Math Physics Lecture Notes

MATH 611/612: Mathematical Physics
Lecture Notes
Alexei Rybkin
University of Alaska Fairbanks
Contents
Part 1.
Complex Analysis
Lecture 1. Complex Numbers. Functions of Complex Variables

1. Imaginary Unit
2. Properties of Complex Numbers
3. Complex Functions
11
11
11
13
Lecture 2. The Concept of Analytic Function. The Cauchy-Riemann Conditions 15

1. Analytic Functions
15
2. Cauchy-Riemann Conditions
16
3. The Cauchy Theorem
18
Lecture 3. The Cauchy Integral
1. Introductory Example
2. Cauchy Integral
Appendix. The Morera Theorem
23
23
23
26
Lecture 4. Complex Series

1. Numerical Series
2. Functional Series
3. Power Series
4. Taylor and Laurent Series
29
29
30
30
31
Lecture 5. Singularities of Analytic Functions and the Residue Theorem

1. Singularities
2. Residues
37
37
38
Lecture 6. Applications of the Residue Theorem to Evaluation of Some Definite

Integrals
43
2
1. Integrals of the type
R(cos , sin )d
43
0
f (x)dx
44

eix f (x)dx
46
Lecture 7. Applications of the Residue Theorem to Some Improper Integrals and

Integrals Involving Multivalued Functions
49
1. Indented Contours and Cauchy Principal Values
49
3
CONTENTS
x1 f (x)dx , 0 < < 1
50
Part 2.
Linear Spaces
53
Lecture 8. Vector Spaces

1. Basic Definitions
2. Bases
3. Coordinates
4. Subspaces
55
55
56
58
58
Lecture 9. Linear Operators

1. Linear Operator
2. Matrices
3. Matrix Representation of Linear Operators
4. Matrix Ring
5. Noncommutative Ring
59
59
59
61
63
64
Lecture 10. Inner Product. Selfadjoint and Unitary Operators

1. Inner Product
2. Adjoint and Selfadjoint Operators
3. Unitary Operators
67
67
70
71
Lecture 11. More about Selfadjoint and Unitary Operators. Change of Basis
1. Examples of Selfadjoint and Unitary Operators
2. Change of Basis
73
73
74
Lecture 12. Eigenvalues and Eigenvectors

1. Eigenvalues
2. Spectral Analysis
77
77
78
Part 3.
81
Hilbert Spaces
Lecture 13. Infinite dimensional spaces. Hilbert Spaces

1. Infinite Dimensional Spaces
2. Basis
3. Normed Spaces
4. Hilbert Spaces
83
83
83
83
85
Lecture 14. L2 spaces. Convergence in Normed Spaces. Basis and Coordinates

in Normed and Hilbert Spaces
1. L2 spaces
2. Convergence in normed spaces
3. Bases and Coordinates
87
87
89
90
Lecture 15. Fourier Series
91
Lecture 16. Some Applications of Fourier Series
95
CONTENTS
1. Signal Processing
2. Solving Some Linear ODEs
95
98
Lecture 17. Linear Operators in Hilbert Spaces

1. General Concepts
2. Selfadjoint and Unitary Operators
3. Spectrum
101
101
105
106
Lecture 18. Spectra of Selfadjoint Operators
109
Lecture 19. Generalized Functions. Dirac function.

1. Physical Motivation
2. Uniform Convergence
3. Weak Convergence
4. Generalized Derivatives
111
111
111
116
119
Lecture 20. Representations of the -function

1. Fourier Representation of the -function
2. Integral Representation of the -function
121
121
123
Lecture 21. The Sturm-Liouville Problem
127
Lecture 22. Legendre Polynomials

Appendix. Frobenius Theory
133
136
Lecture 23. Harmonic Oscillator
143
Lecture 24. The Fourier Transform

1. Fourier Series
2. Fourier Transform
147
147
149
Lecture 25. Properties of the Fourier Transform. The Fourier Operator

1. Properties
2. The Fourier operator
151
151
152
Lecture 26. The Fourier Integral Theorem (Fourier Representation Theorem)

1. The Fourier Integral Theorem
2. Generalization of the Fourier Transform (by a limiting procedure)
155
155
158
Lecture 27. Some Applications of the Fourier Transform

1. Fourier Representation
2. Applications of the Fourier Transform to Second Order Differential
Equations
3. Applications of the Fourier Transform to Higher Order Differential
Equations
159
159
Part 4.
165
Partial Differential Equations
Lecture 28. Wave Equations
160
162
167
CONTENTS
1. The Stretched String

2. The Method of Eigenfunction Expansions (Spectral Method)
Lecture 29. Continuous Spectrum of a Linear Operator
2. Continuous spectrum of selfadjoint operators
167
169
173
173
175
Lecture 30. The Method of Eigenfunction Expansion in the Case of Continuous

Spectrum
179
1. The Schrodinger Operator
179
2. The Wave Equation on the Whole Line
180
3. Propagation of Waves
183
Appendix: Derivation of the dAlembert Formula
187
Lecture 31. The Heat Equation on the Line
189
Lecture 32. Nonhomogeneous Wave and Heat Equations on the Whole Line
1. Nonhomogeneous Heat Equation
2. Nonhomogeneous Wave Equation
Appendix: The Method of Variation of Parameters
195
195
197
203
Lecture 33. Wave and Heat Equations in Nonhomogeneous Media
207
Lecture 34. L2 -spaces on Domains in Multidimensional Spaces
211
Lecture 35. The Laplace Equation in R2 . Harmonic Functions

217
Appendix. Series Representation of the Solution to the Dirichlet Problem for
the Laplace Equation
220
Lecture 36. Uniqueness of the Solution of the Laplace Equation. Laplace
Equation in a Rectangular Domain
1. Well-Posedness of the Laplace Equation
2. Dirichlet Problem in a Rectangle
221
221
222
Lecture 37. The Principle of Conformal Mapping for the Laplace Equation.
Laplace Equation in the Upper Half Plane
Appendix. Change of Coordinate System and the Chain Rule
225
233
Lecture 38. The Spectrum of the Laplace Operator on Some Simple Domains
1. The Spectrum of on a Rectangle
2. The Spectrum of on a Disk
235
235
238
Lecture 39. Nonhomogeneous Laplace Equation
241
Lecture 40. Wave and Heat Equation in Dimension Higher Than One
1. Wave Equation
2. Heat Equation
3. Further Discussions
247
247
248
249
CONTENTS
Part 5.
Greens Function
255
Lecture 41. Introduction to Integral Operators
257
Lecture 42. The Greens Function of u00 + p(x)u0 + q(x)u = f

1. The Greens Function
2. Application to a General Order 2 Linear Differential Operator
261
261
262
Index
267
Part 1
Complex Analysis
LECTURE 1
Complex Numbers. Functions of Complex Variables

The set of real numbers R is too poor to be happy with. Modern science cannot
do without a larger set of numbers, the so-called complex numbers C.
Let us briefly recall some basic notions related to complex numbers.
1. Imaginary Unit
The imaginary unit i is defined as
i2 = 1.
As we all know there is no real number with such a property but we are going to treat
it as a number!
2. Properties of Complex Numbers
A complex number z is defined as
z = x + iy
x, y R
where x is the real part of z also denoted as

x = Re z
and y is the imaginary part of z
y = Im z.
The complex conjugate of z, commonly denoted in math by z, is defined by
z = x iy.
The absolute value (modulus) |z| of z is defined by
p
|z| = x2 + y 2 .
Note that |z| is a non-negative real number.
The sum of two complex numbers z1 , z2 is a complex number z defined as follows
z = z1 + z2 = (x1 + x2 ) + i(y1 + y2 )
that is,
Re(z1 + z2 ) = Re z1 + Re z2
Im(z1 + z2 ) = Im z1 + Im z2 .
11
12
1. COMPLEX NUMBERS. FUNCTIONS OF COMPLEX VARIABLES
The product z of z1 , z2 is defined as

z = z1 z2 = (x1 + iy1 )(x2 + iy2 )
= (x1 x2 y1 y2 ) + i(x1 y2 + x2 y1 )
Observe that

z = x + iy
1)
z = x + iy
R
2) zz = |z|2 .
Next, the division of z1 and z2 is defined by
z=
z1 z2
z1 z2
z1
:=
=
.
z2
z2 z2
|z2 |2
As an exercise check to see

1) z1 z2 = z1 z2
( hence z n = z n )

z1
z1
2)
=
z2
z2
3) z1 + z2 = z1 + z2
z
|z|2
The set of all complex numbers is denoted by C. So
4) 1/z =
C = {z : z = x + iy , x, y R}.
It is convenient to place complex numbers on a plane (called the complex plane).
y
z = x + iy = (x, y)
|z|
x
z = x iy
The argument (phase, polar angle) of a complex number z is the angle between
the x-axis and the vector (x, y). The standard notation for is
= arg z.
3. COMPLEX FUNCTIONS
13
It follows from elementary trigonometry that

z = |z|(cos + i sin ).
Note that
| cos + i sin | = 1
and
z n = |z|n (cos n + i sin n) ,
nN
z1 z2 = |z1 ||z2 | (cos(1 + 2 ) + i sin(1 + 2 ))

z1
|z1 |
=
(cos(1 2 ) + i sin(1 2 )) .
z2
|z2 |
Less trivial is the famous Euler formula
cos + i sin = ei .
Using this formula one gets
z = |z|ei ,
the so-called exponential representation of a complex number. One easily verifies
1) z1 z2 = |z1 ||z2 |ei(1 +2 )
2) z1 /z2 = |z1 |/|z2 |ei(1 2 )
3) z = |z|ei
4) |ei | = 1
5) z n = |z|n ein
If it takes you more than a split second to understand these properties I urge you to
prove them all.
3. Complex Functions
We are now ready to look at complex functions. When considering a domain D, it
should refer to an open set in the complex plane, think of it as an ink blot, a spill
on your carpet.
Definition 1.1. Given a domain D C, a complex valued function f is any
function from D to C of two real variables x, y.
Given z D where z = x + iy, we can always write
f (z) = u(x, y) + iv(x, y)
where u = Re f and v = Im f .
14
1. COMPLEX NUMBERS. FUNCTIONS OF COMPLEX VARIABLES
Example 1.2. Let D = C and let z = x + iy. The following two maps are examples
of complex valued functions.
f (z) = x + iy = z,
g(z) = x2 + iy 2 .
Although both are looking fairly simple, they are profoundly different as we will see next
time.
LECTURE 2
The Concept of Analytic Function. The Cauchy-Riemann

Conditions
1. Analytic Functions
The concept of an analytic function is truly central in math and science.
Definition 2.1. Let f : D C be a complex function of a complex variable z.
The function f (z) is said to be differentiable at z0 D if the following limit (called the
derivative)
f (z0 + z) f (z0 )
lim
f 0 (z0 )
(2.1)
z0
z
exists.
Definition 2.2. If f (z) is differentiable z0 D then f (z) is called analytic or
holomorphic in D.
Let us look again at Example 1.2. We claim that f (z) = z is analytic in C but that
g(z) = x2 + iy 2 is not.
Proof. Consider f . Let z0 C.
f (z0 + z) f (z0 )
z0 + z z0
=
=1
z
z
f (z0 + z) f (z0 )
lim
= 1.
z0
z
So f is differentiable at z0 and analytic in C.

Now for g, rewrite z = x + iy and note:
g(z0 + z) g(z0 )
(x + x)2 + i(y + y)2 x2 iy 2
=
z
x + iy
2xx + (x)2 + 2iyy + i(y)2
.
=
x + iy
case 1: z = x: z0 + z approaches z0 along a horizontal line.
g(z0 + z) g(z0 )
2xx + x2
=
= 2x + x
z
x
g(z0 + z) g(z0 )
lim
= lim 2x + x = 2x.
z0
x0
z
15
16
2. THE CONCEPT OF ANALYTIC FUNCTION. THE CAUCHY-RIEMANN CONDITIONS
case 2: z = iy: z0 + z approaches z0 along a vertical line.
g(z0 + z) g(z0 )
2iyy + iy 2
=
= 2y + y
z
iy
g(z0 + z) g(z0 )
lim
= lim 2y + y = 2y.
z0
y0
z
The limits are different in general so g is not analytic in C. Note that even if
points on the line y = x give the same limit, they do not constitute a domain,
i.e. an open set so g is not analytic.

Exercise 2.3. Show by definition that f (z) = x2 y 2 + 2ixy is analytic on C but
g(z) = x2 + y 2 + 2ixy is not.
We are going to accept the following math language: we write f H(D) if f is
analytic in D, i.e.
f H(D)
f is analytic in D.
As Exercise 2.3 shows, not every complex function f of two variables is analytic.
Moreover, if one picks up such a function arbitrarily, it will most likely not be analytic.
There must be some relations between Re f, Im f .
2. Cauchy-Riemann Conditions
Theorem 2.4. Let f = u + iv be differentiable at z = z0 . Then the partial derivatives ux , uy , vx , vy exist at z0 = x0 + iy0 and are related by
ux (x0 , y0 ) = vy (x0 , y0 )
uy (x0 , y0 ) = vx (x0 , y0 ).
Proof. Since f = u + iv is differentiable then the limit (2.1) exists and is independent of the way z 0. Let us consider two paths z = z0 + x and z = z0 + iy.
In the first case, we have
f (z0 + z) f (z0 )
z0
z
u(x0 + x, y0 ) + iv(x0 + x, y0 ) u(x0 , y0 ) iv(x0 , y0 )
= lim
x0
x
u(x0 + x, y0 ) u(x0 , y0 )
v(x0 + x, y0 ) v(x0 , y0 )
= lim
+ i lim
x0
x0
x
x
= ux (x0 , y0 ) + ivx (x0 , y0 )
ux , vx at z0 = x0 + iy0 .
f 0 (z0 ) = lim
(2.2)
2. CAUCHY-RIEMANN CONDITIONS
17
Similarly, in the second case

f (z0 + z) f (z0 )
f 0 (z0 ) = lim
z0
z
u(x0 , y0 + y) + iv(x0 , y0 + y) u(x0 , y0 ) iv(x0 , y0 )
= lim
y0
iy
u(x0 , y0 + y) u(x0 , y0 )
v(x0 , y0 + y) v(x0 , y0 )
= lim
+ lim
y0
y0
iy
y
1
= uy + vy = vy iuy
i
uy , vy at z0 = x0 + iy0 .
(2.3)
Comparing (2.2) and (2.3) one gets

ux + ivx = vy iuy
ux = vy
uy = vx
at z0 = x0 + iy0 .
Corollary 2.5. f = u + iv H(D)

z = x + iy D
ux = vy
uy = vx .
(2.4)
Conditions (2.4) are called Cauchy-Riemann.

Recall Example 1.2. For f (z) = z, we have u = x, v = y hence ux = 1 = vy
and uy = 0 = vx , so the Cauchy-Riemann conditions are satisfied; whereas for
g(z) = x2 + iy 2 , we have u = x2 , v = y 2 hence ux = 2x, vy = 2y and uy = 0 = vx , so
we find that only points on the line y = x satisfy the Cauchy-Riemann conditions, but
it is not an open set so g is not analytic.
Remark 2.6. The converse of Theorem 2.4 is also true if we assume that the partial
derivatives of u and v are continuous.
Now we are going to define some basic functions of a complex variable.
Example 2.7. f (z) = z n , n N is analytic on the whole C! Indeed, absolutely
similarly to what is done in Calculus I,
(z + z)n z n
= nz n1 ,
lim
z0
z
(although cumbersome to check with Cauchy-Riemann conditions), i.e. this limit exists
z C
Def 2.2
z n H(C).
Example 2.8 (ez ). Lets define ez as follows:

ez = ex+iy = ex eiy = ex (cos y + i sin y) .
(2.5)
Exercise 2.9. Show that ez defined by (2.5) is analytic z C.

If we assume that ez is analytic, we find the derivative along the real line for
example:
ez = ux + ivx = ex cos y + iex sin y = ez .
18
So we find that ez is a true exponential, i.e. the only function (up to a multiplicative constant) whose derivative coincide with itself.
Exercise 2.10.
Prove that the Cauchy-Riemann conditions are equivalent to

f
= 0.
(2.6)
z
Relation (2.6) actually means that a complex function f = u + iv of two real
variables (x, y) is analytic if f (x, y) contains z = x+iy and does not contain z = xiy.
For instance, f (z) = z 2 contains no z but g(z) = |z|2 = zz does contain z. Hence f (z)
is analytic but g(z) is not.
3. The Cauchy Theorem
Definition 2.11. Let be a curve on C and f (z) a function defined on . Then
we define the line integral of f (z) along the path as follows:
f (z)dz = (u + iv)(dx + idy)
= udx vdy + i vdx + udy.
Recall that line integrals are associated with work in physics.

Exercise 2.12. Show that
1. (f1 (z) + f2 (z))dz = f1 (z)dz + f2 (z)dz
2.
f (z)dz =
3.
f (z)dz, is a constant
f (z)dz =
1 2
f (z)dz +
1
f (z)dz
Hint: you are free to use the corresponding properties of line integrals of real functions (Calc III).
Definition 2.13. Let be a path (curve) starting at a point z = z0 and ending at
z = z1 .
z1
z0
We define () as the same curve but starting at z1 and ending at z0 .
3. THE CAUCHY THEOREM
19
z1
z0
f (z)dz =
f (z)dz.
()
Use the hint of Exercise 2.12.

Exercise 2.15.
State and prove the change of variables formula.
The following theorem will play a crucial role in our exposition.

Theorem 2.16. The following analog of the triangle inequality holds

f (z)dz |f (z)| |dz|.

p
Recall |dz| = dx2 + dy 2 , i.e. the arc length dS in Calc III. The triangle inequality
for complex numbers, |z1 + z2 | |z1 | + |z2 | is at the origin of the triangle inequality
for integrals (see proof below).
Proof. It immediately follows from Definition 2.11 that f (z)dz can also be
defined via Riemann partial sums. Namely, if we partition as on the figure
zk
zk1
z1
A = z0
z2
z1
zk
zk+1
zn = B
20
and form the Riemann sum

n
X
k=1
then
f (xk )(zk zk1 )

| {z }
zk
f (z)dz =
n
X
lim
f (zk )zk
maxk |zk |0
k=1

n

X

f (z)dz =
lim
f
(z
)z
k
k

maxk |zk |0

k=1
n
X
triangle inequality
lim
maxk |zk |0
|f (zk )| |zk |
|f (z)| |dz| .
k=1
ffi
Definition 2.17. Let C be a closed curve (contour), then
f (z)dz is called the

C
contour integral of f (z) along C counterclockwise.

Remark 2.18. It follows from Exercise 2.14 that
ffi
f (z)dz = f (z)dz.
C
Unless otherwise stated, we write
ffi
Definition 2.19. A domain D C is said to be simply connected if, roughly

speaking, it consists of one piece and has no holes in it.
Theorem 2.20 (Cauchys Theorem). Let D be a simply connected domain and
f H(D). Then
f (z)dz = 0 , C D.
C
Below are two examples of contours C in a simply connected domain D.
D
C
3. THE CAUCHY THEOREM
21
Exercise 2.21. Prove Theorem 2.20. Hint: use Definition 2.11, Greens formula,
and Cauchy-Riemann conditions.
Remark 2.22. The converse of the Cauchy theorem (known as Moreras theorem)
is also valid. If time permits we will prove it later.
LECTURE 3
The Cauchy Integral

1. Introductory Example
As we know, according to the Cauchy theorem, if f (z) is analytic on a simply
connected domain D then
f (z)dz = 0 ,
contour C D.
The Cauchy theorem fails if Int C contains at least one point at which our function is
not analytic. The following example shows it.
1
. This function is analytic z C \ {z0 }.
Example 3.1. Consider f (z) =
z z0
Let C = {z : |z z0 | = }. Consider the integral
dz
f (z)dz =
.
(3.1)
|zz0 |= z z0
C
Let us compute this integral explicitly using a typical complex variable technique. First
of all, C = {z : |z z0 | = } is a circle in C of radius , centered at z0 . This means
that z z0 = ei where 0 < 2 and by setting = z z0 in (3.1), we have
2 dei 2 i
i e d
dz
d
=
=
=
i
e
|zz0 |= z z0
0
||=
0
ei
2
=i
d = 2i.
Not zero!
0
Actually, we got a formula to be used in the future
dz
= 2i.
|zz0 |= z z0
(3.2)
2. Cauchy Integral
Lets now consider the function
that
f (z)
where f (z) H(D)1 and z0 D. It is clear
z z0
f (z)
/ H(D) but belongs to H(D \ {z0 }). Set up the integral

z z0
1
f (z)
dz , where C is any contour: z0 Int C.
2i C z z0
1recall
that f (z) H(D) means that f (z) is analytic on some domain D.

23
24
3. THE CAUCHY INTEGRAL
This integral is called the Cauchy integral of f (z) along a contour C and is one of the
most important objects in mathematics and theoretical physics.
Theorem 3.2 (Cauchy Formula). If f (z) H(D) and D is a simply connected
domain, then
1
f (z)
dz = f (z0 ),
(3.3)
2i C z z0
for any contour C D and z0 Int C.
This theorem is fundamental and its proof is very typical in complex analysis.
Before we present it we need to discuss one lemma.
Lemma 3.3. If g(z) H(D), where D is the domain between two contours C, C 0 .
Then
g(z)dz =
g(z)dz.
C0
Proof of lemma. Consider a new contour = C 1 (C 0 ) 2 which encloses

the domain D cut along 1 (or 2 ).
C0
C
C0
1 2
C
By condition g(z) H(D) and then by the Cauchy Theorem,
g(z)dz = 0.
(3.4)
On the other hand,
By Exercise 2.14, 2 = 1 and
+
C
(3.5)
. Combining (3.4) and (3.5) we have
g(z)dz
C
+
C0
g(z)dz = 0
C0
and the lemma is proven.

Now we are ready to prove the Cauchy Formula.
2. CAUCHY INTEGRAL
25
f (z)
f (z)
H(D \ {z0 }). Apply Lemma 3.3 to
with C = C, C 0 =
z z0
z z0
{z : |z z0 | = }. We have
f (z)
1
f (z)
1
dz =
dz
2i C z z0
2i |zz0 |= z z0
1
1
f (z) f (z0 )
dz
=
dz + f (z0 )
2i |zz0 |=
z z0
2i |zz0 |= z z0
|
{z
}
Proof.
= f (z0 ) +
1
2i
=1 by (3.2)
|zz0 |=
f (z) f (z0 )
dz
z z0
(3.6)
Let us now evaluate the last integral in (3.6). By Theorem 2.16, we have

1
f (z) f (z0 )
1
f
(z)
f
(z
)
0

|dz|
dz
2i
z z0
2 |zz0 |= z z0
C
1
|dz|
max |f (z) f (z0 )|
.
|zz0 |=
2 |zz0 |= |z z0 |
|
{z
}
1
= 2
|zz0 |=
1 1
|dz|= 2
2=1
So we got

1
f
(z)
f
(z
)
0
max |f (z) f (z0 )|.

dz
(3.7)
|zz0 |=
2i
z z0
C
But is arbitrary and we can make it as small as we want. Since f (z) is continuous,
|f (z) f (z0 )| is small if is small and it follows now from (3.7) that
1
f (z) f (z0 )
lim
dz = 0
(3.8)
0 2i |zz |=
z z0
0
Let now 0 in (3.6). We have
f (z)
f (z) f (z0 )
1
1
dz = lim
dz
f (z0 )
lim
+ lim
0
0 2i |zz |= z z0
0 2i |zz |=
| {z }
z z0
0
0
|
{z
}
|
{z
}
is independent of
is independent of
=0 by (3.8)
and we finally arrive at (3.3).
The Cauchy formula implies

Theorem 3.4. (without proof )
f
(n)
n!
(z0 ) =
2i
f (z)
dz.
(z z0 )n+1
Corollary 3.5 (The Liouville Theorem). If f (z) H(C) and bounded2, then
f (z) = const,
2A
function f (z) is called bounded on D if
z C.
M > 0 : |f (z)| M
z D.
26
3. THE CAUCHY INTEGRAL
Proof. Apply Theorem 3.4 to f (z) with n = 1, C = {z : |z z0 | = R}. We have
1
f (z)
0
f (z0 ) =
dz , z0 C.
(3.9)
2i |zz0 |=R (z z0 )2
(3.9) implies

|f (z)|
1
f (z)
1

|f (z0 )| =
dz
|dz| (By Theorem 2.16).

2
2 |zz0 |=R (z z0 )
2 |zz0 |=R |z z0 |2
0
By condition |f (z)| M and we get
M
M
|dz|
M 1
0
=
2
R
=
.
|f (z0 )|
2 |zz0 |=R |z z0 |2
R
2 R62
R is an arbitrary number. Make R . We have
lim |f 0 (z0 )| = 0
f 0 (z0 ) = 0 z0
f (z0 ) = const.
Done.
Appendix. The Morera Theorem

This theorem is the converse of the Cauchy theorem.
We prove first
Theorem 3.6. Let f (z) be defined and continuous on a simply connected domain
D. If contour C D
f (z)dz = 0
C
then
F (z)
f () d H(D)
z0
and
F 0 (z) = f (z).
Proof. Consider
F (z + z) F (z)
1
=
z
z
1
=
z
z+z
f () d
z0
z+z
f () d
z0
f () d.
z
(3.10)
APPENDIX. THE MORERA THEOREM
27
Note that all the integrals above are independent of specific curves and defined only
by their endpoints. It follows from (3.10) that

z+z
z+z
F (z + z) F (z)

1

d
f (z) =
f () d f (z)

z
|z| z
| z {z }

=z

z+z

1
=
(f () f (z)) d
|z| z
z+z
Theorem 2.16
1
|f () f (z)| |d|
|z| z
z+z
1
max |f () f (z)|
|d|
|z| [z,z+z]
z
| {z }
=|z|
max
|f () f (z)| .
(3.11)
[z,z+z]
Since f (z) is continuous on D

lim
max
z0 [z,z+z]
and (3.11) implies

F (z + z) F (z)
f (z) = 0
lim
0
z
|f () f (z)| = 0
F (z + z) F (z)
= f (z).
z0
z
lim
So f H(D) and F 0 (z) = f (z).
Theorem 3.7 (Moreras Theorem). Let f (z) be continuous on a simply connected

domain D. If contour C D
f (z)dz = 0
C
then f H(D).
Proof. By Theorem 3.6,
f ()d H(D) ,
F (z) =
z0 D
z0
and F 0 (z) = f (z).

By Theorem 3.4, F 0 (z) is analytic if F (z) is analytic. So f (z) is analytic.
LECTURE 4
Complex Series
Series (real or complex) are the main ingredient of not only mathematics but also
physics. Its just one of the very few ways to get to a numerical answer. So, from now
on we are going to deal with series on a regular basis.
1. Numerical Series
Definition 4.1. A sequence of complex numbers
{zk }
k=1 = {z1 , z2 , , zn , } is said to converge to
some z, if for any small > 0
z1
z2
|z zn | <
zn
for sufficiently large n. We write
z3
lim zn = z or zn z, n .
This means that starting from some n, all numbers {zn , zn+1 , } get in a given
disk centered at z.
X
X
Definition 4.2. A formal sum
zn =
zn of numbers {zn }
n=1 is called a
n=1
n1
series.
Definition 4.3. A series
zn is called convergent if a complex number S such
n1
that
Sn S , n ,
P
where Sn = z1 + z2 + X
+ zn = nk=1 zk is a partial sum of the series.
In this case we write
zn = S.
n1
Definition 4.4. A series which is not convergent is called divergent.

X
X
Proposition 4.5. A series
zn is convergent its tail
zk 0, n .
n1
Proof.
1) . S =
kn
zn . For S Sn we have
n1
S Sn =
zn
n
X
n1
k=1
29
zk =
X
k=n+1
zk .
(4.1)
30
4. COMPLEX SERIES
By condition, Sn S. Hence |S Sn | 0, n . So we get

X

|S Sn | =
zk 0.

k=n+1
2) . Read (4.1) backwards.

QED1
zn is said to be absolutely convergent if
n1
|zn | is
n1
convergent.
Actually there is no difference between real and complex series.
2. Functional Series
In applications it is a typical situation when every term of a series is a function of
some variable. This variable can easily be complex.
P
Definition 4.7. A formal sum
fn (z) is called a functional series.
In the future were going to treat functional series with pretty much general {fn (z)}.
But at this point we concentrate on power series.
3. Power Series
Definition 4.8. A power series is
X
n=
an (z z0 )n =
an (z z0 )n +
n1
an (z z0 )n
(4.2)
n0
where {an }
n= = { , an , , a1 , a0 , a1 , , an , } is a sequence of complex
numbers, z0 C.
Note that in Calc II we only consider power series with non-negative powers.
Definition 4.9. The domain of convergence D of a functional series is
X
D = {z C :
fn (z) is convergent }.
Theorem 4.10. The domain of convergence of a power
series is always an annulus.
Proof. Try to prove it yourself.
1QED
means what was to be demonstrated (Latin).
z0
4. TAYLOR AND LAURENT SERIES
31
A good example of a power series is

X
zn = 1 + z + z2 + + zn +
(4.3)
n0
The domain D = {z : |z| < 1} which is a disk of radius 1. Let us compute (4.3).
Sn (z) = 1 + z + z 2 + + z n =
1 z n+1
.
1z
If z D then |z| < 1 and limn z n+1 = 0 limn Sn (z) =

X
zn =
n0
1
1z
and hence
1
.
1z
(4.4)
As we can see, (4.3) represents a function which is clearly analytic in D. In fact,

more generally
Theorem 4.11. Every power series
X
an z n
n0
represents an analytic function on its domain of convergence.

The converse is also true as we will see in
4. Taylor and Laurent Series
Theorem 4.12 (The Taylor Theorem). If f (z) H(D), D = {z : |z z0 | < R},
then
X
f (z) =
an (z z0 )n , z D.
(4.5)
n0
where

1 dn

f
(z)
an =
.

n! dz n
z=z0
Proof. Let z D and consider C = {z : |z z0 | < }

where is chosen so that z Int C .
By the Cauchy formula
1
f ()d
f (z) =
(4.6)
2i C z
1
. We have
Consider z
1
1
1
=
.
0
z
z0 1 zz
z0
z0 R
(4.7)
32
4. COMPLEX SERIES

z z0
< 1, using formula (4.4) we get
Since, by construction,
z0
X z z0 n
1
=
0
z0
1 zz
z0
n0
and (4.7) can be continued

n
1 X z z0
1
=
.
z
z0 n0 z0
Plug now this expression in (4.6) and we get
1
f (z) X (z z0 )n
f (z) =
d
2i C z0 n0 ( z0 )n
X
f ()
1
(z z0 )n d.
=
2i C n0 ( z0 )n+1
(4.8)
Now switch the order of integration and summation. It is a very subtle point and not
that easy to prove. But in this case, its true! So (4.8) can be continued
X 1
f ()d
f (z) =
(z z0 )n .
n+1
2i
(
z
)
0
C
n0 |
{z
}
1 (n)
= n!
f (z0 ) (by Theorem 3.4)
1 (n)
f (z0 ).
n!
Definition 4.13. Series (4.5) is called the Taylor series of f (z).
So (4.5) is proven with an =
QED
Now we turn to Laurent series. Some more terminology.

Definition 4.14. A domain is called path-connected if for every two points z0 , z1
D, there is a path D connecting z0 , z1 .
That is
D is path-connected z0 , z1 D D : z0 , z1 .
Example:
is such a domain.
We associate with path-connected domains multiconnected contours.

stand this concept, just inspect the figure:
C
C0
D
C1
C2
33
To under-
C is a multiconnected contour, i.e. C is

made of three pieces:
C 0 (enclosing contour), oriented counterclockwise and two inner contours C1 , C2 oriented clockwise. The interior of C is the
shaded domain D.
Definition 4.15. We say that a contour C belongs to a domain D if Int C D.

Example:
D
D is a path-connected domain,
C is a multiconnected contour
(two-connected). We claim
that C belongs to D and write
C D since Int C D.
C
Theorem 4.16 (The Cauchy Formula). If f H(D), where D is a path-connected
domain, then
1
f (z)
dz = f (z0 )
2i C z z0
for any C D and z0 Int C.
Exercise 4.17. Prove Theorem 4.16. Hint: use the figure below, some arguments
of Lemma 3.3 and Theorem 3.2.
34
4. COMPLEX SERIES
Theorem 4.18 (The Laurent Theorem). If f H(D), where D is an annulus

D = {z : R1 < |z z0 | < R2 }, then
f (z) =
an (z z0 )n
z D
(4.9)
n=
1
an =
2i
f ()d
( z0 )n+1
R2
R1
where C is any contour in D encircling the disk |z z0 | < R1

and the point z as in the figure.
z0
z
Proof. can be done using the Taylor theorem.

Remark 4.19. Representation (4.9) is not unique. It becomes unique only for a
specified annulus. Check to see it for for yourself for the following series:
X
1
=
z n , 0 < |z| < 1 ,
z(1 z) n1
X
1
=
z n , 1 < |z| < .
z(1 z)
n2
Exercise 4.20. Prove Theorem 4.18.
ez =
X zn
n0
n!
and find its domain of convergence.

X
1
=
zn
1z
n0
and find its domain of convergence.

Example 4.23. Find the Taylor series of f (z) =
1
in |z 1| < 1 and the
z(z 2)
Laurent series in |z 1| > 1.

Solution.
1
f (z) =
2
1
1
z2 z
1
=
2
1
1
1 (z 1) 1 + (z 1)

.
Taylor in |z 1| < 1. (4.10) can be continued

(
)
X
1 X
Exercise 4.22
n
n
n
f (z)
=
(z 1)
(1) (z 1)
2 n0
n0
X
1X
=
(1 (1)n )(z 1)n =

(z 1)2n+1 .
2 n0
n0
(4.10)
35
Laurent in |z 1| > 1. (4.10) can be continued

1
1
1
f (z) =
+
2(z 1) 1 + (z 1)1 1 (z 1)1
)
(
X
X
1
1
4.22
(1)n (z 1)n +
(z 1)n
=
2 z 1 n0
n0
X
1X
=
{(1)n + 1} (z 1)n1 =
(z 1)2n1 .
2 n0
n0
1
(z = 0). Find the Taylor series f (z) in
(1 + z)3
|z| < 1 and the Laurent series in |z| > 1.
z
Exercise 4.25. Let f (z) = 2
(z = 2). Find the Taylor series f (z) in
z 1
|z 2| < 1 and the Laurent series in |z 2| > 3.
Exercise 4.24. Let f (z) =
LECTURE 5
Singularities of Analytic Functions and the Residue Theorem

1. Singularities
Definition 5.1. A point z0 is called an isolated singularity of an analytic function
f (z) if f (z) is analytic in some neighborhood of z0 with the exception of the point z0
itself.
If z0 is a singularity of f (z) then, by definition, f (z) H(D),
with some annulus D = {z : 0 < |z z0 | < R}.
Then, by the Laurent theorem (Theorem 4.18),
X
X
f (z) =
an (z z0 )n +
an (z z0 )n .
n1
R
z0
(5.1)
n0
{z
f1 (z)
{z
f2 (z)
f1 (z) is called the essential or improper part, and f2 (z) is called the correct or proper
part. We can classify the types of singularities.
Definition 5.2.
1) If f1 (z) = 0 then z0 is called a removable singularity.
N
X
an (z z0 )n (i.e. f1 (z) has only a finite number of terms) then
2) If f1 (z) =
n=1
z0 is called a pole of order N .

3) If f1 (z) is an infinite series then z0 is called an essential singular point.
Well come to see that removable singularities are harmless, poles are the usual
singularities we deal with, and essential singularities are nasty but do not usually
occur in physics. There are several alternate mathematical formulations for each type:
1) z0 is removable if an = 0 for n = 1, 2,
2) z0 is a pole if N N : aN 6= 0 and aN 1 = aN 2 = = 0.
3) z0 is an essential singularity if N N , n > N : an 6= 0.
Theorem 5.3.
1) z0 is a removable singularity limzz0 f (z) = a0 .
2) z0 is a pole of order N limzz0 (z z0 )N f (z) = aN .
3) z0 is an essential singularity in any neighborhood of z0 , f (z) takes on
all complex values except for maybe only one particular value.
Proof. Parts 1), 2) are clear. Part 3) is less trivial and we leave it without proof.
37
38
5. SINGULARITIES AND RESIDUE THEOREM
Example 5.4.
sin z
has a removable singularity at z0 = 0. Indeed we have that
1) f (z) =
z
X (1)n
sin z X (1)n 2n
sin z =
z 2n+1 and hence
=
z and no negative
(2n + 1)!
z
(2n + 1)!
n0
n0
powers of z.
sin z
2) f (z) = 3 has a pole of order 2 at z0 = 0. Indeed
z
X (1)n+1
sin z X (1)n 2n2
2
z
=
z
+
z 2n .
=
z3
(2n
+
1)!
(2n
+
3)!
n0
n0
3) f (z) = e1/z has an essential singularity at z0 = 0. Indeed e1/z =
X 1
z n , i.e.
n!
n0
e1/z has an infinite number of terms of negative powers in the Laurent series.
Exercise 5.5. Use Laurent series expansion to show that 0 is
ez 1
1) a removable singularity for f (z) =
;
z
z
e 1
;
2) a double pole for f (z) =
z3
3) an essential singularity for f (z) = e1/z 1.
Definition 5.6. A function f (z) is said to be analytic at infinity if the function
g(z) = f (1/z)
is analytic at z = 0.
It is clear that all the definitions and theorems apply also for z = . Restate all
of them as an exercise for z = .
2. Residues
Definition 5.7. Let z0 be a singular point and C be any contour enclosing z0 . Let
f H(Int C \ {z0 }) then
1
f (z)dz Res{f (z), z0 }
2i C
is called the residue of f (z) at z = z0 . Other notations include Res f , Res{f, z0 }.
z=z0
Remark 5.8. If f H(Int C \{z0 }) where z0 is a

singular point and C a contour enclosing z0 then one
can deform the contour C to C 0 with C 0 D Int C
where
D is an annulus centered at z0 . Then C f =
f and f H(D). So by Theorem 4.18, f admits a

C0
(4.9)
Laurent series expansion with 2ia1 = C 0 f (z)dz.
Hence we also have:
Res{f (z), z0 } = a1 .
C
D
0
C
z0
2. RESIDUES
39
Theorem 5.9. Let z0 be a pole of order N . Then

1
dN 1
N
Res{f (z), z0 } =
(z z0 ) f (z)
(N 1)! dz N 1
z=z0
(5.2)
This is exciting! No integral, instead: a derivative!

Proof. By Lemma 3.3,
1
Res{f (z), z0 } =
2i
1
f (z)dz =
2i
C
f (z)dz
(5.3)
C (z0 )
where C (z0 ) = {z : |z z0 | < }.

Since f (z) H({z : 0 < |z z0 | < }), then by Theorem 4.18
X
X
anN (z z0 )n
an (z z0 )n+N = (z z0 )N
f (z) = (z z0 )N
(5.4)
n0
nN
{z
g(z)
and g(z0 ) 6= 0 since aN 6= 0. Now (5.3) can be continued
1
(z z0 )N g(z)dz.
Res{f (z), z0 } =
2i C (z0 )
Recall the Cauchy formula (Theorem 3.4)
g
(n)
n!
(z0 ) =
2i
g(z)dz
.
(z z0 )n+1
So set N = n + 1, n = N 1 and then

g
(N 1)
(N 1)!
(z0 ) =
2i
g(z)dz
.
(z z0 )N
Hence

g (N 1) (z0 )
1
dN 1

Res{f (z), z0 } =
=
g(z)

(N 1)!
(N 1)! dz N 1
z=z
0
N 1

d
1
(5.4)
(z z0 )N f (z)
.
=
N
1
(N 1)! dz
z=z0
QED
Remark 5.10. The formula above does not apply to essential singularities.
Corollary 5.11. If z0 is a simple pole and f can be represented in some neighborhood of z0 as
(z)
f (z) =
(z)
with some analytic functions , such that (z0 ) 6= 0, (z0 ) = 0 then
Res{f, z0 } =
Proof. is left as an exercise.
(z0 )
.
0 (z0 )
40
5. SINGULARITIES AND RESIDUE THEOREM
ffi
dz
1
, consider f (z) = 2
and note that i
+1
z +1
C1 (i)
are isolated singularities. Furthermore, only i is inside our contour and we can write
1
g(z)
f (z) =
=
(z i)(z + i)
zi
Example 5.12. To evaluate
z2
1
where g(z) = z+i
and thus g(i) = 2i1 is finite and nonzero. So if we expand g(z) in
powers of z i, the first term is nonnegative. So there is only one negative term for f
around z i, i.e. i is a pole of order 1. Hence

1
dz
1
1
= 2i Res 2
= 2i(z i) 2
= 2i
= .
2
z=i z + 1
z + 1 z=i
z + i z=i
C1 (i) z + 1
e1/z dz, we can not use the residue formula since

Example 5.13. To evaluate
C1 (0)
0 is an essential singularity. Instead, we use Remark 5.8, and find Res{e1/z , 0} =

a1 = 1. So
e1/z dz = 2i 1 = 2i.
C1 (0)
Theorem 5.14 (The Residue Theorem). Let f (z) be analytic inside a contour C
except for a finite number of poles {zk }nk=1 = {z1 , z2 , , zn }. Then
n
X
Res{f (z), zk }
(5.5)
f (z)dz = 2i
C
k=1
Proof. Its enough to prove it for n = 2. By Lemma 3.3 we can deform C into
C
2
z1
C 0 = {z : |z z1 | = } 1 {z :
2|
|z z2 | = } 2 where
< |z1 z
(small
2
enough). Since 1 + 2 = 0 we get
f (z)dz =
f (z)dz+
f (z)dz.
z2
|zz1 |=
|zz2 |=
QED
Exercise
5.15. Evaluate the following integrals
ffi
dz
a)
.
2
z +4
|z2i|=1
ffi
cosh z
b)
dz.
z(z 2 + 1)
|z|=2
ffi
c)
ze1/z dz.
|z|=2
2. RESIDUES
ffi
d)
z2
C
z1
dz, where C = {(x, y) : x4 + y 4 = 4}.
+ iz + 2
41
LECTURE 6
Applications of the Residue Theorem to Evaluation of Some

Definite Integrals
In this lecture we are going to show that the residue theorem turns out to be very
helpful in computing some definite integrals which can hardly be computed by any
other means.
2
R(cos , sin )d
0
where R is a rational function, i.e. a quotient a polynomials (in R).

[0, 2] z = ei {z : |z| = 1}, i.e. z runs along the unit circle when
changes from 0 to 2.
|z| = 1
(
z

cos = z+z
z = ei = cos + i sin
2
i
zz
z = e = cos i sin
sin = 2i
0
2
Furthermore, z = 1/z since 1 = |z|2 = zz and dz = iei d = izd. So we have

1
dz
1
1
1
, cos =
d =
z+
, sin =
z
,
iz
2
z
2i
z
and making these substitutions we get
2
e
R(cos , sin )d =
R(z)dz
I :=
0
|z|=1
where R(z)
is a new rational function of z
Pn (z)
e
R(z)
=
, Pn , Qm are polynomials of order n, m respectively.
Qm (z)
e
The function R(z)
is analytic inside {z : |z| = 1} except for a finite number of poles
{z1 , , zN }, N m. By the Residue Theorem
N
n
o
X
e
I = 2i
Res R(z), zk .
k=1
Example 6.1. Compute
I=
0
d
1 + a cos
43
|a| < 1.
44
6. RESIDUE THEOREM - DEFINITE INTEGRALS
Put z = ei . We have
I=
|z|=1
2
=

1
z+
i
iz 1 + a z
dz
dz
az 2 + 2z + a
|z|=1
1
The poles of R(z) =
are solutions to az 2 + 2z + a = 0 z1,2 =
2 + 2z + a
az
r
1
1
1. Note that we can rewrite z1 in various forms:

a
a2
r
a
1
1
1 a2 1
z1 =
=
=
.
2
a
a
a
1 a2 + 1
From this last one, recall |a| < 1 so 1 a2 + 1 > 1 and |z1 | < 1. But z1 z2 = 1
1
|z2 | =
> 1 only z1 is inside of the contour |z| = 1. By the Residue Theorem
|z2 |

1
2
, z1 .
(6.1)
I = 2i Res
i
az 2 + 2z + a
By Theorem 5.9

Res
1
, z1
2
az + 2z + a

= lim
zz1
z z1
z z1
=
lim
az 2 + 2z + a zz1 a (z z1 ) (z z2 )
1
1
1
= q
=
.
a(z1 z2 )
2 1 a2
2a a12 1
So we have for (6.1)

I =6 2 6 i
2
2
1
=
.

