On A Class of Pad e Finite Volume Methods

Journal of Computational Physics 156, 137–180 (1999)
Article ID jcph.1999.6376, available online at http://www.idealibrary.com on
On a Class of Padé Finite Volume Methods

Marcelo H. Kobayashi
Department of Mechanical Engineering, Instituto Superior Técnico, Av. Rovisco Pais, 1049-001 Lisbon, Portugal
E-mail: [email protected]
Received February 2, 1999; revised September 7, 1999
A class of Padé finite volume methods providing an improved spectral reso-

lution is presented and compared with well-known methods. The formulation is
based on the sliding averages of the variables and allows the computation of deriva-
tives of all orders. Using the Fourier analysis, these methods are examined with
respect to (i) order of accuracy, (ii) spectral resolution, (iii) boundary conditions, and
(iv) stability. °c 1999 Academic Press
Key Words: finite volume method; high-order interpolation; Padé interpolation;
boundary conditions.
CONTENTS
1. Introduction.
2. The standard Padé interpolation.
3. Multi-dimensional equations.
4. Optimized phase error and spectral Padé interpolation.
5. Boundary conditions.
6. Stability analysis.
7. Summary.
1. INTRODUCTION
The development and widespread use of high-speed digital computers rendered numerical
simulations of fluid flow problems a common practice in industry (see, for example, [7, 12]).
In the 1990s, numerical simulation has also evolved into a tool for the fundamental study of
turbulence, either by direct numerical simulation or by large eddy simulation. A common
feature in both applied and fundamental developments is the pursuit of highly accurate
methods. The latter provides grid-independent results in coarser grids and sometimes may
be the only way to find a satisfactory solution in a “sensible” grid.
Direct numerical simulation of turbulence and large eddy simulation require solution
of time and length scales that are various orders of magnitude apart. This has led to
137
0021-9991/99 $30.00
Copyright ° c 1999 by Academic Press
All rights of reproduction in any form reserved.
138 MARCELO H. KOBAYASHI
the development of spectral and pseudo-spectral methods for numerical integration of the
Navier–Stokes equations (see, for example, [9, 15, 23]). However, these methods strongly
depend on the grid topology and the type of boundary conditions. Moreover, they may
lose effectiveness in problems with sharp variations over thin shear or boundary layers
(see [2]).
Compact methods, introduced in the 1930s, can ensure spectral-like resolution of short
waves. The implementation of such methods is simple and is not restricted either by the grid
topology or to some types of boundary conditions. The seminal work of Lele [18], building
on previous work on compact finite differences, contributed greatly to the diffusion of
these ideas. Spotz [26] proposed a class of compact finite difference schemes that uses
the governing equations to approximate the leading term in the truncation error. These
schemes can also handle nonuniform grids and have been applied to the stream-function
vorticity formulation of the Navier–Stokes equations, transient problems, and so forth.
Wilson et al. [30] presented and analyzed high-order accurate compact methods for solving
unsteady incompressible Navier–Stokes equations for 2D and 3D fluid flow problems. Such
methods use the primitive variables together with a Poisson equation for the pressure. In
[31], a 4th-order compact method for solving convection–diffusion problems was analyzed
with respect to the effect of boundary closure and cell Peclet number. Carpenter et al.
[4, 5] presented an extensive and illuminating study of explicit and compact high-order
schemes with various boundary conditions. Mahesh [21] introduced a class of compact
schemes that uses a coupled derivative formulation. It is shown that, for the same order of
accuracy, the resulting coupled method displays an improved spectral resolution and that the
work required by the new method is comparable to that required by the standard compact
method.
Despite their differences these methods are all based on the finite difference approach.
Mattiussi [22], using some concepts of the algebraic topology, concluded that for topo-
logical equations, that is, the differential or discrete form of the field equations, a discrete
representation of the geometry, fields, and operators is preferable. This requires the correct
attribution of physical quantities to geometrical objects with appropriate orientation and
dimension. Moreover, the fact that the finite-difference methods do not fulfill this attribute
renders finite volume and finite element methods more qualified for numerical simulation
of field problems. Indeed, finite volume methods are prefered by the computational fluid
dynamics (CFD) practitioners in industry: they are as easy to implement as finite difference
methods and easier than the finite element method; they are based in the weak formulation
of the equations and so require less smoothness of the function; and last, but not least,
they are conservative. This last property is essential for methods aiming at compressible
fluid flow problems (see [12]). Recently, Gaitonde and Shang [8] proposed and analyzed
a class of optimized 4th-order finite volume compact schemes using the reconstruction via
the primitive function.
That said, the objective of the present work is to extend and analyze the class of Padé finite
volume methods proposed by Gaitonde and Shang. The formulation uses the sliding averages
of the variable as the basic variable and the reconstruction via deconvolution. The analysis
includes (i) the study of the truncation error of the spatial interpolation with respect to its
order of accuracy and its spectral resolution, (ii) the study of boundary conditions and their
effects on the quality of the results and stability of the method, and (iii) the stability analysis
of explicit and implicit time marching methods. Throughout the presentation comparisons
with commonly used methods are provided and discussed.
PADÉ FINITE VOLUME METHODS 139
2. THE STANDARD PADÉ INTERPOLATION
In this section we describe a class of (2r )th-order Padé finite volume methods. This class
is standard in the sense that it uses a symmetric stencil, and the resulting interpolating
polynomial provides the maximal order of accuracy for the given stencil.
The method, which can be used as a general interpolating technique, is aimed at transport
equations. So, given the distinct characteristics of transport by convection and diffusion,
we analyze them in turn below.
2.1. Hyperbolic Equations

The linear advection equation
∂φ ∂φ
+c = 0, (1)
∂t ∂x
where c is a real constant and φ is a scalar function, is often used as a model for the descrip-
tion of numerical approximation of hyperbolic systems. This is because many hyperbolic
methods use the Jacobian matrix computed at a convenient point and interpolate in the
characteristic directions. The result is a set of uncoupled linear hyperbolic equations, each
of which are of the form of Eq. (1) (see [29]).
Let {σ j } j ∈ Z , with σ j = [x j , x j+1 ], x j = j h, j ∈ Z be a partition of R, for some grid
parameter h ∈ R. Given a function f ∈ L 1 and h ∈ R its sliding, or moving, average f¯ ∈ L 1
is defined as (see [11])
Z
1 h/2
f¯ = f (x̂ + y) dy
h −h/2
1
= χh ∗ f, (2)
h
where ∗ denotes the usual convolution of functions and
½
1 if |x| < h2 ,
χh (x) =
0 otherwise
is the characteristic function of a cell. We use the convention proposed by Silva [6] that a hat
over a variable denotes a dummy variable. In the finite volume approach we first integrate
Eq. (1) over each control volume, or cell, σ j of a partition {σ j } j ∈ Z , yielding
· ¸
d φ̄ φ j+1 − φ j
+c = 0, j ∈ Z. (3)
dt j+1/2 h
Note that, so far, we have not introduced any approximation in the model. Also, Eq. (3),
which is an evolution equation for the cell averages of φ, requires less smoothness than
Eq. (1). The variables involved in (3) suggest that in the finite volume method we store and
derive the discrete equations for the cell averages. This is the approach used in the present
work.
So, we need to obtain a relation between the values of the variables at cell faces and the
averaged values at the midpoint of each cell. For this purpose we use the Padé interpolation,
which is a solution of the following problem IP:
Given a smooth function φ ∈ C ∞ , and natural numbers m, n ∈ N, find coefficients ai , i =

1, . . . , m and bi , i = 1, . . . , n, such that
X
m X
m X
n X
n
¡ ¢
ai τ−i h φ + φ + ai τi h φ = bi τ(−i+1/2)h φ̄ + bi τ(i−1/2)h φ̄ + o h 2(m+n)−1 .
i=1 i=1 i=1 i=1
(4)
In this expression τs : C → C, with s ∈ R, is the translation, or shift, operator,
(τs f ) (x) = f (x − s), x ∈R
for all f ∈ C. Hence, from (2) the IP problem is actually a deconvolution to o(h 2(m+n)−1 ).
Solution of IP can be easily obtained by fixing a point x 0 ∈ R and expanding the function
φ in a Taylor series around it. We illustrate the procedure for the 4th-order problem, when
m = n = 1. Expanding the function φ around a point x0 ∈ R yields
µ ¶
h 2 00 h 3 000 h 4 (4) h 5 (5)
a φ0 − hφ00
+ φ0 − φ0 + φ0 − φ
2 6 24 120 0
µ ¶
h2 h3 h4 h 5 (5)
+ φ0 + a φ0 + hφ00 + φ000 + φ0000 + φ0(4) + φ0
2 6 24 120
µ 2 3 4 5 ¶
h 0 h 00 h 000 h (4) h (5)
= b φ0 − φ0 + φ0 − φ0 + φ − φ
2 6 24 120 0 720 0
µ ¶
h 0 h 2 00 h 3 000 h 4 (4) h 5 (5)
+ b φ0 + φ0 + φ0 + φ0 + φ + φ + o(h 5 ),
2 6 24 120 0 720 0
where the subscript 0 denotes the value of the corresponding function at x0 . So, a = 14 , b = 34 ,
and the truncation error is o(h 3 ) = (h 4 /120)φ (4) + o(h 5 ). In Table I we summarize the
coefficients and the leading term of the truncation error for the 2nd-, 4th-, 6th-, 8th-, and
12th-order Padé methods, with m + 1 ≥ n ≥ m. Note that for each order of accuracy the
finite volume method requires less smoothness from the function than its finite difference
counterpart (see [18]).
In Table II we list, for the 4th, 6th, 8th, and 12th order of accuracy, the coefficients
and the leading term of the truncation error for the finite volume Lagrange interpolation
scheme. The latter is obtained by taking m = 0 in Eq. (4). As can be seen from this table,
with the exception of the 2nd-order problem, where the interpolation problem is the same,
TABLE I
Coefficients and Leading Term of the Truncation Error
for some Padé Methods
Order a1 a2 a3 b1 b2 b3 Truncation error

1 h 2 (2)
2 — — — 2
— — 6
f
h4
4 1
4
— — 3
4
— — − 120
(4)
f
1 29 1 h 6
(6)
6 3
— — 36 36
— 1260
f
h8
8 4
9
1
36
— 185
216
25
216
— − 22680
(8)
f
h 12
12 9
16
9
100
1
400
1799
2000
973
4000
49
4000
− 4804800
(12)
f
TABLE II
for Some Lagrange Methods
Order b1 b2 b3 b4 b5 b6 Truncation error

4
4 7
12
− 1
12
— — — — − h30 f (4)
h6
6 37
60
− 2
15
1
60
— — — 140
f (6)
h8
8 533
840
− 139
840
29
840
− 1
280
— — − 630
f (8)
h 12
12 18107
27720
− 5653
27720
443
6930
− 107
6930
67
27720
− 1
5544
− 12012
f (12)
the constant in the leading term of the truncation error for the Padé interpolation is always
smaller than the corresponding one for the Lagrange interpolation. The order being the same,
this means that in a Log–Log graph of the interpolation error, the asymptotic curves are
parallel, but the absolute value of the error is smaller for the Padé interpolation. Moreover,
as the order increases, so does the ratio between the constants in the Lagrange and the Padé
methods. For the 12th-order method the constant in the Lagrange interpolation is already
400 times larger than the corresponding one for the Padé interpolation.
The order of the truncation error provides the asymptotic rate of convergence of the
interpolation toward the interpolated function. Still, it conveys no information about the
spectrum of the truncation error. The use of Fourier analysis to characterize the spectrum of
the truncation error is thoroughly described in [29]. It is the classical technique for studying
and comparing approximation methods. In the context of Padé methods it was used, among
others, by Lele [18] and Mahesh [21] to study compact finite difference schemes.
Next, we use the Fourier analysis to study the Padé finite volume method. Let φ ∈ S 0 ,
where S 0 is the space of tempered distributions,1 which is the dual space to the space of
rapidly decreasing smooth functions, or Schwartz space, S; then its Fourier transform exists
and we write
8 = F(φ)
where the operator F: S 0 → S 0 is the Fourier transform2 on S 0 , that is,
F: S 0 → S 0
φ 7→ F(φ)
with
hF(φ),ψi = hφ,F(ψ)i
1
Recall that the space of usual tempered functions 3 belongs to S 0 ; that is, 3 = { f ∈ C: ∃ p ∈ 5 & | f (x)| ≤ p(x),
x ∈ R} ⊂ S 0 , where 5 is the space of polynomials of R. In fact, for all f ∈ S 0 we have f = D p g, for some p ∈ N
and g ∈ 3. A distinguished class of distributions that are not functions in the usual sense, but are in S 0 , comprises
the delta of Dirac and its derivatives δ (n) ∈ S 0 , n ∈ N. R
2
If f ∈ L p , 1 ≤ p ≤ ∞ then f R ∈ S 0 and hF( f ),ψi = R×R f (x)e−i k̂ξ ψ(ξ )d x dξ , for all ψ ∈ S. In particular,
if f, F( f ) ∈ L 1 , then f (x) = 2π1 R F( f )(k)eikx dk, for almost all xR∈ R, with respect to the Lebesgue measure.
Also, for f ∈ L 2 , we have F( f ) ∈ L 2 but, in general, the integral R f (x)e−ikx dk does not converge. In other
words, we do not have an integral representation for the Fourier transform in L 2 (see [25]).
and
Z
F(ψ)(k) = e−ikx ψ(x) d x, k∈R
R
for all ψ ∈ S. We call 8 the spectrum of φ.

