NOTES ON BURGERS’S EQUATION

MARIA CAMERON

Contents
1. Solution of the Burgers equation with nonzero viscosity 1
2. Shock speed 3
3. Characteristics of the Burgers equation 5
4. Weak solutions 6
5. The Riemann problem 6
5.1. Case 1: uL > uR 8
5.2. Case 2: uL < uR 9
6. Numerical methods for hyperbolic conservation laws 9
6.1. Conservative methods for nonlinear problems 10
6.2. Discrete conservation 11
6.3. Consistency 12
6.4. Generalization of methods developed for the advection equation 12
6.5. Convergence 13
6.6. Godunov’s method 15
6.7. Glimm’s method 17
References 17

Burgers’s equation
(1) ut + uux = νuxx
is a successful, though rather simplified, mathematical model of the motion of a viscous
compressible gas, where
• u = the speed of the gas,
• ν = the kinematic viscosity,
• x = the spatial coordinate,
• t = the time.

1. Solution of the Burgers equation with nonzero viscosity

Let us look for a solution of Eq. (1) of the form of traveling wave [1], i.e.,
u(x, t) = w((x − x0 ) − st) ≡ w(y).
1
2 MARIA CAMERON

Then ut = −sw0 , ux = w0 , and uxx = w00 . Plugging this into Eq. (1) we obtain
−sw0 + ww0 = νw00
 2 0
0 w
−sw + = νw00
2
w2
−sw + = νw0 + C.
2
We also impose conditions at the ±∞: w(−∞) = uL , w(∞) = uR , where uL > uR , and
w0 (±∞) = 0. Then we have
u2L u2
−suL + = C = −suR + R .
2 2
Therefore, s must be (uL + uR )/2. Thus, the shock speed is the same as in the case of zero
viscosity. Hence C = −uL uR /2. Then we continue.
w2 uL + uR uL uR
νw0 = − w+
2 2 2
dy dw
= 2
2ν w − (uL + uR )w + uL uR
dy dw
= 2 (uL −uR )2
2ν w2 − (uL + uR )w + (uL +u R)
−
4 4
dy dw
=
uL +uR 2 2
2ν − (uL −uR )


w− 2 4

Integrating the both parts using

dw 1 w−a−b
Z
2 2
= log
(w − a) − b 2b w−a+b
we obtain
uL +uR uL −uR
y 1 w− 2 − 2 1 w − uL 1 uL − w
+C = log uL +uR uL −uR = log = log .
2ν uL − uR w− 2 + 2
uL − uR w − uR uL − uR w − uR
In the last equality we used the fact that uL > w > uR . Hence,
uL − w y(uL −uR )
= e 2ν +C
w − uR
uL − w = weA − uR eA , where A = y(uL − uR )/(2ν) + C
A A
w(e + 1) = (uL + uR e )
uL + uR eA uL − uR 2
w= A
= uR + .
e +1 2 eA + 1
 NOTES ON BURGERS’S EQUATION 3

Multiplying and dividing by exp(−A/2) and using the identity

2e−A/2 eA/2 − e−A/2 A
A/2 −A/2
= 1 − A/2 −A/2
= 1 − tanh
e +e e +e 2
we get
uR + uL uL − uR y(uL − uR )
 
w(y) = − tanh +C .
2 2 4ν
Hence,
uR + uL uL − uR ([x − x0 ] − st)(uL − uR )
 
(2) u(x, t) = − tanh .
2 2 4ν
The profiles w(y) for various values of ν are shown in Fig. 1. As ν → 0, u(x, t) tends to a
step function of the argument x − st.

1
ν=0.25
0.8 ν=0.1
ν=0.02
0.6

0.4

0.2

0
-2 -1.5 -1 -0.5 0 0.5 1 1.5 2

Figure 1. The profiles of the solution of the viscous Burgers equation for
uR = 0, uL = 1, x0 = 0, and ν equal to 0.25, 0.1, and 0.02. Note that w(y)
tends to a step function as ν → 0.

