SCIENTIFIC COMPUTING I

I. Classical PDE - Solutions and properties

1. General definitions
1.1 BVP and IVP
• BVP (Boundary Value Problem): is a problem formed by one or several PDE with boundary
conditions given on the totality of the boundary.

• IVP (Initial Value Problem, or Cauchy Problem): is a time depending problem with an
initial data in time.

1.2 Classification of PDE

Only linear and deterministic problems will be considered in this course. For second order PDE:

auxx + buxy + cuyy + dux + euy + f u = g (1)

computing:
δ = b2 − 4ac (2)
the PDE can be classified as:

• δ < 0: Elliptic.

• δ = 0: Parabolic.

• δ > 0: Hyperbolic.

This definition does not include all PDE. For other PDE, technical algebraic manipulations
must be done.
The problem A(u) = f is well posed if, for all data f

• there exist a unique solution u and

• the solution depends continuously on the data f.

1.3 Boundary Conditions (BC)

• Dirichlet boundary condition: applied on the magnitude (or solution).

• Neumann boundary condition: applied on the derivative (or flux) of the magnitude.

• Fourier (or Robin) boundary condition: linear combination of the two previous ones.

Needless to say, all BC are applied on the domain Ω of the PDE.

1 of 13
2. Elliptic PDE
2.1 Poisson equation
The Poisson equation, which represents the conductivity of a magnitude, is defined by:

−∆u = f x∈Ω (3)

and if the source term f = 0, the Laplace equation is recovered:

∆u = 0 x ∈ Ω (4)

Considering the Poisson problem in 1D:

 00
−u (x) = f (x) for x ∈]0, 1[
(5)
u(0) = u(1) = 0

using the Green function:


y(1 − x) if 0 ≤ y ≤ x
G(x, y) = (6)
x(1 − y) if x ≤ y ≤ 1

the solution of the problem is: Z 1

u(x) = G(x, y)f (y)dy (7)
0
The properties of this solution are:

• The problem is well posed (existence, uniqueness).

• G is positive.

• The maximum principle is satisfied: if f ≥ 0 then u ≥ 0, and if f > 0 then u > 0 on ]0, 1[.

• For a right hand side f located in a small zone, the solution u is not equal to 0 all over the
domain.

• If the right hand side f is locally modified, the solution u is modified globally.

• If f ∈ C 0 ([0, 1]) then u ∈ C 2 ([0, 1]).

If the problem is defined on all the space:

 00
−u (x) = f (x) for x ∈ R3
(8)
u → 0 |x| → ∞

it has a unique solution given by:

Z
1 f (x)
u(x) = dy (9)
4π R3 |x − y|

3. Parabolic PDE
3.1 Heat equation
To represent the heat equation, where u is the time dependent temperature and κ is the heat
conductivity (it represents the diffusion of a magnitude):

ut − κ∆u = 0 (10)

2 of 13
 Considering the 1D initial or boundary problem on I = [0, l] ⊂ R:

ut − κux x = 0 x ∈ I, t > 0
u(x, 0) = φ(x), x ∈ I (11)
u satisfies some BC


where the most common BC are:

• Dirichlet: u(0, t) = 0 = u(l, t).

• Neumann: ux (0, t) = 0 = ux (l, t).

• Robin: ux (0, t) − a0 u(0, t) = 0 and ux (l, t) + a0 (l, t) = 0.

• Periodic: u(−l, t) = u(l, t) and ux (−l, t) = ux (l, t).

3.2 Solution of different BC

Dirichlet BC:
∞  nπ 
X nπ 2
u(x, t) = An sin e−κ( l
) t
(12)
l
n=1

where: Z l
2  nπ 
An = sin φ(x)dx (13)
l 0 l
Periodic BC:
∞ h  nπ   nπ i
X nπ 2
u(x, t) = A0 + An cos + Bn sin e−κ( l ) t (14)
l l
n=1

where:
Z l Z l Z l
1 1  nπ  1  nπ 
A0 = l φ(x)dx An = cos φ(x)dx Bn = sin φ(x)dx (15)
2 −l l −l l l −l l

Solution in R:

Z ∞
(x − y)2
1 −
u(x, t) = √ φ(y)e 4κt dy (t > 0)
4κπt −∞ (16)
lim u(x, t) = φ(x) (t = 0)
t→0+

