Beruflich Dokumente
Kultur Dokumente
Integral Equations
051012 F. Porter
Revision 150928 F. Porter
Introduction
Z b
(1)
We are given functions h(x), g(x), k(x, y), and wish to determine f (x). The
quantity is a parameter, which may be complex in general. The bivariate
function k(x, y) is called the kernel of the integral equation.
We shall assume that h(x) and g(x) are defined and continuous on the
interval a x b, and that the kernel is defined and continuous on a x b
and a y b. Here we will concentrate on the problem for real variables
x and y. The functions may be complex-valued, although we will sometimes
simplify the discussion by considering real functions. However, many of the
results can be generalized in fairly obvious ways, such as relaxation to piecewise continuous functions, and generalization to multiple dimensions.
There are many resources for further reading on this subject. Some of
the popular ones among physicists include the classic texts by Mathews
and Walker, Courant and Hilbert, Whittaker and Watson, and Margenau
and Murphy, as well as the newer texts by Arfken, and Riley, Hobson, and
Bence.
Integral Transforms
g(x) =
(2)
2.1
Fourier Transforms
(3)
(4)
(5)
(6)
(7)
Well forego rigor here and give the physicists demonstration of this:
1 Z ixy Z ix0 y
e g(x0 )dx0
e
dy
2
Z
1 Z
0
g(x0 )dx0
ei(x x)y dy
=
2
g(x) =
(8)
(9)
(10)
= g(x).
(11)
Here, we have used the fact that the Dirac delta-function may be written
1 Z ixy
(x) =
e dy.
2
(12)
The reader is encouraged to demonstrate this, if s/he has not done so before.
It is instructive to notice that the Fourier transform may be regarded as
a limit of the Fourier series. Let f (x) be expanded in a Fourier series in a
box of size [L/2, L/2]:
f (x) =
an e2inx/L .
(13)
n=
(14)
Hence,
1 Z L/2
an =
f (x)e2inx/L dx.
(15)
L L/2
Now consider taking the limit as L . In this limit, the summation
lim
an e2inx/L
(16)
n=
2
g(y)eixy
= lim
L
L
n=
1 Z ixy
e g(y)dy.
=
2
(17)
(18)
Furthermore:
Lan
1 Z
g(y) = =
f (x)eixy dx.
(19)
2
2
We thus verify our earlier statements, including the -function equivalence,
assuming our limit procedure is acceptable.
Suppose now that f (y) is an even function, f (y) = f (y). Then,
Z
0
1
ixy
g(x) =
eixy f (y)dy
e
f (y)dy +
0
2
i
1 Z h ixy
=
e + eixy f (y) dy
2 0
s
2Z
f (y) cos xy dy.
=
0
Z
(20)
(21)
(22)
f (y) =
2Z
g(x) cos xy dx.
0
(23)
f (y) =
2Z
g(x) sin xy dx.
0
(25)
(26)
In electronics we use the Fourier transform to translate time domain problems in terms of frequency domain problems, with xy t. An LCR
circuit is just a complex impedance for a given frequency, hence the integraldifferential time-domain problem is translated into an algebraic problem in
the frequency domain. In quantum mechanics the position-space wave functions are related to momenutm-space wave functions via the Fourier transform.
2.1.1
Example: RC circuit
Suppose we wish to determine the output voltage Vo (t) in the simple circuit
of Fig. 1. The time domain problem requires solving the equation:
1 1
1 Zt
1 Zt
0
0
Vi (t ) dt
+
Vo (t0 ) dt0 .
Vo (t) =
R1 C
C R1 R2
(27)
R1
Vi (t)
R2
Vo(t)
(28)
(29)
(30)
1+
R1
R2
1
,
(1 + iR2 C)
(31)
Vb
(32)
(33)
We are here using the hat ( b ) notation to indicate the integral transform of
the unhatted function. The answer to the problem for general (not necessarily
sinusoidal) input Vi (t) is then:
1 Z b
Vo (t) =
Vo ()eit d
2
1 Z b
eit
=
Vi ()
d.
1
1+ R
(1
+
iR
C)
2
2
R2
(34)
(35)
2.2
Laplace Transforms
f (x)esx dx.
(36)
(37)
where x > 0.
This transform can be useful for some functions where the Fourier transform does not exist. Problems at x + are removed by multiplying by
ecx , where c is a positive real number. Then the problem at is repaired
by multiplying by the unit step function (x):
(x)
if x > 0,
1/2
if x = 0, and
0
if x < 0.
(38)
Thus, we have
g(y) =
=
f (x)(x)ecx eixy dx
(39)
(40)
cx
1 Z
g(y)eixy dy.
