20.

2 The Finite Volume Method 657

in the frameworks of finite volume and of discontinuous Galerkin finite ele-

ment methods; see for instance [178], [169], [170], [175], [176] and [177]. The
comparative study reported in[91] suggests that it is possible to widen the
ADER framework so as to include the method of Harten and collaborators
[242]. A significant step forward in this direction is the recent work of Dumbser
and collaborators who have proposed a rather general solver for the general-
ized Riemann problem; see [173] and [171]. This solver allows, amongst other
things, the proper treatment of stiff source terms, reconciling high accuracy
and stiffness.
In this chapter we give an introduction to the ADER approach in the
framework of finite volume methods. The rest of this chapter is structured as
follows: in section 20.2 we describe the ADER approach in the finite volume
framework; in section 20.3 we construct a second–order ADER method for
a scalar equation with a source term; in section 20.4 we construct ADER
schemes of arbitrary order of accuracy for the model scalar equation with a
linear source term. Sample numerical results are shown in section 20.5 and
conclusions are drawn in section 20.6.

20.2 The Finite Volume Method

The ADER methods studied in this chapter operate in both the frame-
works of finite volumes and discontinuous Galerkin finite element methods.
This introductory chapter deals exclusively with finite volume methods.

20.2.1 The Framework

Here we introduce the main features of the ADER approach as applied to

a scalar conservation law with a source term

∂t q + ∂x f (q) = s(q) . (20.1)

A finite volume scheme to solve (20.1) approximately results from integrating

equation (20.1) on a control or finite volume [xi− 12 , xi+ 12 ] × [tn , tn+1 ] to obtain
the formula
Δt
qin+1 = qin − [f 1 − fi− 12 ] + Δtsi , (20.2)
Δx i+ 2
where qin is the spatial–integral average
xi+ 1
1 2
qin = q(x, tn )dx , (20.3)
Δx xi− 1
2

fi+ 12 is the time–integral average

tn+1
1
fi+ 12 = f (q(xi+ 12 , t))dt , (20.4)
Δt tn
658 20 The ADER Approach

and si is the volume–integral average

tn+1 xi+ 1
1 1 2
si = s(q(x, t))dxdt , (20.5)
Δt Δx tn xi− 1
2

with
Δx = xi+ 12 − xi− 12 , Δt = tn+1 − tn . (20.6)
In (20.3) the integrand is assumed to be the known initial condition at time
tn within the control volume. In (20.4) q(xi+ 12 , t) is the solution of (20.1) at
the volume interface position x = xi+ 12 . In (20.5) the function q(x, t) is the
solution of (20.1) within the volume xi− 12 ≤ x ≤ xi+ 12 , tn ≤ t ≤ tn+1 . Under
these assumptions relation (20.2) is exact.
Numerical methods of the finite volume type can be constructed by in-
terpreting (20.2) as a numerical formula to update approximations to cell
averages qin within the cells Ii = [xi− 12 , xi+ 12 ], with cell or volume centre

1
xi = (x 1 + xi+ 12 ) , (20.7)
2 i− 2
cell interfaces xi− 12 and xi+ 12 , mesh length Δx and time step Δt.
A particular finite volume method is constructed by specifying a particular
approximation to the integrals (20.3)–(20.5). The integral (20.3) is obvious
and results from taking q(x, tn ) as the initial condition inside the volume,
or cell, Ii . Strictly speaking, at the initial time one should convert the given
initial conditions into cells averages via (20.3), but for subsequent times the
updating formula (20.2) takes care of this task automatically. What is crucial
in a finite volume method is the design of a strategy to produce approximations
to the integrals (20.4) and (20.5) that result in approximate expressions for
fi+ 12 , called the numerical flux, and for si , called the numerical source. The
numerical flux and numerical source will still be denoted respectively as fi+ 12
and si , so that (20.2) can be interpreted as a numerical formula.
In what follows we define the ADER approach for prescribing the numerical
flux and the numerical source in the finite volume formula (20.2).

