Partial Differential Equations (Week 2)

First Order PDEs

Gustav Holzegel
January 24, 2019

1 Introduction
Consider the following transport equation

ut + cux = 0 (1)

for a function u : R2 → R depending on x and t (time) and c > 0. As a first

idea, you may try to solve the above using a power series approach. Given
u (x, 0) = h (x) smooth for t = 0, one can indeed compute all partial derivatives
of u on t = 0 from the equation (note all ∂xk -derivatives are already given by
specifying h (x)):
 j+k
j d
∂tj ∂xk u (x, 0) = (−c) h (x) .
dx

The hope then is that

X (−c)j h(j+k) (0)
u (x, t) =! tj xk (2)
j!k!
j,k

converges in a neighborhood of (0, 0).

Exercise 1.1. Show that if h is real analytic and converges for |x| < R, then the
series on the right indeed converges to a solution of the PDE in a neighborhood
m
of
P (0, 0). m!
[Hint: You may want to use the multinomial identity (x1 + . . . xn ) =
α
|α|=m α! x .]

We will talk more about this power series approach when we discuss the
Cauchy-Kovalevskaya theorem for general PDEs (Week 3). Note already, how-
ever, that if h ∈ C ∞ and all derivatives vanish at 0 but not in a neighborhood
of 0 (so h is not analytic), then the series (2) converges, but not to a solution.

1
 In any case, there is a much simpler way of understanding (1). Suppose we
have a classical (i.e. C 1 ) solution u (x, t) of (1). We observe that along the lines
La defined by t = 1c x + a (and parametrized by a) the solution is constant:

d
u (c (s − a) , s) = ut + ux · c = 0 .
ds
This tells us that any C 1 solution u (x, t) will be uniquely determined by speci-
fying a C 1 function h (x) at t = 0 (or, more generally, on any line which is not
one of the La ). Indeed, the solution is simply going to get transported along the
characteristic direction La of the PDE. In particular, the domain of influence
of a point is precisely given by the characteristic line through that point. We
summarize this as
Proposition 1.2. The unique classical solution of (1) with initial data a C 1
function h (x) prescribed at t = 0 is given by u (x, t) = h (x − ct).
Remark 1.3. An interesting question which we will touch below is what happens
for initial functions h (x) which are not C 1 (propagation of singularities along
characteristics).

2 Quasi-linear case
Consider the general quasi-linear first order PDE

a (x, y, u) ux + b (x, y, u) uy = c (x, y, u) , (3)

for C 1 functions a, b, c : R2 × R → R. For simplicity, we will discuss this two-

dimensional case, the higher dimensional generalization being straightforward
(see Exercise 2 below).
Definition 2.1. Let z = u (x, y) be a C 1 -surface in R3 . If z = u (x, y) solves
(3), we will call the surface an integral surface of the PDE.
Suppose we are given an integral surface S, z = u (x, y). Then, the normal of
T
the tangent plane at any point of S is proportional to the vector (ux , uy , −1) by
T
standard multi-variable calculus. If we also interpret (a (x, y, z) , b (x, y, z) , c (x, y, z))
3
as a vector-field in R (defining the characteristic direction of the PDE at each
point), then the PDE (3) says precisely that the characteristic direction at each
point has to be perpendicular to the normal of of the tangent plane. In other
words, the characteristic direction has to lie in the tangent plane to the integral
surface. This suggest to obtain an integral surface via the integral curves along
the characteristic direction.
Suppose we are at a point (x0 , y0 , z0 ) ∈ S. Given the characteristic direction,

2
we can define an integral curve through (x0 , y0 , z0 ), by solving the ODE-system

dx
= a (x, y, z) x (0) = x0
dt
dy
= b (x, y, z) y (0) = y0
dt
dz
= c (x, y, z) z (0) = z0
dt
The corresponding curve, which exists by the ODE theory we developed (Exer-
cise: Check!) is called the characteristic curve.
Lemma 2.2. If a C 1 -surface S is a union of characteristic curves, then it is
an integral surface.
Proof. Through any point P we have a characteristic curve lying in S. Since
the tangent to the characteristic curve is everywhere equal to the characteristic
direction, the latter lies in the tangent plane to S. This means that the normal to
the tangent plane is perpendicular to the characteristic direction, which implies
that the PDE (3) is satisfied.
Proposition 2.3. Given an integral surface S of (3) defined by z = u (x, y)
and a point P = (x0 , y0 , z0 ) together with the characteristic curve γ through P ,
then γ lies entirely in S.
Proof. Let γ (t) = (x (t) , y (t) , z (t)) be the characteristic curve going through
P at t = 0 and consider the quantity

