Chapter 2

Hyperbolic equations
As we have seen, and will further argue below, the hydrodynamics equations are nothing more
than signal-propagation equations. Equations of this kind are called hyperbolic equations. The
equations of hydrodynamics are only a member of the more general class of hyperbolic equa-
tions, and there are many more examples of hyperbolic equations than just the equations of
hydrodynamics. But of course we shall focus mostly on the application to hydrodynamics. In
this chapter we shall study the mathematical properties of hyperbolic equations, and analytic
methods how to solve them for simple linear problems. This background knowledge is not only
important for understanding the nature of hydrodynamic flows, it also lies at the basis of numer-
ical hydrodynamics algorithms. Hence we go into quite some detail here.
2.1 The simplest form of a hyperbolic equation: advection
Consider the following equation:

t
q + u∂
x
q = 0 (2.1)
where q = q(x, t) is a function of one spatial dimension and time, and u is a velocity that
is constant in space and time. This is called an advection equation, as it describes the time-
dependent shifting of the function q(x) along x with a velocity u (Fig. 2.1-left). The solution at
any time t > t
0
can be described as a function of the state at time t
0
:
q(x, t) = q(x −ut, 0) (2.2)
This is a so-called initial value problem in which the state at any time t > t
0
can be uniquely
found when the state at time t = t
0
is fully given. The characteristics of this problem are straight
lines:
x
char
(t) = x
(0)
char
+ ut (2.3)
This is a family of lines in the (x, t) plane, each of which is labeled by its own unique value of
x
(0)
char
.
In this example the initial value problem is quite trivial, yet, as we will see below, this prob-
lem stands at the basis of numerical methods of hydrodynamics and is numerically surprisingly
challenging to solve!
23
24
q(x)
space
u(x)
space
q(x)
u(x)
space
conserved quantity
pure advection
space
Figure 2.1. Advection of a function q(x, t) with constant velocity u (left) and space-varying
velocity u(x) (right). The space-varying velocity problem comes in two versions: the conserved
form (dashed) and the non-conserved simple advection form (dotted).
2.2 Advection with space-dependent velocity
Consider now that u is a function of space: u = u(x). We get:

t
q + u(x)∂
x
q = 0 (2.4)
This is still an advection problem, but the velocity is not constant, and therefore the solution is
somewhat more complex. Define the following variable:
ξ =
_
x
x
0
dx

u(x

)
(2.5)
for some arbitary x
0
. Now the solution is given as:
q(x, t) = q(x(ξ), t) = q(x(ξ −t), 0) (2.6)
where x(ξ) is the value of x belonging to the value of ξ. We see that in regions of low u(x)
the function will be squeezed in x-direction while in regions of high u(x) the function will
be stretched. Fig. 2.1-right, dotted line, shows the squeezing effect for non-constant advection
velocity.
The characteristics are:
x
char
(t) = x(ξ
(0)
char
+ t) (2.7)
In this case the labeling is done with ξ
(0)
char
and, by definition, the curves follow the flow in the
(x, t) plane.
If we define the comoving derivative D
t
as D
t
= ∂
t
+ u(x)∂
x
, Eq.(2.4) translates into:
D
t
q(x, t) = 0 (2.8)
This simply tells that along each flow line (characteristic) the function q remains constant, i.e. the
comoving derivative is zero.
25
2.3 Advection of a conserved quantity with space-dependent velocity
Now consider yet another version of the advection equation:

t
q + ∂
x
[qu(x)] = 0 (2.9)
This is very similar to Eq. (2.4), but this time with u(x) inside the ∂
x
operator. This is a conser-
vation equation. It has the following property of conservation: if u(x) is zero at x
0
and x
1
> x
0
,
then the integral
_
x
1
x
0
qdx is constant in time.
This equation can be written in the previous form, with u(x) outside the operator:

t
q + u(x)∂
x
q = −q∂
x
u(x) (2.10)
This is an advection equation of the type of Eq. (2.4), but this time with a non-zero right-hand-
side. This can be regarded as some kind of source term in the equation (even though strictly
speaking it is not, as the system is conserved). The characteristics are the same as for the case
without this term, but in this case the value of q is no longer constant along these characteristics:
D
t
q(x, t) = −q(x, t)∂
x
u(x) (2.11)
When the characteristics converge, q goes up, and when they diverge, q goes down. This is de-
picted in Fig. 2.1-right, dashed line. This is precisely what one expects to happen if q represents
some kind of density.
2.4 Flux-form of conservation equation and the Jacobian
Let us stick with the equation (2.9), but write it in the following form:

t
q + ∂
x
f(q, x) = 0 (2.12)
where f(q, x) is called the flux of the quantity q. The flux f is entirely determined for any given
value of the function q and position x. We can then use the chain rule to arrive at

t
q +
∂f
∂q
¸
¸
¸
¸
x=const

x
q = −
∂f(q, x)
∂x
¸
¸
¸
¸
q=const
(2.13)
In our case we have:
∂f
∂q
¸
¸
¸
¸
x=const
= u(x) and
∂f
∂x
¸
¸
¸
¸
q=const
= q∂
x
u(x) (2.14)
In this way we also arrive at Eq. (2.10). It is this route to Eq. (2.10) that we will take in the cases
of the more complex non-linear systems of equations to follow. The ∂f/∂q is the Jacobian of
the system.
2.5 Coupled set of equations
All of the above can also be formulated for coupled sets of PDEs. Instead of a state scalar q we
defined a state vector Q = (q
1
, · · · , q
m
). The advection equation of the non-conservative type is
then

t
Q + A∂
x
Q = 0 (2.15)
26
where A is an m×m matrix. The advection equation of conservative type is

t
Q+ ∂
x
(AQ) = 0 (2.16)
The more general conservation equation is:

t
Q + ∂
x
F = 0 (2.17)
where F = F(Q, x) is the flux. Like in the scalar case we have

t
Q +
∂F
∂Q
¸
¸
¸
¸
x=const

x
Q = −
∂F(q, x)
∂x
¸
¸
¸
¸
Q=const
(2.18)
where the J ≡ ∂F/∂Q is the Jacobian of the system, which is an m × m matrix. If F is a
linear function of Q, then one can write F = AQ where A is then the Jacobian, and we arrive at
Eq. (2.16). If A is then also independent of x, then we arrive at Eq. (2.16).
2.6 The wave equation in vector notation: eigenvalues and eigenvectors
In Section 1.7.1 we derived the wave equation for perturbations in an otherwise steady constant-
density constant-velocity background medium. Let us now define:
Q ≡
_
q
1
q
2
_
=
_
ρ
1
ρ
0
u
1
_
(2.19)
We can then write Eqs. (1.56,1.57) as:

t
_
q
1
q
2
_
+
_
u
0
1
C
2
s
u
0
_

x
_
q
1
q
2
_
= 0 (2.20)
We can also write this in flux conservative form:

