7.

The Eigenvalue Problem, December 17, 2009 1

O|x = |x (1)
where is an unknown number and |x is an unknown ket. This equation has
many solutions |x

, which are called eigenstates, or eigenkets, or eigenvectors

, there is a number

, called the eigenvalue

.
I remind you that if

A represents an observable then the eigenstates of

A
are the pure states of that observable. In addition, the spectrum of A (the
values that the observable A can take) consists of the eigenvalues of

A.
Solving the eigenvalue equation for the operators representing observables
is a matter of greatest importance to quantum mechanics. In most cases
encountered in practice, you will know the operator and will want to calculate
some of the eigenvalues and eigenkets.
If one works with the Schrodinger representation (which is the coordinate
representation), the operators of interest are dierential operators acting on
wave functions that are elements of L
2
. For example, the energy operator for
a harmonic oscillator is

H =
h
2
2
d
2
dx
2
kx
2
(2)
and the eigenvalue equation is

H(x)
h
2
2
d
2
(x)
dx
2
kx
2
(x) = E(x) (3)
The eigenvalue equation for the energy of the harmonic oscillator has an
analytic solution (see Metiu, Quantum Mechanics) and so do those equations
for the particle in a box and for the hydrogen atom. This situation is not
general.
One of the most frequently used methods for solving such equations nu-
merically is to turn them into a matrix equation. This is why we dedicate a
chapter to studying the eigenvalue problem for matrices. In practice, we use
computers and software provided by experts to solve the eigenvalue problem
for a given matrix. Because of this, we focus here on understanding various
7. The Eigenvalue Problem, December 17, 2009 2
aspects of the eigenvalue problem for matrices and do not discuss numer-
ical methods for solving it. When we need a numerical solution, we use
Mathematica. For demanding calculations, we must use Fortran, C, or C++,
since Mathematica is less ecient. These computer languages have libraries
of programs designed to give you the eigenvalues and the eigenvectors when
you give them the matrix.
2 Some general considerations. It turns out that the eigenvalue problem
A| = | has many solutions and not all of them are physically mean-
ingful. We are interested in this eigenvalue problem because, if it arises
from an observable, the eigenvalues are the spectrum of the observable and
some of the eigenfunctions are the pure states of the observable. In addition,
the operator representing the observable is Hermitian, because its spectrum,
which consists of all the values the observable can take, just contain only real
numbers.
Before we try to solve the eigenvalue problem, we need to decide how
we recognize which eigenstates correspond to pure states: these are the only
eigenstates of interest to us.
In Chapter 2, we examined the denition of the probability and concluded
that kets |a
i
and |a
j
representing pure states must satisfy the orthonormal-
ization condition
a
i
| a
j
=
ij
(4)
3 Normalization. Let us assume that we are interested in solving the
eigenvalue equation

A| = | (5)
It is easy to see that if is a complex number and | satises Eq. 5 then
| = | satises the same equation. For every given eigenvalue we have
as many eigenstates as complex numbers. The one that is a pure state must
be normalized. We can therefore determine the value of by requiring that
| satises
| = 1 (6)
Using Eq. 5 in Eq. 6 (and the properties of the scalar product) leads to

| = 1 (7)
or

=
1
|
(8)
7. The Eigenvalue Problem, December 17, 2009 3
This gives
=
e
i
_
|
(9)
where is a real number, which cannot be determined by solving Eq. 8. The
ket
| =
e
i
_
|
| (10)
is normalized and therefore is a pure state of

A corresponding to the eigen-
value . If the system is in the state | and we measure A, we are certain
that the result is .
In practice, most computer programs that calculate eigenstates do not
normalize them. If the computer produces | as the eigenket corresponding
to eigenvalue , you must use Eq. 10 to nd the eigenstate | that is a pure
state. Since e
i
in Eq. 10 is an irrelevant phase factor, we can drop it (i.e. take
= 0); the pure state corresponding to is
| =
|
_
|
(11)
4 Orthogonality. The pure states must also be orthogonal. Because an
operator

A, representing an observable A, is always Hermitian, one can prove
that eigenstates corresponding to dierent eigenvalues must be orthogonal.
That is, if

A|
i
=
i
|
i
and

A|
j
=
j
|
j
(12)
and

i
=
j
(13)
then

i
|
j
= 0 (14)
The proof is easy. We have

A
i
|
j

i
|

A
j
= 0 (15)
because

A is Hermitian. Using Eqs. 12 in Eq. 15 gives
(
i

j
)
i
|
j
= 0
7. The Eigenvalue Problem, December 17, 2009 4
When
i
=
j
, we must have

i
|
j
= 0,
which is what we wanted to prove.
So, when you solve an eigenvalue equation for an observable, the eigen-
states corresponding to dierent eigenvalues must be orthogonal. If they are
not, there is something wrong with the program calculating the eigenvectors.
What happens if some eigenvalues are equal to each other? Physicists
call such eigenvalues degenerate, mathematicians call them multiple. If |
i

and |
j
belong to the same eigenvalue (i.e. if

A|
i
= |
i
and

A|
j
= |
j
)
then |
i
and |
j
are under no obligation to be orthogonal to each other (only
to, according to the fact in the previous , all eigenvectors belonging to eigen-
values that dier from ). Most numerical procedures for solving eigenvalue
problems will provide non-orthogonal degenerate eigenvectors. These can be
orthogonalized with the Gram-Schmidt procedure and then normalized. The
resulting eigenvectors are now pure states. You will see how this is done in
an example given later in this chapter.
It is important to keep in mind this distinction between eigenstates and
pure states. All pure states are eigenstates but not all eigenstates are pure
states.
5 The eigenvalue problem in matrix representation: a review. Let |a
1
,
|a
2
, . . . , |a
n
, . . . be a complete orthonormal basis set. This set need not be
a set of pure states of an observable; we can use any complete basis set that
a mathematician can construct. Orthonormality means that
a
i
| a
j
=
ij
for all i, j (16)
Because the set is complete, we have

I
N

n=1
|a
n
a
n
| (17)
Eq. 17 ignores the continuous spectrum and truncates the innite sum in the
completeness relation.
We can act with each a
m
|, m = 1, 2, . . . , N, on the eigenvalue equation
to turn it into the following N equations:
a
m
|