6 i 6 2 1 a2
1 a2
f (x)dx
The computation of this integral is based on

Lemma 6.2. Let f (z) be analytic in C+ {z : Im z > 0} except for a finite number
M
of poles. If |f (z)| 1+ , with some M > 0, > 0 for sufficiently large |z|, then
|z|
lim
f (z)dz = 0 ,
+
CR
CR+ =
R
2. INTEGRALS OF THE TYPE
f (x)dx
45

|dz|
M

f (z)dz
|f (z)| |dz| M
= 1+
|dz|

1+
+
+ |z|
+
CR+

R
CR
CR
CR

M
M
Rd =
= 1+
0 , R .
R
R
0
QED
Theorem 6.3. Let f (z) satisfy the conditions of Lemma 6.2 and let f (z) have no
poles on R. If f (z)|zR = f (x) then

N
X
f (x)dx = 2i
Res{f (z), zk }.
(6.2)
k=1
Proof. By condition f (z) H(C+ \ {zk }N

k=1 ) and by the Residue Theorem
N
X
f (z)dz = 2i
Res{f (z), zk }
(6.3)
CR
k=1
and R is large enough so that all the poles of f (z) get
where CR =
inside of CR :
CR = CR+ (R, R)
z2
zN
z1
R
But
CR
R
R
+
CR
where CR+ as in Lemma 6.2. It follows then from (6.3) that
f (x)dx = 2i
R
N
X
Res{f (z), zk }
k=1
f (z)dz.
+
CR
Pass now to the limit as R in this equation and we get

N
X
f (x)dx = 2i
Res{f (z), zk } lim
k=1
|
and the theorem is proven.
Example 6.4. Prove that
f (z)dz
+
CR
{z
=0 by Lemma 6.2
}
dx
2
=
.
x4 + 1
2
46
M
4
Consider f (z) = z41+1 . Note that |f (z)| |z|
4 and the poles solve z + 1 = 0, i.e.

2k1
zk = ei/4
for k = 1, , 4. Only z1 , z2 are in C+ and since theyre simple zeros
4
of z + 1, theyre simple poles of f (z). Hence by Theorem 6.3 and Corollary 5.11

i 1
dx
1
1
1
=
.
= 2i
+
+
4
4z13 4z23
2 z13 z23
x + 1
But since zk4 = 1 then

follows that:

1
zk3
= zk , and since z1 = ei/4 =
i
i
dx
= (z1 z2 ) =
4
x +1
2
2
2i
1+i
,
2
z2 = e3i/4 =
1+i
,
2
it
2
=
.
2
eix f (x)dx
You can assume > 0 and f : R R. These integrals are important in harmonic
analysis (signal processing) where is the frequency and x is a spatial or temporal
variable.
Lemma 6.5 (Jordans Lemma). Let f (z) be analytic in C+ except for a finite number
of poles and
lim f (z) = 0 , Im z 0.
z
Then
lim
eiz f (z)dz = 0 , if > 0.
R
+
CR
We offer this lemma without proof. Note that CR+ is as before the arc of radius R
in the upper half plane. Note also that z is equivalent to |z| and this can
happen in many ways in the upper half plane:
Theorem 6.6. Let f (z) be subject to the conditions of the Jordan Lemma and have
no poles on R. Then if > 0

N
X

ix
e f (x)dx = 2i
Res eiz f (z), zk .
k=1
Proof. can be done in the very same manner as Theorem 6.3. Do it!
Example 6.7. Compute

cos x
dx
x 2 + a2
>0 ,
a > 0.
eix f (x)dx

a
Show that the integral is

Note that

e
a
47

cos x
eix
dx
=
Re
dx =: Re(I).
2
2
2
2
x + a
x + a
1
+
But f (z) = z2 +a
2 is analytic in C \{ia}, where z1 = ia is a simple pole, and lim f (z) =
z
0. So by Theorem 6.6 and Corollary 5.11

1
= ea .
I = 2i Res eiz f (z), ia = 2iea
2ia
a

cos x
dx = Re(I) = ea .
So
2
2
a
x + a
Observe the following:
when a , the value decays, which makes sense.
when , the integral converges to zero also, but this is not intuitive; it is
caused by lots of cancellations
of high frequencies.

because
1
when 0, we verify that x2 +a2 dx = a1 arctan xa = a . X
when a 0, we verify that cosx2x dx diverges.

Remember that a real integral returns a real value!
Exercise 6.8. Evaluate the following integrals
2
d
, || < 1.
a)
1
+
cos
0
2
cos 3
b)
d.
0 5 4 cos
d
c)
2 .
0 21 + sin
d
d)
.
2
(a
+
b
sin
)
0
2
e)
cosn d.
0
Exercise 6.9. Evaluate the integrals

dx
a)
.
4
4
x
+
a

x2
b)
dx.
6
0 x + 1
dx
c)
.
2
(x + 1)3
cos kx dx
.
d)
(x a)2 + b2
x sin x
e)
dx.
x2 + 1
0
48
f)
(x2
cos x dx
.
+ a2 )(x2 + b2 )
LECTURE 7
Applications of the Residue Theorem to Some Improper

Integrals and Integrals Involving Multivalued Functions
1. Indented Contours and Cauchy Principal Values

sin x
dx = , > 0. Lets prove it!
Example 7.1. I =
x
2
0
sin x
Note that I is improper on both sides but it should be ok since lim
= ; yet
x0
x
we have to find something to do about it.
sin x
is even
Since
x
1
I=
2
1
sin x
dx = Im
x
2
eix
dx.
x
(7.1)
1
has a pole on the real axis. But it is
z
not crucial. First of all, we have to agree upon how we understand (7.1):
Theorem 6.6 does not apply since f (z) =
eix
dx lim
0
x
R
+
R
eix
dx.
x
(7.2)
So we understand it as the principal value by Cauchy. Consider the contour C:

CR+
C+
0=
C
eiz
dz =
z
This is an indented contour: were careful

to drive around the pothole.
eiz
Since
H(D), by the Cauchy theorem
z
R

+
R
R
eix
dx +
x
+
CR
eiz
eiz
dz + y
dz.
z
C+ z
Note that the 0 on the LHS is independent of R, so they can run freely away, R to
infinity and to 0.
49
50
7. RESIDUE THEOREM - IMPROPER INTEGRALS AND MULTIVALUED FUNCTIONS
So it follows from this formula that1

R ix
e
eiz
eiz
lim
+
dx = lim
dz + lim x
dz
0
0 C z
R C + z
x
R
R
|
{z
}
=0 , by Jordans lemma
iei
eiz
e
= lim x
dz = lim
iei d
i
0 C z
0 0
e

(i cos sin )
lim e(i cos sin ) id = i.
e
id =
= lim
0
0 0
0 |
{z
}
=1
Here we switched lim = lim. I swear that in this case it can be justified!
1
So I = Im i = . Done!
2
2
Note the following:
at first, it might look disturbing that no appears in the answer; however, the
result is correct and the absence of is due to the fact that the oscillations
always add up to the same
number;
if = 0 then we have 0 = 0;
if < 0, rewrite sin x = sin()x then the integral is /2;
so a more general result is
2 sin x
dx = sgn .
0
x

x1 f (x)dx , 0 < < 1
0
Before we present the main result of this section, we must introduce a new concept:
branch cuts.
2
Whereas z 2 is analytic, we would like
2z to be the inverse of z , but how can we
define it? There is more to it than just
z) = z because there are two choices. Indeed
(
for z = ei , we naturally think of z = ei/2 . But since we also have z = ei(+2)
note that then z = ei(/2+) = ei/2 .
So
is a different kind of function, but we can figure
it out by choosing z = ei/2 as the principal value.

0 < 2
Is it analytic? No, since it is not continuous when going
around full circle: there is a jump discontinuity; it is like

z=1
youre walking around in your neighborhood then come
z = ei0
back to your house, only its not your house anymore.
z= e2i
You can make it analytic but not on the whole comz = 1
plex plane. You make a cut to separate 0 from 2 this
way, you cant get back: you can cut along [0, ). Note
1x
means that integration is done counterclockwise.
x1 f (x)dx , 0 < < 1
51
that each point on [0, ) is a singularity, but not isolated, so we cant do residue
calculations on it.
A similar issue arises with the logarithmic function.
Note also that the cut along R+ is not the only one possible; any ray is ok since it
precludes any contour/neighborhood from going around the origin.
[, )
Theorem 7.2. Let f (z) be analytic in C except for a finite number of poles off the
positive part of the real axis. Assume |f (z)| |z|M
+ for some > 0, 0 < < 1 and
|z| > r. Then
N
1

2i X
Res
z
f
(z),
z
f (x)dx =
.
k
1 e2i k=1
Note that f (x) = sinx x does not satisfy the conditions above since even though
1
on the real line, this is no longer true for z in the complex plane since sin z
|f (x)| |x|
is unbounded.
Proof. Note first that z 1 cannot be analytic on the whole C but it is analytic
on C \ R+ , R+ = [0, ). Make sure that you understand it!
Consider the following contour C:
CR
0
`+
The function (z) = z 1 f (z) is

analytic inside of C except for the
poles {zk }N
By the Residue
k=1 .
Theorem
N
X
(z)dz = 2i
Res{(z), zk }.
C
k=1
52
7. RESIDUE THEOREM - IMPROPER INTEGRALS AND MULTIVALUED FUNCTIONS
On the other hand,
1
1
1
(z)dz =
z f (z)dz +
z f (z)dz +
z 1 f (z)dz .
z f (z)dz +
C
`
`
C
C
}
| + {z
} | R {z
| {z
}
= R x1 f (x)dx
=:I1
=:I2
But because f (x) is analytic on R+ , it returns the same value along `+ or ` , and
along ` we also have that dz = dx for x from R to , but z 1 = |z|1 e2(1)i =
x1 e2(1)i . Hence the above becomes
R

1
(1)2i
(z)dz =
x f (x)dx + I1 + e
x1 f (x)dx + I2
C
= 1 e(1)2i
R
R
x1 f (x)dx + I1 + I2 .
(7.3)
For I1 we have
R1 Rd
|f (z)| |dz| max |f (z)|
|{z} zCR
0
CR
M
max |f (z)| R 2 2R + 0 , R .
|z|=R
R
lim I1 = 0.
|z|
|I1 |
For I2 in the same manner we get
|I2 | max |f (z)|

1 |dz| = 2
|z|=
max |f (z)|
|z|=
| {z }
0 , 0.
finite since no poles
But lim
0
R
lim I2 = 0.
x1 f (x)dx.
f (x)dx =
QED
x1
dx =
, 0 < < 1.
x+1
sin
0
1
1
Note that f (x) =
admits an analytic continuation f (z) =
with one
x+1
z+1
pole: -1. We also have = 1 > 0 since 0 < < 1. So by Theorem 7.2
1
2i
ei(1)
x
1
(1)
=
2i
dx =
x+1
1 e2i(1)
1 ee2i(1)
0
2i
= i(1)
=
=
.
i(1)
e
e
sin ( 1)
sin
Exercise 7.4. Show that ( 1 < < 3 )

x
(1 )
dx =
.
2
2
(x + 1)
4 cos
0
2
Example 7.3.
Part 2
Linear Spaces
LECTURE 8
Vector Spaces
The concept of a vector space is central in math physics and its going to be a part
of our math language.
Definition 8.1. A vector space E is a set of elements (also called vectors) equipped
with operations + and multiplication by a scalar subject to
1.
2.
3.
4.
5.
6.
7.
8.
9.
10.
X, Y E X + Y E
(closure under addition)
X +Y =Y +X
(commutative law)
(X + Y ) + Z = X + (Y + Z)
(associative law)
0 : X + 0 = X, X E
(existence of zero vector)
X E (X) : X + (X) = 0
(existence of additive inverse)
X E cX E (c is a scalar)
(closure under scalar multiplication)
a(bX) = ab(X) : a, b are scalars
(associative law)
(a + b)X = aX + bX
(distributive law with respect to multiplication)
a(X + Y ) = aX + aY
(distributive law with respect to addition)
1X =X
(invariance with respect to multiplication by unity)
Note, first of all that our usual 3 space (E 3 ) is a space whose elements are the usual
three component vectors with + defined by

x1
y1
x1 + y 1
X + Y = x2 + y 2 = x2 + y 2
x3
y3
x3 + y 3
and multiplication by scalars

x1
cx1
cX = c x2 = cx2
x3
cx3
Verify properties #1 - 10!
Consider now less simple examples.
Example 8.2. Let y 00 +py 0 +qy = 0 be a second order linear homogeneous differential
equation. If y1 , y2 are two solutions, y1 + y2 is a solution too and so is cy1 , cy2 , c is
an arbitrary constant. Operations + and multiplication by a scalar are usual + and .
One can easily verify that the set of all solutions to this differential equation forms a
linear space.
55
56
8. VECTOR SPACES
Example 8.3. Consider the set Pn of all polynomials of order not greater than
n. Its clear that if P1 and P2 are polynomials then P1 + P2 is a polynomial too. I.e,
P1 , P2 Pn P1 + P2 Pn where + is the usual addition. Next, if a is a scalar,
P Pn aP Pn . So Pn forms a linear space.
Example 8.4. Let C[0, 1] be the set of all continuous functions on [0, 1]. As in
Example 8.3, C[0, 1] is a linear space. (Check it!)
These examples show that a linear space is not a weird object.
Definition 8.5. A linear space is called real if all scalars in Def 8.1 are real numbers.
Definition 8.6. A linear space is called complex if all scalars in Def 8.1 are complex numbers.
2. Bases
Some more definitions.
Definition 8.7. Vectors X1 , X2 , . . . , Xn E are called linearly independent if the
equation
c1 X 1 + c2 X 2 + . . . + cn X n = 0
(c1 , . . . , cn are scalars)
holds only for c1 = c2 = . . . = cn = 0.

Otherwise, X1 , X2 , . . . , Xn are called linearly dependent.
Definition 8.8. The system of vectors {e1 , e2 , . . . , enS} E is said to be a basis in
E if {ek }nk=1 is linearly independent but {e1 , e2 , . . . , en } X is linearly dependent for
any X 6= 0 E.
Remark 8.9. Note that a basis is not unique.
Exercise 8.10. If {e1 , e2 , . . . , en } is a basis for E, show that {e1 , e1 +e2 , e3 , . . . , en }
is also a basis.
Definition 8.11. The number n of elements in a basis is called the dimension of
E and denoted n = dim E.

x1
Example 8.12. Consider a space E = x2 : x1 , x2 , x3 R (the space of 3
x3
columns). Consider the following systems of vectors {e1 , e2 , e3 }

1
0
0
e1 = 0 , e2 = 1 , e3 = 0.
0
0
1
We claim it is a basis in E. Indeed,
2. BASES
57

1
0
0
c1 e1 + c2 e2 + c3 e3 = 0 c1 0 + c2 1 + c3 0
0
0
1

c1
c2 = 0 c1 = c2 = c3 = 0.
c3
Hence {e1 , e2 , e3 } are linearly independent.
Now we need to make sure that {e1 , e2 , e3 , X} is linearly dependent with any X 6= 0.

x1
Indeed, let X = x2 with some x1 , x2 , x3 : x21 + x22 + x23 6= 0. We then have
x3
c1 + c4 x 1
c1 e1 + c2 e2 + c3 e3 + c4 X = 0 c2 + c4 x2 = 0
c3 + c4 x 3
which is equivalent to the system
c1 = c4 x1
c2 = c4 x2
c = c x
3
4 3
This system has infinitely many solutions since if we put c4 = t 6= 0 then at least
one of c1 , c2 , c3 is not 0 (x1 , x2 , x3 are not all zeros). Hence {e1 , e2 , e3 , X} is linearly
dependent and by definition, {e1 , e2 , e3 } is a basis in E.
Remark 8.13. The space E in Example 8.12 is actually our 3-space. Commonly,
E is denoted by R3 . So,

x1
x2 : x1 , x2 , x3 R R3
x
3
Example 8.14. Let Pn be the set of all polynomials of order n. We show that
dim Pn = n + 1. Consider {1, x, x2 , . . . , xn }. This system is linearly independent.
Indeed,
n
X
ck xk = 0 ck = 0, k = 0, 1, . . . , n.
k=0
(it means that a polynomial is identically 0 if and only if all of its coefficients are 0.)
We leave it as an exercise to prove that {1, x, x2 , . . . , xn } is a basis in Pn . Hence,
by definition dim Pn = n + 1.
Some more examples:
The set of solutions to an order 2 linear homogeneous differential equation is a
linear space of dimension 2.
58
8. VECTOR SPACES
Similarly, the set of solutions to an order 3 linear homogeneous differential

equation is a linear space of dimension 3.
There are infinite dimensional linear spaces. Function spaces, for example.
C[0, 1] = {f : [0, 1] R | f (x) is continuous on [0, 1]}
H(D) = {f : C C | f is analytic on D}
However, its not so easy to show that the dimension is infinite. We would
need to find an infinite, linearly independent set to form a basis. For example,
{1, x, x2 , . . . , xn , . . .} C[0, 1].
The Taylor expansion would be another candidate basis, however this approach is not valid for all f C[0, 1], only smooth functions. So, is there a
basis for C[0, 1]? Yes, but it belongs to functional analysis.
3. Coordinates
We know that if {e1 , e2 , . . . , en } is a basis in E, then {e1 , e2 , . . . , en , X} is always
linearly dependent. I.e., c1 , c2 , . . . , cn+1 not all zeroes such that
c1 e1 + c2 e2 + . . . + cn en + cn+1 X = 0
P
Note that cn+1 6= 0. Otherwise we had nk=1 ck ek = 0 that would contradict the fact
that {ek }nk=1 is a basis. It follows from (3) that
n
X
ck
X=
ek .
c
n+1
k=1
Coefficients xk ck /cn+1 , k = 1, 2, . . . , n are called the coordinates of a vector X in
a basis {e1 , e2 , . . . , en }.

1
Example 8.15. Let X R3 . Coordinates of X in the basis e1 = 0 , e2 =
0

0
0
1 , e3 = 0 are x1 , x2 , x3 . In physics they commonly use ~,~, ~k for e1 , e2 , e3 .
1
0
Example 8.16. Let E = P4 and {1, x, x2 , x3 , x4 } be a basis in P4 . If P4 (x) =
c0 + c1 x + c2 x2 + c3 x3 + c4 x4 is an arbitrary polynomial of order 4, then its coefficients
{c0 , c1 , . . . , c4 } are coordinates of P4 (x) in the basis {1, x, . . . , x4 }.
Theorem 8.17. Given a vector and a basis, the coordinates are uniquely determined.
4. Subspaces
Definition 8.18. If E1 , E2 are linear spaces and E1 E2 then E1 is called a
subspace of E2 .
Example 8.19.
1) R2 {0} is a subspace of R3 .
2) P3 is a subspace of P4 .
LECTURE 9
Linear Operators
1. Linear Operator
A linear operator is a fundamental object in math physics.
Definition 9.1. Let E1 , E2 be two linear spaces. A mapping A that maps E1 into
E2 (A : E1 E2 ) is called a linear operator if X, Y E1 and scalars , ,
A(X + Y ) = AX + AY.
In most of our cases, E1 = E2 = E and we then say that A acts in E. It means
that A sends every X E into a vector Y E.
Example 9.2. Let E = R2 and A be an operator acting by the following rule,

x1
x1
, where X =
.
AX =
x2
0
It is clear that for all X, Y ;

x1 + y 1
x1 + y1
x1
y
A(X + Y ) = A
=
=
+ 1 = AX + AY ;
x2 + y 2
0
0
0

x1
x1
x
AX = A
=
= 1 = AX.
x2
0
0
Hence, A is a linear operator in R2 . As you can see, A performs an orthogonal
projection of a vector in a 2-space on the x-axis.
Example 9.3. Let E = Pn . Define A by the formula
d
Ap(x) =
p(x) (p(x) Pn ) .
dx
A is clearly a linear operator. This operator is called the operator of differentiation
and will be playing a crucial role in our course.
2. Matrices
Definition 9.4. The following table of numbers
a11 a12 . . . a1n

a21 a22 . . . a2n
A
{aik }m,n
..
..
...
i=1,k=1
...
.
.
am1 am2 . . . amn
is called a m n matrix. If n = m, then the matrix is called square.
59
60
9. LINEAR OPERATORS
Definition 9.5. For any m n matrices A, B, and any C,

1)

a11 a12 . . . a1n
b11 b12 . . . b1n
a21 a22 . . . a2n b21 b22 . . . b2n
+ .
A+B =
..
.. . .
.
..
...
...
. ..
.
.
. ..
bm1 bm2 . . . bmn
am1 am2 . . . amn
a11 + b11 a12 + b12 . . . a1n + b1n

a21 + b21 a22 + b22 . . . a2n + b2n
.
=
..
..
..
..
.
.
.
.
am1 + bm1 am2 + bm2 . . . amn + bmn
That is, matrices add up element-by-element,
(A + B)ik = (A)ik + (B)ik
2)
a11 a12
a21 a22
A =
..
...
.
am1 am2
i.e.,

. . . a1n
a11 a12
. . . a2n a21 a22
= .
..
..
...
. ..
.
. . . amn
am1 am2
. . . a1n
. . . a2n
..
..
.
.
. . . amn
(A)ik = (A)ik
3)
a11
a21
A=
...
a12
a22
..
.
ik
a1n
b11 b12 . . . b1`
b21 b22 . . . b2`
a2n
.
,
B
=
..
.. . .
. ;
..
. ..
.
.
amn
bn1 bn2 . . . bn`

. . . a1n
b11 b12 . . . b1`
. . . a2n b21 b22 . . . b2`
..
..
.. . .
..
...
.
.
.
.
.
am1 am2 . . .
a11 a12
a21 a22
AB =
..
...
.
am1 am2 . . . amn
bn1 bn2 . . . bn`
n
n
n
X
X
X
ak1 b1k
a1k bk2 . . .
a1k bk`
k=1
k=1
k=1
n
n
n
X
X
X
b
a
b
.
.
.
a
b
2k
k1
2k
k2
2k
k`
= k=1
k=1
k=1
..
..
..
..
.
.
.
.
n
n
n
X
X
X
amk bk1
amk bk2 . . .
amk bk`
k=1
1(A)
...
...
..
.
k=1
means the ik th element of the matrix A
k=1
3. MATRIX REPRESENTATION OF LINEAR OPERATORS
61
i.e.,
(AB)ij =
n
X
aik bkj
k=1
Note, that we can mutliply two matrices only in the following case:
n
`
In general, m 6= `.
m
n
Some more
0 0
0 0
. .
.. ..
terminology:
... 0
. . . 0
=0
. . . ..
.
the zero matrix
0 0 ... 0
a11
a22
..
.
a
nn
1
1
..
is called a diagonal matrix
=I
the unit matrix
3. Matrix Representation of Linear Operators

Let A : E E and {ek }nk=1 be a basis in E. Since Aek E it can be represented
as
Aek =
n
X
aik ei , k = 1, 2, . . . , n,
(9.1)
i=1
where {aik }ni=1 are coordinates of the vector Aek in the basis {ek }nk=1 .
Coefficients {aik }ni,k=1 form a matrix. The thing is that this matrix represents the
operator A in the basis {ek }nk=1 . This means that if we know all {aik }ni,k=1 then we can
compute AX for any X E.
62
9. LINEAR OPERATORS
Indeed, let X =
n
X
xk ek . Then,
k=1
AX = A
n
X
x k ek =
k=1
n
by (9.1) X
xk
k=1
n
X
xk Aek
k=1
n
X
n
n
X
X
i=1
i=1
aik ei =
!
aik xk
ei
k=1
i.e.2
(AX)i =
n
X
aik xk ; i = 1, 2, . . . , n
(9.2)
k=1
(9.2) reads
n
X
a1k xk
(AX)1 =
k=1
(AX) =
a2k xk
2
k=1
ank xk
(AX)n =
k=1
x1
x2
n
So if X =
... is given by its coordinates in {ei }i=1 then,
xn
a11 a12
a21 a22
AX =
..
...
.
an1 an2

. . . a1n
x1
. . . a2n x2
. ,
. . . ..
. ..
. . . ann
xn
(9.3)
where the right hand side is understood as a product of two matrices.

Definition 9.6. The matrix on the right in (9.3) is called the matrix of an operator
A in a basis {ei }ni=1 .
So we can always identify an operator A with its matrix:
a11 a12 . . . a1n

a21 a22 . . . a2n
A=
.. . .
. in {e1 , e2 , . . . , en }.
...
. ..
.
an1 an2 . . . ann
2(X)
stands for the ith coordinate of X.
4. MATRIX RING
63
You can think of an operator as a set of houses. You can list these houses using
their street addresses. It would play the role of the matrix representation. Clearly
the set of houses is independent of the way you divide them into blocks. But once a
division is fixed then you identify every house with its street address.
Some examples:
Example
9.7. Consider the operator A as in Example 9.2. Choose a basis e1 =
1
0
, e2 =
.
0
1
Let us find the matrix of this operator in the basis {e1 , e2 }. We have
Ae1 = e2
Ae2 = 0
(9.4)
On the other hand, by (9.1) we have

Ae1 = a11 e1 + a21 e2
Ae2 = a12 e1 + a22 e2
(9.5)
Comparing (9.4) and (9.5), we get

a11 = 1 ,
a21 = 0 ,
a12 = 0 ,
a22 = 0
and finally

A=
1 0
0 0

1
0
in
,
.
0
1
Example 9.8. Let A be as in Example 9.3. Choose in Pn the following basis

{e1 , e2 , . . . , en }, ek = xk1 , k = 1, 2, . . . , n + 1. We have
Aek =
d k1
x
= (k 1)xk2 = (k 1)ek2
dx
k = 1, 2, . . . , n + 1.
On the other hand,

Aek =
n+1
X
i=1
aik ei = a1k e1 + . . . + ak1,k ek1 + . . . + an+1,k en+1

| {z }
| {z }
| {z }
=0
=(k1)ek1
ak1,k = k 1 . So,
... 0
. . . 0
. . ..
in {1, x, x2 , . . . , xn }
. .
. . . n
0 0 0 ... 0
aij = 0 except
0 1 0
0 0 2
. . .
A=
.. .. ..
0 0 0
4. Matrix Ring
The following theorem is very important.
=0
64
9. LINEAR OPERATORS
Theorem 9.9. Let A, B be linear operators in E and
a11 . . . a1n
b11 . . . b1n
e = ... . . . ... , B
e = ... . . . ...
A
an1 . . . ann
bn1 . . . bnn
be their matrix representations in {ei }ni=1 . Then
e+B
e , A
f =A
e
^
1) A
+B =A
g=A
eB
e
2) AB
i.e. when we add two operators their matrices add and when we multiply two operators their matrices multiply.
1) is clear. Show it!
Proof.
2)
A(Bek ) = A
n
X
bjk ej
(by (9.1))
j=1
n
X
bjk Aej =
j=1
n
X
j=1
n
X
n
X
i=1
j=1
bjk
n
X
ei =
n
n
X
X
i=1
(AB)ik =
n
X
aij bjk
!
aij bjk
ei
j=1
(by (9.1) again)
i=1
!
bjk aij
aij ei
{z
=(AB)ik
g=A
eB.
e
AB
}
QED
j=1
5. Noncommutative Ring
Note that multiplication of matrices, and hence operators, is not commutative.
d
Example 9.10. Let E = P1 = {a + bx | a, b R} and A = dx
, B(a + bx) = bx.
Then
AB(a + bx) =
bx = b
dx
AB 6= BA.
d
BA(a + bx) = B (a + bx) = Bb = 0

dx
Exercise 9.11. Let Mn be the set of all n n matrices with + and multiplication
by a scalar given in Def 9.5. Show that Mn is a linear space.
Exercise 9.12. Let A : E E be a linear operator. We define the kernel of A:

ker A := {X E | AX = 0}.
a) Show that ker A is a subspace of E.
b) Find ker A where A is as in Example 9.2.
c) Find ker A where A is as in Example 9.3.
5. NONCOMMUTATIVE RING
Exercise 9.13. Find the matrix of A : E E

a) E = R2 and A rotates each vector X E by 90 counterclockwise.
b) E = Cn and A : AX = X X E where C is a fixed number.
c) E = R3 and A projects each X E onto the xy plane.
Exercise 9.14. Give an example of an operator
a) A 6= 0 , A2 = 0
b) A 6= 0 , A2 6= 0 , A3 = 0.
65
LECTURE 10
Inner Product. Selfadjoint and Unitary Operators

1. Inner Product
The concept of inner (or scalar) product is central in math physics.
Definition 10.1. An inner (scalar) product hX, Y i of two vectors X, Y belonging to
a linear space E is a complex valued function of X, Y subject to the following conditions:
1) Symmetry : hX, Y i = hY, Xi
2) Linearity: hX + Y, Zi = hX, Zi + hY, Zi , C
3) Positivity: hX, Xi 0 and if hX, Xi = 0 X = 0.
Different books adopt different notation for the inner product: e.g. (X, Y ), X Y ,
hX|Y i. The last notation is commonly accepted in Quantum mechanics and called the
Dirac notation. We chose the mixture of (X, Y ) and hX|Y i.
Exercise 10.2. Consider

x1
C3 = x2
x
3

x1 , x2 , x3 C .

Verify that
1) hx, yi =
3
X
xi y i is an inner product on C3 ,
i=1
2) but hx, yi =
3
X
|xi yi |2 is not.
i=1
3) Find another inner product on C3 .

Exercise 10.3. Verify the following
1) hX, Y i = hX, Y i .
2) hZ, X + Y i = hZ, Xi + hZ, Y i .
3) Whereas in real spaces, you have full commutativity for the inner product, if
that were the case in complex spaces, then the positivity condition would fail.
Lemma 10.4. If hX, Y i = 0 for all Y E then X = 0.
Proof.
hX, Y i = 0 Y E
hX, Xi = 0
X = 0.
Corollary 10.5. If hX1 , Y i = hX2 , Y i for all Y E then X1 = X2 .

Proof. Use Definition 10.1 and condition 2) to get hX1 X2 , Y i = 0.
67
68
10. INNER PRODUCT. SELFADJOINT AND UNITARY OPERATORS
Definition 10.6. Vectors X, Y E are called orthogonal if hX, Y i = 0.

Definition 10.7. Let X E.
hX, Xi kXk
is called the norm of X.
(
0 ,
Definition 10.8. ik
1 ,
i 6= k
i=k
is called the Kronecker delta.
Definition 10.9. A basis {ek }nk=1 is called orthogonal and normed (or orthonormal) or ONB if
hei , ek i = 0 ,
i 6= k
kei k = 1 ,
i = 1, 2, , n.
Alternately, we can write: {ek }nk=1 is an ONB

Theorem 10.10. Let
{ek }nk=1
hei , ek i = ik .
be an ONB and X =
n
X
xk ek , Y =
n
X
yk ek . Then
k=1
k=1
hX, Y i =
n
X
xk y k .
k=1
I.e. for each basis, there is only one inner product.

Proof.
hX, Y i =
* n
X
xi ei ,
i=1
1)
n
X
1)
i=1
n
X
=
=
xi
+
2)
yk ek
=
+
yk ek , ei
2)
*
xi
n
X
ei ,
n
X
k=1
n
X
+
yk ek
k=1
xi
i=1
k=1
xi y k hei , ek i =
n
X
i=1
k=1
* n
X
i,k=1
n
X
n
X
n
X
y k hei , ek i
k=1
xk y k hek , ek i +
| {z }
=1
X
i6=k
xi y k hei , ek i
| {z }
=0
xk y k .
QED
k=1
Corollary 10.11 (Parseval Equation).

hX, Xi = kXk2 =
n
X
|xk |2 .
k=1
This seems trivial here, but it is very important in conservation laws.

In R3 the scalar product is also called a dot product.
1. INNER PRODUCT
69
Theorem 10.12. In R3 the inner product can be defined by

[
hX, Y i = kXk kY k cos(X,
Y)
(10.1)
[
where kXk is the length of X and (X,
Y ) is the angle between X and Y .
Proof. Note first that in R3 any inner product must be real-valued. So 1) transforms to
hX, Y i = hY, Xi .
But this equation obviously holds for the inner product defined by (10.1). Property 3)
[
is also clear since cos(X,
X) = 1. Property 2) is the least trivial. Prove first that
hX + Y, Zi = hX, Zi + hY, Zi .
Indeed
X +Y
\
hX + Y, Zi = kX + Y k kZk cos(X+Y,Z)
\
= kZk kX + Y k cos(X+Y,Z).
But geometrically, it is clear that the projection of a sum is the sum of the projections!
So
\ = kXk cos(X,
[
kX + Y k cos(X+Y,Z)
Z)
Z
d
[
Z)
kXk cos(X,
Z) kY k cos(Y,
d
+ kY k cos(Y,
Z)
\
kX + Y k cos(X+Y,Z)
d
[
and hX + Y, Zi = kXk kZk cos(X,
Z) + kY k kZk cos(Y,
Z) = hX, Zi + hY, Zi.
X
X
cos( ) = cos
Observe also that

hX, Zi = hX, Zi , R
(again it is geometrically clear).
QED
Theorem 10.13. Let {ek }nk=1 be an ONB and {xk }nk=1 be the coordinates of a vector
X E in {ek }nk=1 . Then
xk = hX, ek i
Proof.
hX, ek i =
* n
X
k=1
(10.2)
+
xi ei , ek
Pn
k=1
xi hei , ek i = xk .
| {z }
=ik
Definition 10.14. A linear space with an inner product h, i is called a Euclidean

space or inner product space.
70
Every finite dimensional linear space has an inner product (e.g. Rn , Cn ). An infinite
dimensional Euclidean space is called differently: a Hilbert space. But not all infinite
dimensional spaces have an inner product.
2. Adjoint and Selfadjoint Operators
Definition 10.15. Let E be a Euclidean space. An operator A is called the adjoint
operator to an operator A if X, Y E
hAX, Y i = hX, A Y i .
(10.3)
Theorem 10.16. The operator A defined by (10.3) is linear.

Proof. Using the properties of the inner product, we have
hZ, A (X + Y )i
by def of A
hAZ, X + Y i = hAZ, Xi + hAZ, Y i
= hAZ, Xi + hAZ, Y i
by def of A
hZ, A Xi + hZ, A Y i
= hZ, A Xi + hZ, A Y i = hZ, A X + A Y i .

So we get
X, Y, Z E ; , C
hZ, A (X + Y )i = hZ, A X + A Y i .
It follows from Corollary 10.5 that

A (X + Y ) = A X + A Y.
QED
So an adjoint operator A is linear. We will show later how any linear operator in
a Euclidean space has an adjoint.
Definition 10.17. An operator A in a Euclidean space E is called selfadjoint if
A = A.
Lemma 10.18. Let A be a linear operator and let {ek }nk=1 be an ONB, then for the
elements of the matrix of A, we have
aik = hAek , ei i
Proof. By definition (formula (9.1))
n
X
Aek =
ajk ej
(10.4)
(Aek )i = aik .
j=1
On the other hand, by (10.2)

(Aek )i = hAek , ei i .
Theorem 10.19. For the matrices of A , A we have
a11 a12 . . . a1n

a11 a21 . . . an1
a21 a22 . . . a2n
a
a
. . . an2
.
= .12 .22 .
.
.
.
..
..
..
..
..
..
..
..
.
a1n a2n . . . ann
an1 an2 . . . ann
QED
i.e (A )ik = (A)ki .
3. UNITARY OPERATORS
71
Proof. By (10.4)
(A )ik = hA ek , ei i = hek , Aei i = hAei , ek i = (A)ki .
QED
Why is it that every operator in a finite dimensional space has an adjoint? Because
of Theorem 10.19, a constructive proof.
Lemma 10.20. (A ) = A.
Proof.
hA X, Y i = hY, A Xi = hAY, Xi = hX, AY i
(A ) = A.
Theorem 10.21. If A = A then

(A)ik = (A)ki .
Proof. Prove it as an exercise.
Selfadjoint operators play an especially important role in physics. Consider one
important example.
Example 10.22. Let {ek }nk=1 be an ONB in E. Define an operator Pk (projection)
by the formula
Pk X = x k e k ,
th
where xk is the k coordinate of X in {ek }nk=1 . By (10.2), xk = hX, ek i and we have
Pk X = hX, ek i ek .
(10.5)
We claim Pk is selfadjoint. Indeed,

hPk X, Y i = xk hek , Y i = hX, ek i hY, ek i = hX, ek i y k = hX, yk ek i = hX, Pk Y i .
3. Unitary Operators
Definition 10.23. An operator A1 is called the inverse operator to A if
A1 A = AA1 = I.
Definition 10.24. An operator A is called invertible if it has an inverse operator.
Definition 10.25. A linear operator U is called unitary if
U U = U U = I.
Theorem 10.26. If U is unitary, then
1) U 1 = U , i.e. U is invertible and its inverse is U ;
2) kU Xk = kXk , X E.
The first part is marvelous since if you recall in Linear Algebra, computing an
inverse is a lot of work, but not here. The second part explains the name unitary: such
a transformation preserves length in finite dimension.
Proof. By comparing Def 10.23 & 10.25 we conclude that U = U 1 . To prove 2)
consider
kU Xk2 = hU X, U Xi = hX, U U Xi = hX, IXi = kXk2
kU Xk = kXk .
72
Remark 10.27. Equation 2) in Theorem 10.26 means that a unitary operator preserves the norm of a vector.
Norm in physics is often energy (at least in PDE in math physics) so an energy
preserving operator (often of time) leads to a conservation law.
Theorem 10.28. Let U be unitary, and let {uik } be its matrix representation. Then
any two columns and rows are orthogonal (and orthonormal).
Proof. By definition
U U = U U = I.
In the matrix form this equation reads, for any 1 i, k n
n
X
(U U )ik =
(U )ij (U )jk = ik .
(10.6)
j=1
But by Theorem 10.19 (U )jk = (U )kj = ukj and for (10.6) we have
n
X
uij ukj = ik .
(10.7)
j=1
But (ui1 , ui2 , . . . , uin ) Ui is the ith row of {uik } and (uk1 , uk2 , . . . , ukn ) Uk is the
k th row of {uik }. (10.7) then reads
hUi , Uk i = ik
Ui Uk
( and kUi k = 1).
In the same way we prove the orthogonality of columns.
QED
Note that you can use the rows (or columns) to make a new orthonormal basis.
Definition 10.29. Real matrices with orthogonal columns and rows are called orthogonal.
Exercise 10.30. Let U be unitary. Show that its columns are orthonormal.
(AB) = B A .
LECTURE 11
More about Selfadjoint and Unitary Operators. Change of

Basis
1. Examples of Selfadjoint and Unitary Operators
Example 11.1. Let Pn [0, 1] be the space of all polynomials of order n on the
interval [0, 1]. We define a scalar product by the formula
1
(p, q Pn [0, 1]) .
(11.1)
p(x)q(x)dx
hp, qi =
0
It is very easy to see that (11.1) defines a scalar product.

So with this scalar product our space Pn [0, 1] becomes a Euclidean space.
Let P0n [0, 1] = {p Pn [0, 1] : p(0) = p(1) = 0}. Its
a Euclidean space again. Check it!
1 d
0
1
Consider the linear operator A =
. This operai dx
tor is important in quantum mechanics, and is called
the momentum operator:
1
Ap = p0 .
i
It is a selfadjoint operator. Indeed
1
1 0
hAp, qi =
p (x)q(x)dx
0 i
We integrate it by parts
0 1
1
*
1

1
1
1 d

0
hAp, qi = p(x)q(x)
p(x)q (x)dx =
p(x)
q(x)dx = hp, Aqi .

i
i dx
0
0 i
0
A = A. Note that both the boundary conditions and the factor 1i were
crucial in making A selfadjoint. If we just have the operator of differentiation, d/dx
on P0n [0, 1], then this operator is antisymmetric, i.e. hp0 , qi = hp, q 0 i.
1
Exercise 11.2. Show that hp, qi =
p(x)q(x)dx defined on Pn [0, 1] is an inner
product.
Example 11.3 (the operator of rotation). Let a particle rotate about the origin with
an angular velocity . Let (x0 , y0 ) be its initial position. Find a formula for (x(t), y(t))
at any instant of time t.
73
74
11. MORE ABOUT SELFADJOINT AND UNITARY OPERATORS. CHANGE OF BASIS
B (x, y)
Given A, B, construct:
t
C: draw the ray of angle t with the
origin, then C is the projection of B on
the ray;
y
T : the projection of A on the x-axis;
Q: the projection of C on the x-axis;
P : the projection of B on the x-axis;

t
S: the intersection of BP and the horit
zontal line from C.

O xP
Note first that 4OBC = 4OAT
A
(x0 , y0 )
x = OP = OQ P Q = OC cos t SC = |{z}
OT cos t |{z}
BC sin t
=x0
=AT =y0
x(t) = x0 cos t y0 sin t

y = P S + BS = CQ + |{z}
BC cos t = |{z}
OC sin t + y0 cos t
=x0
=AT =y0
y(t) = x0 sin t + y0 cos t
x = x0 cos t y0 sin t
y = x0 sin t + y0 cos t
We claim that

x0
cos t sin t
x
.
=
y0
sin t cos t
y

cos t sin t
U=
sin t cos t
is an orthogonal matrix. Check it!
Hence, U is a unitary operator.
So the solution to our problem can be written as follows
X(t) = U (t)X0

x(t)
x0
where X(t) =
is the position vector at time t, X0 =
is the initial position
y(t)
y0
and

cos t sin t
U (t) =
.
sin t cos t
2. Change of Basis
Note that the xyz-frame is not always practical, e.g. coordinates on Earth, so we
switch to spherical coordinates for example; but switching coordinate system is the
same as switching bases.
Let us have two bases {ei }ni=1 , {gi }ni=1 not necessarily orthogonal. We raise a question: does there exist a linear operator G such that
gi = Gei
i = 1, 2, . . . , n ?
2. CHANGE OF BASIS
The answer is Yes. Here is the contruction.

Represent gk , k = 1, 2, . . . , n, with respect to {ei }ni=1
n
X
gk =
(gk )i ei .
75
(11.2)
i=1
{gik }ni,k=1
: gik (gk )i .
Consider the matrix G
On the other hand by definition, we have
n
n
X
X
Gek =
(Gek )i ei =
gik ei .
| {z }
k=1
(11.3)
i=1
=gik
Comparing (11.2) and (11.3), we get

gk = Gek
where G = {gik }, gik = (gk )i is the i
th
k = 1, 2, . . . , n
coordinate of the k th basis vector.
Definition 11.4. Operators A and B are called similar if there exists an invertible
operator G:
G1 AG = B.
Theorem 11.5. If A and B are similar, i.e.
G1 AG = B,
(11.4)
and {ek }nk=1 is a basis, then the matrix of B in {ek }nk=1 coincides with the matrix of A
in the basis {gk }nk=1 , gk = Gek .
Proof. Left as an exercise.
Remark 11.6. Given G, A, the transformation
A G1 AG
is called a similarity transformation of A.
Theorem 11.5 actually says
(G1 AG)ik =
in {ei }
(A)ik
in {Gei }
and the following important theorem follows.

