Quantum information theory (MAT4430) Spring 2021

Lecture 4: Equivalence between open and closed quantum systems

Lecturer: Alexander Müller-Hermes

In the previous lecture, we have introduced the formalism of quantum theory. We started
with the formalism of closed quantum systems, and to be able to consider quantum systems
embedded in a larger environment we introduced the formalism of open quantum systems.
In particular, we showed how objects from the formalism of open quantum systems arise
from corresponding objects in the formalism of closed quantum systems. Now, we want
to show thatthe two formalisms are actually equivalent. To do this, we will start with
the objects in the formalism of open quantum systems and represent them using objects
from the formalism of closed quantum systems. On the way, we will derive several central
representation theorems for quantum channels and POVMs.

1 Operator-vector correspondence
To study the properties of bipartite quantum systems, it is surprisingly useful to note the
canonical identification
B(HA , HB ) ' HA ⊗ HB ,
given by the isomorphism |φihψ| 7→ |ψi ⊗ |φi for any |ψi ∈ HA and |φi ∈ HB , and extended
linearly. In the following, we will call this isomorphism the operator-vector correspondence.
We will apply it in two ways: First, we may apply it to operators in B (HA , HB ), and second
we may apply it to operators in B(B(HA ), B(HB )), which are linear maps from B(HA ) to
B(HB ). Here, we understand B(HA ) and B(HB ) as Hilbert-Schmidt inner product spaces.
Historically, the isomorphism got different names in these two settings:

Definition 1.1 (Vectorization and Choi-Jamiolkowski isomorphism). Let HA and HB denote

complex Euclidean spaces.

• The vectorization is the operator-vector correspondence vec : B(HA , HB ) → HA ⊗ HB .

• The Choi-Jamiolkowski isomorphism is the operator-vector correspondence

C : B(B(HA ), B(HB )) → B(HA ⊗ HB ),

and we write CL := C(L) for linear maps L ∈ B(B(HA ), B(HB )). The operator CL is
also called the Choi operator or the Choi matrix of L.

Since these isomorphisms might seem abstract, it is helpful to express them in the com-
C
putational basis. For HA = dA and HB = dB we have C
vec (|iB ihjA |) = |jA i ⊗ |iB i,

for any i ∈ {1, . . . , dB } and any j ∈ {1, . . . , dA } (note that vec is antilinear in the ‘A’ part).

1
Using the lexicographic ordering of the tensor product basis, we have
 
x11
 x21 
 
 .. 
 . 
 
xd 1 
B
vec (X) =  x  ,
 
 12 
 . 
 .. 
 
xd 2 
 B 
..
.
for any X ∈ B(HA , HB ), i.e., the vectorization operation stacks the columns of X on top of
each other. In Matlab, this would correspond to using the command “reshape” transforming
the matrix X to a column vector.
Similarly, we can express the Choi-Jamiolkowski isomorphism in the basis of matrix units.
The matrix units on the space B(B(HA ), B(HB )), thought of as bounded operators between
Hilbert-Schmidt inner product spaces, are given by
h i
{X 7→ |iB ihjB | Tr (|kA ihlA |)† X : i, j ∈ {1, . . . , dB } and k, l ∈ {1, . . . , dA }}.

Under the Choi-Jamiolkowski isomorphism the elements of this basis transform as

 h i
C X 7→ |iB ihjB | Tr (|kA ihlA |)† X = |kA ihlA | ⊗ |iB ihjB |,

for any i, j ∈ {1, . . . , dB } and k, l ∈ {1, . . . , dB }. Therefore, given a linear map L : B(HA ) →
B(HB ), we have
 
L(|1ih1|) L(|1ih2|) · · · L(|1ihdA |)
dA
X L(|2ihdA |) L(|2ih2|) · · · L(|2ihdA |) 
CL = |kA ihlA | ⊗ L (|kA ihlA |) =  ,
 