A| = a
m
| , m = 1, 2, . . . , N (18)
7. The Eigenvalue Problem, December 17, 2009 5
Now insert

I, as given by Eq. 17, between

A and | to obtain
N

n=1
a
m
|

A| a
n
a
n
| = a
m
| , m = 1, 2, . . . , N (19)
The complex numbers a
n
| are the coordinates of | in the {|a
n
}
N
n=1
representation. If we know them, we can write | as
| =
N

n=1
|a
n
a
n
| (20)
It is easy to rewrite Eq. 19 in a form familiar from linear matrix algebra.
Let us denote

i
a
i
| , i = 1, 2, . . . , N (21)
and
A
mn
a
m
|

A| a
n
(22)
With this notation, Eq. 19 becomes
N

n=1
A
mn

n
=
m
, m = 1, 2, . . . , N (23)
and Eq. 20,
| =
N

n=1
|a
n

n
(24)
The sum in Eq. 23 is the rule by which the matrix A, having the elements
A
mn
, acts on the vector , having the coordinates
n
. This equation is called
the eigenvalue problem for the matrix A and it is often written as
A = (25)
(which resembles the operator equation

A| = |) or as

A
11
A
12
A
1N
A
21
A
22
A
2N
.
.
.
.
.
.
.
.
.
.
.
.
A
N1
A
N2
A
NN

2
.
.
.

2
.
.
.

(26)
Eqs. 25 and 26 are dierent ways of writing Eq. 23, which, in turn, is the
representation of the equation

A| = | in the basis set {|a
1
, . . . , |a
N
}.
7. The Eigenvalue Problem, December 17, 2009 6
Normally in quantum mechanics we choose the basis set {|a
1
, . . . , |a
n
}
and know the operator

A. This means that we can calculate A
mn
= a
m
|

A| a
n
.
We do not know | or , and our mission is to nd them from

A| = |.
We have converted this operator equation, by using the orthonormal basis
set {|a
1
, . . . , |a
N
}, into the matrix equation Eq. 23 (or Eq. 25 or Eq. 26).
In Eq. 23, we know A
mn
but we do not know = {
1
,
2
, . . . ,
N
} or .
We must calculate them from Eq. 23. Once we know
1
,
2
, . . . ,
N
, we can
calculate | from
| =
N

n=1
|a
n
a
n
|
N

n=1

n
|a
n
(27)
6 Use a computer. Calculating eigenvalues and eigenvectors of a matrix
by hand is extremely tedious, especially since the dimension N is often very
large. Most computer languages (including Fortran, C, C++, Mathematica,
Maple, Mathcad, Basic) have libraries of functions that given a matrix will
return its eigenvalues and eigenvectors.
I will not discuss the numerical algorithms used for nding eigenvalues.
Quantum mechanics could leave that to experts without suering much harm.
The Mathematica le Linear algebra for quantum mechanics.nb shows
how to use Mathematica to perform calculations with vectors and matrices.
Nevertheless, it is important to understand some of the theory related
to the eigenvalue problem, since much of quantum mechanics relies on it. I
present some of it in what follows.
7 The eigenvalue problem and systems of linear equations. The eigenvalue
problem Eq. 23 can be written as
N

n=1
(A
mn

mn
)
n
= 0, m = 1, 2, . . . , N (28)
or

A
11
A
12
A
13
A
1N
A
21
A
22
A
23
A
2N
A
31
A
32
A
33
A
3N
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
A
N1
A
N2
A
N3
A
NN

3
.
.
.

0
0
0
.
.
.
0

d
= 0
(29)
7. The Eigenvalue Problem, December 17, 2009 7
The column in the right-hand side is the zero vector. We can regard
mn
as being the elements of a matrix I, which is called the unit matrix (or the
identity matrix):
I =

1 0 0 0
0 1 0 0
0 0 1 0
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
0 0 0 1

(30)
This has the property
(Iv)
m
=
N

n=1
I
mn
v
n
=
N

n=1

mn
v
n
= v
m
The m-th element of the vector Iv is v
m
; this means that
Iv = v (31)
The denition of I allows us to write Eq. 29 as
(A I) = 0 (32)
A system of equations of the form
Mv = 0, (33)
where M is an N N matrix and v is an N-dimensional vector, is called
a homogeneous system of N linear equations. The name homogeneous is
given because the right-hand side is zero; this distinguishes it from
Mv = u (34)
which is an inhomogeneous system of N linear equations.
Example. Let
A =

3 2 4
2 1.2 3.1
4 3.1 4

(35)
7. The Eigenvalue Problem, December 17, 2009 8
This is a Hermitian matrix with real elements. Let = {
1
,
2
,
3
} be a
vector. Then A = 0 means (use (A)
i
=

3
j=1
A
ij

j
)
A =

A
11

1
+ A
12

2
+ A
13

3
A
21

1
+ A
22

2
+ A
23

3
A
31

1
+ A
32

2
+ A
33

3
1
+ 2
2
+ 4
3
2
1
+ 1.2
2
+ 3.1
3
4
1
+ 3.1
2
+ 4
3

= 0 (36)
If a vector is equal to zero then all its components are zero, and Eq. 36 means
3
1
+ 2
2
+ 4
3
= 0 (37)
2
1
+ 1.2
2
+ 3.1
3
= 0 (38)
4
1
+ 3.1
2
+ 4
3
= 0 (39)
The notation A = 0 is a shorthand for these three equations. This is a
homogeneous system of three linear equations.
Similarly if = {2.1, 6.2, 4.3} then
A = (40)
is shorthand for
3
1
+ 2
2
+ 4
3
= 2.1 (41)
2
1
+ 1.2
2
+ 3.1
3
= 6.2 (42)
4
1
+ 3.1
2
+ 4
3
= 4.3 (43)
This is an inhomogeneous system of three linear equations. If the matrix A
has an inverse A
1
, we can solve Eq. 40 by using
A
1
A = A
1
(44)
and
A
1
A = I = (45)
which together lead to
= A
1
(46)
You have learned in your mathematics courses about Cramers rule, which
allows you to solve Eq. 40. This is tedious to implement and I prefer to
7. The Eigenvalue Problem, December 17, 2009 9
use Mathematica to give me the inverse matrix (see the le WorkBook7 The
eigenvalue problem.nb). Mathematica tells me that
A
1
=