Theorem 11.7. Let {ei } be a basis and {gi } be a new basis. Assume next that
G
{ei } {gi } i.e. gi = Gei . Then the matrix of A in the new basis {gi } is equal to the
matrix of G1 AG in the old basis {ei }.
Exercise 11.8. Prove that if {ek }, {gk } are ONB and gk = Gek then G is unitary.
Note that the above exercise implies that orthogonality is preserved with rotation.
LECTURE 12
Eigenvalues and Eigenvectors

1. Eigenvalues
The concept of eigenvalues and eigenvectors is central in math physics.
Definition 12.1. Let E be a complex linear space and let A be a linear operator
in E. Consider the following equation
AX = X
C.
(12.1)
The values of for which (12.1) has a nontrivial solution are called eigenvalues of A.
The corresponding nontrivial solutions X are called eigenvectors of A.
Equation (12.1) always has a solution X = 0. But we are talking about nontrivial
solutions (i.e. X 6= 0). The curious fact is that (12.1) has nontrivial solutions only for
a finite number of .
Definition 12.2.
p() det(A I)
is called the characteristic polynomial of A and the equation
p() = 0
is called the characteristic equation of A.
Lemma 12.3. Let A be an operator in E, dim E = n. Then the characteristic
equation for A has n roots.
The proof of this fact lies beyond the scope of our consideration.
clear why the number of roots n. Indeed,
a11
a12
a1n
a21
a22
a2n
p() = det(A I) = det
..
..
..
...
.
.
.
an1
an2
But at least its
ann
is a polynomial of order n, and hence the number of roots n.

Definition 12.4. The set of all eigenvalues of A is called the spectrum of A and
is denoted by (A).
An alternate notation is Spec(A).
Theorem 12.5.
(A) = { C : det(A I) = 0}.
77
78
12. EIGENVALUES AND EIGENVECTORS
Proof. Rewrite (12.1) as follows

AX X = 0
(A I)X = 0.
From elementary linear algebra we know that for an equation

(A I)X = 0
to have a nontrivial solution it is necessary that
det(A I) = 0
which is the characteristic equation of A.
QED
Definition 12.6. An eigenvalue 0 is said to be simple if 0 is a simple root of

the characteristic polynomial p().
Definition 12.7. An eigenvalue 0 is said to be of multiplicity m if 0 is a root of
p() = 0 of multiplicity m.
The above is referred to as algebraic multiplicity. There is another type of multiplicity, called geometric multiplicity, but it is never present in selfadjoint operators so
we will not worry about it.
Definition 12.8. The procedure of finding eigenvalues and eigenvectors is called
the spectral analysis of A.
2. Spectral Analysis
Since the most important operators in physics are selfadjoint and unitary (because
of conservation of energy), we will mostly concentrate on the spectral analysis of such
operators.
Theorem 12.9. The spectrum of a selfadjoint operator A is real. Its eigenvectors
are orthogonal. I.e. if A = A then
(A) R
hXi , Xk i = ik .
Proof. Let i , k be two eigenvalues of A and let Xi , Xk be the corresponding

eigenvectors. Compute hAXi , Xk i:
hAXi , Xk i = i hXi , Xk i
k
hXi , AXk i = k hXi , Xk i
So
i hXi Xk i = k hXi , Xk i .
(12.2)
Let i = k. We have
k kXk k2 = k kXk k2
k = k
k R.
So all eigenvalues of A are real. Let now i 6= k. (12.2) implies

(i k ) hXi , Xk i = 0
(i k ) hXi , Xk i = 0
hXi , Xk i = 0.
QED
2. SPECTRAL ANALYSIS
79
Theorem 12.10. If A is a selfadjoint operator and 0 (A), then the set of all
eigenvectors corresponding to 0 forms a subspace, named an eigenspace.
Proof. Let X0 , Y0 be two solutions of (12.1)
AX0 = 0 X0 ,
AY0 = 0 Y0 ,
then X0 + Y0 is a solution to (12.1) too. Indeed
A(X0 + Y0 ) = AX0 + AY0 = 0 X0 + 0 Y0 = 0 (X0 + Y0 ).
QED
Theorem 12.11. Let A = A and let the spectrum (A) be simple (all eigenvalues
are simple). Then the set of its normalized eigenvectors forms an ONB.
Proof. Let dim E = n. By Lemma 12.3 the number of eigenvalues of A is n. By
Theorem 12.9, all eigenvectors X1 , X2 , . . . , Xn are orthogonal. Hence they are linearly
Xk
. Now
independent and therefore {X1 , X2 , . . . , Xn } forms a basis in E. Put ek =
kXk k
QED
{ek }nk=1 is an ONB.
The statement holds without simple for selfadjoint operators, i.e. each eigenspace
is of dimension the multiplicity of its corresponding eigenvalue.
Remark 12.12. The fact that eigenvectors of a selfadjoint operator with a simple
spectrum forms an ONB provides us with a very efficient way of constructing ONBs.
The following theorem is of the utmost importance.
Theorem 12.13 (Diagonalization Theorem). Let A = A and (A) = {k }nk=1 be
simple. Let {ek }nk=1 be the ONB compiled of the eigenvectors of A. Then the matrix of
A in {ek }nk=1 is diagonal.
Again, this is actually true for any selfadjoint operator.
Proof. By definition of (A)ik
Aek
n
X
(A)ik ei
k=1
k
k ek
(A)ik = 0 , i 6= k ; (A)kk = k .
That is, in {ek }nk=1
A=
1
2
..
.
n
QED
80
12. EIGENVALUES AND EIGENVECTORS
Now we are going to answer the following question: Given a selfadjoint matrix A,
find a transformation G that moves the old basis {ek }nk=1 into a basis consisting of the
eigenvectors {gk }nk=1 of A.
So let us find G : Gek = gk , k = 1, 2, . . . , n, where {gk }nk=1 is a basis of normalized
eigenvectors of A.
n
X
Gek =
(G)ik ei
(G)ik = (gk )i where (gk )i is the ith coordinate of gk .
k=1
Therefore,
(g1 )1 (g2 )1
(g1 ) (g2 )
2
2
G=
...
...
(g1 )n (g2 )n
. . . (gn )1

. . . (gn )2
= g1 g2 . . . gn .
... ...
. . . (gn )n
So G is actually a matrix which columns are eigenvectors of A.

Below is the procedure on how to bring a matrix A to the diagonal form
1) Do the spectral analysis of A:
Xk
{1 , 2 , . . . , n } ; {g1 , g2 , . . . , gn } , gk =
kXk k

2) Construct the matrix G = g1 g2 . . . gn
3) G1 AG is diagonal in {g1 , g2 , . . . , gn }.
Remark 12.14. Actually, Theorem 12.13 remains valid not only for the case of a
simple spectrum but for a whole lot more general situations.
Exercise 12.15.
1) Prove the Spectral Theorem: under conditions of Then

X
orem 12.13, the following representation holds: A =
k Pk , where Pk is the
k=1
orthogonal projector on the eigenspace of A corresponding to k .

2) Prove that eigenvalues of a unitary operator lie on the unit circle. Find the
analog of the spectral theorem for a unitary operator.
Note that in the above part 1), Pk is referred to as a spectral projector, and this
representation as a decomposition into projectors holds in infinite dimension. In part
2), observe that the eigenvalues are no longer on the real line.
Part 3
Hilbert Spaces
LECTURE 13
Infinite dimensional spaces. Hilbert Spaces

1. Infinite Dimensional Spaces
Actually, many definitions in the infinite dimensional case remain the same. E.g,
the definition of a linear space does not change. (So Def. 8.1, 8.5, 8.6 remain the
same). The actual difference is
Definition 13.1. A linear space is called infinite dimensional if there is an infinite
system of linearly independent vectors.
Example 13.2. Let P be the set of all polynomials. P is clearly a linear space.
n
Consider the following system {xn }
n=0 . All x P and these vectors are linearly
independent, but their number is infinite. Therefore, dim P = .
Example 13.3. Let C[0, 1] as in Example 8.4. Consider {xn }
n=0 C[0, 1] and
again dim C[0, 1] = .
As a matter of fact, spaces of functions are a good source of infinite dimensional
spaces.
2. Basis
It is natural to ask what a basis is in such a space.
Definition 13.4. Let E be an infinite dimensional space. A system {en }
n=0 E
is called a basis in E if every element X of E can be represented as follows:
X
xn e n .
(13.1)
X=
n=1
Wow! But how to understand this infinite series! Here the main difference between
finite and infinite dimensional spaces starts.
In general, it is a very deep issue. We are not able to treat it here and we restrict
ourselves to very special yet important cases.
3. Normed Spaces
Definition 13.5. A space E is called a normed space if there exists a real valued
function, called a norm, denoted by kXk, of X E subject to
1) kXk 0 ; kXk = 0 X = 0.
2) kXk = || kXk , C.
3) kX + Y k kXk + kY k . (triangle inequality)
83
84
13. INFINITE DIMENSIONAL SPACES. HILBERT SPACES
p
Example 13.6. Let E = R3 , kXk = x21 + x22 + x23 . We claim kXk is a norm.
Indeed,
p
1) kXk 0 p
and if kXk = 0 x21 + x22 +x23 =
p0 x1 = x2 = x3 = 0 X = 0.
2
2
2
2
2) kXk = (x1 ) + (x2 ) + (x3 ) = x21 + x22 + x23 = || kXk .
3)
kX + Y k2 = hX + Y, X + Y i
= hX, Xi + hX, Y i + hY, Xi + hY, Y i
= kXk2 + kY k2 + 2 hX, Y i
[
= kXk2 + kY k2 + 2 kXk kY k cos (X,
Y)
kXk2 + kY k2 + 2 kXk kY k = (kXk + kY k)2
So we get
kX + Y k2 (kXk + kY k)2
kX + Y k kXk + kY k .
So in R the length of X can be chosen as a norm, and hence R3 is a normed space.

Lemma 13.7 (Cauchys Inequality).
|hX, Y i| (hX, Xi)1/2 (hY, Y i)1/2 .
(without a proof )
Theorem 13.8. Every Euclidian space E is normed.
Proof. Let h, i be a scalar product in E. We claim hX, Xi = kXk2 defines a
norm. Based upon Lemma 13.7, we have
kX + Y k2 = hX + Y, X + Y i
= hX, Xi + hX, Y i + hY, Xi + hY, Y i
= kXk2 + kY k2 + hX, Y i + hX, Y i = kXk2 + kY k2 + 2 Re hX, Y i
kXk2 + kY k2 + 2 |hX, Y i|
Cauchy ineq
kXk2 + kY k2 + 2 kXk kY k = (kXk + kY k)2 .

QED
Example 13.9. Let C[0, 1] be as previously defined. Set kf k max |f (x)|. Prove
x[0,1]
that kf k is a norm.
So, C[0, 1] is a normed space.
Example 13.10. P is not a normed space. (try to understand it!)
Example 13.11. Let C 1 [0, 1] be the space of all functions f (x) continuous on [0, 1]
whose derivatives are also continuous on [0, 1]. I.e.,
C 1 [0, 1] = {f C[0, 1] : f 0 C[0, 1]}
Prove that kf k = max |f (x)| + max |f 0 (x)| is a norm.
x[0,1]
x[0,1]
4. HILBERT SPACES
85
4. Hilbert Spaces
We introduce first the concept of a scalar (inner) product in the very same way as
Definition 10.1.
The point here is that not all infinite dimensional spaces have a scalar product.
Example 13.12. C[0, 1] has a scalar product. Indeed, let f, g C[0, 1], then
1
f (x)g(x)dx
hf, gi =
0
satisfies all the properties of a scalar product.

Example 13.13. Let P be the space of all polynomials. There is no scalar product
in P.
Definition 13.14. An infinite dimensional Euclidian space is called a Hilbert space.
Example 13.15. Let L2 (0, 1) be the space of all functions f (x) such that
1
|f (x)|2 dx <
(square integrable functions)
0
2
We claim that L is a Hilbert space.
LECTURE 14
L2 spaces. Convergence in Normed Spaces. Basis and

Coordinates in Normed and Hilbert Spaces
1. L2 spaces
The most typical examples of Hilbert spaces are so-called L2 spaces.
Definition 14.1. 1 L2 (a, b) is the set of all functions f (x) defined on (a, b) and
satisfying the condition
b
(14.1)
|f (x)|2 dx <
a
Theorem 14.2. L2 (a, b) is a Hilbert space.

Proof. Let us check first that L2 (a, b) is a linear space. Its clear that out of the
10 properties of a linear space, only # 1 is in question. Let f, g L2 (a, b). We need to
prove that f + g L2 (a, b).
|f (x) + g(x)|2 = |f (x)|2 + |g(x)|2 + 2 Re f (x)g(x).
But Re f (x)g(x) |f (x)| |g(x)| and hence,
|f (x) + g(x)|2 |f (x)|2 + |g(x)|2 + 2 |f (x)| |g(x)| .
(14.2)
It follows from the obvious inequality

(|f | |g|)2 0
|f |2 + |g|2 2|f ||g|
that
2|f (x)||g(x)| |f (x)|2 + |g(x)|2
and (14.2) becomes
|f (x) + g(x)|2 2|f (x)|2 + 2|g(x)|2

b
b
b
2
2
|f (x) + g(x)| dx 2
|f (x)| dx +2
|g(x)|2 dx
a
| a {z
}
| a {z
}
<
<
f (x) + g(x) L (a, b)
So, L2 (a, b) is a linear space.

1L2 (a, b)
is a typical example of a function space. Other function spaces are C[0, 1], P, etc.
87
88
14. EXAMPLE, CONVERGENCE AND BASIS IN HILBERT SPACES
Clearly,
hf, gi =
f (x)g(x)dx
a
has all the properties of a scalar product and hence L2 (a, b) is a Hilbert space. QED
Remark 14.3. Every function f(x) continuous on (a, b) is in L2 (a, b). So C[a, b]
L2 (a, b). But L2 (a, b) also contains discontinuous functions and even some unbounded.
For example,
1
L2 (0, 1) , log x L2 (0, 1)

4
x
but both functions are unbounded at 0. Note, however, that
sin x
1

/ L2 (0, 1) but f (x) = L2 (0, 1).
x
x
Remark 14.4. In the same way one can define

2
2
L (0, ) = f (x) :
|f (x)| dx <
0
or,
L (, ) L (R) = f (x) :
2

|f (x)| dx < .
2
Exercise 14.5. Prove that:

1
L2 (R) ,
1 + x2
eix 1
L2 (R) ,
x
1
/ L2 (R)
/ L2 (R)
x
eix
/ L2 (R) ,
Definition 14.6. A function f (x) L2 (a, b) is called normalized if

b
2
kf k =
|f (x)|2 dx = 1.
a
2
Two functions f, g L (a, b) are said to be orthogonal if

b
hf, gi =
f (x)g(x)dx = 0.
a
1
Example 14.7. Let fn (x)
sin nx in L2 (0, 2). We prove that kfn k = 1,
fn fm , n 6= m. Indeed,
2
1 2 2
1 2 1 cos 2nx
1 x
2
kfn k =
sin nx dx =
dx =
= 1 kfn k = 1.
0
0
2
2 0
1 2
Now consider hfn , fm i =
sin nx sin mx dx.
0
2. CONVERGENCE IN NORMED SPACES
But sin x sin y = 21 (cos(x y) cos(x + y)). So,

2
2
1
1
hfn , fm i =
cos(n m)x dx
cos(n + m)x dx = 0 ,
2 0
2 0
|
{z
}
|
{z
}
89
n 6= m.
=0
=0 , n6=m
n 6= m.
o
nq
1
sin
nx
play a very important role in Fourier
As we will see, functions
n=1
analysis.
So
hfn , fm i = 0 ,
2. Convergence in normed spaces

Definition 14.8. Let E be a normed space, and {Xn }
n=1 be a sequence of vectors
in E. The sequence {Xn }n1 is said to converge to a vector X E in norm if
lim kX Xn k = 0.
(14.3)
Remark
14.9. If X, Y E, then kX Y k plays the role of a distance between X

p
and Y kX Y k is the distance if X, Y R3 and kXk = x21 + x22 + x23 . In view
of this, (14.3) means that the distance between X and Xn gets smaller and smaller as
n .
Definition 14.10. A sequence {Xn }n1 is said to be a Cauchy sequence if
lim kXn Xm k = 0.
n
m
Definition 14.11. If every Cauchy sequence {Xn } E converges to some element

X of E, the space E is called a complete or Banach space.
The theory of Banach spaces is extremely complex and still has a lot of unanswered
questions. Even elementary theory of Banach spaces requires advanced analysis and
hence lies beyond the scope of our consideration.
Some examples: C[0, 1] is a Banach space, L2 (0, 1) is a Banach space, but P[0, 1]
is not. Every finite dimensional Euclidean space is Banach. Not every normed space
is a Banach space. Actually, the definition of a Hilbert space we gave previously is
incomplete.
A Hilbert space is a Banach space with a scalar product.
As we know, L2 (a, b) is a Hilbert space. I.e., it is also Banach which implies
b
2
if
{fn }n=1 L (a, b) and
lim
|fn (x) fm (x)|2 dx = 0
n
m
then {fn (x)} converges in L2 (a, b) to some function f (x) L2 (a, b).
90
14. EXAMPLE, CONVERGENCE AND BASIS IN HILBERT SPACES
Example 14.12.
It may be proven that

n
X
xk
ex in L2 (0, 1)
k!
k=0
but the proof is not easy though.

Note that C[0, 1] is not a Hilbert space because it is incomplete: we can find a
Cauchy sequence of continuous functions going to a noncontinuous functions.
1
2
1
2
1
n
1
2
1
n
1
2
C[0, 1]
/ C[0, 1]
But L2 (0, 1) is a Hilbert space big enough to contain all the limits of Cauchy
sequences in C[0, 1].
3. Bases and Coordinates
Now we are ready to talk about bases and coordinates in infinite dimensional spaces.
Definition 14.13. Let E be a Banach space. A system {en }
n=1 is called a basis if

n

X

xk ek 0 , n .
X E {xk }k=1 C : X

k=1
The numbers
{xk }
k=1
are called coordinates of X in the basis {ek }

k=1 .
Xk is called convergent to X in E if
k=1

n

X

X
X

k 0 , n .

k=1
Definition 14.15. A basis {ek }

k=1 is called orthogonal and normalized in a Hilbert
space H if
hei , ek i = ik .
Theorem 14.16. If {ek }
k=1 is an ONB then every X H can be represented as
follows
X
hX, ek i ek
X=
k=1
where the series converges in H. Moreover,

X
|xk |2 ,
kXk2 =
k1
xk = hX, ek i .
LECTURE 15
Fourier Series
Now we are going to harvest on the previous abstract results. for example, in the
Hilbert space L2 (, ), there is an easy basis to construct.
Theorem 15.1.
1 einx
2
is an ONB in L2 (, ).
nZ
Proof. Check that this system is orthogonal and normalized. Set en =

If n 6= m,
1
hen , em i =
2
=
i.e.
einx eimx
1
dx =
2
i(nm)x
1 einx .
2

1 ei(nm)x
dx =
2 i(n m)
1
1 ei(nm) ei(nm)
=
sin(n m) = 0.
2i(n m)
(n m)
hen , em i = 0.
1
ken k =
2
einx einx dx = 1.
We leave the fact that {en } is a basis without a proof.

Theorem 15.1 leads to
Theorem 15.2 (Expansion in Fourier Series). Let f be a real valued function such
that f L2 (, ), then

X
1
inx
f (x) =
cn e
, cn =
f (x)einx dx.
(15.1)
2
nZ
For {cn }, the Parseval identity holds,
kf k2 = 2
X
nZ
91
|cn |2 .
(15.2)
92
15. FOURIER SERIES

set en =
1 einx .
2
1 einx
2
is an ONB. For simplicity,
nZ
We have
f=
hf, en i en =
nZ
X

f (x)en (x) dx en
nZ

X 1
1
inx
f (x)e
dx einx
=
2
2
nZ

X 1
X
=
f (x)einx dx einx =
cn einx .
2
nZ |
nZ
{z
}
cn
By Theorem 14.16,
2
kf k =
2
X

|hf, en i| =
2cn
2
and (15.2) follows.
QED
Formula (15.1) is called Fouriers expansion formula.

(
0 , < x < 0
Example 15.3. Let f (x) =
.
1, 0x<
Find its Fourier series.

1
f (x)einx dx
cn =
2

1
1 einx
1 ein 1
inx
=
e
dx =
=
2
2 in 0
2 in
0
1
, n = 1, 3, . . .
in
= 0
, n = 2, 4, . . .
1
, n = 0.
2
So
1
1 X einx
f (x) = +
.
2 i nZ n
n odd
Now clearly,
kf k =
|f (x)| dx =
dx = .
0
But we can also verify this result by Parseval equality:

!

X
1
1
1
1
1 2
2
kf k = 2
+
= 2
+
= .
4 2 n= (2n + 1)2
4 2 4
f (x)
(15.3)
15. FOURIER SERIES
93
Remark 15.4. In general, series (15.1) converges only in L2 (, ), i.e.

2
N

X

inx
cn e dx 0 ,
f (x)

does not imply that

N
X
cn einx f (x)
x (, ).
The reason can be observed from Example 15.3. Series (15.3) converges to f (x) for all
x (, ) but x = 0. At x = 0,

N
1
1 X 1
1
1
1
1
1 1
1
+
= +
. . . 6 1+ 6 1 + . . . + + = .
2 i N n
2 i
N 2
N 2 N
2
N

n odd
So at x = 0 (15.3) converges to 12 .
It can be proven that if f (x) is continuous then its Fourier series converges for all
x (, ). If f (x) is piecewise continuous (Here is a typical function)
then on each interval of continuity of f the Fourier series converges pointwise (i.e. at
every point of this interval) and if x0 is a point of jump discontinuity then the Fourier
series converges to
f (x0 0) + f (x0 + 0)
,
2
f (x0 0) = xx
lim f (x).
0
x<x0
Since in applications most of the functions are at least piecewise continuous, series
(15.1) converges not only in norm but also pointwise.
n
o
1
inx
inx
Note, that e
= cos nx + i sin nx and hence instead of
e
we can
2
nZ
n
o
consider 1 sin nx , 1 cos nx
which is already a real basis. An expansion in
n0
this basis is called a trigonometric series.

Theorem 15.5. Let f L2 (, ), then
94
15. FOURIER SERIES
a0 X
+
(an cos nx + bn sin nx)
2
n1
1
an =
f (x) cos nx dx , n = 0, 1, . . .

1
f (x) sin nx dx , n = 1, 2, . . .
bn =

f (x) =
where
(15.4)
Exercise 15.6. Derive Theorem 15.5 from Theorem 15.2 and establish the Parseval
identity.
Definition 15.7. Functions einx , sin nx, cos nx are called simple harmonics.
So, a Fourier series can be viewed as an expansion in simple harmonics that play
an enormous role in physics.
Remark 15.8. Theorems 15.2 & 15.5 give us expansions of functions defined on
(, ). However if we continue f (x) outside (, ) as a periodic function f (x+2) =
f (x), then the Fourier expansion formulas are valid for all x. To see this, one has only
to observe that einx are 2periodic functions too.
Theorem 15.9. Let f L2 (`, `), then
X
n
1 `
i n
x
f (x) =
cn e `
, cn =
f (x)ei ` x dx
2` `
nZ
n Z.
(15.5)
Exercise 15.10. Derive this theorem from Theorem 15.2 using a suitable change
of variables.
Exercise 15.11. Expand f (x) = x2 on [, ] by
(a) (15.1);
(b) (15.5).
Exercise 15.12. Expand f (x) = |x| on [1, 1] by (15.5).
LECTURE 16
Some Applications of Fourier Series

1. Signal Processing
Consider a black box. Assume that it is a linear system, i.e.
f
input
Lf
L(f1 + f2 ) = Lf1 + Lf2
output
L(f ) = Lf.
In other words, the system works as a linear operator. In physical terms it means
that the system is subject to the principle of superposition.
There are two basic problems:
1) Given input f and L, find the output Lf
2) Given input f and output Lf , find L
(direct problem),
(inverse problem).
We assume that an input f (t) is a periodical function of t. I.e. we study periodical

signals like e.g.
or
and for simplicity put the period 2.

We claim that, to study problems 1), 2), we have to only study how simple harmonics come through this system.
Let {en }nZ , en = eint be a sequence of simple harmonics (see next page).
Then we conduct a laboratory experiment
en
Len
nZ
and form a database: {Len } {gn } and the inverse problem 2) is solved since {Len }nZ
determines our system. Now, given signal f (t) we want to know what the output will
be without actually putting this signal through the system.
What we have to do is represent f (t) by the Fourier formula
X
f (t) =
cn eint
(16.1)
nZ
95
96
16. SOME APPLICATIONS OF FOURIER SERIES
e0
constant signal
e1
first harmonic
e2
second harmonic
e3
third harmonic
1. SIGNAL PROCESSING
97
and compute all {cn }nZ . Then the output (Lf )(t) is
X
X
(Lf )(t) =
cn Leint =
cn Len .
nZ
(16.2)
nZ
So once the system is determined, i.e. {Len } are known, then one can easily compute Lf for any signal f by formulas (16.1), (16.2). So the direct problem 1) is also
solved.
Example 16.1. Consider the following system L
(
en
, n0 n n1
Len =
0
, otherwise.
So, L puts through all simple harmonics with frequencies in [n0 , n1 ] with no change and
cuts off all other harmonics. It is a filter.
Lf =
n1
X
cn e n .
n=n0
high frequency filter

low frequency filter
Example 16.2. Consider L defined by

(
Aen
, n0 n n1
L=
0
, otherwise
A > 1.
L is a band amplifier.
A
1
n0 n1
(n0 , n1 ) frequency amplifier
n0
resonance amplifier
So what does a resonance amplifier do to a signal? It cuts out all the harmonics
but one n0 and increases its amplitude by A.
Exercise 16.3. Put a signal
98
f (t) = t , t ,
through a system L:
(
en
, n = 0, 1, 2
Len =
0
, otherwise.
Graph the output Lf . Use the Fourier trigonometric series expansion for f (x).
2. Solving Some Linear ODEs

We consider just one example from the theory of electrical circuits.
It follows from Kirchhoffs law that for the circuit in the figure
L
where
E(t + T ) = E(t)
is a T -periodic electricity source,
E(t)
the charge Q(t) satisfies the ODE
L
d2 Q
dQ
1
+
R
+
Q = E(t)
dt2
dt
C
(16.3)
assuming that the circuit is steady-state (i.e. its been plugged in for a time long
enough so that all transition processes are over).
We need to find Q(t). I.e. we want to find the response of our linear system to the
periodic external source E(t).
Note that this problem is studied in standard ODE courses. As you may remember,
it can be approached in a few different ways but there is no exact solution unless E(t)
has a very specific form. We offer here yet another solution (a Fourier series solution).
Here is how it goes.
By Theorem 15.9 with ` = T
E(t) =
cn e
int
nZ

2
=
T
(16.4)
and we look for a solution to (16.3) in the form of the Fourier series
Q(t) =
X
nZ
qn eint
(16.5)
2. SOLVING SOME LINEAR ODES
99
where qn are unknown. We have (if we assume that dtd and

X
d
Q=
qn ineint ,
dt
nZ
can be interchanged)
(16.6)
X
d2
Q
=
qn (n2 2 )eint .
dt2
nZ
(16.7)
Plugging (16.4)-(16.7) into (16.3) yields

X
X
1
2 2
qn eint =
cn eint
L(n ) + R(in) +
C
nZ
nZ

1
q n = cn .
n2 2 L + inR +
C
I.e.
qn =
Ccn
.
(1 n2 2 CL) + inCR
But
1
cn =
T
(16.8)
T /2
E(t)eint dt
(16.9)
T /2
are all known (but you have to compute them though) and the problem is solved in
the form of the Fourier expansion (16.5) with qn computed by (16.8), (16.9).
Exercise 16.4. Find the Fourier series solution to (16.3) (assuming a steady-state
solution) in the form (15.5) for
a
(a) E(t) =
2T T
2T
a
(b) E(t) =
T
2
T
2
3T
2
Example 16.5. Take E(t) to be as in Example 15.3, i.e.
We found its Fourier series expansion to be

1 X 1 int
E(t) = +
e .
2
in
nZ
n odd
So using (16.8), we have

q0 =
C
2
qn =
C
n [nCR + i(1 n2 CL)]
n odd integer
100
and the charge Q(t) is given by

X
C
Ceint
Q(t) = +
.
2 CL)]
2
n
[nCR
+
i(1
n
nZ
n odd
Note that the result has only the odd integers again, and trigonometry will cancel
out the imaginary parts.
LECTURE 17
Linear Operators in Hilbert Spaces

The theory of linear operators in infinite dimensional spaces is much more difficult
than the one for finite dimensional spaces. And its still a very active area of research.
1. General Concepts
We will generally denote H a Hilbert space (infinite dimensional) and E a Euclidean
space (finite dimensional). You can not associate a matrix with every linear operator
since not every space has a basis.
The definition of a linear operator in a Hilbert space is absolutely the same as in
the case of a finite dimensional space.
Definition 17.1. The domain Dom A of a linear operator A in a space H is the
set of all elements X H such that AX H, i.e.
Dom A = {X H : AX H}.
In the finite dimensional case, for every linear operator A defined in E, Dom A = E.
For operators in infinite dimensional spaces, its not longer true.
d
Example 17.2. Let H = L2 (0, 1) and A = dx
. Pick up f (x) = 12
computation, kf k = 1 f H. Compute now kAf k:

1
df
1
1
=
Af =
5/4
dx
4 x
2
1
1
1
dx
dx
1
2
but kAf k =
=
=
32 0 (x5/4 )2
32 0 x5/2
4 x.
By a direct
/ Dom
and hence Af
/ L2 (0, 1) and f
A.
d
d
So Dom dx H but Dom dx 6= H!
Example 17.3. Let H = L2 (0, 1) again and define an operator A as follows
(Af )(x) = v(x) f (x)
where v(x) is a bounded function on (0, 1). I.e. max |v(x)| = C < .
x(0,1)
2
We claim that Dom A = H. Indeed, f L (0, 1) we have

1
1
2
2
kAf k =
|v(x)f (x)| dx
max |v(x)|2 |f (x)|2 dx
0
0 x(0,1)
1
= max |v(x)|2
|f (x)|2 dx = C 2 kf k2 < .
x(0,1)
101
(17.1)
102
17. LINEAR OPERATORS IN HILBERT SPACES
So f L2 (0, 1), Af L2 (0, 1) and Dom A = H.

The operator A in Example 17.3 is known in physics as the operator of energy, v
is the potential. If v(x) = x, A is the operator of coordinate. The domain being the
entire space is not always true, it requires v to be bounded.
Exercise 17.4. Show that if v(x) = 1/x, the Coulomb potential, then Dom A 6= H.
It would be natural to ask: what is the difference between the two examples?
Definition 17.5. Let A be a linear operator in H. The norm kAk of A is defined
as
kAk = max kAXk .
kXk1
Definition 17.6. A linear operator A is called bounded if kAk < and unbounded
if kAk = .
Example 17.7. Let A be as in Example 17.3. By Definition 17.5,
1
2
2
kAk = max kAf k = max
|v(x)f (x)|2 dx.
kf k1
By (17.1)
kf k1
(17.2)
|v(x)f (x)|2 dx C 2 kf k2
0
and hence (17.2) can be continued

kAk2 C 2 max kf k2 = C 2
kf k1
kAk C and A is bounded.

Example 17.8. Let A be as in Example 17.2. Pick up f (x) =
kf k = 1, but kAf k = kAk = and A is unbounded.
Theorem 17.9. A is bounded
kAXk C kXk
Proof.
1
1
.
2 4x
We have
C<
,
x H.
(17.3)
1) . Assume A is bounded i.e.

max kAXk kAk < .
kXk1
X
Let X H. Consider e = kXk
. It follows from (17.4) that

X

kAek kAk
A kXk kAk kAXk kAk kXk
and (17.3) follows.
(17.4)
1. GENERAL CONCEPTS
103
2) . Assume we have (17.3). (17.3) implies

kAXk C
kXk 1
and that implies

max kAXk C.
QED
kXk1
Theorem 17.10. All linear operators in finite dimensional spaces are bounded.
Proof. Let A be a linear operator in a finite dimensional space E. X E,

n
n

X
X

xi Aei
kAXk = A
xi e i =

i=1
i=1
v
v
u n
u n
n
X
uX
uX
2
|xi | kAei k t
|xi | t
kAei k2 .
(17.5)
i=1
i=1
i=1
Estimate now kAei k.

n
n
n
X
X
X

|aki |
|aki | kek k =
aki ek
kAei k =

So, kAei k
k=1
k=1
k=1
n
X
kek k = 1.
since
|aki |. Plug it in (17.5) and we have
k=1
v
v
!2
u n
u n
n
uX X
uX
kAXk t
|xi |2 t
|aki |
i=1
k=1
{z
kXk
!
|aki |
kXk
i,k=1
kAk = max kAXk

kXk1
k=1
n
X
n
X
|aki | <
A is bounded.
QED
i,k=1
Exercise 17.11. Consider the Euclidean norm kk2 on Rn .

1) If A is an operator from Rn to Rn , then
p
kAk2 = (A A)
where is the spectral radius, i.e. the largest absolute value |max | of the eigenvalues of A A. Use this property to find the norm of

1 0 2
1 1
3/2
1/2
A=
, B = 1 1 1
, C=
.
2 0
1/2
3/2
0 1 2
2) Now find the norm of A and C again, but using the definition of the norm.
104
So an unbounded operator is a phenomenon of infinite dimensional spaces. The

thing is that operators arising in mathematical physics are often unbounded.
One of the most popular bounded operators in mathematical physics is the Fourier
transform. Some definitions.
n
o
b
1
Definition 17.12. L (a, b) = f : a |f (x)| < is the space of functions integrable on (a, b).
One can easily see that L1 (a, b) is indeed a space:
f, g L1 (a, b) f + g L1 (a, b) where , are scalars.
In applications one often has to deal with
L (R) L (, ) = f :
1

|f (x)| < .
Clearly every function f (x) continuous on [a, b] is in L1 (a, b). Its also obvious that not
every continuous function is in L1 (R). Take, e.g., f (x) = 1.

|f (x)| dx =
dx =
and 1
/ L (R).
1
1
A typical example of an unbounded L1 -function is f (x) = . L1 (0, 1).
x
x
1
1
1
1
However
L1 (R). Next
/ L1 (0, 1), 2 L1 (1, ).
/ L1 (0, ). But
1 + x2
x
x
x
Make sure that you understand these things! Unfortunately, L1 is not a Hilbert
space.
Definition 17.13. The Fourier transform of a function f (x) L1 (R) is a function
f() defined by

1
eix f (x)dx.
F f () = f() =
(17.6)
2
Note that F : L1 (R) L (R) is bounded, but not F : L1 (R) L1 (R). Indeed,

1 ix
kF f k = max
e
f (x)dx .
R
2
And
ix

f (x)dx
ix
e
|f (x)| dx =
| {z }
|f (x)| dx = kf k1
=1
1
kf kL1 and so F : L1 (R) L (R) is bounded.
2
(
1
L x L
But now consider the box function f (x) =
for some fixed L > 0.
0
otherwise
2. SELFADJOINT AND UNITARY OPERATORS
105
We verify that f L1 (R) since
L
kf k1 =
|f | =
1 = 2L < .
L
We next compute
1
F f () =
2
f (x)e
ix
L
1 ix
=
e

2i
L
1
dx =
2
eix dx
L
r
eiL eiL
2 sin L
=
=
.

2i
But
r
r

r
sin x
2 sin L
1 2
2
2
|sin x|

dx =
kF f k1 =
dx.
d =

L R x
L 0
x
R
Using a result from analysis (beyond our scope),
r
r
2 2 X n |sin x|
2 2X
|sin y|
kF f k1 =
dx =
dy
L n=1 (n1) x
L n=1 0 y + (n 1)
r

2 2X 1
sin y dy = .
L n=1 n 0
Hence F : L1 (R) L1 (R) is unbounded.
Theorem 17.14. The Fourier transform defines a bounded operator on L2 (R). I.e.
f L2 (R)
f L2 (R).
Proof. The proof of this theorem requires advanced analysis which lies beyond
the scope of our course.
2. Selfadjoint and Unitary Operators
Selfadjoint and unitary operators are basically defined in the same way as in the
finite dimensional case. There is one difference related to the fact that Dom A 6= H
but we ignore it.
So see Definitions 10.15, 10.17, 10.25.
Note that a selfadjoint operator may be unbounded. Here is a very important
example.
d
Example 17.15. Consider A = 1i dx
in L2 (0, 2) and assume

Dom A = f L2 (0, 2) : f 0 L2 (0, 2) and f (0) = f (2) , || = 1 .
So A is an operator of differentiation with boundary conditions.
106
A is selfadjoint. Indeed let f, g Dom A then

* 0 2

2

1
1 0
1
d

hAf, gi =
f (x)g(x) dx = f (x)g(x)
f (x) g(x) dx

i
i
i
dx
0
0
0
2
1 d
f (x)
=
g(x) dx = hf, Agi .
i dx
0
But as we know, A is unbounded.
However every unitary operator U is bounded and moreover kU k = 1. Indeed, we

have
kU Xk = kXk
XH
max kU Xk = max kXk = 1
kXk1
kXk1
kU k = 1.
A good example of a unitary operator is the Fourier transform. We discuss it in

some detail later.
The boundedness of an operator plays a role for example when you consider applying the operator several times. E.g. A2 where A is as in Example 17.15 is A2 =
d2
dx
, A2 f = A(Af ) for f Dom A. But now f Dom A2 if also f 00 L2 (0, 2)
2
so Dom A2 is narrower. One can show though that A2 is selfadjoint, and hence so is
A2009 but Dom An keeps shrinking.
d2
A2 = dx
2 is also an important operator related to energy: the kinetic energy
(square of the momentum).
3. Spectrum
The situation with the spectra of operators in Hilbert spaces is a whole lot more
difficult. If in the finite case the spectrum of an operator was a finite set of eigenvalues,
the infinite dimensional picture is a lot richer.
Definition 17.16. A1 is called the inverse of A if
A1 A = AA1 = I.
Definition 17.17. Let A be a linear operator and C. Then the operator R (A)
defined by
R (A) (A I)1
is called the resolvent of A at a point .
The resolvent depends on analytically not by the definition given before as a
limit or by the Cauchy-Riemann conditions for matrix-valued functions, but by the
definition from the power series expansion or looking into the complex valued function hR (A)X, Y i. In finite dimensions, all definitions are equivalent, but in infinite
dimensions, one must use the Taylor expansion or the bilinear form.
Theorem 17.18. Let A = A then R (A) is a bounded operator C \ R.
Moreover, R (A) is an analytic function in C \ R.
The proof of this fundamental theorem is beyond our course.
3. SPECTRUM
107
Definition 17.19. The set of singularities of R (A) is called the spectrum of A

and denoted by (A).
So (A) = { C : (A I)1 is not analytic }. Note that in the finite dimensional case, the matrix becomes unbounded at eigenvalues, so for example Theorem
17.18 makes sense in the finite case since by Theorem 12.9, if A = A then (A) R.
Although there exist operators A with (A) = C, there are not common and very
pathological, so the resolvent is generally defined/analytic for at least some values of
. But whereas in the finite dimensional case, the set of eigenvalues is finite and so
the singularities isolated, one can have a branch cut type of singularity in the infinite
dimensional case.
Definition 17.20. The set of poles of R (A) is called the discrete spectrum of A
or eigenvalues of A and denoted d (A).
So d (A) = { C : (A I)1 has a pole }, and in finite dimensions, an operator only has a discrete spectrum, but in infinite dimensions, there exists also a
continuous spectrum.
If 0 d (A) it means that (A I)1 is undefined, i.e. A 0 I does not have an
inverse. It amounts to the following fact
Definition 17.21. If 0 d (A),
f0 H : Af0 = 0 f0
and f0 is called an eigenfunction of A corresponding to 0 .
LECTURE 18
Spectra of Selfadjoint Operators

In the previous lecture we defined the spectrum (A) of an operator A as the set
of C where R (A) is not analytic:
(A) = { C : R (A) is not analytic } .
In particular, the set of poles of R (A) is called the discrete spectrum of A and
denoted by d (A).
It is possible to prove that if 0 d (A) then the equation
AX = 0 X
(18.1)
has a solution X = f0 belonging to H.

Points of d (A) are called eigenvalues and corresponding solutions to (18.1) are
called eigenvectors or eigenfunctions of A.
Unfortunately, the spectrum need not consist of only a discrete spectrum and it
makes the theory a whole lot different from Linear Algebra.
But consider first an example of the case when (A) = d (A), i.e. the spectrum is
purely discrete.
Example 18.1. Let A =
1 d
i dx
in L2 (0, 2) with

Dom A = f L2 (0, 2) : f (0) = f (2)
where || = 1. Let us do the spectral analysis of A. Observe that in Example 17.15 we
showed that A is selfadjoint, so by Theorem 17.18, we expect (A) R. We have to
solve the equation:
1 0
u = u
i
u(0) = u(2)
u0 = iu
u = Ceix L2 (0, 2)
u(0) = C = Ce2i
|| = 1
e2i = 1
= ei for some R and hence we have
ei(2+) = 1 2 + = 2n , n Z
2n
=
n = n
.
2
2

So d (A) = n 2
. Corresponding eigenfunctions are
nZ
un = Cn ei(n 2 )x
109
x (0, 2).
110
18. SPECTRA OF SELFADJOINT OPERATORS
1
Note, if = 1 ( = 0) then d (A) = Z and un (x) = Cn einx . If we choose Cn =
2
then kun k = 1 and {un (x)} forms an ONB.
So let us analyze what we got!
o
n
The set of eigenfunctions of A is the set of simple harmonics 12 einx
we
nZ
studied previously. Its possible to prove that (A) = d (A) and hence the spectrum
of A is purely discrete: d (A) = Z.
So the spectrum of an unbounded operator need not be finite!
d
Example 18.2. Let A = 1i dx
in L2 (, ). In quantum mechanics this operator
is called the operator of momentum. Let us find d (A):
1 0
u = u u0 = iu
(18.2)
i
u(x) = Ceix .

ix 2
2
2
2

But kuk = |C|
e
dx = |C|
dx =
u(x)
/ L (R)
if C 6= 0
d (A) = .
Thus this operator has no discrete spectrum. On the other hand, (18.2) can be
solved R and solutions u (x) = Ceix are bounded functions but not in L2 (R).
Here we have a typical case of the so-called continuous spectrum. Moreover we
have (A) = R. (see Lecture 29 for more on this topic.)
Functions {u }R are called eigenfunctions of the continuous spectrum. It can be
proven that the resolvent R (A) of A is not analytic for R.
Note that we have to have
/ C \ R otherwise u would be unbounded. Now
{u }R is a kind of basis but not in the space, so its a basis for the Fourier transform.
The following theorem is utmost important!
Theorem 18.3. Let A be a sefadjoint operator in H with a purely discrete spectrum
d (A) = {n }. Then all n R and the set of its eigenvectors {Xn } forms an ONB in
H.
Proof. The proof of n R can be done in the same way as for finite dimensional
operators. It is also clear why hXn , Xm i = nm . But why {Xn } is a basis is a difficult
question!
Exercise 18.4.
a) Find eigenvalues and eigenfunctions of A defined in L2 (0, ) as follows
00
Au = u
u0 (0) hu(0) = 0
u0 () hu() = 0
where h is a real parameter.
d2
b) Show that A = 2 in L2 (, ) has no discrete spectrum.
dx
LECTURE 19
Generalized Functions. Dirac function.