(x)f (x) =
2
(41)
f (x)esx dx,
(42)
and
1 Z
F (s)ex(c+iy) dy
2
1 Z c+i
=
F (s)exs ds,
2i ci
f (x)(x) =
(43)
(44)
(45)
f (x)esx dx
(46)
(47)
(48)
(49)
Z
f (x)dx
(y) =
1
(Ff ) (y) + C(y),
iy
(50)
Z x
f (t)dt
(s) =
dxesx
Z x
dtf (t)
f (t)dt
dxesx
1
(Lf ) (s).
s
(51)
4. Translation:
[Ff (x + a)] (y) = eiay [Ff ] (y),
[Lf (x + a)] (s) = eas [Lf ] (s) (a)
(52)
Z a
f (x)esx dx.
(53)
5. Multiplication by an exponential:
{F [eax f (x)]} (y) = (Ff ) (y + ia),
{L [eax f (x)]} (s) = (Lf ) (s a).
(54)
(55)
6. Multiplication by x:
d
(Ff ) (y),
dy
d
{L [xf (x)]} (s) = (Lf ) (s).
ds
(56)
(57)
Z b
a
f1 (y)f2 (x y)dy.
(58)
(59)
(60)
(61)
(62)
(63)
This gives us three equations relating the unknowns Vo (t), i(t), and iC (t),
which we could try to solve to obtain Vo (t). However, the integral in the last
equation complicates the solution. This is where the Laplace transform will
help us.
o (s)
Z
0
(64)
(65)
Vbo (s) =
(66)
(70)
where
R1 R2
C.
R1 + R2
(71)
(72)
Im(s)
x
-1/
Re(s)
s(t)
1 e
R2
= A
lim s +
R1 + R2 s1/
1 + s
1 t
= A
e .
R1 C
(73)
There are many other integral transforms that we could investigate, but the
Fourier and Laplace transforms are the most ubiquitous in physics applications. Rather than pursue other transforms, well look at another example
10
that suggests the breadth of application of these ideas. This is the Laplaces
Method for the solution of ordinary differential equations. This method represents a sort of generalization of the Laplace transform, using the feature of
turning derivatives into powers.
Suppose we wish to solve the differential equation:
n
X
(74)
k=0
where the contour will be chosen to suit the problem (not necessarily the
contour of the Laplace transform). Well insert this proposed solution into
the differential equation. Notice that
f 0 (x) =
F (s)sesx ds.
(77)
Thus,
0=
(78)
where
n
X
U (s) =
k=0
n
X
V (s) =
ak s k ,
(79)
bk s k .
(80)
k=0
sx
sx
c2
]
c1
Z
C
d
[V (s)F (s)] esx ds,
ds
(81)
Z (
C
2
d
sx
sx
U (s)F (s)
[V (s)F (s)] e ds + [V (s)F (s)e ] .
ds
c1
(82)
We assume that we can choose C such that the integrated part vanishes.
Then we will have a solution to the differential equation if
U (s)F (s)
d
[V (s)F (s)] = 0.
ds
11
(83)
ds
V
ds
!
Z
U
d ln V
ln F =
ds
V
ds
Z
U
ds ln V + ln A,
=
V
(84)
(85)
(86)
(87)
(88)
3.1
U (s0 ) 0
ds .
V (s0 )
#
(89)
(90)
(91)
(92)
(93)
(94)
es /4
+1 .
2s
12
(95)
Im(s)
Re(s)
F (s)esx ds
(96)
C
2
es +2sx
= A
ds,
(97)
s+1
C
where A is an arbitrary constant, and where we have let s 2s according
to convention for this problem.
Now we are faced with the question of the choice of contour C. We at
least must require that the integrated part in Eqn. 82 vanish:
Z
2
V (s)F (s)esx
1
es +2sx
= 0.
s 1
(98)
n! Z es +2sx
ds.
(99)
2i C sn+1
This is a powerful integral form for the Hermite polynomials (or, Hermite
functions in general with the branch cut contour of Fig. 3). For example,
Hn (x)
n!
Hn (x) = 2i
residue at s = 0 of
2i
13
es +2sx
.
sn+1
!
(100)
Recall that the residue is the coefficient of the 1/s term in the Laurent series
expansion. Hence,
Hn (x) = n! coefficient of sn in es
That is,
es
2 +2sx
2 +2sx
Hn (x) n
s .
n!
n=0
(101)
(102)
Hn (x) = lim
(103)
(104)
s0
2 +2sx
s0
= 2x,
(105)
and so forth.