U (t) = z (t) − u (x (t) , y (t))

Clearly U (t = 0) = 0. If we can show that U = 0 for all t, then this clearly

implies that γ lies in S. Differentiating U and using that the characteristic
system of ODEs is satisfied, we find
dU
= c (x (t) , y (t) , z (t)) − ux · a (x (t) , y (t) , z (t)) − uy · b (x (t) , y (t) , z (t))
dt
which we may write as
dU
= c (x (t) , y (t) , U (t) + u (x (t) , y (t)))
dt
− ux · a (x (t) , y (t) , U (t) + u (x (t) , y (t)))
− uy · b (x (t) , y (t) , U (t) + u (x (t) , y (t))) .

This is an ODE for U (t), all other quantities being given by the characteristic
curve and the value of u along it. This ODE has the solution U = 0 identically
as is easily checked. However, by the uniqueness theorem for solutions to ODEs,
it is also the only solution.

3
 We conclude that the integral surface is the union of characteristic curves.
Now we want to formulate the appropriate Cauchy-problem inspired by our
initial example of the linear transport equation. The idea is to specify the
solution along a (non-characteristic) curve Γ and then solve the characteristic
ODEs which should provide a unique solution at least in a neighborhood of a
point.
Let the C 1 -curve Γ be represented parametrically by

x = f (s) y = g (s) z = h (s) (4)

and let x0 = f (s0 ), y0 = g (s0 ), z0 = h (s0 ). We are looking for a solution of

the PDE (3) such that we have

h (s) = u (f (s) , g (s))

for all s – this is our initial condition. (Note that in the simplest case (the
curve Γ being “time 0”) we have the familiar x = 0, y = s, z = h (s) = h (y)
and hence h (y) = u (0, y) initially.) We would like to construct a local solution
near x0 = f (s0 ) and y0 = g (s0 ), parametrized locally by x = X (s, t) , y =
Y (s, t) , z = Z (s, t) where X, Y, Z satisfy the characteristic ODEs

Xt = a (X (s, t) , Y, (s, t) , Z (s, t)) X (s, 0) = f (s)

Yt = b (X (s, t) , Y, (s, t) , Z (s, t)) Y (s, 0) = g (s) (5)
Zt = c (X (s, t) , Y, (s, t) , Z (s, t)) Z (s, 0) = h (s)

By existence and uniqueness and continuous dependence on the parameters of

the solution to the ODE system, we obtain a unique solution of class C 1 for
(s, t) near (s0 , 0).
Exercise 2.4. (Week 1 refresher) Show explicitly that the solution depends
continuously on the parameter s.
In order to represent the solution as an integral surface z = u (x, y) we need
to be able to invert the equations x = X (s, t) , y = Y (s, t) to s = S (x, y) and
t = T (x, y), since then u (x, y) = Z (S (x, y) , T (x, y)) is an integral surface.
The implicit function theorem tells us that this is possible locally near (s0 , 0),
provided the determinant of the differential
 ∂X ∂X

∂s (s, t) ∂t (s, t) (6)
∂Y ∂Y
∂s (s, t) ∂t (s, t)

is non-zero at (s0 , 0), which yields the condition

f ′ (s0 ) · b (x0 , y0 , z0 ) − g ′ (s0 ) · a (x0 , y0 , z0 ) 6= 0 (7)

Exercise 2.5. Show that with this condition being satisfied, the function u (x, y) =
Z (S (x, y) , T (x, y)) indeed solves the PDE (3). [Hint: This follows either from
carefully applying the chain rule or from observing that by (5), the parametrized
surface has the characteristic direction in its tangent space.]

