Università degli Studi di Perugia

DIPARTIMENTO DI FISICA E GEOLOGIA

Corso di Laurea Triennale in Fisica

Tesi di Laurea Triennale

Entanglement Entropy

Candidato: Relatore:
Federico Grasselli Prof. Gianluca Grignani
Matricola 256291

Anno Accademico 2013/2014

Sessione di Laurea 30 Settembre 2014
Contents

1 Density Operator 3
1.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.2 Nomenclature . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.3 Pure States . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.4 Mixed States . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
1.5 Properties . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
1.6 Time Evolution of a mixed state . . . . . . . . . . . . . . . . 10
1.7 Density Matrices . . . . . . . . . . . . . . . . . . . . . . . . . 11
1.8 Reduced Density Operator . . . . . . . . . . . . . . . . . . . . 13

2 The Density Operator Picture 15

2.1 The Measurement Postulate of Q.M. . . . . . . . . . . . . . . 15
2.2 Time Evolution and Measurements for Density Operators . . 17
2.3 New Q.M. Postulates . . . . . . . . . . . . . . . . . . . . . . . 18

3 Entanglement and Entanglement Entropy 20

3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
3.2 Pure States . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
3.2.1 Bipartite systems . . . . . . . . . . . . . . . . . . . . . 21
3.2.2 Multipartite systems . . . . . . . . . . . . . . . . . . . 32
3.3 Mixed States . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
3.3.1 Bipartite systems . . . . . . . . . . . . . . . . . . . . . 34
3.3.2 Multipartite systems . . . . . . . . . . . . . . . . . . . 41
3.4 An interesting example . . . . . . . . . . . . . . . . . . . . . . 42
3.5 Monogamy and Area Law . . . . . . . . . . . . . . . . . . . . 44

4 Conclusions 46

Appendices 48

A The generic spin- 12 state 49

B Density Matrix in the Position operator eigenbasis 51

1
C Shannon and von Neumann Entropy 52

Bibliography 54

2
Chapter 1

Density Operator

1.1 Introduction
Let’s consider a quantum system: we usually identify it through its state, i.e.
its wave function. In this way, we assume it is possible to define the system
completely and build the function which represents it. However, this is not
always feasible. For instance, in an electron-target scattering experiment we
use the electron wave function to compute the cross section of the process;
while if we want to estimate it experimentally, an electron beam –made
of many identical electrons prepared in the same initial state– is needed.
Clearly, the experimental apparatus won’t produce electrons in the exact
same conditions, hence considering a single electron as representative of the
whole beam implies a process of idealization. Although such a process is
acceptable in the majority of the cases, sometimes it is not: suppose the
beam was prepared without taking into account electron spin, therefore a
measure of Sz over all electrons will statistically result in 50% h̄/2 and
50% −h̄/2 (neglecting the presence of polarizers). Consequently, we cannot
specify the spin wave function more than:
1
ψ = √ χ+ + eiα χ−


2
where α is unknown. A spin function like the one just written, or in general
a function like
ψ = aχ+ + bχ−
(where a and b are fixed complex numbers satisfying |a|2 +|b|2 = 1), describe1
an electron spin oriented in a particular direction n = n (θ, ϕ) with θ =
2 arccos |a|, ϕ = arg b − arg a. This means that ψ can represent an electron
beam polarized along n while it cannot represent an unpolarized beam
such as the one previously hypothesized. Moreover, since the latter is an
1
proof in appendix A

3
incoherent superposition of states, it cannot be represented by any single
state of which it is composed, contradicting the assumption made at the
beginning.
In these cases, that is when the lack of information about a collection
of single states prevents us from describing them one by one completely, we
can still describe such collection through a statistical analysis, in particular
by means of the density operator [1].

1.2 Nomenclature
• Given a quantum system, if it can be identified by a complete set of
commuting observables, its state is called a pure state. We call pure
states even those states which are a coherent superposition – thus
completely determined– of such states. In other words, every state
which is represented by a definite wave function is a pure state;

• A collection of systems all described by the same pure state is named

pure ensemble;

• A mixture of different pure ensembles is named mixed ensemble, for

example the unpolarized beam.

Now we will go through the description of pure and mixed states made
by the density operator [1].

1.3 Pure States

Let |si be a pure state representing a pure ensemble, normalized to unity. We
already know how to compute expectation values and probabilities regarding
single measurements of an observable F . Given the eigenvalue equation for
F:
F |ϕn i = fn |ϕn i (1.1)
(assuming discrete eigenvalues for simplicity), |si can be expressed as super-
position of the eigenvectors of F :
X
|si = cn |ϕn i (1.2)
n

where
cn = hϕn |si. (1.3)
The probability of obtaining fn as a result of the measure of F on the system
in its state |si is:
P(fn ) = |cn |2 (1.4)

4
which may also be written, introducing the projector Pn of the nth eigen-
vector: Pn = |ϕn ihϕn | , as:

P(fn ) = hs|ϕn ihϕn |si = hs|Pn |si = hPn is . (1.5)

The expectation value of F on the pure state |si is:

X X X
hF is = hs|F |si = |cn |2 fn = hϕn |sihs|ϕn ifn = hϕn |sihs|F |ϕn i
n n n
(1.6)
thus defining the operator:
ρ = |sihs| (1.7)
it can be written: X
hF is = hϕn |ρF |ϕn i. (1.8)
n

The operator (1.7) is the density operator of a pure ensemble represented by

the pure state |si. Note that while a state, although normalized, is defined up
to an arbitrary multiplicative factor, density operators are uniquely defined,
thanks to their definition.
The (1.8) expression shows that the expectation value for F can be
obtained computing the trace of ρF matrix:

hF is = Tr(ρF ) = Tr(F ρ) (1.9)

and since trace is invariant with respect to a change of basis2 , the (1.9)
has general validity. As a matter of fact, computing the trace of ρF in an
arbitrary orthonormal basis {|αk i} we obtain hF is :
X X
Tr(ρF ) = hαk |sihs|F |αk i = hs|F |αk ihαk |si = hs|F |si = hF is (1.10)
k k

(where we used the completeness relation for {|αk i}).

Thanks to (1.9) we can compute P(fn ) in the same way: (choosing
F = Pn )
P(fn ) = hPn is = Tr(ρPn ) = Tr(Pn ρ) (1.11)

1.4 Mixed States

Let’s consider a mixed ensemble, that is, a set of pure ensembles, repre-
sented by their pure state |s(i) i, which constitute the former with different
positive weights pi . Equivalently, we can consider a single quantum system
of the ensemble, which, therefore is in one of a number of states |s(i) i with
2
hence trace is an intrinsic property of an operator, thus we won’t indicate matrix
symbols under trace sign anymore.

5
respective probabilities pi (for i = 1, ..., n). The “state” of such a quantum
system is called a mixed state.
Note that the only assumption on the states |s(i) i is the normalization
to one, with no orthogonality
P required; obviously the statistical weights pi
satisfy: 0 ≤ pi ≤ 1 and i pi = 1.
The definition of the mean value of an observable F over a mixed state
is quite obvious: X
hF i = pi hF ii (1.12)
i

where hF ii = hs(i) |F |s(i) i.

Along the same lines, the probability that a measure of F over a mixed
state provides eigenvalue fn is:
X
P(fn ) = pi P (i) (fn ) (1.13)
i

2 2
where P (i) (fn ) = |cn (i) | = |hϕn |s(i) i| and |s(i) i = n cn (i) |ϕn i.
P
As done in the previous section, we will now define the density operator
of a mixed state by making explicit the expectation value of F in the state
|s(i) i:
X 2 X
hF ii = hs(i) |F |s(i) i = fn |hϕn |s(i) i| = fn hϕn |s(i) ihs(i) |ϕn i (1.14)
n n

which substituted in hF i:
X X X X
hF i = pi fn hϕn |s(i) ihs(i) |ϕn i = hϕn | pi |s(i) ihs(i) |F |ϕn i (1.15)
i n n i

permits us to define the density operator of a mixed state as follows:

X
ρ= pi |s(i) ihs(i) | (1.16)
i

and permits us to achieve the analogous expression of (1.9) in the mixed

state case:
X
hF i = hϕn |ρF |ϕn i = Tr(ρF ) = Tr(F ρ) [ρ is the one defined in (1.16)] (1.17)
n

We point out that:

• As one can see in expression (1.17) all information about the mixture
is factorized and contained in ρ;

• As for the density operator of a pure state, the density operator of a

mixed state is defined in a unique way, since arbitrary phase factors
would vanish if factorized out of the projectors |s(i) ihs(i) |.

6
 By means of the density operator one can also compute the probability
P(fn ) as in (1.11):
X
P(fn ) = pi P (i) (fn )
i
X 2
= pi |hϕn |s(i) i|
i
X
= pi hs(i) |ϕn ihϕn |s(i) i
i
X
= pi hs(i) |Pn |s(i) i
i
X
= pi hPn ii
i
= hPn i [choosing F = Pn ]

= Tr(ρPn ) = Tr(Pn ρ) (1.18)

1.5 Properties
Now let’s list some properties of density operators. Each property is satis-
fied both by density operators of pure and mixed states, unless otherwise
specified.
1. ρ = ρ† : follows from the definition in (1.16) since pi ∈ < ∀i ;
2. Tr(ρ) = 1 :
X X X X 2
Tr(ρ) = hαk | pi |s(i) ihs(i) |αk i = pi |hs(i) |αk i| =
k i i k
2
(Parseval’s equation assures us that k |hs(i) |αk i| = hs(i) |s(i) i = 1)
P
X
= pi = 1;
i

3. ρ is positive-semidefinite: given an arbitrary vector |ai, ha|ρ|ai =

(i) (i) (i) 2
P P
i pi ha|s ihs |ai = i pi |ha|s i| ≥ 0;

4. eigenvalues satisfy 0 ≤ λk ≤ 1 ∀k : since ρ is self-adjoint, one can al-

ways diagonalize ρ and notice that its diagonal elements are its eigen-
values; now, using property 3 one has that every diagonal element, i.e.
every eigenvalue, is non-negative. The eigenvalue equation for ρ is:
ρ|λk i = λk |λk i. If we compute Tr(ρ) P
in the eigenvector basis {|λk i}
we obtain the sum of its eigenvalues: k λk , and thanks to property
2 we have: X
λk = 1. (1.19)
k

7
 In conclusion we have seen that: 0 ≤ λk ≤ 1 ∀k.
N.B.: if the eigenvalue λk̄ is equal to 1, then (1.19) shows that all
the others are 0; this is the particular case of a pure state, as a
matter of fact, writing ρ through its spectral representation: ρ =
P
k λk |λk ihλk | = |λk̄ ihλk̄ | which is the definition of pure state.

5. Tr(ρ2 ) ≤ 1 : spectral representation for ρ2 is:

X
ρ2 = λk 2 |λk ihλk |
k

which in general differs from ρ, computing Tr(ρ2 ) = λk 2 and con-

P
k
sidering property 4 and (1.19), one has:

Tr(ρ2 ) ≤ 1,

where equality occurs for pure states: recalling previous N.B., all eigen-
values are zero except for one which equals 1, hence

ρ2 = λk̄ 2 |λk̄ ihλk̄ | = |λk̄ ihλk̄ | = ρ,

now thanks to property 2, one has:

Tr(ρ2 ) = Tr(ρ) = 1.

Therefore relations

ρ2 = ρ and Tr(ρ2 ) = 1

are both sufficient to determine whether a state is pure.

The following example [2, page 15], in which we will verify property 5,
shows the fundamental difference between coherent superposition of quan-
tum states –realized through state-vector addition– and the mixing of them
–realized through statistical operator addition–:
• A coherent superposition of states leads to a pure state:
1
| %i = √ (| ↑i + | ↓i)
2
and this can be easily verified by computing Tr(P 2 (| %i)) =
Tr(P (| %i)) = 1.