2. Shock speed
If the viscosity ν = 0, or neglected, Eq. (1) can be rewritten as
(3) ut + [ 21 u2 ]x = 0
Eq. (3) is easier to study theoretically and numerically than Eq. (1). From now on, unless
indicated otherwise, we will refer to Eq. (3) as the Burgers equation.
Equation (3) has a solution in the form of the traveling wave [2]
(4) u(x, t) = V (x − x0 − st),
where V (y) is a step function:
 
uL y < 0
(5) V (y) = ,
uR y > 0
4 MARIA CAMERON

where uL > uR . This wave is called the shock wave. s is the speed of propagation of the
shock wave. It can be obtained from the following reasoning. Let M be some large number.
Consider the integral
Z M
u(x, t)dx.
−M
Then

Figure 2. Finding the shock speed using the law of conservation of momentum.

M
d M
Z M
u2 u2 u2
Z
u(x, t)dx = −uux dx = − = L − R.
dt −M −M 2 −M 2 2
On the other hand,
Z M
u(x, t)dx = (M + st)uL + (M − st)uR .
−M
Therefore,
d M
Z
u(x, t)dx = s(uL − uR ).
dt −M
Hence, the speed of propagation of the wave is
 2
uL u2R

uL + uR
(6) s= − /(uL − uR ) = .
2 2 2
Remark The argument above is valid for a more general equation of the form
(7) ut + [f (u)]x = 0.
 NOTES ON BURGERS’S EQUATION 5

Such equations are called hyperbolic conservation laws. The shock speed is given by
f (uL ) − f (uR ) jump in f (u)
(8) s= = .
uL − uR jump in u
This equation is called the Rankine-Hugoniot condition.

3. Characteristics of the Burgers equation

The characteristics of Eq. (3) are given by
dx
(9) = u(x, t).
dt
Let us show that u is constant along the characteristics. Let (x(t), t) be a characteristic.
Then
d ∂u ∂u dx
u(x(t), t) = + = ut + uux = 0.
dt ∂t ∂x dt
Therefore, the solution of Eq. (9) is given by
(10) x(t) = u(x(0), 0)t + x(0) = u0 (x0 )t + x0 , where x0 = x(0), u0 (x) = u(x, 0).
Eq. (10) shows that
• the characteristics are straight lines,
• they may intersect,
• they do not necessarily cover the entire (x, t) space.
This is a new phenomenon in comparison with the linear first order equations ut +a(x, t)ux =
0. For a linear first order equation, there is a unique characteristic passing through every
point of the (x, t) space. Thus, its characteristics never intersect and cover the entire space.
Moreover, even for a smooth initial speed distribution u0 (x) the solution of the Burgers
equation may become discontinuous in a finite time. This happens when u00 (x) is negative
somewhere. Then the characteristics intersect, i.e., the wave breaks. Let us find the break
time. Consider two characteristics x(t) = u0 (x1 )t + x1 and x(t) = u0 (x2 )t + x2 . Then we
equate
x(t) = u0 (x1 )t + x1 = u0 (x2 )t + x2 .
Then the time at which they intersect is
x2 − x1
t=− .
u0 (x2 ) − u0 (x1 )
Therefore, the break time is
x2 − x1
 
1
Tb = min − =−  
x1 ,x2 ∈R u0 (x2 ) − u0 (x1 ) minx1 ,x2 ∈R u0 (x2 )−u0 (x1 )
x2 −x1
1 1
=−  =− .
minx∈R u00 (x)
 R x2
1
minx1 ,x2 ∈R x2 −x1 x1 u00 (x)dx
6 MARIA CAMERON

An example is shown in Figure 3. The initial data u0 (x) = exp(−16x2 ) and the cor-
recponding characteristics of the Burgers equation are shown in Fig. 3 (a) and (b) respec-
tively. The solution at times t = 0.5 and t = 0.8 obtained by the method of characteristics
is shown in Fig. 3 (c) and (e). The numerical solution computed by Godunov’s method
(see Section 6) is shown in Fig. 3 (d) and (f). The solution obtained by the method of
characteristis is triple-valied at some values of x and non-physical in the sense that it is
not the vanishing viscosity solution (see Section 1). In contrast, the solution computed by
Godunov’s method tends to the vanishing viscosity solution as we refine the mesh.

