arXiv:1805.

11965v4 [hep-th] 7 Dec 2018

A Mini-Introduction To Information Theory

Edward Witten

School of Natural Sciences, Institute for Advanced Study

Einstein Drive, Princeton, NJ 08540 USA

Abstract

This article consists of a very short introduction to classical and quantum information theory.
Basic properties of the classical Shannon entropy and the quantum von Neumann entropy are
described, along with related concepts such as classical and quantum relative entropy, conditional
entropy, and mutual information. A few more detailed topics are considered in the quantum case.
Contents

1 Introduction 2

2 Classical Information Theory 2

2.1 Shannon Entropy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2

2.2 Conditional Entropy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4

2.3 Relative Entropy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6

2.4 Monotonicity of Relative Entropy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7

3 Quantum Information Theory: Basic Ingredients 10

3.1 Density Matrices . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10

3.2 Quantum Entropy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14

3.3 Concavity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16

3.4 Conditional and Relative Quantum Entropy . . . . . . . . . . . . . . . . . . . . . . . . 17

3.5 Monotonicity of Relative Entropy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20

3.6 Generalized Measurements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22

3.7 Quantum Channels . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23

3.8 Thermodynamics And Quantum Channels . . . . . . . . . . . . . . . . . . . . . . . . . 26

4 More On Quantum Information Theory 27

4.1 Quantum Teleportation and Conditional Entropy . . . . . . . . . . . . . . . . . . . . . 28

4.2 Quantum Relative Entropy And Hypothesis Testing . . . . . . . . . . . . . . . . . . . . 32

4.3 Encoding Classical Information In A Quantum State . . . . . . . . . . . . . . . . . . . 36

1
1 Introduction

This article is intended as a very short introduction to basic aspects of classical and quantum information
theory.1

Section 2 contains a very short introduction to classical information theory, focusing on the definition
of Shannon entropy and related concepts such as conditional entropy, relative entropy, and mutual
information. Section 3 describes the corresponding quantum concepts – the von Neumann entropy and
the quantum conditional entropy, relative entropy, and mutual information. Section 4 is devoted to
some more detailed topics in the quantum case, chosen to explore the extent to which the quantum
concepts match the intuition that their names suggest.

There is much more to say about classical and quantum information theory than can be found here.
There are several excellent introductory books, for example [1–3]. Another excellent place to start is
the lecture notes [4], especially chapter 10.

2 Classical Information Theory

2.1 Shannon Entropy

We begin with a basic introduction to classical information theory. Suppose that one receives a message
that consists of a string of symbols a or b, say

aababbaaaab · · · (2.1)

And let us suppose that a occurs with probability p, and b with probability 1 − p. How many bits of
information can one extract from a long message of this kind, say with N letters?

For large N, the message will consist very nearly of pN occurrences of a and (1 − p)N occurrences
of b. The number of such messages is

N! NN
∼
(pN)!((1 − p)N)! (pN)pN ((1 − p)N)(1−p)N
1
= pN = 2N S (2.2)
p (1 − p)(1−p)N
1
The article is based on a lecture at the 2018 summer program Prospects in Theoretical Physics at the Institute for
Advanced Study.

2
where S is the Shannon entropy per letter [5]

S = −p log p − (1 − p) log(1 − p). (2.3)

(In information theory, one usually measures entropy in bits and uses logarithms in base 2.)

The total number of messages of length N, given our knowledge of the relative probability of letters
a and b, is roughly
2N S (2.4)
and so the number of bits of information one gains in actually observing such a message is
NS. (2.5)
This is an asymptotic formula for large S, since we used only the leading term in Stirling’s formula to
estimate the number of possible messages, and we ignored fluctuations in the frequencies of the letters.

Suppose more generally that the message is taken from an alphabet with k letters a1 , a2 , · · · , ak ,
where the probability to observe ai is pi , for i = 1, · · · , k. We write A for this probability distribution.
In a long message with N ≫ 1 letters, the symbol ai will occur approximately Npi times, and the
number of such messages is asymptotically
N! NN
∼ Qk = 2 N SA (2.6)
(p1 N)!(p2 N)! · · · (pk N)! i=1 (pi N)
p i N

where now the entropy per letter is

k
X
SA = − pi log pi . (2.7)
i=1

This is the general definition of the Shannon entropy of a probability distribution for a random
variable A that takes values a1 , . . . , ak with probabilities p1 , . . . , pk . The number of bits of information
that one can extract from a message with N symbols is again
NSA . (2.8)
From the derivation, since the number 2N SA of possible messages is certainly at least 1, we have
SA ≥ 0 (2.9)
for any probability distribution. To get SA = 0, there has to be only 1 possible message, meaning that
one of the letters has probability 1 and the others have probability 0. The maximum possible entropy,
for an alphabet with k letters, occurs if the pi are all 1/k and is
k
X
SA = − (1/k) log(1/k) = log k. (2.10)
i=1

3
 P
The reader can provePthis by using the method of Lagrange multipliers to maximize SA = − i pi log pi
with the constraint i pi = 1.

In engineering applications, NSA is the number of bits to which a message with N letters can be
compressed. In such applications, the message is typically not really random but contains information
that one wishes to convey. However, in “lossless encoding,” the encoding program does not understand
the message and treats it as random. It is easy to imagine a situation in which one can make a better
model by incorporating short range correlations between the letters. (For instance, the “letters” might
be words in a message in the English language; then English grammar and syntax would dictate short
range correlations.) A model incorporating such correlations would be a 1-dimensional classical spin
chain of some kind with short range interactions. Estimating the entropy of a long message of N letters
would be a problem in classical statistical mechanics. But in the ideal gas limit, in which we ignore
correlations, the entropy of a long message is just NS where S is the entropy of a message consisting
of only one letter.

Even in the ideal gas model, we are making statements that are only natural in the limit of large N.
To formalize the analogy with statistical mechanics, one could introduce a classical Hamiltonian H whose
value for the ith symbol ai is − log pi , so that the probability of the ith symbol in the thermodynamic
ensemble is 2−H(ai ) = pi . Notice then that in estimating the number of possible messages for large
N, we ignored the difference between the canonical ensemble (defined by probabilities 2−H ) and the
microcanonical ensemble (in which one specifies the precise numbers of occurrences of different letters).
As is usual in statistical mechanics, the different ensembles are equivalent for large N. The equivalence
between the different ensembles is important in classical and quantum information theory.

2.2 Conditional Entropy

Now let us consider the following situation. Alice is trying to communicate with Bob, and she sends a
message that consists of many letters, each being an instance of a random variable2 X whose possible
values are x1 , · · · , xk . She sends the message over a noisy telephone connection, and what Bob receives
is many copies of a random variable Y , drawn from an alphabet with letters y1 , · · · , yr . (Bob might
confuse some of Alice’s letters and misunderstand others.) How many bits of information does Bob gain
after Alice has transmitted a message with N letters?

To analyze this, let us suppose that PX,Y (xi , yj ) is the probability that, in a given occurrence, Alice
sends X = xi and Bob hears Y = yj . The probability that Bob hears Y = yj , summing over all choices
2
Generically, a random variable will be denoted X, P Y, Z, etc. The probability to observe X = x is denoted PX (x), so
if xi , i = 1, · · · , n are the possible values of X, then i PX (xi ) = 1. Similarly, if X, Y are two random variables, the
probability to observe X = x, Y = y will be denoted PX,Y (x, y).

4
of what Alice intended, is X
PY (yj ) = PX,Y (xi , yj ). (2.11)
i

If Bob does hear Y = yj , his estimate of the probability that Alice sent xi is the conditional proba-
bility
PX,Y (xi , yj )
PX|Y (xi |yj ) = . (2.12)
PY (yj )
From Bob’s point of view, once he has heard Y = yj , his estimate of the remaining entropy in Alice’s
signal is the Shannon entropy of the conditional probability distribution. This is
X
SX|Y =yj = − PX|Y (xi |yj ) log(PX|Y (xi |yj )). (2.13)
i

Averaging over all possible values of Y , the average remaining entropy, once Bob has heard Y , is
X X X PX,Y (xi , yj )  
PX,Y (xi , yj )
PY (yj )SX|Y =yj = − PY (yj ) log
j j i
PY (yj ) PY (yj )
X X
=− PX,Y (xi , yj ) log PX,Y (xi , yj ) + PX,Y (xi , yj ) log PY (yj )
i,j i,j

= SXY − SY . (2.14)

Here SXY is the entropy of the joint distribution

P PX,Y (xi , yj ) for the pair X, Y and SY is the entropy of
the probability distribution PY (yj ) = i PX,Y (xi , yj ) for Y only.

The left hand side of eqn. (2.14), which as we see equals SXY − SY , is called the conditional
entropy SX|Y or S(X|Y ); it is the entropy that remains in the probability distribution X once Y is
known. Since it was obtained as a sum of ordinary entropies SX|Y =yj with positive coefficients, it is
clearly positive:
SXY − SY ≥ 0. (2.15)
(The analogous statement is not true quantum mechanically!) Since SX is the total information content
in Alice’s message, and SXY − SY is the information content that Bob still does not have after observing
Y , it follows that the information about X that Bob does gain when he receives Y is the difference or

I(X; Y ) = SX − SXY + SY . (2.16)

Here I(X; Y ) is called the mutual information between X and Y . It measures how much we learn
about X by observing Y .

This interpretation convinces us that I(X; Y ) must be nonnegative. One can prove this directly but
instead I want to deduce it from the properties of one more quantity, the relative entropy. This will
complete our cast of characters.

5
2.3 Relative Entropy

One can motivate the definition of relative entropy as follows. Suppose that we are observing a random
variable X, for example the final state in the decays of a radioactive nucleus. We have a theory that
predicts a probability distribution QX for the final state, say the prediction is that the probability to
observe final state X = xi , where i runs over a set of possible outcomes {1, 2, · · · s}, is qi = QX (xi ). But
maybe our theory is wrong and the decay is actually described by some different probability distribution
PX , such that the probability of X = xi is pi = PX (xi ). After observing the decays of N atoms, how
sure could we be that the initial hypothesis is wrong?

If the correct probability distribution is PX , then after observing N decays, we will see outcome xi
approximately pi N times. Believing QX to be the correct distribution, we will judge the probability of
what we have seen to be3 s
Y N!
P= qipi N Qs . (2.17)
i=1 j=1 (pj N)!

We already calculated that for large N

N! P
Qs ∼ 2−N i pi log pi (2.18)
j=1 (pj N)!
so P
P ∼ 2−N i pi (log pi −log qi )
. (2.19)
This is 2−N S(P ||Q) where the relative entropy (per observation) or Kullback-Liebler divergence is defined
as X
S(PX ||QX ) = pi (log pi − log qi ). (2.20)
i
From the derivation, S(PX ||QX ) is clearly nonnegative, and zero only if PX = QX , that is if the initial
hypothesis is correct. If the initial hypothesis is wrong, we will be sure of this once
NS(PX ||QX ) ≫ 1. (2.21)

The chance of falsely excluding a correct hypothesis, because of a large fluctuation that causes
the data to be more accurately simulated by PX than by QX , decays for large N as 2−N S(PX ||QX ) .
(Later we will more loosely say that the confidence in excluding the wrong hypothesis is controlled by
2−N S(PX ||QX ) .) In this analysis, we have ignored noise in the observations. What we learned earlier
about conditional entropy would give us a start in including the effects of noise.