Z b
(106)
Z b
(107)
If
Lf1 = g1 and
Lf2 = g2 ,
(108)
(109)
(110)
14
(111)
where K|f i indicates here the integral ab k(x, y)f (y)dy. Our linear operator
is then written:
L = I K,
(112)
R
4.1
Z b
(113)
|f (x)|2 dx = 1.
(114)
If there are several solutions, f1 , f2 , f3 , . . . , fn , then any linear combination of these is also a solution. Hence, if we have several linearly independent
solutions, we can assume that they are orthogonal and normalized. If they
are not, we may use the Gram-Schmidt process to obtain such a set of orthonormal solutions. We therefore assume, without loss of generality, that:
Z b
a
fi (x)fj (x)dx = ij .
(115)
(116)
|f ihf | = If ,
(117)
or even
where If is the identity matrix in the subspace spanned by {f }.
A value of for which the homogeneous equation has non-trivial solutions
is called an eigenvalue of the equation (or, of the kernel). Note that the use
of the term eigenvalue here is analogous with, but different in detail from the
15
usage in matrices our present eigenvalue is more similar with the inverse of
a matrix eigenvalue. The corresponding solutions are called eigenfunctions
of the kernel for eigenvalue . We have the following:
Theorem: There are a finite number of eigenfunctions fi corresponding to
a given eigenvalue .
Proof: Well prove this for real functions, leaving the complex case as an
exercise. Given an eigenfunction fj corresponding to eigenvalue , let:
pj (x)
Z b
a
1
fj (x).
(118)
Z b
a
k(x, y)
n
X
(119)
j=1
It must be that D(x) 0 because the integrand is nowhere negative for any
x. Note that the sum term may be regarded as an approximation to the
kernel, hence D(x) is a measure of the closeness of the approximation. With
some manipulation:
D(x) = 2
Z b
[k(x, y)]2 dy 22
Z bX
n
a j=1
2
Z b X
n
+2
pj (x)fj (y) dy
a
Z b
j=1
2
[k(x, y)] dy 2
+2
[pj (x)]2
j=1
n
X
pj (x)
j=1
= 2
n
X
Z b
n
X
pk (x)
Z b
a
k=1
[k(x, y)]2 dy 2
fj (y)fk (y)dy
n
X
[pj (x)]2 .
(120)
j=1
Z b
[k(x, y)]2 dy
n
X
[fj (x)]2 .
(121)
j=1
Z bZ b
a
n Z b
X
j=1 a
n,
16
[fj (x)]2 dx
(122)
Z bZ b
a
(123)
Z b
a
Z b
a
y [a, b],
(124)
x [a, b],
(125)
where U1 and U2 are some fixed upper bounds. We will assume that these
conditions are satisfied in our following discussion. Note that the kernel may
actually be discontinuous and even become infinite in [a, b], as long as these
conditions are satisfied.
4.2
Degenerate Kernels
Definition (Degenerate Kernel ):If we can write the kernel in the form:
k(x, y) =
n
X
i (x)i (y)
(126)
i=1
We may
(or K = ni=1 |i ihi |), then the kernel is called degenerate.
assume that the i (x) are linearly independent. Otherwise we could reduce
the number of terms in the sum to use only independent functions. Likewise
we may assume that the i (x) are linearly independent.
The notion of a degenerate kernel is important due to two facts:
P
k = 0, 1, 2, . . . and z A.
17
(127)
Note that this is a rather strong form of convergence a series may converge
for all z A, but may not be uniformly convergent.
Let us now pursue the second fact asserted above. We wish to solve for
f:
Z
b
g(x) = f (x)
(128)
n
X
i (x)
Z b
a
i=1
i (y)f (y)dy.
(129)
Z b
a
Z b
a
Z b
a
i (x)g(x)dx
(130)
i (x)f (x)dx
(131)
i (x)j (x)dx.
(132)
n
X
cji fi .
(133)
i=1
n
X
fi i (x).
(134)
i=1
Substituting in:
g(x) = g(x) +
n
X
fi i (x)
= g(x) +
i (x) fi
i=1
i (x)
i=1n
i=1
n
X
Z b
a
Z b
a
i (y) g(y) +
i (y) g(y) +
n
X
fj j (y) dy
j=1
n
X
j=1
fj j (y) dy
n
n
X
X
g(x) +
i (x) fi gi +
cij fj
i=1
j=1
= g(x).
(135)
x = f (x)
Z 1
0
18
(136)
g1 =
c11
Z 1
(1 + x)x2 dx =
(137)
(138)
(139)
giving
f1 =
7 1
,
2 6 5
(140)
and
7
x.
