Understanding of Numerical Oscillations

in terms of Modified Equation

Jiequan Li ()

School of Mathematical Sciences, Beijing Normal University

Email:[email protected]
Homepage: http://math.bnu.edu.cn/∼lijiequan/

2011 Int’l Workshop on Numerical Methods and Modeling

for Compressible Multimaterial Flows and Mixing
IAPCM, Sept. 12–15, 2011
Based on:
1. Local Oscillations in Finite Difference Solutions of Hyperbolic
Conservation Laws
Jiequan Li , Huazhong Tang, Gerald Warnecke and Lumei
Zhang, Math. Comp, 78, 1997-2018, 2009.

2. Heuristic Modified Equation Analysis on Oscillations in

Numerical Solutions of Conservation Laws
Jiequan Li and Zhicheng Yang, SIAM J. Numer. Anal., in
press, 2011.

3. The von Neumann Analysis and Modified Equation Approach

for Finite Difference Schemes
Jiequan Li and Zhicheng Yang, preprint, 2011.
Ubiquity of Numerical Oscillations
Gibbs Phenomenon. The n–th partial sum of the Fourier series
has large oscillations near the jump, which might increase the
maximum of the partial sum above that of the function itself. The
overshoot does not die out as the frequency increases, but
approaches a finite limit1 .

(a) Functional approximation (b) Functional approximation

of square wave using 5 har- of square wave using 25 har-
monics monics

1
H. S. Carslaw (1930): Introduction to the theory of Fourier’s series and
integrals.
Lax–Wendroff Method
∂u ∂u
+c = 0,
∂t ∂x
ν n ν2 n
ujn+1 = ujn − [uj+1 n
− uj−1 ] + [uj+1 − 2ujn + uj−1
n
], ν = cτ /h.
2 2
In this lecture, τ is the time step size and h is the grid spacing.

The Lax–Wendroff method has accuracy of second order both in

space and time.
P. Lax further showed that2

when a discontinuous initial value problem for a scalar hyperbolic

equation in one space variable is approximated by a difference
scheme that is more than first order accurate; it leads to
overshoots analogous to the Gibbs phenomenon when
discontinuous functions are approximated by sections of Fourier
series.

Lax’s clarification is very intuitive and algebraic.

It is seemingly impressed that oscillations are present near jumps

only when high order accurate schemes are used to approximate
discontinuous solutions.

2
P. D. Lax, Gibbs phenomena, J. Sci. Comput., Vol. 28, Nos. 2/3 (2006),
445–449.
Local Oscillations in Lax–Friedrichs Solutions34
Numerical results for the advection equation ut + ux = 0, 50 grid
points are used, ν = 0.8

Time=1.0, CFL=0.8, q=1.0 Time=1.0, CFL=0.8, q=1.0

0.18 0.18

0.16 0.16

0.14 0.14

0.12 0.12

0.1 0.1
u

u
0.08 0.08

0.06 0.06

0.04 0.04

0.02 0.02

0 0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
x−axis x−axis

(c) uj0
= 1 for j = 25 and uj0 ≡ (d) uj0
= 1 for j = 25, 26, and
0 otherwise uj0 ≡ 0 otherwise

3
H.-Z. Tang and G. Warnecke, A note on (2K + 1)-point conservative
monotone schemes, M2AN Math. Model. Numer. Anal., 38(2004), 345–357.
4
M. Breuss, The correct use of the Lax-Friedrichs method, M2AN Math.
Model. Numer. Anal. 38(2004), 519–540.
L-F Solutions of Compressible Euler Equations5
ρt +(ρv )x = 0, (ρv )t +(ρv 2 +p)x = 0, (ρE )t +(v (ρE +p))x = 0.
T = 0.25, 100 grid points, CFL = 0.6.
Initial data: (ρ0j , uj0 , pj0 ) = (0.125, 0, 0.1) for j = 49, 50, 51, and
(ρ0j , uj0 , pj0 ) = (1, 0, 1) otherwise.
DENSITY PRESSURE
1 1

