5.

Second-Order linear PDE

Consider a second-order linear PDE

L[u] = auxx + 2buxy + cuyy + dux + euy + f u = g,

(x, y) U

(5.1)

for an unknown function u of two variables x and y. The functions a, b and

c are assumed to be of class C 1 and satisfying a2 + b2 + c2 6= 0. The operator
L0 [u] := auxx + 2buxy + cuyy
consisting of the second order terms of L is called the principal part of L.
Many of the fundamental properties of the solutions of (5.1) are determined
by the sign of the discriminant of L. The discriminant (L)(x, y) is defined
by


b a
(L)(x, y) = det
= b2 ac
c b
where a, b and c are evaluated at the point (x, y).
We are interested in how the PDE is transformed under changes of coordinates. We consider a C 1 - map F (x, y) = ((x, y), (x, y)) whose Jacobian
satisfies


x y
6= 0
detJ(x, y) = det
x y
at each point of (x0 , y0 ) U . The inverse function theorem implies that near
the point (x0 , y0 ) the map F has an inverse F 1 (, ) = (x(, ), y(, )).
The inverse is of class C 1 . Now, assuming that u is a solution of (5.1),
define w(, ) = u(x(, ), y(, )). Then u(x, y) = w((x, y), (x, y)) and,
by the chain rule,
ux = w x + w x
uy = w y + w y
uxx = w (x )2 + 2w x x + w (x )2 + w xx + w xx
uyy = w (y )2 + 2w y y + w (y )2 + w yy + w yy
uxy = uyx = w x y + w (x y + x y ) + w x y + w xy + w xy
Substituting into (5.1), we find that
e
L[w]
= Aw + 2Bw + Cw + Dw + Ew + F w = G
32

(5.2)

e0 [w] = Aw + 2Bw + Cw
with the coefficients of the principal part L
given by
A(, ) = ax2 + 2bx y + cy2
B(, ) = ax x + b(x y + y x ) + cy y
C(, ) =

ax2

+ 2bx y +

(5.3)

cy2

Observe that


 
 
 
t
A B

a b

= x y
x x ,
B C
y y
b c
y y

where t denotes the transpose of the matrix. Recalling that the determinant
of the product of matrices is equal to the product of the determinants of
matrices and that the determinant of a transpose of a matrix is equal to the
determinant of a matrix, we get




a b
A B
(J(x, y))2 .
(5.4)
= det
det
b c
B A
This shows that the discriminant of L has the same sign as the discriminant
of the transformed equation and so it is an invariant of the change of coordinates. Consequently, we can classify equations (5.1) according to the sign
of the discriminant.
Definition 5.1. The equation (5.1) is called
hyperbolic at (x, y) if (L)(x, y) > 0.
parabolic at (x, y) if (L)(x, y) = 0.
elliptic at (x, y) if (L)(x, y) < 0.
Example 5.2.
tion.

(i) The wave equation utt uxx = 0 is a hyperbolic equa-

(ii) The heat equation ut uxx = 0 is a parabolic equation.

(iii) The Laplace equation uxx = uyy = 0 is an elliptic equation.
Example 5.3. Consider the Tricomi equation
yuxx + uyy = 0.

(5.5)

Here a = y, b = 0, c = 1 and d = e = f = g = 0. Its discriminant is equal to





b a
0 y
det
= det
= y.
c b
1 0
33

Hence the equation (5.5) is hyperbolic for y < 0, parabolic when y = 0, and
elliptic for y > 0.
Next we shall show that we can find changes of coordinates in which the
(5.1) takes a simple form.
Hyperbolic equations.
Suppose that equation (5.1) is hyperbolic on the domain U . This means
that b2 ac > 0 at each point of U . We shall show that in this we can
choose (x, y) 7 ((x, y), (x, y)) so that
A(, ) = ax2 + 2bx y + cy2 = 0

(5.6)

C(, ) = ax2 + 2bx y + cy2 = 0.

(5.7)

Under such a change of coordinates and dividing by 2B the hyperbolic equation (5.1) takes its canonical form
e
L[w]
= w + [w] = G,

(5.8)

where is a first-order linear operator and G is function.