.. .. ..
k,l=1
 . . . 
L(|dA ih1|) L(|dA ih2|) · · · L(|dA ihdA |)
where we again made a choice for the ordering of the basis on the tensor product space.
There is another useful way of thinking about the canonical isomorphisms, which directly
relates it to quantum theory. For this, we need to define the (unnormalized) maximally
entangled quantum state:
Definition 1.2 (Maximally entangled state). The vector
d
Cd ⊗ Cd,
X
|Ωd i = vec (1Cd ) = |ii ⊗ |ii ∈
i=1

is called the (unnormalized) maximally entangled state. We will write

ωd = |Ωd ihΩd | ∈ B( Cd)+.
Now, the following lemma is easy to see (and you should check this yourself):
Lemma 1.3 (Operator-vector correspondence via maximally entangled state). Consider
C C
complex Euclidean spaces HA = dA and HB = dB , a linear map L : B(HA ) → B(HB )
and an operator X ∈ B(HA , HB ). Then, we have
vec (X) = (1HA ⊗ X) |ΩdA i,
and
CL = (idA ⊗ L) (ωdA ) .

2
 The next lemma proves two identities for the canonical isomorphisms, which will turn
out to be extremely useful:

Lemma 1.4 (Necklace identities). For complex Euclidean spaces HA = Cd

A and HB = Cd B

we have the following identities:

1. For any operator E ∈ B(HA , HB ) we have

(1HA ⊗ E) |ΩdA i = E T ⊗ 1HB |ΩdB i.

2. For any linear map L : B(HA ) → B(HB ) satisfying L(X)† = L(X † ) for every X ∈
B(HA ) we have

(idA ⊗ L) (ωdA ) = (ϑA ◦ L∗ ◦ ϑB ⊗ idB ) (ωdB ) ,

where ϑA : B(HA ) → B(HA ) and ϑB : B(HB ) → B(HB ) denote the transpose maps in
the computational bases of the respective spaces, i.e., ϑA (X) = X T for X ∈ B(HA ).

The name “necklace identities” comes from a certain graphical calculus that can be used
to better understand matrix and vector operations in multilinear settings: In this calculus all
of the basic objects vectors, matrices and higher-order tensors are represented as boxes with
“legs” corresponding to indices that can be summed over. For example, a matrix A ∈ Md
with entries Aij has two indices i and j and is therefore represented as a box with two legs.
We can then join the legs of two boxes to indicate a contraction, i.e., a sum over the entries
of the objects where the joined indices take the same value. Figure 3 introduces the basic
elements of this way of thinking. Using the graphical calculus, we can write the first necklace
identity as in Figure 1.

Figure 1: First necklace identity.

Another useful observation is, that the canonical isomorphism defines an isometry be-
tween different bipartite Euclidean spaces and Hilbert-Schmidt inner product spaces.

Lemma 1.5 (Isometry). For any X, Y ∈ B(HA , HB ) we have

h i
hvec (X) , vec (Y )i = Tr X † Y .

In particular, the operator-vector correspondence is an isometric isomorphism between the

complex Euclidean space HA ⊗HB and the Hilbert-Schmidt inner product space over B(HA , HB ).

The proof of this lemma can be done using the graphical calculus as in Figure 2. Al-
ternatively, the lemma can also be proved by expanding all operators in the computational
basis. If you want to check the lemma that way, please do so.

3
 Figure 2: Proof by picture.

Figure 3: Graphical calculus.

4
2 Choi-Kraus representation of completely positive maps
For any operator K ∈ B(HA , HB ), we define a linear map AdK : B(HA ) → B(HB ) by

AdK (X) = KXK † .

Note that AdK is positive for any K ∈ B(HA , HB ) since

AdK (Y Y † ) = KY Y † K † = (KY )(KY )† ,

and AdK is even completely positive since idE ⊗ AdK = Ad1HE ⊗K for any Euclidean space
HE . We will now show that the maps AdK can be used to express general completely positive
maps and quantum channels.
Recall the canonical Choi-Jamiolkowski isomorphism C : B(B(HA ), B(HB )) → B(HA ⊗
HB ) associating a linear map L : B(HA ) → B(HB ) to its Choi matrix CL . We have:

Theorem 2.1 (Choi and Kraus). For a linear map T : B (HA ) → B (HB ) the following are
equivalent:

1. The map T is completely positive.

2. The Choi matrix CT is positive.

3. There exist operators {Ki }R

i=1 ⊂ B(HA , HB ) and some R ∈ N such that
R
X
T = AdKi . (1)
i=1

Moreover, we can always choose R = rk (CT ) in 3. above.

