Quantum Optics and Qubit Processors

Lecture Notes
Sophia Economou and Edwin Barnes

Contents
1 Qubits and quantum dynamics 4
1.1 Definition of a qubit . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
1.2 Time-dependent Schrödinger equation . . . . . . . . . . . . . . . . . . 6
1.3 Evolution operator . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
1.4 Pauli matrices . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
1.5 Examples of calculating 2 × 2 evolution operators . . . . . . . . . . . 15
1.5.1 Spin in constant B-field in xz plane . . . . . . . . . . . . . . . 15
1.5.2 Spin in precessing magnetic field . . . . . . . . . . . . . . . . 17
1.5.3 Rabi problem . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
1.5.4 Other analytically solvable problems . . . . . . . . . . . . . . 29
1.6 Pictures and frames . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
1.6.1 Rotating frame (or drive frame) . . . . . . . . . . . . . . . . . 30
1.6.2 Interaction picture . . . . . . . . . . . . . . . . . . . . . . . . 30
1.6.3 General frame transformation . . . . . . . . . . . . . . . . . . 31
1.7 Revisiting the case of a spin in a constant magnetic field . . . . . . . 32
1.8 Spin in precessing magnetic field revisited: . . . . . . . . . . . . . . . 37
1.9 3×3 rotation matrices from 2×2 unitaries: . . . . . . . . . . . . . . . 39
1.10 Rabi problem revisited . . . . . . . . . . . . . . . . . . . . . . . . . . 41
1.11 Adiabatic approximation . . . . . . . . . . . . . . . . . . . . . . . . . 42
1.12 Two-level system evolution operator in adiabatic limit . . . . . . . . . 46

1
 1.13 Berry’s phase . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49

2 The density operator and open quantum systems 52

2.1 Projectors and outer product . . . . . . . . . . . . . . . . . . . . . . 52
2.2 The density operator . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
2.3 Properties of the density operator . . . . . . . . . . . . . . . . . . . . 54
2.4 Bloch vector description of mixed states . . . . . . . . . . . . . . . . 55
2.5 Time evolution of the density operator . . . . . . . . . . . . . . . . . 57
2.6 Rabi problem with dissipation . . . . . . . . . . . . . . . . . . . . . . 57
2.7 Lindblad operators . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61
2.8 Bloch equations, dissipation and dephasing . . . . . . . . . . . . . . . 63

3 Decoherence and dynamical decoupling 66

3.1 Ramsey interferometry: How to measure T2 . . . . . . . . . . . . . . 66
3.2 A simple model of dephasing due to quasistatic noise . . . . . . . . . 68
3.3 Coherence function . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71
3.4 Spin echo and dynamical decoupling . . . . . . . . . . . . . . . . . . 73
3.5 Time-dependent noise . . . . . . . . . . . . . . . . . . . . . . . . . . 76
3.6 Filter functions and multi-pulse dynamical decoupling . . . . . . . . . 80

4 Entanglement 84
4.1 Separable versus entangled two-qubit states . . . . . . . . . . . . . . 84
4.2 Reduced density matrix and von Neumann entropy . . . . . . . . . . 87
4.2.1 Partial trace . . . . . . . . . . . . . . . . . . . . . . . . . . . . 88
4.2.2 Von Neumann entropy . . . . . . . . . . . . . . . . . . . . . . 89
4.2.3 Schmidt decomposition . . . . . . . . . . . . . . . . . . . . . . 90
4.2.4 Purification . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90
4.3 Decoherence as entanglement with the environment . . . . . . . . . . 91
4.3.1 Examples of common operator sum representations for qubits 92

5 Quantum gates 94
5.1 Two-qubit gates . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 94
5.2 Quantum circuits . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 98

2
 5.3 Quantifying entanglement . . . . . . . . . . . . . . . . . . . . . . . . 101
5.3.1 Cartan/KAK decomposition of two-qubit gates . . . . . . . . 101
5.3.2 Concurrence . . . . . . . . . . . . . . . . . . . . . . . . . . . . 104
5.3.3 Makhlin invariants . . . . . . . . . . . . . . . . . . . . . . . . 105

6 Qubit interacting with quantized field 108

6.1 Field quantization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 108
6.2 Jaynes-Cummings model . . . . . . . . . . . . . . . . . . . . . . . . . 115
6.2.1 Free evolution in Jaynes-Cummings model . . . . . . . . . . . 118
6.2.2 Energy spectrum of the Jaynes-Cummings model . . . . . . . 120
6.2.3 Open cavity and lossy qubit . . . . . . . . . . . . . . . . . . . 122
6.3 Physics and uses of the JC Hamiltonian . . . . . . . . . . . . . . . . 123
6.3.1 Photon Blockade . . . . . . . . . . . . . . . . . . . . . . . . . 123
6.3.2 Dispersive readout . . . . . . . . . . . . . . . . . . . . . . . . 123
6.3.3 Cavity-mediated two-qubit interaction . . . . . . . . . . . . . 124
6.3.4 Purcell effect . . . . . . . . . . . . . . . . . . . . . . . . . . . 125
6.4 Spontaneous emission in free space . . . . . . . . . . . . . . . . . . . 126

7 Quantum computing platforms 129

7.1 Semiconductor spin qubits . . . . . . . . . . . . . . . . . . . . . . . . 129
7.1.1 Brief note on semiconductors . . . . . . . . . . . . . . . . . . 129
7.1.2 Quantum dots . . . . . . . . . . . . . . . . . . . . . . . . . . . 130
7.1.3 Donors in silicon . . . . . . . . . . . . . . . . . . . . . . . . . 139
7.2 Superconducting qubits . . . . . . . . . . . . . . . . . . . . . . . . . . 140
7.2.1 LC circuits . . . . . . . . . . . . . . . . . . . . . . . . . . . . 140
7.2.2 A brief note on superconductivity . . . . . . . . . . . . . . . . 142
7.2.3 The Josephson junction . . . . . . . . . . . . . . . . . . . . . 143
7.2.4 Cooper pair box . . . . . . . . . . . . . . . . . . . . . . . . . . 146
7.2.5 Superconducting transmon qubits . . . . . . . . . . . . . . . . 148
7.2.6 Superconducting transmission line resonators . . . . . . . . . . 149
7.2.7 Superconducting qubit coupled to resonator . . . . . . . . . . 151
7.2.8 Driving a qubit in a cavity . . . . . . . . . . . . . . . . . . . . 153

3
 7.2.9 Two-qubit coupling and gates . . . . . . . . . . . . . . . . . . 154
7.3 Trapped ions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 157
7.3.1 Trapping and cooling . . . . . . . . . . . . . . . . . . . . . . . 158
7.3.2 Qubit: definition, readout, single-qubit gates . . . . . . . . . . 161
7.3.3 Normal modes of multiple ions in a trap . . . . . . . . . . . . 163
7.3.4 Coupling trapped ions and generating entangling gates . . . . 167

8 Quantum networks 172

8.1 Quantum communications . . . . . . . . . . . . . . . . . . . . . . . . 172
8.1.1 Quantum key distribution . . . . . . . . . . . . . . . . . . . . 172
8.2 Quantum repeaters . . . . . . . . . . . . . . . . . . . . . . . . . . . . 173
8.2.1 Entanglement swapping . . . . . . . . . . . . . . . . . . . . . 174
8.3 Quantum networks based on defect qubits . . . . . . . . . . . . . . . 175
8.3.1 The NV center in diamond . . . . . . . . . . . . . . . . . . . . 175
8.3.2 Nuclear spin qubits as a memory register . . . . . . . . . . . . 176
8.3.3 Quantum repeaters based on NV centers . . . . . . . . . . . . 177

1 Qubits and quantum dynamics

Present day information technologies such as computers, networks, and communica-
tion systems, are built on the principle that all information can be cast in digital
form, i.e. in terms of strings of 0’s and 1’s. These technologies are thus built up
from small physical systems, such as capacitors or transistors, that have two possible
states (i.e., a capacitor can be charged or uncharged) that can be used to represent
0 and 1. Quantum information science is about developing new information tech-
nologies in which these classical bits of information are replaced by quantum bits
(qubits). A qubit is a quantum two-level system, which unlike classical systems, can
be in superposition states. This fact makes quantum information systems funda-
mentally different from their classical counterparts and allows for novel capabilities.
This course is about understanding the physics of qubits. In particular, what phys-
ical properties must a qubit have for it to be technologically useful, what types of

4
systems can realize these properties, and how can qubit states be manipulated physi-
cally in order to implement quantum information processing? In this first section, we
describe the basic physical criteria that any qubit must satisfy. We also review time-
dependent quantum mechanics in the context of two-level systems since this will be
needed to understand qubit manipulations in most quantum information contexts.

1.1 Definition of a qubit

At the most basic level, a qubit is a quantum two-level system. However, useful
qubits require much more than this. In the late 20th century, David DiVincenzo laid
out a specific set of criteria that a quantum system must satisfy in order for it to
constitute a useful qubit:

• Scalability

• The ability to initialize its state

• Long decoherence times

• The ability to control its state, both alone and jointly with other qubits (uni-
versal set of quantum gates)

• A qubit-specific measurement capability

The first criterion, scalability, means that we need to be able to reliably make a lot of
high-quality qubits (typically thousands or millions depending on the application).
The remaining four criteria define what we mean by “high-quality”. We need to be
able to initialize qubits to some specific state (e.g., one of its two logical states |0i or
|1i) so that they are then ready for whatever computation we want to perform. Good
qubits should have long decoherence times. The decoherence time is a measure of how
long a qubit retains the information stored on it. Unfortunately qubits inevitably
interact with their environment, and through these uncontrolled interactions, the
information can leak out and become lost. The longer the decoherence time, the
better. Of course, simply being able to store a quantum state for a long time is not
that exciting. We also need to be able to change the state to perform information

5
processing. Thus, a good qubit should be controllable in the sense that we can rotate
its state at will. Finally, we need to be able to extract information from the qubits
at the end of a computation or other information task; this means we need to be
able to measure the states of individual qubits with high accuracy. Implicit in these
criteria is a sixth one, which is that we need to be able to couple qubits together and
create quantum entanglement, which is the engine behind many of the most useful
and interesting quantum information applications.
It’s important to note that isolated quantum two-level systems are not readily
available in nature. Quantum systems almost always contain many more levels, and
so in order to turn such systems into qubits, we need to somehow select two levels
and isolate them from the rest. For example, we can imagine a spin 1/2 particle in
a confining potential. If we can somehow fix its spatial state to the ground state of
the potential, then we could try and use the two spin states as a qubit. As another
example, we could imagine trying to use the two lowest levels of an atom as qubit
states. If we can control these states without involving the higher levels, then this
approach could work. Both these approaches are being heavily investigated, along
with several more, less obvious approaches, as we’ll see later.

1.2 Time-dependent Schrödinger equation

Quantum information processing is largely about the dynamics of qubits. This dy-
namics is of course governed by the time-dependent Schrödinger equation. Let’s take
some time to review this equation and solve it in some simple but important cases.
We’ll set ~ = 1 throughout this part. In terms of abstract kets in Hilbert space, the
equation is
∂
i |ψ(t)i = H |ψ(t)i . (1.1)
∂t
Recall that we can choose any representation we want and recast the equation in
various forms. For example, in position space, it becomes
 
∂ψ 1 2
i = Hψ = − ∇ + V (~r) ψ(~r, t). (1.2)
∂t 2m

6
Most introductory courses on quantum mechanics focus on the special case where the
Hamiltonian is time-independent, ∂H∂t
= 0. In this case, we can find solutions in which
the spatial dependence separates from the time-dependence: ψ(~r, t) = e−iEt ϕE (~r).
Plugging this ansatz into Eq. (1.2) reveals that the problem reduces to an eigenvalue
problem, HϕE = EϕE , where we need to solve for the energy eigenstates ϕE and
their corresponding eigenvalues E. This is an example of Noether’s Theorem:

Symmetry ⇒ Conserved quantity

In this case, because the Hamiltonian is time-independent, we have time-translation

symmetry, and so energy is conserved:

∂H
= 0 ⇒ time-translation symmetry
∂t
⇒ energy is conserved

In your first proper course on quantum mechanics, you probably spent most of your
time solving this eigenvalue problem, HϕE = EϕE . This is commonly done for,
e.g., the simple harmonic oscillator and the hydrogen atom in such courses. In this
course, we will often assume we already know all the eigenvalues E and eigenfunc-
tions ϕE (~r) of the system, and we focus on the quantum dynamics that takes place
when the system is not in an energy eigenstate. Note that when the Hamiltonian
is time-dependent (as will often be the case in this course), we cannot even define
energy since it isn’t conserved.
Let’s think a little about what DiVincenzo’s initialization criterion means physically.
Usually, we choose logical qubit states to be energy eigenstates like |ϕE i. For a
concrete example, let’s consider a problem that we’re all familiar with: the infinite
square well. Recall the energy eigenstates and eigenvalues:

7
Example: Infinite square well

r
2  nπx 
ϕn (x) = sin
a a
n2 π 2
En =
2ma2

0 a

Consider the following questions:

If the initial state of the system is ϕ1 (x), what is the state at later times?
What if the initial state is √12 [ϕ1 + ϕ2 ]?
If we prepare the system in an energy eigenstate like ϕ1 (x), then it will remain in
this state indefinitely (unless the particle interacts with something else other than
the square well). On the other hand, if the initial state is a superposition of two
energy eigenstates, then the wave function will evolve and move around inside the
well continuously. The question then becomes: if we want to make a qubit out of the
lowest two energy eigenstates, how can we prepare the system in a specific state like
ϕ1 (x) if it starts off in some random superposition of ϕ1 (x) and ϕ2 (x)? The answer
is that there are two possible ways to perform initialization in this system: We can
either (weakly) couple the system to a reservoir and wait for it to relax to the ground
state, or we can measure the particle’s energy so that we collapse the wavefunction
into the desired state. Note that in the first scenario, we have to bring in a reservoir
because the initial superposition state has more energy than the ground state, and
so the excess energy needs some place to go in order for the particle to relax to the
ground state.

Now let’s think for a moment about another DiVincenzo criterion: We need to be
able to rotate from one qubit state to another. This time, let’s suppose we want to
use two different energy eigenstates of a hydrogen atom as our qubit states.

Recall that the hydrogen atom energy levels are written as |nlmi, where the quantum

8
numbers are

n = 1, 2, 3, . . . (principal)
quantum numbers: l = 0, 1, 2, . . . , n − 1 (orbital)
m = −l, −l + 1, . . . , l − 1, l (orbital z)
We could consider, e.g., |100i and |210i as qubit states. We could also try |100i
| {z } | {z }
1S orbital 2P orbital
and |211i. However, rotating from |100i to |211i requires energy (E100 6= E211 ). How
can we supply this energy? Is there a physical mechanism to go from |100i to |210i
or from |100i to |211i in a controlled way? These operations require driving the
system with external fields. We’ll see how to do this in detail later in this section.

Before we go on to talk about quantum dynamics more systematically, we pause

for a moment and consider the following question: How do you make and operate
a spin qubit? Above, we mentioned that if one can trap a spin 1/2 particle (like in
the infinite square well), then it may be possible to use the spin as a qubit. This
would be nice because there are exactly two spin states, and so it seems like we could
isolate these pretty well. How can we control the spin? The most straightforward
way is to use a time-dependent magnetic field, so that the total Hamiltonian is:

1 2 ~ · S.
~
H=− ∇ + V (~r) + gµB B(t) (1.3)
2m
Here, V (~r) is a confining potential (e.g., simple harmonic oscillator, infinite square
well, Coulomb), g is the Landé g-factor, which is ≈ 2 for electrons, and µB is the
e~
Bohr magneton: µB = 2m ec
. If the spatial part is an energy eigenfunction, then
we can write the total state in separable form: ψ(~r, t) = ϕE (~r) χ(t) , where all the
|{z}
spinor
time-dependence is carried by the spin. Plugging this ansatz into the Schrödinger

9
equation, we can obtain a Schrödinger equation that only involves the spin part:

~ · Sχ
Hψ = EϕE χ + gµB ϕE B ~ = iϕE χ̇
~ · S)χ(t)
⇒ (E + gµB B ~ = iχ̇
↑
Constant energy shift ⇒ ignore

Thus, we see that we can keep the spatial part in the ground state of the confining
potential, then we can just focus on the spin and drive qubit operations using time-
dependent magnetic fields. Trapping electrons in potentials is a popular approach to
realizing spin qubits, as we’ll discuss in depth later in the course.

1.3 Evolution operator

An important tool in describing quantum dynamics is the time evolution operator or
simply evolution operator. The evolution operator is especially important in quantum
information science because it is directly related to the concept of quantum logic gates
and circuits. Let’s recall its definition and properties.
Let’s again start from the time-dependent Schrödinger equation,

Hψ = iψ̇. (1.4)

Given an initial time t0 , the evolution operator is defined to be an operator u(t, t0 )

that transforms the initial state ψ(t0 ) to the state at time t:

ψ(t) = u(t, t0 )ψ(t0 ). (1.5)

Plugging this into Eq. (1.4) yields an equation for u(t, t0 ):

iψ̇(t) = Hψ(t) (1.6)

(ψ̇(t) = u̇ψ(t0 )) (1.7)
iu̇ψ(t0 ) = Huψ(t0 ) ⇒ iu̇ = Hu , u(t0 , t0 ) = 1. (1.8)

10
Thus, the evolution operator also satisfies the time-dependent Schrödinger equation.
However, it is an operator rather than a state, and it has the initial condition that
u(t0 , t0 ) must be the identity operator. This follows directly from Eq. (1.5) by setting
t = 0. Here are some general properties of the evolution operator:

• Unitarity: uu† = u† u = 1

• u(t1 , t2 ) = u† (t2 , t1 )

• u(tn , t1 ) = u(tn , tn−1 )u(tn−1 , tn−2 )...u(t2 , t1 )

It is easy to find an explicit expression for the evolution operator in the special
case where H is independent of time. In this case, for any initial state, we can obtain
the state at later times by exponentiating the Hamiltonian:

ψ(t) = e−iHt ψ(0). (1.9)

Therefore, u(t, 0) = e−iHt if ∂H

∂t
= 0. Also note that if ψ(0) is an energy eigenstate,

Hψ(0) = Hψn = En ψn , (1.10)

then we find ψn (t) = e−iEn t ψn .

However, in the more general case where the Hamiltonian is time-dependent, we

cannot write down a general formula for u(t, t0 ). We must instead try to solve the
operator differential equation directly:

H(t)u(t, t0 ) = iu̇(t, t0 ), u(t0 , t0 ) = 1.

In general, this cannot be done analytically even for a two-level system.

Note that we cannot obtain u by exponentiating H unless H commutes with itself
at different times.

Proof:

11
 Z t
iu̇ = Hu ⇒ u(t) = 1 − i dt1 H(t1 )u(t1 )
0
Z t Z t Z t1
=1−i dt1 H(t1 ) − dt1 dt2 H(t1 )H(t2 )u(t2 )
0 0 0
Z t Z t Z t1
=1−i dt1 H(t1 ) − dt1 dt2 H(t1 )H(t2 )
0 0 0
Z t Z t1 Z t2
+i dt1 dt2 dt3 H(t1 )H(t2 )H(t3 ) + . . .
0 0 0

Define T to be the time-ordering operator and consider the 2nd order term:
Z t Z t1 Z t Z t1
dt1 dt2 H(t1 )H(t2 ) = dt1 dt2 T [H(t1 )H(t2 )]
0 0 0 0

By thinking carefully about the integration region, we can rewrite this term.
t2

Z t Z t
t
= dt2 dt1 T [H(t1 )H(t2 )]
0 t
Z t Z 2t
= dt1 dt2 T [H(t1 )H(t2 )]
0 t1
integration
region

t t1
Z t Z t1 Z t Z t
1
⇒ dt1 dt2 T [H(t1 )H(t2 )] = dt1 dt2 T [H(t1 )H(t2 )]
0 0 2 0 0

12
This then allows us to write u(t) as
Z t Z t Z t
1
u(t) = 1 − i dt1 H(t1 ) − dt1 dt2 T [H(t1 )H(t2 )] + . . .
0 2 0 0

Comparing this with what we get from expanding the exponentiated, integrated
Hamiltonian,
 Z t  Z t Z Z t
1 t
exp −i dt1 H(t1 ) = 1 − i dt1 H(t1 ) − dt1 dt2 H(t1 )H(t2 ) + . . .
0 0 2 0 0

we see that the two expressions are equal only if the Hamiltonian commutes with
itself at different times: [H(t1 ), H(t2 )] = 0, so that T [H(t1 )H(t2 )] = H(t1 )H(t2 )
regardless of whether t1 is larger or smaller than t2 . This completes the proof.

