
m

+
_
m
2

_
= 0,

E
t
+( u)

_
m
m

_
+
m

(E
t
+) =
_
m

(E
t
+)
_
.
(10.9)
In order that this system is meaningful as an actual set of three equations in three
unknowns it is necessary to guarantee that the supplementary equation (10.8) is
indeed an independent equation involving all of the three unknowns and not some
relation between only two them. We will address later the issue of the precise
mathematical conditions assuring the completeness of the nal equation systemand
for the moment being we assume that equation (10.8) represents a relly independent
equation relating the three unknowns.
97
Remark 10.4.1 Meaning of the supplementary condition
The supplementary condition (10.8) involves the mechanical variables m and m
but can be shown to be of a purely thermodynamical content by the following
argument. By expanding the scalar product, equation (10.8) reads

( u) +m
m
( u) +E
t

E
t ( u) = P.
Equation (8.25), namely,
m
(u) = (m/)
E
t (u) allows to eliminate the
partial derivative
m
( u) to give

( u) +
_
E
t
m m/
_

E
t ( u) = P.
Now, the two remaining partial derivatives can be expressed in terms of the deriva-
tive of the pressure equation of state P = P(E, ) by means of relations (8.29)
and (8.30) to give

_
P

(

E, ) +
m
2

2
2
P
E
(

E, )
_
+
_
E
t
m
m

_
P
E
(

E, ) = P,
where

E is the energy density in the intermediate state dened by the solution

h
t
and :

E
def
= E
_

h
t

| u|
2
2
,
_
.
But E
t
= E +(m
2
/2), so we obtain
P

(

E, ) +E P
E
(

E, ) +
_

m
2
2
2
+
_
m
2
2
_
m
m

_
P
E
(

E, ) = P.
The expression within the square brachets vanishes since it is simply the conser-
vation condition for the momentum, so that, in conclusion:
P

(

E, ) +E P
E
(

E, ) = P.
This relation makes it explicit that the supplementary equation has a mere thermo-
dynamic content. 2
The second equation of system (10.9) is a second order equation for the un-
known ratio m/ that will be written as
(
r

)
m
2

2
2(m
r
m

)
m

+
m
2
r

m
2

= 0.
98
When
r
=

this equation reduces to a the linear equation


2(m
r
m

)
m

=
1

=r
_
m
2
r
m
2

_
whose solution is
m

=
m

+m
r
2
=r
=
u

+u
r
2
. (10.10)
By assuming that
r
=

, the second order can be rewritten as


m
2

2

2(m
r
m

m

+
m
2
r
/
r
m
2

= 0,
and its solutions are given by the reduced solution formula
_
m

_
1,2
=
m
r
m

_
(m
r
m

)
2
(
r

)
_
m
2
r

m
2

_
.
The discriminant of this equation is =
(

m
r

r
m

)
2

r
and is a perfect
square, so that there are two real distinct solutions
_
m

_
1,2
=
m
r
m

m
r

r
m

(
r

r
=

r
(m
r
m

) (

m
r

r
m

)
(
r

r
.
Using the following elementary identity

=
_

__

r
+

_
,
we obtain
_
m

_
1,2
=

_
m
r

r
_

_
m

__

r
+

.
The expressions within the parentheses can be simplied with either of the two
possible signs and this yields the two solutions:
_
m

_
1
=
m

+
m
r

r
and
_
m

_
2
=
m

m
r

r
.
99
However, only the rst solution with the positive sign satises the consistency
requirement i i i ) under the request of matchingwiththe solution m/ = (u

+u
r
)/2
in the special case

=
r
. In conclusion, the only physically admissible solution
for the intermediate state is
m

=
m

+
m
r

r
. (10.11)
In terms of the velocities u

and u
r
this solution can be written for the intermediate
velocity u m/
u = u
Roe
= u

r
u
r

r
, (10.12)
which is called a Roe-average or also a -average. It gives the speed of the unknown
state as the speed averaged between the left and right states using the square root
of density as the weight.
Once the momentum equation has been solved, the system (10.7) of Roe lin-
earization problem for the Euler equations will assume the form
_

u
( u) u = P,
m

=
m

+
m
r

r
,

E
t
+( u)

=

_
m

(E
t
+)
_

m

(E
t
+)
m
m

.
(10.13)
We note that the right-hand side of the third equation contains the ratio m/ , which
can be substituted from the solution given explicitly by the second equation. In
this it is possible to obtain the explicit solution of the third equation. To perform
this calculation more easily it is convenient to introduce the variable total specic
enthalpy h
t
as a shorthand of the ratio (E
t
+ (u))/. Thus we rst rewrite the
last term in the right-hand side of the energy equation in the following slightly
different form,

E
t
+( u)

=

_
m
_
E
t
+

_
_

_
E
t
+

_
_
m
m

,
100
we dene the enthalpy of the intermediate state by

h
t

_

E
t
+ ( u)
_
/ , so that
the equation assumes the more compact form

h
t
=

_
m h
t
_

_
h
t
_
m
m

,
Now we can substitute m/ from the second equation. A direct (although slightly
boring) calculation leads to the simple solution

h
t
=

h
t

r
h
t
r

r
, (10.14)
which can be reverted in the expression for the original unknowns

E
t
+( u)

=
E
t

+(u

+
E
t
r
+(u
r
)

r
.
Thus the nal form of the system of three equations for the intermadiate state
u = ( , m,

E
t
) reads
_

u
( u) u = P,
m

=
m

+
m
r

r
,

E
t
+( u)

=
E
t

+(u

+
E
t
r
+(u
r
)

r
.
(10.15)
The solution process of energy equation has been found to involve the same type
of Roe-average applied to the total specic enthalpy as previously encountered for
the velocity. Thus, collected together, the velocity and the total specic enthalpy
of the intermediate state are given by the explicit relations
u = u
Roe
= u

r
u
r

r
,

h
t
= h
t

Roe
= h
t

h
t

r
h
t
r

r
.
(10.16)
101
It is therefore possible and convenient to solve the rst supplementary equation of
the linearization system by choosing a different set of unknown variables, namely
( , u,

h
t
) in place of ( , m,

E
t
). This means that the rst equation will become
(
u
)( , u,

h
t
) u = P, (10.17)
that, thanks to the solution (10.16), is a single equation for the intermediate density
. Under convenient convexity assumption, as we will see in the models presented
in the following, the resulting system has at least one real solution.
10.5 Polytropic van der Waals uid
As an example of the effectiveness of the linearization procedure for a gas endowed
with quite general thermodynamic properties outlined in the previous section, we
consider the uid model described by the equations of states introduced by van
der Waals, that has been described in section 8.2. This model will be referred
as polytropic van der Waals uid since the its specic heat at constant volume
is assumed to be constant. The general procedure can however applieda also to
the more general van der Waals model including the contribution of the molecular
vibrations to the internal energy, which could be referred to as nonplytropic.
For the polytropic van der Waals uid, the pressure function in terms of the
conservative variables reads
(u) = ( 1)
E
t

m
2
2
+a
2
1 b
a
2
, (10.18)
and hence we have

u
(u) = ( 1)
_
_
_
_
_
_
_
_
_
m
2
/(2
2
) +2a +b
_
E
t
m
2
/ a
2
_
(1 b)
2

2a
1

m/
1 b
1
1 b
_
_
_
_
_
_
_
_
_
.
(10.19)
To solve the rst (supplementary) equation of (10.15) let us introduce the following
change of variables
(, m, E
t
) (, u, h
t
), (10.20)
and write the supplementary equation for the intermdiate state in the form
(
u
)( , u,

h
t
) u = P, (10.21)
102
so that it represents an equation for the intermediate density , the other interme-
diate values u and

h
t
being known from the Roe average.
To obtain the supplementary equation with the transformed variables it is nec-
essary to replace in it m with u/ and E
t
with h
t
. The latter substitution requires
to resort to the following equations of state of the van der Waals uid
_

_
P(T, ) =
RT
1 b
a
2
,
e(T, ) =
RT
1
a.
(10.22)
The specic enthalpy h = e +(P/) of the gas is given by
h(T, ) =
b
( 1)(1 b)
RT 2a, (10.23)
from which we had obtained
P(h, ) = ( 1)
h +2a
2
b
a
2
. (10.24)
By eliminating the temperature in the equation of state for the internal specic
energy in favor of the enthalpy we obtain
e(h, ) =
(1 b)(h +2a)
b
. (10.25)
The pressure gradient in terms of the new variables is obtained by substitution of
the new variables and, after some elementary calculations, is found to be
(
u
)(, u, h
t
) = ( 1)
_
_
_
_
_
_
_
_
_

_
u
2
/2 +2a
_
b
_
u
2
h
t
_
(1 b)( b)

2a
1

u
1 b
1
1 b
_
_
_
_
_
_
_
_
_
.
(10.26)
For the polytropic van der Waals model, the supplementary equation is found
to be a polynomial of third order in the variable representing the dimensionless
intermediate density r = 3b = /
c
, in the form
r
3
+ Ar
2
+ Br +C = 0, (10.27)
103
with coefcients dependent on the left and right states as follows:
A =
3b
2a
P

3( +1),
B =
9b
2a
_
( +1)
P

+( 1)
_
u
2


h
t

um E
t

__
9 ( 2),
C =
27b
2a
_
P

( 1)
_
u
2
2

um E
t

__
,
(10.28)
which can be solved exactly by standard formulas. An alternative formulation of
the coefcients above in terms of critical values is found to be the following
A =
1
6

c
P
c
P

3( +1),
B =
1
2

c
P
c
_
( +1)
P

+( 1)
_
u
2


h
t

um E
t

__
9 ( 2),
C =
3
2

c
P
c
_
P

( 1)
_
u
2
2

um E
t

__
.
(10.29)
In particular, in the case = 0, the supplementary equation is found to be linear
in the intermediate dimensionless density and we have
r = 3
P +( 1)(um E
t
)
P
. (10.30)
Remark 10.5.1 The case
c
In the case of very lowdensity (with respect to the critical density) repulsive forces
are nearly absent, i.e.,

c
=
1
3b
1 b 1 (10.31)
and thus this case corresponds to b = 0. By putting b = 0 in the supplementary
equation the latter becomes a simple linear equation in and we obtain that the
intermediate value of the density is given by
=

+
r
2
(10.32)
2
104
Remark 10.5.2 Alternative form of the coefcients
We notice that the expression within the parentheses of the coefcient C can written
as
1

u
2
2
m u +E
t
_
=
1

u
2
2
m u +
_
m
2
2
_
+E
_
.
But the rst three terms within the square brackets are simply the conservation
condition for the momentum which is satised at the intermediate state, so the
considered expression reduces to E/. Therefore, we can use this result also
in the rst coefcient A by dening the intermediate specic enthalpy as

h
def
=

h
t

| u|
2
2
, (10.33)
and write the three coefcients in the slightly more convenient form:
A =
1
6

c
P
c
P

3( +1),
B =
1
2

c
P
c
_
( +1)
P

+( 1)
_
E


h
__
9 ( 2),
C =
3
2

c
P
c
_
P

( 1)
E

_
.
(10.34)
In the case of = 0, we have
r = 3
P ( 1)E
P
. (10.35)
2
Remark 10.5.3 The multidimensional case
By using the multidimensional version of the supplementary equation with the
derivative with respect to m eliminated in terms of
E
t (u), namely

( u)
_
u m E
t
_

E
t ( u) = P (10.36)
where u m/, the proposed method can be extended to the multidimensional
equations By direct computations the coefcients A, B and C of the cubic equation
105
for the dimensionless density reads
A =
3b
2a
P

3( +1),
B =
9b
2a
_
( +1)
_
| u|
2


h
t

u m E
t

__
9 ( 2),
C =
27b
2a
_
P

( 1)
_
| u|
2
2

u m E
t

__
,
(10.37)
where the intermediate vector velocity is found to be
u =

r
u
r

r
. (10.38)
Alternatively, the coefcients of the cubic equation are given by
A =
1
6

c
P
c
P

3( +1),
B =
1
2

c
P
c
_
( +1)
P

+( 1)
_
| u|
2


h
t

u m E
t

__
9 ( 2),
C =
3
2

c
P
c
_
P

( 1)
_
| u|
2
2

u m E
t

__
,
(10.39)
2
The solution of the linearization problem employed for the polytropic van der
Waals uid can provide a guideline for other gases endowed with more complex
thermodynamic properties. The procedure describedfor this type of uidwas based
on the possibility of writing the supplementary equation explicitly and is therefore
applicable in more complicated situations provided the analytical expression of the
thermodynamic equations of state involved is available. When complicated gas
models are addressed, the equation that denes explicitly the intermediate density
is in general to be solved numerically by some iterative procedure.
10.6 Ideal gases
In the previous analysis of the Roe linearization in Jacobian formwe have assumed
that a supplementary equation can be added to complete the system of equations
106
that enforces the conservativity within the linear approximation. This condition
is necessary in order that the resulting system of equations comprises effectively
three equations for the three unknowns of the intermediate state. There is however
another condition to be satised to reach an actual system of three equations: the
two nontrivial equations enforcing the conservation condition must contain three
real unknowns in the sense that their occurrence must be such that they cannot be
reduced to only two effective unknowns. These two requirements are satised for
gas endowed with arbitrarily general thermodynamic properties but can be violated
under certain mathematical conditions that will be examined in the present section.
The physically interesting result is that these conditions are violated for the gas
model represented by the ideal gas, which is the most employed therodynamic
model for a gas. However, surprisingly enough, the consequences of this viola-
tion are found to be quite different depending on the polytropic or nonpolytropic
characteristic of the gas. In particular we will show that the intermediate state
for a nonpolytropic ideal gas cannot be characterized by the Roe average while
for the polytropic ideal gas, by virtue of a mathematical degeneracy of one of its
equations of state, the linearization in Jacobian form is, rather surprisingly, still
possible through Roe average of velocity and specic total enthalpy, as was the
case for a thermodynamically arbitrary gas.
10.6.1 Flux functions homogeneous of degree one
To ascertain whether the supplementary condition is actually needed and can be
included to close the system of the other two nontrivial equations enforcing the
conservation condition, it is necessary establish the homogeneity properties of
the functions that represent the ux components of Euler equations. In fact, by
classical results regarding the homogeneity of functions, if a functions of two or
more independent variables is rst-order (or degree one) homogeneous, its partial
derivatives depends actually on one variable less than the function itself, the new
variables of the reduced set being dened as the ratio of the original ones with one
of them. Therefore, when all ux components of the Euler equations are rst-order
homogeneous, all elements of the Jacobian matrix become functions of only two
independent variables. As a consequence, the two nontrivial equations ensuring
a conservative linearization are enought for determining the two new variables in
the intermediate state. The third unknown for characterizing the intermediate state
completley will remain arbitrary and the linearization problem will possesses a
one-parameter of innite solutions.
Consider the homogeneity condition (8.27) due to Euler in the particular case
of homogeneity of rst-order
u
u
f (u) = f (u) (10.40)
107
and let us apply it to each component of the vector-valued ux function f(u). Taking
into account the explicit form of the Euler uxes (9.8), the rst-order homogeneity
condition translates into an analogous condition for the pressure function (u) of
the conservative variables
u
u
(u) = (u). (10.41)
In other words, all pieces of the Euler uxes are automatically rst-order homoge-
neous but for the pressure function, which must be ascertained independently and
depends on the thermodynamic model of the gas. In fact, the last relation has been
shown in section 8.4 to be equivalent to the following condition for the pressure
equation of state P(e, )