1. Physical Motivation
Consider the following situation. We have a weightless thread with one bead
of mass m. In physics a distribution of mass
x0
is described by a density function.
x
m
It is natural to ask what is the density
0
1
function (x) describing the distribution of this mass. It looks like
(
0
, x 6= x0
(x) =
.
, x = x0
So (x) is identically zero everywhere but one point x = x0 . It means that
x0
1
(x)dx +
(x)dx = 0
(19.1)
x0
On the other hand, by properties of definite integrals,

1
x0
1
(x)dx +
(x)dx =
(x)dx = m
0
x0
(19.2)
Comparing (19.1) and (19.2), we get a contradiction.

This means that (x) is not a usual function and the way we defined it was incorrect
since for a normal function, the value at one point shouldnt matter for the integral.
This function should be defined as some limiting procedure. But its going to take a
while. We need a bunch of definitions.
2. Uniform Convergence
Let f (x) be a bounded function on an interval . We define
kf k max |f (x)| .
x
(19.3)
E.g. if = [0, 1] and f (x) = x2 , then

kf k = max x2 = 1 x2 = 1 on [0, 1].
x[0,1]
As we know, (19.3) defines a norm on C(). This norm is commonly referred to as

the sup-norm.
111
112
19. GENERALIZED FUNCTIONS. DIRAC FUNCTION.
Definition 19.1. A sequence of bounded functions {fn (x)} is said to be uniformly

convergent on to some function f (x) if
lim kf (x) fn (x)k = 0.
In this case they write

fn (x) f (x) on .
What does it actually mean that fn f ? This means that for any > 0 (no matter
how small) there is a number N such that all members of the sequence {fn (x)}
n=N do
not deviate from f (x) by more than , i.e.
kf (x) fn (x)k ,
n N.
Example
19.2.

1
1)
0 on R. Indeed,
x2 + n2 n=1

1
1
1
= max

0
=
0 ,

x2 + n2
xR x2 + n2
n2
(
0 , 1/n < x < 1
2) Let fn (x) =
.
n , 0 < x 1/n
It is clear that
x (0, 1)
lim fn (x) = 0
kfn (x) 0k = max |fn (x)| = n

x(0,1)
n .
n
but
1
n
fn (x) 6 0 on (0, 1).
It is natural to ask why we need uniform convergence? Uniform

convergence is
much better than the usual one! It lets us interchage lim and . Namely,
Theorem 19.3. Let be any interval and {fn (x)} be a sequence of integrable
functions. If fn (x) f (x) and f (x) is integrable then
f (x) dx.
lim fn (x) dx =
lim
fn (x) dx =
n
Proof. (For the case when is bounded.) We need to show

lim
fn (x) dx
f (x) dx = 0.
n
But

fn (x) dx
= (fn (x) f (x)) dx
f
(x)
dx

|fn (x) f (x)| dx

max |fn (x) f (x)| dx
=
kfn f k dx = kfn f k
dx
= kfn f k 0 , n .
QED
2. UNIFORM CONVERGENCE
113
Remark 19.4. Uniform convergence in Thm 19.3 is essential. Indeed, let {fn } be
the sequence defined in Example 19.2 2). It is clear that
1
1
1
0 dx = 0.
lim fn (x)dx =
fn (x)dx = 1 but
0 n
This means you cannot interchange lim and
Definition 19.5. A sequence of bounded functions {n (x)} subject to

1)
n (x) 0
n (x)dx = 1
2)
3) lim n (x) = 0 uniformly on R \ (, ), > 0

n
is called a -sequence.
n
2 2
Example 19.6. n (x) = en x . {n } is a -sequence!
Indeed, n (x) are bounded and continuous n, n (x) 0,

> 0
Finally,
1
n
n (x) = lim n2 2 = 0.
n xR\(,)
n e
lim
n
n (x)dx =
max
n2 x2
1
dx =
ey dy = 1.
Exercise 19.7. Show that {fn (x)} defined in Example 19.2 2) is a -sequence.
Definition 19.8. The support of a function f (x), denoted by Supp f , is the set of
all x : f (x) 6= 0. That is,
Supp f = {x : f (x) 6= 0}.
Example:
f (x)
Supp f
Definition 19.9. A function f (x) is called finitely supported if there exists a finite
interval such that Supp f .
114
Example:
f (x)
Supp f
Definition 19.10. A function f (x) is called smooth on a set if all derivatives
of f are continuous functions.
C () stands for the set of all smooth functions on .
2
Example: f (x) = ex , sin x, x2 x + 1 are smooth functions on R.
We are going to use the following notation:
C0 (R) = {all functions in C (R) having finite support}.
For instance,
f (x) C0 (R)
Functions from C0 (R) will be referred to as test functions.

In a similar way we define C0 () as the set of all smooth functions on having
supports entirely inside .
Example:
is such a function.
Lemma 19.11. If {n (x)} is a -sequence then

lim
n (x)dx = 1 , > 0.
n
Proof. We have
1=
n (x)dx =
n (x)dx +
lim
n (x)dx = 1 lim
n (x)dx
R\(,)
n (x)dx.
R\(,)
2. UNIFORM CONVERGENCE
115
Thm 19.3
But n 0 on R \ (, )
n (x)dx =
0 dx = 0
lim
R\(,)
R\(,)
n (x)dx = 1 ,
lim
> 0.
QED
The following theorem is very important.

Theorem 19.12. Let {n (x)} be a -sequence. Then

n (x)(x)dx = (0) , C0 (R).
lim
n
Proof. For every C0 (R),

n (x) ((x)dx (0)) dx
n (x)(0)dx +
n (x)(x)dx =

= (0) +
n (x) ((x) (0)) dx.
We need to show that the last integral goes to 0 when n . It will take a while!
Let be any positive number. We have

Triangle ineq.
.

(19.4)
+
=

+

R\(,)
R\(,)
Estimate each of these integrals separately.

n (x) |(x) (0)| dx

n (x)dx
max |(x) (0)|
x(,)
|
{z
}
1
max |(x) (0)| .
(19.5)
x(,)
Since C0 (R), lim |(x) (0)| = 0.

x0
That is, choosing > 0 we can make max |(x) (0)| where is as small
x(,)
as we want. Next,
R\(,)
n (x) |(x)
R\(,)
(0)| dx
k(x) (0)k
n (x)dx.
(19.6)
R\(,)
Now plug (19.5) and (19.6) in (19.4).

n (x) ((x) (0)) dx max |(x) (0)| + k(x) (0)k
n (x)dx .
x(,)
R\(,)
|
{z
}
|
{z
}
< (by choosing )
0, >0
(by Lemma 19.11)
116
So,

lim
n (x) ((x) (0)) dx < ,
n

lim
n (x) ((x) (0)) dx = 0.
>0
QED
3. Weak Convergence
Definition 19.13. A sequence {fn (x)}
n=1 of functions integrable
on is said
to
converge weakly if for every C0 () the numerical sequence fn (x)(x)dx n=1

converges.
w
Notations for weak convergence: fn f , or fn * f .

Theorem 19.14. Every -sequence converges weakly.
Proof. By Theorem 19.12, for any -sequence {n },

lim
n (x)(x)dx = (0) , C0 (R).
n
n (x)(x)dx
That is,
converges
Def. 19.13
{n } converges weakly.
QED
n=1
So, we have three types of convergence: usual (pointwise), uniform, and weak.
pointwise
fn f on iff
uniform
fn f on iff
n
weak
fn converges
weakly on iff
lim fn (x) = f (x) max |fn (x) f (x)| 0
fn (x)(x)dx
lim
exists C0
Theorem 19.15. If {fn (x)} converges to f (x) on uniformly then {fn (x)} converges to f (x) weakly.
Proof. We are supposed to prove that
lim
fn (x)(x)dx =
f (x)(x)dx.
n
By condition fn (x) f (x), i.e.

kfn f k 0, n .
We have
fn (x)(x)dx =
(fn (x) f (x)) (x)dx +

|
{z
}
Consider this integral
f (x)(x)dx.
3. WEAK CONVERGENCE

(fn (x) f (x)) (x)dx
|fn (x) f (x)| |(x)| dx

|(x)|dx
max |fn (x) f (x)|
x
= kfn f k
|(x)|dx 0.
lim
fn (x)(x)dx =
f (x)(x)dx.
n
117
QED
Remark 19.16. The converse of Theorem 19.15 is not true. That is, weak convergence does
notimply uniform convergence. We
providea counterexample.
inx
e
einx
Let
on (, ). We claim that
converges weakly to 0.
2 nZ
2
Indeed, let be a test function. Then
einx
(x) dx = cn , n Z ,
2
are Fourier coefficients of (x). Since L2 (, ), by the Fourier Theorem, we

have
X
|cn |2 = kk2 < .
nZ
Hence lim cn = 0, n . So we get
einx
(x)dx 0 , n
cn =
2
n inx o
e
and by definition,
converges weakly to 0.
2
On the other hand,
inx

inx
e

e
0 =
= 1 6= 0
nN
2

2
2
inx
inx
e
e
and hence
does not converge to 0 uniformly. Moreover,
does not
2
2
converge even pointwise.
So,
Uniform convergence
Weak convergence
:
So we know now that a -sequence converges weakly. The function it converges

to is called a generalized function.
Definition 19.17. The limit of a -sequence is called the Dirac -function and
denoted as (x). I.e.,
w
n .
118
We agree to write it in the symbolized form

n (x)(x)dx =
lim
n
(x)(x)dx.
Example 19.18. We explore some properties of the Dirac -function. In the following, is any test function.
1) x(x) = 0, yet neither x, nor (x) are 0 on an interval.

(x)(x(x))dx = 0 (0) = 0.
x(x)(x)dx =
2) (x a)
(x a)(x)dx = (a).
R
a
0
3) (x a) + (x b)
[(x a) + (x b)] (x)dx = (a) + (b).

R
4) c(x)
(x)(c(x))dx = c(0).
c(x)(x)dx =
R
5) 2 (x) is undefined; its not possible to describe it from our definition.

1
6) (cx) = (x), c 6= 0. Use a change of variables:
c

y = cx
1
= lim
n (cx)(x)dx =
(y)(y/c) dy
lim
1

dx = c dy
n
n
c

1
1
= (0) =
(x)(x)dx.
c
c
7) (x a)(x b) = 0, a 6= b (if a = b, DNE)

(x a)(x b)(x)dx = (a b) (a) = 0.
| {z }
Exercise 19.19.
=0
n2 (x)(x)dx does not exist.
Formally prove lim
Surprisingly, is smooth, i.e. infinitely differentiable. You cant square it but you
can differentiate it infinitely many times!
4. GENERALIZED DERIVATIVES
119
4. Generalized Derivatives
Definition 19.20. f 0 (x) is called a generalized derivative of f (x) if
f (x)0 (x)dx
f (x)(x)dx =
for all C0 (R).

1, x 0
(Heaviside function). The generalized
0, x < 0
Theorem 19.21. Let (x) =

derivative of (x) is (x).
Proof. For any C0 (R) we have
(x) (x)dx +
(x)0 (x)dx
(x) (x)dx =
0
0
}
| {z
=0

0
=
(x)dx = (x) = () (0) = (0).
| {z }
0
0
(x) (x)dx =
=0
So
Def.
(0) =
(x) (x)dx =
(x)(x)dx
On the other hand,
0 (x)(x)dx = (0).
n (x)(x)dx.
(0) = lim
Hence, 0 (x) = (x) in the sense of generalized functions.

Exercise 19.22. Prove:
1) (x) = (x),
1
2) (x2 a2 ) =
((x a) + (x + a)),
X2a
3) (sin x) =
(x n).
nZ
Theorem 19.23. Let {n } be a -sequence. Then
n(m) (x)(x)dx = (1)m (m) (0).
lim
QED
120
Proof.

d (m1)
n
(x)(x)dx
dx
:0

by parts (m1)

= n
n(m1) (x)0 (x)dx
(x)(x)

= . . . = (1)m
n (x)(m) (x)dx
m
d
Thm 19.12
lim
n (x)(x)dx = (1)m (m) (0).
n dxm
m

d
Definition 19.24. Let {n (x)} be a -sequence. The weak limit of
n (x)
dxm
is called the mth derivative of (x) and denoted (m) (x).

0 (x)dx = 0. Support is at one point, and the average
Note that in particular,
(m)
n (x)
(x)dx =
is zero. Weird!
n
2 2
But consider for example, the sequence n (x) = ex n ; it is infinitely differen
tiable and converges weakly to (x). Furthermore, one can observe that
3
2
1
10
the derivative at 0 is 0,
and the average of the derivative is indeed 0.
20
30
LECTURE 20
Representations of the -function
1. Fourier Representation of the -function
1
, x
.
Let = 2
1
0
2
, x (, ) \ [, ]
{ } is clearly a -sequence.1
Since L2 (, ) we can expand it into a
Fourier series:

X
1
inx
cn e
, cn =
(x) =
(x)einx dx.
2
nZ
We have

1
1 inx
1
cn =
e
dx =
einx dx
2 2
4

1 einx
1 ein ein
1
=
=
=
sin n.
4 in 2n
2i
2n
So cn =
1 1
sin n, and
2 n
(x) =
1 X 1
sin n einx .
2 nZ n
Take the limit as 0 of the right hand side of (20.1)

X sin n
1 X
sin n inx
1 X inx
1
lim
einx =
lim
e =
e
2 0 nZ n
2 nZ 0
2 nZ
| {zn }
=1
and it looks like

(x) =
1{
}0<<
1 X inx
1
1X
e =
+
cos nx.
2 nZ
2 n1
is more like a -family but no one cares about the difference.

121
(20.1)
122
20. REPRESENTATIONS OF THE -FUNCTION
Unfortunately this series diverges for all x. But it can be understood in the weak
sense. Indeed test function (x),

X
N2
N2
1 X
1
inx
einx (x)dx
e (x)dx =
2
2
n=N
n=N
1
N2
X
1
=
2
n=N1 |
We show now that the partial sum
N2
X
inx
N2
X
cn .
e (x)dx =
N
1
{z
}
(20.2)
=cn
cn converges absolutely.
N1
Lemma 20.1. If C0 (, ) then
|cn | converges.
nZ
Proof. For any n 6= 0,

1
1
deinx
inx
cn =
e
(x)dx =
(x)
2
2
in

: 0 1
1
by parts

inx

(x)e
+
einx 0 (x)dx
=

2in
2in

inx
1
de
1
=
0 (x)
=
einx 00 (x)dx
2in
in
2(in)2

1
(x)dx.
and c0 =
2
Note that all integrated terms are 0 since our function (x) is a test function and
hence Supp (, ). So

1 1
cn =
einx 00 (x)dx
2 n2

1 1 inx 00
|cn |
2
e
(x)dx
2 n

1 1
1
2
|00 (x)| dx C 2 , n 6= 0
2 n
n
X
X 1
|cn | C
< .
QED
n2
n6=0
n6=0
nZ
nZ
So by Lemma 20.1, series (20.2) converges and hence

!

N2
inx
X
X
e
lim
(x)dx =
cn .
N1
2
n=N
nZ
N2
(20.3)
2. INTEGRAL REPRESENTATION OF THE -FUNCTION
On the other hand, by the Fourier Theorem

X
X
X
(x) =
cn einx =
cn einx and (0) =
cn .
nZ
nZ
nZ
So (20.3) transforms into
N2
X
einx
2
n=N
lim
N1 ,N2
(x)(x)dx.
(x)dx = (0) =
and we can conclude that

Theorem 20.2.
N2
X
einx
(x) = w-lim
N1 ,N2
2
n=N
1
and in the weak sense

(x) =
1
1X
+
cos nx.
2 n1
2. Integral Representation of the -function

Lemma 20.3.
1
lim
h 2i
Proof.
a+ih
aih
exz dz
= (x)
z
Heaviside function.
1) For x < 0, consider the following contour:

ih
ezx
H(Int R ) (analytic
z
within R ) then
Since
a+ih
zx
ezx
e
dz = dz.
z
z
aih
ih
ezx
dz =
z
1 dezx
z x
by parts
a+ih
zx
1 zx
1
e
e
+
dz.
zx
x
z2
aih
123
124
The integrated term

a+ih

ihx

1 zx
1 eax+ihx eaxihx
eax
e
eihx
e
=
zx aih x a + ih
a ih
x a + ih a ih
|
{z
}
0 , h
a+ih
1 zx
lim
= 0.
e
h zx
aih
Next,

zx
1

zx
e
|e |
1
1 ee Re z
1 ex Re z
ex Re z

.
dz
|dz|
=
|dz|
R
=

x
z 2 |x|
|x| R2
|x| R62
|x| R
|z|2
So

1
ex Re z
zx
e

dz
.

2
x

z
|x| R
(20.4)
If x < 0 then ex Re z < 1 (since Re z > 0)
zx
e
So lim dz = 0 ,
hz
(20.4) 0 ,
R .
a+ih zx
1
e
x < 0 and lim
dz = 0 ,
h 2i aih
z
x < 0.
2) For x > 0, consider a different contour:

As we did before for x < 0
exz
dz =
z
ih
by parts
0R
ih
eax
x
|
eihx
eihx
1
+
a + ih a ih
x
{z
}
0 , h
| {z }
I1
Consider these integrals separately.
| {z }
I2
ezx
dz.
z2
2. INTEGRAL REPRESENTATION OF THE -FUNCTION
1
|I2 |
|x|
125
zx
x Re z
e
|dz| = 1 e
R.
z2
|x| R62
R
R
Now ex Re z < 1 since x > 0 but Re z 0 and

1
0 , R .
|x| R
zx
e
1
ex Re z |dz|
1
|dz| =
|I1 |
.
z2
|x|
|x|
R2
|I2 |
But since Re z a this estimate can be continued as

|I1 |
exa
1 exa
R
=
0 ,
|x| R62
|x| R
R .
= 0.
So lim
By the Residue Theorem

xz
1
exz
e
zexz
dz = Res
, 0 = lim
= 1.
z0 z
2i 0R z
z

1
2i
k
a+ih
aih
xz
e
1
dz +
z
2i
exz
dz
z
Taking h , we get
a+ih xz
1
e
lim
dz = 1 ,
h 2i aih
z
and the lemma is proven.
QED
Now we are ready to establish

Theorem 20.4.
1
(x) =
2
x>0
eixt dt.
126
Proof. By Theorem 19.21

(x) =
d
(x).
dx
By Lemma 20.3
a+ih zx
e
1
dz.
(x) = lim
h 2i aih
z
Differentiating (x) (formally), we have
a+ih
a+ih
d
1
1
d ezx
(x) = lim
dz = lim
ezx dz.
h 2i aih dx z
h 2i aih
dx
zx
Since a is arbitrary and e analytic

a+ih
ih
h

z = it
zx
zx

e dz
=
e dz =
eixt dt
=i
t
=
iz
aih
ih
h
a=0

1
d
(x) =
eixt dt.
dx
2
QED
LECTURE 21
The Sturm-Liouville Problem

Spectral methods to solve (linear) PDEs call on transforms (Fourier is just one type
of transform). The PDE is solvable if we can reduce it to an ODE.
Definition 21.1. Let p(x), w(x), q(x) be real-valued functions defined on R. The
expression A defined by

d
1 d
p(x) y + q(x)y
Ay =
w(x) dx
dx
is called the Sturm-Liouville operator.
w(x) is the weight function, p(x) > 0, and q(x) is called the potential.
This operator A is clearly linear. But we have to specify a space where A acts.
Definition 21.2.
1
0
(py 0 ) + qy = y
(21.1)
w
where w, p, q are known functions and is a parameter is called the Sturm-Liouville
equation.
Equation (21.1) always comes with some boundary conditions which make it a
Sturm-Liouville problem. Equation (21.1) is considered on a finite interval (a, b), halfline (, a) or (a, ), or the whole line (, ). We start with the finite interval
case. In literature this case is also called regular.
Definition 21.3. The Sturm-Liouville problem on (a, b)
d
d
p(x) y + w(x)q(x)y = w(x)y
dx
dx
y(a) = 0 = y(b)
is called a Dirichlet problem and the condition y(a) = 0 = y(b) is called Dirichlet
conditions.
Definition 21.4.
d
d
dx
dx
y 0 (a) = 0 = y 0 (b)
is called a Neumann problem.
127
128
21. THE STURM-LIOUVILLE PROBLEM
Definition 21.5.
d
d
dx
dx
y 0 (a) + y(a) = 0 = y 0 (b) + y(b) ,
, R
is called a Robin problem.

Note that Def. 21.5 includes 21.3 and 21.4. Indeed, if , = in Def 21.5 then it
transforms into 21.3. Also, Def 21.4 is Def 21.5 with = = 0.
Note also that we cant solve explicitly any of the problems above unless w, p, q are
constant or special functions (like Bessel functions). And the original problem can not
be handled by treating each spectra separately then putting them together.
Before showing seladjointness, we need to know which space and so which inner
product.
Definition 21.6. Let w(x) 0 on (a, b) and let

2
w(x)|f (x)| dx < .
L (, w) f :
2
L2 (, w) is called a weighted L2 -space and w is called a weight function.
hf, gi =
w(x)f (x)g(x) dx
is a weighted scalar product.

Once w 0 then hf, gi is indeed a scalar product.
Theorem 21.7. Let w 0, p(x), q(x) R. The operator
A=
1 d
d
p(x) + q(x)
w(x) dx
dx
defined on

DomA f L2 (, w) : f 0 (a) + f (a) = 0 = f 0 (b) + f (b), , R
is self-adjoint.
129
Proof. Since p, q R we have p = p and q = q so

1 d
0
hAf, gi =
w(x)
p(x)f (x) + q(x)f (x) g(x) dx
w(x) dx

d
0
=
p(x)f (x) g(x) dx +
w(x)q(x)f (x)g(x) dx
dx
|
{z
}
by parts

d
0

= p(x)f (x)g(x) +
w(x)f (x)q(x)g(x) dx
p(x)f (x) g(x) dx +
dx
|
{z
}
by parts

d
0
0

f (x)p(x)g (x) dx +
w(x)f (x)q(x)g(x) dx
= p(x)f (x)g(x) +
dx

d
0
= pf g + f pg 0
f
pg 0 dx +
wf qg dx
dx

1 d 0
0
0
= p f g + f g +
wf
pg + qg dx .
w dx
{z
}
|
{z
} |
0
=hf,Agi
Compute this separately

p f g + f g =

= p(b) f 0 (b) g(b) + f (b) g 0 (b) p(a) f 0 (a) g(a) + f (a) g 0 (a)
|{z}
| {z }
| {z }
| {z }

g(b)
f (b)
g(a)
f (a)
= p(b) f (b)g(b) f (b)g(b) p(a) f (a)g(a) f (a)g(a)

|
{z
}
|
{z
}
=0
=0
= 0.
So, we get hAf, gi = hf, Agi.
QED
Note, that if , C or p, q C, then A is not self-adjoint.

Definition 21.8. The Sturm-Liouville operator with w(x) = p(x) = 1 is called a
Schrodinger operator and is commonly denoted by H, i.e.,
H=
d2
+ q(x).
dx2
Actually any Sturm-Liouville problem can be rewritten with H when put in the
canonical form. But it likes the whole line, so often if we have just an interval, the
operator is called Sturm-Liouville and if its on the whole line, the operator is called
Schrodinger.
130
Theorem 21.9 (Canonical Form of the Sturm-Liouville Operator). Assuming that

p(x), w(x) > 0 on [a, b] and p C1 [a, b], pw C2 [a, b] then the Sturm-Liouville problem
1 d
dy
p(x) + q(x)y = y
(21.2)
w(x) dx
dx
can be tranformed into the Schrodinger problem
d2 u
2 + q(z)u = u
(21.3)
dz
by a suitable substitution.
Proof. Rewrite (21.2) as

dy
1 d
p
+ (q )y = 0.
w dx dx
(21.4)
Let
s
s
1 x w(s)
1 b w(x)
z=
ds , c =
dx
c a
p(s)
a
p(x)
be our new variables. One has
1/2
1/2
1 w
d
d
1 w
=
dx
dz =
c p
dx
c p
dz
and (21.4) reads
1 1
2
c w
|
Introduce = (wp)1/4
1/2
1/2
w
w
dy
d
p
+ (q )y = 0.
p
dz
p
dz
| {z }
{z }
1
(wp)1/2
(21.5)
=(wp)1/2
u = y and (21.5) transforms into

1
d
d
2 2 2 1 u + (q )1 u = 0
c
dz dz
0
1

2 (1 u)0 + c2 (q )u = 0.
(21.6)
But
2 (1 u)0
0
0
= 2 (1 u0 2 u) = (u0 0 u)0
00
00
0 + u00 00 u 0 u
0
=
0 u
= u u
and so for (21.6) one has

1 (u00 00 u) + c2 (q )u = 0
u00 + (1 00 + c2 q) u = |{z}
c2 u.
|
{z
}
=
q
QED
Theorem 21.10. The spectrum of the operator A in Theorem 21.7 is discrete and
simple.
(No proof).
131
Remark 21.11. Since A = A and (A) = d (A) then the set of its eigenfunctions
{yn (x)} forms an ONB in L2 (, w).
LECTURE 22
Legendre Polynomials
d
d
(1x2 ) on L2 (1, 1), i.e. where
dx
dx
2
w(x) = 1 , p(x) = 1 x , q(x) = 0.
Consider a specific Sturm-Liouville operator
Theorem 22.1. The operator

A=
is selfadjoint on
d
d
(1 x2 )
dx
dx

Dom A = y C[1, 1] : Ay L2 (1, 1) .
Note that it is possible to extend this operator to a larger domain and still show
selfadjointness but the proof is more complicated.
Exercise 22.2. Prove that A defined above is selfadjoint for y C 1 [1, 1].
Let us consider the eigenfunction problem for A. The spectrum of A is expected
to be discrete, and we will find that the solutions associated to each eigenvalue in the
spectrum are the so-called Legendre polynomials.
Ay = y
dx

2 d
(1 x ) y = y.
dx
This equation is called the Legendre equation. One can check that it is equivalent
to
d2
dy
(1 x ) 2 y 2x + y = 0,
(22.1)
dx
dx
which is a second order linear homogeneous equation. Let us solve it using power series,
i.e. by the Frobenius method (see Appendix in this Lecture). Remark that x = 1 are
regular singular points since for example for x = 1, (22.1) can be rewritten as
2
y 00 +
q(x)
p(x) 0
y +
y=0
x1
(x 1)2
with
p(x) =
2x
1x
, q(x) =
.
x+1
1+x
But here were not going to use the expansion at the regular singular points but
rather at x0 = 0, i.e. an ordinary point. So we are looking for a solution to (22.1) in
1Recall
that C[1, 1] is the set of continuous functions f on [1, 1]; so in particular lim f (x)
x1
are finite.
133
134
22. LEGENDRE POLYNOMIALS
L2 (1, 1) in the form

y=
cn x n
with
y C[1, 1].
n0
We have
(1 x2 )
n(n 1)cn xn2 2x
n0
ncn xn1 +
n0
n(n 1)cn xn2
n0
cn x n = 0
n0
n(n 1)cn xn
n0
2ncn xn +
n0
cn x n = 0
n0
n
[(n + 2)(n + 1)cn+2 (n(n 1) + 2n )cn ] x = 0.
n0
So by shifting the index, we get

cn =
(n 2)(n 1)
cn2 , n = 2, 3,
n(n 1)
(22.2)
which gives us a recursion formula for {cn }. Separating odd and even powers, one has
X
X
y(x) = A
c2n x2n + B
c2n+1 x2n+1
(22.3)
n0
n0
where A, B are arbitrary constants and for {cn }, and by (22.2), we explicitly have
23
23
, c4 =
c2 =
()
2
43
234
45
(4 5 )(2 3 )()
c6 =
c4 =
,
56
23456
12
,
c1 = 1 , c 3 =
23
From Frobenius theory, we also know that each power series converges for |x| < 1.
Let us check what is going on at the endpoints.
Note that if is an integer of the form n(n + 1), then eventually cn+2 = 0 and then
all the following coefficients will be zero, and we get the Legendre polynomials. We
will show that if not, the corresponding function blows up at one of the endpoints (and
hence is not in the domain of A).
We rewrite the even coefficients as follows
( n1
)
Y
c2n =
1
.
2m(2m
+
1)
2n
m=1
c0 = 1 , c 2 =
Next we use the following theorem offered here without a proof

Y
Theorem 22.3. The infinite product
(1 zm ) converges absolutely if and only
mN
if the series
X
mN
zm converges absolutely.
135
Since
N
X
2N
+1
X
||
(1)k
= ||
,
2m(2m
+
1)
k
m=1
k=2

Y
converges absolutely and hence so does

1
.
then
2m(2m + 1)
2m(2m + 1)
mN
mN
Therefore, for sufficiently large n,
c()
c2n
where c() is a constant depending on .
n
X
But if that is the case, then the part of the solution
c2n x2n is infinite at x = 1.
X
n0
A similar reasoning can be used for the odd terms. So we find that if the sum is
indeed infinite, i.e. if 6= n(n + 1) for some nonnegative integer n, then the solution
y
/ Dom A.
Theorem 22.4. The spectrum (A) of
d
d
A = (1 x2 )
dx
dx
is purely discrete and
(A) = {n(n + 1)}nN0
L2 (1, 1)
on
N0 = {0, 1, 2, } .
0
2
6
Corresponding eigenfunctions, denoted Pn are called Legendre polynomials.
plicitly
P0 (x) = 1 ,
P1 (x) = x
P2 (x) =
3x2 1
2
P3 (x) =
5x3 3x
2
P0 (x) = 1
P1 (x) = x
P2 (x) =
3x2 1
2
3
P3 (x) = 5x 23x
Ex
136
Remark 22.5. Note that we dont have to impose any boundary conditions for
the operator A in Theorem 22.4. The requirement that solutions be in C[1, 1] is a
condition (but not boundary).
There is a nice formula, called Rodrigues formula, for computing Legendre polynomials:
1 dn 2
Pn (x) =
(x 1)n , m N0 .
n!2n dxn
Exercise 22.6. Use Rodrigues formula to compute
1
Pn2 (x)dx
1
where Pn are Legendre polynomials.

Since our operator A is selfadjoint, then {Pn }nN0 forms an orthogonal basis in
L (1, 1). Note that this is not quite an orthonormal basis; indeed the norm depends
on n (see Exercise 22.6).
Nonetheless, this fact plays a fundamental role in physics. Recall that in P[1, 1],
the set of all polynomials, we knew the basis {1, x, x2 , } but it was not an orthogonal
basis. So the Legendre polynomials {Pn } are more useful, and they form a basis for a
bigger space, L2 (1, 1) so
X
f L2 (1, 1) : f (x) =
fn Pn (x).
2
n0
In Numerical Analysis, it is particularly useful to have an orthogonal basis; in physics,

they come from some PDEs, and having a lot of orthogonal bases to choose from allows
us to pick the one related to the problem at hand.
Exercise 22.7. Use binomial series to formally show that (Pn are again Legendre
polynomials)
X
(1 2xt + t2 )1/2 =
Pn (x)tn .
n0
(The expression on the left hand side is called the generating function of Legendre
polynomials.)
Appendix. Frobenius Theory
Consider the foolowing ODE
y 00 + P (x)y 0 + Q(x)y = 0.
(22.4)
We can expand P (x) and Q(x) in power series and solve by the method of undetermined coefficients, but this becomes unbelievably unwieldly.
Example 22.8. Consider y 00 + y = 0. Write
X
y=
cn xn
n0
APPENDIX. FROBENIUS THEORY
137
where we assume this series is absolutely convergent. Hence it is uniformly convergent

inside a ball, so we can differentiate:

X
X
k = n 2 X
00
n2
n2
=
y =
cn n(n 1)x
=
cn n(n 1)x
=
(k + 1)(k + 2)ck+2 xk .
n = k + 2
n0
n2
k0
So we get
X
X
X
(k + 1)(k + 2)ck+2 xk +
ck xk = 0
{(k + 1)(k + 2)ck+2 + ck } xk = 0.
{z
}
|
k0
k0
k0
Now we need to solve this recursive relation

ck
ck+2 =
(k + 1)(k + 2)
=0
k = 0, 1, 2,
Note that
c2 =
c3 =
c0
2
c4 =
c6 =
c4
(1)3 c0
=
56
23456
(1)n
c0
(2n)!
c2n =
c1
23
c2
(1)2 c0
=
34
234
c2n+1 =
c5 =
c3
(1)2 c1
=
45
2345
(1)n
c1 .
(2n + 1)!
So
y = c0
X (1)n
n0
(2n)!
2n
X (1)n
+ c1
x2n+1 = c0 cos x + c1 sin x
(2n + 1)!
n0
as expected!
Note that y is convergent on C, and hence y is entire.

Definition 22.9. x0 is called an ordinary point of (22.4) if P, Q are analytic on
some neighborhood about x0 , DR (x0 ) = {z : |z x0 | < R}.
Theorem 22.10. If x0 is an ordinary point of (22.4), then (22.4) has a power
series solution
X
y(x) =
cn (x x0 )n
(22.5)
n0
absolutely convergent on DR (x0 ). Moreover, the general solution to (22.4) has the form
of (22.5).
Example 22.11. Consider Stokes equation:
y 00 xy = 0.
Note that any point x0 is ordinary since P, Q are entire. Hence we can choose x0 = 0
for simplicity. We have
X
X
y=
cn x n
,
y 00 =
(k + 1)(k + 2)ck+2 xk .
n0
k0
138
Then
X
(k + 1)(k + 2)ck+2 xk
k0
n0
cn xn+1
{z
=0
reindex n+1=k , n=k1
X
X
(k + 1)(k + 2)ck+2 xk
ck1 xk = 0
k0
2c2 +
k1
{(k + 1)(k + 2)ck+2 ck1 } xk = 0.
k1
So c2 = 0 and we get the recursion relation (k+1)(k+2)ck+2 ck1 = 0 for k = 1, 2, .

Reindexing
ck
,
k = 0, 1, 2,
ck+3 =
(k + 2)(k + 3)
Then
c3
c0
c6
c0
c0
, c6 =
=
, c9 =
=
c3 =
23
56
2!
356
89
235689
n1
Y
c0
c3n =
(3k + 1)
(3n)!
k=1
c4
c1
c1
, c7 =
=
c4 =
34
67
3 4! 6 7
n
Y
c1
c3n+1 =
(3k 1)
(3n + 1)!
k=1
c2 = 0
c3n+2 = 0
for any n = 0, 1, 2, Thus

y(x) = c0
c3n x3n + c1
n0
c3n+1 x3n+1
n0
also known as Airy functions A(x), B(x). These are not elementary functions, but they
are special functions. And so sometimes, the Stokes equation is referred to as part of
the Airy family.
Exercise 22.12. Use the power series method to solve
y 00 xy = 1
y(0) = y 0 (0) = 0.
Definition 22.13. If x0 is not an ordinary point of (22.4), i.e.2

y 00 + P (x)y 0 + Q(x)y = 0
then its called a singular point.
What happens around singular points in general? We dont know, we cant offer a
treatment for a generic singular point. We need to pick a good bad guy.
2Recall
that this form is called the standard form; the canonical form would be: y 00 = q(x)y.
139
Definition 22.14. A point x0 is called a regular singular point of (22.4) if (x

x0 )P (x), (x x0 )2 Q(x) are analytic in some neighborhood of x0 . That is, (22.4) can
be represented as
p(x) 0
q(x)
y 00 +
y +
y=0
(22.6)
x x0
(x x0 )2
where p(x) = (x x0 )P (x), q(x) = (x x0 )2 Q(x) are analytic; or alternately
(x x0 )2 y 00 + (x x0 )p(x)y 0 + q(x)y = 0.
(22.7)
Note that there would be no loss of generality to take x0 = 0 since we can change
variables.
Definition 22.15. If x0 is not a regular singular point, then x0 is an irregular
singular point.
Example 22.16. Consider
(x2 4)2 y 00 + (x 2)y 0 + y = 0
(x 2)2 (x + 2)2 y 00 (x 2)y 0 + y = 0

y 00 +
p(x)
q(x)
+
y=0
x 2 (x 2)2
1
where p(x) = q(x) = (x+2)
So x = 2 is a regular singular point, but x = 2 is
2.
an irregular singular point, and all other points are ordinary. There are treatments
for some irregular singular points but they cause severe problems and are not in most
books.
Theorem 22.17 (Frobenius Theorem). If x0 is a regular singular point of (22.7)

then there exists at least one power series solution of the form
X
y(x) = (x x0 )r
cn (x x0 )n
for some r C.
n0
The series converges on x0 < x < x0 + R for some R > 0.

Note that R is related to the neighborhood of analyticity of p, q but it could be
wider. Plus here, x0 < x < x0 + R so x x0 > 0, and its easier to deal with (x x0 )r .
But both solutions need not be of this form as demonstrated by the following two
examples.
Example 22.18 (Two series solutions). Consider the following equation
3xy 00 + y 0 y = 0
1
x
y 00 + y 0 2 y = 0
3x
3x
x = 0 is a regular singular point.
So by the Frobenius theorem,
X
X
X
y(x) = xr
cn xn , y 0 (x) =
cn (n+r)xn+r1 , y 00 (x) =
cn (n+r)(n+r1)xn+r2 .
n0
n0
n0
140
So
0 = 3xy 00 + y 0 y
X
X
X
=
3cn (n + r)(n + r 1)xn+r1 +
cn (n + r)xn+r1
cn xn+r
n0
n0
= c0 xr1 (3r(r 1) + r) +
n0
{3ck (k + r)(k + r 1) + ck (k + r) ck1 } xk+r1
k1
r(3r 2) = 0
ck =
ck1
(k + r)(3k + 3r 2)
k = 1, 2,
The equation r(3r 2) = 0 is referred to as the indicial equation and leads to two
possible solutions: r1 = 2/3 and r2 = 0.
r1 = 2/3
ck1
3ck1
ck1
=
,
=
3k(3k + 2)
k(3k + 2)
k + (3k + 2 2)
c0
c0
ck = Q k
=
k!5 8 (3k + 2)
n=1 {n(3n + 2)}
c1
c0
c0
, c2 =
=
,
c0 6= 0 , c1 =
5
28
258
!
n
X
x
y1 (x) = c0 x2/3 1 +
,
n!5 8 (3n + 2)
n1
ck =
2
3
k = 1, 2,
x R.
r2 = 0
ck =
ck1
k(3k 2)
k = 1, 2,
y2 (x) = c0 1 +
X
n1
xn
n!1 4 7 (3n 2)
!
,
x R.
The powers are different so we have two linearly independent power series solutions!
Remark 22.19. This example says that there could be two series solutions in Frobenius Theorem.
Now the second example seems very similar.
Example 22.20 (One series solution). Consider the equation xy 00 + 3y 0 y = 0.
By the Frobenius theorem,
X
X
X
cn xn , y 0 (x) =
cn (n+r)xn+r1 , y 00 (x) =
cn (n+r)(n+r1)xn+r2 .
y(x) = xr
n0
n0
n0
141
So
0 = xy 00 + 3y 0 y
X
X
=
{(n + r)(n + r 1) + 3(n + r)} cn xn+r1
cn xn+r
n0
n0
= c0 xr1 (r(r 1) + 3r) +
{((n + r)(n + r 1) + 3(n + r))cn cn1 } xn+r1 .
n1
The indicial equation here is r(r + 2) = 0, with solutions: r1 = 0 and r2 = 2.

r1 = 0
cn1
, n = 1, 2,
n(n + 2)
2c0
c0
=
cn = Q n
n!(n + 2)!
k=1 {k(k + 2)}
!
X
xn
y1 (x) = 2c0
n!(n + 2)!
n0
cn =
x R.
r2 = 2. The recursion formula becomes n(n 2)cn cn1 = 0

n=1:
c1 c0 = 0
c0 = c1 = 0
n=2:
0 c2 c1 = 0
cn1
, n = 2, 3,
cn =
n(n 2)
!
!
n
n
X
X
x
2x
= 2c2
y2 (x) = c2 x2
n!(n
2)!
(n + 2)!n!
n0
n2
which is the same as y1 . So there exists only one series solution.
Remark 22.21. Compare Example 22.18 and Example 22.20. There is not much
difference between the two. However, the first one has two series solutions and the
second one only one.
Exercise 22.22 (From a physics qualifying exam). Solve the differential equation
x2 y 00 + 2xy 0 + (x2 2)y = 0
using the Frobenius method. That is, assume a solution of the form
X
y=
cn xn+k
n0
and substitute back into the differential equation.

a) Determine the values of k which are allowed.
b) For each value of k, develop a recursion relation for cn .
c) Discuss the convergence of each solution.
142
Theorem 22.23. If in (22.6)

X
p(x) =
pn (x x0 )n
q(x) =
n0
qn (x x0 )n
n0
then the indicial equation has the form

r(r 1) + p0 r + q0 = 0.
Furthermore, from this emerge three cases:
Case 1 r1 6= r2 and r1 r2
/ Z. Then
X
y1 (x) =
cn (x x0 )n+r1
c0 6= 0
b0 6= 0.
n0
y2 (x) =
bn (x x0 )n+r2
n0
r1 r2 = N , a positive integer. Then

X
y1 (x) =
cn (x x0 )n+r1 , c0 6= 0
Case 2
n0
y2 (x) = Cy1 (x) ln(x x0 ) +
bn (x x0 )n+r2
n0
where C is a constant.
Case 3 r1 = r2
X
y1 (x) =
cn (x x0 )n+r1
c0 6= 0
n0
y2 (x) = y1 (x) ln(x x0 ) +
bn (x x0 )n+r2 .
n1
Exercise 22.24. Prove that the indicial equation is indeed

r(r 1) + p0 r + q0 = 0
under the conditions of Theorem 22.23.
Exercise 22.25. (Two series solutions). Solve
xy 00 + (x 6)y 0 3y = 0.
Exercise 22.26. Find a second solution of
xy 00 + 3y 0 y = 0.
b0 6= 0
LECTURE 23
Harmonic Oscillator
Let us consider the equation
d2
u + x2 u = u , < x < .
(23.1)
dx2
This equation appears in Quantum Mechanics and is a specific case of the Schrodinger
equation
u00 + q(x)u = u.
Equation (23.1) can be viewed as an eigenvalue problem

Au = u ,
d2
d
where A = dx2 + x2 in L2 (R). The term dx
2 represents the kinetic energy, and the
2
term x the potential energy.
Clearly, A is a Sturm-Liouville operator but in L2 on the whole line (, ).
Note at this point that you may wonder what basis would work for L2 (R): polynomials work locally but blow up at so they wont work for R; harmonics (einx ) are
not in L2 (R) so they wont work either; plus they would require some decay at infinity.
So it is not obvious to find something that would work.
Theorem 23.1. A = A .
Proof. Note first, that if f (x) L2 (R) then lim f (x) = 0. For f, g L2 (R), we
x
have
hAf, gi =
00
Af (x)g(x)dx =
x2 f (x)g(x)dx
f (x)g(x)dx +

*0

by parts
0
0
0

= f (x)g(x)
+
f (x)g (x)dx +
x2 f (x)g(x)dx

* 0
by parts

00
f (x)g (x)dx +
x2 f (x)g(x)
= f (x)g(x)

=
f (x) g 00 (x) + x2 g(x) dx = hf, Agi .
QED
d
Now we are going to find the spectrum of A. Recall that H = dx
2 has only a
continuous spectrum with no eigenvalues.
143
144
23. HARMONIC OSCILLATOR
Transform (23.1) into another equation by setting

x2
u(x) = e 2 y(x).
x2
x2
u0 (x) = xe 2 y + e 2 y 0
x2
x2
x2
x2
x2
u00 (x) = (e 2 x2 e 2 )y xe 2 y 0 xe 2 y 0 + e 2 y 00
x2
= e 2 (y 00 2xy 0 (1 x2 )y)
x2
x2
u00 + x2 u = e 2 (y 00 + 2xy 0 + y x2 y + x2 y) = e 2 y
y 00 + 2xy 0 + y = y , or
y 00 2xy 0 + ( 1)y = 0 , x R.
This is known as Hermites equation.

