MATH34032: Greens Functions, Integral Equations and the Calculus of Variations 1

Section 4
Integral Equations: Neumann Series
As we have seen in Section 3, a Fredholm integral equation of the second kind may be
written as
f (x) =
_
b
a
K (x, y) f (y) dy + g (x) . (4.1)
We also saw, in section 3, that if the Kernel is separable, then we may take steps to solve
the integral equation analytically, by reducing it to a system of linear equations. However,
what if the Kernel is not separable? In general these problems are much harder. One
approach exploits the fact that for suciently small values of it looks like we can neglect
the integral term to obtain the approximate solution
f (x) g (x) = f
0
(x) . (4.2)
We can then improve on this approximation by substituting (4.2) into the right hand side
of (4.1), i.e.
f (x) f
0
(x) +
_
b
a
K (x, y) f
0
(y) dy
= f
0
(x) + f
1
(x) . (4.3)
Substituting (4.3) into the right hand side of (4.1) then yields an even better approximation
f (x) f
0
(x) +
_
b
a
K (x, y) [f
0
(y) + f
1
(y)] dy
= f
0
(x) +
_
b
a
K (x, y) f
0
(y) dy +
2
_
b
a
K (x, y) f
1
(y) dy
= f
0
(x) + f
1
(x) +
2
_
b
a
K (x, y) f
1
(y) dy
= f
0
(x) + f
1
(x) +
2
f
2
(x) ,
say. This suggests letting
f (x) =

n=0

n
f
n
(x) , (4.4)
which may be substituted into (4.1),

n=0

n
f
n
(x) =
_
b
a
K (x, y)
_

n=0

n
f
n
(y)
_
dy + g (x)
=

n=0

n+1
_
b
a
K (x, y) f
n
(y) dy + g (x) .
Equating coecients of like powers in gives

0
: f
0
(x) = g(x),

n
, n > 0 : f
n
(x) =
_
b
a
K (x, y) f
n1
(y) dy.
(4.5)
MATH34032: Greens Functions, Integral Equations and the Calculus of Variations 2
Denition 4.1 Formula (4.4), with (4.5), is called the Neumann series for the integral
equation (4.1).
Example 1: Find the rst 3 terms of the Neumann series for the integral equation
u(x) = 1 +
_
1
0
(x y)u(y) dy.
where R.
We note that = and K(x, y) = (x y) so that
u(x) = u
0
(x) + u
1
(x) +
2
u
2
(x) + ....
and it remains for us to determine each term. Clearly neglecting the integral term gives
u
0
(x) = 1, and then
u
1
(x) =
_
1/2
0
(x y)u
0
(y)dy (4.6)
=
_
xy
y
2
2
_
1/2
0
(4.7)
=
_
x
2

1
8
_
(4.8)
and then
u
2
(x) =
_
1/2
0
(x y)u
1
(y)dy (4.9)
=
_
1/2
0
(x y)
_
y
2

1
8
_
dy (4.10)
=
_
x
y
2
4
x
y
8

y
3
6
+
y
2
16
_
1/2
0
(4.11)
=
1
192
(4.12)
so that the Neumann series takes the form
u(x) = 1 +
_
x
2

1
8
_

2
192
+ .... (4.13)
But this approach is clearly only useful if the resulting series converges.....
MATH34032: Greens Functions, Integral Equations and the Calculus of Variations 3
Iterated kernels and convergence
If the Neumann series is convergent it solves (4.1) uniquely. Now
f
n
(x) =
_
b
a
K (x, y
1
) f
n1
(y
1
) dy
1
=
_
b
a
_
b
a
K (x, y
1
) K (y
1
, y
2
) f
n2
(y
2
) dy
2
dy
1
,
replacing y by y
1
in the rst integral, and y
2
in the second integral (for f
n1
). We can
continue replacing f
m
(x) by an integral containing f
m1
(x) for each m down to m = 1,
i.e.
f
n
(x) =
_
b
a
_
b
a
. . .
_
b
a
K (x, y
1
) K (y
1
, y
2
) . . . K (y
n1
, y
n
) f
0
(y
n
) dy
n
. . . dy
2
dy
1
=
_
b
a
__
b
a
. . .
_
b
a
K (x, y
1
) K (y
1
, y
2
) . . . K (y
n1
, y
n
) dy
n1
. . . dy
1
_
f
0
(y
n
) dy
n
=
_
b
a
K
n
(x, y) g (y) dy. (4.14)
Denition 4.2 The nth iterated kernel is
K
n
(x, y) =
_
b
a
. . .
_
b
a
K (x, y
1
) K (y
1
, y
2
) . . . K (y
n1
, y) dy
n1
. . . dy
1
.
Proposition 4.3 The nth iterated kernel satises the following inductive denition
K
1
(x, y) = K (x, y) ,
K
n+1
(x, y) =
_
b
a
K (x, z) K
n
(z, y) dz.
Proof: Using Denition 4.2 with n replaced by n + 1 we have
K
n+1
(x, y) =
_
b
a
. . .
_
b
a
K (x, y
1
) K (y
1
, y
2
) . . . K (y
n1
, y
n
) K (y
n
, y) dy
n
. . . dy
1
.
Now, setting y
1
= z, say, and then relabelling each y
m
, m = 2, 3, . . . n, as y
m1
yields
K
n+1
(x, y) =
_
b
a
K (x, z)
_
b
a
. . .
_
b
a
K (z, y
1
) . . . K (y
n2
, y
n1
) K (y
n1
, y) dy
n1
. . . dy
1
dz,
or
K
n+1
(x, y) =
_
b
a
K (x, z) K
n
(z, y) dz
as required.
From (4.4), (4.14) the Neumann series may be written
f (x) = f
0
(x) +

