6.443/8.371/18.

436 Quantum Information Science Spring 2017

Lecture 2: February 9, 2018

Lecturer: Quantum states and operations Scribe:

2.1 Overview

The dynamics of a closed quantum system is unitary. In other words, if initially the state of a quantum
system is |ψi then in a closed system this can evolve to U |ψi for some unitary matrix U . In terms of density
matrices this is equivalent to the map |ψi hψ| 7→ U |ψi hψ| U † . It is known that quantum mechanics never
creates or destroys information, but this is only true for closed systems.

2.2 Three Definitions of Quantum Operations

Today we discuss how to generalize quantum operation beyond unitaries. Along with unitary evolution, in
general we can add or discard a quantum system

• Add a system: ρ 7→ ρ ⊗ σ
• Discard a subsystem: ρAB 7→ ρA = trB (ρAB )

The latter operation is also known as partial trace:

X
ρA = (IA ⊗ hb|)ρAB (IA ⊗ |bi) = (id ⊗ trB )ρAB
b

Notation: we need the following notation. If A and B are vector spaces, then L(A), L(B) are the space of
linear operations on A and B, and moreover L(A, B) is the space of all linear operations : A → B. We can also
define L(L(A), L(B)) as the space of linear operations : L(A) → L(B). For example the mentioned partial
trace belongs to L(L(A ⊗ B), L(A)). In physics we refer to L(L(A), L(B)) as the space of super-operators.

2.2.1 First Form (Stinespring)

We claim that the only source of non-unitarity is adding or discarding quantum systems. In particular, any
quantum operation can be described by some composition of (1) unitaries (2) adding a system (3) partial
trace.
An isometry V is a map in L(CdA , CdB ) for dB ≥ dA such that V † V = IdA . Note for any sate |ψi,
k |ψi k = kV |ψi k, ∀ |ψi ∈ A. Furthermore V V † = IB iff dA = dB . A key example of an isometry is the map
V |ψi = |ψi ⊗ |0i. In terms of density matrices an isometry maps ρ 7→ V ρV † . For example the previous
example adds V ρV † = ρ ⊗ |0i h0|.
We can combine unitarity and adding quantum systems with isometries. To complete the picture, we have:

• Isometries: N (ρ) = V ρV † , where V is an isometry, including unitaries, adding pure ancillas, etc

2-1
2-2 Lecture 2: February 9, 2018

• Partial Trace: N (ρAB ) = trB (ρAB )

The next simplest thing we can do after the two primitive operations above is combine them: we apply an
isometry and then take a partial trace.
N (ρ) = trE (V ρV † ) (Form 1)

Here, V maps a system in A on to the space B ⊗ E, and the partial trace throws out E, which could be the
environment.
Even though this is simple, we claim it is fully general. This will be further justified later.
Another important thing to notice about this is that it is a composable operation. The composition of two
isometries is another isometry, and taking the partial trace over two systems, E1 and E2 , is just like taking
the partial trace over some larger system. So if we can justify rearranging the operations, we can describe
two quantum operations in Form 1 as a single operation in Form 1.
Church of the Larger Hilbert Space: Whenever you do a partial trace, you can always do it later. Trace just
means I commit to never looking at it again. It is all the same to do it now or later.
So we have the equivalence of the following two operations:

2.2.2 Second Form (Kraus operators)

From
P this perspective, we have another method for looking at channels. We can write isometries as V =
e e ⊗ |ei, where {|ei} is an orthonormal basis and each Ve ∈ L(A, B). This allows Form 1 to be written
V
as:

X
Ve1 ρVe†2 ⊗ |e1 ihe2 |

N (ρ) = trE
e1 ,e2

Evaluating the partial trace yields:

X
N (ρ) = Ve ρVe† (Form 2)
e

This allows us to describe the channel with no reference to the environment. Form 2 is called the Kraus
Decomposition, where Ve are the Kraus Operators of the channel and V is still an isometry.
Lecture 2: February 9, 2018 2-3

What Ve are allowed here? We know their size from the fact that Ve ∈ L(A, B). We also know they are
blocks from a larger matrix, and that larger matrix is an isometry.
Remember the isometry condition I = V † V . This can be expressed as a condition on all of the blocks by
writing it out:

X X
I=( Ve†1 ⊗ he1 |)( Ve2 ⊗ |e2 i)
e1 e2

This gives us the Kraus Operator Conditions:

X
Ve† Ve = I (2.1)
e

It is important to note that these are each positive semidefinite operators, so you don’t get any cancellation.
This also works in reverse: if a set of {Ve } satisfy the Kraus Operator Conditions, then the matrix V satisfies
the isometry condition. That is, if we have a channel in form Form 2, it can also be written in Form 1.
These forms are equivalent.
Here we give some concrete examples of Kraus operators