• Considering instead the normalized sum of the two projectors P (| ↑i)

and P (| ↓i):
1
ρ = (P (| ↑i) + P (| ↓i))
2
we obtain a mixed state, 
confirmed by computing Tr(ρ2 ) =
Tr 14 P 2 (| ↑i) + P 2 (| ↓i) = 14 Tr(P (| ↑i) + P (| ↓i)) = 21 < 1.


8
 Such a difference may be better visualized through an ideal experiment:
consider many identically prepared systems, such as electrons, directed nor-
mally toward a Young double-slit interferometer, and, if not absorbed, sub-
sequently detected on an opaque screen. Let a1 (x) be the probability am-
plitude –i.e. the wave function– corresponding to the passage through the
first slit toward the spatial point x on the measurement screen, which, like
the interferometer, is perpendicular to the direction of the initial beam.
Similarly, let a2 (x) be the probability amplitude corresponding to passage
through the second slit toward the spatial point x. The amplitude for sys-
tems being found at x when both slits are passable, so that either slit
may be entered on the way to the screen, is:
1
a12 (x) = √ (a1 (x) + a2 (x)) (1.20)
2
√
according to the superposition principle (the factor 1/ 2 confers proper
normalization on the total probability). The probability density of finding
the quantum systems at point x on the collection screen upon measurement
is the complex square of the amplitude:
1h i
p12 (x) = |a12 (x)|2 = |a1 (x)|2 + |a2 (x)|2 + 2Re(a1 (x)∗ a2 (x)) (1.21)
2
and refers to the coherent superposition of quantum states a1 (x) and a2 (x),
which clearly interfere.
Suppose now that the experimental apparatus is modified so that only
one slit is available at a time, with both being made available in the
full course of the experiment on the entire ensemble of systems. Assuming
that diffraction from each slit is negligible, interference previously detected
vanishes. As a matter of fact in this case there is no quantum superposition
of states, but rather a mixing of them: the modified apparatus implies that
a certain population p of quantum systems –arbitrarily determined by the
experimentalist– is in state a1 (x), while the remaining population 1 − p is in
state a2 (x). The probability density of finding one of the quantum systems
at point x on the collection screen upon measurement is3 :
X
ρxx = pi |ai (x)|2 (1.22)
i

which yields, taking p = 1 − p = 1/2:

1h i
ρxx = |a1 (x)|2 + |a2 (x)|2 . (1.23)
2
Comparing (1.21) with (1.23), the striking difference between quantum su-
perposition and quantum mixing shows up: the quantum interference char-
acterizing the former is missing in the latter.
3
proof in appendix B

9
 We would like to point out one last important fact regarding density
matrices. When talking about statistical ensembles of quantum systems,
we necessarily deal with two distinct kinds of probability: statistical prob-
abilities pi which appear explicitly in ρ’s definition, and those (P (i) (fn ))
which characterize the expectation valuePof observables measured on each
pure subsystem of the mixture (hF ii = n fn P (i) (fn )) and are implicit in
the vectors |s(i) i. Consequently, when we compute traces in the density op-
erator formalism, we refer to particular representations of it, dealing with
density matrices depending on both kinds of probabilities indistinguishably.
As a result, while one can uniquely define a density operator starting from
a given mixture of states, one cannot associate a definite mixture of sys-
tems to a given density matrix; therefore there can be different system
ensembles described by the same density matrix [1].

1.6 Time Evolution of a mixed state

Given a mixed state described by its density operator:
X
ρ= pi |s(i) ihs(i) | (1.24)
i

in the Schrödinger picture we conclude ρ is time-dependent. Indeed, assum-

ing our ensemble isolated (not interacting with the outside world) statistical
weights pi are time-independent, while states |s(i) i evolve over time accord-
ing to:
d (i) 1
|s (t)i = H|s(i) (t)i. (1.25)
dt ih̄
Deriving (1.24) and using (1.25):
 
dρ X 1 (i) (i) 1 (i) (i)
= pi H|s (t)ihs (t)| − |s (t)ihs (t)|H (1.26)
dt ih̄ ih̄
i

which yields:
dρ 1
= [H, ρ]. (1.27)
dt ih̄
This last equation is called the von Neumann equation or Liouville-von Neu-
mann equation and describes how a density operator evolves in time. This
equation is the quantum analogous of the Liouville equation valid in classical
statistical mechanics:
dρcl
= {H, ρcl },
dt
where ρcl (p, q) is the phase space distribution of the system. Note that,
although identical up to a minus sign, the von Neumann equation and time
evolution of a Heisenberg operator A:
dA 1
= [A, H]
dt ih̄

10
are completely disconnected: while the former was obtained in the Schrödin-
ger picture, the latter arises in a context where operators are, in general,
time-dependent, whereas states are not; this means that in the Heisenberg
picture the density operator –as defined in (1.24)– is time-independent as
well [1].
At last, one can wonder whether (1.27) and the Ehrenfest theorem are
reconcilable, since both belong to the Schrödinger picture. One can imme-
diately verify they are self-consistent by deducing one through the other:
d d
hAi = [using (1.17)] Tr(ρA)
dt   dt
dρ 1
= Tr A = Tr([H, ρ]A)
dt ih̄
1 1
= Tr(HρA − ρHA) = Tr(ρAH − ρHA)
ih̄ ih̄
1 1
= Tr(ρ[A, H]) = h[ρ, H]i
ih̄ ih̄
which is the Ehrenfest theorem.
In case of a time-independent Hamiltonian, the integrated form of the
von Neumann equation is:

ρ(t) = e−i(H/h̄)t ρ(0)ei(H/h̄)t (1.28)

and it can be easily proved by recalling time evolution of a state for time-
independent Hamiltonians:

|s(i) (t)i = e−i(H/h̄)t |s(i) (0)i

and inserting this expression in (1.24):

X
ρ(t) = pi |s(i) (t)ihs(i) (t)|
i
X
= pi e−i(H/h̄)t |s(i) (0)ihs(i) (0)|ei(H/h̄)t
i
−i(H/h̄)t
= e ρ(0)ei(H/h̄)t .

1.7 Density Matrices

In this section we will investigate the physical meaning of the matrix ele-
ments of a density operator and then furnish a couple of notable examples
of density matrices [1].
Given an arbitrary orthonormal basis {|αk i}, the generic matrix element
of ρ is: X
ρkl = hαk |ρ|αl i = pi hαk |s(i) ihs(i) |αl i,
i

11
noting that X
|s(i) i = ck (i) |αk i with ck (i) = hαk |s(i) i
k
we can rewrite it: X ∗
ρkl = pi ck (i) cl (i) . (1.29)
i

2
• The diagonal elements are ρkk = i pi |ck (i) | ; each ρkk is given by the
P
sum of products between probability pi that a generic system belonging
2
to the mixture is in state |s(i) i and probability |ck (i) | of finding the
state |αk i in the state |s(i) i. Hence, ρkk is the global probability of
finding a generic element of the mixture in the state |αk i. For this
reason, diagonal Pelements are called relative populations of the states
|αk i. Obviously k ρkk = 1 due to property 2 of density operators.
∗
• Regarding off-diagonal matrix elements, coefficients ck (i) cl (i) express
interference effects between the states |αk i and |αl i which can arise if
the pure state |s(i) i is a linear coherent superposition of such states.
Therefore, the element ρkl represents the average of these interference
effects over the mixed ensemble, and can result in zero even if singu-
∗
larly each coefficient ck (i) cl (i) is not. In a statistical mixture of states
the averaging can cancel out interference effects. Vice versa, ρkl 6= 0
means that even at the mixture level there are residual effects of co-
herence between |αk i and |αl i. For this reason, off-diagonal matrix
elements are called coherences.

Clearly, both populations and coherences depend on the chosen basis {|αk i}.
Coherences are bounded from above by populations in the following way:

|ρkl |2 ≤ ρkk ρll (1.30)

A remarkable case is when, having a time-independent Hamiltonian,

{|αk i} are its eigenvectors. In this case the population of the state |αk i
expresses the probability of finding a mixture element with energy Ek . Such
a probability is also time-independent; indeed, using (1.28) one has:

ρkk (t) = ρkk (0) (1.31)

i
(El −Ek )t
ρkl (t) = e h̄ ρkl (0) (1.32)

where we also notice that coherences oscillate in time with Bohr frequencies.

It is particularly interesting to analyze the matrix representation of two

opposite density operators: a completely mixed state and a pure state.

12
 1. In a completely mixed state, all states must result equally populated
and there are no coherences. By naming the dimension of H as N
({|αk i} ∈ H), it follows that the population of each state is 1/N ,
hence ρ is the identity matrix multiplied by 1/N (ρkk = 1/N and
ρkl = 0 for l 6= k);

2. Concerning a pure state identified by |si, let’s represent its density

matrix in the basis {|αk i}. Expressing |si as:
X
|si = ak |αk i
k

it follows that: X
ρ= ak al ∗ |αk ihαl |
k,l

which has matrix elements:

ρkl = ak al ∗ . (1.33)

So, in general, density matrix of a pure state has populations and

coherences different from zero. However, if one of the basis vectors
is |si, let’s say |si ≡ |αk̄ i, then ak = δkk̄ , al = δlk̄ , which inserted
in (1.33) return a matrix with all zero elements except for the k̄-th
diagonal element, that equals 1.

1.8 Reduced Density Operator

Given a mixed state C represented by density operator ρ, suppose we want to
measure the observable FA which operates only on A subset of C, and let B
be the complementary set of A in C. What can we say about the expectation
value of FA ? In order to answer, we first define the following orthonormal
bases: {|αk i} ∈ A, {|βl i} ∈ B, {|γkl i} ∈ C where |γkl i ≡ |αk i|βl i. Next
we define an extension of FA which operates in the entire space of C: F =
FA ⊗IB . Then we define the expectation value of FA on C as the expectation
value of F on C:
X
hFA i ≡ hF i = Tr(ρF ) = hγkl |ρF |γkl i = (completeness of {|γkl i})
kl
XX
= hγkl |ρ|γjm ihγjm |FA ⊗ IB |γkl i = (|γkl i ≡ |αk i|βl i)
kl jm
X
= hαk |hβl |ρ|βm i|αj ihβm |βl ihαj |FA |αk i = (sum over m)
kljm
X X
= hαk | hβl |ρ|βl i|αj ihαj |FA |αk i.
kj l

13
Denoting ρA the partial trace computed in the subspace containing subset
B, i.e.: X
ρA = TrB ρ = hβl |ρ|βl i (1.34)
l

we have:
X
hFA i = hαk |ρA |αj ihαj |FA |αk i (completeness of {|αj i})
kj
X
= hαk |ρA FA |αk i = Tr(ρA FA ) (1.35)
k

As one can see, the expectation value of FA has been manipulated achieving
an analogous form of (1.17) by means of the operator ρA , named reduced
density operator. The reduced density operator works only on the subspace
containing subset A and satisfies all the properties already listed for a generic
density operator [1].

14
 Chapter 2

The Density Operator

Picture

In this chapter we will briefly show how all the postulates of Q.M. given
in the state vector language can be reformulated in terms of the density
operator language, since some of them will be useful afterwards. The two
formulations are equivalent and both arise in the context of Schrödinger’s
picture, however it is sometimes much easier to approach problems from one
point of view rather than the other [3].

2.1 The Measurement Postulate of Q.M.

Let’s start by giving a slightly different version of the Measurement postulate
of Q.M. than the one usually encountered in a first study of the state vector
language. Note that it still refers to the state vector picture!