Given φ ∈ S 0 we use the convolution in (2) to define φ̄ ∈ S 0 , and denote
8̄ = F(φ̄)
Now, from its definition and the convolution theorem (see [25, Theorem XV, p. 268]) for
the Fourier transform it follows that
µ ¶
2 k̂h
8̄ = sin 8. (5)
k̂h 2
Applying the Fourier transform to Eq. (4) and taking into account Eq. (5) yields
E = Er 8,
where E ∈ S 0 is the spectrum of the truncation error and Er ∈ S 0 is the relative spectral error
  
 X m µ ¶X n µµ ¶ ¶
2 k̂h 1
Er = 1 + 2 a j cos( j k̂h) − sin b j cos j− k̂h  . (6)
 k̂h 2 2 
j=1 j=1
Given the fact that F(δ) = 1, the relative spectral error is the spectrum of the error associated
with the Dirac distribution. Figure 1 shows the relative spectral error Er for the 4th-, 6th-,
8th-, and 12th-order Padé and Lagrange finite volume methods as a function of the cell
wave number k̃ : 2π k̃ = kh. This figure shows that increasing the order of the Padé method
decreases its spectral error at all wave numbers in the range considered. This effect is
FIG. 1. Plot of relative spectral error versus grid wave number for advection using the Padé method: (a) 4th
order, (b) 6th order, (c) 8th order, (d) 12th order; and the Lagrange method: (e) 4th order, (f) 6th order, (g) 8th
order, (h) 12th order.
more pronounced for wave numbers closer to half grid width. This figure also shows that
while the truncation error of the Lagrange interpolation scheme is of the same order as the
corresponding Padé one, its behaviour in the spectral error is worse for all wave numbers
greater than zero (constant functions). This discrepancy increases with the wave number
and the order of the method.
Inspection of Eq. (6) provides a heuristic argument to explain the improvement in the
Padé interpolation relative to the Lagrange one. Any usual interpolation method has a term
similar to the second one within brackets. This term tends to zero as k̃ → 1/2. However,
any Padé method of 4th order or higher implicitly uses a full interpolating stencil. This
gives rise to the first term within brackets in Eq. (6), which is a truncated expansion of the
spectrum of the variable φ. Note that in Eq. (6) it appears only in the cosine components.
This results from the symmetry of the stencil, which makes Eq. (4) exact for all odd smooth
functions.
Next, we analyze the nature of the error associated with the proposed Padé finite volume
method with respect to both its amplitude and phase.
We start by writing the linear advection equation as
d φ̄ τ−h/2 φ − τh/2 φ
+c = 0, (7)
dt h
where φ ∈ C(R; S 0 ). Then, for each t ∈ R we apply the Fourier transform to (7), yielding
µ ¶
d 8̄ 2ic k̂h
+ sin 8 = 0. (8)
dt h 2
Using Eq. (5) results in
d 8̄
+ ick̂ 8̄ = 0.
dt
Solution of the previous equation provides the law of propagation
8̄(t) = 8̄0 eick̂t , t ∈ R, (9)
where 8̄0 = 8̄(0). This means that the evolution of the sliding average corresponds to a
phase shift and no change in the amplitude. This is expected, as the solution of the linear
advection problem corresponds to a shift in the profile and the averaging commutes with
the shift.
On the other hand, the Padé interpolation gives rise to the following relation between the
spectra 8 and 8̄:
Ã ! µµ ¶ ¶
X
m X
n
1
8 1+2 a j cos( j k̂h) = 28̄ b j cos j− k̂h . (10)
j=1 j=1
2
FIG. 2. Plot of phase speed versus grid wave number for advection using the Padé method: (a) 2nd order,
(b) 4th order, (c) 6th order, (d) 8th order, (e) 12th order, (f) CD 6th order, (g) CD 8th order, (h) CD 12th order;
and the Lagrange method: (i) 4th order, (j) 6th order.
Substituting Eq. (10) into Eq. (8) and solving for 8̄ yields the same law of propagation as
for the exact profile with c in the place of c, where
µ ¶ Pn ¡¡ ¢ ¢
j=1 b j cos j − 2 kh
1
4 kh
c= sin P c, k ∈ R. (11)
kh 2 1 + 2 mj=1 a j cos( jkh)
Figure 2 displays the phase ratio c/c as a function of the grid wave number k̃ for the
2nd-, 4th-, 6th-, 8th-, and 12th-order Padé interpolation method and the 4th-, and 6th-order
Lagrange interpolation methods. This figure shows that, for small grid wave numbers, all
methods perform well. As the grid wave number increases, all methods begin displaying a
retardation error in the phase. It is evident that compared to the standard Lagrange methods
the Padé method stay close to the exact phase velocity over a wider range of cell wave
numbers. In fact, the 4th-order Padé method shows a better spectral resolution than the
6th-order Lagrange method. Again, this is a consequence of the implicit full stencil used in
the Padé interpolation.
To quantify the error in the phase speed Lele [18] and Mahesh [21] used the notions of
resolving efficiency and percentage error as a function of the number of points per wave
(PPW). The latter is related to the cell wave number as follows: PPW = 1/k̃. These are
measures of the ability of each method to resolve a particular range of phase speed. They
show that compact methods are more efficient than the standard Lagrange methods for all
orders of accuracy.
In the present work, we use the accumulated relative spectral error for half wave resolution
(ARSE1/2 ). It is a direct measure of the spectral error provided by the L 1 ([0, 12 ])-norm of
the relative spectral error distribution. That is,
Z 1/2
ARSE1/2 = |Er |(k̃) d k̃.
0
This measure is similar to the one used by Gaitonde and Shang [8]. However, they measured
the norm of the error in the phase velocity, whereas here we measure the error in the
truncation error of the interpolation.
TABLE III
ARSE1/2 for the Padé and Lagrange
Methods Shown in Fig. 1
Method Order ARSE1/2
Padé 4 5.788 × 10−2

Padé 6 2.897 × 10−2
Padé 8 1.109 × 10−2
Padé 12 2.341 × 10−3
Lagrange 4 1.446 × 10−1
Lagrange 6 1.167 × 10−1
Lagrange 8 1.001 × 10−1
Lagrange 12 8.055 × 10−2
The values of the ARSE1/2 for the Padé and Lagrange methods shown in Fig. 1 are
tabulated in Table III. This shows that increasing the order of accuracy of the Padé method
and Lagrange methods causes them to stay close to the exact solution over a progressively
larger range of the spectrum. The values of ARSE1/2 for the Padé methods are always
smaller than the corresponding values for the Lagrange methods. Moreover, as the order
increases, the value of ARSE1/2 drops faster for the Padé method. For example, the ratio
of the value of the ARSE1/2 for the 8th-order to the 12th-order Padé method is 4.7. The
corresponding figure for the Lagrange method is 1.2.
If we apply the inverse of the Fourier transform to Eq. (9) we recover the analytical
solution of Eq. (1), that is,
φ̄(t) = τct φ̄ 0 ,
where φ̄ 0 = φ̄(0) is the initial profile. The reason for its simple form lies in the fact that all
“waves” compounding φ travel at the same speed c.
Now, the nonlinear relation between c and k in Eq. (11) expresses the fact that with the
Padé interpolation the waves travel at different speed, depending on the wave number k̃. The
waves are close to the actual speed at small grid wave numbers and progressively slow down
as k̃ → 12 . Accordingly, applying the inverse of the Fourier transform to law of propagation
of the approximation Padé yields an “averaging” of the various dispersed modes of the
initial profile.
In [29] it is shown that these errors form packets which travel at the group velocity,
also called energy velocity. The authors point out that certain spurious numerical solutions
of the semidiscretizations are affected with a positive phase speed but a negative group
velocity. This may be relevant in simulations of turbulence, because then the small eddies
may propagate in the opposite direction to the true solution.
Next, we compute the group velocity for several Padé and Lagrange methods. We start
by recalling the definition of the group velocity V:
d
V= (k̂c).
dk
Figure 3 shows the group velocity V as a function of the cell wave number for the 2nd-,
4th-, 6th-, 8th-, and 12th-order Padé interpolation methods, and the 4th-, and 6th-order
FIG. 3. Plot of group velocity versus grid wave number for the Padé method: (a) 2nd order, (b) 4th order,
(c) 6th order, (d) 8th order, (e) 12th order; and the Lagrange method: (f) 4th order, (g) 6th order.
Lagrange interpolation methods. The exact solution for a unitary phase speed is V = 1. As
can be seen from this figure, as the cell wave number increases the group velocity starts
deviates from the exact solution. Indeed, all methods display a region with negative group
velocity. The size of this region and the minimum value of V depend on the accuracy, and
also on the method. Hence, increasing the accuracy decreases the region and the minimum.
However, the main effect comes from the method. In fact, the 4th-order Padé method has
a smaller region with negative group velocity than the 6th-order Lagrange method. This
means that energy is transported with a better group velocity for the Padé method. Moreover,
as the order of accuracy increases, the range of wave number with positive group velocity
becomes wider, and the oscillations with typically 2h propagate at a higher speed in the
negative direction.
In [29] the authors also used a wave analysis to study the propagation of the spurious
oscillations. We show that the 4th-order Padé method is equivalent to a three-point semi
discretization of the advection equation. Let A denote the Banach algebra of continuous
linear operators of L 1 , that is, A = L(L 1 ). Clearly, (τ−h + τh )/4 + id ∈ A, for all h ∈ R.
Now, we prove that this operator has an inverse in A.
To see that it is one-to-one, let ψ ∈ L 1 be a function such that
µ ¶
τ−h + τh
+ id (ψ) = 0.
4
Then applying the Fourier transform to this equation yields
µ ¶
1
1+ cos(k̂h) 9 = 0.
2
So, it is one-to-one. The proof that it is onto is analogous, and existence of the inverse
follows from the open mapping theorem (see, for example, [17]).
Consider Eq. (7) for φ ∈ L 1 . Then, using the 4th-order Padé method we obtain
µ ¶µ ¶−1
d φ̄ 3c τ−h/2 − τh/2 τ−h + τh ¡ ¢
+ + id τ−h/2 + τh/2 φ̄ = 0.
dt 4 h 4
FIG. 4. Plot of distribution: t = 0 (top), t = 5h