P(e, )

= P(e, ). (10.42)
The form of this partial differential equation allows for the separation of variables
and the solution is found to be
P(e, ) = g(e) , (10.43)
where g(e) is an arbitrary integration constant. Thus, this is the most general
form of the pressure equation of state such that the Euler uxes are homogeneous
of degree one. Whenever the pressure equation P = P(e, ) has the form just
establishedproportional with mass density no supplementary equation is re-
quired to be introduced in the linearization procedure. Notice that the condition of
homogeneity leaves the form of function g(e) completely arbitrary.
10.6.2 Nonpolytropic ideal gas
A physically relevant example of gas conformal with the equation of state (10.43)
is nonpolytropic ideal gas, described in section 8.3. The form of function g(e) for
this gas depends on the physical model assumed for the contribution to its internal
energy of the molecular vibrations (and possibly also rotations) that are beyond
the interest of the present discussion. We are interested instead to the pressure as
a function of the conservative variables, namely,
(u) = g
_
E
t


m
2
2
2
_
. (10.44)
The gradient of (u) is then given by

u
(u) =
_
_
_
_
_
_
_
g
_
E
t


m
2
2
2
_

_
E
t

m
2

2
_
g

_
E
t


m
2
2
2
_

_
E
t


m
2
2
2
_
g

_
E
t


m
2
2
2
_
_
_
_
_
_
_
_
. (10.45)
108
As anticipated, the rst-order homogeneity implies that the derivatives of the uxes
depend on only two independent variables. We take the gas velocity u = m/ and
the internal specic energy
e =
E
t

m
2
2
2
(10.46)
as independent variables. With this choice, the pressure gradient assumes the form
(
u
)(u, e) =
_
_
_
g(e)
_
e u
2
/2
_
g

(e)
ug

(e)
g

(e)
_
_
_
. (10.47)
In the same vein, the Jacobian matrix of the vector ux f(u) expressed in terms of
the assumed independent variables is
_
f(u)
u
_
(u, e) =
_
_
_
_
0 1 0
u
2
+ g(e)
_
e u
2
/2
_
g

(e) u
_
2 g

(e)
_
g

(e)
u
_
e +u
2
/2 +
_
e u
2
/2
_
g

(e)
_
e +u
2
/2 + g(e) u
2
g

(e) u
_
1 + g

(e)
_
_
_
_
_
.
(10.48)
The two equations of the linearization problem read
_

_
_
u
2
+ g( e)
_
e u
2
/2
_
g

( e)
_
+ u
_
2 g

( e)
_
m
+ g

( e)E
t
=
_
m
2

+
_
u
_
e + u
2
/2 +
_
e u
2
/2
_
g

( e)
_
+
_
e + u
2
/2 + g( e) u
2
g

( e)
_
m
+ u
_
1 + g

( e)
_
E
t
=
_
m

_
E
t
+
_
_
(10.49)
The solution ( u, e) of the system depends on the form of function g(e) and is not
amenable to a Roe average for velocity and total specic enthalpy unless function
g(e) is the proportionality relation
g(e) = e, (10.50)
with a suitable constant. Thus, for a nonpolytropic ideal gas with g(e) = e
the linearization in Jacobian form is possible but it is impossible to obtain the
intermediate state through the classical Roes average.
In the special case g(e) = e, the pressure equation (10.43) applies, repeated
here for convenience,
P(e, ) = e, (10.51)
109
which is the signature of the polytropic ideal gas with = ( 1) and P(e, ) =
( 1)e. In this very special situation, the Jacobian matrix reduces to
_
f(u)
u
_
(u, e) =
_
_
_
_
_
_
0 1 0

3
2
u
2
(3 )u 1
u
_
3
2
u
2
e
_
3 2
2
u
2
+ e u
_
_
_
_
_
_
.
(10.52)
In the next section we investigate this special case of polytropic ideal gas to verify
whether the very simplied form of its Jacobian matrix does not allow the possi-
bility of circumventing the obstacle met by Roes linearization in the more general
case of a truly nonpolytropic ideal gas.
10.6.3 Polytropic ideal gas
For the polytropic ideal gas, introduced in section 8.6.1, the pressure function (u)
assumes the simple form
(u) = ( 1)
_
E
t

m
2
2
_
, (10.53)
that is rst-order homogeneous. By applying the gradient operator to (u) we
obtain

u
(u) =
_
_
_
_
_
_
1
2
m
2

2
( 1)
m

1
_
_
_
_
_
_
. (10.54)
The corresponding Jacobian matrix reads
_
f(u)
u
_
(u, e
t
) =
_
_
_
_
_
_
0 1 0
3
2
u
2
( 3)u 1
( 1)u
3
ue
t
3( 1)
2
u
2
+ e
t
u
_
_
_
_
_
_
. (10.55)
On the other hand, we have
e
t
=
_
e +
1
2
u
2
_
= e +

2
u
2
= h +

2
u
2
= h
t
+
1
2
u
2
. (10.56)
110
By virtue of this result, the Jacobian matrix as a function of variables u and h
t
is
found to be
_
f(u)
u
_
(u, h
t
) =
_
_
_
_
_
_
_
0 1 0
3
2
u
2
( 3)u 1
1
2
u
3
uh
t
2( 1) u
2
+h
t
u
_
_
_
_
_
_
_
. (10.57)
Let us write the momentumequation of Roe conservation conditions, in which the
last term
u
( u) u is evaluated using (10.53) and (10.54), i.e.,
()
_
m

_
2
2(m)
m

+
_
m
2

1
2
_
()
_
m

_
2
2(m)
m

+
_
m
2

_
_
= 0.
(10.58)
It is remarkable that the term within the square brackets of this expression is co-
incident with the rst three terms, of the second equation of Roe linearization
condition. Thus, for a polytropic ideal gas the relation
u
( u) u = is au-
tomatically satised at the intermediate state as a consequence of the conservation
condition for momentum.
Therefore, the solution of Roe linearization problem for the polytropic ideal
gas becomes
_

_
= arbitrary
u =

r
u
r

h
t
=

h
t

r
h
t
r

r
(10.59)
as originally formulated by Roe in his celebrated paper.
To help the practical implementation of Roe linearization for the polytropic
ideal gas we provide the matrices of the eigenvectors expressed as a function of
the two variables u and h
t
which are the leading ladies of the algorithm. First, the
speed of sound as a function of these variables is
c =
_
( 1)
_
h
t

1
2
u
2
_
(10.60)
and the three eigenvalues then read, most simply,

1
= u c,
2
= u,
3
= u +c. (10.61)
111
The matrix R of the right eigenvectors is chosen in the form
R(u, h
t
) =
_
_
_
1 1 1
u c u u +c
h
t
cu u
2
/2 h
t
+cu
_
_
_
(10.62)
while the matrices L of the left eigenrows, normalized so that LR = I, reads
L(u, h
t
) =
1
2c
2
_
_
_
_
u
2
/2 +cu/( 1) u c/( 1) 1
u
2
+2c
2
/( 1) 2u 2
u
2
/2 cu/( 1) u +c/( 1) 1
_
_
_
_
(10.63)
112
Chapter 11
High-resolution schemes for systems
of conservation laws
In this chapter, high-resolution schemes for a system of conservation laws are de-
rived following the analysis performed in the previous chapter for a scalar conser-
vation law. The procedure for the computation of the (vector) integrated numerical
ux including the limiter is also detailed.
11.1 Introduction
Consider a systemof p (nonlinear) conservation laws, that will be written as follows

t
u + f(u) = 0, (11.1)
where u R
p
and f R
p
R
d
and f = (f
x
, f
y
, f
y
)
T
, with f
x
, f
y
, f
z
R
p
. Sans
serif symbols will be used hereinafter to denote vectors with p components, that
is, vectors in the sense of the considered system of equations. The scalar product
between a Cartesian vector c = (c
x
, c
y
, c
z
)
T
R
d
and the tensor array f R
p
R
d
results in a vector of dimension p and is evaluated for d = 3 as
c f = c
x
f
x
+c
y
f
y
+c
z
f
z
, (11.2)
that produces a vector in R
p
. Similarly, we have
f =
x
f
x
+
y
f
y
+
z
f
z
.
The Jacobian three-dimensional array A R
pp
R
d
of the ux tensor function
f R
p
R
d
is dened as
A(u)
def
=
f(u)
u
, (11.3)
113
or, in details, for d = 3,
A(u) = (A
x
(u), A
y
(u), A
z
(u))
T
=
_
f
x
(u)
u
,
f
y
(u)
u
,
f
z
(u)
u
_
T
,
with A
x
, A
y
, A
z
R
pp
. The determination of the boundary values to be used for
calculating the ux on the boundary is based on the analysis presented in the next
chapter 12. For a further discussion of the very important issue of how to enforce
the boundary conditions in hyperbolic systems the interested reader is referred to
[9].
11.2 Multidimensional upwind scheme
for a system of conservation laws
The upwind scheme for a system of conservation laws reads (e.g. [9])
L
i
du
i
dt
=

kN
i,=
_

i k

f
i
+f
k
2

1
2
|

A
i k
| [u
k
u
i
]
_


N
i

, (11.4)
where f
i
= f(u
i
) and

f

= f(u

). Here we have introduced a Roes linearizing


matrix

A
i k
R
pp
[21] satifying the following condition

A
i k
[u
k
u
i
] =
i k

_
f(u
k
) f(u
i
)
_
. (11.5)
The construction of Roes matrices for the Euler equations has been described in
Chapter 10. We recall that, if Roe linearization is sought for in Jacobian form
as described in 10.2.3, the multidimensional linearization problem reduces to the
determination of an intermediate state u(u
i
, u
k
,
i k
) such that

i k
A
_
u(u
i
, u
k
,
i k
)
_
[u
k
u
i
] =
i k

_
f(u
k
) f(u
i
)
_
.
In other words, the linearizing Roe matrix

A
i k
is chosen as

A
i k
=
i k
A
_
u(u
i
, u
k
,
i k
)
_
.
We refer to Chapter 10 and [11] for a complete description of the linearization
problem in the multidimensional case.
The rst-order upwind scheme is now rewritten exploiting the eigenstructure
of Roe matrix, so as to identify the contribution of each characteristic eld to the
integrated numerical ux. The Roe matrix has real eigenvalues and a complete set
of eigenvectors, see section 10.2, that is,

A
i k
=

R
i k

i k

L
i k
,
114
where

R
i k
and

L
i k
R
pp
are the matrices of right and left eigenvectors r
i k
and

l
i k
R
p
, respectively, and

i k
= diag
_

1,i k
,

2,i k
, . . . ,

p,i k
_
,
with

,i k
R representing the -th eigenvalue of

A
i k
, in ascending order.
The matrix |

A
i k
| R
pp
in (11.4) is dened as
|

A
i k
|
def
=

R
i k
|

i k
|

L
i k
,
=

R
i k
diag
_
|

1,i k
|, |

2,i k
|, . . . , |

p,i k
|
_

L
i k
.
(11.6)
The upwind scheme (11.4) is now recast in a slightly different form by projecting
u
k
u
i
onto the basis of the characteristic variable as follows
v
i k
def
=

L
i k
[u
k
u
i
]. (11.7)
Then, by substituting the identity |

A
i k
| [u
k
u
i
] =

R
i k
|

i k
| v
i k
, into the mul-
tidimensional upwind scheme (11.4), we obtain
L
i
du
i
dt
=

kN
i,=
_

i k

f
i
+f
k
2

1
2

R
i k
|

i k
| v
i k
_


N
i

. (11.8)
11.3 Boundary conditions for system
of conservation laws
In this section, the (weak) enforcement of the boundary conditions for a system of
conservation laws is derived by extending the procedure for a scalar conservation
law outlined in Section 2.2. A detailed analysis of the boundary condition proce-
dure for nonlinear hyperbolic systemin one dimension will be provided in Chapter
12. Here we use in anticipation the elements of the analysis for one-dimensional
systems by adapting them to context of the multidimenisonal equations.
The rst observation is that the denition of the inow and outow parts of
the boundary is achieved by resorting to a local linearization of the system of
conservation laws, exactly as in the scalar case. The second observation is that
these partitions can be dened only with reference to each characteristic eld of
the linearized system, for which a (wave) velocity is dened. Therefore, the inow
and outow partitions of different characteristic eld may overlap; in other words,
a portion of the boundary can be an inow boundary for the -th characteristic
eld and, at the same time, an outow boundary for the m-th characteristic eld,
, m < p.
115
To avoid the difculties related to the computation of the boundary integrals
under these inow boundary conditions of vector kind, we seek for a vector u de-
ned on the whole boundary , to be substitued for the trace u

of the unknown
u into the boundary integrals regardless of their inow or outow character, simi-
larly to what done for the scalar equation (see Section 2.2). For simplicity, we will
assume in the following that a vector boundary data b of dimension p is given,
although the actual number of required boundary data depends on the number of
inow conditions to be specied and can be lower than the dimension p of the
unknown, for details see Chapter 12.
Let us introduce the following characteristic vector
v = L

_
b u

_
(11.9)
where L

= L(u

, n), n outward normal, is the matrix of the left eigenvectors of


the ( p p) matrix n A(u

), namely,
L

= n A(u

)
where R

and

= diag (

1
,

2
, . . . ,

p
) are the matrices of the right eigenvectors
and of the eigenvalues, respectively. By making explicit the dependence of v on
time and space, we have
v(s, t ) = L
_
u

(x(s), t ), n(s)
_ _
b(s, t ) u

(x(s), t )
_
.
We now introduce an operator acting on a vector y which selects its components
only when the corresponding eigenvalue is strictly negative, and puts zero other-
wise. Explicitely, we dene SN

the operator selecting the part corresponding to


the strictly negative eigenvalues, as follows
SN

y
_
_
_
_
_
_
_
_
_
_
_
_
_
_
y
1
if

1
< 0
0 if

1
0
y
2
if

2
< 0
0 if

2
0
.
.
.
y
p
if

p
< 0
0 if

p
0
_
_
_
_
_
_
_
_
_
_
_
_
_
_
(11.10)
In terms of this operator, the relation dening the characteristic variation v on the
right extreme of the integration interval can be written in the following compact
form
v
def
= SN

v = SN

_
b u

_
. (11.11)
116
We now dene the boundary value of the unknown satisfying the inlet boundary
condition as
u
def
= u +v = u +SN