One knows that closed quantum systems evolve according to unitary evo-
lution; nevertheless, when an experimentalist observes the system to carry
out a measure of some physical magnitude, the interaction which follows
makes the system no longer closed, thus not subjected to unitary evolution.
The following postulate provides a means for describing the effects of mea-
surements on quantum systems: (number 3 refers to the numeration we will
give at the end of the chapter)

Postulate 3: Quantum measurements are described by a collection {Mm } of mea-

surement operators acting on the state space of the system being mea-
sured. The index m refers to the measurement outcomes that may
occur in the experiment. If the state of the quantum system is |ψi
immediately before the measurement, then the probability that result
m occurs is given by:

p(m) = hψ|Mm † Mm |ψi (2.1)

15
 and the state of the system after the measurement is:

Mm |ψi
|ψ 0 i = p (2.2)
p(m)

Furthermore the measurement operators satisfy the completeness equa-

tion: X
Mm † Mm = I, (2.3)
m

such an equation expresses the fact that probabilities sum to one:

X X
p(m) = hψ|Mm † Mm |ψi = hψ|ψi = 1
m m

(assuming normalized state representing our system).

Obviously this postulate cannot be in contradiction with more widely-known

assumptions of Q.M.; as a matter of fact, we can obtain them in the partic-
ular case of projective measurements.

Projective m.ts: A projective measurement is described by an observable M , a Hermi-

tian operator on the state space of the system being observed. The
observable has a spectral decomposition:
X
M= mPm
m

where Pm = |mihm| is the projector onto the eigenspace of M with

eigenvalue m. These kind of measurements can be understood as a
special case of Postulate 3, that is when Mm = Pm .

Inserting Mm = Pm in (2.1):

p(m) = hψ|Mm † Mm |ψi = (Mm ψ, Mm ψ) = kMm ψk2 = k|mihm|ψik2

= (k|mik2 = 1 and hm|ψi = c(m), being |ψi = m c(m)|mi)
P

= |c(m)|2

which is the well-known probability for outcome m.

Inserting Mm = Pm in (2.2):

Mm |ψi |mihm|ψi
|ψ 0 i = p = = ±|mi (2.4)
p(m) |hm|ψi|

which is, up to a sign, the known final state after a measurement.

16
2.2 Time Evolution and Measurements for Den-
sity Operators
What we are going to show is that the postulates of Q.M. related to unitary
evolution and measurement can be rephrased in the language of density
operators. In the next section we complete this rephrasing by giving an
intrinsic characterization of the density operator that doesn’t rely on the
idea of a state vector.

Time evolution can be easily described in the density operator language:

suppose that the evolution of a closed quantum system is determined by
the unitary operator U . If the system was initially in the state |ψi i with
probability pi , then after the evolution has occurred the system will be in
the state U |ψi i with probability pi . Therefore the evolution of the density
operator is described by the equation:
U
X X
ρ= pi |ψi ihψi | −→ pi U |ψi ihψi |U † = U ρU † (2.5)
i i

Measurements are also readily described in this context: suppose we perform

a measurement defined by measurement operators Mm . If the initial state
was |ψi i, then the probability of getting result m is (2.1):
X
p(m|i) = hψi |Mm † Mm |ψi i = cj ∗ huj |Mm † Mm |ψi i = (cj ∗ = hψi |uj i)
j
X
= huj |Mm † Mm |ψi ihψi |uj i
j

= Tr(Mm † Mm |ψi ihψi |) (2.6)

hence the total probability of getting m in any of the subsystems is:

X
p(m) = p(m|i)pi
i
X
= pi Tr(Mm † Mm |ψi ihψi |)
i
= Tr(Mm † Mm ρ)(which is a generalization of (1.18)) (2.7)

What is the density operator of the system after obtaining the measurement
result m? If the initial state was |ψi i then the state after getting m is (2.2):

Mm |ψi i
|ψi m i = p (2.8)
p(m|i)

Thus, after a measurement which yields the result m, we have an ensemble of

states |ψi m i with respective probabilities p(i|m). The corresponding density

17
operator ρm is:
X
ρm = p(i|m)|ψi m ihψi m | =
i
X Mm |ψi ihψi |Mm †
= p(i|m) =
p(m|i)
i

(by the elementary probability theory: p(m, i) = p(i|m)p(m) = p(m|i)p(i) ,

so: p(i|m) = p(m, i)/p(m) = p(m|i)pi /p(m))
X Mm |ψi ihψi |Mm †
= pi =
p(m)
i
Mm ρMm †
= (2.9)
Tr(Mm † Mm ρ)

Before moving to the next section we want to remark that stating that
a quantum system is prepared in the state P ρi with probability pi means it is
described by the density operator ρ = i pi ρi . Indeed, although prepared
in ρi with probability pi , there is a probability 1 − pi of finding it in many
other states ρi0 with respective probabilities pi0 .

2.3 New Q.M. Postulates

In the previous chapter we have seen that density operators defined by (1.16)
satisfy properties 1,2,3 on page 7. Conversely, we could use these properties
in order to define a density operator to be a Hermitian semidefinite-positive
operator with trace equal to one. Making this definition allows us to refor-
mulate the postulates of Q.M. in the density operator picture, unbinding it
from the state vector idea.

Postulate 1 Associated to any isolated physical system is a Hilbert space

known as the state space of the system. The system is completely described
by its density operator, which is a Hermitian semidefinite-positive operator
with trace equal to one, acting on the state space of the system. If a quantum
system is in
P the state ρi with probability pi , then the density operator for the
system is i pi ρi .

Postulate 2 The evolution of a closed quantum system is described by the

unitary transformation (2.5):

ρ0 = U ρU † (2.10)

18
Postulate 3 (Generalized Measurements) Quantum measurements are
described by a collection {Mm } of measurement operators acting on the state
space of the system being measured. The index m refers to the measurement
outcomes that may occur in the experiment. If the state of the quantum
system is ρ immediately before the measurement, then the probability that
result m occurs is given by (2.7):

p(m) = Tr(Mm † Mm ρ) (2.11)

and the state of the system after the measurement is (2.9):

Mm ρMm †
ρm = (2.12)
Tr(Mm † Mm ρ)

The measurement operators satisfy the completeness equation:

X
Mm † Mm = I. (2.13)
m

Postulate 4 The state space of a composite physical system is the tensor

product of the state spaces of the component physical systems. Moreover, if
we have systems numbered 1 through n, and system number i is prepared in
the state ρi , then the joint state of the total system is ρ1 ⊗ ρ2 ⊗ ... ⊗ ρn .

19
Chapter 3

Entanglement and
Entanglement Entropy

3.1 Introduction
Entanglement is arguably the most fundamental – according to Schrödinger
– and potentially disturbing characteristic distinguishing the quantum from
the classical world. One could describe it as follows: it implies that the
measurement of an observable of a subsystem may affect drastically and
instantaneously the possible outcome of a measurement on another part of
the system, no matter how far apart it is spatially. This is to be distinguished
from the phenomenon of classical correlation, where the distribution of the
possible outcomes of a measurement on one part of the system may depend
on the outcome of a previous measurement elsewhere, but is strictly limited
by the signal propagation speed [7].
In order to better understand the above description, let’s immediately
present a concrete example of entanglement: the singlet state of two spin- 12
particles
1
|s0 i = √ (| ↑i1 ⊗ | ↓i2 − | ↓i1 ⊗ | ↑i2 ) (3.1)
2
where subscripts 1,2 refer respectively to the first and second particle. If an
observer measures the spin of the first particle and obtains a particular value,
say σz 1 = +1/2, the spin value of the second particle is automatically fixed
to σz 2 = −1/2, since after the measurement the system has collapsed into
its eigenstate | ↑i1 ⊗ | ↓i2 . This phenomenon implies a sort of non-classical
correlation between the two particles, i..e. entanglement. The corresponding
state is called entangled. Obviously entanglement is not always guaranteed
in all two-spin systems, for instance in the triplet:

|t1 i = | ↑i1 ⊗ | ↑i2 (3.2)

measuring the spin of the first particle doesn’t give additional information

20
about the state of the second particle, hence the two particles are not en-
tangled, and their state is called separable.

Entanglement can be defined for both pure and mixed states. In the first
case the definition of entanglement poses no difficulty and can be naturally
understood in terms of standard correlations. In the second, however, the
definition is subtle, as correlations do not directly imply entanglement. At
the same time, for two objects, entanglement can be easily defined and quan-
tified, while for many objects this is not so simple: new correlations may
appear, which cannot be reduced to the case of two objects. We will intro-
duce separately entanglement for pure and mixed states, considering in each
the case of bipartite systems (i.e. two objects) and multipartite ones. More
details can be found in [6].

3.2 Pure States

We start out by considering the simplest case, namely the entanglement
present in a many-body quantum pure state Ψ. This is the case, for instance,
at zero temperature (if there’s no degeneracy), or in most applications in
quantum information science. Unfortunately this is an idealized case, since
zero temperature cannot be reached in practice and also interaction with the
environment causes a non-unitary evolution of the pure state into a mixed
one, a process referred to as decoherence.

3.2.1 Bipartite systems

Given two systems A and B, we denote by HA and HB the corresponding
Hilbert spaces, and by {|niX } an orthonormal basis in HX , where n =
1, 2, ...dX with dX = dim(HX ). Unless we state differently, we will always
work with spin- 12 particles, where dA = dB = 2 (generalization to higher
dimensions is straightforward). These systems are also called qubits since
they are widely used in quantum information theory, so we will take as
a basis the standard one: {|0iX , |1iX }. The Hilbert space corresponding
to the whole system is H = HA ⊗ HB (according to Postulate 4). An
orthonormal basis in that space is {|niA ⊗ |miB } that for simplicity will be
written {|n, mi}. For instance, any pure state representing two subsystems
A and B can be written as:
X X
|Ψi = cn,m |n, mi |cn.m |2 = 1.
n,m n,m

We will consider observables for each of the subsystems, which will be

represented as operators acting on the corresponding spaces. For example,
σ1A will stand for σ1A ⊗ I B , which denotes an operator σ1 acting on A and

21
I acting on B. Pauli operators acting on qubits will often appear; they are
defined as follows:

σx = |0ih1| + |1ih0|
σy = i(|0ih1| − |1ih0|)
σz = |1ih1| − |0ih0| = −iσx σy

Definition 1 We say that Ψ ∈ H is a product state if there exist two

vectors ϕ1 ∈ HA and ϕ2 ∈ HB such that |Ψi = |ϕ1 iA ⊗ |ϕ2 iB . Otherwise,
we say that Ψ is an entangled state.

This definition is a particular case of a more general one given by Schrödinger

for entangled pure states of multipartite systems, which we will see in the
next subsection. Examples of product states are those forming the orthonor-
mal basis |n, mi. Examples of entangled states are the so-called Bell states:
1
|Φ± i = √ (|0, 0i ± |1, 1i) (3.3)
2
1
|Ψ± i = √ (|0, 1i ± |1, 0i) (3.4)
2
Quantum correlations. The difference between entangled and product
states is that the first give rise to correlations. If we have a product state
|Ψi = |ϕ1 , ϕ2 i then the expectation value hΨ|σ1A ⊗ σ2B |Ψi factorizes into
hϕ1 |σ1 |ϕ1 ihϕ2 |σ2 |ϕ2 i, therefore the result of measurements in both systems
will be uncorrelated. For an entangled state, on the contrary, there always
exist a couple of observables in A and B whose expectation value does not
factorize and, thus, for which the results of measurements will be correlated.
For instance, let’s consider the Pauli operator along the direction ~n as:

σ~n = nx σx + nz σz (where ~n is a unity vector with ny = 0) (3.5)

One can see that:

hΦ+ |σ~nA ⊗ σm
B +
~ |Φ i = ~
n·m
~ (3.6)
whereas hΦ+ |σ~nA |Φ+ i = 0. Thus, whenever we measure the same Pauli op-
erators in A and B (i.e. when ~n and m~ are parallel) the results are random
but maximally correlated; that is, we have the same outcome in A and B.
This statement, however, was obvious if applied to a measure of σzA and σzB
on Φ+ , due to its definition (3.3). For this reason, Φ+ and the other Bell
states are usually referred to as maximally entangled states; they give rise
to the maximal quantum correlations.