0.98 0.98

0.96
0.96
rho

0.94

p
0.94
0.92
0.92 0.9

0.9 0.88
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
X−AXIS X−AXIS

VELOCITY INTERNAL ENERGY

0.15 2.65

0.1
2.6
0.05

0 2.55
e
v

−0.05 2.5
−0.1
2.45
−0.15

−0.2 2.4
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
X−AXIS X−AXIS

5
J. L. Li, H. T. Tang, G. Warnecke and L. M. Zhang, local oscillations in
finite difference solutions of hyperbolic conservation laws, Math. Comp.,
(2009), Vol. 78, pp. 1997–2018.
Remarks
1. The Lax–Friedrichs scheme is first order accurate, has TVD
property and the maximum numerical viscosity in the class of
TVD schemes. The large dissipation SHOULD suppresses
local oscillations!!! This oscillatory phenomenon is very
counter–intuitive and misleading.6

2. Oscillations in solutions by higher order schemes usually

violate the maximum principle, and they are substantially
different from those by TVD schemes.

6
Nevertheless, the oscillations are just local and do not contradict to
the TVD property because the latter is a global definition while the
oscillations are of local nature. The preservation of the TVD property is
due to the compensation by strong decrease in solution amplitude.
Purposes

Understanding of numerical oscillations in solutions by finite

difference schemes, finite volume, spectral · · ·

Tools: Modified Equation Analysis, plus Discrete Fourier Analysis

Heuristic Observation
Discrete Fourier Decomposition of Initial Data
The initial datum is a square signal

1, 0 < x1 < x2 < 1,
u0 (x) =
0, Otherwise,
and the interval [x1 , x2 ] are partitioned with M grid points.
• Discretization with an odd number of grid points (j1 and p are
parameters related with the grid point indices, ξ = 2kh)
X (−1)k e i ξ(j+j1 ) (1 − e −i ξp )
uj0 = (−1)j+j1 +M/2 h + ph + . (C)
M(1 − e −i ξ )
k6=0,M/2

• Discretization with an even number of grid points

X (−1)k e i ξ(j+j1 −1) [1 − e −i ξ(p−1) ]
uj0 = 0 × (−1)j + (p − 1)h + .
M(1 − e −i ξ )
k6=0,M/2

Highest frequency modes (−1)j are included in discrete initial data

(C) if M is odd.
Chequerboard modes from improper numerical boundary
treatment
Example The leap-frog scheme for vt + vx = 0, x ∈ (0, 1),

ujn+1 = ujn−1 − ν(uj+1

n n
− uj−1 ).

We need numerical boundary conditions both at x = 0 and x = 1.

At x = 1 (j = M), if the (improper) boundary formula is used 7

n+1 1 n n
uM = (u + uM ), n ≥ 1.
2 M−2

This occurs to many central schemes!!!

7
It is well-known to result in instability. An easy remedy is to use

n+1 1 n n
uM = (u + uM )
2 M−1
Chequerboard Mode Solutions
• Generalized Lax–Friedrichs (GLF) scheme with initial datum
uj0 = (−1)j
ν n q n
ujn+1 = ujn − [uj−1 n
−uj+1 ]+ [uj+1 −2ujn +uj−1
n
], 0 < qmin ≤ q ≤ 1.
2 2

1. As q = 1, it is the Lax–Friedrichs scheme, which has the

chequerboard mode solution:
ujn = (−1)j+n .
2. As 0 < qmin ≤ q < 1, there is a damped solution
ujn = (1 − 2q)j+n .
Note that it is the Lax–Wendroff scheme as q = ν 2 .
The same holds true for nonlinear problems:
ν 1 n
ujn+1 = ujn − [f (uj−1
n n
) − f (uj+1 )] + [uj+1 − 2ujn + uj−1
n
].
2 2
• GLF solution with the square signal initial datum
1. Odd discretization case:
M/2−1
1 X
ujn = (1 − 2q)n (−1)j+j1 +M/2 + ckn e i ξj .
M
k=−M/2+1

2. Even discretization case µ

M/2−1
X
ujn n j
= 0 × (1 − 2q) (−1) + ckn e i ξj .
k=−M/2+1

The highest frequency mode is persistent and not damped, which

may be the cause of (local) oscillations.
• MacCormack scheme with initial data uj0 = (−1)j

uj∗ = ujn − λ f (uj+1

 n
) − f (ujn ) ,

1 n
ujn+1 =


uj + uj∗ − λ f (uj∗ ) − f (uj−1
∗
) .
2

8
1. As f (u) is an even function, the solution is

ujn = (−1)j .