t
_
q
1
q
2
_
+ ∂
x
_
f
1
f
2
_
= 0 (2.21)
with the flux F = (f
1
, f
2
) given as:
F ≡
_
f
1
f
2
_
=
_
u
0
1
C
2
s
u
0
__
q
1
q
2
_
=
_
u
0
q
1
+ q
2
C
2
s
q
1
+ u
0
q
2
_
(2.22)
or in other words, the Jacobian matrix is:
J
ik
=
∂f
i
∂q
k
=
_
u
0
1
C
2
s
u
0
_
(2.23)
One of the advantages of writing the wave equation, and lateron many other equations, in
the form of a matrix equation like Eq. (2.20), is that we can use some aspects of linear alge-
bra to solve the equations in an elegant way, which will later turn out to have a very powerful
application in numerical methods of hydrodynamics.
27
let us look again at the Jacobian matrix of the wave equation, Eq. (2.23). This matrix has
the following eigenvectors and eigenvalues:
e
−1
=
_
1
−C
s
_
with λ
−1
= u
0
−C
s
(2.24)
e
+1
=
_
1
+C
s
_
with λ
+1
= u
0
+ C
s
(2.25)
If we include the passive tracer of Eq. (1.7.2) we have:
Q ≡
_
_
q
1
q
2
q
3
_
_
=
_
_
ρ
1
ρ
0
u
1
ϕ
_
_
(2.26)
and the matrix form of the equation of motion becomes:

t
_
_
q
1
q
2
q
3
_
_
+
_
_
u
0
1 0
C
2
s
u
0
0
0 0 u
0
_
_

x
_
_
q
1
q
2
q
3
_
_
= 0 (2.27)
This has the following set of eigenvectors and eigenvalues:
e
−1
=
_
_
1
−C
s
0
_
_
with λ
−1
= u
0
−C
s
(2.28)
e
0
=
_
_
0
0
1
_
_
with λ
−1
= u
0
(2.29)
e
+1
=
_
_
1
+C
s
0
_
_
with λ
+1
= u
0
+ C
s
(2.30)
We can now decompose any vector q into eigenvectors:
Q = ˜ q
−1
e
−1
+ ˜ q
0
e
0
+ ˜ q
+1
e
+1
(2.31)
Then the equation of motion becomes:

t
_
_
˜ q
−1
˜ q
0
˜ q
+1
_
_
+
_
_
u
0
−C
s
0 0
0 u
0
0
0 0 u
0
+ C
s
_
_

x
_
_
˜ q
−1
˜ q
0
˜ q
+1
_
_
= 0 (2.32)
The matrix is here diagonal. This has the advantage that we have now decomposed the problem
into three scalar advection equations:

t
˜ q
i
+ λ
i

x
˜ q
i
= 0 (2.33)
for any i = −1, 0, +1. For these equations we know the solution: they are simply shifts of an
initial function (see Sections 2.1 and 2.2). In the case at hand here we are fortunate that the C
s
and u
0
are constant, so we get:
˜ q
i
(x, t) = ˜ q
i
(x −λ
i
t, 0) (2.34)
28
for any i = −1, 0, +1. The vector-notation and the decomposition into eigenvectors and eigen-
values stands at the basis of much of the theory on numerical algorithms to follow.
The system of equations described here is a hyperbolic set of equations, which is another
way of saying that they describe the motion of signals. We will define hyperbolicity more rigor-
ously below.
2.7 Hyperbolic sets of equations: the linear case with constant Jacobian
Let us consider a set of linear equations that can be written in the form:

t
Q + A∂
x
Q = 0 (2.35)
where Q is a vector of m components and A is an m×m matrix.
This system is called hyperbolic if the matrix A is diagonalizable with real eigenvalues.
The matrix is diagonalizable if there exists a complete set of eigenvectors e
i
, i.e. if any vector
can be written as:
Q =
m

i=1
˜ q
i
e
i
(2.36)
In this case one can write
AQ =
m

i=1
λ
i
˜ q
i
e
i
(2.37)
We can define a matrix in which each column is one of the eigenvectors:
R = (e
1
, · · · , e
m
) (2.38)
Then we can transform Eq. (2.35) into:
R
−1

t
Q + R
−1
ARR
−1

x
Q = 0 (2.39)
which with
˜
Q = R
−1
Q then becomes:

t
˜
Q +
˜
A∂
x
˜
Q = 0 (2.40)
where
˜
A = diag(λ
1
, · · · , λ
m
). Not all λ
i
must be different from each other.
This system of equations has in principle m sets of characteristics. But any set of charac-
teristics that has the same characteristic velocity as another set is usually called the same set of
characteristics. So in the case of 5 eigenvalues, three of which are identical, one typically says
that there are three sets of characteristics.
2.8 Boundary conditions (I)
So far we have always assumed that space is infinite. In real-life applications the domain of
interest is always bound. In some cases these boundaries are real (like a wall or a piston) but
in other cases they have to be somewhat artificially imposed because computing power is not
as infinite as space is and one is limited to a finite volume. It is therefore important to know
how spatial boundary conditions are set. In the numerical chapters we will go into this in far
more detail than here, often going into more practical matters. Here we are concerned with the
mathematical issue.
29
We look at the general hyperbolic equation

t
Q(x, t) + A∂
x
Q(x, t) = 0 (2.41)
where Qis an vector with mcomponents and A is an m×m matrix. Let us define as our domain
of interest the space between x
0
and x
1
> x
0
. At these boundaries we wish to impose the
necessary and sufficient boundary conditions. Since this is a coupled set of m linear equations
we need m boundary conditions. In systems of hyperbolic equations one usually sets Dirichlet
boundary conditions, i.e. one speficies the values of the components of Q at these boundaries.
The best way to do this is to decompose into the eigenvalues of A:

t
˜ q
i
(x, t) + λ
i

x
˜ q
i
(x, t) = 0 (2.42)
For each i ∈ [1, · · · , m] one can then set:
˜ q
i
(x = x
0|1
, t) = ˜ q
(bc)
i
(t) (2.43)
The x
0|1
means that one can either specify that condition for ˜ q
i
at x = x
0
or at x = x
1
. For
instance, for m = 2 one could specify one condition at x
0
for ˜ q
1
and one condition at x
1
for ˜ q
2
.
In total one can impose m conditions, divided over the left and right boundaries. Note, by the
way, that the boundary conditions may be time-dependent!
The question now is: for some ˜ q
i
, at which boundary should we impose the Dirichlet bound-
ary condition? Suppose λ
i
> 0. It is then clear that signals propagate from left to right. Therefore
it is clear that the boundary condition, for ˜ q
i
, must be set at the left boundary, at x = x
0
. The
condition specifies the inflow of the signal into the system through the left boundary. If one were
to specify it at the other boundary (which in a numerical simulation one could simply try to do,
and see what happens), then the installed boundary value is immediately advected off the domain
again and will not influence the solution within the domain at all. In contrast, setting ˜ q
i
at the
left boundary means that this value is advected into the domain and affects the solution there.
In general one can say that at boundary x = x
0
one specifies ˜ q
i
(x = x
0
, t) = ˜ q
(bc)
i
(t) for
all i for which λ
i
> 0 and at boundary x = x
1
one specifies ˜ q
i
(x = x
1
, t) = ˜ q
(bc)
i
(t) for all i for
which λ
i
< 0.
Another possible kind of boundary conditions are periodic boundary conditions. Here we
set, at every time t, for λ
i
> 0:
˜ q
i
(x = x
0
, t) := ˜ q
i
(x = x
1
, t) (2.44)
In this way, the information flowing out of the right boundary flows back into the left boundary.
The system is in this way closed and depends now only on the initial conditions.
2.9 Hyperbolic sets of equations: the linear case with variable Jacobian
So what happens if the matrix A depends on x, i.e. A →A(x)? In this case also the eigenvalues
change with position, like described in Section 2.2. But in addition to this also the eigenvectors
can depend on spatial position x, and hence the diagonalization of the matrix. In other words,
the transformation matrix becomes a local matrix: R → R(x). This means that on cannot
decompose Q into characteristic modes globally, but this has to be done locally. This also means
that the modes get mixed. Consider the following example:

t
_
q
1
q
2
_
+
_
cos x sin x
sin x −cos x
_

x
_
q
1
q
2
_
= 0 (2.45)
30
on a domain limited by x
0
= 0 and x
1
= π. The eigenvalues of the matrix are always ±1, but
the eigenvectors change with x. At both x = x
0
= 0 and x = x
1
= π the eigenvectors can be
written as e
1
= (1, 0) and e
2
= (0, 1), while at, for example, x = π/2 we have e
1
= (1, 1)/