S(PX ||QX ) is an important measure of the difference between two probability distributions PX and
QX , but notice that it is asymmetric in PX and QX . We broke the symmetry by assuming that QX was
our initial hypothesis and PX was the correct answer.
Qs pi N
Here Qs N(p! j N )! is the number of sequences in which outcome xi occurs pi N times, and
3
i=1 qi is the probability
j=1
of any specific such sequence, assuming that the initial hypothesis QX is correct.

6
 Now we will use positivity of the relative entropy to prove positivity of the mutual information. We
consider a pair of random variables X, Y and we consider two different probability distributions. One,
which we will call PX,Y , is defined by a possibly correlated joint probability distribution

PX,Y (xi , yj ). (2.22)

Given such a joint probability distribution, the separate probability distributions for X and for Y are
obtained by “integrating out” or summing over the other variable:
X X
PX (xi ) = PX,Y (xi , yj ), PY (yj ) = PX,Y (xi , yj ). (2.23)
j i

This is an important operation which will frequently recur. We define a second probability distribution
for X, Y by ignoring the correlations between them:

QX,Y (xi , yj ) = PX (xi )PY (yj ). (2.24)

Now we calculate the relative entropy between these two distributions:

X
S(PX,Y ||QX,Y ) = PX,Y (xi , yj )(log PX,Y (xi , yj ) − log(PX (xi )PY (yj )))
i,j
X
= PX,Y (xi , yj )(log PX,Y (xi , yj ) − log PX (xi ) − log PY (yj ))
i,j

=SX + SY − SXY = I(X; Y ). (2.25)

Thus I(X; Y ) ≥ 0, with equality only if the two distributions are the same, meaning that X and Y
were uncorrelated to begin with.

The property
SX + SY − SXY ≥ 0 (2.26)
is called subadditivity of entropy.

2.4 Monotonicity of Relative Entropy

Now there is one more very important property of relative entropy that I want to explain, and this
will more or less conclude our introduction to classical information theory. Suppose that X and Y
are two random variables. Let PX,Y and QX,Y be two probability distributions, described by functions
PX,Y (xi , yj ) and QX,Y (xi , yj ). If we start with a hypothesis QX,Y for the joint probability, then after
many trials in which we observe X and Y , our confidence that we are wrong (assuming that PX,Y is the

7
correct answer) is determined by S(PX,Y ||QX,Y ). But suppose that we only observe X and not Y . The
reduced distributions PX and QX for X only are described by functions
X X
PX (xi ) = PX,Y (xi , yj ), QX (xi ) = QX,Y (xi , yj ). (2.27)
j j

If we observe X only, then the confidence after many trials that the initial hypothesis is wrong is
controlled by S(PX ||QX ).

It is harder to disprove the initial hypothesis if we observe only X, so

S(PX,Y ||QX,Y ) ≥ S(PX ||QX ). (2.28)

This is called monotonicity of relative entropy.

Concretely, if we observe a sequence xi1 , xi2 , . . . xiN in N trials, then to estimate how unlikely this
is, we will imagine a sequence of y’s that minimizes the unlikelihood of the joint sequence

(xi1 , yi1 ), (xi2 , yi2 ), · · · , (xiN , yiN ). (2.29)

An actual sequence of y’s that we might observe can only be more unlikely than this. So observing Y
as well as X can only increase our estimate of how unlikely the outcome was, given the sequence of the
x’s. Thus, the relative entropy only goes down upon “integrating out” some variables and not observing
them.

Hopefully, the reader has found this explanation compelling, but it is also not difficult to give a
proof in formulas. The inequality S(PX,Y ||QX,Y ) − S(PX ||QX ) ≥ 0 can be written
X     
PX,Y (xi , yj ) PX (xi )
PX,Y (xi , yj ) log − log ≥ 0. (2.30)
i,j
QX,Y (xi , yj ) QX (xi )

Equivalently  
X X PX,Y (xi , yj ) PX,Y (xi , yj )/PX (xi )
PX (xi ) log ≥ 0. (2.31)
i j
PX (xi ) QX,Y (xi , yj )/QX (xi )

The left hand side is a sum of positive terms, since it is

X
PX (xi )S(PY |X=xi ||QY |X=xi ), (2.32)
i

where we define probability distributions PY |X=xi , QY |X=xi conditional on observing X = xi :

PY |X=xi (yj ) = PX,Y (xi , yj )/PX (xi ), QY |X=xi (yj ) = QX,Y (xi , yj )/QX (xi ). (2.33)

8
 So this establishes monotonicity of relative entropy.4 An important special case is strong subad-
ditivity of entropy. For this, we consider three random variables X, Y, Z. The combined system has a
joint probability distribution PX,Y,Z (xi , yj , zk ). Alternatively, we could forget the correlations between
X and Y Z, defining a probability distribution QX,Y,Z for the system XY Z by

QX,Y,Z (xi , yj , zk ) = PX (xi )PY,Z (yj , zk ) (2.34)

where as usual
X X
PX (xi ) = PX,Y,Z (xi , yj , zk ), PY,Z (yj , zk ) = PX,Y,Z (xi , yj , zk ). (2.35)
j,k i

The relative entropy is S(PX,Y,Z ||QX,Y,Z ). But what if we only observe the subsystem XY ? Then we
replace PX,Y,Z and QX,Y,Z by probability distributions PX,Y , QX,Y with
X X
PX,Y (xi , yj ) = PX,Y,Z (xi , yj , zk ), QX,Y (xi , yj ) = QX,Y,Z (xi , yj , zk ) = PX (xi )PY (yj ) (2.36)
k k

and we can define the relative entropy S(PX,Y ||QX,Y ). Monotonicity of relative entropy tells us that

S(PX,Y,Z ||QX,Y,Z ) ≥ S(PX,Y ||QX,Y ). (2.37)

But the relation between relative entropy and mutual information that we discussed a moment ago
gives
S(PX,Y,Z ||QX,Y,Z ) = I(X; Y Z) = SX − SXY Z + SY Z (2.38)
and
S(PX,Y ||QX,Y ) = I(X; Y ) = SX − SXY + SY . (2.39)

So
SX − SXY Z + SY Z ≥ SX − SXY + SY (2.40)
or
SXY + SY Z ≥ SY + SXY Z , (2.41)
which is called strong subadditivity. Remarkably, the same statement turns out to be true in quantum
mechanics, where it is both powerful and surprising.

Equivalently, the comparison of eqns. (2.38) and (2.39) gives

I(X; Y Z) ≥ I(X; Y ), (2.42)

4
What we have described is not the most general statement of monotonicity of relative entropy in classical information
theory. More generally, relative entropy is monotonic under an arbitrary stochastic map. We will not explain this here,
though later we will explain the quantum analog (quantum relative entropy is monotonic in any quantum channel).

9
which is called monotonicity of mutual information. The intuition is that what one learns about a
random variable X by observing both Y and Z is at least as much as one could learn by observing Y
only.

We conclude this mini-introduction to classical information theory with one last remark. We re-
peatedly made use of the ability to define a conditional probability distribution, conditional on some
observation. This has no really close analog in the quantum mechanical case5 and it is something of
a miracle that many of the conclusions nonetheless have quantum mechanical analogs. The greatest
miracle is strong subadditivity of quantum entropy.

3 Quantum Information Theory: Basic Ingredients

3.1 Density Matrices

Now we turn to quantum information theory. Quantum mechanics always deals with probabilities, but
the real quantum analog of a classical probability distribution is not a quantum state but a density
matrix. Depending on one’s view of quantum mechanics, one might believe that the whole universe is
described by a quantum mechanical pure state that depends on all the available degrees of freedom.
Even if this is true, one usually studies a subsystem that cannot be described by a pure state.

For an idealized case, let A be a subsystem of interest, with Hilbert space HA . And let B be
everything else of relevance, or possibly all of the rest of the universe, with Hilbert space HB . The
combined Hilbert space is the tensor product HAB = HA ⊗ HB . The simple case is that a state vector
ψAB of the combined system is the tensor product of a state vector ψA ∈ HA and another state vector
ψB ∈ HB :
ψAB = ψA ⊗ ψB . (3.1)
If ψAB is a unit vector, we can choose ψA and ψB to also be unit vectors. In the case of such a product
state, predictions about the A system can be made by forgetting about the B system and using the state
vector ψA . Indeed, if OA is any operator on HA , then the corresponding operator on HAB is OA ⊗ 1B ,
and its expectation value in a factorized state ψAB = ψA ⊗ ψB is
hψAB |OA ⊗ 1B |ψAB i = hψA |OA |ψA ihψB |1B |ψB i = hψA |OA |ψA i. (3.2)

However, a generic pure state ψAB ∈ HAB is not a product state; instead it is “entangled.” If HA
and HB have dimensions N and M, then a generic state in HAB can be presented as an N × M matrix,
for example in the 2 × 3 case  
∗ ∗ ∗
ψAB = . (3.3)
∗ ∗ ∗
5
See, however, [6] for a partial substitute.

10
By unitary transformations on HA and on HB , we can transform ψAB to

ψAB → UψAB V (3.4)

where U and V are N × N and M × M unitaries. The canonical form of a matrix under that operation
is a diagonal matrix, with positive numbers on the diagonal, and extra rows or columns of zeroes, for
example √ 
p1 0 0
√ .
0 p2 0

A slightly more invariant way to say this is that any pure state can be written
X√
ψAB = pi ψAi ⊗ ψBi , (3.5)
i

where we can assume that ψAi and ψBi are orthonormal,

hψAi , ψAj i = hψBi , ψBj i = δ ij (3.6)

and that pi > 0. (The ψAi and ψBi may not be bases of HA or HB , because there may not be enough of
them.) The condition for ψAB to be a unit vector is that
X
pi = 1, (3.7)
i

so we can think of the pi as probabilities. Eqn. (3.5) is called the Schmidt decomposition.

What is the expectation value in such a state of an operator OA that only acts on A? It is
X√
hψAB |OA ⊗ 1B |ψAB i = pi pj hψAi |OA |ψAj ihψBi |1B |ψBj i
i,j
X
= pi hψAi |OA |ψAi i. (3.8)
i

This is the same as

TrHA ρA OA , (3.9)
where ρA is the density matrix X
ρA = pi |ψAi ihψAi |. (3.10)
i

Thus, if we are only going to make measurements on system A, we do not need a wavefunction of the
universe: it is sufficient to have a density matrix for system A.