_
b u

_
, (11.12)
or, making explicit the functional dependence of u
u(s) = u +R
_
u

, n(s)
_
SN

L
_
u

, n(s)
_ _
b u

_
. (11.13)
We notice in passing that, if

> 0, = 1, . . . , p, then u u

, while, if

0, = 1, . . . , p, then u b, two situations which correspond, in the


gasdynamic jargon, to supersonic outow and inow, respectively.
Let us move to the discrete imposition of the boundary conditions. To this
purpose, we introduce the nite element representations of the unknown u and of
the boundary data b, that is,
u(x, t ) u
h
(x, t )
def
=

kN

u
k
(t )
k
(x) and
b(s, t ) b
h
(s, t )
def
=

kN

b
k
(t )

k
(s),
respectively. We take u

= u
|
. We now perform the following reinterpolation
of the ux function f(u),

f(s, t )

kN
f
k
(t )

k
(s)
where
f
k
(t )
def
= f
_
u(s
k
, t )
_
.
Similarly to section 3.2, the functional dependence of

f(s, t ) on the boundary coor-


dinate s is retained because the evaluation of the function u requires the knowledge
of the (local) outward normal n(s). However, assuming linear or bilinear elements,
n is piecewise constant, that is, it is constant over the boundary
e
belonging to
the e-th element, so that we can introduce the following (constant) restriction of
f
k
(s, t ) over the e-th element
f
k
(t ) = f
_
u(s
k
, t )
_
= constant, s
e
k

which is the generalization of f
k
in section 3.2.
Therefore, the complete form of the upwind scheme for a system of conserva-
tion laws reads,
L
i
du
i
dt
=

kN
i,=
_

i k

f
i
+f
k
2

1
2

R
i k
|

i k
| v
i k
_


N
i

, (11.14)
where the inlet boundary conditions are taken into account through the vector
function f

.
117
11.4 High-resolution scheme for steady solutions
Let us write the expression of the node-pair nite element discretization of a mul-
tidimensional system of conservation laws, namely,
L
i
du
i
dt
=

kN
i,=

i k

f
i
+f
k
2

k

N

i,=


N
i

, (11.15)
where the mass lumping approximation has already been adopted. Equation
(11.15) is the system counterpart of the scalar equation (7.7) and it is a system
of p equations in the vector unknown u
i
for each node i of the triangulation.
The high-resolution scheme for steady solutions is obtained by using (11.15) as
the high-order scheme and the rst-order upwind scheme (11.4) as the low-order
integrated numerical ux q
L
i k
associated with the node-pair i -k. To this purpose,
we introduce the following denition of the high-resolution integrated numerical
ux of the system, namely, (cf. (7.1))
q
HR
i k
= q
L
i k
+
_
q
H
i k
q
L
i k
_
limited
. (11.16)
Differently from the scalar denition (7.1), the limiting procedure to be applied to
the high-resolution correction q
H
i k
q
L
i k
of the system is still to be dened.
First, let us focus on the domain contribution to q
HR
. By substituting the
domain contribution of (11.4) and (11.15), namely,
q
L
i k
=
i k

f
i
+f
k
2

1
2
|

A
i k
| [u
k
u
i
] and q
H
i k
=
i k

f
i
+f
k
2
,
we have
q
H
i k
q
L
i k
=
1
2
|

A
i k
| [u
k
u
i
]. (11.17)
In Chapter 7, we introduced a ux limiter function which depends on the ratio of the
centred and upwind variations of the scalar unknown. For a systemof conservation
laws two difculties are encountered in extending the scalar procedure. First, the
upwind direction is not immediately dened because there are two or more
propagation velocities in the case of a system of conservation laws. Second, there
are two or more variable jump ratios (to be limited), since the unknown variable is
a vector.
Following [16], we choose to resort to a local linearization of the nonlinear
hyperbolic systemunder consideration. With this choice, the wave signs and speeds
along a given direction are dened as the sign and magnitude of the eigenvalues
of the linear hyperbolic system, respectively, thus allowing for the selection of the
upwind direction for each characteristic eld. Moreover, for each wave we will
118
consider the jump of the corresponding characteristic variable as the measure of the
wave strength, to be limited independently from the other characteristic variables,
exactly as in the scalar case. We notice in passing that the above procedure depends
strongly on the particular direction that has been chosen for the linearization, unless
the Cartesian components of the Jacobian three-dimensional array of the ux tensor
commute.
For the purpose of introducing the characteristic limitation procedure, let us
rewrite the high-resolution correction by substitutingthe characteristic formulation
(11.8) of the upwind scheme to obtain
q
H
i k
q
L
i k
=
1
2

R
i k
|

i k
| v
i k
(11.18)
The system counterpart of the variable in (7.6) is a vector of dimension p whose
components reads

,i k
def
=
v
upw
,i k
v
cen
,i k
=
v
upw
,i k
v
,k
v
,i
. (11.19)
Correspondingly, we introduce the following diagonal matrix limiter

i k
def
= diag
_
(
1,i k
), (
2,i k
), . . . , (
p,i k
)
_
(11.20)
where the function is the (scalar) limiter function introduced in Chapter 7. In
equation (11.19), the centred variation is dened simply by
v
cen
,i k
= v
,i k
=

l
,i k
[u
k
u
i
],
while the denition of the upwind variation requires to scale the differences u
i
u
i

and u
k
u
k
since the variable of the limiter is a slope ratio and the unstructured
mesh is nonuniform. Thus the upwind variations must be referred to the distance
betweenthe twocentral nodes of the considerednode-pair andthis is easilyobtained
by the following simple scaling:
v
upw
i k,
=
_

l
,i k
[u
i
u
i
]
|x
k
x
i
|
|x
i
x
i
|
if

,i k
0,

l
,i k
[u
k
u
k
]
|x
k
x
i
|
|x
k
x
k
|
if

,i k
< 0.
These denitions of the components of the characteristic variations can be written
compactly in the vector relations
v
cen
i k
=

L
i k
[u
k
u
i
],
v
upw
i k
=

L
i k
SP

ik
[u
i
u
i
]
|x
k
x
i
|
|x
i
x
i
|


L
i k
SN

ik
[u
k
u
k
]
|x
k
x
i
|
|x
k
x
k
|
.
(11.21)
119
The limited characteristic jumps (which correspond to the limiter in Rebays form
of section 7.3) are then obtained as
v

i k
=
i k
v
i k
, (11.22)
which can be directly substituted into (11.18) to obtain the limited domain contri-
bution to the high-resolution integrated numerical ux, that is,
q
L
i k
+
_
q
H
i k
q
L
i k
_
limited
=
i k

f
i
+f
k
2

1
2

R
i k
|

i k
| v
i k
+
1
2

R
i k
|

i k
|
i k
v
i k
=
i k

f
i
+f
k
2

1
2

R
i k
|

i k
| v
i k
+
1
2

R
i k
|

i k
| v

i k
=
i k

f
i
+f
k
2

1
2

R
i k
|

i k
|
_
v
i k
v

i k
_
.
(11.23)
Let us now consider the contribution of the boundary term. For simplicity, we
will consider in the following a boundary condition which does not change kind
on a given node-pair, so that we will use the augmented metric quantities

k
can
be used. The more general case can be dealt with appropriately and it will be
recovered later on without loss of generality.
By substituing the boundary terms in (11.15) and (11.4), we have
_
q
H

k
q
L

k
_

k
f

2
. (11.24)
In order to introduce a limited version of the boundary term contributions, it is
necessary to transform to the characteristic variables. To this aim it is convenint
to resort rst a Roe linearization in the direction

k
(outward normal) satisfying
the following condition

k
[u

k
u

] =

_
f(u

k
) f(u

)
_
, (11.25)
or, whenever the Roe matrix is chosen in Jacobian form,

k
A
_
u(u

, u

k
,

k
)
_
[u

k
u

] =

_
f(u

k
) f(u

)
_
.
The boundary contribution to the high-resolution integrated numerical ux is then
written as
_
q
H

k
q
L

k
_

=
1
2

k
v

k
,
where

R

k
,

L

k
and

k
are the matrices of the right, left eigenvectors and eigen-
values of

A

k
and
v

k
def
=

L

k
[u

k
u

].
120
The limited contribution to the high-resolution ux reads
_
q
H

k
q
L

k
_

limited
=
1
2

k
v

i k
=
1
2

k
v

k
, (11.26)
where

k
is dened as in (11.20) with the understanding that the eigenstructure
corresponding to the boundary node-pair i k is used.
Including the limiting procedure, the complete form of the high-resolution
scheme for steady solutions to hyperbolic systems reads
L
i
du
i
dt
=

kN
i,=
_

i k

f
i
+f
k
2

1
2

R
i k
|

i k
|
_
v
i k
v

i k
_
_

k

N

i,=
1
2

k
v


N
i

.
(11.27)
This can also be recast in the conservation form
L
i
du
i
dt
=

kN
i,=
q
HR
i k


N
i

, (11.28)
by means of the high-resolution integrated numerical ux dened by
q
HR
i k
=
i k

f
i
+f
k
2

1
2

R
i k
|

i k
|
_
v
i k
v

i k
_
+
1
2

k
v

k
. (11.29)
11.5 A general limiter function
for hyperbolic systems
In the previous section, a limitation procedure has been presented to construct
the high-resolution integrated numerical ux for a system of conservation laws.
However, in the procedure detailed above the high-order correction q
H
i k
q
L
i k
is
written in terms of node-pair differences of conservative or characteristic variables.
More precisely, to obtain the characteristic variations v
i k
to be limited, each
term appearing in q
H
i k
q
L
i k
is to be recast as a product of a function times the
difference u
k
u
i
. Remarkably enough, this is not possible, in general, for a given
combination of high/low-order integrated numerical uxes.
In this section, we overcome this restriction by introducing an equivalent lim-
iting procedure based on a limiter full matrix, which allows for other numerical
schemes to be easily encompassed within the high-resolution approach investigated
here.
121
Let us consider the domain contribution to the high-order correction q
H
i k
q
L
i k
as in (11.23), namely,
_
q
H
i k
q
L
i k
_
domain
limited
=
1
2

R
i k
|

i k
|
i k
v
i k
=
1
2

R
i k

i k
|

i k
| v
i k
=
1
2

R
i k

i k

L
i k

R
i k
|

i k
| v
i k
,
where, since |

i k
| and
i k
are diagonal matrices, we have substituted |

i k
|
i k
=
i k
|

i k
| and, in the last relation, we exploited the orthogonality property of the
eigenvectors, that is,

i k

R

i k
= I,
where I is the p p identity matrix. Introducing now the limiter matrix

i k
def
=

R
i k

i k

L
i k
, (11.30)
and coming back to the conservative variable u by substituting v
i k
=

L
i k
[u
k

u
i
], we have
_
q
H
i k
q
L
i k
_
domain
limited
=
1
2

i k

R
i k
|

i k
|

L
i k
[u
k
u
i
]
=
1
2

i k
|

A
i k
| [u
k
u
i
]
=
i k
_
q
H
i k
q
L
i k
_
,
(11.31)
to be compared with dention (7.1) introduced for a single conservation law, that
is, q
HR
i k
= q
L
i k
+
_
q
H
i k
q
L
i k
_
.
It is remarkable that for a system of conservation laws the limiter function is
foundtobe a full matrixof dimension pp, whichis obtainedbytransformingback
to the conservative variable space a vector limiter written in terms of characteristic
differences.
A similar result is obtained when the boundary contribution is considered. In
this case, we have (cf. (11.26))
_
q
H
i k
q
L
i k
_

limited
=
1
2

i k

i k

i k
v

i k
=
1
2

i k

i k

i k
v

i k
=
1
2

i k

i k

L

i k

R

i k

i k
v

i k
.
122
By introducing now the boundary limiter matrix,

i k
def
=

R

i k

i k

L

i k
, (11.32)
and coming back to the conservative variables u, we have
_
q
H
i k
q
L
i k
_

limited
=
1
2

i k

R

i k

i k

L

i k
[u
k
u
i
]
=
1
2

i k

A

i k
[u
k
u
i
]
=

i k

i k

f
k
f
i
2
=

i k
_
q
H
i k
q
L
i k
_

,
(11.33)
in which the denition (11.25) of the boundary Roe matrix has been used.
To conclude, let us rewrite the general form of the high-resolution integrated
numerical ux for a system of conservation laws as (cf. (11.16)) by summing the
domain and boundary contributions
q
HR
i k
= q
L
i k
+
_
q
H
i k
q
L
i k
_
limited
= q
L
i k
+
i k
_
q
H
i k
q
L
i k
_
+

i k
_
q
H
i k
q
L
i k
_

,
(11.34)
in which the domain and boundary limiter matrices
i k
and

i k
are dened as

i k
def
=

R
i k

i k

L
i k
and

i k
def
=

R

i k

i k

L

i k
. (11.35)
11.6 High-resolution scheme for unsteady solutions
We rst write the high order scheme to be used to develop the high-resolution
node-pair TaylorGalerkin discretization for a system of conservation laws. The
second order scheme is the TaylorGalerkin scheme in the node-pair representation
and using the mass lumping approximation:
L
i

t
u
i
t
=

kN
i,=

i k

f
i
+f
k
2

k

N

i,=


N
i

t
2

kN
i,=
K
i k
: [A
i k
A
i k
] [u
k
u
i
]
+
t
2

kN

,=
K

k
: [A
k
A
k
] [u
k
u

].
(11.36)
123
where
t
u
i
= u
n+1
i
u
n
i
. The last summation involves both boundary nodes and
internal ones since the original boundary integral involved the gradient of u, whose
evaluation needs a knowledge of u at some internal points near to boundary.
The meaning of the product indicated by the colon : is that of a double sum
over the index of all of the cartesian components of the tensor, namely,
K : [A A] = K : IB =
d

s=1
d

t =1
k
s,t
B
s,t
.
with the subscript i k dropped for simplicity, and where, for two-dimensional prob-
lems,
IB = A A =
_
A
2
x
A
x
A
y
A
y
A
x
A
2
y
_
Notice that the matrix above contains both products A
x
A
y
and A
y
A
x
, which are
different unless the Jacobian matrices A
x
and A
y
commute. The two summations
leads the addition of four product terms as follows
K : [A A] = k
1,1
A
2
x
+k
1,2
A
x
A
y
+k
2,1
A
y
A
x
+k
2,2
A
2
y
.
For three-dimensional problems, nine product terms are to be summed. In any
case, each product term is the scalar multiple the matrix B
s,t
and is obtained by
multiplying all its elements by one and the same scalar factor. The nal result of
the double tensor contraction is a matrix of order p, namely:
K : IB = C.
To formulate the high-resolution scheme for unsteady solutions, we proceed sim-
ilarly to that developeed for steady solutions in section 11.4. We rst assume a
Roe local linearization of the system with

A
i k
= A
_
u(u
i
, u
k
,
i k
)
_
so that the
TaylorGalerkin scheme (11.36) using mass lumping and node-pair form, will be
rewritten as
L
i

t
u
i
t
=

kN
i,=
q
H
i k


N
i

(11.37)
by exploiting the high-order integrated numerical ux dened by
q
H
i k
=
i k

f
i
+f
k
2
+
t
2
_

A
i k


A
i k
_
[u
k
u
i
]
+

2

t
2

k

N

,=
K

k
:
_

A
k


A
k
_
[u
k
u

].
(11.38)
124
The high-resolution scheme is derived by considering rst the domain contribution
to high-order correction to the ux,
q
H
i k
q
L
i k
=
1
2
_
|

A
i k
| +t K
i k
:
_

A
i k


A
i k
_
_
[u
k
u
i
]
=
1
2
_

R
i k
|

i k
|

L
i k
+t K
i k
:
_

A
i k


A
i k
_

R
i k

L
i k
_
[u
k
u
i
]
=
1
2
_

R
i k
|

i k
| +t K
i k
:
_

A
i k


A
i k
_

R
i k
_

L
i k
[u
k
u
i
].
where in the second line we have substituted |

A
i k
| =

R
i k
|

i k
|

L
i k
and

R
i k

L
i k
=
I. Coming back to the variation of the characteristic variables we have
q
H
i k
q
L
i k
=
1
2
_