2
and e
2
= (−1, 1)/

2. Or more general:
e
1
=
_
cos x
sin x
_
(2.46)
e
2
=
_
−sin x
cos x
_
(2.47)
(the eigenvectors at x
1
are now minus those of x
0
, but there is no difference in the meaning as
the norm of the eigenvalues are irrelevant except in the definition of the norm of ˜ q
i
).
If we set ˜ q
1
(x
0
) = q
1
(x
0
) = f(t) where f(t) is some function of time t, and we set ˜ q
2
(x
1
) =
0 and the initial value of ˜ q at ˜ q
i
(x, t = t
0
), then the signal put into mode ˜ q
1
at the left boundary
initially propagates from left to right, but as it gets to larger x it starts to mix with the ˜ q
2
mode,
which moves in opposite direction.
This example shows that although hyperbolic equations are about signal propagation, this
does not mean that the signals are simply pure waves moving across the domain, but can in-
stead interact with each other even if the equations are linear. However, locally one can always
uniquely divide the state vector Q up into the characteristic modes moving each at their own
characteristic speeds.
2.10 Boundary conditions (II)
When the Jacobian m × m matrix A depends on x, the question of how and where to impose
boundary conditions can become, in some circumstances, a bit more difficult. The limitation
that one must impose precisely m boundary conditions is no longer strictly true. It turns out to
depend on the number of inward-pointing characteristics at each of the boundaries. Consider the
simple example of scalar advection problem between x
0
= −1 and x
1
= 1 with advection speed
A(x) = u(x) = x:

t
q + x∂
x
q = 0 (2.48)
In this case, at x = x
0
= −1 the characteristic is pointing out of the domain. But the same is
true at x = x
1
= 1! So neither at x
0
nor at x
1
must one determine boundary conditions. The
opposite example

t
q −x∂
x
q = 0 (2.49)
requires a Dirichlet boundary condition to be set at both boundaries. Here the information (sig-
nal) flows into the domain on both sides and piles up near x = 0.
In general when the sign of an eigenvalue flips one has the risk that the number of boundary
conditions deviates fromm. The physical interpretation of sign-flips of eigenvalues can be many.
For instance, as we shall see later, the sign of one of the eigenvalues flips in the case of a standing
shock in hydrodynamics. Indeed, signals pile up on both sides of the shock and a catastrophy
can only be avoided by the minute, but essential viscosity in the shock front which (only very
locally) ‘erases’ these converging signals again in a non-hyperbolic manner. But these issues
will be discussed in the chapter on supersonic flows and shocks, chapter ??
31
2.11 Hyperbolic equations versus elliptic equations
As mentioned above, if the Jacobian matrix can be diagonalized and the eigenvalues are real,
then the system is hyperbolic. So what if the eigenvalues are imaginary? Consider the following
system, similar to the equations for waves in hydrodynamics:

t
_
q
1
q
2
_
+
_
0 1
1 0
_

x
_
q
1
q
2
_
= 0 (2.50)
defined on a domain x ∈ [x
0
, x
1
]. This clearly can be written as:

2
t
q
1
−∂
2
x
q
1
= 0 (2.51)
which is a wave equation with eigenvalue λ
−1
= −1 and λ
+1
= 1. If the state Q is given at time
t = t
0
, and the boundary conditions are specified at x = x
0
and x = x
1
in the way explained in
Section 2.8, then the function Q(x, t) within the domain can be computed for all t > t
0
(in fact,
also backward in time).
Now consider the following equation:

t
_
q
1
q
2
_
+
_
0 1
−1 0
_

x
_
q
1
q
2
_
= 0 (2.52)
defined on a domain x ∈ [x
0
, x
1
]. This can be written as:

2
t
q
1
+ ∂
2
x
q
1
= 0 (2.53)
However, this is not a hyperbolic equation. The eigenvalues of the Jacobian matrix are ±i. It is
clear that we cannot do the same trick with eigenvectors and eigenvalues here, because it does not
make sense to move something with a speed ±i. The nature of this equation is therefore entirely
different even though it is merely one minus sign in the Jacobian matrix. In fact, Eq. (2.53)
can be recognized as the Laplace equation. It needs the specification of boundary conditions at
x = x
0
, x = x
1
, t = t
0
and t = t
1
. In other words: the state at some time t depends not only on
the past but also on the future. Clearly this makes not much sense, and the Laplace equation is
usually more used in two (or more) spatial directions instead of space and time.
2.12 Hyperbolic equations: the non-linear case
The above definition for linear hyperbolic sets of equations can be generalized to non-linear sets
of equations. Let us focus on the general conservation equation:

t
Q + ∂
x
F = 0 (2.54)
where, as ever, Q = (q
1
, · · · , q
m
) and F = (f
1
, · · · , f
m
). In general, F is not always a linear
function of Q, i.e. it cannot always be formulated as a matrix A times the vector Q (except if
A is allowed to also depend on Q, but then the usefulness of writing F = AQ is a bit gone).
So let us assume that F is some non-linear function of Q. Let us, for the moment, assume that
F = F(Q, x) = F(Q), i.e. we assume that there is no explicit dependence of F on x, except
through Q. Then according to Eq. (2.18) we get

t
Q+
∂F
∂Q

x
Q = 0 (2.55)
32
Figure 2.2. Characteristics of Burger’s equation. Left: case of diverging characteristics. Right:
case of converging characteristics with the formation of a singularity. Beyond the time of the
creation of the singularity the solution is ill defined unless a recipe is given how to treat the
singularity.
where
∂F
∂Q
is the Jacobian matrix, which depends, in the non-linear case, on Q itself. We can
nevertheless decompose this matrix in eigenvectors (which depend on Q) and we obtain

t
˜
Q+
_
_
_
λ
1
.
.
.
λ
m
_
_
_

x
˜
Q = 0 (2.56)
Here the eigenvalues λ
1
, · · · , λ
m
and eigenvectors (and hence the meaning of
˜
Q) depends on Q.
In principle this is not a problem. The characteristics are now simply given by the state vector Q
itself. The state is, so to speak, self-propagating. We are now getting into the kind of hyperbolic
equations like the hydrodynamics equations, which are also non-linear self-propagating.
2.13 Example of non-linear conservation equation: Burger’s equation
Consider

t
q +
1
2

x
(q
2
) = 0 (2.57)
This is called Burger’s equation. It is a conservation equation in q, with flux f(q) = q
2
/2. The
flux is only dependent on x through q. We have ∂f/∂q = q, so we can write the above equation
as