Schmidt decomposition (SD). In order to analyze entanglement in bi-

partite systems, it is useful to introduce the Schmidt decomposition. Given

22
a generic pure state Φ representing two subsystems A and B it is always
possible to find an orthonormal basis {|un i} in HA and {|vm i} in HB such
that: X
|Φi = dk |uk , vk i (3.7)
k≤dA ,dB

where elements dk are called Schmidt coefficients.

A proof of this theorem ([4]) could be the following: any pure state Φ
representing two subsystems A and B can be written as:
X X
|Ψi = cn,m |niA |miB = |niA |ñiB (3.8)
n,m n

where
P {|niA } and {|miB } are orthonormal bases and having defined |ñiB =
m n,m |miB which are not necessarily orthonormal. Now suppose that
c
{|niA } is an eigenbasis1 of ρA (reduced density operator) with eigenvalues
{dn 2 }. Then spectral representation for ρA is:
X
ρA = dn 2 |niA hn|A (3.9)
n

By definition, we can compute ρA through the partial trace of ρ = |ΦihΦ|

on B:
!
X
ρA = TrB (|ΦihΦ|) = (3.8) = TrB |niA hm|A ⊗ |ñiB hm̃|B
n,m
X X X
= |niA hm|A TrB (|ñiB hm̃|B ) = |niA hm|A (hk|ñihm̃|ki)B
n,m n,m k
X X
= |niA hm|A (hm̃|kihk|ñi)B = (completeness of {|ki})
n,m k
X
= |niA hm|A hm̃|ñiB (3.10)
n,m

Comparing (3.9) with (3.10) one deduces:

hm̃|ñiB = dn 2 δmn (3.11)

which assures us that {|ñiB } is an orthogonal basis. We can normalize it by

rescaling: |n0 iB = d1n |ñiB for all non-zero eigenvalues. From (3.8) we finally
have: X X
|Ψi = |niA |ñiB = dn |niA |n0 iB (3.12)
n n
that proves the theorem.
Some notes ([4]):
1
an orthonormal eigenbasis of ρA can always be found since ρA is a self-adjoint operator.
Furthermore, since it is semidefinite positive, its eigenvalues are non-negative.

23
 1. Every pure state can be decomposed
P according to SD, however while
this decomposition: |Ψi = n,m cn,m |n, mi can be realized with any
orthonormal basis, the basis used in SD depends on the pure state to
be decomposed, as the proof highlights.

2. Computing the reduced density operator of the system B:

X X
ρB = TrA (|ΦihΦ|) = ... = |k̃iB hk̃|B = dn 2 |n0 iB hn0 |B (3.13)
k n

we find out that ρB has the same non-zero eigenvalues of ρA although

dimensions of HA and HB might be different, in which case the number
of null eigenvalues will differ. In conclusion, The square of the
Schmidt coefficients are the non-zero eigenvalues of both ρA
and ρB .

3. Every pure state is associated with its Schmidt number, which is the
number of non-zero eigenvalues of ρA , hence the number of terms com-
posing the state’s SD. If the Schmidt number is equal to 1, then the
state is a product state, also called a separable state. If the Schmidt
number is greater than 1, then the state is entangled, since it cannot
be written as a tensor product of two vectors.

4. It is important to stress the fact that if the pure state representing the
whole system is separable, then the reduced density matrices ρA and
ρB describe pure subsystems (because they are in the form (1.7));
if the pure state is entangled, then the reduced density matrices
describe mixed subsystems (because they are in the form (1.16)).

Consider the following example:

|Φi = cos(θ)|0, 0i + sin(θ)|1, 1i (3.14)

Φ is already written in the the SD form. The eigenvalues of the reduced den-
sity operator are cos(θ)2 , sin(θ)2 ; for θ = 0 we have a product state, whereas
for θ = π/4 we have the Bell state Φ+ , which is a maximally entangled
state, as explained in the previous paragraph. In parallel, the reduced den-
sity operators get more and more mixed as one increases θ from 0 to π/4
(note that the purity of a mixed state ρ is related to the distribution if its
eigenvalues when considered as probabilities).

Entanglement entropy. The previous example shows that entanglement

of a pure state is linked to the mixedness of the reduced density operators.
As a consequence, we could use this relation to introduce a measure of entan-
glement of a pure state by using any measure of mixedness of its subsystems.

24
A natural way of measuring the latter2 is the von Neumann entropy:

S(ρ) = −Tr(ρ ln ρ) (3.15)

Thus, we define the entropy of entanglement of a pure state Φ as:

X
E(Φ) = S(ρA ) = S(ρB ) = − dk 2 ln dk 2 (3.16)
k

For a product state E = 0 (since Schmidt number equals 1 and d =

1), whereas the maximum entanglement is E = log[min(d p A , dB )] which is
reached for the state for which all dk are equal to 1/ min(dA , dB ). These
are thus called maximally entangled states. One can easily verify that the
Bell states are maximally entangled states, as already pointed out.
We want to emphasize the fact that, in order to evaluate the amount
of entanglement of the total pure state Φ, we quantify the v.N. entropy of
its subsystems A and B, rather than the v.N. entropy of the whole state Φ,
which is always zero for pure states. This because entanglement is strictly
related to the correlation between the two subsystems, which can be visual-
ized through a measure of their v.N. entropy.
Finally, the fact that the v.N. entropy of a pure state is zero is quite
obvious, considering that such entropy has been defined to measure the
mixedness of a state. Furthermore, supposing that a quantum pure system
has a pure subsystem, such a system is necessarily separable (as noticed in
the SD paragraph), hence E = 0, which means that the v.N. entropy of its
pure subsystems is null. Alternatively, one could simply compute S(ρ) using
for the pure state ρ the representation given in section 1.7.
Let’s compute the entropy of entanglement of the following parametric
pure state ([3, page 504]):
√ p
|Φi = p| ↑i ⊗ | ↓i + 1 − p| ↓i ⊗ | ↑i (3.17)

which is completely separable for p = 0, 1 and maximally entangled for p =

1/2. The reduced density matrix related to the first subsystem can be easily
obtained, noticing that the vector basis |niA used for subsystem A in the
generic SD (3.12) is the same basis which diagonalizes ρA in (3.9). Therefore,
looking at (3.17) we can immediately write ρA ’s spectral decomposition:

ρA = p| ↑ih↑ | + (1 − p)| ↓ih↓ |

We can then compute:

E(Φ) = S(ρA ) = −p log(p) − (1 − p) log(1 − p) (3.18)

(we don’t specify the logarithm base voluntarily, it could be either 2 or

e). Below we plotted E(Φ) as a function of p; note that the pure state is
2
more info in Appendix C

25
maximally entangled for p = 1/2, as we expected.

Before giving a physical interpretation of the entropy of entanglement, we

will highlight some basic properties of the von Nuemann entropy [3].

Property 1 Suppose a composite system AB is in a pure state. Then

S(A) = S(B).

Property 2 Suppose pi are probabilities, and the states ρi have support on

orthogonal subspaces. Then
!
X X
S pi ρi = H(pi ) + pi S(ρi ).
i i

Property 3 (Joint entropy theorem) Suppose pi are probabilities, |ii

are orthogonal states for a system A, and ρi is any set of density opera-
tors for another system B. Then
!
X X
S pi |iihi| ⊗ ρi = H(pi ) + pi S(ρi ).
i i

Property 4 (Entropy of a tensor product) Suppose ρ is a density op-

erator acting on system A, while σ is a density operator acting on system
B. Then
S(ρ ⊗ σ) = S(ρ) + S(σ).

26
Proof

1. From the SD we know that the eigenvalues of the density operators of

systems A and B are the same. The entropy is determined completely
by these eigenvalues, so S(A) = S(B).

2. Let λji and |eji i be the eigenvalues and corresponding eigenvectors of

ρi . Observe that pi λji and |eji i are the eigenvalues and eigenvectors of
pi ρi , and since different ρi s belong to orthogonal subspaces, pi λji and
|eji i will also be the eigenvalues and eigenvectors of i pi ρi . Thus
P

!
pi λji log pi λji
X X
S pi ρi = −
i ij

λji − λji log λji

X X X X
= − pi log pi pi
i j i j

( j λji = 1, see (1.19))

P
X
= H(pi ) + pi S(ρi )
i

3. As in the previous proof, let pi λji and |eji iB be the eigenvalues and
corresponding eigenvectors of pi ρi . Clearly, the normalized eigenvector
j j
of the operator |iihi| is |iiA with eigenvalue 1. Hence pi λP i and |iiA |ei iB
are the eigenvalues and eigenvectors of the operator i pi |iihi| ⊗ ρi ,
which means that:
!
pi λji log pi λji = H(pi ) +
X X X
S pi |iihi| ⊗ ρi = − pi S(ρi ).
i ij i

4. Suppose pi are probabilities, the states ρi have support on orthogonal

subspaces of a space HA , and σi have support on orthogonal subspaces
of another space HB . Similarly to the last two proofs, let pi λji and
|eji iA be the eigenvalues and corresponding eigenvectors of pi ρi , while
ηik and |gik iB are the eigenvalues and P eigenvectors of σi . Hence pi λji ηik
are the eigenvalues of the operator i ρi ⊗ σi corresponding to the
eigenvectors |eji iA |gik iB . Therefore by the definition of von Neumann

27
 entropy:
!
pi λji ηik log(pi λji ηik )
X X
S pi ρi ⊗ σi = −
i ijk

λji ηik log(λji ηik )

X X X
= − pi log pi − pi
i i jk
 
X j
λi log λji +
X X
= H(pi ) − pi  ηik log ηik 
i j k
X
= H(pi ) + pi (S(ρi ) + S(σi )) (3.19)
i

By calling Γi ≡ ρi ⊗ σi and using property 2 one has that:

! !
X X X
S pi ρi ⊗ σi = S pi Γi = H(pi ) + pi S(Γi ) (3.20)
i i i

Now comparing (3.19) with (3.20) one has that:

S(Γi ) = S(ρi ) + S(σi ) (3.21)
which proves the theorem.
More important properties can be deduced once one defines ([3]):
• Joint entropy for a composite system with two components A and
B in the obvious way: S(A, B) = −Tr(ρAB log ρAB ) where ρAB is the
density matrix of the system AB.
• Conditional entropy by:
S(A|B) = S(A, B) − S(B).

• Mutual information by:

S(A : B) = S(A)+S(B)−S(A, B) = S(A)−S(A|B) = S(B)−S(B|A).

Some properties of the Shannon entropy H(X) fail to hold for the von Neu-
mann entropy, for instance for random variables X and Y , the inequality
H(X) ≤ H(X, Y ) holds. This makes intuitive sense: surely we cannot be
more uncertain about the sate of X than we are about the joint state of X
and Y . This intuition fails for quantum states. Consider a system AB of
two qubits in the entangled state Φ+ . This is a pure state, so S(A, B) = 0.
On the other hand, system A has density operator I/2, and thus has entropy
equal to one. Another way of stating this result is that, for this system, the
quantity S(B|A) = S(A, B) − S(A) is negative. In other words, the subsys-
tems of the entangled system may exhibit more disorder than the system as
a whole. In the classical world this never happens.