R
i k
|

i k
| +t K
i k
:
_

A
i k


A
i k
_

R
i k
_
v
i k
. (11.39)
Therefore, the domain contribution can be limited by applying the same procedure
described in section 11.4, that is, by substituting
v
i k
v

i k
=
i k
v
i k
,
to obtain
_
q
H
i k
q
L
i k
_
limited
=
1
2
_

R

k
|

k
| +t K
i k
:
_

A
i k


A
i k
_

R
i k
_
v

i k
. (11.40)
Coming now to the boundary terms, namely,
_
q
H

k
q
L

k
_

2

t
2

k

N

,=
K

k
:
_

A
k


A
k
_
[u
k
u

],
we notice that the rst term has been already dealt with in section 11.4 while the
second boundary term cannot be expressed in terms of differences of variables
and therefore requires to apply the limiting procedure described in the previous
section. Moreover, there is a nodal contribution which must be separated from the
off diagonal one, since the limiting procedure ought to be applied only to the latter.
The limited boundary contribution to the high-resolution ux therefore reads
_
q
H

k
q
L

k
_

limited
=
1
2

k
v

k

t
2

k

N

,=

i k
K

k
:
_

A
k


A
k
_
[u
k
u

]
=
1
2

k
v

k

t
2

k

N

,=
K

k
:
_

A
k


A
k
_
v

k
.
(11.41)
125
Putting together the low order ux and the limited high order ux, we eventu-
ally obtain the high-resolution integrated numerical ux for the multidimensional
hyperbolic system:
q
HR
i k
=
i k

f
i
+f
k
2

1
2
|

A
i k
| [u
k
u
i
]
+
1
2
_

R
i k
|

i k
| +t K
i k
:
_

A
i k


A
i k
_

R
i k
_
v

i k

t
2

k

N

,=
K

k
:
_

A
k


A
k
_
v

k
+
1
2

k
v

k
.
(11.42)
As in the scalar equation, the boundary contribution to the limited second order
termcan be computed only by introducing ctitious points external to the boundary
in order the node-pair can be extended also for the edges with one point on the
boundary and one points inside the computational region. For this boundary term,
the value of the variables at surface nodes can be taken in the form u
i
, which
includes the treatment of the boundary values (inow or outow) introduced for
taking into account the boundary conditions. The values at the new external points
introduced articially can be dened by extrapolation, e.g., a linear extrapolation,
or eventually a quadratic one.
Using this high-resolution numerical ux, the discrete time advancement of
the hyperbolic system is performed by means of the scheme
L
i

t
u
i
t
=

kN
i,=
q
HR
i k


N
i

. (11.43)
126
Chapter 12
Boundary conditions in nonlinear
hyperbolic systems
12.1 Introduction
The imposition of boundary conditions is a very critical issue in the numerical
solution of nonlinear hyperbolic problems. The rst and most fundamental dif-
culty is caused by the nonlinearity which implies that the parts of the boundary
of the computational domain where boundary values must be specied are not
known a priori. In fact, even in the scalar case, the inow and outow portions
of the boundary depend on the direction of the advection velocity respect to that
of the boundary normal and the former depends on the solution itself evaluated
on the boundary. Therefore, in general it is not possible to specify the boundary
conditions of a conservation law equation without making reference to the actual
boundary values of the solution.
Another very important aspect of the imposition of the boundary conditions in
hyperbolic problems is that their mathematical status is completely different from
that of Dirichlet boundary conditions in elliptic or parabolic problems. Enforcing
a Dirichlet condition corresponds to x the point values of the unknown on the
boundary. On the contrary, imposing a boundary condition for a conservation law
equation means to take into account the prescribed boundary values only through
the computation of the ux. In other words, in the hyperbolic case the unknown
is not required to assume the values specied on the boundary and the nal actual
values assumed by the unknown variable will be only an approximation of the
boundary data.
Athird element of complication is encountered in hyperbolic systems. Here the
unknown has two or more components and a different number of boundary values
can be specied on different parts of the boundary. The two extreme situations are
127
on the one hand the imposition of prescribed values for all the unknowns and on the
other hand no boundary condition at all. In fact, the correct number of boundary
conditions which must be specied at a given boundary point is determined by the
signs of the local eigenvalues of the system written in quasi-linear form. However,
once the right number of boundary conditions is established, one is often still faced
with the additional difculty that some variables prescribed on the boundary do
not correspond to one of the unknowns of the problem.
This chapter is devoted to the problem of imposing the boundary conditions
in nonlinear hyperbolic systems in one dimension. A general procedure recently
suggested by the rst author is presented. It is characterized by the explicit identi-
cation and use of three sets of variables: the conservative unknowns of the system,
the local characteristic variables and the physical variables. The procedure consists
in a number of steps, which are here described for a general hyperbolic system.
The effectiveness of Guardones procedure has been assessed against a test prob-
lem consisting in the reection of a shock-wave by a plane wall by comparing
the numerical solution with the exact analytical solution easily obtainable from
RankineHugoniot relations.
12.2 Conservative, characteristic and
physical variables
Let us consider a system of nonlinear hyperbolic equations in one dimension with
the vector unknown
u =
_
_
_
_
_
u
1
u
2
.
.
.
u
m
_
_
_
_
_
. (12.1)
The m components of the unknown are the conservative or conservation variables
of the system.
For the purposes of a numerical solution method, the system of conservation
laws
t
u +
x
f(u) = 0 is also written in the so-called quasi-linear form
t
u +
A(u)
x
u = 0, where A(u) = f(u)/u is the Jacobian matrix of the ux vector
f(u). In the same context, one considers the eigenvalue problem associated with
this matrix A(u) and, under the condition of strict hyperbolicity, it is standard to
dene the matrix R(u) of right eigenvectors and the matrix L(u) of left eigenrows.
It is standard to normalize the eigenrows on the basis of the right eigenvectors, so
that L(u)R(u) = I. By means of these two matrices, one denes the characteristic
128
variable v = L(u)u, which has m components,
v =
_
_
_
_
_
v
1
v
2
.
.
.
v
m
_
_
_
_
_
. (12.2)
The inverse transformation from the characteristic variables to the conservative
ones is obtained by multiplying v = L(u)u by R(u) and using the orthonormaliza-
tion condition to give u = R(u)v. This relation gives the inverse transformation
only implicitly, since matrix R(u) involves the conservation variable u one is look-
ing for, except when the hyperbolic system is linear. Only in such a linear case,
the characteristic variables have the important properties of being linear combina-
tions of the conservative unknowns which are governed by uncoupled advection
equations, each with its own (constant) propagation velocity.
In view of specifying the boundary values, it is convenient to introduce a third
set of variables, which are called physical or primitive variables and that will be
denoted by
p =
_
_
_
_
_
p
1
p
2
.
.
.
p
m
_
_
_
_
_
. (12.3)
For instance, for the Euler equations of gasdynamics, the three conservative vari-
ables are
u =
_
_
u
1
u
2
u
3
_
_
=
_
_

m
E
t
_
_
, (12.4)
where is the mass density, m is the momentum density and E
t
is total energy
density of the gas, while typical physical variables will be
p =
_
_
p
1
p
2
p
3
_
_
=
_
_

u
P
_
_
, (12.5)
where u is uid velocity and P is pressure.
The relationship between the conservative variables and the physical ones is
a vector-valued function of the type u = u(p) and its inverse p = p(u). Always
with reference to the Euler equations of gasdynamics, the three equations of the
129
rst transformation read
u
1
= = p
1
u
2
= m = u = p
1
p
2
u
3
= E
t
= e
t
=
_
e +
1
2
u
2
_
=
_
e(P, ) +
1
2
u
2
_
= p
1
_
e(p
3
, p
1
) +
1
2
p
2
2
_
,
(12.6)
where e = e(P, ) is an equation of state of the gas, so that
u(p) =
_
_
p
1
p
1
p
2
p
1
_
e(p
3
, p
1
) +
1
2
p
2
2
_
_
_
. (12.7)
For a polytropic ideal gas e(P, ) = P//( 1).
The relations of the inverse transformation p = p(u) will be
p
1
= = u
1
p
2
= u =
m

=
u
2
u
1
p
3
= P = P(e, ) = P
_
E
t

1
2
u
2
,
_
= P
_
u
3
u
1

u
2
2
2u
2
1
, u
1
_
(12.8)
where P = P(e, ) is another equation of state of the gas, so that
p(u) =
_
_
_
_
_
_
_
u
1
u
2
u
1
P
_
u
3
u
1

u
2
2
2u
2
1
, u
1
_
_
_
_
_
_
_
_
. (12.9)
For the polytropic ideal gas P(e, ) = ( 1)e.
12.3 The boundary values for a scalar unknown
Let us consider the problem of imposing the boundary contitions at the point
x = x
right
, which is assumed to be the right end of the integration interval. In a
130
conservative numerical scheme of Godunov type, based on the updating of cell
averages by the accumulation of ux contributions to the each cell, the boundary
values affect the solution only through the evaluation of the ux on the boundary.
The boundary value of the unknown is either an indepedent external datum, to
be provided from outside the computational domain, or is simply the boundary
value of the solution, known from the previous time step. The choice between
these two possibility depends on the inow/ouow nature of the boundary (at the
considere time), at least in the case of a single equation
t
u +
x
f (u) = 0 for a
scalar unknown u. To have a compact notation, let us denote the trace of already
computed solution at time t as follows
u

(t ) u(x, t )
|x=x
right
= u(x
right
, t ). (12.10)
To refer the available boundary data and the known boundary values of the already
computed solution by means of one and the same mathematical symbol let us dene
the function u(t ), dened only on the right boundary, as follows
u(t ) =
_
u

(t ) if a(u(x
right
, t )) 0,
u
ext
(t ) if a(u(x
right
, t )) < 0,
(12.11)
where a(u) = f

(u) is the advection speed. Thus, the alternative corresponds to a
right end which is an outow or an inow boundary. Here the function u
ext
(t ) is
assumed to be given and can be in particular a prescribed constant value.
In a nite volume discretization of the nonlinear conservation law, the boundary
value u(t ) is taken into account through the evaluation of the ux by means of
f = f (u).
12.4 Steps of the boundary procedure for a system
For a system the situation is much more complicated. In fact, there are more
than one variables dened at any boundary point and therefore some of them
are to be taken from outside, as prescribed boundary conditions, and some other
must be extracted from the boundary trace of the solution itself. As anticipated,
the extreme difculty of imposing the boundary conditions in hyperbolic systems
comes from such a hybrid nature. Mathematically speaking, the complication is
due to the difculty in dening the vector quantity to be used to evaluate ux on
the boundary incorporating the correct number of appropriate informations from
the exterior of the computational domain and the same time the proper elements
of the solution already determined at the considered time.
To solve this problem a method was suggested by the rst author. The method
can be implemented in two alternative ways which are described in the following
131
sections. The rst procedure is based on considering variations of the characteristic
variables while the second relies upon the evaluation of the characteristic variables
themselves.
Step 0. Preliminary: Eigenstructure
The preliminary computation is the determination of the eigenstructure of the local
solution on the boundary. This means to compute initially the eigenvalues of the
Jacobiam matrix A(u

(u

) = 1, 2, . . . m, (12.12)
which are assumed to be set in an increasing order. Here and throughout the chapter
the number of equations of the systemis indicated by m instead of p since the letter
p is already engaged here to denote the pressure as well the physical variables.
The eigenvalues are independent of the variables chosen to formulate the eigen-
value problem, so that they can be referred to, collectively, as the vector

1
,

2
, . . . ,

m
). (12.13)
Then, to complete the solution eigenstructure on the boundary, the matrices of the
right and left eigenvectors are computed
R(u

) and L(u

) (12.14)
normalized so that L(u

) R(u

) = I. We will denote the number of strictly negative


eigenvalues by k, with 0 k m. This number is independent of the variables
chosen to formulate the eigenvalue problem In the rst form of Guardones pro-
cedure considered here, the value of k is taken into account only in an implicit
manner through an operator that selects the components of a vector dependening
on the sign of their corresponding eigenvalues, as it will be shown later.
Step 1. Determination of the characteristic variation
Let us suppose that the value of all of the physical variables p is known outside the
right extreme of the integration interval and let us denote the vector of these data
by
p
ext

_
_
_
_
_
p
ext
1
p
ext
2
.
.
.
p
ext
m
_
_
_
_
_
. (12.15)
In practice only some of the components of p
ext
will be needed in any specic
circumstance, but the selection of the proper components that must be actually
taken into account will be achieved automatically, see below.
132
Let us introduce the difference between the solution vector u

and the value of


conservation variable u(p
ext
) corresponding to the external data p
ext
:
u u

u(p
ext
). (12.16)
The variation of the characteristic variables is dened simply by
v L(u

)u = L(u

)
_
u

u(p
ext
)
_
. (12.17)
Step 2. Selection of the characteristic boundary variations
Each component of the characteristic variation is nowdened according to the sign
of the corresponding eigenvalue. A negative sign means that the characteristic line
enters the right extreme of the integration interval, implying that the boundary value
of the corresponding component of the characteristic variation must be chosen from
the vector v. On the contrary a positive sign means that the characteristic line
is sorting out from the right end of the interval so that the boundary value of the
conservative vector u must be chosen from the vector u

of the solution on the


boundary. This corresponds to a zero value for the component of the characteristic
variation. In formula we have
v


_
_
_
v

if

< 0
0 if

0
(12.18)
for = 1, 2, . . . , m. The complete barred characteristic variation results in
v =
_
_
_
_
_
v
1
v
2
.
.
.
v
m
_
_
_
_
_
. (12.19)
We can formalize this step by introducing an operator acting on a vector y which
selects its components only when the corresponding eigenvalue is strictly negative,
and puts zero otherwise. Explicitely, we dene SN

the operator selecting the


part corresponding to the strictly negative eigenvalues, as follows
SN

y
_
_
_
_
_
_
_
_
_
_
_
_
_
y
1
if
1
< 0
0 if
1
0
y
2
if
2
< 0
0 if
2
0
.
.
.
y
m
if
m
< 0
0 if
m
0
_
_
_
_
_
_
_
_
_
_
_
_
_
. (12.20)
133
In terms of this operator, the relation dening the characteristic variation v on
the right extreme of the integration interval can be written in the following form
v = SN

v = SN

L(u

)
_
u

u(p
ext
)
_
. (12.21)
But the characteristic variables are not the unknowns of the nonlinear hyperbolic
problem, thus we needa last steptoreturnbacktothe original conservative variables
of the system.
Step 4. Back transformation to the conservative variables
This step involves the inverse transformation fromthe characteristic variables to the
conservative ones. We will use the matrix R(u

) to performthe back transformation


of the variations in the form
u = R(u

) v. (12.22)
The nal expression for the sought for barred vector of the conservation variables
on the right boundary is
u = u

u
= u

R(u

) v
= u

R(u

) SN

L(u

)
_
u

u(p
ext
)
_
.
(12.23)
By means of the uthe boundary ux f = f(u) can be determined and used eventually
for the time advancement of the discretized hyperbolic system.
The treatment of the boundary condition at the left extreme x = x
left
is similar
and the nal expression for the sought for barred vector of the conservation
variables on the left boundary is
u = u