Employ the Frobenius method. Set
X
X
y=
cn x n y 0 =
cn nxn1
n0
y 00 =
n0
(23.2)
cn n(n 1)xn2 .
n0
By plugging into (23.2), we get

X
X
X
0=
cn n(n 1)xn2
2cn nxn + ( 1)
cn x n
n0
n0
n0
o
Xn
=
cn+2 (n + 1)(n + 2) 2cn n + ( 1)cn xn .
n0
From this, we extract the recursion relation

cn+2 =
2n + 1
cn , n 0
(n + 1)(n + 2)
c0 6= 0 ,
c1 6= 0.
For the general solution of (23.2), we have

X
X
y(x) =
c0n x2n +
c00n x2n+1 .
n0
n0
{z
y1
{z
y2
By the ratio test,

cn+2
2n + 1
= lim
=0
n cn
n (n + 1)(n + 2)
lim
and hence (23.3) converges. Assuming c0 = 1, we can write

2n + 1
2
+1
cn
cn+2 =
cn =
cn
(n + 1)(n + 2)
n + 2 (n + 1)(n + 2)
1 + n/2
X x2n
1
2
c2n
y1 (x)
= ex .
n!
n!
n0
(23.3)
145
Similarly, assuming c1 = 1, we now write

2
+3
2cn
cn+2 =
cn
n + 1 (n + 1)(n + 2)
n+1
2n
X
x
1
2
y2 (x) x
= xex .
c2n+1
n!
n!
n0
2
So, y(x) c0 ex + c1 xex ,
x , and hence
x2
u(x) = e 2 y(x) c0 ex
2 /2
+ c1 xex
2 /2
/ L2 (R).
This means that (23.1) has no L2 -solutions for an arbitrary .

However, if = 2m + 1 , m = 0, 1, . . . then both y1 , y2 in (23.3) terminate at
n = m, n = m + 1, and hence y(x) becomes a polynomial.
So, we arrive at the fact that the harmonic oscillator A has a discrete spectrum,
i.e.
Theorem 23.2. Let A =
d2
+ x2 on L2 (R). Then
dx2
(A) = d (A) = {2m + 1}m0 .

2
Associated eigenfunctions are ex /2 ym (x) where polynomials {ym (x)} are called
Hermite polynomials and usually denoted by Hm (x). Explicitly,
H0 (x) = 1 , H1 (x) = 2x , H2 (x) = 4x2 2 , H3 (x) = 8x3 12x , . . .
2
Since A is selfadjoint, functions um (x) = ex /2 Hm (x) must be orthogonal in L2 (R)

and form a basis in L2 (R). So,

2
ex Hm (x)Hk (x)dx = 0 , m 6= k.
hum , uk i =
This also means that the Hermite polynomials {Hm (x)} are orthogonal in the
2
weighted space L2 (ex , R). So a good basis in L2 (R) looks like a polynomial weighted
2
by ex /2 to ensure decay.
There are some nice formulas for {Hm (x)}.
Hm (x) = (1)m ex
dm x2
e .
dxm
Or, Hm can be computed from the Taylor expansion of the generating function
G(x, t),
X
tm
2
G(x, t) = et +2tx =
Hm (x) .
(23.4)
m!
m0
146

2
Compute L2 (R)-norm of ex /2 Hm (x).

2
dm
2
x2 /2

x2 2
Hm =
e Hm (x)dx = (1)m Hm (x) m ex dx
e
dx
R
R
:0

m1

dm1 x2
d
2

0
m
x
m1

H
(x)
= (1) Hm
(x)
+
(1)
e dx.
m1 e
m

dx
dxm1
R

By a direct computation,
0
Hm
(x) = 2mHm1 (x) ,
m1
and by continued integration by parts, we find

2
x2 /2

Hm = 2m m! .
e
So, for (23.1) we get
um (x) = p
2m m!
ex
2 /2
Hm (x) ,
m = 0, 1, 2, . . .
Exercise 23.3. Prove formula (23.4).

Exercise 23.4. Prove the recursive formula (23.5).
(23.5)
LECTURE 24
The Fourier Transform

The Fourier transform is obtained as a limiting procedure on Fourier series.
1. Fourier Series
As we know from the past, every L2 (, )-function f (x) can be expanded into a
Fourier series (Theorem 15.2):

X
1
inx
cn e , cn =
f (x)einx dx.
(24.1)
f (x) =
2
nZ
It is common to adopt the following notation, cn = fb(n).
By making a scale transformation as seen in Theorem 15.9 and Exercise 15.10,
every function f (x) L2 (T, T ) can be expanded into a Fourier series:
T
X 1
inx
inx
f (x) =
fb(n)e T , fb(n) =
f (x)e T dx.
(24.2)
2T
T
nZ
The coefficients fb(n) represent how much of each harmonic is present in the signal,
the weight of each harmonic. The smoother the function, the faster fb(n) decays.
In other words, a decent function f (x) defined on a finite interval (T, T ) can
be represented by (24.2). It is natural to ask what if f (x) is defined on the whole line
(, ).
Well, if f (x) is periodic with period 2T as in the figure below, then (24.2) remains
valid outside of (T, T ).
3T
But what if f (x) is not periodic, like the example in the figure below?
3T
147
148
24. THE FOURIER TRANSFORM
Let us see whats going on with (24.2) as T . This limiting procedure is not
trivial since none of the formulas (24.2) admit switching T for .
Introduce a new quantity
n
n
,
T
then (24.2) reads
T
X 1
in x
b
b
f (x) =
f (n)e
, f (n) =
f (x)ein x dx.
(24.3)
2T
T
nZ
Note that
1
1
1
=
= (n n1 ) .
{z
}
2T
2 T
2 |
=: n
So, for (24.3) we get

X 1 T n
X 1
1
fbn ein x n ,
f (x) =
ein x n =
fb
2
2
2
nZ
nZ
| {z }
=: fbn
where
1
1
fbn =
2
2
f (x)ein x dx.
T
Let

1 b
1
F (n ) lim fn =
f (x)een x dx.
T
2
2
Note that the above is not rigorous since n 0 for each n when T (hidden
T in n ), but for any large value of T , there are infinitely many n-values still bigger to
compensate for the large T , so since were looking at the overall limit, we press on.
Then,

X 1
1
fbn ein x n
f (x) =
2
2
nZ

X
1 b
1 in x
1
looks like
fn e
= lim
n =
F ()eix d
T
2
2
2
nZ
as in a Riemann sum. So we get

1
f (x) =
eix F ()d , where
2
Inverse Fourier transform
1
F () =
2
eix f (x)dx
(24.4)
Fourier transform
which are continuous analogs of (24.1) or (24.2). Another notation for F () is fb().
This approach is, by no means, rigorous but it prompts a very important concept, the
concept of the Fourier transform.
2. FOURIER TRANSFORM
149
Note that fb() represents how much of the function has frequency, and now with
the continuum of frequencies, we have fb() 0 , , i.e. the relative weight of
harmonics should decrease as frequencies increase.
2. Fourier Transform
Definition 24.1. Let f (x) be a function from L1 (R).
Then the function fb(), R, defined by the formula

1
b
f () =
eix f (x)dx , R
2
is called the Fourier transform of f (x), also denoted fb = F f .
I claim that there is no physicist unaware of this concept!
Theorem 24.2. If f (x) L1 (R) then fb() exists.

Proof. If f (x) L1 (R) then eix f (x) L1 (R) since eix f (x) = |f (x)|. Moreover,

1 ix
b
e
f (x)dx
f () =
2

ix

1
1

e
f (x) dx =
|f (x)| dx.
QED
2
2
We are not proving the above theorem for L2 (R) since its a bit too complicated.
But basically, oscillations at high frequencies will help smaller rates of decay.
Remark 24.3. Equation (24.4) on the previous page suggests that

1
eix fb()d = f (x).

2
(This is to be discussed later.)
Example 24.4. Let f (x) = ea|x| , a > 0. Compute fb().
Since ea|x| L1 (R), fb() exists (Theorem 24.2) and we have

0

1
1
1
ixa|x|
ix+ax
fb() =
e
dx =
e
+
eixax dx
2
2
2 0
0

1
e(ai)x dx +
e(ai)x dx
=
2
0

!
ix ax 0
ix ax
1
e
e
e
e
=
+
a i
a i 0
2

r
1
1
1
2
a
=
+
=
.
2
a + 2
2 a i a + i
(24.5)
150
24. THE FOURIER TRANSFORM
Thus,
r
a|x| () =
e[
2
a
.
a2 + 2
From (24.5) one has
1
2
a
eix
ea|x| =
d
2
a + 2
2 R
or,
a|x|
cos x
e
.
(24.6)
d
=
2
2
a
R a +
A nice, valuable formula for free. (Do you remember seeing this before?)
(
1 , |x| a
Exercise 24.5. Let f (x) =
. Compute fb() and then use (24.5) to
0 , |x| > a
derive a curious integral similar to (24.6).
Note that in Example 24.4 both function and Fourier transform are real. This is
not always the case.
Proposition 24.6.
(i) fb() = fb().
(ii) (symmetry) fb() = fb()
f (x) = f (x).
Exercise 24.7. Prove Proposition 24.6.

Proposition 24.8. fb is real if and only if f (x) = f (x).
Remark 24.9. The proposition above means that f, fb are real
Exercise 24.10. Prove Proposition 24.8.

Example 24.11. Let f L1 (R) and f 0 L1 (R). Compute fb0 ().

1
0
b
f () =
eix f 0 (x)dx
2
:0 1

1 ix
by parts
f (x)(i)eix dx
= e f (x)

2
2

1
= i
eix f (x)dx = ifb()
2
{z
}
|
fb()
i.e.,
fb0 () = ifb().
Exercise 24.12. Show
e 4a
e[ () = .
2a
ax2
f is even.
LECTURE 25
Properties of the Fourier Transform. The Fourier Operator

1. Properties
Start with two obvious ones
(i) (f\
+ g)() = fb() + gb()
(additivity)
[)() = fb()
(ii) (f
(homogeneity)
This means that the Fourier transform can be viewed as a linear operator.
Definition 25.1. The linear operator F defined by

1
F f () =
eix dx
2
is called the Fourier operator. So, by definition, fb = F f .

(iii) \
f (n) () = (i)n fb()
It is enough to prove it for f 0 since you get the rest by induction. So this property
was actually proven in Lecture 24, Example 24.11.
Note that the above means that the Fourier transform converts differentiation to
multiplication by a power of .
Corollary 25.2. If f C n (R) and f (n) L1 , then C > 0 such that

C
b
f () n .
| |
Proof. By Property (iii),
fb() =
where
1 d
f (n) ()
(i)n

1 1
C
b
ix (n)
e
f (x)dx
,
f () =

n
||
||n
2
1
C=
2
(n)
f (x) dx.
QED
This implies that if f is smooth, i.e. infinitely differentiable, then fb decays faster
than any power. Also, unless f is rough, the high frequencies have less of a role so we
can cut them off.
Definition 25.3. Given functions f, g, the convolution f g of these functions is
defined as

1
(f g)(x) =
f (s)g(x s)ds.
2
151
152
25. PROPERTIES OF THE FOURIER TRANSFORM. THE FOURIER OPERATOR
Its clear that f g = g f . Indeed,

x s = t
1

(f g)(x) =
f (s)g(x s)ds =
s = x t
2

1
1
=
f (x t)g(t)d(t) =
f (x t)g(t)d(t)
2
2

1
=
g(t)f (x t)dt = (g f )(x).
2
Another elementary property of the convolution is its linearity with respect to g
(and hence with respect to f ): f (g1 + g2 ) = f g1 + f g2 .
(iv) f[
g = fb gb.
(The convolution theorem)
Proof. Assume f, g L1 (R) and f g L1 (R).

1
[
f g() =
eix (f g)(x)dx
2

1
ix
f (s)g(x s)ds dx
e
=
2

1
=
eix f (s)g(x s)dsdx
2

x s = t
1
ix

=
e
g(x s)dx f (s)ds =
x = t + s
2

1
itis
=
e
g(t)dt f (s)ds
2

1
1
is
it
e
f (s)
e
g(t)dt ds
=
2
2
|
{z
}
g
b()
= fb() gb().
QED
2. The Fourier operator

We now consider the Fourier transform as an operator on a Hilbert space.
The following theorem plays a central role in math physics.
Theorem 25.4. The Fourier operator F is unitary in L2 (R), i.e. hF f, F gi = hf, gi.
Proof. (at the physical level of rigor i.e. we show it for a nicer group of functions)
Let f, g C0 (R) (introduced in Lecture 19). Let us make sure first that fb, gb L2 (R).
2. THE FOURIER OPERATOR
Indeed,
2
b
f =
(1,1)
1
2
|
1

b 2
f () d +
R\(1,1)
153

b 2
f () d
2 1
2

d
1
0
d +
|f (x)|dx
|f (x)|dx
2
2
1
R
R\(1,1)
{z
} |
{z
}
by Theorem 24.2
2
|f (x)|dx
by Corollary 25.2
|f 0 (x)|dx
2
<
d
= 2.
2
R\(1,1)
So fb, gb L2 (R) and the integral
E
D
b
f , gb =
since
fb()b
g ()d
is defined. Let us now compute it.

D
E
1
ix
is
b
e
f (x)dx
e
g(s)ds d.
f , gb =
2
(25.1)
Since all the integrals here are absolutely convergent we can rearrange the order of
integration and (25.1) reads

D
E 1
i(sx)
b
f , gb =
e
d f (x)g(s) dx ds
2

|
{z
}
=(sx) , by Theorem 20.4

f (x)
(s x)g(s) ds
{z
}
dx
=g(x) (follows from Def. of -function, Example 19.18)
f (x)g(x) dx = hf, gi .
D
E
So, we get that for all f, g C0 (R), fb, gb = hf, gi. i.e.,
hF f, F gi = hf, gi .
(25.2)
QED
Observe now that the right side of (25.2) exists not only for C0 -functions but for
any f, g L2 (R). (Indeed, by the Cauchy inequality |hf, gi| kf kkgk.) This suggests
that the left hand side of (25.2) exists too for all f, g L2 (R) which, in turn, means
that the natural domain of the Fourier operator is not C0 (R) or L1 (R) but L2 (R).
Rigorous proofs of this can be found in advanced textbooks.
154
25. PROPERTIES OF THE FOURIER TRANSFORM. THE FOURIER OPERATOR
Remark 25.5. Even though the integral

eix f (x)dx
in general does not converge absolutely for f (x) L2 (R), the following limit
N
eix f (x)dx
lim
N
(25.3)
exists in a certain sense for any function from L2 (R).

Theorem 25.4 implies some important facts.
Corollary 25.6. For all f, g L2 (R),
1) kF f k = kf k ,
2) kF f F gk = kf gk ,
3) hf, F F gi = hf, gi F F = I F 1 = F .
Whereas Property 2) above implies that the Fourier operator preserves distances,
the remaining two are important enough to deserve their own statements.
Theorem 25.7 (Plancherels Theorem).

b 2
f () d =
|f (x)|2 dx.

Theorem 25.9. The Fourier operator is invertible and

1
1
F f (x) =
eix f ()d.
2
(25.4)
LECTURE 26
The Fourier Integral Theorem (Fourier Representation

Theorem)
1. The Fourier Integral Theorem
Theorems 25.4, 25.9, and Remark 25.5 combined imply the following central theorem.
Theorem 26.1 (The Fourier Integral Theorem). Let f L2 (R) be piecewise continuous and differentiable on R, like the one in the figure below.
Then for every point of continuity x, the following representation holds:

1
f (x) =
2
ix
fb()d ,
1
fb() =
2
eix f (x)dx.
(26.1)
Proof. By Theorem 25.9, the Fourier operator F is invertible and hence for all
f L2 (R),
F 1 F f = f.
(26.2)
For simplicity, let f be smooth. Then (26.2), by (25.4), reads:

1
1
ix
is
x R :
e
e
f (s)ds d = f (x)
2
2
which is exactly (26.1).
QED
Note that if f is even or odd, we get only [0, ) with sin or cos. Its still called the
Fourier transform. What about points of irregularity?
Remark 26.2. In view of Remark 25.5, we claim that (26.1) holds under the only
condition f L2 (R) if we understand integrals in (26.1) as

R
= lim
.
(26.3)
155
156
26. THE FOURIER INTEGRAL THEOREM (FOURIER REPRESENTATION THEOREM)
Note, however, that (26.1) holds not for all x R, but, roughly speaking, for those
x for which our function f (x) is defined/differentiable. Points of discontinuity of f (x)
are troublesome, as the next example shows.
Indeed, Carleson showed that if f is continuous, the representation still only converges almost everywhere, although it will be ok if f is differentiable.
(
1, 0x<1
Example 26.3. Let f (x) =
.
0 , otherwise
1
0
1

1

1
1
1
1
ix
ix
ix
e
fb() =
e
f (x)dx =
e
dx =
2
2 0
2 i
0

i
i
1 1e
1
1
e
=
fb() =
.
i
2
2i
By Theorem 26.1, x 6= 0, 1,
1
f (x) =
2
fb()eix d.
(26.4)
Lets see what happens, say, at x = 0. The right side of (26.4) then becomes

1
1
1 ei
1
1 ei

d =
d =
2i
2 2i

i
1
1
1 ei
e 1
=
d() =
d.
2i
2i
Note that we get the same expression when setting x = 1.
This integral is kind of tricky since it is absolutely divergent. We have to use Complex Analysis to evaluate it.
Actually, contour integrals, residue theorem, etc. are usual tools for computing
Fourier integrals.
CR+
On the other hand,
+
CR
By the Residue Theorem (Thm 5.14),

i

i
1
e 1
e 1
d = Res
, 0 = 0 (26.5)
2i
sincen = 0 o
is a removable singularity and hence
ei 1
Res
, 0 = 0.
I1 + I2 + I3 + I4 .
+
R
(26.6)
1. THE FOURIER INTEGRAL THEOREM
Evaluate each of these integrals separately.
1
ei 1
1
ei 1
I1 =
d =
d
2i CR+
2i CR+
|
{z
}
0 , R
157
d
1
2i CR+
|
{z
}
=1/2 (see Example 3.1)
by Jordans Lemma (Lemma 6.5)
and so lim I1 = 1/2.

R
1

|I2 | =
2i

i
e 1
ei 1
1
d
|d|.
2 C
||
Since = ei , < < 0 , d = iei d |d| = d , || = . Also,

ei 1 = ei(cos +i sin ) 1 = e sin +i cos 1

= e sin cos( cos ) + i sin( cos ) 1

sin
= e
cos( cos ) 1 +iei sin sin( cos ) O().
| {z }
|
{z
}
0 , 0
0 , 0

f (x)
where Big O notation is defined by f (x) = O(x) , x 0 if x C , x 0.

So, ei 1 = |O()| 0 , 0, and we get
1
|I2 |
2
|O()|d = |O()| 0 , 0.
So lim I2 = 0. Next,
0
1
I3 + I4 =
2i
(R,R)\(,)
ei 1
d
and it follows from (26.5) and (26.6) that I3 + I4 = I1 I2 and passing to the lim ,
R
we get
1
2i

ei 1
1
d = lim I3 + I4 = lim I1 lim I2 =
R
0
R
2
0
where the integral is understood as in (26.3).

Finally, the Fourier representation theorem for x = 0 gives

1
1

ix
b
f ()e d
= 6= f (0) = 1.

2
2
x=0
So Theorem 26.1 fails for x = 0.

As a matter of fact, the following holds.
158
26. THE FOURIER INTEGRAL THEOREM (FOURIER REPRESENTATION THEOREM)
Theorem 26.4. Let f (x) be in L2 (R) piecewise continuously differentiable. Then

for every point of continuity, (26.1) holds. If x = x0 is a point of discontinuity, then

1
f (x0 + 0) + f (x0 0)
1
ix0 b
e f ()d , fb() =
eix f (x)dx.
=
2
2
2
(No proof).
2. Generalization of the Fourier Transform (by a limiting procedure)
As we already computed in Example 24.4, for a > 0,
r
2
a
a|x| () =
e[
.
2
+ a2
Take the limit as a 0 in (26.7), we get
(
r
2
a
0 , 6= 0
=
.
lim
a0
2 + a2
, =0
(26.7)
a0
Looks like the -function? Yes, indeed!

a
is a -family as a 0.
(a2 + 2 )
(Hint: Introduce n = 1/a and check all three conditions in Definition 19.5.)
Thus, as a 0,the left hand side of (26.7) becomes b
1()1 and the right hand side
of (26.7) becomes 2() and we get
b
1() = 2().
1
b
Note also that ()
=
but this is a generalization of the Fourier transform
2
since
/ L2 (R) (recall that 2 is undefined).
Exercise 26.6.
(Hint: (tan1 x)0 =
1where
BB
Show that:
r ||
3/2
e
1
\
+ ().
tan x = i
2
2
1
.)
1+x2
def
a|x| .
we can set as definition: b
1() = lim e\
a0
LECTURE 27
Some Applications of the Fourier Transform

1. Fourier Representation
Let us discuss first the main reason why the Fourier transform is that important.
As we know ( by property (iii) of Lecture 25)

F f (n) () = (i)n (F f )()
,
n = 1, 2,
(27.1)
In particular, (n = 1)

d
F f
dx
() = i(F f )()
(F f ) () = i(F f )() or

1 d
F
f () = (F f )().
i dx
(27.2)
Definition 27.1. Given f , fb = F f is called the Fourier representation of f (physical terminology).

Next, by Theorem 26.1 (equation (26.2))
F 1 F = I
F 1 F f = f
and (27.2) tranforms into

1 d 1
F
F F f () = (F f )()
i dx
or in the operator form

F
1 d
i dx
F 1 = .
(27.3)
This relation is very profound and all applications of the Fourier theory owe just
to this formula.
Lets try to understand what (27.3) means.
Recollect our old business in Linear Algebra.
1 d
According to Remark 11.6, operators
of differentiation and multiplication by
i dx
are similar.
Multiplying (27.3) by F 1 on the left and F on the right yields

1 d
1 d
1
1
1
F
= F 1 F
(27.4)
| {zF} i dx F
| {zF} = F F
i dx
=I
=I
159
160
27. SOME APPLICATIONS OF THE FOURIER TRANSFORM
1 d
whih reads: the operator of differentiation
in the Fourier representation is equal
i dx
to the operator of multiplication by .
In Quantum Mechanics, it means that coordinate and momentum representations
are similar.
So, once again, (27.3), (27.4) mean that in the Fourier representation the operator
of differentiation becomes the operator of multiplication.
Definition 27.2. The object
A=
n
X
ak (x)
k=0
dk
dxk
is called the differential operator of order n.

Theorem 27.3. Let A be a differential operator with constant coefficients, i.e.
n
X
dk
ak k ,
A=
dx
k=0
then
A = F 1 p()F
where
p() =
, or
n
X
F A = p()F
(27.5)
ak (i)k .
k=0
2. Applications of the Fourier Transform to Second Order Differential

Equations
Let us now see how Theorem 27.3 works.
Example 27.4. Find an L2 -solution to
u + 2u + 20 u = f (t) ,
(27.6)
where , 0 are positive constants with 0 > and f L2 (R).1

Solution. Go over in (27.6) to the Fourier representation. By Theorem 27.3
d2
d
(equation (27.5)) with A = 2 + 2 + 20 , we get
dt
dt

F A = 2 + 2i + 20 F
and (27.6) transforms into

2 + 2i + 20 u
b() = fb() ,
(27.7)
where as usual u
b = F u , fb = F f .
1Such
equations occur in solving differential equations coming from Newtons Second Law of
motion (in particular harmonic oscillators with damping), and the variable is usually temporal. So
here we switched x to t and use dots for derivatives.
2. APPLICATIONS OF THE FOURIER TRANSFORM TO SECOND ORDER DIFFERENTIAL EQUATIONS

161
It follows from (27.7) that

u
b() =
fb()
.
2 + 2i + 20
By Theorem 26.1,
u(t) = F

1
u
b (t) =
2
eit fb()
d.
(20 2 ) + 2i
So the L2 -solution to (27.6) can be obtained by the following formula

1
1
eit fb()
b
u(t) =
d where f () =
eit f (t)dt. (27.8)
2
2 ) + 2i
(
2 0
2
Remark 27.5. We obtained (27.8) under the assumption f (t) L2 (R). Actually,
(27.8) remains true as long as the integrals in (27.8) are defined somehow (e.g. in the
weak sense). For example, one can handle cases like f (t) = (t), f (t) = sin t, etc. All
these functions are not in L2 (R).
Remark 27.6. For the denominator in (27.8) one has
20 2 + 2i = ( 1 )( 2 ) ,
(27.9)
p
where 1,2 = 20 2 + i. Note Im 1,2 > 0.
(
1 , 0t1
Exercise 27.7. Solve (27.6) with f (t) =
.
0 , otherwise
(Hint: use Example 26.3 and equation (27.8).)
Remark 27.8. Lets show another way to handle (27.8). Putting equations (27.8),
(27.9) together we have

1
eit
is
e
f (s)ds d
u(t) =
2 ( 1 )( 2 )

1
ei(ts)
=
d f (s)ds.
(27.10)
2
( 1 )( 2 )
|
{z
}
=I(ts)
Integral I can be computed by Theorem 6.6 since the function
1
is
( 1 )( 2 )
subject to Jordans lemma (Lemma 6.5).

Indeed, if t s > 0 then by Theorem 6.6
i1 (ts)

ei(ts)
e
ei2 (ts)
d = 2i
+
.
(1 2 ) (2 1 )
( 1 )( 2 )
162
If t s < 0 then the function

the lower half plane. But
Theorem
1
is subject to Jordans lemma in
( 1 )( 2 )
1
has no poles in C and hence by Cauchys
( 1 )( 2 )

ei(ts)
d = 0.
( 1 )( 2 )
So, I(t s) = 0 when t s < 0 and (27.10) transforms into

t

1
1
u(t) =
I(t s)f (s)ds
I(t s) f (s)ds
2
2 t | {z }
=0
t

2i
1
ei1 (ts) ei2 (ts) f (s)ds
=

2 1 2
t

i
=
ei2 (ts) ei1 (ts) f (s)ds
1 2
or in the real form
1
u(t) = p 2
0 2

q
2
0 2 (t s) e(ts) f (s)ds.
sin
(27.11)
Looks like a nice formula!? Not particularly, since we can no longer use Complex
Analysis to evaluate this integral.
Remark 27.9. Formula (27.11) implies the so-called principle of causality one
of the basic principle of physics. It says that an effect cannot happen before the cause
has occured. Indeed since the integration in (27.11) is done over (, t), computing
the solution u(t) of equation (27.6) requires the knowledge of the force f (s) on (, t)
and doesnt need any information on f (s) for s > t.
Its the principle of causality for physical processes described by ordinary differential
equations.
3. Applications of the Fourier Transform to Higher Order Differential
Equations
Example 27.10 (A beam on an elastic foundation). Consider an infinite beam with
a force f (x) considered constant over time such as gravity or a load. We measure y(x)
the deflection or displacement.
f (x)
y(x)
0
EIy IV + Cy = f (x)
E, I, C are constants
3. APPLICATIONS OF THE FOURIER TRANSFORM TO HIGHER ORDER DIFFERENTIAL EQUATIONS

163
or
C
1
y=
f (x).
(27.12)
EI
EI
Let us solve (27.12). Note that this equation can be approached by the method of
variation of parameters but its more complicated than the Fourier method well apply
here. Consider y L2 (R), then no boundary conditions are needed (even though we
have a fourth order linear equation) and y 0 , y 00 , y 000 , y IV L2 (R) although these will be
automatically satisfied.
Apply the Fourier transform to (27.12). By (27.5) we get
y IV +
(i)4 yb +
C
1 b
yb =
f.
EI
EI
It follows from here that

1 fb()
EI 4 + 4
By the Fourier Integral Theorem
yb() =
1 1
y(x) =
2 EI
where 4
C
.
EI
eix fb()
d
4 + 4
(27.13)
which is the solution to (27.12) for an arbitrary f (x).

Consider a specific case f (x) = P (x) (P is a constant). In other words, an
external force f (x) is applied at just one point x = 0, that is, an impulse force. But
1
fb() = P and (27.13) transforms into
2

eix
1
d.
y(x) =
2EI 4 + 4
If x < 0, remark that we can write ei()(x) then by a change of variables, since
1
is even, we get the same integral but with x. So we assume x > 0 and close
4 + 4
the contour in the upper half plane.
First we introduce a change of variable: = d = d. Then
1 P
eix
P
eix
y(x) =
d
=
d.
2 EI4 R 4 + 1
2C R 4 + 1
1
The function 4
is subject to Jordans lemma (Lemma 6.5) and then by Theorem
+1
6.6
ix

2
iP X
e
y(x) =
Res
, k
,
, x > 0
C k=1
4 + 1
where 1 , 2 are zeros of 4 + 1 = 0 in C+ , i.e.
1
1
1 = ei/4 = (1 + i)
,
2 = ei3/4 = (1 + i).
2
2
164
1
Since 1 , 2 are simple poles of 4
, by Corollary 5.11 we get
+1

iP ei1 x ei2 x
P
y(x) =
=
i1 ei1 x + i2 ei2 x
+
3
3
C
41
42
4C

iP d
P d i1 x
i 1 (1+i) x
i 1 (1+i) x
i2 x
=
e
+e
=
e 2
+e 2
4C dx
4C dx

P d x
x
P d x i x
x
i
2
2
2
2
=
e
e
e
+e
=
cos
4C dx
2C dx
2

P x
x
x
= e 2 cos + sin
, x>0
2 2C
2
2
and for x < 0, we have

x
P d
x
2
y(x) =
e cos .
2C dx
2
So putting it all together using absolute value, we have
r

x
|x|
P |x|
C
4
y(x) = e 2 cos + sin
, =
EI
2 2C
2
2
an even function. We can also write explicitly
!
r
r
4
C
P e 4EI |x|
C
C
4
4
y(x) =
cos
x + sin
|x| .
4
3
4EI
4EI
2 4EIC
Part 4
Partial Differential Equations
LECTURE 28
Wave Equations
1. The Stretched String
Here we are going to discuss a simple problem that historically led to the wave
equation.
Consider an ideal stretched string as below (finite or infinite):
x
Note that this could also be a water surface.

Ideal means it is a line.
Let us apply to this string a vertical force F (x, t) which makes our string deform
from its free position (as in the figure below). Assume that the force of gravity can
be neglected.
F (x, t)
0
x
x + dx
Zoom in on an arbitrarily small fragment of the string as shown below.

u(x)
T
F dx
dx
x + dx
167
The wave equation is local so

well worry about a small portion of the string.
u(x) is the displacement (deflection) of the string at the
point x.
T is the tension.
dx is small enough so that F dx
is constant over dx.
168
28. WAVE EQUATIONS
Then, by Newtons second law1

2u
(28.1)
t2
where T is the tension, F is the external force per unit length at point x, (x) is the
density of the string (again we assume dx small enough so that (x) stays constant
2u
along it), 2 is the acceleration of the fragment in the transverse direction.
t
We assume that , are small enough, i.e.
T sin + T sin + F dx = (x)dx
sin ' tan ,

sin ' tan ,
cos ' 1 ,
cos ' 1.
Hence,

u
sin ' tan =
,
x
x

u
sin '

x
and
sin sin =
x+dx
u(x + dx) u(x)
.
x
x

sin sin
2u
T
+F = 2
dx
t
|
{z
}
2u
x2
and we arrive at
2u
2u
(x) 2 = T 2 + F (x, t).
(28.2)
t
x
If there is no external force F (x, t) and if (x) = const then (28.2) transforms into
the homogeneous wave equation:
1 2u
2u
=
c2 t2
x2
c2
T
.
(28.3)
The general solution to this equation can be easily found. Indeed, if f (x) is an
arbitrary twice differentiable function, then u(x, t) = f (x ct) is a solution to (28.3)
since if we set z = x ct then
u
f z
=
= f 0 (z) (c)
t
z t
2u
0
f 0 (z) z
=
c
f
(z)
=
c
= (c)2 f 00 (z) = c2 f 00 (z).

t2
t
z t
Similarly,
2u
= f 00 (z) , z = x ct
x2
and hence (28.3) holds.
It can be proven that every solution u(x, t) of (28.3) is
u(x, t) = f (x ct) + g(x + ct)
1i.e.
(28.4)
the sum of the forces equals mass times acceleration, and here projecting on the y-axis.
2. THE METHOD OF EIGENFUNCTION EXPANSIONS (SPECTRAL METHOD)
169
with some twice differentiable functions f, g.

Note that f (x ct) plays the role of the wave traveling in the negative direction of
the x-axis and g(x + ct) in the positive direction.
However, equation (28.3), also called the free wave equation, is not of particular
interest without some physical restrictions on its solution.
Example 28.1. A finite string is fixed at points x = 0 , x = `.
The natural restrictions then are
u(0, t) = 0 = u(`, t).
0
Such conditions are called boundary conditions.
Example 28.2. The string is infinite but the initial shape of the string and the
distribution of initial velocities are given:

u
u(x, 0) = (x) ,
= (x).
t
t=0
Such conditions are called initial conditions.

One may well have a combination of the above conditions (especially if you have an
infinite string). Remember that you need 4 conditions so we may have both boundary
and initial conditions. All this makes formula (28.4) of a little interest.
2. The Method of Eigenfunction Expansions (Spectral Method)
Consider equation (28.3) on a finite interval (0, ) with Dirichlet boundary conditions and some initial conditions. For simplicity set c = 1. So we have
utt uxx = 0
(28.5a)
(BC)
u(0, t) = u(, t) = 0
(IC)
u(x, 0) = (x) , ut (x, 0) = (x)
which is often referred to as an initial value Dirichlet problem for the free wave equation in dimension one or the boundary initial value (BIV) problem for the homogeneous
wave equation.
For solving our problem at hand, there are two steps, the first one being called
spectral analysis.
It is reasonable to assume that the solution u(x, t) of (28.5a) belongs to L2 (0, )
(as a function of x) and (28.5a) can then be viewed as
utt + Au = 0
(28.5b)
d2
is the operator of kinetic energy (Schrodinger operator), with bounddx2
ary conditions u(0) = u() = 0. For now we will ignore t. Let us perform the spectral
where A =
170
28. WAVE EQUATIONS
analysis of the operator A on L2 (0, ).

(
Ay = y
y(0) = 0 = y()
One can easily see that the general solution is

y(x) = aei
+ bei
Further,
(
y(0) = a +b = 0
y() = aei + bei = 0
= n = n2 , n N
(
b = a
sin = 0
and yn (x) = Cn sin nx are eigenfunctions.

Normalize yn (x): kyn k = 1 kyn k2 = 1 so
1=
|yn (x)| dx =
0
r
2
Cn =
.
Cn2
C2
sin nxdx = n
2
(1 cos 2nx)dx =
0
Cn2
2
And finally, the result of the spectral analysis of A is

r
n = n2
en (x) =
2
sin nx ,
n N.
The idea at this point is to note that since A is selfadjoint and its spectrum purely
discrete (A) = {n2 }, the resulting eigenfunctions {en (x)} form a basis, and so the
second step is to consider the solutions in this basis. I.e.
By Theorem 18.3, we can represent any solution of (28.5a) in the form
u(x, t) =
un (t)en (x) ,
where
un (t) = hu, en i
(28.6)
n1
(t appears here as a parameter).

Differentiate (28.6) in t twice
utt =
X 2
u (t) en (x)
2 n
t
|
{z
}
n1
=
un (t)
(28.7)
2. THE METHOD OF EIGENFUNCTION EXPANSIONS (SPECTRAL METHOD)
171
and plug (28.6), (28.7) into (28.5b)

X
X
0=
un (t)en + A
un (t)en
n1
n1
X
un (t)en +
n1
un (t)n en
n1

un (t) + n un (t) en .
n1
So we get

X
un (t) + n un (t) en = 0
un (t) + n un (t) = 0 , n N.
(28.8)
n1
So, our original partial differential equation (28.5a) broke into the infinite chain
of linear ordinary differential equations (28.8). This is the crux of this approach:
reduce the partial differential equation (PDE) to infinitely many ordinary differential
equations (ODE) which are hopefully simple enough to solve. Here indeed, each of
these equations can be trivially solved
un (t) = an ei
n t
+ bn ei
n t
and recollecting that n = n2 we get

un (t) = an eint + bn eint
n N.
(28.9)
Now we need to find {an , bn }. It should be found from the initial conditions:

X
X

un (0)en (x) = (x)
u(x, 0) =
un (t)en (x) =

n1
n1
t=0
.
(28.10)

X
X

0
0
un (0)en (x) = (x)
ut (x, 0) =
un (t)en (x) =

n1
t=0
n1
But by Theorem 18.3,

(x) =
(x) =
n en (x) ,
n = h, en i
n en (x) ,
n = h, en i
n1
(28.11)
n1
and hence, comparing (28.9), (28.10) yields

un (0) = n
u0n (0) = n .
172
28. WAVE EQUATIONS
So we get
or
un (t) = an eint + bn eint u0n (t) = inan eint inbn eint

(
un (0) = an + bn = n
u0n (0) = inan inbn = in(an bn ) = n
(
(

an + bn = n
an = 12 n + inn
.
(an bn ) = inn
bn = n an = 12 n inn
So

n int
n int
n +
e + n
e
in
in
)
(

1
n int
n int
e + n +
e
=
n +
2
in
in

n int
n
= Re n +
e = Re
n +
(cos nt + i sin nt)
in
in
n
= n cos nt +
sin nt.
n
Now we are able to present the solution to (28.5a)
X
u(x, t) =
un (t)en (x)
(28.12)
1
un (t) =
2

n1
where
n
sin nt , n =
un (t) = n cos nt +
n
r
2
en (x) =
sin nx ,
n N.
(x)en (x)dx , n =
0
(x)en (x)dx ,
0
Exercise 28.3. Adapt this method to the nonhomogeneous wave equation (i.e. derive formulas similar to (28.12))
utt uxx = f (x, t)

.
u(0, t) = u(, t) = 0
u(x, 0) = (x) , u (x, 0) = (x)

t
Note that there are other wave equations such as the Gordon-Klein equation where
there is a potential q and which has applications in plasma physics:
utt = uxx + qu.
LECTURE 29
Continuous Spectrum of a Linear Operator

Continuous spectrum happens only in infinite dimensional spaces. Relevant examples in physics include crystals with bands of continuous spectrum and particles in
quantum mechanics where we have bound states corresponding to eigenvalues and then
the particle gets stuck there a while (in the order of nanoseconds) and a continuous
spectrum corresponding to positions where the particle doesnt stay any amount of
time.
The concept of a continuous spectrum is fundamental in math physics and, ALAS,
not easy at all. The way its usually defined requires Advanced Analysis. Well try to
detour these difficulties somehow.1
We recall that a number C is an eigenvalue of a linear operator A in a Hilbert
space H (usually written actually in gothic letter: H) if the equation
Au = u
(29.1)
has a nontrivial solution u from H.

The set of all eigenvalues is called the point spectrum or discrete (actually there
is a difference between point and discrete spectrum but physicists dont usually care).
We denote it by d (A).
Symbolically,
d (A)
u H , u 6= 0 :
Au = u.
This means that if d (A) then the operator A I is not invertible.

(If AI were invertible then equation (29.1) would have only a trivial solution u = 0.)
The spectrum was first described to you with the resolvent:
R() = (A I)1
which has poles at the eigenvalues of A. But the spectrum in general is the set of all
singularities of R(). The poles will be the discrete spectrum, but there can be all
sorts of other singularities.
Rewrite (29.1) as
(A ) u = 0
(29.2)
and ask the following question. If
/ d (A) (and hence equation (29.2) has no solution
in H), whether (29.2) has a solution in some sense?
1Dont
hold me accountable for some semirigorous shortcuts.

173
174
29. CONTINUOUS SPECTRUM OF A LINEAR OPERATOR
Definition 29.1. Let A be a linear operator on a Hilbert space H. A scalar C

is said to be a point in the spectrum (A) of A if there exists a sequence {un }, called
a Weyl sequence, such that un H , kun k = 1 and
lim k(A ) un k = 0.
(29.3)
It is clear that d (A) (A): pick un = u for all n (or rather a normalized version)
where u solves Au = u.
Here is the standard classification of the spectrum.
Definition 29.2. A scalar (A) is said to be from the discrete spectrum d (A)
if {un } can be chosen convergent to some element u H i.e.
d (A) {un } , un H , kun k = 1 and un u H such that (29.3) holds.
Definition 29.3. A scalar (A) is said to be from the continuous spectrum
c (A) if {un } can be chosen divergent in H, i.e. un doesnt converge to any element of
H.
Theorem 29.4.
(A) = d (A) c (A).
Proof. Let (A), then by Definition 29.1 there exists a Weyl sequence {un }
of elements un H, kun k = 1 such that (29.3) holds. Each such sequence can be
represented as {u0n } {u00n } (as a union of two subsequences2) such that
(i) u0n u H , n
(ii) u00n does not converge to any element of H.
Consider case (i). Let us make up a new sequence {un }, un = u (yes, it consists of
the same element u), (29.3) then reads
k(A )uk = 0
(A )u = 0
Au = u
and so d (A).
Consider case (ii). By Definition 29.3, c (A).
So if (A) then d (A) or c (A) i.e. d (A) c (A).
QED
Recall that the spectrum could also be empty or the whole complex plane.
Remark 29.5. d (A) c (A) need not be empty! I.e. there may exist eigenvalues
embedded into the continuous spectrum.
For such , you can find a divergent sequence where you can pick a divergent
subsequence (and then c (A)) and a convergent subsequence (and so d (A)).
It seems like a weird case, but it happens all the time in physics with bound states.
Lemma 29.6. If (A) then there exists a Weyl sequence {un }:
lim hAun , un i = .
2{u0
n}
or {u00n } may actually be empty.
2. CONTINUOUS SPECTRUM OF SELFADJOINT OPERATORS
175
Proof.
|hAun , un i | = |h(A )un , un i|
Cauchy
k(A )un k 0 , n .
QED
There is a whole theory behind the concept of the continuous spectrum. We concentrate mainly on the spectral theory only of some specific operators of mathematical
physics.
2. Continuous spectrum of selfadjoint operators
Theorem 29.7. Let A = A then (A) R.
Proof. By Lemma 29.6, for (A), there exists a Weyl sequence {un } such
that
lim hAun , un i = .
n
But if A = A , then for each n N

hAun , un i hun , Aun i
hAun , un i hAun , un i
=
2i
2i
hAun , un i hAun , un i
=
= 0.
2i
So we must have 0 = lim Im hAun , un i = Im . Therefore the spectrum is real. QED
Im hAun , un i =
Exercise 29.8. State and prove Theorem 29.7 for unitary operators.
(Hint: use Lemma 29.6.)
Let us consider a very important specific operator.
1 d
, the operator of momentum, on L2 (R), then
i dx
(A) = c (A) = R.
(29.4)
Theorem 29.9. Consider A =
Proof. Note first3 that A = A . Weve also proved in Example 18.2 that
d (A) =
(A) = c (A).
Furthermore, by Theorem 29.7, (A) R. We prove now that (A) = R.

