Chapter 4

Open quantum systems

4.1 Quantum operations

Let’s go back for a second to the basic postulates of quantum mechanics. Recall
that when we first establish the theory, we begin by postulating that a system can be
represented by an abstract state | i. Then we also postulate that the time evolution of
| i must be given by a map which is (i) linear and (ii) preserves probability, h t | t i =
const. This is the entry point for the unitaries: any evolution in quantum mechanics
can be represented by a unitary operator

| i ! | 0 i = U| i. (4.1)

However, after a while we realized that the state | i is not the most general state of a
system. Instead, the general state is the density matrix ⇢.
We can then rethink the evolution postulate: what is the most general evolution
which is (i) linear and (ii) maps density matrices into density matrices? We already
saw that unitary evolutions are translated to density matrices as maps of the form

⇢ ! ⇢0 = U⇢U † . (4.2)

This is certainly a linear map and if ⇢ is a valid density matrix, then so will ⇢0 . But is it
the most general kind of map satisfying these properties? The answer is no. The most
general map is actually called a quantum operation, E(⇢), and has the form:

X X
⇢ ! ⇢0 = E(⇢) = Mk ⇢Mk† , with Mk† Mk = 1. (4.3)
k k

This way of representing the map E(⇢) in terms of a set of operators Mk is called the
operator-sum representation. If there is only one Mk then it must be unitary and
P
we recover (4.2). A set of operators {Mk } satisfying k Mk† Mk = 1 are called Kraus
operators.

104
 The take-home message I want you to keep is that quantum operations are the most
general evolution map a density matrix can have. This chapter will be all about quan-
tum operations and their ramifications, so we will have quite a lot to discuss about this.
But for now let us start slow. In this section we will do two things: first I will show you
that quantum operations are the natural language for describing open quantum systems.
Any evolution of a system connected to an external environment can be written as a
quantum operation. Second, we will prove the claim surrounding Eq. (4.3); that is, that
any linear map which takes density matrices into density matrices can be written in the
form (4.3).

Example: amplitude damping

Consider a qubit system and let
! p !
1 p0 0
M0 = , M1 = , (4.4)
0 1 0 0

with 2 [0, 1]. This is a valid set of Kraus operators since M0† M0 + M1† M1 = 1. Its
action on a general qubit density matrix reads:
0 1 0 p 1
BBB p q CCC BBB + p(1 ) q 1 CCC
⇢ = BB@ CA ! ⇢ = BBB
C 0
p CCC . (4.5)
q⇤ 1 p @ A
q⇤ 1 (1 )(1 p)

If = 0 nothing happens, ⇢0 = ⇢. Conversely, if = 1 then

!
1 0
⇢ ! ⇢0 = . (4.6)
0 0

This is why this is called an amplitude damping: no matter where you start, the p map
tries to push the system towards |0i. It does so by destroying coherences, q ! q 1 ,
and by a↵ecting the populations, p ! +p(1 ). The larger the value of , the stronger
is the e↵ect.

Amplitude damping from a master equation

Consider a quantum master equation of the form
d⇢  1
= +⇢ { + , ⇢} . (4.7)
dt 2
We have briefly touched upon this type of equation in Secs. 2.2 and 3.7. And we will
have a lot more to say about it below. Applying this equation to a general density
matrix yields the pair of equations
dp t t
= (1 p) ! p(t) = p0 e + (1 e ),
dt
dq q t/2
= ! q(t) = q0 e .
dt 2

105
Comparing this with Eq. (4.5) we see that the solution of the di↵erential Eq. (4.7) can
be viewed, at any given time t, as a map
X
⇢(t) = Mk ⇢(0)Mk† , (4.8)
k

with the same Kraus operators (4.4) and

t
=1 e . (4.9)
If t = 0 then = 0 and nothing happens. If t ! 1 then ! 1 and the system collapses
completely towards |0i, as in Eq. (4.6).