Extension in RN : Z
u(x, t) = H(x − y, t)φ(y)dy (17)
RN
where H(x, t) is the fundamental solution of the heat equation:

|x|2

1 −


H(x, t) = N/2
e 4κt t>0 (18)
 (4πκt)

0 t<0

3 of 13
3.3 Parabolic regularization
Assuming φ ∈ C(RN ) ∩ L∞ (RN ), then the solution of the heat equation satisfies:
u ∈ C ∞ (RN × (0, ∞)) and lim u(x, t) = φ(x0 ) (19)
(x, t) → (x0 , t)
x0 , x ∈ R, t > 0
This theorem represents an asymptotic behaviour: the solution converges to 0 when t → 0
(which makes sense, since heat is dissipated with time).
It is also important to remark that for φ > 0, then u(x, ε) > 0, ∀x ∈ Ω and ε > 0, which
means that for an initial point which is cold (φ(x)  1) and very far from the heat source, it
becomes immediately hot. Thus we have a propagation of the heat with infinite speed, which
means that this is not a realistic model.
Also, for t ≤ 0 the problem is not well posed, since a solution cannot be obtained. This is
modelled with the backward heat equation (not well posed).

4. Hyperbolic PDE
4.1 Transport and waves equation
The transport equation represents the convection of a magnitude:
ut + c.∇u = 0 x, c ∈ RN t>0 (20)
and the propagation of waves is expressed as:
utt − c2 ∇u = 0 x, c ∈ RN t>0 (21)
Considering the transport equation in 1D:
ut + cux = 0 x, c ∈ R t > 0 (22)
the solution is constant along the characteristic curves:
x + ct → x(t) = ct + x(0) (23)

4.2 The Cauchy Problem

The solution of the general Cauchy Problem in N dimensions is:
x, c ∈ RN t > 0

ut + c.∇u = f (x, t)
(24)
u(x, 0) = u0 (x) x ∈ RN
Z t
u(x, t) = u0 (x − ct) + f (x + (s − t)c, s)ds (25)
0

4.3 The Initial Boundary Problem (IBV)

The solution of: 
ut + cux = f (x, t) x > 0, c > 0, t > 0
u(x, 0) = u0 (x) x∈R (26)
u(0, t) = u1 (t) t>0


is equal to:  Z t
 u0 (x − ct) + f (x + (s − t)c, s)ds

 x > ct
u(x, t) = 0Z x  (27)
 x 
1 s 
u1 1 − +c f x − s, t − ds x < ct


c 0 c
The interpretation is that the solution corresponds to translate u0 with the quantity ct.

4 of 13
4.4 Regularisation and conservation properties
For ∀t > 0, u(., t) has the same regularity as u0 , so there is not an instantaneous regularisation
like for the parabolic problem, but there is a propagation of a singularity. Two theorems can be
applied:

• Conservation of the ‘mass’:

Z Z
u(x, t)dx = u0 (x − ct)dx (28)
R R

• Conservation of the infinite norm:

max |u(x, t)| = max |u0 (x)| (29)

x∈R x∈R

Reversibility in time: For t < 0, the formula u(x, t) = u0 (x − ct) is still satisfied, so there is a
solution for the transport problem for negative time. This problem is associated to a propagation
with speed −c.

II. Finite Difference Method

1. Discretization
• Space: 1D uniform mesh grid on Ω = [0, 1]:

x0 = 0 < x1 < ... < xN < xN +1 = 1 →

1 (30)
→ xj = j∆x with ∆x = 0≤j ≤N +1
N +1

• Time: for t ∈ [0, T ]:

1
tn = n∆t with ∆t = 0≤n≤T (31)
T +1

To simplify notation:
unj ≈ u(xj , tn ) (32)
where unj is the approximated solution given by the scheme and u(xj , tn ) is the exact solution.