Take
1
un (x) = eix e|x|/n , R.
n
We have
1
2 2x/n
2
2|x|/n
e
dx =
kun k =
e
dx = 1.
n R
n 0
3Weve
proved this fact for the operator of momentum on two different spaces in Examples 11.1
and 17.15. The proof here would go similarly.
176
Also
1 d
1
sgn x
un un =
i
un un
i dx
i
n
(
1 ,
where sgn x =
1,
x<0
x>0
i
sgn x un

n
1 d
1
= ksgn x un k = 1 kun k = 1 0 , n .

u
u
n
n
i dx
n
n | {z } n
=1

d
d
By Definition 29.1, 1i dx
. Since is arbitrary, we have 1i dx
= R. QED
=
Remark 29.10. It is not difficult to show that u(x) = eix is a solution of

1 du
u = 0
i dx
which is obviously not from L2 (R). However, u(x) is a weak solution4 to this equation.
Such solutions are commonly referred to as eigenfunctions of the continuous spectrum
or generalized eigenfunctions.
Remark 29.11. Theorem 29.9 sheds some light on the reasonable question: why
is the continuous spectrum called so? Roughly speaking, the continuous spectrum is
always made of intervals and no isolated point can be from this set.
Definition 29.12. Two operators A, B on a Hilbert space are called unitary equivalent if there exists a unitary operator U such that
B = U 1 AU = U AU.
(29.5)
Observe, by the way, that (29.5) implies also

A = U BU 1 = U BU .
(29.6)
Indeed multiply (29.5) by U from the right and U from the left.
Do you remember that we dealt with Definition 29.12 while considering finite dimensional spaces? It was then called similarity (see Definition 11.4) but also goes
by equivalence. Here we consider unitary equivalence which is much more useful and
powerful.
Theorem 29.13. If A, B are unitary equivalent then
(A) = (B).
In other words unitary equivalence preserves the spectrum.
Proof. By (29.6) we have

k(A )uk = U BU 1 I u = U BU 1 U U 1 u

= U (B )U 1 u = (B )U 1 u since U is unitary.
4The
term weak solution does not mean that the Weyl sequence converges weakly to the weak
solution. In the example above, one can easily show that {un } converges weakly to 0, not eix .
2. CONTINUOUS SPECTRUM OF SELFADJOINT OPERATORS
177
Set v = U 1 u. So
k(A )uk = k(B )vk .
(29.7)
Note v H and kvk = kuk since U is unitary. So, (29.7) means that if (A),
then by Definition 29.1, there exists a Weyl sequence {un } for A in the Hilbert space
H, such that kun k = 1 and
lim k(A ) un k = 0.
n
Then
lim k(B ) vn k = 0
and {vn } is a Weyl sequence for B. Hence (B).

Similarly, (B) (A).
QED
Theorem 29.13 is very important to spectral analysis. I.e., if you need to find (B)
of an operator B but you know that B is unitary equivalent to another operator A for
which the spectrum (A) is known then we simply have
(B) = (A).
Definition 29.14. Let H = L2 (R). The operator B acting by the rule
u(x) L2 (R)
Bu(x) = x u(x)
is called the operator of multiplication by an independent variable x, or the operator of

coordinate (following Quantum Mechanics terminology).
Theorem 29.15. The spectrum of the operator B of multiplication on L2 (R) is
purely continuous and fills out the whole real line. I.e.
(B) = c (B) = R.
Proof. By Theorem 25.4, the Fourier operator F is unitary in L2 (R). By property
(iii) of F in Lecture 25 we have
1 d
u = F u
F
i dx
which also reads
F Au = F u = BF u ,
u L2 (R)
or
F AF 1 = B
A = F 1 BF.
where A is the operator of momentum.

This means that B is unitary equivalent to the operator of momentum A in L2 (R).
By Theorem 29.13 then
(B) = (A).
Now by Theorem 29.9
(A) = c (A) = R
and the theorem is proven.
QED
Remark 29.16. By recalling (from Example 19.18 1)) that for all y R, y(y) = 0
then we find that (x ) is a so-called weak solution to Bu = u for any R since
we always have (x )(x ) = 0 but (x )
/ L2 (R).
178
Once again, eigenfunctions of the continuous spectrum of the momentum operator

1 d
are eix and of the operator of multiplication by x are (x ).
i dx
So roughly speaking, if c (A) then there exists a solution to
Au = u
(29.8)
which is not from H, and basically to perform the spectral analysis of A we have to
find those s for which (29.8) has a solution (from H or not). Then {} will be (A)
and solutions will be eigenfunctions of A (from the discrete or continuous spectrum).
In general, the theory behind is very involved. But for differential operators everything
is a whole lot simpler.
Exercise 29.17. Prove Theorem 29.9 without using Theorem 29.13. Instead modify
the proof of Theorem 29.15 using
n
un (x) =
p
n (x )
where
n (x) =
1
2n
1
2n
LECTURE 30
The Method of Eigenfunction Expansion in the Case of

Continuous Spectrum
1. The Schr
odinger Operator
d2
We start out from the Schrodinger operator A = 2 of the second derivative on
dx
the whole line R, also called kinetic energy.
Theorem 30.1. Let A =
d2
on L2 (R). Then
dx2
(A) = c (A) = [0, ).
(30.1)
I.e. kinetic energy can take on any nonnegative value.

1 d
on R. By Theorem
Proof. Consider first the operator of momentum B =
i dx
29.9, we have
(B) = c (B) = R
and by Remark 29.10, functions
eix
e(x, ) =
, R
2
are the eigenfunctions of the continuous spectrum of B, i.e.
Be = e ,
R.
(30.2)
Apply the operator B to both sides of (30.2)

B 2 e = Be = 2 e.
But on the other hand,
1 d 1 d
B =
=
i dx i dx
2
2 2
1
d
d2
=
=A
i
dx2
dx2
and (30.3) reads

Ae = 2 e , R.
That is e(x, ) is a solution to the eigenvalue problem
Ae = e
1
with = 2 , e(x, ) = ei x .
2
179
(30.3)
180
30. EIGENFUNCTION EXPANSION WITH CONTINUOUS SPECTRUM
Since = 2 , < < , then 0 and

(A) = c (A) = [0, ).
QED
We also obtained
d2
Theorem 30.2. Let A = 2 on L2 (R). Then
dx

1 ix
1 ix
e
, e
2
2
[0,)
(30.4)
is the set of eigenfunctions of the continuous spectrum of A.

Remark 30.3. (30.4) says that to each eigenvalue [0, ) of the continuous
spectrum of A there correspond two eigenfunctions. This means that the spectrum (A)
is not simple but of (algebraic) multiplicity two. This fact results in many complications
to be overcome.
2. The Wave Equation on the Whole Line
I.e.,
utt = uxx , u(x, 0) = (x) , ut (x, 0) = (x).
(30.5)
Although generally not explicitly stated we assume decay at , and generally
consider , L2 (R). These functions are called respectively the initial profile for
and the distribution of initial velocity for .
We approach this problem as in Lecture 28.
d2
1 d
utt + Au = 0 , A = 2 = B 2 , B =
dx
i dx
and so we have
1 d
utt + B 2 u = 0 , B =
.
(30.6)
i dx
It follows from Theorem 29.9 and Remark 29.10 that the set
1
{e(x, )}R , e(x, ) = eix
(30.7)
2
is the set of all eigenfunctions of B. By the Fourier Integral Theorem (Theorem 26.1),
every function u(x) L2 (R) can be represented by
eix
1
u(x) =
u
b() d , u
b() =
eix u(x)dx.
2
2 R
R
Due to (30.7), we rewrite it as
u(x) =
u
b()e(x, )d , u
b() =
u(x)e(x, ) .
R
R
|
{z
}
looks like hu,ei?
Even though e(x, )

/ L2 (R), we agree to write
u(x)e(x, ) = hu, ei .
R
2. THE WAVE EQUATION ON THE WHOLE LINE
So we get
181
u(x) =
R
u
b()e(x, )d ,
u
b = hu, ei
(30.8)
which looks similar to

u(x) =
u
bn = hu, en i .
u
bn en (x) ,
So (30.8) can be treated as expansion in eigenfunctions of operator B. Note in

particular that we have
he(x, ), e(x, )i = ( )
which is analogous to the discrete case and its Kronecker-delta function.
Now represent the solution u(x, t) of (30.6) in the form
u(x, t) =
u
b(, t)e(x, )d
R
and plug it in (30.6): for any x
2
2
u
b,t e(x, ) d + B
u
b(, t) e(x, )d
0= 2
| {z }
t R
R | {z }
no t
no x
=
u
btt (, t)e(x, )d + u
b(, t)
B 2 e(x, )
| {z }
R
R
=2 e(x,) due to (30.3)
and we get for any x

2
u
btt (, t) + u
b(, t) e(x, )d = 0
R
u
btt (, t) + 2 u
b(, t) = 0
u
b(, t) = a()eit + b()eit .
In the above, in a way we have u

btt + 2 u
b orthogonal to each of the e(x, ); and
because the e(x, ) are a basis then u
btt + 2 u
b has to be zero, but the proof is not
obvious.
Now convert the initial conditions into the frequency domain:
b
u
b(, 0) = ()
b
, u
bt (, 0) = ()
and we get
(
()
b
= a() + b()
or
b
() = ia() ib()
!
1
b
a=
b+
, b=
ba =
2
i
So
1
u
b(, t) =
2
b
()
()
b
+
i
1
eit +
2
a + b =
b
b
a b =
!i
b
1
b
.
2
i
b
()
()
b
!
eit
182
and finally
u(x, t) = F 1 a()eit + b()eit
or more explicitly
u(x, t) =
u
b(, t)e(x, )d , where
!
!
b
b
1
()
()
1
u
b(, t) =
()
b
+
eit +
()
b
eit
2
i
2
i
b
(x)e(x, )dx , () =
()
b
=
(x)e(x, ) ,
R
,
(30.9)
1
e(x, ) = eix .
2
In this form, the solution is difficult to visualize and analyze. Furthermore, historically, the wave equation was solved directly through a clever change of variable by
dAlembert. So we present here dAlemberts solution, but in Appendix to this lecture
we present how we can derive the dAlembert formula from (30.9).
Consider
(
utt = uxx
.
u(x, 0) = (x) , ut (x, 0) = (x)
Introduce the following canonical variables:
=x+t
= x t.
By applying the chain rule,

(
(
ut = u u
utt = (u )t (u )t = u u (u u )
ux = u + u
uxx = (u )x + (u )x = u + u + u + u
(
utt = u 2u + u
u 2u + u = u + 2u + u
uxx = u + 2u + u
u = 0
u(x, t) = F (x + t) + G(x t).
u = F ()
u(, ) = F () + G()
This solution looks simpler but how do we express F, G with respect to , ? In

this case, it is simple too, but this approach cannot always be generalized.
(
u(x, 0) = F (x) + G(x) = (x)
ut (x, 0) = F 0 (x) + G0 (x) = (x)
3. PROPAGATION OF WAVES
183
Now integrate the second equation from a fixed point x0 :
x
1
1
C
x
(s)ds +
F (x) = (x) +
2
2 x0
2
(s)ds + C
F (x) G(x) =
x
1
C
1
x0
(s)ds
G(x) = (x)
2
2 x0
2
x+t
xt
1
1
C 1
1
C
u(x, t) = (x + t) +
(s)ds + + (x t)
(s)ds
2
2 x0
2
2
2 x0
2
1 x+t
1
(x t) + (x + t) +
(s)ds.
u(x, t) =
2
2 xt
So we summarize our result as
Theorem 30.4. The solution to the initial value problem
(
utt = uxx
.
u(x, 0) = (x) , ut (x, 0) = (x)
can be represented by the dAlembert formula
(x t) + (x + t) 1
u(x, t) =
+
2
2
x+t
(s)ds.
(30.10)
xt
3. Propagation of Waves
Formula (30.10) describes the phenomenon of wave propagation.
1) Consider first the case with no initial velocity; imagine that the string is pinched
then released.
(x) =
,
a
0.
Then (30.10) becomes

u(x, t) =
(x t) + (x + t)
.
2
(x)
2
(xt)
2
a + t t
a+t
184
(x t)
represents a bump initially supported on (a, a) moving with a
2
speed 1 in the positive direction of the x-axis.1
Similarly,
i.e.
(x+t)
2
(x)
2
a t t a t
i.e.
1.
(x + t)
is a bump moving in the opposite direction with the same speed
2
1
1
Since u(x, t) = (x t) + (x + t) we get
2
2
u(x, t)
t = 4a
t = 3a
t = 2a
5a
3a
4a
3a
2a
3a
t=a
2a
2a
t=0
5a
4a
3a
2a
Figure 1
1Some
students have trouble understanding why subtracting t from x moves the graph to the
right. But watch:
Supp = (a, a)
a x t a
a + t x a + t.
3. PROPAGATION OF WAVES
185
2) Now let the string be initially flat, but we give it a boost.
(x) = 0
(x) =
.
a
Now (30.10) becomes
1
Set g(x) =
2
1
u(x, t) =
2
x+t
(s)ds.
xt
(s)ds, then
0
h/2
a
g(x) =
1
h
2
(s)ds
a
h/2
and u(x, t) = g(x + t) g(x t).
For g(x t) we have
h/2
g(x t)
g(x)
a+t
h/2
a + t
and hence
a
h/2
a+t
g(x)
a + t
x
g(x t)
h/2
Figure 2
For g(x + t), see Figure 3 on the next page.
The total picture, Figure 4, is the result of the addition of Figure 2 and
Figure 3.
186
h/2
g(x + t)
g(x)
a
at
h/2
a t
Figure 3
u(x, t)
t = 4a
5a
t = 3a
4a
t = 2a
5a
3a
t=a
4a
3a
2a
2a
g(x + t) g(x t)
t=0
Figure 4
3) If 6= 0 , 6= 0 then the picture will be the superposition of Figure 1 and
Figure 4 since we have a linear equation.
Exercise 30.5. Solve the wave equation for u(x, t) with
2
(x) =
,
1
(x) =
Graph the solution u(x, t) for various values of time.
APPENDIX: DERIVATION OF THE DALEMBERT FORMULA
187
Exercise 30.6. Solve ( < x < , t > 0)

utt = uxx
u(x, 0) = 0 ,
ut (x, 0) = xex
and graph u(x, t) for various values of t.

Exercise 30.7. Assuming Theorem 30.4, modify equation (30.10) for
utt = c2 uxx
where c > 0 is a constant.
Appendix: Derivation of the dAlembert Formula
From (30.9), we get
!
!
)
(
b
b
1
1
()
()
u(x, t) =
()
b
+
eit +
()
b
eit e(x, )d
2
i
2
i
R
!
!
b
b
()
()
1
1
()
b
+
eit+ix d +
()
b
eit+ix d
=
i
i
2 2 R
2 2 R
!
!
b
b
1
()
1
()
=
ei(x+t) d +
ei(xt) d
()
b
+
()
b
i
i
2 2 R
2 2 R

1
(s) is
=
(s) +
e
ds ei(x+t) d
4 R
i
R

1
(s) is
+
(s)
e
ds ei(xt) d.
(30.11)
4 R
i
R
Change now the order of integration in (30.11)
i(x+ts)

1
e
i(x+ts)
u(x, t) =
d + (s)
d ds
(s) e
4 R
i
R
R

i(xts)
1
e
i(xts)
+
(s) e
d (s)
d ds.
4 R
i
R
R
(30.12)
In order to give this nasty expression a nice look, we basically have to evaluate
ia
e
ia
e d
,
d
R
R i
with a = x + t s a = x t s. But it has been already done. By Theorem 20.4 we
have
eia d = 2(a).
(30.13)
We have also computed (even twice!) the other integral in Example 7.1:
ia
ia
e
e
d = i
d =
(a > 0).

R i
(30.14)
188
If a < 0 then taking the complex conjugation of (30.14) we get

ia
i(a)
ia
e
e
e
d =
d =
d = since a > 0.
i
R i
R
R i
So,
(
ia
e
, a>0
d =
= sgn a.
, a < 0
i
(30.15)
Plug (30.13) and (30.15) into (30.12)

1
(s)
u(x, t) =
sgn(x + t s) ds
(s)(x + t s) +
2 R
2

1
(s)
sgn(x t s) ds
+
(s)(x t s)
2 R
2
1
1
= (x + t) +
(s) sgn(x + t s)ds
2
4
R
|{z}
x+t
=
+
x+t
1
1
+ (x t)
(s) sgn(x t s)ds
2
4
R
|{z}
xt
=
+
xt
(x + t) + (x t) 1
1
=
+
(s) (1) ds +
(s) (1) ds
2
4
4 x+t
1 xt
1 xt
(s)ds
(s)(1)ds
4
4
(x + t) + (x t) 1 x+t
1 xt
=
(s)ds
(s)ds
+
2
4
4
|
{z
}
x+t
1
=
(s)ds
4 xt
1
1
(s)ds +
(s)ds
4 x+t
4 xt
{z
}
|
x+t
1
=
(s)ds
4 xt
(x + t) + (x t) 1 x+t
=
+
(s)ds.
2
2 xt
x+t
LECTURE 31
The Heat Equation on the Line

We are concerned with the initial value problem
(
ut = uxx
, < x <
.
u(x, 0) = (x)
(31.1)
Note that represents the initial distribution of heat; its physical so it is reasonable
to assume L2 (R) since the total energy is finite.
Our approach is going to be absosulety the same as that for the wave equation.
d2
Namely, we introduce the operator A = 2 and (31.1) rewrites
dx
ut + Au = 0.
Next, A = B 2 , B =
1 d
and we have
i dx
ut + B 2 u = 0.
By Theorem 29.9 and Remark 29.10, (B) = R and

{e(x, )}R
1
e(x, ) = eix
2
is the set of all eigenfunctions of B. By Theorem 26.1 we have
u(x, t) =
u
b(, t)e(x, )d , u
b(, t) =
u(x, t)e(x, )dx.
R
(31.2)
Plug (31.2) into (31.1)
ut (x, t) =
u
bt (, t)e(x, )d
B u(x, t) =
R
u
b(, t)B e(x, )d =
2 u
b(, t)e(x, )d.
Since ut + B 2 u = 0 we get

u
bt (, t)+2 u
b(, t) d = 0 , x R u
bt +2 u
b=0
R
189
u
b(, t) = C()e t .
190
31. THE HEAT EQUATION ON THE LINE
Now we need to find C(). But
u(x, 0) =
u
b(, 0)e(x, )d =
C()e(x, )d
R
||
(x)
()e(x,
b
)d
(x)e(x, )dx
where ()
b
=
It follows from here that

C() = ()
b
and problem (31.1) is solved:
1
1
2 t+ix
()e
b
d where ()
b
=
u(x, t) =
(x)eix dx.
2 R
2 R
(31.3)
However, this answer cannot be considered as final (we cant see if its real!).

1
2
is
u(x, t) =
(s)e
ds e t+ix d
2 R
R
and changing the order of integration, we get

1
i(xs)2 t
u(x, t) =
e
d (s)ds
2 R
R
|
{z
}
(31.4)
=:I(x,s,t)
where ei(xs) t is known as the dispersion term.

Lets evaluate I(x, s, t) separately. Complete the square

2

2
2
xs
(x s)2
xs
xs
2
t + i(x s) = i t +
= t+
.
4t
2 t
2i t
2 t
Then
1
I(x, s, t) =
2

2
2
t+ xs
xs
e
R
2i t
1
d =
e
2
xs
2 t
2

2
t+ xs
2i t
{z
d .
}
(31.5)
Evaluate now this integral
xs
dz
Make a substitution t + = z, then dz = td d = and
2i t
t
2

1
2
t+ xs
2i t
e
d =
ez dz ,
t C
R

xs
where C = z C : z = t i , < < .
2 t
(31.6)
191
t
R
xs
2 t
Evaluate
z 2
ez dz ,
dz = lim
CR
where CR is the piece of C bounded by vertical lines R.

Consider the contour :
R
+
h=
CR
xs
.
2 t
Since ez is analytic inside of , by the Cauchy Theorem one gets
R
z 2
0 = e dz =
+
+
Prove now that lim
CR
= 0.
R2 2iRy+y 2
R2
ez dz = 0.
lim
e2iRy ey dy.
h
So,
dy = ie
Taking its absolute value we get

0

2
2
2
2
z
R
2iRy
y
R

e dz = e
e
e dy e

+
(31.7)
Indeed, on + , z = R + iy and
z 2
(R+iy)2
e dz =
e
idy = i
+
ey dy eR heh
0.
192
Similarly,
ez dz = 0
lim
and passing in (31.7) to the limit as R , we get

0
0
R

lim
+ lim
lim
=0
+lim
CR
+
R
| {z
| {z }
}
and finally
z 2
ez dz =
dz =
(Gauss integral)
and hence plugging this into (31.6) yields

r

2
t+ xs
2i t
e
d =
.
t
R
Insert it into (31.5) we get
(xs)2

2 r
4t
1 xs
I(x, s, t) =
e 2 t
=
2
t
4t
and (31.4) becomes
(xs)2
1
e 4t (s)ds.
u(x, t) =
2 t R
So we proved
Theorem 31.1. The solution of (31.1) can be represented in the following form:
(xs)2
1
u(x, t) =
(31.8)
e 4t (s)ds.
2 t R
Remark 31.2. The right hand side of (31.8) is clearly undefined at t = 0. But as
you know1

1
n2 (xs)2
ne
nN
1
forms a -sequence and hence, setting n = , we find that
2 t
2
(xs)
1
n
2
2
w-lim e 4t = w-lim en (xs) = (x s)
t0 2 t
n
and from (31.8) we have
u(x, 0) =
(x s)(s)ds = (x)
R
1Recall
Example 19.6.
193
which is completely consistent with the initial condition u(x, 0) = (x).

The above will be true whether is a test function or not as illustrated in the
example below. The result will hold even if
/ L2 (R) even though the only method
used is the Fourier method.
Example 31.3. Let us solve (31.1) with (x) = (x x0 ). Physically it means that
we heated one point x = x0 with a laser beam.
By (31.8) in Theorem 31.1 we have
(xx0 )2

4t
(xs)2
1
e
u(x, t) =
e 4t (s x0 )ds =
.
2 t
2 t
u(x, t)
t1
t2
t3
x0
t1 < t2 < t3
Note that {u(x, t)}t0
forms a -sequence.
Remark 31.4. In the wave equation, the speed of propagation is constant and finite.
But in the heat equation, in a split second, tails spread from to (since exponentials are positive for all x). This means that the heat propagates instantaneously, i.e.
the speed of propagation is infinite. This of course is not quite true in real situations;
it shouldnt go faster than the speed of light. So recall that this is a model, and far
enough the change is infinitesimally small.
Remark 31.5. Note also that the solution to ut = uxx , u(x, 0) = (x) is the convolution of and the solution to the heat equation with an impulse initial distribution.
This is known as Duhamels principle.
Exercise 31.6. Solve (31.1) and draw a picture of u(x, t) for
(
1
1 , |x| a
(x) =
=
.
0 , |x| > a
a
a
x
2
2
(Hint: use (x) =
eu du, the error function and its properties.)
0
Exercise 31.7. Solve
(
ut = cuxx bux x R , t > 0
2
u(x, 0) = ex
where b, c are constants and c > 0. Graph the solution for various values of t > 0.
(Hint: use the substitution = x bt to reduce this eq. to the standard heat eq.)
LECTURE 32
Nonhomogeneous Wave and Heat Equations on the Whole

Line
In the nonhomogeneous case, we will treat the heat equation first because it is
simpler than the wave equation.
1. Nonhomogeneous Heat Equation
Consider
(
ut = uxx + f (x, t)
u(x, 0) = (x)
(32.1)
on the whole line. Assume f, L2 (R) with respect to x so that we can use the
spectral, here Fourier method. The function f is independent of u and is referred to
as the forcing term in general. Here it corresponds to heat generation (a heat source)
or dispersion.
We are going to apply the same eigenfunction expansion as we did in Lecture 30,
31.
We write (32.1) as
1 d
.
(32.2)
ut + B 2 u = f , B =
i dx
As previously, write the solution of (32.1) as
1
u(x, t) =
u
b(, t)e(x, )d , u
b(, t) =
u(x, t)e(x, )dx , e(x, ) = eix .
2
R
R
Since
2
ut =
u
bt (, t)e(x, )d
;
B u(, t) =
2 u
b(, t)e(x, )d
R
R
f (x, t) =
fb(, t)e(x, )d
,
fb(, t) =
f (x, t)e(x, )dx ,
R
by inserting the above in (32.2), equation (32.1) becomes in the frequency domain:
u
bt (, t) + 2 u
b(, t) = fb(, t).
(32.3)
It is a linear equation of order 1 and by applying the variation of parameters method

(see Appendix), we obtain that its general solution is
t

2 t
2
b
u
b(, t) = e
f (, )e d + C() .
0
195
196
32. NONHOMOGENEOUS WAVE AND HEAT EQUATIONS
By passing the initial conditions to the frequency domain, we have:

()
b
=u
b(, 0) = C()
and hence
u
b(, t) = e
|
2 t
2
2
fb(, )e d + ()
b
e t .
| {z }
{z
}
u
b0 (,t)
u
b1 (,t)
If we set
u0
R
u
b0 (, t)e(x, )d
and compare this with (31.3), we note that u0 is the solution of equation (32.1) with
f 0, i.e. the homogeneous equation (32.1). By Theorem 31.1 then
(xs)2
1
u0 (x, t) =
e 4t (s)ds
2 t R
and we arrive at
u(x, t) = u0 (x, t) +
u
b1 (, t)e(x, )d .
|
{z
}
R
u1 (x,t)
So now we only have to evaluate u1 (x, t).

t
1
2
2 t
fb(, )e d eix d.
u1 (x, t) =
e
2 R
0
1
Since fb(, ) =
f (s, )eis ds, (32.4) can be continued as
2 R
t
1
2
2 t
u1 (x, t) =
e
f (s, )eis ds e d eix d.
2 R
0
R
Now rearrange the terms and the order of integration.
!

t
1
2
ei(xs) (t ) d f (s, )d ds.
u1 (x, t) =
2
R
R
0
|
{z
}
(32.4)
(32.5)
=I(x,s,t )
But I(x, s, t ) was evaluated in Lecture 31. We have

(xs)2
e 4(t )
I(x, s, t ) =
.
2 t
Plugging it into (32.5) one has
n
o
2
(xs)2
t (xs)
t
exp
4(t )
e 4(t )
u1 (x, t) =
f (s, )d ds =
ds
d p
f (s, ).
2 (t )
R
0 2 t
R
0
2. NONHOMOGENEOUS WAVE EQUATION
197
The second expression is equivalent and is just another representation sometimes

used in research papers to associate more easily the variable with each integral when
multiple ones are present.
Theorem 32.1. The solution to (32.1) can be represented in the following form
(xs)2
t
(xs)2
1
e 4(t )
p
f (s, )d ds.
u(x, t) =
e 4t (s)ds +
2 t R
R
0 2 (t )
In this case, we have again a sort of superposition principle where the first term
corresponds to the homogeneous solution and the second term is a particular solution
for the initial condition 0.
This formula is not particularly pleasant but Im not aware of any better derivation
of it.
Exercise 32.2. Show that the solution to the heat equation for (x) = 0 and
f (x, t) = (x) , t is
r

t x2
x
x
u(x, t) =
e 4t erfc
2
2 t

2
2
where erfc(x) =
eu du, the complimentary error function.
x
2. Nonhomogeneous Wave Equation
Consider the initial value problem:
utt = uxx + f (x, t)

u(x, 0) = (x)
ut (x, 0) = (x)
(32.6a)
(32.6b)
(32.6c)
where f is the forcing function on the string, or a perturbation on the water.

Our arguments are going to be absolutely similar to those in Section 1 and we will
not repeat them.
In place of (32.3) one gets
u
btt (, t) + 2 u
b(, t) = fb(, t).
(32.7)
Its a nonhomogeneous order 2 linear differential equation with constant coefficients.

We apply the method of variation of parameters (see Appendix). This means that we
are looking for a particular solution to (32.7) in the following form
u
b(, t) = C1 (, t)eit + C2 (, t)eit
where C1 , C2 satisfy

eit
eit
it
ie
ieit
0
0
C1
= b
0
C2
f
and C10 , C20 are the (partial) derivatives of respectively C1 , C2 with respect to t.
198
Solving this system, one gets

C10 =
fbeit
W
C20 =
fbeit
,
W
(32.8)

where W is the Wronskian of eit , eit , i.e.

it
e
eit
= 2i.
W = det
ieit ieit
It now follows from (32.8) that
b
f (, t)eit
C1 (, t) =
dt
2i
b
f (, t)eit
C2 (, t) =
dt.
2i
(32.9)
Note that these integrals are indefinite so far.

For the general solution to (32.7), we have
t b
fb(, )ei
f (, )ei
it
u
b(, t) = e
d + e
d
2i
2i
0
0
+ C1 ()eit + C2 ()eit .
t
it
(32.10)
We note that C1 (), C2 () are functions of only as opposed to C1 (, t), C2 (, t)

defined by (32.9).
Now lets find C1 (), C2 (). We have, by (32.6b),
u
b(, 0)e(x, )d
u(x, 0) =
R
||
(x)
()e(x,
b
)d
where
()
b
=
(x)e(x, )dx
R
u
b(, 0) = ().
b
(32.11)
Similarly,
b
b
b
ut (x, 0) =
()e(x, )d , () =
(x)e(x, )dx and u
bt (, 0) = ().
(32.12)
R
But on the other hand, by (32.10)
u
b(, 0) =
|0
0 b
fb(, )ei
f (, )ei
d +
d +C1 () + C2 ()
2i
2i
0
{z
} |
{z
}
=0
u
b(, 0) = C1 () + C2 ().
=0
Next, differentiate (32.10) in t

t b
i
b(, )ei
f
(,
)e
f
u
bt (, t) = ieit
d + eit

2i
2i
0
=t

t b
i
i
b
f (, )e
f (, )e
d eit
+ ieit

2i
2i
0
199
=t
it
it
+ iC1 ()e iC2 ()e

t b
t b
f (, )ei
f (, )ei
fb(, t)
fb(, t)
it
it
=e
d +
+e
d
2
2i
2
2i
0
0
+ iC1 ()eit iC2 ()eit
eit t b
eit t b
i
d +
f (, )e
f (, )ei d
=
2 0
2
0
it
it
+ iC1 ()e iC2 ()e
and compute it at t = 0
u
bt (, 0) = iC1 () iC2 () = i (C1 () C2 ()) .
By (32.11), (32.12) we have
C 1 + C 2 =
b
b
C 1 C 2 =
i
C 2 = 2
C 1 = 2
!
b
b
i
!
b
b+
i
b
b
b + /i)
, C2 () = 21 (
b /i)
and finally,
That is C1 () = 12 (
(
!)
t b
i
b
f
(,
)e
1
()
u
b(, t) = eit
d +
()
b
+
2i
2
i
0
(
!)
t
b(, )ei
b
f
1
()
+ eit
d +
()
b
2i
2
i
0
eit t b
eit t b
i
=u
b0 (, t) +
f (, )e
d
f (, )ei d
2i 0
2i 0
|
{z
}
u
b1 (,t)
where
1
u
b0 (, t) =
2
b+
i
1
eit +
2
b
i
Comparing this to (30.9) in Lecture 30 we derive that
u0 (x, t)
u
b0 (, t)e(x, t)d
R
!
eit .
200
is the solution of equation (32.6) with f 0, i.e. the homogeneous equation (32.6).
By Theorem 30.4 then
(x t) + (x + t) 1 x+t
u0 (x, t) =
+
(s)ds
2
2 xt
and we arrive at
u(x, t) = u0 (x, t) +
u
b1 (, t)e(x, )d .
|
{z
}
R
u1 (x,t)
Let us treat u1 (x, t). Putting everything together one has

it t
it t
e
e
i
i
u1 (x, t) =
fb(, )e
d
fb(, )e d e(x, )d
2i 0
2i 0
R

ix
it t
e
e
is
i
f (s, )e
ds e
d
d
=
2 0
2i
R
R

ix
it t
e
e
is
i
f (s, )e
ds e d
d.
2 0
2i
R
R
(32.13)
So we got two triple integrals! Let us integrate with respect to d first.

)
( t i(t +xs)
1
e
u1 (x, t) =
d f (s, )d ds
2 R
2i
R
0
|
{z
}
( t
I(t +xs)
1
2
)

ei(t+ +xs)
d f (s, )d ds.
2i
{z
}
I(t+ +xs)
We need to evaluate I(a) where

1
I(a) =
2
eia
d.
2i
But its been done (Lecture 30, equation (30.15))

I(a) =
1
sgn a
4
and (32.14) can be continued

t
1
sgn(t + x s) f (s, )d ds
u1 (x, t) =
4 R
0

t
1
sgn(t + + x s) f (s, )d ds.

4 R
0
(32.14)
201
Since sgn(a) = sgn(a) we can continue

1
u1 (x, t) =
4
ds
t
0
sgn ((s + ) (x + t)) + sgn ((s ) (x t)) f (s, )d.

{z
}
=(s, )
(32.15)
Let us now figure out the support of (s, ).
s+ =x+t
s =xt
t
I
II
III (x, t)
xt
x+t x+t
Since (0, t) it is enough to check only regions I, II, III.

I. sgn((s ) (x t)) = 1 , sgn((s + ) (x + t)) = 1
II. sgn((s ) (x t)) = 1 , sgn((s + ) (x + t)) = 1
III. sgn((s ) (x t)) = 1 , sgn((s + ) (x + t)) = 1
(s, ) 0.
(s, ) 0.
(s, ) 2.
So (s, ) = 2 for (s, ) III and zero otherwise and (32.15) becomes
u1 (x, t) =
1 x
1x
f (s, )dsd =
f (s, )dsd
2
2
III
where
(x, t) III.
(x,t)
We are now able to state the final result.

Theorem 32.3. The solution u(x, t) to the initial value problem
utt = uxx + f (x, t)u(x, 0) = (x)ut (x, 0) = (x)
(32.16)
can be represented as
1To
see this, note that a point in I has a y-coordinate that is above the curve s = x t and
below s + = x + t, i.e. > s (x t) and < s (x t).
202
(x t) + (x + t) 1
u(x, t) =
+
2
2
x+t
(s)ds +
xt
1 x
f (s, )dsd
2
(32.17)
(x,t)
s+ =x+t
s =xt
where (x, t) =
0
xt
x+t x+t
Note that u(x, t) can also be written explicitly as

(x t) + (x + t) 1 x+t
1 t +(x+t)
u(x, t) =
+
(s)ds +
f (s, )dsd.
2
2 xt
2 0 +(xt)
Definition 32.4. (x, t) is called the characteristic triangle.
One can easily see that (x, t) expands as t increases for every fixed x.
Example 32.5. Consider the case when (x) 0 (x) and
T
(
1 , (x, t)
f (x, t) = (x, t)
0 , (x, t)
/
where =
a
i.e. = { (x, t) : a x a , 0 t T } .
That is the perturbation f (x, t) acts like a blast concentrated on [a, a] and lasting
T seconds.
Then the solution of (32.16) takes the form:
1 x
1 x
1
u(x, t) =
(s, )dsd =
dsd = Area (x, t).
2
2
2
(x,t)
(x,t)
Let us interpret this formula. Fix x = x0

/ (a, a).
As it follows from Figure 1 on the next page, u(x0 , t) = 0 as long as (x0 , t) and
do not intersect, i.e. the blast has yet to reach x0 (t < t0 ). Then u(x, t) = shaded area
for t0 < t < ta and when t ta , (x0 , t) then u(x0 , t) = 21 2a T = aT (see Figure
2).
Exercise 32.6. Find an analytic expression for u(x0 , t) on Figure 2. Consider 3
cases: x0 > a , a < x0 < a, x0 < a. Compute also ta in each of these cases.
APPENDIX: THE METHOD OF VARIATION OF PARAMETERS
203
t
(x0 , ta )
(x0 , t)
T
a
a = x0 t0
x0
Figure 1
u(x0 , t)
aT
ta
t0 = x0 a
Figure 2
Appendix: The Method of Variation of Parameters

Although the method can be generalized to higher order nonhomogeneous linear
ordinary differential equations, we will review here the cases of first and second order
nonhomogeneous linear ODEs.
1) First order nonhomogeneous linear ODE: Consider
y 0 + p(x)y = f (x).
Rewrite y = uv where u is a solution to the homogeneous equation
u0 + p(x)u = 0 ,
u = e
p(x)dx
Note that this method is called variation of parameters because whereas

the
0
p(x)dx
general solution to the homogeneous equation y + p(x)y = 0 is y = Ce
for some constant C, here we consider C to vary by being a function of x (and
called it v).
204
Now our original equation becomes

u0 v + uv 0 + puv = f
(u0 + pu) v + uv 0 = f
| {z }
=0
uv = f
v 0 = f u1 .
Hence

x
x

1
1
(s)ds + C
fu
(s)ds + C y = u(x)
fu
v=
0
0

x
p
p
fe
(s)ds + C .
y=e
0
2) Second order nonhomogeneous linear ODE: Consider

y 00 + p(x)y 0 + q(x)y = f (x).
Solve first the corresponding homogeneous equation
y 00 + p(x)y 0 + q(x)y = 0
by the method of your choice and obtain two linearly independent solutions
y1 , y2 .
Now we are looking for a particular solution of the form
y = C1 (x)y1 + C2 (x)y2 .
We will further assume that
C10 (x)y1 + C20 (x)y2 = 0.
Then by differentiating, we obtain
(
y 0 = C10 (x)y1 + C1 (x)y10 + C20 (x)y2 + C2 (x)y20 = C1 (x)y10 + C2 (x)y20
y 00 = C10 (x)y10 + C20 (x)y20 + C1 (x)y100 + C2 (x)y200
(32.18)
(32.19)
Plugging (32.19) in the original ODE leads to

f = C10 (x)y10 + C20 (x)y20 + C1 (x)y100 + C2 (x)y200

+ p(x) C1 (x)y10 + C2 (x)y20 + q(x) C1 (x)y1 + C2 (x)y2

0
0
0
0
0
00
= C1 (x)y1 + C2 (x)y2 + C1 (x) y1 + p(x)y1 + q(x)y1
{z
}
|
=0

00
0
+ C2 (x) y2 + p(x)y2 + q(x)y2
|
{z
}
=0
f=
C10 (x)y10
C20 (x)y20 .
Combining (32.18) and (32.20) leads to the system of equations

0
y1 y2
C1 (x)
0
=
.
0
0
0
y1 y2
C2 (x)
f
(32.20)
APPENDIX: THE METHOD OF VARIATION OF PARAMETERS
Using Cramers rule, we find that

W1
C10 (x) =
W
where W is the Wronskian of y1 , y2 , i.e.
C20 (x) =
,

y y
W = det 10 20
y1 y2
and
0 y2
W1 = det
f y20

= f y2
f y2
dx
W
and the general solution to our ODE is:
,

y1 0
W2 = det 0
= f y1 .
y1 f
f y1
C2 (x) =
dx
W
C1 (x) =
y = yc + yp
W2
W
Hence
205
yc = C1 y1 + C2 y2
yp = y1
x0
f y2
dx + y2
W
x0
(32.21)
f y1
dx
W
where C1 , C2 are constants and x0 is arbitrary.

Exercise 32.7. Find the general solution of
y 00 + y = tan t
0<t<
.
2
Exercise 32.8. Show that the solution to the IVP

Ly = y 00 + p(t)y 0 + q(t)y = g(t)
y(t0 ) = y0
y 0 (t0 ) = y00
can be written as
y =u+v
where u, v solve
Lu = 0
u(t0 ) = y0
Lv = g
v(t0 ) = 0 ,
u0 (t0 ) = y00
v 0 (t0 ) = 0
respectively.
(Hint: use (32.21) and choose C1 , C2 to satisfy the ICs. Then u = C1 y1 + C2 y2 and
v is the other part.)
Exercise 32.9.
a) Show that the solution of the IVP

(
y 00 + y = g(t)
y(t0 ) = 0 , y 0 (t0 ) = 0
is
sin(t s)g(s)ds.
y(t) =
t0
b) Find the solution of

y 00 + y = g(t)
y(0) = y0
y 0 (0) = y00 .
LECTURE 33
Wave and Heat Equations in Nonhomogeneous Media

So far we have considered the wave and heat equations of the form utt = uxx +
f , ut = uxx + f which basically describe the wave or heat propagation in an even
media. But
when doing acoustic imaging via the wave equation for the Earth for example,
we have a nonhomogeneous media;
using the heat equation for ice with cracks, there are some discontinuities, and
the media is not homogeneous.
So more general models lead to
utt = uxx q(x)u + f
(33.1)
ut = uxx q(x)u + f
where q(x) is an optical (heat) potential respectively.
Depending on the physical meaning, the first equation for example can take different
forms. In the above, it is referred to as the wave equation with potential or the plasma
wave equation. It is also known in relativistic quantum mechanics as the Klein-Gordon
equation.1 But other forms include
utt = c2 (x)uxx
utt
((x)ux ) = 0.
x
(Helmholtz equation)
(33.2)
None of these forms is considered canonical. Since q(x) is not constant, this is a
different ballgame (although we can handle boxes to get explicit solutions).
By denoting uxx + q(x)u Hu, (33.1) transforms into
utt + Hu = f
.
(33.3)
ut + Hu = f
Recall that
H=
d2
+ q(x)
dx2
(33.4)
is called the Schrodinger operator.

Note that if q is a real-valued function then H is selfadjoint. If q is complex-valued,
the problem is a whole lot more complicated.
1In
nonrelativistic quantum mechanics, they use the nonstationary Schrodinger equation:

iut = uxx qu.
207
208
33. WAVE AND HEAT EQUATIONS IN NONHOMOGENEOUS MEDIA
Our goal is to adjust Fouriers method to this setting. In general, a lot now depends
on the properties of the optical potential q(x) and things get a lot more complex. But
the general idea of eigenfunction expansions does work and as previously, while studying
the case q 0, we start from the spectral analysis of operator H. Well, its easier said
than done since if q 6= 0 the Schrodinger operator is a very difficult object and we are
not in a position to present its theory at any level.
First of all, a lot depends on whether we deal with a finite or infinite interval.
1
Finite interval. That is

(
utt + Hu = f
a x b a, b <
.
boundary + initial conditions
(33.5)
Typically, the spectrum of H is purely discrete but infinite. I.e. the equation
(
Hu = u
boundary conditions
has L2 (a, b)-solutions {en (x)}, i.e. eigenfunctions, corresponding to a discrete
set of {n }, the eigenvalues. Then by Theorem 18.3, {en (x)} forms an orthonormal basis in L2 (a, b) and we apply the procedure of Lecture 28.
Namely any solution to (33.5) can be represented as
X
u(x, t) =
u
bn (t)en (x) , u
bn (t) = hu(x, t), en (x)i
where hf, gi =
f (x)g(x)dx.
a
Plug this expression into (33.5)

X
X
X
fbn (t)en (x)
u
b00n (t)en (x) + H
u
bn (t)en (x) =
X
X
X
fbn (t)en (x)
u
b00n (t)en (x) +
n u
bn (t)en (x) =
u
b00n + n u
bn = fbn .
Then we solve this second order linear nonhomogeneous equation by variation of parameters finding constants of integration from meeting the initial
conditions.
1
bn , hence, are no longer Fourier
Note that en (x) are no longer einx and u
2
coefficients. But {b
un } are commonly called generalized Fourier coefficients.
Example 33.1. Consider
utt =
(1 x2 )ux
x
x [1, 1]
the wave equation in the form (33.2).