1.4 Pauli matrices

Before we proceed to compute evolution operators for different examples of two-level
Hamiltonians, we first review some basic properties of the Pauli matrices. These
matrices form a basis of 2×2 Hermitian operators and, as such, play a very important
role throughout quantum information science. They will also help facilitate our
subsequent discussion of two-level dynamics.
Here are the Pauli matrices and some of their properties:

# "
0 1
σx = X = σ0 = 1
1 0
" #
0 −i
σy = Y =
i 0
" #
1 0
σz = Z =
0 −1

13
 Properties
[σx , σy ] = 2iσz (& c.p.)
σj2 = 1 {σi , σj } = 21δij
Tr(σj ) = 0 j 6= 0

Practice problem: What are the eigenvalues of Y?

Practice problem: What are the eigenvalues of cos(φ)X + sin(φ)Y ?

 
0 a + ib
The most general form of a Hermitian 2×2 is H = .
a − ib 1
Question: How do we show we can express H in terms of 1, σx , σy , σz ?

H = C0 1 + Cx σx + Cy σy + Cz σz

We can answer this using the traceless property of X, Y, Z:

• Take trace: Tr(H) = 2C0 = 0 + 1

• Multiply by σz & trace: Tr(σz H) = 2Cz = 0 − 1

• Multiply by σx & trace: Tr(σx H) = 2Cx = 2a

• Multiply by σy & trace: Tr(σy H) = 2Cy = −2b

Because any two-level Hamiltonian can be expanded as a linear combination of Pauli

matrices with real coefficients, we see that any quantum mechanical 2-level system
can be thought of as a spin in a B-field. This motivates taking a closer look at the
latter. We’ll do this next.

14
1.5 Examples of calculating 2 × 2 evolution operators
1.5.1 Spin in constant B-field in xz plane
~ = (Sx , Sy , Sz ).
Consider a spin 1/2 in a B-field, and define the spin operators as S
The Hamiltonian is
~
H = −~µ · B
e ~ ~
=g S·B
X2m
= ωj Sj .
j

Here we have again set ~ = 1, so that frequency and energy have the same units.
Sj = ~2 σj , we can also write the Hamiltonian in terms of the Pauli matrices:
SinceP
H = j 12 ωj σj .

Suppose the B-field has components in the z and x directions. The Hamiltonian
is then H = 12 ωz σz + 12 ωx σx . What is u in this case? We know from the discussion
from above that if the B-field is time-independent, then H is also, and so we can just
1
exponentiate H: u = e−iHt = e−i 2 (ωz σz +ωx σx )t .

Question: What does it mean to have an exponential of an operator (matrix)?

Answer: The exponential of an operator or matrix is defined in terms of the Taylor
expansion of the exponential function: ex = 1 + x + 21 x2 + 3!1 x3 + ... More generally,
a function of an operator is defined in terms of the Taylor expansion of the function.
? 1 1
Question: Can we factorize the exponential, i.e., e−iHt = e−i 2 ωz tσz e−i 2 ωx tσx
Answer: No, because σx and σz don’t commute.

How can we obtain an explicit 2×2 matrix expression for the evolution operator
from u = e−i(ωz σz +ωx σx )t/2 ? One way is to rewrite the matrix in the exponent in a
nice way so that if we perform a Taylor series expansion of the exponential function,
we can figure out how to resum the terms and get something in closed form. We can
do this by first rewriting the Hamiltonian.p Start by multiplying and dividing H by
λ ωz ωx
λ: H = 2 ( λ σz + λ σx ). Choosing λ = ωz2 + ωx2 , we have
p !
ωz2 + ωx2 ω ωx
H= p z σz + p σx . (1.11)
2 2
ωz + ωx2 ωz2 + ωx2

15
If we now define cos(θ) = √ ω2z ⇒ sin(θ) = √ ω2x , the Hamiltonian becomes
ωz +ωx2 ωz +ωx2

λ λ
H= (cos(θ)σz + sin(θ)σx ) = σθ . (1.12)
2 2
From the basic properties of the Pauli matrices listed above, we can easily check
that σθ2 = 1. This observation makes it easy to resum the Taylor expansion of the
exponential:
X∞  k X 1  λ k X 1  λ k
1 λ
u = e −i λ
2
tσθ
= k
−i t σθ = −i t 1 + −i t σθ
k=0
k! 2 k even
k! 2 k odd
k! 2
= cos(λt/2)1 − i sin(λt/2)σθ . (1.13)

We thus obtain an explicit 2 × 2 matrix expression for the evolution operator u(t).

There is also an alternative way to obtain the same result that does not require
resumming a Taylor series. In particular, we can first diagonalize H, perform the
matrix exponentiation (which is now just an ordinary exponential of each diago-
nal entry of H, and then transform back to the original basis in which H is not
diagonal. We can perform the diagonalization either starting from the expression
u = e−i(ωz σz +ωx σx )t/2 , or we can start from the time-dependent Schrödinger equation
for u(t) and perform the diagonalization directly in this equation. Let’s see how the
latter approach works first.
Define T to be a time-independent unitary transformation. Starting from Hu =
iu̇, we can multiply both sides on the left by T :
d
T † T u = iT u̇ ⇒ (T HT † )(T u) = i
T Hu = iT u̇ ⇒ T H |{z} (T u). (1.14)
dt
1

Note that we used the fact that T is unitary to insert a T † T , which is just the
identity operator. We also used that T is time-independent to bring it inside the
time derivative. Now choose T such that T HT † is diagonal, and call this diagonal
matrix Hd = T HT † . Let’s also define ũ = T u. Then we have Hd ũ = i dt d
ũ ⇒ ũ(t) =
e−iHd t
ũ(0). The initial condition for ũ follows from that of u. Recall that u(0) = 1,
implying that ũ(0) = T . Therefore, we find

u = T † ũ = T † e−iHd t T. (1.15)

16
 The next step is to find an explicit expression for T . We can start from the
following general parameterization:



cos φ2
sin φ2

T = . (1.16)
− sin φ2 cos φ2

Applying this expression to both sides of H, we find

   
1 ωz ωx † 1 ωz cos(φ) + ωx sin(φ) ωx cos(φ) − ωz sin(φ)
Hd = T T = .
2 ωx −ωz 2 ωx cos(φ) − ωz sin(φ) −ωz cos(φ) − ωx sin(φ)
(1.17)
We need this to be diagonal, which can be accomplished p by choosing cos(φ) =
ωz ωx 2 2
A
, sin(φ) = A
. Since sin φ + cos φ = 1, we must have A = ωz2 + ωx2 :
2 2
! p 
1 ωz +ω x
0 1 ωz2 + ωx2 0
Hd = A
2 2 = p . (1.18)
2 0 − ωz +ω
A
x 2 0 − ωz2 + ωx2

Given Hd and T , we can then find an explicit expression for u using Eq. (1.15) and
the fact that √ !
−i ωz2 +ωx2 t/2
e 0
√ 2 2
e−iHd t = . (1.19)
0 ei ωz +ωx t/2
You should check that this gives the same result as Eq. (1.13). We will take a closer
look at the physics of this solution later on in Sec. 1.7.

1.5.2 Spin in precessing magnetic field

Next, let’s consider the more interesting case of a spin in a precessing magnetic field
given by
~ = B0 cos(ωt)x̂ + B0 sin(ωt)ŷ.
B (1.20)
In this case, the Hamiltonian from the Zeeman interaction is now time-dependent:

~ =g e ~ ~ geB0
H = −~µ · B S·B = (cos(ωt)σx + sin(ωt)σy ) (1.21)
2m 2m
| {z }
b
   
0 b cos(ωt) − ib sin(ωt) 0 be−iωt
= = . (1.22)
b cos(ωt) + ib sin(ωt) 0 beiωt 0

17
Because of this time-dependence, we now have to be very careful about exponenti-
ating the Hamiltonian to obtain the evolution operator. As we proved above, this
only works if the Hamiltonian commutes with itself at different times.
Question: Does the Hamiltonian in Eq. (1.21) commute with itself at different
times?
Answer: No! Therefore, we cannot exponentiate the Hamiltonian to obtain u:
−i 0t H(t0 )dt0
R
u 6= e .

Question: Can we still obtain an expression for u analytically? Answer: In this

case, yes!

It just so happens that for this particular Hamiltonian, we can find a unitary
transformation that brings the Hamiltonian into a time-independent form. This
then allows us to find the evolution operator in the new frame by exponentiating the
Hamiltonian in that frame. After we do so, we can then undo the transformation to
obtain an explicit expression for u in the original frame. Let’s see how this works in
detail.
Starting from the time-dependent Schrödinger equation, Hu = iu̇, apply a time-
dependent transformation R to both sides:

RHR† Ru = iRu̇, (1.23)

where R is unitary (R† = R−1 ). Note that we cannot follow the same steps as in
Eq. (1.14) because unlike T in that equation, here R is time-dependent. We cannot
for instance bring R inside the time derivative. Therefore, defining ũ = Ru, we
instead obtain the following expression for the time derivative:
˙
u̇ = Ṙ† ũ + R† ũ. (1.24)

Using this expression in Eq. (1.23), we find

˙ = iRṘ† ũ + iũ˙
RHR† ũ = iR(Ṙ† ũ + R† ũ) (1.25)
⇒ (RHR† − iRṘ† ) ũ = iũ˙ ⇒ H̃ ũ = iũ˙ . (1.26)
| {z }
≡H̃

Thus, we obtain what looks like a regular time-dependent Schrödinger equation, but
where the Hamiltonian is now given by H̃. This is called the effective Hamiltonian

18
in this new frame defined by R.
We just saw how to perform a general, time-dependent unitary transformation
to the Schrödinger equation, but how does this help us with our precessing B-field
problem? To solve that problem, we need to find a transformation R such that H̃ is
time-independent. This can be accomplished by making the following choice:
 iωt/2   −iωt/2 
e 0 † e 0
R= , R = . (1.27)
0 e−iωt/2 0 eiωt/2

Applying this to the original lab frame Hamiltonian H, we find

 
† 0 b
RHR = , (1.28)
b 0

while the other term in H̃ evaluates to

   
† iω −1 0 −ω/2 0
− iRṘ = −i = . (1.29)
2 0 1 0 ω/2

Combining these two results, we thus obtain the following result for the effective
Hamiltonian:  
−ω/2 b
H̃ = , (1.30)
b ω/2
which is indeed time-independent. Not only that, this is precisely the same problem
we solved in the previous subsection when we considered a static B-field in the
xz plane! Therefore, in the rotating frame, we can use either of the two approaches
described in that subsection to obtain the evolution operator from the exponentiated
Hamiltonian. We won’t bother to repeat those steps here, and we’ll just keep it
abstract: Rt 0 0
ũ(t) = e−i 0 H̃(t )dt ũ(0). (1.31)
Using that the initial condition is ũ(0) = R(0)u(0) = 1, and transforming back to
the lab frame, we finally obtain
Rt
H̃(t0 )dt0
u(t) = R† (t)e−i 0 . (1.32)

We will study the physics of this solution later on in Sec. 1.8.

19
1.5.3 Rabi problem
As a third example, we consider the famous Rabi problem. This problem is a classic
example of quantum two-level dynamics, and as such plays a very important role in
quantum information science and beyond. We will solve this problem in two different
ways: (i) using time-dependent perturbation theory and (ii) solving it exactly. We
will begin by reviewing time-dependent perturbation theory and then show how it
can be applied to solve the Rabi problem approximately at the limit of weak driving.
We will then show how to obtain the solution for the evolution operator exactly.

Time-dependent perturbation theory

Suppose we have a Hamiltonian that is the sum of two terms,

H(t) = H0 + H1 (t) (1.33)

where H0 is a static Hamiltonian whose eigenvalue problem

H0 |ni = En |ni , (1.34)

has already been solved, and where all the time-dependence is carried by H1 (t),
which is in some sense small compared to H0 . The question we want to answer is:
Assuming H1 (t) is weak, what is the probability that there will be a transition from
the initial state |ii to a given state |ki?.
At this point, you might be asking yourself, what do we mean by “weak”. This
point is a little subtle since H0 and H1 are operators and not simple scalars. Some-
times, “weak” is taken to mean that |Ef − Ei |  hf |H1 (t)|ii, however, this is a
somewhat problematic definition for reasons we won’t get into here. A more robust
definition of “weak” is to require that the probability of a transition from the initial
to the final state should be small: Pi→f  1. This seems natural, because if this
probability is large, then it really doesn’t make sense to call H1 (t) weak, as it is
clearly having a dramatic effect on the dynamics of the system.
Assuming that H1 (t) satisfies this definition of weak (which we can only really
confirm after we compute the transition probability and check whether it is indeed
small), then the task is to perform a perturbative expansion of the Schrödinger
equation in powers of H1 (t). Since H1 (t) is an operator, it’s a little awkward to
talk about an expansion in this quantity; to avoid confusion, it helps to pull out a

20
dimensionless parameter λ from H1 :

H1 (t) = λH10 (t), (1.35)

and to perform a perturbative expansion in λ assuming λ  1. Even if the matrix

elements of H10 are comparable to those of H0 , this expansion will still make sense so
long as λ is small enough.
Now consider the Schrödinger equation with initial state |ψ(0)i = |ii:
∂
i ∂t |ψ(t)i = [H0 + λH10 (t)] |ψ(t)i . (1.36)

We can expand |ψ(t)i in terms of unperturbed energy eigenstates (recall that we

assume we already know what these are):
X
|ψ(t)i = cn (t)e−iEn t |ni . (1.37)
n

Combining (1.36) & (1.37), we have

X X
i (ċn (t) − iEn cn (t))e−iEn t |ni = (H0 + λH10 (t))cn (t)e−iEn t |ni
n n
X
= (En + λH10 (t))cn (t)e−iEn t |ni
n

We can drop the En terms from both sides since these are the same, obtaining
X X
i ċn (t)e−iEn t |ni = λcn (t)e−iEn t H10 (t) |ni . (1.38)
n n

Next, project both sides onto the state |ki that we’re interested in:
X
iċk (t)e−iEk t = λ cn (t)e−iEn t hk| H10 (t) |ni . (1.39)
n

We then arrive at the following equation for the coefficients ck (t):

X
ċk (t) = −iλ cn (t)e−i(En −Ek )t hk| H10 (t) |ni . (1.40)
n

So far this is an exact system of coupled equations; no approximations have been

21
made yet. To proceed, we expand ck in a power series in λ:
(0) (1) (2) (0)
ck = ck + λck + λ2 ck + ..., iċk = 0. (1.41)

so that we can match terms at each order in λ. At first order, we obtain

(1)
X
−i(En −Ek )t
iċk = c(0)
n e hk| H10 (t) |ni (First order solution). (1.42)
n

(0)
Since the initial state is chosen to be |ii, we should have ci (t) = 1, while all other
coefficients are initially zero. The probability to end up in state |ki at later times is
given by
2
(1)
Pi→k = λck . (1.43)
(1)
The coefficient ck can be obtained by integrating Eq. (1.42):
H1 (t)
Z t z }| {
(1) )t0
λck = −i dt0 e−i(Ei −Ek hk|λH10 (t0 )|ii , (1.44)
t0

and so the probability to transition from state |ii to state |ki is

Z t 2
0 −i(Ei −Ek )t0 0
Pi→k = dt e hk|H1 (t )|ii . (1.45)
t0

This expression is perturbative, so it can only be valid if Pi→k  1.

Perturbation theory applied to the Rabi problem

Now we will apply the above perturbation theory formalism to the Rabi problem.
Consider a weak, oscillating electric field with a “square” pulse profile. That is, the
field suddenly switches on at time t0 = 0, and it suddenly switches off at time τ :
~ = êE0 (t) cos(ωt),
E (1.46)

22
with (
E0 for 0 < t < τ
E0 (t) = . (1.47)
0 otherwise
This profile looks like this:
E0(t)
"Square Pulse"
Envelope

0 τ t
For now, we will use the expression from classical physics for the coupling between
the field and a 2-level system:

H1 (t) = −d~ · E(r

~ 0 , t), (1.48)

where d~ is the dipole (in quantum mechanics, this is an operator) and E(r
~ 0 , t) is
the field at r0 , the position of the dipole. We’ll come back later and derive this
Hamiltonian rigorously. We also define −d~ · êE0 = W . This interaction can cause
transitions between the two states:
|1>

|0>

The amplitude we found perturbatively above now becomes

Z τ Z
(1) −i(Ei −Ek )t 1 τ

λci→k = dt e hk|W |ii cos(ωt) = dt e−i(Ei −Ek )t eiωt + e−iωt hk|W |ii .
0 2 0
(1.49)
We also define Ek − Ei = ω0 (transition frequency) (for simplicity we take Ek > Ei ).

23
Now we can integrate:
Z τ Z τ 
(1) hk|W |ii i(ω0 +ω)t i(ω0 −ω)t
λci→k = dt e + dt e
2 0 0
 
hk|W |ii ei(ω0 +ω)τ − 1 ei(ω0 −ω)τ − 1
= + . (1.50)
2 i(ω0 + ω) i(ω0 − ω)

Notice the quantities that appear in the denominators are ω0 −ω & ω0 +ω. Typically,
the radiation is near-resonant, meaning that |ω0 − ω|  ω. Then the first term is
much smaller than the second, and we can drop it:

(1) ei(ω0 −ω)τ /2 sin[(ω0 − ω)τ /2]

λci→k ≈ hk|W |ii . (1.51)
ω0 − ω
And the probability is

| hk|W |ii|2 sin2 [(ω0 − ω)τ /2]

Pi→k = |ci→k |2 ≈ . (1.52)
(ω0 − ω)2

We also define the concept of detuning, which is the difference between the pulse and
transition frequencies:
∆ = ω − ω0 . (1.53)
In terms of the detuning, we can express the transition probability as

| hk|W |ii|2 sin2 (∆τ /2)

Pi→k = . (1.54)
∆2
Let’s note some important properties of this result:
• As a function of frequency, it has a maximum at ∆ = 0, since this is a sinc
function.
• At zero detuning, the maximum probability is | hk|W |ii|2 τ 2 /4 6= 1. Does this
make sense?