t
q + q∂
x
q = 0 (2.58)
So the advection velocity is, in this important example, the to-be-advected quantity q itself! The
quantity propagates itself with u = q. If we use the comoving derivative, then we obtain the
following equation:
D
t
q(x, t) = 0 (2.59)
which appears to be identical to Eq. (2.8). The difference lies in the definition of D
t
which is
defined with respect to a given function u(x) in Eq. (2.8) and with respect to q(x, t) in Eq. (2.59).
Eq. (2.59) shows that along a characteristic the value of q does not change, or in other
words: the characteristic speed (slope in the x, t-plane) does not change along a characteristic.
This means that the characteristics are straight lines in the (x, t)-plane, as in the case of the
example of Section 2.3. The difference to that example lies in the fact that in this case not all
straight line characteristics are parallel.
The interpretation of Burger’s equation is that of the motion of a pressureless fluid. Often
it is said to be the equation describing the motion of dust in space, as dust clouds do not have
33
pressure and each dust particle moves along a straight line. This is only partially correct, as we
shall show below, but it does describe roughly the point.
Since any non-parallel straight lines in (x, t) must have a crossing at some point in space
we can immediately derive that for converging flows there will be a point at which Burger’s
equations break down.
Now to come back at the difference of Burger’s equation with the motion of dust. In
Burger’s equation the assumption is that at any time and any position there exists only one ve-
locity. In case of dust flows this is not necessary: since the particles do not interact, at any given
time and position one can have dust particles flowing left, right at various velocities. If two
dust cloud approach, in Burger’s equation the equations produce shocks (i.e. a breakdown of the
pure inviscid equation). In the case of dust the particles would not feel each other and the cloud
simply go through each other.
Perhaps a more adequate physical interpretation of Burger’s equation is that of a pressure-
less fluid. If the characteristics converge a shock will happen and Burger’s equation will no
longer be valid and will be replaced by the more generally valid hydrodynamics equation with
pressure.
→ Exercise: As we have shown in Section 2.3 a conserved quantity tends to grow in regions
of converging flow and it tends to diminish in regions of diverging flow. In case of Burger’s
equation it appears that this is not the case, as the equation amounts to D
t
q = 0. Give a
modified version of Burger’s equation that does have this property of increasing value for
converging flow, and note which is the characteristic speed.
2.14 Isothermal hydrodynamic equations
The topic of this lecture is hydrodynamics, so let’s express the equations of hydrodynamics in
the above form. Let’s take the isothermal equations for simplicity. We have

t
ρ + ∂
x
(ρu) = 0 (2.60)

t
(ρu) + ∂
x
(ρu
2
+ ρc
2
s
) = 0 (2.61)
Let us define
q
1
≡ ρ , q
2
≡ ρu (2.62)
Then we can write the above equations as

t
_
q
1
q
2
_
+ ∂
x
_
q
2
q
2
2
q
1
+ q
1
c
2
s
_
= 0 (2.63)
or in other words: Q = (q
1
, q
2
) and F = (q
2
, q
2
2
/q
1
+q
1
c
2
s
). This can be written with the Jacobian:

t
_
q
1
q
2
_
+
_
0 1
_
c
2
s

q
2
2
q
2
1
_
2
q
2
q
1
_

x
_
q
1
q
2
_
= 0 (2.64)
The eigenvalues are
λ
±
=
q
2
q
1
±c
s
= u ±c
s
(2.65)
and the eigenvectors are:
e
±
=
_
1
λ
±
_
(2.66)
34
One sees that both the eigenvalues and the eigenvectors depend on the state vector (q
1
, q
2
) itself
and are therefore space- and time-dependent. The state vector determines for itself how it should
be decomposed. Modes mix in two different ways: a) the eigenvectors change in space and time,
and b) each mode influences the other mode due to the non-linearity.
→ Exercise: In Section 2.6 we derived the set of eigenvectors and eigenvalues for the pertur-
bation equation of hydrodynamics with an adiabatic equation of state. If we replace γP
0

0
with the isothermal sound speed c
2
s
, then we obtain the results for isothermal waves. The
funny thing is, however, that the Jacobain matrix in that case (Eq. 2.23) does not appear to
be the linearized form of the Jacobian matrix derived in the present section (the one used in
Eq. 2.64). Explain this in terms of how the linearization is done in Section 2.6.
2.15 Non-isothermal hydrodynamic equations
Now let us turn to the generalization of the isothermal hydrodynamics equations: the non-
isothermal hydrodynamics equations. Note that this is not equal to the adiabatic hydrodynamics
equations, because the adiabatic hydrodynamics equations assume that all of the gas lies on the
same adiabat, or in other words: that the gas is isentropic. In constrast, we would now like to
relax any assumption of the entropy of the gas, and allow the entropy of the gas to vary arbitrar-
ily in space. This means necessarily that we must include a third equation: the energy equation.
Now the equations become significantly more complex.
We have now:

t
ρ + ∂
x
(ρu) = 0 (2.67)

t
(ρu) + ∂
x
(ρu
2
+ P) = 0 (2.68)

t
(ρe
tot
) + ∂
x
[(ρe
tot
+ P)u] = 0 (2.69)
Let us define
q
1
≡ ρ , q
2
≡ ρu , q
3
≡ ρe
tot
(2.70)
Then we can write the above equations as

t
_
_
q
1
q
2
q
3
_
_
+ ∂
x
_
_
f
1
f
2
f
3
_
_
= 0 (2.71)
in which
_
_
f
1
f
2
f
3
_
_
=
_
_
ρu
ρu
2
+ P
(ρe
tot
+ P)u
_
_
=
_
_
_
q
2
(γ −1)q
3
+
_
3−γ
2
_
q
2
2
q
1
γ
q
3
q
2
q
1
+
_
1−γ
2
_
q
3
2
q
2
1
_
_
_
(2.72)
where we used
u = q
2
/q
1
(2.73)
P = (γ −1)
_
q
3

1
2
q
2
2
q
1
_
(2.74)
Eq. 2.71 with the above expressions for (f
1
, f
2
, f
3
) can be written with the Jacobian:

t
_
_
q
1
q
2
q
3
_
_
+
_
_
_
0 1 0
γ−3
2
q
2
2
q
1
(3 −γ)
q
2
q
1
(γ −1)

_
γ
q
3
q
2
q
2
1
+ (γ −1)
q
3
2
q
3
1
_ _
γ
q
3
q
1
+
3
2
(1 −γ)
q
2
2
q
2
1
_
γ
q
2
q
1
_
_
_

x
_
_
q
1
q
2
q
3
_
_
= 0 (2.75)
35
It can be useful to rewrite the Jacobian using the primitive variables:

t
_
_
q
1
q
2
q
3
_
_
+
_
_
0 1 0
γ−3
2
ρu
2
(3 −γ)u (γ −1)
−{γe
tot
u + (γ −1)u
3
}
_
γe
tot
+
3
2
(1 −γ)u
2
_
γu
_
_

x
_
_
q
1
q
2
q
3
_
_
= 0
(2.76)
The eigenvalues are
λ

= u −C
s
(2.77)
λ
0
= u (2.78)
λ
+
= u + C
s
(2.79)
(2.80)
with eigenvectors:
e

=
_
_
1
u −C
s
h
tot
−C
s
u
_
_
(2.81)
e
0
=
_
_
1
u
1
2
u
2
_
_
(2.82)
e
+
=
_
_
1
u + C
s
h
tot
+ C
s
u
_
_
(2.83)
where h
tot
= e
tot
+ P/ρ is the total specific enthalpy and C
s
=
_
γP/ρ is the adiabatic sound
speed.
2.16 Traffic flow equations
As already mentioned, hyperbolic equations are more general than only the equations of hydro-
dynamics. Here is an example of a model of traffic flow, as was first discussed in papers by
Lighthill, Whitham and Richards ( for the references, see book LeVeque from which this exam-
ple is taken). Let us assume a single lane road with a density of cars q (the number of cars per
car length). We assume that 0 ≤ q ≤ 1 (because we cannot have more than 1 car per car length)
and we verify a-posteriori if this condition is satified. The conservation equation is:

t
q + ∂
x
(qu) = 0 (2.84)
where u is the speed of the cars at time t and position x. Suppose that there is a speed limit of
u
max
and that if the road is nearly empty, the cars drive at the speed limit u = u
max
. If this was all,
then the advection equation simply moves the density of cars linearly toward larger x. However,
if the road gets more crowded drivers naturally slow down. Let us for simplicity assume that
u(q) = u
max
(1 −q) (2.85)
The flux of cars is then
f = q(1 −q)u
max
(2.86)
36
The flux of cars is greatest when q = 1/2. For lower q the flux is lower because the density
of cars is lower, while for higher q the flux is lower because the speed of the cars goes down
(congestion).
We can now write the traffic flow equation as

t
q + u
max
(1 −2q)∂
x
q = 0 (2.87)
This shows that the characteristic velocity of the system is:
λ = u
max
(1 −2q) (2.88)
which can range from −u
max
to u
max
. It is very important to note here that the characteristic
speed is not equal to the speed of propagation of the cars! It is the speed at which information is
propagating, not the speed of the advected quantity q itself. This is one of the peculiar features
of non-linear hyperbolic equations, and it is very similar to the peculiarities of the Burger’s
equation. The traffic flow equation is a particularly nice example of a non-linear hyperbolic
equation because it is a very simple equation, yet has very interesting solution properties.
2.17 Hyperbolic equations in 2-D and 3-D
So far we have done everything only in 1-D. But what we learned can also be generalized to
higher dimensions, by use of the concept of operator splitting. In 2-D we get

t
Q + ∂
x
F(Q) + ∂
y
G(Q) (2.89)
which can be written as:

t
q
i
+
∂f
i
∂q
k

x
q
k
+
∂g
i
∂q
k

y
q
k
(2.90)
By operator splitting we can focus our attention to one of the space dimensions only. If we split
it as

t
q
i
+
∂f
i
∂q
k

x
q
k
= −
∂f
i
∂q
k

y
q
k
(2.91)
then we focus on the advection in x-direction, and consider the y-advection as a source term,
while if we write

t
q
i
+
∂f
i
∂q
k

y
q
k
= −
∂f
i
∂q
k

x
q
k
(2.92)
we focus on advection in y-direction and consider the x-direction as a source term. In numer-
ical methods this operator splitting is often done to reduce the full 2-D or 3-D problem into
consecutive 1-D problems which are much easier to handle.

2. t) plane.5) for some arbitary x0 . The characteristics are: (0) xchar (t) = x(ξchar + t) (2. the curves follow the flow in the (x. t) = q(x(ξ − t).2 Advection with space-dependent velocity Consider now that u is a function of space: u = u(x). but the velocity is not constant.8) (0) This simply tells that along each flow line (characteristic) the function q remains constant.4) This is still an advection problem. t) = q(x(ξ).7) In this case the labeling is done with ξchar and.1. If we define the comoving derivative Dt as Dt = ∂t + u(x)∂x . dotted line. 0) (2. shows the squeezing effect for non-constant advection velocity. t) = 0 (2. Now the solution is given as: q(x. The space-varying velocity problem comes in two versions: the conserved form (dashed) and the non-conserved simple advection form (dotted). We get: ∂t q + u(x)∂x q = 0 (2. and therefore the solution is somewhat more complex.1-right.(2. . by definition.4) translates into: Dt q(x.e.24 q(x) q(x) conserved quantity pure advection space u(x) u(x) space space space Figure 2. i. t) with constant velocity u (left) and space-varying velocity u(x) (right).6) where x(ξ) is the value of x belonging to the value of ξ. the comoving derivative is zero. We see that in regions of low u(x) the function will be squeezed in x-direction while in regions of high u(x) the function will be stretched. Fig. Advection of a function q(x. 2. Define the following variable: x ξ= x0 dx′ u(x′ ) (2. Eq.

5 Coupled set of equations All of the above can also be formulated for coupled sets of PDEs.10). The ∂f /∂q is the Jacobian of the system. x) is called the flux of the quantity q. (2. This equation can be written in the previous form. q goes up. but this time with a non-zero right-handside. but this time with u(x) inside the ∂x operator. The advection equation of the non-conservative type is then ∂t Q + A∂x Q = 0 (2. 2.11) When the characteristics converge.15) . and when they diverge. t) = −q(x. 2. (2.1-right.4). (2. 2.10) that we will take in the cases of the more complex non-linear systems of equations to follow. with u(x) outside the operator: ∂t q + u(x)∂x q = −q∂x u(x) (2. as the system is conserved). It is this route to Eq. It has the following property of conservation: if u(x) is zero at x0 and x1 > x0 .12) where f (q. x) ∂x (2.9). Instead of a state scalar q we defined a state vector Q = (q1 .4). We can then use the chain rule to arrive at ∂t q + In our case we have: ∂f ∂q = u(x) x=const ∂f ∂q x=const ∂x q = − ∂f (q.14) In this way we also arrive at Eq. x) = 0 (2. This is precisely what one expects to happen if q represents some kind of density. The flux f is entirely determined for any given value of the function q and position x.3 Advection of a conserved quantity with space-dependent velocity Now consider yet another version of the advection equation: ∂t q + ∂x [qu(x)] = 0 (2. qm ).25 2. but write it in the following form: ∂t q + ∂x f (q. x then the integral x01 qdx is constant in time.4 Flux-form of conservation equation and the Jacobian Let us stick with the equation (2. dashed line. · · · . This is depicted in Fig.9) This is very similar to Eq. (2. but in this case the value of q is no longer constant along these characteristics: Dt q(x.13) q=const and ∂f ∂x = q∂x u(x) q=const (2.10) This is an advection equation of the type of Eq. t)∂x u(x) (2. The characteristics are the same as for the case without this term. This is a conservation equation. q goes down. This can be regarded as some kind of source term in the equation (even though strictly speaking it is not.

1 we derived the wave equation for perturbations in an otherwise steady constantdensity constant-velocity background medium. Like in the scalar case we have ∂t Q + ∂F ∂Q ∂x Q = − ∂F (q. f2 ) given as: F ≡ f1 f2 = u0 1 2 Cs u 0 q1 q2 = u0 q1 + q2 2 Cs q1 + u0 q2 (2. in the form of a matrix equation like Eq.18) Q=const (2.26 where A is an m × m matrix. (2. then one can write F = AQ where A is then the Jacobian. (2. (1. The advection equation of conservative type is ∂t Q + ∂x (AQ) = 0 The more general conservation equation is: ∂t Q + ∂x F = 0 where F = F (Q. and lateron many other equations.57) as: ∂t q1 q2 + q u0 1 ∂x 1 2 q2 Cs u 0 =0 (2. then we arrive at Eq. (2. is that we can use some aspects of linear algebra to solve the equations in an elegant way.6 The wave equation in vector notation: eigenvalues and eigenvectors In Section 1. Let us now define: Q≡ We can then write Eqs. which is an m × m matrix.21) or in other words.17) x=const where the J ≡ ∂F/∂Q is the Jacobian of the system. and we arrive at Eq.23) One of the advantages of writing the wave equation. which will later turn out to have a very powerful application in numerical methods of hydrodynamics. .22) q1 q2 + ∂x f1 f2 =0 (2. x) ∂x (2.7.20).19) We can also write this in flux conservative form: ∂t with the flux F = (f1 . If F is a linear function of Q.56.16). 2.16) (2.16).1. x) is the flux.20) q1 q2 = ρ1 ρ0 u1 (2. If A is then also independent of x. the Jacobian matrix is: Jik = ∂fi = ∂qk u0 1 2 Cs u 0 (2.