2. As f (u) is an odd function, the solution is

 
f (b) ′
ujn+1 = (−1)j+n b 1 − 2λ2 f (ξ) , ξ ∈ (b, b + 2λf (b))
b

f (b) ′
if ujn = (−1)j+n b. Amplification factor: 1 − 2λ2 f (ξ).
b
f (b) ′
The amplitude of solution is amplified if |1 − 2λ2 f (ξ)| > 1.
b
8
Private communication with Huazhong Tang
• MacCormack scheme for ut + (u α )x = 0, 0 < α < 1
The initial data is uj0 = (−1)j . α = 13 , t = 1.19, CFL=0.9.

20
f=u**(1/3)

15

10

-5

-10

-15

-20
0 0.5 1 1.5 2

Oscillations are crazily amplified.

Brief Summary
1. High frequency modes are often present in data, which may
contaminate solutions.

2. High frequency modes may be persistent in solutions, not

damped or be even amplified if there is no sufficiently strong
dissipation.

Question: How do we quantify dissipation effects of numerical

schemes so that the cause of oscillations could be well understood?
Discrete Fourier Analysis on Linear Problems (GLF)

ujn = λnk e ijξ , i 2 = −1, ξ = kh.

• Low Frequency Modes ξ ≈ 0 The amplitude λk is

λ(k) = 1 + q(cos ξ − 1) − iνa sin ξ,

and

|λk |2 = |λ(k)|2 = 1 + 4(ν 2 − q) sin2 (ξ/2) + 4(q 2 − ν 2 ) sin4 (ξ/2).

The amplitude error is

(
O(ξ), for monotone schemes
|λk | − 1 =
2
O(ξ ), for L-W q = ν 2 .

The relative phase error is

arg λ(k)
Ep (k) := − − 1 = O(ξ 2 ).
νξ
Common understanding: The numerical dissipation of monotone
schemes O(ξ) suppresses relative phase errors, while the dissipation
of L–W is too weaker O(ξ 2 ) and oscillations are triggered.

Moreover, for discontinuous L–W solutions, it can be shown that

û(ξ) ∼ 1/|ξ|, as ξ → 0,

where û is the Fourier representation of u.

This implies that low frequency modes can cause oscillations in the
L-W solutions.
• High Frequency Modes ξ ≈ π.
Introduce ξ ′ = k ′ h ≈ 0 such that ξ = π + ξ ′
′ ′
ujn = λnk e ijξ = λnk e i (π+ξ )j = (−1)j+n λnk ′ e i ξ j .

λk and λk ′ have the same amplitude modulus:

|λk ′ |2 = 1 + 4(ν 2 − q) cos2 (ξ ′ /2) + 4(q 2 − ν 2 ) cos4 (ξ ′ /2)

The amplitude error at ξ ′ = 0 is

|λk |2 − 1 = −4q(1 − q),

which vanishes at q = 0. GLF becomes much less dissipative for high

frequency modes as the parameter q increases.
The relative phase error at ξ = π is

π(1 − ν) 2(q − 1)νξ ′

Ep (k) := − + + · · · = O(1).
νξ (2q − 1)νξ

The relative phase errors of high frequency modes O(1) are huge and are
difficult to be suppressed by the numerical dissipation.
Brief Summary for GLF
We distinguish low and high frequency Fourier modes ujn = λnk e ijξ ,
ξ = kh. They behave differently.

1. For the low frequency modes (ξ ≈ 0), the relative phase error
is of order O(ξ 2 ), and the amplitude error (dissipation)
becomes smaller as the parameter q decreases.
1.1 The order of amplitude error is O(ξ) for the monotone
schemes and thereby offsets the relative phase errors.
1.2 The order of amplitude error is O(ξ 2 ) for the LW scheme and
the relative phase errors cannot be offset. Hence oscillations
are observable.
2. For the high frequency modes (ξ ≈ π), the relative phase error
is of order O(1), the amplitude error becomes larger as the
parameter q is closer to 1/2.
In particular, as q = 1 (LF), the amplitude error vanishes and
there is no dissipation on the highest frequency mode. This
explains why oscillations in LF solutions are observable.
Application: Composite Scheme 9

u n+1 = LF ◦ (LW )m u n .