c
(middle), t = 16h
c
(bottom).
The operators in the second term of the LHS are all Toeplitz and so commute. Hence,
µ ¶ µ ¶
τ−h + τh 2 d φ̄ τ−h − τh
+ id +c φ̄ = 0,
3 3 dt 2h
as desired. In particular we see that the “oscillatory” part travels at the speed −3c, in
agreement with the group velocity V( 12 ) for the 4th-order Padé method shown in Fig. 3.
Figure 4 illustrates the effect of the discretization in the error propagation. The problem
involves the time evolution of a periodic square profile
½
1 if |x01 − 1/2| < 1/8,
φ0 (x) =
0 otherwise,
where x 01 = x − [x], with [x] the integer part of x. The 4th-order accurate Padé finite
volume method is used, with periodic boundary conditions, in a grid comprising 64 control
volumes. Time marching is implemented with the 4th-order Runge–Kutta (RK) method,
using a value of c1t
h
= 0.1 for the CFL number. This level of CFL guarantees that the error
is predominantly due to spatial discretization. The numerical solution shown in Fig. 4 is as
predicted by the group velocity (note that V(1/3) = 0).
Remark. In the linear hyperbolic constant speed problem considered, the conservative
and nonconservative equations give rise to equivalent numerical approximations using the
finite volume and the finite difference methods. However, in the general case, when the
speed is a function of the space coordinate, the conservative and nonconservative formu-
lations, although equivalent from the analytical point of view, yield different numerical
approximations. To investigate the influence of both forms of the equation in the numerical
approach, we consider the hyperbolic equation
∂φ ∂ xφ
+ =0 (12)
∂t ∂x
or, in the non-conservative form,
∂φ ∂φ
+ + φ = 0. (13)
∂t ∂x
In the finite volume method, as usual, we consider Eq. (12) in integral form:
∂ φ̄ τ−h/2 − τh/2
+ xφ = 0. (14)
∂t h
The analytic solution of Eq. (12) is readily obtained using the method of characteristics
(see, for example, [13]). It is given by
φ(t) = e−t φ0 (x̂e−t ), t ≥ 0.
So, the characteristics are exponential curves in the (t, x)-plane and on these curves the
function decreases exponentially. That is, the initial distribution spreads and fades.
Now, applying the Fourier transform to Eq. (13) we obtain
∂8 ∂8
−k = 0. (15)
∂t ∂k
Hence, the spectrum is conserved on curves k = k0 e−t . Analogously, we have for the
spectrum of the sliding averages
µ µ ¶¶
∂ 8̄ ∂ 8̄ kh kh
−k = 1− cot 8̄. (16)
∂t ∂k 2 2
The characteristics are the same as for the 8 equation. However, 8̄ changes over these
curves as
µ ¶
dz kh kh
=1− cot
dt 2 2
and dz
dt
≥ 0, 0 ≤ k̃ ≤ 12 ; that is, the spectrum 8̄ increases over the characteristics. Therefore,
the spectra 8 and 8̄ have different evolution laws, and it is expected that their numerical
approximations have different characteristics too.
Indeed, using the Padé finite difference yields an equation similar to (15) with κfd in the
place of k, where
3 sin(2k̃π )
κfd = k
2k̃π(2 + cos(2k̃π ))
and a change over the characteristics of
3 + 6 cos(2k̃π )
ż fd = −1 + 2
.
(2 + cos(2k̃π ))
FIG. 5. Plot of the error in the change rate for Padé finite volume (thicker line) and Padé finite difference
(thinner line).
The Padé finite volume method yields
κfv = κfd
and a change over the characteristics of
3 cos(2k̃π ) sin2 (k̃π )

ż fv = 2
.
(2 + cos(2k̃π ))
The error in the rate of change is shown in Fig. 5. Both methods can resolve wave
components with small cell wave length, but start to deviate from the correct decay as
k̃ → 12 . However, it is interesting to note that the Padé finite volume method yields a smaller
error in the decay, as compared with the finite difference method.
Remark. The leading term of the truncation error of an 2r th-order Padé interpolation
method, with r = m + n, is proportional to the 2r th derivative of the (smooth) function.
So, it is exact for all polynomials of degree smaller than 2r . We can use this fact to write
a general matrix equation for the coefficients ai , i = 1, . . . , m, bi , i = 1, . . . , n. Given the
linearity of Eq. (4), we can work with a basis of the vector space 52r −1 of polynomials
of degree smaller than 2r . The natural basis being the monomials x j , j = 1, . . . , 2r − 1,
substituting each basic element x j , j = 1, . . . , 2r − 1 into Eq. (4) and considering the origin
yields
 
X m Xn
b
 a j jl −
j
( j l+1 − ( j − 1)l+1 ) (1 + (−1)l ) = 0 (17)
j=1 j=1
l + 1
for all l = 1, . . . , 2r − 1. This is a 2r × r system of linear equations. However, due to the

symmetry of the stencil and the equality of symmetric coefficients, Eq. (17) is trivial for
l odd. The latter system of equations is equivalent to the matrix equation
· ¸
a
[A B] = d, (18)
b
where the block matrices are defined as

 
1 1 ··· 1 1
1 ··· (m − 1)2 m2 
 4 
 .. .. .. .. .. 
A=
. . . . . 

 2r −4 
 1 2 · · · (m − 1)2r −4 m 2r −4 
1 22r −2 · · · (m − 1)2r −2 m 2r −2
 
1 1 ··· 1 1
 1 
 7
··· (n − 1)3 − (n − 2)3 n 3 − (n − 1)3 
 3 3 3 3 
 .. .. .. .. .. 
B=
 . . . . . 

 1 (n − 1)2r −3 − (n − 2)2r −3 n 2r −3 − (n−1)2r − 3 
 2r − 3 1
(22r −3 − 1) · · · 
 2r − 3 2r − 3 2r −3 
(n − 1)2r −1 − (n − 2)2r −1 n 2r −1 − (n − 1)2r −1
1
2r − 1
1
2r − 1
(22r −1 − 1) · · · 2r − 1 2r − 1
 
− 12
 
 0 
d=
 ..
.

 . 
0
Next we prove existence and uniqueness of a solution of (18). The proof is divided into
two steps. In the first, we prove that the finite difference Padé interpolation problem can
be written as a Hermite interpolation problem. Then, in the second, we prove that IP is
equivalent to a Padé finite difference problem for the primitive of φ.
The finite difference Padé interpolation problem can be stated as follows:
Given a smooth function φ ∈ C ∞ , and natural numbers m, n ∈ N find coefficients ai , i =
1, . . . , m and bi , i = 1, . . . , n, such that
Ã n !
X
m X
m
1 X Xn
¡ ¢
0 0 0
ai τ−i h φ + φ + ai τi h φ = bi τ−i h φ − bi τi h φ + o h 2(m+n)−1 . (19)
i=1 i=1
h i=1 i=1
Consider the real numbers xi , φi(k) , k = 0, . . . , n i − 1, i = 0, . . . , m 0 with
x0 < x1 < · · · < xm 0 .
The Hermite interpolation problem IH for these data consists of determining a polynomial
Pm 0
P ∈ 5n 0 where n 0 + 1 = i=0 n i , which satisfies the interpolation conditions
P (k) (xi ) = φi(k) , k = 0, . . . , n i−1 , i = 0, . . . , m 0 .
There is a theorem (see, for example, [27]) which gives existence and uniqueness of a
solution of IH. It is given by
X i −1
m 0 nX
P(x) = φi(k) L ik (x), (20)
i=0 k=0
where L ik ∈ 5n 0 are generalized Lagrange polynomials. They are defined as follows: starting
with the polynomials
m0 µ ¶
(x − xi )k Y x − xj nj
lik (x) = , 0 ≤ i ≤ m 0, 0 ≤ k < ni ,
k! j=0
xi − x j
j6=i
put
L i,ni −1 (x) = li,ni −1 (x), i = 0, . . . , m 0
and recursively for k = n i − 2, . . . , 0,
i −1
nX
ν
L ik (x) = lik − lik (xi )L iν (x).
ν=k+1
Consider the points x±i = ±i, i = 1, . . . , n.

Claim.
0
ai = −L i1 (0), i = 1, . . . , m
and
0
bi = L i0 (0), i = 1, . . . , n.
To prove this we note that n 0 = 2r − 1, with r = n + m. Then, Eq. (19) is exact for
the polynomial P given by Eq. (20). Substituting P in Eq. (19) and taking into account the
uniqueness of the Hermite polynomial proves the claim.
For the second part of the proof we first observe that given φ ∈ L 1 we have
τ−h/2 2 − τh/2 2
φ̄ = ,
h
where 2 ∈ C is the primitive of φ,

Z x
2(x) = φ(u) du, x ∈ R,
x0
for some x0 ∈ R.
We conclude that the IP is equivalent to the following problem:
1, . . . , m, and b̃i , i = 1, . . . , n, such that
Ã n !
X
m X
m
1 X Xn
¡ ¢
0 0 0
ai τi h 2 + 2 + ai τi h 2 = b̃i τ−i h 2 + b̃i τi h 2 + o h 2(m+n)−1 . (21)
i=1 i=1
h i=1 i=1
Using the result of the claim we compute ai , i = 1, . . . , m, and b̃i , i = 1, . . . , n. Then

an easy computation shows that
bn = b̃n
bn−1 − bn = b̃n−1
.. (22)
.
b1 − b2 = b̃1 .
Equations (21), (22) represent the reconstruction via a primitive function, which is used
in [8]. Because of the stencil selection in the ENO approach (see [11]) the reconstruction
via the deconvolution and the primitive function yield different methods, the deconvolution
being more accurate in some cases. So, it is expected that the same may occur with the Padé
method, if an ENO-like stencil selection is used.
Remark. We have performed the Fourier analysis in the space S 0 (R). Another possibility
is to work with the Fourier transform on S 0 (T1 ), where T1 = R/Z is the 1-torus.3 That
corresponds to work with periodic distributions on R. Then, to each distribution φ ∈ S 0 (T1 )
we may associate its Fourier series (see [25]):
X
φ= 8(k)(φ)eik x̂ .
k∈Z
Formally, the analysis proceeds as above, with the distinction that now the distributions
have a “discrete” spectrum; that is, we should use k ∈ Z instead of k ∈ R in the expressions
like (11).
As remarked in [29], the semidiscrete equation, Eq. (3), can be studied using a band-
limited function (see also [18]). This fits into the present framework, since it corresponds
to periodic functions with a finite discrete spectrum.
2.2. Parabolic Diffusion Equations

Similarly to the previous subsection, we use the linear diffusion equation
∂φ ∂ 2φ
=ν 2 (23)
∂t ∂x
as a model for the description of the numerical approximation. Again, in the finite volume
approach we first integrate Eq. (23) over each control volume σ j of a partition {σ j } j∈Z ,
yielding
· ¸
d φ̄ φ 0j+1 − φ 0j−1
=ν , j ∈ Z. (24)
dt j+1/2 h
The corresponding IP can be stated as follows:
3
Clearly T1 is diffeomorphic to S 1 , where S 1 denotes the unit circle. However, in higher dimensions the
generalization is Tn , not S n .
TABLE IV
Coefficients and Leading Term of the Truncation Error for Some Padé Methods
Order a1 a2 a3 b1 b2 b3 Truncation error

4
4 1
10
— — 6
5
— — − 200
h
f (5)
2 51 3 23h 6 (7)
6 11
— — 44 44
— 55440
f
8
8 344
1179
23
2558
— 265
262
155
786
— − 79h
2971080
(9)
f
38223h 12
12 329913
725308
18387
362654
619
725308
813155
1087962
835345
2175924
49483
2175924
− 290413323200
(13)
f

1, . . . , m, and bi , i = 1, . . . , n, such that
Xm Xm
ai τ−i h φ 0 + φ 0 + ai τi h φ 0
i=1 i=1
Ã n !
1 X X ¡ ¢
n
= bi τ(−i+1/2)h φ̄ − bi τ(i−1/2)h φ̄ + o h 2(m+n)−1 . (25)
h i=1 i=1
The solution of IP above can be computed as above using either the Taylor series expan-
sion or the second derivative of corresponding Hermite interpolating polynomial. Table IV
lists the coefficients and the leading term of the truncation error for the 4th-, 6th-, 8th-, and
12th-order Padé interpolations with m + 1 ≥ n ≥ m.
The Fourier analysis for the parabolic equation proceeds as in the analysis of hyperbolic
equation. The exact relation yields
d 8̄
= −νk 2 8̄,
dt
so,
8̄(t) = 8̄0 eν k̂ t ,
2
t ∈ R. (26)
This means that the spectrum of the sliding average decays at an exponential rate, which is
proportional to the square of the wave number. The Padé interpolation provides an analogous
law of evolution with a decay ratio νh k 2 , where
v
u µ ¶ Pn ¡¡ ¢ ¢
³ ν ´1/2 2 u kh b sin j − 1
kh
h
= tsin j=1
Pm
j 2
, k∈R (27)
ν kh 2 1 + 2 j=1 a j cos( jkh)
Figure 6 shows (νh /ν)1/2 as a function of the grid wave number for the 2nd-, 4th-, 6th-,
8th-, and 12th-order methods. As expected, increasing the order of the method improves
the spectral resolution of the method.
2.3. Transport Equations