(141)
2 6 5
The reader is encouraged to check that this is a solution to the original
equation, and that no solution exists if = 6/5.
To investigate this special value = 6/5, consider the homogeneous equation:
Z 1
x(1 + y)f (y)dy.
(142)
f (x) =
f (x) = x2 +
We may use the same procedure in this case, except now g1 = 0 and we find
that
5
f1 1
= 0.
(143)
6
Either f1 = 0 or = 6/5. If f1 = 0, then f (x) = g(x) + f1 1 (x) = 0. If
6= 6/5 the only solution to the homogeneous equation is the trivial one. But
if = 6/5 the solution to the homogeneous equation is f (x) = ax, where a
is arbitrary. The value = 6/5 is an (in this case the only) eigenvalue ofthe
integral equation, with corresponding normalized eigenfunction f (x) = 3x.
This example suggests the plausibility of the important theorem in the
next section.
4.3
Z b
a
19
(144)
with given , possesses a unique continuous solution f (x) for each continuous function g(x) (and in particular f (x) = 0 if g(x) = 0), or the
associated homogeneous equation
f (x) =
Z b
(145)
K =
|i ihi |
(146)
i=1
|f i = |gi +
n
X
|i ihi |f i,
(147)
i=1
and let
gi hi |gi
fi hi |f i
cij hj |i i.
Then,
f j = gj +
n
X
cji fi ,
(148)
(149)
(150)
(151)
i=1
or
g1
g2
g=
.. = (I C) f ,
.
(152)
gn
where C is the matrix formed of the cij constants.
Thus, we have a system of n linear equations for the n unknowns {fi }.
Either the matrix I C is non-singular, in which case a unique solution f
exists for any given g (in particular f = 0 if g = 0), or I C is singular, in
which case the homogeneous equation f = Cf possesses a finite number of
linearly independent solutions. Up to some further considerations concerning
continuity, this proves the theorem for the case of a degenerate kernel.
We may extend the proof to arbitrary kernels by appealing to the fact
that any continuous funciton k(x, y) may be uniformly approximated by degenerate kernels in a closed interval (for example, see Courant and Hilbert).
There is an additional useful theorem under Fredholms name:
20
Z b
(153)
Z b
k(y, x)t(y)dy
(154)
Z b
k(y, x)t(y)dy
(155)
Z b
a
g (x)ti (x)dx = 0,
i = 1, 2, . . . , n.
(156)
Practical Approaches
We turn now to a discussion of some practical tools of the trade for solving
integral equations.
5.1
Degenerate Kernels
If the kernel is degenerate, we have shown that the solution may be obtained
by transforming the problem to that of solving a system of linear equations.
21
5.2
Volterras Equations
Z x
(158)
f (x) = g(x) +
Z x
(159)
are called Volterras equations of the first and second kind, respectively.
One situation where such equations arise is when k(x, y) = 0 for y > x:
k(x, y) = (x y)`(x, y). Thus,
Z b
Z x
(160)
(161)
We may use this to transform the equation of the first kind to:
Z x
dg
k
(x) = k(x, x)f (x) +
(x, y)f (y)dy.
dx
a x
(162)
This is now a Volterras equation of the second kind, and the approach to
solution may thus be similar.
Notice that if the kernel is independent of x, k(x, y) = k(y), then the
solution to the equation of the first kind is simply:
f (x) =
1 dg
(x).
k(x) dx
(163)
x =1+
Z x
xyf (y)dy.
(164)
Rx
1
yf (y)dy. Then
x2 1
.
x
(165)
2x = x f (x) +
Z x
(166)
22
(167)
(168)
The Volterras equation readily lends itself to a numerical approach to solution on a grid (or mesh or lattice). We note first that (absorbing the
factor into the definition of k for convenience):
f (a) = g(a) +
Z x=a
= g(a).
(169)
n = 0, 1, . . . , N,
xa
.
N
(170)
We have here defined a uniform grid, but that is not a requirement. Now let
gn = g(xn )
fn = f (xn )
knm = k(xn , xm ).
(171)
(172)
(173)
Note that f0 = g0 .
We may pick various approaches to the numerical integration, for example, the trapezoidal rule gives:
Z xn
a
n1
X
1
1
k(xn , y)f (y) dy
kn0 f0 +
knm fm + knn fn .
2
2
m=1
(174)
23
n = 1, 2, . . . N.
(175)
gn +
1
k f
2 n0 0
Pn1
m=1
knm fm
k
2 nn
n = 1, 2, . . . N.
(176)
For example,
f 0 = g0 ,
g1 + 2 k10 f0
f1 =
,
1 2 k11
f2 =
g2 +
(177)
(178)
k f + k21 f1
2 20 0
,
1 2 k22
(179)
and so forth.