1. N (ρ) = U ρU † with a single Kraus operator U .

√
2. N (ρ) = e pe Ue ρUe† with the set of Kraus operators { pe Ue }e .
P

3. N (ρAB ) = trρAB with the set of Kraus operators {Vb = IA ⊗ hb|}b .

4. Cooling: N (ρ) = |0i h0| for qubit input. The set of Kraus operators is {|0i h0| , |0i h1|}.
5. Noise
√ Map: A bit flip with probability p. N (ρ) = (1 − p)ρ + pXρX. The Kraus operators are simply
√
( 1 − p)I and pX.
6. Phase Damping: N (ρ) = (1 − p)ρ + pZρZ, with similar Kruas operators.
7. Depolarizing Channel: N (ρ) = (1 − p)ρ + p4 (ρ + XρX + Y ρY + ZρZ) = (1 − p)ρ + p2 I. In this
channel, with probability p, the state is thrown away and replaced with the maximally mixed state.
8. Amplitude damping (problem set)

Vb† Vb =
P P
Note for the partial trace (item 3) we have the normalization b b IA ⊗ |biB hb| = IA ⊗ IB .
As a practical note, in a lot of physical systems that represent qubits, if we define it in terms of energy level,
having a phase error is more likely than a bit flip. Maximal phase damping occurs when p = 12 and this
corresponds to measurement, which we will now discuss.

2.2.3 Third Form: axiomatic approach

The previous two forms have given models of quantum channels, but we have not yet shown that they can
represent any general quantum operation. For this form, we will start from another perspective by asking
”what properties should a general quantum operator satisfy?”

0. Hermeticity Preserving - A hermitian input should lead to a hermitian output.

2-4 Lecture 2: February 9, 2018

1. Trace Preserving - Just as unitaries preserve length, our quantum operations should preserve trace.
2. Completely positive - Just like the nonnegativity condition on stochastic maps. Positive means that if
ρ is nonnegative, then N (ρ) is nonnegative. However, we need a stronger condition for it to be correct.
We need to stipulate that if we act on any part of ρ it should stay positive. That is if ρAR is positive
semidefinite, then (N ⊗ idR )(ρ) should also be positive semidefinite.
(As a comparison, transpose is positive but not completely positive, as a side note: the partial positive
transpose test (PPT) is one test of an entangled state: if PPT fails, it must be an entangled state).

These conditions give us the third form of quantum channels: a Trace Preserving, Completely Positive
(TPCP) map. Since this is derived from the axioms of what a quantum channel should have, it describes
the most general possible channel.
In fact, this turns out to be equivalent to the Kraus Form. We will only show that quantum channels in the
Kraus Form are TPCP channels, and the converse will be left for the problem set. By taking the daggerP of the
Kraus condition, it is clear that it is hermeticity preserving. Looking at the trace tr(N (ρ)) = tr( e Ve ρVe† )
and using the cyclic property of the trace and the Kraus Operator Condition, it can be seen that it is trace-
preserving. Finally, we show
P that it is completely positive. If ρ is positive semidefinite, it can be written as
W W † . Then (N ⊗ I)ρ = e (Ve ⊗ I)W W † (Ve† ⊗ I). This is a sum of positive semidefinite things, operators
times their adjoints, and thus itself is positive semidefinite.

2.3 Measurement as a Quantum Operation

Measurement can be thoughtPof as a quantum operation where the input is any quantum state and the
output is classical, N (ρ) = x px |xihx| (diagonal density matrix). The probabilities, px , should depend
on the state. Further, the px should be a linear function of ρ = tr(ρMx ). From this we can work out the
properties that Mx should obey:
P P
1. Normalization: tr( x Mx ρ) = 1 meaning x Mx = I.
2. Positive Semidefinite: tr(Mx ρ) > 0 and ρ is positive semidefinte.

These are similar to the Kraus Operator Conditions.

Just as we discussed the most general possible quantum operators, these are the most general measurements:
X
Mx = I
x

Mx  0
These conditions still leave room for noisy measurements, etc.
We can also talk about non-demolition measurements, which do not discard the quantum systems after
measurements. Consider the following quantum channel
X X Ve ρV †
Ņ(ρ) = Ve ρVe† ⊗ |ei he| = e
⊗ trVe ρVe† |ei he|
e e trVe ρVe†

This channel has the following interpretation: with probability pe := trVe ρVe† = trVe† Ve ρ =: trMe ρ the state
V ρV †
of the system after the application of this channel is ρe = trVe ρVe † . This is similar to having a measurement
e e
that outputs the post measured state ρe with probability pe .