Note that if a and c are equal to 0, then the equation (5.1) is already in its
canonical form (just divide by 2b). Hence without loss of generality we may
assume that a 6= 0. Note also that the equation (5.7) for is the same as
(5.6) for . Hence it suffices to consider only one of the equations, say (5.6).
It can be written as a product
# "
#
"

b + b2 ac
b b2 ac
y x
y = 0
a x
a
a
and so, we need to solve the following two linear equations
x 1 y = 0

(5.9)

x 2 y = 0,

(5.10)

and
where we have abbreviated

b b2 ac
1 =
a

and 2 =

b2 ac
.
a

Note that 1 and 2 are the real solutions of the equation

a2 + 2b + c = 0.
34

(5.11)

In order to obtain a nonsingular map (x, y) 7 ((x, y), (x, y)), we

choose to be the solution of (5.9) and to be the solution of (5.10).
To solve (5.9), we use the method of characteristics (except that we dont
specify the initial condition). The characteristic equations are
dx
= 1,
dt

dy
= 1 ,
dt

dz
= 0.
dt

The last equation says that the solution is constant along each of the characteristics (x(t), y(t)). In view of the first two equations, the characteristics
can be obtain as curves y = y(x) solving
dy
=
dx

dy
dt
dx
dt

= 1 ,

(5.12)

and then the solution is constant at points (x, y(x)). Similarly, one solves
(5.10) to obtain .
In summary, to choose and one solves the (5.11) to obtain two real
roots 1 and 2 . Then, denoting by
f (x, y) = C1

and

g(x, y) = C2

the solutions of characteristics equations

dy
= 1 (x, y
dx

and

dy
= 2 (x, y),
dx

(5.13)

the variables and are defined by

(x, y) = f (x, y),

(x, y) = g(x, y).

The solutions of both equations in (5.13) are called the two families of characteristics of (5.1).
Example 5.4. Consider
yuxx + uyy = 0
In the region where y < 0, the equation is hyperbolic. Solving y2 + 1 = 0,
one finds two real solutions
1 =

1
(y)1/2

and 2 =

We look for two real families of characteristics,

dy
1
=0

dx (y)1/2

and
35

1
(y)1/2

dy
dx

+ 1 = 0 and

dy
1
= 0.
+
dx (y)1/2

dy
dx

+ 2 = 0,

The solutions of the equations are

2
(y)3/2 + x = C1
3

2
(y)3/2 + x = C2 .
3

and

Therefore, we set
2
= (y)3/2 + x
3

and

2
= (y)3/2 + x.
3

The derivatives of and are,

x = 1,

y = (y)1/2 ,

x = 1,

y = (y)1/2 .

With u(x, y) = v((x, y), (x, y), one gets

ux = v + v
uy = (y)1/2 v + (y)1/2 v

uxx = v + 2v + v

1
uyy = yv + 2yv yv + (y)1/2 [v v ].
2
Substituting into the equation, one obtains
0 = yuxx + uyy



1
1
1/2
3/2
[v v ] = 4y v (y)
(v v ) .
= 4yv + (y)
2
8

Since = 43 (y)3/2 , one concludes that

v

1
(v v ) = 0.
6( )

Parabolic equations.
Suppose that (5.1) is parabolic on the domain U . Hence b2 ac = 0 at
each point of U . As before assume that a 6= 0 on U . We find a map
(x, y) 7 (x, y), (x, y)) so that B(, ) = A(, ) = 0. It suffices to make
A = 0 since 0 = B 2 AC = B 2 implies that B(, ) = 0. Under such
a change of coordinates the parabolic equation (5.1) can be brought to its
canonical form
e
L[w]
= w + [w] = G(, )

where is a first-order linear operator and G is function.

36

To do this we look for (x, y) 7 (x, y) so that

A(, ) = ax2 + 2bx y + cy2 = 0.
Since b2 = ac, we have


b
b2
ax2 + 2bx y + cy2 = a x2 + 2 x y + 2 y2 = a (x y )2 ,
a
a
where = ab is the double root of
a2 + 2b + c = 0.
So, we look for the solution of the first-order linear equation,
x y = 0.
The solution is constant along each characteristic which is determined by
the equation
b
dy
= = .
(5.14)
dx
a
For the map (x, y) 7 (x, y) we can take any map so that x y x 6= 0.
In summary, in the parabolic case to choose and one solves the (5.11)
to obtain double root = b/a. Then, denoting by
f (x, y) = C
the solution of characteristics equation
dy
=
dx

(5.15)

the variable is defined by

(x, y) = f (x, y)
and the variable is chosen so that x y x 6= 0.
Example 5.5. Reduce the following equation to its canonical form and then
find the general solution,
x2 uxx 2xyuxy + y 2 uyy + xux + yuy = 0
for x > 0.
37

The discriminant is equal to


xy x2
=0
det
y 2 xy


so that the equation is parabolic. The quadratic equation x2 2 2xy+y 2 =

0 has exactly one solution, = xy . Next we look for characteristics. These
are solutions of the equation
dy
= ,
dx

i.e.,

dy
y
= .
dx
x

The family of solutions is given by xy = C. Therefore, we define (x, y) = xy

and we take as the second independent variable (x, y) = x. Then
x = y,

y = x,

x = 1,

y = 0.