Rabi problem - exact solution for the square pulse

24
 Now let’s resolve the same problem, but this time without using perturbation
theory. We again consider a two-level system with levels |0i, |1i and energies E0 , E1
respectively. The transition energy is then E1 − E0 :
|1>

E1-E0

|0>

 
E0 0
The free (unperturbed) Hamiltonian is H0 = . We consider again a square
0 E1
pulse, so that the Hamiltonian is H(t) = H0 + H1 (t), where

H1 (t) = Ω(t) cos(ωt)(|0i h1| + |1i h0|), (1.55)

and (
Ω0 for t ∈ (0, τ )
Ω(t) = . (1.56)
0 otherwise
Therefore, we can write
 
E0 Ω0 cos(ωt)
H= , for t ∈ (0, τ ). (1.57)
Ω0 cos(ωt) E1

We can redefine the zero of energy to obtain a more symmetric form of the Hamil-
tonian:  E0 −E1   
2
Ω 0 cos(ωt) −ω 0 /2 Ω0 cos(ωt)
H= E1 −E0 = . (1.58)
Ω0 cos(ωt) 2
Ω0 cos(ωt) ω0 /2
Now consider a general state in the two-level Hilbert space:
 
a0 (t)
|ψ(t)i = a0 (t) |0i + a1 (t) |1i = (1.59)
a1 (t)

25
Plugging this into the Schrödinger equation, we obtain
   
a0 (t) ȧ0 (t)
H =i , (1.60)
a1 (t) ȧ1 (t)

from which we find that the coefficients obey the following pair of differential equa-
tions:
ω0
iȧ0 = − a0 + Ω0 cos(ωt)a1 , (1.61)
2
ω0
iȧ1 = a1 + Ω0 cos(ωt)a0 . (1.62)
2
Define b0 (t), b1 (t) through the relations

a0 = eiωt/2 b0 , a1 = e−iωt/2 b1 . (1.63)

Then the first equation in (1.61) becomes

 ω  ω0
i i b0 + ḃ0 eiωt/2 = − b0 eiωt/2 + Ω0 cos(ωt)e−iωt/2 b1 , (1.64)
2 2
which reduces to
ω − ω0
iḃ0 = b0 + Ω0 cos(ωt)b1 e−iωt . (1.65)
2
Recognizing ω − ω0 = ∆ as the detuning, we can rewrite this as
∆ Ω0  iωt  ∆ Ω0  
iḃ0 = b0 + b1 e + e−iωt e−iωt = b0 + b1 1 + e−2iωt . (1.66)
2 2 2 2
Similar to what we saw in the time-dependent perturbation theory calculation, we
notice the “rapidly rotating” term e−2iωt , which we understand will average out
when we integrate. Dropping this term is called the “rotating wave approximation”
(RWA), which is valid in the near-resonant case. Following a similar procedure for
Eq. (1.62), we arrive at the following equations:

∆ Ω0
iḃ0 = b0 + b1 , (1.67)
2 2
−∆ Ω0
iḃ1 = b1 + b0 . (1.68)
2 2

26
We can think of these as coming from an effective Schrödinger equation of the form
   
b0 d b0
Hef f =i , (1.69)
b1 dt b1

with  
∆/2 Ω0 /2
Hef f = . (1.70)
Ω0 /2 −∆/2
What have we achieved? A tremendous simplification! A hard time-dependent
problem has just been reduced to a time-independent one! Notice that this problem
is only time-independent (and thus analytically solvable) for the square pulse case.

Question: What happens for Ω0 → Ω0 (t) of arbitrary shape? Why can we no

longer solve the problem in this case?

Now let’s ask the same question as in the case of perturbation theory: If we start
from state |0i at t = 0, what is the probability to be in state |1i at time t? After we
find the evolution operator u using the methods discussed previously, we can then
apply it to the initial state to obtain the state at later times:
   
1 u00 (t)
u(t) = . (1.71)
0 u10 (t)

The transition probability can then be obtained from the off-diagonal component of
u:
P0→1 (t) = |u10 (t)|2 . (1.72)
Carrying out the details of the calculation, we find that
q 
2 Ω20 2 2
P0→1 (t) = |a1 (t)| = 2 sin Ω0 + ∆2 t/2 . (1.73)
Ω0 + ∆2

This is the exact solution for the transition probability (within the RWA).

Let’s examine special cases:

• When the driving is resonant with the 2-level system, ∆ = 0, we see that

27
 P0→1 (t) has a maximum amplitude of 1, and it oscillates with frequency Ω0 .
• ∆p = Ω0 . In this√case, the amplitude is 1/2, while the frequency of oscillation
is Ω20 + ∆2 = 2Ω0 > Ω0 .
p
• ∆  Ω0 . This is the weak pulse limit, which implies Ω20 + ∆2 ≈ ∆. In this
Ω2

case, we obtain P ≈ ∆02 sin2 ∆t
2
, which recovers the perturbative result from
before.
Some remarks on the Rabi problem:
• What conditions have to be met to have a transition with 100% probability?

1. First of all, we need ∆ = 0, so that the amplitude of P0→1 (t) is 1. This

makes physical sense because in this case, the frequency of the driving
field exactly matches the transition energy, and so there is a conservation
of energy. Recall that the energy of a photon in the field is ω, so we
can think of this as a photon being absorbed by the 2-level system. This
point will become clearer later on when we quantize the electric field. For
this reason, sometimes the counter-rotating terms in RWA are also called
energy non-conserving terms.
p
2. We also need to set the time to t = π/ Ω20 + ∆2 in order to get a perfect
transition. This is sometimes called a π-pulse.

• Note that this solution applies to a wide range of physical systems and fre-
quency scales, including lasers driving atoms, B-fields driving electron spins,
B-fields driving neutron spins (MRI), etc.
Finally, we note that in order to obtain the exact result for u, we used the close
relation between the state and the evolution operator. It’s
 important
 to be able to
1
go between the two in general. Take for example ψ(0) = . Then we have
0
 
u11 (t) u12 (t)
u(t) = , (1.74)
u21 (t) u22 (t)

and so we have  
u11 (t)
ψ(t) = u(t)ψ(0) = . (1.75)
u21 (t)
Therefore, by solving for this initial condition, we can find u11 , u21 . Remember, how-
ever, that for a two-level system, u22 = u11 ∗ and u21 = −u12 ∗. This means that we

28
can fully obtain u(t) by solving for ψ(t) for one initial condition.

Question: Does this work for quantum systems with more than 2 levels? Why or
why not?

1.5.4 Other analytically solvable problems

We just saw the classic Rabi problem and its analytical solution. As we discussed, the
Schrödinger equation for a general time-dependent Hamiltonian is not analytically
solvable, even in the 2-level case. Some exceptions are:
• Hamiltonians that commute with themselves at all times; e.g.,

H(t) = f (t)σx . (1.76)

Here, we can obtain the evolution operator

R by integrating
R the Hamiltonian and
−i Hdt −iσx f (t)dt
exponentiating the result: u(t) = e =e .

• What about the following Hamiltonian:

ω
H= σz + Ω(t) cos(ωt)σx ?
2
[H(t1 ), H(t2 )] = iωΩ(t2 ) cos(ωt2 )σy − iωΩ(t1 ) cos(ωt1 )σy 6= 0

Since the Hamiltonian doesn’t commute with itself at different times, it seems
like we shouldn’t be able to find the evolution operator analytically. However,
if we perform the transformation

e−iωt/2 0
T = and take the RWA, we can get it in the form (1.76) and
0 eiωt/2
then solve it.
• All resonant pulses are analytically solvable.
It also sometimes happens that the time-dependent Schrödinger equation can be
rewritten as a second-order differential equation whose solutions have been studied
in depth. Examples include:
• Rosen-Zener, 1932: Hyperbolic secant envelope

29
 • Landau-Zener, 1932: Linearly varying energies.

1.6 Pictures and frames

1.6.1 Rotating frame (or drive frame)
In the Rabi problem, we replaced

|ψ(t)i =a0 (t) |0i + a1 (t) |1i with

eiωt/2 b0 |0i + e−iωt/2 b1 |1i

This allowed us to cancel the pulse time-dependence. This transformation takes us

to the “rotating frame” of the driving field. This frame rotates with frequency ω (the
driving frequency) relative to the original lab frame. We saw that this frame transfor-
mation (together with the RWA) gave us a Hamiltonian that was time-independent,
and thus we could solve the Schrödinger equation in this new frame analytically.

1.6.2 Interaction picture

We can also transform to a frame that rotates with frequency ω0 (the transition
frequency) relative to the lab frame. This frame is generally known as the “interaction
picture”. For a general Hamiltonian with arbitrarily many levels, the transformation
to the interaction picture works as follows:

Hu = iu̇ ; H = H0 + H1
Define u0 as sln to H0 : H0 u0 = iu̇0
Then take u = u0 u1 (no assumptions/approx’s)

iu̇ = iu̇0 u1 + iu0 u̇1 = H0 u0 u1 + H1 u0 u1

 
†
iu0 u̇1 = H1 u0 u1 ⇒ iu̇1 = u0 H1 u0 u1

iu̇1 = H̃1 u1 in interaction picture

30
1.6.3 General frame transformation
The rotating frame and interaction picture are common examples of a more general
notion of frames/pictures. There are infinitely many possible frames that we can
transform into to make the math a little easier. Regardless of which frame we choose
to work in, all the physical results we obtain will be the same. Starting from the
time-dependent Schrödinger equation in the lab frame,

Hu = iu̇,

we go to a new frame by defining a unitary transformation operator T (t), which can

in general be time-dependent.

To find the evolution operator in the new frame, we first multiply by T †

T † Hu = iT † u̇ ⇒ T † HT |{z}
T † u = iT † u̇
ũ

We can then rewrite the equation as a Schrödinger equation for a new evolution
operator ũ:
 
T † u ≡ ũ ⇒ u = T ũ ⇒ u̇ = Ṫ ũ + T ũ˙

  h i
⇒ T † HT ũ = i T † Ṫ ũ + iT † T ũ˙ ⇒ iũ˙ = T † HT − iT † Ṫ ũ

We can refer to the operator in the square brackets as an effective Hamiltonian

in this new frame:
Hef f = T † HT − iT † Ṫ
If we manage to find a T such that Hef f commutes with itself at different times, then
we can easily solve the Schrödinger equation for ũ. Afterward, we can transform back
to the lab frame to obtain u. However, there is no general criterion that tells us if
such a T exists or how to find it. Given a new time-dependent Hamiltonian, you
might try the popular rotating frame or interaction picture first to see if either one
leads to a nice Hef f . Beyond that, you will likely resort to numerics to find u unless
you are very lucky and happen to guess a nice T .

31
1.7 Revisiting the case of a spin in a constant magnetic field
Previously, we saw that we can obtain the evolution operator for a constant magnetic
field by diagonalizing the Hamiltonian, exponentiating, and then applying the inverse
transformation:
~ = Bx x̂ + Bz ẑ
B Bx = const, Bz = const
ωx ωz
H= σx + σz
2 2 p 
1 ωx2 + ωz2 0
u=T e † −iHd t
T, Hd = p
2 0 − ωx2 + ωz2


 p
cos 2θ
sin 2θ
cos(θ) = ωz /p ωx2 + ωz2
T =
− sin 2θ cos 2θ sin(θ) = ωx / ωx2 + ωz2

Alternatively, we can use the Taylor series expansion of the exponential together
with the fact that n̂ · ~σ = 1 for any unit vector n̂:

σθ = cos(θ)σz + sin(θ)σx
u = e−iλtσθ /2 p
λ = ωx2 + ωz2
 2  3
λt 1 λt 1 λt
u = 1 − i σθ − 1+i σθ + . . .
2 2 2 6 2
   
λt λt
= cos 1 − i sin σθ
2 2
 
1
Suppose χ(t = 0) = = |0i. What is χ(t)?
0
      
λt 1 λt 1
χ(t) =u(t)χ(0) = cos − i sin σθ
2 0 2 0
     
1 1 0
σθ = cos(θ) + sin(θ)
0 0 1
 λt

λt


cos 2 − i sin
2 cos(θ)
⇒ χ(t) = normalized? Yes!
−i sin λt
2
sin(θ)

32
 To understand more intuitively what the state is doing, we’ll make use of the spin
vector representation. First, let’s introduce this representation.

Spin vector representation

A nice way to visualize what the state of a 2-level system is doing is to use the “spin
vector” representation. The spin vector is a collection of three expectation values,
one for each Pauli matrix:
 
hψ(t)|σx |ψ(t)i
~ ≡  hψ(t)|σy |ψ(t)i
S
hψ(t)|σz |ψ(t)i

33
 Here are some popular qubit states along with their spin vector representation:
 
0
|ψi = |0i , ~ = 0 = ẑ
S
1 
0
|ψi = |1i , ~
S=  0  = −ẑ
−1
1
|ψi = |0i+|1i
√ , ~ = 0 = x̂
S
2
0 
−1
|ψi = |0i−|1i
√ , ~
S=  0  = −x̂
2
 0
0
|ψi = |0i+i|1i
√
2
, S = 1 = ŷ
~ 
0 
0
|ψi = |0i−i|1i
√ , ~ = −1 = −ŷ
S
2
0 


sin(θ) cos(φ)
~ =  sin(θ) sin(φ) 
|ψi = cos 2θ |0i + eiφ sin 2θ |1i , S
cos(θ)

The last case above is a general qubit state. We see that every qubit state maps to
a unique unit spin vector in 3D Euclidean space (ignoring possible global phases on
the qubit state). Therefore, given S,~ we can find |ψi. Because S ~ is a unit vector,
the tip of the spin vector lives on a unit sphere called the Bloch sphere. Thus, each
point on the Bloch sphere represents a different state of the two-level system. The
Bloch sphere is the Hilbert space of a two-level system.

Coming back to our spin in a constant magnetic field problem, where the state
at time t is 


cos λt
2
− i sin

λt
2
cos(θ)
χ(t) = ,
−i sin λt2
sin(θ)

34
we can compute the spin vector components:
 
† 0 1
hσx iχ(t) = χ (t) χ(t)
1 0






−i
sin λt sin(θ)
λt λt
= cos 2 + i sin 2 cos(θ) +i sin 2 sin(θ) λt 2

cos λt2
− i sin λt
cos(θ)
      2   
λt λt λt λt λt
= −i cos sin sin(θ) + sin2 sin(θ) cos(θ) + i cos sin sin(θ)
2 2 2 2 2
 
2 λt
+ sin sin(θ) cos(θ)
2
 
2 λt
= sin sin(2θ)
2

 
† 0 −i
hσy iχ(t) = χ (t) χ(t)
i 0






− sin λt
sin(θ)
= cos λt + i sin λt cos(θ) +i sin λt sin(θ)
2 λt

2 2 2 i cos λt
2
+ sin 2 cos(θ)
         
λt λt 2 λt λt λt
= − cos sin sin(θ) − i sin sin(θ) cos(θ) − cos sin sin(θ)
2 2 2 2 2
 
2 λt
+ i sin sin(θ) cos(θ)
2
= − sin(λt) sin(θ)

 
† 1 0
hσz iχ(t) = χ (t) χ(t)
0 −1







cos λt − i sin λt cos(θ)
λt λt λt
= cos 2 + i sin 2 cos(θ) +i sin 2 sin(θ) 2
2
i sin λt
2
sin(θ)
     
λt λt λt
= cos2 + sin2 cos2 (θ) − sin2 sin2 (θ)
2 2 2
   
2 λt 2 λt
= cos + sin cos(2θ)
2 2

35
Putting these together, we then have the following spin vector:


sin2 λt
2
sin(2θ)
~=
S −
sin(λt) sin(θ) 
2 λt 2 λt


cos 2 + sin 2 cos(2θ)

Before analyzing this in detail, we should first check that it is properly normalized:
       
~ ·S~ = sin 4 λt 4 λt 2 λt 2 λt
S + cos + 2 cos sin cos(2θ)
2 2 2 2
+ sin2 (λt) sin2 (θ)
| {z } | {z }
4 sin2 ( λt 2 λt 1 (1−cos(2θ))
2 ) cos ( 2 ) 2

=1 X

Let’s see what the spin vector is doing at a few particular times:
z  
B 0
~
S(0) = 0 = ẑ
1
θ  
0
x ~ = 2π ) =? 0
S(t
λ
1
 
sin(2θ)
~ = π ) =?  0 
S(t
λ
-y cos(2θ)
What is the spin vector doing as time evolves? It precesses uniformly around the
magnetic field direction, maintaining a constant angle θ between the spin vector and
magnetic field.
What should θ be if I want to do a bit flip, i.e. |0i → |1i? This can be achieved by
finding a time at which the evolution operator is equal to a NOT gate. Above, we

36
found that the evolution operator is
   
λt λt
uconst B (t) = cos 1 − i sin σθ = e−iλtσθ /2
2 2




cos λt
2
− i sin λt

2 cos(θ) −i
sin λt
2
sin(θ)


=
−i sin λt 2
sin(θ) cos λt2
+ i sin λt2
cos(θ)

How should we choose θ, t to get a NOT gate?

NOT = e−iπσx /2 = −iσx

2π π π  
⇒t= = , θ= ~ = B0 x̂
B
λ2 λ 2

What if we want a Rz (φ) gate instead?

Rz (φ) = e−iφσz /2

φ  
⇒t= , θ=0 ~ = B0 ẑ
B
λ

What if we want a Hadamard gate?

 
1 1 1
Hadamard = √
2 1 −1
 
π π ~ B0
⇒t= , θ= B = √ (x̂ + ẑ)
λ 4 2

(This gives us a Hadamard gate up to a global phase, which we don’t care about.)

1.8 Spin in precessing magnetic field revisited:

~ = B0 cos(ωt)x̂ + B0 sin(ωt)ŷ
B

37
  
0 be−iωt geB0
H= iωt b=
be 0 2me
Previously found u(t) = R† (t)e−iH̃t
 −iωt/2   ω 
† e 0 −2 b
R = H̃ = ω
0 eiωt/2 b 2

e−iH̃t =? Write H̃ = λ2 (cos(θ)σz + sin(θ)σx )

H̃
q is just like a Hamiltonian for a constant B field in the xz-plane, but now λ =
ω2
b2 + 4
, tan(θ) = − 2b
ω
   
λt λt
e−iH̃t
=e −iλtσθ /2
= cos 1 − i sin σθ
2 2
σθ = cos(θ)σz + sin(θ)σx
 −iωt/2 



e 0 cos λt2
− i sin λt

2 cos(θ) −i
sin λt
2
sin(θ)


u(t) =
0 eiωt/2 −i sin λt 2
sin(θ) cos λt 2
+ i sin λt 2
cos(θ)





cos λt
2
− i sin
λt
2
cos(θ) e−iωt/2 −i
sin λt2
sin(θ)e


−iωt/2


=
−i sin λt 2
sin(θ)eiωt/2 cos λt
2
+ i sin λt 2
cos(θ) eiωt/2

Does this have the correct structure for a 2x2 unitary matrix? Yes!

How does the spin vector evolve in this case?

χ(t) = u(t)χ(0)
 
1
Let’s again consider χ(0) = .
0  

1
~
To find S(t), we could compute Sj = 1 0 u (t)σj u(t) †
like before.
0
 −iωt/2 
e 0
However, here u(t) = u (t), and we already found the spin
0 eiωt/2 const B
vector for  
1
uconst B (t) = χconst B (t)
0

38
  −iωt/2 
e 0
Can we use the spin vector for χconst B (t) to find the spin vector for χ(t) = χ (t)?
0 eiωt/2 const B
 −iωt/2 
e 0 ~const B to S.
~
Yes! iωt/2 should rotate S
0 e
There is some 3x3 rotation matrix Mrot that does this. How can we find it?

1.9 3×3 rotation matrices from 2×2 unitaries:

u(t)χ(0) ⇔ M (t)S(0)~
  
1

† 1
   2 1 1 u (t)σx u(t) 1   
1    M (t)
1
† 1  xx
M (t)0 =   
 2 1 1 u (t)σy u(t) 1  = Myx (t)

0    Mzx (t)
1
† 1 
2
1 1 u (t)σz u(t)
1
 
1

† 1
Mky (t) = 2 1 −i u (t)σk u(t)
 i
Similarly,
† 1
Mkz (t) = 1 0 u (t)σk u(t)
0
 −iωt/2 
† e 0
In the present problem, we want to map the 2x2 unitary R =
0 eiωt/2
†
to the 3x3 rotation matrix MR† . First, we compute Rσk R :
 iωt/2     
† e 0 0 1 e−iωt/2 0 0 eiωt
Rσx R = = −iωt
0 e−iωt/2 1 0 0 eiωt/2 e 0
 
0 −ieiωt
Rσy R† =
ie−iωt 0
Rσz R† = σz

39
     
1
0 eiωt 1 1
eiωt
MR† ,xx = 1 1 = 1 1 = cos(ωt)
2 e−iωt 0 1 2 e−iωt
    
1
0 eiωt 1 1
ieiωt
MR† ,xy = 1 −i = 1 −i = − sin(ωt)
2 e−iωt 0 i 2 e−iωt
 

1
MR† ,zz = 1 0 σz =1
0
etc.
 
cos(ωt) − sin(ωt) 0
⇒ MR† =  sin(ωt) cos(ωt) 0
0 0 1

Therefore, the total spin vector at time t is

~ = MR† (t)S
S(t) ~const B (t)
 
 q
cos(ωt) − sin(ωt) 0 sin2 λt sin(2θ) 2
2
λ = b2 + ω4
=  sin(ωt) cos(ωt) 0 −
sin(λt) sin(θ)



0 0 1 2 λt 2 λt
cos 2 + sin 2 cos(2θ) tan(θ) = − 2b
ω

What happens when ω  b? In this limit, θ → 0, so that S ~const B (t) → ẑ, and
~
similarly for S(t). Therefore, we see that when the magnetic field precesses very
fast, the spin remains locked in its initial state, |0i.

