Study material for Week-4 on

Partial Differential Equations

S. Sivaji Ganesh, IIT Bombay.

Distributed to students of MA 515, Autumn 2021.

104 Chapter 2. First order PDE

point (s1 , s2 ).

2.6 Burgers equation

In this section, we study properties of solutions to initial value problems
for Burgers equation. We investigate into propagation aspects of one of the
simplest class of initial profiles, which are given by monotonic functions.
Burgers equation is one of the simplest examples of a quasilinear equa-
tions whose solutions exhibit ‘typical behaviours’ of solutions to general
quasilinear equations. Our study forewarns us of difficulties and problems
awaiting us in studies of general quasilinear equations.
For our study, we admit initial profiles, i.e., the functions h used to pre-
scribe the initial conditions u(x, 0) = h(x), which are discontinuous or
piecewise linear functions apart from differentiable functions. For discon-
tinuous or piecewise linear functions h, the local existence and uniqueness
theorem Theorem 2.36 is not applicable. However, such initial profiles are
physically relevant, for example, in the study of traffic near a traffic sig-
nal. Of course, solutions must be interpreted in a generalized sense. An
advantage with such initial profiles is that computations are very easy, and
we find formal expressions (formulae) for solutions, and then ask when do
they actually define a solution. We would like to reiterate that observa-
tions made in these examples continue to hold with smoother functions h,
where analytical computations are almost impossible.
Thus, in summary, we are going to use such functions h, even though
they are not smooth.
In the following example, a solution to the Cauchy problem for Burgers
equation is found for a general initial profile h. In Burgers equation, the
variable y has the interpretation of time, thus justifying the usage of the
term initial profile for h.

Example 2.63 (Burgers equation).

Consider the Cauchy Problem for Burgers equation given by

uy + u u x = 0, u(x, 0) = h(x) for x ∈ R, y > 0.

Let us parametrize the given Cauchy data as

Γ : x = s, y = 0, z = h(s), s ∈ R.

Sivaji IIT Bombay

2.6. Burgers equation 105

The characteristic system of ODE for the given equation is

dx dy dz
= z, = 1, = 0. (2.106)
dt dt dt
The characteristic equations (2.106) is a system of linear ODE and hence
every IVP associated to it has a unique solution defined for all t ∈ R.
The unique solution of the characteristic system of ODE (2.106) satis-
fying the initial conditions

x(0) = s, y(0) = 0, z(0) = h(s).

is given by

x = X (t , s ) = h(s)t + s , y = Y (t , s) = t , z = Z(t , s ) = h(s ) , (2.107)

defined for (t , s) ∈ R × R. Eliminating t , s, we get

u = Z(T (x, y), S(x, y)) = h(x − y u).

Thus the solution to Cauchy problem is given implicitly by

u = h(x − y u) (2.108)

provided we can express u as a function of (x, y).

Note that the equation (2.108) is meaningful even if h is not differen-
tiable. We may be able to find a function u of (x, y) defined on some sub-
domain of Ω2 , the function thus obtained may turn out to be differentiable
on a further subdomain and hence is a solution to Burgers equation.

Comments on base characteristic curves

Note that the solution u(x, y) is a constant given by h(s), on the base
characteristic curve given parametrically by

x = h(s)t + s, y = t , t ∈ R

which is a straight line passing through the point (s, 0) on the x-axis, and
we represent this straight line by L s .
In cartesian coordinates, the equation for a base characteristic curve L s
(corresponding to a fixed s ∈ I ) is given by

L s : x = h(s)y + s.

August 9, 2021 Sivaji

106 Chapter 2. First order PDE

1
Note that the slope of the straight line L s is . If base characteris-
h(s)
tic curves passing through distinct points (s1 , 0) and (s2 , 0) on x-axis inter-
sect, then there is ambiguity in what should be the value of a solution at
the intersection point, since a solution is constant along each of the base
characteristic curves. This naturally leads to the investigation of points of
intersection of distinct base characteristic curves.
Note that the base characteristics L s1 and L s2 corresponding to distinct
s1 and s2 intersect at a point (x0 , y0 ) if and only if the following system has
a solution     
1 −h(s1 ) x0 s1
= .
1 −h(s2 ) y0 s2
The determinant of the matrix in the above equation is h(s1 ) − h(s2 ). If
h(s1 ) = h(s2 ), then the lines L s1 and L s2 are parallel lines passing through
(s1 , 0) and (s2 , 0) respectively. Thus L s1 and L s2 do not intersect.
If h(s1 ) 6= h(s2 ), then both straight lines meet at a unique point given by
   
x0 1 s2 h(s1 ) − s1 h(s2 )
= ,
y0 h(s1 ) − h(s2 ) s2 − s1

leading to an ambiguity over the value of u(x0 , y0 ): whether it should be

h(s1 ) or h(s2 ).

Comments on the transversality condition

Note that the Jacobian J (t , s ) is given by

 
∂ (X , Y )  h(s) h 0 (s)t + 1 

J (t , s) = =  = −1 − h 0 (s)t .
∂ (t , s)  1 0 
Based on our experience with Cauchy problems so far, we expect troubles
at the points (t , s ) where 1+h 0 (s)t = 0, i.e., at (t , s) satisfying the condition
1
t = − h 0 (s) . In terms of (x, y)-coordinates (by the equation (2.107)), these
1
points correspond to y = − h 0 (s) . Thus if h 0 (s) < 0 for some s, then we
expect singularities at some y > 0, and the smallest negative value of h 0 (s)
corresponds to the smallest value of y at which singularities in solutions are
expected. Indeed solution breaks down for the first time at such a point,
see (2.111).

Sivaji IIT Bombay

2.6. Burgers equation 107

2.6.1 Propagation of monotonic initial profiles

We consider four examples of Cauchy problems for Burgers equation, with

initial profiles which are monotonic functions that are either piecewise
linear or piecewise constant. Note that such functions are differentiable
throughout their domain except for a few points.
However, explicit expression for u(x, y) can be obtained using the im-
plicit relation (2.108). The function u exhibits the effects of nonlinear na-
ture of Burgers equation, and the monotonicity of the initial profile. Thus
these examples highlight the role played by nonlinearity of a PDE in solu-
tions to Cauchy problems.
In the first example, the method of characteristics fails to determine a
solution in some region of the upper half-plane. This is due to the discon-
tinuity in the piecewise constant, non-decreasing initial profile.

Example 2.64 (Base characteristics do not reach some points).

Consider Burgers equation with Cauchy data given by

−1 if x < 0,
h(x) =
1 if x ≥ 0.

Equation for the family of base characteristics (indexed by s ∈ R) is given

by
1 s
Ls : y = x− .
h(s) h(s)
Thus for s < 0, the base characteristic curve L s is given by the line y =
−x + s. Thus L s (s < 0) is a family of lines parallel to y = −x. The solution
u = −1 along each of these straight lines, since the solution is constant
along each of the base characteristics by (2.107). Similarly, for s ≥ 0 the
base characteristic curve L s is given by the line y = x − s. Thus L s (s ≥ 0)
is a family of lines parallel to y = x. The solution u = 1 along each of these
straight lines. See Figure 2.14 for an illustration.
However, no base characteristic curve passes through the V -shaped re-
gion in the upper half-plane which is bounded by the lines y = x and
y = −x as illustrated in Figure 2.14. Base characteristic curves carry infor-
mation from Cauchy data (prescribed on x-axis in this case) into rest of the
upper half-plane. Since there are no base characteristics passing through
the V -shaped region, we may say that information from the Cauchy data
does not reach this region. Due to the jump discontinuity of the function
h, the base characteristics L s also have a jump in their slopes (from −1 to

August 9, 2021 Sivaji

108 Chapter 2. First order PDE

No base characteristics.
Solution is undetermined.

y
=

=
−

x
y
u(x, y) = −1 u(x, y) = 1

−4 −3 −2 −1 0 1 2 3 4

Figure 2.14. Example 2.64: base characteristics and the solution

1) across s = 0, resulting in the V -shaped region. In Section 2.7, we ob-

tain a solution of the Cauchy problem in the V -shaped region, but in a
generalized sense.
Despite the discontinuity in the initial profile h, we could determine
solution to the Cauchy problem in two subsets of the upper half-plane,
namely
(i) On the domain {(x, y) ∈ R2+ : x + y < 0}, the function u defined by
u(x, y) = −1 is a solution to the Cauchy problem.
(ii) On the domain {(x, y) ∈ R2+ : x − y > 0}, the function u defined by
u(x, y) = 1 is a solution to the Cauchy problem.

In the next example, solution becomes multi-valued in some region of

the upper half-plane. This is due to the interaction of nonlinearity of Burg-
ers equation with non-increasing Cauchy data.
Example 2.65 (Intersecting base characteristics). Consider Burgers equa-
tion with Cauchy data given by

1 if x < 0,
h(x) =
−1 if x ≥ 0.

Equation for the family of base characteristics (indexed by s ∈ R) is given

by
1 s
Ls : y = x− .
h(s) h(s)
Sivaji IIT Bombay
2.6. Burgers equation 109

Intersecting base characteristics.

Solution becomes multi-valued.

u(x, y) = 1 u(x, y) = −1

−4 −3 −2 −1 0 1 2 3 4

Figure 2.15. Example 2.65: base characteristics and the solution

Thus for s < 0, the base characteristic curve L s is given by the line y =
x − s. Thus L s (s < 0) is a family of lines parallel to y = x. The solution
u = 1 along each of these straight lines, since the solution is constant along
each of the base characteristics by (2.107). Similarly, for s ≥ 0 the base
characteristic curve L s is given by the line y = −x + s . Thus L s (s ≥ 0) is
a family of lines parallel to y = −x. The solution u = −1 along each of
these straight lines. See Figure 2.15 for an illustration.
However, two base characteristics pass through each point of the V -
shaped region that is bounded by the lines y = x and y = −x in the upper
half-plane: one coming from the negative x-axis, and another from positive
x-axis (see Figure 2.15). As a consequence, multiplicity of information is
reaching this V -shaped region, and they are in conflict (base characteristics
coming from negative, and positive x-axis carry the information u = 1,
and u = −1 respectively), and thus solution becomes multi-valued in the
V -shaped region. The question is how to make sense of solution at those
points? This question will be discussed in Section 2.7.
Note that despite the non-differentiability of the initial profile h, we
could determine solution to the Cauchy problem in two subsets of the up-
per half-plane, namely

(i) On the domain {(x, y) ∈ R2+ : x + y < 0}, the function u defined by
u(x, y) = 1 is a solution to the Cauchy problem.
(ii) On the domain {(x, y) ∈ R2+ : x − y > 0}, the function u defined by
u(x, y) = −1 is a solution to the Cauchy problem.

August 9, 2021 Sivaji

110 Chapter 2. First order PDE

In the next example, the initial profile is a linear polynomial.

Example 2.66. Consider Burgers equation with Cauchy data given by

h(x) = 1 − x, for x ∈ R.

Note that the Cauchy data is a linear function, and thus it is continuously
differentiable.
Equation for the family of base characteristics (indexed by s ∈ R) is
given by
1 s
Ls : y = x−
h(s) h(s)
for s 6= 1, and L1 is the y-axis.
Note that any two distinct base characteristic curves intersect at the
point (x, y) = (1, 1) as illustrated in Figure 2.16, and hence solution be-
comes multi-valued at the point (1, 1).
The solution itself may be obtained from equation (2.108) as

1− x
u(x, y) = for x ∈ R, y 6= 1.
1−y
Note that the domain of definition of u consists of two disjoint sets {(x, y) ∈
R2+ : y < 1} and {(x, y) ∈ R2+ : y > 1}. The function u represents a solu-
tion to Burgers equations in each of these domains. However if we want
to talk about a solution of Cauchy problem, it is defined only on the set
{(x, y) ∈ R2+ : y < 1} as the set {(x, y) ∈ R2+ : y > 1} is not in touch with
x-axis where initial conditions are prescribed.
It appears from Figure 2.16 that there are regions in the upper half-plane
through which no base characteristic curve passes. But this impression is
wrong, since the figure is drawn only for a limited range of values of s.
Base characteristics emanating from points on the positive (respectively,
negative) x-axis fill-up the second (respectively, first) quadrant.