4. Weak solutions
We have seen in the previous section that a solution to the Burgers equation can become
discontinuous even if the initial data are smooth. Then the discontinuity travels with a
certain speed, the shock speed s, given by Eq. (6). In Section 2 we found s using the
underlying integral conservation law. However, at this point, we can only call the step
function given by Eqs (4), (5), and (6) a solution of the integral conservation law rather
than the solution of Eq. (3). In order to validate discontinuous solutions for differential
equations, the concept of the weak solutions was introduced (see e.g. [3]). This extension
of the concept of the solution must satisfy the following requirements:
• a smooth function is a weak solution iff it is a regular solution,
• a discontinuous function can be a weak solution,
• only those discontinuous functions which satisfy the associated integral equation
can be weak solutions.
Motivation. Let φ(x, t) be an infinitely smooth function with a compact support, i.e.,
it is different from zero only within some compact subset of the space (x, t) = R × [0, +∞).
Let u(x, t) be a smooth solution of a hyperbolic conservation law given by Eq. (7). Then
Z ∞Z ∞ Z ∞ ∞ Z ∞ Z ∞
(11) 0 = (ut + [f (u)]x )φdxdt = φudx − φt u + φx f (u)dxdt.
0 −∞ −∞ 0 0 −∞
Therefore, here is the definition [3].
Definition 1. u(x, t) is a weak solution of the conservation law ut + [f (u)]x = 0 if for any
infinitely differentiable function φ(x, t) with a compact support
Z ∞Z ∞ Z ∞ ∞ Z ∞
(12) φt u + φx f (u)dxdt = φudx =− φ(x, 0)u(x, 0)dx.
0 −∞ −∞ 0 −∞

Such a function φ(x, t) is called a test function.

5. The Riemann problem

In this section, we consider the following initial value problem for the Burgers equation:
 
uL x < 0
u(x, 0) = .
uR x > 0
This problem is called the Riemann problem. We will consider two cases.
 NOTES ON BURGERS’S EQUATION 7

(a) (b)

(c) (d)

(e) (f)

Figure 3. (a) The initial data u0 (x) = exp(−16x2 ). (b) The corresponding
characteristics of the Burgers equation.(c-d) The solution at time t = 0.5
obtained by the method of characteristics (c) and computed numerically
(by Godunov’s method) (d). (e-f) The solution at time t = 0.8 obtained by
the method of characteristics (e) and computed numerically (by Godunov’s
method) (f).
8 MARIA CAMERON

5.1. Case 1: uL > uR . In this case, the characteristics cover the entire (x, t) space but also
cross. Hence the construction of the solution using only the characteristics is ambiguous.
Let us show that in this case there exists a unique weak solution given by
 
uL x < st uL + uR
(13) u(x, t) = , where s = .
uR x > st 2
The characteristics for this solution are chown in Fig. 4(a).

(b)

Figure 4. (a) The characteristics for the shock wave. (b) Illustration for
the proof that the shock wave is the unique weak solution.