We note that we dont even have to explicitly solve a system of linear
equations, as we did for Fredholms equation with a degenerate kernel. There
are of order
!
O
N
X
n = O N2
(180)
n=1
5.3
Often an exact closed solution is elusive, and we resort to approximate methods. For example, one common approach is the iterative solution. We start
with the integral equation:
f (x) = g(x) +
Z b
(181)
(182)
Z b
k(x, y)g(y)dy.
(183)
Z b
"
k(x, y) g(y) +
Z b
a
24
#
0
(184)
This may be continuted indefinitely, with the nth iterative solution given in
terms of the (n 1)th:
Z b
fn (x) = g(x) +
a
Z b
= g(x) +
+2
Z bZ
a
(185)
k(x, y)g(y)dy
(186)
a
b
+...
n1
Z b
Z b
(187)
This method is only useful if the series converges, and the faster the better. It will converge if the kernel is bounded and lambda is small enough.
We wont pursue this further here, except to note what happens if
Z b
k(x, y)g(y)dy = 0.
(188)
In this case, the series clearly converges, onto solution f (x) = g(x). However,
this solution is not necessarily unique, as we may add any linear combination
of solutions to the homogeneous equation.
5.4
Fredholm Series
Z b
(189)
Let
2 Z Z k(x, x) k(x, x0 )
0
k(x0 , x) k(x0 , x0 ) dxdx
2!
a
k(x, x)
k(x, x0 ) k(x, x00 )
Z
Z
Z
3
D() = 1
Z b
k(x, x)dx +
and let
2
D(x, y; ) = k(x, y)
Z
k(x, y)
k(z, y)
25
k(x, z)
dz
k(z, z)
2!
Z Z k(x, y)
k(z, y)
k(z 0 , y)
k(x, z) k(x, z 0 )
k(z, z) k(z, z 0 ) dzdz 0
k(z 0 , z) k(z 0 , z 0 )
+...,
(191)
Note that not everyone uses the same convention for this notation. For
example, Mathews and Walker defines D(x, y; ) to be 1/ times the quantity
defined here.
We have the following:
Theorem: If D() 6= 0 and if the Fredholms equation has a solution, then
the solution is, uniquely:
f (x) = g(x) +
Z b
a
D(x, y; )
g(y)dy.
D()
(192)
The homogeneous equation f (x) = ab k(x, y)f (y)dy has no continuous non-trivial solutions unless D() = 0.
R
n
X
k(x, xi )f (xi ),
(193)
i=1
26
the repeated variables. A heavy dot on a segment breaks the segment into
two meeting segments, according to the above rule, and furthermore means
integration over the repeated variable with a factor of . For illustration of
these rules:
k(x, y)
k(x, x)
k(x, y)k(y, z)
k(x, x)dx
Thus,
D(x, y; )
=
1
2!
...,
(194)
and
D() = 1
1
3!
1
+ 2!
+ ...
(195)
Let us try a very simple example to see how things work. Suppose we
wish to solve:
Z 1
f (x) = x +
xyf (y)dy.
(196)
0
Of course, this may be readily solved by elementary means, but let us apply
our new techniques. We have:
= k(x, y) = xy
Z 1
0
k(x, x)dx =
(197)
27
(198)
Z 1
0
2
Z 1Z 1
0
xz =
3
!2
(199)
2
(200)
n
dots
Z 1
dx . . . dx(n) x2 . . . x(n)
i2
Z 1
n
(201)
n
dots
n
(202)
= 1
= xy.
(203)
(204)
Z 1
0
D(x, y; )
g(y)dy
D()
3 Z 1 2
= x+
xy dy
3 0
3
=
x.
3
(205)
n
X
i (x)i (y),
i=1
28
(206)
Z 1
k(x, y)g(y)dy + 2
y 2 dy + 2 x
= x
X
n=0
Z 1Z 1
0
Z Z
y 2 (y 0 )2 dydy 0 + . . .
!n
(207)
f (x) = 3
Z 1
xyf (y)dy.
(209)
Indeed, f (x) = Ax solves this equation. The roots of D() are the eigenvalues
of the kernel. If the kernel is degenerate we only have a finite number of
eigenvalues.
Symmetric Kernels
[k(x, y)]2 dy M,
(210)
Z b"
k
a
#2
(x, y)
dy M 0 ,
(211)
Note that, since we are assuming real functions in this section, we do not put a complex
conjugate in our scalar products. But dont forget to put in the complex conjugate if you
have a problem with complex functions!
