The density operator in quantum mechanics

1
D. E. Soper
2
University of Oregon
20 April 2012
I oer here some background for Chapter 3 of J. J. Sakurai, Modern
Quantum Mechanics.
1 The density operator
Consider an ensemble of identical quantum systems. The system has proba-
bility w
i
to be in quantum state

i
_
. Here

i
_
= 1, but the states

i
_
are not necessarily orthogonal to each other. That means that out of all the
examples in the ensemble, a fraction w
i
are in state

i
_
. We must have
w
i
> 0 ,

i
w
i
= 1 .
(1)
We dont allow w
i
= 0 because there is no reason to include state

i
_
in the
description unless there is a non-zero number of ensemble members in that
state.
The expectation value for the result of a measurement represented by a
self-adjoint operator A is
A =

i
w
i

i
_
. (2)
(Sakurai uses [A] instead of A.)
We can write the expectation value in a dierent way using a basis

K
_
1
Copyright, 2012, D. E. Soper
2
soper@uoregon.edu
1
as
A =

i
w
i

i
_
=

i
w
i

J,K

J
_
J

K
_
K

i
_
=

J,K

i
w
i

i
_

J
_
J

K
_
=

J,K

J
_
J

K
_
= Tr[A] ,
(3)
where is the operator
=

i
w
i

i
_

. (4)
We call the density operator. It provides a useful way to characterize the
state of the ensemble of quantum systems.
In what follows, we will speak simply of a system with density operator
. That always means that we imagine having many copies of the system
an ensemble of systems.
2 Properties of the density operator
Several properties of follow from its denition. First, its trace is 1 since
1 must equal 1:
Tr[] = 1 . (5)
Second, it is self-adjoint:

= . (6)
Because it is self-adjoint, it has eigenvectors

J
_
with eigenvalues
J
and the
eigenvectors form a basis for vector space. Thus has a standard spectral
representation
=

J
_
J

. (7)
We can express
J
as

J
=

J
_
=

i
w
i

i
_

2
(8)
2
Since w
i
0 and

i
_

2
0, we see that each eigenvalue must be non-
negative

J
0 . (9)
The trace of is the sum of its eigenvalues, so

J
= 1 . (10)
Since each eigenvalue is non-negative, this implies

J
1 . (11)
It is also useful to consider the trace of
2
:
Tr[
2
] =

2
J
1

J
= 1 . (12)
3 Pure states and mixed states
In Eq. (4), we dont include any terms with w
i
= 0. Also, suppose there were
two states

i
_
, call them

a
_
and

b
_
that are the same up to a phase.
Then

a
_

b
_

. That means that the i = a term and the i = b

term in Eq. (4) can be combined in a trivial way. Lets suppose that we
always combine them if necessary.
With this understanding of Eq. (4), if there is more than one term and
not all of the states

i
_
are the same (up to a phase), then we say that
the ensemble represents a mixed state. If there is only one term, so that
=

for a single state , then we say that the ensemble represents a

pure state. Then
2
= . Thus for a pure state we have Tr[
2
] = 1.
If Tr[
2
] = 1, can we conclude that represents a pure state? Yes. If
we have
J
0,

J

J
= 1 and

J

2
J
= 1, then all of the
J
must vanish
except for one, call it
0
for J = 0, with
0
= 1. Then =

0
_
0

, so is a
projection onto the state

0
_
and thus represents a pure state.
Could we then be sure that there is not another representation of the
form (4) in which is a mixed state? Also yes, but we have to think a little
more carefully. Suppose that there is another representation of the form (4).
Then for any J with J ,= 0 we have
0 =
J
=

0
_
=

i
w
i

i
_

2
(13)
3
Since we dont allow w
i
= 0, this means that for each i that occurs in Eq. (4),

i
_
= 0. That is,

i
_
is orthogonal to each basis vector

J
_
except for

0
_
. That implies that

i
_

0
_
and

i
_

0
_
0

(14)
That is,
=

i
w
i

0
_
0

. (15)
We have agreed that if all of the

i
_
are the same up to a phase, we combine
the like terms and call the ensemble a pure state, not a mixed state. With
this understanding, we see that Tr[
2
] = 1 implies that represents a pure
state.
4 Mixing ensembles
We can mix two ensembles. If one ensemble is represented by a density
operator
1
and another is represented by a density operator
2
, then we can
make another ensemble by taking a random member of ensemble 1 a fraction
f
1
of the time and taking a random member of ensemble 2 a fraction f
2
of
the time. Here we need f
1
+ f
2
= 1. The resulting ensemble has density
operator

tot
= f
1

1
+f
2

2
. (16)
5 Spin 1/2 example
A spin 1/2 system provides a nice example of the density operator. Let be
a density operator for a spin 1/2 system. Since