What happens when ω  b? Now θ → −π/2, and we have

 
0
~const B (t) →  sin(bt) ,
S
cos(bt)

which is a precession about the x axis. MR† adds an additional rotation about the z
axis on top of this:
 
− sin(bt) sin(ωt)
~ →  sin(bt) cos(ωt) .
S(t) (1.77)
cos(bt)

The spin vector quickly rotates back and forth between the north and south poles
of the Bloch sphere at rate b. Notice that the spin vector comes back to its initial

40
state along the ẑ axis periodically with period T = 2π/b. In addition, these looping
pole-to-pole trajectories slowly rotate around the z axis at rate ω. See Fig. 1. This

Figure 1: Trajectory of the tip of the spin vector on the Bloch sphere in the case of
a spin interacting with a slowly precessing magnetic field. Slowly precessing means
ω  b, which corresponds to the limiting case given in Eq. (1.77).

trajectory makes sense because it describes a spin that quickly precesses around each
orientation of the magnetic field as the field slowly rotates in the xy plane.

1.10 Rabi problem revisited

The Rabi problem is mathematically almost identical to the spin in precessing B field
problem. Recall from the homework that the evolution operator in the lab frame is

u(t) = eiωtσz /2 e−iλtσθ /2

The only difference compared to the evolution operator in the previous subsection
is that here, the sign in front of ω has changed, which effectively replaces R† with
R. Thus, we can immediately write down the spin vector using the result from the

41
previous subsection:
~ = MR (t)S
S(t) ~eff (t)
 

cos(ωt) sin(ωt) 0 sin2 λt
2
sin(2θ)
= − sin(ωt) cos(ωt) 0 −
sin(λt) sin(θ)



0 0 1 cos2 λt
2
+ sin2 λt 2
cos(2θ)

Here, S~eff (t) is the spin vector in the rotating frame. Of course λ and θ are different
in the Rabi problem:
q
λ = Ω20 + ∆2
Ω0
tan θ =
∆
In the large detuning (or weak driving) limit, ∆  Ω0 , we have θ → 0, and the spin
vector remains in its initial orientation along the ẑ axis. In the opposite limit as
∆ → 0 (approaching resonance), we have θ → π/2, and the spin vector becomes
 
− sin(Ω0 t) sin(ωt)
~ → − sin(Ω0 t) cos(ωt).
S(t) (1.78)
cos(Ω0 t)

This follows a similar trajectory on the Bloch sphere as shown in Fig. 1, except
that now the period is T = 2π/Ω0 , and the slow rotation about the ẑ axis is in the
opposite direction. This is consistent with our previous finding that we need ∆ = 0
in order to induce a complete transition from |0i to |1i (i.e., rotate the spin vector
from the north pole to the south pole), and that the time it takes to perform this
transition is 2π/Ω0 .

1.11 Adiabatic approximation

Next we will see that in the case of a slowly varying Hamiltonian we can find an
approximate solution for the evolution operator.

Consider the Schrödinger equation for an explicitly time-dependent H(t):

E
i ψ̇(t) = H(t) |ψ(t)i (1.79)

42
We can find instantaneous eigenvalues & eigenvectors:

H(t) |n(t)i = En (t) |n(t)i (1.80)

Let’s think about this equation. Here time is a parameter. Thus, Eq. (1.80) is a
static problem for fixed t.

Q: Is |n(t)i the solution of Eq. (1.79)? No! |n(t)i is defined to be the solution
of the eigenvalue problem in Eq. (1.80). There’s no reason why it should also solve
the differential equation in Eq. (1.79).

Because they are the eigenvectors of a Hermitian operator, the {|n(t)i} form a com-
plete orthonormal basis:
hn(t)|m(t)i = δnm (1.81)
Q: Are |n(t)i, |m(t0 )i orthogonal? No! These are eigenvectors of two different opera-
tors, H(t) and H(t0 ). There’s no reason there should be any special relation between
them.

Because the {|n(t)i} form a complete orthonormal basis, we can express ψ(t) in
terms of the |n(t)i basis: X
|ψ(t)i = bn (t) |n(t)i (1.82)
n

Plug (1.82) into (1.79):

X X X
i ḃn (t) |n(t)i + i bn (t) |ṅ(t)i = H(t) bn (t) |n(t)i
n n n
X
= bn (t)En (t) |n(t)i
n

and apply hm(t)| to both sides:

43
 X X X
i ḃn (t) hm(t)|(n(t)i +i bn (t) hm(t)|ṅ(t)i = En (t)bn (t) hm(t)|n(t)i
n
| {z } n n
| {z }
δmn δmn
X
iḃm (t) + i bn (t) hm(t)|ṅ(t)i = bm (t)Em (t)
n

Note that if H were time-independent, then {|ni} and En would be time-independent,

and we’d have bm (t) = e−iEm t bm (0), which is consistent with what we found for time-
independent Hamiltonians in Lecture 1.

Now let’s break the second term in the LHS into two parts: n = m and n 6= m:
P
iḃm (t) + ibm (t) hm(t)|ṁ(t)i + i n6=m bn (t) hm(t)|ṅ(t)i = bm (t)Em (t) (1.83)

To make progress with the terms involving |ṅ(t)i we can go to Eq. (1.80) and differ-
entiate both sides:

Ḣ(t) |n(t)i + H(t) |ṅ(t)i = Ėn (t) |n(t)i + En (t) |ṅ(t)i

Apply hm(t)| with m 6= n:

nm δ
z }| {
hm(t)|Ḣ(t)|n(t)i + hm(t)| H(t) |ṅ(t)i = Ėn (t) hm(t)|n(t)i +En (t) hm(t)|ṅ(t)i
| {z }
hm(t)|Em (t)

Since by assumption m 6= n, the first term in the RHS vanishes.

Then we have:

hm(t)|Ḣ(t)|n(t)i
hm(t)|ṅ(t)i = En (t)−Em (t)
(notice that we assume En (t) 6= Em (t) ∀t)

Plugging this into Eq. (1.83) we get:

X hm(t)|Ḣ(t)|n(t)i
iḃm (t) + ibm (t) hm(t)|ṁ(t)i + i bn (t) = bm (t)Em (t). (1.84)
n6=m
En (t) − Em (t)

Note that Eq. (1.84) is exact—no approximation has been made yet. The approx-
imation happens at the next step, where we assume that the Hamiltonian is slowly

44
varying. This means that we will throw out the term involving Ḣ, which leaves us
with
iḃm (t) + ibm (t) hm(t)|ṁ(t)i = bm (t)Em (t) (1.85)
At this stage we can go to a different frame by defining the coefficients cn (t)
through:
Rt
Em (t0 )dt0
bm (t) = cm (t)e−i 0

Rt
and define 0
Em (t0 ) dt0 = θm (t), which is called the dynamical phase.

Then we have that ḃm (t) = ċm (t)e−iθm (t) + cm (t)Em (t)e−iθm (t) (−i)
| {z }
−iEm (t)bm (t)
Plugging this into (1.85) we have

iċm (t)e−iθm (t) + Em (t)bm (t) + icm (t)e−iθm (t) hm(t)|ṁ(t)i = Em (t)bm (t)

So we are left with the equation

ċm (t) = − hm(t)|ṁ(t)i cm (t)

Q: What is the solution of this differential equation?

Rt 0 0 0
cm (t) = cm (0)e− 0 dt hm(t )|ṁ(t )i

Plugging this into Eq. (1.82) we finally have:

X
|ψ(t)i = cn (0) e−iθn (t) eiγn (t) |n(t)i
n
Z t
θn (t) = En (t0 ) dt0 : dynamical phase
0
Z t
γn (t) = i hn(t0 )| |ṅ(t0 )i dt0 : geometric phase
0

The adiabatic theorem tells us that if the rate of change of the Hamiltonian is
slow enough, the states don’t mix. Equivalently, if the initial state is an instanta-

45
neous eigenstate, it remains such throughout the evolution.

Q: Suppose the system is in an arbitrary state |ψ(t)i. How does the probability to
find it in the instantaneous eigenstate |n(t)i change as a function of time? Answer: It
doesn’t. The probability to find the system in |n(t)i is |hn(t)|ψ(t)i|2 = |cn (0)|2 =const.

1.12 Two-level system evolution operator in adiabatic limit

Following the procedure for general frame transformations discussed at the end of
Lecture 3, we transform the general time-dependent Schrödinger equation using a
time-dependent transformation Λ(t):

Hu = iu̇
Λ† HΛΛ† u = iΛ† u̇


Λ† HΛ ũ = iΛ† Λ̇ũ + iũ˙
 
Λ† HΛ − iΛ† Λ̇ ũ = iũ˙
| {z }
Hef f

Q: What happens to the first term if we take Λ to be a matrix made of the instanta-
neous eigenstates of H? The first term will be diagonal. We will take advantage of
this in the following analysis.

General 2-level system: B field with arbitrary time dependence

Consider the Hamiltonian for a spin 1/2 interacting with a general, time-dependent
~
magnetic field B(t) = b(t)[sin θ(t) cos φ(t)x̂ + sin θ(t) sin φ(t)ŷ + cos θ(t)ẑ]:
 
1~ b(t) cos θ(t) e−iφ(t) sin θ(t)
H(t) = B(t) · ~σ = (1.86)
2 2 eiφ(t) sin θ(t) − cos θ(t)

Here, we absorbed the Bohr magneton and Landé g factor into b. This Hamiltonian
in fact describes any time-dependent 2-level system (and thus any driven qubit).
For example, we can turn this into a Rabi-like problem (but with an arbitrary pulse
envelope shape) by choosing b(t) = b0 / cos θ(t) and φ(t) = ωt. We then have a Rabi-
like Hamiltonian where b0 is the transition frequency, ω is the driving frequency, and
(b0 /2) tan θ(t) is the (now time-dependent) pulse envelope. In this case, we would
want θ(0) = 0 or π to describe a proper pulse that starts at zero amplitude. From

46
the point of view of the time-dependent magnetic field, this corresponds to a field
that is initially lying along the ẑ axis.

The Λ(t) that diagonalizes H(t) has the instantaneous eigenvectors as its columns:
!
sin θ(t)
2
e−iφ(t)
cos θ(t)
2
Λ(t) = (1.87)
−eiφ(t) cos θ(t)
2
sin θ(t)
2

We then find that

!
− b(t) 0
Λ† HΛ = 2
b(t) diagonalizes H as it should
0 2

If H were time-independent, Λ would also be time-independent, and we’d be done.

However, it is not and we have the −iΛ† Λ̇ term. After some algebra, we find
 
† 1 (1 + cos θ)φ̇ ie−iφ (θ̇ + iφ̇ sin θ)
− iΛ Λ̇ = (1.88)
2 −ieiφ (θ̇ − iφ̇ sin θ) −(1 + cos θ)φ̇
 
1 −b + (1 + cos θ)φ̇ ie−iφ (θ̇ + iφ̇ sin θ)
⇒ Hef f = (1.89)
2 −ieiφ (θ̇ − iφ̇ sin θ) b − (1 + cos θ)φ̇
This Hef f is written in the basis of instantaneous eigenvectors.

The adiabatic approximation amounts to throwing out the off-diagonal terms

(recall that these are the terms involving matrix elements of Ḣ that we discarded in
the general analysis above), so that Hef f ≈ Had with
 
1 −b + (1 + cos θ)φ̇ 0
Had = .
2 0 b − (1 + cos θ)φ̇
This Hamiltonian commutes with itself at different times, so we can exponentiate
its integral to obtain the evolution operator in this frame:
Rt
dt0 Had (t0 )
⇒ uad (t) = e−i 0 = e−iχ(t)σz

with Z t
1
χ(t) = dt0 [−b(t0 ) + (1 + cos θ(t0 ))φ̇(t0 )].
2 0

The first term in χ(t) is the dynamical phase since it is just the integral of the in-

47
stantaneous energy. The second term must therefore be the geometric phase (up to
a minus sign that depends on which of the two instantaneous eigenstates we want to
focus on). We can see this easily from the fact that the inner products hn(t)|ṅ(t)i
are just the diagonal entries of Λ† Λ (see Eq. (1.88)) since the instantaneous eigenvec-
tors |n(t)i are the columns of Λ(t). We’ll return to why this is called the geometric
phase below when we talk about Berry’s phase, which is another name for this phase.

It is tempting to equate (approximately) ũ with uad since uad is an evolution

operator in the frame defined by Λ. However, we need to be a little careful here.
In general if the transformation matrix Λ(t) is defined such that Λ(0) 6= 1, then we
need to adjust the initial condition of ũ accordingly. Recall from above that the lab
frame evolution operator is u(t) = Λ(t)ũ(t), and it must always be the case that
u(0) = 1. This means that ũ(0) = Λ† (0), and this is not equal to the identity matrix
unless θ(0) = π, i.e., the magnetic field is initially pointing in the −ẑ direction. On
the other hand, the expression for uad (t) we found above satisfies uad (0) = 1 instead.
Therefore, we must multiply this result on the right by Λ† (0) to get ũ:

ũ(t) = uad (t)Λ† (0)

Note that we can always multiply an evolution operator on the right by a constant
unitary matrix, and the result will still satisfy the same time-dependent Schrödinger
equation. This freedom is just the usual freedom we have in choosing initial condi-
tions when we solve an ordinary differential equation. We can then write down the
evolution operator in the lab frame under the adiabatic approximation. In general,
we have

u ≈ Λ(t)uad Λ† (0)
Applied to our time-dependent magnetic field problem, this formula gives a some-
what messy result, but if we restrict attention to the case where θ(0) = π, we find a
fairly simple result:
 −iχ/2 
e sin 2θ e−i(φ−χ/2) cos 2θ
u≈
−ei(φ−χ/2) cos 2θ eiχ/2 sin 2θ

Q: Suppose the qubit is initially in the state |0i. What is its state at later times?
What is special about this state?

48
1.13 Berry’s phase
Let’s return to the question of what is geometric about the geometric phase. For the
2-level system problem above, we found that the geometric phases are
Z
1 t 0
γ0 (t) = − dt [1 + cos θ(t0 )]φ̇(t0 )
2 0
Z
1 t 0
γ1 (t) = dt [1 + cos θ(t0 )]φ̇(t0 )
2 0
These are called geometric because they depend on the path traced out on the unit
sphere by the evolving magnetic field, but they do not depend on how quickly this
path is traversed. To see this, notice that we can rewrite the integral as
Z t Z φ(t)
0 0 0
dt [1 + cos θ(t )]φ̇(t ) = dφ0 {1 + cos[Θ(φ0 )]},
0 0

where Θ(φ) is the path on the unit sphere. If we speed up or slow down the rate
at which the magnetic field moves along this path, this integral remains the same.
We could do this for instance by rescaling time: t → λt for some constant λ. This
only changes the upper integration limit in the second integral: φ(t) → φ(λt). So
long as the same path is followed, this has no effect on the value of the integral.
This is in contrast to the dynamical phase, which depends on both the path and the
rate at which the path is traversed. For example, if the magnitude of the magnetic
R t 0 0 then the dynamical phase for the instantaneous eigenstate |0(t)i is
field is constant,
θ0 = − 0 dt b(t ) = −bt, which clearly depends on the total evolution time.
Although the geometric phases are “rate-independent”, they do generally depend
on how we choose global phases in our instantaneous eigenstates. Consider the
eigenstate |n(t)i with geometric phase
Z t
γn (t) = i dt0 hn(t0 )|ṅ(t0 )i
0

Obviously the state |ñ(t)i = eif (t) |n(t)i, where f (t) is some real time-dependent
function, is also an instantaneous eigenstate of H(t), and we could have just as well
chosen to work with it instead of |n(t)i. Thus, there is some arbitrariness in how we
choose our instantaneous eigenstates. Moreover, the geometric phase for this state

49
is different from the one for |n(t)i:
Z t
γñ (t) = i ˙ 0 )i = γn (t) − f (t)
dt0 hñ(t0 )|ñ(t
0

This time-dependent global phase transformation is an example of a gauge transfor-

mation. It has close mathematical connections to the notion of gauge transforma-
tions in electromagnetism. We have just seen that geometric phases are generally
not gauge-invariant.

Back in 1984, Michael Berry came up with a way to make this notion of geometry-
only-dependent phases a bit more formal and general. In addition, he showed that
if the geometrical path traversed by the system is closed, then the corresponding
geometric phases are gauge-invariant. Berry’s key insight was to rewrite the phase
parametrically, that is, suppose the Hamiltonian depends on time through its depen-
~
dence on some set of parameters R(t) (here we collect these parameters together into
~
a real vector). For example, in our 2-level magnetic field problem above, R(t) ~
= B(t).
We can then rewrite the geometric phase in terms of R: ~
Z t Z ~
R(t)
d ~
hn| |ni dt = hn(R)|∇R |n(R)i · dR
0 dt ~0
R

For a closed curve in (multi) parameter space, this becomes a loop integral:
I
~
hn(R)|∇R |n(R)i · dR

Let’s revisit the spin in a general time-dependent magnetic field problem, where
we found the instantaneous eigenstates
θ θ
|0(R)i = sin |0i − eiφ cos |1i
2 2
θ θ
|1(R)i = e−iφ cos |0i + sin |1i
2 2
Since we have R~ = B,
~ our parameters are b, θ, and φ, which can be thought of
as spherical coordinates (with r = b as the radial coordinate). We can then compute
∇R |n(R)i using the expression for the gradient in spherical coordinates:

50
 ∂ 1 ∂ 1 ∂
∇R |n(R)i = |n(R)i r̂ + |n(R)i θ̂ + |n(R)i φ̂
∂r r ∂θ r sin θ ∂φ
Let’s consider |0(R)i for concreteness. The first term in the gradient vanishes
since |0(R)i does not depend on r = b. The second term gives

1 ∂ 1h θ θ i
|0(R)i θ̂ = cos |0i + eiφ sin |1i θ̂
r ∂θ 2r 2 2
while the third gives

1 ∂ ieiφ cos 2θ
|0(R)i φ̂ = − |1i φ̂
r sin θ ∂φ r sin θ

The term proportional to θ̂ is orthogonal to h0(R)| and so it doesn’t contribute

to the geometric phase. On the other hand, the φ̂ term gives a nonzero contribution:

1 ∂ i cos2 2θ
h0(R)| |0(R)i φ̂ = φ̂
r sin θ ∂φ r sin θ
The Berry phase is then
I I I
~ =i i cos2 2θ 1
γ0 = i h0(R)|∇R |0(R)i · dR r sin θdφ =− dφ(1 + cos θ)
r sin θ 2
which is consistent with what we obtained previously. This integral can be recognized
as half the solid angle enclosed by the path. For instance, if we consider the case
~
where B(t) precesses around the ẑ axis at a fixed polar angle θ, then the Berry phase
becomes
γ0 = −π(1 + cos θ)

51
2 The density operator and open quantum sys-
tems
2.1 Projectors and outer product
Given that a matrix is written in an orthonormal basis, we have for a general matrix
M
X
M= Mij |iihj|.
ij

Let’s see why this works. Take a matrix element of M using Dirac notation:
X X X
hk|M |`i = Mij hk|(|iihj|)|`i = Mij hk|iihj|`i = Mij δik δj` = Mk` .
ij ij ij

Let’s consider a few examples. For a diagonal matrix, we will clearly have a sum
of weighted projectors. The identity operator in particular, which has 1’s in the
diagonal and 0’s everywhere else can be expressed as

X
L
1L = |kihk|. (2.1)
k=1

Q: How would we express the Pauli operators?