29
(212)
Z bZ b
a
k(x, y)(x)(y)dxdy,
(213)
where is any (piecewise) continuous function in [a, b]. Well assume |a|, |b| <
for simplicity here; the reader may consider what additional criteria must
be satisifed if the interval is infinite.
Our quadratic integral form is analogous with the quadratic form for
systems of linear equations:
A(x, x) =
n
X
aij xi xj = ( x )
x ,
(214)
i,j=1
Z Z
(215)
where
u(x, y) k(x, y)
v(x, y) (x)(y),
(216)
(217)
we have defined a scalar product between the vectors u and v. We are thus
led to consider its square,
[J(, )]2 =
dx
dy
dx0
Z Z
[(x)(y)]2 dxdy
30
Z Z
(219)
[(x)]2 dx = 1,
(220)
we see that |J(, )| is bounded, since the integral of the squared kernel is
bounded.
Furthermore, we can have J(, ) = 0 for all if and only if k(x, y) = 0.
The if part is obviously true; let us deal with the only if part. This
statement depends on the symmetry of the kernel. Consider the bilinear
integral form:
J(, ) = J(, )
Z Z
k(x, y)(x)(y)dxdy.
(221)
We have
J( + , + ) = J(, ) + J(, ) + 2J(, ),
(222)
for all , piecewise continuous on [a, b]. We see that J(, ) = 0 for all
only if it is also true that J(, ) = 0, , .
In particular, let us take
(y) =
k(x, y)(x)dx.
(223)
Then
0 = J(, ) =
=
Z Z
dx
dyk(x, y)(x)
2
dy.
k(x, y)(x)dx
(224)
(226)
We wish to show that this bound is actually achieved for a suitable (x).
Let us suppose that the kernel is uniformly approximated by a series of
degenerate symmetric kernels:
An (x, y) =
an
X
(n)
i,j=1
31
(227)
(n)
(n)
where cij = cji and hi |j i = ij , and such that the approximating kernels
are uniformly bounded in the senses:
Z b
a
Z b"
An
(228)
#2
(x, y)
dy MA0 ,
(229)
Z Z
Z
an
X
(n)
=
=
An (x)(y)dxdy
cij
i,j=1
an
X
i (x)(x)dx
j (y)(y)dy
(n)
cij ui uj ,
(230)
i,j=1
(x)
an
X
ui i (x)
dx 0,
(231)
i=1
an
X
u2i .
(232)
i=1
u2i = 1.
(233)
i=1
an
X
ui i (x),
(234)
i=1
32
maximum value in the domain. Suppose that {u} is the appropriate vector.
Then
a
n
X
(n)
cij ui uj = 1n
(235)
i,j=1
cij uj = 1n ui ,
i = 1, 2, . . . , an .
(236)
j=1
(237)
where {u} is now our (normalized) vector for which the quadratic form is
maximal. The normalization hn |n i still holds. Apply the approximate
kernel operator to this function:
Z
an
X
(n)
cij i (x)
i,j=1
an
X
an
X
(n)
i=1
j=1
i (x)
= 1n
an
X
j (y)n (y)dy
cij uj
ui i (x)
i=1
= 1n n (x).
(238)
x, y [a, b].
(239)
Thus,
2
[J(, ) Jn (, )]
Z Z
|h|i|
2
Z Z
Z bZ b
a
2
(Schwarz),
dxdy
2 (b a)2 .
(240)
33
Thus, the range of Jn may be made arbitrarily close to the range of J by taking n large enough, and hence, the maximum of Jn may be made arbitrarily
close to that of J:
lim 1n = 1 .
(241)
n
Now, by the Schwarz inequality, the functions n (x) are uniformly bounded
for all n:
[n (x)]
1n
2
21n hn |n i
(242)
(243)
whenever || < . This may be seen as follows: First, we show that 0n (x) is
uniformly bounded:
2
[0n (x)]
"
1n
An
(x, y)n (y)dy
x
#2
Z "
An
(x, y)
x
21n MA0 .
21n
#2
dy
(Schwarz)
(244)
Or, [0n (x)]2 MA00 , where MA00 = MA0 max 21n . With this, we find:
2
|n (x + ) n (x)|
=
=
Z x+
2
0
n (y)dy
x
Z
b
[(y x) (y
a
Z b
2
x )] 0n (y)dy
Z b
[(y x) (y x )]2 dy
||(b a)MA00
< ,
[0n (y)] dy
(245)
34
Theorem: (Arzela) If f1 (x), f2 (x), . . . is a uniformly bounded equicontinuous set of functions on a domain D, then it is possible to select a
subsequence that converges uniformly to a continuous limit function in
the domain D.