28
Property 5 Suppose |ABi is a pure state of a composite system. Then
|ABi is entangled ⇐⇒ S(B|A) < 0.

Property 6 (Subadditivity) Suppose distinct quantum systems A and B

have a joint state ρAB . Then the joint entropy for the two systems satisfies
the inequalities:

S(A, B) ≤ S(A) + S(B) S(A, B) ≥ |S(A) − S(B)|.

Property 7 (Concavity) The von Neumann entropy is a concave function

of itsP
inputs. That is, given probabilities pi – non-negative real numbers such
that i pi = 1– and corresponding density operators ρi , the entropy satisfies
the inequality: !
X X
S pi ρi ≥ pi S(ρi ).
i i
(note that in this case we didn’t make any assumption on the operators ρi ,
contrarily to property 2).

Proof
5. S(A, B) = S(B, A) = 0 since |A, Bi is pure. If |ABi is entangled,
then ρA is a mixed state, for which S(A) > 0, consequently S(B|A) =
S(B, A) − S(A) = −S(A) < 0.
If S(B|A) < 0 then S(A) > 0 indicating that ρA is a mixed state,
which means that |A, Bi is entangled.

6. This proof has been omitted.

7. Suppose the ρi are states of a system A. Introduce an auxiliary system

B whose state space has an orthonormal basis {|ii} corresponding to
the index i on the density operators ρi . Define a joint state of AB by:
X
ρAB = pi ρi ⊗ |iihi| (3.22)
i

To prove concavity we use the sub-additivity of the entropy. Note that

for the density matrix ρAB we have:
!
X
S(A) = S(ρA ) = S(TrB (ρAB )) = S pi ρi (3.23)
i
S(B) = S(ρB ) = S(TrA (ρAB )) (3.24)
!
X
= S pi |iihi| = (property 2) = H(pi )
i
X
S(A, B) = H(pi ) + pi S(ρi ) (thanks to property 3)
i

29
Applying the sub-additivity inequality S(A, B) ≤ S(A) + S(B), we
obtain: !
X X
pi S(ρi ) ≤ S pi ρi
i i
which proves the theorem.
However we still have to justify equations (3.23) and (3.24). For the
former we can say that:
S(A) = S(ρA ) = S(TrB (ρAB )) where
!
X
TrB (ρAB ) = TrB pi ρi ⊗ |iihi| =
i
(we choose basis {|ii} to compute the partial trace)
XX XX
= pi ρi hk|iihi|ki = pi ρi δik 2 =
k i k i
X
= pi ρi .
i

For the latter:

S(B) = S(ρB ) = S(TrA (ρAB )) where
!
X
TrA (ρAB ) = TrA pi ρi ⊗ |iihi| =
i

we choose the eigenbasis {|eki0 i} ⊂ HA of a fixed operator ρi0 , with cor-

responding eigenvalues λki0 , to compute the partial trace over system
A:
X
= heki0 |pi ρi |eki0 i|iihi| =
ik
XX X
= heki0 |pi ρi |eki0 i|iihi| + λki0 pi0 |i0 ihi0 |
k i6=i0 k
XX
= heki0 |pi ρi |eki0 i|iihi| + pi0 |i0 ihi0 |
k i6=i0

After iterating this procedure N times (N being the number of the

operators ρi and thus the dimension of the auxiliary system B) and
each time choosing an eigenbasis of a different ρi , we then add all left
sides of the equations obtained as well as the right sides:
N
!
X XX
N TrA (ρAB ) = pi |iihi| + N TrA (ρAB ) − heki0 |pi ρi |eki0 i|iihi|
i k i
N
X
= pi |iihi| + (N − 1)TrA (ρAB )
i

30
 which yields X
TrA (ρAB ) = pi |iihi|
i

that concludes the proof.

Entanglement distillation and dilution. In order to give a physical

interpretation to the entropy of entanglement, we will now define two key
concepts in the context of quantum information: entanglement distillation
and dilution ([3] and [6]). As far as the former is concerned, let us consider
the state (3.14) and let us assume that our goal is to create a maximally
entangled state (i.e. a state in the same form, but with θ = π/4) by acting
locally on each of the particles. In order to do that, we can try to perform
a generalized measurement on the first particle. Hence we choose {A0 , A1 }
as our collection of measurement operators acting on the first particle, with
1/2
A0 = tan(θ)|0ih0| + |1ih1| and A1 = (1 − A0 † A0 ) . Those two operators
fulfill (2.13). In case we measure and obtain the outcome associated with
A0 we will achieve our goal; as a matter of fact, the state of the system after
the measurement is the Bell state Φ+ :
A0 ρA0 †
ρ0 = (recall chapter 2) = = ... = |Φ+ ihΦ+ |.
p0

where ρ = |ΦihΦ| and p0 is our probability of succeeding: p0 = Tr(A0 ρA0 † ) =

... = 2 sin2 (θ) (note that θ ∈ (0, π/4)). On the contrary, if we obtain the
outcome associated with A1 , we have produced a product state. We point
out that, if A and B are spatially separated and are held by Alice and Bob,
respectively, in order to know if the measurement has been successful, the
outcome (i.e. classical bit of information) has to be transmitted from Al-
ice to Bob. One says that by local operations and classical communication
(LOCC) one can distill a maximally entangled state out of the state Φ with
probability p0 . One may wonder if there is another generalized measurement
applied to A and B (individually) giving a higher probability of success. In
fact, this is not the case since the generalized measurement we just chose
is the optimal one. Now one can consider the case in which Alice and Bob
possess two identical copies of the state Φ, and they try to obtain maximally
entangled states by LOCC (in which joint measurements on both qubits of
Alice, or both qubits of Bob, are authorized). In general, they may obtain as
outcome a maximally entangled state in a Hilbert space – of entangled states
– of dimension d = 2, 3, 4. For instance, if they are completely successful,
they will get two copies of a maximally entangled state –that is a Bell state–
which is equivalent to a single copy of such a state in a space of dimension
d = 4. Or if they get a single copy, they will have d = 2 (for example, the
Hilbert spaces {|Φ+ i, |Φ− i} or {|Ψ+ i, |Ψ− i}). Now we can consider what
happens when we take n copies and allow for the optimal LOCC in order to

31
optimize
2n
1X
Ē = pd log2 (d)
n
d=1

where pd is the probability they end up with an entangled state in a space

of dimension d (for d = 1 they end up in a product state). The logarithm is
the right quantity such that n copies of a maximally entangled state (which
corresponds to a dimension dn ) exactly give a factor n. In this case Alice
and Bob try to maximize quantity Ē (note that Ē ≤ 1)) in order to obtain
as much Bell states as possible out of the n copies of state Φ. It turns
out that in the limit n → ∞, the result, called entanglement of distillation,
precisely coincides with the entropy of entanglement E(Φ). This occurs
not only for qubits, but for any d-level systems. Thus, the entanglement
entropy is nothing but the optimal (i.e. maximized) averaged amount of
entanglement (i.e. of Bell states) that we can distill out of Φ by LOCC
in the asymptotic limit where we have a large number of copies. In other
words, the physical purpose of entanglement entropy E(Ψ) is to quantify
the amount of entanglement contained in the pure state |Ψi.
One may consider the opposite process, called entanglement dilution:
given n maximally entangled states, and by applying the optimal LOCC,
how many copies, m, of a state |Ψi we can obtain. The ratio n/m in the
limit n → ∞ is called entanglement of formation and turns out to coincide
again with E(Ψ), giving another physical meaning to the latter.

Concurrence and Fidelity. We finish this section by mentioning other

quantities that are usually employed to quantify entanglement. One is con-
currence, which in the case of pure states reduces to the product of the
Schmidt coefficients. Another one is the fidelity with a maximally entangled
state:
2
F (Ψ) = max |hΦ+ |(UA ⊗ VB )|Ψi|
where the maximization is with respect to the unitary operators U and V .
It measures in a sense how close we are to a maximally entangled state; U
and V just correspond to a basis change.

3.2.2 Multipartite systems

Entanglement in multipartite systems becomes more complicated than in
bipartite ones. First of all, one can have a situation where certain objects
are entangled to others, but not to all of them. Second, the quantification
gets harder, since it is not known if a property like the inter-convertibility
of states by distillation and dilution exists [3].

Definition 2 We say that a state Ψ of systems A, B, ...Z is a product state

32
if there exist |ϕX i ∈ HX such that

|Ψi = |ϕA i ⊗ |ϕB i ⊗ ... ⊗ |ϕZ i

otherwise we say that we have an entangled state. This definition was given
by Schrödinger.

Although a pure state Ψ is entangled, it may happen that some of the

subsystems are still disentangled. In order to characterize the entanglement,
we consider all possible partitions of the subsystems and for each of them
we apply the above definition, with an element of the partition replacing a
single subsystem (note that an element of the partition can contain more
than one subsystem). For instance, for three parties we can have: (i) they
are in a product state; (ii) only A and B are entangled; (iii) only A and C
are entangled; (iv) only B and C are entangled; (v) all are entangled. These
cases are mutually disjoint. An example of case (i) is |0, 0, 0i, of case (ii)
|Φ+ iAB ⊗ |0iC , and of case (iii) the states:

|0, 0, 1i + |0, 1, 0i + |1, 0, 0i

|W i = √
3
|0, 0, 0i + |1, 1, 1i
|GHZi = √
2
These two last examples illustrate the difficulty of quantifying entanglement
in many-body quantum systems; it is not clear which of those states is
“more” entangled. Furthermore, in this case it is not possible to convert the
state |W i into the state |GHZi (or vice versa) by LOCC in the asymptotic
limit without losing copies (equivalently, entanglement), and thus we cannot
assign a quantity like the entanglement entropy to them.
What we can still do is consider bipartite partitions, in which we compute
the entropy of entanglement of two disjoint sets of subsystems. Another
possibility is to look at fidelities with respect to certain particular states.
For instance, one can define the fidelity with respect to a GHZ or to a W
state; or to products of Bell states.
Moreover, along with the problem of quantifying entanglement, there is
also the problem of its detection: while in bipartite systems SD allows us to
immediately detect whether a pure state is entangled, in multipartite ones
SD is typically not available.

3.3 Mixed States

Pure states are hardly representative of the majority of physical situations.
As a matter of fact, a system very soon interacts with a number of systems
so that, even if it was prepared in a pure state, it is typically described by
a mixed state characterized by a certain probability of finding the above

33
mentioned state within the pure state in which it was prepared. Hence
the definition of entanglement has broadened beyond Schrödinger’s original
definition in order to include such mixed states [2].

3.3.1 Bipartite systems

We consider ([6]) again two subsystems, A and B.

Definition 3 We say that ρ represents a product state whenever we can

find ρA , ρB operators acting on HA , HB such that:

ρ = ρA ⊗ ρB

Definition 4 A state ρ is separable if it can be written as a statistical

mixture of product states; that is, if it can be written as a convex combination
of products of subsystem states:
X
ρ= pi ρA i ⊗ ρB i
i

Without loss of generality, we can assume that ρA i and ρB i are all pure
ensembles of the appropriate subsystems, i.e. ρA i = |aiha|i and ρB i = |bihb|i ;
indeed, if they were mixed states, we could make them explicit through (1.16)
obtaining a sum, over 3 indexes, of pure states and we could then reduce such
sum to only one index:
X X
ρ= pi |aiha|i ⊗ |bihb|i = pi |ai bi ihai bi |
i i

Otherwise, we say that ρ represents an entangled state.