Proof. To show that 1. implies 2. we use Lemma 1.3 to conclude that

CT = (idA ⊗ T ) (ωdA ) > 0

for any completely positive map T , since ωdA ∈ B (HA ⊗ HA )+ . To show that 2. implies 3.
we apply the spectral decomposition to the Hermitian operator CT . We have
R
X
CT = |ψi ihψi |,
i=1

for R = rk (CT ) and some (unnormalized) vectors |ψi i ∈ HA ⊗ HB . Using the inverse of the
vectorization isomorphism and Lemma 1.3 we have

|ψi i = vec (Ki ) = (1HA ⊗ Ki ) |ΩdA i,

for some operators Ki ∈ B(HA , HB ). Combining the previous equations, we find that
R R
(1dA ⊗ Ki ) ωdA (1dA ⊗ Ki )† =
X X
CT = CAdKi = CPR AdKi .
i=1
i=1 i=1

Applying the inverse of the Choi-Jamiolkowski isomorphism finishes the proof. Showing that
3. implies 1. is easy since each summand in (1) (and hence the whole sum) is completely
positive.

5
 The decomposition (1) is called the Choi decomposition in the mathematics literature,
and the Kraus decomposition in the physics literature1 . Throughout this course, we will call
it the Choi-Kraus decomposition as a compromise. It should be noted that this decomposition
is not unique in general.
Theorem 2.2. Let H denote a complex Euclidean space and consider sets of operators
{Kn }N M
n=1 ⊂ B(H) and {Lm }m=1 ⊂ B(H). The following are equivalent:
1. We have
N
X M
X
AdKn = AdLm .
n=1 m=1

2. There exists a unitary U ∈ U CR
for R = max(M, N ) such that

R
X
Kn = Unm Lm .
m=1

Here the smaller of the two sets is extended by zero operators to make them the same size.
Proof. By the Choi-Jamiolkowski isomorphism this theorem is equivalent to the statement
proved in Exercise 4 on Sheet 2.
As a corollary we obtain the following characterization of quantum channels:
Corollary 2.3 (Characterization of quantum channels). A linear map T : B (HA ) → B (HB )
is a quantum channel if and only if it can be written as
R
X
T = AdKi ,
i=1

for operators {Ki }R

i=1 ⊂ B(HA , HB ) satisfying
R
Ki† Ki = 1HA .
X
(2)
i=1
Proof. By Theorem 2.1, we only have to check that (2) holds if and only if the map T is
trace-preserving. Indeed, this is easy to check since
" R # " R ! #
†
X X †
Tr Ki XKi = Tr Ki Ki X ,
i=1 i=1
for any X ∈ B (HA ) by cyclicity of the trace.
If you are reading these lecture notes in order, then you might remember that the Choi-
Kraus decomposition has already appeared. When computing a concrete example of an
instrument in the third lecture, it appeared in the definition of the linear maps Tn . In light
of the previous discussion, the general definition of instruments should feel more natural
now:
Definition 2.4 (Instruments). A set {Tn }N
P of completely positive maps Tn : B (HA ) →
n=1
HB is called an instrument if the sum T = Nn=1 Tn is a quantum channel.
In this general form, instruments contain both quantum channels (if N = 1) and POVMs
(if dim (HB ) = 1) as special cases. We could therefore choose to formulate quantum theory
in terms of instruments alone. However, we will not do so. Historically, quantum information
theory was build on the notions of quantum channels and POVMs, and for many applications
this is sufficient and considerably simpler. Therefore, we will put emphasize on these concepts
and we keep general instruments in mind for when they are really neccessary.
1
Historically, it was discovered by Choi, but popularized and extended by Kraus.

6
3 Equivalence of closed and open quantum formalism
In the third lecture we have seen how certain pure quantum states, time-evolutions and
PVMs acting on a composite quantum system ‘AE’ give rise to quantum states, quantum
channels and POVMs when restricted to the subsystem ‘A’. Now, we will go in the opposite
direction and show that every quantum state, quantum channel and POVM acting on a
quantum system ‘A’ arises from the corresponding type of object acting on some composite
system ‘AE’. This implies that the postulates of closed quantum systems and the postulates
of open quantum systems are actually equivalent. Mathematically, the extension of an object
to a larger object with some additional structure is called a dilation, and the theorems proved
in this section give some examples of this general principle.