2.2 The density operator

An observable in quantum mechanics corresponds to a Hermitian operator the mean
value of which can be measured experimentally. For example, all the Pauli matrices
correspond to quantities that can be measured. The mean value then corresponds to
the average outcome if we were to make a large number of measurements on identical
copies of the quantum state. For example, consider a qubit in state |ψi and let’s
assume we are making a measurement of X (in practice, this is the same as measuring
the qubit in the |±i basis. The mean value of X in state |ψi is

X̄ = hψ| X |ψi . (2.2)

Let us first consider the case |ψi = |0i. The operator X flips |0i to |1i, and the
overlap of that with |0i is 0, so we find that X̄ = 0 in this case. A different way to

52
think of this is to consider a large number of qubits, each in the state |0i. Every time
we make a measurement in the |±i basis (the eigenbasis of X) we find either |+i or
|−i, each with probability 0.5. Since the former corresponds to the +1 eigenvalue and
the latter to the −1, the average is 0. Next consider the case |ψi = |−i. Measuring
in the |±i basis now gives us outcome |−i with probability 1. The mean value of
X in this case should therefore be −1, which is the eigenvalue of state |−i. We can
easily check that this is the case, as h−| X |−i = −1. For later use, notice that we
could also express this as

Tr(X |−i h−|). (2.3)

Now we will consider more general states, which physically will correspond to
some degree of ignorance about them, similarly to using probability distributions,
as in statistical mechanics. Consider for example the situation where we have N
qubits in a box and we are told that half of them are in state |ψi and the other
half in state |χi. We want to be able to make predictions about measurements for
this system. Let us assume we are measuring an observable A, which is again a
Hermitian operator. What is the mean value of this operator for this many-qubit
state? In this case, we have to take a statistical (classical) average of the quantum
mechanical mean values. Specifically, we will have for this example:

Ā = 0.5 hψ| A |ψi + 0.5 hχ| A |χi . (2.4)

More generally, if p out of the N qubits are in state |ψi and N − p are in state |χi,
we have
p N −p
Ā = hψ| A |ψi + hχ| A |χi , (2.5)
N N
where Np and NN−p add to 1, as they represent (classical) probabilities for the qubits
to be in each of the two states. Notice that when all the qubits are in the same state
(i.e., when |χi = |ψi) then we get the expression for a single qubit (Ā = hψ| A |ψi).

Let us assume for simplicity that |ψi = |0i and |χi = |1i. Then Eq. 2.5 becomes

p N −p
Ā = h0| A |0i + h1| A |1i . (2.6)
N N
Notice that this expression looks like the trace of some operator, i.e., we can rewrite
it as
Ā = Tr(Aρ), (2.7)

53
where ρ = Np |0i h0| + NN−p |1i h1|. The operator ρ, called the density operator or
density matrix, represents the state of a qubit in the ensemble. While we discussed
the example above as a collection of N qubits, we could also recast the discussion in
terms of a single qubit with (classical) probability Np to be in quantum state |0i and
N −p
N
to be in quantum state |1i.

Notice from our example above with the X operator that we can express any
quantum state in terms of a density operator. In the case where the qubit is in state
|−i for example, ρ = |−i h−|. Any state that can be written as a vector |ψi, or
equivalently, as a projector |ψi hψ| is called a pure state. Any state that cannot be
written that way is called a mixed state. The qubit state for which we have minimal
information is 0.5 |0i h0| + 0.5 |1i h1| and is called a maximally mixed state.

2.3 Properties of the density operator

The density operator has the following properties:

• It is Hermitian: ρ† = ρ

• Its diagonal elements represent probabilities. As a result they are positive real
numbers between 0 and 1 and

• the trace of ρ is one: Tr(ρ) = 1

• ρ is a pure state iff ρ2 = ρ

• Tr(ρ2 ) ≤ 1. The equality only holds for pure states

Since the density operator is a Hermitian matrix, it has real eigenvalues and an
orthonormal eigenbasis. In fact, these correspond to the probabilities to be in each
of the corresponding eigenstates.

Notice that for a maximally mixed state of a qubit we have Tr(ρ2 ) = 1/2, whereas
in a d-dimensional Hilbert space we would have Tr(ρ2 ) = 1/d for a maximally mixed
state.

54
2.4 Bloch vector description of mixed states
For pure states, the density operator and the state vector descriptions are completely
equivalent. We know that a general pure state can be parameterized as
θ θ
|ψi = cos |0i + eiφ sin |1i .
2 2
We can immediately write down the density operator for this state:
θ θ θ θ θ θ
ρ = |ψi hψ| = cos2 |0i h0|+sin2 |1i h1|+e−iφ cos sin |0i h1|+eiφ cos sin |1i h0| .
2 2 2 2 2 2
(2.8)
Remember that the components of the Bloch vector are Sz = cos θ, Sx = sin θ cos φ
and Sy = sin θ sin φ. These are in fact the mean values of the operators Z, X, and
Y . Let’s verify that.
X̄ = Tr(Xρ) = sin θ cos φ = Sx , (2.9)
etc. In the case of pure states, this is the same as

X̄ = hψ| X |ψi . (2.10)

So, the mean of the three Pauli operators give the Bloch vector components. Recall
also that we can express any Hermitian matrix in terms of the Pauli matrices and
the identity with real coefficients. We thus have for the density operator:

ρ = a0 1 + ax X + ay Y + az Z. (2.11)

Multiplying with each of the Paulis and taking the trace leads to the assignment
a0 = 1/2 and aj = Sj /2. We can therefore express an arbitrary density operator for
a qubit as
1
ρ = (1 + Sx X + Sy Y + Sz Z) , (2.12)
2
which in shorthand can also be written as ρ = 21 (1 + S ~ · ~σ ), where ~σ is a vector of
the Pauli matrices. Notice that, unlike pure quantum states, where two parameters
suffice to parameterize a state, for mixed states we need three parameters. This will
become visually obvious when we sketch the vectors using the Bloch sphere.

Let’s look into some examples. For each of these, put the density operator in the
form of Eq. 2.11.

55
Example 1
A qubit has 50% probability to be in state |0i and 50% probability to be in state |1i:

ρ1 = 0.5 |0i h0| + 0.5 |1i h1| (2.13)

Notice that we can immediately see that ρ1 = 0.5 1. This is a maximally mixed
state. It has Sj = 0 for each j = x, y, z, meaning that it corresponds to a point in
the center of the Bloch sphere.

Example 2
A qubit has 50% probability to be in state |+i and 50% probability to be in state
|−i:
ρ2 = 0.5 |+i h+| + 0.5 |−i h−| (2.14)
There are two ways to work this out. One way—the simplest—is to realize imme-
diately that |+i h+| + |−i h−| = 1, which means that ρ2 = 0.5 1 = ρ1 . The other
approach is to write the |±i basis in terms of the computational basis.

Example 3
A qubit has 50% probability to be in state |+i and 50% probability to be in state
|0i:
ρ3 = 0.5 |+i h+| + 0.5 |0i h0| . (2.15)
√
Using the fact that |+i = (|0i + |1i)/ 2, we have

ρ3 = 0.75 |0i h0| + 0.25 |1i h1| + 0.25 |0i h1| + 0.25 |1i h0| . (2.16)
√
The spin vector is halfway between x and z with length 0.5.

Example 4
A qubit has 60% probability to be in state |+i and 40% probability to be in state
|−i:
ρ4 = 0.6 |+i h+| + 0.4 |−i h−| (2.17)

56
2.5 Time evolution of the density operator
Since the density operator can describe a general quantum state, we should be able
to find an equation of motion for it (the analog of the Schrödinger equation). Here
is how we go about doing that; simply start with the definition of ρ in terms of some
basis states and differentiate both sides with respect to time:

d d X
ρ= λi |ψi ihψi |
dt dt i
X d  X 
d

= λi |ψi i hψi | + λi |ψi i hψi |
i
dt i
dt
X
= λi [−iH |ψi ihψi | + i |ψi ihψi | H]
i
= −iHρ + iρH = i[ρ, H]
⇒ ρ̇ = i[ρ, H] Liouville-von Neumann eqn.

and ρ(t) = u(t)ρ(0)u† (t)

2.6 Rabi problem with dissipation

ρ̇ = i[ρ, H]
Let’s look at the Rabi problem:
 
∆/2 Ω/2
HRF = (H in rotating frame)
Ω/2 −∆/2

57
Define ρ as the DM in the RF: ρ̇ = i[ρ, HRF ]

ρ̇gg = i(ρge Heg + ρgg Hgg − Hgg ρgg − Hge ρeg )

Ω
= i (ρge − ρeg )
2
Ω
ρ̇ee = i(ρeg Hge − Heg ρge ) = i (ρeg − ρge )
2
ρ̇ge = iHge (ρgg − ρee ) + iρge (Hee − Hgg )
Ω
= i (ρgg − ρee ) − i∆ρge
2
(Which we can solve, as we solved the Rabi problem)

Case I: When Ω = 0 (no drive)

ρ̇gg = 0; ρ̇ee = 0; ρ̇ge = −i∆ρge

ρgg (t) = ρgg (0); ρee (t) = ρee (0); ρge (t) = ρge (0)e−i∆t
→ phase is in the off-diagonal matrix element.
Q: How do you explain physically the fact that the diagonal elements don’t change
in time in this case? Answer: The Hamiltonian is time-independent, so energy is
conserved.

For Ω 6= 0 : ρgg (t), ρee (t) also change in time.

Let’s analyze the phyiscs of the Rabi problem.

Two types of transitions:
|1> |1>

absorption emission

|0> |0>

Both of these are implemented w/ a π-pulse

58
  
0 −i
u(π) = Reversible (as it should be for a unitary)
−i 0
Which process we have then depends on initial state
|1>

More accurate schematic:

|0>

What other processes can we have?

Consider a hydrogen atom in state |2pi: what will happen?
|2p>

spontaneous
emission
(irreversible)

|1s>

How do we describe such processes?

We need to include non-unitary terms.

For a two-level system, the simplest equations to describe spontaneous emission

are given by (we will give these without proof, which would require us to consider a
quantized bosonic bath):

ρ̇ee = i[ρ, H]ee − γ1 ρee

ρ̇gg = i[ρ, H]gg + γ1 ρee ← this follows from conservation of probability
γ1
ρ̇eg = i[ρ, H]eg − γ2 ρeg ; γ2 ≥ ← physics of this constraint?
2
The constraint γ2 ≥ γ1 /2 comes from the requirement Tr(ρ2 ) ≤ 1 (or equivalently,
det(ρ) > 0 since Tr(ρ2 ) = 1−2det(ρ) for a 2-level system). Consider an initial density

59
matrix subject to both relaxation and dephasing:
 
1 − ρee e−γ1 t ρ∗eg e−γ2 t
ρ(t) =
ρeg e−γ2 t ρee e−γ1 t

We have
det(ρ) = ρee e−γ1 t − ρ2ee e−2γ1 t − |ρeg |2 e−2γ2 t
This must be nonnegative at all times in order for ρ(t) to be physically sensible.
Suppose that γ2 < γ1 /2. Then at very late times, the first two terms become small
compared to the last, in which case det(ρ) becomes negative, which is problematic.
Therefore, it must be that γ2 ≥ γ1 /2. Physically, this constraint makes sense because
the relaxation process from the excited state to the ground state necessarily destroys
superpositions and hence the coherence terms in the density matrix must decay.

The inverses of γ1 and γ2 have units of time. These time scales are called T1 and
T2 : T1 = 1/γ1 and T2 = 1/γ2 . T1 is commonly referred to as the energy relaxation
time (or just relaxation time), while T2 is called the coherence (or sometimes deco-
herence) time. T1 and T2 are two of the main characteristics people use to describe
the quality of a qubit. In terms of these time scales, the constraint γ2 ≥ γ1 /2 be-
comes T2 ≤ 2T1 .

Qualitative difference between γ1 & γ2 ? (energy relaxation vs decoherence).

Let’s look at an example.

√1 (|gi
ρee (0) = ρgg (0) = 1/2
Initial state |ψ(0)i = |+i = + |ei) ⇒
2 ρeg (0) = ρge (0) = 1/2
No drive, just free evolution & γ1 , γ2
Lab frame:
  Going to interaction picture:
Eg 0  
HL = 0 0
0 Ee † †
u HL u + iu̇ u = =H
0 0
u = e−iHL t
ρL : Lab frame density matrix
ρ: interaction picture density matrix. ρ = u† ρL u

60
 Notice that the dissipation equations also need to be transformed

ρ̇ = i[ρ, H] + L(ρ)

In this process:
1
ρ̇ee = −γ1 ρee ⇒ ρee (t) = e−γ1 t ρee (0) = e−γ1 t
2
Z t
ρ̇gg = γ1 ρee ⇒ ρgg = γ1 ρee (t0 ) dt0 + ρgg (0)
Z0 t
γ1 0
= dt0 e−iγ1 t + ρgg (0)
2 0
γ1 1 −γ1 t
=− (e − 1) + ρgg (0)
2 γ1
1 1 1
= + (1 − e−γ1 t ) = 1 − e−γ1 t
2 2 2
Consistency check: ρee + ρgg = 1
1
ρ̇ge = −γ2 ρge ⇒ ρge (t) = e−γ2 t ρge (0) = e−γ2 t
2
Go back to Schrödinger picture / lab frame

1
ρL,ge = uρu†= ugg ρge u†ee = e−i(g −e )t e−γ2 t
ge 2
Long time behevior? (t → ∞): ρee → 0, ρgg → 1, ρge → 0

Let’s see what the purity of ρ(t) is in this problem:

ρ(0) pure
ρ(t → ∞) pure
ρ(t), t ∈ (0, ∞) mixed

2.7 Lindblad operators

In the general case we can express dissipative terms (in the Markovian limit) using
so-called Lindblad operators and have for the equation of motion for the density

61
operator:
X 1 † 1 †

ρ̇ = i[ρ, H] + Li ρL†i − Li Li ρ − ρLi Li , (2.18)
i
2 2

where the Li are sometimes called jump or transition operators. The Lindblad
terms are designed to guarantee that ρ(t) obeys all the physical constraints it is
supposed to. As an example, consider the two-level system case we have already
seen and set γ1 = 2γ2 . This case is captured by choosing the following Lindblad
√
operator: L1 = γ1 |gi he|. Plugging this into the Lindblad equation gives the
following dissipation terms:
 
† 1 † 1 † γ1 ρee −γ1 ρ∗eg /2
L1 ρL1 − L1 L1 ρ − ρL1 L1 =
2 2 −γ1 ρeg /2 −γ1 ρee

which is precisely what we found earlier.

In this example, we saturated the inequality constraint between the decoherence
and relaxation rates: γ2 = γ1 /2. In this case, the decoherence is due entirely to
relaxation. However, the system can of course decohere even faster than this de-
pending on the nature of the dissipation, in which case we would have γ2 > γ1 /2 (or
T2 < 2T1 ). The extra decoherence that comes from additional dissipation mecha-
nisms (beyond relaxation processes) is generally called dephasing. People often talk
about a separate dephasing rate γφ to characterize these processes, so that we have:

γ2 = γ1 /2 + γφ

In terms of time scales, this is

1/T2 = 1/(2T1 ) + 1/Tφ

In our two-level system example, we can include dephasing in the Lindblad equation
by adding a second Lindblad operator:
p
Lφ = 2γφ |gi hg|

The corresponding dissipation term in the evolution equation is then

 
† 1 † 1 † 0 −γφ ρ∗eg
Lφ ρLφ − Lφ Lφ ρ − ρLφ Lφ =
2 2 −γφ ρeg 0

62
To include both relaxation and dephasing in our qubit evolution, we then need two
Lindblad terms: one with Li = L1 and another with Li = Lφ . The full dissipation
term is then
   
γ1 ρee −(γ1 /2 + γφ )ρ∗eg γ1 ρee −γ2 ρ∗eg
=
−(γ1 /2 + γφ )ρeg −γ1 ρee −γ2 ρeg −γ1 ρee

2.8 Bloch equations, dissipation and dephasing

Now we will use the spin vector representation to find the steady state of a driven
two-level system under relaxation and dephasing. We parametrize the density matrix
as  
ρ00 ρ01
ρ=
ρ10 ρ11
The spin vector components are then
ρ10 − ρ01
Sx = ρ10 + ρ01 Sy = Sz = 2ρ00 − 1
i
or conversely
1 + Sz Sx + iSy Sx − iSy 1 − Sz
ρ00 = ρ10 = ρ01 = ρ11 =
2 2 2 2
Consider again the following Rabi-type Hamiltonian:
 
∆/2 Ω/2
H=
Ω/2 −∆/2

The Lindblad equation of motion with both relaxation and dephasing is

1 1 1 1
ρ̇ = i[ρ, H] + L1 ρL†1 − L†1 L1 ρ − ρL†1 L1 + Lφ ρL†φ − L†φ Lφ ρ − ρL†φ Lφ
2 2 2 2

63
For our driven two-level system this reduces to the following four equations:
Ω
ρ̇00 = i (ρ01 − ρ10 ) + γ1 ρ11
2
Ω
ρ̇11 = i (ρ10 − ρ01 ) − γ1 ρ11
2
Ω
ρ̇01 = i (ρ00 − ρ11 ) − i∆ρ01 − γ2 ρ01
2
Ω
ρ̇10 = −i (ρ00 − ρ11 ) + i∆ρ10 − γ2 ρ10
2
We want to rewrite these as equations for the spin vector:

Ṡz = ρ̇00 − ρ̇11 = Ω i(ρ01 − ρ10 ) +2γ1 ρ11

| {z } |{z}
Sy 1−Sz
2

= ΩSy − γ1 Sz + γ1
Ṡx = ρ̇10 + ρ̇01 = −∆ i(ρ01 − ρ10 ) −γ2 (ρ01 + ρ10 )
| {z } | {z }
Sy Sx

= −∆Sy − γ2 Sx
ρ̇10 − ρ̇01
Ṡy = = i(ρ̇01 − ρ̇10 )
 i 
Ω Ω
= i −i∆ρ01 + i (ρ00 − ρ11 ) − i∆ρ10 + i (ρ00 − ρ11 ) − γ2 (ρ01 − ρ10 )
2 2
= ∆ (ρ10 + ρ01 ) −Ω (ρ00 − ρ11 ) −γ2 i(ρ01 − ρ10 )
| {z } | {z } | {z }
Sx Sz Sy

= ∆Sx − ΩSz − γ2 Sy

To summarize:

Notice asymmetry

Ṡz = ΩSy − γ1 Sz + γ1 • between z & (x, y) in terms of γ1

term
Ṡx = −∆Sy − γ2 Sx
Ṡy = ∆Sx − ΩSz − γ2 Sy • between x & y in terms of Sz
↑ Q: B field along x,z
Can combine these together into the following form (analogous to magnetization

64
equation):
~˙ = B
S ~ ×S ~ + ~Γ(S)
~
where B~ = Ωx̂ + ∆ẑ and ~Γ(S)
~ = −γ2 Sx x̂ − γ2 Sy ŷ + γ1 (1 − Sz )ẑ

To obtain the steady state, we set Ṡj = 0:

ΩSy − γ1 Sz + γ1 = 0 (2.19)
−∆Sy − γ2 Sx = 0 (2.20)
∆Sx − ΩSz − γ2 Sy = 0 (2.21)
   2 
∆ ∆
(2.20), (2.21) ⇒ ∆ − Sy − γ2 Sy = ΩSz ⇒ −Sy + γ2 = ΩSz
γ2 γ2

Sy = − ∆γ22+γ
Ω
2 Sz
2
 
Ω2
Plug into (2.19) ⇒ − ∆γ22+γ 2 + γ1 Sz + γ1 = 0
2

γ1 γ1 (∆2 + γ22 )
Sz = Ω2
=
γ1 + γ2 γ1 (∆2 + γ22 ) + γ2 Ω2
∆2 +γ22

γ1 = 2γ
Take the simple case γ2 = γ1 /2 = γ
γ2 = γ

2γ(∆2 + γ 2 )
Sz =
2γ(∆2 + γ 2 ) + γΩ2
2(∆2 + γ 2 )
=
2(∆2 + γ 2 ) + Ω2
2(∆2 +γ 2 )
Steady-state Sz = 2(∆2 +γ 2 )+Ω2

65
 1 − Sz 1 ∆2 + γ 2
ρss
11 = = −
2 2 2(∆2 + γ 2 ) + Ω2
2(∆2 + γ 2 ) + Ω2 − 2(∆2 + γ 2 )
=
2(2(∆2 + γ 2 ) + Ω2 )
Ω2
=
4(∆2 + γ 2 ) + 2Ω2