4
 The previous exercise ensures existence of a solution to the PDE near (x0 , y0 ).
This solution is also the unique C 1 solution realizing the “data” prescribed on
Γ: Indeed, in view of Proposition 2.3, any integral surface through Γ has to
contain all characteristic curves through Γ and hence locally agrees with the
integral surface we constructed.
Let us finally investigate what happens if condition (7) is violated. In the
latter case we have

f ′b − g′a = 0 , h′ = u x f ′ + g ′ u y , c = aux + buy (8)

at (s0 , 0), which implies that the vectors (f ′ , g ′ , h′ ) and (a, b, c) are parallel
(Exercise), so Γ has to be characteristic at s = s0 in this case.
In summary, we have proven the following theorem
Theorem 2.6. Consider the PDE (3), a point P = (x0 , y0 , z0 = u (x0 , y0 )) in
R3 and a parametrized C 1 curve Γ : s → (f (s) , g (s) , h (s)) going through P
at s0 , which is also non-characteristic at P (i.e. (7) holds). Then, there exists
a small neighborhood U ⊂ R2 of (x0 , y0 ) and a unique C 1 function u : U → R
which solves the PDE (3) in U and satisfies h (s) = u (f (s) , g (s)) along Γ.
Remark 2.7. It is important to note that in the quasi-linear case, the condition
(7) of being non-characteristic at a point depends on the value of the solution
at that point.

3 Examples
3.1 The transport equation revisited
Let us go back to (1) with C 1 initial data h (x) prescribed at t = 0. The initial
data on the curve Γ is then given by t = 0, x = s, z = h (s) = h (x). The
characteristic equations are
dt dx dz
=1 , =c , =0 (9)
dτ dτ dτ
with initial conditions t (s, 0) = 0 (hence t (s, τ ) = τ ), x (s, 0) = s (hence
x (s, τ ) = c · τ + s) and z (s, 0) = h (s) (hence z (s, τ ) = h (s)). From this
we easily obtain u (x, t) = h (x − c · t) in agreement with what we found earlier.

3.2 Inviscid Burger’s equation

Consider the non-linear first order PDE
u2
 
ut + uux = ut + =0 (10)
2 x

with initial data prescribed at t = 0. This equation arises in fluid dynamics

and is the simplest model exhibiting shock formation. Recall that the equation

5
describing a free particle in Newtonian mechanics is ẍ (t) = 0. Defining the
velocity field u (x (t) , t) = ẋ (t), the chain rule implies that u (t, x (t)) satisfies
Burger’s equation. We will elucidate this relation further once we gained insight
on the structure of solutions to (10).
In the notation of the previous section, the initial data is prescribed on the
curve Γ given by t = 0, x = s, z = h (s) = h (x) (say again with h being C 1 ).
The characteristic equations are
dt dx dz
=1 , =z , =0 (11)
dτ dτ dτ
with initial conditions t (s, 0) = 0 (hence t (s, τ ) = τ ), z (s, 0) = h (s) (hence
dz
z (s, τ ) = h (s) and x (s, 0) = s (hence x (s, τ ) = z · τ + s noting that dτ = 0).
From this we obtain the implicit formula u (x, t) = h (x − u (x, t) · t).
To understand what general solutions look like, we depict the projected
characteristics of the solution in the (x, t) plane. Fixing a point (s, 0), where
u (s, 0) = h (s) we have that the solution is constant (equal to h (s)) along the
1
line x = h (s) · t + s (Exercise: Why?). In other words, h (s) (or rather h(s) )
determines the slope of the characteristic going through each point. From the
Newtonian particle point of view, we can think of h (x) as the initial velocity
distribution of particles. Every particle moves along its characteristic. It is
already clear graphically (see the figures below) that faster particles will even-
tually overtake slower ones and hence that the evolution will break down (unless
h (s) is monotonically increasing). More precisely, two characteristics C1 and
C2 will intersect at some (x, t) where
s2 − s1
t=− . (12)
h (s2 ) − h (s1 )
At the point of intersection, the solution u can no-longer be uni-valued. To
visualize what happens it is instructive to plot different snapshots of a solution
arising from data of compact support (Exercise). This breakdown of the solution
in finite time is characteristic of non-linear PDEs. Here, the geometry of the
break-up is simple enough to be understood.
It is clear that if we want to continue the solution beyond the shock, we
have to give up our requirement that the solution should be C 1 , i.e. classical.
However, if we do this, then, how should the PDE (10) be understood? The
key is the introduction of the notion of a so-called weak solution (below called
an integral solution) of (10). This will concern us in greater detail later in the
course once we have developed some distribution theory. For now we will only
provide a first taste by discussing the present example
The basic idea is simple. Suppose first that we have a C 1 solution of (10)
with initial datum u (x, 0) = h (x). Multiply the PDE by a smooth test function
v : R × [0, ∞) → R of compact support and integrate over spacetime. This
produces the identity
Z ∞ Z ∞  Z ∞ 
u2
 