LF has large dissipation on low frequency modes; in contrast, LW

has large dissipation on high frequency modes. When they are
combined to form a composite scheme, their merits are preserved
and some shortcomings are canceled out.

9
R. Liska and B. Wendroff, Composite schemes for conservation laws. SIAM
J. Numer. Anal. 35 (1998), no. 6, 2250õ 2271.
• Remark on Discrete Fourier Analysis

1. Merits: Most often–used approach to investigate the stability

of a scheme
2. Limitation: Linear problems with constant coefficients;
difficult to extend to nonlinear problems or even problems
with variable coefficients.
Modified Equation Approach

1. Linear Part: Traditional modified equation at any wave

number
References:
(a) F. Warming and B. J. Hyett, The modified equation approach
to the stability and accuracy analysis of finite difference
methods, J. Comp. Phys. 14, 159-179, 1974.
(b) Jiequan Li and Zhicheng Yang, The von Neumann Analysis
and Modified Equation Approach for Finite Difference
Schemes, preprint, 2011.
2. Nonlinear Part: Heuristic modified equations (not rigorously)
Reference: Jiequan Li and Zhicheng Yang, Heuristic modified
equation analysis on oscillations in numerical solutions of
conservation laws, SINUM, in press, 2011.
Linear Modified Equations
1. Time-dependent PDE with constant coefficients
K
X
ut = Lx (u), Lx (u) = αs ∂xs u,
s=1

where αs , s = 1, · · · , K are constant.

2. Two-level schemes take the form

N2
X M2
X
n+1
Bu n+1 = Au n , Bp uj+p = n
Ap uj+p ,
p=−N1 p=−M1

where B is invertible, the numerical solution ujn approximates

the PDE solution u(xj , tn ),

ujn ∼ u(xj , tn ), xj = jh, tn = nτ.

Warming & Hyett’s Approach (1974)
The modified equation is the actual partial differential equation
that the difference scheme solves (ujn → u(xj , tn )). The Taylor
series expansion yields the modified equation
X
∂t u = βs (h, τ )∂xs u,
s

which implies the consistency to the equation if

βs = αs , s = 1, · · · , K ;

and the accuracy of order q if

(i ) βs = αs , s = 1, · · · , K ;
(ii ) βs = 0, s = K + 1, · · · , K + q − 1;
(iii ) |βK +p (h, τ )| ≤ Chq , C is constant.
Modified equation from discrete Fourier analysis (Classical!!!)
Write
λk = e i ω(k)τ
where ω(k) is the frequency. Expand ω(k) at k = 0,
∞ ∞
X ω (s) (0) 1 X [ln λ(0)](s) s
i ω(k) = i ks = k , ω(0) = 0.
s! τ s!
s=0 s=0

Formally transform ik → ∂x , i ω → ∂t to obtain the modified

equation at k = 0,
∞
1 X i −s
∂t u = [ln λ(0)](s) ∂xs u.
τ s!
s=0

Note that (e.g., for explicit scheme)

X
λ(s) (0) = i s hs p s Ap .
p

Can write out βs (h, τ ) explicitly only in terms of λ(0).

Example: Linear Advection Equation
∂u ∂u
+c = 0.
∂t ∂x
Consistency conditions are
X X
Ap = 1, pAp = ν,
p p

which implies that the modified equation should be

∞
∂u ∂u 1 X i −s
+c = [ln λ(0)](s) ∂xs u.
∂t ∂x τ s!
s=2

The term of second order derivative is the numerical viscosity,

which vanishes for L-W.
L-W:
∂u ∂u h2 ∂3u
+c = − (1 − ν 2 ) 3 + · · · .
∂t ∂x 6 ∂x
L-F:
∂u ∂u h2 ∂2u
+c = (1 − ν 2 ) 2 + · · · .
∂t ∂x 2τ ∂x
Modified equation at any wave-number
Motivated by the von Neumann analysis, we make an ansatz,

ujn = αnj ũjn , αnj = e i (nω0 τ +jk0 h) , ω0 = ω(k0 ).