From the definition X

ρA = pi |ψAi ihψAi | (3.11)
i

11
 P
we see that ρA is hermitian and positive semi-definite. Because i pi = 1, ρA has trace 1:
TrHA ρA = 1. (3.12)
Conversely, every matrix with those properties can be “purified,” meaning that it is the density matrix
of some pure state on some “bipartite” (or two-part) system AB. For this, we first observe that any
hermitian matrix ρA can be diagonalized, meaning that in a suitable basis it takes the form of eqn.
(3.11); moreover, if ρA ≥ 0, then the pi are likewise positive (if one of the pi vanishes, we omit it from
the sum). Having gotten this far, to realize ρA as a density matrix we simply introduce another Hilbert
space HB with orthonormal states ψBi and observe that ρA is the density matrix of the pure state
X√
ψAB = pi ψAi ⊗ ψBi ∈ HA ⊗ HB . (3.13)
i

In this situation, ψAB is called a “purification” of the density matrix ρA . The existence of purifications
is a nice property of quantum mechanics that has no classical analog: the classical analog of a density
matrix is a probability distribution, and there is no notion of purifying a probability distribution.

The purification ψAB of a density matrix ρA is far from unique (even if the auxiliary system B is
specified), because there is freedom in choosing the orthonormal states ψBi in eqn. (3.13). However,
any other set of orthonormal vectors in HB can be obtained from a given choice ψBi by a unitary
transformation of HB , so we learn the following important fact: any two purifications of the same
density matrix ρA on system A by pure states of a bipartite system AB are equivalent under a unitary
transformation of system B.

If there is more than one term in the expansion

X√
ψAB = pi ψAi ⊗ ψBi ∈ HA ⊗ HB , (3.14)
i

we say that systems A and B are entangled in the state ψAB . If there is only one term, the expansion
reduces to
ψAB = ψA ⊗ ψB , (3.15)
an “unentangled” tensor product state. Then system A can be described by the pure state ψA and the
density matrix is of rank 1:
ρA = |ψA ihψA |.
If ρA has rank higher than 1, we say that system A is in a mixed state. If ρA is a multiple of the identity,
we say that A is maximally mixed.

In the general case X

ρA = pi |ψAi ihψAi | (3.16)
i

one will describe all measurements of system A correctly if one says that system A is in the state ψAi
with probability pi . However, one has to be careful here because the decomposition of eqn. (3.16) is

12
not unique. It is unique if the pi are all distinct and one wants the number of terms in the expansion to
be as small as possible, or equivalently if one wants the ψAi to be orthonormal. But if one relaxes those
conditions, then (except for a pure state) there are many ways to make this expansion. This means that
if Alice prepares a quantum system to be in the pure state ψAi with probability pi , then there is no way
to determine the pi or the ψAi by measurements, even if one is provided P with many identical copies to
measure. Any measurement of the system will depend only on ρA = i pi |ψAi ihψAi |. There is no way to
get additional information about how the system was prepared.

So far, when we have discussed a bipartite system AB, we have assumed that the combined system
is in a pure state ψAB , and we have discussed density matrices ρA and ρB for systems A and B. More
generally, we should allow for the possibility that the combined system AB is described to begin with by
a density matrix ρAB . Consideration of this situation leads to the following very fundamental definition.

Just as for classical probability distributions, for density matrices we can always “integrate out” an
unobserved system and get a reduced density matrix for a subsystem. Classically, given a joint prob-
ability distribution PX,Y (xi , yj ) for a bipartite system XY , we “integrated out” Y to get a probability
distribution for X only: X
PX (xi ) = PX,Y (xi , yj ). (3.17)
j

The quantum analog of that is a partial trace. Suppose that AB is a bipartite system with Hilbert
space HA ⊗ HB and a density matrix ρAB . Concretely, if |iiA , i = 1, . . . , n are an orthonormal basis of
HA and |αiB , α = 1, . . . , m are an orthonormal basis of HB , then a density matrix for AB takes the
general form X
ρAB = cii′ αα′ |iiA ⊗ |αiB A hi′ | ⊗ B hα′ |. (3.18)
i,i′ ,α,α′

The reduced density matrix for measurements of system A only is obtained by setting α = α′ , replacing
|αiB B hα| by its trace, which is 1, and summing:
X
ρA = ci,i′ ,α,α |iiA A hi′ |. (3.19)
i,i′ ,α

In other words, if we are going to measure system A only, we sum over all of the unobserved states of
system B. This is usually written as a partial trace:

ρA = TrHB ρAB , (3.20)

the idea being that one has “traced out” HB , leaving a density operator on HA . Likewise (summing
over i to eliminate HA )
ρB = TrHA ρAB . (3.21)

Before going on, perhaps I should give a simple example of a concrete situation in which it is
impractical to not use density matrices. Consider an isolated atom interacting with passing photons. A

13
photon might be scattered, or absorbed and reemitted, or might pass by without interacting with the
atom. Regardless, after a certain time, the atom is again alone. After n photons have had the chance to
interact with the atom, to give a pure state description, we need a joint wavefunction for the atom and
all the outgoing photons. The mathematical machinery gets bigger and bigger, even though (assuming
we observe only the atom) the physical situation is not changing. By using a density matrix, we get
a mathematical framework for describing the state of the system that does not change regardless of
how many photons have interacted with the atom in the past (and what else those photons might have
interacted with). All we need is a density matrix for the atom.

3.2 Quantum Entropy

The von Neumann entropy6 of a density matrix ρA is defined by a formula analogous to the Shannon
entropy of a probability distribution:

S(ρA ) = −Tr ρA log ρA . (3.22)

As an immediate comment, we note that S(ρA ) is manifestly invariant under a unitary transformation

ρA → UρA U −1 . (3.23)

Quantum conditional and relative entropy, which will be introduced in section 3.4, are similarly invariant
under a suitable class of unitaries.

By a unitary transformation, we can diagonalize ρA , putting it in the form

X
ρA = pi |ψAi ihψAi |, (3.24)
i

with ψAi being orthonormal and pi > 0. Then in an obvious basis

 
p1 log p1
 p2 log p2 
 
ρA log ρA =  p3 log p3  (3.25)
 
..
.

and so X
S(ρA ) = − pi log pi , (3.26)
i

the same as the Shannon entropy of the probability distribution {pi }.

6
The von Neumann entropy is the most important quantum entropy, but generalizations such as the Rényi entropies
1
Sα (ρA ) = 1−α log Tr ρα
A can also be useful.

14
 An immediate consequence is that, just as for the Shannon entropy,

S(ρA ) ≥ 0, (3.27)

with
P equality only for a pure state (one of the p’s being 1 and the others 0). The formula S(ρA ) =
− i pi log pi also implies the same upper bound that we had classically for a system with k states

S(ρA ) ≤ log k, (3.28)

with equality only if ρA is a multiple of the identity:

 
1

1 1 

ρA =  1 . (3.29)
k 
..
.

In this case, we say that A is in a maximally mixed state. In fact, the von Neumann entropy has many
properties analogous to the Shannon entropy, but the explanations required are usually more subtle and
there are key differences.

Here is a nice property of the von Neumann entropy that does not have a classical analog. If a
bipartite system AB is in a pure state
X√
ψAB = pi ψAi ⊗ ψBi ∈ HA ⊗ HB , (3.30)
i

then the density matrices of systems A and B are

X
ρA = pi |ψAi ihψAi |, (3.31)
i

and likewise X
ρB = pi |ψBi ihψBi |. (3.32)
i

The same constants pi appear in each, so clearly

S(ρA ) = S(ρB ). (3.33)

Thus a system A and a purifying system B always have the same entropy. Note that in this situation,
since the combined system AB is in a pure state, its entropy SAB vanishes.

15
3.3 Concavity

The von Neumann entropy – like its antecedents in classical thermodynamics and statistical mechanics
– has the important property of concavity. Suppose that ρ1 and ρ2 are two density matrices, and set
ρ(t) = tρ1 + (1 − t)ρ2 , for 0 ≤ t ≤ 1. We will write ρ̇(t), ρ̈(t) for dρ(t)/dt, d2 ρ(t)/dt2 . Then
d2
S(ρ(t)) ≤ 0. (3.34)
dt2
To prove this, we first compute that7
d
S(ρ(t)) = −Tr ρ̇ log ρ. (3.35)
dt
Then as Z ∞  
1 1
log ρ = ds − (3.36)
0 s + 1 s + ρ(t)
and ρ̈ = 0, we have Z ∞
d2 1 1
S(ρ(t)) = − dsTr ρ̇ ρ̇ . (3.37)
dt2 0 s + ρ(t) s + ρ(t)
The integrand is positive, as it is Tr B 2 , where B is the self-adjoint operator (s+ρ(t))−1/2 ρ̇(t)(s+ρ(t))−1/2 .
d2
So dt2 S(ρ(t)) ≤ 0.

In other words, the function S(ρ(t)) is concave. Like any concave function, S(ρ(t)) has the property
that the straight line connecting two points on its graph lies below the graph. Explicitly, this gives

tS(ρ1 ) + (1 − t)S(ρ2 ) ≤ S(tρ1 + (1 − t)ρ2 ) = S(ρ(t)). (3.38)

More generally, let ρi , i = 1, . . . , n be density matrices and pi , i = 1, . . . , n nonnegative numbers with
P
i pi = 1. Then by induction starting with (3.38), or because this is a general property of concave
functions, we have X X
pi S(ρi ) ≤ S(ρ), ρ = pi ρi . (3.39)
i i
This may be described by saying that entropy can only increase under mixing. The nonnegative quantity
that appears here is known as the Holevo information or Holevo χ [7]:
X
χ = S(ρ) − pi S(ρi ). (3.40)
i

7
For this, consider an arbitrary density matrix ρ and a first order perturbation ρ → ρ + δρ. After diagonalizing ρ, one
observes that to first order in δρ, the off-diagonal part of δρ does not contribute to the trace in the definition of S(ρ + δρ).
Therefore, S(ρ(t)) can be differentiated assuming that ρ and ρ̇ commute. P So it suffices to check (3.35) for a diagonal
family of density matrices ρ(t) = diag(λ1 (t), λ2 (t), · · · , λn (t)), with i λi (t) = 1. Another approach is to use (3.36) to
substitute for log ρ(t) in the definition S(ρ(t)) = −Tr ρ(t) log ρ(t). Differentiating with respect to t, observing that ρ(t)
commutes with 1/(s + ρ(t)), and then integrating over s, one arrives at (3.35). In either approach, one uses that Tr ρ̇ = 0
since Tr ρ(t) = 1.

16
 An interesting special case is the following. Let ρ be any density matrix on a Hilbert space H.
Pick a basis of H, and let ρD be the diagonal density matrix obtained in that basis by dropping the
off-diagonal matrix elements from ρ and keeping the diagonal ones. Let ρ(t) = (1 − t)ρD + tρ. We see
that 
d 
S(ρ(t)) = 0, (3.41)
dt t=0

by virtue of (3.35), because ρ(0) and log ρ(0) are diagonal while the diagonal matrix elements of dρ/dt
vanish at t = 0. When we combine this with d2 S(ρ(t))/dt2 ≤ 0, we get S(ρ(1)) ≤ S(ρ(0)) or

S(ρD ) ≥ S(ρ). (3.42)

Thus, dropping the off-diagonal part of a density matrix (in any basis) can only increase the entropy.
d2 
Eqn. (3.42) is a strict inequality unless ρ = ρD , because eqn. (3.37) shows that dt2 S(ρ(t)) is strictly
t=0
negative unless ρ = ρD .