Amplitude damping from system-environment interactions

Let us now label our system S and suppose it interacts with an environment ancilla
E by means of the Hamiltonian
S E S E
H = g( + + + ), (4.10)
where g is some parameter. The corresponding unitary evolution matrix will be
0 1
BBB1 0 0 0CC
C
BBB0 cos gt i sin gt 0CCCC
U = e iHt = BBBB C. (4.11)
B@B0 i sin gt cos gt 0CCCAC
0 0 0 1
Suppose that the ancila starts in the state |0iE whereas the system starts in an arbitrary
state ⇢S . Then we compute

⇢0S E = U ⇢S ⌦ |0iE h0| U † ,

and finally take the partial trace over E to obtain ⇢0S = trE ⇢0S E . I will leave this task for
you as an exercise. The result is
0 1
BB p + (1 p) sin2 (gt) q cos(gt) CCC
⇢0S = BBB@ CCA . (4.12)
q⇤ cos(gt) (1 p) cos2 (gt)
Comparing this with the amplitude damping result (4.5) we see that this is also a quan-
tum operation, again with the same Kraus operators (4.4), but with
= sin2 (gt). (4.13)
Thus, the evolution of two qubits, when viewed from the perspective of only one of
them, will behave like a quantum operation. But unlike in the master equation example
above, here the amplitude damping parameter will not increase monotonically, but
will rather oscillate in time. If you happen to interrupt the evolution when gt is an
integer multiple of ⇡ then it will look like a complete damping. But if we wait a
bit longer it will seem that less damping occurred. This is what happens when the
environment is small (in this case it is only one qubit). If your environment had 1023
qubits, which is what Eq. (4.7) tries to model, you would not observe these revivals.

106
Amplitude damping and spontaneous emission
The amplitude damping process is also what happens if you have an atom in the
excited state interacting with the electromagnetic vacuum. In this case, the atom may
fall down to the ground-state and emit a photon, a process we call spontaneous emis-
sion. To have a toy model to describe this, suppose that the atom only interacts with
one mode of the electromagnetic field, whose frequency ! matches that of the atom
⌦. In that case the Hamiltonian reduces to the resonant Jaynes-Cummings model [cf.
Eq. (3.23)].
⌦
H = ⌦a† a + z + g(a
†
+ a + ). (4.14)
2
In the resonant case we can move to the interaction picture and still get a time-independent
Hamiltonian
H̃ = g(a† + a + ). (4.15)
Suppose now that the electromagnetic mode starts in the vacuum, |0iE , whereas the
atom starts in an arbitrary state ⇢S . In principle, this Hamiltonian will act on the full
Hilbert space, which is spanned by |0, niS E and |1, niS E , where n = 0, 1, 2, . . . is the
number of photons in the mode a. But since the Jaynes-Cummings Hamiltonian pre-
serves the total number of quanta [Eq. (3.24)] and since the electromagnetic mode
started in the vacuum, at any time there will be either 0 or 1 photons in the mode.
Thus, the only basis elements that will matter to us are |0, 0iS E , |0, 1iS E and |1, 0iS E .
The matrix elements of H̃ in these states are
0 1
BBB0 0 0CCC
BB C
H̃ = BB0 0 gCCCC .
B
@ A
0 g 0

Hence, the time-evolution operator will be

0 1
BBB1 0 0 CCC
B C
U=e iH̃t
= BBBB0 cos(gt) i sin(gt)CCCC . (4.16)
@ A
0 i sin(gt) cos(gt)

I wrote down this result just so you could have a look at it. But the truth is we don’t
need it. Since we are restricting the dynamics to this sub-space, the problem is exactly
identical to that generated by the Hamiltonian (4.10) (except for a phase factor, which
makes no di↵erence). Indeed, if you now repeat the steps of computing ⇢0S E and then
⇢0S , you will find as a result exactly the state (4.12).
This example serves to show that many Hamiltonians may lead to the same quan-
tum operation. The quantum operation describes a dynamical evolution from the per-
spective of the system’s density matrix and has no information on what exactly gen-
erated that evolution. It could have been one qubit, one electromagnetic mode, 1023
water molecules in a bucket of water or a swarm of killer bees armed with machine
guns. From the perspective of the map, they may all lead to the same result.
The above paragraph is a common source of confusion. You may immediately
protest and say “How can a one qubit environment lead to the same evolution as a
1023 -atom environment?”. They don’t! They lead to the same map, not the same

107
evolution. That’s the point. If we analyze the evolution as a function of time, both will
be completely di↵erent. But if we are only interested in the map that takes you from
one state to another, then this map can be engineered by a single qubit or by 1023 of
them.