The Jacobian of the map (x, y) 7 ((x, y), (x, y)) is nonzero. Let v(, ) =
u(x(, ), y(, )), that is, u(x, y) = v((x, y), (x, y)). Using the chain rule,
ux = yv + v
uy = xv
uxx = y 2 v + 2yv + v
uxy = xyv + xv + v
uyy = x2 v .
Substituting into the equation, one gets
x2 v + xv = 0,
so that, using x = ,

1
v + v = 0.

To solve this equation introduce the function w = v . Then

1
w = w

which has the general solution w = 1 A(). So,

v =

1
A()

38

which after integration gives

v(, ) = A() ln + B().
Since (x, y) = xy and (x, y) = x and u(x, y) = v((x, y), (x, y)), the
general solution of the equation has the form
u(x, y) = A(xy) ln x + B(xy)
where A, B are two arbitrary functions of class C 2 .
Elliptic equations.
Suppose that (5.1) is elliptic on the domain U . Then b2 ac < 0 at each
point of U . This time we look for the map (x, y) 7 (x(x, y), (x, y)) so
that
A(, ) = ax2 + 2bx y + cy2 = C(, ) = ax2 + 2bx y + cy2
B(, ) = ax x + b(x y + y x ) + cy y = 0.
Under such a change of coordinates and dividing by A, the elliptic equation (5.1) can be brought to its canonical form
e
L[w]
= w + w + [w] = G(, )

where is a first-order linear operator.

The above system consists of two nonlinear first-order equations. Subtracting C(, ) form A(, ) and multiplying B(, ) by 2i leads to the
following system,
a(x2 x2 ) + 2b(x y x y ) + c(y2 y2 ) = 0

ax (2ix ) + b(x (2iy ) + y (2ix )) + cy (2iy ) = 0.

Consequently, setting = + i, one finds that the above system is equivalent to
a2x + 2bx y + c2y = 0.
This can be written as a product,
a [x 1 y ] [x 2 y ] = 0,
where we have abbreviated

b i ac b2
1 =
a

and
39

b i ac b2
2 =
.
a

These are complex roots of

a2 + 2b + c = 0.
Note that 1 and 2 are conjugated, i.e., 1 = 2 . As in the hyperbolic case
we solve the characteristics equations
dy
= 1
dx

and

dy
= 2 .
dx

(5.16)

This time the solutions are complex. If f (x, y) = C1 and g(x, y) = C2 are
complex solutions of (5.16), then
(x, y) = f (x, y)
Then set
=

1
( + )
2

and (x, y) = g(x, y).

and =

1
( ).
2i

Example 5.6. Consider

yuxx + uyy = 0
In the region where y > 0, the equation is elliptic. Solving y2 + 1 = 0, one
finds two complex solutions
1 =

i
y 1/2

and

2 =

i
y 1/2

dy
+1 = 0 and
We look for two complex families of characteristics, dx
0,
dy
dy
i
i
+ 1/2 = 0 and
1/2 = 0.
dx y
dx y

The solutions of the equations are

2 3/2
y + ix = C1
3

and

2 3/2
y ix = C2 .
3

and

2
= y 3/2 x,
3

Therefore, we set
2
= y 3/2 + x
3
and then
1
2
= ( + ) = y 3/2
2
3

and
40

1
( ) = x.
2i

dy
dx +2

The derivatives of and are,

x = 0,

y = y 1/2 ,

x = 1,

With u(x, y) = v((x, y), (x, y), one gets

ux = v
uy = y 1/2 v
uxx = v
1
uyy = yv + y 1/2 v .
2
Substituting into the equation, one obtains
v + v +

1
v = 0.
2y 3/2

Finally, since = 23 y 3/2 , the equation becomes

3
v + v + v = 0.

41

y = 0.