So here we have the density function (x) = 1 x2 on [1, 1]. By Theorem
22.4, (H) = {n(n + 1)}n0 and the Legendre polynomials {Pn (x)} are the
209
associated eigenfunctions and form an orthogonal basis in L2 (1, 1). We have

1 dn 2
Pn (x) =
(x 1)n .
n!2n dxn
Then we need only solve
un = 0
u
b00n + n(n + 1)b
where the derivatives are now in t.
On a finite interval, we can get tons of orthogonal polynomial bases by

this method by taking a different density function (Chebyshev, etc.). Ambartsumyan also famously showed in 1929 that for a finite interval, if (H) = {n2 }
then q 0. This will not be the case for an infinite interval.
Infinite interval That is
(
utt + Hu = f
< x <
.
(33.6)
initial conditions
Typically, H is selfadjoint (i.e. q real-valued), and hence (H) R and it
has a continuous component. I.e.
(H) = d (H) c (H).
There are now a lot of cases for the spectrum: the spectrum can still be
purely discrete, it can be discrete on top of continuous, you could have negative
eigenvalues, and there could be infinitely many of them, an accumulation to
zero, the continuous spectrum may extend past zero, etc.
It is also quite typical that c (H) is not simple but of multiplicity two,
i.e. if c (H) then there are exactly two different (linearly independent)
eigenfunctions of the continuous spectrum + (x, ), (x, ).
Then the complete system of eigenfunctions of the discrete and continuous
spectrum is
{n (x), (x, )}
where n are the eigenfunctions of the discrete spectrum and we have
hn , m i = nm ,
(x, ) (x, ) = ( ) ,
+ (x, ) (x, ) = 0.
Note that even with infinitely many n , they will not form a basis by themselves; the are needed to cover the rest of the space as we will see in the
next theorem.
d2
Recall also that for q 0, i.e. H = 2 , by Theorem 30.2 we have
dx
1
d (H) = , c (H) = [0, ) , and (x, ) = ei x .
2
It is no longer true if q 6= 0. Furthermore, if (H) = [0, ), this does
not imply that q(x) 0 (counterexamples abound). But, amazingly enough,
Fouriers Integral Theorem remains valid in the following edition.
210
d2
+ q(x) on L2 (R) and q(x) is such that
2
dx
(H) = d (H) c (H). Let {n (x)} be eigenfunctions of d (H) and let
{ (x, )} be eigenfunctions of c (H). Then any u L2 (R) can be represented
as
X
u(x) =
u
bn n (x) +
u
b+ ()+ (x, )d +
u
b () (x, )d ,
(33.7)
Theorem 33.2. Let H =
c (H)
where u
bn = hu, n i , u
b () =
c (H)
u(x) (x, )dx.

R
Note that the above becomes the Fourier transform for q = 0 with a square
root:

i x
u
b+ () =
u(x)e
dx , u
b () =
u(x)ei x dx
0
(and the two are put together if u is even). So u

b is referred to as the generalized
Fourier transform.
Note that again by considering various qs, one can create lots of special
functions corresponding to . For example, you can obtain Bessel functions,
and then u
b is referred to as the Fourier-Bessel transform.
Now how does one handle (33.6) in this situation?
The approach developed in Lecture 32 works perfectly! (with some modifications though)
Namely, we are looking for the solution to (33.6) in the form (33.7). Next,
X
2
2
00
u
b
(,
t)
(x,
)d
+
u
b (, t) (x, )d ,
utt =
u
bn (t)n (x) +
+
+
2
2
c (H) t
c (H) t
X
u
b (, t) (x, )d.
u
b+ (, t)+ (x, )d +
Hu =
n u
bn (t)n (x) +
c (H)
c (H)
Plugging it into (33.6), we get

(
bn (t) = fbn (t)
u
b00n (t) + n u
u
b00 (, t) + b
u (, t) = fb (, t)
where
fbn (t) =
f (x, t)n (x)dx
(33.8)
fb (, t) =
f (x, t) (x, )dx.

R
Then we solve the solve the differential equations (33.8) by variation of

parameters using the initial conditions. So we get u
bn (t), u
b (, t) and write the
solution u(x, t) of (33.6) in the form (33.7). Done!
LECTURE 34
L2 -spaces on Domains in Multidimensional Spaces

We are finally moving into two and higher dimensional spaces. There is a 3dimensional version of the Fourier transform, but here well introduce a new tool:
the Laplace operator. The method however will carry over.
Definition 34.1. Let be a domain in Rn . The set of all functions f (X) , X =
(x1 , x2 , , xn ) , subject to
2
2
|f (X)| dV
|f (x1 , x2 , , xn )|2 dx1 dx2 dxn <
kf k
is called the set of square integrable functions on and commonly denoted by L2 ().
y
Typical notation:
2
1) [0, 1] [0, 1] is a rectangle in R2 :
x
2) [0, 1] [0, 1] [0, 1] [0, 1]3 is a unit cube in R3 :
y
x
[0, 1]3 {(x, y, z) R3 : 0 x 1, 0 y 1, 0 z 1}.

z
S2
3) S 2 is a unit sphere in R3 :
y
x
S = {(x, y, z) R : x + y + z = 1}
z
4) T2 is a unit torus1 in R3 :
y
1
2
x
1I.e.
a bagel, and the unit torus T1 in R2 is the unit circle.

211
[0, 1]3
3
2
34. L2 -SPACES ON DOMAINS IN MULTIDIMENSIONAL SPACES
212
5) [0, 1]n is a unit cube in Rn :

[0, 1]n = {(x1 , x2 , , xn ) Rn : 0 xi 1 , i = 1, 2, , n}
6) B n is a unit ball in Rn :
B n = {(x1 , x2 , , xn ) Rn : x21 + x22 + + x2n 1}
7) S n1 is a unit sphere in Rn , i.e. S n1 = B n , the boundary of B n ,
S n1 = {(x1 , x2 , , xn ) Rn : x21 + x22 + + x2n = 1}.
Note that S n1 is an object of dimension n 1 in a space of dimension n.
Theorem 34.2. L2 () is a Hilbert space.
Proof. Let us show first that if f, g L2 () then f + g L2 ().
|f + g|2 = (f + g)(f + g) = |f |2 + |g|2 + 2 Re f g
|f |2 + |g|2
.
Re f g |f g|
2
So |f + g|2 2 |f |2 + 2 |g|2 and hence
2
2
|f + g| dV 2 |f | dV + 2 |g|2 dV < .
So L () is a linear space. We have to show now that L2 () has a scalar product.

Set
hf, gi =
f (X)g(X)dV.
It clearly has all the properties of a scalar product

hf1 + f2 , gi = hf1 , gi + hf2 , gi
hf, gi = hg, f i
hf, f i 0
and hence L2 () is a Hilbert space.
QED
Actually, L2 on domains in Rn is not harder than L2 on intervals.

Example 34.3. L2 ([0, 1]3 ). Its the set of all functions of three variables defined on
the unit cube such that
1 1 1
|f (x, y, z)|2 dx dy dz < .
(34.1)
0
Any continuous function is subject to (34.1).

More generally, any bounded function meets (34.1).
Lets look at unbounded functions in L2 ([0, 1]3 ). Consider
1
f (x, y, z) =
,
, , > 0.
x y z
Note that f (X) , X 0.
dx dy dz
x2 y 2 z 2
|f (x, y, z)| dx dy dz =
[0,1]3
[0,1]3
213
dx
x2
dz
2
0
0
0 z
1
1
1
x2+1
y 2+1
z 2+1
=
.
2 + 1 0 2 + 1 0 2 + 1 0
=
dy
y 2
(34.2)
So if 2 + 1 > 0 , 2 + 1 > 0 , 2 + 1 > 0, or , , < 1/2 then every factor in

(34.2) is finite and
1
|f (x, y, z)|2 dx dy dz =
< .
(1 2)(1 2)(1 2)
[0,1]3
Therefore, if , , < 1/2 then f (x, y, z) L2 ([0, 1]3 ); otherwise f

/ L2 ([0, 1]3 ).
So L2 () contains unbounded functions.
(
Example 34.4. L2 (R3 ) =
f (x, y, z) :
)
2
|f (x, y, z)| dx dy dz < .
1
1
2
3
L
(R
),
/ L2 (R3 ),
3 xyz
1 + (x2 + y 2 + z 2 )
L2 (R3 ), but xyze|x||y|
/ L2 (R3 ).

xyze|x||y||z|
Example 34.6. L2 (B 3 ).
z
Spherical coordinate system:
x = r sin cos
y = r sin sin
z = r cos
z
(x, y, z)
0 r < , 0 2 , 0 .
y
r
Or by setting r = r sin , we can write:
x = r cos
y = r sin
z = r cos
214
Compute the absolute value of the Jacobian2 of (x, y, z) (r, , ):

x y z

r r r

x y z = r2 sin .
|J(r, , )| = det

x y z

Hence if f L2 (B 3 )
y
|f (x, y, z)| dx dy dz =
0
B3

2

f (r, , ) sin d d r2 dr
(34.3)
where f(r, , ) = f (r sin cos , r sin sin , r cos ).

1
Let us show that f (x, y, z) = 2
L2 (B 3 ).
2
(x + y + z 2 )1/2
Indeed by (34.3)

2
1
1 2
1
2
2
sin d d r dr =
sin d
d
dr = 22 = 4
kf k =
2
0
0
0
0
0
0 r
i.e. f L2 (B 3 ).
This example is significant since f (X) represents the Coulomb potential. Note that
1
it likes higher dimensions since f (x) =
/ L2 (B 1 ) = L2 (1, 1), and f (x, y) =

|x|
1
p
/ L2 (B 2 ) since
2
2
x +y
2 1
1
1
2
kf k =
dx dy =
r dr d =
2
2
2
B2 x + y
0
0 r
but weve just seen that f (x, y, z) L2 (B 3 ).
Note that exact computation of norms are rare. Most of the time, youll need to
use estimates to prove membership into L2 ().
Definition 34.7. The operator defined in L2 () as follows
u
2u
2u 2u
+
+
+
x21 x22
x2n
u L2 ()
is called the Laplace operator.

2The
sign.
Jacobian can be rearranged some other ways too, but it will lead at most to a change of
215
d2
, and in R3 we have:
2
dx
2u 2u 2u
+
+
i.e. = x2 + y2 + z2 .
u =
x2 y 2 z 2
Clearly, is a linear operator.
So the Laplace operator in L2 (R) is
Definition 34.8. Let u(x, y, z) be a function of 3 variables, then

u
hu, ni , n R3 , knk = 1
n
is called the directional derivative with respect to n.
Note that is the gradient and n is often a normal vector to some surface.
Theorem 34.9 (Greens Formula). Let be a connected domain in R3 and be
its boundary, then for any u, v

y
x v
u
(uv vu)dx dy dz =
u
v
dS
n
n
where n is the external normal vector to .

This formula is a 3-space analog of integration by parts.
Definition 34.10. The Laplace operator in L2 () with domains

n
o

1) Dom = u L2 : u L2 , u = 0

u

2) Dom = u L2 : u L2 ,
=0
n

u

2
2
3) Dom = u L : u L ,
+ hu = 0 , for some h R
n
are called, respectively,

1) Dirichlet Laplacian
2) Neumann Laplacian
3) Robin Laplacian .
Theorem 34.11. All Laplacian (1)-(3) are selfadjoint operators.
Proof. Consider only case (1). Let u, v Dom ,
hu, vi =
u v dx dy dz
Thm 34.9
:
0
x u

v

uv dx dy dz +
v
u
dS
n
n

= hu, vi .
QED
Exercise 34.12. Prove Theorem 34.11 for Neumann and Robin Laplacians.
As you can guess, our goal will be the spectral analysis of .
But we start from the Laplace equation
u = 0.
LECTURE 35
The Laplace Equation in R2 . Harmonic Functions

In this lecture we are going to deal with 2-space, i.e. deal with the Laplace equation
uxx + uyy = 0.
(35.1)
In R there is a remarkable relation between this equation and Complex Analysis.
2
Definition 35.1. A function u(x, y) is harmonic on a domain R2 if u is

subject to (35.1).
Theorem 35.2. Let be a domain in R2 .
u = 0 , (x, y)
u(x, y) = Re f (z) for some function f analytic on .
Proof.
1) . Let f (z) = u(x, y) + iv(x, y) H(). This means that u, v
are subject to the Cauchy-Riemann conditions:
ux = vy
uy = vx
(x, y) .
(35.2)
Differentiate the first equation in (35.2) in x and the other in y:

uxx = vyx
uyy = vxy .
Since vyx = vxy then uxx + uyy = 0 (x, y) .

2) . Let u be harmonic, i.e. u = 0. We need to show that there exists a real
function v(x, y) such that f (z) = u(x, y) + iv(x, y) H().
Consider uy dx + ux dy. By the Green formula (Theorem 34.9), for any
contour C
x
(uy dx + ux dy) =
(uxx (uyy ))dx dy = 0.
C
Int C
Therefore, uy dx + ux dy is the total differential of some function v, i.e.

dv = uy dx + ux dy.
(35.3)
But on the other hand

dv = vx dx + vy dy.
(35.4)
Comparing (35.3) & (35.4) one has

ux = vy
uy = vx
f = u + iv H()
since its just the Cauchy-Riemann conditions.

QED
217
35. THE LAPLACE EQUATION IN R2 . HARMONIC FUNCTIONS
218
This statement is too general to be useful although it gives us the general structure
of a solution to u = 0: the real part of an analytic function.
Consider the Dirichlet problem for the Laplace equation on the unit disk D
{(x, y) : x2 + y 2 < 1}:
(
u
=0
.
(35.5)

u = () , 0 2
D
We can apply Theorem 35.2 to find a series representation for the above problem,
and we present such a solution in Appendix to this lecture.
But for now, let us find an integral representation of the solution to (35.5). We
need some ingredients.
Lemma 35.3. Let u be harmonic on the unit disk then
2

1
u(0) =
(35.6)
u eit dt.
2 0
Proof. By the Cauchy formula (Theorem 3.2) with z0 = 0,
2
2

1
f (z)
1
f (eit )
1
it
f (0) =
dz =
f eit dt
ie dt =

2i D z
2i 0
2 0
eit
2
2

1
1
Re f eit dt =
u eit dt.
QED
u(0) = Re f (0) =
2 0
2 0
Remark 35.4. Lemma 35.3 is called the mean value theorem.
Lemma 35.5. The function (z) defined by
z z0
(35.7)
(z) =
1 z0z
where is unimodular, i.e. || = 1, and z0 D, transforms the unit disk onto itself.
This transformation is known as a Mobius transform or linear fractional transformation.

Proof. We need only to prove that (35.7) transforms the unit circle D = {z :
|z| = 1} onto itself.1 Indeed, as long as |z| = 1,

2
z z0 2
2

= |z z0 | = (z z0 )(z z 0 )
| (z)| =
1 z0z
(1 z 0 z)(1 z0 z)
|1 z 0 z|2
=
1One
|z|2 + |z0 |2 2 Re zz 0 |z|=1 1 + |z0 |2 2 Re zz 0

=
= 1.
1 + |z 0 z|2 2 Re zz 0
1 + |z0 |2 2 Re zz 0
QED
of the properties of conformal mapping is that the boundary of the domain is mapped onto
the boundary of the image of the domain. Then using the fact that for z = 0, (0) = z0 , one finds
that (0) = |z0 | < 1 since z0 D and so the image is inside the circle, i.e. it is the disk.
Take = 1 in (35.7). Then (z) =
219
z0 z
and z0 = 1 (0).
1 z0z
Change variables in (35.6) as follows

(By Lemma 35.5 its possible!)
eit = ei

1
Then t = ln ei and we have (z = ei )
i
1 dz d
dt
d 1
=
ln ei =
ln (z)
d
d i
i d dz

d
d
1
1
z0
ln(z0 z)
ln(1 z 0 z) = z
+
= iz
i
dz
dz
z z0 1 z 0 z
|z|2 =zz=1
1
z 0
z +
zz0 |z0 |2
1 |z0 |2
1 |z0 |2
=
=
z(z z0 )(1 z 0 z)
(1 z 0 z)(1 z 0 z)
(|z|2 zz0 )(1 z 0 z)
1 |z0 |2
1 |z0 |2
1 |z0 |2
=
=
.
|1 z 0 z|2
|z z 0 |2
|z z0 |2
The function (z0 rei0 )

1 1 |z0 |2
1
1 dt
1 r2
=
=
Pz0 ()
(35.8)
2 d
2 |1 z 0 z|2
2 1 2r cos(0 ) + r2
is called the Poisson kernel.
One can easily check (Exercise 35.7) that the function u ( 1 (z)) is harmonic in D
and then by Lemma 35.3 one has
2
2
change of variables formula 1
dt

1
1 it
1
u (e ) dt
=
u ei
d
u (0) =
2 0
2 0
d
2
1 dt
u ei
=
dt
2 d
0
which by (35.8) is
2

1
u ei Pz0 ()d.
u (0) =
0
Since as we know, 1 (0) = z0 we have

2

u(z0 ) =
Pz0 ()u ei d
0

i
but u e
= () (our boundary condition) and we finally arrive at the following
(where we write z instead of z0 )
Theorem 35.6 (Poisson Formula). The solution of problem (35.5) can be represented by the Poisson formula
2
1 |z|2
1
()d , z = x + iy
(35.9)
u(z) =
2 0 |ei z|2
for any (x, y) D.
220
So we got an explicit formula in integral form but this is rare. To go further to an

algebraic form may be possible if is analytic by using residues.
Note though that the kernel diverges at the boundary. However, we will get a

-sequence, allowing us to have u = satisfied.
D
Exercise 35.7. Prove that if u(z) is harmonic (i.e. u = 0) on the unit disk
D = {z : |z| < 1} then so is u ( 1 (z)) where (z) is defined by (35.7).
Exercise 35.8. Graph Poisson kernels as a function of for different 0 r < 1.
Prove that Poisson kernels represent a -family r ( 0 ) as r 1.
Appendix. Series Representation of the Solution to the Dirichlet Problem
for the Laplace Equation
Consider the Laplace equation as a Dirichlet problem on the unit disk, i.e. (35.5):

u = 0 , u = 0.
D
By Theorem 35.2, there exists a function f (z), analytic in D such that u = Re f .

On the other hand, by the Taylor theorem (Theorem 4.12)
X
X

f (z) =
an z n =
an rn ein
z = rei
n0
n0
+ an ein
u = Re f =
r
2
n0
1 X n in
1 X n in
+ Re a0 +
=
r an e
r an e
2 n1
2 n1
1 X |n|
1 X n in
=
r an ein + Re a0 +
r an e .
2 n1
2 n1
X
n an e
in
So we get a series representation for (35.5)

X
u rei =
An r|n| ein
0r<1
nZ
an
2
where An = Re a0
1 an
2
Exercise 35.9.
n 1
n=0
n1
(35.10)
.
Show that (35.9) implies (35.10) and vice versa.
LECTURE 36
Uniqueness of the Solution of the Laplace Equation. Laplace

Equation in a Rectangular Domain
1. Well-Posedness of the Laplace Equation
In an electric field, we can measure the potential on the boundary and using the
solution to the Laplace equation, we rebuild the inside. But we havent asked ourselves
some fundamental questions about PDEs: given a PDE is there a solution? is this
solution unique? does this solution offer some continuity properties with respect to
variations in initial or boundary conditions?
These three questions together constitute well-posedness: existence, uniqueness and
continuity.
As far as existence is concerned, so far weve given constructive proofs, but overall
its a difficult question especially for nonlinear equations. Wellposedness is not always
fully proven, for example it is not for the famous Navier-Stokes equation. Here we will
address the uniqueness of solution in the case of the Dirichlet problem for the Laplace
equation.
Theorem 36.1 (Uniqueness Theorem). Let be a domain in R2 . Then there is
only one solution to
(
u
=0
.
(36.1)

u =
Proof. Suppose that there exist two solutions u1 , u2 to (36.1). I.e.

(
(
u 1 = 0
u 2 = 0
.

u1 =
u2 =
It follows from here that u u1 u2 is a solution to

(
u
=0
.

u =
(36.2)
We are going to show that u 0.

Lemma 36.2 (Greens First Identity).
x
v
u ds =
(uv + ux vx + uy vy )dxdy.
n
(No proof).
221
(36.3)
222
36. LAPLACE EQUATION. UNIQUENESS; RECTANGULAR DOMAIN
Consider now (36.3) for u = v:
x
x
u
u ds =
(u |{z}
u +u2x + vy2 )dxdy =
(u2x + vy2 )dxdy.
n
=0

But u = 0. Therefore
x
(u2x + vy2 )dxdy = 0 u2x + vy2 0 in u = const in .

But since u = 0 this const = 0.
So we showed that (36.2) has only a trivial solution, hence u1 = u2 and that means
uniqueness.
QED

u
Note that if now we have Neumann conditions, i.e.
= 0, then solutions can
n
only differ by a constant.
2. Dirichlet Problem in a Rectangle
y
(x, y) = f (x)
b
=0
Consider the following problem:

(
u
=0

u =
=0
(36.4)
For example, you can imagine 3 walls

grounded and one charged.
a x
0
=0
Let us look for a solution to (36.4) in the form1
u(x, y) = X(x)Y (y)

then
ux = X 0 Y
,
uxx = X 00 Y
0
uy = XY
,
uyy = XY 00
and u = X 00 Y + XY 00 = 0. Dividing by XY yields
X 00 Y 00
+
=0
X
Y
or
X 00 (x)
Y 00 (y)
=
.
X(x)
Y (y)
So for some R (since we are looking for real solutions)

00
X (x) = X(x)
Y 00 (y) = Y (y)
(36.5a)
(36.5b)
These are second order ordinary linear homogeneous equations.

Let us find now boundary conditions for X, Y .
1This
particular method is known as separation of variables; more generally, in physics, one

often comes up with an Ansatz and see if it is now solvable, then uniqueness takes care of the rest.
2. DIRICHLET PROBLEM IN A RECTANGLE
From (36.4) we have
u(x, 0) = 0
u(0, y) = 0
u(a, y) = 0
u(x, b) = f (x)
X(x)Y (0) = 0
X(0)Y (y) = 0
X(a)Y (y) = 0
X(x)Y (b) = f (x)
223
Y (0) = 0
X(0) = 0
X(a) = 0
and hence from (36.5a) we have

X 00 = X
X(0) = X(a) = 0.
We now solve this equation2
X = A sin x + B cos x
X(0) = B cos x = 0 B = 0
X(a) = A sin x = 0
a = n ,
n 2
=
.
a
So by choosing A = 1, we get
Xn (x) = sin
nZ
n
x.
a
Then it follows from (36.5b) that

Y 00 = Y
Y (y) = Ae
Y (0) = 0
+ Be
Y (0) = A + B = 0
, where =
n 2
a
B = A
and for Y , by choosing A = 1/2 we get

Yn (y) =
n
y
a
e
2
n
y
a
or
Yn (y) = sinh
n
y.
a
So finally we get that for any n = 1, 2,

n
n
x sinh
y
a
a
is a solution to the Laplace equation vanishing at x = 0.
Functions un (x, y) are called rectangular harmonics.
The general solution to (36.4) is then
un (x, y) = Xn (x)Yn (y) = sin
u(x, y) =
X
n=1
2We
cn sin
n
ny
x sinh
.
a
a
find that we need 0 since if < 0 then X(x) = Ae

X(0) = X(a) = 0 unless X 0.
+ Be
cannot satisfy
224
36. LAPLACE EQUATION. UNIQUENESS; RECTANGULAR DOMAIN
Let us find now {cn }. Since u(x, b) = f (x) we have

X
nb
n
f (x) =
cn sinh
sin
x.
a
a
n=1 |
{z
}
bn
But on the other hand, {bn } are the sin-Fourier coefficients of f (x) so
2 a
n
bn =
f (x) sin
xdx
a 0
a
||
1
cn sinh nb
cn = sinh nb
bn
a
a
and
sinh ny
n
a
sin
x
u(x, y) =
bn
nb
a
sinh a
n=1
2 a
n
where bn =
f (x) sin
xdx.
a 0
a
LECTURE 37
The Principle of Conformal Mapping for the Laplace

Equation. Laplace Equation in the Upper Half Plane
We are going to demonstrate how Complex Analysis helps to solve Laplace equations in R2 .
Definition 37.1. A univalent function produces two different values for two different points.
Definition 37.2. Let f (z) be an analytic univalent function on a simply connected
domain and f 0 (z) 6= 0 z . Then the image 0 of under the transformation
f (z) is called the conformal mapping of under f and we write
0 = f () = {f (z) : z }.
The condition f 0 (z) 6= 0 is very important.
Theorem 37.3 (Riemann Theorem on Conformal Mapping). Given two simply
connected domains1 , 0 C,
f H() : 0 = f () is a conformal mapping of .
(No proof).
f
0
This is not constructive...

The following theorem clarifies why conformal mapping is so important.
Theorem 37.4. The Laplace equation is invariant under a conformal mapping.
1Rigorous
statements of this theorem can be found in advanced books on complex analysis.

225
226
37. CONFORMAL MAPPING. APPLICATION TO THE LAPLACE EQUATION
The exact meaning of this statement will be explained a bit later, but recall notions
of invariance: time invariance means some quantity doesnt depend on time, for example the energy of a system; here the equation is invariant with respect to conformal
mapping, i.e. you can change the coordinate system (via an analytic function2), and
the equation has still the same form.
Proof. Consider the Laplace equation on some domain :
2u 2u
u =
+
=0
x2 y 2
(x, y) .
Let f (z) be an analytic function performing the conformal mapping of onto

another domain 0 , i.e.
w = f (z) 0
w = + i.
We will prove that

2U
2U
+
= 0 , (, ) 0
(37.1)
2
2
where U (, ) u (x (, ) , y (, )). Since f : 0 is a one-to-one correspondence,
f 1 (w) maps 0 onto and we have
(
x = x (, )
.
y = y(, )
So what we have is a change of coordinate system, and rules for partial derivatives
have been put for review in Appendix to this lecture, but basically, we have by the
chain rule
u
U U
=
+
= U x + U x
x
x
x
2u
=
(U x + U x ) = (U )x x + U xx + (U )x x + U xx
2
x
x
= (U x + U x ) x + U xx + (U x + U x ) x + U xx
= U x2 + U x x + U xx + U x x + U x2 + U xx
uxx = U x2 + 2U x x + U x2 + U xx + U xx .
(37.2)
Similarly, one has

uyy = U y2 + 2U y y + U y2 + U yy + U yy .
(37.3)
Add (37.2), (37.3) up. We get

0 = uxx + uyy

= U x2 + y2 + 2U (x x + y y ) + U x2 + y2
+ U (xx + yy ) +U (xx + yy ) .
| {z }
| {z }
=
2This
condition is important, otherwise it screws up the Laplace equation.
(37.4)
227
Now our function f (z) = (x, y) + i(x, y) is analytic. So by Theorem 35.2, and
are harmonic = 0 , = 0. Also since f is analytic, the Cauchy-Riemann
conditions apply
(
x x + y y = x y + y x = 0
x = y , y = x
x2 + 22 = x2 + y2
and (37.4) reads
uxx + uyy = (U + U ) x2 + y2
||
0
U + U = 0
since x2 + y2 6= 0 and we are done!
QED
Let us now figure out what Theorem 37.4 actually means. It allows one to solve
Laplace equations for various domains, not only disks or rectangles.
Indeed, consider the Dirichlet problem
y
z
(
u
=0

u =
(37.5)
where =
x
By the Riemann Theorem (Theorem 37.3), there exists an analytic

function f (z) that conformally maps onto a unit disk D. Problem
(37.5) then reads ( U (w) = u (f 1 (w)) )
(
w
U
(w)
=
0

(37.6)

U = f 1 ei
f (z)
0 = D
since the image of the boundary

is the boundary of the image in
conformal mapping, i.e. gets
mapped to D.
w = + i
We solve next (37.6) by Poissons formula (35.9).
2

1
1 |w|2
U (w) =
f 1 ei d.
2
2 0 |ei w|
We now come back to our original variable z and have
u(z) = U (w)
and problem (37.5) is solved!
where
w = f (z)
(37.7)
228
This is a top-of-the-line equation since it is explicit. But the drawback is finding

the conformal map. This is big business for aerodynamics (in aircraft building) and is
the object of contemporary research.
We demonstrate this algorithm on a particular example.
Example 37.5. The Dirichlet problem on a half plane.
(
u
=0

u =
where =
(37.8)
x
Physically, this problem can be interpreted in the following way. We have some
electrostatic field u(x, y) in the upper half plane C+ . We know the potential (x) on
its boundary R. How to recover the potential u(x, y) in the whole C+ ?
Solution. According to the procedure oulined above, we need to find a conformal
mapping that maps C+ onto D.
This conformal mapping is well-known3
w
z
z=i
1w
.
1+w
(37.9)
Let us check it:

zz
1 1w
1w
1 1w 1w
Im z =
=
i
+i
=
+
2i
2 1+w
1+w
2 1+w 1+w
=
1 1 + w
w |w|2 + 1 w +
w |w2 |
1 |w|2
=
0
2
|1 + w|2
|1 + w|2
with equality for |w| = 1. So, indeed (37.9) maps the unit circle onto the real line, and
onto
.
w=
3Note
iz
f (z)
i+z
(37.10)
that a conformal mapping is not unique; here for example, we could multiply by a unimodular factor, i.e. do a rotation.
and plugging it in (37.7)

1
u(z) =
2

1 |f (z)|2
f 1 ei d.
2
|ei f (z)|
229
(37.11)
1 ei
From (37.9) one has i
= t R and if runs through (0, 2) then t runs
1 + ei
through (, ).
So lets make the following change of variables
t=i

1 ei
i
1
e
=
f
1 + ei
ei =
it
.
i+t
Then
2i dt
2 dt
2 dt
(i + t)(1) (i t)(1)
dt d = it
=
=
.
2
2
2
(i + t)
1 t
1 + t2
i i+t (i + t)

Since t = f 1 ei , (37.11) then reads
iz 2

2 dt
dt
1 |f (z)|2
1 1 i+z
1
(t)
=
u(z) =

2 (t)
2
2 |f (t) f (z)|
1 + t2
it iz
1 + t2
i+t
i+z
dt
1
|z + i|2 |z i|2

2

|i
+
t|
=

2 (t)

1+
t2
|(i t)(i + z) (i z)(i + t)|
1
(z + i)(z i) (z i)(z + i)
=
2 (t)dt
|1 + iz it
+ 1

zt
zt|
it + iz +
1 |z|2 + iz iz + 1 |z|2 iz + iz 1
=
(t)dt

|2i(z t)|2
1 zz
1 Im z
1 2i(z z)
(t)dt =
(t)dt =
(t)dt
=
4 |t z|2
2i |t z|2
|t z|2

y
z=x+iy 1
=
(t)dt.
(t x)2 + y 2
iei d =
So the solution to (37.8) is

1
u(x, y) = y
(t) dt
(t x)2 + y 2
y 0 , x R.
(37.12)
Done!
The function
1
y
is also known as the Poisson kernel for the upper half
(x t)2 + y 2
plane.
Remark 37.6. Note that the solution (37.8) represents the convolution of the Poisson kernel and the boundary condition.
Note a couple of potential issues:
230
if (t) |t|, then the integral is not convergent. But has physical meaning,
so in practice, this will not be an issue;
how to reproduce u(x, y) when y 0? well actually, we have again a
-sequence, so this will be satisfied too (see Exercise 37.10).
In practice, even the simple integral (37.12) is hard to evaluate by hand. However,
in some cases we can avoid computing (37.12).
Indeed, u(x, y) is the real (or imaginary) part of some analytic function f (z) , z =
x + iy by Theorem 35.2. Moreover, (x) = Re f (x) (or Im f (x)). By the look of (x)
sometimes its possible to tell what f (x) and even f (z) is.
Example 37.7. In the numbered examples below, we consider the Dirichlet problem
in the upper half plane:
Note that since its an order 2 equation, we would need an extra condition besides
the boundary one given. This generally comes from physics. So here, we will rather
simply look for a continuation of our function in the upper half plane and check its
analyticity. Furthermore, we make no claim of uniqueness since here the domain is
infinite.
(
u
=0
1
.
(37.13)

u = 1
R
This problem can easily be solved by (37.12). But on the other hand, there is
an obvious analytic function f (z) in C+ which is identically 1 on R and whose
real part is a bounded function.4 We have f (z) = 1. That is
u(x, y) = 1
2
(x, y) C+ .
(
u
=0
.
(37.14)

u = ln |x|
R
Note here that the boundary condition is not bounded, but one can easily
figure out that the function f (z) = ln z is analytic on C+ and equals ln |x| on
R. Indeed,
1
y
ln(x2 + y 2 ) + i arctan .
2
x
1
If y = 0 then f (x) = f (x + i0) = ln x2 = ln |x|.
2
ln z = ln |z| + i arg z =
4Indeed,
one could argue that f (z) = 1 iCz for some real constant C is also analytic and
identically 1 on R. But it is not bounded, and hence u(x, y) = 1 + Cy does not make sense physically
if were looking for a finite energy field.
231
So the solution to (37.14) is
u(x, y) =
1
ln(x2 + y 2 )
2
(x, y) C+ .
equipotential
curves
u(x, y) = ln r
x2 + y 2 = r 2
u = 0

1
u =
x
R
Note that f (z) =
(37.15)
1
is analytic on C+ , and one easily has
z
u(x, y) = Re
x
1
= 2
.
z
x + y2
Exercise 37.8. Draw equipotential curves for (37.15).

The Riemann theorem (Theorem 37.3) is not constructive. That is, it does not
provide any explicit construction of f (z). However, there are some well-known domains
for which f (z) is available.
Example 37.9.
f (z) = z 4
/4
232
1
f (z) =
2

1
z+
z
the so-called Zhukovskys function.5

This function appears and was studied in the context of aerodynamics and
is part of early 20th -century research.
3
f (z)
f (z) = C
(a, b)
a
z
(a, b)
dt
elliptic integral
(1 t2 )(1 k 2 t2 )
where C, k are defined from solving the equations
1
1/k
dt
dt
p
p
=b
,
a
(t2 1)(1 k 2 t2 )
(1 t2 )(1 k 2 t2 )
0
1
b
C = 1/k
.
dt
p
(t2 1)(1 k 2 t2 )
1
This formula is a particular case of the more general Schwarz-Christoffel
conformal mappings which describe how to transform the upper half plane into
polygons. As one may imagine, such formulas get more and more complicated,
but at least they are known.
One can now figure out why we could solve in Lecture 36 the Laplace equation
on a rectangular domain only with some restrictions.
From the above, you may also appreciate the following funny story: a physicist needed to solve a Laplace problem and decided to use a simple model for
the boundary. So he asked the mathematician to solve his problem for a square.
After much complicated calculations, the mathematician gave his answer, and
then the physicist said that now he needed to have it altered to fit his original
data: a circle!
0
5The
transform itself is often referred to as the Joukowsky transform after the same person since
Russian names can be transcribed in the Latin alphabet in multiple ways.
APPENDIX. CHANGE OF COORDINATE SYSTEM AND THE CHAIN RULE
233
Exercise 37.10.
1
Prove that the Poisson kernel on the upper half plane is a -sequence for y 0.
Note that you must show among other things that:
1
y dt
= 1.
(x t)2 + y 2
Show that if (t) in (37.12) is odd then

y
1
1
(t)dt.
u(x, y) =
0
(x t)2 + y 2 (x + t)2 + y 2
Show also that it solves the following Dirichlet problem
, x>0, y>0
u = 0
.
u(0, y) = 0
, y>0
u(x, 0) = (x) , x 0
Solve the following Dirichlet problem

y > 0:
u = 0
u(x, 0) = (x)
u(0, y) = (y)
by assuming that u is bounded for x > 0,

,
,
,
x, y > 0
.
x0
y0
Appendix. Change of Coordinate System and the Chain Rule

We present here as a review an important chain rule related to a change of variables
in the case of two dimensions.
Let f (x, y) be a function of (x, y) and
(
x = x(, )
y = y(, )
a change of variables. If
F (, ) := f (x(, ), y(, ))
is the function f in the new variables , then
F (, ) F (, )
f (x, y) =
+
x
x
x
F (, ) F (, )
f (x, y) =
+
y
y
y
or in a compact form
fx = F x + F x
fy = F y + F y
234
or in the matrix form

fx
x x
F
=
.
fy
y y
F
| {z }
Jacobis matrix
And for the second partial derivative (in x), since fx = F x + F x , one has
fxx
(F x )x
(F )x x
F x2
(F x )x
F xx
F x x
F xx
(F )x x
F x x
F x2
Adding up the circled expressions, we get

fxx = F x2 + 2F x x + F x2 + F xx + F xx .
LECTURE 38
The Spectrum of the Laplace Operator on Some Simple

Domains
1. The Spectrum of on a Rectangle
Let us solve the following eigenfunction problem
(
u
= u

(Dirichlet eigenfunction problem)

u = 0,
(38.1)
where is
a
0
Set u(x, y) = X(x)Y (y), then (38.1) reads
d2
d2
X
Y = XY
dx2
dy 2
1 d2 X 1 d2 Y

=
separable equation!
2
2
| X{zdx } | Y{zdy }
1
2
00
X = 1 X
.
Y 00 = 2 Y
+ =
1
2
(38.2)
Lets find now boundary conditions for X, Y .

u(0, y) = X(0)Y (y) = 0
X(0) = 0 ,
u(a, y) = X(a)Y (y) = 0
X(a) = 0 ,
u(x, 0) = X(x)Y (0) = 0
Y (0) = 0 ,
u(x, b) = X(x)Y (b) = 0
Y (b) = 0 .
So we got
(
X 00 = 1 X
X(0) = 0 = X(a)
235
X = Aei
1 x
+ Bei
1 x
236
38. SPECTRUM OF THE LAPLACE OPERATOR
Find now A, B.
(
X(0) = A + B = 0 B = A
X(a) = Aei 1 a Aei 1 a = 0
p
e2i 1 a = 1 2 1 a = 2 n , n Z
2
n
n
n
or 1,n =
, n N = {1, 2, } ; Xn (x) = A ei a x ei a x .
a
1
we finally get
Taking A =
2i
n
Xn (x) = sin
x , n N.
a
In a very similar way,
Ym (y) = sin
m
y
b
m N.
So all the solutions to (38.1) are

un,m (x, y) = sin
m
n
x sin
y
a
b
and the set of all possible in (38.1) is

n 2 m 2
2
= 1 + 2 =
+
a
b
So we proved
n, m N.
Theorem 38.1. Let D be the Dirichlet Laplace operator on a rectangle =

[0, a] [0, b], i.e.

D u = u & u = 0.
Then (D ) is purely discrete and

n 2 m 2
2
(D ) =
+
; n, m N
a
b
and corresponding eigenfunctions are
n
m
un,m (x, y) = sin
x sin
y.
a
b
Theorem 38.1 can be easily generalized to R3 .
Theorem 38.2. Let D be the Dirichlet Laplace operator on a domain =
[0, a] [0, b] [0, c], i.e.

D u = u & u = 0.
Then
(
(D ) = d (D ) =
n 2
a
m 2
b
)
2 !
k
+
; n, m, k N
c
1. THE SPECTRUM OF ON A RECTANGLE
237

n
m
k
x sin
y sin
z.
a
b
c
Discussions. So eigenvalues of the Dirichlet Laplace operator are now parameterized by 2 (or even 3 integers). If we order the set (D ), it looks like

1
1
2
1 =
n=m=1
+
a2 b 2
2 !
1
2
2 = 2
+
n = 1, m = 2
2
a
b
!
2
2
1
2
3 =
+ 2
n = 2, m = 1
a
b
2 2 !
2
2
4 = 2
+
n=m=2
a
b
..........................
In the simplest case of a = b = we get
un,m,k (x, y, z) = sin
1 = 1 + 1
=2
2
2 = 1 + 2 = 2 + 1
2
3 = 2 + 2
=5
=8
4 = 1 + 32 = 32 + 1
= 10
5 = 22 + 32 = 32 + 22 = 13
6 = 1 + 42 = 42 + 1
2
7 = 3 + 3
= 18
8 = 2 + 4 = 4 + 2
= 17
= 20
9 = 3 + 4 = 4 + 3 = 24
.............................
3 4
5
6 7
10
...
20
2
So the spectrum is a lot more dense than for the operator

n2 = {1, 4, 9, 16, 25, }:
1
2
d
on [0, ] which was
dx2
5
...
0
10
20
In math physics they ask the following question: Can you hear the shape of a drum?
In our case it would be a box. This means: if we know all the tones (proper
frequencies) of a box can we then recover the shape of the box?
This problem belongs to so-called inverse problems and the answer is negative.
238
2. The Spectrum of on a Disk

Similarly to (38.1) we have
D
(
u
= u

u = 0
(38.3)
1
This problem is easily solved in the polar coordinate system

(
x = r cos
.
y = r sin
Equation (38.3) then reads
1
r r

U
1 2U
r
2 2 = U
r
r
(38.4)
where U (r, ) = u(r cos , r sin ).

Exercise 38.3. Prove that in the polar coordinate system (38.3) transforms into
(38.4).
Note that variables in (38.4) separate and we set
U (r, ) = R(r)().
Plugging it into (38.4) one has
1 00 ()
1 (rR0 (r))0
+ 2
=
r R(r)
r ()
r
1
0
(rR0 (r)) + r2 +
00 () = 0
R(r)
()
|
{z
} | {z }
or
(
00 () = ()
r(rR0 (r))0 r2 R(r) = R(r)
00 =
1 d
d
r R + 2 R = R(r)
r dr
dr
r
(38.5)
Note that the second equation above is a Sturm-Liouville problem!

Let us now recompute the boundary conditions.

U (r, )
= u(r cos , r sin )
=0
[0, 2)
r=1
U (1, ) = 0
||
R(1)() = 0
r=1
R(1) = 0.
2. THE SPECTRUM OF ON A DISK
239
Next, we must have (due to continuity)

U
U
(r, 0)=
(r, 2)
, 0 (0) = 0 (2)
U (r, 0)= U (r, 2) ,

(0) = (2)
and it now follows from (38.5) that

(
00 =
(0) = (2) , 0 (0) = 0 (2)

1 d r dR + R = R
r dr
dr
r2
.
2
R(1) = 0 , R L (0, 1)
(38.6)
(38.7)
Solve (38.6) first

() = Aei
+ Bei
(0) = A + B = (2) = Aei 2 + Bei 2

0 (0) = (A B)i = i Aei 2 Bei 2
=

ei 2 = 1

2
2n
= n2 , n N.
So the solutions to (38.6) are

n () = An ein + Bn ein .
Now (38.7) reads

2
1 d r d R + n R = R , 0 r 1
r dr
dr
r2
R(1) = 0
Sturm-Liouville problem
or
1 0
n2
00
R R + 2 R = R
r
r
or finally we get

1 0
n2
R + R + 2 R=0 ,
r
r
00
0r1
All bounded solutions to this equation are
Rn (r) = AJn ( r).
The boundary condition at r = 1 requires
Jn ( ) = 0.
The equation has infinitely many solutions:
the Bessel equation
240
J0 (k)
J1 (k)
J2 (k)
k01
0
k11
k02
k12
k21
k03
k22
k13
k04
10
k23
k14
k05
k24
k15
k06
k25
k16
There is no nice formula for the roots but they can be computed numerically. We
call them knm (counted by two indices n, m). So
2
= knm
and we arrive at
Theorem 38.4. Let D be the Dirichlet Laplace operator on the unit disk D. The
spectrum (D ) is discrete and infinite and
2
(D ) = {knm
}
where knm are roots of the equations

Jn (k) = 0
n = 0, 1,
LECTURE 39
Nonhomogeneous Laplace Equation

We first state without proof the following fundamental theorem on the spectrum of
the Laplace operator on a bounded domain.
Theorem 39.1. Let be the Laplace operator with Dirichlet or Neumann or
mixed conditions on a bounded domain . Then the spectrum () is purely discrete
and consists of infinitely many positive eigenvalues {n } accumulating at .
The set of corresponding eigenfunctions forms an orthonormal basis in L2 ().
Remark 39.2. This problem:
(
u
= u

u = 0
is known as Helmholtz equation (from the wave equation in nonhomogeneous media).

