Lecture3iDensitymatri

Recommended reading Nielsen chapter Z 9

2
Feymann chapter
 Mathematicians before we start with the physics

let us review the concept of trace of an operator

1.10 The Trace

The trace of an operator is defined as the sum of its diagonal entries:
X
tr(A) = hi|A|ii. (1.71)
i

It turns out that the trace is the same no matter which basis you use. You can see that
using completeness: for instance, if |ai is some other basis then
X XX XX X
hi|A|ii = hi|aiha|A|ii = ha|A|iihi|ai = ha|A|ai.
i i a i a a

Thus, we conclude that

X X
tr(A) = hi|A|ii = ha|A|ai. (1.72)
i a

The trace is a property of the operator, not of the basis you choose. Since it does not
matter which basis you use, let us choose the basis | i i which diagonalizes the operator
A. Then h i |A| i i = i will be an eigenvalue of A. Thus, we also see that

X
tr(A) = i = sum of all eigenvalues of A . (1.73)
i

Perhaps the most useful property of the trace is that it is cyclic:

tr(AB) = tr(BA). (1.74)

I will leave it for you to demonstrate this. Simply insert a convenient completeness
relation in the middle of AB. Using the cyclic property (1.74) you can also move
around an arbitrary number of operators, but only in cyclic permutations. For instance:

tr(ABC) = tr(CAB) = tr(BCA). (1.75)

Note how I am moving them around in a specific order: tr(ABC) , tr(BAC). An

example that appears often is a trace of the form tr(UAU † ), where U is unitary operator.
In this case, it follows from the cyclic property that

tr(UAU † ) = tr(AU † U) = tr(A)

Thus, the trace of an operator is invariant by unitary transformations. This is also in

line with the fact that the trace is the sum of the eigenvalues and unitaries preserve
eigenvalues.
Finally, let | i and | i be arbitrary kets and let us compute the trace of the outer
product | ih |: X X
tr(| ih |) = hi| ih |ii = h |iihi| i
i i

19
The sum over |ii becomes a 1 due to completeness and we conclude that

tr(| ih |) = h | i. (1.76)

Notice how this follows the same logic as Eq. (1.74), so you can pretend you just used
the cyclic property. This formula turns out to be extremely useful, so it is definitely
worth remembering.

20
 Matrices and thermal status
Density
Let us review a bit the Gibbs formalisms we causielen a

Hamiltonian H with eigenstuff

system described by
a

i
H Em Im Lm

then the probability to find

If this system is in equilibrium
it in state In will be given by
PEm 2
e
In
2

Moreover in equilibrium the expectation value of any observable

A is given by
A CmlAtm In 3

is there a quantum
We
may even ask the following question
allow to write 3 as
state 143 which would
us

A 441A14 Not possible ca

For instance one could naively try a skate

the answer in mo

E FI
like in In CST

But then 4 gives

E Smt Alm Im
XIA 143
m m
which is not like 3 It will coincide with 3
if A happens to be

But in that case

diagonal in the basin Im
only
the in that it is not possible to attribute
The moral of story
is
net in so a thermal state The only exception of course
a

at zero temperature where the system tends to the ground state

In order to attribute a
quantum state to the system we

is
the motion of beets The generalization
have to generalize
write f the
called a
density matrix
which we
usually as

the thermal state is defined as

density matrix of

f fImlm o

in 3 can then be written using

the expectation value of 4A
as
the trace
tr C
az

in the as tr ga
which same
 will to better motivate why an object such as
Below I
try
But before let just show you
6 makes sense doing so me

H in i makes it
something neat the decomposition of as

H For instance
convenient to study functions of

HE Ent in soul
f
H3 Em EI in Kmt

then
and so on Thus if f se is some
arbitrary function
8
f En 1m3cm
f H f
projectors
we just apply f C to the eigenvalues and multiply by the
In Kmt
G we see that
now look at
If we