Proof. Let φ(x, t) be a test function. First suppose that support U lies entirely in one of
the sets {x < st} or {x > st}. Then since u(x, t) is constant in each of these sets, it satisfies
the Burgers equation on the support of φ. Then using Eq. (11) we conclude that Eq. (12)
holds.
Now suppose that the support U of φ is divided by the line x = st into two sets UL and
UR (Fig. 4(b)). Then we have
Z ∞Z ∞ Z Z Z Z
φt u + φx f (u)dx = φt u + φx f (u)dxdt + φt u + φx f (u)dxdt
0 −∞ UL UR

Applying the Green identity

ZZ Z
(14) (Px − Qt )dxdt = P dt + Qdx
D ∂D
 NOTES ON BURGERS’S EQUATION 9

and noting that u is constant within UL and within UR , hence (φu)t = φt u and (φf (u))x =
φx f (u), we continue
 2  Z  2 
uL uR
Z
= φ dt − uL dx + φ dt − uR dx
∂UL 2 ∂UR 2
 2   2  Z 0 Z ∞
uL uR
Z Z
= φ − uL dx − φ − uR dx − φuL dx − φuR dx
x=st 2s x=st 2s −∞ 0
 2 Z ∞
uL − u2R
Z 
= − (uL − uR ) φdx − φ(x, 0)u(x, 0)dx
x=st 2s −∞
The first integral in the last equality is zero for any test function φ iff s = (uL + uR )/2.
Thus, the solution given by Eq. (13) is the unique weak solution for the Riemann problem
in the case uL > uR . 
The discussion in Section 1 indicates that the solution given by Eq. (13) is the vanishing
viscosity solution, i.e. the limit of the solutions of Eq. (1) as ν → 0.

5.2. Case 2: uL < uR . Then the characteristics do not cross but do not cover the entire
space (x, t). There are many weak solutions. Two of them are shown in Fig. 5. The one
in Fig. 5(a) called the rarefaction fan or the transonic rarefaction is given by
 
 uL , x < uL t 
(15) u(x, t) = x/t, uL t ≤ x ≤ uR t .
uR , x > uR t
 

Another solution, shown in Fig. 5 (b), is called the rarefaction shock. Despite there are
many weak solutions, only one of them is “physical”, i.e., the vanishing viscosity solution.
It is the rarefaction fan solution. One can reject all of the nonphysical weak solutions
by analyzing Eq. (1). However, the analysis of the equation with nonzero viscosity is
harder than the analysis of the one with zero viscosity. Then an additional simpler-to-
verify condition, the so called entropy condition were introduced to eliminate nonphysical
weak solutions. There are several variations of the entropy condition. We will state only
the simplest one.
Definition 2. A discontinuity propagating with speed s given by Eq. (8) satisfies entropy
condition if f 0 (uL ) > s > f 0 (uR ).
For the Burgers equation this entropy condition reduces to the requirement that if a
discontinuity is propagating with speed s then uL > uR .

6. Numerical methods for hyperbolic conservation laws

Numerical solution of the Burgers equation is a challenging problem because perfectly
consistent and stable schemes might propagate discontinuities with wrong speeds and hence
fail to converge to the physically correct vanishing viscosity solution with the mesh re-
finement. In order to address this issue, additional requirements have been imposed on
 10 MARIA CAMERON

(a) (a)

Figure 5. The characteristics for the rarefaction fan (a) and the rarefaction
shock (b). Both of these are weak solutions. However, the rarefaction wave
is the physical vanishing viscosity solution, while the rarefaction shock is
not.

numerical schemes for conservation laws that guarantee that they propagate discontinuities
with the right speeds [2].
6.1. Conservative methods for nonlinear problems. A specific difficulty in comput-
ing discontinuous solutions of hyperbolic conservation laws can be illustrated by the fol-
lowing simple example.
Example 1 Consider the Burgers equation written in the quasi-linear
form
ut + uux = 0.
Let (
1, x < 0,
u(x, 0) = u0 (x) =
0, x ≥ 0.
A generalization of the left upwind method for the case where the speed
a = u gives
k
Ujn+1 = Ujn − Ujn (Ujn − Uj−1
n
).
h
The initial condition gives Uj0 = 1 for j < 0 and Uj0 = 0 for j ≥ 0. Then
(
1 1 − hk 1(1 − 1) = 1, j < 0,
Uj =
0 − hk 0(0 − Uj−1
0 ) = 0, j ≥ 0.
 NOTES ON BURGERS’S EQUATION 11