In this subsection we consider the linear transport equation
∂φ ∂φ ∂ 2φ
+c = ν 2. (28)
∂t ∂x ∂x
Integration of Eq. (28) over each control volume σ j of a partition {σ j } j ∈ Z yields
· ¸
d φ̄ φ j+1 − φ j−1 φ 0j+1 − φ 0j−1
+c =ν , j ∈ Z. (29)
dt j+1/2 h h
FIG. 6. Plot of decay ratio versus grid wave number for diffusion using the Padé method: (a) 4th order,
(b) 6th order, (c) 8th order, (d) 12th order, (e) CD 6th order, (f) CD 8th order, (g) CD 12th order.
One obvious way to solve Eq. (29) is to apply the Padé interpolation, Eqs. (4) and (25), to
its convection and diffusion components, respectively. However, Mahesh in a recent paper
[21] proposed a coupled-derivative (CD) approach to the finite difference compact scheme,
in which both derivatives are computed simultaneously. The resulting method showed a
better spectral resolution than the usual uncoupled method.
In the CD approach the IP problem for the convection term can be stated as follows:
Given a smooth function φ ∈ C ∞ , and natural numbers m, n, o ∈ N, find coefficients
ai , i = 1, . . . , m, bi , i = 1, . . . , n, and ci , i = 1, . . . , o, such that
X
m X
m
ai τ−i h φ + φ + ai τi h φ
i=1 i=1
Ã o !
X
n X
n X X
o
= bi τ(−i+1/2)h φ̄ + bi τ(i−1/2)h φ̄ + h ci τ−i h φ 0 − ci τi h φ 0
i=1 i=1 i=1 i=1
¡ ¢
+o h 2(m+n+o)−1
. (30)
For the diffusion term we have:

Given a smooth function φ ∈ C ∞ , and natural numbers m, n, o ∈ N find coefficients
di , i = 1, . . . , m, ei , i = 1, . . . , n, and f i , i = 1, . . . , o, such that
X
m X
m
di τ−i h φ 0 + φ 0 + di τi h φ 0
i=1 i=1
Ã n !
1 X X X X
n o o
= ei τ(−i+1/2)h φ̄ − ei τ(i−1/2)h φ̄ + f i τ−i h φ − f i τi h φ
h i=1 i=1 i=1 i=1
¡ ¢
+ o h 2(m+n+o)−1 . (31)
In Tables V and VI we list the coefficients and the leading term in the truncation order for
the 6th-, 8th-, and 12th-order Padé interpolations with n − 1 ≤ m = o ≤ n. A comparison of
TABLE V
for Some CD Padé Methods: Convection
Order a1 a2 b1 b2 c1 c2 Truncation error

7 15 1 h6
6 16
— 16
— 16
— 5040
f (6)
h8
8 17
36
— 53
54
− 1
108
1
12
— − 90720
f (8)
h 12
12 4
9
− 23
432
2755
2592
− 445
2592
1
27
− 1
216 97297200
f (12)
the constant in the leading term of the truncation error for the CD scheme with the uncoupled
one shows a smaller asymptotic error for the CD approach, for all orders.
Figures 2 and 6 show the phase speed for the convection and the decay ratio for the
diffusion term, respectively. As can be seen from these figures, the CD approach shows a
noticeably smaller error than the standard Padé method. A distinct feature of the CD method
is that it displays a decay rate higher than the true one for high cell wave numbers. This is
in contrast to the standard Padé, which exhibits a decay rate smaller than the true one for
all cell wave numbers.
Clearly, CD Padé methods may be applied to pure advection or pure diffusion problems.
However, in this case they require the computation of a “useless” derivative. This computa-
tion doubles the work required for the evaluation of the derivative, making it less attractive
than the uncouple Padé method for problems of pure advection or pure diffusion problems.
Remark. Equations involving second and higher derivatives can be discretized analo-
gously. For instance, the Korteweg–de Vries (KdV) equation
∂u ∂u 2 ∂ 3u
+ + 3 =0
∂t ∂x ∂x
requires the evaluation of the second derivative. This can be done using the CD Padé
approach solving the following IP problems:
ai , i = 1, . . . , m, bi , i = 1, . . . , n, and ci , i = 1, . . . , o, such that
X
m X
m
ai τ−i h φ + φ + ai τi h φ
i=1 i=1
Ã o !
X
n X
n X X
o
00 00
= bi τ(−i+1/2)h φ̄ + bi τ(i−1/2)h φ̄ + h 2
ci τ−i h φ + ci τi h φ
i=1 i=1 i=1 i=1
¡ ¢
+ o h 2(m+n+o)−1 (32)
TABLE VI
for Some CD Padé Methods: Diffusion
Order d1 d2 e1 e2 f1 f2 Truncation error

h6
6 − 1
8
— 3 — − 9
8
— 2016
f (7)
79h 8
8 − 16 — 13
4
− 108
1
− 23
18
— − 16329600 f (9)
12
12 − 8
27
1
216
15
4
− 63
324
− 40
27
25
432
− 73556683200
11273h
f (13)
TABLE VII
Coefficients and Leading Term of the Truncation Error of Some
CD Padé Methods for KdV Equation: Convection
Order a1 a2 b1 b2 c1 c2 Truncation error

9 25 1 h6
6 32
— 32
— 96
— 2520
f (6)
h8
8 11
48
— 109
144
− 361 1
48
— − 36288 f (8)
12
12 − 209
81
− 1153
2592
− 7955
15552
− 31315
15552
7
27
− 1
216
− 194594400
29h
f (12)
and
di , i = 1, . . . , m, ei , i = 1, . . . , n, and f i , i = 1, . . . , o, such that
X
m X
m
00 00
di τ−i h φ + φ + di τi h φ 00
i=1 i=1
Ã n !
1 X X X X
n o o
= 2 ei τ(−i+1/2)h φ̄ − ei τ(i−1/2)h φ̄ + f i τ−i h φ + f 0 φ + f i τi h φ
h i=1 i=1 i=1 i=1
¡ ¢
+ o h 2(m+n+o)−1 . (33)
Note the inclusion of the term f 0 φ in Eq. (33). In Tables VII and VIII we list the
coefficients and leading term of the truncation error for the 6th-, 8th-, and 12th-order CD
Padé methods.
Remark. The computational cost of the Padé finite volume method is the same as for
the Padé finite difference methods. Indeed, writing the IP problems in matrix form
A f (n) = B f, (34)
we see that the matrix A is the same as for the finite difference method. The common
practice is to perform an LU decomposition of the matrix A only once, and store the L and
U matrices. Hence, computation of the derivatives involves evaluating the RHS of Eq. (34)
followed by forward and backward substitution. Lele [18], using a Cholesky decomposition
of the symmetric part A for a tridiagonal scheme, gave an operation count of 5N , N , and
5N for multiply, divide, and add/subtract, respectively. A radix 2 FFT (see [3]) requires
TABLE VIII
Coefficients and Leading Term of the Truncation Error of Some CD Padé
Methods for KdV Equation: Dispersion
Order d1 d2 e1 e2 f0 f1 f2 Truncation error

h6
6 − 1
32
— 315
32
— −15 − 75
32
— 40320
f (8)
h8
8 − 201 — 421
40
− 401 − 159
10
− 51
20
— − 1039500 f (10)
12
12 − 34
261
7
8352
6486715
601344
− 330565
601344
− 1935
116
− 1570
783
10855
100224
− 1519333200
h
f (14)
2N log2 N multiplies and 3N log2 N adds. Also, as pointed out by Mahesh [21], the cost
incurred by the use of CD Padé methods is essentially the same as for the uncoupled one.
3. MULTI-DIMENSIONAL EQUATIONS
Higher dimensional equations require the average of the variables at the cell faces. For
example, the two-dimensional convection problem can be written as
µ x x ¶ µ x x ¶
d φ̄ τ−h 1 /2 φ̄ 2 − τh 1 /2 φ̄ 2 τ−h 2 /2 φ̄ 1 − τh 2 /2 φ̄ 1
+ c1 h 2 + c2 h 1 = 0,
dt h1 h2
where
Z
1
φ̄ = χh 1 ,h 2 ∗ φ,
h1h2 R×R
and the superscripts x 1 and x2 indicate the one-dimensional averaging in the x1 and x2
directions, respectively; that is, for example,
Z
x2 1
φ̄ = χh 2 ∗ φ.
h2 R
The choice between sliding averages or point values for methods that are 2nd order or
lower is immaterial, since both variables differ by a 2nd-order term only. However, for
higher order methods this choice can influence the stencil for the interpolation. Take for
instance the usual QUICK scheme of Leonard [19] and the 4th-order method proposed in
[20]. Both methods are based in point values. Now, because cross derivatives appear in the
Taylor series, as the averaging is needed in the transverse direction to the interpolation, they
require 13 and 25 points in the stencil to obtain 3rd- and 4th-order accuracy, respectively.
Hence, these methods need points out of the Cartesian product of the one-dimensional
stencils for their one-dimensional convection counterpart. All methods developed above
use the sliding average of the dependent variable for interpolation. The integration in its
definition is performed in both directions, which accounts for the variations in the transverse
direction.
In conclusion, the stencil of the methods developed above for higher dimensional prob-
lems is the Cartesian product of the stencils of their one-dimensional counterparts. However,
when using point values we need points out of the latter product in order to take into account
transverse variation of the variables.
Another feature, unique to the interpolation on higher dimensional domains, is the
anisotropy in the interpolation of the convection term. In fact, the symmetry to the group
of rotations of the Laplacian operator means that it is isotropic, or, in other words, that a
sinusoidal wave propagating in any direction will experiment the same rate of decay. The
discretization of the Laplacian operator presented earlier possesses this property as well.
However, the convection term has the velocity direction as a preferred direction (see [29]).
It may be shown that
1
cα = (cos(α) sin(kg cos(α))c(k cos(α)) + sin(α) sin(kg sin(α))c(k sin(α))),
kh
FIG. 7. Polar plot of phase speed anisotropy for advection using (a) 4th order Lagrange, (b) 4th order Padé,
(c) 6th order Lagrange, (d) 6th order Padé, (e) 6th order CD Padé, (f) 8th order Padé, (g) 8th CD Padé, (h) 12th
order Padé, and (i) 12th order CD Padé, for cell wave number k̃ = 201 , 202 , . . . , 10
20
.
where cα is the phase velocity and α is the angle between the direction of propagation and
the x1 axis.
The anisotropic propagation is displayed in Fig. 7 for several methods. The curves in this
figure are polar plots of cα /c at fixed cell grid wave numbers. For each curve, the radial
distance at an angle α represents cα obtained for a wave propagating in that direction. The
curves corresponds to k̃ = 201
, . . . , 10
20
; the outmost curves correspond to small k̃. For these
waves the propagation is nearly isotropic and the phase speed is close to exact. For larger k̃
the waves have a small phase speed and the propagation is anisotropic. It may be seen that
in the Padé methods the anisotropy error is noticeable in a narrower range of short waves
lengths, when compared to the Lagrange methods. Moreover, the CD methods display an
even smaller anisotropy error for all orders and wave lengths.
Remark. The notion of group velocity applies to multi-dimensional equations as well.

The definition is the same as above, with trivial adaptations (for instance, now we must
consider the wave vector k). In the general case, group velocity depends on the wave
number, and its direction can be different from that of the wave vector.
Remark. In [24] it is mentioned that experimental evidence suggests that acoustic waves
are strogly coupled to many mechanisms encountered in turbulence. The error in wave
equation describing the propagation of the acoustic with Padé methods is predominantly in
the high wave number range. So, except for the unusual case of vanishing sound speed, it is
of high frequency and thus analogous to sound rays. In this situation, we can apply the theory
of geometric acoustics (see [16]) to the propagation of error “rays” in nonhomogeneous
medium, that is, where c is not constant. It can be shown that the magnitude of the wave
vector varies according to the simple law k = ω/c, while its direction changes according to
1
K = − (n · grad c),
c
where K is the curvature of the curve described by the ray and n is the unitary vector of
the principal normal to the curve. In other words, the error rays turn in the sense of the
diminishing of the speed c.
4. OPTIMIZED PHASE ERROR AND SPECTRAL PADÉ INTERPOLATION
In addtion to maximal order schemes, classes of parametrized nonmaximal order Padé