dt dx uvt + vx + hvdx = 0. (13)

0 −∞ 2 −∞ 
t=0

6
Although we arrived at (13) assuming that u was classical, this identity makes
sense for much less regular u (and data h (x)). Therefore, we define
Definition 3.1. We say that u ∈ L∞ (R × (0, ∞)) is an integral solution of (10)
with data u (x, 0) = h (x) provided that (13) holds for any smooth test function
v : R × [0, ∞) → R of compact support.
Exercise 3.2. Show that if u is an integral solution that is also C 1 , then it is
a classical solution.
Of course, the difficult question to answer is: Does there always exist a global
weak solution of (10) and if so, is it unique?
We will not be able to answer this question in fully generality but will gain
further insight by studying a particular class of integral solutions. Let us inves-
tigate what happens for a piecewise smooth solution u given in some connected
region V ⊂ R×(0, ∞) which is discontinuous along a smooth curve C separating
V into Vl and Vr . We assume u is an integral solution and that moreover u and
its first derivatives are uniformly continuous in Vl and Vr .
Choosing a test function compactly supported in Vl (or Vr respectively) we
easily find that we must have

ut + uux = 0

in both Vl and Vr (recall Exercise 3.2). However, if the test function doesn’t
vanish on the curve
(but say still vanishes near the boundary of V ) we will find
(with ν = ν 1 , ν 2 denoting the unit-normal to C pointing from Vl into Vr )

u2 u2
ZZ Z  
uvt + vx dxdt = ul ν 2 + l ν 1 v dl
Vl 2 C 2

u2 u2r 1
ZZ Z  
2
uvt + vx dxdt = − ur ν + ν v dl
Vr 2 C 2
where ul and ur denote the limits from the left and the right respectively. Since
u is assumed to be an integral solution we must have
u2 − u2r 1
Z  
(ul − ur ) ν 2 + l ν v dl = 0 (14)
C 2
for all test functions, and hence
u2l − u2r 1
(ul − ur ) ν 2 + ν =0
2
along C. Suppose that the curve C is parametrized by γ (t) = (x = s (t) , t) for
some smooth function s : [0, ∞) → R. The tangent
is then proportional to (ṡ, 1)
1
and hence for the normal we obtain ν = ν 1 , ν 2 = √1+ ṡ2
(1, −ṡ). Therefore,
we have

u2l − u2r = 2ṡ (ul − ur ) . (15)

7
This is the Rankine-Hugoniot condition relating the jump in u to the speed ṡ.
Note that ul , ur and ṡ all change along C, however, the relation above must
always hold.
Exercise 3.3. Repeat the above computations for the more general PDE
ut + (F (u))x = 0

to conclude the Rankine-Hugoniot condition

F (ul ) − F (ur ) = ṡ (ul − ur ) . (16)

We will illustrate the above computation by examples. We consider the
evolution of the three initial-data sets

 1 if x ≤ 0,
h1 (x) = 1 − x if 0 ≤ x ≤ 1 (17)
0 if x > 1.


1 if x < 0,
h2 (x) = (18)
0 if x ≥ 0.

0 if x < 0,
h3 (x) = (19)
1 if x ≥ 0.
For h1 (piecewise smooth), the method of characteristics yields the solution
depicted below in (a).

x=0 x=1

(a) (b)

(c) (d)

8
Indeed, up to time t < 1 we have

 1 if x ≤ t and 0 ≤ t < 1,
1−x
u1 (x, t) = if t ≤ x ≤ 1 and 0 ≤ t < 1 (20)
 1−t
0 if x ≥ 1 and 0 ≤ t < 1.