On resonance:
Ω2 Ω2
ρss
11 (∆ = 0) = =
4γ 2 + 2Ω2 γ12 + 2Ω2
For strong drive Ω  γ1 , ∆:
ρss
11 → 1/2

Excited state population saturates at 1/2

Distinct from Rabi oscillations (with γ = 0)

3 Decoherence and dynamical decoupling

3.1 Ramsey interferometry: How to measure T2
Consider again the case of an undriven two-level system with Hamiltonian
 
E0 0
H=
0 E1

In the presence of both relaxation and dephasing, the Lindblad equation for the
density matrix is  
γ1 ρ11 −γ2 ρ01
ρ̇ = i[ρ, H] +
−γ2 ρ10 −γ1 ρ11

66
where γ2 = γ1 /2 + γφ . The Liouville term is
   
0 (E1 − E0 )ρ01 0 ω0 ρ01
i[ρ, H] = i =i
−(E1 − E0 )ρ10 0 −ω0 ρ10 0

where ω0 = E1 − E0 is the transition frequency of the qubit. The full set of evolution
equations for the density matrix are then

ρ̇00 = γ1 ρ11
ρ̇11 = −γ1 ρ11
ρ̇01 = iω0 ρ01 − γ2 ρ01
ρ̇10 = −iω0 ρ10 − γ2 ρ10

The general solution is

ρ11 (t) = ρ11 (0)e−γ1 t

ρ00 (t) = 1 − ρ11 (t) = 1 − ρ11 (0)e−γ1 t
ρ01 (t) = ρ01 (0)eiω0 t−γ2 t
ρ10 (t) = ρ10 (0)e−iω0 t−γ2 t

How can we measure γ2 ? This requires isolating its effects from those of γ1 . γ2
destroys superpositions, so to see its effect in an experiment, we should prepare a
superposition state and let it evolve. Suppose we start off in state |0i, which we can
initialize by just waiting for the qubit to relax to its ground state, which we assume
is |0i. We can then create a superposition by applying
√ a π/2 rotation about the y
axis. This produces the state |+i = (|0i + |1i)/ 2. We’ll take the density matrix
right after the rotation to be our initial condition:
 
1 1 1
ρ(0) = |+i h+| =
2 1 1

Using our general solution for ρ(t) above, we see that the density matrix at later

67
times has components
1
ρ00 (t) = 1 − e−γ1 t
2
1
ρ11 (t) = e−γ1 t
2
1 iω0 t−γ2 t
ρ01 (t) = e
2
1 −iω0 t−γ2 t
ρ10 (t) = e
2
After we let the system evolve for some, we then apply a second rotation, this time
by −π/2 rotation about y. The density matrix after this rotation, ρ0 , is given by
 
0 i π4 σy −i π4 σy 1 ρ00 + ρ11 + ρ01 + ρ10 −ρ00 + ρ11 + ρ01 − ρ10
ρ =e ρ(t)e =
2 −ρ00 + ρ11 − ρ01 + ρ10 ρ00 + ρ11 − ρ01 − ρ10

From this result, we see that the probability to find the qubit in the |0i state after
the second y rotation is
1 1 1
ρ000 = [ρ00 (t) + ρ11 (t) + ρ01 (t) + ρ10 (t)] = + [ρ01 (t) + ρ10 (t)]
2 2 2
Using our solution for ρ(t) from above, this is
1 1 −γ2 t iω0 t 1 1
ρ000 = + e (e + e−iω0 t ) = + e−γ2 t cos(ω0 t)
2 4 2 2
We see that this probability oscillates with frequency ω0 and decays with rate
γ2 —there is no dependence on γ1 (see Fig. 2). Measurements of the decay of this
probability then yield T2 . To summarize, the full procedure for performing Ramsey
interferometry to measure T2 is to first prepare the qubit in state |0i, rotate by π/2
about y to create the state |+i, let the system evolve for a time t, perform a second
rotation by −π/2 about y, and then measure the probability to be in state |0i. The
decay of this probability over time gives T2 .

3.2 A simple model of dephasing due to quasistatic noise

Where does γφ come from? How do we know dephasing causes an exponential de-
cay of the off-diagonal components of the density matrix? So far we have simply
been assuming this is true. In reality, decoherence does not always cause an expo-

68
 1.0
0.8
probability ρ'00 γ2=0
0.6
γ2=0.03ω0
0.4 γ2=0.1ω0
0.2 γ2=0.3ω0
0.0 γ2=ω0
0 5 10 15 20
time ω0t

Figure 2: Probability to be in the |0i state at the end of a Ramsey interferometry

experiment.

nential decay—it depends on the details of the dephasing mechanism. However, we

can still define the time scales T2 and Tφ even in cases where dephasing is not de-
scribed by a simple exponential decay of the off-diagonal density matrix components.

A common way in which dephasing arises is through slow fluctuations of the qubit
energy splitting caused by interactions with the environment. We can model this
process using much of the analysis we employed to describe Ramsey interferometry
in the previous section. Suppose again that we have an undriven qubit, but this time
there is a small fluctuation  in the transition frequency caused by the environment:
 ω0 + 
− 2 0
H= ω0 +
0 2

Consider again the Ramsey interferometry experiment, but now let’s ignore relax-
ation so that γ1 = 0, and let’s set γ2 = γφ = 0, because now we want to see how 
can give rise to γφ . Using the results from the previous section, the probability to
find the qubit in its ground state is
1 1
ρ000 = + cos[(ω0 + )t]
2 2
Now in order to measure this probability, we need to run the experiment many times
(for a fixed t), measure the state of the qubit each time, and count how many times
we find it in |0i. Suppose that  is constant during each run of the experiment,

69
but it varies a little from one run to the next. To get the actual probability that is
measured, we need to average over all possible values of , weighted by the probability
p() that a given value of  occurs:
Z ∞ Z
0 1 1 ∞
measured probability = dp()ρ00 = + dp() cos[(ω0 + )t]
−∞ 2 2 −∞
 Z Z ∞ 
1 1 iω0 t ∞ it −iω0 t −it
= + e dp()e + e dp()e
2 4 −∞ −∞

We can recognize that the  integrals are the Fourier transform of p(). For simplicity
(and as is often the case in reality) let’s assume p(−) = p(). Then both  integrals
are equal and can be replaced by p̃(t), the Fourier transform of p(). We then have
1 1
measured probability = + p̃(t) cos(ω0 t)
2 2
We see that p̃(t) carries the effect of the qubit energy fluctuations caused by the
environment. If we compare this result to what we found in the previous section,
where we simply assumed that the environment causes exponential decay of off-
diagonal density matrix components with rate γφ , we see that in that case we would
need p̃(t) = e−γφ t to reproduce those results. (Note that here, γ2 = γφ since we are
assuming γ1 = 0.) Performing the inverse Fourier transform of this function, we see
that this corresponds to a Lorentzian energy fluctuation distribution:
γφ 1
p() =
π 2 + γφ2

Thus, we see that if the environment causes energy fluctuations that follow a Lorentzian
distribution, then superpositions will decay exponentially with a rate γφ that depends
on the width of the Lorentzian. Later on when we talk about time-dependent noise,
we’ll see that there is another way to get a simple exponential decay—white noise.
Lindblad terms are actually meant to represent white noise (which has a very dif-
ferent physical origin corresponding to an equal mix of fast and slow fluctuations)
rather than the slow Lorentzian noise considered here. The fact that both types of
noise give the same decay profile is coincidental.

Is a Lorentzian the only possibility for the case of slow noise we are focusing on
here? No! In general different sources of dephasing (e.g., magnetic field fluctuations,
charge fluctuations, ...) will be described by different distributions p(). For example,

70
these fluctuations might be described by a Gaussian distribution (which is quite
common and naturally referred to as Gaussian noise):
1 2 2
p() = √ e− /(2σ )
2πσ
But the Fourier transform of a Gaussian is a Gaussian:
2 2 /2
p̃(t) = e−σ t

and so the decay in the measured probability from the Ramsey interferometry ex-
periment does not look like a simple exponential decay. However, we can still pick
out a characteristic dephasing time:
√
Tφ = 2/σ

which depends on the width of the Gaussian energy fluctuation distribution.

The simple model of dephasing we have just described is called the quasistatic
bath approximation, and the assumption about the slowness of the energy fluctua-
tions is referred to as quasistatic noise. The model can be generalized to include
multiple, slowly fluctuating parameters in the Hamiltonian. Quasistatic noise de-
scribes the majority of the noise haunting today’s most popular solid-state qubits,
including superconducting circuits and semiconductor spin qubits. However, faster
fluctuations in Hamiltonian parameters also matter in quantum computing—we’ll
come back to this later in the section on time-dependent noise. It’s also worth men-
tioning that people often use the symbol T2∗ to refer to the dephasing time associated
with quasistatic noise. A little later when we talk about spin echo and dynamical
decoupling, we’ll see that this type of noise can be dealt with by applying carefully
designed pulses on the qubit. So you can also think of the ∗ as meaning “treatable
with dynamical decoupling”—it’s not as bad dephasing caused by time-dependent
noise.

3.3 Coherence function

Let’s return to the simple model of dephasing from above, where dephasing comes
from energy fluctuations  caused by the environment:
 ω0 + 
− 2 0
H= ω0 +
0 2

71
We saw that in a Ramsey interferometry experiment, the probability to find the
qubit in the state |0i in a single run of the experiment is
1 1
ρ000 = + cos[(ω0 + )t]
2 2
assuming that  is constant over one run. However, when we average over many runs
of the experiment, we have to take into account that  can change from one run to
the next by averaging over  with respect to some probability distribution p() that
depends on the noise source. This gives rise to dephasing.
It is possible to study dephasing by looking directly at the off-diagonal compo-
nent of the density matrix, ρ01 , before the second Ramsey rotation, instead of looking
at ρ000 . On the one hand, this makes√sense because if ρ describes a pure superposi-
tion state like |ψi = (|0i + eiφ |1i)/ 2, then ρ01 keeps track of the phase φ in this
superposition:
 
1 1 eiφ
ρ = |ψi hψ| = ⇒ ρ01 = eiφ /2
2 e−iφ 1

Dephasing is all about how information about φ gets lost to the environment, so it
seems like dephasing studies should focus on ρ01 . However, when we introduced the
simple dephasing model above, we focused on ρ000 because it is a physical observable—
it is the probability to find the qubit in state |0i. The whole point of the Ramsey
rotations is to convert the phase information in ρ01 into something we can measure,
namely ρ000 . Furthermore, because it’s an observable, it makes sense to average ρ000
over .
Although ρ01 is not a probability, there is still a sense in which we can measure
it. This is sort of what we’re doing in the Ramsey experiment. After all, we found
that
1 1
ρ000 = (1 + ρ01 + ρ10 ) = + Re[ρ01 ]
2 2
so it seems like Re[ρ01 ] is measurable, and this is actually the quantity we average
over  to get the dephased Ramsey signal. We could also do a slightly different
experiment in which we replace the second rotation (the one by angle −π/2 about
the y axis) by a −π/2 rotation about x instead. This would then give a slightly
different result for the 00 component of the final density matrix:
1 1
ρ0000 = [1 − i(ρ01 − ρ10 )] = + Im[ρ01 ]
2 2

72
By combining results from these two experiments, we effectively get a measurement
of ρ01 (the ordinary Ramsey experiment gives the real part of ρ01 , while this modified
Ramsey gives the imaginary part). In this sense, we can think of ρ01 as a physical
observable, but measuring this quantity is more complicated. If all we want is to
measure T2 , it’s easier to just measure ρ000 in the ordinary Ramsey experiment like
we described above. However, some things are easier to understand by looking at ρ01
directly, as we’ll see. People call the average of ρ01 over noise the coherence function
W (t):
|hhρ01 (t)ii|
W (t) =
|hhρ01 (0)ii|
The double angular brackets here are not Dirac notation. They represent an average
over noise. For our simple model of dephasing from above, we have
Z ∞ Z ∞
1
hhρ01 (t)ii = dp()ρ01 (t) = dp() ei(ω0 +)t
−∞ −∞ 2

If p() is a Lorenztian, this becomes

Z ∞
γφ 1 1 i(ω0 +)t 1 iω0 t −γφ t
hhρ01 (t)ii = d e = e e
−∞ π 2 + γφ2 2 2

The coherence function is then

W (t) = e−γφ t

which precisely captures the loss of quantum coherence we expect to see due to
dephasing. In general, the closer W (t) stays to 1, the more the qubit is retaining
the information stored on it. Note that since ρ01 = Sx + iSy , the noise-averaged
components of the spin vector in the xy plane also provide a measure of how coherent
the qubit state is: p
hhSx (t)ii2 + hhSy (t)ii2
W (t) = p
hhSx (0)ii2 + hhSy (0)ii2

3.4 Spin echo and dynamical decoupling

In the case of slow noise fluctuations like in our simple model of dephasing, there is
something simple we can do to slow down decoherence. Again suppose we start in
the state ρ(0) = |+i h+|. Since H = −(ω0 + )σz /2 is time-independent, the state at

73
later times is  
−iHt iHt 1 1 ei(ω0 +)t
ρ(t) = e ρ(0)e =
2 e−i(ω0 +)t 1
Suppose that at time t = τ we quickly rotate the qubit by π about the x axis. The
state immediately after this rotation is
   
−i π2 σx 1 1 ei(ω0 +)τ i π2 σx 1 1 e−i(ω0 +)τ
ρ(τ ) = e e =
2 e−i(ω0 +)τ 1 2 ei(ω0 +)τ 1

Notice what happened to the signs in the off-diagonal phases! They effectively
switched. This has important repercussions as we’re about to see. If we let the
state continue to evolve under our simple dephasing Hamiltonian, we find
 
−iH(t−τ ) iH(t−τ ) 1 1 ei(ω0 +)(t−2τ )
ρ(t > τ ) = e ρ(τ )e =
2 e−i(ω0 +)(t−2τ ) 1

Surprisingly, we see that at time t = 2τ , the qubit ends up back in its initial state,
no matter what  is. It follows immediately that

W (t = 2τ ) = 1

no matter what p() is. The qubit’s coherence is completely restored at time t = 2τ !
This is easy to understand from the point of view of the spin vector representation.
See Fig. 3.
With a single pulse, we can undo the damage done by dephasing, provided the
noise parameter  is constant during each run of the experiment. Of course, an
infinitely fast π pulse is not physically realizable, but if the pulse is fast compared to
the precession frequency ω0 + , this is often good enough. There also exist schemes
based on slower, shaped pulses that are more experimentally feasible, but designing
these requires dealing with the time-dependent Schrödinger equation, and the way
in which  cancels out of the final state is more complicated in these cases. Also, this
cancellation only occurs up to the first few orders in  (and not to all orders like for
the spin echo δ-function pulse), and so these methods only work if  is sufficiently
small. This trick of using pulses to make noise errors cancel out so that coherence is
restored is called dynamical decoupling. We are using pulses to effectively decouple
the qubit from its environment.

74
 y

s(𝜏-)

ϕ
x
ϕ

s(𝜏+)
Figure 3: Hahn spin echo. The spin vector starts at S(0) ~ = x̂ and precesses coun-
terclockwise in the xy plane by angle φ = (ω0 + )τ . At time t = τ a π rotation
about x is performed, flipping the spin vector about the x axis from S(τ ~ − ) to S(τ
~ + ).
Afterward the spin vector precesses for an additional timespan τ , at which point it
returns to its initial orientation along x̂. Thus, the spin vector returns to its starting
point no matter what  is.

75
3.5 Time-dependent noise
What happens if the noise fluctuation  is not constant during each run of the
experiment but instead has some nontrivial time-dependence? In this case, we have
to consider a time-dependent dephasing Hamiltonian:

H(t) = −[ω0 + (t)]σz /2

The good news is that this Hamiltonian commutes with itself at different times,
[H(t1 ), H(t2 )] = 0, so we can still obtain the evolution operator analytically:
Rt i
Rt
dt0 H(t0 ) dt0 (t0 )]σz
u(t) = e−i 0 = e− 2 [ω0 t+ 0

We can also write down the state at later times, again assuming ρ(0) = |+i h+|.
We’re interested in the coherence function W (t), so let’s just look at the off-diagonal
component of ρ(t) = u(t)ρ(0)u† (t):
1 Rt 0 0
ρ01 (t) = ei[ω0 t+ 0 dt (t )]
2
To compute W (t), we need to average this over some noise distribution. What
does such a distribution look like when (t) is time-dependent? As in the case of
quasistatic noise, this is determined by what the noise source is. The type of noise
can be quantified by noise correlation functions. For example, we can consider the
two-point correlation function:

S(t, t0 ) = hh(t)(t0 )ii

This function provides information about how the strength of the fluctuation  at
time t0 is related to that at time t (i.e., if the fluctuation is big at time t0 , is it more
likely to also be big at time t, or are the fluctuation strengths at different times
completely independent of each other?). For a completely general noise source, one
needs all of the correlation functions (two-point, three-point, four-point, etc.) to fully
characterize the noise. It is of course generally not practical to measure infinitely
many correlation functions though. Fortunately, there exists a large class of noise
sources called Gaussian noise sources that are (i) fully characterized by the two-
point correlation function only and (ii) do a good job of describing noise sources
that afflict a number of popular qubits, including semiconductor spin qubits and
superconducting qubits. It is also usually the case that the two-point correlation

76
function only depends on the time difference:

S(t, t0 ) → S(t − t0 )

We can think of this condition as saying that the noise source obeys some sort of
time-translation symmetry: the correlations do not depend on any absolute time,
only on time differences. This condition could be violated if the noise source itself
were evolving over time, but this is not something people normally worry about in
qubit systems. It is also common practice to convert S(t) to frequency space:
Z ∞
S(ω) = dteiωt S(t)
−∞

Here, S(ω) is the Fourier transform of S(t), and it is called the noise power spectral
density or noise power spectrum. You can think of it as a function that tells you how
strong the noise fluctuations are at each frequency ω. For example, if the noise is
slow like in the quasistatic noise model of dephasing we discussed earlier, then S(ω)
is large near ω = 0 and small away from zero. S(ω) is often more convenient to work
with than S(t), as we’ll see.

OK, suppose we know we have a source of dephasing noise that has two-point
correlator S(t) and power spectrum S(ω). How can we predict what W (t) will look
like and what Tφ (or equivalently T2 ) will be? This requires averaging ρ01 over all
possible functions (t) that are consistent with the given S(ω). This averaging thus
involves a functional integration. In the case of Gaussian noise, we use a Gaussian
functional integral to average a physical observable A((t)):
Z
1 1 ∞
R R∞ −1
hhA((t))ii = D e− 2 −∞ dt1 −∞ dt2 (t1 )S (t1 ,t2 )(t2 ) A((t))
N

Here, S −1 (t1 , t2 ) is the functional inverse of S(t):

Z ∞
dt3 S −1 (t1 , t3 )S(t3 − t2 ) = δ(t1 − t2 )
−∞

and N is the normalization factor for the distribution:

Z
1 ∞
R R∞ −1
N = D e− 2 −∞ dt1 −∞ dt2 (t1 )S (t1 ,t2 )(t2 )

In general, functional integrals are nasty things, and many mathematicians don’t

77
consider them well defined objects. However, in the case of a Gaussian functional
integral like this one, we can define it as the continuum limit of a mathematically
rigorous multivariate integral where  can be thought of as a vector of integration
variables, and S −1 is a matrix acting on this vector. This integral is a simple general-
ization of the usual Gaussian integral (see the Wikipedia page for Gaussian integral
for details). To compute W (t), we just need the following result:
R∞ 1
R∞ R∞ 1
R∞
dωS(ω)|g̃(ω)|2
hhei −∞ dt(t)g(t)
ii = e− 2 −∞ dt1 −∞ dt2 g(t1 )S(t1 −t2 )g(t2 )
= e− 4π −∞

We can think of the middle expression as theR functional analog of “completing the
1 2 2
square” to compute a Gaussian integral like dxe− 2S x +igx ∼ e−Sg /2 . In the final
expression, we used Parseval’s theorem to rewrite the integral in terms of Fourier
transforms.