160.333 146.667 46.6667

(47)
Acting with A
1
on gives (see WorkBook7.nb)
= A
1
= {773.3, 705.0, 228.0} (48)
End of Example
Exercise 1 Prove that if a matrix H is Hermitian then so is its inverse.
Note a very important fact. If A
1
exists then A = has a solution.
This also means that the solution of A = 0 is = 0 (because A
1
A =
A
1
0 = 0 and A
1
A = I = ). The solution = 0 is of no interest,
and we say that if A
1
exists then A = 0 does not have a solution.
How does this relate to the eigenvalue problem? The eigenvalue equation
written in the form
(A I) = 0 (49)
is a homogenous system of equations. If the operator
(

A

I)
1
(50)
exists then (AI) = 0 does not have a solution.
Therefore (A I) = 0 has a solution if and only if (A I)
1
does
not exist, which happens exactly when
det(A I) = 0 (51)
Note that this condition does not involve the unknown vector ; it depends
only on (which we dont know) and on the matrix elements A
ij
(which we
do know). This is an equation for . One can prove, based on the properties
of determinants, that the left-hand side of Eq. 51 is a polynomial of order
N in , which is called the characteristic polynomial of the matrix A. The
fundamental theorem of algebra tells us that Eq. 51 must therefore have
7. The Eigenvalue Problem, December 17, 2009 10
N solutions. Because they are eigenvalues of A and A is Hermitian, these
solutions must all be real numbers.
I will assume here that you have heard of determinants. If not, take heart:
computers can calculate them for you. Mathematica has the function Det[A]
that returns the determinant of the matrix A.
8 Back to physics. If A is an observable then the eigenvalues of the
matrix A are the values that we observe when we measure the observable A.
The quantization of the values that the observable can take occurs because
(AI) = 0 has solutions only for those special values of for which the
matrix AI does not have an inverse. These are the values of for which
det[A I] = 0. Note that the spectrum of A depends only on the matrix
A, as it should.
If we use N functions in our orthonormal, complete basis set, the matrix
A can have only N eigenvalues. The are not necessarily distinct, because
some roots of the characteristic polynomial might be equal to each other. As
we increase the number of functions in the orthonormal basis set, we increase
the number of eigenvalues. If we use an innite basis set, we will have an
innite number of eigenvalues. These eigenvalues will be discrete, however.
We cannot possible use an innite basis set and therefore we never get
the exact eigenvalues and eigenfunctions. By luck, or perhaps because of
some mathematical theorem unknown to me, if we are intelligent in our
choice of basis set, we get a good approximation to the smallest eigenvalues
even with a relatively small basis set. As youll see in an example given later,
using roughly 40 basis functions gives accurate values for about the 20 lowest
eigenvalues.
9 An example of an eigenvalue problem: nondegenerate eigenvalues. Let
us examine the eigenvalue problem for the matrix A given by Eq. 35. The
characteristic equation is (see Eq. 51)
det

3 2 4
2 1.2 3.1
4 3.1 4

= 0 (52)
If you nd tedious algebra soothing, you can calculate this determinant.
I prefer to use Mathematica, and the result is (see Section 2 of the le
WorkBook7 The eigenvalue problem.nb)

3
+ 8.2
2
+ 9.21 0.03 = 0 (53)
7. The Eigenvalue Problem, December 17, 2009 11
This is the characteristic polynomial of the matrix A. It is a third-order
polynomial because we used a three-dimensional basis set and therefore A is
a 3 3 matrix.
The roots of this equation are the eigenvalues of A. They are (see Section
2 of WorkBook7)

1
= 1.00391

2
= 0.0032479

3
= 9.20007

(54)
The eigenvalues are real numbers because A is Hermitian.
Exercise 2 Use Mathematica to check that det A =
1

3
.
How do we nd the eigenvectors? For the eigenvector (1) corresponding
to the eigenvalue
1
, insert
1
in the equation A = to obtain
A(1)
1
(1) = 0, (55)
We have modied the notation so that (1) = {
1
(1),
2
(1),
3
(1)} is the
eigenvector corresponding to the eigenvalue
1
. If I insert
2
in A = , I
get a dierent equation
A(2)
2
(2) = 0 (56)
for a dierent eigenvector, (2), corresponding to the eigenvalue
2
. Simi-
larly,
A(3)
3
(3) = 0 (57)
gives the third eigenvector (3). Because I have three values for , I have
three distinct equations for , and their solutions are (1), (2), and (3).
Each symbol (1), (2), (3) represents a vector in R
3
. The components
of (), = 1, 2, 3, are denoted by
() = {
1
(),
2
(),
3
()} (58)
With this notation,
(A
1
I)(1) =

3
1
2 4
2 1.2
1
3.1
4 3.1 4
1

1
(1)

2
(1)

3
(1)

= 0
7. The Eigenvalue Problem, December 17, 2009 12
This is the system of equations
(3
1
)
1
(1) + 2
2
(1) + 4
3
(1) = 0 (59)
2
1
(1) + (1.2
1
)
2
(1) + 3.1
3
(1) = 0 (60)
4
1
(1) + 3.1
2
(1) + (4
1
)
3
(1) = 0 (61)
We seem to have a diculty here. If {
1
(1),
2
(1),
3
(1)} is a solution
then {
1
(1),
2
(1),
3
(1)} is also a solution. This is not a real problem,
since, as noted earlier, we can x by requiring that the eigenvector be
normalized. This is the only meaningful eigenvector in quantum mechanics.
I can look at Eqs. 5961 and see right away that = {0, 0, 0} is a
solution. This is not physically acceptable, because it cannot represent a
state of the system. Are there other solutions besides {0, 0, 0}? Since this
is a homogeneous system of equations, the answer is yes: we have chosen
=
1
, which means that det[A I] = 0 and the system has a solution.
If we divide Eqs. 5961 by
3
(1), we obtain
(3
1
)

1
(1)

3
(1)
+ 2

2
(1)

3
(1)
= 4
2

1
(1)

3
(1)
+ (1.2
1
)

2
(1)

3
(1)
= 3.1
4

1
(1)

3
(1)
+ 3.1

2
(1)

3
(1)
=
1
4
The homogeneous system of three equations and three unknowns has become
an inhomogeneous system of three equations and two unknowns (
1
(1)/
3
(1)
and
2
(1)/
3
(1)).
Mathematica gives (see WorkBook7, Section 2, Cell 2)