20.2.2 The Numerical Flux

The numerical flux fi+ 12 results from approximating the time–integral av-
erage (20.4) of the flux function at the interface. Where convenient we intro-
duce local coordinates in which x is replaced by x − xi+ 12 and t is replaced by
t − tn . In the ADER approach the numerical flux is obtained by evaluating
the integral
Δt
1
fi+ 12 = f (qi+ 12 (τ ))dτ , (20.8)
Δt 0
 20.2 The Finite Volume Method 659

with a suitably high order numerical integration scheme, where the function
of time qi+ 12 (τ ) at the interface position is the solution of a generalized Rie-
mann problem. The generalized Riemann problem GRPK for a given system
with source terms has initial conditions with a piece–wise high–order spatial
representation and whose solution at the interface has a high–order temporal
representation. The GRPK for (20.1) is the Cauchy problem
⎫
PDE: ∂t q + ∂x f (q) = s(q) , ⎪
⎪
⎪
⎪
⎧ ⎬
⎨ pi (x) if x < 0 , (20.9)
⎪
⎪
IC : q(x, 0) = ⎪
⎪
⎩ ⎭
pi+1 (x) if x > 0 .

Here pi (x) and pi+1 (x) are assumed to be polynomials of degree K obtained
from a conservative and non–oscillatory reconstruction procedure. The solu-
tion of (20.9) at the fixed interface position xi+ 12 , or x = 0 in local coordinates,
denoted by qi+ 12 (τ ), is a function of time and has the form

K '
% ( τk
(k)
q i+ 12 (τ ) = q(0, 0+ ) + ∂t q(0, 0+ ) , (20.10)
k!
k=1

where
q(0, 0+ ) = lim q(0, t)
t→0+

and K is the order of the approximation. See chapter 19 for details on solu-
tion methods for the generalized Riemann problem GRPK . With the solution
qi+ 12 (τ ) available one can compute the numerical flux according to (20.8). In
the absence of source terms the resulting finite volume scheme (20.2) is of
order K + 1 in both space and time.

20.2.3 The Numerical Source

In the presence of source terms the ADER approach determines the nu-
merical source si by approximating the space–time integral (20.5) within the
space–time volume [xi− 12 , xi+ 12 ] × [0, Δt] to obtain the average

Δt xi+ 1
1 1 2
si = s(qi (x, t))dxdt , (20.11)
Δt Δx 0 xi− 1
2

where qi (x, t) denotes an approximation to the solution of (20.1) inside the

volume [xi− 12 , xi+ 12 ] × [0, Δt], with qi (x, 0) = pi (x) in cell i, at the initial
time t = 0. To obtain the required values of qi (x, t) for xi− 12 ≤ x ≤ xi+ 12 ,
0 < t ≤ Δt we use the Cauchy–Kowalewski procedure, introduced in chapter
19.
660 20 The ADER Approach

A possible version of the method considers fixed integration points x̂ within

the cell i at which one constructs a Taylor series expansion in time. By means
of the Cauchy–Kowalewski procedure one converts time derivatives to space
derivatives using the PDE (20.1). The space derivatives in cell i are all readily
available from the smooth reconstructed function pi (x) and thus the solution
at each integration point x̂ is known as an explicit function of time and there-
fore the integration can be carried out to obtain the numerical source si to
the desired order of accuracy. The resulting finite volume scheme (20.2) for
the advection equation with a source term (20.1) is of order K + 1 in both
space and time
Two pending issues are the reconstruction procedure to obtain the smooth
functions pi (x) that form the initial condition of the generalized Riemann
problem (20.9) and the application of the Cauchy–Kowalewski procedure to
find qi (x, t) in (20.11). In the following subsection we give a very succinct
overview of the reconstruction problem.