31) (2.2) we have:     q1 ρ1 Q ≡ q2  = ρ0 u1  q3 ϕ (2.25) and the matrix form of the equation of motion becomes:       q1 u0 1 0 q1 2 q2  + Cs u0 0  ∂x q2  = 0 ∂t q3 0 0 u0 q3 If we include the passive tracer of Eq.27) This has the following set of eigenvectors and eigenvalues:   1 e−1 = −Cs  with λ−1 = u0 − Cs 0   0 0 e0 = with λ−1 = u0 1   1 e+1 = +Cs  with λ+1 = u0 + Cs 0 We can now decompose any vector q into eigenvectors: Q = q−1 e−1 + q0 e0 + q+1 e+1 ˜ ˜ ˜ Then the equation of motion becomes:       q−1 ˜ u 0 − Cs 0 0 q−1 ˜ ˜ u0 0  ∂x  q0  = 0 ˜ ∂t  q0  +  0 q+1 ˜ 0 0 u 0 + Cs q+1 ˜ (2. For these equations we know the solution: they are simply shifts of an initial function (see Sections 2.1 and 2.34) .27 let us look again at the Jacobian matrix of the wave equation. 0.2). t) = qi (x − λi t. so we get: qi (x. +1.29) (2. (2.26) (2.28) (2. 0) ˜ ˜ (2. Eq.32) The matrix is here diagonal. (1. In the case at hand here we are fortunate that the Cs and u0 are constant.24) (2. This matrix has the following eigenvectors and eigenvalues: e−1 = e+1 = 1 −Cs 1 +Cs with with λ−1 = u0 − Cs λ+1 = u0 + Cs (2.30) (2.7. This has the advantage that we have now decomposed the problem into three scalar advection equations: ∂t qi + λi ∂x qi = 0 ˜ ˜ (2.23).33) for any i = −1.

i. em ) Then we can transform Eq. λm ). The system of equations described here is a hyperbolic set of equations. So in the case of 5 eigenvalues.36) In this case one can write AQ = λi qi ei ˜ i=1 (2. But any set of characteristics that has the same characteristic velocity as another set is usually called the same set of characteristics.7 Hyperbolic sets of equations: the linear case with constant Jacobian Let us consider a set of linear equations that can be written in the form: ∂t Q + A∂x Q = 0 (2. if any vector can be written as: m Q= i=1 qi ei ˜ m (2.35) where Q is a vector of m components and A is an m × m matrix. 2. · · · .40) (2. In some cases these boundaries are real (like a wall or a piston) but in other cases they have to be somewhat artificially imposed because computing power is not as infinite as space is and one is limited to a finite volume. · · · .8 Boundary conditions (I) So far we have always assumed that space is infinite. 2.e. +1. This system of equations has in principle m sets of characteristics. In the numerical chapters we will go into this in far more detail than here. It is therefore important to know how spatial boundary conditions are set. The vector-notation and the decomposition into eigenvectors and eigenvalues stands at the basis of much of the theory on numerical algorithms to follow. In real-life applications the domain of interest is always bound. which is another way of saying that they describe the motion of signals. 0.35) into: R−1 ∂t Q + R−1 ARR−1 ∂x Q = 0 ˜ which with Q = R−1 Q then becomes: ˜ ˜ ˜ ∂t Q + A∂x Q = 0 (2. (2. one typically says that there are three sets of characteristics. often going into more practical matters. .37) We can define a matrix in which each column is one of the eigenvectors: R = (e1 . Here we are concerned with the mathematical issue. This system is called hyperbolic if the matrix A is diagonalizable with real eigenvalues. We will define hyperbolicity more rigorously below.38) ˜ where A = diag(λ1 . three of which are identical. Not all λi must be different from each other. The matrix is diagonalizable if there exists a complete set of eigenvectors ei .28 for any i = −1.39) (2.

at x = x0 . m] one can then set: qi (x = x0|1 . This also means that the modes get mixed.45) . t) = qi ˜ ˜ (bc) (2.43) The x0|1 means that one can either specify that condition for qi at x = x0 or at x = x1 . In other words. like described in Section 2. but this has to be done locally. · · · . must be set at the left boundary. But in addition to this also the eigenvectors can depend on spatial position x. Consider the following example: ∂t q1 q2 + q cos x sin x ∂x 1 q2 sin x − cos x =0 (2. the transformation matrix becomes a local matrix: R → R(x). If one were to specify it at the other boundary (which in a numerical simulation one could simply try to do. In systems of hyperbolic equations one usually sets Dirichlet boundary conditions. for qi . by the way. It is then clear that signals propagate from left to right. At these boundaries we wish to impose the necessary and sufficient boundary conditions. divided over the left and right boundaries. and see what happens). that the boundary conditions may be time-dependent! The question now is: for some qi . one speficies the values of the components of Q at these boundaries. t) = 0 ˜ ˜ For each i ∈ [1. Let us define as our domain of interest the space between x0 and x1 > x0 . Another possible kind of boundary conditions are periodic boundary conditions. (bc) In general one can say that at boundary x = x0 one specifies qi (x = x0 .9 Hyperbolic sets of equations: the linear case with variable Jacobian So what happens if the matrix A depends on x. The best way to do this is to decompose into the eigenvalues of A: ∂t qi (x. t) = qi (t) for all i for ˜ ˜ which λi < 0. Note. at which boundary should we impose the Dirichlet bound˜ ary condition? Suppose λi > 0. i.e. the information flowing out of the right boundary flows back into the left boundary. t) := qi (x = x1 . t) ˜ ˜ (2. In contrast. 2. For ˜ instance. for m = 2 one could specify one condition at x0 for q1 and one condition at x1 for q2 . Since this is a coupled set of m linear equations we need m boundary conditions. ˜ ˜ In total one can impose m conditions.44) In this way. This means that on cannot decompose Q into characteristic modes globally. setting qi at the ˜ left boundary means that this value is advected into the domain and affects the solution there. for λi > 0: qi (x = x0 . The ˜ condition specifies the inflow of the signal into the system through the left boundary. t) = 0 (2.2.29 We look at the general hyperbolic equation ∂t Q(x.41) where Q is an vector with m components and A is an m × m matrix. then the installed boundary value is immediately advected off the domain again and will not influence the solution within the domain at all. A → A(x)? In this case also the eigenvalues change with position. Here we set. and hence the diagonalization of the matrix. Therefore it is clear that the boundary condition. t) + A∂x Q(x. t) = qi (t) for ˜ ˜ (bc) all i for which λi > 0 and at boundary x = x1 one specifies qi (x = x1 . at every time t.e. i.42) (t) (2. The system is in this way closed and depends now only on the initial conditions. t) + λi ∂x qi (x.