= and Tr[] = 1, we can

write in the form
=
1
2
[1 +a ] . (17)
The eigenvalues of n for a unit vector n are 1. Thus the eigenvalues of
are

=
1
2
[1 [a[] . (18)
Since

cannot be negative, we must have

[a[ 1 . (19)
4
This automatically ensures that
+
1.
We have
Tr[
2
] =
2
+
+
2

=
1
2
_
1 +a
2
_
. (20)
We see that Tr[
2
] = 1 if a
2
= 1. Thus the density operators that represent
pure states have a
2
= 1, while the density operators that represent mixed
states have a
2
< 1.
The value of a tells the expectation values of
x
,
y
, and
z
. We have

j
= Tr[
j
]
=
1
2
Tr[
j
] +
1
2
a
i
Tr[
i

j
]
= 0 +a
i

ij
= a
j
.
(21)
Thus
=a . (22)
We can mix two ensembles
1
=
1
2
[1 +a
1
] and
2
=
1
2
[1 +a
2
]. The
mixture with fraction f
1
of the rst ensemble and fraction f
2
of the second
ensemble is

tot
= f
1

1
+f
2

2
=
1
2
_
1 +a
tot

, (23)
with
a
tot
= f
1
a
1
+f
2
a
2
. (24)
That is, a
tot
lies on the line from a
1
to a
2
a fraction f
2
of the way from a
1
to
a
2
. This illustrates that there is more than one way to get a given density
operator .
The ensemble with = 1/2, that is a = 0, has = 0. We say that this
state is unpolarized.
Exercise 5.1 Consider a statistical ensemble of spin 1/2 particles such that
the density operator written in matrix form (in the conventional basis in
which J
z
is diagonal) is
=
_
1/2 1/2
1/2 1/2
_
(25)
What are the expectation values of J
x
, J
y
, and J
z
for this ensemble? Is this
a pure state or a mixture? Why?
5
6 Density operator for higher j
One can have a density operator for the spin space for spin j with j > 1/2.
However, it is not so simple. With spin j, there are N = 2j + 1 dimensions.
Thus the matrix representing is an N N self-adjoint matrix, which can
be characterized with N
2
real numbers. Since we need Tr[] = 1, we can
characterize with N
2
1 real numbers. Thus for spin 1, we have N = 3
and N
2
1 = 8. Thus we need more than S
x
, S
y
, and S
z
to characterize
.
7 Entropy
With a mixed state, we have less than perfect knowledge of what the quantum
state is. One uses the entropy to describe how much less. We dene the
entropy by
S = Tr[ log()] . (26)
Actually, the denition is S = k
B
Tr[ log()] where k
B
is the Boltzmann
constant. We can get rid of k
B
by using units with k
B
= 1. Then, for
instance, we measure temperatures in eV.
In terms of the eigenvalues
J
of , this is
S =

J
[
J
log(
J
)] . (27)
We should understand here that in the case of an eigenvalue zero we take
lim
0
log() = 0 . (28)
For a pure state, all of the
J
vanish except for one, which equals 1. Then,
since log(1) = 0, we have S = 0. Mixed states have S > 0 since log() > 0
for < 1.
For instance, for our spin 1/2 example, we have
S =
+
log(
+
)

log(

)
=
1
2
[1 +[a[] log
_
1
2
[1 +[a[]
_

1
2
[1 [a[] log
_
1
2
[1 [a[]
_
.
(29)
This varies from S = 0 at [a[ = 1 to S = log 2 at a = 0.
6
8 Products of vector spaces
If we have one vector space that describes one set of quantum variables and
another vector space that describes another set of quantum variables, we can
form the tensor product of the two vector spaces. A simple example is that
we have a vector space 1
A
that describes particle A and another vector space
1
B
that describes particle B. Then we can form the tensor product space
1
A
1
B
. Then if

A
_
1
A
and

B
_
1
B
, then we can form

A
,
B
_
=

A
_

B
_
1
A
1
B
. (30)
Since 1
A
1
B
is a vector space, any linear combination of states of the form

A
_

B
_
is also in 1
A
1
B
. Note that not every vector in 1
A
1
B
has
the form of a product of a vector from 1
A
times a vector from 1
B
.
The inner product for product states is dened by

A
,
B

A
,
B
_
=

A
_

B
_
. (31)
Then the inner product for states that are linear combinations of product
states is dened by the property that the inner product is linear in its ket
vector and antilinear in its bra vector. We can construct a basis for 1
A
1
B
from product states. If the vectors

i
_
A
form a basis for 1
A
and the vectors

j
_
B
form a basis for 1
B
, then the vectors

i
_
A

j
_
B
form a basis for 1
A
1
B
.
If O
A
is an operator on 1
A
and O
B
is an operator on 1
B
, then we can
dene an operator O
A
O
B
on 1
A
1
B
by
(O
A
O
B
)
_