The next example is the most complicated compared to the three exam-
ples presented so far. The initial profile h is defined in three pieces, constant
on two of the pieces as in Examples 2.64 and 2.65, and linear on the third
piece as in Example 2.66.

Sivaji IIT Bombay

2.6. Burgers equation 111

1−x
u(x, y) = 1−y

y =1

−4 −3 −2 −1 0 1 2 3 4

Figure 2.16. Example 2.66: base characteristics and the solution

Example 2.67. Consider Burgers equation with Cauchy data given by



 1 if x ≤ 0,
h(x) = 1 − x if 0 ≤ x ≤ 1,


0 if x ≥ 1.

In this example, there are three distinct families of base characteristic

curves (see Figure 2.17). They are

(i) F1 family: base characteristics emanating from points (s, 0) where

s ≤ 0. These are lines parallel to y = x, and along them solution
u = 1.
(ii) F2 family: base characteristics emanating from points (s, 0) where
0 < s < 1. This family has varied slopes, as s → 1 the slopes increase
from 1 (at s = 0) to ∞ (at s = 1).
(iii) F3 family: base characteristics emanating from points (s, 0) where
s ≥ 1. These are lines parallel to x = 1, and along them u = 0.

Since the function h is non-increasing function, the slopes of base charac-

teristics increase with s ∈ R, some of them would intersect. Since solution
is constant along each base characteristic, at points of intersection of two
base characteristics on which solution is different, solution becomes multi-
valued, and loses its meaning.
Since a unique characteristic curve passes through every point of the da-
tum curve, the base characteristics (which are projections of characteristic
curves) do not intersect on x-axis. Thus we are interested in knowing the

August 9, 2021 Sivaji

112 Chapter 2. First order PDE

largest positive value of yc such that no two base characteristics intersect

at any point whose y-coordinate is less than yc . This leads us to finding
points of intersection of the base characteristics.
Let L s1 and L s2 be two base characteristics emanating from the points
(s1 , 0) and (s2 , 0) respectively, with s1 < s2 . Recall that L s1 and L s2 are
straight lines, and are given by the equations

L s1 : x = h(s1 )y + s1 ,
L s2 : x = h(s2 )y + s2 .

If L s1 and L s2 intersect at a point (x0 , y0 ), then we have

h(s1 )s2 − h(s2 )s1 s2 − s1

x0 = , y0 = .
h(s1 ) − h(s2 ) h(s1 ) − h(s2 )

We have the following observations regarding L s1 and L s2 for different

choices of s1 , s2 :

(i) L s1 and L s2 are from F1 family: If s1 < s2 ≤ 0, then L s1 and L s2 do not

intersect as both of them are parallel to the line y = x. Thus no two
members of the F1 family intersect.
(ii) L s1 is from F1 family, and L s2 is from F2 family: If s1 ≤ 0 < s2 < 1,
then L s1 and L s2 intersect at

s2 − s1 + s1 s2 s2 − s1
x0 = , y0 = .
s2 s2

We are interested in finding out the minimum value of y0 as s1 varies

in (−∞, 0], and s2 varies in (0, 1). In other words, as y is interpreted as
time, we are interested in finding out the first time that some member
of the family F1 intersects a member of the family F2. Minimum
value of y0 is obtained when s1 = 0 and s2 is arbitrary, and equals 1.
The corresponding value of x0 is then given by x0 = 1. In other words,
each L s2 intersects L0 at the point (1, 1), as illustrated in Figure 2.17.
(iii) L s1 is from F1 family, and L s2 is from F3 family: If s1 ≤ 0 and
s2 ≥ 1, then L s1 and L s2 intersect at

x0 = s2 , y0 = s2 − s1 .

The minimum value of y0 as s1 varies in (−∞, 0], and s2 varies in

[1, ∞) is the first time that some member of the family F1 intersects

Sivaji IIT Bombay

2.6. Burgers equation 113

a member of the family F3, is obtained when s1 = 0 and s2 = 1, and

is equal to 1. The corresponding value of x0 is then given by x0 = 1.
In other words, L1 intersects L0 at the point (1, 1), as illustrated in
Figure 2.17.
(iv) L s1 and L s2 are from F2 family: If 0 < s1 < s2 < 1, then L s1 and L s2
intersect at

x0 = 1, y0 = 1.

That is, all the characteristics from the family F2 intersect at the point
(1, 1), as illustrated in Figure 2.17.
(v) L s1 is from F2 family, and L s2 is from F3 family: If 0 < s1 < 1 and
s2 ≥ 1, then L s1 and L s2 intersect at

s2 − s1
x0 = s2 , y0 = .
1 − s1

Note that y0 ≥ 1. The minimum value of y0 as s1 varies in (0, 1), and

s2 varies in [1, ∞) is the first time that some member of the family F2
intersects a member of the family F3, is obtained when s2 = 1 and s1
is arbitrary, and equals 1. The corresponding value of x0 is then given
by x0 = 1. In other words, every L s1 from family F2 intersects L1 at
the point (1, 1), as illustrated in Figure 2.17.
(vi) L s1 and L s2 are from F3 family: If 1 ≤ s1 < s2 , then L s1 and L s2 do not
intersect as both of them are parallel to the line x = 1.

From the above observations on base characteristic curves, note that no

two base characteristics intersect in the open region bounded by x-axis and
the line y = 1. The solution is uniquely determined and is given by

 1 if x ≤ y,

1−x
u(x, y) = if y ≤ x ≤ 1,


1−y
0 if x ≥ 1.

Observe that u is not a differentiable function on the line segments

x = y, and x = 1,

with 0 < y < 1.

Note that a solution to the Cauchy problem is also determined in two
subsets of the upper half-plane, namely

August 9, 2021 Sivaji

114 Chapter 2. First order PDE

Intersecting base characteristics

Solution becomes multi-valued

y =1
u(x, y) = 1 u(x, y) = 0
1−x
u= 1−y

−4 −3 −2 −1 0 1 2 3 4

Figure 2.17. Example 2.67: base characteristics and the solution

(i) On the domain { (x, y) ∈ R2+ : (x < 0, x < y < 1) or (x < 1, y ≥

1) }, the function u defined by u(x, y) = 1 is a solution to the Cauchy
problem.
(ii) On the domain { (x, y) ∈ R2+ : (x > 1, y ≤ 1) or (x > y > 1) }, the
function u defined by u(x, y) = −1 is a solution to the Cauchy prob-
lem.

A generalized solution to this problem will be discussed in Section 2.7.

Intersecting base characteristics and Gradient catastrophe

In Example 2.67, it was observed that

(i) exactly one base characteristic passes through every point (x, y) with
y < 1.
(ii) for each y ≥ 1, there exists a point (x, y) through which two or more
base characteristics pass.

Thus y = 1 is the critical value as far as crossing of base characteristics is

concerned. The graphs of u(x, y) are shown for y = 0, 0.25, 0.5, 0.75, 0.9
in Figure 2.18, from which we observe that the graph of x 7→ u(x, y) be-
comes steeper near x = 1, as y → 1−. This is due to the reduction in the
lengths of the intervals within which the function u(· , y) transits from 1
to 0, as y → 1−. This phenomenon is referred to as gradient catastrophe.
In Example 2.67, gradient catastrophe arises as a result of intersecting base

Sivaji IIT Bombay

2.6. Burgers equation 115

Graph of u(x, 0)
−4 −3 −2 −1 0 1 2 3 4

Graph of u(x, 0.25)

−4 −3 −2 −1 0 1 2 3 4

Graph of u(x, 0.5)

−4 −3 −2 −1 0 1 2 3 4

Graph of u(x, 0.75)

−4 −3 −2 −1 0 1 2 3 4

Graph of u(x, 0.9)

−4 −3 −2 −1 0 1 2 3 4

Figure 2.18. Example 2.67: Development of a Gradient catastrophe at y = 1.

characteristics on which u carries distinct values. For instance, the base

characteristics y = x, x = 1 meet at the point (1, 1), and u takes the values
1 and 0 respectively on the base characteristics y = x and x = 1.
Let s1 < s2 , and let L s1 and L s2 be two base characteristics corresponding
to Burgers equation (posed on the upper half-plane i.e., y > 0) with Cauchy
data u(x, 0) = h(x). We know that L s1 and L s2 are straight lines, and are
given by the equations

L s1 : x = h(s1 )y + s1 ,
L s2 : x = h(s2 )y + s2 .

If L s1 and L s2 intersect at a point (x0 , y0 ), then

h(s1 )s2 − h(s2 )s1 s2 − s1

x0 = , y0 = .
h(s1 ) − h(s2 ) h(s1 ) − h(s2 )

Case (i). h is an increasing function.

The function h is increasing means that h(s1 ) < h(s2 ) whenever s1 <
s −s1
s2 . As s1 < s2 , we have h(s 2)−h(s )
< 0. This rules out the possibility of
1 2

August 9, 2021 Sivaji

116 Chapter 2. First order PDE

intersecting base characteristics (as seen in Example 2.64 ) in the upper half-
plane.

Case (ii). h is a decreasing function.

s −s1
For a decreasing function h, we have h(s 2)−h(s > 0. Thus base charac-
1 2)
teristics L s1 and L s2 intersect. This results in a gradient catastrophe which
is corroborated with explicit computations given below.
Differentiating both sides of u(x, y) = h(x − y u(x, y)) w.r.t. x gives

u x (x, y) = h 0 (x − y u(x, y))(1 − y u x (x, y)).

Re-arranging terms in the last equation yields

h 0 (x − y u(x, y))
u x (x, y) = (2.109)
1 + y h 0 (x − y u(x, y))

When the point (x, y) lies on a base characteristic curve L s , then the equa-
tion (2.109) takes the form, in view of (2.107),

h 0 (s)
u x (x, y) = .
1 + t h 0 (s)

Thus u x (x, y) will become infinite if (x, y) ∈ L s and

1
y =− . (2.110)
h 0 (s )

If h is a decreasing function, then h 0 takes negative values, and there exists

a y > 0 satisfying the equation (2.110). Thus gradient catastrophe occurs
at such a point (x, y). The smallest value of y such that there exists a point
(x, y) at which gradient catastrophe occurs, is called breaking time of solu-
tion u, and is denoted by yc . Indeed, yc is given by
1
yc = − . (2.111)
inf h 0 (s)
s∈R

In terms of existence of solutions, it is easier to observe that yc has the

following qualitative characterization

yc = sup {y ∈ (0, ∞) : u(x, y) is defined for all x ∈ R} .

Note that in Example 2.66, h 0 (s) = −1, and thus gradient catastrophe oc-

Sivaji IIT Bombay

2.6. Burgers equation 117

curs at yc = 1, and the derivative of the solution is given by

1
u x (x, y) = − ,
1−y

from which we conclude that gradient catastrophe occurs only at y = 1.

2.6.2 Domain of dependence, Domain of influence

Base characteristic curves are carriers of information from the Cauchy data
into the domain Ω2 . It may happen that the information from the Cauchy
data may not reach all points of the domain Ω2 (as in Example 2.64), or
excessive information reaches some points (as in Example 2.65). These ob-
servations lead us to seek answers to the following questions:

(i) Given a point (x, y) in the domain Ω2 , which part of the Cauchy data
determines solution at (x, y)?
(ii) Given a piece of the initial data, on which part of the domain Ω2 does
it influence the solution?