Coming back to our dephasing problem, we have

Rt Rt
dt0 (t0 )] dt0 (t0 )
W (t) = |hhei[ω0 t+ 0 ii| = |hhei 0 ii|

Comparing this to the general formula above, we see that in this case

g(t0 ) = Θ(t − t0 )Θ(t0 )

where Θ(t) is the step function (which is 1 for t > 0 and 0 otherwise). The Fourier
transform of this is
Z ∞ Z t
0 iωt0 0 0 0 1
g̃(ω) = dt e Θ(t − t )Θ(t ) = dt0 eiωt = (eiωt − 1)
−∞ 0 iω

and so
4 sin2 (ωt/2)
|g̃(ω)|2 =
ω2
The coherence function is then
 Z 
1 ∞ sin2 (ωt/2)
W (t) = exp − dωS(ω)
π −∞ ω2

If we know the noise power spectrum S(ω), then we can use this formula to predict
how the coherence decays as a function of time t.

Consider again the case of quasistatic noise, where the noise is very slow. This

78
means that all the noise is concentrated at ω = 0. We can describe this case by
choosing
S(ω) = αδ(ω)
for some constant α. We can then easily do the integral in the coherence function to
find  
1 αt2
W (t) = exp −
π 4
This is the same sort of Gaussian decay like we saw for Gaussian quasistatic noise,
which seems consistent. To reproduce the exact same decay profile that we found
for this case earlier, we need to set

α = 2πσ2
2 2
1
Recall that σ is the width of the Gaussian distribution p() = √2πσ e− /(2σ ) that

we considered when we first discussed quasistatic noise. In fact, the integrand in
the Gaussian functional integral reduces to p() when S(ω) = 8πσ2 δ(ω). Taking the
inverse Fourier transform of this, we have
Z ∞
1
S(t) = dωe−iωt S(ω) = σ2
2π −∞

The functional inverse of this is a product of two Dirac delta functions:

1
S −1 (t1 , t2 ) = δ(t1 )δ(t1 − t2 )
σ2

Plugging this into the functional integral integrand, we get

1
R∞ R∞
dt2 (t1 )S −1 (t1 ,t2 )(t2 ) 2 /(2σ 2 )
e− 2 −∞ dt1 −∞ = e− 

where we replaced (0) → . The functional integral thus reduces to precisely the
same ordinary integral over  we considered earlier when we first talked about qua-
sistatic Gaussian noise.

Another important case is when the noise is constant across all frequencies:
S(ω) = β = const. This is called white noise. The coherence function in this
case is  Z   
β ∞ sin2 (ωt/2) βt
W (t) = exp − dω = exp −
π −∞ ω2 2

79
This is a simple exponential decay. In fact, if we set β = 2γφ , then we end up
with precisely the same coherence function we found for quasistatic noise with a
Lorenztian distribution. However, here the physics is completely different. Instead
of being constant like in the quasistatic noise case, here the two-point correlation
function is a delta function:

S(t − t0 ) ∼ δ(t − t0 )

People call the noise source or bath Markovian in this case, because the noise fluc-
tuations do not retain any “memory” of past fluctuations. In some sense this is
the opposite of quasistatic noise. And unlike quasistatic noise, white noise is not
amenable to dynamical decoupling. It is actually white, Markovian noise that peo-
ple usually have in mind when they introduce Lindblad terms to describe dephasing.
This is why these terms produce a simple exponential decay of the decoherence.

3.6 Filter functions and multi-pulse dynamical decoupling

How well does dynamical decoupling work for time-dependent noise? This depends
on the noise power spectrum and on the number, type, and timing of the pulses
we apply. Understanding the effect of a particular pulse sequence on a given noise
spectrum is facilitated by using a function called the filter function. We can also
exploit filter functions to indirectly measure a noise power spectrum through its
effect on the qubit.
We again start with the time-dependent dephasing Hamiltonian H(t) = −[ω0 +
(t)]σz /2 and consider the initial state |ψ(0)i
R t 0= 0|+i. The evolution operator is
u(t) = exp{iφ(t)σz /2}, where φ(t) ≡ ω0 t + 0 dt (t ). The state at later times is

1  
|ψ(t)i = u(t) |ψ(0)i = √ eiφ(t)/2 |0i + e−iφ(t)/2 |1i
2
Applying a π about x effectively swaps the two coefficients and multiplies the whole
state by −i:
i  
Rx (π) |ψ(t)i = − √ e−iφ(t)/2 |0i + eiφ(t)/2 |1i
2
because Rx (π) = −iσx . Suppose we apply this rotation at time t = τ1 . The state at
later times t = τ1 + τ is then
i  
|ψ(t > τ1 )i = u(τ )Rx (π)u(τ1 ) |+i = − √ ei[φ(τ )−φ(τ1 )]/2 |0i + e−i[φ(τ )−φ(τ1 )]/2 |1i
2

80
Suppose we apply a second π pulse at time t = τ1 + τ2 . After the second pulse, the
state at time t = τ1 + τ2 + τ is

|ψ(t > τ1 + τ2 )i = u(τ )Rx (π)u(τ2 )Rx (π)u(τ1 ) |+i

1  
= − √ ei[φ(τ )−φ(τ2 )+φ(τ1 )]/2 |0i + e−i[φ(τ )−φ(τ2 )+φ(τ1 )]/2 |1i
2
The τk are the intervals between pulses. We’ll reserve Latin symbols like t, t0 for
absolute times. The pattern continues, so that after n pulses, the state is

(−i)n  iξ(t)/2 
|ψ(t)i = √ e |0i + e−iξ(t)/2 |1i
2
P
where t = τ + n τn and

ξ(t) = φ(τ ) − φ(τn ) + φ(τn−1 ) − . . . + (−1)n φ(τ1 )

Z t
= dt0 [ω0 + (t0 )]f (t0 )
0

Here, f (t0 ) = ±1 is a piecewise-constant function that switches back and forth be-
tweenP+1 and -1 at each time a pulse occurs. Let’s call these times tk . Clearly
tk = km=1 τm . In terms of the step function, we can express f (t0 ) as

X
n
f (t0 ) = (−1)k Θ(tk+1 − t0 )Θ(t0 − tk )
k=0

with tn+1 ≡ t and t0 ≡ 0. Notice that for a given value of t0 , only one term in the
sum is nonzero—the one corresponding to the value of k for which tk < t0 < tk+1 .
The value of f (t0 ) in this range is (−1)k .
Since ρ(t) = |ψ(t)i hψ(t)|, we see that
1
ρ01 (t) = eiξ(t)
2
and so the coherence function is
  Z t 
|hhρ01 (t)ii| 0 0 0
W (t) = = |hheiξ(t) ii| = exp i dt (t )f (t )
|hhρ01 (0)ii| 0

In the last step, we discarded the terms in ξ(t) involving ω0 since these just contribute

81
an overall phase that can be brought outside the noise average and then eliminated
by the modulus. If we assume that the noise is Gaussian, we can then use our
Gaussian functional averaging identity where now the function g is given by g(t0 ) =
f (t0 )Θ(t − t0 )Θ(t0 ). We then just need to compute the Fourier transform, g̃(t, ω), of
this function and plug the result into
 Z ∞ 
1 2
W (t) = exp − dωS(ω)|g̃(t, ω)|
4π −∞

The function F (t, ω) = |g̃(t, ω)|2 is called a filter function (although sometimes people
2
define the filter function to be F (t, ω) = ω2 |g̃(t, ω)|2 —this is a matter of convention).
F (t, ω) carries information about the pulse sequence. The goal of dynamical decou-
pling in the case of time-dependent noise is to choose a sequence such that F (t, ω)
is small whenever S(ω) is large so that the integral in W (t) is suppressed and W (t)
remains close to 1.

If we don’t apply any pulses at all, then f (t0 ) = 1 at all times, and we return to
the case we already considered earlier with g(t0 ) = Θ(t − t0 )Θ(t0 ). In this case, we
found that the filter function was F (t, ω) = 4 sin2 (ωt/2)/ω 2 . We saw that both for
quasistatic noise and white noise, W (t) quickly decays to zero (like a Gaussian in
the former case, and exponentially in the latter).
What happens when we apply a spin echo pulse in the presence of time-dependent
noise? In spin echo, there is a single π pulse at time t = t1 = τ1 , so f (t0 ) is

f (t0 ) = Θ(t1 − t0 )Θ(t0 ) − Θ(t − t0 )Θ(t0 − t1 )

and therefore g(t0 ) is

g(t0 ) = f (t0 )Θ(t − t0 )Θ(t0 ) = Θ(t − t0 )Θ(t1 − t0 )Θ(t0 ) − Θ(t − t0 )Θ(t0 − t1 )Θ(t0 )
= Θ(t1 − t0 )Θ(t0 ) − Θ(t − t0 )Θ(t0 − t1 )

In the second line, we eliminated some redundant step functions. The Fourier trans-

82
form of this is
Z ∞
0
g̃(t, ω) = dt0 eiωt g(t0 )
−∞
Z t1 Z t
0 iωt0 0
= dt e − dt0 eiωt
0 t1
1 1
= (eiωt1 − 1) − (eiωt − eiωt1 )
iω iω
With some algebra, it is straightforward to show that
2
|g̃(t, ω)|2 = [3 + cos(ωt) − 2 cos(ω(t − t1 )) − 2 cos(ωt1 )]
ω2
If we follow the spin echo recipe of setting t = 2t1 (so that the pulse is applied
halfway through the evolution), this reduces to
16 4
|g̃(2t1 , ω)|2 = sin (ωt1 /2)
ω2
Let’s consider the same two types of noise from before: quasistatic noise and white
noise. For quasistatic noise, S(ω) ∼ δ(ω), so the exponent in the coherence function
depends on the integral
Z ∞ Z ∞
16 4 16
dωS(ω) 2 sin (ωt1 /2) ∼ dωδ(ω) 2 sin4 (ωt1 /2) = 0
−∞ ω −∞ ω

Therefore, W (2t1 ) = 1, i.e., the coherence is completely restored at time t = 2t1 .

This is exactly what we found from the quasistatic noise model.
What about white noise? Now we have S(ω) = 2γφ is constant, and so we have
 Z 
γφ ∞ 2
W (2t1 ) = exp − dω|g̃(2t1 , ω)|
4π −∞
 Z 
2γφ ∞ 16 4
= exp − dω sin (ωt1 /2) = e−2γφ t1
4π −∞ ω 2

This is precisely the decay we found without the spin echo pulse. We see that spin
echo does nothing against white noise. In fact, no matter what pulse sequence we
apply, the dephasing due to white noise is unaffected. You’ll prove this in the next
homework.
For a general noise power spectrum S(ω), spin echo and other dynamical decou-

83
 only the low-frequency
and the decay
4 of W (t
1
SE noise spectrum. For
PDD ation it can be shown
(a) SE 0.8 (a) a comparison of more kinds of DDCPMG
techniques applied to
CDD tion (although slightly
the spin bath problem). It is defined
UDD by δk =k/(n + 1).
Eq. (18) gives
0.6
n=5 Although it looks very similar to CPMG (the difference

W(t)
PDD
0.4
being only a small offset of the initial and final delay
χ(t)
CPMG times), below we will show that CPMG visibly outper-
0.2 forms PDD when considering the realistic small n. The
CDD key difference is that while in the limit of very HerefastCap-α is a sequenc
0 plication of pulses both PDD and CPMG decouple one. This theresult can b
UDD
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 thetimefollowing
qubit from the environment (during the whole of way. Its
the evolution), ! t/(2n+2)
c for realistic small n the CPMG sequence
a periodic train of nar
is much better at refocusing the coherence atWe theapproximate
final these
CDD n=10 0 ∆z, which can be der
time t = t.
Figure 4: Effect of multi-pulse dynamical decoupling sequences on S(ω) ∼ 1/ω noise. (Eq. (20)). Then, to t
(b) Recently, a new family of DD protocols involving con-
0
10
Left:UDD Temporal spacing of Rx (π) pulses for several popular dynamical decoupling ∆z/z p = 1/n we get fo
catenating (recursively embedding the sequences within

"(t)=−ln W(t)
17
sequences. SE stands for spin echo. t=t PDD −5is “periodic themselves) has been proposed.
dynamical decoupling”, For the UDD purpose of com-
t=0
` ` 10
(b)is “Uhrig dynamical decoupling”, CDD is “concatenated dynamical decoupling”. See on con- Cα = π 2 (2
bating the pure dephasing, we will concentrate
catenations of the echo sequence, and for simplicity we
f(t;t’)
Phys. Rev. B 77, 174509
CPMG n=2 (2008) for more10details. −10 will Right:
refer to itThe coherence
as simply CDD. The function for
CDD sequence at l-th
1 which
1/ω
f(t;t’<0)=0 noise when various dynamical decoupling
f(t;t’>t)=0 sequences
order of are
concatenation applied.and CPMG
for total gives
evolution time for t isα = 1 gives
defined as CDD (t). CDD (t) is free evolution is for
the time
Riemann zeta f
the best performancet’=t/4 for this type oft’=tnoise.
t’=3t/4 t’
l 0
UDD one can use a s
t. 10 The
−1 l-th order of concatenation 10
0 is then recursively
-1 ! t/(2n+2) FnUDD (z) = an Θ(z − (2n
defined by c motivated by the fact
FIG.pling 1: (a)sequences
The illustrationwork betterpulse
of various when the noise
sequences with is more concentrated t near ω = 0. The
    looks similar to a ran
t
application times of π pulses
most prevalent types marked. Spin in
of noise echo (SE) is shown
solid-state qubit CDD
systems l (t) ≡haveCDDl−1 a power −spectrum
π − CDDl−1that − sumπ , of(15)periodic funct
along with the PDD, CPMG, CDD and UDD sequences with FIG. 3: (Color online) The dependence 2 of (a) W (t) and (b) 2 riods (see Eqs. (16) a
n = scales
5 pulses like the inverse
(for CDD of the tofrequency:
this corresponds χ(t)=−
the 3rd order ofln W∼
S(ω) (t) 1/ω.
for SE and This is commonly
higher-order referred
(n = 5) sequences for to a = π(n + 1), and t
(4th order of
1/f noise with so A that,
0 /ωc = CDD1. UDD1 (t) is andtheCPMGSE with give W a (t)
π pulse
≈1 atnt/2 and
as 1/f noise.
concatenation). This
10 pulse UDDnoiseand is
CDDin between quasistatic
for
con-
t < (2n + 2)/ω and. white
Echo signal noise;
(SE) most
(equivalent
CDDc2 (t) is the same as n = 2 CPMG sequence.ofFor
of to the
PDD noise
or
with results of numeric
catenation) are also shown. (b) Function f (t; t0 ) defined in thel filter.
>2 The an
Eq. is concentrated at sequence.
low frequencies, but there isthea long tail at forhigh thefrequencies. Thesequences
CPMG with n=1) is also
concatenationsshown ofcomparison.
echo give us new of + 1)2α+1 ]
(14) for 2 pulse CPMG Cα = π 2 /[(α
effect of commonly used dynamical decouplingnon-trivially sequences spaced on 1/fπ noise pulses, is theillustrated
performanceinofthe which
denominator in
has been investigated theoretically in the case and of the fromnu-numerical c
in Fig. 4.
specific infra-red clear cutoff
spin bath. ωir . 31,32,33
However, Note that the
in typical numberracy,
experi- of pulses
we find that CP
than 1/t (the quasi-static shifts of the qubit precession mental situations for theWl-th FID
decaysofmuch
order faster thanisWnSE
concatenation ≈.2l . other
The CDD sequences (more
frequency). This, of course, gives the complete For cancel-a given n, PDD is clearly the least17 effective approach
sequences were argued to be more tolerant ittois imple- enough to implem
lation of the static randomness in qubit splittings at Ω allintimes.mentation
4 Entanglement
the measurement of an ensemble of qubits (inhomoge- UDD is orders
As expected,
of magnitude
errorsfor and short
moretimes
betternumber
than theofother
t  2n/ω
efficient c,
(in terms of perfor-
qubit
pulsethan the PDD
coherence in thi
mance for the same pulses)
neous broadening). In the considered here case of sequences, the SC see Fig. 3b. However,
Thus, UDD is the ideal sequence
sequence. the theoretical comparisons17 with
qubits,4.1 Separable versus entangled
one always deals with repeated measurements of two-qubit
for maintaining othera DD low protocols
level of states
decoherence,
were done in i.e.the high fi-
quantum mechani-
C. Noise spectrosc
a single qubit (a time ensemble), and SE is a very efficient delity, whichcal is setting,
a necessary condition for quantum error
using the Magnus expansion of the evolution
technique whicha suppresses thestate
low-frequency partcorrection.
of the However,
Consider two qubit operators,if or theso-called
goal is simply“average to increase the
Hamiltonian” theory.
The time 13
dependenc
1/f noise, leading to a substantial increase in the Tcharacteristic 2 time decoherence time T2 , defined by χ(T2 )=1,
5,11,23 20,21 CDD is designed to cancel, with each order ofnalconcate- pulse sequences ca
in superconducting charge and flux qubits. then for a given n, the CPMG sequence is the best strat-
nation, successive orders of the qubit-bath interaction for extractingin the noise
The multiple-pulse extension of the echo is theegy. Carr- the Magnus expansion. It also cancels ing. orders
successive The idea of using
Purcell-Meiboom-Gill (CPMG) sequence, |Ψ12s idefined
= |ψIt1 iby is⊗interesting
|ψ2ofi .intra-bath
to note interaction.
that the CDD 32 sequence does (4.1) a priori was intro
However, it is notspectroscopy
δk=(k −1/2)/n. This sequence leads to periodic echo-like not offer strong advantages compared
clear whether the advantages of CDD are going to CPMG and been torealized
also experime
revivals This of thestate
coherence.
is calledWhile it was used or
a separable forUDD.manyAt short
product state. times itsuch
hold for For
gives smaller the
the case of states, dephasing due
χ(t) qubits
comparedare titative method for ex
to classical noise.
years in NMR experiments performed on ensembles to CPMG, of but the difference is not as dramatic as in contributing
independent, and any measurements we make The most recent development in suppressing the pure to pure d
spins, it was recently predicted that it should bethe case of on
highly UDD. oneAt qubit
longer will times,not CPMG affect the
is better other.
and than thebynoise leading
effective at prolonging the coherence of a spin qubit givesin- larger dephasing seemswas
T2 . It 18,51 thatthe theintroduction
benefits of using of aCDD new sequence
ticular, for t<1/ωc w
expected Uhrig,
the regime which we term here UDD. This are sequence was
teracting with the nuclear bath.30 Whereas CPMG is inoptimized for
of quantum
pure
bath
dephasing
dynamics
due to a bosonic Eq. (24)), where Mk =
environ-
largely lost when dealing with classical noise. of the spectral density
best understood as a refocusing sequence, in recent years 84 ment or classical Gaussian noise, but later its surprising
a lot of attention has been devoted to the idea ofA dy- different situation arises when the cutoff cannot be
universal character was discovered in Then,
in principle, reconstru
a general two-pulse
quantum-UDD (equiv
namical decoupling of the qubits from its environment by reached, i.e. we cannot apply pulses
19 fast enough.
mechanical setting. UDD is defined by
pulses.14,15,16,17 In this approach the pulses are used to
average out the influence of the environment on the qubit, δk = sin2 [πk/(2n + 2)] , (16)
which can be fully achieved only in the limit of very fast
repetition of pulses. Out of many types of dynamical and in the next Section we will explain in what sense it is
decoupling (DD) protocols, we concentrate here on de- “optimal” for the case of the Gaussian noise. Originally
terministic periodic (PDD) sequence (see e.g. Ref. 33 for the sequence was applied18 to the case of the environment
For example, imagine that we measure the first qubit in the computational basis and
we find the post-measurement state to be |0i. Then we would have

|Ψs i → |0i h0| |ψ1 i ⊗ |ψ2 i , (4.2)

and after normalization the post-measurement state would be

|Ψ0 i = |0i ⊗ |ψ2 i . (4.3)

If the first qubit was instead found in state |1i, the post measurement state of the
two qubits would be

|Ψ0 i = |1i ⊗ |ψ2 i . (4.4)

In either case, the second qubit has not changed its state. This is true no matter
what basis we measure the first qubit in.