How should u1 be defined for t ≥ 1? The RH-condition suggests that ṡ = 21

should hold along the curve separating the two regions where the solution is still
smooth. Hence we define s (t) = 1+t
2 and

1 if x < s (t) and t ≥ 1,
u1 (x, t) = (21)
0 if x > s (t) and t ≥ 1.

We have thus found a global integral solution to our PDE and it also turns out
to be unique. In understanding the problem for h1 , we have understood it for
h2 because the situation at t = 1 for h1 is precisely the situation for h2 at time
zero (only shifted).
The evolution of h3 is more tricky because of a failure of uniqueness. Note
that a-priori, the solution is not determined by the characteristics in a small
wedge opening up at zero. It is easy to check that both

0 if x < 2t ,

u3c (x, t) = (22)
1 if x > 2t ,

depicted in (c) above, and


 1 if x > t,
x
u3d (x) = t if 0 ≤ t ≤ x (23)
0 if x < 0.


depicted in (d), define an integral solution in R× [0, ∞) to (10) with data h3 (x).
This example shows that we can, in general, not expect to find a unique weak
solution of (10) unless we prescribe additional information (in the above case,
the rarefaction wave (d) turns out to be physically correct).
The additional information ensuring the uniqueness of weak solutions is
typically given by so-called entropy conditions. Suppose we are in the class
of piecewise-smooth integral solutions. We may hope that starting from a ar-
bitrary point P in the (t, x)-plane, we will not see any crossing characteristics
when moving towards the past along the characteristic through that point. This
requires ul > ṡ > ur to hold along a curve of discontinuities. It turns out (as
suggested by the examples) that the above condition along any shock curve suf-
fices to ensure uniqueness of weak solutions. For our example, it singles out the
“rarefaction wave” u3d as the unique weak solution.1
We will not go into the existence theory and uniqueness theory of global
weak solutions at this point. For further discussion, see Chapter 3 of Evans.
1 This is also the solution one obtains from the viscous Burger’s equation in the zero

viscosity limit.

9
4 The fully non-linear case
The geometry of fully non-linear first order PDEs is considerably more compli-
cated but remarkably, the local problem still reduces to that of solving ODEs.
I will first (Section 4.1) explain the geometric intuition leading to the system
of characteristic ODEs (30)–(32), (36), (37). The impatient reader may jump
immediately to this ODE system and – after doing the computation of Lemma
4.1 – continue with Section 4.2. See the book of Fritz John for a more detailed
treatment.

4.1 Heuristic derivation of the ODE system

We start with the PDE

F (x, y, z, p = ux , q = uy ) = 0 (24)

with (say) F being smooth. At a given point P = (x0 , y0 , z0 ), the tangent plane
to an integral surface z = u (x, y) is given by

z − z0 = p (x − x0 ) + q (y − y0 )

with p and q satisfying F (x0 , y0 , z0 , p, q) = 0. Suppose that we found an explicit

(p0 , q0 ) with F (x0 , y0 , z0 , p0 , q0 ) = 0. Then, provided Fq (x0 , y0 , z0 , p0 , q0 ) 6= 0,
we expect that for p near p0 we can solve for q (p) and thereby obtain an entire
family of possible tangent planes at P , parametrized by p.
How do we, from this, single out a characteristic direction? The key idea is
to use the envelope of the family of planes (i.e. the surface which is tangent to all
members of the family). The envelope touches each plane in a certain generator;
therefore, once we have chosen p and q at the point P (hence a particular tangent
plane), this determines a particular direction: the characteristic direction.2 The
envelope is typically a cone and one speaks of the field of Monge-cones associated
with the PDE (24).
Let us improve the above heuristics by some formulae. The envelope is
defined as follows. If

G (x, y, z, p) := p (x − x0 ) + q (p) (y − y0 ) − (z − z0 ) = 0 (25)

is our family of tangent planes, then its envelope is obtained by solving

∂p G (x, y, z, p) = 0 (26)

for p = p (x, y, z) and plugging it back into (25). The generator of intersection
of the envelope and the plane is given by solving (25) and (26) for fixed p. It is
convenient to work in the language of differential forms. For fixed p we have

dz = pdx + q (p) dy (27)

2 For the quasi-linear problem, this generator will be the same for any plane, as the envelope

degenerates to a line in this case, as can be checked by a quick computation.