R(u

) SP

L(u

)
_
u

u(p
ext
)
_
. (12.24)
where one has introduced the operator SP

that selects the components of the


vector argument when the corresponding eigenvalue is strictly positive and sets
zero otherwise.
12.5 Alternative boundary procedure
An alternative implementation of the procedure suggested by the rst author relies
on the characteristic variables rather than on their variation. This choice allows a
certain degree of control in the explict selection of the external value to be provided
to the program depending on the number of the negative eigenvalues on the right
end (or the positive eigenvalues on the left end).
134
Step 0. Preliminaries
A rst preliminary computation is the determination of the boundary values of the
primitive variables from the known solution restricted to the boundary:
p

p(u

). (12.25)
A second preliminary calculation is the determination of the eigenstructure of the
local solution on the boundary. This means to compute initially the eigenvalues of
the Jacobian matrix A(u

(u

) = 1, 2, . . . m, (12.26)
which are assumed to be set in an increasing order. These values are independent
of the variables chosen to formulate the eigenvalue problem, so that they can be
referred to, collectively, as the vector

1
,

2
, . . . ,

m
). (12.27)
Then, to complete the solution eigenstructure on the boundary, the matrices of the
right and left eigenvectors are computed
R(u

) and L(u

) (12.28)
normalized so that L(u

) R(u

) = I. We will denote the number of strictly negative


eigenvalues by k, with 0 k m. This number is independent of the variables
chosen to formulate the eigenvalue problem
After the eigenproblem has been solved, the characteristic variables on the
boundary can be determined by the explicit evaluation
v

L(u

)u

= 1, 2, . . . m, (12.29)
This vector is therefore only an alternative representation of the same information
contained in the trace of the computed solution. At this point we have the three
vectors u

, p

and v

. Moreover, the number k of negative eigenvalues is known,


together with the two matrices R(u

) and L(u

).
Step 1. Collecting the BVs of the physical variables
On the right end of the interval, the number k of negative eigenvalues gives the
number of boundary values that must be provided from the outside. Just to x the
ideas, suppose that the available data are the rst k components of the primitive
variable vector p, denoted as p
ext
1
, p
ext
2
, . . . , p
ext
k
. Then, the remaining mk values
135
will be taken from the last m k components of p

, namely, p

k+1
, . . . , p

m
. Thus,
one can collect these two sets of values in the primitive vector:
p
_
_
_
_
_
_
_
_
_
_
_
_
p
ext
1
p
ext
2
.
.
.
p
ext
k
p

k+1
.
.
.
p

m
_
_
_
_
_
_
_
_
_
_
_
_
. (12.30)
We use the special accent _ to denote the collected variables. This notation em-
phasizes that these values are different form the barred boundary values needed
to generalize the scalar procedure to the system and which must specied for the
characteristic variables. Thus to obtain the latter values the physical variables
vector p must be transformed into its counterpart for the characteristic variables.
We must notice that the choice of the k components xed externally is arbitrary,
and they are not obliged to be the leading one. As a matter of facts, the choice
depends on some physical property of the boundary: for instance, on a xed wall
the velocity is specied and must be zero.
Step 2. From the collected values to conservative variables
To arrive at the characteristic variables it is rst necessary an intermediate passage
throught the conservative variables. This is achieved by the explicit evaluation
u u( p). (12.31)
This transformation is nonlinear and can involve also the complex functional de-
pendence associated with the thermodynamic equation of state.
Step 3. From the collected values to characteristic variables
The collected values of the characteristic variables could be provided by the matrix
multiplication L( u) u. But this relation involves a matrix different from L(u

) that
was used to establish the number k of boundary data. It is therefore more convenient
not to depart from the basic eigenstructure used in other part of the procedure and
to approximate the transformation above by means of of the following one
v L(u

) u. (12.32)
At this point we have also the three sets of collected values p, u and v, with the
latter affected by an approximation error. We are ready to dene the barred
charateristic variables.
136
Step 4. Selection of BVs for the characteristic variables
Each component of the characteristic vector is now dened according to the sign
of the corresponding eigenvalue. A negative sign means that the characteristic line
enters the right extreme of the integration interval, implying that the boundary value
of the corresponding characteristic component must be chosen from the vector v
of the collected characteristic values. On the contrary a positivs sign means that
the characteristic line is sorting out from the right end of the interval so that the
boundary value must be chosen from the vector v

of the characteristic solution


on the boundary. In formula we have
v


_
_
_
v

if

< 0,
v

if

0.
(12.33)
The complete barred characteristic vector is given by
v =
_
_
_
_
_
v
1
v
2
.
.
.
v
m
_
_
_
_
_
. (12.34)
We can formalize this step by introducing an operator acting on a vector y which
selects its components only when the corresponding eigenvalue is strictly negative,
and puts zero otherwise. Explicitely, we dene SN

the operator selecting the part


corresponding to the strictly negative eigenvalues, as follows
SN

y
_
_
_
_
_
_
_
_
_
_
_
_
_
y
1
if
1
< 0
0 if
1
0
y
2
if
2
< 0
0 if
2
0
.
.
.
y
m
if
m
< 0
0 if
m
0
_
_
_
_
_
_
_
_
_
_
_
_
_
. (12.35)
In the same vein, we dene the operator NN

acting on y which selects the


components of y only when the corresponding eigenvalue is not negative (i.e.,
137
positive or zero):
NN

y
_
_
_
_
_
_
_
_
_
_
_
_
_
0 if
1
< 0
y
1
if
1
0
0 if
2
< 0
y
2
if
2
0
.
.
.
0 if
m
< 0
y
m
if
m
0
_
_
_
_
_
_
_
_
_
_
_
_
_
(12.36)
In terms of the two operators, the relation dening the characteristic vector v on
the right extreme of the integration interval can be written in the following form
v = SN

v +NN

. (12.37)
Expressing v = L(u

) u = L(u

)u( p) from the previous Step 3 and using the


denition of the characteristic vector v

= L(u

)u

, we have
v = SN

L(u

)u( p) +NN

L(u

)u

. (12.38)
But the characteristic variables are not the unknowns of the nonlinear hyperbolic
problem, thus we needa last steptoreturnbacktothe original conservative variables
of the system.
Step 5. Back transformation to the conservative variables
This step involves the inverse transformation fromthe characteristic variables to the
conservative ones. In the present case, it would read u = R(u)v, but, as already
mentioned, this relation constitutes only an implicit denition of the unknown
u. Thus, we are somewhat obliged to approximate the relation by the explicit
denition
u R(u

)v (12.39)
in which the same matrix R(u

) of the procedure is employed. The nal expres-


sion for the sought for barred vector of the conservation variables on the right
boundary is
u = R(u

)
_
SN

L(u

)u( p) +NN

L(u

)u

_
right extreme. (12.40)
By means of the uthe boundary ux f = f(u) can be determined and used eventually
for the time advancement of the discretized hyperbolic system.
The treatment of the boundary condition at the left extreme x = x
left
is similar
and the nal expression for the sought for barred vector of the conservation
variables on the left boundary is
u = R(u

)
_
SP

L(u

)u( p) +NP

L(u

)u

_
left extreme. (12.41)
138
Here one has introduced the operators SP

and NP

that select the components


of the vector argument when the corresponding eigenvalue is, respectively, strictly
positive and not positive (i.e., negative or zero).
0.2 2 2.2
2.843
3
4
4.651
1
2.843
3
4
4.651
u
x
11 00 0.2 2 2.2
0.3063
2
3
3.368
2
0.3063
2
3
3.368
u
x
1
1
1
1
0
0
0
0
Figure 12.1: Comparison of density and momentumof the numerical solution with
the exact solution in the shock reection problem
139
Appendix A
Discrete differential operators
in node-pair weak form
For the readers convenience, in this appendix the node-pair representation of
selected dicrete differential operators is recalled. Both scalar and vector piecewise
linear functions are considered, to be intended as FEM approximations of scalar
and vector functions dened over , respectively, as
s(x, t ) s
h
(x, t )
def
=

kN
s
k
(t )
k
(x),
and
v(x, t ) v
h
(x, t )
def
=

kN
v
k
(t )
k
(x).
It is remarkable that the node-pair representation described below is strictly equiv-
alent to its FEM counterpart but for the last relations, namely, the node-pair repre-
sentation of a diffusive operator for a nonconsatnt diffusion coefcient .
Gradient in weak form, s R
_

i
s
h
=

kN
i,=
s
i
+s
k
2

i k
+

kN

i,=
s
k
s
i
2

i k
+ s
i

i
Divergence in weak form, v R
d
_

i
v
h
=

kN
i,=
v
k
+v
i
2

i k
+

kN

i,=
v
k
v
i
2

i k
+ v
i

i
140
Curl in weak form, v R
d
_

i
v
h
=

kN
i,=
v
k
+v
i
2

i k

kN

i,=
v
k
v
i
2

i k
v
i

i
Laplacian operator, s R, constant diffusion coefcient
_

i
(
i
) s
h
=

kN
i,=
K
i k
(s
k
s
i
) +

k(N

i
N

N
)
_
_

ik

i

k
_
b
k
Laplacian operator, s R, nonconstant diffusion coefcient
_

i
(
i
) s
h

kN
i,=

i k
K
i k
(s
k
s
i
) +

k(N

i
N

N
)
_
_

ik

i

k
_
b
k
,
where

i k
def
=

k
+
i
2
141
Appendix B
Nomenclature

i
Shape function associated with node i .
Computational domain.
Boundary of .

i
Support of the shape function
i
.

i k
Intersection of the supports
i
and
k
of the shape functions
i
and
k
:
i k
=
i

i
.

i k
Intersection of boundary
i
and
k
of the supports
i
and
k
of the shape functions
i
and
k
:
i k
=
i

i
.

e
Subset of pertaining to the e-th element.
N Set of all mesh nodes.
N

Set of all nodes belonging to the boundary .


N
i
Set of nodes belonging to the support
i
of
i
.
N
i,=
Set of nodes belonging to the support
i
of
i
, excluding node i ,
namely N
i,=
def
= N
i
{i }.
142
N
e
Set of nodes belonging to the element
e
.
E Set of all elements belonging to the nite element mesh.
E
i
Set of the elements belonging to the support
i
of
i
.
M
i k
Generic element of the consistent mass matrix.
K
i k
Generic element of the stiffness matrix.
L
i
Generic element of the diagonally lumped mass matrix, namely
L
i
=

kN
i
M
i k
.

i k
_

ik
(
i

k

k

i
).