finite volume methods may offer optimization opportunities which go beyond the maximal
order of accuracy. In [18], a class of compact finite difference schemes with spectral-like
resolution is proposed. The methods are obtained by constraining the error to be of 4th
order, and the phase speed to match the exact value of the phase speed at some predefined
locations. In [8] the authors define a family of parametrized schemes which minimizes the
L 1 -norm of the error in the phase velocity for a given interval of cell wave numbers. That
is, they minimize the dispersion error,
Z k̃ max
E d (a1 , k̃ max ) = |c(k̃) − c| d k̃,
0
with the restriction that the method must be of at least 4th order of accuracy. Both methods
in [18] and [8] use the stencil of the 6th-order Padé method referred to above.
In [8] the authors concluded that a1,opt takes values in 13 ≤ a1,opt ≤ 0.431 (the value
of a1,opt = 13 corresponds to the 6th-order Padé method). They also considered optimal
methods with respect to anisotropic errors, the values of the optimum coefficient being in
1
3
≤ a1,opt ≤ 0.408.
Other target functions may be considered; for example, the standard Padé interpolation
optimizes the order of accuracy in a given stencil. It is evident that the optimum solution
depends on the target function. Consider the ability of a method to handle discontinuities,
the occurrence of which in compressible inviscid flows is relevant in applications. In [29]
it is shown that
Z π/ h ¯ ¯2 ¯ µ ¶¯2
¯ ¯ ¯ ¯
k²k22 = ¯ h ¯ · ¯ sin u(c(u) − c)t ¯ du ,
¯ sin(uh/2) ¯ ¯ 2 ¯ π
0
where ² is the global error of a semidiscrete approximation to this problem. We look for the
optimum methods as t → 0 and t → ∞. Using sin(u(c−c)t/2) ∼ u(c−c)t/2, for sufficiently
small t ∈ R, it follows that the former minimizes the L 2 -norm of the consistency error
Z π/ h
h/2 du
Th = iu(c − c)eiux .
−π/ h i sin(uh/2) 2π
When t → ∞, the least error is obtained with the method which maximizes the order of
accuracy. To show this we first perform a change of coordinate ū = ut; then,
Z ¯ ¯2 ¯ µ ¶¯2
¯
πt/ h ¯ ¯ ¯
k²k22 = ¯ h ¯ · ¯sin ū(c − c) ¯ du
¯ sin(ūh/2t) ¯ ¯ 2 ¯ tπ
0
Z πt/ h ¯ ¯ ¯ µ ¶¯2
¯ h ¯2 ¯ ¯
∼ ¯ ¯ · ¯sin ū(c − c) ¯ d ū
¯ ūh/2t ¯ ¯ 2 ¯ tπ
0
Z π/ h ¯ ¯ ¯ µ ¶¯2
¯ h ¯2 ¯ ¯
= ¯ ¯ · ¯sin u(c(u) − c)t ¯ du .
¯ uh/2 ¯ ¯ 2 ¯ π
0
The result follows from the fact that, in the topology of S 0 , we have
sin(n x̂)
lim = π δ. (35)
n→∞ x̂
Then, for sufficiently large values of t ∈ R the value of k²k22 is dominated by the error in
the phase velocity near the origin. A Taylor series expansion shows that this error decays
at the same rate as the truncation error.
To prove (35) we first use integration by parts to show that if f ∈ C 1 ∩ C0 , where C0 =
{ f ∈ C: limx→±∞ f (x) = 0}, f 0 ∈ L 1 , and for all n ∈ N the integral
Z
sin(nx) f (x) d x
R
exists; then
Z
lim sin(nx) f (x) d x = 0.
n→∞ R
Now, given a ψ ∈ S it is easy to see that the function
ψ(y) − ψ(0)
f : x ∈ R 7→ lim ∈R
y→x y
belongs to C 1 ∩ C0 and f 0 ∈ L 1 (just use the Maclaurin series and the fact that ψ 0 ∈ L 1 ).
The result follows if we observe that sin(n
x̂
x̂)
∈ 3, so
¿ À Z
sin(n x̂) sin(n x̂)
,ψ = ψd x → π ψ(0)
x̂ R x̂
R
for all ψ ∈ S, where we have used the fact that R
sin(nx)
x
d x = π (as an improper integral
of Riemann).
Spectral methods are based on representing the solution of a problem as a truncated

series of global smooth functions. The latter are often trigonometric functions for a pe-
riodic boundary condition, or Chebyshev or Legendre polynomials for general boundary
conditions. A prominent characteristic of these methods is exponential convergence.
Using the fact that the Padé method corresponds to the use of the Hermite polynomial,
we relax the requirement of being a global smooth function and define a piecewise spectral
Padé method (SPM) as follows:
In a mesh comprising N control volumes, use the (N/8th)-order Padé method.
Then, the error decays at a rate of O(N N /8 ), which is faster than any finite power of
N −1 . Also, as the Fourier analysis have shown, the phase speed using the SPM will stay
progressively close to the exact one, with respect to both the cell wave number and the
anisotropy error. In actual computations we use N = 2 p for p = 4, . . ..
EXAMPLE 4.1. We close this section with an example comparing several Padé methods.
We resolve the problem
∂φ ∂φ
+ = 0, x ∈ [0, 1], t ≥ 0
∂t ∂x
with periodic boundary conditions and initial profile
½
K e−1/((x−a)(b−x)) if x ∈ [a, b],
φ(0, x) =
0 otherwise,
where a, b, K ∈ R are constants. In all cases, a = 14 , b = 34 , K = e1/16 (to make the peak
unitary) and the temporal discretization is accomplished with the 4th-order RK algorithm.
This bell-shaped profile corresponds to a Cc∞ -distribution with high gradients in the fore and
aft portions of the bell and a high curvature region close to the peak. It can serve as test case
for peak resolution and/or monotonicity. A final time T = 14 is used in all simulations for
all grids and methods at CFL = 10 1 4
. At this CFL level the errors due to time discretization
are negligible compared with the spatial terms. The maximum error at T is calculated and
is shown in Fig. 8.
The optimized method of Gaitonde and Shang display a smaller error than the 6th-order
method for grids with less than 32 control volumes. In this range of grid parameter, the
errors are still in the region of relatively high cell wave numbers and so where the optimized
method has a better performance. After that, most of the monochromatic waves with small
cell wavelength are resolved, the error is dominated by the asymptotic order, and the 6th
becomes superior. The exponential decay in the error for the SPM is also evident.
Remark. All methods considered above present an error in the phase, but no error in
amplitude. This is the case for any Padé method with a symmetric stencil for the sliding
averages and the variable (it is immediate from the Fourier analysis). This is valid for the
infinite domain or for periodic boundary conditions. However, in general, the boundary
4
To avoid any contamination from time errors, the result for the SPM in the grid with 128 control volumes is
obtained with the exact solution of the semidiscrete equation. That is, writing the latter as d φ̄/dt = (σc /1t)C φ̄ we
compute φ̄(T ) = ecσ C(T /1t) φ̄ 0 = Seσc 3(T /1t) S −1 φ̄ 0 , where σc is the CFL number, C is defined in Section 6, S is the
matrix with the eigenvectors of C in its columns, and 3 is the corresponding diagonal matrix with the eigenvalues
in its diagonal.
FIG. 8. Plot of maximum error versus grid parameter for pure advection using various Padé methods. Solid
thin gray lines are asymptotic curves Ah r , with A a constant and r the corresponding order of accuracy. The solid
black line is the spectral method. The thick solid gray line is the optimum 4th method of Gaitond and Shang with
k̃ max = 38 . Points in increasing gray levels are the numerical solutions corresponding to 2th, 4th, 6th, and 8th Padé
methods, respectively.
conditions in a finite domain will break this symmetry and the numerical solution will
display some damping effect (see [4], and also Section 5).
5. BOUNDARY CONDITIONS
Boundary conditions represent a particular case of the general IH problem. However,

boundary effects may affect the accuracy of the solution as well as its stability. In this
section we study their effect on the accuracy, and we defer to Section 6 our analysis of their
effect on the stability.
Some authors have suggested that the approximation on the boundary, if one order of
accuracy below that at interior interpolation, may not affect the global (uniform) order of
accuracy (see, for example, [4, 24, 29]). In [29] this idea is supported using the notion of
the reflection ratio. We use this concept to show that in the case of the Padé finite volume
method, the boundary approximation has a determining effect on the global accuracy.
To compute the reflection ratio, we need to find normal, or fundamental, solutions of the
time-Fourier transform of the semidiscrete equation. We write the latter as
d φ̄ φ j+1 − φ j
+c = 0, j ∈ Z. (36)
dt j+1/2 h
Then, we define the time-Fourier transform of φ as

Z
8= φe−i ω̂t dt.
R
Note that we use the same letter for the time-Fourier transform as for the space-Fourier
transform. Since we do not use both at the same time, this should not cause confusion. We
take the time-Fourier transform of (36) and obtain
8 j+1 − 8 j
iω8̄ j+1/2 + c = 0.
h
Using the 4th-order Padé method and rearranging terms we obtain
(κ −1 + 1 + κ)is = 3(κ − κ −1 ),
where s = ωh/c, and 8̄ j+1/2 /8̄ j−1/2 = κ, j ∈ Z. The characteristic roots are
√ √
−2 s − i 3 3 − s 2
κ1 =
s − 3i
√ √
−2 s + i 3 3 − s 2
κ2 = .
s − 3i
We consider the following approximations at the boundary:
1. 3rd-order Padé
1¡ ¢
φ N + 2φ N −1 = 5φ̄ N −1/2 + φ̄ N −3/2 .
2
2. 4th-order Padé
1¡ ¢
φ N + 3φ N −1 = 17φ̄ N −1/2 + 8φ̄ N −3/2 − φ̄ N −5/2 .
6
At the control volume closest to the boundary Eq. (36) with the 3rd-order Padé method
at boundary yields
¡ ¢ ¡5 ¢ ¡ ¢
5
2
+ 1
2 κ1
3 1+ 1
κ1
ρ 2
+ 1
2 κ2
3ρ 1 + 1
κ2
is(1 + ρ) = − + − + ,
1 + κ21 4+ 1
κ1
+ κ1 1+ 2
κ2
4+ 1
κ2
+ κ2
whence, solving for the reflection ratio, we find

√ √
3(−6i + s) 3 − s 2 − 3(−6i + s + is 2 )
ρ(s) = √ √ .
3(−6i + s) 3 − s 2 + 3(−6i + s + is 2 )
Analogously, we find for the 4th-order Padé method at the boundary

√ √
3 3 − s 2 (24 + s(−6i + s)) − 3(24 − s(6i + (3 − is)s))
ρ(s) = √ √ .
3 3 − s 2 (24 + s(−6i + s)) + 3(24 − s(6i + (3 − is)s))
In Figs. 9 and 10 the absolute value and the phase of the reflection ratio, respectively, are
√ the Padé method, the region of propagating waves enlarges from 0 ≤ s ≤ 1
shown. Using
to 0 ≤ s ≤ 3. Moreover, the spurious oscillations generated at the boundary by the 4th-
order Padé method, independent of the boundary approximation, are much smaller than
those generated by the 2nd-order Padé finite difference method. The difference between the
absolute value of the reflection ratio for the 3rd- and 4th-order boundary conditions does
not seem significant (note that the difference in the phase of the reflection ratio is not small).
However, a Taylor series expansion of the reflection ratio yields
1. 3rd-order Padé
is 3
ρ(s) = − + o(s 3 ).
72
FIG. 9. Plot of the absolute value of the reflection ratio versus specific frequency s using the Padé method:
(a) 2nd-order finite difference with 1st order at boundary; (b) 4th order with 3rd order at boundary, (c) 4th order
with 4th order at boundary.
2. 4th-order Padé
s4
ρ(s) = − + o(s 4 ).
96
We conclude that the implicit character of the Padé method makes it more sensitive to the
boundary closure. Indeed, for a 4th-order Padé method for the inner points and the 3rd-order
boundary closure, the global accuracy of the method is of 3rd order.
EXAMPLE 5.1. We resolve the problem
∂φ ∂φ
+ = 0, x ∈ [0, 1], t ≥ 0
∂t ∂x
φ(t, 0) = 0
½ −1/((x−a)(b−x))
Ke if x ∈ [a, b],
φ(0, x) =
0 otherwise,
FIG. 10. Plot of the phase of the reflection ratio versus specific frequency s using the Padé method: (a) 2nd-
order finite difference with 1st order at boundary, (b) 4th order with 3rd order at boundary, (c) 4th order with 4th
order at boundary.
TABLE IX
Effect of the Boundary Closure on the Maximum Error of the 4th-Order
Accurate Padé Finite Volume Method
3rd order 4th order

Boundary
grid k²k∞ Numerical order k²k∞ Numerical order
8 1.3831 × 10−1 — 1.0871 × 10−1 —