1
(1) = 0.542
3
(1) (62)

2
(1) = 0.914
3
(1) (63)
so the eigenvector (1) is
(1) = {0.542
3
(1), 0.914
3
(1),
3
(1)}
=
3
(1){0.542, 0.914, 1} (64)
Of course,
3
(1) is not yet determined. The unknown
3
(1) is a multiplicative
constant, which I determine by imposing the normalization requirement
1 = (1) | (1) = 2.13 |
3
(1)|
2
(65)
7. The Eigenvalue Problem, December 17, 2009 13
I solved Eq. 65 for
3
(1) and I got two solutions, 0.685 and 0.685. They are
equivalent, since they dier by a factor whose absolute value is 1; such phase
factors make no dierence in any calculation of an observable quantity. I
use the positive root:

3
(1) = 0.685 (66)
and with this, I have
(1) = 0.685{0.542, 0.914, 1} = {0.371, 0.626, +0.685} (67)
If you are the kind of person who worries about the mysteries of mathe-
matics, you might wonder what happens if you use Eqs. 60 and 61, or perhaps
Eqs. 59 and 61. I did this (see WorkBook7, Section 2, Cell 2) and no matter
which pair of equations I pick, I get the same result for
1
(1) and
2
(1)! This
must be so, if the theory is consistent, but it seems rather miraculous that
it happens. However, no miracle is involved. Let us assume that we did not
calculate the eigenvalue and left it undetermined in Eqs. 5961. We can
take Eqs. 59 and 60 and solve for the ratios
1
(1)/
3
(1) and
2
(1)/
3
(1).
We obtain two solutions that depend on . Next we take Eqs. 59 and 61
and solve for
1
(1)/
3
(1) and
2
(1)/
3
(1), obtaining two new solutions for
those ratios. But the ratios obtained in the rst calculation must be equal to
the ratios obtained from the second calculation. This equality gives me an
equation for . I calculate from it and then I solve Eqs. 60 and 61, with this
value of , for
1
(1)/
3
(1) and
2
(1)/
3
(1). The result I get must be equal
to the one obtained previously. This demand for consistency of the system
of homogeneous equations, Eqs. 5961, is fullled only for special values of
. These are the eigenvalues.
Exercise 3 Implement the recipe outlined above for the homogeneous sys-
tem
(3.2 )
1
+ 2.4
2
= 0 (68)
2.4
1
+ (3.6 )
2
= 0 (69)
Solve each equation for
1
/
2
and demand that the ratio (
1
/
2
) obtained
from Eq. 68 be equal to the ratio obtained from Eq. 69. This gives you an
equation for ; call it EQ. Then calculate the characteristic polynomial of
the matrix and test whether it is the same as EQ. After that, go ahead and
nd the eigenvalues and the eigenvectors.
7. The Eigenvalue Problem, December 17, 2009 14
Section 7.2. Degenerate eigenvalues
10 Introduction. The eigenvalues of a matrix A are the roots of the
characteristic polynomial det[A I]. There is no law that prevents some
of these roots from being equal. When two or more eigenvalues are equal,
we say that they are degenerate and the corresponding eigenvectors are also
called degenerate. This name suggests a certain moral depravity for these
eigenvalues and I am sure that they nd it oensive. I could not nd who
introduced this terminology. Mathematicians are more polite and use instead
the terms multiple and single eigenvalues, which is simpler and friendlier.
One would think that nding operators that have identical eigenvalues
should be as rare as nding by accident polynomials with multiple roots.
But this is not so. Degenerate eigenvalues are fairly frequent and here I use
two examples to suggest why.
11 Particle in a cubic box with innite potential-energy walls. One of
the simplest models in quantum mechanics is a particle in a rectangular
parallelepiped whose walls are innitely repulsive. The Hamiltonian is

H =

K
x
+

K
y
+

K
z
(70)
where

K
x
=
h
2
2m

2
x
2
(71)
is the operator representing the kinetic energy of the motion in direction x;

K
y
and

K
z
are dened similarly. In classical mechanics, the total kinetic
energy is
p
2
2m
=
p
2
x
2m
+
p
2
y
2m
+
p
2
z
2m
(72)
where p is the momentum vector and p
x
is the momentum in the direction
x (and analogously for p
y
, p
z
). The operator

K
x
is the quantum analog of
p
2
x
/2m.
The total energy of the particle is discrete and is given by (see H. Metiu,
Physical Chemistry, Quantum Mechanics, page 104)
E
n,j,k
=
h
2

2
2m
_
n
L
x
_
2
+
h
2

2
2m
_
j
L
y
_
2
+
h
2

2
2m
_
k
L
z
_
2
(73)
7. The Eigenvalue Problem, December 17, 2009 15
where n, j, k can take any of the values
n, j, k = 1, 2, 3, . . . (74)
Here L
x
, L
y
, and L
z
are the lengths of the box edges.
The ket representing the pure states (the eigenstates of

H) is denoted by
|n, j, k. For example, the symbol |2, 1, 1 means the state with n = 2, j = 1,
k = 1 and the energy
E
2,1,1
=
h
2

2
2m

_
2
L
x
_
2
+
_
1
L
y
_
2
+
_
1
L
z
_
2

(75)
Exercise 4 Write down the energy of the states |1, 2, 1 and |1, 1, 2.
If L
x
= L
y
= L
z
, the states are not degenerate (except if accidentally
_
n
L
x
_
2
+
_
k
L
y
_
2
+
_
m
L
z
_
2
=
_
n

L
x
_
2
+
_
k

L
y
_
2
+
_
m

L
z
_
2
in which case |n, j, m is
degenerate with |n

, j

, m

).
But consider the case of the cubic box, when
L
x
= L
y
= L
z
= L (76)
In this case the states |2, 1, 1, |1, 2, 1, and |1, 1, 2 have the same energy,
equal to
h
2