E.PT
PFmtmkml a
g.szqe Z

writhing the thermal

This is a neat and very powerful way of
very to any
nice about it is that it makes no reference
state what is

writes the state directly as a function

basin it
simply
specific
of the
Hamiltonian
 be written in a neat way
partition function
The can also

z trCEP Go

because PEM
my e PH Im
e z
EPH

value 77 also be written as

Thus the expectation may

tIf p
a7 C
 for thermal skates but the idea of a
The results above are

absolutely general To motivate the reasoning

density matrix is

behind this I introduce below a construction of the density

matrix which is mere general Hopefully it will clarify the logic
behind these objects

Chapter 2

Density matrix theory

2.1 The density matrix

A ket | i is actually not the most general way of defining a quantum state. To
motivate this, consider the state |n+ i in Eq. (1.47) and the corresponding expectation
values computed in Eq. (1.48). This state is always poiting somewhere: it points at the
direction n of the Bloch sphere. It is impossible, for instance, to find a quantum ket
which is isotropic. That is, where h x i = h y i = h z i = 0. That sounds strange. The
solution to this conundrum lies in the fact that we need to also introduce some classical
uncertainty to the problem. Kets are only able to encompass quantum uncertainty.
The most general representation of a quantum system is written in terms of an
operator ⇢ called the density operator, or density matrix. It is built in such a way
that it naturally encompasses both quantum and classical probabilities. But that is
not all. We will also learn next chapter that density matrices are intimately related to
entanglement. So even if we have no classical uncertainties, we will also eventually
find the need for dealing with density matrices. For this reason, the density matrix is
the most important concept in quantum theory. I am not exaggerating. You started this
chapter as a kid. You will finish it as an adult. :)
To motivate the idea, imagine we have a machine which prepares quantum systems
in certain states. For instance, this could be an oven producing spin 1/2 particles, or a
quantum optics setup producing photons. But suppose that this apparatus is imperfect,
so it does not always produces the same state. That is, suppose that it produces a state
| 1 i with a certian probability p1 or a state | 2 i with a certain probability p2 and so
on. Notice how we are introducing here a classical uncertainty. The | i i are quantum
states, but we simply don’t know which states we will get out of the machine. We
can have as many p’s as we want. All we need to assume is that satisfy the properties
expected from a probability:
X
pi 2 [0, 1], and pi = 1 (2.1)
i

Now let A be an observable. If the state is | 1 i, then the expectation value of A

will be h 1 |A| 1 i. But if it is | 2 i then it will be h 2 |A| 2 i. To compute the actual

21
expectation value of A we must therefore perform an average of quantum averages:
X
hAi = pi h i |A| i i (2.2)
i

We simply weight the possible expectation values h i |A| i i by the relative probabilities
pi that each one occurs.
What is important to realize is that this type of average cannot be writen as h |A| i
for some ket | i. If we want to attribute a “state” to our system, then we must generalize
the idea of ket. To do that, we use Eq. (1.76) to write

h i |A| i i = tr A| i ih i |

Then Eq. (2.2) may be written as

X  ⇢ X
hAi = pi tr A| i ih i | = tr A pi | i ih i |
i i

This motivates us to define the density matrix as

X
⇢= pi | i ih i | (2.3)
i

Then we may finally write Eq. (2.2) as

hAi = tr(A⇢) (2.4)

which, by the way, is the same as tr(⇢A) since the trace is cyclic [Eq. (1.74)].
With this idea, we may now recast all of quantum mechanics in terms of density
matrices, instead of kets. If it happens that a density matrix can be written as ⇢ = | ih |,
we say we have a pure state. And in this case it is not necessary to use ⇢ at all. One
may simply continue to use | i. For instance, Eq. (2.4) reduces to the usual result:
tr(A⇢) = h |A| i. A state which is not pure is usually called a mixed state. In this case
kets won’t do us no good and we must use ⇢.