16 2.5453 × 10−1 −0.88 2.2546 × 10−1 −1.06
32 8.8983 × 10−2 1.52 7.7016 × 10−2 1.55
64 5.5606 × 10−3 4.00 5.3655 × 10−3 3.84
128 4.5039 × 10−4 3.63 3.4711 × 10−4 3.95
256 4.9249 × 10−5 3.19 2.1577 × 10−5 4.01
where a, b, K ∈ R are constants. In all cases, a = 14 , b = 34 , K = e1/16 , and the temporal

discretization is accomplished with the 4th-order RK algorithm. A final time T = 12 is used
in all simulations for all grids and methods at CFL = 10 1
. This value of T corresponds to the
critical instant when the peak crosses the outlet. At this CFL level the errors due to time
discretization are negligible compared with the spatial terms. The error at T is calculated
and reported as the L ∞ -norm. Table IX shows the effect of the boundary approximation on
the accuracy of the 4th-order Padé finite volume method.
As expected the global order of accuracy is determined by the boundary closure: it is of
3rd order for the 3rd-order accurate closure, and it is of 4th order for the 4th-order accurate
closure. Moreover, the asymptotic order is only achieved for relatively fine grids.
The effect of the boundary closure on global accuracy of solution in a transport problem
using the CD approach may be even more pronounced. Indeed, in this case, not only is the
problem of an elliptic character, but the coupling in the interpolation problem makes the
implicit nature of the method even stronger. To illustrate this we consider.
EXAMPLE 5.2. We solve the transport problem
dφ 1 d 2φ
= , x ∈ [0, 1],
dx Pe d x 2
φ(0) = 0
ePe Pe
φ 0 (1) = .
ePe − 1
The exact solution is
1 − ePe x
φ(x) = , x ∈ [0, 1].
1 − ePe
In all grids we use Pe = 10 and compute the maximum error in the sliding average, the inter-
face values of the variable and its first derivative. Tables X, XI, and XII show the effects of the
boundary approximation on the accuracy of the sliding average, interface values, and inter-
face derivative, using the 6th-order CD Padé finite volume method. The boundary closures
considered are:
TABLE X
Effect of the Boundary Closure on the Maximum Error of the Sliding Average
Using the CD 6th-Order Accurate Padé Finite Volume Method
5th order 6th order

Boundary
8 5.3693 × 10−3 — 2.9127 × 10−3 —

16 3.6119 × 10−4 3.89 1.2802 × 10−4 4.51
32 1.8785 × 10−5 4.27 3.8778 × 10−6 5.05
64 8.5455 × 10−7 4.46 9.4831 × 10−8 5.67
128 3.2238 × 10−8 4.73 1.8570 × 10−9 5.67
256 1.1071 × 10−9 4.86 3.2359 × 10−11 5.84
1. Inlet: 6th-order Dirichlet

µ ¶
1 49 28 1447 67 5 1
φ00 − 4φ10 = − φ0 − φ1 + φ̄ 1/2 − φ̄ 3/2 + φ̄ 5/2 − φ̄ 7/2 .
h 6 3 72 24 2 72
2. Outlet: von Neumann

(i) 5th-order
8 h 31 1
φ N + φ N −1 = (φ N0 − 2φ N0 −1 ) + φ̄ N −1/2 − φ̄ N −3/2 .
7 7 14 14
(ii) 6th-order
27 3h 0 325 13 1
φN + φ N −1 = (φ − 3φ N0 −1 ) + φ̄ − φ̄ + φ̄ .
23 23 N 138 N −1/2 69 N −3/2 138 N −5/2
Again, the imposition of lower order boundary conditions at the outflow (and only there)
lowers the overall accuracy. It is also noteworthy that in the Padé method the high order of
accuracy for the sliding averages is the same for the interface values and the derivatives at
the interfaces. This contrasts with usual methods for which the order of the latter two values
decrease one order of accuracy for each derivative (see, for example, [11]).
TABLE XI
Effect of the Boundary Closure on the Maximum Error of the Interface Values
5th order 6th order

Boundary
8 7.0614 × 10−3 — 3.7353 × 10−3 —

16 5.5274 × 10−4 3.68 1.9197 × 10−4 4.28
32 2.8681 × 10−5 4.27 5.7913 × 10−6 5.05
64 1.1643 × 10−6 4.62 1.2720 × 10−7 5.51
128 4.1542 × 10−8 4.81 2.3622 × 10−9 5.75
256 1.3877 × 10−9 4.90 4.0085 × 10−11 5.88
TABLE XII
Effect of the Boundary Closure on the Maximum Error of the Interface Derivative
5th order 6th order

Boundary
8 6.8994 × 10−2 — 3.6472 × 10−2 —

16 5.5270 × 10−3 3.64 1.9191 × 10−3 4.25
32 2.8682 × 10−4 4.27 5.7910 × 10−5 5.05
64 1.1643 × 10−5 4.62 1.2719 × 10−6 5.51
128 4.1544 × 10−7 4.81 2.3622 × 10−8 5.75
256 1.3878 × 10−8 4.90 4.0083 × 10−10 5.88
Remark. This effect of the boundary condition on the global accuracy of the Padé method
is not exclusive of the finite volume formulation. In the finite difference method the point
closest to the boundary has a similar equation. Add to this the unavoidable approximation
of the evolution equation at the boundary and the finite difference method would be at least
as sensitive to the boundary approximation as is the finite volume method.
Remark. At the inlet, the finite difference and the finite volume methods require different
approaches too. The finite difference method needs a downwind extrapolation from the
interior points to the boundary, and as is well known this is highly unstable. In contrast, for
the finite volume method no approximation is needed since it can use the exact prescribed
value at the inlet, both in the discretized equation and in the Padé method. Thus, for example,
a 4th-order Padé method does not require any special treatment for the inlet. Sixth and higher
order methods require some modifications at points close to the inlet, which depends on the
specific method (see Section 6 for an example with the 6th-order method.)
Remark. In a pure diffusion problem and using the standard Padé method, the derivative
at the boundary does not involve the prescribed value of the variable. This problem can be
overcome by considering the value of the function available at the boundary in the inter-
polation problem. For instance, at the right boundary, and for a 4th-order accurate boundary
condition, we have
µ ¶
1 5 89 127 4
φ N0 + 6φ N0 −1 = φ B + φ̄ N −1/2 − φ̄ N −3/2 + φ̄ N −5/2 ,
h 3 18 18 9
where φ B is the prescribed value of the variable at the boundary and N is the number of
control volumes in the grid.
Remark. Poinsot and Lele [24] proposed characteristic boundary conditions for Euler
and Navier–Stokes equations. They advanced the solution in time on the boundaries by us-
ing a characteristic system. In this system they used a 3rd-order one-sided space derivative.
This methodology can be used in the finite volume framework as well. Moreover, using the
Hermite polynomial, which interpolates the sliding average, and whose derivative interpo-
lates the interface values, we can perform a complete characteristic evolution. Indeed, this
polynomial may be used as a reconstruction polynomial and we can directly apply either
a local Cauchy–Kowalewski procedure or a characteristic method (see [11]). For example,
using the characteristic method for the linear advection problem with the 4th-order Padé
method and a Simpson’s rule yields

Z t+1t
1t 0
φi dt ∼
= (Hi (xi , t) + 4Hi0 (xi − c1t/2, t) + Hi0 (xi − c1t, t))
t 6
for all i = 1, . . . , N , where H is the Hermite polynomial
¡ n n ¢
Hi (x) = φ̄ i−1/2 + φ̄ i+1/2 hηi (x) + φi+1
n
η̃i+1 (x) + φi−1
n
η̃i−1 (x)
for all i = 1, . . . , N − 1, x ∈ R, with

0
ηi (x) = (1 − 2li+1 (xi+1 )(x − xi+1 ))(li+1 (x))2 , i = 1, . . . , N − 1, x ∈ R
η̃i±1 = (x − xi±1 )(li±1 (x)) ,
2
i = 1, . . . , N − 1, x ∈ R
and
x − xi∓1
li±1 (x) = , i = 1, . . . , N − 1, x ∈ R.
±2h
At the boundary we use the Hermite polynomial HN −1 , that is, HN = HN −1 .
This procedure requires less storage than the RK one and is faster for Padé methods since
it needs only one evaluation of the interface values per time step. However, the Hermite
polynomial approximation is, except for the center point used in the Padé method, one order
of accuracy lower than that for the sliding average. So, the accuracy of the resulting method
is one order of accuracy lower than that using the RK method.
6. STABILITY ANALYSIS
We now turn to the analysis of the stability of the proposed methods for both explicit and
implicit temporal discretization. There are many definitions of stability in the literature of
CFD (see, for example, [4, 10]). In the present work, we consider the notions of Lyapunov
and asymptotic stability.
Recall (see, for example, [1]) that a stationary solution of an autonomous dynamical
system is said to be Lyapunov stable if all solutions of the equation, with initial conditions
in a sufficiently small neighborhood of the equilibrium point, are defined for all positive time
and converge uniformly with respect to time to the stationary solution as the initial conditions
tend to the equilibrium point. A stationary solution is said to be asymptotically stable if it
is Lyapunov stable and if, in addition, all solutions with initial conditions sufficiently close
to the equilibrium point under consideration tend to this equilibrium point as t → +∞.
We start with the stability of the linear hyperbolic equation with periodic initial conditions,
discretized with a Padé finite volume method. The semidiscrete equation can be written as
d φ̄ σc
= C φ̄, (37)
dt 1t
where σc = c1t
h
is CFL number, and
 
φ̄ 1/2
 . 
[φ̄] = 
 . .
. 
φ̄ N −1/2
The operator C ∈ End(R N ) is given locally as
(Cφ)i+1/2 = φi − φi+1 , i = 0, . . . , N − 1,
where φ is obtained with the Padé discretization method.

Consider the implicit Euler method. In this case the fully discrete equation can be written
as
n+1 n n+1
φ̄ = φ̄ + σc C φ̄ .
It is easy to see that any centered scheme is unconditionally stable. Indeed, given the fact
that C is skew-adjoint, all of its eigenvalues are pure imaginary (one or two of them are
zero; see Eq. (39) below):
° ° 1
°(id − σc C)−1 ° = p
2
λmin (id − (σc C)2 )
= 1.
This proves the assertion.

Now, we turn to the analysis of the 4th RK algorithm for the solution of the semidiscrete
problem. Given the symbolic polynomial
1 l
RKl (z) = 1 + z + · · · + z, l ∈ N,
l!
the associate discrete dynamical system is given by
n+1 n
φ̄ = RK4 (σc C)φ̄ ,
where, in the present linear case, RK l (C) ∈ End(R N ), l ∈ N is the Runge–Kutta operator,
which corresponds to the l-truncated series of the exponential of σc C, that is,
1
RKl (σc C) = id + σc C + · · · + (σc C)l , l ∈ N. (38)
l!
The operator C is skew-adjoint, so the origin is a stable equilibrium point of the semidis-
crete. Indeed, this follows from the fact that
2 µ ¶
d φ̄ C φ̄
= 2σc φ̄ · = 0,
dt 1t
and so, the solution will stay at the sphere containing the initial condition.
This means that any instability in the discrete formulation comes from the time discretiza-
tion. And, in general, increasing the order of the time discretization improves the stability
of the method. This contrasts with the spatial discretization, which loses stability with the
increase in accuracy. However, this is a consequence of the fact that the latter converges to
an unbounded operator.
Analogously to the semidiscrete equation, we compute
n+1 2 n σc6 3 n 2 σ8 n
(φ̄ ) = (φ̄ )2 − (C φ̄ ) + c (C 4 φ̄ )2 .
72 576
Note that any centered scheme would lead to this equation, and so, for a sufficiently small
CFL number any centered scheme can be made stable for the 4th-order RK.
To determine how small the CFL number must be, we solve
1 3 2 σ2
− (C v) + c (C 4 v)2 ≤ 0, v ∈ RN .
72 576
Then a necessary and sufficient condition for stability is
√
2 2
σc ≤ σc,max = .
kCk2
Given A ∈ End(R N ), let ρ(A) denote its spectral radius. By definition
kCk2 = sup kCvk2 = ρ(C),