2
2mL
2
_
2
2
+ 1 + 1
_
(77)
They are degenerate states.
Note the connection between degeneracy and symmetry. If L
x
= L
y
= L
z
,
there is no degeneracy, except perhaps by accident. However, if L
x
= L
y
=
L
z
, the box is symmetric and almost all states are degenerate.
Exercise 5 Is the state |1, 1, 1 degenerate? How about |3, 2, 1? Enumerate
the states having the same energy as |3, 2, 1.
7. The Eigenvalue Problem, December 17, 2009 16
In the Schr odinger representation, the state |n, j, k is (see Metiu, Eq. 8.31)
x, y, z | n, j, k
n,j,k
(x, y, z)
=
_
2
L
x
sin
_
nx
L
x
_
_
2
L
y
sin
_
jy
L
y
__
2
L
z
sin
_
kz
L
z
_
(78)
For the example of the cube,
x, y, z | 2, 1, 1 =
_
8
L
3
sin
_
2x
L
_
sin
_
y
L
_
sin
_
z
L
_
(79)
and
x, y, z | 1, 2, 1 =
_
8
L
3
sin
_
x
L
_
sin
_
2y
L
_
sin
_
z
L
_
(80)
A similar equation holds for x, y, z | 1, 1, 2. These three states have the same
energy but they are dierent states. What is the physical dierence between
them?
Let us look at the kinetic energy in the x-direction:

K
x
x, y, z | 2, 1, 1 =
h
2
2m

2
x
2

2,1,1
(x, y, z)
=
h
2
2m
_
2
L
_
2

2,1,1
(x, y, z) (81)
I obtained the second line from the rst by taking the second derivative of

2,1,1
(x, y, z) given by Eq. 79. We see that |2, 1, 1 is an eigenstate of

K
x
,
with the eigenvalue
h
2
2m
_
2
L
_
2
. In the same way, you can show that |2, 1, 1 is
an eigenstate of

K
y
and

K
z
, with the eigenvalues
h
2
2m
_
1
L
_
2
. The total energy
is the sum of those three kinetic energies. In the state |2, 1, 1, the particle
has higher kinetic energy in the x-direction. We can analyze in the same way
the states |1, 2, 1 and |1, 1, 2. Table 1 gives the result. The three degenerate
states have the high (i.e.
h
2
2m
_
2
L
_
2
) kinetic energy in dierent directions.
Can we distinguish these three states by some measurement? We can.
The particle in the state |2, 1, 1 will emit a photon in a dierent direction
than will a particle in |1, 2, 1 or |1, 1, 2.
The degenerate states |2, 1, 1, |1, 2, 1, and |1, 1, 2 are dierent physical
states, even though they have the same energy.
7. The Eigenvalue Problem, December 17, 2009 17
state K
x
K
y
K
z
E
|2, 1, 1
h
2
2m
_
2
L
_
2
h
2
2m
_
1
L
_
2
h
2
2m
_
1
L
_
2
6 h
2

2
2mL
2
|1, 2, 1
h
2
2m
_
1
L
_
2
h
2
2m
_
2
L
_
2
h
2
2m
_
1
L
_
2
6 h
2

2
2mL
2
|1, 1, 2
h
2
2m
_
1
L
_
2
h
2
2m
_
1
L
_
2
h
2
2m
_
2
L
_
2
6 h
2

2
2mL
2
Table 1: Kinetic and total energy
We have also found a curious thing. The degenerate states |2, 1, 1,
|1, 2, 1, and |1, 1, 2 are eigenstates of

H but also of

K
x
,

K
y
, and

K
z
. Is
this a coincidence? No. Except for accidental degeneracies, we always nd
that the states of an observable (H) that are degenerate are so because they
are also eigenstates of other observables (K
x
, K
y
, K
z
).
Finally, note that

K
x
,

K
y
, and

K
z
all commute with

H. Is this a coin-
cidence? No. We will prove in a later chapter that operators that commute
have common eigenfunctions.
Exercise 6 Take L
x
= L
y
= L and L
z
= L. Does the system have degener-
ate states? Are these states eigenstates of

K
x
,

K
y
, and

K
z
? How about the
case L
z
= L
y
= L
z
?
12 The hydrogen atom. The hydrogen atom is another system whose
energy eigenstates are degenerate. Its energy is (see Metiu, Quantum Me-
chanics, page 300)
E
n
=
1
n
2
e
4
2(4
0
)
2
h
2
(82)
and the states |n, , m are labeled by three quantum numbers: n controls the
energy, controls the square of the angular momentum (essentially, this is the
rotational energy), and m controls the projection of the angular momentum
on the OZ axis. The states |n, , m are eigenvectors of three operators. These
7. The Eigenvalue Problem, December 17, 2009 18
operators are (1) the Hamiltonian

H, for which

H|n, , m = E
n
|n, , m, (83)
(2) the angular momentum squared

L
2
, for which

L
2
|n, , m = h
2
( + 1)|n, , m, (84)
and (3) the projection

L
z
of the angular momentum on the OZ axis, for
which

L
z
|n, , m = hm|n, , m, (85)
The states |2, 0, 0, |2, 1, 1, |2, 1, 0, and |2, 1, 1 (remember that takes
only the values 0, 1, . . . , n 1 and m takes only the values , +1, . . . , )
all have the energy
E
2
=
1
4
e
4
2(4
0
)
2
h
2
(86)
They have the same energy, but they are dierent states. The state |2, 0, 0
has no rotational energy. The states |2, 1, 1, |2, 1, 0, and |2, 1, 1 have
the same rotational energy but dier through the orientation of the angular-
momentum vector. Are there experiments that distinguish these states? Yes.
A hydrogen atom in state |2, 1, 1 emits photons of the same frequency as
one in the state |2, 1, 0 but in a dierent direction.
We have the same situation as for the particle in a cubic box. The de-
generate states of the Hamiltonian have the same energy but dier through
other properties. The degenerate states are eigenstates of

H but also of other
operators (here

L
2
and

L
z
play roles similar to

K
x
,

K
y
, and

K
z
).
There is also a connection to symmetry: the degeneracy appears because
the system is spherically symmetric (the Coulomb attraction V (r) is). Had
the interaction been of a form V (x, y, z) with no spherical symmetry, the
system would not have had degenerate states.
13 Degenerate eigenvalues: an example. I hope you have accepted that
degeneracy is not a freakish accident and you need to understand the math-
ematical properties of degenerate eigenstates. These are a bit more compli-
cated than those of the non-degenerate states.
Let us start with an example. Mathematica tells me (see WorkBook7,
Section 3, Cell 1) that the matrix
M =