Proof of the operator-sum representation

After this warm-up, we are now ready to prove Eq. (4.3). But let us be very precise
on what we want to prove. We define E(⇢) as a map satisfying
1. Linearity: E(↵⇢1 + ⇢2 ) = ↵E(⇢1 ) + E(⇢2 ).
2. Trace preserving: tr[E(⇢)] = tr(⇢).
3. Completely positive: if ⇢ 0 then E(⇢) 0.
There is a subtle di↵erence between a map that is positive and a map that is completely
positive. Completely positive means E(⇢) 0 even if ⇢ is a density matrix living in
a larger space than the one E acts on. For instance, suppose E acts on the space of a
qubit. But the ⇢ it is acting on could mean the density matrix of 2 entangled qubits,
even though the map acts on only one of them. If even in this case the resulting ⇢0 is
positive semi-definite, we say it is completely positive.1 A map satisfying properties 1,
2 and 3 above is called a Completely Positive Trace Preserving (CPTP) map.
Our goal is now to show that any CPTP map can be written as an operator-sum
representation [Eq. (4.3)] for some set of operators {Mk }. The proof of this claim
is usually based on a powerful, yet abstract, idea related to what is called the Choi-
Jamiolkowski isomorphism. Let S denote the space where our map E acts and define
an auxiliary space R which is an exact copy of S. Define also the (unnormalized) Bell
state X
|⌦i = |iiR ⌦ |iiS , (4.17)
i
where |ii is an arbitrary basis and from now on I will always write the R space in the
left and the S space in the right. We now construct the following operator:

⇤E = (IR ⌦ ES )(|⌦ih⌦|). (4.18)

This is called the Choi matrix of the map E. Note how it is like a density operator. It
is the outcome of applying the map ES on one side of the maximally entangled Bell
state of R+S.
The most surprising thing about the Choi matrix is that it completely determines
the map E. That is, if we somehow learn how our map E acts on |⌦ih⌦| we have
completely determined how it will act on any other density matrix. This is summarized
by the following formula:
⇢
E(⇢) = trR (⇢T ⌦ IS )⇤E . (4.19)
1 There aren’t many examples of maps that are positive but not completely positive. The only example I

know is the partial trace (see, for instance, Box 8.2 of Nielsen and Chuang).

108
I know what you are thinking: this is really weird! Yes, it is. But it is true. Note that
here ⇢T is placed on the auxiliary space R in which the trace is being taken. Conse-
quently, the result on the left-hand side is still an operator living on S. To verify that
Eq. (4.19) is true we first rewrite (4.18) as
X
⇤E = |iiR h j| ⌦ E(|iih j|). (4.20)
i, j

Then we get
⇢ X ⇢ 
trR (⇢T ⌦ IS )⇤E = trR (⇢T ⌦ IS ) |iih j| ⌦ E(|iih j|)
i, j
X
= h j|⇢T |iiE(|iih j|)
i, j
✓X ◆
=E ⇢i, j |iih j|
i, j

= E(⇢).
T
Here I used the fact that h j|⇢ |ii = hi|⇢| ji = ⇢i, j . Moreover, I used our assumption that
E is a linear map.
We are now in the position to prove our claim. As I mentioned, the Choi matrix
looks like a density matrix on R+S. In fact, we are assuming that our map E is CPTP.
Thus, since |⌦ih⌦| is a positive semi-definite operator, then so will ⇤E (although it will
not be normalized). We may then diagonalize ⇤E as
X
⇤E = k | k ih k |,
k

where | k i are vectors living in the big R+S space and k 0. For the purpose of what
we are going to do next, it is convenient to absorb the eigenvalues into the eigenvectors
(which will no longer be normalized) and define
X p
⇤E = |mk ihmk |, |mk i = k | k i, (4.21)
k
p
Note that here CPTP is crucial because it implies that k 0 so that hmk | = h k | k .
To finish the proof we insert this into Eq. (4.19) to get
X ⇢
E(⇢) = trR (⇢T ⌦ IS )|mk ihmk | . (4.22)
k

The right-hand side will still be an operator living in S, since we only traced over R. All
we are left to do is convince ourselves that this will have the shape of the operator-sum
representation in Eq. (4.3).
To do that things will get a little nasty. The trick is to connect the states |mk i
of the Choi matrix ⇤E with the Kraus operators Mk appearing in the operator-sum
representation (4.3): X
E(⇢) = Mk ⇢Mk† .
k