1)/ 2. But the same is true at x = x1 = 1! So neither at x0 nor at x1 must one determine boundary conditions. The limitation that one must impose precisely m boundary conditions is no longer strictly true. However. but the eigenvectors change with x. as we shall see later. while at. The physical interpretation of sign-flips of eigenvalues can be many. Here the information (signal) flows into the domain on both sides and piles up near x = 0. but as it gets to larger x it starts to mix with the q2 mode. The eigenvalues of the matrix are always ±1. Or more general: e1 = e2 = cos x sin x − sin x cos x (2. the question of how and where to impose boundary conditions can become. and we set q2 (x1 ) = ˜ ˜ 0 and the initial value of q at qi (x. signals pile up on both sides of the shock and a catastrophy can only be avoided by the minute. In general when the sign of an eigenvalue flips one has the risk that the number of boundary conditions deviates from m. ˜ If we set q1 (x0 ) = q1 (x0 ) = f (t) where f (t) is some function of time t. a bit more difficult. 1). this does not mean that the signals are simply pure waves moving across the domain. At both x = x0 = 0 and x = x1 = π the eigenvectors can√ be written as e1 = (1.49) requires a Dirichlet boundary condition to be set at both boundaries. but essential viscosity in the shock front which (only very locally) ‘erases’ these converging signals again in a non-hyperbolic manner. It turns out to depend on the number of inward-pointing characteristics at each of the boundaries. This example shows that although hyperbolic equations are about signal propagation. Consider the simple example of scalar advection problem between x0 = −1 and x1 = 1 with advection speed A(x) = u(x) = x: ∂t q + x∂x q = 0 (2. but can instead interact with each other even if the equations are linear. 0) and e2 = (0. 1)/ 2 √ and e2 = (−1. But these issues will be discussed in the chapter on supersonic flows and shocks. t = t0 ). Indeed. for example. ˜ which moves in opposite direction. locally one can always uniquely divide the state vector Q up into the characteristic modes moving each at their own characteristic speeds. x = π/2 we have e1 = (1. chapter ?? . The opposite example ∂t q − x∂x q = 0 (2. at x = x0 = −1 the characteristic is pointing out of the domain. For instance.47) (the eigenvectors at x1 are now minus those of x0 . but there is no difference in the meaning as the norm of the eigenvalues are irrelevant except in the definition of the norm of qi ).48) In this case.46) (2. in some circumstances.10 Boundary conditions (II) When the Jacobian m × m matrix A depends on x.30 on a domain limited by x0 = 0 and x1 = π. the sign of one of the eigenvalues flips in the case of a standing shock in hydrodynamics. then the signal put into mode q1 at the left boundary ˜ ˜ ˜ initially propagates from left to right. 2.

i.50) defined on a domain x ∈ [x0 . Let us.53) However. It needs the specification of boundary conditions at x = x0 . F is not always a linear function of Q. and the Laplace equation is usually more used in two (or more) spatial directions instead of space and time. t = t0 and t = t1 . i. except through Q. x1 ]. similar to the equations for waves in hydrodynamics: ∂t q1 q2 + 0 1 q ∂x 1 1 0 q2 =0 (2. but then the usefulness of writing F = AQ is a bit gone). assume that F = F (Q. qm ) and F = (f1 . 2. because it does not make sense to move something with a speed ±i. (2. it cannot always be formulated as a matrix A times the vector Q (except if A is allowed to also depend on Q. So what if the eigenvalues are imaginary? Consider the following system.53) can be recognized as the Laplace equation. In general. t) within the domain can be computed for all t > t0 (in fact. we assume that there is no explicit dependence of F on x. (2. Eq. The eigenvalues of the Jacobian matrix are ±i. as ever. · · · . x) = F (Q). Clearly this makes not much sense.52) defined on a domain x ∈ [x0 .55) . then the function Q(x. The nature of this equation is therefore entirely different even though it is merely one minus sign in the Jacobian matrix.12 Hyperbolic equations: the non-linear case The above definition for linear hyperbolic sets of equations can be generalized to non-linear sets of equations. In other words: the state at some time t depends not only on the past but also on the future. x1 ].8.e.18) we get ∂t Q + ∂F ∂x Q = 0 ∂Q (2. also backward in time). Then according to Eq. So let us assume that F is some non-linear function of Q.e. if the Jacobian matrix can be diagonalized and the eigenvalues are real.54) where. In fact. This clearly can be written as: 2 2 ∂t q1 − ∂x q1 = 0 (2. fm ). and the boundary conditions are specified at x = x0 and x = x1 in the way explained in Section 2.31 2. Q = (q1 . this is not a hyperbolic equation.11 Hyperbolic equations versus elliptic equations As mentioned above. x = x1 . then the system is hyperbolic. Let us focus on the general conservation equation: ∂t Q + ∂x F = 0 (2. It is clear that we cannot do the same trick with eigenvectors and eigenvalues here. This can be written as: 2 2 ∂t q1 + ∂x q1 = 0 (2. If the state Q is given at time t = t0 . Now consider the following equation: ∂t q1 q2 + q 0 1 ∂x 1 q2 −1 0 =0 (2. · · · .51) which is a wave equation with eigenvalue λ−1 = −1 and λ+1 = 1. for the moment.

2. t-plane) does not change along a characteristic. The difference lies in the definition of Dt which is defined with respect to a given function u(x) in Eq. as in the case of the example of Section 2. t) in Eq. or in other words: the characteristic speed (slope in the x..59). ˜  ∂t Q +  (2. In principle this is not a problem. (2.59) which appears to be identical to Eq. Characteristics of Burger’s equation. The interpretation of Burger’s equation is that of the motion of a pressureless fluid. The difference to that example lies in the fact that in this case not all straight line characteristics are parallel. as dust clouds do not have .59) shows that along a characteristic the value of q does not change. Left: case of diverging characteristics. ˜ Here the eigenvalues λ1 . If we use the comoving derivative. t) = 0 (2. λm 2. which depends.8) and with respect to q(x. The characteristics are now simply given by the state vector Q itself. The flux is only dependent on x through q. (2. so we can write the above equation as ∂t q + q∂x q = 0 (2. Right: case of converging characteristics with the formation of a singularity.58) So the advection velocity is.13 Example of non-linear conservation equation: Burger’s equation Consider 1 ∂t q + ∂x (q 2 ) = 0 (2.57) 2 This is called Burger’s equation. (2. the to-be-advected quantity q itself! The quantity propagates itself with u = q. in the non-linear case. so to speak. Beyond the time of the creation of the singularity the solution is ill defined unless a recipe is given how to treat the singularity. · · · . λm and eigenvectors (and hence the meaning of Q) depends on Q.3. We are now getting into the kind of hyperbolic equations like the hydrodynamics equations. in this important example. self-propagating.56)  ∂x Q = 0 . where ∂F is the Jacobian matrix. on Q itself. (2.8). It is a conservation equation in q.32 Figure 2. This means that the characteristics are straight lines in the (x. which are also non-linear self-propagating. Eq. then we obtain the following equation: Dt q(x. We can ∂Q nevertheless decompose this matrix in eigenvectors (which depend on Q) and we obtain   λ1  ˜ . We have ∂f /∂q = q. The state is. t)-plane. Often it is said to be the equation describing the motion of dust in space. with flux f (q) = q 2 /2.