An alternative proof of eqn. (3.42), again using the inequality (3.39), is as follows. For an N state
system, there are 2N matrices that are diagonal matrices (in some chosen basis) with diagonal matrix
elements that are all ±1. Let Ui be any of these and set ρi = Ui ρUi−1 . Of course, ρi is also a density
matrix, since Ui is unitary. The average of the ρi , over all 2N choices of Ui , is the diagonal density
matrix ρD . So eqn. (3.39) says that the average of S(ρi ) is less than or equal to S(ρD ). But S(ρi ) is
independent of i and equal to S(ρ), since the von Neumann entropy is invariant under conjugation by
a unitary matrix such as Ui . So in fact the average of the S(ρi ) is just S(ρ) and the inequality (3.39)
becomes S(ρ) ≤ S(ρD ).

Somewhat similarly to what we have explained here, concavity of the function f (q) = −q log q could
have been used in the classical arguments in section 2, though we circumvented this by using Stirling’s
formula instead.

3.4 Conditional and Relative Quantum Entropy

It is now possible to formally imitate some of the other definitions that we made in the classical case.
For example, if AB is a bipartite system, we define what is called quantum conditional entropy

S(A|B) = SAB − SB . (3.43)

This name is potentially misleading because there is not a good quantum notion of conditional probabil-
ities. Unlike the classical case, quantum conditional entropy is not an entropy conditional on something.
Nevertheless, in section 4.1, we will discuss at least one sense in which quantum conditional entropy
behaves in a way analogous to classical conditional entropy.

17
 There is also a fundamental difference from the classical case: quantum mechanically, S(A|B) can
be negative. In fact, suppose that system AB is in an entangled pure state. Then SAB = 0 but as
system B is in a mixed state, SB > 0. So in this situation S(A|B) < 0.

Another classical definition that is worth imitating is the mutual information. Given a bipartite
system AB with density matrix ρAB , the mutual information is defined just as it is classically:

I(A; B) = SA − SAB + SB . (3.44)

Here, however, we are more fortunate, and the quantum mutual information is nonnegative:

I(A; B) ≥ 0. (3.45)

Moreover, I(A; B) = 0 if and only if the density matrix factorizes, in the sense that

ρAB = ρA ⊗ ρB . (3.46)

Positivity of mutual information is also called subadditivity of entropy. To begin with, quantum mutual
information is a formal definition and it is not obvious how it is related to information that one can
gain about system A by observing system B. We will explore at least one aspect of this question in
section 4.3.

Before proving positivity of mutual information, I will explain an interesting corollary. Although
conditional entropy S(A|B) can be negative, the possibility of “purifying” a density matrix gives a lower
bound on S(A|B). Let C be such that ABC is in a pure state. Remember that in general if XY is in
a pure state then SX = SY . So if ABC is in a pure state then SAB = SC and SB = SAC . Thus

SAB − SB = SC − SAC ≥ −SA , (3.47)

where the last step is positivity of mutual information. So

S(A|B) = SAB − SB ≥ −SA . (3.48)

Reversing the roles of A and B in the derivation, we get the Araki-Lieb inequality [8]

SAB ≥ |SA − SB |. (3.49)

It is saturated if SAB = 0, which implies SB = SA . What has just been explained is a typical argument
exploiting the existence of purifications.

Just as in the classical case, to understand positivity of the mutual information, it helps to first
define the relative entropy [9]. Suppose that ρ and σ are two density matrices on the same Hilbert space
H. The relative entropy can be defined by imitating the classical formula:

S(ρ||σ) = Trρ(log ρ − log σ). (3.50)

18
For now, this is just a formal definition, but we will learn in section 4.2 that S(ρ||σ) has the same
interpretation quantum mechanically that it does classically: if one’s hypothesis is that a quantum
system is described by a density matrix σ, and it is actually described by a different density matrix ρ,
then to learn that one is wrong, one needs to observe N copies of the system where NS(ρ||σ) >> 1.

Just as classically, it turns out that S(ρ||σ) ≥ 0 for all density matrices ρ, σ, with equality precisely
if ρ = σ. To prove this, first diagonalize σ. In general ρ is not diagonal in the same basis. Let ρD be
the diagonal density matrix obtained from ρ by dropping the off-diagonal matrix elements in the basis
in which σ is diagonal, and keeping the diagonal ones. Since Tr ρ log σ = Tr ρD log σ, it follows directly
from the definitions of von Neumann entropy and relative entropy that

S(ρ||σ) = S(ρD ||σ) + S(ρD ) − S(ρ). (3.51)

This actually exhibits S(ρ||σ) as the sum of two nonnegative terms. We showed in eqn. (3.42)
that S(ρD ) − S(ρ) ≥ 0. As for S(ρD ||σ), it is nonnegative, because if σ = diag(q1 , . . . , qn ), ρD =
diag(p1 , . . . , pn ), then X
S(ρD ||σ) = pi (log pi − log qi ), (3.52)
i

which can be interpreted as a classical relative entropy and so is nonnegative. To get equality in these
statements, we need σ = ρD and ρD = ρ, so S(ρ||σ) vanishes only if ρ = σ.

Now we can use positivity of the relative entropy to prove that I(A; B) ≥ 0 for any density matrix
ρAB . Imitating the classical proof, we define

σAB = ρA ⊗ ρB , (3.53)

and we observe that

log σAB = log ρA ⊗ 1B + 1A ⊗ log ρB , (3.54)
so

S(ρAB ||σAB ) = TrAB ρAB (log ρAB − log σAB )

= TrAB ρAB (log ρAB − log ρA ⊗ 1B − 1B ⊗ log ρB )
= SA + SB − SAB = I(A; B). (3.55)

So just as classically, positivity of the relative entropy implies positivity of the mutual information
(which is also called subadditivity of entropy).

The inequality (3.39) that expresses the concavity of the von Neumann entropy can be viewed as a
special case of the positivity of mutual information. Let B be a quantum system with density matrices
ρiB and let C be an auxiliary system C with an orthonormal basis |iiC . Endow CB with the density
matrix: X
ρCB = pi |iiC C hi| ⊗ ρiB . (3.56)
i

19
The mutual information between C and B if the combined system is described by ρCB is readily com-
puted to be X
I(C; B) = S(ρB ) − pi S(ρiB ), (3.57)
i

so positivity of mutual information gives our inequality.

3.5 Monotonicity of Relative Entropy

So relative entropy is positive, just as it is classically. Do we dare to hope that relative entropy is also
monotonic, as classically? Yes it is, as first proved by Lieb and Ruskai [10], using a lemma of Lieb [11].
How to prove strong subadditivity will not be described here; this has been explored in a companion
article [12], sections 3 and 4.

Monotonicity of quantum relative entropy is something of a miracle, because, as there is no such thing
as a joint probability distribution for general quantum observables, the intuition behind the classical
statement is not applicable in any obvious way. Rather, strong subadditivity is ultimately used to prove
that quantities such as quantum conditional entropy and quantum relative entropy and quantum mutual
information do have properties somewhat similar to the classical case. We will explore some of this in
section 4.

There are different statements of monotonicity of relative entropy, but a very basic one (and actually
the version proved in [10]) is monotonicity under partial trace. If AB is a bipartite system with two
density matrices ρAB and σAB , then we can take a partial trace on B to get reduced density matrices
on A:
ρA = TrB ρAB , σA = TrB σAB . (3.58)
Monotonicity of relative entropy under partial trace is the statement that taking a partial trace can
only reduce the relative entropy:
S(ρAB ||σAB ) ≥ S(ρA ||σA ). (3.59)
(This is also called the Data Processing Inequality.)

By imitating what we said classically in section 2, one can deduce strong subadditivity of quantum
entropy from monotonicity of relative entropy. We consider a tripartite system ABC with density
matrix ρABC . There are reduced density matrices such as ρA = TrBC ρABC , ρBC = TrA ρABC , etc., and
we define a second density matrix
σABC = ρA ⊗ ρBC . (3.60)
The reduced density matrices of ρABC and σABC , obtained by tracing out C, are

ρAB = TrC ρABC , σAB = TrC σABC = ρA ⊗ ρB . (3.61)

20
 Monotonicity of relative entropy under partial trace says that

S(ρABC ||σABC ) ≥ S(ρAB ||σAB ). (3.62)

But (as in our discussion of positivity of mutual information)

S(ρABC ||σABC ) = S(ρABC ||ρA ⊗ ρBC ) = I(A; BC) = SA + SBC − SABC (3.63)

and similarly
S(ρAB ||σAB ) = S(ρAB ||ρA ⊗ ρB ) = I(A; B) = SA + SB − SAB . (3.64)

So eqn. (3.62) becomes monotonicity of mutual information

I(A; BC) ≥ I(A; B) (3.65)

or equivalently strong subadditivity [10]

SAB + SBC ≥ SB + SABC . (3.66)

All of these steps are the same as they were classically. Using purifications, one can find various
equivalent statements. If ABCD is in a pure state then SAB = SCD , SABC = SD so the inequality
becomes
SCD + SBC ≥ SB + SD . (3.67)
So for instance S(C|D) = SCD − SD can be negative, or S(C|B) = SBC − SB can be negative, but

S(C|D) + S(C|B) ≥ 0. (3.68)

(This is related to “monogamy of entanglement”: a given qubit in C can be entangled with D, reducing
SCD , or with B, reducing SBC , but not both.)

Classically, the intuition behind monotonicity of mutual information was explained in section 2; one
learns at least as much about system A by observing B and C as one could learn by observing B only.
Quantum mechanically, it is just not clear a priori that the formal definition I(A; B) = SA − SAB + SB
will lead to something consistent with that intuition. The rather subtle result of monotonicity of relative
entropy [10] shows that it does.

In general, strong subadditivity (or monotonicity of relative entropy) is the key to many interesting
statements in quantum information theory. Many of the most useful statements that are not more
elementary are deduced from strong subadditivity.