Hence, Uj1 = Uj0 . Therefore, Ujn = Uj0 for all j. This means that the method
propagates the discontinuity with a wrong speed s = 0.
In order to resolve this issue, let us recall what a conservation law is. A hyperbolic conser-
vation law
(16) ut + [f (u)]x = 0
means that the conserved quantity
Z R
u(x, t)dx
L

can only change due to the flux through the boundaries, i.e.
Z R Z R Z k Z k
u(x, t + k)dx − u(x, t)dx = f (u(L, t + τ ))dτ − f (u(R, t + τ ))dτ.
L L 0 0

Definition 3. A numerical method is in a conservation form if it can be rewritten in the

form
k
Ujn+1 = Ujn − n n n n

(17) F (Uj−p , . . . , Uj+q ) − F (Uj−p−1 , . . . , Uj+q−1 ) ,
h
for some function F which is called the numerical flux. A method that can be written in a
conservation form is called conservative.
6.2. Discrete conservation. The basic principle of a conservation law is that the con-
served quantity in a given interval [L, R] can change only due to the flux through the
boundaries, i.e.
Z R Z R Z t2 Z t2 
u(x, t2 )dx = u(x, t1 )dx − f (u(R, t))dt − f (u(L, t))dt .
L L t1 t1

A similar identity holds for conservative methods due to telescoping the sums. If we sum
a conservative method (17) from j = jL to j = jR we get
jR
X jR
X jR
X
h Ujn+1 =h Ujn −k [F (U n ; j) − F (U n ; j − 1)] .
j=jL j=jL j=jL

Here, for brevity, we have used the notation F (U n ; j) ≡ F (Uj−p

n , . . . , U n ). The sum
j+q
telescopes, and only boundary fluxes remain as a result:
jR
X jR
X
h Ujn+1 = h Ujn − k [F (U n ; jR ) − F (U n ; jL − 1)] .
j=jL j=jL

Therefore, the numerical solution, like the exact solution, allows the conserved quantity
h jj=j Ujn to change only due to the flux through the boundaries.
PR
L
12 MARIA CAMERON

6.3. Consistency. A method of the form (17) is consistent with the original conservation
law (16) if the numerical flux function F reduces to the true flux function f (the one in
the equation ut + [f (u)]x = 0) for the case of constant flow. I.e., if u(x, t) ≡ ū then
(18) F (ū, . . . , ū) = f (ū).
Furthermore, if the arguments of F approach some constant value ū, F should approach
f (ū) smoothly, i.e.
(19) lim F (v1 , . . . , vr ) = f (ū).
v1 ,...,vr →ū

It is sufficient to require the Lipschitz continuity of F in order to satisfy the smoothness

condition given by Eq. (19). Recall that F (v1 , . . . , vr ) is Lipschitz continuous if
(20) |F (v1 , . . . , vr ) − F (w1 , . . . , wr )| ≤ K max {|v1 − w1 |, . . . , |vr − wr |} ,
where K is a constant depending on F but not on its arguments called the Lipschitz
constant.
Therefore, sufficient consistency conditions are given by
(21) F (ū, . . . , ū) = f (ū),
(22) |F (v1 , . . . , vr ) − F (ū, . . . , ū)| ≤ K max {|v1 − ū|, . . . , |vr − ū|} .
6.4. Generalization of methods developed for the advection equation. Now we
generalize some methods developed for the advection equation ut + aux = 0 for hyper-
bolic conservation laws (16) making sure that these generalizations are conservative and
consistent.
Lax-Friedrichs
The generalization of the Lax-Friedrichs method to Eq. (16) takes the form
1 k
Ujn+1 = n n n n



(23) Uj−1 + Uj+1 − f (Uj+1 ) − f (Uj−1 ) .
2 2h
Observe that this method can be rewritten in the conservative form as
k
Ujn+1 = Ujn − n
, Ujn ) − F (Ujn , Uj−1
n