10
describing the plane, while condition (26) yields
dq
dx + dy = 0 . (28)
dp
Finally, from differentiating F (x0 , y0 , z0 , p, q) = 0 we obtain
dq
Fp + Fq = 0 . (29)
dp
This suggests that we should solve the system of ODEs
dx
= Fp (x, y, z, p, q) , (30)
dt
dy
= Fq (x, y, z, p, q) , (31)
dt
dz
= pFp (x, y, z, p, q) + qFq (x, y, z, p, q) . (32)
dt
However, this is not a closed system unless we are in the quasi-linear case! In
general, only if we are given the integral surface we can determine p = ux
and q = uy and hence solve the above ODEs. Remarkably, we can actually
complement the above system by ODEs for p and q to obtain a closed system.
Here is how. Differentiate the PDE with respect to x and y to produce
Fx + Fz p + Fp uxx + Fq uxy = 0 ,
Fy + Fz q + Fp uxy + Fq uyy = 0 . (33)
On the other hand, along a characteristic we have
dp d
= (ux (x (t) , y (t))) = uxx Fp + uxy Fq , (34)
dt dt

dq d
= (uy (x (t) , y (t))) = uyx Fp + uyy Fq . (35)
dt dt
Combining these equations it becomes apparent that we should complement the
ODE system (30)–(32) above with the equations
dp
= −Fx (x, y, z, p, q) − pFz (x, y, z, p, q) , (36)
dt
dq
= −Fy (x, y, z, p, q) − qFz (x, y, z, p, q) , (37)
dt
thereby achieving closure. We will call the system (30)–(32), (36), (37) the
system of characteristic ODEs. The following is easy to check:
Lemma 4.1. The function F (x, y, z, p, q) is an integral of the system of char-
acteristic ODEs (30)–(32), (36), (37) in that
d
[F (x (t) , y (t) , z (t) , p (t) , q (t))] = 0
dt
holds along a solution.

11
Proof. Straightforward computation.
Geometrically, compared with the quasi-linear case, we are now propagat-
ing plane elements (x, y, z, p, q) in R5 (i.e. a point in R3 and a tangent plane)
according to the system of ODEs. We will call (x (t) , y (t) , z (t) , p (t) , q (t)) a
characteristic strip if the tangent to the curve defined by (x (t) , y (t) , z (t)) lies
in the tangent plane determined by the plane element. This is indeed ensured if
dy
the characteristic ODEs are satisfied, as they imply that dz dx
dt = p dt + q dt holds.

4.2 The Cauchy problem

We proceed with the Cauchy problem. As in the quasi-linear case, we would
like to specify data on a curve Γ parametrized by x = f (s), y = g (s), z = h (s)
with functions of class C 1 near s0 . However, to obtain appropriate data for the
characteristic ODEs, we now need to complete Γ into a strip, i.e. we need to
determine functions φ (s) and ψ (s) satisfying

h′ (s) = φ (s) f ′ (s) + ψ (s) g ′ (s) , (38)

0 = F (f (s) , g (s) , h (s) , φ (s) , ψ (s)) . (39)

In the quasi-linear case, these equations are linear in φ (s) and ψ (s); there-
fore, φ (s) and ψ (s) are determined uniquely near s0 (from f , g and h) in view of
the non-degeneracy condition (7). In the present fully non-linear case, the equa-
tions (38) and (39) may have no, one or many solutions. To ensure uniqueness,
we will need to invoke the implicit function theorem as follows:
Lemma 4.2. Given C 1 functions f (s), g (s), h (s) of class C 1 near s0 and
p0 , q0 ∈ R such that

h′ (s0 ) = p0 f ′ (s0 ) + q0 g ′ (s0 ) (40)