Note that the above two questions are questions of cause and effect,
and are dual to each other. They give rise to the concepts of domain of
dependence and region of influence, which are special to hyperbolic partial
differential equations. We revisit these concepts in the context of wave
equation in Section ??. We use a simple example to illustrate these concepts.

Example 2.68. Consider the following Cauchy problem

u x + u t = 0, (2.112a)
u(x, 0) = sin x, for x ≥ 0. (2.112b)

The general solution of u x + u t = 0 is given by u(x, y) = F (x − t ) where

F : R → R is an arbitrary differentiable function. Note that F (x) = u(x, 0).
Thus, the solution of the Cauchy problem (2.112) is given by

u(x, t ) = sin(x − t ),

for (x, t ) belonging to the set {(x, t ) ∈ R2 : t ≤ x}.

Note that no base characteristic curve passes through the points in the
x t -plane which lie to the left side of the line t = x. Thus the information
from the initial data does not reach those points, and hence Cauchy data
does not determine a solution at these points. In other words, domain of

August 9, 2021 Sivaji

Chapter 3

Classification of Second
order PDEs

In this chapter, the second order quasilinear PDEs are classified into three
types, namely hyperbolic, parabolic, elliptic and reducing them to canonical
forms is discussed. The classification is based on the coefficients of all the
second order derivatives of the unknown function appearing in the equa-
tion, and is invariant under change of coordinate systems.
A major part of this chapter is devoted to PDEs where the number of
independent variables is two for two reasons: (i). the notational clutter
will be minimum, and (ii). finding canonical forms is either difficult or
impossible when the number of independent variables exceeds two.
The classification is based on the notion of a characteristic surface (char-
acteristic curve in the case of two independent variables) that arises very
naturally in the analysis of second order quasilinear PDEs. We observe
that a certain geometrical object appears in at least two different contexts
dealing with second order quasilinear PDEs, and serves as a motivation
for defining the notion of a characteristic surface and investigate the PDEs
using this notion.
Consider a Cauchy problem posed for a quasilinear second order PDE,
where Cauchy data is prescribed on a curve Γ in R2 (a surface in Rd ). A
well-known formal procedure for solving the Cauchy problem is to pro-
pose a candidate solution via a formal Taylor series around a point P on
the curve Γ . In order to propose the Taylor series, all the derivatives of
the unknown function at P need to be determined, using the PDE and the
Cauchy data. The computation of derivatives of the unknown function
proceeds from finding first order derivatives to higher order derivatives.
Computation of first order derivatives is easy and is related to the regular-
ity of the curve Γ alone. Computation of the second order derivatives hits
a roadblock if the curve Γ is special to the PDE, and the details are found in

143
144 Chapter 3. Classification of second order PDEs

Section 3.1. These curves which are special to the PDE also turn out to be
possible curves along which a piecewise smooth solution to the quasilinear
PDE admits discontinuities in second order derivatives as demonstrated in
Section 3.2.
The above discussion leads us to the natural question, namely, does ev-
ery second order quasilinear PDE has special curves associated to it? Based
on the answer to this question, the classification of the quasilinear second
order PDEs is performed in Section 3.3. Canonical forms for the linear
PDEs are derived in Section 3.4. The ideas of classification and canonical
forms are extended to the class of second order PDEs in more than two
independent variables in Section 3.5.
Roughly speaking, the classification is similar to that of conic sections
in the plane geometry, while canonical forms are similar to the standard
forms of the conic sections.
Recall that a second order quasilinear PDE in two independent variables
is of the form

a(x, y, u, u x , uy )u x x + 2b (x, y, u, u x , uy )u x y + c(x, y, u, u x , uy )uy y

+d (x, y, u, u x , uy ) = 0. (3.1)

where a, b , c, d are functions defined on a domain Ω5 in R5 . Denote the

projections of Ω5 to x y-plane by Ω2 , and to x y z-space by Ω3 . The following
special case of (3.1), namely the linear equation, will also be considered in
the sequel.

a(x, y)u x x + 2b (x, y)u x y + c(x, y)uy y + d (x, y)u x + e(x, y)uy
+ f (x, y)u + g (x, y) = 0. (3.2)

3.1 Cauchy problem for quasilinear PDE

In this section, Cauchy problem for a second order quasilinear PDE in two
independent variables is introduced. In an attempt to construct a formal
solution to the Cauchy problem, obstacles in computing partial derivatives
of a possible solution from the PDE and the Cauchy data are analyzed.

Cauchy problem

Let Γ2 ⊂ R2 denote a regular parametric curve in the plane. That is,

Sivaji IIT Bombay

3.1. Cauchy problem for quasilinear PDE 145

(i) (Parametric smooth curve) There exist an interval I in R, and func-

tions f , g ∈ C 1 (I ) such that Γ2 is described parametrically by

Γ2 : x = f (s), y = g (s), s ∈ I ,

(ii) (Regularity) and for each s ∈ I

f 0 (s), g 0 (s ) 6= (0, 0)

holds.

Geometric significance of the regularity assumption on Γ2 is that the curve

possesses a well-defined tangent at each of its points. Thus, we are in a
position to define a unit normal n( f (s), g (s)) at the point ( f (s), g (s )) ∈ Γ2
(for s ∈ I ) by

(− g 0 (s), f 0 (s))
n( f (s), g (s )) := p .
( f 0 (s ))2 + ( g 0 (s))2

Statement of the Cauchy problem

Given functions h, χ ∈ C 1 (I ), Cauchy problem for the quasilinear sec-

ond order PDE (3.1) consists of finding a twice continuously differentiable
function u : D → R where D is a subset of Ω2 , such that
(i) the function u is a solution to the PDE (3.1),
(ii) and satisfies the two conditions

u( f (s ), g (s)) = h(s), (3.3a)

∂u
( f (s), g (s )) = χ (s) (3.3b)
∂n
for s belonging to a subinterval of I .
The functions h, χ are called Cauchy data, and the conditions (3.3) are
called Cauchy conditions.
Geometrically speaking, if we define a space curve Γ ⊂ R3 parametri-
cally by

Γ : x = f (s), y = g (s), z = h(s), s ∈ I ,

then the planar curve Γ2 is the projection of Γ to xy-plane, and the surface

August 9, 2021 Sivaji

146 Chapter 3. Classification of second order PDEs

z = u(x, y) contains a part of Γ . Note that the domain of the solution u,

namely D, contains a part of the curve Γ2 .

Computation of derivatives of a solution to Cauchy problem

A classical strategy, which is the essential idea behind the proof of Cauchy-
Kowalewski theorem [31], to solve the Cauchy problem in a neighbour-
hood of the point P0 (x0 , y0 ) = ( f (s0 ), g (s0 )) ∈ Γ2 is as follows:

(i) Determinie all the derivatives of a possible solution at P0 using the

Cauchy conditions (3.3) and the PDE (3.1).
(ii) Propose a cadidate solution defined by a formal Taylor series around
the point P0 .

The above strategy hopes that the formal Taylor series converges in a neigh-
bourhood of the point P0 and is a solution to the Cauchy problem.
In what follows, we explore the possibility of determining all the partial
derivatives of a solution at P0 using the PDE (3.1) and the Cauchy condi-
tions (3.3); and identify possible obstructions if any.
Of course, to implement the strategy outlined above, we require that all
the functions involved in the Cauchy problem, namely a, b , c, d , f , g , h, χ ,
have derivatives of all orders.

Computation of first order derivatives

Let u be a solution to the PDE (3.1) defined in a neighbourhood D of

the point P0 , and satisfying the Cauchy conditions (3.3). Without loss of
generality, assume that Γ2 ⊂ D.
For brevity in the presentation, let us introduce the functions defined
for s ∈ I by

p(s) := u x ( f (s), g (s)), (3.4a)

q(s) := uy ( f (s), g (s )). (3.4b)

In terms of p and q, the condition (3.3b) takes the form

− p(s )g 0 (s) + q(s) f 0 (s)

p = χ (s). (3.5)
( f 0 (s))2 + ( g 0 (s ))2

Since u also satisfies the condition (3.3a), we have h(s) = u( f (s), g (s))

Sivaji IIT Bombay

3.1. Cauchy problem for quasilinear PDE 147

for all s ∈ I . On differentiating this equation w.r.t. s, we get for all s ∈ I ,

d
h 0 (s) = (u( f (s), g (s))) = u x ( f (s), g (s)) f 0 (s) + uy ( f (s), g (s))g 0 (s)
ds
In view of the notations (3.4), the last equation reduces to

h 0 (s) = p(s ) f 0 (s ) + q(s )g 0 (s). (3.6)

The linear system of equations (3.5) - (3.6) has a unique solution for p(s )
and q(s ), since ( f 0 (s))2 + (g 0 (s))2 6= 0. Thus both the first order derivatives
have been determined at all the points of Γ2 , under the assumption of reg-
ularity on the curve Γ2 .

Remark 3.1 (on the computation of first order derivatives). The above
arguments prove that for any function u defined on a neighbourhood of Γ2 ,
where Γ2 is a regular curve in R2 , the computation of first order derivatives
at points of Γ2 requires the knowledge of the function u and its normal
derivative along Γ2 . This is not surprising in view of the following obser-
vations.

(i) The Cauchy data (3.3) contains the information on directional deriva-
tives of u in two independent directions, namely tangential ((3.3a))
and normal directions ((3.3b)).
(ii) The two partial derivatives are derivatives in the two coordinate di-
rections.
(iii) For a differentiable function, knowledge of any two directional deriva-
tives, where directions are linearly independent, is enough to deter-
mine any other directional derivative as the map

v 7−→ Dv u(P )
is a linear functional on R2 , which is fully determined once its values
on a basis is known. Once the total derivative is known, any direc-
tional derivative can be determined.
(iv) At any point on Γ2 , the tangential and normal directions are always
linearly independent.
(v) Thus prescribing the function and its normal derivative along a reg-
ular curve Γ2 (different from any of the lines parallel to x-axis or y-
axis) is equivalent to prescribing the function and its first order partial
derivatives.

August 9, 2021 Sivaji

148 Chapter 3. Classification of second order PDEs

Computation of second order derivatives

On differentiating the equations (3.4), we get

p 0 (s ) = u x x ( f (s), g (s)) f 0 (s) + u xy ( f (s), g (s ))g 0 (s) (3.7a)

q 0 (s) = u xy ( f (s), g (s )) f 0 (s ) + uy y ( f (s), g (s))g 0 (s). (3.7b)

Since u solves the PDE (3.1), for every s ∈ I we have

a(ζ (s))u x x ( f (s), g (s )) + 2b (ζ (s ))u xy ( f (s), g (s ))

+c(ζ (s ))uyy ( f (s), g (s)) = −d (ζ (s)). (3.8)

where ζ (s) := ( f (s ), g (s), h(s), p(s), q(s)). The equations (3.7)-(3.8) may
be written as the linear system
    
f 0 (s ) g 0 (s) 0 u x x ( f (s), g (s )) p 0 (s )
    
 0 f 0 (s) g 0 (s)   u xy ( f (s), g (s))  =  q 0 (s)  .
a(ζ (s)) 2b (ζ (s)) c(ζ (s)) uyy ( f (s ), g (s)) − d (ζ (s))
(3.9)
Thus, the equations (3.7)-(3.8) determine all the second order partial
derivatives of u along the curve Γ2 uniquely if
 
 f 0 (s) g 0 (s) 0 
 
 0 0 
∆(s) :=  0 f (s) g (s)  6= 0.
 
 a(ζ (s)) 2b (ζ (s)) c(ζ (s)) 

On the other hand, if ∆(s) = 0 for some s ∈ I , then there exists at least
one of the second order partial derivatives that is either undeterminable
or not uniquely determinable. On expanding the determinant, we get the
following expression for ∆(s)

2
2
∆(s) = c(ζ (s)) f 0 (s) − 2b (ζ (s )) f 0 (s )g 0 (s) + a(ζ (s )) g 0 (s) (3.10)

Computation of third and higher order derivatives

Recall that the computation of second order partial derivatives of the func-
tion u along Γ2 relied on the following three important observations.