62) (2. In the case of dust the particles would not feel each other and the cloud simply go through each other.61) + ∂x 2 q2 q1 =0 (2. Let’s take the isothermal equations for simplicity. at any given time and position one can have dust particles flowing left. This is only partially correct. as the equation amounts to Dt q = 0. q2 /q1 +q1 c2 ). in Burger’s equation the equations produce shocks (i.63) 2 or in other words: Q = (q1 . q2 ≡ ρu q2 + q1 c2 s (2. → Exercise: As we have shown in Section 2. Since any non-parallel straight lines in (x. and note which is the characteristic speed.60) (2.66) . In Burger’s equation the assumption is that at any time and any position there exists only one velocity.e. If two dust cloud approach. 2. In case of Burger’s equation it appears that this is not the case. q2 ) and F = (q2 . but it does describe roughly the point. right at various velocities.14 Isothermal hydrodynamic equations The topic of this lecture is hydrodynamics. This can be written with the Jacobian: s ∂t The eigenvalues are q1 q2 + 0 c2 s − 2 q2 2 q1 1 q2 2 q1 ∂x q1 q2 =0 (2. In case of dust flows this is not necessary: since the particles do not interact. If the characteristics converge a shock will happen and Burger’s equation will no longer be valid and will be replaced by the more generally valid hydrodynamics equation with pressure. a breakdown of the pure inviscid equation).33 pressure and each dust particle moves along a straight line.65) (2.3 a conserved quantity tends to grow in regions of converging flow and it tends to diminish in regions of diverging flow. as we shall show below. Give a modified version of Burger’s equation that does have this property of increasing value for converging flow. Perhaps a more adequate physical interpretation of Burger’s equation is that of a pressureless fluid. We have ∂t ρ + ∂x (ρu) = 0 ∂t (ρu) + ∂x (ρu2 + ρc2 ) = 0 s Let us define Then we can write the above equations as ∂t q1 q2 q1 ≡ ρ . Now to come back at the difference of Burger’s equation with the motion of dust. so let’s express the equations of hydrodynamics in the above form.64) λ± = and the eigenvectors are: q2 ± cs = u ± cs q1 e± = 1 λ± (2. t) must have a crossing at some point in space we can immediately derive that for converging flows there will be a point at which Burger’s equations break down.

71) in which where we used       q2 f1 ρu 2 f2  =  ρu2 + P  = (γ − 1)q3 + 3−γ q2   2 q1  q3 f3 (ρetot + P )u γ q3 q2 + 1−γ 2 2 q1 2 q1 (2.34 One sees that both the eigenvalues and the eigenvectors depend on the state vector (q1 . we would now like to relax any assumption of the entropy of the gas. q2 ≡ ρu .74) Eq. 2.23) does not appear to be the linearized form of the Jacobian matrix derived in the present section (the one used in Eq. or in other words: that the gas is isentropic. The state vector determines for itself how it should be decomposed. f3 ) can be written with the Jacobian:       0 1 0 q1 q1 2 γ−3 q2 q2 (3 − γ) q1 (γ − 1) ∂x q2  = 0 (2.6. f2 . however. This means necessarily that we must include a third equation: the energy equation. Note that this is not equal to the adiabatic hydrodynamics equations. and allow the entropy of the gas to vary arbitrarily in space. and b) each mode influences the other mode due to the non-linearity. q3 ≡ ρetot (2. 2. 2.71 with the above expressions for (f1 .67) (2. q2 ) itself and are therefore space. In constrast.15 Non-isothermal hydrodynamic equations Now let us turn to the generalization of the isothermal hydrodynamics equations: the nonisothermal hydrodynamics equations. → Exercise: In Section 2. The s funny thing is. Now the equations become significantly more complex.64).and time-dependent.70) (2. 2.73) (2. Explain this in terms of how the linearization is done in Section 2. Modes mix in two different ways: a) the eigenvectors change in space and time.69) (2. If we replace γP0 /ρ0 with the isothermal sound speed c2 .72) u = q2 /q1 P = (γ − 1) q3 − 2 1 q2 2 q1 (2. then we obtain the results for isothermal waves. that the Jacobain matrix in that case (Eq.6 we derived the set of eigenvectors and eigenvalues for the perturbation equation of hydrodynamics with an adiabatic equation of state.75) q2  +  ∂t   2 q1 3 2 q2 q2 q3 q2 q3 q2 3 q3 q3 − γ 2 + (γ − 1) 3 γ + (1 − γ) 2 γ q1 q1 q1 2 q1 q1 . because the adiabatic hydrodynamics equations assume that all of the gas lies on the same adiabat.68) (2. We have now: ∂t ρ + ∂x (ρu) = 0 ∂t (ρu) + ∂x (ρu2 + P ) = 0 ∂t (ρetot ) + ∂x [(ρetot + P )u] = 0 Let us define Then we can write the above equations as     q1 f1 q2  + ∂x f2  = 0 ∂t q3 f3 q1 ≡ ρ .

then the advection equation simply moves the density of cars linearly toward larger x. if the road gets more crowded drivers naturally slow down. We assume that 0 ≤ q ≤ 1 (because we cannot have more than 1 car per car length) and we verify a-posteriori if this condition is satified.82) e+ (2.78) (2. Let us assume a single lane road with a density of cars q (the number of cars per car length). see book LeVeque from which this example is taken).80) with eigenvectors: e−  1 =  u − Cs  htot − Cs u   1 =  u  1 2 u 2  1 =  u + Cs  htot + Cs u  (2.77) (2.85) .16 Traffic flow equations As already mentioned. the cars drive at the speed limit u = umax .81) e0 (2. γP/ρ is the adiabatic sound 2.84) where u is the speed of the cars at time t and position x. hyperbolic equations are more general than only the equations of hydrodynamics. as was first discussed in papers by Lighthill. However.35 It can be useful to rewrite the Jacobian using the primitive variables:       0 1 0 q1 q1 γ−3 2 q2  +   ∂x q2  = 0 ∂t ρu (3 − γ)u (γ − 1) 2 3 3 2 q3 q3 − {γetot u + (γ − 1)u } γetot + 2 (1 − γ)u γu (2.86) (2. If this was all.79) (2.76) The eigenvalues are λ− = u − C s λ0 = u λ+ = u + C s (2. Suppose that there is a speed limit of umax and that if the road is nearly empty. The conservation equation is: ∂t q + ∂x (qu) = 0 (2. Whitham and Richards ( for the references.83) where htot = etot + P/ρ is the total specific enthalpy and Cs = speed. Here is an example of a model of traffic flow. Let us for simplicity assume that u(q) = umax (1 − q) The flux of cars is then f = q(1 − q)umax (2.

while if we write ∂fi ∂fi ∂t qi + ∂y qk = − ∂x qk (2. while for higher q the flux is lower because the speed of the cars goes down (congestion).92) ∂qk ∂qk we focus on advection in y-direction and consider the x-direction as a source term. and consider the y-advection as a source term. If we split it as ∂fi ∂fi ∂t qi + ∂x qk = − ∂y qk (2.89) (2. not the speed of the advected quantity q itself. It is very important to note here that the characteristic speed is not equal to the speed of propagation of the cars! It is the speed at which information is propagating. We can now write the traffic flow equation as ∂t q + umax (1 − 2q)∂x q = 0 This shows that the characteristic velocity of the system is: λ = umax (1 − 2q) (2. and it is very similar to the peculiarities of the Burger’s equation. yet has very interesting solution properties. In 2-D we get ∂t Q + ∂x F (Q) + ∂y G(Q) which can be written as: ∂t qi + ∂fi ∂gi ∂x qk + ∂y qk ∂qk ∂qk (2. But what we learned can also be generalized to higher dimensions.88) (2.90) By operator splitting we can focus our attention to one of the space dimensions only. by use of the concept of operator splitting.87) which can range from −umax to umax . This is one of the peculiar features of non-linear hyperbolic equations.91) ∂qk ∂qk then we focus on the advection in x-direction.17 Hyperbolic equations in 2-D and 3-D So far we have done everything only in 1-D. 2. In numerical methods this operator splitting is often done to reduce the full 2-D or 3-D problem into consecutive 1-D problems which are much easier to handle. The traffic flow equation is a particularly nice example of a non-linear hyperbolic equation because it is a very simple equation.36 The flux of cars is greatest when q = 1/2. . For lower q the flux is lower because the density of cars is lower.

Sign up to vote on this title
UsefulNot useful