20.2.4 Reconstruction

At any given time tn one has a set of cell averages {qin } that are approx-
imations to (20.3). This set represents a piece–wise constant distribution of
the data. Note that the spatial variation of the solution within each cell has
been lost in the averaging procedure. The task at hand is to recover the lost
information by using the cells averages qin to construct, or reconstruct, smooth
functions pi (x) that locally represent the spatial variation of the solution of
(20.1). The reconstruction procedure to find the function pi (x) must be con-
servative, that is, the integral average of pi (x) inside cell i is identically qin .
In general, the functions pi (x) have jump discontinuities at the cell interfaces
x = xi+ 12 . Fig. 20.1 depicts typical reconstructed functions for three consec-
utive finite volumes i − 1, i and i + 1. In this chapter we assume that the
reconstructed functions are polynomials of a specified degree.
In principle, any of the classical interpolation techniques can be employed
in the reconstruction procedure. In particular, one can construct polynomials
on a fixed stencil formed by a set of averages
n n n
qi−k L
, qi−k L +1
, . . . , qi−1 , qin , qi+1
n n
, . . . , qi+k R
, (20.12)
where kL and kR and two non–negative integers.
A reconstruction based on a fixed stencil will be called a linear reconstruc-
tion. The resulting numerical schemes will be called linear schemes. Recall
that from Godunov’s theorem, see chapter 13, linear schemes of accuracy
greater than one are oscillatory. If avoiding spurious oscillations is an impor-
tant requirement on the numerical methods, then one must design non–linear
schemes, even when applied to linear PDEs. A necessary condition for design-
ing non–linear schemes is that the reconstruction be implemented on variable
or adaptive stencils, rather than fixed stencils. Such reconstructions will be
called non–linear reconstructions.
 20.2 The Finite Volume Method 661
q(x, tn )
pi (x) pi+1 (x)
pi−1 (x)

n
qi−1 qin n
qi+1

xi−1 xi xi+1 x

Fig. 20.1. Cell averages {qin } define a piece–wise constant distribution {pi (x)} on
each cell i. Illustration of reconstructed polynomial functions pi−1 (x), pi (x) and
pi+1 (x) in cells i − 1, i and i + 1, respectively.

The simplest non–linear reconstruction is based on the Total Variation Di-

minishing, or TVD, criterion studied in chapter 13. Higher–order non–linear
ADER schemes can be constructed on the basis of other types of reconstruc-
tions, such as ENO and WENO interpolations.
The ENO (Essentially Non–Oscillatory) interpolation method [245] is suit-
able for approximating functions that contain discontinuities or exhibit large
gradients. For a given interval [a, b] and a set of samples (xi , h(xi )), for
i = 0, . . . , N , classical interpolation methods construct a global polynomial
of order N that interpolates all points, using for example, the Newton form
of the polynomial. Instead, in the ENO approach one constructs polynomials
of order less than N that have three distinguishing features:
• they are local, that is, they are defined on sub–intervals Ii = [xi− 12 , xi+ 12 ]

• they are conservative, that is

xi+ 1
1 2
pi (x)dx = qin , (20.13)
Δx xi− 1
2

• they are adaptive, their coefficients depend on the behaviour of the data.
The local and adaptive nature of the interpolation method results in a
representation polynomial that is essentially non–oscillatory (ENO).
A trivial example of a set of reconstructed polynomials is given by the set
of constant polynomials
⎧ n
⎨ qi if x ∈ Ii = [xi− 12 , xi+ 12 ] ,
pi (x) = (20.14)
⎩
0 otherwise ,
662 20 The ADER Approach

resulting in a global, piece–wise constant, polynomial pi (x), for all x in a given

interval I. This reconstruction satisfies trivially all three conditions above.
A first–degree polynomial in cell Ii = [xi− 12 , xi+ 12 ] is

(1) (1)
pi (x) = qin + (x − xi )Δi , (20.15)

where the slope, given as

(1) (1)
Δi = Δi (Δi− 12 , Δi+ 12 ) , (20.16)

is a function of the two neighbouring gradients

qin − qi−1
n
q n − qin
Δi− 12 = , Δi+ 12 = i+1 . (20.17)
Δx Δx
(1)
A very simple choice for Δi is the mean value