109
This is done by noting that since |mk i lives on the R+S space, it can be decomposed as
X
|mk i = (Mk ) j,i |iiR ⌦ | jiS , (4.23)
i, j

where (Mk ) j,i are a set of coefficients which we can interpret as a matrix Mk . To estab-
lish this connection we first manipulate (4.22) to read
XX
T
E(⇢) = R hi|⇢ | jiR R h j|mk ihmk |iiR .
k i, j

Then we insert Eq. (4.23) to find

XXX
E(⇢) = ⇢ j,i (Mk ) j0 , j (Mk⇤ )i0 ,i | j0 ihi0 |
k i, j i0 , j0
XXX
= | j0 ih j0 |Mk | jih j|⇢|iihi|Mk† |i0 ihi0 |.
k i, j i0 , j0
X
= Mk ⇢Mk† ,
k

and voilá!
In conclusion, we have seen that any map which is linear and CPTP can be de-
scribed by an operator-sum representation, Eq. (4.3). I like this a lot because we are
not asking for much: linearity and CPTP is just the basic things we expect from a
physical map. Linearity should be there because everything in quantum mechanics is
linear and CPTP must be there because the evolution must map a physical state into a
physical state. When we first arrived at the idea of a unitary, we were also very relaxed
because all we required was the conservation of ket probabilities. The spirit here is the
same. For this reason, the quantum operation is really just a very natural and simplistic
generalization of the evolution of quantum systems, using density matrices instead of
kets.

4.2 Stinespring dilations

In the previous section we defined quantum operations based on the idea of a gen-
eral map that takes density matrices to density matrices. We also showed that these
maps may arise in di↵erent circumstances, such as from a master equation or from the
unitary interaction of a qubit with a one-qubit environment. This last idea is very pow-
erful and is related to the concept of a dilation. That is, the representation of a quantum
operation as larger unitary between our system and some environment, as illustrated in
Fig. 4.1. It turns out that this dilation idea is always possible and it works in both ways:
• Given a S+E unitary, the corresponding map in terms of S will be given by a
quantum operation.
• Given a quantum operation, we can always find a global S+E unitary represent-
ing it (in fact, there is an infinite number of such unitaries, as we will see).

110
 ρ ℰ(ρ)

�
ρ�

Figure 4.1: Idea behind a Stinespring dilation: a quantum operation E(⇢) can always be con-
structed by evolving the system together with an environment, with a global unitary
U, and then discarding the environment.

More precisely, a dilation is described as follows. Our quantum system, with den-
sity matrix ⇢, is put to interact via a global unitary U with an environment (which can
be of any size) having an initial density matrix ⇢E . After the interaction we throw
away the environment. The result, from the perspective of the system, is a quantum
operation. This can be summarized by the expression:

⇢
E(⇢) = trE U(⇢ ⌦ ⇢E )U † . (4.24)

We will now demonstrate that this is indeed a quantum operation.

Top-down, easy case

Let |ek i denote a basis for the enviroment. To warm up assume the initial state of
the environment is pure, ⇢E = |e0 ihe0 |. Then Eq. (4.24) becomes
X
E(⇢) = hek |U⇢|e0 ihe0 |U † |ek i,
k

which is similar to a calculation we did in Sec. 2.10. Since ⇢ and |e0 i live on di↵erent
Hilbert spaces, we may define2

Mk = hek |U|e0 i, (4.25)

with which we arrive at the usual formula for a quantum operation.

X
E(⇢) = Mk ⇢Mk† . (4.26)
k
2 Remember that what this formula really means is
 
Mk = 1S ⌦ hek | U 1S ⌦ |e0 i .

111
We can also check that the Mk in Eq. (4.25) form a valid set of Kraus operators:
X X
Mk† Mk = he0 |U † |ek ihek |U † |e0 i = 1
k k

Each term in this sum cancelled sequentially: first a completeness relation of the |ek i,
then the unitarity of U, then he0 |e0 i = 1. The result is still an identity on the space of S.