i k
_

ik

k
n, for any i, k N

i
_


i
n, for any i N

,e
i k
_

ik

e

i

k
n, for any i, k N

,e
i
_

e

i
n, for any i N

.
[
h

h
]
i k
_

ik

i

k
.
f Flux vector.
a Advection velocity vector.
n Outward normal unit vector.
143
Appendix C
Algorithms
As an example, we write here the FORTRAN 90 code that computes the right
hand side and updates the solution to a conservation law for a scalar unknown u
by means of the TG scheme with mass lumping in node-pair formulation. The
vector-valued functions dening the ux f (u) and its derivative d f (u)/du are
assumed to be available. The values u at all boundary nodes and are assumed to be
given in a known vector indicated by ubar. The list of boundary node is assumed
to be augmented, as described in section 4.4.1, to allow for different boundary
conditions to be imposed at different elements on the boundary. The identier
qc denotes the integrated ux q at the interface or their contributions, while rhs
indicates the right-hand side to be used in the solution update.
rhs = 0
! ===================
! DOMAIN CONTRIBUTION
! ===================
DO c = 1, Number_of_Domain_node_pairs
! Extract nodes i and k of the node-pair
! couple c from node <-- node-pair connectivity
i = j_c_D(1,c); k = j_c_D(2,c)
ui = uu(i); uk = uu(k)
! Contributions of node-pair couple c ...
144
qc = - SUM(eta(:,c) * (flux(ui) + flux(uk))/2
ac = (dflux_du(ui) + dflux_du(uk))/2
a2c_T = tensor_product(ac, ac)
qc = qc - (dt/2) * SUM(K_T(:,:,c) * a2c_T) * (uk - ui)
rhs(i) = rhs(i) + qc
rhs(k) = rhs(k) - qc
ENDDO
! ==========================
! BOUNDARY-PAIR CONTRIBUTION
! ==========================
! Boundary node-pairs must be ARTIFICIALLY DOUBLED ONLY IN 3D
! ===========================================================
DO c = 1, Number_of_Boundary_node_pairs
! Extract nodes i and k of the node-pair
! couple c from node <-- node-pair connectivity
i = j_c_D(1,c); k = j_c_D(2,c)
! Extract nodes i and k of the node-pair
! couple c in the boundary numbering
i_ = j_c_B(1,c); k_ = j_c_B(2,c)
ubi_ = ubar(i_); ubk_ = ubar(k_)
! Flux-difference contribution
! ============================
! The contribution of node-pair couple c ...
qc = - SUM(chi(:,c) * (flux(ubk_) - flux(ubi_)))/2
145
! [In 3D: chi <--- chi_hat]
! ... is accumulated in the right components of rhs
rhs(i) = rhs(i) + qc
rhs(k) = rhs(k) - qc
! Boundary contribution of the TG term
! ====================================
! NOT YET IMPLEMENTED
ENDDO
! ======================================
! BOUNDARY NODAL CONTRIBUTION TO THE RHS
! ======================================
! The boundary node must be ARTIFICIALLY DOUBLED
! IN 2D OR ARTIFICIALLY MULTIPLICATED IN 3D
! ==============================================
DO i_ = 1, Number_of_Boundary_nodes
! Extract the corresponding node i
! in the domain numbering
i = jD_jB(i_)
! The contribution of node i is accumulated in rhs
rhs(i) = rhs(i) - SUM(xi_hat(:,i_) * flux(ubar(i_)))
ENDDO
! SOLUTION UPDATE
! ================
uu = uu + dt * rhs
146
We provide also an alternative implementation of the same algorithm in which the
interface integrated uxes are memorized in vector qc for all interfaces before
proceding to accumulate them in the update phase.
! ============================================
! DOMAIN CONTRIBUTION TO THE INTEGRATED FLUXES
! ============================================
DO c = 1, Number_of_Domain_node_pairs
i = j_c_D(1,c); k = j_c_D(2,c)
ui = uu(i); uk = uu(k)
qc(c) = - SUM(eta(:,c) * (flux(ui) + flux(uk))/2
ac = (dflux_du(ui) + dflux_du(uk))/2
a2c_T = tensor_product(ac, ac)
qc(c) = qc(c) - (dt/2) * SUM(K_T(:,:,c) * a2c_T) * (uk - ui)
ENDDO
! ===================================================
! BOUNDARY-PAIR CONTRIBUTION TO THE INTEGRATED FLUXES
! ===================================================
DO c = 1, Number_of_Boundary_node_pairs
i = j_c_D(1,c); k = j_c_D(2,c)
i_ = j_c_B(1,c); k_ = j_c_B(2,c)
ubi_ = ubar(i_); ubk_ = ubar(k_)
qc(c) = qc(c) - SUM(chi(:,c) * (flux(ubk_) - flux(ubi_)))/2
! Boundary contribution of the TG term
! ====================================
! NOT YET IMPLEMENTED
ENDDO
147
! ====================================
! SOLUTION UPDATE FOR INTERFACE FLUXES
! ====================================
DO c = 1, Number_of_Domain_node_pairs
i = j_c_D(1,c); k = j_c_D(2,c)
uu(i) = uu(i) + dt * qc(c)
uu(k) = uu(k) - dt * qc(c)
ENDDO
! ==============================================
! BOUNDARY NODAL CONTRIBUTION TO SOLUTION UPDATE
! ==============================================
DO i_ = 1, Number_of_Boundary_nodes
i = jD_jB(i_)
uu(i) = uu(i) - dt * SUM(xi_hat(:,i_) * flux(ubar(i_)))
ENDDO
148
As a second example, the FORTRAN 90 code is provided for solving a system of
p conservation laws for a vector unknown u with p components by means of the
same TG scheme with mass lumping in node-pair formulation used in the previous
scalar problem. The vector unknown u is denoted by vv. The ux array f(u)
and its Jacobian array A(u) = f(u)/u are evaluated at u by the array-valued
functions fv and Av, respectively, and are then memorized in the arrays fi and
Ai. The vector values u at all boundary nodes and are assumed to be given in a
known array indicated by vbar. The algorithmis implemented in the version with
the update of the solution done after the construction of the uxes at all interfaces.
! ============================================
! DOMAIN CONTRIBUTION TO THE INTEGRATED FLUXES
! ============================================
DO c = 1, Number_of_Domain_node_pairs
i = j_c_D(1,c); k = j_c_D(2,c)
vi = vv(:,i); vk = vv(:,k)
fi = fv(vi); fk = fv(vk)
! Ai = Av(vi); Ak = Av(vk)
Ac = (Av(vi) + Av(vk))/2 ! Ac(d, p,p)
A2c_T = matrix_tensor_product(Ac, Ac) ! A2c_T(d,d, p,p)
DO l = 1, p; DO l1 = 1, p
K_A2c(l,l1) = SUM(K_T(:,:,c) * A2c_T(:,:, l,l1))
ENDDO; ENDDO
DO l = 1, p
qc(l,c) = - SUM(eta(:,c) * (fi(:,l) + fk(:,l)))/2 &
- (dt/2) * SUM(K_A2c(l,:), vk - vi)
ENDDO
ENDDO
149
! ===================================================
! BOUNDARY-PAIR CONTRIBUTION TO THE INTEGRATED FLUXES
! ===================================================
DO c = 1, Number_of_Boundary_node_pairs
i = j_c_D(1,c); k = j_c_D(2,c)
i_ = j_c_B(1,c); k_ = j_c_B(2,c)
vbi_ = vbar(i_); vbk_ = vbar(k_)
fbi_ = fv(vbi_); fbk_i = fv(vbk_)
! Flux-difference contribution
! ============================
DO l = 1, p
qc(l,c) = qc(l,c) &
- SUM(chi(:,c) * (fbk_(:,l) - fbi_(:,l)))/2
ENDDO
! Boundary contribution of the TG term
! ====================================
! NOT YET IMPLEMENTED
ENDDO
! ====================================
! SOLUTION UPDATE FOR INTERFACE FLUXES
! ====================================
DO c = 1, Number_of_Domain_node_pairs
i = j_c_D(1,c); k = j_c_D(2,c)
vv(:,i) = vv(:,i) + dt * qc(:,c)
vv(:,k) = vv(:,k) - dt * qc(:,c)
ENDDO
! ==============================================
150
! BOUNDARY NODAL CONTRIBUTION TO SOLUTION UPDATE
! ==============================================
DO i_ = 1, Number_of_Boundary_nodes
i = jD_jB(i_)
fbi_ = fv(vbar(i_))
DO l = 1, p
vv(l,i) = vv(l,i) - dt * SUM(xi_hat(:,i_) * fbi_(:,l)))
ENDDO
ENDDO
151
Appendix D
Error analysis of the
TaylorGalerkin method
In this appendix we will rst describe the TaylorGalerkin scheme of third-order
temporal accuracy and recall its basic numerical properties as far as amplitude and
phase-speed errors are concerned. After pointing out the difculties faced by the
application of this scheme to nonlinear hyperbolic equations, a two-step version of
the third-order TGscheme is described which can be easily used to solve nonlinear
problems without lowering the high phase accuracy of its single-step counterpart
and which presents some stability advantages in multidimensional problems. This
method has been introduced by Selmin [22]. The two-step strategy is then exploited
a step further by introducing two newfourth-order accurate TGschemes, the second
of which is found to be characterized by a domain of numerical stability which is
exactly isotropic. Finally, the nonlinear vector equation governing a vector eld
which advects itself is examined and the second-order accurate TaylorGalerkin
scheme proposed for its solution by Laval [14] (see also Laval and Quartapelle
[15]) is described. This scheme is here used to construct a fourth-order accurate
two-step scheme for dealing with the advection part of the momentum equation.
D.1 Basic third-order TG scheme
The TaylorGalerkin scheme of third-order temporal accuracy has been introduced
by Donea [3] to solve transient advection problems by nite elements. This scheme
is derived here to have at hands its amplication factor for subsequent comparison
with two-step versions of TG schemes of high-order accuracy.
In contrast with the procedure commonly followed in the nite element ap-
proach for transient problems, wherein the spatial approximation precedes the
time discretization, in the TaylorGalerkin method the time discretization is taken
152
as the rst step of the discretization process and is prearranged so as to match
the high spatial accuracy achieved by linear elements (Donea [3]). Consider the
explicit Euler time-stepping applied to the linear advection equation

t
u +a
x
u = 0 (D.1)
with the spatial variable x left continuous and where
t
and
x
denote the differen-
tiation with respect to time and space variabls t and x. A Taylor series expansion
in the time step provides
u
n+1
= u
n
+t
t
u
n
+
1
2
(t )
2

2
t
u
n
+
1
6
(t )
3

3
t
u
n
+O
_
(t)
4
_
.
Now, the governing equation and its successive derivatives with respect to time
imply that
t
u = a
x
u,
2
t
u = a
2

2
x
u and
3
t
u = a
2

2
x

t
u. The derivative in
the right-hand side of the third equation is written in mixed temporalspatial form
to avoid third-order spatial derivatives which would require C
1
continuity of the
interpolation space.
Substituting these expressions into the Taylor series above and rearranging
some terms provides the following semidiscrete equation
_
1
1
6
a
2
(t )
2

2
x
_
u
n+1
u
n
t
= a
x
u
n
+
1
2
a
2
t
2
x
u
n
. (D.2)
The standard Galerkin formulation is now applied using a basis of linear elements
{
i
(x)} over a uniform mesh of size h. The solution u
n
(x) at the time level t
n
is
approximated by means of the expansion
u
n
(x) u
n
h
(x) =

kN
u
n
k

i
(x),
where u
n
h
= {u
n
k
} is the vector of the nodal values of the unknown. Substituting
the expansion in the semidiscrete equation and rendering the resulting equation
orthogonal to the space of the basis functions provides, after integrating by parts
the terms with the second-order spatial derivative,
_
1 +
1
6
(1
2
)
2
_
_
u
n+1
h
u
n
h
_
=
0
u
n
h
+
1
2

2
u
n
h
, (D.3)
where = at /h is the Courant number and where the standard notations for the
difference operators, namely,
_

0
{u
i
, i }
_
k
=
1
2
(u
k+1
u
k1
),
_

2
{u
i
, i }
_
k
= u
k1
2u
k
+u
k+1
,
153
have been used.
The amplication factor of this TaylorGalerkin scheme of third-order accuracy
is
G
3
TG
(, ) = 1 +
i sin 2
2
sin
2 1
2

1
2
3
(1
2
) sin
2 1
2

,
where = kh is the dimensionless wave number (k is the dimensional wave
number of the Fourier mode e
ikx
). The numerical properties of the scheme are
shown in Figure D.1 which contains the polar diagrams of the modulus of the
amplication factor |G(, )| and the phase speed ratio

num

ex
=
arg(G(, ))

.
The condition of numerical stability is found to be || 1. It coincides, of course,
with the necessary condition provided by the modied equation (Warming and
Hyett [28]) associated with the discrete equation (D.3)

t
w +a
x
w =
1
24
a
2
h
2
t (1
2
)
4
x
w
+
1
180
ah
4
(1 5
2
+4
4
)
5
x
w + .
The third-order TGmethodis referredtoas TGor TG3scheme. (It has alsobeen
calledEuler/TaylorGalerkinandLaxWendroff/TaylorGalerkinscheme whenits
numerical properties have been compared with those of other schemes derived from
the Euler and LaxWendroff time-stepping algorithms.) A detailed comparison of
this TG method with other nite element methods for solving transient advection
problems, including Semi-Lagrangian, Characteristic and PetrovGalerkin meth-
ods, is given by Donea and Quartapelle [4].
The basic distinguishing features of the TaylorGalerkin method are pointed
out by the following remarks.
1. The scheme has no free or adjustable parameter.
2. The scheme possesses the so-called unit CFLproperty, i.e., the exact solution
is obtained when the characteristic lines pass through the nodes.
3. The third-order term transforms the consistent mass matrix M = 1 +
1
6

2
into the TaylorGalerkin mass matrix
M
TG
() = 1 +
1
6
(1
2
)
2
.
Such a generalized mass matrix depends on the Courant number but is still
symmetric. For one-dimensional equations the TGmass matrix is tridiagonal
154
|G|

num
/
ex
0.5
0.75
1
0.25
0.75
1
0.25
0.50
Figure D.1: Numerical properties of the third-order TaylorGalerkin scheme
TG for some values of the Courant number .
and in multidimensional problems it has the same prole of the stiffness
matrix. Generally, the matrix is characterized by a diagonal dominance
that allows an approximate but accurate solution of the corresponding linear
system with a few Jacobi iterations or by other iterative methods.
4. Despite the appearances, the last term in equation (D.3) is not a (second-
order) numerical diffusion inherent to the scheme. In fact, in the solution of
transient problems the second-order termis only an element of the improved
approximation to
t
u with respect to the simple explicit Euler algorithm.
This does not mean that the scheme is free from numerical diffusion but
only that it has a fourth-order numerical diffusion, as shown explicitly by
the modied equation above.
5. The weak variational formulation of the semidiscrete equation (D.2) reads
_
,
u
n+1
u
n
t
_
+
1
6
t
_
a
x
,
x
(u
n+1
u
n
)
_
=
_
+
1
2
at
x
, a
x
u
n
_
,
where denotes the weighting function. Therefore, the TG scheme can
also be interpreted as a kind of PetrovGalerkin method in which the spatial
derivative term of the equation is weighted by means of the modied test
function

modif
= +
1
2
t a
x
.
It must be noted, however, that in the TG method the time derivative term is
weighted in a different manner, since the weighting of the various terms of
the equation results from, and is dictated by, the analysis of the truncation
155
error of the time-stepping algorithm. This is in contrast with most Petrov
Galerkin methods where the same modied test function is used consistently
for all terms of the equation.
6. The TaylorGalerkin method is applicable with very simple modications
also in the presence of a source term s, as in the equation
t
u +a
x
u = s.
Other TaylorGalerkin schemes for linear advection equations can be obtained
starting from the leap-frog and CrankNicolson time-stepping algorithms (see
Donea et al. [5]).
D.2 Two-step third-order TG scheme
Let us now examine the computational difculties which are encountered by the
application of the TaylorGalerkin method (D.3) to nonlinear problems. Consider,
for instance, the following scalar conservation-law equation in one dimension:

t
u +
x
f (u) = 0, (D.4)
where the ux f (u) is a known nonlinear function of u (hyperbolic equation written
in conservation form). Introducing the advection velocity a(u) = d f (u)/du, the
second and third time derivative of the unknown u can be expressed as follows

2
t
u =
t
[
x
f (u)] =
x
[
t
f (u)] =
x
[a(u)
t
u] =
x
[a(u)
x
f (u)],

3
t
u = (t )
1
_

x
_
a(u
n+1
)
x
f (u
n+1
)
_

x
_
a(u
n
)
x
f (u
n
)
_
_
.
Substituting them in the Taylor series yields the following third-order accurate
time-integration scheme for the nonlinear equation (D.4)
u
n+1
u
n
t

1
6
t
x
_
a(u
n+1
)
x
f (u
n+1
)
_
=
x
f (u
n
) +
1
3
t
x
_
a(u
n
)
x
f (u
n
)
_
.
The scheme would require to solve a nonlinear equation at each time step. How-
ever, this is not the only possibility for assuring third-order time accuracy in the
integration of the conservation-law equation (cf. Selmin et al.[24]). In fact, one
could also express the third-order time derivative according to the following sub-
156
stitutions

3
t
u =
t
_

x
[a(u)
t
u]
_
=
x
_

t
[a(u)
t
u]
_
=
x
_

t
a(u)
t
u +a(u)
2
t
u
_
=
x
_

t
a(u)
t
u +a(u)
x
[a(u)
t
u]
_
=
x
_

t
a(ut
t
u +a(u)[
x
a(u)
t
u +a(u)
2
t x
u]
_
=
x
_

t
a(u)
t
u +a(u)
x
a(u)
t
u +a(u)a(u)
2
t x
u
_
=
x
_
a

(u)
x
f (u)
t
u +a(u)
x
a(u)
t
u +a(u)a(u)
2
t x
u
_
.
Therefore, factoring out the quantity involving the time derivative gives

3
t
u =
x
__
a

(u)
x
f (u) +a(u)
x
a(u) +[a(u)]
2

x
_

t
u
_
.
The approximation
t
u = (u
n+1
u
n
)/t can be used to express the third-order
term so as to introduce it into the Taylor series and to obtain an alternative scheme
for the integration of the nonlinear equation. The nal equation obtained after
the spatial approximation can be written as a linear system of equations for the
incremental unknowns u
n+1
k
= u
n+1
k
u
n
k
. The three correction terms assuring
the third-order accuracy depend on u
n
i
, so that the matrix of the linear system
changes at each time level making the computational cost of the scheme very high.
This drawback is still present if the third-order correction is written in another
different form, using the governing equation in the rst two terms (which do not
involve the double space differentiation) appearing in the previous expression, to
give

3
t
u =
x
_
a

(u)[
x
f (u)]
2
a(u)
x
a(u)
x
f (u) +[a(u)]
2

x

t
u
_
.
Here, only one term is to be taken into account implicitly. In both cases, however,
the various implicit or explicit third-order correction terms are rather complicated
to evaluate, especially when dealing with system of conservation-law equations,
not to speak of multidimensional problems. In conclusion, we see that, in order to
achieve a third-order time accuracy in nonlinear problems with smooth solutions,
it is necessary to include complicated generalizations of the consistent mass ma-
trix which contain several nonsymmetric terms dependent on the solution at the
previous time level. For these reasons it would be attractive to have an alterna-
tive approach to assure the high-order temporal accuracy of the TaylorGalerkin
method in these situations. As a matter of fact, these difculties can be circum-
vented by formulating the TaylorGalerkin scheme according to two-step strategy
157
which can preserve the excellent phase-accuracy properties of the basic TGscheme
(Selmin [22]).
A third-order TG scheme suitable for nonlinear advection problems can be
obtained by considering the following two-step procedure
u
n
= u
n
+
1
3
t
t
u
n
+(t )
2