Purifications of quantum states: Is any quantum state the reduced density operator of
some pure state? To answer this question, we will need the following definition:

Definition 3.1 (Purification). A pure quantum state |ψAB ihψAB | ∈ Proj (HA ⊗ HB ) is called
the purification of a quantum state ρA ∈ D (HA ) if we have

ρA = TrB [|ψAB ihψAB |] .

We will now show that any quantum state has a purification. For this we will need the
canonical isomorphism vec : HA ⊗ HB → B(HA , HB ) introduced in the beginning of this
lecture. We have the following theorem:

Theorem 3.2 (Existence of purifications). For a density operator ρA ∈ D (HA ) consider

the pure quantum state |ψA0 A ihψA0 A | ∈ Proj (HA ⊗ HA ) given by
√
|ψA0 A i = vec ( ρA ) ∈ HA ⊗ HA ,
√
where ρA denotes the unique positive square root of the positive operator ρ. Then, we have

ρA = (TrA0 ⊗idA ) [|ψA0 A ihψA0 A |] ,

showing that |ψA0 A ihψA0 A | is a purification of ρ.

Proof. By Lemma 1.5, we have

√ √
hψAA0 |ψAA0 i = h ρA , ρA iHS = Tr [ρA ] = 1.

This shows that |ψAA0 ihψAA0 | is a pure quantum state. Using Lemma 1.3, we have
  √ √
(TrA0 ⊗idA ) [|ψA0 A ihψA0 A |] = TrA0 ⊗ Ad√ρ (ωdA ) = ρA 1HA ρA = ρA ,

since
(TrA0 ⊗idA ) (ωdA ) = 1HA .
This finishes the proof.

The previous theorem shows how to relate each quantum state to a pure state by vec-
torizing a matrix. Since the map vec : HA ⊗ HB → B(HA , HB ) is an isomorphism, we can
also reverse these ideas and go from vectors to matrices: Starting with a pure quantum state
|ψAB ihψAB | ∈ Proj (HA ⊗ HB ), we have

|ψAB i = vec (Xψ ) ,

7
for some operator Xψ ∈ B(HA , HB ) such that the reduced density operators ρA = TrB [|ψAB ihψAB |]
and ρB = TrA [|ψAB ihψAB |] can be expressed as

ρA = XψT Xψ and ρB = Xψ Xψ† .

Let dA = dim (HA ) and dB = dim (HB ) denote the dimensions of the involved Euclidean
√
spaces. When dA = dB , the previous theorem shows that we can choose Xψ = ρB . Even
if dA 6= dB there is a close connection between purifications and reduced density operators.
Next, we will express this in a concrete form: Using the singular value decomposition, we
can then find orthonormal bases {|ai i}di=1
A
and {|bi i}di=1
B
of HA and HB , respectively, such
that
min(dA ,dB )
X p
Xψ = λi |bi ihai |,
i=1
√
R
where we have denoted the singular values of Xψ by λi ∈ + for reasons that will become
clear in a moment. Since vectorization is linear, we have the following theorem:

Theorem 3.3 (Schmidt decomposition). Consider Euclidean spaces HA and HB with dim (HA ) =
dA and dim (HB ) = dB . For any pure state |ψAB i ∈ HA ⊗ HB there exist orthonormal bases
{|ai i}di=1
A
and {|bi i}di=1
B
of HA and HB , respectively, such that

min(dA ,dB )
X p
|ψAB i = λi |ai i ⊗ |bi i,
i=1
√
R
for numbers λi ∈ + called the Schmidt coefficients of |ψAB i. The number of non-zero
Schmidt coefficients is called the Schmidt rank of the pure state |ψAB i.

The Schmidt decomposition makes the correspondence between purifications and reduced
density operators quite clear. In particular, we have the following:

• The rank of a density operator equals the Schmidt rank of its purification, and the
Schmidt rank of a pure state equals the rank of its reduced density operators.

• The minimal dimension of a Euclidean space HE such that there is a purification

|ψAE i ∈ HA ⊗ HE of a given density operator ρA ∈ D (HA ) equals rk (ρA ).