F (Uj+1 ) , where
h
h 1
(24) F (Uj+1 , Uj ) = (Uj − Uj+1 ) + (f (Uj ) + f (Uj+1 )) .
2k 2
Consistency check:
h 1
F (U, U ) = (U − U ) + (f (U ) + f (U )) = f (U ),
2k 2 
h h 1 1 
|F (V, W ) − F (U, U )| =  (V − U ) −
 (W − U ) + (f (V ) − f (U )) + (f (W ) − f (U ))
2k 2k 2 2
 
h
≤ + Kf max {|V − U |, |W − U |} ,
k
where Kf is the Lipschitz constant for the true flux f (we have assumed that f is Lipschitz-
continuous).
 NOTES ON BURGERS’S EQUATION 13

Lax-Wendroff
Recall that the Lax-Wendroff method for the advection equation ut + aux = 0 is given by

ak
a2 k 2
Ujn+1 = Ujn − n n n
− 2Ujn + Uj−1
n


Uj+1 − Uj−1 + 2
Uj+1 .
2h 2h

The most straightforward generalization requires the evaluation of the Jacobian f 0 (u)
(in the multidimensional case this is especially unpleasant). Two alternative extensions
were developed by Richtmyer and MacCormac. Both of them are two-step procedures.
Richtmyer:

n+1/2 1 k 
Ujn + Uj+1
n n
) − f (Ujn ) ,


Uj+1/2 = − f (Uj+1
2 2h
kh n+1/2 n+1/2
i
(25) Ujn+1 n
= Uj − f (Uj+1/2 ) − f (Uj−1/2 ) ,
h
MacCormack:
k
Uj∗ = Ujn − n
) − f (Ujn ) ,

f (Uj+1
h
1 k 
Ujn+1 Uj + Uj∗ −
n
f (Uj∗ ) − f (Uj−1
∗


(26) = ) ,
2 2h
Exercise Show that the methods (25) and (26) reduce to the Lax-Wendroff method for
f (u) ≡ au. Show that methods (25) and (26) are second-order accurate on smooth solu-
tions. Determine numerical flux functions for the methods (25) and (26) and show that
they are conservative.

6.5. Convergence. Lax and Wendroff have proven that a consistent and conservative
method converges to a weak solution of the conservation law almost everywhere as k, h → 0
and k/h satisfies stability conditions.
However, as we know from Section 5, a weak solution might be non-unique in the case
where uL < uR at the discontinuity. The entropy condition 2 allows us to reject all
physically irrelevant weak solutions and select the physically correct one, which is consistent
with the limit of the solutions of the viscous equation ut +[f (u)]x = uxx as  → 0. There is
a danger that the numerical solution of ut +[f (u)]x = 0 by a conservative method converges
to a weak but physically irrelevant solution. The following example demonstrates that a
seemingly reasonable method can fall into this trap.
Example 2 Consider the Burgers equation ut + [u2 /2]x = 0 with the
initial data
(
−1, x < 0,
u0 (x) =
1, x > 0.
14 MARIA CAMERON

The physically relevant solution to this problem (the vanishing viscosity

solution) is a transonic rarefaction given by

−1, x < −t,

u(x, t) = x/t, −t ≤ x ≤ t,

1, x > t.


The stationary discontinuity u(x, t) = u0 (x) is another weak solution (note,

the shock speed is zero: s = (uL + uR )/2 = (1 + (−1))/2 = 0).
Let us set the numerical initial velocity to
(
−1, j ≤ 0,
Uj0 =
1, j>0
and consider a conservative method obtained by a generalization of the
upwind scheme:
k
Ujn+1 = Ujn − n
, Ujn ) − F (Ujn , Uj−1
n

F (Uj+1 ) , where
( h
f (v), if (f (v) − f (w))/(v − w) ≥ 0,
F (v, w) = .
f (w), if (f (v) − f (w))/(v − w) < 0,