(1) 1
Δi = (Δ 1 + Δi+ 12 ) . (20.18)
2 i− 2
(1)
In the ENO approach one chooses the slope Δi adaptively. For this first–
degree polynomial one takes
⎧
⎨ Δi− 12 if |Δi− 12 | ≤ |Δi+ 12 | ,
(1)
Δi = (20.19)
⎩
Δi+ 12 if |Δi− 12 | > |Δi+ 12 | .

Then, a global, piece–wise linear polynomial is defined as

⎧
⎨ pi (x) if x ∈ Ii = [xi− 12 , xi+ 12 ] ,
h(x) = (20.20)
⎩
0 otherwise .

This reconstruction is equivalent to have considered two candidate polynomi-

als qL (x), qR (x), with corresponding two candidate stencils

SL = {qi−1
n
, qin } and SR = {qin , qi+1
n
} (20.21)

and then choosing the one with the smallest first divided difference. Higher–
order ENO polynomials for each cell Ii can be constructed in an analogous
manner.
Another choice for stencil–adaptive reconstructions is the WENO ap-
proach [331], [270] (Weigthed Essentially Non–Oscillatory), whereby the sought
polynomial pi (x) is a weighted average of all polynomials that correspond to
the candidate stencils in the ENO method. Further details of WENO are found
in [493]. A modified WENO reconstruction method applicable to unstructured
meshes in two and three space dimensions is proposed in [175], [176].
 20.3 Second–Order Scheme for a Model Equation 663

20.3 Second–Order Scheme for a Model Equation

Concerning accuracy, ADER is a unified approach, in that the same frame-
work allows the choice of any desired order of accuracy. Here we illustrate the
application of the ADER approach to construct a scheme of the form (20.2)
of second order of accuracy. We follow the full general framework, as would be
done for constructing an ADER scheme of any order of accuracy. The specific
second–order scheme is presented in terms of the model advection equation
with a source term
∂t q + λ∂x q = βq . (20.22)
The flux function is f (q) = λq and the source function is s(q) = βq, where λ
and β ≤ 0 are known constants. We need to compute the numerical flux fi+ 12
and the numerical source si for the scheme (20.2).

20.3.1 Numerical Flux and Numerical Source

First we assume a reconstruction of the data in terms of first–degree poly-

nomials pi (x) of the form
(1) (1)
pi (x) = qin + Δi (x − xi ) , pi+1 (x) = qi+1
n
+ Δi+1 (x − xi+1 ) , (20.23)

(1) (1)
with the slopes Δi and Δi+1 chosen as in (20.19), for example. Then, to
compute the numerical flux fi+ 12 according to (20.8) we solve the generalized
Riemann problem GRP1
⎫
PDE: ∂t q + λ∂x q = βq , ⎪
⎪
⎪
⎪
⎧ ⎬
⎨ pi (x) if x < 0 , (20.24)
⎪
⎪
IC: q(x, 0) = ⎪
⎪
⎩ ⎭
pi+1 (x) if x > 0 .

The solution of (20.24) takes the form

qi+ 12 (τ ) = q(0, 0+ ) + τ ∂t q(0, 0+ ) (20.25)

and is completely determined by performing the following steps:

Step (I): The leading term. To compute the leading term q(0, 0+ ) one
solves the classical Riemann problem
⎫
PDE: ∂t q + λ∂x q = 0 , ⎪
⎪
⎪
⎪
⎧ ⎪
⎬
⎪ n 1 (1)
⎨ pi (0) ≡ qi + 2 ΔxΔi if x < 0 , (20.26)
⎪
⎪
IC: q(x, 0) = ⎪
⎪
⎩ p (0) ≡ q n − 1 ΔxΔ(1) if x > 0 , ⎪
⎪ ⎭
i+1 i+1 2 i+1
664 20 The ADER Approach

in which the initial condition consists of the boundary extrapolated val-

ues pi (0) from the left and pi+1 (0) from the right. The similarity solution
(0)
di+ 1 (x/t) of this problem is
2