• Any purification |ψAB i ∈ HA ⊗ HB of a density operator ρA ∈ D (HA ) arises from a

purification |ψAE i ∈ HA ⊗ HE for dE = rk (ρA ) as |ψAB i = (1HA ⊗ V )|ψAE i for an
isometry V : HE → HB .

• The eigenvalues of a density operator equals the square roots of the Schmidt coefficients
of its purification, and squaring the Schmidt coefficients of a pure state gives the
eigenvalues of its reduced density operators.

• In particular, the eigenvalues of the two reduced density operators of the same bipartite
pure state coincide.

The last of these facts will turn out to be of particular importance later. We have shown
that every density operator arises as a reduced density operator of some pure state, and that
every pure state arises as the purification of some density operator. We will now see how to
find similar correspondences between the other objects of the formalisms of closed and open
quantum systems.

8
The Naimark and Stinespring dilations: The following theorem is the foundation of
the dilation theorems of both POVMs and quantum channels:

Theorem 3.4 (Naimark-Stinespring dilation). For any set {Kn }Nn=1 ⊂ B(HA , HB ) satisfy-
PN †
ing n=1 Kn Kn = 1HA , there exists a Euclidean space HE with dimension dE = N and an
isometry V : HA → HB ⊗ HE such that the following two statements hold:

1. For any m ∈ {1, . . . , N } we have

V † (1HB ⊗ |mihm|E ) V = Km
†
Km .

2. We have
N
X h i
Kn XKn† = TrE V XV † ,
n=1

for any X ∈ B (HA ) .

In the first case, the isometry V is called the Naimark dilation of the POVM {Kn† Kn }N n=1
and in the second case it is called the Stinespring dilation of the quantum channel with Kraus
operators {Kn }Nn=1 .

Proof. We set HE = CN and define an operator V : HA → HB ⊗ HE by

N
X
V = Kn ⊗ |niE .
n=1

†
Since N
P
n=1 Kn Kn = 1HA the operator V defines an isometry and the two statements in the
theorem can be checked easily.

The dilations in the previous theorem have PN neat† physical interpretations: For a set of
operators {Kn }N
n=1 ⊂ B(H A , H B ) satisfying n=1 Kn Kn = 1HA we denote by T : B(HA ) →
P N †
B(HB ) the quantum channel with T (X) = n=1 Kn XKn for any X ∈ B(HA ), and by
†
{Qn }N +
n=1 ⊂ B(HA ) the POVM given by Qn = Kn Kn .

• If the Euclidean space HE in the Stinespring dilation T (X) = TrE V XV † of a quan-

 

tum channel T : B(HA ) → B(HB ) has dimension dE > 1, then it means that some
information about the system ‘A’ got leaked to the environment system ‘E’.

• We can implement the measurement of {Qn }N +

n=1 ⊂ B(HA ) on any quantum state
ρA ∈ D(HA) by the following process: First, we apply the quantum channel X 7→
TrB V XV † to the state ρA and then we measure the outcome using the PVM
{|nihn|E }N
n=1 .

To make this a bit more precise, we can state the following definition:

Definition 3.5 (Complementary channel). If the isometry V : HA → HB ⊗ HE is the

Stinespring dilation of a quantum channel T : B(HA ) → B(HB ), then the complementary
channel T c : B(HA ) → B(HE ) is given by
 
T c (X) = TrB V XV † .

The complementary channel describes the channel as it is ‘seen by the environment’.

Combining, the two points from above, we get the following physical interpretation of instru-
N †
PNDefine the instrument {Tn }n=1 by Tn (X) = Kn XKn for any X ∈ B(HA ) such that
ments:
T = n=1 Tn . The POVM performed by this instrument is exactly the PVM {|nihn|E }N n=1

9
on the output of the complementary channel T c , and if we forget (or do not know) the out-
come of this measurement the final quantum state can be obtained by applying the channel
T to the input.
We will later express many information-processing tasks in terms of the channel and
its complementary channel together. Intuitively, this becomes clear when considering a
cryptographic protocol involving the quantum channel T : Any information leaked to the
environment (i.e., obtained by applying the complementary channel to the data) could in
principle be retrieved by some eavesdropper. Private communication protocols, therefore,
have to use clever encodings to make the complementary channel as noisy as possible, so
that no useful information can be extracted from it.