kvk2 =1
where we have used the fact that C is skew-adjoint. Taking into account that
{eik(n+1/2)2π/N }n=0,...,N −1 , for k = 0, . . . , N − 1 are eigenvectors of C with eigenvalues
µ ¶ Pn ¡¡ ¢ ¢
kπ j=1 b j cos j − 12 kπ/N
λk = 4i sin P , k = 0, . . . , N − 1, (39)
N 1 + 2 mj=1 a j cos( jkπ/N )
we obtain
¯ µ ¶ Pn ¡¡ ¢ ¢¯
¯ j − 12 kh ¯¯
¯ kh j=1 b j cos
kCk2 = sup ¯4i sin P ¯. (40)
|kh|≤π ¯ 2 1 + 2 mj=1 a j cos( jkh) ¯
Table XIII shows the CFL limit of stability (independent of the grid) for the 4th-order
Lagrange method and for the 4th-, 6th-, 8th-, and 12th-order Padé methods. As expected,
increasing the accuracy of the spatial discretization decreases the limit in the CFL for the
stability.
Equation (40) is the L ∞ -norm of the spectral function of the operator C. So, the stability
limit coincides with that obtained by the von Neumann method. Note that it is not an
immediate consequence of the Lyapunov theorem (see, for example, [1]), for k = 0 implies
|λ| = 1.
A simple calculation shows that a fully discrete algorithm is von Neumann stable if
Sσ = σc F(C)([0, 1/2]) ⊂ S, where S is the stability region (see [29]). For RK methods we
have Sk = RK−1 k [B1 ], k ∈ N, where B1 = {z ∈ C : |z| ≤ 1}. For completeness, the stability
zone for the 1st-, 2nd-, 3rd-, 4th-, 6th-, and 8th-order RK methods is shown in Fig. 11.
It is interesting to note that any centered method is unstable for the 1st- and 2nd-order RK
methods. This follows at once from the fact that the region of stability of these methods is
TABLE XIII
CFL Stability Limit for the 4th-Order RK Method
with Some Spatial Padé Discretization Schemes
4th Lagrange 4th Padé 6th Padé 8th Padé 12th Padé
√
σc,max 2.0612 2 2/3 1.4217 1.2829 1.1640
FIG. 11. Stability region for the RK methods of order 1, 2, 3, 4, 6, and 8.
tangent to the imaginary axis and that for any centered method <(F(C)) = 0. Also interesting
is the fact that for centered methods the 4th-order RK method provides an improved stability
as compared to the 6th-order method. In other words, from the stability viewpoint, and for
centered methods, it does not pay off to move from the 4th- to the 6th-order RK.
In Fig. 12 we plot Sσ for the 1st-order upwind scheme, the 3rd-order QUICK scheme,
√
the 4th-order Lagrange scheme, and the 4th-order Padé scheme, all for σc = 2 2/3. At this
CFL the QUICK scheme is stable for RK4 (its stability limit for the 3rd RK is, within three
√
digits of accuracy, 1.625, which is slightly smaller than 2 2/3), the Lagrange method is
also stable for RK4 , and, interestingly enough, with this CFL number the 1st-order upwind
scheme is not stable even for a 4th-order RK method. The dissipation generated by this
scheme, which helps to stabilize it, for instance, in the Euler explicit method, is excessive
for RK4 . Also, the Padé method is more restrictive in CFL terms than is the Lagrange
method. However, this gain in stability is more than compensated by the improved spectral
resolution of the Padé method.
FIG. 12. Sσ for (a) 4th-order Padé, (b) 4th-order Lagrange, (c) 1st-order upwind, and (d) 3rd-order QUICK.
TABLE XIV
Diffusion CFL Stability Limit for the 4th-
Order RK Method with Some Spatial CD
Padé Discretization Schemes
σc 6th 8th 12th
0.5 0.2901 0.2849 0.2823

1 0.2407 0.1912 0.1422
We proceed with the stability limits of the discrete transport equations for the 4th RK with
the CD Padé 6th order with periodic boundary conditions. Now, the semidiscrete equation
is
d φ̄ 1
= (σc C φ̄ + σd Dφ̄), (41)
dt 1t
where σd = ν1t/ h 2 is the CFL number for diffusion, and
0
(Dφ)i+1/2 = φi+1 − φi0 , i = 0, . . . , N − 1
with φ 0 obtained with the CD Padé discretization method.

In Table XIV are listed the stability limits of the diffusion CFL number for the 4th-order
RK method with CD 6th-, 8th-, and 12th-order CD Padé methods and for two values of
the convection CFL number: σc = 0.5 and σc = 1. Again, as expected, increasing the order
of the method decreases the maximum value of the diffusion CFL number, while for the
same method a smaller σc improves the stability limit. Figure 13 shows Sσ for some limit
combinations of σc and σd . Note that simultaneously using the limit values of σc = 1.3304
for pure convection with σd = 0.2901 for pure diffusion in a transport equation yields an
FIG. 13. Stability region for σc = 1.3304, σd = 0 (thick black); σc = 1, σd = 0.2407 (thick dark gray);
σc = 0.5, σd = 0.2901 (thick gray); and σc = 0, σd = 0.2901 (horizontal thick light gray).
unstable method. In other words, the stability analysis for the transport equations cannot be
undertaken separately.
We now proceed to the study of the effect of the boundary condition on stability. In [10,
14, 28] is introduced the notion of GKS stability and an analytical method to establish the
GKS stability of discrete methods with boundary approximation taken into account. In [4]
this technique is applied to some compact finite difference methods, together with several
boundary approximations. In [5] the authors concluded that in practical computations only
those schemes for which the semidiscrete equation is Lyapunov stable are of any usefulness
for long-time integrations. Thus, we use the spectrum of the semidiscrete equations to study
the effect of the boundary closure.
Let us start with the hyperbolic equation. First, we consider the effect of the order of the
accuracy of the boundary closure on stability. The semidiscrete equation is formally the same
as Eq. (37) with C modified in order to take the boundary closure into account. We remark
that for the stability analysis it is sufficient to consider homogeneous boundary conditions.
In fact, we can think of stability analysis as the study of the evolution of perturbations of
the solution with a fixed boundary condition.
Figure 14 shows the spectrum that results from 4th-order Padé finite method closed at the
outlet boundary with the 3rd- and the 4th-order schemes presented in the previous section.
The grids comprise 32, 64, and 128 control volumes. The spectra are very similar, with
the 3rd-order approximation closer to the imaginary axis. This finding for the Padé finite
volume method is in clear contrast to the findings for the Padé finite difference method (see
[4]). The latter is stable for the 3rd-order closure, but is unstable for the 4th. This is due to
the fact that the finite difference approach requires a downwind extrapolation at the inlet
and, as is well known, a downwind extrapolation is unstable. On the other hand, the finite
volume method does not need any approximation at the inlet and thus avoids the potentially
unstable mode.
As mentioned earlier a 6th-order or higher Padé method requires some approximation at
points close to the boundaries. We consider the 6th-order Padé method with the following
FIG. 14. Spectrum of the semidiscrete 4th-order Padé method in pure advection for (a) 3rd- and (b) 4th-order
boundary conditions at outlet: (■) 32, (▲) 64, and (◆) 128 control volumes.
FIG. 15. Spectrum of the semidiscrete 6th-order Padé method in pure advection for (a) 5th- and (b) 6th-order
boundary conditions at outlet: (■) 32, (▲) 64, and (◆) 128 control volumes.
6th-order approximation close to the boundaries:
1. At inlet
1 3 43 41 5 1
φ0 + φ1 + φ3 = φ̄ 1/2 + φ̄ 3/2 + φ̄ 5/2 − φ̄ 7/2 .
8 4 96 32 32 96
2. At outlet
197 37 13 11 1
5φ N −1 + φ N = φ̄ N −1/2 + φ̄ N −3/2 − φ̄ N −5/2 + φ̄ N −7/2 − φ̄ N −9/2 .
60 10 10 30 20
The Padé method for the point N − 1 is computed analogously to that for the point 1 at
inlet.
Figure 15 shows the spectrum of the 6th-order Padé method with the boundary closure
mentioned above. This figure corroborates the previous findings, namely, that the Padé finite
volume method is stable for a boundary closure of the same order of accuracy as the inner
points.
Because the effect of the boundary is to create some damping in the high-frequency
modes, the limit value computed with the von Neumann analysis is still valid, and both
approximations are stable with this limit. To verify this assertion, in Fig. 16 is shown the
spectrum that results from the the explicit 4th-order RK and 4th-order Padé finite method,
which is closed at the outlet boundary with the 4th-order scheme. The grids comprise 32
√
and 128 control volumes for σc = 2 2/3. This figure confirms the fact that the CFL limit
obtained with the von Neumann analysis can be used for the advection problem with a
prescribed inlet boundary condition.
√
FIG. 16. Spectrum of the discrete 4th RK, 4th-order Padé method in pure advection for σc = 2 2/3: (a) 32
and (b) 128 control volumes.
FIG. 17. Spectrum of the semidiscrete 6th-order Padé method in pure advection for the 6th-order boundary
closure at outlet: (■) 32, (▲) 64, and (◆) 128 control volumes.
Finally, we study the effect of the boundary closure on the CD Padé methods for a general
transport equation. We start with the critical limit of pure advection.
When both boundary conditions are of Dirichlet type the operator C is skew-adjoint,
whence the eigenvalues are pure imaginary. Thus, we only need to consider the case with
a combined Dirichlet and a von Neumann boundary condition. Figure 17 shows the result-
ing spectrum for the CD 6th-order Padé semidiscrete method with the 6th-order closure.
The boundary condition corresponds to a Dirichlet boundary condition at inlet and a von
Neumann at outlet.
Again, the finite volume method does not require any lowering of the order at the bound-
ary. In this case the finite volume method represents a major gain in stability. Indeed, in
[21] it is shown that the CD Padé finite difference method is unstable even for a boundary
closure of 4th order of accuracy.
An interesting feature of the general transport equation is the effect of the boundary
closure for the diffusion flux. Figure 18 shows the spectrum that results for the discrete
4th-order RK method with the 6th-order CD Padé method for the transport equations with
Dirichlet BC at inlet and von Neumann BC at outlet. The spectrum has been calculated
with the same values of CFL as in Fig. 13. For small CFL of diffusion the limit determined
by the von Neumann analysis is valid. However, as diffusion becomes more important, the
spectrum starts to display a reduction in the diffusion limit. For example, for σc = 1 the limit
for the 6th-order closure is σd = 0.2125, which is smaller than the limit for periodic bound-
ary conditions of σd = 0.2407. Contrary to the pure convection case, where the spectrum
converges towards the spectral function, in the transport equation the effect of the boundary
closure on diffusion is permanent (see Fig.18b). A possible explanation for this different
behavior of the spectrum for pure advection and diffusion may lie in the fact that the former
has a hyperbolic character while the latter has an elliptic character. Thus, the effect of the
boundary is swept away as the grid is refined in the hyperbolic equation, while it is not
in the elliptic one. Notwithstanding the reduction in the stability limit of σd , it should be
stressed that it is necessary neither to reduce the order of accuracy at the boundary to ensure
the stability of the method nor to enlarge the stencil close to the boundary, to accommodate
stability.
FIG. 18. Spectrum of the discrete 4th-order RK method with the 6th-order CD Padé method for the transport
equations with Dirichlet BC at inlet and von Neumann BC at outlet for (a) 64 and (b) 128 control volumes:
(◆) σc = 1.3304, σd = 0; (★) σc = 1, σd = 0.2407; (■) σc = 0.5, σd = 0.2901; and (▲) σc = 0, σd = 0.2901.
Remark. Equation (39) shows en passant that for an odd number of control volumes
dim ker(C) = 1, that is, only constant functions have a zero convective flux. In contrast, for
an even number of control volumes dim ker(C) = 2. Thus, in addition to the constant mode,
we also have the so called “checkerboard” mode or the “odd–even” decoupling.
Using two first derivative operators to represent the Laplacian operator on a nonstaggered
grid may lead to the checkerboard mode. A remedy for this, put forward in [30], is to use
an approximation for the second derivative. However, as the previous observation shows,
an alternative solution may be the use of an odd number of control volumes.
Remark. The limits of stability determined above also works for time-periodic solutions.
Indeed, it is enough to observe that a time-periodic solution is a fixed point of the operator
(RKk )l , where l ∈ N is the discrete period, and that ρ((RKk )l ) = (ρ((RKk ))l .
Remark. For nonlinear problems the previous stability limits also hold true for the
stability of fixed points or periodic orbits. Indeed, consider a general conservation law
∂φ ∂ f (φ)
+ = 0,
∂t ∂x
where f is some smooth function. Then, the linearized equation, for example, in the fixed
point φ0 can be written as
· ¸
dδ φ̄ f 0 ((φ0 )i+1 )δφi+1 − f 0 ((φ0 )i )δφi
+ = 0.
dt i+1/2 h
Hence, for σc < σc,max the result follows at once from the Lyapunov theorem. As a fi-
nal comment, we remark that in the nonlinear case, the Lyapunov theorem guarantees the
convergence to the critical point only in a neighborhood of it, while in the linear case the con-
vergence is for any point. In other words, the stable manifold is the whole space in the linear
case, but in general, it is not so for the nonlinear case.
Remark. The above analysis considered constant coefficients in a uniform grid. The
following argument suggests that for a sufficiently fine grid, the results above hold true, if
applied locally.
We start with the nonconstant case. The conservation law in intrinsic form is
∂φ
+ div(cφ) = 0.
∂t
Let c x ∈ C ∞ ∩ L ∞ be a smooth limited function. Then, in a Cartesian grid it is
∂φ ∂c x φ
+ =0
∂t ∂x
and in a general coordinate ξ ,