21
3.6 Generalized Measurements

Once we start using density matrices, there are a few more tools we should add to our toolkit. First let
us discuss measurements. Textbooks begin with “projective measurements,” which involve projection
onto orthogonal subspaces of a Hilbert space H of quantum states. We pick orthogonal hermitian
projection operators πs , s = 1, · · · , k obeying
X
πs = 1, πs2 = πs , πs πs′ = 0, s 6= s′ . (3.69)
s

A measurement of a state ψ involving these projection operators has outcome s with probability

ps = hψ|πs |ψi. (3.70)

P P
These satisfy s ps = 1 since s πs = 1. If instead of a pure state ψ the system is described by a
density matrix ρ, then the probability of outcome s is

ps = TrH πs ρ. (3.71)

After the measurement is made, if outcome s has been found, the system can be described by a new
density matrix
1
ρs = πs ρπs . (3.72)
ps

But Alice can make a more general type of measurement using an auxiliary system C (sometimes
called an ancillary system) with Hilbert space C. We suppose that C is k-dimensional with a basis of
states |si, s = 1, · · · , k. Alice initializes C in the state |1i. Then she acts on the combined system
C ⊗ H with a unitary transformation U, which she achieves by suitably adjusting a time-dependent
Hamiltonian. She chooses U so that for any ψ ∈ H
k
X
U(|1i ⊗ ψ) = |si ⊗ Es ψ (3.73)
s=1

for some linear operators Es . (She does not care what U does on other states.) Unitarity of U implies
that
Xk
Es† Es = 1, (3.74)
s=1

but otherwise the Es are completely arbitrary.

Then Alice makes a projective measurement of the system C ⊗ H, using the commuting projection
operators
πs = |sihs| ⊗ 1, (3.75)

22
which have all the appropriate properties. The probability of outcome s is
ps = |Es |ψi|2 = hψ|Es† Es |ψi. (3.76)
More generally, if the system H is described initially by a density matrix ρ, then the probability of
outcome s is
ps = Tr Es† Es ρ. (3.77)
P P
The numbers ps are nonnegative because Es† Es is nonnegative, and s ps = 1 because s Es† Es = 1.
But the Es† Es are not orthogonal projection operators; they are just nonnegative hermitian operators
that add to 1. What we have described is a more general kind of quantum mechanical measurement of
the original system. (In the jargon, this is a “positive operator-valued measurement” or POVM.)

According to eqn. (3.72), after Alice’s measurement, if the outcome s has been found, then the
combined system C ⊗ H can be described by the density matrix p1s |sihs| ⊗ Es |ψihψ|Es†. Taking the
partial trace over system C, we learn that, after the generalized measurement, the original system H
can be described by the density matrix
X 1
Es |ψihψ|Es† (3.78)
s
ps
or more generally (if the original system was in a mixed state with density matrix ρ)
X 1
Es ρEs† . (3.79)
s
p s

One can slightly generalize this construction as follows.8 Suppose that the initial system actually
had for its Hilbert space a direct sum H ⊕ H′ , but it is known that the initial state of the system is
valued in H, in other words the initial state ψ has the form χ ⊕ 0 with χ ∈ H, and 0 the zero vector in
H′ . Then Alice couples H ⊕ H′ to her auxiliary system C, so she describes the combined P system by a
Hilbert space C ⊗ (H ⊕ H′ ). Now she picks U so that it maps a vector |1i ⊗ (χ ⊕ 0) to s |si ⊗ (0 ⊕ Es χ),
where Es is a linear transformation Es : H → H′ . (As before, Alice does not care what U does on other
vectors.) After applying U, Alice makes a projective measurement using the same projection operators
πs = |sihs| ⊗ 1 as before (of course, 1 is now the identity on H ⊕ H′ ). The linear transformations Es
still obey eqn. (3.74), the probability of outcome s is still given by eqn. (3.77), and the density matrix
after a measurement that gives outcome s is still given by eqn. (3.78).

3.7 Quantum Channels

Now let us view this process from another point of view. How can a density matrix evolve? The usual
Hamiltonian evolution of a state ψ is ψ → Uψ for a unitary operator U, and on the density matrix it
8
The following paragraph may be omitted on first reading. It is included to make possible a more general statement
in section 3.7.

23
corresponds to
ρ → UρU −1 . (3.80)
As we remarked earlier (eqn. (3.23)), such unitary evolution preserves the von Neumann entropy of a
density matrix, and similarly it preserves the relative entropy between two density matrices.

But let us consider Alice again with her extended system C ⊗ H. She initializes the extended system
with the density matrix
ρb = |1ih1| ⊗ ρ (3.81)
where ρ is a density matrix on H. Then she applies the same unitary U as before, mapping ρb to
k
X
′
ρb = U ρbU −1
= |sihs′ | ⊗ Es ρEs†′ . (3.82)
s,s′ =1

The induced density matrix on the original system H is obtained by a partial trace and is
k
X
ρ′ = TrC ρb′ = Es ρEs† . (3.83)
s=1

We have found a more general way that density matrices can evolve. The operation
k
X X
ρ→ Es ρEs† , Es† Es = 1 (3.84)
s=1 s

is called a “quantum channel,” and the Es are called Kraus operators. Unitary evolution is the special
case in which there is only one Kraus operator.

The notion of a quantum channel is axiomatized in more complete treatments than we will give
here.9 The upshot of a general analysis is that the most general physically sensible evolution of a
density matrix takes the form (3.84), provided one allows the generalization described at the end of
section 3.6 in which the Es are linear transformations from one Hilbert space H to another Hilbert
space H′ .

Now let ρ and σ be two different density matrices on H. Let us ask what happens to the relative
entropy S(ρ||σ) when we apply a quantum channel, mapping ρ and σ to
X X
ρ′ = Es ρEs† , σ′ = Es σEs† . (3.85)
s s

The first step of initialization, replacing ρ and σ by |1ih1| ⊗ ρ and |1ih1| ⊗ σ, does not change anything.
The second step, conjugating by a unitary matrix U, also does not change anything since relative entropy
9
In the most general case, a quantum channel is a “completely positive trace-preserving” (CPTP) map from density
matrices on one Hilbert space H to density matrices on another Hilbert space H′ .

24
is invariant under conjugation. Finally, the last step was a partial trace, which can only reduce the
quantum relative entropy. So relative entropy can only go down under a quantum channel:

S(ρ||σ) ≥ S(ρ′ ||σ ′ ).

This is the most general statement of monotonicity of quantum relative entropy.

We conclude this section with some exercises to familiarize oneself with quantum channels.

(1) Let ψ be any pure state of a given system. Find Kraus operators of a quantum channel that
maps any density matrix ρ to |ψihψ|. (One way to implement this is to turn on a Hamiltonian for which
ψ is the ground state, and wait until the system relaxes to its ground state by releasing energy to the
environment.)

(2) Find Kraus operators of a quantum channel that maps any density matrix for a given system
(with finite-dimensional Hilbert space) to a maximally mixed one, a multiple of the identity. (This can
arise as the outcome of sufficiently random interaction of the system with its environment.)

(3) Do the same for a quantum channel that, in a given basis, maps any k ×k density matrix ρ = (ρij )
to the corresponding diagonal density matrix ρD = diag(ρ11 , ρ22 , · · · , ρkk ). (An idealized description of
a physical realization is as follows. A cavity is probed by atoms. Denote as |ni the state of the cavity
when it contains n photons. Suppose that n is unchanged when an atom passes through the cavity, but
the final state of the atom depends on n. The probability to find the cavity in state |ni is unchanged by
the interaction with a passing atom, so in the basis {|ni}, the diagonal elements of the density matrix
are unchanged. After many atoms have passed through the cavity, an observation of the atoms would
reveal with high confidence the number of photons in the cavity. Therefore, tracing over the atomic
states, the final density matrix of the cavity is diagonal in the basis {|ni}. Regardless of what state the
cavity begins in, it will end up with high probability in an eigenstate of the photon number operator,
though one cannot say what the eigenvalue will be.)

(4) Show that the composition of two quantum channels is a quantum channel. If the first channel
has Kraus operators Es , s = 1, · · · , p, and the second has Kraus operators Et′ , t = 1, · · · , q, what are
the Kraus operators of the composite channel?

(5) This and the next exercise involve quantum channels that map one Hilbert space to another.
The goal is to show that natural operations that are well-motivated in other ways can also be viewed
as special cases of the evolution described in eqn. (3.84). First, given a Hilbert space H, construct a
rather trivial quantum channel that maps density matrices on H to density matrices on a 1-dimensional
Hilbert space H0 . Note that, since a density matrix is hermitian, positive-definite, and of trace 1, there
is a unique density matrix on H0 , namely the unit density matrix 1. Thus, given a Hilbert space H,
find Kraus operators Es : H → H0 for a quantum channel that maps any density matrix ρ on H to
the density matrix 1 on H0 . Once you have done this, show that a partial trace is a quantum channel
in the following sense. If AB is a bipartite system with Hilbert space HA ⊗ HB , find Kraus operators

25
Es : HA ⊗ HB → HA that implement thePpartial trace ρABP→ ρA = TrB ρAB . In other words, find
operators Es : HA ⊗ HB → HA , satisfying s Es† Es = 1 and s Es ρAB Es† = TrB ρAB , for any ρAB .

(6) Let A be a quantum system with Hilbert space HA , and let B be a second quantum system with
Hilbert space HB and some given density matrix ρB . Find Kraus operators Es : HA → HA ⊗ HB for a
quantum channel that combines a quantum system A with some other system B by mapping any given
density matrix ρA on A to the density matrix ρA ⊗ ρB on AB. (You might want to consider first the
trivial case that HA is 1-dimensional.) An example of this is what happens whenever a system A under
study is combined with some experimental apparatus B, which has been initialized in the state ρB .

3.8 Thermodynamics And Quantum Channels

As an example of these considerations, let us suppose that σ is a thermal density matrix at some
temperature T = 1/β
1
σ = exp(−βH). (3.86)
Z
So log σ = −βH − log Z and therefore the relative entropy between any density matrix ρ and σ is

S(ρ||σ) =Tr ρ(log ρ − log σ) = −S(ρ) + Trρ(βH + log Z)

=β(E(ρ) − T S(ρ)) + log Z (3.87)

where the average energy computed in the density matrix ρ is

E(ρ) = Tr ρH. (3.88)

We define the free energy

F (ρ) = E(ρ) − T S(ρ). (3.89)
The log Z term in eqn (3.87) is independent of ρ and gives a constant that ensures that S(σ||σ) = 0. So

S(ρ||σ) = β(F (ρ) − F (σ)). (3.90)

Now consider any evolution of the system, that is any quantum channel, that preserves thermal
equilibrium at temperature β. Thus, this channel maps σ to itself, but it maps ρ to a generally different
density matrix ρ′ . The relative entropy can only go down under a quantum channel, so

S(ρ||σ) ≥ S(ρ′ ||σ), (3.91)

and therefore
F (ρ) ≥ F (ρ′ ). (3.92)

26
 In other words, a quantum channel that preserves thermal equilibrium can only reduce the free
energy. This is an aspect of the second law of thermodynamics. If you stir a system in a way that maps
thermal equilibrium at temperature T to thermal equilibrium at the same temperature, then it moves
any density matrix closer to thermal equilibrium at temperature T .