Substituting it into the (explicit) scheme yields

X
e i ω0 τ ũjn+1 = Ap α0p ũj+p
n
, α0p = e ipk0 h .
p

Replace ũjn by the value of a smooth function ũ(xj , tn ) to obtain

X
e i ω0 τ ũ(xj , tn+1 ) = Ap α0p ũ(xj+p , tn ).
p

Then we can obtain the modified equation for ũ(x, t)

Theorem. The modified equation for ũ at k0 is
∞
ln |λ(k0 )| 1 X i −s
∂t ũ = ũ + [ln λ(k0 )](s) ∂xs ũ, i 2 = −1.
τ τ s!
s=1
Interpretation of modified equation at any wave number
1. Connection between ũ and u. At k = 0, ũ = u. Their amplitudes
are the same.
2. Full connection with the von Neumann analysis.
∞
X ω̃ (s) (k0 )
ω̃ = ω̃(k0 + k̃) = k̃ s .
s=1
s!

Then we transform iω → ∂t and i k̃ → ∂x to obtain formally the

modified equation.
3. The propagation of wave packets and dispersion. Introduce the
group velocity
dω(k) i d ln λ(k)
G (k) = − = .
dk τ dk
Then the modified equation can be rewritten as
∞
X i −(s−1)
ln |λ(k0 )|
∂t ũ + G (k0 )∂x ũ = ũ − G (s−1) (k0 )∂xs ũ.
τ s=1
s!
This gives the PDE interpretation of group velocity in terms of
modified equation.
4. Numerical Dissipation. We distinguish the dissipation as

ln |λ(k0 )|
4.1 Numerical damping: ∂t ũ ∼ ũ. Such solution behaves
τ
ln |λ(k0 )|
ũ(t) ∼ ũ(0)e αt , α = .
τ
As α < 0 (⇐⇒ |λ| < 1), the solution decays exponentially.
As |λ| ≤ 1 + C τ (von Neumann stability condition), the
solution is bounded

|ũ| ≤ |ũ(0)|E C .
′′
4.2 Numerical diffusion: ∂t ũ ∼ ǫ∂ 2 ũ, ǫ = ω̃ 2(k0 ) . The solution
behaves
ũ(t) ∼ O(t −1/2 ),
if ǫ > 0. Note that:
(i) As k0 = 0, |λ(0)| = 1 is the scheme is consistent. Then the
numerical diffusion (viscosity) dominates the dissipation.
(ii) As k0 6= 0, generally speaking |λ(k0 )| < 1. Hence the
numerical damping dominates the dissipation.
Nonlinear Modified Equations
• Conservative Schemes
τ n
ujn+1 = ujn − [Fj+ 1 −F
n
j− 12
],
h 2

n
where numerical fluxes Fj+ n n
1 = F (uj−q , · · · , uj+p ) is consistent
2
with f (u)
F (u, · · · , u) = f (u).
The classical modified equation of monotone schemes is (Harten,
Hyman & Lax, CPAM, 1976),

ut + f (u)x = τ ∂x (β(u, ν)∂x u), β(u, ν) > 0,

where the “red” term is the numerical viscosity, which suppresses

oscillations near shocks.
Note that such dissipation just takes effect on oscillations caused
by low frequency modes. Stronger dissipation is needed to suppress
the oscillations caused by high frequency modes.
Ansatz for oscillatory modes 10

ujn = (−1)j+n ũjn .

We obtain a heuristic modified equation for ũ,

Γ(ũ)
Dt ũ + G (ũ)x = −2 ũ + τ · (β(ũ, λ)ũx )x + τ · δ(ũ, λ)ũx2 + O(τ 2 ),
τ

where Dt is the forward difference operator in time

Dt ũ = (ũ(t + τ, ·) − ũ(t, ·))/τ , G , Γ, β and δ are all functions of
ũ, λ = τ /h.
The damping factor Γ(ũ) is computable

Γ(ũ) := 1 + (−1)n+j+1 λ F ((−1)n+j−q+1 ũ, · · · , (−1)n+j+p ũ)



−F ((−1)n+j−q ũ, · · · , (−1)n+j+p−1 ũ) /(2ũ).