1. The PDE (3.1) is a condition on a combination of second order deriva-

tives of u along Γ2 .

Sivaji IIT Bombay

3.1. Cauchy problem for quasilinear PDE 149

2. Knowledge of first order derivatives of u on Γ2 yields another two

relations among second order derivatives of u along Γ2 .
3. Assumption that ∆(s ) 6= 0 for s ∈ I .

In order to determine all the third order derivatives of u along Γ2 , we

would like to follow a similar strategy as for the determination of second
order derivatives. To this end, we need to get some new information from
the knowledge of second order derivatives along Γ2 , and a PDE satisfied by
third order derivatives.
Differentiating the PDE (3.1) w.r.t. x yields the PDE

∂ ∂
a u x x x + 2b u x x y + c u xy y + u x x a(x, y, u, u x , uy ) + 2u xy b (x, y, u, u x , uy )
∂x ∂x
∂ ∂
+uyy c(x, y, u, u x , uy ) + d (x, y, u, u x , uy ) = 0 (3.11)
∂x ∂x
where the argument of the functions a, b , c is (x, y, u, u x , uy ). Note that in
the equation (3.11), the coefficients of the third order derivatives are exactly
the same as in the equation (3.1). The above computation shows that the
coefficients of highest order derivatives continue to be the same as in (3.1)
in all the equations that can be obtained from (3.1) on repeated differenti-
ations. Along Γ2 , these coefficients are known as the first order derivates
are known. Rest of the terms in the equation (3.11) involve derivatives of
the function u up to second order, and thus are known along Γ2 . For, for
any function φ ∈ {a, b , c, d }, we have


∂ 

φ(x, y, u, u x , uy ) = φ x (ζ (s)) + φ z (ζ (s))u x ( f (s), g (s ))
∂x 
(ζ (s))
+ φ p (ζ (s))u x x ( f (s ), g (s )) + φq (ζ (s))u x y ( f (s), g (s)),

where ζ (s) := ( f (s), g (s), h(s), p(s), q(s )). Thus,


∂ 

φ(x, y, u, u x , uy ) is a known function of s.
∂x 
(ζ (s))

Thus, along Γ2

a(ζ (s))u x x x + 2b (ζ (s))u x x y + c(ζ (s ))u xy y is a known function of s. (3.12)

The following system of equations holds for third order derivatives at

August 9, 2021 Sivaji

150 Chapter 3. Classification of second order PDEs

the point ( f (s ), g (s)) ∈ Γ2 :

d
(u ( f (s), g (s))) = u x x x f 0 (s) + u x x y g 0 (s ) (3.13a)
d s xx
d

u xy ( f (s), g (s )) = u x xy f 0 (s) + u xy y g 0 (s ) (3.13b)
ds
d

uy y ( f (s ), g (s )) = u xy y f 0 (s) + uyy y g 0 (s ). (3.13c)
ds
The system of linear equations given by (3.13a), (3.13b), (3.12) may be writ-
ten as

    
f 0 (s) g 0 (s ) 0 u x x x ( f (s), g (s )) ψ1 (s)
    
 0 f 0 (s) g 0 (s)
  u x xy ( f (s), g (s))  =  ψ2 (s)  ,
a(ζ (s)) 2b (ζ (s )) c(ζ (s)) u x yy ( f (s), g (s)) ψ3 (s)
(3.14)
where ψ1 , ψ2 , ψ3 are known functions of s ∈ I . The coefficient matrix in
the linear system (3.14) is the same as in (3.9), and thus has a unique solution
if ∆(s) 6= 0. This however does not determine uy y y along Γ2 for which we
need to use the PDE resulting from differentiating the PDE (3.1) w.r.t. y,
and the equations (3.13b), (3.13c).
The procedure described above can be continued indefinitely, and all
higher order derivatives of u may be determined. There is no need to im-
pose any more assumptions on Γ2 than requiring ∆(s ) 6= 0. Of course, we
require that the functions a, b , c, d , f , g , h are infinitely differentiable.
Let Γ2 be such that ∆(s ) = 0 for every s ∈ I . That is, for each s ∈ I ,

2
2
c(ζ (s )) f 0 (s) − 2b (ζ (s )) f 0 (s)g 0 (s) + a(ζ (s )) g 0 (s) = 0. (3.15)

If Γ2 is defined by the equation y = φ(x) for x ∈ I , i.e., Γ2 the graph of a

function φ : I → R, then the equation (3.15) takes the form

2
c(ζ (x)) − 2b (ζ (x))φ 0 (x) + a(ζ (x)) φ 0 (x) = 0,

where ζ (x) := (x, φ(x), u(x, φ(x)), u x (x, φ(x)), uy (x, φ(x))). On the other
hand, if Γ2 is defined by the equation x = ψ(y) for y ∈ I where ψ : I → R ,
then the equation (3.15) takes the form

2
c(ζ (y)) ψ0 (y) − 2b (ζ (y))ψ0 (y) + a(ζ (y)) = 0,

Sivaji IIT Bombay

3.2. Characteristics as carriers of discontinuities 151

where ζ (y) := (ψ(y), y, u(ψ(y), y), u x (ψ(y), y), uy (ψ(y), y).

3.2 Characteristics as carriers of discontinuities

In Section 3.1, we came across an equation satisfied by regular paramet-
ric curves which are curves of indeterminacy of second and higher order
derivatives of solutions to Cauchy problems. It also turns out that the
curves of discontinuities in second order derivatives of a ‘piecewise C 2 so-
lution’ to a second order quasilinear PDE also satisfy the same equation.
This is the content of the next result.

Theorem 3.2. Let Ω be an open and connected subset of R2 , and γ be a curve

in R2 that divides Ω into two parts such that Ω \ γ is composed of two disjoint
regions Ω1 and Ω2 . For any function defined piecewise by
 (1)
v (x, y) if (x, y) ∈ Ω1 ,
v(x, y) :=
v (2) (x, y) if (x, y) ∈ Ω2 ,

where v (1) (x, y) ∈ C (Ω1 ) and v (2) (x, y) ∈ C (Ω2 ), denote by [v], the jump in
the values of v across γ defined by

[v](x, y) := v (2) (x, y) − v (1) (x, y) for (x, y) ∈ γ .

Let u, u x , uy be continuous on Ω, and the restrictions of u to Ωi , denoted by

u (i) , belong to C 2 (Ωi ) for i = 1, 2. Let γ be given by γ : x = φ(y), where φ is
a continuously differentiable function. The following statements hold true.

(i) The jumps in the quantities u x x , u xy , uy y across γ are not independent of

one another. Denoting λ(y) := [u x x ](φ(y), y), we have

[u x y ](φ(y), y) = −λ(y)φ 0 (y), (3.16a)

2
[uyy ](φ(y), y) = λ(y) φ 0 (y) (3.16b)

at each point (φ(y), y) on the curve γ .

(ii) Let u (1) and u (2) solve the quasilinear equation (3.1) in the regions Ω1
and Ω2 respectively. Let y be such that λ(y) 6= 0. Then

2
c(ζ (y)) φ 0 (y) − 2b (ζ (y))φ 0 (y) + a(ζ (y)) = 0, (3.17)

where ζ (y) := (φ(y), y, u(φ(y), y), u x (φ(y), y), uy (φ(y), y). If λ(y) 6= 0
for every y, then the function y 7→ φ(y) is a solution to the ODE (3.17).

August 9, 2021 Sivaji

152 Chapter 3. Classification of second order PDEs

Ω1 Ω2

γ:
x=
φ (y
)
(iii) Let u (1) and u (2) solve the linear equation (3.2) in the regions Ω1 and Ω2
respectively. Let u possess third order derivatives and have well-defined
jumps across γ . Let y be such that λ(y) 6= 0. Then the following equation
holds.
€
2 Š
0 = 2(b − cφ 0 )λ0 + a x − 2b x φ 0 + c x φ 0 + d − eφ 0 − cφ 00 λ. (3.18)

In addition, if λ(y) = 0 holds at isolated points, then λ is a solution to

the ODE (3.18).

Proof. Proof of (i): Since u, u x , uy are assumed to be continuous on Ω,

these functions are continuous at points of γ also. As a consequence, we
have along the curve γ

[u] := u (2) (φ(y), y) − u (1) (φ(y), y) = 0 (3.19a)

[u x ] := u x(2) (φ(y), y) − u x(1) (φ(y), y) = 0 (3.19b)
[uy ] := uy(2) (φ(y), y) − uy(1) (φ(y), y) = 0 (3.19c)

On differentiating the equations (3.19b) and (3.19c) w.r.t. y, we get

u x(2)x (φ(y), y)φ 0 (y) − u x(1)x (φ(y), y)φ 0 (y) + u xy

(2) (1)
(φ(y), y) − u xy (φ(y), y) = 0
(2)
u xy (φ(y), y)φ 0 (y) − u xy
(1)
(φ(y), y)φ 0 (y) + uy(2)y (φ(y), y) − uyy
(1)
(φ(y), y) = 0

which can be written as (in terms of jumps)

Sivaji IIT Bombay

3.2. Characteristics as carriers of discontinuities 153

[u x x ](φ(y), y)φ 0 (y) + [u x y ](φ(y), y) = 0, (3.20a)

[u xy ](φ(y), y)φ 0 (y) + [uyy ](φ(y), y) = 0. (3.20b)

From the equations (3.20), the relations (3.16) follow.

Proof of (ii):
We now assume that u (1) and u (2) are solutions of the quasilinear second
order PDE

au x x + 2b u xy + c uyy + d (x, y, u, u x , uy ) = 0

in the regions Ω1 and Ω2 respectively. That is,

0 = au x(i)x + 2b u xy
(i )
+ c uy(i)y + d i = 1, 2, (3.21)

where the functions a, b , c, d are evaluated at the 5-tuple

(i) (i)
(x, y, u (i) (x, y), u x (x, y), uy (x, y)), (x, y) ∈ Ωi . Note that the two equa-
tions in (3.21) are meaningful for (x, y) ∈ γ in view of the assumptions on
the function u. For (x, y) ∈ γ , on subtracting one of the two equations in
(3.21) from the other, we get

0 = a[u x x ] + 2b [u x y ] + c[uy y ] along γ .

Denote ζ (y) := (φ(y), y, u(φ(y), y), u x (φ(y), y), uy (φ(y), y). Using the re-
lations (3.16) in the last equation, we get

2
0 = a(ζ (y))λ(y) − 2b (ζ (y))λ(y)φ 0 (y) + c(ζ (y))λ(y) φ 0 (y)
‚  2 Œ
dx dx
= λ(y) a(ζ (y)) − 2b (ζ (y)) + c(ζ (y))
dy dy

Since λ(y) 6= 0, we get

 2
dx dx
a(ζ (y)) − 2b (ζ (y)) + c(ζ (y)) = 0.
dy dy

If λ(y) 6= 0 for every y, then the above equation holds for every y. Note
that the above equation is the (x, y)-coordinate version of ∆(s) = 0 for
s ∈ I.

