Lecture Notes on Hyperbolic Conservation Laws

Alberto Bressan
Department of Mathematics, Penn State University,
University Park, Pa. 16802, USA.
[email protected]

May 21, 2009

Abstract

These notes provide an introduction to the theory of hyperbolic systems of conser-

vation laws in one space dimension. The various chapters cover the following topics:
1. Meaning of the conservation equations and definition of weak solutions. 2. Shocks,
Rankine-Hugoniot equations and admissibility conditions. 3. Genuinely nonlinear and
linearly degenerate characteristic fields. Solution to the Riemann problem. Wave interac-
tion estimates. 4. Weak solutions to the Cauchy problem, with initial data having small
total variation. Proof of global existence via front-tracking approximations. 5. Continu-
ous dependence of solutions w.r.t. the initial data, in the L1 distance. 6. Uniqueness of
entropy-admissible weak solutions. 7. Approximate solutions constructed by the Glimm
scheme. 8. Vanishing viscosity approximations. 9. Counter-examples to global existence,
uniqueness, and continuous dependence of solutions, when some key hypotheses are re-
moved. The survey is concluded with an Appendix, reviewing some basic analytical tools
used in the previous chapters.

1 Introduction

A scalar conservation law in one space dimension is a first order partial differential equation
of the form
ut + f (u)x = 0 . (1.1)
Here u is called the conserved quantity while f is the flux. Equations of this type often describe
transport phenomena. Integrating (1.1) over a given interval [a, b] one obtains
Z b Z b Z b
d
u(t, x) dx = ut (t, x) dx = − f (u(t, x))x dx
dt a a a

= f (u(t, a)) − f (u(t, b)) = [inflow at a] − [outflow at b] .

In other words, the quantity u is neither created nor destroyed: the total amount of u contained
inside any given interval [a, b] can change only due to the flow of u across boundary points
(fig. 1).
Using the chain rule, (1.1) can be written in the quasilinear form
ut + a(u)ux = 0, (1.2)

1
 u

a b ξ x

Figure 1: Flow across two points.

where a = f 0 is the derivative of f . For smooth solutions, the two equations (1.1) and (1.2)
are entirely equivalent. However, if u has a jump at a point ξ, the left hand side of (1.2)
will contain the product of a discontinuous function a(u) with the distributional derivative
ux , which in this case contains a Dirac mass at the point ξ. In general, such a product is
not well defined. Hence (1.2) is meaningful only within a class of continuous functions. On
the other hand, working with the equation in divergence form (1.1) allows us to consider
discontinuous solutions as well, interpreted in distributional sense. More precisely, a locally
integrable function u = u(t, x) is a weak solution of (1.1) provided that
ZZ
{uφt + f (u)φx } dxdt = 0 (1.3)

for every differentiable function with compact support φ ∈ Cc1 .

Example 1 (Traffic flow). Let u(t, x) be the density of cars on a highway, at the point x
at time t. For example, u may be the number of cars per kilometer. In first approximation,
we shall assume that u is continuous and that the speed s of the cars depends only on their
density, say
ds
s = s(u), with < 0.
du
Given any two points a, b on the highway, the number of cars between a and b therefore varies
according to the law
Z b Z b
d
ut (t, x) dx = u(t, x) dx = [inflow at a] − [outflow at b]
a dt a
(1.4)
Z b
= s(u(t, a)) · u(t, a) − s(u(t, b)) · u(t, b) = − [s(u) u]x dx .
a

Since (1.4) holds for all a, b, this leads to the conservation law

ut + [s(u) u]x = 0,

where u is the conserved quantity and f (u) = s(u)u is the flux function.

2
1.1 Strictly hyperbolic systems

The main object of our study will be the n × n system of conservation laws



∂ ∂

 u1 + [f1 (u1 , . . . , un )] = 0,



∂t ∂x


· · · (1.5)







 ∂ un + ∂ [fn (u1 , . . . , un )]

= 0.
∂t ∂x
For simplicity, this will still be written in the form (1.1), but keeping in mind that now
u = (u1 , . . . , un ) is a vector in IRn and that f = (f1 , . . . , fn ) is a map from IRn into IRn .
.
Calling A(u) = Df (u) the n × n Jacobian matrix of the map f at the point u, the system
(1.5) can be written in the quasilinear form

ut + A(u)ux = 0. (1.6)

We say that the above system is strictly hyperbolic if every matrix A(u) has n real, distinct
eigenvalues: λ1 (u) < · · · < λn (u). In this case, one can find bases of left and right eigenvectors
of A(u), denoted by l1 (u), . . . , ln (u) and r1 (u), . . . , rn (u), with
½
1 if i = j,
li (u) · rj (u) =
0 if i =
6 j.

Example 2 (Gas Dynamics). The Euler equations describing the evolution of a non viscous
gas take the form


 ρt + (ρv)x = 0 (conservation of mass)
(ρv)t + (ρv 2 + p)x = 0 (conservation of momentum)

 (ρE) + (ρEv + pv)
t x =0 (conservation of energy)

Here ρ is the mass density, v is the velocity while E = e + v 2 /2 is the energy density per
unit mass. The system is closed by a constitutive relation of the form p = p(ρ, e), giving the
pressure as a function of the density and the internal energy. The particular form of p depends
on the gas under consideration.

1.2 Linear systems

We consider here two elementary cases where the solution of the Cauchy problem can be
written explicitly.

The linear homogeneous scalar Cauchy problem with constant coefficients has the form

ut + λux = 0, u(0, x) = ū(x), (1.7)

with λ ∈ IR. If ū ∈ C 1 , one easily checks that the travelling wave

u(t, x) = ū(x − λt) (1.8)

3
provides a classical solution to (1.7). In the case where the initial condition ū is not differen-
tiable and we only have ū ∈ L1loc , the function u defined by (1.8) can still be interpreted as a
solution, in distributional sense.

Next, consider the homogeneous system with constant coefficients

ut + Aux = 0, u(0, x) = ū(x), (1.9)

where A is a n × n hyperbolic matrix, with real eigenvalues λ1 < · · · < λn and right and left
eigenvectors ri , li , chosen so that
½
1 if i = j,
li · rj =
0 if i =
6 j,
.
Call ui = li · u the coordinates of a vector u ∈ IRn w.r.t. the basis of right eigenvectors
{r1 , · · · , rn }. Multiplying (1.9) on the left by l1 , . . . , ln we obtain
.
(ui )t + λi (ui )x = (li u)t + λi (li u)x = li ut + li Aux = 0, ui (0, x) = li ū(x) = ūi (x).

Therefore, (1.9) decouples into n scalar Cauchy problems, which can be solved separately in
the same way as (1.7). The function
n
X
u(t, x) = ūi (x − λi t) ri (1.10)
i=1

now provides the explicit solution to (1.9), because

n
X
ut (t, x) = −λi (li · ūx (x − λi t))ri = − Aux (t, x).
i=1

Observe that in the scalar case (1.7) the initial profile is shifted with constant speed λ. For
the system (1.9), the initial profile is decomposed as a sum of n waves, each travelling with
one of the characteristic speeds λ1 , . . . , λn .
As a special case, consider the Riemann initial data
½
ū(x) = u− if x < 0,
u+ if x > 0.
The corresponding solution (1.10) can then be obtained as follows.
Write the vector u+ − u− as a linear combination of eigenvectors of A, i.e.
n
X
u+ − u− = cj rj .
j=1

Define the intermediate states

. X
ωi = u− + cj rj , i = 0, . . . , n,
j≤i

4
 t x/t=λ
2

x/t= λ x/t=λ
1
3
ω
ω 2
1

ω = u− ω3 = u+
0

0 x

Figure 2: Solution of the Riemann problem for a linear system.

so that each difference ωi − ωi−1 is an i-eigenvector of A. The solution then takes the form
(fig. 2): 

 ω0 = u− for x/t < λ1 ,



 ...
u(t, x) = ωi for λi < x/t < λi+1 , (1.11)



 ...


ωn = u+ for x/t > λn .

u(T)
u(t)
u(0)

x
Figure 3: If the wave propagation speed depends on u, the profile of the solution changes in
time, eventually leading to shock formation.

In the general nonlinear case (1.6) where the matrix A depends on the state u, new features
will appear in the solutions.

(i) Since the eigenvalues λi now depend on u, the shape of the various components in the
solution will vary in time (fig. 3). Rarefaction waves will decay, and compression waves will
become steeper, possibly leading to shock formation in finite time.

(ii) Since the eigenvectors ri also depend on u, nontrivial interactions between different waves
will occur (fig. 4). The strength of the interacting waves may change, and new waves of

5
different families can be created, as a result of the interaction.

linear nonlinear
t

Figure 4: Left: for the linear hyperbolic system (1.9), the solution is a simple superposition
of traveling waves. Right: For the non-linear system (1.5), waves of different families have
nontrivial interactions.

The strong nonlinearity of the equations and the lack of regularity of solutions, also due to
the absence of second order terms that could provide a smoothing effect, account for most
of the difficulties encountered in a rigorous mathematical analysis of the system (1.1). It
is well known that the main techniques of abstract functional analysis do not apply in this
context. Solutions cannot be represented as fixed points of continuous transformations, or in
variational form, as critical points of suitable functionals. Dealing with vector valued functions,
comparison principles based on upper or lower solutions cannot be used. Moreover, the theory
of accretive operators and contractive nonlinear semigroups works well in the scalar case [20],
but does not apply to systems. For the above reasons, the theory of hyperbolic conservation
laws has largely developed by ad hoc methods, along two main lines.

1. The BV setting, considered by J. Glimm [27]. Solutions are here constructed within a
space of functions with bounded variation, controlling the BV norm by a wave interaction
functional.

2. The L∞ setting, considered by L. Tartar and R. DiPerna [24], based on weak convergence
and a compensated compactness argument.

Both approaches yield results on the global existence of weak solutions. However, the method
of compensated compactness appears to be suitable only for 2×2 systems. Moreover, it is only
in the BV setting that the well-posedness of the Cauchy problem could recently be proved, as
well as the stability and convergence of vanishing viscosity approximations. Throughout the
following we thus restrict ourselves to the study of BV solutions, referring to [24] or [43, 48]
for the alternative approach based on compensated compactness.

1.3 Loss of regularity

A basic feature of nonlinear systems of the form (1.1) is that, even for smooth initial data,
the solution of the Cauchy problem may develop discontinuities in finite time. To achieve a

6
global existence result, it is thus essential to work within a class of discontinuous functions,
interpreting the equations (1.1) in their distributional sense (1.3).

Figure 5: At time T when characteristics start to intersect, a shock is produced.

Example 3. Consider the scalar conservation law (inviscid Burgers’ equation)

à !
u2
ut + = 0 (1.12)
2 x
with initial condition
1
u(0, x) = ū(x) = .
1 + x2
For t > 0 small the solution can be found by the method of characteristics. Indeed, if u is
smooth, (1.12) is equivalent to
ut + uux = 0. (1.13)
By (1.13) the directional derivative of the function u = u(t, x) along the vector (1, u) vanishes.
Therefore, u must be constant along the characteristic lines in the t-x plane:
µ ¶
t
t 7→ (t, x + tū(x)) = t, x + .
1 + x2
. √
For t < T = 8/ 27, these lines do not intersect (fig. 5). The solution to our Cauchy problem
is thus given implicitly by µ ¶
t 1
u t, x + 2
= . (1.14)
1+x 1 + x2
√
On the other hand, when t > 8/ 27, the characteristic lines start to intersect. As a result,
the map
t
x 7→ x +
1 + x2
is not one-to-one and (1.14) no longer defines a single valued solution of our Cauchy problem.
An alternative point of view is the following (fig. 3). As time increases, points on the graph of
u(t, ·) move horizontally with speed u, equal to their distance from the x-axis. This
√ determines
.
a change in the profile of the solution. As t approaches the critical time T = 8/ 27, one has
n o
lim inf ux (t, x) = −∞,
t→T − x∈IR

7
and no classical solution exists beyond time T . The solution can be prolonged for all times
t ≥ 0 only within a class discontinuous functions.

2 Weak Solutions

2.1 Basic Definitions

A basic feature of nonlinear hyperbolic systems is the possible loss of regularity: solutions
which are initially smooth may become discontinuous within finite time. In order to construct
solutions globally in time, we are thus forced to work in a space of discontinuous functions,
and interpret the conservation equations in a distributional sense.

Definition 2.1. Let f : IRn 7→ IRn be a smooth vector field. A measurable function u =
u(t, x), defined on an open set Ω ⊆ IR × IR and with values in IRn , is a distributional solution
of the system of conservation laws
ut + f (u)x = 0 (2.1)
if, for every C 1 function φ : Ω 7→ IRn with compact support, one has
ZZ
{u φt + f (u) φx } dxdt = 0. (2.2)
Ω

Observe that no continuity assumption is made on u. We only require u and f (u) to be locally
integrable in Ω. Notice also that weak solutions are defined up to L1 equivalence. A solution
is not affected by changing its values on a set of measure zero in the t-x plane.

An easy consequence of the above definition is the closure of the set of solutions w.r.t. con-
vergence in L1loc .

Lemma 1. Let (um )m≥1 be a uniformly bounded sequence of distributional solutions of (2.1).
If um → u in L1loc , then the limit function u is itself a distributional solution.

Indeed, the assumption of uniform boundedness implies f (um ) → f (u) in L1loc . For every
φ ∈ Cc1 we now have
ZZ ZZ
{u φt + f (u) φx } dxdt = lim {um φt + f (um ) φx } dxdt = 0.
Ω m→∞ Ω

In the following, we shall be mainly interested in solutions defined on a strip [0, T ] × IR, with
an assigned initial condition
u(0, x) = ū(x). (2.3)

8
Here ū ∈ L1loc (IR). To treat the initial value problem, it is convenient to require some additional
regularity w.r.t. time.

Definition 2.2. A function u : [0, T ] × IR 7→ IRn is a weak solution of the Cauchy problem
(2.1), (2.3) if u is continuous as a function from [0, T ] into L1loc , the initial condition (2.3)
holds and the restriction of u to the open strip ]0, T [ ×IR is a distributional solution of (2.1).

2.2 Rankine-Hugoniot conditions

Let u be a weak solution of (2.1). If u is continuously differentiable restricted to an open

domain Ω0 , then at every point (t, x) ∈ Ω0 , the function u must satisfy the quasilinear system
ut + A(u)ux = 0 , (2.4)
.
with A(u) = Df (u). Indeed, from (2.2) an integration by parts yields
ZZ
[ut + A(u)ux ] φ dxdt = 0.

Since this holds for every φ ∈ Cc1 (Ω0 ), the identity (2.4) follows.

Next, we look at a discontinuous solution and derive some conditions which must be satisfied
at points of jump. Consider first the simple case of a piecewise constant function, say
½ +
u if x > λt,
U (t, x) = (2.5)
u− if x < λt,
for some u− , u+ ∈ IRn , λ ∈ IR.

x
x=λt

n−
Ω+

n+

t
Ω−

Figure 6: Deriving the Rankine-Hugoniot equations.

Lemma 2. If the function U in (2.5) is a weak solution of the system of conservation laws
(2.1), then
λ (u+ − u− ) = f (u+ ) − f (u− ). (2.6)

Proof. Let φ = φ(t, x) be any continuously differentiable function with compact support. Let
Ω be a ball containing the support of φ and consider the two domains
. .
Ω+ = Ω ∩ {x > λt}, Ω− = Ω ∩ {x < λt} ,

9
 .
as in fig. 6. Introducing the vector field v = (U φ, f (U )φ), the identity (2.2) can be rewritten
as ZZ
div v dxdt = 0. (2.7)
Ω+ ∪Ω−

We now apply the divergence theorem separately on the two domains Ω+ , Ω− . Call n+ , n−
the outer unit normals to Ω+ , Ω− , respectively. Observe that φ = 0 on the boundary ∂Ω.
Denoting by ds the differential of the arc-length, along the line x = λt we have

n+ ds = (λ, − 1) dt n− ds = (−λ, 1) dt ,
ZZ Z Z
+
0 = div v dxdt = n · v ds + n− · v ds
Ω+ ∪Ω− ∂Ω+ ∂Ω−
Z Z
= [λu+ − f (u+ )] φ(t, λt) dt + [ − λu− + f (u− )] φ(t, λt) dt .

Therefore, the identity

Z
[λ(u+ − u− ) − f (u+ ) + f (u− )] φ(t, λt) dt = 0

must hold for every function φ ∈ Cc1 . This implies (2.6).

The vector equations (2.6) are the famous Rankine-Hugoniot conditions. They form a set
of n scalar equations relating the right and left states u+ , u− ∈ IRn and the speed λ of
the discontinuity. An alternative way of writing these conditions is as follows. Denote by
A(u) = Df (u) the n × n Jacobian matrix of f at u. For any u, v ∈ IRn , define the averaged
matrix Z
. 1
A(u, v) = A(θu + (1 − θ)v) dθ (2.8)
0
and call λi (u, v), i = 1, . . . , n, its eigenvalues. We can then write (2.6) in the equivalent form
Z 1
+ − + −
λ (u −u ) = f (u )−f (u ) = Df (θu+ +(1−θ)u− )·(u+ −u− ) dθ = A(u+ , u− )·(u+ −u− ).
0
(2.9)
In other words, the Rankine-Hugoniot conditions hold if and only if the jump − is an u+ u−
eigenvector of the averaged matrix A(u+ , u− ) and the speed λ coincides with the corresponding
eigenvalue.

Remark 1. In the scalar case, one arbitrarily assign the left and right states u− , u+ ∈ IR and
determine the shock speed as
f (u+ ) − f (u− )
λ = . (2.10)
u+ − u−
Geometrically, this means that the shock speed is the slope of the secant line through the
points (u− , f (u− )) and (u+ , f (u+ )) on the graph of the flux function f .

We now consider a more general solution u = u(t, x) of (2.1) and show that the Rankine-
Hugoniot equations are still satisfied at every point (τ, ξ) where u has an approximate jump,
in the following sense.

10
Definition 2.3. We say that a function u = u(t, x) has an approximate jump discontinuity
at the point (τ, ξ) if there exists vectors u+ 6= u− and a speed λ such that, defining U as in
(2.5), there holds
Z r Z r
1
lim |u(τ + t, ξ + x) − U (t, x)| dxdt = 0. (2.11)
r→0+ r 2 −r −r

Moreover, we say that u is approximately continuous at the point (τ, ξ) if the above relations
hold with u+ = u− (and λ arbitrary).

Observe that the above definitions depend only on the L1 equivalence class of u. Indeed,
the limit in (2.11) is unaffected if the values of u are changed on a set N ⊂ IR2 of Lebesgue
measure zero.

Example 4. Let g − , g + : IR2 7→ IRn be any two continuous functions and let x = γ(t) be a
. d
smooth curve, with derivative γ̇(t) = dt γ(t). Define the function
½ −
. g (t, x) if x < γ(t),
u(t, x) =
g + (t, x) if x > γ(t).
. .
At a point (τ, ξ), with ξ = γ(τ ), call u− = g − (τ, ξ), u+ = g + (τ, ξ). If u+ = u− , then u is
continuous at (τ, ξ), hence also approximately continuous. On the other hand, if u+ 6= u− ,
then u has an approximate jump at (τ, ξ). Indeed, the limit (2.11) holds with λ = γ̇(τ ) and
U as in (2.5).

We now prove the Rankine-Hugoniot conditions in the more general case of a point of approx-
imate jump.

Theorem 1. Let u be a bounded distributional solution of (2.1) having an approximate jump

at a point (τ, ξ). In other words, assume that (2.11) holds, for some states u− , u+ and a speed
λ, with U as in (2.5). Then the Rankine-Hugoniot equations (2.6) hold.

Proof. For any given θ > 0, one easily checks that the rescaled function
.
uθ (t, x) = u(τ + θt, ξ + θx)

is also a solution to the system of conservation laws. We claim that, as θ → 0, the convergence
uθ → U holds in L1loc (IR2 ; IRn ). Indeed, for any R > 0 one has
Z R Z R Z θR Z θR
1
lim |uθ (t, x) − U (t, x)| dxdt = lim |u(τ + t, ξ + x) − U (t, x)| dxdt = 0
θ→0 −R −R θ→0 θ 2 −θR −θR

because of (2.11). Lemma 1 now implies that U itself is a distributional solution of (2.1),
hence by Lemma 2 the Rankine-Hugoniot equations (2.6) hold.

11
2.3 Admissibility conditions

To motivate the following discussion, we first observe that the concept of weak solution is
usually not stringent enough to achieve uniqueness for a Cauchy problem. In some cases,
infinitely many weak solutions can be found.

Example 5. For Burgers’ equation

ut + (u2 /2)x = 0 , (2.12)
consider the Cauchy problem with initial data
½
1 if x ≥ 0,
u(0, x) =
0 if x < 0.
For every 0 < α < 1, a weak solution is

0 if x < αt/2,
uα (t, x) = α if αt/2 ≤ x < (1 + α)t/2, (2.13)

1 if x ≥ (1 + α)t/2 .
Indeed, the piecewise constant function uα trivially satisfies the equation outside the jumps.
Moreover, the Rankine-Hugoniot conditions hold along the two lines of discontinuity {x =
αt/2} and {x = (1 + α)t/2}.

From the previous example it is clear that, in order to achieve the uniqueness of solutions and
their continuous dependence on the initial data, the notion of weak solution must be supple-
mented with further “admissibility conditions”, possibly motivated by physical considerations.
Some of these conditions will be presently discussed.

Admissibility Condition 1 (Vanishing viscosity). A weak solution u of (2.1) is admissible

in the vanishing viscosity sense if there exists a sequence of smooth solutions uε to
uεt + A(uε )uεx = εuεxx (2.14)
which converge to u in L1loc as ε → 0+ .

Unfortunately, it is very difficult to provide uniform estimates on solutions to the parabolic

system (2.14) and characterize the corresponding limits as ε → 0+. From the above condition,
however, one can deduce other conditions which can be more easily verified in practice. A
standard approach relies on the concept of entropy.

Definition 2.4. A continuously differentiable function η : IRn → 7 IR is called an entropy for

n
the system of conservation laws (2.1), with entropy flux q : IR → 7 IR, if for all u ∈ IRn there
holds
Dη(u) · Df (u) = Dq(u). (2.15)

For n × n systems, (2.15) can be regarded as a first order system of n equations for the two
scalar variables η, q. When n ≥ 3, this system is overdetermined. In general, one should

12
thus expect to find solutions only in the case n ≤ 2. However, there are important physical
examples of larger systems which admit a nontrivial entropy function.
An immediate consequence of (2.15) is that, if u = u(t, x) is a C 1 solution of (2.1), then

η(u)t + q(u)x = 0. (2.16)

Indeed,

η(u)t + q(u)x = Dη(u)ut + Dq(u)ux = Dη(u)( − Df (u)ux ) + Dq(u)ux = 0.