Note also that if is unbounded, then we get a continuous spectrum.
For example, in the case of an infinite cylinder, eigenvalues are embedded in the continuous spectrum.
Why is Theorem 39.1 so important? It allows us to solve nonhomogeneous Laplace
equations as well as wave and heat equations.
Problem 39.3. Let be a bounded domain in Rm . Consider
(
u
=f

, where f L2 ().

u = 0
(39.1)
Solution. By Theorem 39.1, the spectrum of is discrete and made of eigenvalues {n }. Corresponding eigenfunctions {en (X)} form an ONB in L2 ().
Hence every function u L2 () can be represented as
X
u(X) =
u
bn en (X)
where
X = (x1 , x2 , , xm ) ,
n
u
bn = hu, en i
u(X)en (X) dx1 dx2 dxm .
241
242
39. NONHOMOGENEOUS LAPLACE EQUATION
Plug this into (39.1)
u
bn en
fbn en
X ||
X
u
bn (en ) =
n u
bn en
n
X
n u
bn fbn en = 0
{en } is an ONB
n u
bn fbn = 0.
But its just an algebraic equation!

u
bn =
fbn
.
n
And for the solution to (39.1) one has

u(X) =
X fbn
en (X)
n
n
X = (x1 , x2 , , xm ).
Note however that there is no way to generalize from dimension 1 how to obtain
n , en (X). So lets consider a particular example where this algorithm can be performed
by hand (explicitly).
Example 39.4. Consider the problem
b
(
u
=f

u = 0
where
and
0
f (x, y) = x + y.
By Theorem 38.1,

n 2 m 2
() =
+
; n, m N
a
b
un,m (x, y) = sin
nx
ny
sin .
a
b
We next normalize them:

b a
nx
my
2
kun,m k =
sin2
sin2
dx dy
a
b
0
0
a

b

2
2 nx
2 my
=
sin
dx
sin
dy =
ab.
a
b
2
0
0
|
{z
}
|
{z
}
a
2nx
b
1 cos a
1
dx = a
2
2
2
0
(39.2)
So kun,m k2 =
We now set
2
2
243
ab.
2
nx
my
en,m = sin
sin
a
b
ab
and {en,m } are all normalized by 1, i.e.
ken,m k = 1.
By Theorem 39.1, {en,m } is an ONB in L2 () and hence any u L2 () can be
represented as
u(x, y) =
u
bn,m en,m (x, y)
Double Sine Fourier series.

x
where u
bn,m =
u(x, y)en,m (x, y) dx dy
For f (x, y) one has

X
fbn,m en,m (x, y)
,
where
f (x, y) =
b a
2
my
nx
fbn,m =
sin
dx dy
f (x, y) sin
a
b
ab 0
0

b a
2
my
nx
=
dx sin
dy
(x + y) sin
a
b
ab 0
0

b a
2
nx
my
a
=
dy
sin
(x + y) d cos
n
a
b
ab 0
0
)
b (
a

0
:

a

2a
nx
nx
my

dx
=
(x + y) cos
sin
dy
+ cos

a 0
a
b
n ab 0
0
r b
2
my
a
{(a + y) |cos{zn} +y} sin
=
dy
n b 0
b
n
(1)
and by parts again

r b
a
my b
(y (a + y)(1)n ) d cos
b 0
b m
2
my b
n
=
ab (y (a + y)(1) ) cos

nm
b 0
2 ab
=
{(b (a + b)(1)n ) (1)m + a(1)n }
nm

2 ab
(1)m b + a(1)n+m + b(1)n+m a(1)n
=
nm
2 ab
=
{(1)n ((1)m 1) a + (1)m ((1)n 1) b} .
nm
2
fbn,m =
n
244
So
2
ab
fbn,m =
{(1)n ((1)m 1) a + (1)m ((1)n 1) b} .
nm
Now the solution is
(39.3)
X fbn,m
en,m (x, y)
(Double Fourier series)
n,m1 n,m
n 2 m 2
n,m =
+
,
a
b
2 ab
fbn,m =
{(1)n ((1)m 1) a + (1)m ((1)n 1) b} ,
nm
2
mb
nx
en,m (x, y) = sin
sin
a
y
ab
u(x, y) =
where
and the problem is completely solved.

Note that we verify that the fbn,m (generalized Fourier coefficients of f ) decay for
2
2
n, m and then we divide by na + mb , so the rate of decay for u
bn,m is like
3
3
n , m . It makes sense that it should be better since u has already two derivatives
(because of the order 2 equation).
Note also that by Galilean transformation could be rescaled to a square to
make it simpler.
One now may wonder what if in Problem 39.3 the boundary condition is not zero.
I.e. we have
Problem 39.5.
(
u
=f

u =
This problem can be reduced to other problems considered previously. Indeed,

we have a linear equation so the superposition principle applies, and therefore let
u = u1 + u2 where
(
(
u
=
0
u
1

2=f

u1 :
u
:
2
u1 =
u2 = 0
homogeneous equation,
nonhomogeneous boundary conditions
(I)
nonhomogeneous equation,
homogeneous boundary conditions
(II)
Equation (I) was the content of Lectures 35-37 (with conformal mappings and
such).
Equation (II) was considered in Problem 39.3, equation (39.1).
So the solution to Problem 39.5 now is
u = u1 + u2 where u1 is the solution to (I)
and
u2 is the solution to (II).
245
In the very same manner one can treat other than Dirichlet boundary conditions.

2
Exercise 39.6. Solve u = 1 in = [0, 1] subject to u = 0.
LECTURE 40
Wave and Heat Equation in Dimension Higher Than One

1. Wave Equation
Recall that this equation appears in various physical contexts: sound, electromagnetic waves, drum, surface waves, gravitational waves (even if undetected so far).
Consider the following initial value Dirichlet problem on a domain for the wave
equation (homogeneous media):
utt = u

u = 0
u(0, X) = (X)
,
ut (0, X) = (X)
(40.1)
X
We employ the eigenfunction expansion method. We only consider the case of a

bounded domain .
By Theorem 39.1, the spectrum of the Dirichlet Laplacian operator

D , u
=0
is discrete and the eigenfunctions form as basis for L2 ().

Let {n } be eigenvalues and {en } eigenfunctions. We now represent the solution of
(40.1) as
u=
u
bn en =
u
bn (t)en (X) ,
where u
bn = hu, en i
are the (generalized) Fourier coefficients.

We plug it back into the wave equation (40.1)
X
2 X
u
b
e
=
u
bn en
n
n
t2 n
n
b
u00n = n u
bn
u
b00n en =
u
bn = An cos
247
X
n
u
bn en =
X
n
p
p
n t + Bn sin n t.
u
bn (n )en
248
40. WAVE AND HEAT EQUATION IN HIGHER DIMENSIONS
Lets now find An , Bn . In order to do so we need to use the initial conditions.

Represent , as
X
X
=
bn en
,
=
bn en

X
X

u =
u
bn en =
u
bn en
t=0
t=0
X
X t=0
=
An en = =
bn en An =
bn

X
X

ut =
u
b0n en =
u
b0n en
t=0
t=0
t=0
X
Xp
n Bn en = =
bn en
=
bn
Bn =
n
and the problem is completely solved. Indeed, from solving the eigenfunction problem
(
u
= u

u = 0
we obtain the spectrum {n } and eigenfunctions {en }. The solution to (40.1) is then
(X )
X
p
p
u(X, t) =
An cos n t + Bn sin n t en (X)
!
X
p
p
bn
=
bn cos n t + sin n t en (X).

n
So
u(X, t) =
bn cos
!
b
p
n
n t + sin n t en (X) ,
n
X .
(40.2)
Done!
Remark 40.1. One has probably already observed that once we know the spectrum of
the eigenfunction expansion method works in the very same way in any dimension
(1, 2, 3 or even higher).
2. Heat Equation
Exercise 40.2. As it was done for the wave equation, solve
ut = u

u = 0
u(0, X) = (X) , X
and get a formula similar to (40.2).
3. FURTHER DISCUSSIONS
249
3. Further Discussions
We will explore two further topics: nodal lines and how to handle unbounded
domains.
1
Nodal lines
Consider the wave equation with Dirichlet and initial conditions for =
[0, ]2 :
utt = u
u = 0

.

ut=0 = (X)
u = (X)
t
t=0
By Theorem 38.1, we have n,m = n2 +m2 and by Example 39.4, normalized

eigenfunctions are
2
en,m (X) = sin nx sin my.
Now by (40.2), the solution is

!
X
p
p
bn,m
bn,m cos n,m t + p

u(X, t) =
sin n,m t en,m (X).
n,m
n,m
Nodal lines are lines where en,m (x, y) = 0, i.e. they are lines which remain
at rest while other parts vibrate. So assuming all coefficients equal 1 (or 0), we
have the following nodal lines on [0, ]2 :
simple harmonics
e1,1
e1,2
e2,1
e1,3
e3,1
e3,3
e1,2 + e2,1
e1,3 + e3,1
e1,4 + e4,1
double harmonics
250
All come from solving a trigonometric equation in two variables. Simple

harmonics are easy to understand. For example, for e1,3 , we need to solve:
sin x sin 3y = 0 for (x, y) [0, ]2 . So we obtain:
x = 0, and y [0, ] or x [0, ] and 3y = k , k Z , y [0, ]
2
x = 0, and y [0, ] or x [0, ] and y = 0, , , .
3 3
Hence the nodal lines are x = 0, x = , y = 0, y = (the boundary as
expected) plus y = 3 , y = 2
.
3
For double harmonics, well examine each graph produced:
e1,2 + e2,1 . We have
sin x sin 2y + sin 2x sin y = 0
2 sin x sin y cos y + 2 sin x cos x sin y = 0
sin x sin y(cos y + cos x) = 0.
Note that sin x = 0, sin y = 0 give us the boundary. Then since we have
cos( x) = cos x, the other solution in [0, ]2 is the line y = x (the
diagonal).
e1,3 + e3,1 . We will use the following, derived from trigonometric identities:
sin 3x = sin x(2 cos 2x + 1). Then we have:
sin x sin y(2 cos 2y + 1) + sin x(2 cos 2x + 1) sin y = 0
sin x sin y(2 cos 2y + 2 cos 2x + 2) = 0
sin x sin y(cos 2y + cos 2x + 1) = 0.
Again the first two factors give us the boundary, and for the third, note
that we solve: cos 2y = 1 cos 2x so in order to get a solution, we need
cos 2x 0 for x [0, ], i.e. x 4 , 3
. Then we must have y [0, ]
4
and
2y = arccos(1 cos 2x) or 2y = 2 arccos(1 cos 2x)
1
1
y = arccos(1 cos 2x) or y = arccos(1 cos 2x).
2
2
Note that the equation cos 2x + cos 2y = 1 is the most convenient to
solve to plot, but it takes several other forms: cos2 x + cos2 y = 12 or
cos(x + y) cos(x y) = 32 for example.
e1,4 + e4,1 . Here one can check that sin 4x = 4 sin x cos x cos 2x. Then
4 sin x sin y cos y cos 2y + 4 sin x cos x cos 2x sin y = 0
sin x sin y(cos y cos 2y + cos x cos 2x) = 0.
Well have sin x = 0, sin y = 0 give us the boundary, and noting the
symmetry in cos x cos 2x + cos y cos 2y = 0, one can see that y = x is
251
again a solution since then cos y = cos x while cos 2y = cos 2x. So we
can also factor cos x + cos y:
0 = cos x cos 2x + cos y cos 2y
= (cos x + cos y)(cos 2x + cos 2y) cos x cos 2y cos y cos 2x
= (cos x + cos y)(cos 2x + cos 2y) cos x(2 cos2 y 1) cos y(2 cos2 x 1)
= (cos x + cos y)(cos 2x + cos 2y 2 cos x cos y + 1)
= (cos x + cos y)(2 cos2 x 1 + 2 cos2 y 1 2 cos x cos y + 1)

1
2
2
= 2(cos x + cos y) cos y cos x cos y + cos x
.
2
But the second term is a quadratic equation in Y = cos y, X = cos x. In
order to plot the nodal line, we further solve the equation, noting that
using X as a parameter, the equation has a real solution for

1
2
2
= (X) 4 X
= 2 3X 2 0,
2
r !
r !
2
2
i.e. arccos
x arccos
.
3
3
Then on [0, ], we have for suitable x-values:

cos x 2 3 cos2 x
y = arccos
.
2
Remark 40.3. Note that for the eigenvalue 5 for example, there are 2 corresponding eigenfunctions e1,2 and e2,1 . And thus, = 5 has multiplicity two.
But then the number of integer solutions to = n2 + m2 , known as a general
quadratic Diophantine equation in 2 variables, depends on the prime number
decomposition of , and although this study belongs to Algebra, it is important
to note that as a result, the multiplicity of is unbounded.
Remark also that the superposition of harmonics is not trivial if coefficients
are changed.
In addition, if we deal with a round membrane, we now have Bessel functions
instead of the double sine functions so wed get different pictures for the nodal
lines too.
The study of nodal lines have applications in construction for example, to
make sure certain parts wont crack under the vibrations (recall that nodal
lines on the membrane will stay put).
Actually, there is a whole field of study concerned with nodal lines and
eigenfunctions involving the Laplacian: in particular in seismology for the wave
equation on a sphere, called the Laplace-Beltrami equation (or spherical Laplacian); or in quantum mechanics since the Schrodinger equation iut = u + qu
is involved (not quite the heat equation though; i makes a big difference!). We
252
could also consider the Klein-Gordon or plasma equation

utt = u + qu
but now in 2D or 3D. Sadly at this point, very little is known about this
equation in 2D or 3D compared to what is known for one dimension.
Unbounded domains

For domains like = Rn , the condition u = 0 is automatically satisfied
since L2 functions must decay at . But there are many ways for a domain
to be unbounded, for example an infinite strip in R2 or an infinite rod in R3
have only one dimension carrying the unboundedness.
There is no general approach since we there are so many kinds of possible
spectrum. What we will have is that is positive definite, i.e. hu, ui 0
for all u. Hence we can only have nonnegative eigenvalues.
We will consider the specific case of Rn . We need to redefine the Fourier
transform in higher dimensions so that it still works. Recall in one dimension,
we have:
1
fb() =
eix f (x)dx.
2 R
How can we move to a vector form? We have: f (x) f (X) and dx dX.
So we introduce Rn , called a wave vector and
1
b
f () =
f (X)eih,Xi dX
direct Fourier Transform,
(2)n/2 Rn
1
fb()eih,Xi d
inverse Fourier Transform.
f (X) =
(2)n/2 Rn
n
Note
the inner product here is the dot product in R . We also verify
that

that fb() = kf (X)k, we have an inverse, and the uniqueness property is
also satisfied. Furthermore,
e (X) = eih,Xi
is an eigenfunction of the continuous spectrum with
e (X) = kk2 e (X).
This can be easily checked in R2 (and generalization to Rn will be obvious):
e (X) = ei(1 x1 +2 x2 )
=
2 i(1 x1 +2 x2 )
2 i(1 x1 +2 x2 )
e
+
e
x21
x22
= (i21 )ei(1 x1 +2 x2 ) + (i2 )2 ei(1 x1 +2 x2 )

= (21 + 22 )e (X).
So to solve utt = u, one passes to the Fourier transform form, then we
are left with an ordinary linear order 2 equation which once solved is passed
through the inverse Fourier transform.
Now
253
1
1
eih,Xi =
eih,Xi
n/2
(2)
(2)n/2
where is a unit directional vector; and we now have infinitely many directions
for each eigenvalue (instead of just before), and the multiplicity of each
eigenvalue is therefore infinite.
Part 5
Greens Function
LECTURE 41
Introduction to Integral Operators

So far, weve studied a lot of differential operators, but dealt with really only one
integral operator: the Fourier transform. But in a way, differentiation and integration
are inverses of each other. I.e.
dF (x)
d
F (x) = f (x) F (x) = f (x)dx =
dx.
dx
dx
But we have non-uniqueness here since for each f (x) there exists an infinite number
of antiderivatives F (x) + C , C is a constant. If we add an initial condition, then C
can be fixed and we get uniqueness. That is
x
d
F (x) = f (x)
F (x) = C +
f (t)dt.
dx
F (x ) = C
x
0
0
x
The operator Au =
u(t)dt is the simplest integral operator.
x0
Integration is nicer than differentiation since the resulting function is smoother

than the original (because its at least differentiable). Think in electricity of the difference between an integral circuit (smoother) and a differential circuit (sharper). So
it may be best to rewrite a system with integrals.
Now the general definition
Definition 41.1. Let = (a, b) be a finite or infinite (a, b could be ) interval
and K(x, y) a function on (a, b) (a, b). The formal expression
b
(Kf )(x) =
K(x, y)f (y)dy =
K(x, y)f (y)dy , x = (a, b)
is called an integral operator. The function K(x, y) is called the kernel of the integral
operator K.
x
Exercise 41.2. Rewrite the operator of integration Au =
u(t)dt formally as an
x0
integral operator.
Theorem 41.3. An integral operator is linear. That is if K is an integral operator

then
K(f1 + f2 ) = Kf1 + Kf2 ,
K(f ) = Kf.
Proof. Trivial.
257
258
41. INTEGRAL OPERATORS
Example 41.4. Consider the Fourier operator F:

1
(Ff )() =
eix f (x)dx.
2
1
Its an integral operator with kernel K(, x) = F (, x) = eix .
2
Theorem 41.5. Let K1 , K2 be integral operators with interval and kernels K1 , K2 .
1) If K = c1 K1 + c2 K2 where c1 , c2 C then
K(x, y) = c1 K1 (x, y) + c2 K2 (x, y).
2) If K = K1 K2 then
K(x, y) =
K1 (x, z)K2 (z, y)dz.
Proof.
2)
1) trivial.
(Kf )(x) = (K1 K2 f )(x) = (K1 (K2 f ))(x)
=
K1 (x, z) (K2 f )(z)dz

K2 (z, y)f (y)dy dz

K1 (x, z)
=

K1 (x, z)K2 (z, y)dz f (y)dy
=
|
{z
}
=K(x,y)
where we switched the order of integration for the last line.
QED
Theorem 41.6. Let K be an integral operator on L2 (), where is an interval on

R. I.e.
(Kf )(x) =
K(x, y)f (y)dy , f L2 ()
where K(x, y) is the kernel of K. Then the kernel K (x, y) of the adjoint operator K
is given by
K (x, y) = K(y, x).
Exercise 41.7. Prove Theorem 41.6
Corollary 41.8. For a self adjoint integral operator, we have
K = K
K(x, y) = K(y, x)
Remark 41.9. Theorems 41.5 & 41.6 show us that one can understand an integral
operator as analogous to a continuous matrix. Indeed, addition is a linear combination
of the corresponding components. The product of integral operator looks like a matrix
product, where we take dz = 1 and i, j become the continuous variables x, y (like we did
when considering Fourier transform vs Fourier series). Similarly, the adjoint kernel is
a continuous adjoint matrix analog since we had (A )ij = (A)ji for discrete indices i, j.
259
Exercise 41.10. If F is the Fourier transform, find the kernel of its adjoint.
Let A be the integral operator on L1 (0, 1) defined by
x
f (y)dy.
(Af )(x) =
0
Exercise 41.11. Find the kernel and adjoint of A.

Theorem 41.12. A is a linear bounded operator on L1 (0, 1).
Proof. Linearity is obvious. Lets prove boundedness. We need to show that
kAf k C kf k
for any f with kf k 1.
Indeed the norm in L1 (0, 1) is
|f (x)| dx.
kf k1
0
For an arbitrary u L1 (0, 1) we have

1 x
1

u(t)dt dx
|Au| dx =
kAf k1 =

0
0
0
1 x
1 1
|u(t)| dt dx
|u(t)| dt dx
0
0
0
0
|
{z
}
=kuk1
dx = kuk1 .
= kuk1
0
That is
kAuk1 kuk1
and by definition A is bounded and
kAk 1.
QED
Note that it can be proven that kAk = 1 but we dont care at this point.
Remark also that the Fourier operator is bounded on L2 (R) since it is unitary:
kFf k2 = kf k2 .
Exercise 41.13. Show that the operator B defined on L2 (0, 1) by the formula
1 x
Bu =
u(t)dt
i 0
1
on functions u with a zero mean (that is
u(x)dx = 0) is selfadjoint.
0
x
d
(Hint: if v L1 (0, 1) then
v(t)dt = v(x).)
dx 0
260
Definition 41.14. An integral operator K is called Hilbert-Schmidt if

x
|K(x, y)|2 dxdy < .
2
Theorem 41.15. A Hilbert-Schmidt operator is bounded on L2 ().

Proof. Note
k(Kf )(x)k22
2

K(x, y)f (y)dy dx.
|Kf (x)| dx =

2
But by the Cauchy-Schwarz inequality, we have
2
2
k(Kf )(x)k2
dx
|K(x, y)| dy
|f (y)|2 dy
| {z
}
=kf k22
|K(x, y)|2 dxdy kf k22
|K(x, y)|2 dxdy is finite since K is Hilbert-Schmidt. So
where
|K(x, y)| dxdy
k(Kf )(x)k2
1/2
kf k2
and K is bounded.
QED
1/2

|K(x, y)|2 dxdy
is called the Hilbert-Schmidt norm of an
Note that
integral operator.
Example 41.16 (good example). Consider K(x, y) = e(x+y) , x, y (0, 1). Then
1
1
1 1
2x
2(x+y)
e2y dy
e dx
e
dxdy =
0
0
0
0
1 !2
1
2
2

1
1 e2
2x
2x
e dx = e
=
=
.
2
2
0
0
So K is Hilbert-Schmidt.
The converse of Theorem 41.15 is not true as illustrated in the example below.
Example 41.17 (bad example). Let F be the Fourier transform, which is bounded
on L2 (R). Is it Hilbert-Schmidt?

1 ix 2
ddx = 1
e
ddx =

2
2 R R
R R
so F is not Hilbert-Schmidt.
Usually, Hilbert-Schmidt does not like infinite intervals. It could be that it converges in x but not in y (or vice versa).
LECTURE 42
The Greens Function of u00 + p(x)u0 + q(x)u = f

1. The Greens Function
The most important physical operators are one of the following:
multiplication by a function,
differentiation,
integration.
So integral operators play an enormous role in physics. The main reason for this
is that, roughly speaking, the inverse operator of a differential operator is an integral
operator.
To show this we need
Definition 42.1. Let A be a differential operator on a Hilbert space H. Then the
solution G(x, y) to the problem
AG = (x y)
(42.1)
where x is a variable and y is a parameter is called the Greens function of the differential operator A.
Note that this means that the Greens function is the kernel of an important integral
operator. The operator A need not be doing only differentiation, e.g. we could consider
d2
the Schrodinger operator A = 2 + q(x).
dx
One may wonder why equation (42.1) has a solution. It is a very difficult question
which we are unable to answer in this course. But the answer is affirmative for most
of the differential operators of mathematical physics. But finding G needs to be done
on a case-by-case basis.
Why is the Greens function that important?
The reason is: if we know the Greens function G(x, y) of an operator A then we
can easily solve the equation
Au = f
(42.2)
for any f .
Indeed, let us show that
u(x) =
G(x, y)f (y)dy

261
(42.3)

42. THE GREENS FUNCTION OF THE SCHRODINGER
OPERATOR
262
is a solution to (42.2). Note that the limits of integration depend on the problem at
hand. We have
(Au)(x) = A G(x, y)f (y)dy =

AG(x, y) f (y)dy
| {z }
=(xy) by (42.1)
(x y)f (y)dy = f (x).
So (42.3) is a solution to (42.2). If (42.2) has a unique solution then (42.3) is the
solution.
Remark 42.2. One can view the Greens function as the continuum matrix of the
inverse of A. Indeed as a parallel to Linear Algebra, one would have:
n
X
1
u = A f = Gf
,
ui =
Gik fk
k=1
and AG = In = ij becomes AG = (x y).

Note also that the computations above are not formal; rigorous justifications are
difficult questions of the theory.
One can see also that x could become X in higher dimension spaces.
Exercise 42.3. Verify that
G(x, y) =
ei|xy|
2i
is the Greens function of the operator

A=
d2
2
2
dx
on L2 (R).
(Hint: verify (42.1))

2. Application to a General Order 2 Linear Differential Operator
Consider
(
u00 + p(x)u0 + q(x)u = f (x) ,
(BC) u(a) = 0 = u(b)
x (a, b)
(42.4)
i.e. we have Dirichlet boundary conditions on a finite interval.

Let us find the Greens function, i.e. the function G(x, y) satisfying
G00 + pG0 + qG = (x y).
We need to use variation of parameters for the nonhomogeneous equation (42.4)
(see also Appendix to Lecture 32).
First we need to know the solution to the homogeneous equation: u00 + pu0 + qu = 0.
It could be special functions: Bessel,... depending on p, q.
We will assume that we know the set of fundamental solutions {u1 (x), u2 (x)} to the
homogeneous equation, and we further choose u1 , u2 such that u1 (a) = 0 , u2 (b) = 0.
2. APPLICATION OF THE GREENS FUNCTION
263
Note that we cant have u1 (b) = 0 otherwise the Wronskian is 0 and u1 , u2 would
be linearly dependent. Similarly, u2 (a) 6= 0.
Now we look for a particular solution to (42.4) in the form
up (x) = C1 (x)u1 (x) + C2 (x)u2 (x).
For C1 , C2 we have

u1 u2
u01 u02
0
0
C1
=
.
f
C20
Solving this, one has

C10
f u2
=
W
C20
f u1
=
W

u1 u2
,
where W = det 0
u1 u02
the Wronskian of (u1 , u2 ). Note that even if p = 0 then W = constant 6= 0. So

x
x
f u2
f u1
C1 (x) =
(t)dt , C2 (x) =
(t)dt.
W
W
x0
x1
The general solution to (42.4) is
u(x) = C1 (x)u1 (x) + C2 (x)u2 (x) + C1 u1 (x) + C2 u2 (x)
where C1 , C2 are constants to be found using the BC on u, u1 , u2 . Since u2 (a) 6=
0 , u1 (b) 6= 0,
(
:0

:0

1
u(a) =
C1(a)u
(a) + C2 (a)u2 (a) +
C1u
1 (a) + C2 u2 (a) = 0
:0

0
:

2 (b) + C1 u1 (b) +
u(b) = C1 u1 (b) +
C2(b)u
C2u
2 (b) = 0
(
C2 = C2 (a)
C1 = C1 (b)
and

u(x) = C1 (x) C1 (b) u1 (x) + C2 (x) C2 (a) u2 (x).
But
C1 (x) C1 (b) =
x0
and
C2 (x) C2 (a) =
x1
f u2
W
f u1
W
b

(t)dt +
x0
(t)dt
x1
f u2
W
f u1
W
b

(t)dt =

(t)dt =
(42.5)
f u2
W
f u1
W

(t)dt ,

(t)dt.
Finally, for (42.5) we have
u(x) = u1 (x)
x
f (t)u2 (t)
dt + u2 (x)
W (t)
f (t)u1 (t)
dt.
W (t)
(42.6)

42. THE GREENS FUNCTION OF THE SCHRODINGER
OPERATOR
264
By definition of the Greens function, G(x, y) is computed by (42.6) for f (t) =

(t y) where y is an arbitrary point on (a, b). We have
x
b
(t y)u1 (t)
(t y)u2 (t)
dt + u2 (x)
dt.
G(x, y) = u1 (x)
W (t)
W (t)
a
x
Consider two cases (1) x > y, (2) x < y :
u2 (x)u1 (y)
(1) x > y then G(x, y) =
,
W (y)
u1 (x)u2 (y)
(2) x < y then G(x, y) =
.
W (y)
Alternately, we can read G(x, t) directly from (42.6):
b
b
x
u1 (x)u2 (t)
u2 (x)u1 (t)
f (t)dt +
f (t)dt =
G(x, t)f (t)dt,
u(x) =
W (t)
W (t)
x
a
a
where
1
G(x, t) =
W (t)
(
u2 (x)u1 (t) ,
u1 (x)u2 (t) ,
atx
.
x<tb
So we come up with the following theorem

Theorem 42.4. The Greens function G(x, y) of a general order 2 linear differential
d
d2
+
p(x)
+ q(x) on L2 (a, b) with Dirichlet boundary conditions can be
operator
dx2
dx
represented by
u1 (x)u2 (y)
,
x<y
W (y)
G(x, y) = u (x)u (y)
2
1
,
x>y
W (y)
where u1 , u2 are solutions of u00 + pu0 + qu = 0 with conditions u1 (a) = 0 , u2 (b) = 0,
respectively, and W is the Wronskian of (u1 , u2 ), that is
W = u1 u02 u01 u2 .
But we need a fundamental set... thats the hard part. If we have Neumann
conditions or others, we still follow the same procedure, but we figure out different BC
to impose on the fundamental set so that we can get a solution in the simplest form.
Note that the above theorem can easily be restated for the Schrodinger operator.
Exercise 42.5. Derive the expression for Greens function G(x, y) of the operator
d2
A = 2 + 2 on L2 (R) for > 0.
dx
e|xy|
Answer: G(x, y) =
.
2
(Hint: modify the arguments of this section to treat (, ).)
Note that q(x) could be some form of spectral parameter as above. Then we can
write: G (x, y).
2. APPLICATION OF THE GREENS FUNCTION
265
What about higher dimensions? For example the free Laplacian:

(
u 2 u = f
.
u L2 (Rn )
The solution was exponential in one dimension. Here variables appear in the denominator and there are infinitely many linearly independent solutions for the fundamental
set (not just 2).
Index
properties, 13
real part, 11
Complex valued function, 13
Conformal mapping, 225, 228
Joukowsky transform, 232
Mobius transform, 218
Riemann theorem, 225, 231
Schwarz-Christoffel formula, 232
Connected
multiconnected contour, 33
path-connected, 32
simply connected domain, 225
Continuous spectrum, 107, 110, 173, 174, 176,
180, 209
Contour
multiconnected, 33
Contour integral, 20
Convergence
absolute, 30
normed spaces, 89
of a sequence, 29
series, 90
uniform, 112
weak, 116
Convolution, 151, 193, 229
Convolution theorem, 152
Coordinates, 90
Coordinates of a vector, 58
Coulomb potential, 214
Adjoint operator, 70, 258

Amplifier, 97
Analytic
at infinity, 38
Analytic function, 217
definition, 15
power series, 31
Banach space, 89
Basis, 56
infinite dimension, 83
orthogonal, 136
orthonormal, 68, 110
Bessel equation., 239
Bessel function, 210, 251
Boundary conditions, 105, 169
Bounded linear operator, 102, 103
Branch cut, 50
Cauchy formula, 24, 25, 33, 218
Cauchy inequality, 84
Cauchy integral, 24
Cauchy Principal Value, 49
Cauchy sequence, 89
Cauchys theorem, 20, 162, 191
Cauchy-Riemann conditions, 16, 217, 227
Cauchy-Schwarz inequality, 260
Change of variables, 233
Characteristic equation, 77
Characteristic polynomial, 77
Characteristic triangle, 202
Complex numbers
argument, 12
conjugate, 11
properties, 12
exponential representation, 13
imaginary part, 11
imaginary unit, 11
inverse, 12
modulus, 11
DAlembert formula, 182, 183

-function, see also Dirac -function
-sequence, 113, 116, 117, 220, 230
Derivative, 15
Differential operator, 59, 160
Dimension
infinite, 83
Dirac -function, 117
Fourier series representation, 123
integral representation, 125
267
268
Directional derivative, 215

Dirichlet
Laplacian, 215, 241
Dirichlet problem, 127, 221, 222, 241, 247
Discrete spectrum, 107, 109, 173, 174
Disk
unit, 218
Domain
simply connected, 225
Double sine Fourier series, 243
Duhamel principle, 193
Eigenfunction, 107, 109, 110, 241
expansion, 181
Eigenfunction expansion, 208, 247
Eigenfunctions, 135, 180
generalized, 176
Eigenspace, 79
Eigenvalue, 77, 107, 109, 173, 241
multiplicity, 78, 251
simple, 78
Eigenvector, 77, 109
Elliptic integral, 232
Euclidean space, 69, 84
Euclidian norm, 84
Euler formula, 13
Filter, 97
Fourier coefficient
generalized, 208, 244, 247
Fourier integral theorem, 155, 163, 209
Fourier operator, 151, 154, 177, 258, 259
Fourier representation, 159, 160
Fourier series, 244
complex form, 91
double sine, 243
expansion, 91
trigonometric form, 93
Fourier transform, 104, 148, 149, 151, 159, 258
generalized, 210
in Rn , 252
Frobenius method, 133
Frobenius theorem, 139
Gauss integral, 192
Generalized derivative, 119
Generalized function, 117
Gordon-Klein equation, 172
Greens first identity, 221
Greens formula, 215, 217
Greens function, 261, 264
Harmonic function, 227
INDEX
Harmonics
rectangular, 223
Heat equation, 193, 207, 241
Helmholtz equation, 207, 241
Hermite polynomials, 145
Hermites equation, 144
Hilbert space, 85, 212, 261
examples
L2 (a, b), 87
orthonormal basis, 90
Hilbert-Schmidt integral operator, 260
Hilbert-Schmidt norm, 260
Holomorphic function, 15
Homogeneous wave equation, 168
Indicial equation, 140142
Initial conditions, 169
Inner product, 67, 85, 212
in R3 , 69
Inner product space, see also Euclidean space
Integral operator, 257, 261
Hilbert-Schmidt, 260
Inverse
of an operator, 106
Inverse operator, 71
Invertible operator, 71
Irregular singular point, 139
Jacobi matrix, 234
Jordans lemma, 46, 161, 163
Kernel
integral operator, 257
Klein-Gordon equation, 207, 252
Kronecker delta, 68, 181
L1 space, 104
L2 space, 87
Laplace equation, 215, 217, 221, 232
nonhomogeneous, 241
Laplace operator, 211, 214, 241, 247
Laplace-Beltrami equation, 252
Laplacian, 265
Dirichlet, 215, 241, 247
Neumann, 215, 241
Robin, 215
Laurent theorem, 34
Legendre equation, 133
Legendre polynomials, 133, 135, 209
Line integral, 18
Linear operator, 59, 173
bounded, 102
differentiation, 59, 105
INDEX
domain, 101
inverse, 106
kernel, 64
momentum, 110
norm, 102
resolvent, 106
spectrum, 107
unbounded, 102
Linear space, 55
basis, 56, 90
complex, 56
convergence of series, 90
dimension, 56
infinite dimensional, 83
norm, 83
real, 56
Linearly dependent, 56
Linearly independent, 56
Liouville theorem, 25
M
obius transform, 218
Matrix, 59
addition, 60
diagonal matrix, 61
multiplication, 60
orthogonal, 72
scalar multiplication, 60
square, 59
unit matrix, 61
zero matrix, 61
Matrix representation of an operator, 62
Mean value theorem, 218
Momentum operator, 73, 105, 110, 160, 175,
177, 179, 189
Moreras theorem, 27
Multiplicity, 78
Neumann
Laplacian, 215, 241
Neumann problem, 127, 222, 241
Nodal lines, 249
Norm, 83
induced by an inner product, 68
of a linear operator, 102
sup-norm, 111
Normalized, 88
Normed space, 83
Operator
adjoint, 70, 258
coordinate, 177
coordinates, 160
269
differential, 261
differentiation, 160
Fourier, 177, 258
integral, 257, 261
inverse, 71
invertible, 71
Laplace, 211, 214, 241, 247
momentum, 73, 105, 160, 175, 177, 179, 189
multiplication, 160, 177
rotation, 73
Schrodinger, 179, 189, 261, 264
selfadjoint, 70, 133
similar, 75, 159
unitary, 71
unitary equivalent, 176, 177
Ordinary point, 137
Orthogonal basis, 136
Orthogonal functions, 88
Orthogonal matrix, 72
Orthogonal vectors, 68
Orthonormal basis, 68, 110, 241
Parseval equation, 68, 91
Path
connected, 32
Plancherel theorem, 154
Point
irregular singular, 139
ordinary, 137
regular singular, 139
singular, 138
Point spectrum, 173
Poisson formula, 219, 227
Poisson kernel, 219
upper half plane, 229
Pole, 37
Potential, 207
Coulomb, 214
Power series, 30, 133
Power series solution, 137
Purely discrete spectrum, 110
Rectangular harmonics, 223
Regular singular point, 139
Residue, 38, 39
Residue formula, 39
Residue theorem, 40
Resolvent, 106, 173
Robin
Laplacian, 215
Robin problem, 128
Rodrigues formula, 136
270
INDEX
Rotation operator, 73
Triangle inequality, 19
Scalar product, see also Inner product

Schr
odinger operator, 129, 169, 179, 189, 207,
261, 264
Schr
odinger equation, 143
Schwarz-Christoffel formula, 232
Selfadjoint operator, 70, 133, 215
spectrum, 175
Series
convergence, 29
definition, 29
divergence, 29
functional, 30
domain of convergence, 30
power, 30
Similar operators, 75, 159, 176
Similarity transformation, 75
Simple harmonics, 94, 95, 110
Simply connected domain, 20
Singular point, 138
Singularity, 37
essential, 37
isolated, 37
pole, 37
removable, 37
Smooth, 114
Spectral analysis, 78, 169, 170, 177, 178, 208
Spectral theorem, 80
Spectrum, 77, 107, 135, 174, 176
continuous, 107, 110, 173, 174, 176, 180, 209
discrete, 107, 109, 173, 174, 240
point, 173
purely discrete, 110
selfadjoint operator, 175
Sturm-Liouville
Dirichlet problem, 127
equation, 127, 208
Neumann problem, 127
operator, 127
canonical form, 129
Robin problem, 128
Sturm-Liouville equation, 238
Sturm-Liouville operator, 133
Subspace, 58
Support, 113
finitely supported, 113
Uniform convergence, 112

Unit disk, 218
Unitary equivalent, 176, 177
Unitary operator, 71
Univalent function, 225
Taylor series, 31, 32

Taylor theorem, 31, 220
Tension, 168
Test function, 114
Variation of parameters, 163, 195, 197, 203,

208, 210, 262
Vector space, see also Linear space
Wave equation, 168, 193, 207, 241, 247
free, 169
Gordon-Klein, 172
Helmholtz, 207
homogeneous, 168
Klein-Gordon, 207
nonhomogeneous, 172
Wave propagation, 183
Wave vector, 252
Weak convergence, 116
Weak solution, 176, 177
Weighted L2 -space, 128
Weyl sequence, 174
Zhukovsky function, 232

Math Physics Lecture Notes

Uploaded by

Math Physics Lecture Notes

Uploaded by

MATH 611/612: Mathematical Physics

Lecture 1. Complex Numbers. Functions of Complex Variables

Lecture 2. The Concept of Analytic Function. The Cauchy-Riemann Conditions 15

Lecture 4. Complex Series

Lecture 5. Singularities of Analytic Functions and the Residue Theorem

Lecture 6. Applications of the Residue Theorem to Evaluation of Some Definite

3. Integrals of the type

Lecture 7. Applications of the Residue Theorem to Some Improper Integrals and

x1 f (x)dx , 0 < < 1

2. Integrals of the type

Lecture 8. Vector Spaces

Lecture 9. Linear Operators

Lecture 10. Inner Product. Selfadjoint and Unitary Operators

Lecture 12. Eigenvalues and Eigenvectors

Lecture 13. Infinite dimensional spaces. Hilbert Spaces

Lecture 14. L2 spaces. Convergence in Normed Spaces. Basis and Coordinates

Lecture 15. Fourier Series

Lecture 16. Some Applications of Fourier Series

Lecture 17. Linear Operators in Hilbert Spaces

Lecture 18. Spectra of Selfadjoint Operators

Lecture 19. Generalized Functions. Dirac function.

Lecture 20. Representations of the -function

Lecture 21. The Sturm-Liouville Problem

Lecture 22. Legendre Polynomials

Lecture 23. Harmonic Oscillator

Lecture 24. The Fourier Transform

Lecture 25. Properties of the Fourier Transform. The Fourier Operator

Lecture 26. The Fourier Integral Theorem (Fourier Representation Theorem)

Lecture 27. Some Applications of the Fourier Transform

Partial Differential Equations

Lecture 28. Wave Equations

1. The Stretched String

Lecture 30. The Method of Eigenfunction Expansion in the Case of Continuous

Lecture 33. Wave and Heat Equations in Nonhomogeneous Media

Lecture 34. L2 -spaces on Domains in Multidimensional Spaces

Lecture 35. The Laplace Equation in R2 . Harmonic Functions

Lecture 39. Nonhomogeneous Laplace Equation

Lecture 41. Introduction to Integral Operators

Lecture 42. The Greens Function of u00 + p(x)u0 + q(x)u = f

Complex Numbers. Functions of Complex Variables

where x is the real part of z also denoted as

1. COMPLEX NUMBERS. FUNCTIONS OF COMPLEX VARIABLES

The product z of z1 , z2 is defined as

As an exercise check to see

It follows from elementary trigonometry that

z1 z2 = |z1 ||z2 | (cos(1 + 2 ) + i sin(1 + 2 ))

1. COMPLEX NUMBERS. FUNCTIONS OF COMPLEX VARIABLES

The Concept of Analytic Function. The Cauchy-Riemann

So f is differentiable at z0 and analytic in C.

2. THE CONCEPT OF ANALYTIC FUNCTION. THE CAUCHY-RIEMANN CONDITIONS

case 2: z = iy: z0 + z approaches z0 along a vertical line.

Similarly, in the second case

Comparing (2.2) and (2.3) one gets

Corollary 2.5. f = u + iv H(D)

Conditions (2.4) are called Cauchy-Riemann.

Example 2.8 (ez ). Lets define ez as follows:

Exercise 2.9. Show that ez defined by (2.5) is analytic z C.

2. THE CONCEPT OF ANALYTIC FUNCTION. THE CAUCHY-RIEMANN CONDITIONS

Prove that the Cauchy-Riemann conditions are equivalent to

f (z)dz = (u + iv)(dx + idy)

= udx vdy + i vdx + udy.

Recall that line integrals are associated with work in physics.

1. (f1 (z) + f2 (z))dz = f1 (z)dz + f2 (z)dz

3. THE CAUCHY THEOREM

Use the hint of Exercise 2.12.

State and prove the change of variables formula.

The following theorem will play a crucial role in our exposition.

2. THE CONCEPT OF ANALYTIC FUNCTION. THE CAUCHY-RIEMANN CONDITIONS

and form the Riemann sum

f (xk )(zk zk1 )

f (z)dz is called the

contour integral of f (z) along C counterclockwise.

Unless otherwise stated, we write

Definition 2.19. A domain D C is said to be simply connected if, roughly

Below are two examples of contours C in a simply connected domain D.

3. THE CAUCHY THEOREM