⎧
⎪ (1)
⎨ qin + 12 ΔxΔi if x/t < λ ,
(0)
di+ 1 (x/t) = (20.27)
2 ⎪
⎩ qn (1)
i+1 − 1
2 ΔxΔi+1 if x/t > λ ,

and the leading term is the Godunov state (x/t = 0)

⎧
⎪ (1)
⎨ qin + 12 ΔxΔi if λ > 0 ,
(0)
q(0, 0+ ) = di+ 1 (0) = (20.28)
2 ⎪
⎩ q n − 1 ΔxΔ(1) if λ < 0 .
i+1 2 i+1

Step (II): The higher order term. To compute the second term in (20.25)
we do the following:
1. Time derivative in terms of the spatial derivative. The Cauchy–Kowalewski
procedure uses the PDE (20.22) to express the time derivative in terms
of the space derivative and the source term,

∂t q(x, t) = −λ∂x q + βq , (20.29)

with the function q(x, t) evaluated at the leading term, so that the expan-
sion (20.25) reads

qi+ 12 (τ ) = q(0, 0+ ) + [−λ∂x q(0, 0+ ) + βq(0, 0+ )]τ . (20.30)

2. Evolution equation for the space derivative. To determine the space deriva-
tive on the right–hand side of (20.30) we first construct an evolution equa-
tion for ∂x q(x, t); it is easily seen that in fact ∂x q(x, t) obeys the original
evolution advection–reaction equation (20.22), namely

∂t (∂x q(x, t)) + λ∂x (∂x q(x, t)) = β∂x q(x, t) . (20.31)

3. Riemann problem for the spatial derivative. We simplify (20.31) by ne-

glecting the source term for the gradient. Then we pose the classical ho-
mogeneous Riemann problem for the spatial derivative:
⎫
PDE: ∂t (∂x q(x, t)) + λ∂x (∂x q(x, t)) = 0 , ⎪
⎪
⎪
⎧ ⎪
⎬
⎪
⎨ Δi
(1)
if x < 0 , (20.32)
⎪
⎪
IC: ∂x q(x, 0) =
⎪ ⎪
⎪
⎩ Δ(1) if x > 0 . ⎭
i+1

We solve this Riemann problem to obtain the similarity solution

 20.3 Second–Order Scheme for a Model Equation 665
⎧
⎪ (1)
⎨ Δi if x/t < λ ,
(1)
di+ 1 (x/t) = (20.33)
2 ⎪
⎩ Δ(1) if x/t > λ
i+1

and set ⎧
⎪ (1)
⎨ Δi if λ > 0 ,
(1)
∂x q(0, 0+ ) = di+ 1 (0) = (20.34)
2 ⎪
⎩ Δ(1) if λ < 0 .
i+1

Step (III): Solution of GRPK . The final expression for the sought solution
(20.25), (20.30) is
⎧ ' (
⎪ (1) (1) (1)
⎨ qi + 2 ΔxΔi + τ −λΔi + β(qi + 2 ΔxΔi ) ,
⎪ n 1 n 1
λ>0,
qi+ 12 (τ ) = ' (
⎪
⎪
⎩ qn (1) (1)
− 12 ΔxΔi+1 ) + τ −λΔi+1 + β(qi+1
n (1)
− 12 ΔxΔi+1 ) , λ < 0 .
i+1
(20.35)