August 9, 2021 Sivaji

154 Chapter 3. Classification of second order PDEs

Proof of (iii):
Since λ(y) = [u x x ](φ(y), y), we have

λ(y) = u x(2)x (φ(y), y) − u x(1)x (φ(y), y) (3.22)

Differentiating on both sides of the equation (3.22) w.r.t. y gives

dλ
(y) = u x(2)x x (φ(y), y)φ 0 (y)−u x(1)x x (φ(y), y)φ 0 (y)+u x(2)x y (φ(y), y)−u x(1)x y (φ(y), y).
dy

That is,
dλ
(y) = [u x x x ](φ(y), y)φ 0 (y) + [u x x y ](φ(y), y). (3.23)
dy
Similarly, we get

d [u xy ]
(y) = [u x x y ](φ(y), y)φ 0 (y) + [u xy y ](φ(y), y) (3.24)
dy

In view of the relation [u xy ] = −λφ 0 , the equation (3.24) becomes

d

−λφ 0 = [u x x y ](φ(y), y)φ 0 (y) + [u xy y ](φ(y), y). (3.25)
dy

If u (i) (i = 1, 2) solves the linear equation in the domain Ωi , then we also

have
(i)
au x(i)x + 2b u xy + c uy(i)y + d u x(i ) + e uy(i ) + f u (i ) + g = 0, i = 1, 2. (3.26)

First differentiate each of the two equations in (3.26) w.r.t. x, and then
subtract one from another to get

a[u x x x ] + 2b [u x x y ] + c[u x y y ] + d [u x x ] + e[u x y ] + a x [u x x ]

+2b x [u x y ] + c x [uyy ] = 0

Using the relations (3.16) in the last equation, we get

2
a[u x x x ] + 2b [u x xy ] + c[u x y y ] + d λ − eλφ 0 + a x λ − 2b x λφ 0 + c x λ φ 0 = 0.
(3.27)
Eliminating the jumps in the third order derivatives of u across γ from
the equations (3.23),(3.25), (3.27), and using the relation a−2b φ 0 +c (φ 0 )2 =
0, we get the following first order ODE satisfied by λ:

Sivaji IIT Bombay

3.3. Classification of quasilinear PDEs 155

€
2 Š
0 = 2(b − cφ 0 )λ0 + a x − 2b x φ 0 + c x φ 0 + d − eφ 0 − cφ 00 λ

at every y such that λ(y) 6= 0. Since φ is a continuously differentiable

function, the above equation holds at every point y whenever the equation
λ(y) = 0 holds at isolated points.

Remark 3.3.

(i) Intensity of jumps : We may interpret λ(y) := [u x x ](φ(y), y) as the

intensity of jump in u x x across γ [31]. If λ(y0 ) = 0 for some y0 , and
if the initial value problem for the ODE (3.18) has a unique solution
(which is the case if b − cφ 0 6= 0 along γ ), then λ(y) ≡ 0. That is,
if u x x is continuous at some point of γ , then u x x is continuous at all
points of γ , and as a consequence of the relations (3.16), all second
order derivatives of u are continuous across γ .
(ii) Location and speed of jumps : If the variable y is given the inter-
pretation of time, then the equation x = φ(y) gives the location of
the jump in u x x for varioius times. The speed of propagation of dis-
dx
continuities is given by φ 0 (y), which is equal to d y , and satsifies the
differential equation
 2
dφ dφ
a(ζ (y)) − 2b (ζ (y)) + c(ζ (y)) = 0,
dy dy

where ζ (y) := (φ(y), y, u(φ(y), y), u x (φ(y), y), uy (φ(y), y)).

3.3 Classification of quasilinear PDEs

The discussion of the Section 3.1, we are interested in knowing if there
are regular curves along which ∆(s ) = 0, and their description in (x, y)-
coordinates.
Recall from Section 3.1 that if ∆ 6= 0 for all s ∈ I , all the partial deriva-
tives of a solution (if exist) may be determined along Γ2 , from the Cauchy
data. However if ∆ = 0 for all s ∈ I , then we cannot implement the pro-
cedure (described in Section 3.1) to solve the Cauchy problem. This moti-
vates the following definition.

Definition 3.4 (characteristic curve).

August 9, 2021 Sivaji

156 Chapter 3. Classification of second order PDEs

This necessary and sufficient condition is violated at every point of the

curve Γ2 if ∆(s) = 0 for every s ∈ I . The equations (3.15),(3.16),(3.17) were
obtained, depending on the description of the curve Γ2 .
In Section 3.2, it was observed that planar curves satisfying the equation
(3.16) appear as curves of discontinuities for piecewise smooth solutions of
the PDE (3.1).
In view of the above observations, the curves which satisfy the equation
(3.15) (equivalently, (3.16) or (3.17)) are important to the study of the PDE
(3.1). This motivates the following definition.

Definition 3.4 (characteristic curve). Let Γ2 denote a regular planar curve

described parametrically by

Γ2 : x = f (s ), y = g (s), s ∈ I ,

where I is an interval in R, f , g ∈ C 1 (I ), ( f 0 (s ), g 0 (s )) 6= (0, 0) for every

s ∈ I.

(i) Γ2 is said to be a characteristic curve for the PDE (3.1) w.r.t. a given
solution u of (3.1) if

∆(s ) = 0 for all s ∈ I .

(ii) Γ2 is said to be a non-characteristic curve for the PDE (3.1) w.r.t. a

given solution u of (3.1) if

∆(s ) 6= 0 for all s ∈ I .

Remark 3.5 (characteristic curves for semilinear equations).

When the PDE (3.1) is semilinear, ∆(s) = 0 reduces to

2
2
a( f (s), g (s)) g 0 (s) −2b ( f (s), g (s)) f 0 (s)g 0 (s)+ c( f (s ), g (s)) f 0 (s ) = 0,

since the coefficients a, b , c in (3.1) depend only on x, y.

If Γ2 is given by y = ψ(x), then the above equation becomes
 ‹2
dy dy
a(x, y) − 2b (x, y) + c(x, y) = 0. (3.30)
dx dx
dy
We may solve for dx
from the equation (3.30) and obtain

Sivaji IIT Bombay

3.3. Characteristic curves associated to quasilinear PDEs 157

p p
d y 2b ± 4b 2 − 4ac b ± b 2 − ac
= =
dx 2a a
The solutions to the equation (3.30) are characteristic curves, and hence
the equation (3.30) is called the characteristic (differential) equation for the
semilinear equation.

(a) If b 2 − ac > 0, two families of real characteristics exist which are

transversal to each other.
(b) If b 2 − ac = 0, there is only one family of real characteristic curves.
(c) If b 2 − ac < 0, there are no real characteristic curves.

Coming back to the case of the quasilinear PDE (3.1), if the curve Γ2 is
given by y = ψ(x), then the equation ∆(s) = 0 reduces to
 ‹2
dy dy
a(x, y, u(x, y)) − 2b (x, y, u(x, y)) + c(x, y, u(x, y)) = 0, (3.31)
dx dx
where u is a given solution to the quasilinear PDE (3.1).
dy
We may solve for d x from the equation (3.31), and obtain
p p
d y 2b ± 4b 2 − 4ac b ± b 2 − ac
= =
dx 2a a
In general the differential equation (3.31) is called the characteristic equa-
tion for the quasilinear equation (3.1) for a given integral surface. The so-
lutions of the characteristic equation are characteristic curves.
If b 2 − ac > 0, the equation (3.31) gives rise to two families of real char-
acteristics which are transversal to each other. If b 2 − ac = 0, there is only
one family of real characteristic curves. If b 2 − ac < 0, there are no real
characteristic curves.

Remark 3.6. In Subsection 2.2.1, three examples of Cauchy problems were

presented each of which exhibited different behaviours w.r.t. the number
of solutions. Reason for such behaviour was explained in Remark 2.50,
wherein we noted that number of solutions that a Cauchy problem ad-
mits, depends on whether the datum curve is a characteristic curve. The
present remark is in a similar spirit.
Consider the following two Cauchy problems for the same PDE u x x =
0, whose general solution is given by

u(x, y) = K1 (y)x + K2 (y), (3.32)

August 17, 2021 Sivaji

158 Chapter 3. Classification of second order PDEs

where K1 , K2 are twice continuously differentiable functions. It is easy to

observe that the functions K1 , K2 satisfy

u(0, y) = K2 (y), u x (0, y) = K1 (y). (3.33)

(i) Cauchy problem 1: In view of (3.33), we have freedom to prescribe

the Cauchy data along y-axis, namely the functions u(0, y), u x (0, y),
arbitrarily. This Cauchy problem would have a unique solution.
(ii) Cauchy problem 2: If we would like to prescribe the Cauchy data
along x-axis, namely u(x, 0), uy (x, 0), then such a Cauchy problem
has either infinitely many solutions or no solutions. Indeed, the form
of the general solution (3.32) suggests that the function u(x, 0) must be
an affine function which is of the form a x+b for some a, b ∈ R. Once
again, the form of the general solution (3.32) suggests that the function
uy (x, 0) must be an affine function which is of the form c x + d for
some c, d ∈ R. Note that the Cauchy problem with u(x, 0) = ax + b
and uy (x, 0) = c d + d has infinitely many solutions. For, the Cauchy
data determines the values of the functions K1 , K2 along with their
first derivatives at the point y = 0, which are given by

K1 (0) = a, K2 (0) = b , K10 (0) = c, K20 (0) = d .

Since there are an infinite number of functions K1 , K2 satisfying the

above conditions, the Cauchy problem has infinitely many solutions.
On the other hand if u(x, 0) is prescribed as a non-affine function, and
uy (x, 0) arbitrarily, the Cauchy problem does not admit a solution.

Note that x-axis is a charactertistic curve associated to the PDE u x x = 0,

which follows from the characteristic equation (3.30), while y-axis is a non-
characteristic curve. Thus it is not surprising that the Cauchy problem 1
has a unique solution, and the Cauchy problem 2 has either infinitely many
solutions or no solution.

Sivaji IIT Bombay

3.4. Classification of quasilinear PDEs 159

3.4 Classification of quasilinear PDEs

Motivated by the discussion in Section 3.3, we introduce a classification
scheme for second order quasilinear PDEs.

Definition 3.7. Let u be a solution of the quasilinear PDE

a(x, y, u, u x , uy )u x x + 2b (x, y, u, u x , uy )u xy + c(x, y, u, u x , uy )uyy

+d (x, y, u, u x , uy ) = 0.

Define a function δ by

δ(x, y) = (b 2 − ac)(x, y, u(x, y), u x (x, y), uy (x, y))

With respect to the integral surface z = u(x, y), we say that the given quasi-
linear equation is
(i) of hyperbolic type at the point (x, y) if δ(x, y) is positve.
(ii) of parabolic type at the point (x, y) if δ(x, y) is zero.
(iii) of elliptic type at the point (x, y) if δ(x, y) is negative.

Remark 3.8. (i) The above classification is symmetric in a and c, which

indicates that there is no preferred variable between x and y.
(ii) The type of a quasilinear PDE of second order depends on the terms
involving second order derivatives only.
(iii) In view of the law of trichotomy for real numbers, at every point (x, y)
the quasilinear PDE (with a given integral surface) must be of one of
the three types.
(iv) The type of a quasilinear PDE might vary from point to point, de-
pending on the coefficients.
(v) If a, b , c are constant functions, then the equation is of the same type
at every point (x, y).
(vi) If the quasilinear equation is semilinear, then the type of the equation
depends only on the point (x, y) in the plane. For such equations,
the characteristic curves are called characteristic curves for the PDE
without any reference to a given integral surface.
Example 3.9. (a) The wave equation u t t − u x x = 0 is of hyperbolic type
at every point (x, t ), since δ(x, t ) = 1.
(b) The heat equation u t − u x x = 0 is of parabolic type at every point
(x, t ), since δ(x, t ) = 0.