2. Numerical analysis
2.1 Some definitions
A general finite difference scheme is formally defined by:
n o 
n+m
F∆t,∆x uj+k − + −
=0 (33)
m ≤m≤m ,k ≤k≤k+

where the integers m− , m+ , k − and k + define the width of the stencil. The stencil is the point
defined by the couple (j, n).
A FDS is linear if F (u) = 0 is linear with respect to un+m
j+k .
Consistency

5 of 13
The FDS is consistent with the PDE defined by F (u) = 0 if, for every sufficiently regular solution
of this equation, the truncation error of the system, defined by:
 
F∆t,∆x {u(t + m∆t, x + k∆x)}m− ≤m≤m+ ,k− ≤k≤k+ = 0 (34)

tends to zero uniformly with respect to (t, x) as ∆t and ∆x tend to zero independently. In
n+m
practice, we calculate the truncation error of a scheme by replacing uj+k with its correspondent
Taylor expansion.
Accuracy
The scheme has an accuracy of order p in space and of order q in time if the truncation error
tends to zero as:
o(∆xp + ∆tq ) (35)
when ∆x and ∆t tend to zero.

2.2 Lp Norm
Considering the norm on RN for the numerical solution:
 1
N p
X
n
||u ||p =  ∆x|unj |p  for 1 ≤ p ≤ +∞ (36)
j=1

In practice, p = 2 and p = +∞ are the ones that are used. When p = +∞:

||un ||∞ = max |unj | (37)

i≤j≤N

2.3 Stability analysis

A FDS is stable for the norm ||.|| if there exists a constant K > 0 independent on ∆t and ∆x
such that for every arbitrary initial data u0 :

||un || ≤ K||u0 || for all n ≥ 0 (38)

If this condition only holds for steps ∆t and ∆x defined under certain inequalities, the scheme
is conditionally stable.
A two level linear scheme can be written as:

un+1 = Aun (39)

where A is the amplification matrix (value for purely implicit and explicit schemes; such schemes
will be defined hereafter). Since un = An u0 , the stability is:

||An u0 || ≤ K||u0 || → ||An || ≤ K ∀n ≥ 0 (40)

3. FDS for Elliptic PDE

3.1 Finite Difference Formula based on Taylor expansion
• First derivative:

– Centered scheme (2nd order approximation):

 
∂u uj+1 − uj−1
(xj ) ≈ (41)
∂x 2∆x

6 of 13
 – Forward and backward schemes (1st order):
   
∂u uj+1 − uj ∂u uj − uj−1
(xj ) ≈ (xj ) ≈ (42)
∂x ∆x ∂x ∆x

• Second derivative (centered) (3rd order):

 2 
∂ u uj+1 − 2uj + uj−1
2
(xj ) ≈ (43)
∂x ∆x2

3.2 FDS for Laplace problem

For Dirichlet boundary problem (u0 and uN +1 known), the Laplace 1D problem is:
 2 
∂ u uj+1 − 2uj + uj−1
− 2
(xj ) = fj → − = fj (44)
∂x ∆x2
so there is a linear system AU = b:
 u0 
  u  f1 −
2 −1 0 1  ∆x2 
−1 2 −1  u 2
  f 2

1    
..  =  ..

. . .  (45)

 .   .

∆x2 
   
−1 2 −1 uN −1   fN −1 
 
0 −1 2 uN
 uN +1 
fN −
∆x2
Maximum principle and convergence
• The matrix A is tridiagonal, symmetric and inversible.

• If all components of b are negative or equal to zero, then all components of the solution U
are negative or equal to zero.

• For u ∈ C 4 ([0, 1]), the solution of the continuous problem and U the approximated solution
given by the linear system:

∆x2
max |u(xi ) − U (xi )| ≤ max |u(4) (x)| (46)
1≤i≤N 96 x∈[0,1]
This convergence is obtained thanks to consistency and stability.

4. FDS for Parabolic PDE

4.1 Discretization
The space-derivatives discretization will be the same than for the elliptic case. For the time
derivatives, the following will be considered:
• Centered scheme (2nd order):

un+1 − ujn−1
 
∂u j
(xj , tn ) ≈ (47)
∂t 2∆t

• Forward and backward schemes (1st order):

un+1 − unj unj − ujn−1

   
∂u n j ∂u
(xj , t ) ≈ (xj , tn ) ≈ (48)
∂t ∆t ∂t ∆t

7 of 13
4.2 FDS for Heat equation
Considering the heat equation:

∂u ∂2u
−ν 2 =0 (x, t) ∈ (0, 1) × R+


 f or ∗
∂t ∂x (49)

 u(0, x) = u0 (x) (IC)
u(t, 0) = u(t, 1) = 0 (BC)

We consider u0j = u0 (xj ) (IC) and un0 = unN +1 = 0 (BC). There are two basic types of
schemes:

• Explicit: information at a given time step can be obtained straightforward from the infor-
mation (or solutions) for previous times.