Examples
Let’ s play with some examples. To start, suppose a machine tries to produce qubits
in the state |0i. But it is not very good so it only produces |0i with probability p. And,
with probability 1 p it produces the state |1i. The density matrix would then be.
!
p 0
⇢ = p|0ih0| + (1 p)|1ih1| = .
0 1 p

22
 p
Or it could be such that it produces either |0i or |+i = (|0i + |1i)/ 2. Then,
!
1 1+ p 1 p
⇢ = p|0ih0| + (1 p)|+ih+| = .
2 1 p 1 p
Maybe if our device is not completely terrible, it will produce most of the time |0i and
every once in a while, a state | i = cos 2✓ |0i + sin 2✓ |1i, where ✓ is some small angle.
The density matrix for this system will then be
!
p + (1 p) cos2 2✓ (1 p) sin 2✓ cos 2✓
⇢ = p|0ih0| + (1 p)| ih | = 2
(1 p) sin 2✓ cos 2✓ (1 p) sin 2✓

Of course, the machine can very well produce more than 2 states. But you get the idea.
Next let’s talk about something really cool (and actually quite deep), called the
ambiguity of mixtures. The idea is quite simple: if you mix stu↵, you generally loose
information, so you don’t always know where you started at. To see what I mean,
consider a state which is a 50-50 mixture of |0i and |1i. The corresponding density
matrix will then be !
1 1 1 1 0
⇢ = |0ih0| + |1ih1| = .
2 2 2 0 1
Alternatively, consider a 50-50 mixture of the states |±i in Eq. (1.11). In this case we
get !
1 1 1 1 0
⇢ = |+ih+| + | ih | = .
2 2 2 0 1
We see that both are identical. Hence, we have no way to tell if we began with a 50-50
mixture of |0i and |1i or of |+i and | i. By mixing stu↵, we have lost information.

2.2 Properties of the density matrix

The density matrix satisfies a bunch of very special properties. We can figure them
P
out using only the definition (2.3) and recalling that pi 2 [0, 1] and i pi = 1 [Eq. (2.1)].
First, the density matrix is a Hermitian operator:

⇢† = ⇢. (2.5)

Second, X X X
tr(⇢) = pi tr(| i ih i |) = pi h i | i i = pi = 1. (2.6)
i i i

This is the normalization condition of the density matrix. Another way to see this is
from Eq. (2.4) by choosing A = 1. Then, since h1i = 1 we again get tr(⇢) = 1.
We also see from Eq. (2.8) that h |⇢| i is a sum of quantum probabilities |h | i i|2
averaged by classical probabilities pi . This entails the following interpretation: for an
arbitrary state | i,

h |⇢| i = Prob. of finding the system at state | i given that it’s state is ⇢ (2.7)

23
 Besides normalization, the other big property of a density matrix is that it is positive
semi-definite, which we write symbolically as ⇢ 0. What this means is that its
sandwich in any quantum state is always non-negative. In symbols, if | i is an arbitrary
quantum state then X
h |⇢| i = pi |h | i i|2 0. (2.8)
i
Of course, this makes sense in view of the probabilistic interpretation of Eq. (2.7).
Please note that this does not mean that all entries of ⇢ are non-negative. Some of
them may be negative. It does mean, however, that the diagonal entries are always
non-negative, no matter which basis you use.
Another equivalent definition of a positive semi-definite operator is one whose
eigenvalues are always non-negative. In Eq. (2.3) it already looks as if ⇢ is in di-
agonal form. However, we need to be a bit careful because the | i i are arbitrary states
and do not necessarily form a basis (which can be seen explicitly in the examples given
above). Thus, in general, the diagonal structure of ⇢ will be di↵erent. Notwithstanding,
⇢ is Hermitian and may therefore be diagonalized by some orthonormal basis | k i as
X
⇢= k | k ih k |, (2.9)
k
for certain eigenvalues k . Since Eq. (2.8) must be true for any state | i we may choose,
in particular, | i = | k i, which gives
k = h k |⇢|ki 0.
Thus, we see that the statement of positive semi-definiteness is equivalent to saying
that the eigenvalues are non-negative. In addition to this, we also have that tr(⇢) = 1,
P
which implies that k k = 1. Thus we conclude that the eigenvalues of ⇢ behave like
probabilities: X
k 2 [0, 1], k = 1. (2.10)
k
But they are not the same probabilities pi . They just behave like a set of probabilities,
that is all.
For future reference, let me summarize what we learned in a big box: the basic
properties of a density matrix are

Defining properties of a density matrix: tr(⇢) = 1 and ⇢ 0. (2.11)

Any normalized positive semi-definite matrix is a valid candidate for a density matrix.
I emphasize again that the notation ⇢ 0 in Eq. (2.11) means the matrix is positive
semi-definite, not that the entries are positive. For future reference, let me list here
some properties of positive semi-definite matrices:
• h |⇢| i 0 for any state | i;
• The eigenvalues of ⇢ are always non-negative.
• The diagonal entries are always non-negative.
p
• The o↵-diagonal entries in any basis satisfy |⇢i j |  ⇢ii ⇢ j j .