Quantum channels and POVMs as interactions with the environment: Does any
quantum channel arise as the reduction of some unitary time-evolution, and does any POVM
arise from some PVM? In this section, we will answer these two questions in the affirmative.
For this, we will apply the Naimark-Stinespring dilation theorem and a trick of how to embed
an isometry into a unitary operator.
To explain the trick, consider an isometry V : HA → HB ⊗ HE and pure quantum
states |ψBE ihψBE | ∈ Proj (HB ⊗ HE ) and |φA ihφA | ∈ Proj (HA ). Now, define two sets
{|xi i}di=1
A
, {|yi i}di=1
A
⊂ HA ⊗ HB ⊗ HE of orthogonal vectors by setting

|xi i = |φA i ⊗ V |iiA ,

and
|yi i = |iiA ⊗ |ψBE i.
Since the two sets of orthogonal vectors have the same size, there exists a unitary operator
U ∈ U (HA ⊗ HB ⊗ HE ) such that

U |yi i = |xi i for any i ∈ {1, . . . , dA },

and we can obtain the isometry V by

(hφA | ⊗ 1HBE ) U (|xiA ⊗ |ψBE i) = V |xi. (3)

We will use this trick in two instances to express quantum channels and POVMs as reductions
of unitary time-evolutions and PVMs, respectively. We start with quantum channels:
Theorem 3.6 (Open system representation of quantum channels). Consider a quantum
channel T : B(HA ) → B(HB ) and denote by HE the Euclidean space appearing in its
Stinespring dilation. For any pure state |ψBE ihψBE | ∈ Proj (HB ⊗ HE ) there exists a unitary
U ∈ U (HA ⊗ HB ⊗ HE ) such that
 
T (X) = TrAE U (X ⊗ |ψBE ihψBE |) U † ,

for any X ∈ B (HA ).

Proof. By Theorem 3.4, we
can consider the Stinespring dilation V : HA → HB ⊗ HE such
that T (X) = TrE V XV † for any X ∈ B (HA ). Let U ∈ U (HA ⊗ HB ⊗ HE ) denote a
unitary operator such that (3) holds with |ψBE i and some pure state |φA ihφA | ∈ Proj (HA ).
It is then easy to check that
h i h i
TrAE U (|iihj|A ⊗ |ψBE ihψBE |) U † = TrAE |φA ihφA | ⊗ V |iihj|A V † = T (|iihj|A ) ,

for any i, j ∈ {1, . . . , dA } and the proof is finished by linearity.

10
 The previous theorem shows that any quantum channel can be written in the concrete
form introduced in Lecture 3 for some unitary time-evolution and some environment state.
Next, we show the same statement for POVMs:
+
Theorem 3.7 (Environment induced measurements). Consider a POVM {Qi }N n=1 ⊂ B (HA )
and denote by HE the Euclidean space appearing in its Naimark dilation. For any pure state
|ψE ihψE | ∈ Proj (HE ) there exists a PVM {Pn }N
n=1 ⊂ Proj (HA ⊗ HE ) such that

Tr [Qn ρA ] = Tr [Pn (ρA ⊗ |ψE ihψE |)] ,

for any n ∈ {1, . . . , N } and any ρA ∈ D (HA ) .

Proof. By Theorem 3.4, we can consider the Naimark dilation V : HA → HA ⊗ HE such that

Qn = V † (1HA ⊗ |nihn|E ) V,

for any n ∈ {1, . . . , N }. Let U ∈ U (HA ⊗ HE ) denote a unitary operator such that (3) holds
with the pure state |ψE ihψE | ∈ Proj (HE ) (where the system ‘B’ has dimension 1), i.e., such
that
U (|xA i ⊗ |ψE i) = V |xA i,
for any |xA i ∈ HA . For each n ∈ {1, . . . , N } we define the projector

Pn = U † (1HA ⊗ |nihn|E ) U.

Then, we compute
h i
Tr [Pn (ρA ⊗ |ψE ihψE |)] = Tr (1HA ⊗ |nihn|E ) U (ρA ⊗ |ψE ihψE |) U †
h i
= Tr (1HA ⊗ |nihn|E ) V ρA V †
= Tr [Qn ρA ] .
PN
It is easy to check that n=1 Pn = 1HA ⊗ 1HE , and the proof is finished.

11