In other words, for a smooth solution u, not only the quantities u1 , . . . , un are conserved but
the additional conservation law (2.16) holds as well. However one should be aware that, when
u is discontinuous, the quantity η(u) may not be conserved.
Example 6 Consider Burgers’ equation (2.12). Here the flux is f (u) = u2 /2. Taking η(u) = u3
and q(u) = (3/4)u4 , one checks that the equation (2.15) is satisfied. Hence η is an entropy
and q is the corresponding entropy flux. We observe that the function
½
1 if x < t/2,
u(0, x) =
0 if x ≥ t/2

is a (discontinuous) weak solution of (2.12). However, it does not satisfy (2.16) in distribution
sense. Indeed, calling u− = 1, u+ = 0 the left and right states, and λ = 1/2 the speed of the
shock, one has h i
q(u+ ) − q(u− ) 6= λ η(u+ ) − η(u− ) .

We now study how a convex entropy behaves in the presence of a small diffusion term. Assume
η, q ∈ C 2 , with η convex. Multiplying both sides of (2.14) on the left by Dη(uε ) one finds
n o
[η(uε )]t + [q(uε )]x = εDη(uε )uεxx = ε [η(uε )]xx − D2 η(uε ) · (uεx ⊗ uεx ) . (2.17)

Observe that the last term in (2.17) satisfies

n
X ∂ 2 η(uε ) ∂uεi ∂uεj
D2 η(uε )(uεx ⊗ uεx ) = · ≥ 0,
i,j=1
∂ui ∂uj ∂x ∂x

because η is convex, hence its second derivative at any point uε is a positive semidefinite
quadratic form. Multiplying (2.17) by a nonnegative smooth function ϕ with compact support
and integrating by parts, we thus have
ZZ ZZ
ε ε
{η(u )ϕt + q(u )ϕx } dxdt ≥ − ε η(uε )ϕxx dxdt.

If uε → u in L1 as ε → 0, the previous inequality yields

ZZ
{η(u)ϕt + q(u)ϕx } dxdt ≥ 0 (2.18)

whenever ϕ ∈ Cc1 , ϕ ≥ 0. The above can be restated by saying that η(u)t + q(u)x ≤ 0 in
distribution sense. The previous analysis leads to:

13
Admissibility Condition 2 (Entropy inequality). A weak solution u of (1.1) is entropy-
admissible if
η(u)t + q(u)x ≤ 0 (2.19)
in the sense of distributions, for every pair (η, q), where η is a convex entropy for (2.1) and q
is the corresponding entropy flux.

For the piecewise constant function U in (2.5), an application of the divergence theorem shows
that η(U )t + q(U )x ≤ 0 in distribution if and only if

λ [η(u+ ) − η(u− )] ≥ q(u+ ) − q(u− ). (2.20)

More generally, let u = u(t, x) be a bounded function which satisfies the conservation law
(2.1). Assume that u has an approximate jump at (τ, ξ), so that (2.11) holds with U as in
(2.5). Then, by the rescaling argument used in the proof of Theorem 1, one can show that
the inequality (2.20) must again hold.

Of course, the above admissibility condition can be useful only if some nontrivial convex
entropy for the system (2.1) is known. In the scalar case where u ∈ IR, convex entropies are
easy to construct. In particular, for each k ∈ IR, consider the functions

η(u) = |u − k|, q(u) = sgn(u − k) · (f (u) − f (k)).

It is easily checked that η, q are locally Lipschitz continuous and satisfy (2.15) at every u 6= k.
Although η, q ∈ / C 1 , we can still regard η as a convex entropy for (2.1), with entropy flux q.
Following Volpert [53], we say that a bounded measurable function u is an entropic solution
of (2.1) if ZZn o
|u − k|ϕt + sgn(u − k)(f (u) − f (k))ϕx dxdt ≥ 0 (2.21)

for every constant k ∈ IR and every C 1 function ϕ ≥ 0 with compact support.

Based on (2.21), one can show that a shock connecting the left and right states u− , u+ is
admissible if and only if
f (u∗ ) − f (u− ) f (u+ ) − f (u∗ )
≥ (2.22)
u∗ − u− u+ − u∗
for every u∗ = αu+ + (1 − α)u− , with 0 < α < 1. The above inequality can be interpreted
as a stability condition. Indeed, let u∗ ∈ [u− , u+ ] be an intermediate state and consider a
slightly perturbed solution (fig. 7), where the shock (u− , u+ ) is decomposed as two separate
jumps, (u− , u∗ ) and (u∗ , u+ ), say located at γ − (t) < γ + (t) respectively. By the the Rankine-
Hugoniot conditions (2.10), the two sides of (2.22) yield precisely the speeds of these jumps.
If the inequality holds, then γ̇ − ≥ γ̇ + , so that the backward shock travels at least as fast as
the forward one. Therefore, the two shocks will not split apart as time increases, and the
perturbed solution will remain close to the original solution possessing a single shock.
Observing that the Rankine-Hugoniot speed of a shock in (1.19) is given by the slope of the
secant line to the graph of f through the points u− , u+ , the condition (1.33) holds if and only
if for every α ∈ [0, 1] one has
(
f (αu+ + (1 − α)u− ) ≥ αf (u+ ) + (1 − α)f (u− ) if u− < u+ ,
(2.23)
f (αu+ + (1 − α)u− ) ≤ αf (u+ ) + (1 − α)f (u− ) if u− > u+ .

14
 u+ _
u

u*
u*
_
u u+
x x

Figure 7: The solution is stable if the speed of the shock behind is greater or equal than the
speed of the one ahead.

In other words, when u− < u+ the graph of f should remain above the secant line (fig. 8,
left). When u− > u+ , the graph of f should remain below the secant line (fig. 8, right).

f f

u - u* u+ u+ u* u-

Figure 8: Left: an admissible upward shock. Right: an admissible downward shock.

Remark 2. One can show that the inequality (2.22) holds for every intermediate state
u∗ = αu+ + (1 − α)u− if and only if

f (u∗ ) − f (u− ) f (u+ ) − f (u− )

≥ . (2.24)
u∗ − u− u+ − u−
In other words, the speed of the shock joining u− with u+ must be less or equal to the speed of
every smaller shock, joining u− with an intermediate state u∗ ∈ [u− , u+ ]. While the condition
(2.22) is meaningful only in the scalar case, the equivalent condition (2.24) admits a natural
extension to the vector-valued case. This leads to the Liu admissibility condition, which will
be discussed in Chapter 3.

x = γ (t) x = αt / 2

Figure 9: Left: a shock satisfying the Lax condition. Characteristics run toward the shock,
from both sides. Right: the two shocks in the weak solution (2.13) violate this condition.

An alternative admissibility condition, due to P. Lax, is particularly useful because it can be

applied to any system and has a simple geometrical interpretation. We recall that a function

15
u has an approximate jump at a point (τ, ξ) if (2.11) holds, for some U defined as in (2.5). In
this case, by Theorem 1 the right and left states u− , u+ and the speed λ of the jump satisfy the
Rankine-Hugoniot equations. In particular, λ must be an eigenvalue of the averaged matrix
A(u− , u+ ) defined at (2.8), i.e. λ = λi (u− , u+ ) for some i ∈ {1, . . . , n}.

Admissibility Condition 3 (Lax condition). A solution u = u(t, x) of (2.1) satisfies the

Lax admissibility condition if, at each point (τ, ξ) of approximate jump, the left and right
states u− , u+ and the speed λ = λi (u− , u+ ) of the jump satisfy

λi (u− ) ≥ λ ≥ λi (u+ ). (2.25)

To appreciate the geometric meaning of this condition, consider a piecewise smooth solution,
having a discontinuity along the line x = γ(t), where the solution jumps from a left state u−
.
to a right state u+ . According to (2.9), this discontinuity must travel with a speed λ = γ̇ =
λi (u− , u+ ) equal to an eigenvalue of the averaged matrix A(u− , u+ ). If we now look at the
i-characteristics, i.e. at the solutions of the O.D.E.

ẋ = λi (u(t, x)),

we see that the Lax condition requires that these lines run into the shock, from both sides.

3 The Riemann Problem

In this chapter we construct the solution to the Riemann problem, consisting of the system of
conservation laws
ut + f (u)x = 0 (3.1)
with the simple, piecewise constant initial data
½
.
u(0, x) = ū(x) = u− if x < 0, (3.2)
u+ if x > 0.
This will provide the basic building block toward the solution of the Cauchy problem with
more general initial data. Throughout our analysis, we shall adopt the following standard
assumption, introduced by P. Lax [33].

(H) For each i = 1, . . . , n, the i-th field is either genuinely nonlinear, so that Dλi (u)·ri (u) > 0
for all u, or linearly degenerate, with Dλi (u) · ri (u) = 0 for all u.

Notice that, in the genuinely nonlinear case, the i-th eigenvalue λi is strictly increasing along
each integral curve of the corresponding field of eigenvectors ri . In the linearly degenerate
case, on the other hand, the eigenvalue λi is constant along each such curve. With the above
assumption we are ruling out the possibility that, along some integral curve of an eigenvector
ri , the corresponding eigenvalue λi may partly increase and partly decrease, having several
local maxima and minima.

16
Example 7 (the “p-system”, modelling isentropic gas dynamics). Denote by ρ the
density of a gas, by v = ρ−1 its specific volume and by u its velocity. A simple model for
isentropic gas dynamics (in Lagrangian coordinates) is then provided by the system
(
vt − u x = 0 ,
(3.3)
ut + p(v)x = 0 .

Here p = p(v) is a function which determines the pressure in terms of of the specific volume.
An appropriate choice is p(v) = kv −γ , with 1 ≤ γ ≤ 3. In the region where v > 0, the Jacobian
matrix of the system is µ ¶
. 0 −1
A = Df = .
p0 (v) 0
The eigenvalues and eigenvectors are found to be
q q
λ1 = − −p0 (v) , λ2 = −p0 (v) , (3.4)
   
1 −1
r1 =  p , r2 =  p . (3.5)
−p0 (v) −p0 (v)
It is now clear that the system is strictly hyperbolic provided that p0 (v) < 0 for all v > 0.
Moreover, observing that
p00 (v)
Dλ1 · r1 = p 0 = Dλ2 · r2 ,
2 −p (v)
we conclude that both characteristic fields are genuinely nonlinear if p00 (v) > 0 for all v > 0.

As we shall see in the sequel, if the assumption (H) holds, then the solution of the Riemann
problem has a simple structure consisting of the superposition of n elementary waves: shocks,
rarefactions or contact discontinuities. This considerably simplifies all further analysis. On
the other hand, for strictly hyperbolic systems that do not satisfy the condition (H), basic
existence and stability results can still be obtained but at the price of heavier technicalities.

3.1 Shock and rarefaction waves

Fix a state u0 ∈ IRn and an index i ∈ {1, . . . , n}. As before, let ri (u) be an i-eigenvector of
the Jacobian matrix A(u) = Df (u). The integral curve of the vector field ri through the point
u0 is called the i-rarefaction curve through u0 . It is obtained by solving the Cauchy problem
in state space:
du
= ri (u), u(0) = u0 . (3.6)
dσ
We shall denote this curve as
σ 7→ Ri (σ)(u0 ). (3.7)
Clearly, the parametrization depends on the choice of the eigenvectors ri . In particular, if we
impose the normalization |ri (u)| ≡ 1, then the rarefaction curve (3.7) will be parameterized
by arc-length.

17
Next, for a fixed u0 ∈ IRn and i ∈ {1, . . . , n}, we consider the curve of states u which can be
connected to the right of u0 by an i-shock, satisfying the Rankine-Hugoniot equations

λ(u − u0 ) = f (u) − f (u0 ) . (3.8)

As in (2.9), we can write these equations in the form

A(u, u0 )(u − u0 ) = λ(u − u0 ) , (3.9)

showing that u − u0 must be a right i-eigenvector of the averaged matrix

Z 1
.
A(u, u0 ) = A(su + (1 − s)u0 ) ds .
0

By a theorem of linear algebra, this holds if and only if u − u0 is orthogonal to every left
j-eigenvector of A(u, u0 ), with j 6= i. The Rankine-Hugoniot conditions can thus be written
in the equivalent form

lj (u, u0 ) · (u − u0 ) = 0 for all j 6= i, (3.10)

together with λ = λi (u, u0 ). Notice that (3.10) is a system of n − 1 scalar equations in n

variables (the n components of the vector u). Linearizing at the point u = u0 one obtains the
linear system
lj (u0 ) · (w − u0 ) = 0 j 6= i,
whose solutions are all the points w = u0 + cri (u0 ), c ∈ IR. Since the vectors lj (u0 ) are linearly
independent, we can apply the implicit function theorem and conclude that the set of solutions
of (3.10) is a smooth curve, tangent to the vector ri at the point u0 . This will be called the
i-shock curve through the point u0 and denoted as

σ 7→ Si (σ)(u0 ). (3.11)

Using a suitable parametrization (say, by arclength), one can show that the two curves Ri , Si
have a second order contact at the point u0 (fig. 10). More precisely, the following estimates
hold. (
Ri (σ)(u0 ) = u0 + σri (u0 ) + O(1) · σ 2 ,
(3.12)
Si (σ)(u0 ) = u0 + σri (u0 ) + O(1) · σ 2 ,
¯ ¯
¯ ¯
¯Ri (σ)(u0 ) − Si (σ)(u0 )¯ = O(1) · σ 3 , (3.13)
³ ´ σ
λi Si (σ)(u0 ), u0 (Dλi (u0 )) · ri (u0 ) + O(1) · σ 2 .
= λi (u0 ) + (3.14)
2
Here and throughout the following, the Landau symbol O(1) denotes a quantity whose absolute
value satisfies a uniform bound, depending only on the system (3.1).

3.2 Three special cases

Before constructing the solution of the Riemann problem (3.1)-(3.2) with arbitrary data
u− , u+ , we first study three special cases:

18
 Si

Ri

ri (u 0)
u
0

Figure 10: Shock and rarefaction curves.

1. Centered Rarefaction Waves. Let the i-th field be genuinely nonlinear, and assume
that u+ lies on the positive i-rarefaction curve through u− , i.e. u+ = Ri (σ)(u− ) for some
σ > 0. For each s ∈ [0, σ], define the characteristic speed

λi (s) = λi (Ri (s)(u− )).

Observe that, by genuine nonlinearity, the map s 7→ λi (s) is strictly increasing. Hence, for
every λ ∈ [λi (u− ), λi (u+ )], there is a unique value s ∈ [0, σ] such that λ = λi (s). For t ≥ 0,
we claim that the function
 −
u if x/t < λi (u− ),
u(t, x) = R (s)(u− ) if x/t = λi (s) ∈ [λi (u− ), λi (u+ )], (3.15)
 +i
u if x/t > λi (u+ ),
is a piecewise smooth solution of the Riemann problem, continuous for t > 0. Indeed, from
the definition it follows
lim ku(t, ·) − ūkL1 = 0.
t→0+

Moreover, the equation (3.1) is trivially satisfied in the sectors where x < tλi (u− ) or x >
tλi (u+ ), because here ut = ux = 0. Next, assume x = tλi (s) for some s ∈ ]0, σ[ . Since u is
constant along each ray through the origin {x/t = c}, we have
x
ut (t, x) + ux (t, x) = 0. (3.16)
t
We now observe that the definition (3.15) implies x/t = λi (u(t, x)). By construction, the vector
.
ux has the same direction as ri (u), hence it is an eigenvector of the Jacobian matrix A(u) =
Df (u) with eigenvalue λi (u). On the sector of the t-x plane where λi (u− ) < x/t < λi (u+ )
we thus have
ut + A(u)ux = ut + λi (u)ux = 0 ,
proving our claim. Notice that the assumption σ > 0 is essential for the validity of this
construction. In the opposite case σ < 0, the definition (3.15) would yield a triple-valued
function in the region where x/t ∈ [λi (u+ ) , λi (u− )].

2. Shocks. Assume again that the i-th family is genuinely nonlinear and that the state u+ is
.
connected to the right of u− by an i-shock, i.e. u+ = Si (σ)(u− ). Then, calling λ = λi (u+ , u− )
the Rankine-Hugoniot speed of the shock, the function
½ −
u if x < λt,
u(t, x) = (3.17)
u+ if x > λt,

19
provides a piecewise constant solution to the Riemann problem. Observe that, if σ < 0, than
this solution is entropy admissible in the sense of Lax. Indeed, since the speed is monotonically
increasing along the shock curve, recalling (3.14) we have

λi (u+ ) < λi (u− , u+ ) < λi (u− ). (3.18)

Hence the conditions (2.25) hold. In the case σ > 0, however, one has λi (u− ) < λi (u+ ) and
the admissibility conditions (2.25) are violated.
If η is a strictly convex entropy with flux q, one can also prove that the entropy admissibility
conditions (2.19) hold for all σ ≤ 0 small, but fail when σ > 0.

3. Contact discontinuities. Assume that the i-th field is linearly degenerate and that
the state u+ lies on the i-th rarefaction curve through u− , i.e. u+ = Ri (σ)(u− ) for some
σ. By assumption, the i-th characteristic speed λi is constant along this curve. Choosing
λ = λ(u− ), the piecewise constant function (3.17) then provides a solution to our Riemann
problem. Indeed, the Rankine-Hugoniot conditions hold at the point of jump:
Z σ
f (u+ ) − f (u− ) = Df (Ri (s)(u− )) ri (Ri (s)(u− )) ds
0
Z σ (3.19)
= λ(u− ) ri (Ri (s)(u− )) ds = λi (u− ) · [Ri (σ)(u− ) − u− ].
0

In this case, the Lax entropy condition holds regardless of the sign of σ. Indeed,

λi (u+ ) = λi (u− , u+ ) = λi (u− ). (3.20)

Moreover, if η is a strictly convex entropy with flux q, one can show that the entropy admis-
sibility conditions (2.19) hold, for all values of σ.
Observe that, according to (3.19), for linearly degenerate fields the shock and rarefaction
curves actually coincide, i.e. Si (σ)(u0 ) = Ri (σ)(u0 ) for all σ.

The above results can be summarized as follows. For a fixed left state u− and i ∈ {1, . . . , n}
define the mixed curve
½
− Ri (σ)(u− ) if σ ≥ 0,
Ψi (σ)(u ) = (3.21)
Si (σ)(u− ) if σ < 0.
In the special case where u+ = Ψi (σ)(u− ) for some σ, the Riemann problem can then be
solved by an elementary wave: a rarefaction, a shock or a contact discontinuity.

Remark 3. The shock admissibility condition, derived in (2.24) for the scalar case, has a
natural generalization to the vector valued case. Namely:
Admissibility Condition 4 (Liu condition). Given two states u− ∈ IRn and u+ =
Si (σ)(u− ), the weak solution (3.17) satisfies the T. P. Liu condition if
³ ´ ³ ´
λi Si (σ)(u− ), u− ≤ λi Si (s)(u− ), u− for all s ∈ [0, σ] . (3.22)

20
In other words, every i-shock joining u− with some intermediate state Si (s)(u− ) must travel
with speed greater or equal than the speed of our shock (u− , u+ ). A very remarkable fact
is that, for general strictly hyperbolic systems (without any reference to the assumptions
(H) on genuine nonlinearity or linear degeneracy), the Liu conditions completely characterize
vanishing viscosity limits. Namely, a weak solution to (3.1) with small total variation can be
obtained as limit of vanishing viscosity approximations if and only if all of its jumps satisfy
the Liu admissibility condition.

3.3 General solution of the Riemann problem

Relying on the previous analysis, the solution of the general Riemann problem (3.1)-(3.2) can
be obtained by finding intermediate states ω0 = u− , ω1 , . . . , ωn = u+ such that each pair of
adiacent states ωi−1 , ωi can be connected by an elementary wave, i. e.
ωi = Ψi (σi )(ωi−1 ) i = 1, . . . , n. (3.23)
This can be done whenever u+ is sufficiently close to u− . Indeed, for |u+ − u− | small, the
implicit function theorem provides the existence of unique wave strengths σ1 , . . . σn such that
u+ = Ψn (σn ) ◦ · · · ◦ Ψ1 (σ1 )(u− ). (3.24)
In turn, these determine the intermediate states ωi in (3.23). The complete solution is now
obtained by piecing together the solutions of the n Riemann problems
½
ωi−1 if x < 0,
ut + f (u)x = 0, u(0, x) = (3.25)
ωi if x > 0,
on different sectors of the t-x plane. By construction, each of these problems has an entropy-
admissible solution consisting of a simple wave of the i-th characteristic family. More precisely:

CASE 1: The i-th characteristic field is genuinely nonlinear and σi > 0. Then the solution
of (3.25) consists of a centered rarefaction wave. The i-th characteristic speeds range over the
interval [λ− +
i , λi ], defined as
. .
λ−
i = λi (ωi−1 ), λ+i = λi (ωi ).

CASE 2: Either the i-th characteristic field is genuinely nonlinear and σi ≤ 0, or else the
i-th characteristic field is linearly degenerate (with σi arbitrary). Then the solution of (3.25)
consists of an admissible shock or a contact discontinuity, travelling with Rankine-Hugoniot
speed
. + .
λ−i = λi = λi (ωi−1 , ωi ).

The solution to the original problem (3.1)-(3.2) can now be constructed (fig. 11) by piecing
together the solutions of the n Riemann problems (3.25), i = 1, . . . , n. Indeed, for σ1 , . . . , σn
sufficiently small, the speeds λ− +
i , λi introduced above remain close to the corresponding eigen-
− −
values λi (u ) of the matrix A(u ). By strict hyperbolicity and continuity, we can thus assume
that the intervals [λ− +
i , λi ] are disjoint, i.e.

λ− + − + − +
1 ≤ λ1 < λ2 ≤ λ2 < · · · < λn ≤ λn .

21
 x= λ−2 =λ+2
t
x=λ−3
+
x=λ−1 =λ+1 x= λ3
ω
2
ω
1

ω = u− ω = u+
0 3

0 x

Figure 11: Typical solution of a Riemann problem, with the assumption (H).