A
_

B
__
=
_
O
A

A
__

_
O
B

B
__
. (32)
There is also is a natural denition of O
A
as an operator on 1
A
1
B
. We
can understand O
A
acting on 1
A
1
B
as O
A
1:
O
A
_

A
_

B
__
=
_
O
A

A
__

B
_
. (33)
A simple example is for two spinless particles, A and B. Particle A has
position x
A
and particle B has position x
B
. We often describe a vector

A
_
for particle A by giving its wave function
A
(x
A
) =

x
A

A
_
. Here we use
the fact that the states

x
A
_
A
form a basis for 1
A
. Similarly we use the basis
states

x
B
_
B
of 1
B
to dene the wave functions
B
(x
B
) =

x
B

B
_
. For
both particles together, we use basis states

x
A
, x
B
_
=

x
A
_
A

x
B
_
B
(34)
7
to dene wave functions
(x
A
, x
B
) =

x
A
, x
B

_
. (35)
Sometimes the wave function that represents both particles together is a
product, (x
A
, x
B
) =
A
(x
A
)
B
(x
B
). Then

_
=

A
_
A

B
_
B
. But
note that not all functions (x
A
, x
B
) are products of two functions in this
fashion.
Another simple example is the product of a space that carries the spin
j
1
representation of the rotation group and another that carries the spin j
2
representation. The tensor product of these spaces is a vector space with
basis elements

j
1
, j
2
, m
1
, m
2
_
=

j
1
, m
1
_

j
2
, m
2
_
. (36)
The vectors in this space represent physically the combined j
1
and j
2
systems.
The general vector in the tensor product space is a linear combination of these
basis states.
9 Why do we get mixed states?
If we deliberately make a statistical ensemble of pure quantum states, then
the ensemble can be described using a density operator. But cant we just
stick with pure quantum states and avoid mixed states? Apparently not,
because the pure states have zero entropy and we know from statistical me-
chanics that there is a tendency for entropy to increase. How does that
happen?
The simplest answer is that we get mixed states because we dont look
at everything. Consider two electrons, thinking of just their spins. Each
electron has spin 1/2. A basis for the states of the two electrons together is

1/2, 1/2, m
A
, m
B
_
=

1/2, m
A
_
A

1/2, m
B
_
B
. (37)
Here m
A
and m
B
can be either +1/2 or 1/2. A particularly interesting
state of this system is the so-called spin singlet state:

_
=
1

2
_

1/2, +1/2
_
A

1/2, 1/2
_
B

1/2, 1/2
_
A

1/2, +1/2
_
B
_
.
(38)
Suppose that the A electron is sent to Alice and the B electron is sent to
Bob. If Bob measures S
z
for his electron he gets +1/2 with probability 1/2
8
and 1/2 with probability 1/2. If Alice measures S
z
for her electron she gets
+1/2 with probability 1/2 and 1/2 with probability 1/2. However, if they
both make a measurement, then Alice always gets the opposite of what Bob
gets.
In fact, this state is a bit more special. Let n be any unit vector. If Bob
measures

S n for his electron he gets +1/2 with probability 1/2 and 1/2
with probability 1/2. If Alice measures

S n for her electron she gets +1/2
with probability 1/2 and 1/2 with probability 1/2. As with the case that
n is along the z axis, if they both make a measurement, then Alice always
gets the opposite of what Bob gets. To see this, one just needs to write the
vectors in the basis of eigenvectors of

S n.
Now notice that Alice and Bob are dealing with a pure quantum state
the spin singlet state of two electron spins.
But what if Bob abandons his laboratory and joins a circus? Now nobody
is looking at Bobs electron. What does Alice see? No mater how she adjusts
the axis n of her Stern-Gerlach apparatus, she sees that her electron has a
probability 1/2 to have

S n = +1/2 and a probability 1/2 to have

S n =
1/2. Thus her electron is unpolarized, described by a statistical ensemble
with = 1/2.
This is common. We may have a quantum system that we subject to
experiments, but our quantum system is typically interacting with the en-
vironment. Thus the quantum state of our system becomes entangled with
the quantum state of the environment. This means that the quantum state
of both together is not just a product of the state of our system and the
state of the environment, just as the states of Alices and Bobs electrons is
not just something with the form