10
K.W. Morton and D.F.Mayers, Numerical solution of partial differential
equations, 2nd edition, Cambridge University Press, 2005.
Numerical Damping Determines Dissipation!!!

Γ(ũ)
Dt ũ = −2 ũ.
τ
1. |ũ(t + τ, ·)| < |ũ(t, ·)| if 0 < Γ(ũ) < 1;

2. |ũ(t + τ, ·)| > |ũ(t, ·)| if Γ(ũ) < 0 or Γ(ũ) > 1;

3. |ũ(t + τ, ·)| = |ũ(t, ·)| if Γ(ũ) = 0 or Γ(ũ) = 1.

Stability Criterion from Heuristic Modified Equation

Proposition. The heuristic modified equation implies the criterion

of the stability of the oscillatory solution:
Γ(ũ)
(i) Damping. As 0 < Γ(ũ) < 1, the term −2 ũ plays the role
τ
of numerical damping that dissipates the oscillatory part and
suppresses oscillations eventually.
(ii) Neutrality. As Γ = 0 or Γ = 1, the dissipation effect depends
on the numerical viscosity (the term of second order derivative).
For oscillatory modes, no dissipation effect exists and the
oscillations persist.
(iii) Amplification. As Γ < 0 or Γ > 1, the factor Γ(ũ) amplifies
the magnitude of the oscillatory modes, which may lead to the
instability.
Monotone Schemes

ujn+1 = H(uj−q
n n
, · · · , uj+p ) = ujn − λ[Fj+
n
1 −F
n
j− 1
],
2 2

where n
Fj+ = n
F (uj−q+1 ,··· n ).
, uj+p The monotonicity means that
1
2

∂H
(v−q , · · · , vℓ , · · · , vp ) ≥ 0,
∂vℓ
for all −q ≤ ℓ ≤ p.
Proposition. If the schemes are monotone, then there holds,

0 ≤ Γ(ũ) ≤ 1,

which implies that oscillatory modes are not amplified.

Specific Examples
1. Generalized Lax–Friedrichs Schemes
n
The numerical flux Fj+ 1 takes the form
2

n
n
f (uj+1 ) + f (ujn ) Q n
Fj+1/2 = F (ujn , uj+1
n
)= − (u − ujn ),
2 2λ j+1
where 0 < Qmin ≤ Q ≤ 1. It is easy to calculate
Γ(ũ) = 1 + (−1)n+1 λ F ((−1)n ũ, (−1)n+1 ũ)


−F ((−1)n−1 ũ, (−1)n ũ) /(2ũ)

= 1 − Q.
Hence we have 0 ≤ Γ(ũ) ≤ 1, and the generalized Lax-Friedrichs
scheme does have the damping effect around the highest frequency
mode if 0 < Qmin < Q < 1.
As Q = 1, it is the Lax–Friedrichs scheme and Γ(ũ) = 1. This
explains why local oscillations in the solution obtained by the
Lax–Friedrichs scheme are observed once the initial data is polluted
by the oscillatory modes.
Remark. This is consistent with Breuss’s conclusion that schemes
with a large viscosity coefficient are prone to oscillations at data
extrema.
2. The Richtmyer Scheme. A two–step Lax–Wendroff scheme with
 n+1/2
n n n
 Fj+1/2 = F (uj , uj+1 ) = f (uj+1/2 ),


 u n+1/2 = 1 (u n + u n ) − λ f (u n ) − f (u n ) .


j+1/2 j j+1 j+1 j
2 2

1 + (−1)n+1 λ f − λ2 (f ((−1)n+1 ũ) − f ((−1)n ũ))



Γ(ũ) =

−f − λ2 (f ((−1)n ũ) − f ((−1)n−1 ũ)) /(2ũ).

◮ Odd flux case, f (−u) = −f (u).

f (λf (ũ))
Γ(ũ) = 1 − λ .
ũ
◮ Even flux case, f (−u) = f (u).