2
t
u
n
, (D.5a)
u
n+1
= u
n
+t
t
u
n
+
1
2
(t )
2

2
t
u
n
, (D.5b)
where the value of the parameter is left unspecied for the time being. In fact,
while the other coefcients in the equations (D.5) assume the value necessary for
a third-order accuracy of the two discretizations combined together, the parameter
enters only the coefcient of the fourth-order term in the overall series. As a
consequence, its value affects the amplication factor only within the accuracy of
the fourth order. Now, a convenient way of taking full advantage of this available
degree of freedomconsists in imposing that the phase speed of the two-step scheme
be coincident with that of the basic TaylorGalerkin scheme (D.3) in the linear case
(Selmin [22]). The fully discrete version of equations (D.5) in the case of the linear
advection equation in one dimension is
_
1 +
1
6

2
_
_
u
n
h
u
n
h
_
=
1
3

0
u
n
h
+
2

2
u
n
h
, (D.6a)
_
1 +
1
6

2
_
_
u
n+1
h
u
n
h
_
=
0
u
n
h
+
1
2

2
u
n
h
. (D.6b)
The amplication factor of this two-step linear scheme containing the parameter
is
G
3
TTG
(, ; ) = 1 +
i sin 2
2
sin
2 1
2


G(, ; )
1
2
3
sin
2 1
2

,
where

G(, ; ) = 1 +
i
1
3
sin 4
2
sin
2 1
2

1
2
3
sin
2 1
2

.
It is immediate to verify that
G
3
TTG
_
, ;
1
9
_
= G
3
TG
(, ) R(, ),
where R(, ) is the real function given by
R(, ) =
_
1
2
3
(1
2
) sin
2 1
2

_ _
1
2
3
(1 +
2
) sin
2 1
2

_
_
1
2
3
sin
2 1
2

_
2
.
Therefore, for =
1
9
the two-step procedure (D.6) reproduces exactly the excel-
lent phase-speed characteristics of the single-step third-order TG scheme. The
158
|G|
num
/
ex
0.75
0.5
0.25
0.25
0.75
0.5
Figure D.2: Numerical properties of the two-step third-order TaylorGalerkin
scheme TTG for some values of the Courant number .
condition of numerical stability is easily found to be ||

3/2. Figure D.2


contains the diagrams of the phase and amplitude errors of the two-step scheme
TTG (the interruption of the curve for = 0.75 is due to a jump in the evalua-
tion of an inverse trigonometric function by the computer). The new scheme is
more dissipative than its one-step counterpart but this difference is negligible for
all wavelengths except when is very close to the stability limit.
The modied equation associated with the two-step TG scheme is easily found
to be

t
w +a
x
w =
1
72
a
2
h
2
t (3
2
)
4
x
w
+
1
360
ah
4
(2 15
2
+8
4
)
5
x
w + .
The formassumedbythe two-stepscheme inthe case of the nonlinear conservation-
law equation
t
u +
x
f (u) = 0 is easily found to be
u
n
u
n
t
=
1
3

x
_
f (u
n
)
1
3
t a(u
n
)
x
f (u
n
)
_
, (D.7a)
u
n+1
u
n
t
=
x
_
f (u
n
)
1
2
t a(u
n
)
x
f (u
n
)
_
. (D.7b)
We notice the simplicity of the two equations as compared to any of the possible
equations provided by the single-step TG scheme. In particular, the favourable
effect of the complicated, solution-dependent, augmented mass matrix is achieved
simply through a double application of the usual consistent mass matrix.
The important point nowis that the phase properties of the one-step third-order
TG scheme can be reproduced exactly by the two-step procedure also in two and
three dimensions. To investigate these favourable numerical properties of the two-
step TG scheme applied to multidimensional equations, we consider the linear
159
advection equation

t
u + a u = 0. (D.8)
The single-step third-order accurate TG scheme for this equation is obtained with-
out difculty in the form
_
1
1
6
(t )
2
(a )
2
_
u
n+1
u
n
t
= a u
n
+
1
2
t (a )
2
u
n
. (D.9)
Let us assume that this equation is solved by means of nite elements, using a
bilinear approximation of the unknown u
n+1
over a uniform mesh. The ampli-
cation factor of the fully discrete equations of the single-step third-order scheme
is
G
3
TG
(, ) = 1 +
A(, ) +
1
2
K(, )
M
TG
(, )
,
M
TG
(, ) = M()
1
6
K(, ),
where is the dimensionless wave vector number and = (t /h)a is the vector
counterpart of the Courant number . The quantities A(, ) and K(, ) are the
Fourier transformof the operators t (a ) and t
2
(a )
2
, respectively, whereas
M() is the Fourier transform of the consistent mass matrix.
The application of the two-step procedure to the multidimensional advection
equation gives immediately the equations
u
n
u
n
=
1
3
t a u
n
+
1
9
(t )
2
(a )
2
u
n
, (D.10a)
u
n+1
u
n
= t a u
n
+
1
2
(t )
2
(a )
2
u
n
. (D.10b)
The amplication factor of the two-step TG scheme is easily found to be
G
3
TTG
(, ) = 1 +
A(, ) +
1
2
K(, )

G(, )
M()
,
where

G(, ) = 1 +

1
3
A(, ) +
1
9
K(, )
M()
.
It is not difcult to verify that the amplication factors of the two-step and single-
step third-order accurate TG schemes are connected by the following relationship
G
3
TTG
(, ) = G
3
TG
(, ) R(, ),
where the factor R(, ) is the real quantity
R(, ) = 1
_
1
6
K(, )
M()
_
2
.
160
Therefore, also in two and three dimensions the two-step procedure reproduces ex-
actly the phase-speed properties of the single-step third-order accurate TGscheme.
Moreover, the multidimensional two-step TG scheme presents the advantage
that its stability domain is bounded by essentially the same value as the scheme for
the equation in one dimension. For instance, the domain of numerical stability of
the two-step scheme for the equation in two dimensions is found to be very nearly
circular: the bound of the domain of numerical stability is found to be almost
coincident with a circle of radius || =

3/2 = 0.866, but for a slight attening


in the diagonal direction
x
=
y
on which the stability limit reads || < 0.854
(Selmin [22]).
This result is in contrast with what happens to the other single-step nite ele-
ment or difference schemes, such as the (single-step) LaxWendroff nite differ-
ence method, the second-order TaylorGalerkin scheme TG2 (that is the scheme
with M
TG
replaced by M) and also the third-order TG scheme: all these schemes
suffer from the same drastic reduction of the stability domain in multi dimensions.
For instance, due to its dependence on the direction of , the condition of nu-
merical stability of both LW/FD and TG schemes is || < 1/(2

2) in 2D and
|| < 1/(3

3) in 3D. These values are collected in the Table together with those
of other two-step TGschemes of fourth-order accuracy to be introduced in the next
section.
To illustrate the rle played by the surface integrals in the weak formulation of
the two-step scheme, we report here the weak equations obtained in the case of the
nonlinear conservation-lawequation u
t
+ f (u) = 0 in two or three dimensions:
_
,
u
n
u
n
t
_
=
1
3
_
, f (u
n
)
1
3
t a(u
n
) f (u
n
)
_

1
3
_

out
v n
_
f (u
n
)
1
3
t a(u
n
) f (u
n
)
_
,
_
,
u
n+1
u
n
t
_
=
_
, f (u
n
)
1
2
t a(u
n
) f (u
n
)
_

out
n
_
f (u
n
)
1
2
t a(u
n
) f (u
n
)
_
.
In conclusion, the two-step formulation of the TaylorGalerkin method, be-
sides making high-order accuracy accessible for truly nonlinear problems, offers
the additional advantage of giving an almost isotropic stability domain in multidi-
mensional problems.
161
Bounds of the Courant number
for numerical stability
1D 2D 3D
LW/FD 1
1
2

2
= 0.353
1
3

3
= 0.192
TG2, O
_
(t )
2
_
1

3
= 0.577
1
2

6
= 0.204
1
9
= 0.111
TG, O
_
(t )
3
_
1
1
2

2
= 0.353
1
3

3
= 0.192
TTG, O
_
(t)
3
_
0.854

=

3
2
= 0.866
TTG-4A, O
_
(t )
4
_
1
TTG-4B, O
_
(t )
4
_
0.84738
Table D.1:
D.3 Two-step fourth-order TG schemes
The two-stepstrategycanbe pursuedfurther toobtainfourth-order accurate Taylor
Galerkin schemes suitable for nonlinear problems. Consider the Taylor series
including terms up to the fourth-order derivative, namely,
u
n+1
= u
n
+t
t
u
n
+
1
2
(t )
2

2
t
u
n
+
1
6
(t )
3

3
t
u
n
+
1
24
(t )
4

4
t
u
n
+ O
_
(t )
5
_
.
A rst method achieving the fourth-order temporal accuracy of this expansion
consists in advancing the solution from u
n
to u
n+1
by means of the following
two-step procedure
u
n
= u
n
+
1
3
t
t
u
n
+
1
12
(t )
2

2
t
u
n
, (D.11a)
u
n+1
= u
n
+t
t
u
n
+
1
2
(t )
2

2
t
u
n
. (D.11b)
The application of the Galerkin method to the two equations gives, in the particular
case of the linear advection equation in one dimension solved by means of linear
162
|G|
num
/
ex
1
0.5 0.25
0.75
0.25
0.5
1
0.75
Figure D.3: Numerical properties of the two-step fourth-order Taylor
Galerkin scheme TTG-4A for some values of the Courant number .
elements,
_
1 +
1
6

2
_
_
u
n
h
u
n
h
_
=
1
3

0
u
n
h
+
1
12

2
u
n
h
, (D.12a)
_
1 +
1
6

2
_
_
u
n+1
h
u
n
h
_
=
0
u
n
h
+
1
2

2
u
n
h
. (D.12b)
The amplication factor of this two-step fourth-order TaylorGalerkin scheme,
which will be denoted as TTG-4A, is found to be
G
4A
TTG
(, ) = 1 +
i sin 2
2
sin
2 1
2


G
A
(, )
1
2
3
sin
2 1
2

,
where

G
A
(, ) = 1 +
i
1
3
sin
1
3

2
sin
2 1
2

1
2
3
sin
2 1
2

.
The scheme has a stability condition || 1 and its numerical properties can be
deduced from the polar plot of |G| and of the ratio
num
/
ex
given in Figure D.3.
In comparison with the TTG scheme, the fourth-order scheme appears to reach the
stability limit of 1 at the expense of lowering appreciably the phase accuracy of
the propagation of the intermediate and small wavelengths as || 1.
There is still another possibility of reaching the fourth-order temporal accuracy
within the two-step strategy. In fact, one can evaluate the rst-order term in the
second step in terms of the solution predicted in the rst step. In this case to
assure the fourth-order accuracy it is necessary to introduce some parameters in
163
the equations of the two steps, as follows:
u
n
= u
n
+t
t
u
n
+(t )
2

2
t
u
n
,
u
n+1
= u
n
+t
t
u
n
+ (t )
2

2
t
u
n
.
The values of the parameters , and are determined after substituting u
n
from
the rst equation into the second one, which gives
u
n+1
= u
n
+t
_

t
u
n
+t
2
t
u
n
+(t )
2

3
t
u
n
_
+ (t )
2
_

2
t
u
n
+t
3
t
u
n
+(t )
2

4
t
u
n
_
,
that is
u
n+1
= u
n
+t
t
u
n
+( + )(t )
2

2
t
u
n
+( + )(t)
3

3
t
u
n
+(t )
4

4
t
u
n
.
In order for this expansion to coincide with the Taylor series truncated after the
fourth-order term, the parameters in it must satisfy the following system of three
equations
+ =
1
2
, + =
1
6
, =
1
24
,
which gives the third-order algebraic equation for

1
2

2
+
1
6

1
24
= 0.
This equation has two complex solutions and one real solution which is given by
= ( +1)/6 with
=
_
1
2
_
5 +

29
__
1
3
+
_
1
2
_
5

29
__
1
3
= 1.1541715.
Thus the three parameters are determined as a function of from the relations

=
1
6
(2 ),

=
1
4( +1)
,

=
1
6
( +1),
and their numerical values are found to be

= 0.1409714,

= 0.1160538,

= 0.3590284.
For a linear advection equation in one dimension the resulting fourth-order accurate
TaylorGalerkin scheme (TTG-4B) assumes the form
u
n
u
n
=

t a
x
u
n
+

(t)
2
a
2

2
x
u
n
, (D.13a)
u
n+1
u
n
= t a
x
u
n
+

(t )
2
a
2

2
x
u
n
. (D.13b)
164
|G|
num
/
ex
0.75
0.5
0.25
0.25
0.75
0.5
Figure D.4: Numerical properties of the two-step fourth-order Taylor
Galerkin scheme TTG-4B for some values of the Courant number .
Its amplication factor for a Galerkin approximation over a uniformmesh of linear
elements is given by
G
4B
TTG
(, ) = 1 +
i sin 4

2
sin
2 1
2

1
2
3
sin
2 1
2

G
B
(, ),
where

G
B
(, ) = 1 +
i

sin 4

2
sin
2 1
2

1
2
3
sin
2 1
2

.
The stability limit of the TTG-4B scheme is found to be || 0.84738. The
numerical properties of the scheme can be inferred from the polar plots of |G|
and
num
/
ex
given in Figure D.3. We observe that the (limited) reduction of the
stability limit has permitted to recover very good phase response properties in the
full range of the Courant number.
The scheme can be applied to multidimensional nonlinear problems with no
difculty. Considering, for an illustrative purpose, the linear advection equation

t
u + a u = 0, (D.14)
it can be integrated in time according to the two-step scheme
u
n
u
n
=

t a u
n
+

(t)
2
(a )
2
u
n
, (D.15a)
u
n+1
u
n
= t a u
n
+

(t)
2
(a )
2
u
n
. (D.15b)
It can be noted that such a scheme is simpler to code than the previous two-step
TG schemes since the equation of the second step is a replica of that of the rst
step, but for the change of the constants which multiply the terms of the equations.
165
The amplicationfactor of the Galerkinapproximationtothis multidimensional
equation can be shown to be
G
4B
TTG
(, ) = 1 +
A(, ) +

K(, )
M()

G
B
(, ),
where

G
B
(, ) = 1 +

A(, ) +

K(, )
M()
.
The quantities M(), A(, ) and K(, ) are given in Appendix E. The stability
study of the TTG-4Bscheme shows that the domain of numerical stability is exactly
isotropic.
The latter scheme represents a rather important achievement in the eld of
numerical methods for advection problems in that it is fourth-order accurate both
in space and in time, is (slightly) dissipative, has a domain of numerical stability
exactly isotropic and is simpler to implement than the other two-step schemes.
D.4 Vector advection equation
We can nowexamine more closely the problemof solving the momentumequation
numerically. The difculties faced by nonlinear advection problems which have
been mentioned in section D.2 are here enhanced by the vector character of the
advection equation, which reads

t
u +(u )u = 0. (D.16)
In order to apply high-order two-step TG schemes to such an equation, a second-
order TaylorGalerkin method for its advancement in time is necessary. Such
a scheme has been introduced by Laval [14, 15] and is described here briey.
Consider the Taylor series expansion in the time step t, up to second order:
u
n+1
= u
n
+t
t
u
n
+
1
2
(t )
2