A
_

B
_
. Then if we dont measure
the quantum state of the environment, we nd that the system that we are
interested in must be described as a mixed state with a density operator.
In statistical mechanics, we often describe the environment as a heat
bath. The heat bath exchanges energy with the system of interest.
10 The canonical ensemble
In quantum statistical mechanics, one nds in certain standard situations.
One of the most important is the following. Let E = H be the expectation
value of the energy of the system. Lets say that we know what E is, but we
dont know anything else about the system. Then we look for the density
9
operator with the biggest entropy subject to the two conditions
Tr[] = 1 ,
Tr[H] = E .
(39)
The solution of this problem is
=
1
Tr[e
H
]
e
H
. (40)
This is the canonical ensemble.
To get this is not so easy. The derivation in Sakurai is not really complete.
Here is a derivation. First, dene an operator
X = log() (41)
so that
= e
X
. (42)
Consider a small variation X of X. There is a corresponding small variation
of . We want Tr[X] to be maximum subject to the two conditions:
Tr[] = 1 and Tr[H] = E. That is
Tr[()X] + Tr[(X)] = 0 (43)
for every variation X that satises
Tr[] = 0 ,
Tr[()H] = 0 .
(44)
Now, we would like to say that = X, but that isnt right if X does not
commute with . (The calculus of variations is harder with operators than
with numbers.) Instead, we can write
e
X+X
= e
(X+X)/N
e
(X+X)/N
e
(X+X)/N
, (45)
where N is very big and there is a product of N factors of e
(X+X)/N
. Here
e
(X+X)/N
is equivalent to 1 +(X +X)/N because N is large. Then to rst
order in X we have
e
X+X
e
X
= [X/N]e
X/N
e
X/N
+e
X/N
[X/N] e
X/N
+e
X/N
e
X/N
[X/N] .
(46)
10
Here the factor [X/N] appears in the rst place, then in the second place,
then in the third place, etc. We can write this as
e
X
=
N1

J=0
1
N
_
e
X/N
_
NJ1
X
_
e
X/N
_
J
. (47)
We have a sum here with many small terms. We can replace this by an
integral over a parameter that is approximated by a sum with step size
= 1/N:
e
X
=
_
1
0
d e
(1)X
Xe
X
. (48)
This seems a little complicated, but it is the price that we must pay to take
into account that X, which could be a variation in any direction in operator
space and might not commute with X.
Using Eq. (48),we have
0 =
_
1
0
d Tr[e
(1)X
Xe
X
X] + Tr[e
X
X] (49)
for every variation X that satises
0 =
_
1
0
d Tr[e
(1)X
Xe
X
] ,
0 =
_
1
0
d Tr[e
(1)X
Xe
X
H] .
(50)
Since Tr[AB] = Tr[BA], X commutes with itself, and
_
1
0
d = 1, Eq. (49)
can be written
0 = Tr[Xe
X
X] + Tr[e
X
X] . (51)
With the same sort of manipulations, we see that the required conditions
(50) on the variations X are
0 = Tr[e
X
X] ,
0 = Tr
_
X
_
1
0
d e
X
He
(1)X
_
.
(52)
11
Finally, since the rst of the two conditions is that the second term in Eq. (51)
vanishes, our revised statement of the problem is as follows. We are to nd
X such that
0 = Tr[e
X
X X] . (53)
whenever the variation X is such that
0 = Tr[e
X
X] ,
0 = Tr
_

H X
_
.
(54)
where

H =
_
1
0
d e
X
He
(1)X
. (55)
Now we have a classic problem in calculus of variations. We want to nd
a vector

F such that

F x = 0 whenever

A x = 0 and

B x = 0. In
the classic calculus of variations problem,

F,

A and

B are the gradients of
certain functions. The solution is that

F must be a linear combination of

A
and

B. The coecients are known as Lagrange multipliers. In our case, the
role of the dot product between vectors, as in

F x, is taken by the trace of
a product of operators, Tr[F(X)].
We conclude that e
X
X must be a linear combination of e
X
and

H:
e
X
X = e
X

H . (56)
That is
X =

H , (57)
where

H = e
X

H =
_
1
0
d e
(1)X
He
(1)X
. (58)
Now I claim that
X = H , (59)
solves our problem. That is because with this X we have [X, H] = 0, so that

H = H. Thus our solution is

= e

e
H
(60)
The requirement that Tr[] = 1 xes :
e

= 1/Tr[e
H
] . (61)
12
Finally, Tr[H] = E xes the value of (which is 1/T where T is the tem-
perature of the system, or 1/(k
B
T) if you dont set k
B
= 1).
This gives the canonical ensemble. There is a derivation in Sakurai, but
that derivation assumes that the operator has the same eigenvectors as
H and also has the same eigenvectors. Then all the operators commute
and everything is simple. However, that doesnt tell us what to do for a
general and doesnt tell us how to deduce that and H have the same
eigenvectors.
13