Γ(ũ) = 1.
◮ General cases, f (u) = fodd (u) + feven (u).

fodd (λfodd (ũ))

Γ(ũ) = 1 − λ .
ũ
3. Zwas–Abarbanel Third Order Scheme A third order scheme with
 
n+1 n 1 n n δ n n n n
uj = uj − λ (fj+1 − fj−1 ) − (fj+2 − 2fj+1 + 2fj−1 − fj−2 )
2 12
λ2 h n n
i
+ aj+ 1 (fj+1 − fj n ) − aj−
n n n
1 (fj − fj−1 )
2 2 2

ǫλ3  n
a (f n − fj n ) − 2ajn (fj+1 n n n
(fj n − fj−2
n

+ − fj−1 ) + aj−1 ) ,
12 j+1 j+2
As δ = ǫ = 0, back to L-W.
◮ Odd flux case.

Γ(ũ) = 1 − λ2 a(0)f (ũ)/ũ.

◮ Even flux case.

Γ(ũ) = 1.
◮ General cases.

Γ(ũ) = 1 − λ2 a(0)fodd (ũ)/ũ.

Numerical Examples
◮ Damping. The initial data

1, j = 49, 50, 51,

(
Dodd : uj0 =
0, otherwise.

0.5 0.4
odd odd
0.45 even even
0.35

0.4
0.3
0.35

0.25
0.3

0.25 0.2

0.2
0.15

0.15
0.1
0.1

0.05
0.05

0 0
0 1 2 3 4 5 6 7 8 9 10 0 1 2 3 4 5 6 7 8 9 10

(e) The L–F scheme with (f) The generalized L–F

Q = 1.0 scheme with Q = 0.9

Figure: The propagation of a single square signal. The flux

u
f (u) = √1+u 2
4. Neutrality. The initial data

uj0 = Cj + 0.0001 × (Dodd )j ,

where Cj is the highest frequency mode Cj = (−1)j .

1.5 1.5
Lax−Fridrichs MacComack
1 1

0.5 0.5

0 0

−0.5 −0.5

−1 −1

−1.5 −1.5
3 4 5 6 7 3 4 5 6 7

1.5 1.5
Richtmyer ZA
1 1

0.5 0.5

0 0

−0.5 −0.5

−1 −1

−1.5 −1.5
3 4 5 6 7 3 4 5 6 7

Figure: The neutrality of schemes on high frequency modes. The

number of computation steps is 10. The Burgers equation is used.
5. Amplification.
ΓLxF (1) = 0, ΓMac (1) = 0.6994,
ΓRich (1) = −0.9188, ΓZA (1) = −2.6204.

0.5
1 Lax−Friedrichs MacCormack

0.5

0 0

−0.5

−1
−0.5
3 4 5 6 7 3 4 5 6 7

Richtmyer ZA
2 5

0 0

−2 −5

3 4 5 6 7 3 4 5 6 7

Figure: One first step computation: The flux function is

u
f (u) = √ .
1 + u2
 1.1 0.4
Lax−Friedrichs MacCormack
1.05 0.3

|u|

|u|
1 0.2

0.95 0.1

0.9 0
1 2 3 4 1 2 3 4
step step

20 200
Richtmyer ZA
log10|u| 15 150

log10|u|
10 100

5 50

0 0
1 2 3 4 1 2 3 4
step step

Figure: Amplification or damping trend of oscillatory modes.

Summary and Discussions
What we have done:
1. Both discrete Fourier analysis and modified equation
approaches are used to analyze oscillatory modes. A heuristic
modified equation is derived.
2. Numerical dissipation is distinguished as numerical damping
ornumerical diffusion. Numerical damping plays a dominant
role in suppressing oscillations caused by oscillatory modes.
3. A heuristic stability criterion is proposed to check the
numerical damping effect on oscillatory modes.
What we want to do µ
1. Rigorously derive the modified equation in order to analyze
oscillatory modes.
2. Extend the results in a general setting, not only for scalar
conservation laws.
3. Use the results to design new schemes.
4. · · ·
THANK YOU VERY MUCH