24
2.3 Purity
Next let us look at ⇢2 . The eigenvalues of this matrix are 2
k so
X
tr(⇢2 ) = 2
k 1 (2.12)
k

The only case when tr(⇢2 ) = 1 is when ⇢ is a pure state. In that case it can be written
as ⇢ = | ih | so it will have one eigenvalue p1 = 1 and all other eigenvalues equal to
zero. Hence, the quantity tr(⇢2 ) represents the purity of the quantum state. When it is
1 the state is pure. Otherwise, it will be smaller than 1:

Purity = P := tr(⇢2 )  1 (2.13)

As a side note, when the dimension of the Hilbert space d is finite, it also follows
that tr(⇢2 ) will have a lower bound:
1
 tr(⇢2 )  1 (2.14)
d
This lower bound occurs when ⇢ is the maximally disordered state
Id
⇢= (2.15)
d
where Id is the identity matrix of dimension d.

2.4 Bloch’s sphere and coherence

The density matrix for a qubit will be 2 ⇥ 2 and may therefore be parametrized as
0 1
BBB p q CCC
⇢ = BB@ CCA , (2.16)
q⇤ 1 p

where p 2 [0, 1] and I used 1 p in the last entry due to the normalization tr(⇢2 ) = 1.
If the state is pure then it can be written as | i = a|0i + b|1i, in which case the density
matrix becomes !
|a|2 ab⇤
⇢ = | ih | = ⇤ . (2.17)
a b |b|2
This is the density matrix of a system which is in a superposition of |0i and |1i. Con-
versely, we could construct a state which can be in |0i or |1i with di↵erent probabilities.
According to the very definition of the density matrix in Eq. (2.3), this state would be
!
p 0
⇢ = p|0ih0| + (1 p)|1ih1| = . (2.18)
0 1 p

25
Pauli matrices on Yj ry f i re IE

This is a classical state, obtained from classical probability theory. The examples in
Eqs. (2.17) and (2.18) reflect well the di↵erence between quantum superpositions and
classical probability distributions.
Another convenient way to write the state (2.16) is as

0 1
1 1 BBBB 1 + sz sx isy CCC
CCA .
⇢ = (1 + s · ) = B@ (2.19)
2 2 s x + isy 1 sz

where s = (s x , sy , sz ) is a vector. The physical interpretation of s becomes evident from

the following relation:
si = h i i = tr( i ⇢). (2.20)
The relation between these parameters and the parametrization in Eq. (2.16) is
h xi = q + q⇤ ,
h yi = i(q q⇤ ),
h zi = 2p 1.
Next we look at the purity of a qubit density matrix. From Eq. (2.19) one readily
finds that
1
tr(⇢2 ) = (1 + s2 ). (2.21)
2
Thus, due to Eq. (2.12), it also follows that
s2 = s2x + s2y + s2z  1. (2.22)
When s2 = 1 we are in a pure state. In this case the vector s lies on the surface of
the Bloch sphere. For mixed states s2 < 1 and the vector is inside the Bloch sphere.
Thus, we see that the purity can be directly associated with the radius in the sphere.
This is pretty cool! The smaller the radius, the more mixed is the state. In particular,
the maximally disordered state occurs when s = 0 and reads
!
1 1 0
⇢= . (2.23)
2 0 1
In this case the state lies in the center of the sphere. A graphical representation of pure
and mixed states in the Bloch sphere is shown in Fig. 2.1.