• Implicit: information at a given time depends on itself, so a linear system needs to be

solved. Are generally more stable than explicit schemes, but require more computational
costs.

Explicit scheme (or Forward Time Centered Space FTCS)

un+1
j − unj unj−1 − 2unj + unj−1
−ν =0 (50)
∆t ∆x2
Although stability and accuracy will be analysed after, it is advanced that:

• Accuracy: second order in space and first order in time.

• Stability: stable in L2 and stable in L∞ if 2ν∆t ≤ ∆x2 .

Implicit scheme
un+1
j − unj un+1 n+1
j−1 − 2uj + un+1
j−1
−ν =0 (51)
∆t ∆x2
• Accuracy: second order in space and first order in time.

• Stability: stable in L2 and stable in L∞ .

DuFort-Frankel
un+1
j − un−1
j unj−1 − 2unj + unj−1
−ν =0 (52)
2∆t ∆x2
• Accuracy: second order in space and second order in time.

• Stability: stable in L2 if ∆t/∆x2 is bounded and stable in L∞ if 2ν∆ ≤ ∆x2 .

Gear scheme
3un+1
j − 4unj + un−1
j unj−1 − 2unj + unj−1
−ν =0 (53)
2∆t ∆x2
• Accuracy: second order in space and first order in time.

• Stability: stable in L2 if ν∆t ≤ ∆x2 (L∞ ?).

θ-scheme for 0 ≤ θ ≤ 1

un+1
j − unj un+1 n+1
j−1 − 2uj + un+1
j−1 unj−1 − 2unj + unj−1
− θν − ν(1 − θ) =0 (54)
∆t ∆x2 ∆x2
• Accuracy: second order in space and first order in time.

8 of 13
 • Stability: stable in L2 if 2(1 − 2θ)ν∆t ≤ ∆x2 (for 0 ≤ θ ≤ 1/2) and unconditionally stable
in L2 if 1/2 ≤ θ ≤ 1 (L∞ ?).

• If θ = 0: explicit scheme.

• If θ 6= 0: implicit scheme.

• If θ = 1/2: Crank-Nicholson scheme.

– Accuracy: second order in space and second order in time.

– Stability: stable in L2 and stable in L∞ if ν∆t ≤ ∆x2 .

4.3 Stability in L∞ norm

A FDS satisfies a discrete maximum principle if ∀n ≥ 0 and 1 ≤ j ≤ N :

min u0j ≤ unj ≤ max (55)

0≤j≤N +1 0≤j≤N +1

for all arbitrary initial data u0 .

The explicit scheme satisfies a discrete maximum principle; it is only stable in the L∞ norm
if and only if the condition 2ν∆t ≤ ∆x is satisfies. This condition is called CFL condition:
ν∆t
CF L = (56)
∆x2
The implicit scheme is stable in the L∞ norm for any time and space step; thus it is uncon-
ditionally stable.

4.4 Stability in L2 norm

Many schemes do not satisfy the maximum discrete principle, but are nevertheless good schemes,
since they are L2 stable. The amplification factor is obtained using the Fourier analysis, where
u is considered to be periodic for the domain Ω and thus can be expressed as:
 n
uj+1 = uj ei2πk∆x

n n i2πkx
uj = ûk e → (57)
unj−1 = uj e−i2πk∆x

Considering the spacial discretization, it can be obtained the spacial eigenvalue, substituting
the previous expressions and using the Euler’s relations:
d n
û = λ(k)ûnk (58)
dt k
so the time discretization is expressed only as a function of λ(k). Computing the amplification
matrix:  n+1   n 
u u
= A n−1 (59)
un u
it will be expressed as a function of λ(k). The stability criterion will be:

ρ(A) ≤ 1 (60)

where ρ represents the spectral radius of the matrix (maximum absolute value of the eigenvalues
of the matrix). This stability criterion is called Von-Neumann stability criterion. Such stability
is studied in the homework (see procedure there).