The proof of this is similar to the proof of the Bolzano-Weierstrass theorem,
which it relies on. We start by selecting a set of points x1 , x2 , . . . that is
everywhere dense in [a, b]. For example, we could pick successive midpoints
of intervals. By the Bolzano-Weierstrass theorem, this sequence of numbers contains a convergent subsequence. Now select an infinite sequence of
functions (out of {f }) a1 (x), a2 (x), . . . whose values at x1 form a convergent
sequence, which we may also accomplish by the same reasoning. Similarly,
select a convergent sequence of functions (out of {a}) b1 (x), b2 (x), . . . whose
values at x2 form a convergent sequence, and so on.
Now consider the diagonal sequence:
q1 (x)
q2 (x)
q3 (x)
= a1 (x)
= b2 (x)
= c3 (x)
...
(246)
We wish to show that the sequence {q} converges on the entire interval [a, b].
Given > 0, take M large enough so that there exist values xk with
k M such that |x xk | () for every point x of the interval, where ()
is the in our definition of equicontinuity. Now choose N = N () so that for
m, n > N
|qm (xk ) qn (xk )| < , k = 1, 2, . . . , M.
(247)
By equicontinuity, we have, for some k M :
|qm (x) qm (xk )| < ,
|qn (x) qn (xk )| < .
(248)
(249)
35
(251)
(252)
1
> 0.
1
(254)
Note also that, just as in the principal axis problem, additional eigenvalues (if
any exist) can be found by repeating the procedure, restricting to functions
orthogonal to the first one. If k(x, y) is degenerate, there can only be a finite
number of them, as the reader may demonstrate. This completes the proof
of the theorem stated at the beginning of the section.
Well conclude this section with some further properties of symmetric kernels. Suppose that we have found all of the positive and negative eigenvalues
and ordered them by absolute value:
|1 | |2 | . . .
(255)
(256)
They form an orthonormal set (e.g., if two independent eigenfunctions corresponding to the same eigenvalue are not orthogonal, we use the GramSchmidt procedure to obtain orthogonal functions).
We now note that if there are only a finite number of eigenvalues, then
the kernel k(x, y) must be degenerate:
k(x, y) =
n
X
i (x)i (y)
i=1
(257)
n
X
i (x)i (y)
i=1
(258)
J (, ) =
Z Z
k 0 (x, y)(x)(y)dxdy.
(259)
The maximum (and minimum) of this form is zero, since the eigenvalues of
P
i (y)
k(x, y) equal eigenvalues of ni=1 i (x)
. Hence k 0 (x, y) = 0.
i
We also have the following expansion theorem for integral transforms
with a symmetric kernel.
Theorem: Every continuous function g(x) that is an integral transform with
symmetric kernel k(x, y) of a piecewise continuous function f (y),
g(x) =
36
(260)
g(x) =
gi i (x),
(261)
i=1
where gi = hi |gi.
We notice that for series of the form:
k(x, y) =
X
i (x)i (y)
i=1
(262)
X
i (x) Z
g(x) =
i=1
i (y)f (y)dy
gi i (x),
(263)
i=1
Z b
(264)
Z b
(265)
Assuming f (y) is at least piecewise continuous (hence, f g must be continuous), the expansion theorem tells us that f (x) g(x) may be expanded
in the absolutely convergent series:
f (x) g(x) =
ai i (x),
(266)
i=1
where
ai = hi |f gi
=
=
=
Z Z
Z
f (y)dy
hi |f i.
i
37
hi |f i =
(268)
hi |gi.
i
(269)
i (x)
i=1
hi |gi
.
i
(270)
6.1
Z b
R(x, y; )g(y)dy.
(271)
Then
R(x, y; ) =
X
i (y)i (x)
i=1
1 D(x, y; )
D()
(272)
(273)
(274)
1
g.
1 K
(275)
38
(276)
kKk max
2
k(x, y)(x)(y)dxdy
< 1.
(277)
Z Z
(278)
The reader is invited to compare this notion with the condition for convergence of the Neumann series in Whittaker and Watson:
|(b a)| max
|k(x, y)| < 1.
x,y
6.2
(279)
Example
f (x) = sin x +
Z 2
(280)
1 2
1
,
2 1 2 cos(x y) + 2
|| < 1.
(281)
hi |gi
i (x),
i=1 i
(282)
With some inspection, we realize that the constant 1/ 2 is an (normalized) eigenfunction. This is because the integral:
I0
Z 2
0
dx
1 2 cos(x y) + 2
(283)
is simply a constant, with no dependence on y. In order to find the corresponding eigenvalue, we must evaluate I0 . Since I0 is independent of y,
evaluate it at y = 0:
I0 =
Z 2
0
dx
.