Step (IV): Numerical flux. Finally, the numerical flux in (20.2) is obtained
according to (20.8), as
⎧  
⎪
⎪ λ q n
+ 1
(1 − c)ΔxΔ
(1)
+ 1
r(q n
+ 1
ΔxΔ
(1)
) if λ > 0 ,
⎨ i 2 i 2 i 2 i
fi+ 12 =  
⎪
⎪
⎩ λ qn − 1
(1 + c)ΔxΔ
(1)
) + 1
r(q n
− 1
ΔxΔ
(1)
) if λ < 0 ,
i+1 2 i+1 2 i+1 2 i+1
(20.36)
where c = λΔt/Δx is the Courant number and r = Δtβ is a dimensionless
reaction number.
Remark: the numerical flux (20.36) includes a contribution due to the source
term. In other words, the flux knows of the source term, as one would expect.
But note that there is still another contribution to the scheme resulting from
the numerical source, still to be specified.

Step (V): Numerical source. To compute the numerical source si in the

finite volume scheme (20.2) we need to select the function qi (x, t) in the inte-
grand of the volume integral (20.11) and an integration scheme. For a second
order method we may apply the mid–point integration rule in space and time,
so that one only requires qi (xi , 12 Δt). At time t = 0, within the cell i we have
the reconstructed polynomial pi (x) given by (20.15)–(20.16). Taylor expand-
ing the solution in time at the point xi and applying the Cauchy–Kowalewski
procedure we obtain
⎫
qi (xi , τ ) = qi (xi , 0+ ) + τ ∂t q(xi , 0+ ) ⎬
(20.37)
⎭
= qi (xi , 0+ ) + τ (−λ∂x qi (xi , 0+ ) + βqi (xi , 0+ )) .
666 20 The ADER Approach

In the case considered we have

(1)
qi (xi , 0+ ) = qin = pi (xi ) , ∂x qi (xi , 0+ ) = Δi )
so that the numerical source from (20.11) is
 
1 (1)
si = β qin + Δt(−λΔi + αqin ) . (20.38)
2

20.3.2 The Scheme

The finite volume scheme (20.2) to solve the model advection–reaction
equation (20.22) to second–order of accuracy is now completely determined,
the numerical flux is given by (20.36) and the numerical source is given by
(20.38). In full, for the case of λ > 0 the scheme reads
' (⎫
) + 12 (1 − c)(Δi − Δi−1 ) ⎪
(1) (1)
qin+1 = qin −c (qin − qi−1
n
⎪
⎪
⎪
⎪
⎪
⎪
' ( ⎬
(1) (1)
−cr 2 (qi − qi−1 ) + 2 Δx(Δi − Δi−1 )
1 n n 1
(20.39)
⎪
⎪
⎪
⎪
' ( ⎪
⎪
(1) ⎪
⎭
+r (1 + r)q − cΔxΔ
2
1
i
n 1
2 i .
Remark: Note that if the slopes are all set to zero and we also set to zero the
contribution of the source to the numerical flux, then we produce a first–order
numerical scheme for (20.22) that consists of the first–order upwind method
of Godunov for the advection part, along with an upwind discretization of the
source term, namely
 
1 n 1
qin+1 = qin − c(qin − qi−1
n
) − r cqi−1 + (1 − c)qin . (20.40)
2 2
This corresponds to one of the upwind schemes analyzed in [412] and [51] for
solving advection equations with source terms.
(1)
Remark: By choosing the slopes Δi appropriately, we can reproduce some
well–known schemes, when β = 0. For example, for the case λ > 0 in (20.22),
the three classical second–order schemes of Warming–Beam, Lax–Wendroff
and Fromm are identically reproduced by (20.39) by choosing the slope as
follows ⎧
⎪ Δi− 12 Warming–Beam ,
⎪
⎪
⎪
⎪
⎨
(1)
Δi = Δi+ 12 Lax–Wendroff , (20.41)
⎪
⎪
⎪
⎪  
⎪
⎩ 1 Δ 1 + Δ 1 Fromm .
i− 2 i+ 2
2
Remark: the second–order ADER scheme with ENO reconstruction (20.15)–
(20.19) is effectively a blend of the Warming–Beam and the Lax–Wendroff
schemes.