Therefore, a piecewise smooth solution u : [0, ∞) × IR 7→ IRn is well defined by the assignment




u− = ω0 if x/t ∈ ] − ∞, λ−
1 [,







 Ri (s)(ωi−1 ) if x/t = λi (Ri (s)(ωi−1 )) ∈ [λ− +
i , λi [ ,
u(t, x) = (3.26)




 ωi if x/t ∈ [λ+ −
i , λi+1 [ ,





 u+ = ωn if x/t ∈ [λ+
n , ∞[ .

Observe that this solution is self-similar, having the form u(t, x) = ψ(x/t), with ψ : IR 7→ IRn
possibly discontinuous.

3.4 Error and interaction estimates

We conclude this section by proving two types of estimates, which will play a key role in the
analysis of front tracking approximations.
.
Fix a left state u− and consider a right state u+ = Rk (σ)(u− ) on the k-rarefaction curve. In
general, shock and rarefaction curves do not coincide; hence the function
½ −
. u if x < λk (u− ) ,
u(t, x) = (3.27)
u+ if x > λk (u− ) ,

is not an exact solution of the system (3.1). However, we now show that the error in the
Rankine-Hugoniot equations and the possible increase in a convex entropy are indeed very
small. In the following, the Landau symbol O(1) denotes a quantity which remains uniformly
bounded as u− , σ range on bounded sets.

Lemma 3 (error estimates). For σ > 0 small, one has the estimate
h h i
λk (u− ) Rk (σ)(u− ) − u− ] − f (Rk (σ)(u− )) − f (u− ) = O(1) · σ 2 . (3.28)

22
Moreover, if η is a convex entropy with entropy flux q, then
h i h i
λk (u− ) η(Rk (σ)(u− )) − η(u− ) − q(Rk (σ)(u− )) − q(u− ) = O(1) · σ 2 . (3.29)

Proof. Call E(σ) the left hand side of (3.28). Clearly E(0) = 0. Differentiating w.r.t. σ at
the point σ = 0 and recalling that dRk /dσ = rk , we find
¯
dE ¯¯
= λk (u− )rk (u− ) − Df (u− )rk (u− ) = 0 .
dσ ¯σ=0

Since E varies smoothly with u− and σ, the estimate (3.28) easily follows by Taylor’s formula.
The second estimate is proved similarly. Call E 0 (σ) the left hand side of (3.29) and observe
that E 0 (0) = 0. Differentiating w.r.t. σ we now obtain
¯
dE 0 ¯¯
= λk (u− )Dη(u− )rk (u− ) − Dq(u− )rk (u− ) = 0 ,
dσ ¯σ=0

because Dηλk rk = DηDf rk = Dq rk . Taylor’s formula now yields (3.29).

σj
σi σk
σi
σk

σ’ σ"
ur σ’ σ"
ul

um ul ur
m
u

Figure 12: Wave interactions. Strengths of the incoming and outgoing waves.

Next, consider a left state ul , a middle state um and a right state ur (fig. 12). Assume that the
pair (ul , um ) is connected by a j-wave of strength σ 0 , while the pair (ul , um ) is connected by an
i-wave of strength σ 00 , with i < j. We are interested in the strength of the waves (σ1 , . . . , σn )
in the solution of the Riemann problem where u− = ul and u+ = ur . Roughly speaking, these
are the waves determined by the interaction of the σ 0 and σ 00 . The next lemma shows that
σi ≈ σ 00 , σj ≈ σ 0 while σk ≈ 0 for k 6= i, j.
A different type of interaction is considered in fig. 8. Here the pair (ul , um ) is connected by an
i-wave of strength σ 0 , while the pair (ul , um ) is connected by a second i-wave, say of strength
σ 00 . In this case, the strengths (σ1 , . . . , σn ) of the outgoing waves satisfy σi ≈ σ 0 + σ 00 while
σk ≈ 0 for k 6= i. As usual, O(1) will denote a quantity which remains uniformly bounded as
u− , σ 0 , σ 00 range on bounded sets.

23
Lemma 4 (interaction estimates). Consider the Riemann problem (3.1)-(3.2).

(i) Recalling (3.21), assume that the right state is given by

u+ = Ψi (σ 00 ) ◦ Ψj (σ 0 )(u− ). (3.30)

Let the solution consist of waves of size (σ1 , . . . , σn ), as in (3.24). Then

X
|σi − σ 00 | + |σj − σ 0 | |σk | = O(1) · |σ 0 σ 00 | . (3.31)
k6=i,j

(ii) Next, assume that the right state is given by

u+ = Ψi (σ 00 ) ◦ Ψi (σ 0 )(u− ), (3.32)

Then the waves (σ1 , . . . , σn ) in the solution of the Riemann problem can be estimated by
X
|σi − σ 0 − σ 00 | + |σk | = O(1) · |σ 0 σ 00 |(|σ 0 | + |σ 00 |) . (3.33)
k6=i

.
Proof. (i) For u− , u+ ∈ IRn , k = 1, . . . , n, call Ek (u− , u+ ) = σk the size of the k-th wave in
the solution of the corresponding Riemann problem. By our earlier analysis, the functions Ek
are C 2 with Lipschitz continuous second derivatives.
For a given left state u− , we now define the composed functions Φk , k = 1, . . . , n by setting
³ ´
.
Φi (σ 0 , σ 00 ) = σi − σ 00 = Ei u− , Ψj (σ 0 ) ◦ Ψi (σ 00 )(u− ) − σ 00 ,
³ ´
.
Φj (σ 0 , σ 00 ) = σj − σ 0 = Ej u− , Ψj (σ 0 ) ◦ Ψi (σ 00 )(u− ) − σj0 ,
³ ´
.
Φk (σ 0 , σ 00 ) = σk = Ek u− , Ψj (σ 0 ) ◦ Ψi (σ 00 )(u− ) if k 6= i, j.

The Φk are C 2 functions of σ 0 , σ 00 with Lipschitz continuous second derivatives, depending

continuously also on the left state u− . Observing that

Φk (σ 0 , 0) = Φk (0, σ 00 ) = 0 for all σ 0 , σ 00 , k = 1, . . . , n,

by Taylor’s formula (see Lemma A.2 in the Appendix) we conclude that

Φk (σ 0 , σ 00 ) = O(1) · |σ 0 σ 00 |

for all k = 1, . . . , n. This establishes (3.31).

(ii) To prove (3.33), we consider the functions

³ ´
.
Φi (σ 0 , σ 00 ) = σi − σ 0 − σ 00 = Ei u− , Ψi (σ 00 ) ◦ Ψi (σ 0 )(u− ) − σ 0 − σ 00 ,
³ ´
.
Φk (σ 0 , σ 00 ) = σk = Ek u− , Ψi (σ 00 ) ◦ Ψi (σ 0 )(u− ) if k 6= i .

24
In the case where σ 0 , σ 00 ≥ 0, recalling (3.21) we have

u+ = Ri (σ 00 ) ◦ Ri (σ 0 )(u− ) = Ri (σ 00 + σ 0 )(u− ) .

Hence the Riemann problem is solved exactly by an i-rarefaction of strength σ 0 +σ 00 . Therefore

Φk (σ 0 , σ 00 ) = 0 whenever σ 0 , σ 00 ≥ 0 , (3.34)

for all k = 1, . . . , n. As before, the Φk are C 2 functions of σ 0 , σ 00 with Lipschitz continuous

second derivatives, depending continuously also on the left state u− . We observe that

Φk (σ 0 , 0) = Φk (0, σ 00 ) = 0 . (3.35)

Moreover, by (3.34) the continuity of the second derivatives implies

∂ 2 Φk
(0, 0) = 0 . (3.36)
∂σ 0 ∂σ 00
By (3.35)-(3.36), using the estimate (10.16) in the Appendix we now conclude

Φk (σ, σ 0 ) = O(1) · |σ 0 σ 00 |(|σ 0 | + |σ 00 |)

for all k = 1, . . . , n. This establishes (ii).

3.5 Two examples

Example 8 (chromatography). The 2 × 2 system of conservation laws

· ¸ · ¸
u1 u2
[u1 ]t + = 0, [u2 ]t + = 0, u1 , u2 > 0, (3.37)
1 + u1 + u2 x 1 + u1 + u2 x

arises in the study of two-component chromatography. Writing (3.37) in the quasilinear form
(1.6), the eigenvalues and eigenvectors of the corresponding 2 × 2 matrix A(u) are found to be
1 1
λ1 (u) = , λ2 (u) = ,
(1 + u1 + u2 )2 1 + u1 + u2
µ ¶ µ ¶
−1 u1 1 1
r1 (u) = q · , r2 (u) = √ · .
u21 + u22 u2 2 −1

The first characteristic field is genuinely nonlinear, the second is linearly degenerate. In this
example, the two shock and rarefaction curves Si , Ri always coincide, for i = 1, 2. Their
computation is easy, because they are straight lines (fig. 13):

R1 (σ)(u) = u + σr1 (u), R2 (σ)(u) = u + σr2 (u). (3.38)

Observe that the integral curves of the vector field r1 are precisely the rays through the
origin, while the integral curves of r2 are the lines with slope −1. Now let two states u− =
(u− − + + +
1 , u2 ), u = (u1 , u2 ) be given. To solve the Riemann problem (3.1)-(3.2), we first compute
an intermediate state u∗ such that u∗ = R1 (σ1 )(u− ), u+ = R2 (σ2 )(u∗ ) for some σ1 , σ2 . By
(3.38), the components of u∗ satisfy

u∗1 + u∗2 = u+ +
1 + u2 , u∗1 u− − ∗
2 = u1 u2 .

25
 u2

r1 r2

u1

Figure 13: Integral curves of the eigenvectors r1 , r2 , for the system (3.37).

The solution
q of the Riemann
q problem thus takes two different forms, depending on the sign
− 2 − 2
of σ1 = (u1 ) + (u2 ) − (u∗1 )2 + (u∗2 )2 .

CASE 1: σ1 > 0. Then the solution consists of a centered rarefaction wave of the first
family and of a contact discontinuity of the second family:


 u− if x/t < λ1 (u− ),

 su∗ + (1 − s)u− if x/t = λ1 (su∗ + (1 − s)u− ), s ∈ [0, 1],
u(t, x) = (3.39)

 u∗ if λ1 (u∗ ) < x/t < λ2 (u+ ),


u+ if x/t ≥ λ2 (u+ ).

CASE 2: σ1 ≤ 0. Then the solution contains a compressive shock of the first family (which
vanishes if σ1 = 0) and a contact discontinuity of the second family:

 − x/t < λ1 (u− , u∗ ),
 u if
u(t, x) = u∗ if λ1 (u− , u∗ ) ≤ x/t < λ2 (u+ ), (3.40)

 u+ if x/t ≥ λ2 (u+ ).

Observe that λ2 (u∗ ) = λ2 (u+ ) = (1 + u+ + −1

1 + u2 ) , because the second characteristic field is
linearly degenerate. In this special case, since the integral curves of r1 are straight lines, the
shock speed in (3.40) can be computed as
Z 1
λ1 (u− , u∗ ) = λ1 (su∗ + (1 − s)u− ) ds
0
Z 1
− −2
= [1 + s(u∗1 + u∗2 ) + (1 − s)(u−
1 + u2 )] ds
0

1
= .
(1 + u∗1 + u∗2 )(1 + u1− + u−
2)

Example 9 (the p-system). Consider again the equations for isentropic gas dynamics (in

26
Lagrangian coordinates) (
vt − u x = 0 ,
(3.41)
ut + p(v)x = 0 .
We now study the Riemann problem, for general initial data
½
U − = (v − , u− ) if x < 0,
U (0, x) = (3.42)
U + = (v + , u+ ) if x > 0,

assuming, that v − , v + > 0.

By (3.5), the 1-rarefaction curve through U − is obtained by solving the Cauchy problem
du q 0
= −p (v), u(v − ) = u− .
dv
This yields the curve
n Z v q o
R1 = (v, u); u − u− = −p0 (y) dy . (3.43)
v−

Similarly, the 2-rarefaction curve through the point U − is

n Z v q o
R2 = (v, u); u − u− = − −p0 (y) dy . (3.44)
v−

Next, the shock curves S1 , S2 through U − are derived from the Rankine-Hugoniot conditions

λ(v − v − ) = −(u − u− ), λ(u − u− ) = p(v) − p(v − ). (3.45)

Using the first equation in (3.45) to eliminate λ, these shock curves are computed as
( )
− 2 − − . u − u−
S1 = (v, u); −(u − u ) = (v − v )(p(v) − p(v )), λ=− <0 , (3.46)
v − v−
( )
− 2 − − . u − u−
S2 = (v, u); −(u − u ) = (v − v )(p(v) − p(v )), λ=− >0 . (3.47)
v − v−

Recalling (3.4)-(3.5) and the assumptions p0 (v) < 0, p00 (v) > 0, we now compute the directional
derivatives
p00 (v)
(Dλ1 )r1 = (Dλ2 )r2 = p 0 > 0. (3.48)
2 −p (v)
By (3.48) it is clear that the Riemann problem (3.41)-(3.42) admits a solution in the form of
a centered rarefaction wave in the two cases U + ∈ R1 , v + > v − , or else U + ∈ R2 , v + < v − .
On the other hand, a shock connecting U − with U + will be admissible provided that either
U + ∈ S1 and v + < v − , or else U + ∈ S2 and v + > v − .
Taking the above admissibility conditions into account, we thus obtain four lines originating
from the point U − = (v − , u− ), i.e. the two rarefaction curves

σ 7→ R1 (σ), R2 (σ) σ ≥ 0,

27
and the two shock curves
σ 7→ S1 (σ), S2 (σ) σ ≤ 0.
In turn, these curves divide a neighborhood of U − into four regions (fig 14):

Ω1 , bordering on R1 , S2 , Ω2 , bordering on R1 , R2 ,
Ω3 , bordering on S1 , S2 , Ω4 , bordering on S1 , R2 .

u Ω2
R
R2 1

Ω1
Ω4 U −= (u −, v −)

v
S1 S
2
Ω3

Figure 14: Shocks and rarefaction curves through the point U − = (v − , u− ).

For U + sufficiently close to U − , the structure of the general solution to the Riemann problem
is now determined by the location of the state U + , with respect to the curves Ri , Si (fig 15).

CASE 1: U + ∈ Ω1 . The solution consists of a 1-rarefaction wave and a 2-shock.

CASE 2: U + ∈ Ω2 . The solution consists of two centered rarefaction waves.

CASE 3: U + ∈ Ω3 . The solution consists of two shocks.

CASE 4: U + ∈ Ω4 . The solution consists of a 1-shock and a 2-rarefaction wave.

Remark 4. Consider a 2 × 2 strictly hyperbolic system of conservation laws. Assume that

the i-th characteristic field is genuinely nonlinear. The relative position of the i-shock and the
i-rarefaction curve through a point u0 can be determined as follows (fig. 10). Let σ 7→ Ri (σ)
be the i-rarefaction curve, parameterized so that λi (Ri (σ)) = λi (u0 ) + σ. Assume that, for
some constant α, the point

Si (σ) = Ri (σ) + (ασ 3 + o(σ 3 ))rj (u0 ) (3.49)

lies on the i-shock curve through u0 , for all σ. Here the Landau symbol o(σ 3 ) denotes a
higher order infinitesimal, as σ → 0. The wedge product of two vectors in IR2 is defined as

28
 R S
U+
U- Case 1

U+
R R

U- Case 2

S S
U-
Case 3

U+

+ S R
U
U-
Case 4

Figure 15: Solution to the Riemann problem for the p-system. The four different cases.
µ ¶ µ ¶
a c .
∧ = ad − bc. We then have
b d
. h i h ³ ´ i
Ψ(σ) = Ri (σ) + (ασ 3 + o(σ 3 ))rj (u0 ) − u0 ∧ f Ri (σ) + (ασ 3 + o(σ 3 ))rj (u0 ) − f (u0 )
.
= A(σ) ∧ B(σ) ≡ 0 .

Indeed, the Rankine-Hugoniot equations imply that the vectors A(σ) and B(σ) are parallel.
According to Leibnitz’ rule, the fourth derivative is computed by
à ! à ! µ ¶ à ! à !
d4 d4 d3 d d2 d2
Ψ = A ∧ B + 4 A ∧ B +6 A ∧ B
dσ 4 dσ 4 dσ 3 dσ dσ 2 dσ 2
µ ¶ Ã ! Ã !
d d3 d4
+4 A ∧ B +A∧ B
dσ dσ 3 dσ 4
d
By the choice of the parametrization, dσ λi (Ri (σ)) ≡ 1. Hence
d d
f (Ri (σ)) = λi (Ri (σ)) Ri (σ) ,
dσ dσ
d2 d d2
f (R i (σ)) = R i (σ) + λi (R i (σ)) Ri (σ) ,
dσ 2 dσ dσ 2
d3 d2 d3
f (Ri (σ)) = 2 R i (σ) + λ i (R i (σ)) Ri (σ) .
dσ 3 dσ 2 dσ 3
.
For convenience, we write ri • rj = (Drj )ri to denote the directional derivative of rj in the
direction of ri . At σ = 0 we have
d d2
A = B = 0, Ri = ri (u0 ) , Ri = (ri • ri )(u0 ) .
dσ dσ 2

29
Using the above identities and the fact that the wedge product is anti-symmetric, we conclude
¯ Ã ! µ ¶ Ã ! Ã !
d4 ¯¯ d3 d d2 d d2
Ψ¯ = 4 Ri + 6αrj ∧ λi Ri + 6 Ri ∧ Ri + λi 2 Ri
dσ 4 ¯σ=0 dσ 3 dσ dσ 2 dσ dσ

µ ¶ Ã !
d d2 d3
+4 Ri ∧ 2 2 Ri + λi 3 Ri + 6αλj rj
dσ dσ dσ

= 24α(λi − λj )(rj ∧ ri ) − 2(ri • ri ) ∧ ri = 0 .

The i-shock curve through u0 is thus traced by points Si (σ) at (3.49), with

(ri • ri ) ∧ ri
α = . (3.50)
12(λi − λj )(rj ∧ ri )

The sign of α in (3.50) gives the position of the i-shock curve, relative to the i-rarefaction
curve, near the point u0 . In particular, if (ri • ri ) ∧ ri 6= 0, it is clear that these two curves do
not coincide.

4 Front tracking approximations

4.1 Global existence of entropy weak solutions

In this chapter we describe in detail the construction of front tracking approximations and
give the main application of this technique, proving the global existence of weak solutions.
Consider the Cauchy problem
ut + f (u)x = 0, (4.1)
u(0, x) = ū(x). (4.2)
where the flux function f : IRn 7→ IRn is smooth, defined on a neighborhood of the origin. We
always assume that the system is strictly hyperbolic, and that the assumption (H) holds, so
that each characteristic field is either genuinely nonlinear or linearly degenerate. The following
basic theorem provides the global existence of an entropy weak solution, for all initial data
with suitably small total variation.

Theorem 2. Assume that the system (4.1) is strictly hyperbolic, and that each characteristic
field is either linearly degenerate or genuinely nonlinear. Then there exists a constant δ0 > 0
such that, for every initial condition ū ∈ L1 (IR; IRn ) with

Tot.Var.{ū} ≤ δ0 , (4.3)

the Cauchy problem (4.1)-(4.2) has a weak solution u = u(t, x) defined for all t ≥ 0. This
solution also satisfies the admissibility conditions

η(u)t + q(u)x ≤ 0 , (4.4)

30
for every convex entropy η with entropy flux q.

The above theorem was first proved in fundamental paper of J. Glimm [27], where approximate
solutions are constructed by piecing together Riemann solutions on the nodes of a fixed grid.
Here we shall describe an alternative construction, based on wave-front tracking. This method
was developed in [21], [23], and in [7] respectively for scalar, 2 × 2, and general n × n systems.
Further versions of this construction can also be found in [3, 30, 47].

t
t4

t3
xβ
xα

t2
t1
σ σ’

0 x

Figure 16: An approximate solution obtained by front tracking.

Roughly speaking, a front tracking ε-approximate solution (fig. 16) is a piecewise constant
function u = u(t, x) whose jumps are located along finitely many segments x = xα (t) in
the t-x plane. At any given time t > 0, all these jumps should approximately satisfy the
Rankine-Hugoniot conditions, namely
X ¯¯ ¯
¯
¯ẋα [u(t, xα +) − u(t, xα −)] − [f (u(t, xα +)) − f (u(t, xα −))]¯ = O(1) · ε.
¯ ¯
α

Moreover, if η is a convex entropy with flux q, at each time t > 0 in view of (2.18) we also
require the approximate admissibility conditions
X½ ¾
[q(u(t, xα +)) − q(u(t, xα −))] − ẋα [η(u(t, xα +)) − η(u(t, xα −))] ≤ O(1) · ε.
α

The proof of Theorem 2 will be given in two steps. In the first part of the proof we describe
a naive front tracking algorithm and derive uniform bounds on the total variation of the
approximate solutions. Relying on Helly’s compactness theorem, we also prove that a suitable
subsequence of front tracking approximations converges to an entropy weak solution. This first
step thus contains all the “heart of the matter”. It would provide a complete proof, except for
one gap: we are not considering here the possibility that infinitely many wave-fronts appear
within finite time (fig. 22, left). If this happens, the naive front tracking algorithm breaks
down, and solutions cannot be further prolonged in time.
The only purpose of the second part of the proof is to fix this technical problem. We show
that a suitable modified front tracking algorithm will always produce approximate solutions
having a finite number of wave-fronts for all times t ≥ 0. By extending the previous estimates
to this modified construction, one can complete the proof of Theorem 2.

31
4.2 A naive front tracking algorithm

Let the initial condition ū be given and fix ε > 0. We now describe an algorithm which
produces a piecewise constant approximate solution to the Cauchy problem (4.1)-(4.2). The
construction (fig. 16) starts at time t = 0 by taking a piecewise constant approximation u(0, ·)
of ū, such that
Z
Tot.Var.{u(0, ·)} ≤ Tot.Var.{ū} , |u(0, x) − ū(x)| dx ≤ ε . (4.5)

Let x1 < · · · < xN be the points where u(0, ·) is discontinuous. For each α = 1, . . . , N , the
Riemann problem generated by the jump (u(0, xα −), u(0, xα +)) is approximately solved on a
forward neighborhood of (0, xα ) in the t-x plane by a piecewise constant function, according
to the following procedure.