To specialize further, take the temperature T = ∞, β = 0. (This makes sense for a system with a
finite-dimensional Hilbert space.) The thermal density matrix σ is then maximally mixed, a multiple of
the identity. For T → ∞, F (ρ) ∼ −T S(ρ). So in this case, reducing the free energy means increasing
the entropy. Thus a quantum channel that maps a maximally mixed density matrix to itself can only
increase
P the entropy. The condition that a channel maps a maximally mixed density matrix to itself
†
is
P s†Es Es = 1. (A channel satisfying this condition is called unital. By contrast, the condition
s Es Es = 1 is satisfied by all quantum channels.)

An example of a quantum channel that maps a maximally mixed density matrix to itself is the
channel that maps any density matrix ρ to the corresponding diagonal density matrix ρD (in some
chosen basis). The fact that the entropy can only increase under such a channel implies the inequality
S(ρ) ≤ S(ρD ) (eqn. (3.42)).

4 More On Quantum Information Theory

From this point, one could pursue many different directions toward a deeper understanding of quantum
information theory. This article will conclude with three topics that the author found helpful in gaining
insight about the meaning of formal definitions such as quantum conditional entropy and quantum rela-
tive entropy. These concepts were defined by formally imitating the corresponding classical definitions,
and it is not really clear a priori what to expect of such formal definitions.

A secondary reason for the choice of topics is to help the reader appreciate the importance of
monotonicity of quantum relative entropy – and its close cousin, strong subadditivity. At several points,
we will have to invoke monotonicity of relative entropy to prove that quantities like quantum mutual
information and quantum relative entropy that have been defined in a formal way do behave in a fashion
suggested by their names.

The three topics that we will consider are quantum teleportation and conditional entropy, relative
entropy and quantum hypothesis testing, and the use of a quantum state to encode classical information.

27
4.1 Quantum Teleportation and Conditional Entropy

We start with quantum teleportation [13]. For a first example, imagine that Alice has in her pos-
session a qubit A0 , a quantum system with a two-dimensional Hilbert space. Alice would like to help
Bob create in his lab a qubit in a state identical to A0 . However, it is too difficult to actually send a
qubit; she can only communicate by sending a classical message over the telephone. If Alice knows the
state of her qubit, there is no problem: she tells Bob the state of her qubit and he creates one like it
in his lab. If, however, Alice does not know the state of her qubit, she is out of luck. All she can do is
make a measurement, which will give some information about the prior state of qubit A0 . She can tell
Bob what she learns, but the measurement will destroy the remaining information about A0 and it will
never be possible for Bob to recreate A0 .

Suppose, however, that Alice and Bob have previously shared a qubit pair A1 B1 (Alice has A1 , Bob
has B1 ) in a known entangled state, for example
1
ΨA1 B1 = √ (|0 0i + |1 1i)A1 B1 . (4.1)
2
Maybe Alice created this pair in her lab and then Bob took B1 on the road with him, leaving A1 in
Alice’s lab. In this case, Alice can solve the problem. To do so she makes a joint measurement of her
system A0 A1 in a basis that is chosen so that no matter what the answer is, Alice learns nothing about
the prior state of A0 . In the process, she also loses no information about A0 , since she had none before.
But as we will see, after getting her measurement outcome, she can tell Bob what to do to recreate A0 .

To see how this works, let us describe a specific measurement that Alice can make on A0 A1 that will
shed no light on the state of A0 . She can project A0 A1 on the basis of four states
1 1
√ (|0 0i ± |1 1i)A0A1 and √ (|0 1i ± |1 0i)A0A1 . (4.2)
2 2
To see the result of a measurement, suppose the unknown state of qubit A0 is α|0i + β|1i. So the initial
state of A0 A1 B1 is
1
ΨA0 A1 B1 = √ (α|0 0 0i + α|0 1 1i + β|1 0 0i + β|1 1 1i)A0 A1 B1 . (4.3)
2

Suppose that the outcome of Alice’s measurement is to learn that A0 A1 is in the state
1
√ (|0 0i − |1 1i)A0A1 . (4.4)
2
After the measurement, B1 will be in the state (α|0i − β|1i)B1 . Knowing this, Alice can tell Bob that
he can recreate the initial state by acting on his qubit by
 
1 0
ΨB1 → ΨB1 (4.5)
0 −1

28
in the basis |0i, |1i. The other cases are similar, as the reader can verify.

We will analyze a generalization, but first it is useful to formalize in a different way the idea that
Alice is trying to teleport an arbitrary unknown quantum state. For this, we add another system R,
to which Alice and Bob do not have access. We assume that R is maximally entangled with A0 in a
known state, say
1
ΨRA0 = √ (|0 0i + |1 1i)RA0 . (4.6)
2
In this version of the problem, Alice’s goal is to manipulate her system A0 A1 in some way, and then tell
Bob what to do to his system B = B1 so that in the end the system RB1 will be in the same state
1
ΨRB1 = √ (|0 0i + |1 1i)RB1 (4.7)
2
that RA0 was previously – with R never being touched. In this version of the problem, the combined
system RAB1 = RA0 A1 B1 starts in a pure state ΨRAB1 = ΨRA0 ⊗ ΨA1 B1 . The solution of this version
of the problem is the same as the other one: Alice makes the same measurements and sends the same
instructions as before.

We can understand better what is happening if we take a look at the conditional entropy of the
system AB = A0 A1 B1 . Since A1 B1 is in a pure state, it does not contribute to SAB , so SAB = SA0 = 1
(A0 is maximally mixed, since it is maximally entangled with R). Also SB = 1 since B = B1 is
maximally entangled with A1 . Hence

S(A|B) = SAB − SB = 1 − 1 = 0. (4.8)

It turns out that this is the key to quantum teleportation: teleportation, in a suitably generalized sense,
is possible when and only when
S(A|B) ≤ 0. (4.9)

Let us explain first why this is a necessary condition. We start with an arbitrary system RAB in
a pure state ΨRAB ; Alice has access to A, Bob has access to B, and neither one has access to R. For
teleportation, Alice might measure her system A using some rank 1 orthogonal projection operators
πi . (If she makes a more general measurement, for example using projection operators of higher rank,
the system RB does not end up in a known pure state and she will not be able to give appropriate
instructions to Bob.) No matter what answer she gets, after the measurement, system A is in a pure
state and therefore RB is also in a pure state χRB , generally entangled. For teleportation, Alice has
to choose the πi so that, no matter what outcome she gets, the density matrix ρR of R is the same as
before. If this is so, then after her measurement, the state χRB of RB is a purification of the original ρR .
Since she knows her measurement outcome, Alice knows which entangled state is χRB and can convey
this information to Bob. Bob is then in possession of part B of a known purification χRB of system R.
He makes in his lab a copy A′ of Alice’s original system A, initialized in a known pure state ΩA′ , so
now he has part A′ B of a known purification Ψ e RA′ B = ΩA′ ⊗ χRB of ρR . By a unitary transformation of

29
system A′ B, which Bob can implement in his lab, Ψe RA′ B can be converted into any other pure state of
′
RA B that purifies the same ρR . (This was explained following eqn. (3.13).) So Bob can convert Ψe RA′ B
to a copy of the original ΨRAB .

But do there exist projection operators of Alice’s system with the necessary properties? The initial
state ΨABR is pure so it has
SAB = SR . (4.10)
Bob’s density matrix at the beginning is

ρB = TrRA ρRAB (4.11)

where ρRAB is the initial pure state density matrix. By definition

SB = S(ρB ). (4.12)

If Alice gets measurement outcome i, then Bob’s density matrix after the measurement is
1
ρiB = TrRA πi ρRAB . (4.13)
pi

Note that X
ρB = pi ρiB , (4.14)
i
P
since i πi = 1. After the measurement, since A is in a pure state, RB is also in a pure state ΨiRB , so
S(ρiB ) = SR . But by hypothesis, the measurement did not change ρR , so SR is unchanged and so equals
the original SAB . Hence
S(ρiB ) = SAB . (4.15)
If all this is possible X
SAB = S(ρiB ) = pi S(ρiB ). (4.16)
i

The concavity
P inequality (3.39) or equivalently positivity of the Holevo information (3.40) says that if
i
ρB = i pi ρB then X
S(ρB ) ≥ pi S(ρiB ). (4.17)
i

So if teleportation can occur, X

SAB = pi S(ρiB ) ≤ S(ρB ) = SB (4.18)
i

and hence S(A|B) = SAB − SB ≤ 0.

Actually, S(A|B) ≤ 0 is sufficient as well as necessary for teleportation, in the following sense [14].
(In this generality, what we are calling teleportation is known as state merging.) One has to consider the

30
problem of teleporting not a single system but N copies of the system for large N. (This is a common
device in quantum information theory. It is a rough analog of the fact that to get simple statements in
the classical case in section 2, we had to consider a long message, obtained by sampling N times from
a probability distribution.) So one takes N copies of system RAB for large N, thus replacing RAB
by R⊗N A⊗N B ⊗N . This multiplies all the entropies by N, so it preserves the condition S(A|B) ≤ 0.
Now Alice tries to achieve teleportation by making a complete projective measurement on her system
A⊗N . It is very hard to find an explicit set of projection operators πi with the right properties, but it
turns out, remarkably, that for large N, a random choice will work (in the sense that with a probability
approaching 1, the error in state merging is vanishing for N → ∞). This statement actually has strong
subadditivity as a corollary [14]. This approach to strong subadditivity has been described in sections
10.8-9 of [4].

We actually can now give a good explanation of the meaning of quantum conditional entropy S(A|B).
Remember that classically S(A|B) measures how many additional bits of information Alice has to send
to Bob after he has already received B, so that he will have full knowledge of A. We will find a quantum
analog of this, but now involving qubits rather than classical bits. Suppose that S(A|B) > 0 and Alice
nevertheless wants to share her state with Bob. Now we have to assume that Alice is capable of quantum
communication, that is of sending a quantum system to Bob while maintaining its quantum state, but
that she wishes to minimize the amount of quantum communication she will need. She first creates
some maximally entangled qubit pairs and sends half of each pair to Bob. Each time she sends Bob half
of a pair, SAB is unchanged but SB goes up by 1, so S(A|B) = SAB − SB goes down by 1. So S(A|B),
if positive, is the number of such qubits that Alice must send to Bob to make S(A|B) nonpositive and
so make teleportation or state merging possible without any further quantum communication.

If S(A|B) is negative, teleportation or state merging is possible to begin with and −S(A|B) is the
number of maximally entangled qubit pairs that Alice and Bob can be left with afterwards [14]. This
may be seen as follows. Alice creates an auxiliary system A′ A′′ , where A′ consists of n qubits that
are completely entangled with another set of n qubits that comprise system A′′ . Alice considers the
problem of teleporting to Bob the combined system A = A′′ A, while leaving A′ untouched. Since
S(A|B) = n + S(A|B), Alice observes that S(A|B) < 0 provided n < −S(A|B). Given this inequality,
Alice can teleport A = A′′ A to Bob, keeping A′ in reserve. At the end of this, Alice and Bob share n
maximally entangled qubit pairs, namely Alice’s system A′ and Bob’s copy of A′′ . This description is a
shorthand; it is implicit that at each stage, we are free to replace the system under consideration by the
tensor product of N copies of itself, for some large N. As a result, integrality of n is not an important
constraint. A more precise statement of the conclusion is that for large N, after teleportation to Bob of
part A⊗N of a composite system A⊗N B ⊗N , Alice and Bob can be left with up to −NS(A|B) maximally
entangled qubit pairs.