Note that while every mixed state can be written as:

X
ρ= pi |Ψi ihΨi |
i

requiring that ρ is separable means that every operator |Ψi ihΨi |, which acts
on HAB , is a product state: |Ψi ihΨi | = ρA i ⊗ ρB i , or equivalently that every
state |Ψi i is a product state |ai i|bi i. However it is not always possible to
tell whether or not a given mixed state is separable; this problem is known
as the separability problem.
What differentiates pure and mixed states is that non-entangled mixed
states may contain correlations, while non-entangled pure states do not. For
example, the mixed separable state:
1 1
ρ = (|0ih0| ⊗ |0ih0| + |1ih1| ⊗ |1ih1|) = (|0, 0ih0, 0| + |1, 1ih1, 1|)
2 2

34
fulfills hσzA ⊗σzB i = Tr(σzA ⊗σzB ρ) = ... = 1 whereas hσzA,B i = Tr(σzA,B ρA,B ) =
... = 0. These correlations are, however, very trivial. If we had a classical
system we could also have them. Only entangled states may display non-
classical correlations.
A subtle point is that a state may look entangled even though it is
separable. Let us take, for instance:
1
σ = (|Φ+ ihΦ+ | + |Φ− ihΦ− |)
2
According to this formula, we can prepare σ by mixing two maximally en-
tangled states. However, there is another way of preparing the same state
σ which does not require using entangled states at all. This immediately
follows from the fact that σ = ρ (one has just to replace the definition of
the Bell states in this formula), and thus one can prepare it by mixing two
product states. Therefore σ is a separable state as well. This simple exam-
ple illustrates the difficulty of finding out whether a state is entangled or
not: we have to check all the possible decompositions and only if none of
them involves exclusively product states, we will have an entangled state.
Unfortunately there exist infinite decompositions, so that this task is hope-
less. Fortunately, in some special cases there are shortcuts, which can give
us the right answer with much less effort. We will now go through them.

Entanglement witnesses. An entanglement witness is an observable

which detects the presence of entanglement. Given an operator W = W † we
say that it is a witness if for all product states |a, bi we have ha, b|W |a, bi ≥ 0,
but W possesses negative eigenvalues. From the definition of separable states
it is clear that if Tr(ρW ) < 0, then ρ must be entangled; as a matter of fact
if ρ is separable, then Tr(ρW ) ≥ 0:
X X
Tr(ρW ) = hkl| pi |ai bi ihai bi |W |kli =
kl i
(being {|kli} an orthonormal basis in HAB )
P (i)
= (writing |ai bi i in such basis: |ai bi i = nm cnm |nmi)
(i) ∗
X X
= pi hkl|c(i)
nm |nmicnm hnm|W |kli =
kli nm
X 2
= |c(i)
nm | pi δkn δlm hnm|W |kli =
klinm
(i) 2
X
= pi |ckl | hkl|W |kli ≥ 0
kli

(i) 2
since pi ≥ 0, |ckl | ≥ 0, hkl|W |kli ≥ 0. Thus, a negative expectation value
of a witness indicates the presence of entanglement. Note, however, that
the converse is not necessarily true: if the expectation value of a witness is

35
positive, this does not imply that the corresponding state is not entangled.
Hence the important question one can ask about entanglement witnesses
regards their optimality. We say that an entanglement witness W1 is finer
than W2 if and only if the entanglement of any ρ detected by W2 is also
detected by W1 . Horodecki showed ([5]) that for any entangled state there
always exists a witness that detects it. Sadly, there is no simple way of find-
ing out such a witness, which makes the problem of detecting entanglement
in mixed states rather non-trivial.
Let’s consider the following example. For a system of two qubits, W =
2 − S, where S = σ1A ⊗ (σ1B + σ2B ) + σ2A ⊗ (σ1B − σ2B ) and the sigma’s are
Pauli operators, is an entanglement witness. This can be shown by noticing
that for a product state: |hSi| ≤ |hσ1B i + hσ2B i| + |hσ1B i − hσ2B i| ≤ 2, since
hσi ≤ 1. By choosing the sigma’s as follows: σ1 = σ~n , σ2 = σm ~ (where σ~ n,m
~
are those in subsection (3.2.1)), we see that for ρ(p) = (1 − p) 14 + p|Φ+ ihΦ+ |,
entanglement is detected for p > √12 .

PPT criterion. A very strong necessary condition for separability has

been proved by Peres, called the positive partial transpose (PPT) criterion
(more details in [5]). It says that if ρAB is separable, then the new operator
TB
ρAB , with matrix elements defined in some fixed product basis as

(ρTAB
B
)mµ,nν = hm|hµ|ρTAB
B
|ni|νi ≡ hm|hν|ρAB |ni|µi

is a density operator (i.e. has a non-negative spectrum), which means that

ρTAB
B
is also a quantum state. Thus, if ρTAB B
has any negative eigenvalue,
then ρAB must necessarily be entangled. It also guarantees the positivity of
ρTAB
A
defined in an analogous way. The operation TB , called a partial trans-
pose, corresponds to transposition of indexes corresponding to the second
subsystem. A fundamental fact pointed out by Horodecki is that the PPT
condition is not only a necessary but also a sufficient condition for separa-
bility of the 2 ⊗ 2 and 2 ⊗ 3 cases. Thus it gives a complete characterization
of separability in those cases.
We will briefly prove the PPT criterion:

if ρAB is separable, then it can be written

X
ρAB = pi ρA i ⊗ ρB i
i

In this case, the effect of the partial transposition is trivial:

X
ρTAB
B
= pi ρA i ⊗ ρB Ti
i

As the transposition map preserves eigenvalues –the eigenvalues

of an operator A are equal to the eigenvalues of its transpose since

36
 they share the same Characteristic polynomial: det(A − λI) =
det (A − λI)T = det(AT − λI)– the spectrum of ρTAB
B
is the same
TB
as the spectrum of ρAB , and in particular ρAB must still be
positive semidefinite, which concludes the demonstration.

Entanglement measures and mutual information. For mixed states,

entanglement measures similar to entanglement entropy can be defined, al-
though they are much harder to evaluate. For instance, we can consider
the distillation procedure as before, but now with mixed states. That is,
we may try to distill out of n copies of a state ρ the maximal number m of
Bell states using LOCC. The ratio m/n in the limit n → ∞ is called entan-
glement of distillation D(ρ). Analogously, we may consider the process of
entanglement dilution and define the entanglement cost Ec (ρ). In general,
D(ρ) < Ec (ρ) so that we cannot first distill and then get back the same state
as before [6]. Moreover, there exist very few examples where these quantities
can be evaluated. One could wonder why entanglement cannot be evaluated
by the entropy of entanglement –that is the von Neumann entropy of the
reduced density operator – in the case of bipartite mixed states. In order
to answer this question, let’s list the axioms that a satisfactory measure of
entanglement over mixed systems E(ρ) should fulfill ([4]):

Axiom 1 E(ρ) reduces to the entropy of entanglement ES (ρ) in case ρ is a

pure state.

Axiom 2 E(ρ) = 0 for separable states.

Axiom 3 A measure of entanglement is not increasing under LOCC:

E(U ρU −1 ) ≤ E(ρ). Since a measure of entanglement gives a measure about
the amount of quantum correlations between two subsystems, it cannot in-
crease if we modify a subsystem (LO) or if the two subsystems exchange
classical information (CC). Focusing on local operations, since they are im-
plemented by unitary operators, we can express this by writing: E(U ρU † ) ≤
E(ρ).

Axiom 4 (Convexity) Entanglement should not increase by mixing var-

ious density matrices, since a convex combination of density matrices is
a purely classical statistical superposition of states: E(λρ1 + (1 − λ)ρ2 ) ≤
λE(ρ1 ) + (1 − λ)E(ρ2 ).

Axiom 5 Similarly entanglement should not increase if we simply put to-

gether two states describing different subsystems: E(ρ ⊗ σ) ≤ E(ρ) + E(σ).

Comparing these properties with the ones fulfilled by the von Neumann
entropy, one immediately notices that while S(ρ) is concave, E(ρ) is con-
vex. This striking difference is basically the reason for which the entropy

37
of entanglement fails to hold as a satisfactory measure of entanglement over
mixed states.
This statement
P can be better visualized by considering a generic mixed state
ρ: ρ = i pi |Ψi ihΨi | and by evaluating its entanglement through the en-
tropy of entanglement:

ES (ρ) = S(ρA ) where

X X
ρA = TrB (ρ) = pi TrB (|Ψi ihΨi |) ≡ pi ρA i
i i
!
X
⇒ S(ρA ) = S pi ρA i
i

and due to the concavity of the von Neumann entropy:

X
ES (ρ) = S(ρA ) > pi S(ρA i ) (3.25)
i

(equality occurs only when all ρA i are equal). The last inequality is clearly
in contradiction with the axiom of convexity:
!
X X
E(ρ) = E pi |Ψi ihΨi | ≤ pi E(Ψi )
i i

recalling axiom 1: E(Ψi ) = ES (Ψi ) = S(ρA i ), which yields

X
E(ρ) ≤ pi S(ρA i ) (3.26)
i

Inequalities (3.25) and (3.26) show that ES (ρ) 6= E(ρ), hence the entropy
of entanglement can not be taken as a measure of entanglement for mixed
states.

A possible measure of entanglement that satisfies the above axioms and that
can be determined in practice is the entanglement of formation:
X
EF (ρ) = min pi E(Ψi )
i

where E(Ψ) is the entropy of entanglement and the minimization is done

with respect to all decompositions of ρ. This quantity is related to the
entanglement cost through EF (ρ⊗n )/n → Ec (ρ) in the limit n → ∞. Note
that in the pure state case, E(Ψ) = EF (Ψ) = Ec (Ψ) = D(Ψ).
Another way of measuring entanglement is through the fidelity with a
maximally entangled state, as the one already defined, but now with:
X
F (ρ) = maxhΦ+ |(U ⊗ V )ρ(U † ⊗ V † )|Φ+ i = pi F (Ψi )
i

38
where F (Ψ) is the fidelity previously defined and pi are the coefficients of
the convex combination representing ρ.
Finally, entanglement can also be measured using the definition of partial
transposition. One defines negativity as

N (ρ) = max kρTA k1 − 1, 0

(3.27)

where the 1-norm is given by the sum of the absolute values of the eigen-
values. Negativity can be positive only if we have an entangled state, so it’s
a sufficient condition but it is not necessary: there exist entangled states
for which it is zero. In fact, we are going to prove that N (ρ) > 0 ⇒ ρ is
entangled:

N (ρ) > 0 ⇒ kρTA k1 − 1 > 0 ⇒ kρTA k1 > 1. (3.28)

Now since Tr(ρ) = 1 and
X
Tr(ρTA ) = hm|hµ|ρTA |mi|µi =
mµ
X
= hm|hµ|ρ|mi|µi =
mµ
= Tr(ρ)

We have that Tr(ρTA ) = 1 ⇒ k λk = P 1 (being ρTA |ki = λk |ki).

P
Now considering (3.28) we have that k |λk | > 1, which com-
pared with the previous equation means that there exists at least
one k for which λk < 0, thus ρTA has at least a negative eigen-
value. Applying the PPT criterion we conclude that ρ is an
entangled state.