9 of 13
4.5 Convergence of the scheme: Lax theorem
Assuming that the scheme is linear, two level, consistent and stable for a norm ||.||, then the
scheme is convergent in the sense that:
!
∀T > 0 lim sup ||en || =0 (61)
∆t,∆x→0 tn ≤T

where enj = unj − u(tn , xj ) is the error vector. If the scheme has an accuracy of order p in space
and q in time, then for all T > 0 there exists a constant CT such that:

sup ||en || ≤ CT (∆xp + ∆tq ) (62)

tn ≤T

5. FDS for Hyperbolic PDE

5.1 FDS for Transport equation
Considering the transport equation problem:

ut + cux = f (x, t) x > 0, c > 0, t > 0
(63)
u(x, 0) = u0 (x) x∈R

with solution u(x, t) = u0 (x − ct), there are some properties to take into account:

• The solution satisfies the maximum principle (L∞ stable):

inf u0 (x) ≤ u(x, t) ≤ sup u0 (x) ∀x ∈ R, t ≥ 0 (64)

x∈R x∈R

• The Total Variation (TV) of the solution is constant:

Z Z
∂u
T V (u(., t)) = (x, t) dx = |u00 (x)|dx = T V (u0 ) = cte (65)
R ∂x R

since both the derivative of the solution and the initial condition are the same. The TV
measures the variation of the first derivative, so for first class PDE of this type is constant.
For nonlinear PDE, the TV is not constant, but decreases.

For the transport equation, the CFL condition is not the same as for the parabolic equation;
it is defined as:
c∆t
CF L = (66)
∆x
Explicit scheme
un+1
j − unj unj+1 − unj−1
+c =0 (67)
∆t 2∆x
• Accuracy: second order in space and first order in time.

• Stability: unstable in L2 .

Crank-Nicholson scheme

un+1 un+1 n+1

!
j − unj c∆t j+1 − uj−1 unj+1 − unj−1
+ + =0 (68)
∆t ∆x 2 2

• Accuracy: second order in space and second order in time.

10 of 13
 • Stability: the amplification factor depens on the spatial eigenvalue:

1 + 12 λ∆t
A= (69)
1 − 21 λ∆t

Implicit scheme
un+1
j − unj un+1 n+1
j+1 − uj−1
+c =0 (70)
∆t 2∆x
• Accuracy: second order in space and first order in time.

• Stability: stable in L2 .

Non-centered explicit scheme

Backward scheme (for c > 0):

un+1
j − unj unj − unj−1
+c =0 (71)
∆t ∆x
• Accuracy: first order in space and first order in time.

• Stability: stable in L2 if CF L ≤ 1, and stable in L∞ if c > 0 and under the same CFL
condition.

Forward scheme (for c < 0):

un+1
j − unj unj+1 − unj
+c =0 (72)
∆t ∆x
• Accuracy: first order in space and first order in time.

• Stability: stable in L2 if CF L ≤ 1, and stable in L∞ if c < 0 and under the same CFL
condition.

Leap-Frog scheme
un+1
j − un−1
j unj+1 − unj−1
+c =0 (73)
2∆t 2∆x
• Accuracy: second order in space and second order in time.

• Stability: stable in L2 if CF L ≤ 1.

Lax-Friedrichs scheme
un n
j−1 +uj+1
un+1
j − 2 unj+1 − unj−1
+c =0 (74)
∆t 2∆x
• Accuracy: second order in space and first order in time.

• Stability: stable in L2 if CF L ≤ 1, and satisfies the discrete maximum principle (stable in

L∞ ). The Total Variation (TV) decreases (T V n+1 ≤ T V n ).

Lax-Wendroff scheme
un+1
j − unj unj+1 − unj−1 ∆t unj+1 − 2unj + unj−1
+c − c2 =0 (75)
∆t 2∆x 2 ∆x2
• Accuracy: second order in space and second order in time.

• Stability: stable in L2 if CF L ≤ 1. The TV is not preserved.

11 of 13
6. FDS and Boundary Conditions
Considering the Elliptic (Laplace) equation (although the reasoning is valid for every equation):
 00
−u (x) = f (x) for x ∈]0; 1[ uj+1 − 2uj + uj−1
→− = fj (76)
u(0) = 0 u0 (1) = 0 ∆x2

• For i = 0: u0 = 0, so u0 is known.