2.5 Schrödinger and von Neumann

We will now talk about how states evolve in time. Kets evolve according to Schrödinger’s
equation. When Schrödinger’s equation is written for density matrices, it then goes by
the name of von Neumann’s equation. However, as we will learn, von Neumann’s
equation is not the most general kind of quantum evolution, which is what we will call
a Quantum Channel or Quantum Operation. The theory of quantum operations is
awesome. Here I just want to give you a quick look at it, but we will get back to this
many times again.

26
 The Bloch sphere helps us understand why thermal estates have

to be mined stakes Take for instance

h Iz O
H Koz
z
o wz

state in then
The thermal

1 EP fpnn
s EPH
Z 2cosh 13h12

and as we saw in lecture s Tz

S
s

03h
coz tanh PI I

a too then hot o S thin in the north pole of Beech'ssphere

If ph Now comes
a 0 then coz so which in the equator
But when ph the
then when Liz o spin
if the state pure
was
the key point A pure
to be pointing somewhere on
Ny plane
the
would have
same where this is something
state of gerbil is always painting
a
Bet it's not what
would be easily detectable experimentally
that in that when ph 00
the contrary
what we find
we find On
the
which is only possible if
the spin doesn't paint
anywhere
shows I
in the center of Bloch's sphere this neatly
spin is mixed
think why thermal estates are naturally
 Thetooolimity

Remember our previous discussions about infinite temperatures

d when T ooo all storks became
If the system has dimension

with
equally einely
1 iz
In
d

This man becomes much more obvious in terms of density

matrices pH
e i3
f
1 r eBH

o 0 and thus
If T ooo p

e
PH I identity matrix Cia

dimension d
Moreover for a system with
I d 15

Thes ein f I 76
de
T ooo

ruined state
which in the unanimally
in arbitrary state
finding the system an
The prob of
2 8 above wow becomes
to Eq
I 771
0 1814 I 4410 de
de

to be found in state
At T ooo the system is equally likely every
of the Hiebert space
Enample qubits
us futrits

The density matrices for a Z level and 3 level system arreuning

would have the form
they are diagonal
Po

a I a
f Pr

is
If we think about it a
diagonal state of a Z level system
thermal i thatgiven Po
in and P l Po we can always
always
find a value of P such that
pEo PE l Po
e
Po
e P
PEO PEI e PEE EPI
e e

need co But still we can always view e

If p Po we
may p
thermal For qutrits however this is not
state as
qubit diagonal
the case we now have 2 independent probabilities
in general
and in general we cannot find a

2 because Pz l Po P
early
number p which fits both
single
all this to call attention to the special
I'm just saying because the
estates thermal estates are special
structure of thermal related to the
specific proportion
populations appear
in a
very
value of p
F
narnple.ro
characterized creation and annihilation
The QH 0 in by
such that
operators at They
and a are

ftammmm
3 CI8

and fmI Int

at pm 19
s
a In FT lm

Hamiltonian then reads

the
2e
H tw ata Yz

in lectures and reeds

The partition function
was computed
e
ptswk 2
2 a

l Btw
e

matrix becomes
thus the thermal density

bwje
pt.hu Zz
geq eB
The generalization of the Shannan entropy to density matrices

is called the von Neumann entropy

of 6 Pauli states, Ek = p1 | k ih k |, with

6
! !
1 0
| 1i = |z+ i = , | 2i = |z i = ,
0 1
! !
1 1 1 1
| 3 i = |x+ i = p , | 4 i = |x i = p , (2.52)
2 1 2 1
! !
1 1 1 1
| 5 i = |y+ i = p , | 6 i = |y i = p .
2 i 2 i
This POVM is not minimal: we have more elements than we need in principle. But
from an experimental point of view that is actually a good thing, as it means more data
is available.