Another quantity of interest is the already known quantum mutual informa-

tion S(A : B) = S(A) + S(B) − S(A, B). This does not measure entangle-
ment, but rather correlations. In fact, it is the finest measure of correlations,
in the sense that it detects them even when correlation functions do not [6].
It satisfies the following properties:

1. S(A : B) ≥ 0; S(A : B) = 0 ⇔ ρ = ρA ⊗ ρB

2. S(A : B) ≤ S((aA) : B) ≤ S(A : B) + 2S(a)

The first indicates that it is zero only for product states, i.e. when there are
no correlations. The second indicates that it decreases whenever we discard
a subsystem (in this case a), but it cannot decrease by more than twice the
entropy of such a system.
Proof

39
1. At first let’s show that there are no correlations for product states
such as ρ = ρA ⊗ ρB :
X
hA ⊗ Bi = Tr(ρA ⊗ B) = hkl|ρA ⊗ B|ki|li =
kl
X
= hkl|ρA ⊗ ρB |A|kiB|li =
kl
X
= hkl|ρA A|kiρB B|li =
kl
X
= hk|ρA A|kihl|ρB B|li =
kl
= Tr(ρA A)Tr(ρB B) =
obviously for product states the reduced density matrix 3 for system
A,B is ρA,B , resulting in:
hA ⊗ Bi = Tr(ρA A)Tr(ρB B) = hAihBi (3.29)
which confirms the absence of correlations.
In order to prove the property we will use the so called strong sub-
additivity of the von Neumann entropy: given three subsystems X, Y, Z
then
S(X, Y ) + S(X, Z) ≥ S(X, Y, Z) + S(X)
If we choose Y = A, Z = B and X a system which is pure (S(X) = 0)
and disentangled from the others (i.e. XY and XZ are product states)
we immediately obtain the first property, since S(X, Y ) = S(ρX ⊗
ρY ) = recall the von Neumann property number 4 = S(X) + S(Y ) =
S(Y ) and S(X, Z) = S(ρX ⊗ρZ ) = S(X)+S(Z) = S(Z). The equality
holds when ρ = ρA ⊗ ρB : S(A, B) = S(ρA ⊗ ρB ) = S(ρA ) + S(ρB ) =
S(A) + S(B).
2. By choosing X = A, Y = B and Z such that the whole state Ψ is
pure and the reduced state in AB is our state ρ, we obtain S(A, B) ≥
S(A)−S(A, Z); using property 1 of the von Neumann entropy we have
S(A, Z) = S(B) (being ABZ a pure state), which yields: S(A, B) ≥
S(A) − S(B).
We prove S(A : B) ≤ S((AZ) : B), which is equivalent to S(A) +
S(B) − S(A, B) ≤ S(A, Z) + S(B) − S(Z, A, B), by simply choosing
X = A, Y = B in the strong sub-additivity inequality.
We prove S((AZ) : B) ≤ S(A, B)+2S(Z) making it explicit: S(A, Z)+
S(B) − S(A, Z, B) ≤ S(A) + S(B) − S(A, B) + 2S(Z) and using in-
equalities: S(A, Z, B) ≥ S(A, B) − S(Z) (obtained from S(A, B) ≥
S(A) − S(B)) and S(Z) + S(A) ≥ S(A, Z) (obtained from property 6
of the von Neumann entropy).
3
TrB (ρ) = TrB (ρA ⊗ ρB ) = ρA TrB (ρB ) = ρA

40
Moreover, quantum mutual information for pure states Ψ reduces to the
entropy of entanglement: S(A, B) = 0 and S(A) = S(B), which yield:
S(A : B) = 2S(A) = 2E(Ψ).

3.3.2 Multipartite systems

The description of multipartite entanglement must still confront several chal-
lenges for mixed states. On the one hand, the definitions of product and
separable states are straightforward:

Definition 5 The state ρA1 ...Am of m subsystems A1 , . . . , Am is a product

state whenever we can find ρA1 , . . . , ρAm operators acting on HA1 , . . . , HAm
such that:
ρA1 ...Am = ρA1 ⊗ ... ⊗ ρAm

Definition 6 The state ρA1 ...Am of m subsystems A1 , . . . , Am is completely

separable if and only if it can be written in the form:
X
ρA1 ...Am = pi ρA1 i ⊗ . . . ⊗ ρAm i
i

On the other hand we have to consider again different partitions in order to

characterize entanglement, since it may happen that some of the subsystems
are disentangled, even if the whole state is entangled.

Definition 7 The state ρA1 ...Am of m subsystems A1 , . . . , Am is entangled

with respect to a given partition {I1 , . . . , Ik }, where Ii are disjoint subsets
of the indices I = {1, . . . , m}, ∪kj=1 Ij = I, if and only if it can not be written
in the form: X
ρA1 ...Am = pi ρ1 i ⊗ . . . ⊗ ρk i
i

We end up with a table in which for each partition we declare whether the
state is entangled or not. In order to check the entanglement for each parti-
tion, we have to, for example, find the appropriate witness, whose definition
follows very naturally the one for bipartite systems. Such a table may con-
tain some redundancies since, for instance, if a tripartite state is completely
separable, i.e. separable with respect to the partition (A)(B)(C), it is auto-
matically separable for any other partition, and thus not entangled at all.
This is obviously a general fact: if a multipartite state is completely
separable, then it is not entangled with respect to any possible
partition.

41
3.4 An interesting example
In this section we will show that the von Neumann entropy and the state
function entropy known in thermodynamics have something in common. In
particular, we will compute ([8, page 188]) the von Neumann entropy of an
ensemble in thermal equilibrium and show ([4]) that it coincides with the
above-mentioned thermodynamic entropy.
First we obtain the density operator ρ describing the ensemble in thermal
equilibrium. The basic assumption we make is that nature tends to maximize
the von Neumann entropy of the ensemble S(ρ), subject to the constraint
that the ensemble average of the Hamiltonian has a certain prescribed value,
since it is in equilibrium. Once thermal equilibrium is established, we expect:
dρ
=0 (3.30)
dt
otherwise hHi = Tr(ρH) would be time-dependent, in contradiction with
the hypothesis of equilibrium. Because of (1.27), this means that ρ and
H can be simultaneously
P diagonalized. Therefore, the eigenstates |ki used
to trace: S(ρ) = − k xk log xk , which are ρ’s eigenstates (and xk are its
eigenvalues), are chosen to be also energy eigenstates. With this choice, the
fractional population of the state |ki expresses the probability of finding a
mixture element with energy Ek .
Thus we are trying to maximize S(ρ), taking into account the constraints:

• The ensemble average of H has, at equilibrium, a prescribed value U :

hHi = Tr(ρH) = U , which yields:
X
Tr(ρH) − U = xk Ek − U = 0 (3.31)
k

• For any density operator its eigenvalues sum to one:

X
xk − 1 = 0 (3.32)
k

We can most readily accomplish this by using Lagrange multipliers, where

S(ρ) is a function of N variables xk and the two constraints (3.31) , (3.32)
introduce two Lagrange multipliers, which we call λ1 = β and λ2 = γ. We
obtain:
log xk + 1 + βEk + γ = 0 ⇒ xk = e−βEk −γ−1
The constant γ can be eliminated using (3.32):

X N
X
−βEk −γ−1
e =1 ⇒Z≡ e−βEk = eγ+1
k k

42
where we recognize the partition function Z, that can also be written Z =
Tr(e−βH ); in conclusion
e−βEk
xk = (3.33)
Z
where the constant β can also be determined in terms of U through the first
constraint.
Had we attempted to maximize S(ρ) without the internal-energy con-
straint, we would have obtained instead:
1
xk = (3.34)
N
which is the density matrix eigenvalue appropriate for a completely random
ensemble, as we stated in section 1.7. Comparing (3.33) with (3.34), we
infer that a completely random ensemble can be regarded as the β → 0
limit (physically the high-temperature limit) of a canonical ensemble.
Now, since ρ and H are simultaneously diagonalized with respect to basis
{|ki}, one can write ρ through its spectral decomposition:
X X e−βEk
ρ = xk |kihk| = |kihk| =
Z
k k
X e−βH e−βH X
= |kihk| = |kihk| =
Z Z
k k
e−βH
= (3.35)
Z
The last expression for ρ, with β = 1/kB T , is known in statistical mechanics
to describe a canonical ensemble.

Now it’s time to compute the von Neumann entropy of the canonical ensem-
ble: !
e−H/kB T e−H/kB T
S(ρcan ) = −Tr log
Z Z
−H/k T
where log e Z B = log e−H/kB T − log Z and log e−H/kB T = −H/kB T as far
as the exponent is a Hermitian operator. Thus:
1
S(ρcan ) = Tr(ρcan H) + log ZTr(ρcan ) =
kB T
recalling that Helmholtz free energy F = −kB T log Z
hHi − F
=
kB T
recalling F ’s definition: F = U − T Stherm and that hHi = U
= Stherm /kB
We have shown that, up to a factor, the von Neumann entropy of a canonical
ensemble is the state function entropy in the thermodynamic sense!

43
3.5 Monogamy and Area Law
In this section we will present, from a qualitative point of view, two interest-
ing properties fulfilled by many-body quantum systems: the area law ([6])
and the monogamy of entanglement ([4]).

The area law seems to be satisfied by all systems in thermal equilibrium

which interact with short range interactions in lattices.
For instance, we consider a spin system on a lattice in d spatial dimensions.
The system at temperature T is described by the density operator:

e−H/kB T
ρ=
Z
where the Hamiltonian H expresses the short range interaction involving
only a few spins close to each other. The area law can be obtained both at
zero and finite temperature.
In the first case, the state ρ reduces to a projector onto the ground
subspace, whose vector basis {|Ψi i} fulfills H|Ψi i = E0 |Ψi i where E0 is the
ground state energy. If the ground state is not degenerate, which we will
assume, we will just use a pure state |Ψ0 i to denote the ground state. We
consider a connected region A of the lattice with a smooth boundary and the
complementary region B, and concentrate on the amount of entanglement
between these two regions, and its growth as we make region A larger and
larger. Since we are dealing with a bipartite system of a pure state |Ψ0 i,
we can evaluate such amount through the entropy of entanglement E(Ψ0 ) =
S(ρA ). In general, since entropy is an extensive quantity, one would expect
that it scales with the number of spins in region A. However, for ground
states of Hamiltonians as we are considering here, this seems not to be the
case. Instead, the entanglement (entropy of entanglement) scales not with
the volume of region A, but with its boundary area, and thus the name area
law.
In the second case, the one with finite temperature, ρ is in general a
bipartite mixed state, hence its amount of entanglement cannot be mea-
sured by the entropy of entanglement (see discussion in subsection 3.3.1).
Nevertheless, one can still derive an analogous area law by measuring the
amount of correlations between regions A and B through the quantum mu-
tual information S(A : B). Note that at zero temperature, i.e. pure states,
the quantum mutual information reduces (up to a factor two) to the entan-
glement entropy, providing us with a suitable generalization of the previous
area law.

Monogamy is one of the most fundamental properties of entanglement and

further differentiates the notion of entanglement from classical correlations.
In many-body systems with local interactions, classical correlations typically

44
decay with distance, but there’s no upper bound to the total number of
correlations a single particle can establish with the others of the system. In-
stead, in 2006 Osborne and Verstraete proved the Coffman-Kundu-Wootters
(CKW) monogamy inequality, which states that given a multipartite system
with subsystems (qubits) A, B1 , B2 , . . . then it results:

C 2 (A|B1 ) + C 2 (A|B2 ) + . . . ≤ 4∆A ≤ 1 (3.36)

where C 2 (A|Bk ) is the concurrence of the bipartite mixed system of qubits

A|Bk (concurrence can also be defined for mixed states, and it still represents
a measure of entanglement as in the pure case); ∆A is the determinant of
the reduced density matrix ρA and it quantifies the entanglement between
A and the rest of the system.
It appears clearly that it is not possible for A to be strongly entangled
with several subsystems B. If, for instance, A is strongly entangled with B1
then necessarily it can not be equally entangled, nor entangled at all, with
the other subsystems. This is the reason why it is said that the entanglement
bond is monogamous and why it differs from what happens in classical
correlations.