Top-down, general case

It turns out that the assumption that the environment started in a pure state is not at
all restrictive. After all, we can always purify the mixed state ⇢E . That is, we can always
say the environment actually lives on a larger Hilbert space in which its state is pure.
Notwithstanding, it is still useful, from a practical point of view, to generalize (4.25)
for general mixed states. In this case the trick is to choose the environment basis |ek i
as the eigenbasis of ⇢E . That is,
X
⇢E = pk |ek ihek |.
k

We now write Eq. (4.24) as

X
E(⇢) = hek |U⇢pq |eq iheq |U † |ek i.
k,q

And, instead of (4.25), we define the Kraus operators as

p
Mk,q = pq hek |U|eq i. (4.27)

Then the map becomes X

†
E(⇢) = Mk,q ⇢Mk,q . (4.28)
k,q

At first it seems we are cheating a bit because we have two indices. But if we think
about (k, q) as a collective index ↵, then we go back to the usual structure of the quan-
tum operation.

Bottom-up
Now let’s go the other way around. Suppose we are given a quantum operation
of the form (4.26), with a given set of Kraus operators {Mk }. We then ask how to
construct a global S+E unitary with some environment E, such as to reproduce this
quantum operation. That turns out to be quite simple.
First let us ask what should be the dimension of the environment. If we had a vector
of dimension d, we all now that the most general linear operation would be given by
a d ⇥ d matrix. In our case our system has dimensions d, but we want operations on
a density matrix, which is already a d ⇥ d matrix. However, recall that matrices also
form a vector space, so the quantum operation can be thought of as an operation on a
vector with d2 entries. The only point is that this vector is displaced like a matrix, so

112
things become messy because we have to multiply it on both sides. Notwithstanding,
we can infer from this argument that we need at most d2 Kraus operators Mk in order to
fully describe a quantum operation. But we have already seen from Eq. (4.25) that the
number of k values is related to the number of basis elements |ek i of the environment.
Hence, we conclude that any quantum operation on a d-dimensional system may be
reproduced by a dilation with an environment of dimension d2 . This fact is quite re-
markable. In many cases we are interested in what happens when a system S interacts
with a very large environment E. But this argument shows that, as far as the map is
concerned, we can always reproduce it with an environment that is only d2 .
Suppose now that the environment starts in some state |e0 i. We then construct a
unitary U such as to obtain the Kraus operators in Eq. (4.25). This unitary is more
easily mocked up if we consider the Hilbert space structure HE ⌦ HS (that is, the
environment on the left). Then the unitary that does the job can be written in Block
form as 0 1
BBB M0 . . . . . . . . .CC
C
BBBB M1 . . . . . . . . .CCCC
BBBB M C
. . . . . . . . .CCCCC .
U = BBB 2 (4.29)
BBB . CCCC
BBB .. . . . . . . . . . CCC
@ A
Md 2 1 . . . . . . . . .
where the remainder of the matrix should be filled with whatever it needs to make U
an actual unitary. The reason why this works is actually related all the way back to
the matrix definition of the Kronecker product, Eq. (2.43). The operator Mk is just the
matrix element Uk,0 in the basis of the environment.
As an example, consider the unitary in the two-qubit example (4.11). In this case
the left blocks are
! !
1 0 0 i sin(gt)
M0 = , M1 = .
0 cos(gt) 0 0

This is the same as the amplitude damping Kraus operators in Eq. (4.4), with =
sin2 (gt) [Eq. (4.13)]. There is an extra weird factor of i, but that doesn’t matter because
it vanishes when we do M1 ⇢M1† .

Interpretation in terms of measurements

There is a nice way to picture a quantum operation within this Stinespring dilation
setting. You of course noticed that what we are doing here is somewhat similar to the
generalized measurement scenario discussed in Sec. 2.10. In fact, there we said that a
generalized measurement was also described by a set of Kraus operators {Mk } and was
such that the probability of obtaining measurement outcome k was

pk = tr(Mk ⇢Mk† ).

Moreover, if outcome k was obtained, the state would collapse to

Mk ⇢Mk†
⇢ ! ⇢k = .
pk

113
We can therefore interpret a quantum operation of the form (4.26) as
X X
E(⇢) = Mk ⇢Mk† = pk ⇢k .
k k

That is, we can view it as just a random sampling of states ⇢k with probability pk .
The total e↵ect of the quantum operation is a convex combinations of the possible
outcomes with di↵erent probability weights. Of course, we don’t really need to do a
measurement. Is just how the system behaves from the eyes of S.