2.8 The von Neumann Entropy

The concept of entropy plays a central role in classical and quantum information
theory. In its simplest interpretation, entropy is a measure of the disorder (or mixed-
ness) of a density matrix, quite like the purity tr(⇢2 ). But with entropy this disorder
acquires a more informational sense. We will therefore start to associate entropy with
questions like “how much information is stored in my system”. We will also introduce
another extremely important concept, called relative entropy which plays the role of
a “distance” between two density matrices.
Given a density matrix ⇢, the von Neumann entropy is defined as
Xu we
In terms of
simply get the
X
S (⇢) = tr(⇢ log ⇢) = k log k, (2.53)
Shannon entropy
k

where k are the eigenvalues of ⇢. Working with the logarithm of an operator can be
awkward. That is why in the last equality I expressed S (⇢) in terms of them. The
logarithm in Eq. (2.53) can be either base 2 or base e. It depends if the application is
more oriented towards information theory or physics (respectively). The last expression
in (2.53), in terms of a sum of probabilities, is also called the Shannon entropy.
The entropy is seen to be a sum of functions of the form p log(p), where p 2 [0, 1].
The behavior of this function is shown in Fig. 2.3. It tends to zero both when p ! 0
and p ! 1, and has a maximum at p = 1/e. Hence, any state which has pk = 0 or
pk = 1 will not contribute to the entropy (even though log(0) alone diverges, 0 log(0) is
well behaved). States that are too deterministic therefore contribute little to the entropy.
Entropy likes randomness.
Since each p log(p) is always non-negative, the same must be true for S (⇢):

S (⇢) 0. (2.54)

35
Moreover, if the system is in a pure state, ⇢ = | ih |, then it will have one eigenvalue
p1 = 1 and all others zero. Consequently, in a pure state the entropy will be zero:

The entropy of a pure state is zero. (2.55)

In information theory the quantity log(pk ) is sometimes called the surprise. When an
“event” is rare (pk ⇠ 0) this quantity is big (“surprise!”) and when an event is common
(pk ⇠ 1) this quantity is small (“meh”). The entropy is then interpreted as the average
surprise of the system, which I think is a little bit funny.

���

��� �/�
-� ��(�)

���

���

���
�/�
���
��� ��� ��� ��� ��� ���
�

Figure 2.3: The function p log(p), corresponding to each term in the von Neumann en-
tropy (2.53).

As we have just seen, the entropy is bounded from below by 0. But if the Hilbert
space dimension d is finite, then the entropy will also be bounded from above. I will
leave this proof for you as an exercise. What you need to do is maximize Eq. (2.53) with
P
respect to the pk , but using Lagrange multipliers to impose the constraint k pk = 1.
Or, if you are not in the mood for Lagrange multipliers, wait until Eq. (??) where I will
introduce a much easier method to demonstrate the same thing. In any case, the result
is
I
max(S ) = log(d). Occurs when ⇢ = . (2.56)
d
The entropy therefore varies between 0 for pure states and log(d) for maximally disor-
dered states. Hence, it clearly serves as a measure of how mixed a state is.
Another very important property of the entropy (2.53) is that it is invariant under
unitary transformations:
S (U⇢U † ) = S (⇢). (2.57)
This is a consequence of the infiltration property of the unitaries U f (A)U † = f (UAU † )
[Eq. (1.58)], together with the cyclic property of the trace. Since the time evolution
of closed systems are implemented by unitary transformations, this means that the
entropy is a constant of motion. We have seen that the same is true for the purity:
unitary evolutions do not change the mixedness of a state. Or, in the Bloch sphere

36
picture, unitary evolutions keep the state on the same spherical shell. For open quantum
systems this will no longer be the case.
As a quick example, let us write down the formula for the entropy of a qubit. Recall
the discussion in Sec. 2.4: the density matrix of a qubit may always beqwritten as in
Eq. (2.19). The eigenvalues of ⇢ are therefore (1 ± s)/2 where s = s2x + s2y + s2z
represents the radius of the state in Bloch’s sphere. Hence, applying Eq. (2.53) we get
✓1 + s◆ ✓1 + s◆ ✓1 s◆ ✓1 s◆
S = log log . (2.58)
2 2 2 2
For a pure state we have s = 1 which then gives S = 0. On the other hand, for a
maximally disordered state we have s = 0 which gives the maximum value S = log 2,
the log of the dimension of the Hilbert space. The shape of S is shown in Fig. 2.4.