Accurate Riemann Solver. Consider the general Riemann problem at a point (t̄, x̄),
½ −
u if x < x̄,
vt + f (v)x = 0, v(t̄, x) = (4.6)
u+ if x > x̄,
Recalling (3.21), let ω0 , . . . , ωn be the intermediate states and σ1 , . . . , σn be the strengths of
the waves in the solution, so that

ω0 = u− , ωn = u+ , ωi = Ψi (σi )(ωi−1 ) i = 1, . . . , n. (4.7)

If all jumps (ωi−1 , ωi ) were shocks or contact discontinuities, then this solution would be
already piecewise constant. In general, the exact solution of (4.6) is not piecewise constant,
because of the presence of centered rarefaction waves. These will be approximated by piecewise
constant rarefaction fans, inserting additional states ωi,j as follows.

x i,j
Ri

ω i =ωi,4 ωi−1 ωi
ωi,3
ωi,2
ωi,1
ωi−1 =ωi,0 _
x

Figure 17: Replacing a centered rarefaction wave by a rarefaction fan.

If the i-th characteristic field is genuinely nonlinear and σi > 0, we divide the centered i-
rarefaction into a number pi of smaller i-waves, each with strength σi /pi . Since we want
σ/pi < ε, we choose the integer
.
pi = 1 + [[σi /ε]], (4.8)
where [[s]] denotes the largest integer ≤ s. For j = 1, . . . , pi , we now define the intermediate
states and wave-fronts (fig. 17)

ωi,j = Ψi (jσi /pi )(ωi−1 ), xi,j (t) = x̄ + (t − t̄)λi (ωi,j−1 ). (4.9)

32
For notational convenience, if the i-th characteristic field is genuinely nonlinear and σi ≤ 0, or
.
if the i-th characteristic field is linearly degenerate (with σi arbitrary), we define pi = 1 and
set
ωi,1 = ωi , xi,1 (t) = x̄ + (t − t̄)λi (ωi−1 , ωi ). (4.10)
Here λi (ωi−1 , ωi ) is the Rankine-Hugoniot speed of the jump connecting ωi−1 with ωi , so that

λi (ωi−1 , ωi ) · (ωi − ωi−1 ) = f (ωi ) − f (ωi−1 ). (4.11)

As soon as the intermediate states ωi,j and the locations xi,j (t) of the jumps have been
determined by (4.9) or (4.10), we can define a piecewise constant approximate solution to the
Riemann problem (4.6) by setting (fig. 18)
 −

 u if x < x1,1 (t),





 +

u if x > xn,pn (t),
v(t, x) = (4.12)



 ωi (= ωi,pi ) if xi,pi (t) < x < xi+1,1 (t),






ωi,j if xi,j (t) < x < xi,j+1 (t) (j = 1, . . . , pi − 1).

We observe that the difference between this function v and the exact solution is only due to
the fact that every centered i-rarefaction wave is here divided into equal parts and replaced
by a rarefaction fan containing pi wave-fronts. Because of (4.8), the strength of each one of
these fronts is < ε.
x 2,1
t

x 3,j
x1,1 ω
2

ω1

ω ω
0 3

0 x

Figure 18: A piecewise constant approximate solution to the Riemann problem.

We now resume the construction of a front tracking solution to the original Cauchy problem
(4.1)-(4.2). Having solved all the Riemann problems at time t = 0, the approximate solution
u can be prolonged until a first time t1 is reached, when two wave-fronts interact (fig. 16).
Since u(t1 , ·) is still a piecewise constant function, the corresponding Riemann problems can
again be approximately solved within the class of piecewise constant functions. The solution

33
u is then continued up to a time t2 where a second interaction takes place, etc. . . We remark
that, by an arbitrary small change in the speed of one of the wave fronts, it is not restrictive
to assume that at most two incoming fronts collide, at each given time t > 0. This will
considerably simplify all subsequent analysis, since we don’t need to consider the case where
three or more incoming fronts meet together.

4.3 Bounds on the total variation

In this section we derive bounds on the total variation of a front tracking approximation
u(t, ·), uniformly valid for all t ≥ 0. These estimates will be obtained from Lemma 4, using
an interaction functional.
We begin by introducing some notation. At a fixed time t, let xα , α = 1, . . . , N , be the
locations of the fronts in u(t, ·). Moreover, let |σα | be the strength of the wave-front at xα ,
say of the family kα ∈ {1, . . . , n}. Consider the two functionals
. . X
V (t) = V (u(t)) = |σα | , (4.13)
α

measuring the total strength of waves in u(t, ·), and

. . X
Q(t) = Q(u(t)) = |σα σβ | , (4.14)
(α,β)∈A

measuring the wave interaction potential. In (4.14), the summation ranges over the set A of
all couples of approaching wave-fronts:

Definition 4.1. Two fronts, located at points xα < xβ and belonging to the characteristic
families kα , kβ ∈ {1, . . . , n} respectively, are approaching if kα > kβ or else if kα = kβ and at
least one of the wave-fronts is a shock of a genuinely nonlinear family.

Roughly speaking, two fronts are approaching if the one behind has the larger speed (and
hence it can collide with the other, at a future time).

Now consider the approximate solution u = u(t, x) constructed by the front tracking algorithm.
It is clear that the quantities V (u(t)), Q(u(t)) remain constant except at times where an
interaction occurs. At a time τ where two fronts of strength |σ 0 |, |σ 00 | collide, the interaction
estimates (3.31) or (3.33) yield
.
∆V (τ ) = V (τ +) − V (τ −) = O(1) · |σ 0 σ 00 |, (4.15)
.
∆Q(τ ) = Q(τ +) − Q(τ −) = − |σ 0 σ 00 | + O(1) · |σ 0 σ 00 | · V (τ −). (4.16)
Indeed (fig. 19), after time τ the two colliding fronts σ 0 , σ 00 are no longer approaching. Hence
the product |σ 0 σ 00 | is no longer counted within the summation (4.14). On the other hand, the
new waves emerging from the interaction (having strength O(1) · |σ 0 σ 00 |) can approach all the
other fronts not involved in the interaction (which have total strength ≤ V (τ −) ).

34
 σ
k
t σi σj

τ
σα
σ’ σ"

Figure 19: Estimating the change in the total variation at a time where two fronts interact.

If V remains sufficiently small, so that O(1) · V (τ −) ≤ 1/2, from (4.16) it follows

|σ 0 σ 00 |
Q(τ +) − Q(τ −) ≤ − . (4.17)
2
By (4.15) and (4.17) we can thus choose a constant C0 large enough so that the quantity
.
Υ(t) = V (t) + C0 Q(t)
decreases at every interaction time, provided that V remains sufficiently small.
We now observe that the total strength of waves is an equivalent way of measuring the total
variation. Indeed, for some constant κ one has
Tot.Var.{u(t)} ≤ V (u(t)) ≤ κ · Tot.Var.{u(t)} . (4.18)
Moreover, the definitions (4.13)-(4.14) trivially imply Q ≤ V 2 . If the total variation of the
initial data u(0, ·) is sufficiently small, the previous estimates show that the quantity V + C0 Q
is nonincreasing in time. Therefore
Tot.Var.{u(t)} ≤ V (u(t)) ≤ V (u(0)) + C0 Q(u(0)) . (4.19)
This provides a uniform bound on the total variation of u(t, ·) valid for all times t ≥ 0.
An important consequence of the bound (4.19) is that, at every time τ where two fronts
interact, the corresponding Riemann problem can always be solved. Indeed, the left and right
states differ by the quantity
|u+ − u− | ≤ Tot.Var.{u(τ )} ,
which remains small.
Another consequence of the bound on the total variation is the continuity of t 7→ u(t, ·) as a
function with values in L1loc . More precisely, there exists a Lipschitz constant L0 such that
Z ∞
|u(t, x) − u(t0 , x)| dx ≤ L0 |t − t0 | for all t, t0 ≥ 0 . (4.20)
−∞

Indeed, if no interaction occurs inside the interval [t, t0 ], the left hand side of (4.20) can be
estimated simply as
P
ku(t) − u(t0 )kL1 ≤ |t − t0 | α |σα | |ẋα |

≤ |t − t0 | · [total strength of all wave fronts] · [maximum speed] (4.21)

≤ L0 · |t − t0 |,

35
for some uniform constant L0 . The case where one or more interactions take place within [t, t0 ]
is handled in the same way, observing that the map t 7→ u(t, ·) is continuous across interaction
times.

4.4 Convergence to a limit solution

Fix any sequence εν → 0+. For every ν ≥ 1 we apply the front tracking algorithm and
construct an εν -approximate solution uν of the Cauchy problem (4.1)-(4.2). By the previous
analysis, the total variation of uν (t, ·) remains bounded, uniformly for all t ≥ 0 and ν ≥ 1.
Moreover, by (4.20) the maps t 7→ uν (t, ·) are uniformly Lipschitz continuous w.r.t. the L1
distance. We can thus apply Helly’s compactness theorem (see Theorem A.1 in the Appendix)
and extract a subsequence which converges to some limit function u in L1loc , also satisfying
(4.20).
By the second relation in (4.5), as εν → 0 we have uν (0) → ū in L1loc . Hence the initial
condition (4.2) is clearly attained. To prove that u is a weak solution of the Cauchy problem,
it remains to show that, for every φ ∈ Cc1 with compact support contained in the open half
plane where t > 0, one has
Z ∞Z ∞
φt (t, x)u(t, x) + φx (t, x)f (u(t, x)) dxdt = 0. (4.22)
0 −∞
Since the uν are uniformly bounded and f is uniformly continuous on bounded sets, it suffices
to prove that
Z ∞Z ∞ n o
lim φt (t, x)uν (t, x) + φx (t, x)f (uν (t, x)) dxdt = 0. (4.23)
ν→0 0 −∞

Choose T > 0 such that φ(t, x) = 0 whenever t ∈ / [0, T ]. For a fixed ν, at any time t call
x1 (t) < · · · < xN (t) the points where uν (t, ·) has a jump, and set
. .
∆uν (t, xα ) = uν (t, xα +) − uν (t, xα −), ∆f (uν (t, xα )) = f (uν (t, xα +)) − f (uν (t, xα −)).

n
Supp. φ
xα
Γj

0 x

Figure 20: Estimating the error in an approximate solution obtained by front tracking.

Observe that the polygonal lines x = xα (t) subdivide the strip [0, T ] × IR into finitely many
regions Γj where uν is constant (fig. 20). Introducing the vector
.
Φ = (φ · uν , φ · f (uν )),

36
by the divergence theorem the double integral in (4.23) can be written as
X ZZ XZ
div Φ(t, x) dtdx = Φ · n dσ. (4.24)
j Γj j ∂Γj

Here ∂Γj is the oriented boundary of Γj , while n denotes an outer normal. Observe that
ndσ = ±(ẋα , −1)dt along each polygonal line x = xα (t), while φ(t, x) = 0 along the lines
t = 0, t = T . By (4.24) the expression within square brackets in (4.23) is computed by
Z T Xh i
ẋα (t) · ∆uν (t, xα ) − ∆f (uν (t, xα )) φ(t, xα (t)) dt . (4.25)
0 α

Here, for each t ∈ [0, T ], the sum ranges over all fronts of uν (t, ·). To estimate the above
integral, let |σα | be the strength of the wave-front at xα . If this wave is a shock or or contact
discontinuity, by construction the Rankine-Hugoniot equations are satisfied exactly, i.e.

ẋα (t) · ∆uν (t, xα ) − ∆f (uν (t, xα )) = 0. (4.26)

On the other hand, if the wave at xα is a rarefaction front, its strength will satisfy σα ∈ ]0, εν [ .
Therefore, the error estimate (3.28) yields
¯ ¯
¯ ¯
¯ẋα (t) · ∆uν (t, xα ) − ∆f (uν (t, xα ))¯ = O(1) · |σα |2 ≤ C · εν |σα | (4.27)

for some constant C. Summing over all wave-fronts and recalling that the total strength of
waves in uν (t, ·) satisfies a uniform bound independent of t, ν, we obtain
¯ ¯
¯X h i ¯
¯ ¯
lim sup ¯ ẋα (t) · ∆uν (t, xα ) − ∆f (uν (t, xα )) φ(t, xα (t))¯
ν→∞ ¯ α ¯
³ ´ nX o (4.28)
≤ max |φ(t, x)| · lim sup C εν |σα |
t,x ν→∞
α∈R
= 0.

The limit (4.23) is now a consequence of (4.28). This shows that u is a weak solution to the
Cauchy problem.

To conclude the proof, assume that a convex entropy η is given, with entropy flux q. To prove
the admissibility condition (4.4) we need to show that
Z ∞Z ∞ n o
lim inf η(uν )ϕt + q(uν )ϕx dxdt ≥ 0 (4.29)
ν→∞ 0 −∞

for every nonnegative ϕ ∈ Cc1 with compact support contained in the half plane where t > 0.
Choose T > 0 so that ϕ vanishes outside the strip [0, T ] × IR. Using again the divergence
theorem, for every ν the double integral in (4.29) can be computed as
Z T Xh i
ẋα (t) · ∆η(uν (t, xα )) − ∆q(uν (t, xα )) ϕ(t, xα ) dt (4.30)
0 α

where, for each t ∈ [0, T ], the sum ranges over all jumps of uν (t, ·). We use here the notation
. .
∆η = η(uν (t, xα +)) − η(uν (t, xα −)), ∆q = q(uν (t, xα +)) − q(uν (t, xα −)).

37
To estimate the integral (4.30), let |σα | be the strength of the wave-front at xα . If this wave
is a shock or or contact discontinuity, by construction it satisfies the entropy condition

ẋα (t) · ∆η(uν (t, xα )) − ∆q(uν (t, xα )) ≥ 0. (4.31)

On the other hand, if the wave at xα is a rarefaction front, its strength will satisfy σα ∈ ]0, εν [ .
Therefore, the estimate (3.29) yields

ẋα (t) · ∆η(uν (t, xα )) − ∆q(uν (t, xα )) = O(1) · σα2 ≥ − C · εν |σα | (4.32)

for some constant C. Summing over all wave-fronts and recalling that the total strength of
waves in uν (t, ·) satisfies a uniform bound independent of t, ν, we obtain
Xh i
lim inf ẋα (t) · ∆η(uν (t, xα )) − ∆q(uν (t, xα )) ϕ(t, xα ) dt
ν→∞
α

µ ¶ ( ) (4.33)
X
≥ max ϕ(t, x) · lim inf − C εν |σα | ≥ 0.
t,x ν→∞
α∈R

The relation (4.29) now follows from (4.33).

4.5 A modified front tracking algorithm

For general n × n systems, the analysis given in the previous sections does not provide a
complete proof of Theorem 2, because the naive front tracking algorithm can generate an
infinite number of wave-fronts, within finite time. If this happens, the whole construction
breaks down.
accurate Riemann solver simplified Riemann solver

Figure 21: Two different ways of approximately solving the Riemann problem.

We thus need to modify the algorithm, to ensure that the total number of fronts remains
uniformly bounded. For this purpose, we shall use two different procedures for solving a
Riemann problem within the class of piecewise constant functions: the Accurate Riemann
Solver (fig. 21, left), which introduces several new wave-fronts, and a Simplified Riemann
Solver (fig. 21, right), which involves a minimum number of outgoing fronts. In this second
case, all new waves are lumped together in a single non-physical front, travelling with a fixed
speed λ̂ strictly larger than all characteristic speeds. The main feature of this algorithm is

38
Figure 22: The use of the simplified Riemann solver prevents the number of wave fronts
becoming infinite.

illustrated in fig. 22. If all Riemann problems are solved accurately, the number of wave-fronts
can become infinite (fig. 22, left). On the other hand, if at a certain point we use the Simplified
Riemann Solver, the total number of fronts remains bounded for all times (fig. 22, right).
The Accurate Riemann Solver was described at (4.12). Next, we introduce a simplified way
of solving a Riemann problem, with ≤ 3 outgoing fronts. Throughout the following, λ̂ will
denote a fixed constant, strictly larger than all characteristic speeds λj (u).

σ’ σ σ+σ’ σ

~u ~
r ur u~ r

ul ul ur ul ur
ur
um
um
um
σ σ’ σ
σ σ’

Figure 23: Three cases of the Simplified Riemann Solver.

Simplified Riemann Solver. Consider again the Riemann problem (4.6) at a point (t̄, x̄),
say generated by the interaction of two incoming fronts, of strength σ, σ 0 . We distinguish two
cases.

CASE 1: Let j, j 0 ∈ {1, . . . , n} be the families of the two incoming wave-fronts, with j ≥ j 0 .
Assume that the left, middle and right states ul , um , ur before the interaction are related by

um = Ψj (σ)(ul ), ur = Ψj 0 (σ 0 )(um ). (4.34)

Define the auxiliary right state

 0
 Ψj (σ) ◦ Ψj 0 (σ )(ul ) if j > j 0 ,
ũr = (4.35)

Ψj (σ + σ 0 )(ur ) if j = j 0 .
Let ṽ = ṽ(t, x) be the piecewise constant solution of the Riemann problem with data ul , ũr ,
constructed as in (4.12). Because of (4.25), the piecewise constant function ṽ contains exactly

39
two wave-fronts of sizes σ 0 , σ if j > j 0 , or a single wave-front of size σ + σ 0 if j = j 0 . It is
important to observe that ũr 6= ur , in general. We let the jump (ũr , ur ) travel with the fixed
speed λ̂, strictly bigger than all characteristic speeds. In a forward neighborhood of the point
(t̄, x̄), we thus define an approximate solution v as follows (fig. 23 left, center)


 ṽ(t, x) if x − x̄ < (t − t̄)λ̂,
v(t, x) = (4.36)


ur if x − x̄ > (t − t̄)λ̂.

Notice that this simplified Riemann solver introduces a new non-physical wave-front, joining
the states ũr , ur and travelling with constant speed λ̂. In turn, this front may interact with
other wave-fronts. One more case of interaction thus needs to be considered.

CASE 2: A non-physical front hits from the left a wave front of the i-the family (fig. 23
right), for some i ∈ {1, . . . , n}.
Let ul , um , ur be the left, middle and right state before the interaction. If

ur = Ψi (σ)(um ), (4.37)

define the auxiliary right state

ũr = Ψi (σ)(ul ). (4.38)
Call ṽ the solution to the Riemann problem with data ul , ũr , constructed as in (4.12). Because
of (4.38), ṽ will contain a single i-wave with size σ. Since ũr 6= ur in general, we let the jump
(ũr , ur ) travel with the fixed speed λ̂. In a forward neighborhood of the point (t̄, x̄), the
approximate solution u is then defined again according to (4.36).

By construction, all non-physical fronts travel with the same speed λ̂, hence they never interact
with each other. The above cases therefore cover all possible interactions between two wave-
fronts.

2 3
2 1
1

t 2

1 1
1

Figure 24: Generation order of various wave-fronts.

We observe that, in a non-physical front, the left and right states u− , u+ are essentially

40
arbitrary. Therefore, the Rankine-Hugoniot conditions will not be satisfied, not even approx-
imately. Calling
.
σ = |u+ − u− | (4.39)
the strength of the non-physical front, we only have the trivial estimate
h i
λ̂(u+ − u− ) − f (u+ ) − f (u− ) = O(1) · σ. (4.40)
To ensure that our modified front tracking approximations uν converge to a solution of the
original equations (4.1), the total strength of all non-physical fronts in each uν must approach
zero, as ν → ∞. This can be accomplished by carefully choosing which Riemann solver is
used at any given interaction.
For this purpose, to each wave-front we attach a generation order, counting how many inter-
actions were needed to produce it.

Definition 4.2. The generation order of a front is inductively defined as follows (fig. 24).
– All fronts generated by the Riemann problems at the initial time t = 0 are the “primal
ancestors” and have generation order k = 1.
– Let two incoming fronts interact, say of the families i, i0 ∈ {1, . . . , n}, with generation orders
k, k 0 . The orders of the outgoing fronts are then defined as follows.

CASE 1: i 6= i0 . Then
- the outgoing i-wave and i0 -wave have the same orders k, k 0 as the incoming ones.
- the outgoing fronts of every other family j 6= i, i0 have order max{k, k 0 } + 1.

CASE 2: i = i0 . Then
- the outgoing front of the i-th family has order min{k, k 0 },
- the outgoing fronts of every family j 6= i have order max{k, k 0 } + 1.

In the above, we tacitly assumed that at most two incoming wave-fronts interact at any given
time. This can always be achieved by an arbitrarily small change in the speed of the fronts.
We shall also adopt the following important provision:

(P)In every Riemann Solver, rarefaction waves of the same family as one of the incoming
fronts are never partitioned

In other words, if one of the incoming fronts belongs to the i-th family, we set pi = 1 regardless
of the strength of the i-th outgoing front. This guarantees that every wave-front can be
uniquely continued forward in time, unless it gets completely cancelled by interacting with
other fronts of the same family and opposite sign.

Given an integer ν ≥ 1, to construct the approximate solution uν (·), at each point of interaction
we choose:

41
• The Accurate Riemann Solver, if the two incoming fronts both have generation order < ν.

• The Simplified Riemann Solver, if one of the incoming fronts has generation order ≥ ν, or
if it is a non-physical front.

By implementing the above algorithm, one obtains a globally defined front tracking solution
to the Cauchy problem. For future use, we record the following definition.

Definition 2. Given ε > 0, we say that u : [0, ∞[ 7→ L1 (IR; IRn ) is anε-approximate front
tracking solution of (4.1) if the following holds:

1. As a function of two variables, u = u(t, x) is piecewise constant, with discontinuities

occurring along finitely many lines in the t-x plane. Only finitely many wave-front interactions
occur, each involving exactly two incoming fronts. Jumps can be of three types: shocks (or
contact discontinuities), rarefactions and non-physical waves, denoted as J = S ∪ R ∪ N P.

.
2. Along each shock (or contact discontinuity) x = xα (t), α ∈ S, the values u− = u(t, xα −)
.
and u+ = u(t, xα +) are related by
u+ = Skα (σα )(u− ), (4.41)
for some kα ∈ {1, . . . , n} and some wave size σα . If the kα -th family is genuinely nonlinear,
then the entropy admissibility condition σα < 0 also holds. Moreover, the speed of the shock
front satisfies
|ẋα − λkα (u+ , u− )| ≤ ε. (4.42)

3. Along each rarefaction front x = xα (t), α ∈ R, one has

u+ = Rkα (σα )(u− ), σα ∈ ]0, ε] (4.43)
for some genuinely nonlinear family kα . Moreover,
|ẋα (t) − λkα (u+ )| ≤ ε. (4.44)

4. All non-physical fronts x = xα (t), α ∈ N P have the same speed:

ẋα (t) ≡ λ̂, (4.45)
where λ̂ is a fixed constant strictly greater than all characteristic speeds. The total strength of
all non-physical fronts in u(t, ·) remains uniformly small, namely
X
|u(t, xα +) − u(t, xα −)| ≤ ε for all t ≥ 0. (4.46)
α∈N P

If, in addition, the initial value of u satisfies

ku(0, ·) − ūkL1 < ε, (4.47)

42
we say that u is an ε-approximate solution to the Cauchy problem (4.1)-2.2).