1 2 cos x + 2
39
(284)
We turn this into a contour integral on the unit circle, letting z = eix .
Then dx = dz/iz and 2 cos x = z + z1 . This leads to:
I0 = i
z 2
dz
.
(1 + 2 )z +
(285)
i I
dz
.
(z )(z 1/)
(286)
Only the root at is inside the contour; we evaluate the residue at this pole,
and hence determine that
2
I0 =
.
(287)
1 2
ieiy I
dz
,
(288)
and the relevant pole is at eiy . We thence notice that we know a whole class
of integrals:
1 2 ieiy I
z n dz
= n einy ,
iy
iy
2
(z e )(z e /)
n 0.
(289)
dz
ieiy I
,
n
iy
z (z e )(z eiy /)
n 0.
(290)
1
1
The residue at pole z = eiy is 1
2 n einy . We need also the residue at z = 0.
It is coefficient A1 in the expansion:
X
eiy
=
Aj z j .
z n (z eiy )(z eiy /) j=
(291)
eiy
X
1
jn i(j+1)y
j+1
=
z
e
.
z n (z eiy )(z eiy /)
1 2 j=0
j+1
40
(292)
n
n
.
1 2
(293)
Thus,
2
n einy .
(294)
2
1
We summarize the result: The normalized eigenfunctions are n (x) =
einx
hn | sin2 xi
n (x)
n
n=1
"
#
2 20 (x) 2 (x) + 2 (x)
2
= sin x +
4
1
2
1
1
= sin2 x +
2
cos 2x .
2 1
f (x) = sin2 x +
(295)
Exercises
1. Given an abstract complex vector space (linear space), upon which we
have defined a scalar product (inner product):
ha|bi
(296)
(297)
hc|ci.
(298)
41
(299)
"
(300)
(a) In class we found (in an equivalent form) that the Fourier Transform of a Gaussian of mean zero was:
2 2
(y; 0, ) = 1 exp y .
N
2
2
"
(301)
exp
dx = 1.
2 2
2
(303)
Typically, the constants and 2 are real. However, we have encountered situations where they are complex. Determine the domain in
(, 2 ) for which this integral is valid. Try to do a careful and convincing demonstration of your answer.
6. In class we
consider the three-dimensional Fourier transform of er /r,
where r = x2 + y 2 + z 2 . What would the Fourier transform
in two
2 be
2
dimensions (i.e., in a two-dimensional space with r = x + y )?
42
V(t)
R1
Vc (t)
R2
X
i=1
43
i (x)i (y).
(306)
(307)
f (x) = e +
Z x
1+y
f (y)dy.
x
(308)
f (x) = x +
exy f (y)dy.
(309)
In particular, estimate f (1), using one, two, and three intervals (i.e.,
N = 1, N = 2, and N = 3). [Were only doing some low values so you
dont have to develop a lot of technology to do the computation, but
going to high enough N to get a glimpse at the convergence.]
15. Another method we discussed in section 3 is the extension to the
Laplace transform in Laplaces method for solving differential equations. Ill summarize here: We are given a differential equation of the
form:
n
X
(310)
k=0
F (s)esx ds,
(311)
n
X
k=0
n
X
ak s k
(312)
bk s k ,
(313)
k=0
Z
A
exp
V (s)
44
U (s0 ) 0
ds ,
V (s0 )
(314)
(315)
f (x) = sin x +
(316)
X
sin nx sin ny
, 0 x, y .
=
n
n=1
(317)
1
1 2
,
2 1 2 cos(x y) + 2
(318)
f (x) = e +
(319)
Z x
(y x)f (y)dy,
(320)
45
Z x
f (y)dy,
(321)
Rx
0
R
ex + 2x + x eyx f (y) dy.
0
V(t)
L
V0(t)
R
0<t<T
(322)
otherwise.
Make a sketch of V (t) for (a) 2RC > LC; (b) 2RC < LC; (c)
2RC = LC.
V (t) =
V
0
46
f (x) = e +
Z 2
xyf (y)dy.
(323)
(b)
f (x) =
(324)
f (x) = x +
Z 1
0
47
xyf (y)dy.
(325)
(a) Find the Neumann series solution, to order 2 . For what values
of do you expect the Neumann series to be convergent? If you
arent sure from what you have done so far, try doing the rest of
the problem and come back to this.
(b) Find the Fredholm series solution, to order 2 .
(c) This is a degenerate kernel, so find the solution according to our
method for degenerate kernels.
48