���
��(�)
���
���
���
�(ρ)

���
���
���
���
��� ��� ��� ��� ��� ���
�

Figure 2.4: The von Neumann entropy for a qubit, Eq. (2.58), as a function of the Bloch-sphere
radius s.

The quantum relative entropy

Another very important quantity in quantum information theory is the quantum
relative entropy or Kullback-Leibler divergence. Given two density matrices ⇢ and ,
it is defined as

S (⇢|| ) = tr(⇢ log ⇢ ⇢ log ). (2.59)

This quantity is important for a series of reasons. But one in particular is that it satisfies
the Klein inequality:

S (⇢|| ) 0, S (⇢|| ) = 0 i↵ ⇢ = . (2.60)

The proof of this inequality is really boring and I’m not gonna do it here. You can find
it in Nielsen and Chuang or even in Wikipedia.

37
 Eq. (2.60) gives us the idea that we could use the relative entropy as a measure of
the distance between two density matrices. But that is not entirely precise since the
relative entropy does not satisfy the triangle inequality

d(x, z)  d(x, y) + +d(y, z). (2.61)

This is something a true measure of distance must always satisfy. If you are wondering
what quantities are actual distances, the trace distance is one of them2
q
T (⇢, ) = ||⇢ ||1 := tr (⇢ )† (⇢ ). (2.62)

But there are others as well.

Entropy and information

Define the maximally mixed state ⇡ = Id /d. This is the state we know absolutely
nothing about. We have zero information about it. Motivated by this, we can define the
amount of information in a state ⇢ as the “distance” between ⇢ and ⇡; viz,

I(⇢) = S (⇢||⇡).

But we can also open this up as

S (⇢||1/d) = tr(⇢ log ⇢) tr(⇢ log(Id /d)) = S (⇢) + log(d).

I know it is a bit weird to manipulate Id /d here. But remember that the identity matrix
satisfies exactly the same properties as the number one, so we can just use the usual
algebra of logarithms in this case.
We see from this result that the information contained in a state is nothing but

I(⇢) = S (⇢||⇡) = log(d) S (⇢). (2.63)

This shows how information is connected with entropy. The larger the entropy, the less
information we have about the system. For the maximally mixed state S (⇢) = log(d)
and we get zero information. For a pure state S (⇢) = 0 and we have the maximal
information log(d).
As I mentioned above, the relative entropy is very useful in proving some mathe-
matical relations. For instance consider the result in Eq. (2.56). If we look at Eq. (2.63)
and remember that S (⇢|| ) 0, this result becomes kind of obvious: S (⇢)  log(d) and
S (⇢) = log(d) i↵ ⇢ = 1/d, which is precisely Eq. (2.56).

2 The fact that ⇢ is Hermitian can be used to simplify this a bit. I just wanted to write it in a more
general way, which also holds for non-Hermitian operators.

38
 Respansefenctions
in statistical mechanics is that of
A
very
common problem
to a certain external
measuring the response of a system
formulated as follows
We have a system
stimulus This is usually which couples
stimulus x
Hamiltonian Ho we then apply a
with
such that the Hamiltonian changes to
to some operator A

t
231
H

like this But this kind

doesn't have to be
of course
the coupling
terms out to be quite
common the most important
of coupling
magnetic field
so
to a
is that of a spin system ecerpled
example
G
that Ho h Sz
H

now though because I want to

I will work with
237 for
valid beyond spin systems
that the results are also
emphasize
evaluate in the response of the
the quantity we naw want to

to the stimulus x
system
tr A e
P't 25
ay e

tr ePH

in that this can be written as

what I want to show

sA 26

where F T en Z
 You already showed 26 in the problem set for the ease of spin

But here I would like to go through a careful proof because

mere

subtlety involved related to whether or not A

there is actually a

and Ho commute

is that in order to write something like 26 we

The reason

need to earn pete

2 e
pH
2h

then we can write

If Ho A 0

pH PHO
e e epXA

In thin ease
P't e
P'to epia
e

A
and 2 ERA p A EP EP PA
22

Then
since A commutes with itself
A ePH when Ho A o
BH zz
2,2 e p

251 to 26 in easy
so that going from
Z
Fent If
a
f
trivial Quite
2 ePHI 2x in not