Lemma 5. Let the assumptions of Theorem 2 hold. If the initial data ū has sufficiently small
total variation, then for every ε > 0 the Cauchy problem (4.1)-(4.2) admits an ε-approximate
front tracking solution.

Using the above lemma, we construct a sequence of front tracking approximate solutions uν ,
with εν → 0, having uniformly small total variation. As shown in Section 4.4, by taking a
subsequence uν converging to some function u = u(t, x) in L1loc , one obtains an entropy-weak
solution to the Cauchy problem. For further details we refer to [9].

5 A semigroup of solutions

We always assume that the system (4.1) is strictly hyperbolic, and satisfies the hypotheses
(H), so that each characteristic field is either linearly degenerate or genuinely nonlinear. The
analysis in the previous chapter has shown the existence of a global entropy weak solution
of the Cauchy problem for every initial data with sufficiently small total variation. More
precisely, recalling the definitions (4.13)-(4.14), consider a domain of the form
n o
.
D = cl u ∈ L1 (IR; IRn ); u is piecewise constant, Υ(u) = V (u) + C0 · Q(u) < δ0 , (5.1)

where cl denotes closure in L1 . With a suitable choice of the constants C0 and δ0 > 0, the
proof of Theorem 2 for every ū ∈ D, one can construct a sequence of ε-approximate front
tracking solutions converging to a weak solution u taking values inside D. Observe that, since
the proof of convergence relied on a compactness argument, no information was obtained on
the uniqueness of the limit. The main goal of the present chapter is to show that this limit is
unique and depends continuously on the initial data. In the scalar case, this follows from the
classical results of Volpert [53] and Kruzhkov [32]. For strictly hyperbolic systems we have:

Theorem 3. For every ū ∈ D, as ε → 0 every sequence of ε-approximate solutions uε :

[0, ∞[ 7→ D of the Cauchy problem (4.1)-(4.2), obtained by the front tracking method, converges
.
to a unique limit solution u : [0, ∞[ 7→ D. The map (ū, t) 7→ u(t, ·) = St ū is a uniformly
Lipschitz semigroup, i.e.:
S0 ū = ū, Ss (St ū) = Ss+t ū, (5.2)
kSt ū − Ss v̄kL1 ≤ L · (kū − v̄kL1 + |t − s|) for all ū, v̄ ∈ D, s, t ≥ 0. (5.3)

This result was first proved in [11] for 2 × 2 systems, then in [12] for general n × n systems,
relying on a (lengthy and technical) homotopy method. Relying on some original ideas intro-
duced by T.P.Liu and T.Yang in [41, 42], the paper [16] provided a much simpler proof of the
continuous dependence result, which will be described here. An extension of the above result
to initial-boundary value problems for hyperbolic conservation laws has recently appeared in
[25]. All of the above results deal with solutions having small total variation. The existence

43
of solutions, and the well posedness of the Cauchy problem for large BV data was studied
respectively in [46] and in [34].

To prove the uniqueness of the limit of front tracking approximations, we need to estimate
the distance between any two ε-approximate solutions u, v of (4.1). For this purpose we
introduce a functional Φ = Φ(u, v), uniformly equivalent to the L1 distance, which is “almost
decreasing” along pairs of solutions. Recalling the construction of shock curves at (3.11), given
two piecewise constant functions u, v : IR 7→ Rn , we consider the scalar functions qi defined
implicitly by
v(x) = Sn (qn (x)) ◦ · · · ◦ S1 (q1 (x))(u(x)). (5.4)

ω3 = v(x)
λ3 σα xα
q v
3
ω2

λ1 ω1 u
q
1
ω0= u(x)
x xα

Figure 25: Decomposing a jump (u(x), v(x)) in terms of n (possibly non-admissible) shocks.

Remark 5. If we wanted to solve the Riemann problem with data u− = u(x) and u+ = v(x)
only in terms of shock waves (possibly not entropy-admissible), then q1 (x), . . . , qn (x) would
be the sizes of these shocks (fig. 25). It is thus useful to think of qi (x) as the strength of the
i-th component in the jump (u(x), v(x)).
On a compact neighborhood of the origin, we clearly have
Xn
1
· |v(x) − u(x)| ≤ |qi (x)| ≤ C1 · |v(x) − u(x)| (5.5)
C1 i=1

for some constant C1 . We now consider the functional

n Z
. X ∞
Φ(u, v) = |qi (x)|Wi (x) dx, (5.6)
i=1 −∞

where the weights Wi are defined by setting:

.
Wi (x) = 1 + κ1 · [total strength of waves in u and in v which approach the i-wave qi (x)]
+κ2 · [wave interaction potentials of u and of v]
.
= 1 + κ1 Ai (x) + κ2 [Q(u) + Q(v)].
(5.7)
The amount of waves approaching qi (x) is defined as follows. If the i-shock and i-rarefaction
curves coincide, we simply take
 
. X X
Ai (x) =  +  |σα |. (5.8)
xα <x, i<kα ≤n xα >x, 1≤kα <i

44
The summations here extend to waves both of u and of v. The definition (5.8) applies if
the i-th field is linearly degenerate, or else if all i-rarefaction curves are straight lines. On
the other hand, if the i-th field is genuinely nonlinear with shock and rarefactions curves not
coinciding, our definition of Ai will contain an additional term, accounting for waves in u and
in v of the same i-th family:
 
.  X X 
Ai (x) =  +  |σα |
α∈J (u)∪J (v) α∈J (u)∪J (v)
xα <x, i<kα ≤n xα >x, 1≤kα <i

  



  X X 

  +  |σα | if qi (x) < 0, (5.9)



 kα =i kα =i

 α∈J (u), xα <x α∈J (v), xα >x

+  





  X X 

  +  |σα | if qi (x) > 0.



 kα =i kα =i
α∈J (v), xα <x α∈J (u), xα >x

Here and in the sequel, J (u) and J (v) denote the sets of all jumps in u and in v, while
.
J = J (u) ∪ J (v). We recall that kα ∈ {1, . . . , n + 1} is the family of the jump located at xα
with size σα . Notice that the strengths of non-physical waves do enter in the definition of Q.
Indeed, a non-physical front located at xα approaches all shock and rarefaction fronts located
at points xβ > xα . On the other hand, non-physical fronts play no role in the definition of Ai .
The values of the large constants κ1 , κ2 in (5.7) will be specified later. Observe that, as soon
as these constants have been assigned, we can then impose a suitably small bound on the total
variation of u, v so that
1 ≤ Wi (x) ≤ 2 for all i, x. (5.10)
From (5.5), (5.6) and (5.10) it thus follows
1
· kv − ukL1 ≤ Φ(u, v) ≤ 2C1 · kv − ukL1 . (5.11)
C1
.
Recalling the definition Υ(u) = V (u) + C0 Q(u), the basic L1 stability estimate for front
tracking approximations can now be stated as follows.

Theorem 4. For suitable constants C2 , κ1 , κ2 , δ0 > 0 the following holds. Let u, v be ε-

approximate front tracking solutions of (4.1), with small total variation, so that
Υ(u(t)) < δ0 , Υ(v(t)) < δ0 for all t ≥ 0. (5.12)
Then the functional Φ in (5.6)–(5.9) satisfies
Φ(u(t), v(t)) − Φ(u(s), v(s)) ≤ C2 ε(t − s) for all 0 ≤ s < t. (5.13)

Relying on the above estimate, a proof of Theorem 3 can be easily worked out. Indeed,
let ū ∈ D be given. Consider any sequence {uν }ν≥1 , such that each uν is a front tracking
εν -approximate solution of Cauchy problem (4.1)-(4.2), with
lim εν = 0, Υ(uν (t)) < δ0 for all t ≥ 0, ν ≥ 1.
ν→∞

45
For every µ, ν ≥ 1 and t ≥ 0, by (5.11) and (5.13) it now follows

kuµ (t) − uν (t)kL1 ≤ C1 · Φ(uµ (t), uν (t))

h i
≤ C1 · Φ(uµ (0), uν (0)) + C2 t · max{εµ , εν } (5.14)

≤ 2C12 kuµ (0) − uν (0)kL1 + C1 C2 t · max{εµ , εν }.

Since the right hand side of (5.14) approaches zero as µ, ν → ∞, the sequence is Cauchy
and converges to a unique limit. The semigroup property (5.2) is an immediate consequence
of uniqueness. Finally, let ū, v̄ ∈ D be given. For each ν ≥ 1, let uν , vν be front tracking
εν -approximate solutions of (2.1) with

kuν (0) − ūkL1 < εν , kvν (0) − v̄kL1 < εν , lim εν = 0. (5.15)
ν→∞

Using again (5.11) and (5.13) we deduce

kuν (t) − vν (t)kL1 ≤ C1 · Φ(u

h ν (t), vν (t)) i
≤ C1 · Φ(uν (0), vν (0)) + C2 tεν (5.16)
≤ 2C12 kuν (0) − vν (0)kL1 + C1 C2 tεν .

Letting ν → ∞, by (5.15) it follows

ku(t) − v(t)kL1 ≤ 2C12 · kū − v̄kL1 . (5.17)

This establishes the uniform Lipschitz continuity of the semigroup with respect to the initial
data. Recalling (4.21), the Lipschitz continuity with respect to time is clear. This completes
the proof of Theorem 3.

In the remainder of this chapter we will sketch the main ideas in the proof of Theorem 4. The
key point is to understand how the functional Φ evolves in time. In connection with (5.4), at
each x define the intermediate states ω0 (x) = u(x), ω1 (x), . . . , ωn (x) = v(x) by setting
.
ωi (x) = Si (qi (x)) ◦ Si−1 (qi−1 (x)) ◦ · · · ◦ S1 (q1 (x))(u(x)). (5.18)

As remarked earlier, these are the intermediate states appearing in the solution of the Riemann
problem with data u− = u(x) and u+ = v(x) obtained in terms of n (possibly not entropy-
admissible) shock waves (fig. 25). Moreover, call
.
λi (x) = λi (ωi−1 (x), ωi (x)) (5.19)

the speed of the i-shock connecting ωi−1 (x) with ωi (x). A direct computation now yields

XX n n o
d
Φ(u(t), v(t)) = |qi (xα −)|Wi (xα −) − |qi (xα +)|Wi (xα +) · ẋα
dt α∈J i=1
(5.20)
n n
XX o
= |qiα+ |Wiα+ (λα+
i − ẋα ) − |qiα− |Wiα− (λα−
i − ẋα ) ,
α∈J i=1

46
with obvious meaning of notations. We regard the quantity |qi (x)|λi (x) as the flux of the i-th
component of |v − u| at x. For xα−1 < x < xα , one clearly has
(α−1)+ (α−1)+ (α−1)+
|qi | λi Wi = |qi (x)|λi (x)Wi (x) = |qiα− | λα− α−
i Wi .

Since for each t > 0 the functions u(t, ·), v(t, ·) are in L1 and piecewise constant, we must have
u(t, x) = v(t, x) = 0, qi (t, x) = 0 for all x outside a bounded interval. This allowed us to add
and subtract the above terms in (5.20), without changing the overall sum.
In connection with (5.20), for each jump point α ∈ J and every i = 1, . . . , n, define
.
Eα,i = |qiα+ |Wiα+ (λα+
i − ẋα ) − |qiα− |Wiα− (λα−
i − ẋα ). (5.21)

Our main goal will be to establish the bounds

n
X
Eα,i ≤ O(1) · |σα | α ∈ N P, (5.22)
i=1

n
X
Eα,i ≤ O(1) · ε|σα | α ∈ R ∪ S. (5.23)
i=1

Here R, S, and N P denote respectively the sets of rarefaction, shock, and non-physical fronts
in u and v. As usual, by the Landau symbol O(1) we denote a quantity whose absolute value
satisfies a uniform bound, depending only on the system (2.1). In particular, this bound does
not depend on ε or on the functions u, v. It is also independent of the choice of the constants
κ1 , κ2 in (5.7).
From (5.22)-(5.23), recalling (2.9) and the uniform bounds (5.12) on the total strength of
waves, one obtains the key estimate
d
Φ(u(t), v(t)) ≤ O(1) · ε. (5.24)
dt
If the constant κ2 in (5.7) is chosen large enough, by the interaction estimates (3.31), (3.33)
all weight functions Wi (x) will decrease at each time τ where two fronts of u or two fronts of
v interact. Integrating (5.24) over the interval [s, t] we therefore obtain

Φ(u(t), v(t)) ≤ Φ(u(s), v(s)) + O(1) · ε(t − s), (5.25)

proving the theorem.

All the remaining work is thus aimed at establishing (5.22)-(5.23).

If α ∈ N P, calling σα the strength of this jump as in (4.39), for i = 1, . . . , n one has the easy
estimates
qiα+ − qiα− = O(1) · σα ,

λα+
i − λα−
i = O(1) · σα ,

Wiα+ − Wiα− = 0 if qiα+ · qiα− > 0.

47
Therefore, writing
³ ´
Eα,i = |qiα+ | − |qiα− | Wiα+ (λα+
i − ẋα ) + |qiα− |(Wiα+ − Wiα− )(λα+
i − ẋα )
α− α− α+ α−
(5.26)
+ |qi |Wi (λi − λi ),

the estimate (5.22) is clear.

The proof of (5.23) requires more work. Instead of writing down all computations, we will try
to convey the main ideas with the help of a few pictures. For all details we refer to [9].
Given two piecewise constant functions u, v with compact support, for i = 1, . . . , n we can
define the scalar components ui , vi by induction on the jump points of u, v. We start by setting
ui (−∞) = vi (−∞) = 0. If xα ∈ J (u) is a jump point of u, then we let vi be constant across
xα and set
.
ui (xα +) = ui (xα −) − [qi (xα +) − qi (xα −)].
On the other hand, if xα ∈ J (v) is a jump point of v, then we let ui be constant across xα
and set
.
vi (xα +) = vi (xα −) + [qi (xα +) − qi (xα −)].
These definitions trivially imply

qi (x) = vi (x) − ui (x) for all x ∈ IR, i = 1, . . . , n.

Observe that the functional Φ in (5.6)-(5.7) can also be written in the equivalent form
n Z ∞
X
Φ(u, v) = [1 + κ2 Q(u) + κ2 Q(v)] · |qi (x)| dx
i=1 −∞
(5.27)
X XZ
+ κ1 · |σα | · |qi (x)| dx .
i qi (x) approaches σα
σα ∈J (u)∪J (v)

According to the definition (5.9), the i-waves in u and v which approach qi (x) are those located
within the thick portions of the graphs of ui , vi in fig. 26. Viceversa, for a given i-wave σα
located at xα , the regions where the jumps qi (x) approach σα are represented by the shaded
areas in in fig. 27.

vi

q
i
ui

Figure 26: The i-waves in u and in v which are approaching the jump qi (x).

48
 σα

ui

vi

xα

Figure 27: The jumps qi in the shaded regions are the ones which approach the i-wave σα .

Now let v have a wave-front at xα with strength σα , in the genuinely nonlinear k-th family.
To fix the ideas, assume that vk (xα ±) > uk (xα ). In connection with this front, for every i < k
the functional Φ(u, v) in (5.27) contains a term of the form (fig. 28)
.
Aα,i = κ1 · |σα | · [area of the region between the graphs of ui and vi , to the right of xα ].

vk
σ
α

uk

vi

q
ui i

x
α

Figure 28: The jumps qi in the shaded regions are the ones which approach the k-wave σα .

By strict hyperbolicity, the i-th and k-th characteristic speeds are strictly separated, say
λk − λi ≥ c > 0. If each component ui , vi , i = 1, . . . , n, were an exact solution to a scalar
conservation law:
(ui )t + Fi (ui )x = 0, (Fi0 = λi )
uncoupled from all the other components, then we would have the estimate
dAα,i
≤ − κ1 |σα ||qiα+ |(ẋα − λα+ α+
i ) ≤ − cκ1 |σα ||qi |. (5.28)
dt
.
Here λα+
i = λi (ui (xα +), vi (xα +)) is a speed of an i-shock of strength qiα+ . In general, the
estimate (5.28) must be supplemented with coupling and error terms, whose size is estimated
as  
X
O(1) · ε + |qkα+ |(|qkα+ | + |σα |) + |qjα+ | |σα |. (5.29)
j6=k
A detailed computation thus yields
 
X
Eα,i ≤ O(1)· ε + |qkα+ |(|qkα+ | + |σα |) + |qjα+ | |σα |−cκ1 |qiα+ ||σα | i 6= k. (5.30)
j6=k

49
Next, according to (5.9) the functional Φ(u, v) in (5.27) also contains a term of the form
.
Aα,k = κ1 · |σα | · [area of the region between the graphs of uk and vk , to the right of xα ].

xα vk vk
σα
P P
α−
qα− λ k
k

Q uk Q P’ uk
x

Figure 29: Due to genuine nonlinearity, the total amount of qk approaching the k-wave σα
(shaded region) decreases in time.

If the components uk , vk were exact solutions of a genuinely nonlinear scalar conservation law,
say
(uk )t + Fk (uk )x = 0, (Fk0 = λk ) (5.31)
with Fk00 ≥ c0 > 0, then one would have the estimate

dAα,k c0 α+
≤ − κ1 |σα | |qkα+ |(ẋα − λα+
k ) ≤ − κ |σ
1 α | |qk
α+
| · (|q | + |σα |). (5.32)
dt 2 k
The decrease of this area is illustrated in fig. 29. In general, the estimate (5.32) must be
supplemented with coupling and error terms, whose size is again estimated as (5.29). A
detailed computation thus yields
 
³ ´ X c0 κ1 α+ ³ ´
Eα,k ≤ O(1)·ε + |qkα+ | |qkα+ | + |σα | + |qjα+ | |σα |− |qk ||σα | |qkα+ |+|σα | . (5.33)
j6=k
2

Choosing κ1 sufficiently large, (5.30) and (5.33) together yield (5.23).

A different estimate is needed in the case where the jump in vk crosses the graph of uk , say
vk (xα +) < uk (xα ) < vk (xα −). To fix the ideas, assume

|qkα+ | = |vk (xα +) − uk (xα )| ≤ |vk (xα −) − uk (xα )| = |qkα− |. (5.34)

In this case, the estimates (5.30) remain valid. In connection with the k-th field, the functional
Φ contains a term of the form (fig. 30):
.
Aα,k = [area of the region between the graphs of uk and vk ], (5.35)

where the above area includes points both to the right and to the left of xα .
If the components uk , vk provided an exact solution to the genuinely nonlinear scalar conser-
vation law (5.31), due to genuine nonlinearity we would have
dAα,k c0 c0
≤ − |qkα− | · |qkα+ | ≤ − |σα ||qkα+ |. (5.36)
dt 2 4

50
 vk σα
α−
λ
qα− k
k
xα
uk
α+
qα+ λk
k

xα

Figure 30: Due to genuine nonlinearity, also in this case the total amount of qk approaching
the k-wave σα (shaded area) decreases in time.

Indeed, by (5.34),
|σα | = |qkα− | + |qkα+ | ≤ 2|qkα− |.
In general, the estimate (5.36) must be supplemented with coupling and error terms, whose
size is again estimated as (5.29). A detailed computation thus yields
 
X c0 α+
Eα,k ≤ O(1) · ε + |qkα+ |(|qkα+ | + |σα |) + |qjα+ | |σα | − |q ||σα |. (5.37)
j6=k
4 k

Assuming that the total strength of waves remains sufficiently small, we have
c0
|qkα+ | + |σα | << ,
4
hence (5.30) and (5.37) together yield (5.23). For details we again refer to [9].

6 Uniqueness of solutions.

According to the analysis in the previous chapters, the solution of the Cauchy problem (4.1)-
(4.2) obtained as limit of front tracking approximations is unique and depends Lipschitz con-
tinuously on the initial data, in the L1 norm. This basic result, however, leaves open the
question whether other weak solutions may exist, possibly constructed by different approxi-
mation algorithms. We will show that this is not the case: indeed, every entropy admissible
solution, satisfying some minimal regularity assumptions, necessarily coincides with the one
obtained as limit of front tracking approximations.
As a first step, we estimate the distance between an approximate solution, obtained by the
front tracking method, and the exact solution of the Cauchy problem (4.1)-(4.2), given by
the semigroup trajectory t 7→ u(t, ·) = St ū . Let uε : [0, T ] 7→ D be an ε-approximate front
tracking solution, according to Definition 2. We claim that the corresponding error can then
be estimated as
kuε (T, ·) − ST ūkL1 = O(1) · ε(1 + T ). (6.1)

51
To see this, we first estimate the limit

kuε (τ + h) − Sh uε (τ )kL1
lim
h→0+ h
at any time τ ∈ [0, T ] where no wave-front interaction takes place. Let u(τ, ·) have jumps
at points x1 < · · · < xN . Call S the set of indices α ∈ {1, . . . , N } such that u(τ, xα −) and
u(τ, xα +) are connected by a shock or by a contact discontinuity, so that (2.4)-(2.5) hold.
Moreover, call R the set of indices α corresponding to a rarefaction wave of a genuinely
nonlinear family, so that (2.6)-(2.7) hold. Finally, call N P the set of indices α corresponding
to non-physical fronts.

S R
S
NP

τ xα (τ )

Figure 31: The exact solution (dotted lines) which, at time τ , coincides with the value of a
piecewise constant front tracking approximation.

For each α, call ωα the self-similar solution of the Riemann problem with data u± = u(τ, xα ±) .
We observe that, for h > 0 small enough, the semigroup trajectory h 7→ Sh u(τ ) is obtained
by piecing together the solutions of these Riemann problems (fig. 31). Using the properties
(4.41)–(4.46) one now derives the bound

ku(τ + h) − Seh u(τ )kL1

lim
h→0+ h
X µ Z xα +ρ ¶
1
= lim |u(τ + h, x) − ωα (h, x − xα )| dx
h→0+ h xα −ρ
α∈R∪S∪N P (6.2)
X X ³ ´ X
= O(1) · ε|σα | + O(1) · |σα | |σα | + ε + O(1) · |σα |
α∈S α∈R α∈N P

= O(1) · ε .