2
t
u
n
+O
_
(t )
3
_
.
The rst-order and second-order time derivatives in this series are then expressed
from the governing equation in the form
t
u = (u )u and

2
t
u =
t
[(u )u]
= (
t
u )u (u )
t
u
= ([(u )u] )u +(u )[(u )u],
166
respectively. Substituting the two expressions into the previous Taylor series gives
u
n+1
u
n
t
= (u
n
)u
n
+
1
2
t
__
[(u
n
)u
n
]
_
u
n
+(u
n
)[(u
n
)u
n
]
_
.
(D.17)
Equation (D.17) represents a time discretization of LaxWendroff type for the
vector advection equation, with second-order accuracy. As for the scalar linear
equation, the second-order spatial derivative terms are not a (second-order) numer-
ical diffusion or viscosity inherent to the scheme. These terms represent instead
an element of the improved nite difference approximation to the time derivative
with respect to that given by the explicit Euler algorithm.
In order to employ nite elements with only C
0
continuity (linear and multilin-
ear elements), the second-order spatial derivatives appearing in the time-discretized
equation (D.17) are nowtransformed into rst derivatives using the variational for-
mulation. The equation is multiplied by the vector-valued weighting function (x)
belonging to a suitable space, and the terms containing second-order derivatives
are integrated by parts. To simplify the calculation the weak expression of the
terms within the braces is written as
_

2
t
u =
_

_
(a )u +(u )a
_
,
where, by denition, a = (u )u. Now, the two terms under the integral on the
right-hand side are transformed by means of the vector identity
(a b) = (a )b +(b )a + ab + ba,
which holds for arbitrarily differentiable vector elds a and b, with b = u, to give
_

2
t
u =
_
[au ua +(a u)].
Integrating by parts the last two terms on the right-hand side provides
_

2
t
u =
_
au +
_
a (u)
+
_
n a(u)
_
a u +
_
n a u.
The termcontaining (u) is then transformed by means of the vector identity
(ab) = a b b a +(b )a (a )b
167
with a = u and b = , and the double cross product a(u) in the surface inte-
gral is expressed through the elementary identity a(bc) = (a c)b (a b)c.
The second-order term becomes nally
_

2
t
u =
_
au

_
a [ u +(u ) ( )u] +
_
a n u.
According to this result, the weak variational formulation of the complete advection
equation (D.17) yields, after some rearrangements,
_
, u
n+1
u
n
_
/t =
_
, (u
n
)u
n
_

1
2
t
_
_
(u
n
), (u
n
)u
n
_

_
(u
n
)u
n
n u
n
_
+
1
2
t
__
u
n
+( )u
n
u
n
_
, (u
n
)u
n
_
.
(D.18)
This expression represents the variational statement of the LaxWendroff scheme
for the nonlinear vector advection equation (D.16).
The various terms ensuring the second-order temporal accuracy admit the fol-
lowing interpretation. The most important term is the rst one, namely,
_
(u
n
), (u
n
)u
n
_
,
which is present even with the linearized version of the equation, with a uniform
advection velocity u
n
= constant. This term is block diagonal, i.e., the Cartesian
components of the vector equation are left uncoupled by the term. In the spatially
discrete case, each block gives rise to a symmetric matrix if u and are chosen to
belong to the same space (Galerkin method). Furthermore, each scalar block is of
the form
_
u
n
v, u
n
u
n
_
and is the weak form of the second directional derivative of u
n
in the direction
of u
n
(times |u
n
|
2
): therefore this term introduces a correction only when u
n
has
a nonzero curvature in the direction of the advection eld u
n
. By this property,
the scheme (D.17), with the third term on the right-hand side omitted, can be
interpreted as a second-order characteristic method for linear advection in multi
dimensions, for details see Laval [15] and also Donea and Quartapelle [4]. The
surface integral
_
[ (u
n
)u
n
] n u
n
168
resulting from the integration by parts must be evaluated only on the part
out
of the boundary where n u
n
> 0 (outow boundary). In fact, the boundary
condition associated with (D.16) is u
n+1
|

in
= b
n+1
, where
in

out
=
, so that the weighting functions satisfy the boundary condition |

in
=
0. By its very structure, such boundary term takes into account the outow of
uid momentum along the streamlines of the advection eld and guarantees good
absorbing properties at outow boundaries in transient calculations.
Coming now to the last volume term in (D.18), namely,
__
u
n
+( )u
n
u
n
_
, (u
n
)u
n
_
,
it vanishes when the velocity eld is uniform, and thus it represents a correction
caused by the spatial variations of the advection eld. The correction depends on
the non uniformity of u
n
measured by its three vector derivatives, namely, u
n
(vorticity), ( )u
n
(variation along the direction of ) and u
n
(dilatation).
The second-order TaylorGalerkin scheme (D.18) can be used to construct a
fourth-order accurate TaylorGalerkin method for the nonlinear advection equation
(D.16) by means of a two-step procedure described in the previous section, namely,
u
n
= u
n
+

t
t
u
n
+

(t)
2

2
t
u
n
, (D.19a)
u
n+1
= u
n
+t u
n
+

(t )
2

2
t
u
n
. (D.19b)
The two-step scheme applied to the vector advection problem of interest here
requires to determine a rst-step solution u
n
from the equation
_
,u
n
u
n
_
/t =

_
, (u
n
)u
n
_

t
_
_
(u
n
), (u
n
)u
n
_

out
(u
n
)u
n
n u
n
_
+

t
__
u
n
+( )u
n
u
n
_
, (u
n
)u
n
_
,
(D.20)
and then to evaluate the unknown u
n+1
from the second-step equation
_
, u
n+1
u
n
_
/t =
_
, (u
n
)u
n
_

t
_
_
(u
n
), (u
n
)u
n
_

out
(u
n
)u
n
n u
n
_
+

t
__
u
n
+( )u
n
u
n
_
, (u
n
)u
n
_
.
(D.21)
In both steps, the boundary integral has to be evaluated only on
out
and the
weightingfunctions are suchthat |

in
= 0. This methodhas beenimplemented
and used successfully to solve the incompressible Euler equations by means of the
fractional-step method and using bilinear/constant nite elements to approximate
velocity and pressure. The numerical results will be reported elsewhere.
169
D.5 Mass conservation equation
The TaylorGalerkin method can be applied also to conservation-law equations.
In this section we consider the equation expressing mass conservation, namely,

t
+ (u) = 0, (D.22)
where is the unknown mass density while u = u(r, t ) is a known velocity
eld, and derive both the second-order TG2 scheme and two-step third-order TTG
scheme appropriate for such an equation. As usual, we start with the Taylor series

n+1
=
n
+t
t

n
+
1
2
(t )
2

2
t

n
+O
_
(t )
3
_
,
up to second order. From the governing equation we have
t

n
= (u
n

n
),
where u
n
= u(r, t
n
). The second time derivative is obtained from

2
t
=
t
(u) =
_

t
(u)
_
=
_
u
t
+
t
u
_
=
_
u (u) +
t
u
_
=
_
u (u)
_

_

t
u
_
.
Passing to a time discretized version of this expression, we have

2
t

n
=
_
u
n
(u
n

n
)
_

n
u
n
u
n1
t
_
.
As a consequence, the second-order time-stepping for the mass conservation equa-
tion reads

n+1

n
t
= (u
n

n
) +
1
2
t
_
u
n
(u
n

n
)
_

1
2

n
(u
n
u
n1
)
_
,
which can be rewritten also as

n+1

n
t
=
_
_
3
2
u
n

1
2
u
n1
_

n
_
+
1
2
t
_
u
n
u
n

n
_
+
1
2
t
_
u
n

n
u
n
_
.
(D.23)
The weak Galerkin form of the scheme is, omitting the boundary terms for sim-
plicity,
_
,

n+1

n
t
_
=
_
,
_
3
2
u
n

1
2
u
n1
_

n
_

1
2
t
_
u
n
, u
n

n
_

1
2
t
_
, u
n

n
u
n
_
.
170
The last scalar product is evaluated after computing preliminarly the weak diver-
gence of u
n
. This means to determine the scalar function D
n
(r) expanded on the
basis of the linear interpolation by solving the mass matrix problem
_
, D
n
_
=
_
, u
n
_
.
Then, the weak Galerkin equation becomes
_
,

n+1

n
t
_
=
_
,
_
3
2
u
n

1
2
u
n1
_

n
_

1
2
t
_
u
n
, u
n

n
_

1
2
t
_
, u
n

n
D
n
_
.
We noware introduce an approximation in the rst and third termon the right-hand
side, consisting in replacing the product of the functions by their linear interpolation
based on the product of the nodal values. In other words, we introduce the nodal
(vector) quantities
F
n
k
=
_
3
2
u
n
k

1
2
u
n1
k
_

n
k
,

n
k
= u
n
k
D
n
k

n
k
,
and then dene the linear interpolations
F
n
(r) =

k
F
n
k

k
(r),

n
(r) =

n
k

k
(r).
The fully discrete equation assumes the simpler form
_
,

n+1

n
t
_
=
_
, F
n
_

1
2
t
_
u
n
, u
n

n
_

1
2
t
_
,
n
_
,
which can be simplied further to
_
,

n+1

n
t
_
=
_
, G
n
_

1
2
t
_
u
n
, u
n

n
_
(D.24)
by introducing
G
n
(r) = F
n
(r)
1
2
t
n
(r).
The construction of the third-order accurate two-step scheme TTG proceed in
a similar way. The weak equation for for the rst step is
_
,

n

n
t
_
=
_
, G
n

1
3
t
_
u
n
, u
n

n
_
, (D.25)
171
where
G
n

(r) = F
n

(r)
1
3
t
n
(r),
F
n

(r) =

k
_
4
3
u
n
k

1
3
u
n1
k
_

n
k

k
(r).
The equation of the second step is instead
_
,

n+1

n
t
_
=
_
,

G
n
_

1
2
t
_
u
n
, u
n

n
_
, (D.26)
with

G
n
dened as follows:

G
n
(r) =

F
n
(r)
1
2
t

n
(r),

F
n
(r) =

k
_
u
n
k
_

n
k
+
1
2

n
k
_

1
2
u
n1
k

n
k
_

k
(r),

n
(r) =

k
u
n
k
D
n
k

n
k

k
(r).
172
Bibliography
[1] Ambrosi D., and Quartapelle L. A TaylorGalerkin method for sim-
ulating nonlinear dispersive water waves. J. Comput. Phys. 146, 2 (1998),
546569.
[2] Callen H.B. Thermodynamics and an Introduction to Thermostatistics.
Wiley, 2nd edition, 1985.
[3] Donea J. A TaylorGalerkin method for convective transport problems.
Int. J. Numer. Meth. Eng. 20 (1984), 101119.
[4] Donea J., and Quartapelle L. An introduction to nite element meth-
ods for transient advection problems. Comput. Methods Appl. Mech. Eng. 95
(1992), 169203.
[5] Donea J., Quartapelle L., and Selmin V. An analysis of time dis-
cretization in the nite element solution of hyperbolic problems. J. Comput.
Phys. 70, 2 (1987), 463499.
[6] Donea J., Selmin V., and Quartapelle L. Recent developments of the
TaylorGalerkin methods for the numerical solutions of hyperbolic problems.
In Numerical Methods for Fluid Dynamics (1988), Morton K.W. and Baines
M.J., Eds., Oxford Univeristy Press.
[7] Ern A., and Guermond J. L. Elments nis: thorie, applications, mise
en uvre. Mathmatiques & Applications 36, Springer-Verlag, Berlin, 2002.
[8] Galgani L. and Scotti A. On subadditivity and convexity properties of
thermodynamic functions. Pure and Applied Chemistry, 22:229235, 1970.
[9] Godlewski E., and Raviart P. A. Numerical Approximation of Hyper-
bolic Systems of Conservation Laws. Springer-Verlag, New York, 1995.
[10] Guardone A. Nuovi risultati sui metodi per il calcolo di ussi comprimibili
su reticoli non strutturati. Tesi di laurea, Politecnico di Milano, Italy, 1998.
173
[11] Guardone A., and Vigevano L. Roe linearization for the van der Waals
gas. J. Comput. Phys. 175 (2002), 5078.
[12] Guardone A., and Quartapelle L. Exact Roes linearization for the
van der Waals gas. Proceedings of Godunov70 Conference, October 1999,
Oxford, England. Godunov Methods: Theory and Application, E. F. Toro,
Ed., Kluwer/Plenum Academic Press, 2001, 419424
[13] Landau L.D. and Lifshitz E.M. Course of Theoretical Physics, Volume
6, Fluid Mechanics. Pergamon Press, Oxford, 1959.
[14] Laval H. TaylorGalerkin solution of the time-dependent Navier-Stokes
equations. In International Conference on Computational Methods in Flow
Analysis (1988), Niki H. and Kawagara M, Eds.
[15] Laval H., and Quartapelle L. A fractional-step Taylor-Galerkin
method for unsteady incompressible ows. Int. J. Numer. Meth. Fluids 11
(1990), 501513.
[16] LeVeque R. J. Numerical Methods for Conservation Laws. Birkhuser,
Basel, 1992.
[17] Mottura L., Vigevano L., and Zaccanti M. An evaluation of Roes
scheme generalizations for equilibrium real gas ows. J. Comput. Phys. 138
(1997), 354399.
[18] Peraire J., Morgan K., and Peiro J. Numerical grid generation. Lec-
ture Series 06, von Karman Institute of Fluid Dynamics, 1990.
[19] Quartapelle L. Numerical Solution of the Incompressible Navier-Stokes
Equations. International Series of Numerical Mathematics. Birkhuser, 1993.
[20] Quarteroni A., and Valli A. Numerical Approximation of Partial Dif-
ferential Equations. No. 23 in Springer Series in Computational Mathematics.
Springer-Verlag, 1994.
[21] Roe P. L. Approximate Riemann solvers, parameter vectors, and difference
schemes. J. Comput. Phys. 43 (1981), 357372.
[22] Selmin V. Third-order nite element schemes for the solution of hyperbolic
problems. Report 707, INRIA, 1987.
[23] Selmin V. The node-centred nite volume approach: bridge between nite
differences and nite elements. Comput. Meths. Appl. Mech. Eng. 102 (1993),
107138.
174
[24] Selmin V., Donea J., and Quartapelle L. Finite element methods for
nonlinear advection. Comp. Meths. Appl. Mech. Eng. 52 (1985), 817845.
[25] Selmin V., and Formaggia L. Unied construction of nite element
and nite volume discretisations for compressible ows. Int. J. Numer. Meth.
Eng. 39 (1996), 132.
[26] Sweby P.K. High resolution schemes using ux limiters for hyperbolic
conservation laws. SIAM J. Num. Anal. 21 (1984), 9851011.
[27] van der Waals J.D. On the continuity of the gaseous and liquid states,
volume XIV. North-Holland, Amsterdam, 1988. reprinted.
[28] Warming R.F., and Hyett B.J. The modied equation approach to the
stability and accuracy analysis of nite-difference methods. J. Comput. Phys.
14 (1974), 159179.
175

You might also like