0 = F (f (s0 ) , g (s0 ) , h (s0 ) , p0 , q0 ) (41)

and also
f ′ (s0 ) g ′ (s0 )
 
det 6= 0
Fp (f (s0 ) , g (s0 ) , h (s0 ) , p0 , q0 ) Fq (f (s0 ) , g (s0 ) , h (s0 ) , p0 , q0 )
(42)

holds. Then, there exist unique C 1 -functions φ (s) and ψ (s) with φ (s0 ) = p0
and ψ (s0 ) = q0 satisfying both (38) and (39).
Proof. Standard application of the implicit function theorem.
We will call a quintuple (f (s) , g (s) , h (s) , p0 , q0 ) such that the conditions
(40)–(42) of Lemma 4.2 are satisfied an admissible data set of the PDE (24).
By the Lemma, an admissible data set determines uniquely a quintuple of func-
tions (f (s) , g (s) , h (s) , φ (s) , ψ (s)) of class C 1 near s0 which serves as a one-
parameter initial data set for the characteristic system of ODEs. Solving the

12
characteristic ODEs with this data we obtain unique C 1 functions

x = X (s, t) , y = Y (s, t) , z = Z (s, t) , p = P (s, t) , q = Q (s, t) . (43)

d
Since F = 0 holds identically along Γ and moreover dt F = 0 along the solutions
of the system of ODEs (cf. Lemma 4.1) we have that F (X (s, t) , ...., Q (s, t)) = 0
in a neighborhood of (s0 , 0). As any integral surface through Γ (if it exists)
contains its characteristic curves through Γ, it must must locally agree with the
surface we constructed (uniqueness).3
Conversely, (to show existence) let us show that the parametrized surface
(43) (obtained from an admissible data set) represents a solution of the Cauchy
problem for the PDE in a neighborhood of (x0 , y0 , z0 ). To do this, note first
that can invert the relation X (s, t) and Y (s, t) locally near (s0 , 0) in view of
the non-degeneracy condition (42) to obtain

s = S (x, y) t = T (x, y) . (44)

Then

u (x, y) := Z (S (x, y) , T (x, y)) (45)

defines a function which is C 1 is a neighborhood of (x0 = f (s0 ) , y0 = g (s0 ))

and agrees with the data prescribed on Γ. Observe that equivalently one may
write

Z (s, t) = u (X (s, t) , Y (s, t)) . (46)

We want to show that the u defined above actually solves the PDE. We claim
that for this it is sufficient to establish the two identities ux (X (s, t) , Y (s, t)) =
P (s, t) and uy (X (s, t) , Y (s, t)) = Q (s, t), where the right hand side is defined
by the solution of the characteristic ODEs. Indeed, given these identities, the
known (from solving the ODE system) fact that

F (X (s, t) , Y (s, t) , Z (s, t) , P (s, t) , Q (s, t)) = 0 holds near (s0 , 0) (47)

is equivalent to

F (x, y, u (x, y) , ux (x, y) , uy (x, y)) = 0 (48)

holding in a neighborhood of (x0 = f (s0 ) , y0 = g (s0 )). To finally prove the

desired identites note that the chain rule applied to (46) yields

Zs (s, t) = ux (X (s, t) , Y (s, t)) · Xs (s, t) + uy (X (s, t) , Y (s, t)) · Ys (s, t) (49)
Zt (s, t) = ux (X (s, t) , Y (s, t)) · Xt (s, t) + uy (X (s, t) , Y (s, t)) · Yt (s, t) (50)
3 If this is not clear, repeat the proof of Proposition 2.3 adapted to the fully non-linear
case.

13
for all (s, t) in a neighborhood of (s0 , 0). We are done if we can also establish
the identities (why?)
Zs (s, t) = P (s, t) Xs (s, t) + Q (s, t) Ys (s, t) , (51)
Zt (s, t) = P (s, t) Xt (s, t) + Q (s, t) Yt (s, t) . (52)
Now (52) follows directly from the characteristic system of ODEs. For (51), we
use a familiar trick: Defining A = Zs − P Xs − QYs we derive an ODE for At to
conclude that A = 0 on Γ implies A = 0 for all times (Exercise). Summarizing,
we have proven
Theorem 4.3. Let F (x, y, u, ux , uy ) = 0 be a PDE for a smooth function F :
R5 → R. Let (f (s) , g (s) , h (s) , p0 , q0 ) be an admissible data set for the PDE
(i.e. the assumptions of Lemma 4.2 are satisfied). Then there exists a unique
solution z = u (x, y) of the PDE in a neighborhood of (x0 = f (s0 ) , y0 = g (s0 )).
Of course, we have also obtained an explicit representation for the solution
in the above proof.