Here ρ can be any suitably small positive number. From the assumption (4.47), the bound
(6.2) and the error formula (10.6) in the Appendix, we finally obtain

kuε (T, ·) − ST ūkL1 ≤ kST uε (0, ·) − ST ūkL1 + kuε (T, ·) − ST uε (0, ·)kL1
Z T½ ¾
kuε (τ + h) − Sh uε (τ )kL1
≤ L · kuε (0, ·) − ūkL1 + L · lim inf dτ
0 h→0+ h

= O(1) · ε + O(1) · εT.

52
We shall now state our main uniqueness theorem. For sake of clarity, a complete set of
assumptions is listed below.

(A1) (Conservation Equations) The function u = u(t, x) is a weak solution of the Cauchy
problem (4.1)-(4.2), taking values within the domain D of a semigroup S. More precisely,
u : [0, T ] 7→ D is continuous w.r.t. the L1 distance. The initial condition (2.2) holds, together
with ZZ
(uϕt + f (u)ϕx ) dxdt = 0 (6.3)

for every C 1 function ϕ with compact support contained inside the open strip ]0, T [ ×IR.

(A2) (Lax Entropy Conditions) Let u have an approximate jump discontinuity at some
point (τ, ξ) ∈ ]0, T [×IR. More precisely, let there exists states u− , u+ ∈ Ω and a speed λ ∈ IR
such that, calling ½
. u− if x < λt,
U (t, x) = (6.4)
u+ if x > λt,
there holds Z r Z r ¯ ¯
1 ¯ ¯
lim ¯u(τ + t, ξ + x) − U (t, x)¯ dxdt = 0. (6.5)
r→0+ r 2 −r −r

Then, for some i ∈ {1, . . . , n}, one has the inequalities

λi (u− ) ≥ λ ≥ λi (u+ ). (6.6)

(A3) (Tame Oscillation Condition) For some constants C, λ̂ the following holds. For
every point x ∈ IR and every t, h > 0 one has
n o
|u(t + h, x) − u(t, x)| ≤ C · Tot.Var. u(t, ·) ; [x − λ̂h, x + λ̂h] . (6.7)

(A4) (Bounded Variation Condition) There exists δ > 0 such that, for every space-like
curve {t = τ (x)} with |dτ /dx| ≤ δ a.e., the function x 7→ u(τ (x), x) has locally bounded
variation.

γ
t+h

Figure 32: The tame oscillation and the bounded variation condition.

Remark 6. The condition (A3) restricts the oscillation of the solution. An equivalent, more
intuitive formulation is the following. For some constant λ̂ larger than all characteristic speeds,

53
 .
given any interval [a, b] and t ≥ 0, the oscillation of u on the triangle (fig. 32) ∆ = {(s, y) : s ≥
t, a + λ̂(s − t) < y < b − λ̂(s − t)}, defined as
.
Osc{u; ∆} = sup |u(s, y) − u(s0 , y 0 )|,
(s,y),(s0 ,y 0 )∈∆

is bounded by a constant multiple of the total variation of u(t, ·) on [a, b].

Assumption (A4) simply requires that, for some fixed δ > 0, the function u has bounded
variation along every space-like curve γ which is “almost horizontal” (fig. 32). Indeed, the
condition is imposed only along curves of the form {t = τ (x); x ∈ [a, b]} with

|τ (x) − τ (x0 )| ≤ δ|x − x0 | for all x, x0 ∈ [a, b].

One can prove that all of the above assumptions are satisfied by weak solutions obtained as
limits of Glimm or wave-front tracking approximations [9]. The following result shows that
the entropy weak solution of the Cauchy problem (4.1)-(4.2) is unique within the class of
functions that satisfy either the additional regularity condition (A3), or (A4).

Theorem 5. Let the map u : [0, T ] 7→ D be continuous (w.r.t. the L1 distance), taking values
in the domain of the semigroup S generated by the system (4.1). If (A1), (A2) and (A3) hold,
then
u(t, ·) = St ū for all t ∈ [0, T ]. (6.8)
In particular, the weak solution that satisfies these conditions is unique. The same conclusion
holds if the assumption (A3) is replaced by (A4).

The main steps of the proof are given below.

1. Since u takes values inside the domain D of the semigroup, the total variation of u(t, ·)
remains uniformly bounded. From the basic equation (4.1), it follows that u is Lipschitz
continuous with values in L1 , namely

ku(t) − u(s)kL1 ≤ L · |t − s| (6.9)

for some Lipschitz constant L. More precisely, if M and λ∗ are constants such that

Tot.Var.{u} ≤ M for all u ∈ D,

|f (ω) − f (ω 0 )| ≤ λ∗ |ω − ω 0 | whenever |ω|, |ω 0 | ≤ M,

.
as Lipschitz constant in (6.9) one can take L = λ∗ M . As a consequence, u = u(t, x) can
be regarded as a BV function of the two variables t, x, in the sense that the distributional
derivatives Dt u, Dx u are Radon measures. By a well known structure theorem [26], there
f ⊂ ]0, T [ ×IR of 1-dimensional Hausdorff measure zero such that, at every point
exists a set N
(τ, ξ) ∈ f, u either is approximately continuous or has an approximate jump discontinuity.
/ N
Taking the projection of N f on the t-axis, we conclude that there exists a set N ⊂ [0, T ] of
measure zero, containing the endpoints 0 and T , such that, at every point (τ, ξ) ∈ [0, T ] × IR
. .
with τ ∈/ N , setting u− = u(τ, ξ−), u+ = u(τ, ξ+) the following property holds.

54
(P) Either u+ = u− , in which case (6.4)-(6.5) hold with λ arbitrary. Or else u+ 6= u− , in which
case (6.4)-(6.5) hold for some particular λ ∈ IR. In this second case, for some i ∈ {1, . . . , n}
the Rankine-Hugoniot equations and the Lax entropy condition hold:

λi (u− , u+ ) · (u+ − u− ) = f (u+ ) − f (u− ), λi (u+ ) ≤ λi (u− , u+ ) ≤ λi (u− ). (6.10)

2. According to Theorem A.2 in the Appendix, at any fixed time T > 0 the distance between
the solution u(T ) and the unique semigroup trajectory ST ū is estimated by
Z T½ ¾
kuε (τ + h) − Sh uε (τ )kL1
kuε (T, ·)−ST ūkL1 ≤ L·kuε (0, ·)−ūkL1 +L· lim inf dτ . (6.11)
0 h→0+ h
We will establish (6.8) by showing that the integrand on the right hand side of (6.11) vanishes
at each time t ∈ / N . Because of the finite speed of propagation, it actually suffices to show
that, for each t ∈
/ N , ε > 0 and every interval [a, b], there holds
Z b¯ ¯
1 ¯ ¯
lim sup ¯u(t + h, x) − (Sh u(t))(x)¯ dx = O(1) · ε. (6.12)
h→0+ h a

3. Let u = u(t, x) be as in Theorem 5. In the following, for any given point (τ, ξ), we denote
]
by U ] = U(τ,ξ) the solution of the Riemann problem
½ + .
u = u(τ, ξ+) if x > ξ,
wt + f (w)x = 0, w(τ, x) = . (6.13)
u− = u(τ, ξ−) if x < ξ.

By the property (P), apart from the trivial case where u+ = u− , this solution consists of
a single entropy admissible shock. The function U ] provides a good approximation to the
solution u in a forward neighborhood of the point (τ, ξ). More precisely, using the Lipschitz
continuity (6.9), one can prove that
Z ξ+hλ̂ ¯ ¯
1 ¯ ] ¯
lim ¯u(τ + h, x) − U(τ,ξ) (τ + h, x)¯ dx = 0, (6.14)
h→0+ h ξ−hλ̂

for every λ̂ > 0 and every (τ, ξ) with τ ∈

/ N.

4. Next, for a given point (τ, ξ) we denote by U [ = U(τ,ξ)

[ the solution of the linear system
with constant coefficients
e x = 0,
wt + Aw w(τ, x) = u(τ, x), (6.15)
.
where Ae = A(u(τ, ξ)). As in the previous step, we need to estimate the difference between u
and U [ , in a forward neighborhood of the point (τ, ξ). Consider any open interval ]a, b[ con-
taining the point ξ and fix a speed λ̂ strictly larger than the absolute values of all characteristic
speeds. For t ≥ τ define the open intervals
.
J(t) = ]a + (t − τ )λ̂, b − (t − τ )λ̂[ (6.16)

and the region

.
Γ(t) = {(s, x); s ∈ [τ, t], x ∈ J(s)}. (6.17)

55
With the above notation, for every τ 0 ≥ τ one can prove the estimate
Z ¯ ¯ Z τ0
¯ ¯
¯u(τ 0 , x)−U [ (τ 0 , x)¯ dx = O(1)· sup |u(t, x)−u(τ, ξ)|· Tot.Var.{u(t, ·); J(t)} dt.
J(τ 0 ) (t,x)∈Γ(τ 0 ) τ
(6.18)

5. Given τ ∈ / N , ε > 0 and a < b, using either one of the assumptions (A3) or (A4) we
can cover a neighborhood of the interval [a, b] with finitely many points ξi and open intervals
.
Ij =]aj , bj [ such that the following conditions hold (fig. 33).
(i) Each point x is contained in at most two of the open intervals Ij .
(ii) The total variation of u(τ, ·) on each Ij is ≤ ε.
.
(iii) For some τ 0 > τ , calling ζj = (aj + bj )/2 and
. n o
Γj = (t, x); t ∈ [τ, τ 0 ], aj + (t − τ )λ̂ < x < bj − (t − τ )λ̂ ,

there holds
sup |u(t, x) − u(τ, ζj )| ≤ ε. (6.19)
(t,x)∈Γj

τ+ h Γj
τ
a aj ζ bj ξ b
j i

Figure 33: Covering the interval [a, b] with subintervals where the solutions u(τ + h) and
Sh u(τ ) can both be compared with the approximations U ] or U [ .

]
6. We now construct a function U = U (t, x) which coincides with U(τ,ξ i)
near each point (τ, ξi )
[
and with U(τ,ζ in a forward neighborhood of each point (τ, ζj ).
j)

 ]

 U(τ,ξi ) (t, x) if |x − ξi | ≤ (t − τ )λ̂,
.
U (t, x) =

 U[ S
(τ,ζj ) (t, x) if (t, x) ∈ Γj \ k<j Γk .

By (6.14) and (6.18), this function U provides a good approximation of u, for times t = τ + h

56
with h > 0 small. Indeed, recalling (6.19) and the property (i) of the covering, we have
Z b¯ ¯
1 ¯ ¯
lim sup ¯u(τ + h, x) − U (τ + h, x)¯ dx
h→0+ h a
X Z ξi +hλ̂ ¯ ¯
1 ¯ ] ¯
≤ lim sup ¯u(τ + h, x) − U(τ,ξi ) (τ + h, x)¯ dx
i h→0+ h ξi −hλ̂

X Z bj −hλ̂ ¯ ¯
1 ¯ [ ¯
+ lim sup ¯u(τ + h, x) − U(τ,ζ j)
(τ + h, x)¯ dx
j h→0+ h aj +hλ̂

  (6.20)
 X ε Z τ +h 
≤ 0 + lim sup O(1) · Tot.Var.{u(t); Ij } dt
h→0+  h τ 
j

( Z τ +h )
ε
≤ lim sup O(1) · 2 · Tot.Var.{u(t); IR} dt
h→0+ h τ

= O(1) · ε.

.
7. We now observe that the semigroup trajectory v(t, ·) = St ū is also an entropy weak solution
of the Cauchy problem (4.1)-(4.2), and satisfies all the assumptions (A1)–(A3). In particular,
the total variation of v(t, ·) remains uniformly bounded, and its oscillation on each domain
Γ(t) of the form (6.17) is bounded by

sup |v(t, x) − v(τ, ξ)| = O(1) · Tot.Var.{v(τ ); ]a, b[ }. (6.21)

(t,x)∈Γ(τ 0 )

As a consequence, we can repeat the estimate (6.20) with v in the role of u and obtain
Z b¯ ¯
1 ¯ ¯
lim sup ¯v(τ + h, x) − U (τ + h, x)¯ dx = O(1) · ε. (6.22)
h→0+ h a

Together, (6.20) and (6.22) imply (6.12). Since ε > 0 and the interval [a, b] were arbitrary,
this achieves a proof of Theorem 5.

7 The Glimm scheme

The fundamental paper of Glimm [27] contained the first rigorous proof of existence of global
weak solutions to hyperbolic systems of conservation laws. For several years, the Glimm ap-
proximation scheme has provided the foundation for most of the theoretical results on the
subject. We shall now describe this algorithm in a somewhat simplified setting, for systems
where all characteristic speeds remain inside the interval [0, 1]. This is not a restrictive as-
sumption. Indeed, consider any hyperbolic system of the form

ut + A(u)ux = 0,

57
and assume that all eigenvalues of A remain inside the interval [−M, M ]. Performing the
linear change of independent variables

y = x + M t, τ = 2M t,

we obtain a new system

. 1 1
uτ + A∗ (u)uy = 0, A∗ (u) = A(u) + I
2M 2
where all eigenvalues of the matrix A∗ now lie inside the interval [0, 1].

To construct an approximate solution to the Cauchy problem

ut + f (u)x = 0, u(0, x) = ū(x), (7.1)

we start with a grid in the t-x plane having step size ∆t = ∆x, with nodes at the points
.
Pjk = (tj , xk ) = (j∆t, k∆x) j, k ∈ ZZ .

Moreover, we shall need a sequence of real numbers θ1 , θ2 , θ3 , . . . uniformly distributed over

the interval [0, 1]. This means that, for every λ ∈ [0, 1], the percentage of points θi , 1 ≤ i ≤ N
which fall inside [0, λ] should approach λ as N → ∞, i.e.:

#{j ; 1 ≤ j ≤ N, θj ∈ [0, λ] }
lim =λ for each λ ∈ [0, 1]. (7.2)
N →∞ N
By #I we denote here the cardinality of a set I.

At time t = 0, the Glimm algorithm starts by taking an approximation of the initial data ū,
which is constant on each interval of the form ]xk−1 , xk [ , and has jumps only at the nodal
.
points xk = k ∆x. To fix the ideas, we shall take

u(0, x) = ū(xk ) for all x ∈ [xk , xk+1 [ . (7.3)

For times t > 0 sufficiently small, the solution is then obtained by solving the Riemann
problems corresponding to the jumps of the initial approximation u(0, ·) at the nodes xk .
Since by assumption all waves speeds are contained in [0, 1], waves generated from different
nodes remain separated at least until the time t1 = ∆t. The solution can thus be prolonged
on the whole time interval [0, ∆t[ . For bigger times, waves emerging from different nodes
may cross each other, and the solution would become extremely complicated. To avoid this
unpleasant situation, a restarting procedure is adopted. Namely, at time t1 = ∆t the function
u(t1 −, ·) is approximated by a new function u(t1 +, ·) which is piecewise constant, having
.
jumps exactly at the nodes xk = k ∆x. Our approximate solution u can now be constructed
on the further time interval [∆t, 2∆t[ , again by piecing together the solutions of the various
Riemann problems determined by the jumps at the nodal points xk . At time t2 = 2∆t, this
solution is again approximated by a piecewise constant function, etc. . .
.
A key aspect of the construction is the restarting procedure. At each time tj = j ∆t, we need
to approximate u(tj −, ·) with a a piecewise constant function u(tj +, ·), having jumps precisely
at the nodal points xk . This is achieved by a random sampling technique. More precisely, we

58
look at the number θj in our uniformly distributed sequence. On each interval [xk−1 , xk [ , the
old value of our solution at the intermediate point x∗k = θ1 xk + (1 − θ1 )xk−1 becomes the new
value over the whole interval:

u(tj +, x) = u(tj −, θj xk + (1 − θj )xk−1 ) for all x ∈ [xk−1 , xk [ . (7.4)

An approximate solution constructed in this way is illustrated in fig. 34. The asterisks mark
the points where the function is sampled, assuming θ1 = 1/2, θ2 = 1/3.

* * * θ2 = 1/3

∆t * * * θ 1 = 1/2

∆x x

Figure 34: An approximate solution constructed by the Glimm scheme.

For a strictly hyperbolic system of conservation laws, satisfying the hypotheses (H) in Sec-
tion 3, the fundamental results of J.Glimm [27] and T.P.Liu [38] have established that

1. If the initial data ū has small total variation, then an approximate solution can be con-
structed by the above algorithm for all times t ≥ 0. The total variation of u(t, ·) remains
small.

2. Letting the grid size ∆t = ∆x tend to zero and using always the same sequence of numbers
θj ∈ [0, 1], one obtains a sequence of approximate solutions uν . By Helly’s compactness
theorem, one can extract a subsequence that converges to some limit function u = u(t, x) in
L1loc .

3. If the numbers θj are uniformly distributed over the interval [0, 1], i.e. if (7.2) holds, then
the limit function u provides a weak solution to the Cauchy problem (7.1).

The importance of the sequence θj being uniformly distributed can be best appreciated in the
following example.

Example 10. Consider a Cauchy problem of the form (7.1). Assume that the exact solution
consists of exactly one single shock, travelling with speed λ ∈ [0, 1], say
½
U (t, x) = u+ if x > λt,
u− if x < λt.

Consider an approximation of this solution obtained by implementing the Glimm algorithm

.
(fig. 35). By construction, at each time tj = j∆t, the position of the shock in this approximate
solution must coincide with one of the nodes of the grid. Observe that, passing from tj−1 to

59
 t *
*
u=u−
*
*
*
*
u=u+
*
*
*
∆t
*

∆x x

Figure 35: Applying the Glimm scheme to a solution consisting of a single shock.

tj , the position x(t) of the shock remains the same if the j-th sampling point lies on the left,
while it moves forward by ∆x if the j-th sampling point lies on the right. In other words,
½
x(tj−1 ) if θj ∈ ]λ, 1],
x(tj ) = (7.5)
x(tj−1 ) + ∆x if θj ∈ [0, λ].
.
Let us fix a time T > 0, and take ∆t = T /N . From (7.5) it now follows

x(T ) = #{j ; 1 ≤ j ≤ N, θj ∈ [0, λ] } · ∆t

#{j ; 1 ≤ j ≤ N, θj ∈ [0, λ] }
= ·T .
N
It is now clear that the assumption (7.2) on the uniform distribution of the sequence {θj }j≥1 is
precisely what is needed to guarantee that, as N → ∞ (equivalently, as ∆t → 0), the location
x(T ) of the shock in the approximate solution converges to the exact value λT .

Remark 7. At each restarting time tj we need to approximate the BV function u(tj −, ·)

with a new function which is piecewise constant on each interval [xk−1 , xk [ . Instead of the
sampling procedure (7.4), an alternative method consists of taking average values:
Z xk
. 1
u(tj +, x) = u(tj −, y) dy for all x ∈ [xk−1 , xk [ . (7.6)
∆x xk−1

Calling ujk the constant value of u(tj +) on the interval [xk−1 , xk [ , an application of the
divergence theorem on the square Γjk (fig. 36) yields

uj+1,k = uj,k + [f (uj,k−1 ) − f (uj,k )] (7.7)

Indeed, all wave speeds are in [0, 1], hence

u(t, xk−1 ) = uj,k−1 , u(t, xk ) = uj,k for all t ∈ [tj , tj+1 [ .

60
 u j+1, k

f(u j, k−1 ) f(u j, k )

Γj, k

u j, k

Figure 36: Approximations leading to the Godunov scheme.

The finite difference scheme (7.6) is the simplest version of the Godunov (upwind) scheme. It
is very easy to implement numerically, since it does not require the solution of any Riemann
problem. Unfortunately, it is very difficult to obtain a priori bounds on the total variation of
solutions constructed by the Godunov method. Proving the convergence of these approxima-
tions remains an outstanding open problem.

The remaining part of this chapter will be concerned with error bounds, for solutions generated
by the Glimm scheme.
Observe that, at each restarting time tj = j ∆t, the replacement of u(tj −) with the piecewise
constant function u(tj +) produces an error measured by
ku(tj +) − u(tj −)kL1
As the time step ∆t = T /N approaches zero, the total sum of all these errors does not converge
to zero, in general. This can be easily seen in Example 10, where we have
N
X N
X
ku(tj +) − u(tj −)kL1 ≥ |u+ − u− | · ∆t · min {(1 − λ), λ}
j=1 j=1

= |u+ − u− | · T · min {(1 − λ), λ}.

This fact makes it difficult to obtain sharp error estimates for solutions generated by the
Glimm scheme. Roughly speaking, the approximate solutions converge to the correct one not
because the total errors become small, but because, by a sort of randomness in the sampling
choice, small errors eventually cancel each other in the limit.

Clearly, the speed of convergence of the Glimm approximate solutions as ∆t, ∆x → 0 strongly
depends on how well the sequence {θi } approximates a uniform distribution on the interval
[0, 1]. In this connection, let us introduce

Definition 3. Let a sequence of numbers θj ∈ [0, 1] be given. For fixed integers 0 ≤ m < n,
the discrepancy of the set {θm , . . . , θn−1 } is defined as
¯ ¯
. ¯ #{j ; m ≤ j < n, θj ∈ [0, λ] } ¯¯
Dm,n = sup ¯¯λ − ¯. (7.8)
λ∈[0,1] n−m

61
We now describe a simple method for defining the numbers θj , so that the corresponding
discrepancies Dm,n approach zero as n − m → ∞ at a nearly optimal rate.

Fix an integer r ≥ 2. Every integer k ≥ 1 can then be uniquely written in base r, i.e. as a
sum of powers of r of the form

k = k0 + k1 r + . . . + kM r M 0 ≤ ki ≤ r − 1. (7.9)

In connection with (7.9) we then define

. k0 k1 kM
θk = + 2 + . . . + M +1 ∈ [0, 1]. (7.10)
r r r
For example, writing integer numbers in the usual basis r = 10, the computation of θk is as
follows. One simply has to invert the order of the digits of k and put a zero in front:

θ1 = 0.1 , ... , θ759 = 0.957 , ... , θ39022 = 0.22093 , ...

For the sequence (7.9)-(7.10), one can prove that the discrepancies satisfy

1 + ln(n − m)
Dm,n ≤ C · for all n > m ≥ 0 (7.11)
n−m
for some constant C = C(r). See [17] for details. For approximate solutions constructed
in terms of the above sequences (θj ), using the restarting procedures (7.3)-(7.4), sharp error
estimates can be proved.