• For i = 1: the scheme becomes:

2u1 − u2
= f (x1 ) (77)
∆x2
• For i = N + 1: u0 (1) needs to be approximated using FDS:

– First order:
uN +2 − uN +1 uN +1 − uN
u0 (1) = = 0 → uN +1 = uN +2 → = f (xN +1 ) (78)
∆x ∆x2

– Second order (Taylor expansion around uN +2 ):

∆x2
uN +2 = uN +1 + u0 (1)∆x + u00 (1) (79)
2
and since u0 (1) = 0 and u00 (1) = −f (1) = −f (xN +1 ):

∆x2
uN +2 = uN +1 − f (xN +1 ) (80)
2
Substituting in the scheme:
∆x2
uN +1 − − 2uN +1 + uN
2 f (1) 2(uN +1 − uN )
− = f (1) → = f (xN +1 ) (81)
∆x2 ∆x2
The final scheme can be expressed as:
2u1 − u2

 = f (x1 )
∆x2



 uj+1 − 2uj + uj−1
− = fj (82)
 ∆x2
u − u 2(uN +1 − uN )

 N +1 N

or = f (xN +1 )

∆x 2 ∆x2
The scheme of order 1 introduces an error of order 1 at the BC. Even if the scheme is of order
2, an approximation of order 1 at the boundary implies a global first order error. Therefore, one
must choose a BC conditions that at least the order of the system.
For E = U − Uexact the error vector, it can be proven that:

• For scheme 1: |E|∞ ≤ C∆x, so the scheme is of order 1.

• For scheme 2: |E|∞ ≤ C∆x2 , so the scheme is of order 2.

where C is a constant depending only on f , u and the derivatives of u. The scheme 1 is also not
consistent, since the truncation error near the right boundary condition does not tend to zero
when ∆x → 0.

12 of 13
7. FDS for multidimensional cases
The Heat equation in 2D problem will be considered (although the concept is applicable to every
equation):  2 2u
 
∂u ∂ u ∂
−ν + 2 = 0 for (x, y, t) ∈ Ω × R+

∗

∂t ∂x2 ∂y

(83)

 u(0, x, y) = u0 (x, y) for (x, y) ∈ Ω
u(t, x, y) = 0 for t ∈ R+ (x, y) ∈ ∂Ω

∗
which scheme (explicit) is represented as:
un+1 n
j,k − uj,k unj+1,k − 2unj,k + unj−1,k unj,k+1 − 2unj,k + unj,k−1
−ν −ν =0 (84)
∆t ∆x2 ∆y 2
The new CFL condition for L∞ stability is:
 
1 1
ν∆t + ≤ CF L2D (85)
∆x2 ∆y 2
Two cases will be considered:
• ∆x ∼ ∆y: the solutions in both directions are of the same order, which makes:
2ν∆t CF L1D 1
2
= 2CF L1D ≤ CF L2D → CF L2D = = (86)
∆x 2 2
so the CFL condition in 2D is more restrictive than the CFL in 1D.
• ∆x  ∆y: one solution in one direction is way greater than the other:
ν∆t
= CF L1D ≤ CF L2D → CF L2D = CF L1D = 1 (87)
∆x2
so the 1D scheme is recovered.

8. Equivalent equation
The equivalent equation of a scheme is the equation obtained by adding the principal part of the
truncation error to the model studied (that is, the term with dominant order). As an example,
lets take the Lax-Friedrichs scheme. The original equation and the truncation error are:
un n
j−1 +uj+1
un+1
j − unj+1 − unj−1
2
ut + cux = 0 → +c =0→
∆t  2∆x (88)
∆x2 (c∆t)2

→E=− 1− uxx = νuxx
2∆t 2∆x2
and the equivalent equation:
ut + cux = νuxx → ut + cux − νuxx = 0 (89)
The coefficient of diffusion ν is called numerical diffusion. If it is large, the scheme is diffusive
(or dissipative). The typical behaviour of a diffusive scheme is its tendency to artificially spread
out the initial data in the course of time. The schemes that are too diffusive are therefore bad
schemes. If ∆t is very small, ν may be very large and the scheme is bad as it is too weighted
to the diffusion. The Lax-Friedrichs scheme is a good approximation (order 2) of the equivalent
equation where the coefficient ν is small.
In the case of a dispersive scheme, if the principal part of the truncation error is added to the
equation, then this scheme is not only consistent with this new equivalent equation, but it is also
strictly more accurate. The equivalent equation of the Lax-Wendroff scheme contains a third
order term, called dispersive. The typical behaviour of a dispersive scheme is that it produces
oscillations when the solution is discontinuous. The coefficient of this dispersive term can be
neglected compared to the coefficient of diffusion for the diffusive schemes.

13 of 13