4.3 Example: The Eikonal Equation

As an interesting application of the above techniques, we will solve the eikonal
equation, which is central in geometric optics describing the motion of wave
fronts.

c2 u2x + u2y = 1


(53)

with c > 0 a constant, which can be interpreted as the speed. Geometrically, the
family of possible tangent planes at each point envelopes a cone with opening
angle 2θ0 = arctan c (Exercise: Why?). From
1 2 2
c p + c2 q 2 − 1


F (x, y, z, p, q) = (54)
2
we obtain the characteristic ODEs
dx dy dz dp dq
= c2 p , = c2 q , = c2 p + c2 q = 1 , = 0, = 0. (55)
dt dt dt dt dt
Given a curve Γ parametrized by x = f (s), y = g (s) and z = h (s), we need to
find φ (s), ψ (s) satisfying (cf. Lemma 4.2)

h′ (s) = φ (s) f ′ (s) + ψ (s) g ′ (s) and φ2 (s) + ψ 2 (s) = c−2 (56)
It is easy to see that there can be no solutions if (f ′ )2 + (g ′ )2 < c2 (h′ )2 (corre-
sponding to Γ making an angle θ < θ0 with the z-axis, “timelike”) and that in
the case (f ′ )2 + (g ′ )2 > c2 (h′ )2 there are two solutions differing in sign.
In the special case that x = f (s), y = g (s) and z = 0 (prescribing the
intersection of the integral surface with the xy-plane) we obtain
x (s, t) = f (s) + c2 φ (s) t , y (s, t) = g (s) + c2 ψ (s) t , z (t) = t , (57)

14
while p (s, t) = φ (s) and q (s, t) = ψ (s). It is instructive to draw some pictures
of the evolution. In particular, we can illustrate Huygen’s principle by depicting
the level surfaces of constant u (= constant t) in the xy-plane (Exercise). This
is an instructive example of the duality between a particle description (ODEs,
characteristics) and a wave description (PDE for u).

5 Exercises
1. Consider the PDE ut − iux = 0 for u (t, x) ∈ C. Identifying the (t, x) plane
appropriately with C, show that the solution u has to be holomorphic.
Conclude that the initial value problem can only be solved for analytic
data. Compare and contrast with the transport equation.
2. Generalize the quasi-linear theory developed for in two dimensions to any
dimension. In particular, state an appropriate Cauchy problem (cf. The-
orem 2.6 and Theorem 4.3).
3. (F. John’s PDE book; Picone) Let u be a solution of

a (x, y) ux + b (x, y) uy = −u

of class C 1 in the closed unit disk Ω in the xy-plane. Let a (x, y) x +

b (x, y) y > 0 hold along the boundary ∂Ω. Prove that u vanishes identi-
cally. (Hint: Show that maxΩ u ≤ 0 and minΩ u ≥ 0.)
4. (F. John’s PDE book) Show that the function u (x, t) defined for t ≥ 0 by
2 p 
u=− t + 3x + t2 for 4x + t2 > 0
3
u=0 for 4x + t2 < 0

is an integral solution of (10).

5. Find a global weak solution of Burger’s equation for t > 0 arising from
the initial data 
 0 if x ≤ 0,
h (x) = 1 if 0 ≤ x ≤ 1 (58)
0 if x > 1.


prescribed at t = 0.
6. (F. John’s PDE book) Solve the initial value problem for the PDE

ux + uy = u2

with initial data u (x, 0) = h (x).

15
7. (F. John’s PDE book) Solve the initial value problem for the PDE

u2x + u2y = u2

with the following initial data prescribed on the unit circle in the xy-plane:
u (cos s, sin s) = 1 for alls ∈ [0, 2π]. Is the
solution unique?
p
[Answer: u (x, y) = exp ± 1 − x + y 2 2 .]

16