Theorem 6. Given any initial data ū ∈ L1 with small total variation, call uexact (t, ·) = St ū
the exact solution of the Cauchy problem (7.1). Moreover, let uGlimm (t, ·) be the approximate
solution generated by the Glimm scheme, in connection with a grid of size ∆t = ∆x and a
fixed sequence (θj )j≥0 satisfying (7.11). For every fixed time T ≥ 0, letting the grid size tend
to zero, one has the error estimate

kuGlimm (T, ·) − uexact (T, ·)kL1

lim √ = 0. (7.12)
∆x→0 ∆x · | ln ∆x|

√
In other words, the L1 error tends to zero faster then ∆x · | ln ∆x|, i.e. slightly slower than
the square root of the grid size.
To prove Theorem 6, one first constructs a front tracking approximate solution u = u(t, x)
that coincides with uGlimm at the initial time t = 0 and at the terminal time t = T . This
construction is based on a fundamental lemma of T.P.Liu [38]. The L1 distance between
u(T, ·) and the exact solution ST ū can then be estimated using the error formula (10.6). For
all details we refer to [17].
Extensions to more general hyperbolic systems, without the assumption of genuine nonlinear-
ity, have recently appeared in [2, 29].

62
8 The Vanishing Viscosity Approach

In view of the previous uniqueness and stability results, it is natural to expect that the entropy
weak solutions of the hyperbolic system
ut + f (u)x = 0 (8.1)
should coincide with the unique limits of solutions to the parabolic system
uεt + f (uε )x = ε uεxx (8.2)
letting the viscosity coefficient ε → 0. For smooth solutions, this convergence is easy to
show. However, one should keep in mind that a weak solution of the hyperbolic system (8.1)
in general is only a function with bounded variation, possibly with a countable number of
discontinuities. In this case, as the smooth functions uε approach the discontinuous solution
u, near points of jump their gradients uεx must tend to infinity (fig. 37), while their second
derivatives uεxx become even more singular. Therefore, establishing the convergence uε → u
is a highly nontrivial matter. In earlier literature, results in this direction relied on three
different approaches:

1 - Comparison principles for parabolic equations. For a scalar conservation law, the
existence, uniqueness and global stability of vanishing viscosity solutions was first established
by Oleinik [44] in one space dimension. The famous paper by Kruzhkov [32] covers the more
general class of L∞ solutions and is also valid in several space dimensions.

2 - Singular perturbations. This technique was developed by Goodman and Xin [28],
and covers the case where the limit solution u is piecewise smooth, with a finite number of
non-interacting, entropy admissible shocks. See also [52] and [45], for further results in this
direction.

3 - Compensated compactness. With this approach, introduced by Tartar and DiPerna

[24], one first considers a weakly convergent subsequence uε * u. For a class of 2 × 2 systems,
one can show that this weak limit u actually provides a distributional solution to the non-
linear system (8.1). The proof relies on a compensated compactness argument, based on the
representation of the weak limit in terms of Young measures, which must reduce to a Dirac
mass due to the presence of a large family of entropies.

ε
u
u

Figure 37: A discontinuous solution to the hyperbolic system and a viscous approximation.

Since the hyperbolic Cauchy problem is known to be well posed within a space of functions with
small total variation, it seems natural to develop a theory of vanishing viscosity approximations

63
within the same space BV. This was indeed accomplished in [5], in the more general framework
of nonlinear hyperbolic systems not necessarily in conservation form. The only assumptions
needed here are the strict hyperbolicity of the system and the small total variation of the
initial data.

Theorem 7. Consider the Cauchy problem for the hyperbolic system with viscosity

uεt + A(uε )uεx = ε uεxx uε (0, x) = ū(x) . (8.3)

Assume that the matrices A(u) are strictly hyperbolic, smoothly depending on u in a neigh-
borhood of the origin. Then there exist constants C, L, L0 and δ > 0 such that the following
holds. If
Tot.Var.{ū} < δ , kūkL∞ < δ, (8.4)
then for each ε > 0 the Cauchy problem (8.3)ε has a unique solution uε , defined for all t ≥ 0.
.
Adopting a semigroup notation, this will be written as t 7→ uε (t, ·) = Stε ū.
In addition, one has:

BV bounds : Tot.Var.{Stε ū} ≤ C Tot.Var.{ū} . (8.5)

L1 stability : kStε ū − Stε v̄kL1 ≤ L kū − v̄kL1 , (8.6)

³ √ √ ´
kStε ū − Ssε ūkL1 ≤ L0 |t − s| + | εt − εs | . (8.7)

Convergence: As ε → 0+, the solutions uε converge to the trajectories of a semigroup S

such that
kSt ū − Ss v̄kL1 ≤ L kū − v̄kL1 + L0 |t − s| . (8.8)

These vanishing viscosity limits can be regarded as the unique vanishing viscosity solutions
of the hyperbolic Cauchy problem

ut + A(u)ux = 0, u(0, x) = ū(x) . (8.9)

In the conservative case A(u) = Df (u), every vanishing viscosity solution is a weak solution
of
ut + f (u)x = 0, u(0, x) = ū(x) , (8.10)
satisfying the Liu admissibility conditions.
Assuming, in addition, that each characteristic field is genuinely nonlinear or linearly degen-
erate, the vanishing viscosity solutions coincide with the unique limits of Glimm and front
tracking approximations.

In the genuinely nonlinear case, an estimate on the rate of convergence of these viscous ap-
proximations was provided in [19]:

64
Theorem 8. For the strictly hyperbolic system of conservation laws (8.10), assume that
every characteristic field is genuinely nonlinear. At any time t > 0, the difference between the
corresponding solutions of (8.3) and (8.10) can be estimated as
√
kuε (t, ·) − u(t, ·)kL1 = O(1) · (1 + t) ε| ln ε| Tot.Var.{ū} .

The following remarks relate Theorem 7 with previous literature. We recall again a standard
assumption on the characteristic fields, introduced by P.Lax:

(H) For each i = 1, . . . , n, the i-th field is either genuinely nonlinear, so that Dλi (u)·ri (u) > 0
for all u, or linearly degenerate, with Dλi (u) · ri (u) = 0 for all u.

(i) Concerning the global existence of weak solutions to the strictly hyperbolic system of
conservation laws (8.10), Theorem 7 contains the famous result of J.Glimm [27]. It also
slightly extends the existence theorem of T.P.Liu [39], which does not require the assumption
(H) and is satisfied by “generic” hyperbolic systems. The main novelty of Theorem 7 lies
in the case where the system (8.9) is non-conservative. However, one should be aware that
for non-conservative systems the limit solution might change if we choose a viscosity matrix
different from the identity, say
ut + A(u)ux = ε (B(u)ux )x .
The analysis of viscous solutions, with a more general viscosity matrix B(u), is still a wide
open problem.

(ii) Concerning the uniform stability of entropy weak solutions, the results previously available
for n×n hyperbolic systems [11, 12, 16] always required the assumption (H). For 2×2 systems,
this condition was greatly relaxed in [1]. On the other hand, Theorem 7 establishes this uniform
stability for completely general n × n strictly hyperbolic systems, without any reference to the
assumption (H).

(iii) For the viscous system (8.11), previous results in [40, 50, 51, 52] have established the
stability of special types of solutions, such as travelling viscous shocks or viscous rarefactions,
w.r.t. suitably small perturbations. Taking ε = 1, our present theorem yields at once the uni-
form Lipschitz stability of all viscous solutions with sufficiently small total variation, w.r.t. the
L1 distance.

(iv) For piecewise smooth solutions of the hyperbolic system (8.1), convergence of vanishing
viscosity approximations was studied in [28], using singular perturbation techniques. In this
special case, better convergent rates are established. On the other hand, Theorem 8 applies
to arbitrary solutions, with the only assumption of small total variation.

We give below a general outline of the proof of Theorem 7. For details, wee [5] or the lecture
notes [10].

1 - Rescaling. As a preliminary, we observe that uε is a solution of (8.3) if and only if the

65
 .
rescaled function u(t, x) = uε (εt, εx) is a solution of the parabolic system with unit viscosity

ut + A(u)ux = uxx , (8.11)

with initial data u(0, x) = ū(εx). Clearly, the stretching of the space variable has no effect
on the total variation. Notice however that the values of uε on a fixed time interval [0, T ]
correspond to the values of u on the much longer time interval [0, T /ε]. To obtain the desired
BV bounds for the viscous solutions uε , we can confine all our analysis to solutions of (8.11),
but we need estimates uniformly valid for all times t ≥ 0 , depending only on the total variation
of the initial data ū.

2 - Parabolic estimates. Consider a solution of (8.11), with initial data

u(0, x) = ū(x) (8.12)

having small total variation, say

Tot.Var.{ū} ≤ δ0 .
Setting
. .
u∗ = lim ū(x) , A∗ = A(u∗ ) , (8.13)
x→−∞

we can rewrite the parabolic equation (8.11) in the form

ut + A∗ ux − uxx = [A∗ − A(u)]ux . (8.14)

Call G∗ (t) the Green function corresponding to the linear homogeneous equation

ut + A∗ ux − uxx = 0 . (8.15)

The matrix valued function G∗ can be explicitly computed. If w solves (8.15), then the i-th
.
component wi = li∗ · w satisfies the scalar equation

wi,t + λ∗i wi,x − wi,xx = 0 .

Therefore wi (t) = G∗i (t) ∗ wi (0), where the convolution kernel is

( )
1 (x − λ∗i t)2
G∗i (t, x) = √ exp − .
2 πt 4t

We observe that this Green kernel G∗ = G∗ (t, x) satisfies the bounds

κ κ
kG∗ (t)kL1 ≤ κ , kG∗x (t)kL1 ≤ √ , kG∗xx (t)kL1 ≤ , (8.16)
t t
for some constant κ and all t > 0.
The solution of the non-homogeneous Cauchy problem (8.14)-(8.12) now admits the represen-
tation Z ³ ´
t
u(t) = G(t) ∗ ū + G(t − s) ∗ [A∗ − A(u(s))]ux (s) ds . (8.17)
0

Notice that, if the total variation of u(s, ·) remains small, the right hand side of (8.17) can be
regarded as a small perturbation of the homogeneous equation. We thus expect that similar
regularity estimates will hold.

66
Indeed, one can then prove that the Cauchy problem (8.11)-(8.12) has a unique solution,
defined on a sufficiently small time interval [0, t̂], with t̂ ≈ δ0−2 . For t ∈ [0, t̂] the total
variation remains small:

Tot.Var.{u(t)} = kux (t)kL1 = O(1) · δ0 ,

while all higher derivatives decay quickly. In particular

δ0 δ0
kux (t)kL∞ , kuxx (t)kL1 = O(1) · √ , kuxx (t)kL∞ , kuxxx (t)kL1 = O(1) · .
t t
As long as the total variation remains small, one can prolong the solution also for larger times
t > t̂. In this case, uniform bounds on higher derivatives remain valid (fig. 38).

δ0
ux L
1

u xx 1
L
u xxx 1
L
0 ^t ~ δ− 2 t
~ 0

Figure 38: On an initial time interval, the higher order norms of the solution decay rapidly.
For t large, these norms remain small as long as the total variation of the solution is small.

3 - Wave decomposition. To establish uniform bounds on the total variation, valid for arbi-
trarily large times, we decompose the gradient along a basis of unit vectors r̃i = r̃i (u, ux , uxx ),
say X
ux = vi r̃i (8.18)
i
We then derive an evolution equation for these gradient components, of the form

vi,t + (λ̃i vi )x − vi,xx = φi i = 1, . . . , n , (8.19)

A crucial point in the entire analysis is the careful choice of the unit vectors r̃i . A natural
guess would be to take r̃i = ri (u) as the i-th eigenvector of the hyperbolic matrix A(u).
Unfortunately, this choice does not work. Indeed, the corresponding equation (8.19) would
then contain source terms φi which are NOT integrable:
Z ∞Z ∞
|φi (t, x)| dx dt = + ∞.
0 −∞

67
The key idea, introduced in [5], is to decompose the gradient ux as a sum of gradients of
visocus travelling waves, namely
X X
ux (x) = vi r̃i = Ui0 (x) . (8.20)
i i

We recall that, for the parabolic system (8.11), a travelling wave profile is a solution of the
special form
u(t, x) = U (x − σt).
Of course, the function U must then satisfy the second order O.D.E.

U 00 = (A(U ) − σ)U 0 .

At each point x, looking at the third order jet (u, ux , uxx ) we seek travelling profiles U1 , . . . , Un
such that
Ui00 = (A(Ui ) − σi )Ui0 (8.21)
for some speed σi close to the characteristic speed λi , and moreover

Ui (x) = u(x) i = 1, . . . , n , (8.22)

X X
Ui0 (x) = ux (x) , Ui00 (x) = uxx (x) . (8.23)
i i
It turns out that this decomposition is unique provided that the travelling profiles are chosen
within suitable center manifolds. We let r̃i be the unit vector parallel to Ui0 , so that Ui0 = vi r̃i
for some scalar vi . One can show that r̃i remains close to the eigenvector ri (u) of the Jacobian
.
matrix A(u) = Df (u), but r̃i 6= ri (u) in general. The first equation in (8.23) now yields the
decomposition (7.11). If u = u(t, x) is a solution of (8.11), we can think of vi as the density
of i-waves in u. The remarkable fact is that these components satisfy a system of evolution
equations (8.19) where the source terms φi on the right hand side are INTEGRABLE over the
whole domain {x ∈ IR , t > 0}. Indeed, we can think of the sources φi as new waves produced
by interactions between viscous waves. Their total strength is controlled by means of viscous
interaction functionals, somewhat similar to the one introduced by Glimm in [G] to study the
hyperbolic case.

4 - Bounds on the source terms. A careful examination of the source terms φi in (8.19)
shows that they arise mainly because of interactions among different viscous waves. Their
total strength can thus be estimated in terms of suitable interaction potentials. This will
allow us to prove the implication
XZ T Z
kux (t)kL1 ≤ δ0 for all t ∈ [t̂, T ], |φi (t, x)| dxdt ≤ δ0
i t̂
(8.24)
Z TZ
=⇒ |φi (t, x)| dxdt = O(1) · δ02 i = 1, . . . , n .
t̂

Since the left hand side of (8.19) is in conservation form, for every t ∈ [t̂, T ] one has
X Xµ Z tZ ¶
kux (t)kL1 ≤ kvi (t)kL1 ≤ kvi (t̂)kL1 + |φi (s, x)| dxds . (8.25)
i i t̂

68
By the implication (8.24), for δ0 sufficiently small we obtain uniform bounds on the total
strength of the source terms φi , and hence on the BV norm of u(t, ·).

5 - Analysis of the linearized variational equation. Similar techniques can also be

applied to a solution z = z(t, x) of the variational equation

zt + [DA(u) · z]ux + A(u)zx = zxx , (8.26)

which describes the evolution of a first order perturbation to a solution u of (8.12). Assuming
that the total variation of u remains small, one proves an estimate of the form

kz(t, ·)kL1 ≤ L kz(0, ·)kL1 for all t ≥ 0 , (8.27)

valid for every solution u of (8.11) having small total variation and every L1 solution of the
corresponding system (8.26).

6 - Lipschitz continuity w.r.t. the initial data. Relying on (8.27), a standard homotopy
argument yields the Lipschitz continuity of the flow of (8.11) w.r.t. the initial data, uniformly
in time. Indeed, let any two solutions u, v of (8.11) be given (fig. 39). We can connect them
by a smooth path of solutions uθ , whose initial data satisfy
.
uθ (0, x) = θu(0, x) + (1 − θ)v(0, x) , θ ∈ [0, 1] .

The distance ku(t, ·) − v(t, ·)kL1 at any later time t > 0 is clearly bounded by the length of the
path θ 7→ uθ (t). In turn, this can be computed by integrating the norm of a tangent vector.
.
Calling z θ = duθ /dθ, each vector z θ is a solution of the corresponding equation (8.26), with u
replaced by uθ . Using (8.28) we thus obtain
Z 1° ° Z 1
°d θ °
ku(t, ·) − v(t, ·)kL1 ≤ ° °
° dθ u (t)° dθ ≤ kz θ (t)kL1 dθ
0 L1 0
(8.28)
R1
≤ L 0 kz θ (0)kL1 dθ = L ku(0, ·) − v(0, ·)kL1 .

v(t)

zθ (t)
v(0) u θ(t)

zθ(0)
θ
u (0) u(t)
u(0)

Figure 39: Estimating the distance between two viscous solution by a homotopy method.

By the simple rescaling of coordinates t 7→ εt, x 7→ εx, all of the above estimates remain valid
for solutions uε of the system (8.3)ε . In particular this yields the bounds (8.5) and (8.6).

69
7 - Existence of a vanishing viscosity limit. By a compactness argument, these BV
bounds imply the existence of a strong limit uεm → u in L1loc , at least for some subsequence
εm → 0. In the conservative case where A = Df , it is now easy to show that this limit u
provides a weak solution to the Cauchy problem (8.10).

8 - Characterization and uniqueness of the vanishing viscosity limit. At this stage,

it only remains to prove that the limit is unique, i.e. it does not depend on the choice of the
sequence εm → 0. For a system in conservative form, and with the standard assumption (H)
that each field is either genuinely nonlinear or linearly degenerate, we can apply the uniqueness
theorem in [13] and conclude that the limit of vanishing viscosity approximations coincides
with the limit of Glimm and of front tracking approximations.

9 - The non-conservative case. To handle the general non-conservative case, some addi-
tional work is required. We first consider Riemann initial data and show that in this special
case the vanishing viscosity solution is unique and can be accurately described. Then we prove
that any weak solution obtained as limit vanishing viscosity approximations is also a “viscosity
solution” in the sense that it satisfies the local integral estimates (6.14)-(6.18), where U ] is now
the unique non-conservative solution of a Riemann problem. By an argument introduced in
[8], a Lipschitz semigroup is completely determined as soon as one specifies its local behavior
for piecewise constant initial data. Characterizing its trajectories as “viscosity solutions” we
thus obtain the uniqueness of the semigroup of vanishing viscosity limits.

9 Counterexamples

In the previous chapters, the global existence and uniqueness theory for weak solutions to
systems of conservation laws was developed under certain key assumptions. In particular:

- The system is strictly hyperbolic.

- The initial data is small in BV.

In this final chapter, we consider three examples showing that:

1. If the system is not strictly hyperbolic, BV solutions may not depend continuously on the
initial data, in the L1 norm.

2. If the initial data is in L∞ but with possibly unbounded variation, then the solution may
become measure-valued in finite time. In such case, the uniqueness and continuous dependence
of solutions may also fail.

70
3. If the total variation of the initial data is large, then the L∞ norm of the solution may
blow up in finite time. In particular, no global solution exists in BV .

Example 11. Consider the following simple model of traffic flow. Denote by x the space
coordinate along a highway. Call k(x) the number of lanes open to traffic. Typically, one will
have k(x) = 2 for most x, except along some short stretches of the highway where k(x) = 1
because one lane is closed due to repair work. Calling u = u(t, x) the density of cars (= number
of cars per kilometer), we consider the system
(
ut + f (k, u)x = 0, .
f (k, u) = 8ku − u2 . (9.1)
kt = 0,

Consider first the family of stationary solutions

½ ½
ε 3 if x ∈ [0, ε], 1 if x ∈ [0, ε],
u (t, x) = kε (t, x) =
1 if x∈/ [0, ε], 2 if x∈/ [0, ε],
Clearly, as ε → 0, the functions uε , k ε converge in L1loc to the constant solutions

u(t, x) ≡ 1, k(t, x) ≡ 2.

ω ω
1 3
ω2

ω ω
0 0

0 ε x

Figure 40: For a system which is not strictly hyperbolic, the solution may not depend contin-
uously on the initial data.

On the other hand, consider now the family of (time-dependent) solutions (fig. 40)
 √

 ω0 if x/t < −6/(3
√ + 2 3),

ω

 1 if −6/(3 + 2 3) < x/t < 0,
(k ε , uε )(t, x) = ω2 if 0 < x < ε, √



 ω3 if 0 < (x − ε)/t < 6/(2 3 − 3),

 √
ω0 if 6/(2 3 − 3) < (x − ε)/t,
where (fig. 41)
√ √
ω0 = (2, 2), ω1 = (8 + 4 3, 2), ω2 = (4, 1), ω3 = (8 − 4 3, 2).

Observe that, as ε → 0, the initial data converges to the constant function

ū(x) ≡ 2, k̄(x) ≡ 2.

71
However, for times t > 0, the solutions uε do not converge to the corresponding constant
solution
u(t, x) ≡ 2, k(t, x) ≡ 2.

64

f(2,u)

ω0

ω3 ω2 ω1
16
f(1,u)

4 8 16 u

Figure 41: Different flux functions corresponding to k = 1 and k = 2 .

Example 12. We now consider the Cauchy problem for a strictly hyperbolic 3 × 3 system of
the form  

 ut + ux = 0, 
 u(0, x) = ū(x),
vt − vx = 0, v(0, x) = v̄(x), (9.2)

 w + [φ(u, v) · w] 
 w(0, x) = w̄(x).
t x = 0,
for some smooth function φ : IR2 7→ [−1/2, 1/2]. Observe that the above system is strictly
hyperbolic, with characteristic speeds
λ1 = −1, λ2 = φ(u, v), λ3 = 1.
All fields are linearly degenerate.
ϕ

Figure 42: The initial data considered at (9.4).

If the initial data ū, v̄ are in BV , then by a result in [6] it follows that the O.D.E. for charac-
teristics
.
ẋ = ψ(t, x) = φ(u(t, x), v(t, x)) (9.3)

72
 / BV and φ ∈ C ∞ such
has unique solutions. On the other hand, it is possible to choose ū, v̄ ∈
that the corresponding O.D.E. (9.3) has multiple (forward and backward) solutions through
a given point (fig. 43). In this case, if we look at the conservation equation for w, we see
that all the mass initially located within the interval [a, b] will be concentrated at the single
point at time τ . The L∞ norm of the solution thus blows up, and the solution itself becomes
measure-valued. After time τ , the solution may be prolonged as a measure-valued solution in
many different ways. Indeed, the Dirac mass at P may be split among different characteristics
emerging from P , or else continued as a single Dirac mass moving along any one particular
characteristic originating from P .
An example where this happens is the following. Take (fig. 42)
½
. 1 if 2−2n−1 < |x| < 2−2n
ū(x) = v̄(x) = ϕ(x) = (9.4)
−1 if 2−2n < |x| < 2−2n+1

. 1 − uv
φ(u, v) = .
6
In this case, it is easy to see that the O.D.E. (9.3) admits infinitely many solutions through
the origin (fig. 43).

Figure 43: Since the total variation of the solution to (9.2)-(9.4) is unbounded, the O.D.E. for
characteristics can have multiple forward and backward solutions through a given point.