For the problem above, we have f (v) = v 2 /2 and

F (1, 1) = f (1) = 1/2, F (−1, −1) = f (−1) = 1/2, F (1, −1) = f (1) = 1/2.
Therefore,
k
Uj1 = Uj0 −[1/2 − 1/2] = Uj0 .
h
Hence Ujn = Uj0 for all n and j. Thus, the numerical solution will converge
to the physically irrelevant “stationary discontinuity” weak solution. On
the other hand, if we set

−1, j < 0,

0
Uj = 0, j = 0,

1, j > 0,


then the numerical solution by the same method will be the transonic rar-
efaction. Check this.
In order to outlaw conservative methods that might produce discrete approximations
to physically irrelevant weak solutions, we need a discrete analog of the entropy condition
2. Unfortunately, the simple version of the entropy condition given in Definition 2 is not
extended to the discrete case. Another, more involved version, requiring an introduction
of an entropy function, has a discrete analog. I will not discuss this topic here. Instead, I
refer curious students to Ref. [2], Sections 3.8.1, 12.5, and 13.4.
 NOTES ON BURGERS’S EQUATION 15

In the next section, we will discuss different type of methods for conservation laws for
which the version of the entropy condition involving an entropy function can be readily
verified (we will not verify it, if you wish, see Section 13.4 in [2]).

6.6. Godunov’s method. The idea of Godunov’s method (1959) is the following. Let
Ujn be a numerical solution on the n-th layer (i.e., at time tn = kn). Then we define a
function ũ(x, t) for tn ≤ t ≤ tn+1 as follows. At t = tn ,

ũ(x, tn ) = Ujn , xj − h/2 < x < xj + h/2, j = 2, . . . , n − 1.

Then ũ(x, t) is the solution of the collection of the Riemann problems on the interval
[tn , tn+1 ]. If the time step k is small enough so that the characteristics starting at the
points xj + h/2 do not intersect within this interval (i.e., the CFL condition is satisfied),
then ũ(x, tn+1 ) is determined unambiguously. Then the numerical solution on the next
layer, Ujn+1 is defined by averaging ũ(x, tn+1 ) over the intervals xj − h/2 < x < xj + h/2:
xj +h/2
1
Z
(27) Ujn+1 = ũ(tn+1 )dx.
h xj −h/2

This idea is illustrated in Fig. 6.

In practice, the cell averages (27) can be easily calculated using the integral form of
conservation law:
1 xj +h/2 1 xj +h/2
Z Z
ũ(x, tn+1 )dx = ũ(x, tn )dx
h xj −h/2 h xj −h/2
Z tn+1 Z tn+1 
1
− f (ũ(xj + h/2, t)) dt − f (ũ(xj − h/2, t)) dt .
h tn tn

Observing that ũ(xj +h/2, t) and ũ(xj −h/2, t) are constant over the time interval [tn , tn+1 ]
we obtain
k
Ujn+1 = Ujn − F (Ujn , Uj+1
n n
, Ujn ) ,

) − F (Uj−1
h
(28) where F (uL , uR ) = f (u∗ (uL , uR )) .

Therefore, Godunov’s method is conservative.

The analysis below will be done for a scalar conservation law, i.e., for the case where
is a scalar function. The value u∗ (uL , uR ) in the numerical flux function F (uL , uR ) =
f (u∗ (uL , uR )) is defined so that the entropy condition is satisfied and hence the weak
solution, to which the numerical solution converges, is the vanishing viscosity solution. If
f (u) is convex (if f is twice differentiable then f 00 (u) > 0), the following four cases must
be considered:
(1) f 0 (uL ) ≥ 0 and f 0 (uR ) ≥ 0. Then u∗ = uL .
(2) f 0 (uL ) ≤ 0 and f 0 (uR ) ≤ 0. Then u∗ = uR .
16 MARIA CAMERON