1 1 1
1 1 1
1 1 1

(87)
7. The Eigenvalue Problem, December 17, 2009 19
has the eigenvalues

1
= 3 (88)

2
= 0 (89)

3
= 0 (90)
with corresponding eigenvectors
v(1) = {1, 1, 1} (91)
v(2) = {1, 0, 1} (92)
v(3) = {1, 1, 0} (93)
I calculated v(i) | v(j) and found
v(1) | v(1) = 3, v(2) | v(2) = 2, v(3) | v(3) = 2 (94)
The eigenvectors are not normalized. I also calculated that
v(1) | v(2) = 0, v(1) | v(3) = 0, v(2) | v(3) = 1 (95)
This means that the degenerate eigenvectors v(2) and v(3) are not orthog-
onal. There are an innite number of eigenvectors, and Mathematica has
picked three of them by mysterious rules. They are not the best choice for
physics, however, since we need them to be orthonormal.
We have seen earlier a theorem stating that the eigenvectors of a Her-
mitian matrix corresponding to dierent eigenvalues are orthogonal. The
matrix M is Hermitian, and v(1) (corresponding to the eigenvalue 3) is or-
thogonal to v(2) and v(3) (corresponding to the eigenvalue 0). This is in
agreement with the theorem. Since the eigenvalues corresponding to v(2)
and v(3) are equal to each other, v(2) and v(3) are under no obligation to be
orthogonal and we found that they are not. There is, however, a troublesome
problem. If the Hermitian matrix represents an observable, we have assumed
that its eigenstates form a complete, orthonormal system. They dont seem
to in this case but, as youll see soon, we can turn these eigenvectors into an
orthonormal set of eigenvectors by using Gram-Schmidt orthogonalization.
14 Completeness. Does this system provide a complete basis set in the
space of three-dimensional vectors? I am asking whether it is possible to
write any vector x in our three-dimensional space in the form
x =
1
v(1) +
2
v(2) +
3
v(3) (96)
7. The Eigenvalue Problem, December 17, 2009 20
where
1
,
2
,
3
are suitably chosen numbers.
The answer will be yes if the three vectors are linearly independent, which
means that there are no numbers
1
,
2
,
3
such that

1
v(1) +
2
v(2) +
3
v(3) = 0 (97)
(the selection
1
=
2
=
3
= 0 does not count). Another way to put this is
that none of v(1), v(2), v(3) can be written as a linear combination of the
others. How do we establish that this is true? Calculating the left-hand side
of Eq. 97 for all possible numbers would take too long. Fortunately, there
is a short-cut we can use, based on the theory of determinants. Form the
matrix
B =

v
1
(1) v
2
(1) v
3
(1)
v
1
(2) v
2
(2) v
3
(2)
v
1
(3) v
2
(3) v
3
(1)

(98)
where v
i
() is the i-th component of the vector v().
If a row can be written as a linear combination of the other rows, then
the determinant of B is zero (and vice versa). Therefore
det B = 0 (99)
ensures that the three vectors are linearly independent. Straightforward cal-
culation (or Mathematica) tells me that in this case
det B = 3 (100)
We conclude that the three eigenvectors are linearly independent.
Exercise 7 Suppose three 3-dimensional vectors lie in the same plane. Are
they linearly independent? What about three vectors for which two are in
the same plane but not parallel and the third is perpendicular to that plane?
Because these three vectors are linearly independent, any three-dimensional
vector x can be written as
x =
1
v(1) +
2
v(2) +
3
v(3) (101)
where
1
,
2
, and
3
are numbers. So v(1), v(2), v(3) do form a complete
basis set. But the three vectors making up the basis are not orthonormal.
7. The Eigenvalue Problem, December 17, 2009 21
15 Orthogonalization. We can use the Gram-Schmidt orthogonalization
procedure (see Chapter 3, 17) to convert the degenerate eigenvectors v(2)
and v(3) into two orthonormal vectors x(2) and x(3). These are
x(2) = v(2) (102)
x(3) = v(3)
x(2)
x(2) | x(2)
x(2) | v(3) (103)
The calculation was performed in Section 3, Cell 3 of WorkBook7 and the
result is
x(2) = v(2) = {1, 0, 1} (104)
x(3) = {
1
2
, 1,
1
2
} (105)
It is easy to verify that Mx(2) = 0 = 0 x(2) and Mx(3) = 0 = 0
x(3), which means that x(2) and x(3) are eigenstates of M corresponding
to the eigenvalue equal to zero. Also, we can verify that v(1) | x(2) =
v(1) | x(3) = x(2) | x(3) = 0; the vectors x(i) are orthogonal, as expected.
Next we normalize x(1), x(2), and x(3), by using
s(1) =
x(1)
_
x(1) | x(1)
(106)
s(2) =
x(2)
_
x(2) | x(2)
(107)
s(3) =
x(3)
_
x(3) | x(3)
(108)
The orthonormal vectors are therefore (see WorkBook7)
s(1) =
_
1

3
,
1

3
,
1

3
_
(109)
s(2) =
_

1

2
, 0,
1

2
_
(110)
s(3) =

6
,
_
2
3
,
1

(111)
7. The Eigenvalue Problem, December 17, 2009 22
Exercise 8 Verify that (1) Ms(i) =
i
s(i), i = 1, 2, 3, with
1
= 3,
2
= 0,

3
= 0; and (2) s(i) | s(j) =
ij
, i, j = 1, 2, 3.
What you have seen in this example is general.
1. The eigenvectors of a Hermitian matrix are linearly independent.
2. The eigenvectors corresponding to dierent eigenvalues are automati-
cally orthogonal.
3. The degenerate eigenvectors are not necessarily orthogonal to each
other. One can always convert them (by Gram-Schmidt) into a set
of degenerate eigenvectors that are orthogonal to each other and to all
other eigenvectors.
4. Any nonzero vector can be normalized by dividing it by its norm
(x/

x x is normalized).
16 Find the eigenvalues of the matrix M (Eq. 87) by hand. You have
learned enough to be able to use the computer and generate orthonormal
eigenvectors. It would not hurt, however, to understand how degeneracy
appears, by doing an eigenvalue calculation, step by step, to look at the
details.
The eigenvalue problem for matrix M is
Mx = x (112)
For the matrix given by Eq. 87, this is the same as the system of equations
(1 )x
1
+ x
2
+ x
3
= 0 (113)
x
1
+ (1 )x
2
+ x
3
= 0 (114)
x
1
+ x
2
+ (1 )x
3
= 0 (115)
where x
1
, x
2
, x
3
are the components of the vector x.
The characteristic polynomial is
det[M I] = det