surprisingly
But if H A to competing
267 continues to hold
Let's see
why
though Eq
 to have the following BCH expansion in mind
It is useful
then
let Kcx denote our arbitrary operator

d 3 e
4 K
eN dd f K dd tz
us

the exponentials in
verify this formula by
series expanding
you can

both sides in a
Taylor series
in that to compete de it matters whether
what 287 shows DX
do then it's easy If they
with K If they
or not dI commutes
DX series
don't then we set a tin principle infinite
21 pa we then get
KC x pHo
t
p XA so
For our problem 2x

PH Ca
CK CK PA 1 e
e P't PA I k PA
a
let's earmpute
But now comes the fun part
B't
P't PH tr Ex pa e
e tr pa e Iz
tr If 30
P
what we
want
tr Ex pa e
J t
 the trace we can write the End
Using the cyclic property of
team as

tr
tr Capa e tr
f k pA e PA ke

tr pa e k tr panek
0

the same is also true for all other

since k and e earmmute
nice conclusion that
Thus we reach the quite
serums in 128

AEBH csi
tr eBH ptr
A and
126 therefore renren.ms unaltered even if
the derivation of
Ho do not commute
 Susceptibility
the susceptibility is defined as the sensitivity of a to

changes in
2C A 2 132
X 2 2

CA together with
Ilugging in the definition of
P
X I trfA
epH
2X
fr

ad tr ca eBH trae
z 2
22
2X

PH 2
CA
tr CA e p
Iz
0 we can use 277 to
If Ho A
The first term in a bit tricky

get PH a 2
tr AZ e p
X f
Z

133
pfcA7

It shows that the susceptibility the

thin result is quite nice
in X is actually related to the
sensitivity of CA to changes
a ie to the fluctuations of A ins equilibrium
variance of
 Ho A to longer holds In this ease it
when Eq 33 no

in useful to use the following Feymmann integral tricky

Hs
B pojds
PHS
e Aep e PH gu
e
a

1287 It is
consequence of

demonstrated
Thin identity is actually a

below we then get

in the appendix

EPH PH ca 2
as tr A e PASA e p
X
Tf
or

EPAeP
x.io
A 1352

like this it
reduces to 133 But written
If Ho A a o thin clearly
how He A f o affects X
clearly shows is
one
X However suite often
The susceptibility
in defined for any
this is linear response theory
the value of X gov X ro
interested in how the
and analyze
we apply an infinitesimal perturbation order we
may in thin ease compete
leading to it To
system responds
unperturbed state Ho i
the
X from

A ePHOSA
EPH's AI go
X e
p f jds
 tr C EP Izo in a thermal average over the

unperturbed
where So
Hamiltonian Ho
 fymmmti
gwihded.tw
App

e discussed in Eq 267 the following BCH expansion

j.cn a If 33 i e ca
i

dote dd ECK IF
due to
we can rewrite this in terms of an integral identity
To do that recall the
be quite useful
traditional
which can
Feynman
BC it expansions

x y X Ex Y CA 2
e y e Yt

that the coefficients are a bit off

comparing with CA 1 we see

Yz multiplies K K whereas
For instance in CA 7 the factor of

it multiplies x EX YJ
in ca 2

We can fix this by introducing

some
armillary integrals
s
first
K
K K Jds s K
f 0

Continuing in the same way

since ds s 42

K Ek K Jds K Cri k

with
I'm notdoing anything with the operators just playing
the coefficients
 then be rewritten as
Eg CA I may

u SEK k i k En K e
ds K
daff

thin with CA 27 we see that the

were
and if we earmpane
in the usual BCH forum
is precisely
integrand
s Ks k
d Jds e de e e
di

write this as
s s s we earn alternatively
changing
K KS KS
dd
e fois e DI e
0
di

Thus to summarize

ddzek ejodseksdye ksek ekfdseksdd.IE ca

s with e KS leading
can cancel out e
0 then we
K K
If
K
to DI e
dde DX

stick with A3
we have to
otherwise