ũ(x, tn )

Ujn Ujn n
1 Uj+1

xj 1 xj 1
2
xj xj+ 12 xj+1
t
tn+1

tn
xj 1 xj 1 xj xj+ 12 xj+1
ũ(x, tn+1 ) 2

xj 1 xj 1
2
xj xj+ 12 xj+1

Figure 6. Top: The function ũ(tn ) obtained by setting its value in the
intervals xj − h/2 < x < xj + h/2 to Ujn . Middle: The diagram on the
(x, t)-plane showing lines of discontinuity and the critical characteristics. It
allows us to obtain ũ(tn+1 ). Bottom: ũ(tn+1 ) obtained using the diagram
on the (x, t)-plane.

(3) f 0 (uL ) ≥ 0 ≥ f 0 (uR ). Then

( f (uL )−f (uR )
∗ uL , if uL −uR > 0,
(29) u = f (uL )−f (uR )
uR , if uL −uR < 0.

(4) f 0 (uL ) < 0 < f 0 (uR ). Then u∗ = us (transonic rarefaction), where the value us is
such that f 0 (us ) = 0. It is called the sonic point. For example, for the Burgers
equation ut + [u2 /2]x = 0, us = 0.
In the first three cases, the value u∗ is either uL and uR , and it can be simply determined
by Eq. (29). Note that in Cases 1 and 2, u∗ is the same whether the physically correct
weak solution to the Riemann problem is a shock wave or a rarefaction. Only in Case 4,
the transonic rarefaction, the value of u∗ differs from the one determined by Eq. (29). This
is the value of u for which the characteristic speed is zero.
 NOTES ON BURGERS’S EQUATION 17

Note that the numerical flux determined by Cases 1 - 4 can be rewritten more compactly
as
(
minuL ≤u≤uR f (u), if uL ≤ uR ,
(30) F (uL , uR ) =
maxuL ≤u≤uR f (u), if uL > uR .
It was proven that the numerical flux given by Eq. (30) gives the physically correct flux
for scalar conservation laws even if f (u) is non-convex.
If you perform some numerical experiments computing discontinuous solutions using
Godunov’s method, you observe that Godunov’s method adds some artificial smearing to
the solution. Glimm’s method, considered in the next Section, is designed to preserve sharp
discontinuities.
6.7. Glimm’s method. Like Godunov’s method, Glimm’s method is also based on solv-
ing a collection of Riemann problems on the interval [tn , tn+1 ]. But instead of averaging
procedure for getting the numerical solution on the next layer a random choice procedure
is used. Glimm’s time step proceeds in two stages
n+1/2
Uj+1/2 = ξ(Ujn , Uj+1
n
),
n+1/2 n+1/2
Ujn+1 = ξ(Uj−1/2 , Uj+1/2 ),
where ξ(uL , uR ) is a random variable taking values either uL or uR with probabilities
proportional to lengths of the corresponding intervals ( Fig. 7 ):
( )
uL , p = h/2+sk/2
h
ξ(uL , uR ) = ,
uR , p = h/2−sk/2
h
where s = [f (uR ) − f (uL )]/(uR − uL ) is the shock speed. We can rewrite the definition of
ξ as follows.
uL , p = 21 (1 + sλ)
 
ξ(uL , uR ) = ,
uR , p = 12 (1 − sλ)
where λ = k/h.

References
[1] G. I. Barenblatt, Scaling, self-similarity, and intermediate asymptotics, Cambridge University Press,
1996
[2] R. J. LeVeque, Numerical Methods for Conservation Laws, Birkhauser, Basel, Boston, Berlin, 1992
[3] Lawrence C. Evans, Partial Differential Equations, AMS, Providence, RI, 1998
18 MARIA CAMERON

tn+1

h sk
2 2
tn+ 12
h sk
+
2 2

tn
xj 1 xj 1
2
xj xj+ 12 xj+1

Figure 7. An illustration for Glimm’s method.