1 1 1
1 1 1
1 1 1

= 3
2

3
(116)
7. The Eigenvalue Problem, December 17, 2009 23
The eigenvalues are the roots of this polynomial, i.e. the solutions of the
equation
3
2

3
= 0 (117)
They are (see WorkBook7.nb, Section 3, Cell 4)

1
= 3 (118)

2
= 0 (119)

3
= 0 (120)
We can calculate the eigenvector x(1) corresponding to
1
= 3 as we did
before (this eigenvector is not degenerate). The equation (M
1
I)x(1) = 0
is shorthand for the system of linear equations:
(1 3)x
1
(1) + x
2
(1) + x
3
(1) = 0 (121)
x
1
(1) + (1 3)x
2
(1) + x
3
(1) = 0 (122)
x
1
(1) + x
2
(1) + (1 3)x
3
(1) = 0 (123)
These equations have a solution dierent from x
1
(1) = x
2
(1) = x
3
(1) = 0
because we use = 3, which makes the determinant in Eq. 116 equal to
zero. To nd this solution, we take two equations and solve for x
2
(1) and
x
3
(1). The result will depend on x
1
(1), which is left as a parameter to be
determined later.
I solved Eqs. 121 and 122 and obtained
x
2
(1) = x
1
(1) (124)
x
3
(1) = x
1
(1) (125)
The eigenvector x(1) corresponding to the eigenvalue
1
= 3 is therefore
x(1) = {x
1
(1), x
1
(1), x
1
(1)} = x
1
(1){1, 1, 1} (126)
This is a perfectly ne result. The presence of x
1
(1) in it does not bother me
since it can be determined by forcing x(1) to be normalized. The normalized
eigenvector is
s(1) =
x(1)
_
x(1) | x(1)
=
_
1

3
,
1

3
,
1

3
_
(127)
Notice that x
1
(1) disappears from this expression.
7. The Eigenvalue Problem, December 17, 2009 24
We have
(M 0I)x(2) = 0 (128)
and
(M 0I)x(3) = 0 (129)
because the eigenvalues
2
and
3
are equal to zero. However, we must have
x(2) = x(3) because they are dierent eigenstates. This can happen only if
all three equations in the system (M 0I)x = 0 are identical. If they were
not, we could proceed as we did when we calculated x(1). Take two of the
equations in the system (M 0I)x = 0 and solve for x(2) to get a solution
of the form
x(2) = x
3
(2){a, b, 1}
where a and b are numbers whose values are irrelevant for this discussion.
Normalizing this vector xes x
3
(2). But the system for x(3) is the same as
that for x(2). Therefore when we solve it we obtain x(3) = x(2). This is not
acceptable and therefore all three equations in the system must be the same.
Indeed when we insert
2
= 0 in Eqs. 113115 we obtain
x
1
(2) + x
2
(2) + x
3
(2) = 0 (130)
x
1
(2) + x
2
(2) + x
3
(2) = 0 (131)
x
1
(2) + x
2
(2) + x
3
(2) = 0 (132)
The equations for x
1
(3), x
2
(3), and x
3
(3) are identical to Eqs. 130132 be-
cause
2
=
3
. Inserting
3
= 0 in Eqs. 113115 yields three identical
equations (i.e. Eqs. 130132).
All I can conclude from Eqs. 130132 is that
x
1
(2) = x
2
(2) x
3
(2) (133)
Therefore, the eigenvectors corresponding to
2
= 0 and
3
= 0 are both of
the form
x() = {x
2
() x
3
(), x
2
(), x
3
()}, = 2, 3 (134)
Mathematics does not further determine x
2
and x
3
; any vector of this form
satises the equation Mx = 0x, as you can easily verify. In addition, any
vector of the form in Eq. 134 is orthogonal to s(1) (see WorkBook7) because
s(1) corresponds to a dierent eigenvalue of M.
7. The Eigenvalue Problem, December 17, 2009 25
Exercise 9 For x = {a b, a, b}, test that Mx = 0x and x s(1) = 0.
To get two non-zero eigenvectors corresponding to the eigenvalue = 0,
I rst pick arbitrarily x
2
(2) = 1 and x
3
(2) = 2, and obtain (from the general
form in Eq. 134)
x(2) = {3, 1, 2} (135)
Then I pick a second vector by choosing arbitrarily
1
x
2
(3) = 0 and x
3
(3) = 1
and obtain
x(3) = {1, 0, 1} (136)
Physics requires that these two vectors be normalized and orthogonal. The
Gram-Schmidt procedure will make them orthogonal, and the orthogonalized
eigenvectors are
v(2) = x(2) = {3, 1, 2} (137)
v(3) = x(3)
v(2)
v(2) | v(2)
v(2) | x(3) =
_
1
14
,
5
14
,
2
7
_
(138)
Exercise 10 Show that s(1) | v(2) = s(1) | v(3) = v(2) | v(3) = 0 and
Mv(i) = 0v(i) for i = 2, 3.
The vectors v(2) and v(3) are not normalized. The respective normalized
eigenvectors are
s(2) =
v(2)
_
v(2) | v(2)
=
_

14
,
1

14
,
2

14
_
(139)
s(3) =
v(3)
_
v(3) | v(3)
=
_
1

42
,
5

42
,
2

21
_
(140)
These vectors s(2) and s(3) are degenerate eigenvectors corresponding to the
eigenvalues
2
=
3
= 0. The system {s(1), s(2), s(3)} is orthonormal.
How does the degenerate problem dier from the non-degenerate one?
For a non-degenerate eigenvalue, we lose an equation and we can solve the
1
or almost so: make sure not to choose a multiple of the rst vector
7. The Eigenvalue Problem, December 17, 2009 26
remaining ones. The eigenvector given by them has an undetermined compo-
nent (since we lost an equation) and we nd it by imposing normalization. If
an eigenvalue is doubly degenerate, we lose two equations. The eigenvectors
have two undetermined components and we determine them by forcing the
degenerate eigenvectors to be orthonormal. If you have a 9 9 matrix and
an eigenvalue is six-fold degenerate, you lose seven equations; and so forth.
Appendix 7.1. A few properties of determinants
1. The determinant of a Hermitian or a unitary matrix is equal to the
product of its eigenvalues. Therefore the determinant of a Hermitian
matrix is a real number (even when the matrix has complex elements)
and the absolute value of the determinant of a unitary matrix is equal
to 1.
2. We have
det
_
A
1
_
=
1
det A
(141)
where by writing A
1
we are asserting that it exists.
3. The determinant of a product is the product of the determinants:
det(AB) = (det A)(det B) (142)
4. If a row of a matrix is a linear combination of other rows, then the
determinant of the matrix is zero.
5. If a column of a matrix is a linear combination of other columns, then
the determinant of the matrix is zero.
Exercise 11 Prove Eq. 141 when A is the matrix representation of