n=1
_

n
_
b
a
K
n
(x, y) g (y) dy
_
MATH34032: Greens Functions, Integral Equations and the Calculus of Variations 4
= f
0
(x) +
_
b
a
_

n=1

n1
K
n
(x, y)
_
g (y) dy
= g (x) +
_
b
a
R(, x, y) g (y) dy. (4.15)
Thus, comparing with our previous denition of the resolvent kernel (see Denition 3.11)
we get the following proposition.
Proposition 4.4 The resolvent kernel R(, x, y) may be written
R(, x, y) =

n=1

n1
K
n
(x, y) .
Theorem 4.5: If it exists, R(, x, z) is unique and solves (4.1) via (4.15).
No proof given.
Theorem 4.6: The Neumann series for the Fredholm equation
f (x) =
_
b
a
K (x, y) f (y) dy + g (x) ,
where K (x, y) and g (x) are bounded and absolutely integrable, converges absolutely for all
values of such that
|| <
1
M (b a)
,
where
M = sup
axb
ayb
|K (x, y)| .
No proof given.
Example 2: Determine for which values of the Neumann series converges in Example
1.
Using Theorem 4.6, we see that b = 1, a = 0 and
M = sup
axb
ayb
|K (x, y)|
= sup
axb
ayb
|x y|
= 1
so that the series converges for || < 1.
MATH34032: Greens Functions, Integral Equations and the Calculus of Variations 5
Theorem 4.7: The Neumann series for the Volterra equation
f (x) =
_
x
a
K (x, y) f (y) dy + g (x) ,
where K (x, y) is bounded and g (x) is absolutely integrable, is absolutely convergent for all
values of .
No proof given.
Theorem 4.8: The mth iterated kernel for a Volterra integral equation is given by
K
m
(x, y) =
_ _
x
y
K (x, z) K
m1
(z, y) dz if y < x,
0 if y x.
Proof: We know from the denition of Volterra kernels that K
1
(x, y) = K(x, y) is zero
for y > x. So, start with m = 2:
K
2
(x, y) =
_
b
a
K (x, z) K (z, y) dz.
The rst term in the integrand, K(x, z), is zero when z > x and the second, K(z, y), is
zero when z < y. Hence
K
2
(x, y) =
_
x
y
K (x, z) K
1
(z, y) dz.
The proof may now be completed by induction.
Example 3: Solve the Volterra integral equation
f (x) =
_
x
0
f (y) dy + g (x)
and write down the resolvent kernel.
Solution: Here, K
1
(x, y) = K (x, y) = 1.
K
2
(x, y) =
_
x
y
K (x, z) K
1
(z, y) dz =
_
x
y
dz = x y,
K
3
(x, y) =
_
x
y
K (x, z) K
2
(z, y) dz =
_
x
y
1. (z y) dz
=
_
1
2
(z y)
2
_
x
y
=
1
2
(x y)
2
,
K
4
(x, y) =
_
x
y
K (x, z) K
3
(z, y) dz =
_
x
y
1.
1
2
(z y)
2
dz
=
_
1
3!
(z y)
3
_
x
y
=
1
3!
(x y)
3
.
MATH34032: Greens Functions, Integral Equations and the Calculus of Variations 6
Suppose that, in general,
K
n
(x, y) =
(x y)
n1
(n 1)!
. (4.16)
Assume true for n = k, then for n = k + 1
K
k+1
=
_
x
y
K (x, z) K
k
(z, y) dz =
_
x
y
(z y)
k1
(k 1)!
dz
=
_
(z y)
k
k!
_
x
y
=
(x y)
k
k!
.
Hence, (4.16) follows by induction. Therefore, from (4.15)
f (x) = g (x) +

n=1
_

n
_
x
0
K
n
(x, y) g (y) dy
_
= g (x) +
_
_
x
0

n=1

n1
K
n
(x, y) g (y) dy
_
= g (x) +
_
x
0
_

n=1

n1
(x y)
n1
(n 1)!
_
g (y) dy
= g (x) +
_
x
0
exp [ (x y)] g (y) dy.
Note, by comparison with (4.15), that the resolvent kernel is exp [ (x y)].
The theory for degenerate kernels and for Neumann series may be used to prove the
following theorem, valid for an arbitrary kernel.
Theorem 4.9: The Fredholm Alternative for integral equations
Either the integral equation
f

Kf = g, f (x)
_
b
a
K (x, y) f (y) dy = g (x) , (4.17)
for xed has precisely one solution f (x) for each arbitrary g (x); in particular
f = 0 when g = 0.
The associated adjoint integral equation
f (x)
_
b
a
K (y, x) f (y) dy = g (x) ,
also has a unique solution. Note that the adjoint equation has kernel K(y, x) not
K(x, y).
In this case is called a regular value.
MATH34032: Greens Functions, Integral Equations and the Calculus of Variations 7
Or the associated homogeneous equation
f =

Kf, f (x) =
_
b
a
K (x, y) f (y) dy,
has m > 0, m N, linearly independent solutions f
1
, f
2
, . . . , f
m
.
The associated homogeneous adjoint integral equation
f (x) =
_
b
a
K (y, x) f (y) dy,
also has m linearly independent solutions h
1
, h
2
, . . . , h
m
.
In this case equation (4.17) has a solution if and only if g (x) satises
g, h
i
=
_
b
a
g (y) h
i
(y) dy = 0
for i = 1, . . . , m, and the solution of (4.17) is then given by
f (x) = f
0
(x) +
m

i=1

i
f
i
(x)
where the
i
are arbitrary constants.
In this case is called a characteristic value or eigenvalue.