Example 13. Following [31], consider a 3 × 3 system of the form

Ut + F (U )x = 0, (9.5)

where U = (u, v, w) and

 
uv + w
F (U ) = F (u, v, w) =  g(v)  (9.6)
2
u(1 − v ) − vw

73
We assume that the scalar function g is strictly convex, with
g(0) = 0, g(v) = g(−v), −1 < g 0 (v) < 1 (9.7)
for all v. It is then clear that the system is strictly hyperbolic in the region where |v| < 1.
Indeed  
v u 1
.
A(U ) = DF (u, v, w) =  0 g 0 (v) 0 
2
1 − v 2uv − w −v
is a matrix with real distinct eigenvalues
λ1 = −1, λ2 = g 0 (v), λ3 = 1.
The corresponding right eigenvectors are
     
1 0 1
r1 =  0  r2 =  1  r3 =  0  .
−1 − v 0 1−v
The first and third characteristic fields are linearly degenerate, the second is genuinely non-
linear.

4
β

3
β

2
β

β
0 x

Figure 44: Structure of a solution where the total variation blows up in finite time.

By a suitable choice of the function g, the analysis in [31] shows that one can construct a
solution U = U (t, x) of the following form (fig. 44). The initial data contains two approaching
2-shocks, and a 1-wave between them, of unit strength. Subsequent interactions produce
alternatively a 3-wave and a 1-wave, with strengths given as in fig. 44, with β > 1. The L∞
norm of the solution, and hence also its total variation, approach infinity as t → T −, the time
where the two 2-shocks hit each other.
This example shows that, in general, if the initial data are large enough, the solution can blow
up in finite time. One should remark, however, that the system (9.6) does not come from any
realistic physical model. In particular, it does not admit any strictly convex entropy. It is a
major open problem to establish whether blow up in the total variation norm can occur for
solutions to the equations of gas dynamics.

74
10 Appendix

We collect here some results of basic mathematical analysis, which were used in previous
sections.

A.1 - A compactness theorem

We state below a version of Helly’s compactness theorem, which provides the basic tool in the
proof of existence of weak solutions.

Theorem A.1. Consider a sequence of functions uν : [0, ∞[ ×IR 7→ IRn with the following
properties.
Tot.Var.{uν (t, ·)} ≤ C, |uν (t, x)| ≤ M for all t, x, (10.1)
Z ∞
|uν (t, x) − uν (s, x)| dx ≤ L|t − s| for all t, s ≥ 0, (10.2)
−∞
for some constants C, M, L. Then there exists a subsequence uµ which converges to some
function u in L1loc ([0, ∞) × IR; IRn ). This limit function satisfies
Z ∞
|u(t, x) − u(s, x)| dx ≤ L|t − s| for all t, s ≥ 0. (10.3)
−∞

The point values of the limit function u can be uniquely determined by requiring that
.
u(t, x) = u(t, x+) = lim u(t, y) for all t, x. (10.4)
y→x+

In this case, one has

Tot.Var.{u(t, ·)} ≤ C, |u(t, x)| ≤ M for all t, x. (10.5)

A detailed proof can be found in [9].

A.2 - An elementary error estimate

Let D be a closed subset of a Banach space E and consider a Lipschitz continuous semigroup
S : D × [0, ∞[ 7→ D. More precisely, assume that

(i) S0 u = u, Ss St u = Ss+t u,

(ii) kSt u − Ss vk ≤ L · ku − vk + L0 · |t − s|.

Given a Lipschitz continuous map w : [0, T ] 7→ D, we wish to estimate the difference between
w and the trajectory of the semigroup S starting at w(0).

75
Theorem A.2. Let S : D × [0, ∞[ 7→ D be a continuous flow satisfying the properties (i)-(ii).
For every Lipschitz continuous map w : [0, T ] 7→ D one then has the estimate
Z T½ ¾
kw(t + h) − Sh w(t)k
kw(T ) − S w(0)k ≤ L · lim inf dt. (10.6)
T 0 h→0+ h

Remark 8. The integrand in (10.6) can be regarded as the instantaneous error rate for w at
time t. Since the flow is uniformly Lipschitz continuous, during the time interval [t, T ] this
error is amplified at most by a factor L (see fig. 45).

w(T)

w(t+h) S w(t+h)
T−t−h
S w(t)
w(t) S w(t) T−t
h

S w(0)
T
w(0)

Figure 45: Comparing the approximate solution w with the trajectory of the semigroup having
the same initial data.

Proof. We first show that the integrand function

. kw(t + h) − Sh w(t)k
φ(t) = lim inf
h→0+ h
is bounded and measurable. Indeed, for every h > 0, the function

. kw(t + h) − Sh w(t)k
φh (t) = (10.7)
h
is continuous. By the continuity of the maps h 7→ φh (t), we have

φ(t) = lim inf φh (t),

ε→0+ h∈Q∩ ]0,ε]

where the infimum is taken only over rational values of h. Hence the function φ is Borel
measurable. Its boundedness is a consequence of the uniform Lipschitz continuity of w and of
all trajectories of the semigroup S.
Next, define
.
Φ(t) = kST −t w(t) − ST w(0)k, (10.8)
Z t
.
z(t) = Φ(t) − L φ(s) ds. (10.9)
0
We need to show that z(T ) ≤ 0. Since z is Lipschitz continuous and z(0) = 0, it suffices to
show that the time derivative satisfies ż(s) ≤ 0 for almost every s ∈ [0, T ]. By the theorems

76
of Rademacher and of Lebesgue, there exists a null set N ⊂ [0, T ] such that, for each s ∈/ N,
both functions z, Φ are differentiable at time s while φ is quasicontinuous at s. For s ∈
/ N we
thus have
z(s + h) − z(s) Φ(s + h) − Φ(s)
ż(s) = lim = lim − L · φ(s). (10.10)
h→0 h h→0 h
By the definition of Φ and the properties of S it follows

Φ(s + h) − Φ(s) = kST −s−h w(s + h) − ST w(0)k − kST −s w(s) − ST w(0)k

≤ kST −s−h w(s + h) − ST −s w(s)k

≤ kST −s−h w(s + h) − ST −s−h Sh w(s)k

≤ L · kw(s + h) − Sh w(s)k .

Therefore
Φ(s + h) − Φ(s) kw(s + h) − Sh w(s)k
lim ≤ L · lim inf = L · φ(s).
h→0+ h h→0+ h
By (10.10) this implies
ż(s) ≤ 0 for all s ∈
/ N.
Hence z(T ) ≤ 0, proving the theorem.

A.3 - Taylor estimates

This final section contains some useful bounds on the size of a function Ψ of several variables,
valid in a neighborhood of the origin. The assumption that Ψ vanishes on certain sets, such
as the coordinate axes or the diagonals, implies that some of the Taylor coefficients of Ψ
at the origin are equal to zero. Writing a Taylor approximation and using this additional
information, one obtains the desired estimates. As usual, the Landau notation O(1) here
indicates a function whose absolute value remains uniformly bounded.

Lemma A.1. Let Φ : IRm 7→ IRn be differentiable with Lipschitz continuous derivative. If
∂Φ
Φ(0) = 0, (0) = 0, (10.11)
∂x
then one has
Φ(x) = O(1) · |x|2 . (10.12)

Lemma A.2. Let Φ : IRp ×IRq 7→ IRn be twice differentiable, with Lipschitz continuous second
derivatives. If
Φ(x, 0) = Φ(0, y) = 0 for all x ∈ IRp , y ∈ IRq , (10.13)
then
Φ(x, y) = O(1) · |x| |y| (10.14)

77
for all x, y in a neighborhood of the origin. If in addition there holds

∂2Φ
(0, 0) = 0, (10.15)
∂x∂y
then one has the sharper estimate

Φ(x, y) = O(1) · |x| |y| (|x| + |y|). (10.16)

Proof. Observe that (10.13) implies

Z 1µ ¶
∂Φ
Φ(x, y) = (x, θy) · y dθ
0 ∂y
Z 1Z 1Ã 2 !
∂ Φ (10.17)
= (θ x, θy) · (x ⊗ y) dθ0 dθ
0
0 0 ∂x∂y

= O(1) · |x| |y|.

This gives (10.14). If (10.15) also holds, then the Lipschitz continuity of the second derivative
implies ¯ ¯
¯ ∂2Φ ¯
¯ ¯
¯ (x, y)¯ = O(1) · (|x| + |y|).
¯ ∂x∂y ¯

Using this bound inside the double integral in (10.17), we obtain (10.16).

Lemma A.3. Let Φ = Φ(q̃, q ∗ , σ) be a C 2 mapping from IRn−1 ×IR ×IR into IR, with Lipschitz
continuous second derivatives. Assume that, for all q̃ ∈ IRn−1 , s, q ∗ , σ ∈ IR there holds

Φ(q̃, q ∗ , 0) = Φ(0, s, −s) = Φ(0, 0, σ) = 0. (10.18)

Then one has the estimate

Φ(q̃, q ∗ , σ) = O(1) · (|q̃| |σ| + |q ∗ | |σ| |q ∗ + σ|). (10.19)

If, in addition, for all q ∗ , σ there holds

Φ(0, q ∗ , σ) = 0, (10.20)

then one has

Φ(q̃, q ∗ , σ) = O(1) · |q̃| |σ|. (10.21)

Proof. To derive (10.19) we write

Φ(q̃, q ∗ , σ) = Φ(0, q ∗ , σ) + [Φ(q̃, q ∗ , σ) − Φ(0, q ∗ , σ)]. (10.22)

By (10.18), for all q̃, q ∗ one has

∂Φ
(q̃, q ∗ , 0) = 0.
∂ q̃

78
Hence, by the Lipschitz continuity of the first derivative,
∂Φ
(q̃, q ∗ , σ) = O(1) · |σ|.
∂ q̃
Using this last estimate we obtain
Z 1µ ¶
∂Φ ∗
Φ(q̃, q ∗ , σ) − Φ(0, q ∗ , σ) = (θq̃, q , σ) · q̃ dθ
0 ∂ q̃
(10.23)
= O(1) · |q̃| |σ|.

To establish (10.19) it thus remains to prove

Φ(0, q ∗ , σ) = O(1) · |q ∗ | |σ| |q ∗ + σ|. (10.24)

Observe that (10.18) implies

∂2Φ
(0, 0, 0) = 0.
∂q ∗ ∂σ
In the case where q ∗ and σ have the same sign, one has |q ∗ + σ| = |q ∗ | + |σ|. Therefore (10.24)
follows from Lemma A.2, being equivalent to the estimate (10.16).
To see what happens when q ∗ and σ have opposite signs, to fix the ideas assume q ∗ < 0 < σ,
|q ∗ | < σ, the other cases being entirely similar. We consider two possibilities:
If |q ∗ | < σ/2, then
|σ| |q ∗ | + |σ|
|q ∗ + σ| ≥
≥
2 3
and (10.24) again follows from (10.16) in Lemma A.2.
In the remaining case where σ/2 ≤ |q ∗ | ≤ σ, we write
Z q∗
∂Φ
Φ(0, q ∗ , σ) = Φ(0, −σ, σ) + (0, s, σ) ds
−σ ∂q ∗
Z q∗ Z σ
∂2Φ
= 0+ (0, s, s0 ) dsds0
−σ 0 ∂q ∗ ∂σ
Z q∗ Z σ
= O(1) · (|s| + |s0 |) dsds0
−σ 0

= O(1) · |σ + q ∗ | |σ| (|σ| + |σ|).

This yields again (10.24), completing the proof of (10.19).

If the additional assumption (10.20) holds, then the estimate (10.21) is an immediate conse-
quence of (10.22) and (10.23).

Lemma A.4. Let Φ = Φ(q̃, q ∗ , σ) be a C 1 mapping from IRn−1 ×IR ×IR into IR, with Lipschitz
continuous derivatives.

79
(a) If Φ satisfies the assumptions
∂Φ ∂Φ 1
∗
(0, 0, 0) = (0, 0, 0) = , Φ(0, s, −s) = 0 for all s. (10.25)
∂q ∂σ 2
then one has the estimate
q∗ + σ ³ ´
Φ(q̃, q ∗ , σ) = + O(1) · |q̃| + |q ∗ + σ|(|q ∗ | + |σ|) . (10.26)
2

(b) If instead Φ satisfies the assumptions

∂Φ 1
∗
(0, 0, 0) = , Φ(0, 0, σ) = 0 for all σ, (10.27)
∂q 2
then one has the estimate
q∗ ³ ´
Φ(q̃, q ∗ , σ) = + O(1) · |q̃| + |q ∗ |(|q ∗ | + |σ|) . (10.28)
2

Proof. We write Φ in the form

Φ(q̃, q ∗ , σ) = [Φ(qe, q ∗ , σ) − Φ(0, q ∗ , σ)] + Φ(0, q ∗ , σ).

The first term on the right hand side clearly satisfies

Φ(q̃, q ∗ , σ) − Φ(0, q ∗ , σ) = O(1) · |q̃|. (10.29)

.
To estimate the second term, in case (a) we define s = (q ∗ − σ)/2 and compute
Z (q∗ +σ)/2 · ¸
d
Φ(0, q ∗ , σ) = Φ(0, s, −s) + Φ(0, s + θ, − s + θ) dθ
0 dθ
Z (q∗ +σ)/2 h i
= 1 + O(1) · (|q ∗ | + |σ|) dθ (10.30)
0

q∗ + σ
= + O(1) · |q ∗ + σ| (|q ∗ | + |σ|).
2
Combining (10.29) with (10.30) one obtains (10.26).
In case (b), we write
Z q∗ · ¸
d
Φ(0, q ∗ , σ) = Φ(0, θ, σ) dθ
0 dθ
Z q∗ h i
1 (10.31)
= + O(1) · (|q ∗ | + |σ|) dθ
0 2

q∗
= + O(1) · |q ∗ | (|q ∗ | + |σ|).
2
Combining (10.29) with (10.31) one obtains (10.28).

80
Remark 9. In the above lemmas, assume that the functions Φ and their derivatives depend
continuously on an additional parameter ω and satisfy the given assumptions for all values of
ω. Then all of the above estimates are still valid, for quantities O(1) which remain uniformly
bounded as ω ranges on compact sets.

References

[1] F. Ancona and A. Marson, Well-posedness for general 2 × 2 systems of conservation laws.
Memoir Amer. Math. Soc. 169 (2004), no. 801.

[2] F. Ancona and A. Marson: Sharp convergence rate of the Glimm scheme for general
nonlinear hyperbolic systems. Preprint 2009.

[3] P. Baiti and H. K. Jenssen, On the front tracking algorithm, J. Math. Anal. Appl. 217
(1998), 395-404.

[4] S. Bianchini, On the Riemann problem for non-conservative hyperbolic systems, Arch.
Rat. Mach. Anal. 166 (2003), 1-26.

[5] S. Bianchini and A. Bressan, Vanishing viscosity solutions to nonlinear hyperbolic sys-
tems, Annals of Mathematics 161 (2005), 223-342.

[6] A. Bressan, Unique solutions for a class of discontinuous differential equations, Proc.
Amer. Math. Soc. 104 (1988), 772-778.

[7] A. Bressan, Global solutions of systems of conservation laws by wave-front tracking, J.

Math. Anal. Appl. 170 (1992), 414-432.

[8] A. Bressan, The unique limit of the Glimm scheme, Arch. Rational Mech. Anal. 130
(1995), 205-230.

[9] A. Bressan, Hyperbolic Systems of Conservation Laws. The One Dimensional Cauchy
Problem, Oxford University Press, 2000.

[10] A. Bressan, BV solutions to systems of conservation laws by vanishing viscosity, C.I.M.E.

course in Cetraro, 2003, P. Marcati ed. Springer Lecture Notes in Math. 1911 (2007),
pp.1–78.

[11] A. Bressan and R. M. Colombo, The semigroup generated by 2 × 2 conservation laws,

Arch. Rational Mech. Anal. 133 (1995), 1-75.

[12] A. Bressan, G. Crasta, and B. Piccoli, Well posedness of the Cauchy problem for n × n
systems of conservation laws, Amer. Math. Soc. Memoir 694 (2000).

81
[13] A. Bressan and P. Goatin, Oleinik type estimates and uniqueness for n × n conservation
laws, J. Differential Equations, 156 (1999), 26-49.

[14] A. Bressan and P. LeFloch, Uniqueness of weak solutions to hyperbolic systems of con-
servation laws. Arch. Rational Mech. Anal. 140 (1997), 301-317.

[15] A. Bressan and M. Lewicka, A uniqueness condition for hyperbolic systems of conservation
laws, Discr. Con. Dynam. Syst. 6 (2000), 673-682.

[16] A. Bressan, T. P. Liu and T. Yang, L1 stability estimates for n × n conservation laws,
Arch. Rational Mech. Anal. 149 (1999), 1-22.

[17] A. Bressan and A. Marson, Error bounds for a deterministic version of the Glimm scheme,
Arch. Rat. Mech. Anal. 142 (1998), 155-176.

[18] A. Bressan and W. Shen, Uniqueness for discontinuous O.D.E. and conservation laws,
Nonlinear Analysis, T. M. A. 34 (1998), 637-652.

[19] A. Bressan and T. Yang, On the convergence rate of vanishing viscosity approximations,
Comm. Pure Appl. Math 57 (2004), 1075-1109.

[20] M. G. Crandall, The semigroup approach to first order quasilinear equations in several
space variables. Israel J. Math. 12 (1972), 108–132.

[21] C. Dafermos, Polygonal approximations of solutions of the initial value problem for a
conservation law, J. Math. Anal. Appl. 38 (1972), 33-41.

[22] C. Dafermos, Hyperbolic Conservation Laws in Continuum Physics, Springer-Verlag,

1999.

[23] R. J. DiPerna, Global existence of solutions to nonlinear hyperbolic systems of conserva-

tion laws, J. Differential Equations 20 (1976), 187-212.

[24] R. DiPerna, Convergence of approximate solutions to conservation laws, Arch. Rational

Mech. Anal. 82 (1983), 27-70.

[25] C. Donadello and A. Marson, Stability of front tracking solutions to the initial and bound-
ary value problem for systems of conservation laws. Nonlinear Differential Equations Appl.
14 (2007), 569–592.

[26] L. C. Evans and R. F. Gariepy, Measure Theory and Fine Properties of Functions,
C.R.C. Press, 1992.

[27] J. Glimm, Solutions in the large for nonlinear hyperbolic systems of equations, Comm.
Pure Appl. Math. 18 (1965), 697-715.

82
[28] J. Goodman and Z. Xin, Viscous limits for piecewise smooth solutions to systems of
conservation laws, Arch. Rational Mech. Anal. 121 (1992), 235-265.

[29] J. Hua, Z. Jiang, and T. Yang, A new Glimm functional and convergence rate of Glimm
scheme for general systems of hyperbolic conservation laws. Arch. Rational Mech. Anal.,
to appear.

[30] H. Holden and N. H. Risebro, Front Tracking for Hyperbolic Systems of Conservation
Laws, Springer-Verlag, New York, 2002.

[31] H. K. Jenssen, Blowup for systems of conservation laws, SIAM J. Math. Anal. 31 (2000),
894-908.

[32] S. Kruzhkov, First-order quasilinear equations with several space variables, Mat. Sb. 123
(1970), 228–255. English transl. in Math. USSR Sb. 10 (1970), 217–273.

[33] P. D. Lax, Hyperbolic systems of conservation laws II, Comm. Pure Appl. Math. 10
(1957), 537-566.

[34] M. Lewicka, Well-posedness for hyperbolic systems of conservation laws with large BV
data. Arch. Rational Mech. Anal. 173 (2004), 415–445.

[35] Ta-Tsien Li Global classical solutions for quasilinear hyperbolic systems. Wiley, Chich-
ester, 1994.

[36] T. P. Liu, The Riemann problem for general systems of conservation laws, J. Diff. Equat.
18 (1975), 218-234.

[37] T. P. Liu, The entropy condition and the admissibility of shocks, J. Math. Anal. Appl.
53 (1976), 78-88.

[38] T. P. Liu, The deterministic version of the Glimm scheme, Comm. Math. Phys. 57 (1977),
135-148.

[39] T. P. Liu, Admissible solutions of hyperbolic conservation laws, Amer. Math. Soc. Memoir
240 (1981).

[40] T. P. Liu, Nonlinear stability of shock waves for viscous conservation laws, Amer. Math.
Soc. Memoir 328 (1986).

[41] T.-P. Liu and T. Yang, L1 stability of conservation laws with coinciding Hugoniot and
characteristic curves, Indiana Univ. Math. J. 48 (1999), 237-247.

[42] T.-P. Liu and T. Yang, L1 stability of weak solutions for 2 × 2 systems of hyperbolic
conservation laws, J. Amer. Math. Soc. 12 (1999), 729-774.

83
[43] Y. Lu, Hyperbolic Conservation Laws and the Compensated Compactness Method. Chap-
man & Hall/CRC, Boca Raton, FL, 2003.

[44] O. Oleinik, Discontinuous solutions of nonlinear differential equations, Amer. Math. Soc.
Transl. 26, 95-172.

[45] F. Rousset, Viscous approximation of strong shocks of systems of conservation laws. SIAM
J. Math. Anal. 35 (2003), 492–519.

[46] S. Schochet, Sufficient conditions for local existence via Glimm’s scheme for large BV
data, J. Differential Equations 89 (1991), 317-354.

[47] S. Schochet, The essence of Glimm’s scheme, in Nonlinear Evolutionary Partial Differen-
tial Equations, X. Ding and T. P. Liu Eds., American Mathematical Society - International
Press (1997), 355-362.

[48] D. Serre, Systemes de Lois de Conservation I, II, Diderot Editeur, 1996.

[49] J. Smoller, Shock Waves and Reaction-Diffusion Equations, Springer-Verlag, New York,
1983.

[50] A. Szepessy and Z. Xin, Nonlinear stability of viscous shocks, Arch. Rational Mech. Anal.
122 (1993), 53-103.

[51] A. Szepessy and K. Zumbrun, Stability of rarefaction waves in viscous media, Arch.
Rational Mech. Anal. 133 (1996), 249-298.

[52] S. H. Yu, Zero-dissipation limit of solutions with shocks for systems of hyperbolic con-
servation laws, Arch. Rational Mech. Anal. 146 (1999), 275-370.

[53] A. I. Volpert, The spaces BV and quasilinear equations, Math. USSR Sbornik 2 (1967),
225-267.

Several recent papers on conservation laws are also available on the preprint server
http://www.math.ntnu.no/conservation/

84