31
4.2 Quantum Relative Entropy And Hypothesis Testing

In a somewhat similar way, we can give a physical meaning to the relative entropy S(ρ||σ) between
two density matrices ρ, σ. Recall from section 2.3 that classically, if we believe a random variable is
governed by a probability distribution Q but it is actually governed by a probability distribution P ,
then after N trials the ability to disprove the wrong hypothesis is controlled by
2−N S(P ||Q). (4.19)

A similar statement holds quantum mechanically: if our initial hypothesis is that a quantum system
X has density matrix σ, and the actual answer is ρ, then after N trials with an optimal measurement
used to test the initial hypothesis, the confidence that the initial hypothesis was wrong is controlled in
the same sense by
2−N S(ρ||σ) . (4.20)

Let us first see that monotonicity of relative entropy implies that one cannot do better than that [15].
A measurement is a special case of a quantum channel, in the following sense. To measure a system X,
one lets it interact quantum mechanically with some other system Y C where Y is any quantum system
and C is the measuring device. After they interact, one looks at the measuring device and forgets the
rest. Forgetting the rest is a partial trace that maps a density matrix βXY C to βC = TrXY βXY C . If C
is a good measuring device with n distinguishable quantum states, this means that in a distinguished
basis |αi, α = 1, · · · , n, its density matrix βC will have a diagonal form
X
βC = bα |αihα|. (4.21)
α

The “measurement” converts the original density matrix into the probability distribution {bα }.

So when we try to distinguish ρ from σ, we use a quantum channel plus partial trace (or simply a
quantum channel, since a partial trace can be viewed as a quantum channel) that maps ρ and σ into
density matrices for C X X
ρC = rα |αihα| σC = sα |αihα|, (4.22)
α α

and thereby into classical probability distributions R = {rα } and S = {sα }. We can learn that ρ and σ
are different is by observing that R and S are different, a process controlled by
2−N Scl (R||S) , (4.23)
where Scl (R||S) is the classical relative entropy between R and S.

This is the same as the relative entropy between ρC and σC :

S(ρC ||σC ) = Scl (R||S). (4.24)

32
And monotonicity of relative entropy gives

S(ρ||σ) ≥ S(ρC ||σC ). (4.25)

So if we follow this procedure, then S(ρ||σ) gives a bound on how well we can do:

2−N Scl (R||S) ≥ 2−N S(ρ||σ) . (4.26)

Actually, quantum mechanics allows us to do something more sophisticated than making N repeated
measurements of the system of interest. We could more generally make a joint measurement on all N
copies. Taking N copies replaces the Hilbert space H of the system under study by H⊗N , and replaces
the density matrices σ and ρ by σ ⊗N and ρ⊗N . All entropies and relative entropies are multiplied by
N. A joint measurement on N copies would convert a density matrix σ ⊗N or ρ⊗N to a probability
distribution S [N ] or R[N ] . We will not learn much from a single joint measurement on N copies, since it
will just produce a random answer. But given NN ′ copies of the system, we could repeat N ′ times a joint
measurement of N copies. The ability to distinguish S [N ] from R[N ] in N ′ tries is controlled for large
′ [N] [N] ′ [N] [N] ′ ⊗N ⊗N
N ′ by 2−N Scl (R ||S ) . The monotonicity of relative entropy gives 2−N Scl (R ||S ) ≥ 2−N S(ρ ||σ ) =
b
2−N S(ρ||σ) , where Nb = NN ′ . So also with such a more general procedure, the ability to disprove in N b
b S(ρ||σ)
−N
trials an initial hypothesis σ for a system actually described by ρ is bounded by 2 .

b , it is actually possible to saturate this bound, as follows [16,17]. If ρ is diagonal

In the limit of large N
in the same basis in which σ is diagonal, then by making a measurement that involves projecting on
1-dimensional eigenspaces of σ, we could convert the density matrices ρ, σ into classical probability
distributions R, S with S(ρ||σ) = Scl (R||S). The quantum problem would be equivalent to a classical
problem, even without taking many copies. As usual the subtlety comes because the matrices are not
simultaneously diagonal. By dropping from ρ the off-diagonal matrix elements in some basis in which
σ is diagonal, we can always construct a diagonal density matrix ρD . Then a measurement projecting
on 1-dimensional eigenspaces of σ will give probability distributions R, S satisfying

S(ρD ||σ) = Scl (R||S). (4.27)

This is not very useful, because it is hard to compare S(ρD ||σ) to S(ρ||σ). That is why it is necessary
to consider a joint measurement on N copies, for large N, which makes possible an easier alternative to
comparing S(ρD ||σ) to S(ρ||σ), as we will see.

Let us recall the definition of relative entropy:

S(ρ⊗N ||σ ⊗N ) = Tr ρ⊗N log ρ⊗N − Tr ρ⊗N log σ ⊗N . (4.28)

The second term Tr ρ⊗N log σ ⊗N is unchanged if we replace ρ⊗N by its counterpart ρ⊗N
D that is diagonal
in the same basis as σ ⊗N . So

S(ρ⊗N ||σ ⊗N ) − S(ρ⊗N

D ||σ
⊗N
) = Trρ⊗N log ρ⊗N − Trρ⊗N ⊗N
D log ρD . (4.29)

33
For large N, we will be able to get a useful bound on the right hand side.

Roughly speaking, there is simplification for large N because group theory can be used to simulta-
neously put ρ⊗N and σ ⊗N in a block diagonal form with relatively small blocks. This will make possible
the comparison we need. In more detail, the group SN of permutations of N objects acts in an obvious
way on H⊗N . It commutes with the action on H⊗N of U(k), the group of unitary transformations of
the k-dimensional Hilbert space H. Schur-Weyl duality gives the decomposition of H⊗N in irreducible
representations of SN × U(k). Every Young diagram Y with N boxes and at most k rows determines an
irreducible representation λY of SN and an irreducible representation µY of U(k). The decomposition
of H⊗N in irreducibles of SN × U(k) is

H⊗N = ⊕Y λY ⊗ µY . (4.30)

The λY of distinct Y are non-isomorphic, and the same is true of the µY . Let aY and bY be, respectively,
the dimension of λY and of µY . The maximum value of bY is bounded10 by a power of N:

bmax ≤ (N + 1)k(k−1)/2 . (4.31)

The important point will be that bmax grows only polynomially for N → ∞, not exponentially. In
contrast, the numbers aY can be exponentially large for large N.

Eqn. (4.30) gives a decomposition of H⊗N as the direct sum of subspaces of dimension aY bY . Since
⊗N
ρ and σ ⊗N commute with SN , they are block diagonal with respect to this decomposition. But more
specifically, the fact that ρ⊗N and σ ⊗N commute with SN means that each aY bY × aY bY block is just
the direct sum of aY identical blocks of size bY × bY . So ρ⊗N has a decomposition
 
p1 ρ1
 p2 ρ2 
 
ρ⊗N =  p3 ρ3  (4.32)
 
..
.

in blocks of size
P bY ⊗ bY , with each such block occurring aY times, for all possible Y. (The total number
of blocks is Y aY .) The ρi are density matrices and the pi are nonnegative numbers adding to 1. In
10
See eqn. (6.16) of [17]. One approach to this upper bound is as follows. In general, the highest weight of an
irreducible representation of the group SU (k) is a linear combination of certain fundamental weights with nonnegative
integer coefficients ai , i = 1, · · · , k − 1. In the case of a representation associated to a Young diagram with N boxes, the ai
are bounded by N . The dimension of an irreducible representation with highest weights (a1 , a2 , · · · , ak−1 ) is a polynomial
in the ai of total degree k(k − 1)/2, so if all ai are bounded by N , the dimension is bounded by a constant times N k(k−1)/2 .
One way to prove that the dimension is a polynomial in the ai of the stated degree is to use the Borel-Weil-Bott theorem.
k−1 ai
According to this theorem, a representation with highest weights (a1 , a2 , · · · , ak−1 ) can be realized as H 0 (F, ⊗i=1 Li ),
k−1
where F = SU (k)/U (1) is the flag manifold of the group SU (k) and Li → F are certain holomorphic line bundles.
k−1 ai
Because F has complex dimension k(k − 1)/2, the Riemann-Roch theorem says that the dimension of H 0 (F, ⊗i=1 Li ) is
a polynomial in the ai of that degree.

34
the same basis, σ ⊗N has just the same sort of decomposition:
 
q1 σ1
 q2 σ2 
 
σ ⊗N =  q3 σ3 . (4.33)
 
..
.

We can furthermore make a unitary transformation in each block to diagonalize σ ⊗N . This will generi-
cally not diagonalize ρ⊗N . But because ρ⊗N is block diagonal with relatively small blocks, its entropy
can be usefully compared with that of the diagonal density matrix (ρ⊗N )D that is obtained by setting
to 0 the off-diagonal matrix elements of ρ⊗N in a basis in which σ ⊗N is diagonal within each block and
keeping the diagonal ones:
 
p1 ρ1,D
 p2 ρ2,D 
 
(ρ⊗N )D =  p ρ . (4.34)
 3 3,D 
..
.

One finds then

X
Trρ⊗N log ρ⊗N − Tr(ρ⊗N )D log(ρ⊗N
D ) = pi (S(ρiD ) − S(ρi )). (4.35)
i
P
It is important that a potentially large term i pi log pi cancels out here. Any density matrix on an
n-dimensional space has an entropy S boundedP by 0 ≤ S ≤ log n. Because the sizes of the blocks are
bounded above by bmax ∼ N k(k−1)/2 , and i pi = 1, the right hand side11 of eqn. (4.35) is bounded by
log bmax ∼ 12 k(k − 1) log N, which for large N is negligible compared to N.

Combining this with eqns. (4.27) and (4.29), we see that for large N, a measurement that projects
onto 1-dimensional eigenspaces of σi within each block maps the density matrices ρ⊗N and σ ⊗N to
classical probability distributions R[N ] and S [N ] such that the quantum relative entropy S(ρ⊗N ||σ ⊗N ) and
the classical relative entropy S(R[N ] ||S [N ]) are asymptotically equal. To be more precise, S(ρ⊗N ||σ ⊗N ) =
NS(ρ||σ) is of order N for large N, and differs from S(R[N ] ||S [N ] ) by at most a constant times log N.
In other words  
1 ⊗N ⊗N 1 [N ] [N ] log N
S(ρ||σ) = S(ρ ||σ ) = S(R ||S ) + O . (4.36)
N N N

Once we have identified a measurement that converts the quantum relative entropy (for N copies of
the original system) to a classical relative entropy, we take many copies again and invoke the analysis
of classical relative entropy in section 2.3. In more detail, consider a composite system consisting of
11
The right hand side is actually positive because of the inequality (3.42).