August 17, 2021 Sivaji

160 Chapter 3. Classification of second order PDEs

(c) The Laplace equation u x x + uyy = 0 is of elliptic type at every point

(x, y), since δ(x, y) = −1.

Example 3.10 (Tricomi equation). The linear PDE

uy y − y u x x = 0

is known as Tricomi equation. For this equation, note that a = −y, b =

0, c = 1. Thus b 2 −ac = y, and hence the Tricomi equation is of hyperbolic
type in the upper half plane, is of parabolic type on x-axis, and is of elliptic
type in the lower half plane.

3.4.1 On the necessity of classification

The standard answer to the question “What are reasons for classifying PDEs?"
is that the types of problems that are well-posed for a given PDE depends
on the type of the PDE. In the present context of second order PDEs, ini-
tial or initial-boundary value problems are well-posed for hyperbolic and
parabolic equations and are generally ill-posed for elliptic equations. On
the other hand, boundary value problems are well-posed for elliptic equa-
tions while they are ill-posed for hyperbolic equations. In the chapters on
wave equation, Laplace equation, and heat equation we will discuss them
in detail.
Recall that Hadamard’s definition of well-posedness has three require-
ments: Existence, Uniqueness, Continuous dependence on the data. Of
these three requirements, only existence and uniqueness are mathematical
requirements. The third requirement is added citing the inaccuracies of
measurements, modelling etc. and as such is not a mathematical require-
ment.
Hellwig [28] asks if the classification necessitated by the way the math-
ematical problem is posed, or is it because we have “different methods for
different types of equations" which could be due to our inability to find a
common method to deal with all the equations?
Since the questions posed above are purely mathematical, it would be
ideal to have answers based on the first two requirements, namely existence
and uniqueness, of Hadamard. Hellwig cites an example of Perron [?]

Theorem 3.11. For a ∈ R, a necessary and sufficient condition for the exis-

Sivaji IIT Bombay

3.5. Canonical forms for linear PDEs 161

tence of a solution u, v ∈ C 1 to the problem

u x − uy − vy = 0, auy − v x + vy + f (x + y) = 0,

u(0, y) = 0, v(0, y) = 0

(i) is f ∈ C 0 if a > 0.
(ii) is f ∈ C 2 if a = 0.
(iii) is f is analytic if a < 0.

Moreover, when the necessary and sufficient conditions are met, the corre-
sponding solutions are unique.

There is a classification scheme for first order linear systems, according

to which the system is hyperbolic if a > 0 , parabolic if a = 0 , elliptic if
a < 0. This example once again illustrates that the idea of Classification is
for mathematical reasons.

3.5 Canonical forms for linear PDEs

In this section we will obtain canonical forms for the second order linear
PDE given by

a(x, y)u x x + 2b (x, y)u xy + c(x, y)uyy + d (x, y)u x

+e(x, y)uy + f (x, y)u + g (x, y) = 0. (3.34)

For the purposes of deriving canonical forms, we assume that a, b , c are

continuous functions.
Canonical forms for hyperbolic, parabolic, and elliptic types of equa-
tions are modeled after wave, heat, and laplace equations respectively. In
other words, a change of coordinates from (x, y) to (ξ , η) will be intro-
duced so that when the equation (3.34) is written in the new coordinate
system, terms with second order derivatives look like those in wave, heat,
or laplace equations depending on the type of the equation in the region.
We have the following result which asserts that the type of the equa-
tion in a region is independent of the coordinate system w.r.t. which the
equaion is written.

Lemma 3.12. Suppose that we have a change of coordinates from (x, y) to

(ξ , η), and vice versa, given by

August 17, 2021 Sivaji

162 Chapter 3. Classification of second order PDEs

ξ = φ(x, y), η = ψ(x, y); (3.35a)

x = Φ(ξ , η), y = Ψ(ξ , η). (3.35b)

Then in the (ξ , η) coordinates, the equation (3.34) transforms to an equation

of the form

Awξ ξ + 2B wξ η + C wηη + D wξ + E wη + F w + G = 0, (3.36)

where the functions A, B, C are given by


€ Š 
A(ξ , η) := aφ x2 + 2b φ x φy + cφy2  (3.37a)

(x,y)=(Φ(ξ ,η), Ψ(ξ ,η))




B(ξ , η) := aφ x ψ x + b φ x ψy + φy ψ x + cφy ψy  (3.37b)

(x,y)=(Φ(ξ ,η), Ψ(ξ ,η))

€ Š 
C (ξ , η) := aψ2x + 2b ψ x ψy + cψ2y  (3.37c)

(x,y)=(Φ(ξ ,η), Ψ(ξ ,η))

The type of the equation does not change under a change of coordinates. In
fact, the following equality holds:

€
2
Š 
B 2 − AC = φ x ψy − φy ψ x (b 2 − ac)  . (3.38)

(x,y)=(Φ(ξ ,η), Ψ(ξ ,η))

Proof. Under the change of coordinates (3.35), a function u(x, y) gets trans-
formed to a function w(ξ , η) and vice versa by

w(ξ , η) = u (Φ(ξ , η), Ψ(ξ , η), ) , (3.39a)

u(x, y) = w (φ(x, y), ψ(x, y)) . (3.39b)

In order to find an equation in (ξ , η) variables, we will compute first and

second order partial derivatives of u in terms of new coordinates, and then
substitute in the given equation (3.34). This computation is done by choos-
ing an appropriate identity from (3.39). Applying one of the most funda-
mental and important results in differential calculus, namely, chain rule,

Sivaji IIT Bombay

3.5. Canonical forms for linear PDEs 163

to
u(x, y) = w (φ(x, y), ψ(x, y))
we get

u x (x, y) = wξ (φ(x, y), ψ(x, y)) φ x (x, y) + wη (φ(x, y), ψ(x, y)) ψ x (x, y),
uy (x, y) = wξ (φ(x, y), ψ(x, y)) φy (x, y) + wη (φ(x, y), ψ(x, y)) ψy (x, y).

Differentiating the above set of equations once more, we get

u x x (x, y) = wξ ξ (φ, ψ) φ x2 + 2wξ η (φ, ψ) φ x ψ x + wηη (φ, ψ) ψ2x

+wξ (φ, ψ) φ x x + wη (φ, ψ) ψ x x ,

u x y (x, y) = wξ ξ (φ, ψ) φ x φy + wξ η (φ, ψ) φ x ψy

+wξ η (φ, ψ) φy ψ x + wηη (φ, ψ) ψ x ψy
+wξ (φ, ψ) φ xy + wη (φ, ψ) ψ xy ,

uyy (x, y) = wξ ξ (φ, ψ) φy2 + 2wξ η (φ, ψ) φy ψy + wηη (φ, ψ) ψ2y

+wξ (φ, ψ) φyy + wη (φ, ψ) ψyy ,

where the argument of all the functions, namely (x, y), is omitted for brevity.
Substituting the expressions for the first and second order derivatives of
u obtained above into the equation (3.34), we get an equation for w of the
form (3.36).
Note that we did not compute D, E, F , G explicitly as they do not play
any role in determining type of the equation. It is left to the reader to check
that (3.36) and (3.38) hold.
Note that the factor

φ x (x, y)ψy (x, y) − φy (x, y)ψ x (x, y)

appearing in the equation (3.38) is the determinant of the Jacobian matrix

corresponding to the change of coordinates, and hence is non-zero. Thus
the type of the equation does not change under a change of coordinates
follows from the equation (3.38).

Remark 3.13 (Utility of canonical forms). We shall see in later sections

that equations of hyperbolic, parabolic, and elliptic types can be locally re-
duced to a canonical form. Since such results are of local nature, a question

August 17, 2021 Sivaji

164 Chapter 3. Classification of second order PDEs

of their utility arises. We believe that reduction into a canonical form will
be of some use if the change of coordinates are global. It is not clear if such
a thing can be achieved while dealing with semilinear equations with vari-
able coefficients. Even if it is achieved, we do not have methods to solve
such equations explicitly. Reader is advised to reflect on this later on.

3.5.1 Hyperbolic equations

Theorem 3.14. Let the equation (3.34) be hyperbolic in a region Ω of the

xy-plane. Let (x0 , y0 ) ∈ Ω. Then there exists a change of coordinates (x, y) 7→
(ξ , η) in an open set containing the point (x0 , y0 ) such that the equation (3.34)
is transformed in the (ξ , η) variables into

wξ η + D(ξ , η)wξ + E(ξ , η)wη + F (ξ , η)w + G(ξ , η) = 0. (3.40)

Proof. If a(x0 , y0 ) = c(x0 , y0 ) = 0, then we introduce a change of coordi-

nates (x, y) 7→ (x̃, ỹ) where x̃ = X (x, y) = x + y, ỹ = Y (x, y) = x − y.
Under this change of coordinates the given equation (3.34) takes a form
where the coefficients of u x̃ x̃ and uỹ ỹ are non-zero at the point x̃ = x̃0 =
x0 + y0 , ỹ = ỹ0 = x0 − y0 . Thus we may assume that at least one of the
two quantities a(x0 , y0 ) and c(x0 , y0 ) is non-zero. Without loss of general-
ity, assume that a(x0 , y0 ) 6= 0. As a consequence, there exists an open set U
containing (x0 , y0 ) such that a(x, y) 6= 0 for all (x, y) ∈ U .
We know that under a change of coordinates the equation (3.34) trans-
forms to (3.36). Thus for proving the theorem, it is sufficient to find a
system of coordinates (ξ , η) so that A(ξ , η) = C (ξ , η) = 0 where A, C are
given by (3.37). Thus we need to find φ, ψ satisfying the equations

aφ x2 + 2b φ x φy + cφy2 = 0, (3.41a)

aψ2x + 2b ψ x ψy + cψ2y = 0. (3.41b)

Note that the equations for both φ and ψ are the same. Thus we need
to solve for φ, ψ using only one equation. Note that the equation (3.41a)
may be factorized as

1 p  p 
aφ x + (b − b − ac)φy aφ x + (b + b − ac)φy = 0.
2 2 (3.42)
a
Sivaji IIT Bombay
3.5. Canonical forms for linear PDEs 165

A function φ satisfies the equation (3.42) whenever it satisfies either of the

first order linear PDEs given by

p
aφ x + (b − b 2 − ac)φy = 0, (3.43a)
p
aφ x + (b + b 2 − ac)φy = 0. (3.43b)

We choose φ to be a solution of (3.43a), and ψ to be a solution of (3.43b)

as (ξ , η) = (φ(x, y), ψ(x, y)) is required to define a coordinate change trans-
formation. Both the equations (3.43) may be solved by the method of char-
acteristics. The characteristics corresponding to (3.43a) are governed by the
characteristic ODEs given by

dx dy p dz
= a, = b − b 2 − ac, = 0.
dt dt dt
Thus φ is constant along each of the base characteristic curves (x(t ), y(t )).
On eliminating the parameter t , we conclude that φ(x, y) = k represents a
one parameter family of solutions to the ODE
p
dy b (x, y) − b 2 (x, y) − a(x, y)c(x, y)
= ,
dx a(x, y)

which are the base characteristic curves. On differentiating the equation

φ(x, y(x)) = k

w.r.t. x, we get

dy
φ x (x, y(x)) + φy (x, y(x)) (x) = 0.
dx
From the last equation, we get
p
φ dy b (x, y) − b 2 (x, y) − a(x, y)c(x, y)
− x (x, y) = = . (3.44)
φy dx a(x, y)

Similarly the function ψ will be constant along the base characteristic curves
determined by the ODE corresponding to (3.43b)
p
dy b (x, y) + b 2 (x, y) − a(x, y)c(x, y)
= ,
dx a(x, y)

August 17, 2021 Sivaji

166 Chapter 3. Classification of second order PDEs

and the following relation holds:

p
ψ dy b (x, y) + b 2 (x, y) − a(x, y)c(x, y)
− x (x, y) = = . (3.45)
ψy dx a(x, y)

Since b 2 − ac > 0, we get

 
 φ x (x, y) φy (x, y) 
 
  6= 0
 ψ (x, y) ψ (x, y) 
x y

in view of the relations (3.44) and (3.45). Thus (ξ , η) = (φ(x, y), ψ(x, y))
defines a coordinate transformation near the point (x0 , y0 ), by the inverse
function theorem.