Freedom in the operator-sum representation

There is a reason why we distinguish between the terms “quantum operation” and
“operator-sum representation”. As the name of the latter implies, when we write a
quantum operation in terms of the Kraus operators, like in Eq. (4.26), we are really
introducing a representation for the map. And the point I wish to make now is that
this representation is not unique: there is a freedom in how we choose the Kraus oper-
ators which lead to the same quantum operation. The same happens for unitaries: two
unitaries U and U 0 = ei✓ U are physically equivalent so multiplying by a global phase
changes nothing. For quantum operations the freedom is even larger.
Let {Mk } be a set of Kraus operators and consider the quantum operation (4.26).
Now define a new set of Kraus operators {N↵ } as

X X
⇤
N↵ = V↵,k Mk , Mk = V↵,k N↵ , (4.30)
k ↵

where V is a unitary matrix.3 Substituting (4.30) in (4.26) we find

X
E(⇢) = ⇤
V↵,k V ,k N↵ ⇢N † .
k,↵,

The trick now is to do the sum over k first. Since V is unitary

X X
⇤
V↵,k V ,k = V ,k (V † )k,↵ = ,↵ .
k k

Hence we conclude that

X X
E(⇢) = Mk ⇢Mk† = N↵ ⇢N↵† . (4.31)
k ↵

Thus, two sets of Kraus operators connected by a unitary transformation lead to the
same quantum operation. It is cool that this even works when the two sets have a
3 I know this can sound strange at first. Here M are operators (maybe there are 7 of them). But we can
k
arrange them to form a list. What we are doing is writing each element in this list as a linear combination
of another set of operators N↵ . However, we are choosing the coefficients of this linear combinations Vk,↵ to
form a unitary matrix, VV † = V † V = 1.

114
di↵erent number of elements. For instance, suppose {Mk } has 5 elements, M0 , . . . , M4 ,
and {N↵ } has 3 elements, N0 , . . . , N2 . Then we can add to the list {N↵ } two zero elements
N3 = 0 and N4 = 0. Now both have the same number of elements and we can construct
a unitary connecting the two sets.
The next interesting question is what is the origin of this freedom. It turns out it
is related to local operations on the environment. Recall that, as shown in Eq. (4.25),
Mk = hek |U|e0 i. Now suppose that before we finish the evolution, we perform a unitary
VE ⌦ 1S on the environment. Then the new set of Kraus operators will be
X X
N↵ = he↵ |(V ⌦ 1)U|e0 i = he↵ |V|ek ihek |U|e0 i = V↵,k Mk ,
k k

which is exactly Eq. (4.30). Thus, we can view this freedom of choice as a sort of
“post-processing” on the environment, which has no e↵ect on the system.

Partial trace as a quantum operation

So far we have considered quantum operations that map a given Hilbert space to
the same space. However, the entire framework generalizes naturally to maps taking
a density matrix in a given subspace H1 to another subspace H2 . In this case all that
changes is that the condition on Kraus operators become
X
Mk† Mk = I1 (4.32)
k

That is, with the identity being on the space H1 . An example of such an operation is
the partial trace. Suppose our system S is actually a bipartite system AB. The partial
trace over B is written, as we know, as
X X
trB (⇢) = hbk |⇢|bk i = (1A ⌦ hbk |)⇢(1A ⌦ |bk i). (4.33)
k k

If we define the Kraus operators

Mk = 1A ⌦ hbk |, (4.34)

then the partial trace can be identified with the quantum operation
X
trB (⇢) = Mk ⇢Mk† . (4.35)
k

Moreover we see that

X X
Mk† Mk = 1A ⌦ |bk ihbk | = 1AB .
k k

That is, the identity on the original space.

115
 We can also do the opposite. That is, we can define a quantum operation which
adds a state to the system. For instance, suppose we have a system S and we want to
add an environment ancilla E in a state |e0 i. Then we can define the Kraus operators

M0 = 1S ⌦ |e0 iE . (4.36)

The corresponding quantum operation will then be

M0 ⇢M0† = ⇢ ⌦ |e0 ihe0 |. (4.37)

Moreover,
M0† M0 = 1S .
Of course, if we want to add an ancilla in a more general state, all we need to do is
construct a larger set of Kraus operators. With these ideas we can actually cover all
types of quantum operations. That is, any map can always be described by quantum
operations mapping the same Hilbert space, combined with partial traces and adding
ancillas.

116