√
∂φ 1 ∂ gcξ φ
+√ = 0, (42)
∂t g ∂ξ
where g is the determinant of the metric in the coordinate ξ . Suppose c x ∈ C ∞ does not
change sign. Then, in the coordinate ξ given by
∂ξ 1
=
∂x c
Eq. (42) simplifies to
∂ φ̃ ∂ φ̃
+ = 0,
∂t ∂ξ
√
where φ̃ = gφ. The stability criterion for the previous equation is
1t
≥ σc,max .
hξ
Using h ξ = h x /c + o(h x ) we find, for a sufficiently small grid, the following stability
criterion in the “physical” grid:
1tkck∞
≤ σc,max .
hx
The case when c = 0 is trivial. Consider it is not zero for some points and consider a
δ-neighborhood of the complement of the support of cU . For sufficiently small δ the max-
imum of the speed occurs outside U . We can apply the previous result for each connected
component of the complement of U . And since it is uniform with respect to δ we obtain the
result by passing to the limit as δ → 0.
FIG. 19. Spectrum of the discrete 4th-order RK method with the 4th-order Padé method for pure advection
and periodic boundary conditions for varying local CFL: (left) σc (x) = σc,max sin(2π x), x ∈ [0, 1] and (right)
σc (x) = σc,max (cos(32π x) + cos(30π x)), x ∈ [0, 1].
To verify this remark we compute the spectrum of the 4th-order Padé method with a
4th RK method for time evolution for a pure convection problem with periodic boundary
conditions and varying CFL numbers. Figure 19 shows the resulting spectra for two CFL pro-
files: σc (x) = σc,max sin(2π x), x ∈ [0, 1], and σc (x) = σc,max (cos(32π x) + cos(30π x)), x ∈
[0, 1]. The first is a slowly varying function, while the second is a highly oscillatory profile
with a “beating.” Figure 19 shows that the spectrum converges toward Sσ and also that this
convergence depends on the resolution of the monochromatic components of the profile.
In the case of general transport equations the results are valid if σd is sufficiently small.
This is not so rare, and in fact, it is true for most convection-dominated problems. For
example, for a Reynolds number of the order of 106 and a grid with about 100 control
volumes we have σd /σc ' 10−4 .
The case of a nonuniform grid can be handled analogously by noting that, from Eq. (42),
the nonuniform grid can be seen as a particular case of the nonconstant speed.
7. SUMMARY
A class of Padé finite volume methods for the evaluation of derivatives and interpolation
have been presented and analyzed. From the analysis, the following conclusions can be
drawn:
1. The use of the sliding averages in the finite volume formulation requires a smaller
stencil in multi-dimensional problems, as compared with point values. Moreover, in time-
dependent problems the former appears explicit, while the latter requires interpolation.
2. For pure convection, or pure diffusion problems, the standard (uncoupled) Padé
method has the highest order of accuracy in the given stencil and requires less CPU time
than the CD Padé interpolation. For general transport equations, the latter presents an
improved spectral resolution and no significant additional cost.
3. Padé methods convey energy better than usual interpolation.

4. The spectral Padé method shows spectral-like resolution and exponential conver-
gence.
5. Using the notion of reflection ratio it is shown that the order of the boundary closure,
if lower than or equal to the order of accuracy of the inner points, determines the uniform
order of accuracy of the Padé method.
6. The limits of stability computed by the von Neumman procedure are valid for the
pure advection problem or for convection-dominated problems. Moreover, contrary to the
finite difference method, the finite volume method requires neither the lowering of the order
of accuracy at the boundary closure nor any change in the stencil of the point at the boundary
to accommodate stability.
ACKNOWLEDGMENTS
The author is grateful to Professor. J. C. F. Pereira for bringing the problem of the Padé finite volume interpolation
to his attention, to Eng◦ J. M. C. Pereira for many useful discussions, and to the referees for raising a number of
issues left unaddressed in the first version of the paper.
REFERENCES
1. D. V. Anosov, S. Kh. Aranson, V. I. Arnold, I. U. Bronshtein, V. Z. Grines, and Yu. S. Il’yashenko, Ordinary
Differential Equations and Smooth Dynamical Systems (Springer-Verlag, Berlin, 1997).
2. C. Basdevant, M. Deville, P. Haldenwang, J. M. Lacroix, J. Ouazzani, R. Peyret, P. Orlandi, and A. T. Patera,
Spectral and finite-difference solutions of the burgers equation, Comput. Fluids 14(1), 23 (1986).
3. E. O. Brigham, The Fast Fourier Transform (Prentice Hall, Englewood Cliffs, NJ, 1974).
4. M. H. Carpenter, D. Gottlieb, and S. Abarbanel, The stability of numerical boundary treatments for compact
high-order finite-difference schemes, J. Comput. Phys. 108, 272 (1993).
5. M. H. Carpenter, D. Gottlieb, and S. Abarbanel, Time-stable boundary conditions for finite-difference schemes
solving hyperbolic systems: Schemes, J. Comput. Phys. 111, 220 (1994).
6. J. Sebastião e Silva, Sur l’ espace de fonctions holomorphes a croissance lente a droite, Portugal. Math. 17,
435 (1958).
7. J. H. Ferziger and M. Perić, Computational Methods for Fluid Dynamics (Springer-Verlag, Berlin, 1997).
8. D. Gaitonde and J. S. Shang, Optimized compact-difference-based finite-volume schemes for linear wave
phenomena, J. Comput. Phys. 138, 617 (1997).
9. D. Gottlieb and S. A. Orszag, Numerical Analysis of Spectral Methods (SIAM, Philadelphia, 1977).
10. B. Gustafsson, H.-O. Kreiss, and A. Sundström, Stability theory of difference approximations for mixed initial
boundary value problems, II, Math. Comput. 26(119), 649 (1972).
11. A. Harten, B. Engquist, S. Osher, and S. R. Chakravarthy, Uniformly high order accurate essentially non-
oscillatory schemes, III, J. Comput. Phys. 71, 231 (1987).
12. C. Hirsch, Numerical Computation of Internal and External Flows (Wiley, West Sussex, 1994).
13. F. John, Partial Differential Equations, Applied Mathematical Sciences, Vol. 1, 4th ed. (Springer-Verlag, New
York, 1982).
14. H.-O. Kreiss, Stability theory of difference approximations for mixed initial boundary value problems, I,
Math. Comput. 22, 703 (1968).
15. H. C. Ku, R. S. Hirsh, and T. D. Taylor, A pseudospectral method for solution of the three-dimensional
incompressible Navier–Stokes equations, J. Comput. Phys. 70, 439 (1987).
16. L. Landau and E. Lifchitz, Mécanique des Fluides, Physique Théorique, Vol. 6, 2nd ed. (Mir, Moscow, 1989).
17. S. Lang, Real and Functional Analysis, Graduate Texts in Mathematics, Vol. 142, 3rd ed. (Springer-Verlag,
New York, 1993).
18. S. K. Lele, Compact finite-difference schemes with spectral-like resolution, J. Comput. Phys. 103, 16 (1992).
19. B. P. Leonard, A stable and accurate convective modelling procedure based on quadratic upstream interpola-
tion, Comput. Methods Appl. Mech. Eng. 19, 59 (1979).
20. Z. Lilek and M. Perić, A fourth-order finite-volume method with collocated variable arrangement, Comput.
Fluids 24(3), 239 (1995).
21. K. Mahesh, A family of high order finite difference schemes with good spectral resolution, J. Comput. Phys.
145, 332 (1998).
22. B. Mattiussi, An analysis of finite-volume, finite-element, and finite-difference methods using some concepts
from algebraic topology, J. Comput. Phys. 133(2), 289 (1997).
23. S. A. Orzag, Spectral methods for problems in complex geometries, J. Comput. Phys. 37, 70 (1980).
24. T. J. Poinsot and S. K. Lele, Boundary conditions for direct simulations of compressible viscous flows,
J. Comput. Phys. 101, 104 (1992).
25. L. Schwartz, Théorie des Distribuitions (Hermann, Paris, 1966).
26. F. Spotz, High Order Compact Finite Difference Schemes for Computational Mechanics (Ph.D. thesis,
University of Texas at Austin, 1995).
27. J. Stoer and R. Bulirsch, Introduction to Numerical Analysis (Springer-Verlag, New York, 1980).
28. J. C. Strikwerda, Initial boundary value problems for the method of lines, J. Comput. Phys. 34, 94 (1980).
29. R. Vichnevetsky and J. B. Bowles, Fourier Analysis of Numerical Approximations of Hyperbolic Equations,
Studies in Applied Mathematics (SIAM, Philadelphia, 1982).
30. R. V. Wilson, A. O. Demuren, and M. H. Carpenter, Higher-order compact schemes for numerical simulation
of incompressible flows, ICASE Report 98-13, ICASE, Langley Reseach Center, 1998.
31. I. Yavneh, Analysis of a fourth-order compact scheme for convection-diffusion, J. Comput. Phys. 133, 361
(1997).

On A Class of Pad e Finite Volume Methods

Uploaded by

Copyright:

Available Formats

On A Class of Pad e Finite Volume Methods

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

On A Class of Pad e Finite Volume Methods

Uploaded by

Copyright:

Available Formats

Journal of Computational Physics 156, 137–180 (1999)

Article ID jcph.1999.6376, available online at http://www.idealibrary.com on

On a Class of Padé Finite Volume Methods

Received February 2, 1999; revised September 7, 1999

A class of Padé finite volume methods providing an improved spectral reso-

2. THE STANDARD PADÉ INTERPOLATION

2.1. Hyperbolic Equations

Given a smooth function φ ∈ C ∞ , and natural numbers m, n ∈ N, find coefficients ai , i =

(τs f ) (x) = f (x − s), x ∈R

Order a1 a2 a3 b1 b2 b3 Truncation error

Order b1 b2 b3 b4 b5 b6 Truncation error

where the operator F: S 0 → S 0 is the Fourier transform2 on S 0 , that is,

for all ψ ∈ S. We call 8 the spectrum of φ.

Using Eq. (5) results in

Solution of the previous equation provides the law of propagation

8̄(t) = 8̄0 eick̂t , t ∈ R, (9)

Method Order ARSE1/2

Padé 4 5.788 × 10−2

Then applying the Fourier transform to this equation yields

FIG. 4. Plot of distribution: t = 0 (top), t = 5h

φ(t) = e−t φ0 (x̂e−t ), t ≥ 0.

The Padé finite volume method yields

and a change over the characteristics of

3 cos(2k̃π ) sin2 (k̃π )

for all l = 1, . . . , 2r − 1. This is a 2r × r system of linear equations. However, due to the

where the block matrices are defined as

Consider the real numbers xi , φi(k) , k = 0, . . . , n i − 1, i = 0, . . . , m 0 with

x0 < x1 < · · · < xm 0 .

P (k) (xi ) = φi(k) , k = 0, . . . , n i−1 , i = 0, . . . , m 0 .

L i,ni −1 (x) = li,ni −1 (x), i = 0, . . . , m 0

and recursively for k = n i − 2, . . . , 0,

Consider the points x±i = ±i, i = 1, . . . , n.

where 2 ∈ C is the primitive of φ,

Using the result of the claim we compute ai , i = 1, . . . , m, and b̃i , i = 1, . . . , n. Then

2.2. Parabolic Diffusion Equations

The corresponding IP can be stated as follows:

Order a1 a2 a3 b1 b2 b3 Truncation error

Given a smooth function φ ∈ C ∞ , and natural numbers m, n ∈ N, find coefficients ai , i =

2.3. Transport Equations

For the diffusion term we have:

Order a1 a2 b1 b2 c1 c2 Truncation error

Order d1 d2 e1 e2 f1 f2 Truncation error

Order a1 a2 b1 b2 c1 c2 Truncation error

Order d1 d2 e1 e2 f0 f1 f2 Truncation error

Remark. The notion of group velocity applies to multi-dimensional equations as well.

4. OPTIMIZED PHASE ERROR AND SPECTRAL PADÉ INTERPOLATION

In addtion to maximal order schemes, classes of parametrized nonmaximal order Padé

Now, given a ψ ∈ S it is easy to see that the function

Spectral methods are based on representing the solution of a problem as a truncated

Boundary conditions represent a particular case of the general IH problem. However,

Then, we define the time-Fourier transform of φ as

Using the 4th-order Padé method and rearranging terms we obtain

whence, solving for the reflection ratio, we find

Analogously, we find for the 4th-order Padé method at the boundary

3rd order 4th order

8 1.3831 × 10−1 — 1.0871 × 10−1 —

where a, b, K ∈ R are constants. In all cases, a = 14 , b = 34 , K = e1/16 , and the temporal

The exact solution is

5th order 6th order