Remark 3.15. We may make a further change of variables x 0 = ξ + η,

y 0 = ξ − η. The equation (3.40) then transforms into

vy 0 y 0 − v x 0 x 0 + 2D 0 (x 0 , y 0 )v x 0 + 2E 0 (x 0 , y 0 )vy 0 + F 0 (x 0 , y 0 )v + G 0 (x 0 , y 0 ) = 0,
(3.46)
on verifying that

A0 (x 0 , y 0 ) = 1, B 0 (x 0 , y 0 ) = 0, C 0 (x 0 , y 0 ) = −1.

Both the equations (3.40) and (3.46) are known as canonical forms of an
equation that is of hyperbolic type in a region Ω.

Example 3.16. Consider the following linear PDE

x 2 u x x − 2xy u xy − 3y 2 uy y + uy = 0. (3.47)

For this equation a = x 2 , b = −xy, c = −3y 2 . Thus b 2 − ac = 4x 2 y 2 .

Thus the equation (3.47) is of hyperbolic type at every point (x, y) such
that x y 6= 0, which means that the equation is of hyperbolic type at all
points of xy-plane except for the coordinate axes x = 0 and y = 0. At
points on the coordinate axes, the equation is of parabolic type.
Let us transform the equation (3.47) into its canonical form in the first
quadrant. In order to find the new coordinate system (ξ , η), we need to

Sivaji IIT Bombay

3.5. Canonical forms for linear PDEs 167

solve the ODEs

p
dy b (x, y) ± b 2 (x, y) − a(x, y)c(x, y) −xy ± 2|xy| −y ± 2y
= = = .
dx a(x, y) x2 x

Thus we need to solve the two ODEs

dy y d y −3y
= , and = ,
dx x dx x
whose solutions are given by x −1 y = constant, and x 3 y = constant respec-
tively. We introduce the following change of coordinates

ξ = φ(x, y) = x −1 y, and η = ψ(x, y) = x 3 y.

On differentiating the equation

u(x, y) = w(φ(x, y), ψ(x, y)) = w(x −1 y, x 3 y)

w.r.t. x and y we obtain

u x = −x −2 ywξ + 3x 2 ywη ,
u x x = x −4 y 2 wξ ξ − 6y 2 wξ η + 9x 4 y 2 wηη + 2x −3 wξ + 6xywη ,
u xy = −x −3 ywξ ξ + 2xywξ η + 3x 5 ywηη − x −2 wξ + 3x 2 wη ,
uy = x −1 wξ + x 3 wη ,
uy y = x −2 wξ ξ + 2x 2 wξ η + x 6 wηη .

On substituting these values in the equation (3.47), we get

−16x 2 y 2 wξ η + 5x −1 wξ + x 3 wη = 0. (3.48)

The last equation can be written in the variables ξ , η completely, on ex-

pressing x and y as functions of ξ , η. Indeed we have
s Æ
η
x = Φ(ξ , η) = 4 , y = Ψ(ξ , η) = ξ 3 η.
4

Thus the equation (3.48) becomes

v v
uξ u ‹3
t t η
−16ξ ηwξ η + 5 wξ + wη = 0.
4 4
(3.49)
η ξ

August 17, 2021 Sivaji

168 Chapter 3. Classification of second order PDEs

On simplification, the last equation (3.49) takes the form

5 1 1 1
wξ η − p wξ − p w = 0,
16 ξ 3 η5
4
16 4 ξ 7 η η

which is a canonical form of the given PDE.

3.5.2 Parabolic equations

Theorem 3.17. Let the equation (3.34) be parabolic in a region Ω of the xy-
plane. Let (x0 , y0 ) ∈ Ω. Then there exists a change of coordinates (x, y) 7→
(ξ , η) in an open set containing the point (x0 , y0 ) such that the equation (3.34)
is transformed in the (ξ , η) variables into

wηη + D(ξ , η)wξ + E(ξ , η)wη + F (ξ , η)w + G(ξ , η) = 0. (3.50)

Proof. Recall that we are assuming that at least one of the three functions
a, b , c is non-zero at every point of Ω. As a consequence, at least one of
a(x0 , y0 ) and c(x0 , y0 ) is not zero. For, if both a(x0 , y0 ) = 0 and c(x0 , y0 ) =
0 hold, then it follows that b (x0 , y0 ) = 0 since the equation (3.34) is of
parabolic type in Ω.
For the purposes of proving this theorem, we need a(x0 , y0 ) 6= 0 and
c(x0 , y0 ) 6= 0, and there is no loss of generality in assuming this. For, this
can be arranged by a change of variables given by x̃ = X (x, y) = x, ỹ =
Y (x, y) = x + y (if a(x0 , y0 ) 6= 0 and c(x0 , y0 ) = 0; and similarly for the
other case) so that when the equation (3.34) is transformed in terms of the
new coordinates, the coefficients of u x̃ x̃ and uỹ ỹ at the point x̃ = x̃0 = x0 ,
ỹ = ỹ0 = x0 + y0 are non-zero.
By continuity of the functions a, b , it follows that a(x, y) 6= 0 and b (x, y) 6=
0 in some open set U containing the point (x0 , y0 ). As a consequence of
parabolicity of the equation (3.34), we conclude that b (x, y) 6= 0 in U .
We know that under a change of coordinates the equation (3.34) trans-
forms to (3.36). Thus for proving the theorem, it is sufficient to find a
system of coordinates (ξ , η) so that A(ξ , η) = B(ξ , η) = 0 where A, B are
given by (3.37). Thus we need to find φ, ψ satisfying

Sivaji IIT Bombay

3.5. Canonical forms for linear PDEs 169

aφ x2 + 2b φ x φy + cφy2 = 0, (3.51a)


aφ x ψ x + b φ x ψy + φy ψ x + cφy ψy = 0. (3.51b)

Since the equation (3.51a) involves only φ, we may solve for φ as in the
proof of Theorem 3.14 to get that φ is constant along the base characteristic
curves of the equation (3.43a). Indeed the equation (3.43a) reduces in the
case of parabolic equations to aφ x + b φy = 0, and the corresponding base
characteristic curves are given by the ODE

dy b (x, y)
= .
dx a(x, y)

Thus whatever may be the choice of ψ, we have A(ξ , η) ≡ 0 and conse-

quently B(ξ , η) ≡ 0 due to the invariance of type of the equation under
change of coordinates. One can also check that B(ξ , η) ≡ 0 holds. Note
that the equation (3.51b) reduces to (b φ x +cφy )ψy = 0. Since φ was chosen
as a solution to aφ x + b φy = 0, we have
 
b2
(b φ x + cφy )ψy = − + c φy ψy = 0.
a

Thus we need to choose ψ so that (x, y) 7→ (φ(x, y), ψ(x, y)) gives rise to
a nonsingular transformation. Applying the inverse function theorem, we
conclude that (ξ , η) = (φ(x, y), ψ(x, y)) defines a new coordinate system
near (x0 , y0 ). To achieve this we need φ, ψ to satisfy
 
 φ x (x, y) φy (x, y) 
 
  6= 0.
 ψ (x, y) ψ (x, y) 
x y

φ ψ
In other words, the quantities φx (x, y) and ψx (x, y) are not equal. Note that
y y

there are infinitely many such choices for ψ.

Since we know that
φx b (x, y)
(x, y) = − ,
φy a(x, y)

August 17, 2021 Sivaji

170 Chapter 3. Classification of second order PDEs

we choose ψ satisfying

ψx a(x, y)
(x, y) = ,
ψy b (x, y)

and this means that the curves corresponding to φ = constant and ψ =

constant are orthogonal families of curves. With these choices of φ, ψ,
the inverse function theorem guarantees that (ξ , η) defines a coordinate
transformation near the point (x0 , y0 ). The equation (3.50) is known as the
canonical form of a parabolic equation.

Example 3.18. Consider the following linear PDE

x 2 u x x − 2xy u x y + y 2 uyy = 0. (3.52)

For this equation a = x 2 , b = −xy, c = y 2 . Thus b 2 − ac = 0, and hence

the equation (3.52) is of parabolic type at every point (x, y) ∈ R2 . However,
note that we need to avoid the point (x, y) = (0, 0) as the coefficients of all
the three second order derivatives of u are zero at the point (x, y) = (0, 0).
Thus we try to determine a canonical form in any domain not containing
the origin.
Let us transform the equation (3.52) into its canonical form in the right
half-plane. This restriction is a technical one as the method followed here
is valid only if we avoid x = 0 or y = 0. However, the canonical form may
be obtained in any of the half-planes defined by x-axis or y-axis.
In order to find the new coordinate system (ξ , η) = (φ(x, y), ψ(x, y)),
we need to solve the ODE
p
dy b (x, y) ± b 2 (x, y) − a(x, y)c(x, y) y
= =− .
dx a(x, y) x

to find φ, and then we need to choose ψ so that (ξ , η) = (φ(x, y), ψ(x, y))
gives rise to a coordinate system. Thus we need to solve the ODE

dy y
=− ,
dx x
whose solution is given by xy = constant. We choose φ(x, y) = xy, and

Sivaji IIT Bombay

3.5. Canonical forms for linear PDEs 171

ψ(x, y) = x so that the Jacobian

   
 φ (x, y) φ (x, y)   y x 
   
J = x y
=  = x 6= 0. (3.53)
 ψ x (x, y) ψy (x, y)   1 0 

Thus we introduce the following change of coordinates

ξ = φ(x, y) = xy, and η = ψ(x, y) = x.

On differentiating the equation

u(x, y) = w(φ(x, y), ψ(x, y)) = w(xy, x)

w.r.t. x and y we obtain

u x = ywξ + wη , uy = xwξ ,
u x x = y 2 wξ ξ + 2y wξ η + wηη ,
u xy = xywξ ξ + xwξ η + wξ ,
uyy = x 2 wξ ξ .

On substituting for the first and second order derivatives of u in the equa-
tion (3.52), we get
x 2 wηη − 2xy wξ = 0. (3.54)
The last equation can be written in the variables ξ , η completely, on ex-
pressing x and y as functions of ξ , η. Indeed we have

ξ
x = Φ(ξ , η) = η, y = Ψ(ξ , η) = .
η

Thus the equation (3.54) becomes

η2 wηη − 2ξ wξ = 0. (3.55)

On simplification, the last equation (3.55) takes the form

ξ
wηη − 2 w = 0, (3.56)
η2 ξ

which is a canonical form of the given PDE.