A =

| |.
Exercise 12 Show that det(A
1
BA) = det B.
7. The Eigenvalue Problem, December 17, 2009 27
Appendix 7.2. Matrix diagonalization
The following theorem is often used to simplify proofs and calculations.
Theorem. For any Hermitian matrix

H there is a unitary matrix U such that
U
1

HU is a diagonal matrix (143)
and the eigenvalues of H are the diagonal elements of .
Whenever we talk about equations involving matrices, their elements are
calculated with the same basis set {|e(i)}
N
i=1
. Since

H is Hermitian, it has
a set of eigenkets {|x(i)}
N
i=1
that satisfy the eigenvalue equation

H|x(i) =
i
|x(i), i = 1, . . . , N (144)
Dene the operator

U through
|x(i)

U|e(i) (145)
Because {|e(i)}
N
i=1
and {|x(i)}
N
i=1
are both orthonormal, complete basis sets,

U is unitary (see Property 14 in Chapter 4). Replace |x(i) in the eigenvalue

U
1

H

U|e(i) =
i
|e(i), i = 1, . . . , N (146)
Act on this with e(j)| and use e(j) | e(i) =
ji
to obtain
e(j) |

U
1

H

U | e(i) =
i

ji

ji
, i, j = 1, . . . , N (147)
The expression e(j) |

U
1

H

U | e(i) is the matrix element of the operator

U
1

H

U in the {|e(j)} representation. The matrix (having elements
ji
=

ji
) is diagonal:
=

1
0 0 0
0
2
0 0
0 0
3
0
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
0 0 0
N

(148)
7. The Eigenvalue Problem, December 17, 2009 28
The left-hand side of Eq. 147 can be written as (use

p
|e(p) e(p)| =

I)
N

p=1
N

k=1
e(j) |

U
1
| e(p) e(p) |

H | e(k) e(k) |

U | e(i)
=
N

p=1
N

k=1
_
U
1
_
jp
H
pk
U
ki
=
_
U
1
HU
_
ji
(149)
where we have denoted the matrix elements by
_
U
1
_
jp
e(j) |

U
1
| e(p)
H
pk
e(p) |

H | e(k)
U
ki
e(k) |

U | e(i)
and used the observation that the second double sum in Eq. 149 is the rule
for matrix multiplication. This allows us to write the last term (i.e. U
1
HU)
as the product of the matrices U
1
, H, and U. Combining Eq. 147 with
Eq. 149 proves the theorem.
We can go a step further and construct the matrix U from the eigenvectors
of

H. We have dened

U through Eq. 145,
|x(i) =

U|e(i)
Acting with e(j)| on this gives
e(j) | x(i) = e(j) |

U | e(i) = U
ji
(150)
Using the completeness relation for the basis set {|e(i)}
N
i=1
, we have
|x(i) =
N

j=1
|e(j) e(j) | x(i) (151)
This means that e(j) | x(i) are the components of the eigenket |x(i) in the
basis {|e(i)}
N
i=1
. They are the numbers we obtain when we nd the eigenvec-
tors of the matrix H; the eigenvectors are x(i) = {e(1) | x(i), e(2) | x(i),
. . . , e(N) | x(i) }. The matrix U is (see Eq. 150)
U =

e(1) | x(1) e(1) | x(2) e(1) | x(N)

(152)
7. The Eigenvalue Problem, December 17, 2009 29
Remember that if U
ij
is a matrix element, the rst index (here, i) labels the
rows and the second index labels the columns. A simple way to construct U
according to Eq. 152 is to take the rst eigenvector of

H and use it as the
rst column in U, then use the second eigenvector as the second column of
U, etc.
To construct U, we need to solve the eigenvalue problem for H. Because
of this, constructing U is not a shortcut for nding the eigenvectors (from
Eq. 145) or the eigenvalues (from Eq. 143). The theorem is however very
useful for simplifying some equations and as an intermediate step in some
mathematical proofs.
In Section 4, Cells 67, of the le linear algebra for quantum mechanics.nb,
I give an example of the construction of a unitary matrix U that diagonalizes
a Hermitian matrix M. In the following displays, the numbers have been
rounded to two signicant digits.
M =

1.1 2.3 + 0.022i 0.67 + 3.0i 3.5 + 5.0i

(153)
The eigenvectors are (see Cell 7 of linear algebra for quantum mechanics.nb)
x(1) = {0.33 + 0.26i, 0.26 + 0.34i, 0.024 + 0.53i, 0.60} (154)
x(2) = {0.25 + 0.42i, 0.40 + 0.18i, 0.50 + 0.26i, 0.50} (155)
x(3) = {0.026 0.50i, 0.74 + 0.024i, 0.16 + 0.11i, 0.41} (156)
x(4) = {0.18 0.54i, 0.12 + 0.26i, 0.36 + 0.49i, 0.47} (157)
Construct U by using the eigenvectors as columns:
U =

0.33 + 0.26i 0.25 + 0.42i 0.026 0.50i 0.18 0.54i

(158)
We calculate that (see linear algebra for quantum mechanics.nb)
U
1
MU =

11.63 0 0 0
0 9.87 0 0
0 0 4.16 0
0 0 0 0.77

(159)
7. The Eigenvalue Problem, December 17, 2009 30
The numbers on the diagonal are the eigenvalues of M.
Exercise 13 Show that if M is Hermitian and A has an inverse, then A
1
MA
and AMA
1
have the same eigenvalues as M. Try to nd at least two meth-
ods of proof.