45
Chapter 4

Conclusions

In this paper we have briefly given insight into the world of entanglement
and entanglement entropy.
In the first chapter we introduced the notion of density operator and ana-
lyzed its properties in order to show, in the following chapter, how Quantum
Mechanics could be reformulated in the density operator picture. As we
highlighted, this new formulation is equivalent to the most-known one; how-
ever, the density operator approach really shines for two applications: the
description of quantum systems whose state is not known, and the descrip-
tion of subsystems of a composite quantum system. Since the entanglement
of a system arises as a quantum correlation between two or more subsystems,
the density operator picture is fundamental in its description.
In the third chapter we presented the basic concepts regarding entangle-
ment of bipartite and multipartite systems, both for pure and mixed states.
As we have argued, whereas entanglement for bipartite systems is well es-
tablished, for multipartite ones there exist many possibilities of defining
entanglement measures. It may turn out that one state is more entangled
according to one measure, but not according to a different one.
Nevertheless, we have highlighted two measures of entanglement/correlations
which have a clear physical meaning. The first one, the entropy of entan-
glement, applies to bipartite pure states and is given by the von Neumann
entropy of the reduced density operator of one of the subsystems. The
second one, the quantum mutual information, measures the correlations be-
tween the two subsystems and for pure states it reduces to the entropy of
entanglement. One can apply these two measures to multi-spin systems in
lattices by separating all the spins into two disjoint regions which are then
considered as a bipartite system. In the particular case where we deal with
ground or thermal states of short range interaction Hamiltonians, the appli-
cation of those measurements gives rise to area laws. These laws state that
the quantum mutual information, between a region A and its complemen-
tary, scales with the number of spins at the boundary of A (and not, as one

46
would expect, with the total number of spins in A).
Another peculiar characteristic of entanglement that distinguishes it from
classical correlations is its monogamy, according to which each subsystem
can be strongly entangled only with one other subsystem, as there is an
upper bound to the sum of its quantum correlations.
Entanglement and its quantification have been primary issues in the
study of black holes by means of quantum field theory and in the science
of quantum information, for which entanglement is an important resource.
Since these two subjects naturally overlap in the context of many-body
systems (in particular close to criticality), it soon became apparent that
there is a rich structure to be uncovered by studying entanglement entropy
in this context, and by putting together ideas from field theory and quantum
information science; much has been done, much more is yet to come.

47
Appendices

48
Appendix A

The generic spin- 12 state

The most general spin state of a 12 -spin particle is represented, in the eigen-
basis of operator Sz , by the spinor:
 
a
χ= = aχ+ + bχ− (A.1)
b

where a and b are two arbitrary complex numbers. Requiring the normal-
ization to one, a and b have to satisfy the relation: |a|2 + |b|2 = 1. The
remarkable fact is that spinor (A.1) represents a state in which the spin is
parallel to a precise direction, defined by the versor n̂ = n̂(ϑ, ϕ), where ϑ
and ϕ are obtained as follows:

ϑ = 2 arccos |a| = 2 arccos |b| ϕ = arg b − arg a (A.2)

In order to verify the validity of this assertion, let’s start by observing that
it’s not very meaningful to talk about a precise direction defined by the
two coefficients a and b, since they are probability amplitudes. What we
can actually show is that the spinor χ is an eigenvector of the operator
σn = ~σ · n̂ with corresponding eigenvalue +1. This means that measuring σn
(i.e. the spin) along direction n̂ yields the result +1 with √ 100% probability
of success; however, being the modulus of the spin vector 31 , we can only
conclude that the spinor is situated on a cone with axis n̂, and not precisely
along n̂. Consequently it is not χ’s direction that is fixed, but rather the
direction on which the projection of the spin vector has the fixed value +1.
We could interpret n̂ as the average direction of spinor χ.

Let’s determine the eigenvector of σn and show that it coincides with (A.1).
Given the following components for versor n̂:

nx = sin ϑ cos ϕ, ny = sin ϑ sin ϕ, nz = cos ϑ

1
note that σ 2 = 3I

49
and using the explicit form of Pauli’s matrices, it results:
cos ϑ sin ϑe−iϕ
 
σn =
sin ϑeiϕ − cos ϑ
Note that σn 2 = I, hence according to Dirac’s theorem the operator σn has
eigenvalues λ = ±1. This means that measuring the spin of a 12 -particle
along an arbitrary direction n̂, the only possible results are +h̄/2 and −h̄/2.
We define the eigenspinor associated to the eigenvalue λ = 1 as:
 
+ a
χn = (A.3)
b
and impose that:
cos ϑ sin ϑe−iϕ
    
a a
=
sin ϑeiϕ − cos ϑ b b
which implies the following equalities:
a cos ϑ + b sin ϑe−iϕ = a
a sin ϑeiϕ − b cos ϑ = b
and therefore:
sin ϑ iϕ sin(ϑ/2) iϕ
b=a e =a e (A.4)
1 + cos ϑ cos(ϑ/2)
This, substituted in the normalization condition yields:
|a|2 (1 + tan2 (ϑ/2)) = 1
and thus:
|a| = cos(ϑ/2) (A.5)
We then can choose a = cos(ϑ/2) which implies b = sin(ϑ/2)eiϕ thanks
to (A.4). In this way, up to an arbitrary phase factor, the spinor (A.3) is
completely determined:
χ+ iϕ
n = cos(ϑ/2)χ+ + sin(ϑ/2)e χ−

and satisfies (A.2), which ends the proof.

The fact that any spinor like (A.1) represents a spin state with a definite
polarization has an obvious explanation: the spinor is identified by two
complex numbers, or equivalently four real numbers, but only two of these
are essential in defining the state described by χ. One is eliminated by
the normalization condition, and another is incorporated in an arbitrary
phase factor and doesn’t play any role. The two remaining parameters can
always be put in biunivocal correspondence with two polar angles ϑ and
ϕ, associating to the state a definite polarization direction identified by the
versor n̂ = n̂(ϑ, ϕ) ([1] for further reading).

50
Appendix B

Density Matrix in the

Position operator eigenbasis

A significant representation ([1]) of the density operator ρ is the one given

in the eigenbasis of the position operator x̂: {δ(x − x0 )} with correspond-
ing eigenvalues {x0 } (supposing that scalar products are computed in the
Coordinate space). For simplicity, we will indicate |xi the eigenvector of
the position operator corresponding to eigenvalue x. It is clear that given
a pure state |s(i) i, the scalar product: hx|s(i) i = ψi (x) outputs the spatial
component of that pure state evaluated at x. Given the density operator:
X
ρ= pi |s(i) ihs(i) |
i

its representation with respect to the eigenbasis {|xi} is:

X
ρxx0 ≡ ρ(x, x0 ) = pi hx|s(i) ihs(i) |x0 i =
i
X
= pi ψi (x)ψi∗ (x0 )
i

in particular, for diagonal elements we have:

X
ρ(x, x) = pi |ψi (x)|2 (B.1)
i

Diagonal elements coincide with the weighted average of the probability

densities, this confirms once again the physical meaning of the fractional
populations: they express the probability density of finding a generic element
of the mixture in the state |xi; in this particular case, the probability density
of finding such generic element in the position x.

51
Appendix C

Shannon and von Neumann

Entropy

An entangled pure state is related to mixed subsystems, whose density ma-

trices can be written as classical superpositions (1.16) of density matrices of
pure states, with real non-negative statistical weights (the ones defined at
the beginning of section 1.4):
X
ρ (density operator of a mixed subsystem) = pk ρk (C.1)
k
where X
ρk = |kihk| pk = 1 0 ≤ pk ≤ 1 ∀k (C.2)
k
These classical weights define a classical probability distribution, which as-
sociates to each state ρk its classical probability pk . This classical nature is
evident in the expectation value of an operator O which acts on the subsys-
tem considered:
X
hOi = Tr(ρO) = pk Tr(ρk O) = (ρk = |kihk|)
k
X
= pk Tr(|kihk|O) = (Tr(|kihk|O) = hk|O|ki, see (1.9))
k
X
= pk hOik
k
where the pk s play the role of classical weights, since they quantify the weight
of each state |ki in the mean value of O.
To each classical probability distribution X = {p1 , p2 , ...pn } we can al-
ways associate a function called Shannon Entropy, which measures how
much a distribution is not deterministic, i.e. the amount of uncertainty
about X before we learn its value. Shannon Entropy is defined as:
X
H(X) = H(p1 , p2 , ...pn ) = − pk ln pk (C.3)
k

52
(in quantum information theory the natural logarithm is substituted with
log2 ). For instance, a constant distribution maximizes H(X) since each
outcome is equally probable, while a delta distribution pk = δk,k0 minimizes
H(X) since there’s only one possible issue for a measure. An event that can
never occur should not contribute to the entropy, so by convention we agree
that 0 ln 0 = 0. We could immediately use Shannon Entropy to quantify the
uncertainty about the distribution of pure states composing the mixed state
ρ, that is, the “amount of mixedness” of such state, hence making Shan-
non Entropy a measure of the entanglement of the whole system of which ρ
represents a subsystem. However, since the same operator ρ can be written
in many different ways –due to the non-uniqueness of the decomposition
of mixed states as convex combinations of pure states– we would face the
problem of deciding which decomposition provides the classical probability
distribution employed in computing the Shannon Entropy of the selected
mixed state. This problem is solved by defining a generalization of Shan-
non Entropy that is suitable for quantum systems, with density operators
replacing probability distributions, the von Neumann Entropy:

S(ρ) = −Tr(ρ ln ρ)

(in quantum information theory the natural logarithm is substituted with

log2 ). Von Neumann Entropy is the quantum counterpart of Shannon En-
tropy, in the sense that it measures the mixedness of the state ρ on which
it is applied; as a matter of fact, it coincides with Shannon Entropy when
the latter is computed on the probability distribution related to ρ’s spectral
decomposition:
X
ρ= dk 2 |kihk| (spectral decomposition)
k

where the eigenvalues are the square of the Schmidt coefficients;

X = {d1 2 , d2 2 , ...dn 2 }
X
H(X) = − dk 2 ln dk 2
k
S(ρ) = −Tr(ρ ln ρ) = (choosing for ρ the diagonalized matrix: δkj dk 2 )
X
= − dk 2 ln dk 2
k
⇒ S(ρ) = H(X)

For instance, the completely mixed density operator in a d-dimensional space

defined in section 1.7, I/d, has entropy log d.
For a more complete study of the properties of the von Neumann entropy,
we refer the reader to [3, chapter 11].

53
Bibliography

[1] Cesare Rossetti. Rudimenti di Meccanica Quantistica. Seconda Edi-

zione. Levrotto & Bella Libreria Editrice Universitaria, 2011.

[2] Gregg Jaeger. Entanglement, Information, and the Interpretation of

Quantum Mechanics. Springer, 2009.

[3] M. A. Nielsen and I. L. Chuang. Quantum Computation and Quantum

Information. Cambridge University Press, 2010.

[4] C. Degli Esposti Boschi. Introduzione alla Teoria dell’Informazione

Quantistica. Lecture Notes, 2014.
http://www.bo.imm.cnr.it/users/degliesposti/TIQ.pdf.

[5] R. Horodecki, P. Horodecki, M. Horodecki, K. Horodecki. Quantum

entanglement. Rev. Mod. Phys., Vol. 81, No. 2, April–June 2009.

[6] J. Ignacio Cirac. Entanglement in many-body quantum systems.

arXiv:1205.3742v1 [quant-ph] (16 May 2012).

[7] P. Calabrese, J. Cardy, B. Doyon. Entanglement entropy in extended

quantum systems. (2009) J. Phys. A: Math. Theor. 42 500301

[8] Sakurai, Napolitano. Modern Quantum Mechanics. Second Edition.

Addison-Wesley, 2011.

54