35
N copies of the original system. Suppose that we observe N ′ copies of this composite system (making
NN ′ copies of the original system), for very large N ′ . On each copy of the composite system, we make
the above-described measurement. This means that we sample N ′ times from the classical probability
distribution S [N ] (if the original hypothesis σ was correct) or R[N ] (if the original system was actually
described by ρ). According to the classical analysis in section 2.3, the ability to distinguish between R[N ]
′ [N] [N]
and S [N ] in N ′ trials is controlled by 2−N S(R ||S ) . According to eqn. (4.36), this is asymptotically
′ ⊗N ⊗N ′
the same as 2−N S(ρ ||σ ) = 2−N N S(ρ||σ) . In short, we learn that after a suitable measurement on
Nb = NN ′ copies of the original system, we can distinguish between the hypotheses σ and ρ with a
power
b
2−N S(ρ||σ) , (4.37)
saturating the upper bound (4.26) (with the total number of trials now being N b rather than N). In the
exponent, there are errors of order N ′ log N (from the logarithmic correction in (4.36)) and N log N ′
(coming from the fact that the classical analysis of section 2.3, which for instance used only the leading
term in Stirling’s formula, has corrections of relative order N1′ log N ′ ).

This confirms that quantum relative entropy has the same interpretation as classical relative entropy:
it controls the ability to show, by a measurement, that an initial hypothesis is incorrect. A noteworthy
fact [16] is that the measurement that must be made on the composite system to accomplish this depends
only on σ (the initial hypothesis) and not on ρ (the unknown answer).

At the outset, we assumed monotonicity of relative entropy and deduced from it an upper bound
(4.20) on how well one can distinguish two density matrices in N trials. Actually, now that we know
that the upper bound is attainable, one can reverse the argument and show that this upper bound
implies monotonicity of relative entropy. Suppose that AB is a bipartite system with density matrices
ρAB , σAB that we want to distinguish by a measurement. One thing that we can do is to forget system
B and just make measurements on A. The above argument shows that, after taking N copies, the
reduced density matrices ρA = TrB ρAB , σA = TrB σAB can be distinguished at the rate 2−N S(ρA ||σA) .
But since measurements of subsystem A are a special case of measurements of AB, this implies that
ρAB and σAB can be distinguished at the rate 2−N S(ρA ||σA ) . If therefore we know the bound (4.20),
which says that ρAB and σAB cannot be distinguished at a faster rather than 2−N S(ρAB ||σAB ) , then the
monotonicity inequality S(ρAB ||σAB ) ≥ S(ρA ||σA ) follows. In [18], monotonicity of relative entropy has
been proved by giving an independent proof of the upper bound on how well two density matrices can
be distinguished.

4.3 Encoding Classical Information In A Quantum State

Finally, we will address the following question: how many bits of information can Alice send to Bob by
sending him a quantum system X with a k-dimensional Hilbert space H? (See [4], especially section
10.6, for more on this and related topics.)

36
 One thing Alice can do is to send one of k orthogonal basis vectors in H. Bob can find which one
she sent by making a measurement. So in that way Alice can send log k bits of information. We will
see that in fact it is not possible to do better.

We suppose that Alice wants to encode a random variable that takes the values xi , i = 1, . . . , n with
probability pi . When the value is xi , she writes down this fact in her notebook C and creates a density
matrix ρiX on system X. If |ii is the state of the notebook when Alice has written the value xi , then on
the combined system CX, Alice has created the density matrix
X
ρCX = pi |iihi| ⊗ ρiX (4.38)
i

Then Alice sends the system X to Bob. Bob’s task is to somehow extract information by making a
measurement.

Before worrying about what Bob can do, let us observe that the density matrix ρCX of the system
CX is the one (eqn. (3.56)) that was used earlier in discussing the entropy inequality for mixing. It
is sometimes
P called a classical-quantum density matrix. The reduced density matrix of X is ρX =
TrC ρCX = i pi ρiX . As before, the mutual information between C and X is the Holevo information
X
I(C; X) = S(ρX ) − pi S(ρiX ). (4.39)
i

Since S(ρiX ) ≥ 0 and S(ρX ) ≤ log k, it follows that

I(C; X) ≤ log k. (4.40)

If we knew that quantum mutual information has a similar interpretation to classical mutual information,
we would stop here and say that since I(C; X) ≤ log k, at most log k bits of information about the
contents of Alice’s notebook have been encoded in X. However, we aim to demonstrate that quantum
mutual information behaves like classical mutual information, at least in this respect, not to assume it.
As we will see, what we want is precisely what monotonicity of mutual information says, in the present
context.

What can Bob do on receiving system X? The best he can do is to combine it with some other
system which may include a quantum system Y and a measuring apparatus C ′ . He acts on the combined
system XY C ′ with some unitary transformation or more general quantum channel and then reads C ′ .
The combined operation is a quantum channel. As in our discussion of relative entropy, the outcome of
the channel is a density matrix of the form
r
X
ρC ′ = qα |αihα|, (4.41)
α=1

where |αi are distinguished states of C ′ – the states that one reads in a classical sense. The outcome of
Bob’s measurement is a probability distribution {qα } for a random variable whose values are labeled by

37
α. What Bob learns about the contents of Alice’s notebook is the classical mutual information between
Alice’s probability distribution {pi } and Bob’s probability distribution {qα }. Differently put, what Bob
learns is the mutual information I(C; C ′ ).

To analyze this, we note that before Bob does anything, I(C; X) is the same as I(C; XY C ′ ) because
Y C ′ (Bob’s auxiliary quantum system Y and his measuring apparatus C ′ ) is not coupled to CX. In
more detail, the initial description of the combined system CXY C ′ is by the tensor product of a density
matrix ρCX for CX and a density matrix ρY C ′ for Y C ′ . As one can deduce immediately from the
definitions, the mutual information between C and XY C ′ if the full system CXY C ′ is described by
ρCX ⊗ ρY C ′ is the same as the mutual information between C and X if the subsystem CX is described
by ρCX . Bob then acts on XY C ′ with a unitary transformation, or maybe a more general quantum
channel, which can only reduce the mutual information. Then he takes a partial trace over XY , which
also can only reduce the mutual information, since monotonicity of mutual information under partial
trace tells us that
I(C; XY C ′ ) ≥ I(C; C ′ ). (4.42)
So
log k ≥ I(C; X) = I(C; XY C ′ )before ≥ I(C; XY C ′ )after ≥ I(C; C ′)after , (4.43)
where “before” and “after” mean before and after Bob’s manipulations. More briefly, any way that
Bob processes the signal he receives can only reduce the mutual information. Thus Alice cannot encode
more than log k bits of classical information in an k-dimensional quantum state, though it takes strong
subadditivity (or its equivalents) to prove this.

The problem that we have discussed also has a more symmetrical variant. In this version, Alice and
Bob share a bipartite state AB; Alice has access to A and Bob has access to B. The system is initially
described by a density matrix ρAB . Alice makes a generalized measurement of A and Bob makes
a generalized measurement of B. What is the maximum amount of information that Alice’s results
may give her about Bob’s measurements, and vice-versa? An upper bound is given by the mutual
information I(A; B) in the initial density matrix ρAB . Alice’s measurements amount to a quantum
channel mapping her system A to her measurement apparatus C; Bob’s measurements amount to a
quantum channel mapping his system B to his measurement apparatus C ′ . The mutual information
between their measurement outcomes is simply the mutual information I(C; C ′) in the final state.
Monotonicity of mutual information in any quantum channel says that this can only be less than the
initial I(A; B).

A more subtle issue is the extent to which these upper bounds can be saturated. For an introduction
to such questions see [4], section 10.6.

Research supported in part by NSF Grant PHY-1606531. I thank N. Arkani-Hamed, J. Cotler, B.

Czech, M. Headrick, and R. Witten for discussions. I also thank M. Hayashi, as well as the referees, for
some explanations and helpful criticisms and for a careful reading of the manuscript.

38
References
[1] M. A. Nielsen and I. L. Chuang, Quantum Computation And Quantum Information (Cambridge
University Press, 2000).
[2] T. M. Cover and J. A. Thomas, Elements Of Information Theory (John Wiley & Sons, second
edition, 2006).
[3] M. M. Wilde, Quantum Information Theory (Cambridge University Press, second edition, 2017).
[4] J. Preskill, lecture notes, available at http://www.theory.caltech.edu/~ preskill/ph219/index.html#lec
[5] C. E. Shannon, “A Mathematical Theory Of Communication,” Bell System Technical Journal 27
(1918) 379-423 and 623-656.
[6] M. F. Leifer and R. W. Spekkens, “Towards A Formulation Of Quantum Theory As A Causally
Neutral Theory of Bayesian Inference,” Phys. Rev. A88 (2013) 052130, arXiv:1107.5849.
[7] A. S. Holevo, “Bounds For The Quantity of Information Transmitted By A Quantum Communi-
cation Channel,” Problems of Information Transmission 9 (1973) 177-83.
[8] H. Araki and E. H. Lieb, “Entropy Inequalities,” Commun. Math. Phys. 18 (1970) 160-70.
[9] H. Umegaki, “Conditional Expectation in an Operator Algebra,”’ Kodai Math. Sem. Rep. 14 (1962)
59-85.
[10] E. H. Lieb and M. B. Ruskai, “Proof Of The Strong Subadditivity Of Quantum Mechanical En-
tropy,” J. Math. Phys. 14 (1973) 1938.
[11] E. H. Lieb, “Convex Trace Functions and the Wigner-Yanase-Dyson Conjecture,” Adv. Math. 11
(1973) 267-88.
[12] E. Witten, “Notes On Some Entanglement Properties Of Quantum Field Theory,” Rev. Mod. Phys.
90 (2018) 045003, arXiv:1803.04993.
[13] C. H. Bennett, G. Brassard, C. Crépeau, R. Jozsa, A. Peres, and W. K. Wootters, “Teleporting an
Unknown Quantum State Via Dual Classical and Einstein-Podolsky-Rosen Channels,” Phys. Rev.
Lett. 70 (1993) 1895-9.
[14] M. Horodecki, J. Oppenheim, and A. Winter, “Quantum State Merging And Negative Information,”
Commun. Math. Phys. 269 (2007) 107-36, arXiv:quant-ph/0512247.
[15] F. Hiai and D. Petz, “The Proper Formula For Relative Entropy And Its Asymptotics in Quantum
Probability,” Commun. Math. Phys. 143 (1991) 99-114.
[16] M. Hayashi, “Asymptotics of Quantum Relative Entropy From Representation Theoretical View-
point,” J. Phys. A34 (2001) 3413-20.
[17] M. Hayashi, A Group Theoretic Approach To Quantum Information (Springer, 2017).
[18] I. Bjelakovic and R. Siegmund-Schultze, “Quantum Stein’s Lemma Revisited, Inequalities For
Quantum Entropies, and a Concavity Theorem of Lieb,” quant-ph/0307170.

39