Remark 3.19 (on Example 3.18). (i) Note that at the point (x, y) = (0, 0)

August 17, 2021 Sivaji

172 Chapter 3. Classification of second order PDEs

the equation (3.52) reduces to 0 = 0 due to vanishing of all the coeffi-

cients of derivatives of the unknown u. In any differential equation,
vanishing of coefficients of highest order derivative(s) at a point may
lead to singularities in the solutions. In the context of ordinary differ-
ential equations, such points are called singular points and bevaiour
of solutions near singular points is studied.
(ii) Note that (x, y) = (0, 0) is a singular point for the equation (3.52).
The change of coordinates employed in the previous example will con-
tinue to be a coordinate change transformation as long as x 6= 0 (see
(3.53)), and hence the canonical form (3.56) is valid in the left half-
plane as well.

3.5.3 Elliptic equations

Theorem 3.20. Let the equation (3.34) be elliptic in a region Ω of the xy-
plane, and the coefficients a, b , c be real-analytic functions in Ω. Let (x0 , y0 ) ∈
Ω. Then there exists a change of coordinates (x, y) 7→ (ξ , η) in an open set
containing the point (x0 , y0 ) such that the equation (3.34) is transformed in
the (ξ , η) variables into

wξ ξ + wηη + D(ξ , η)wξ + E(ξ , η)wη + F (ξ , η)w + G(ξ , η) = 0, (3.57)

Proof. Since (b 2 − ac)(x0 , y0 ) < 0, both a(x0 , y0 ) and c(x0 , y0 ) are non-zero.
By continuity of the function a, there exists an open set U containing
(x0 , y0 ) such that a(x, y) 6= 0 for all (x, y) ∈ U . The desired canonical form
(3.57) requires the change of coordinates transformation (x, y) 7→ (ξ , η) to
satisfy A = C , and B = 0, i.e.,

aφ x2 + 2b φ x φy + cφy2 = aψ2x + 2b ψ x ψy + cψ2y (3.58a)

aφ x ψ x + b φ x ψy + φy ψ x + cφy ψy = 0. (3.58b)

Note that (3.58) is a coupled system of first order nonlinear PDEs for
φ, ψ. Recall that in the hyperbolic case, the equations for φ and ψ were
decoupled. In the parabolic case, the equations for φ and ψ were weakly
coupled i.e., equation for φ did not involve ψ. But for elliptic case, the
system (3.58) is strongly couple. We can overcome this difficulty by using
the assumption of real-analyticity of a, b , c and complex variables. The

Sivaji IIT Bombay

3.5. Canonical forms for linear PDEs 173

system of equations (3.58) may be rewritten as

a(φ x2 − ψ2y ) + 2b (φ x φy − ψ x ψy ) + c(φy2 − ψ2y ) = 0, (3.59a)

i aφ x ψ x + i b φ x ψy + φy ψ x + i cφy ψy = 0, (3.59b)
p
where i = −1. Define a complex valued function Φ = φ + iψ. The
system (3.59) is equivalent to

aΦ2x + 2b Φ x Φy + cΦ2y = 0. (3.60)

Note that the equation (3.60) is exactly the same equation that we solved
while determining the canonical form for hyperbolic equations. However,
factorizing the equation (3.60) leads to PDEs with complex coefficients
given by

p
aΦ x + (b + i ac − b 2 ) Φy = 0, (3.61a)
p
aΦ x + (b − i ac − b 2 ) Φy = 0. (3.61b)

Since the coefficients a, b , c are real-analytic functions, it is known (by

a complex version of Cauchy-Kowalewski theorem) that the system (3.61)
has solutions near the point (x0 , y0 ). If Φ is a solution of (3.61a), then Ψ := Φ
is a solution of (3.61b).
Thus Φ, Ψ are constant on the two complex characteristics given by the
equation(s)
p
dy b (x, y) ± i a(x, y)c(x, y) − b 2 (x, y)
=
dx a(x, y)

corresponding to + and − respectively.

Now define φ and ψ by φ(x, y) = Re Φ(x, y), and ψ(x, y) = Im Φ(x, y)
where Re Φ and Im Φ are real and imaginary parts of the complex valued
function Φ. Define a coordinate transformation by ξ := φ(x, y), η :=
ψ(x, y) in a neighbourhood of the point (x0 , y0 ), which is possible since the
∂ (φ,ψ)
Jacobian ∂ (x,y) (x0 , y0 ) 6= 0. For, if the Jacobian were zero, then it follows
that A = C = 0 and contradicts the ellipticity of the given equation.
Note that ξ and η satisfy the system of equations (3.58), and thus the
equation (3.34) reduces to the required canonical form (3.57).
For a detailed exposition, the reader may consult [21].

August 17, 2021 Sivaji

174 Chapter 3. Classification of second order PDEs

Example 3.21. Consider the following linear PDE

u x x + (1 + y 2 )2 uyy − 2y(1 + y 2 )uy = 0. (3.62)

For this equation a = 1, b = 0, c = (1 + y 2 )2 . Thus b 2 − ac = −(1 + y 2 )2 .

Thus the equation (3.62) is of elliptic type at every point (x, y) ∈ R2 .
Let us transform the equation (3.62) into its canonical form. In order
to find the new coordinate system (ξ , η), we need to solve the ODEs
p
dy b (x, y) ± b 2 (x, y) − a(x, y)c(x, y)
= = ±i (1 + y 2 ).
dx a(x, y)

Thus we need to solve the ODE

dy
= i (1 + y 2 ),
dx
whose solution is given by tan−1 y − i x = constant. We introduce the fol-
lowing change of coordinates

ξ = φ(x, y) = x, and η = ψ(x, y) = tan−1 y.

On differentiating the equation

u(x, y) = w(φ(x, y), ψ(x, y)) = w(x, tan−1 y)

w.r.t. x and y we obtain

1
u x = wξ , u y = w , u = wξ ξ ,
1 + y2 η xx
1 2y
uy y = wηη − w .
(1 + y 2 )2 (1 + y 2 )2 η

On substituting these expressions for the derivatives of u in the equation

(3.62), we get

wξ ξ + wηη − 4(tan η) wη = 0,

which is a canonical form of the given PDE.

Sivaji IIT Bombay

Exercises 185

Exercises
Origin of second order PDEs

3.1. [52] Find PDEs whose general solutions are of the form
(i) u(x, y) = φ(x + y) + ψ(x − 2y)
(ii) u(x, y) = xφ(x + y) + yψ(x + y)
€xŠ
(iii) u(x, y) = φ(xy) + ψ y

Classification of PDEs

3.2. Give an example of a second order linear partial differential equation

such that
(a) the equation is of elliptic type at points of a domain Ω and is of
parabolic type on R2 \ Ω.
(b) the equation is of hyperbolic type at points of a domain Ω and
is of parabolic type on R2 \ Ω.
(c) the equation is of elliptic type at points of a domain Ω and is of
hyperbolic type on R2 \ Ω.
Justify each of your answers.

Canonical forms for PDEs in two variables

3.3. [52] Classify and find canonical forms of the following PDEs
(i) u x x − 2u xy − 3uyy + uy = 0



(ii) 1 + x 2 u x x + 1 + y 2 uyy + x u x + y uy = 0


(iii) e 2x u x x + 2e x+y u xy + e 2y uy y + e 2y − e x+y uy = 0
(iv) u x x − 2 sin x u xy − cos2 x uy y − cos x uy = 0
3.4. Find the canonical forms for the Tricomi equation

uy y − y u x x = 0

in the upper and lower half planes, where Tricomi equation is of hy-
perbolic and elliptic types respectively. (Answer: For the upper half
plane, with ξ = 3x − 2y 3/2 and η = 3x + 2y 3/2 , the canonical form is
1
wξ η − (w + wη ) = 0.
6(η − ξ ) ξ
August 17, 2021 Sivaji
186 Chapter 3. Classification of second order PDEs

3 3
For the lower half plane, with ξ = 2 x and η = (−y) 2 , the canonical
form is
1
wξ ξ + wηη + w = 0.)
3η η

3.5. [32] Find the general solution of the PDE

x 2 u x x − 2xy u xy − 3y 2 uyy = 0

by first reducing it to a canonical form. Then find a particular solu-

tion that satisfies u(x, 1) = φ(x) and uy (x, 1) = ψ(x) where φ, ψ are
smooth functions.
3.6. [32] Solve the following Cauchy problem

u x x + 2 cos x u xy − sin2 x uyy − sin x uy = 0,

u(x, sin x) = φ(x), uy (x, sin x) = ψ(x),

by first reducing the PDE to a canonical form. Here assume that φ, ψ

are smooth functions.

Canonical forms for PDEs in several variables

3.7. Let Γ : φ(x) = 0 be a hypersurface. Find the equation satisfied by φ

so that Γ is a characteristic surface for the wave equation in n space
variables given by
€ Š
u t t − c 2 u x1 x1 + u x2 x2 + · · · + u xn xn = 0.

Find a solution of the equation thus obtained. (Hint: )

3.8. [52] Classify and find canonical forms of the following PDEs
1
(i) 4u x x + uy y + 3 u z z − 2u xy + 2u x z = 0
2
(ii) 3u x x + 4uyy − 3 u z z + 4u xy + 2uy z = 0

Sivaji IIT Bombay

Exercises 187

List of problems
Classify and find canonical forms of the following PDEs

1. u x x − 2u xy − 3uy y + uy = 0



2. 1 + x 2 u x x + 1 + y 2 uy y + x u x + y uy = 0


3. e 2x u x x + 2e x+y u xy + e 2y uy y + e 2y − e x+y uy = 0
4. u x x − 2 sin x u xy − cos2 x uy y − cos x uy = 0
5. 6u x x − 5u x y + uy y = 0
6. u x x + 2u xy + uyy + 3uy + 9u = 0
7. 5u x x − 2u x y + 2uyy + u = 0
8. y 2 u x x − uy y = 0
9. y 2 u x x + uy y = 0
10. y 2 u x x + 2y u xy + uy y = 0
11. 6y u x x + (3 + 2y)u xy + uyy = 0
12. (1 + x 2 )(4 + x 2 )u x x + (5 + 2x 2 )u xy + uy y = 0
13. u x x + 2u xy + uyy = 0 and solve.
14. (cos2 x)u x x + 2(sin x)u xy − uyy = 0 and solve.
15. y 2 u x x − 2xy u xy + x 2 uy y = 0 and solve.
16. u x x + 2u xy + 3uy y + 4u x + 5uy + u = e x
17. 2u x x − 4u x y + 2uyy + 3u = 0
18. u x x + 5u xy + 4uy y + 7uy = sin x
19. u x x + uyy + 2u x + 8uy + u = 0
20. u x x + 2uy y + 9u x + uy = 2
21. 6u x x − u xy + u = y 2
22. u xy + u x + uy = 3x
23. uyy − 9u x + 7uy = cos y
24. x 2 u x x − y 2 uyy − u x = 1 + 2y 2

August 17, 2021 Sivaji

188 Chapter 3. Classification of second order PDEs

1
25. u x x + y uy y + 2 uy + 4y u x = 0
26. x 2 y 2 u x x + 2x y u xy + uy y = 0
27. u x x + y uy y = 0
28. y u x x − x uy y = 0, x > 0, y > 0

29. u x x + (sech4 x)uy y = 0

30. y 2 u x x + x 2 uyy = 0

31. u x x − (sech4 x)uy y = 0

32. u x x + 6u x y + 9uy y + 3y uy = 0
33. y 2 u x x + 2xy u x y + 2x 2 uyy + x u x = 0
34. u x x − 2(cos x)u x y + (1 + cos2 x)uyy + u = 0
35. u x x + 2(cosec y)u x y + (cosec2 y)uyy = 0
36. u x x − 2u x y + uyy + 3u x − u + 1 = 0
37. u x x − y 2 uy y + u x − u + x 2 = 0
38. u x x + y uy y − x uy + y = 0
39. u x x + y uy y + sin(x + y) = 0

Sivaji IIT Bombay