Now consider a different two-qubit state

√
1 2
|Ψe i = √ |0i ⊗ |0i + √ |1i ⊗ |1i . (4.5)
3 3
If we now make a measurement on the first qubit, with probability 1/3:

|Ψe i → |0i ⊗ |0i , (4.6)

and with probability 2/3:

|Ψe i → |1i ⊗ |1i . (4.7)

Clearly, this state is qualitatively different than the first type of state, which was
separable. For this state, the post-measurement state of the second qubit depends
on the measurement outcome of the first qubit. Any pure state such as this, that
cannot be written as a tensor product of single-qubit states, is called entangled. This
type of correlation is purely quantum (has no classical analog). Another point to
make here is that the post-measurement state is separable.

A note on how we more generally find the post-measurement state: Knowing

the measurement outcome for the measured qubit, we apply to the whole state a
projector onto that measurement outcome tensored with an identity on the remaining

85
qubits (which was implicit in the above). Let’s look into a less trivial example to
demonstrate this. Consider again the state from Eq. (4.5) and assume that now we
make a measurement in the |±i basis. If the outcome we find is |+i, then

|Ψe i → (|+i h+| ⊗ 1) |Ψe i

√ !
1 2
= (|+i h+| ⊗ 1) √ |0i ⊗ |0i + √ |1i ⊗ |1i
3 3
√
1 2
= √ |+i h+| |0i ⊗ |0i + √ |+i h+| |1i ⊗ |1i . (4.8)
3 3
√
The inner products appearing in Eq. (4.8) are h+| |0i = h+| |1i = 1/ 2. Replac-
ing we get
√ !
1 2
|Ψe i → |+i ⊗ √ |0i + √ |1i . (4.9)
6 6

Notice that the state is not normalized (which is expected, since a projector is not
a unitary operator), so we have to normalize it. Doing that, we find the post-
measurement state
√ !
1 2
|Ψ0 i = |+i ⊗ √ |0i + √ |1i . (4.10)
3 3

Again, we find that the post-measurement state of the second qubit is different from
what it was before the measurement and different from the post-measurement state
for other basis measurement choices and other outcomes.

We can therefore say that qubits in entangled states do not have well defined
properties. The global system does, but not the constituents. What is impressive and
counterintuitive is that these quantum correlations persist even for two qubits that
are separated by an arbitrary distance. This effect of a measurement on a system
apparently affecting another system that is separated by a superluminal distance
is what unsettled Einstein and what he termed a ‘spooky action at a distance’.
Besides its fundamental importance in terms of how nature behaves at the quantum
level, entanglement is a crucial resource for quantum computing and more generally
quantum technologies. As we will see below (for the case of two qubits), we can
quantify entanglement. Four maximally entangled states of two qubits that hold a

86
special place in quantum information science are
1
Φ+ = √ (|00i + |11i) (4.11)
2
1
Φ− = √ (|00i − |11i) (4.12)
2
1
Ψ+ = √ (|01i + |10i) (4.13)
2
1
Ψ− = √ (|01i − |10i) . (4.14)
2
These states, which form an orthonormal basis, are called Bell states. The first of
these states, |Φ+ i is often used as a unit of entanglement.

Note that entanglement cannot increase by local operations on the qubits. How-
ever, it can certainly decrease that way. For example, if we make a projective mea-
surement on one of the qubits of a two-qubit system, then the post-measurement
state is a product state, which means that any entanglement initially present has
been destroyed.

4.2 Reduced density matrix and von Neumann entropy

Next we address two questions: Given that a qubit in an entangled state does not
have well-defined properties, how would we describe it mathematically if we had to
make predictions about measurements of it? For example, imagine that we have a
qubit that is part of an entangled state and we want to make a prediction about the
mean value of a certain operator if we were to measure that operator. For example,
let’s look at the state |Ψe i from above and assume we are interested in the mean
value of operator Z on the first qubit. Let’s calculate this:
1
hΨe | Z ⊗ 1 |Ψe i = − . (4.15)
3
We can also do the same for X and Y :

hΨe | X ⊗ 1 |Ψe i = 0 (4.16)

hΨe | Y ⊗ 1 |Ψe i = 0. (4.17)

Remember that these three numbers together form the spin vector in the Bloch

87
sphere. Since we find that the length of this vector is < 1 for this example, that
means that if we only had access to one of the qubits, any measurements we made
would be consistent with it being described by a density matrix for a mixed state.
Below we see how to obtain this density operator.

4.2.1 Partial trace

We can calculate the density matrix of a subsystem by tracing out the other qubit
(or, in the case of a multi-qubit system, all the qubit(s) that we are not interested in
or have no access to). This operation is called a partial trace. In the example of two
qubits, A and B, we would find the reduced density operator of qubit A by tracing
out qubit B. The procedure is as follows:
• Consider a pure state of two qubits |ΨAB i
• Create the density operator of the two-qubit system ρ = |ΨAB i hΨAB |
• Trace out qubit B to obtain the density operator describing qubit A, ρA =
TrB |ΨAB i hΨAB |
Remember that the trace is basis-independent, so that we can write, for example,

ρA = TrB ρ =B h0| ρ |0iB +B h1| ρ |1iB .

What kind of mathematical object is this? Notice that because this is a partial trace,
it is an operator, not a scalar.

Let’s work out the example of a Bell state and see what we obtain for the reduced
density operator of qubit A. First, let’s find explicitly the density operator for the
total system AB.
1
ρ= (|01i h01| + |10i h10| − |01i h10| − |10i h01|) . (4.18)
2
When we take the partial trace wrt B, only the first two terms contribute. The
reduced density operator for qubit A is then
1
ρ= (|0i h0| + |1i h1|) , (4.19)
2
which is a maximally mixed state. Notice that in this example if we traced out qubit
A instead we would find the same density operator for qubit B.

88
 Q: What is the reduced density matrix for qubits A and B in the case of the
two-qubit entangled state we saw above |Ψe i?
Answer: For both qubits, the reduced density matrix is 1/3 |0i h0| + 2/3 |1i h1|.

Q: What is the reduced density matrix for qubits A and B in the case of the
two-qubit state |Ψi = |φi ⊗ |χi?
Answer: The reduced density matrix for qubit A is ρA = |φi hφ|, while that of
qubit B is ρB = |χi hχ|.

We notice that when the two-qubit state is entangled, the reduced density oper-
ator of each of the individual qubits is mixed. For maximally entangled states, the
reduced density operator is maximally mixed. On the other extreme, for product
states, the reduced density matrices are pure. Therefore for a pure two-qubit state
the purity of the reduced density operator quantifies the entanglement between the
two qubits.

4.2.2 Von Neumann entropy

A pure state is a state of maximal information, whereas a completely mixed state is
a state of minimal information (complete ignorance, for example in a binary case, is
50-50%). We can define various measures of the purity of a state. We have already
seen that
P = Tr(ρ2 )
gives a measure of the degree of purity of a state. We can therefore use that to
quantify the information contained in the state.
Alternatively, since information is related to entropy, we can define entropy for
a density operator. In analogy to (classical) statistical mechanics and information
theory (Shannon entropy), we can define the von Neumann entropy as

S(ρ) ≡ −Tr(ρ log(ρ)) (4.20)

If we define the eigenvalues of ρ as λj , then we can re-express the entropy as

X
S(ρ) = − λj log λj , (4.21)
j

where we define 0 log 0 ≡ 0.

Q: What is the entropy for a pure state?

89
Answer: Notice that for a pure state, we have all eigenvalues λj = 0 except one, let’s
call it λ0 , which is 1. In that case we would have that S = 0, which is consistent
with thinking of pure states as states of maximal information.

Basic properties of von Neumann entropy

1. The entropy is non-negative, and it is zero iff the state is pure.
2. In a d-dimensional Hilbert space, the entropy is at most log d, and equal to that
iff the state is maximally mixed, i.e., for ρ = d1 1.
3. If a composite system (subparts A and B) is in a pure state then S(A) = S(B).

4.2.3 Schmidt decomposition

For any pure state |Ψi of a composite system AB there exists an orthonormal basis
{|iA i} for system A and an orhonormal basis {|iB i} for system B such that
X
|Ψi = λi |iA i ⊗ |iB i , (4.22)
i

with real λi ≥ 0. The λi ’s are called Schmidt coefficients, and the number of nonzero
λi ’s is called the Schmidt number. The Schmidt number gives us information about
the entanglement of the two subsystems. For example, a Schmidt number of 1
corresponds to a pure state.
The Schmidt decomposition is an important concept in quantum information
and also in condensed matter physics. One immediate consequence P 2 is that the
reduced P density matrices for subsystems A and B are ρA = i λi |iA i hiA | and
2
ρB = i λi |iB i hiB | respectively. From this we can immediately see that the eigen-
values of the reduced density matrices are the same.

4.2.4 Purification
Suppose we are given a state ρA of a quantum system A. It is possible to introduce
another system, which we denote R, and define a pure state |ΨAR i for the joint system
AR such that ρA = TrR |ΨAR i hΨAR |. This is a purely mathematical procedure,
known as purification, which allows us to associate pure states with mixed states. The
system R (also called a reference system) is a fictitious system, without necessarily
a direct physical significance.

90
4.3 Decoherence as entanglement with the environment
The connection between entanglement and the resulting mixed reduced density ma-
trix of a subsystem, as well as the idea of purification, naturally lead to the idea of
decoherence, or loss of purity, originating from the interaction with the environment.
While in previous sections we included environmental effects by introducing stochas-
tic variables in the Hamiltonian, a more exact description (which is often intractable)
is to take into account the quantum mechanical nature of the environment. To make
things simpler we will assume that the system (e.g., a single qubit or a collection
of qubits) and the environment are originally in a product state. However, we do
not need to assume that this state is pure. Indeed, the environment will typically
have many degrees of freedom and will be in a high-entropy state (e.g., in a thermal
state). As a result, it is more reasonable to consider the general case where both the
system and the environment are described each by their own density operator prior
to interacting. In most cases of practical interest, the system will begin in a pure
state and the environment in a mixed state, but for now we’ll keep things general
and write the total initial state as

ρSE = ρS ⊗ ρE . (4.23)

We then allow the system and environment to interact. This interaction is de-
scribed by a unitary operator U , such that the state after the interaction is

ρ0SE = U (ρS ⊗ ρE ) U † . (4.24)

Since we are only interested in the system, and in practice we only have access
to the system (e.g., we can typically only control and measure the system not the
environment), we trace out the environment to find the reduced density operator of
the system:


ρ0S = TrE U (ρS ⊗ ρE ) U † = E(ρS ). (4.25)

Let’s try to express this a bit more compactly. First let’s write the initial state of
the environment as a pure state |e0 i, so that ρE = |e0 i he0 |. Note that this is general
enough, as we could start with a mixed state and then purify that, as described in the
previous section, by introducing a fictitious additional part to the environment. This
would not change the dynamics of the system, since everything labeled environment
would be traced out anyway. We will also use the basis {|ek i} to trace out the

91
environment. Then Eq. (4.25) becomes
X

E(ρS ) = hek | U (ρS ⊗ |e0 i he0 |) U † |ek i
k
X
= Ek ρS Ek† , (4.26)
k

where Ek ≡ hek | U |e0 i is an operator acting on the system. Eq. (4.26) is called
the operator sum representation and the Ek are also known as Kraus operators.
Sometimes this is also called a quantum map. In order to guarantee that the trace
of the density operator is 1, the constraint on the Kraus operators is
X †
Ek Ek = 1. (4.27)
k

Clearly, when there is only one operator we go back to unitary evolution. Also
note that the Ek are generally not unique. For a given non-unitary evolution, there
are typically many possible choices of Kraus operators that can be used to implement
it.

4.3.1 Examples of common operator sum representations for qubits

1. Spontaneous emission
When we solved the Rabi problem we found a unitary evolution opearator that
described the time dynamics. If instead we have a qubit that undergoes spontaneous
emission (decay), then we need Kraus operators to describe its evolution. First we
can think of the case where we don’t explicitly include time as a parameter, but think
about what would happen after some time, which is fixed. Then the two operators
describing spontaneous emission (also called amplitude damping in this context) are
 
1 √ 0
E0 =
0 1−p
 √ 
0 p
E1 = . (4.28)
0 0

Let’s verify that this does what it’s supposed to. When this OSR acts on an
initial state in |0i, the state doesn’t change. This is consistent with what we know
from spontaneous emission. On the other hand, if the initial state is |1i, then the

92
state becomes
 
p 0
0
ρ = E0 |1i h1| E0† + E1 |1i h1| E1† = . (4.29)
0 1−p

If we wanted to make these operators explicitly time dependent and recover our re-
sult from Section 6, we could set p = 1 − e−γt .

2. Dephasing
For pure dephasing of a qubit, the Kraus operators are
 
√ 1 0
E0 = p
0 1
 
p 1 0
E1 = 1−p . (4.30)
0 −1

p is the probability that nothing happens to the qubit. Therefore, p = 1 corresponds

to the limit of no dephasing. The qubit is completely dephased for p = 1/2 and ends
up in a completely mixed state. When p = 0, the qubit undergoes a z rotation and
remains pure.

3. Depolarization
Interactions with an environment often lead to what is called depolarization,
meaning that the qubit ends up in a completely mixed state. If this happens with
probability p, then for an initial density matrix ρ, the final density matrix ρ0 is
 
0 1 1 0
ρ =p + (1 − p)ρ. (4.31)
2 0 1

The first term says that with probability p we end up with a completely mixed
state, while the second term says that with probability 1 − p, nothing happens to ρ.
Eq. (4.31) is not in operator sum form. However, using the identity
 
1 1 0 1
= [ρ + XρX + Y ρY + ZρZ] , (4.32)
2 0 1 4

which holds for any density matrix ρ, we can put it in such a form:
 
0 3p p
ρ = 1− ρ + [XρX + Y ρY + ZρZ] . (4.33)
4 4

93
We see that we can describe this process using four Kraus operators:
r √ √ √
3p p p p
E0 = 1 − I, E1 = X, E2 = Y, E3 = Z. (4.34)
4 2 2 2
The operator sum representation is an extremely useful tool in quantum infor-
mation science. It is widely used in error correction to describe errors.

5 Quantum gates
5.1 Two-qubit gates
To create entangled states of two qubits, we typically allow them to interact either
directly or via some intermediary quantum system. In the next section, we will
discuss physical interactions that can create entanglement. Here, we focus on the
gates that create entanglement before worrying about how we actually implement
these gates. This approach is somewhat different from how we introduced single-
qubit gates. The reason for this choice is that for the majority of qubits, the standard
way of doing single-qubit gates is driving the qubit via a (magnetic or electric) field
so that it undergoes Rabi oscillations. Two-qubit gate implementations are more
varied, so we will first present the form of the relevant gates and later examine how
we would go about realizing them experimentally.

Controlled-NOT (CNOT) gate

The most well-known two-qubit gate which can prepare maximally entangled states
is the CNOT gate (controlled-NOT). The definition of this gate is that one of the
qubits (the target) undergoes a NOT gate (X rotation by π) only if the other qubit
(the control) is in state |1i. This means that under the CNOT, the computational
basis states would transform in the following way:

|00i → |00i
|01i → |01i
|10i → |11i
|11i → |10i

94
so that we would write, in Dirac notation, the following operator

CNOT = |00i h00| + |01i h01| + |10i h11| + |11i h10| (5.1)

and in matrix form:  

1 0 0 0
0 1 0 0
CNOT = 
0

0 0 1
0 0 1 0
We can create a Bell state by first putting the control qubit in an equal superpo-
sition of the two computational states followed by a CNOT gate:
1  
CNOT |+0i = √ CNOT |00i + |10i
2
1
= √ (|00i + |11i) = Φ+ = maximally entangled
2
On the other hand, for if we apply CNOT to other product states, we can also get
back another product state:

CNOT |00i = |00i unentangled

and
1
CNOT |++i = CNOT(|00i + |01i + |10i + |11i) = |++i
2
Therefore, we see that only some product states become entangled. This is generally
true for any entangling operation.

Controlled-Z (CZ) gate

Another very commonly used gate is the controlled-Z or CZ gate. Like the CNOT,
CZ performs an operation on one qubit (the target) only if the other qubit (the
control) is in state |1i. However, now the operation that is peformed is a Z gate
(Z rotation by π). Here is a summary of how the CZ gate acts on each two-qubit

95
computational state:

|00i → |00i
|01i → |01i
|10i → |10i
|11i → − |11i

In terms of Dirac notation and a matrix representation, we can write this as

 
1 0 0 0
0 1 0 0 
CZ = |00ih00| + |01ih01| + |10ih10| − |11ih11| = 
0 0 1 0 


0 0 0 −1

CZ also produces maximally entangled states when it acts on certain product states.
However, notice that it entangles a different set of initial product states compared
to CNOT:
1
CZ |+0i = √ CZ(|10i + |00i) = |+0i no entanglement generated
2
CZ |00i = |00i
1
CZ |++i = CZ(|00i + |01i + |10i + |11i)
2
1
= (|00i + |01i + |10i − |11i)
2
1
= √ (|0+i + |1−i) entangled
2

SWAP
Another important two-qubit gate is the so-called SWAP gate. As its name suggests,
the SWAP gate swaps the states of two qubits:

SWAP |ψ1 i |ψ2 i = |ψ2 i |ψ1 i

96
This definition immediately leads to the following set of rules for how SWAP acts on
the computational states:

|00i → |00i
|01i → |10i
|10i → |01i
|11i → |11i

In Dirac notation this is

SWAP = |00ih00| + |10ih01| + |01ih10| + |11ih11|

In terms of a matrix, we can write this as

 
1 0 0 0
0 0 1 0
 
0 1 0 0
0 0 0 1

Note that from its very definition, SWAP can never produce entanglement from an
initial product state. We can see this in detail if we expand out an arbitrary product
state in terms of the computational states:

|ψ1 i = α |0i + β |1i

|ψ2 i = γ |0i + δ |1i

|ψ1 i |ψ2 i = (α |0i + β |1i)(γ |0i + δ |1i) = αγ |00i + αδ |01i + βγ |10i + βδ |11i

Applying SWAP effectively exchanges the coefficients in front of |01i and |10i. The
resulting state can again be factorized, but this time γ and δ end up in the first qubit
part of the state, while α and β end up in the second qubit part:

(γ |0i + δ |1i)(α |0i + β |1i)

| {z }
product state

The result is of course a product state, and so SWAP never entangles. We will see
later on that if√we stop the SWAP gate halfway through, we obtain an entangling
gate known as SWAP.

97
5.2 Quantum circuits
A quantum circuit is a convenient notation that is used to describe a sequence of
gates acting on a set of qubits. We draw a horizontal line for each qubit, and we
place symbols on one or more lines to represent quantum logic gates. Schematically,
a quantum circuit looks something like this:

Init. states

|ψ1 i Measurement

|ψ2 i

..
.

- gate
−−−→
time

Each qubit is initialized in a state on the left, and gates are applied to the qubits
in sequence from left to right, i.e., time flows from left to right. Qubit measurements
are represeted by boxes containing a dial symbol.
Each common gate has a symbol associated with it:

98
 Pauli X, Y, Z X Rx (π), Ry (π), Rz (π)
 
1 0
Phase gate S = ∼ Rz (π/2)
0 i
 
1 0
π/8 gate T = ∼ Rz (π/4) = e−iπZ/8
0 eiπ/4
 
√1
1 1
Hadamard H = 2 1 −1
| {z }
X+Z

• control
CNOT
target

• Z •
CZ = =
Z • •

SWAP ×
×
A
generic controlled gate
•

Note that in the case of the CNOT gate, the symbol is asymmetric between the
control and target because the gate acts differently depending on which is which.
This is not true for the CZ gate—CZ looks the same regardless of which qubit is
the control and which the target. Because of this symmetry, the symbol for CZ is
symmetric between the two qubits. The same is true of the symbol for SWAP. On
the other hand, like CNOT, a generic controlled two-qubit operation should have
an asymmetric symbol to reflect the general asymmetry between control and target
qubits.
We can combine basic gates to make more complex n-qubit unitaries that can be
represented with quantum circuits. For example, we can make a controlled two-qubit

99
operation, CU, which implements an arbitrary single-qubit unitary U on the target if
the control is in |1i, by combining two CNOT gates together with some single-qubit
gates:
 
1 0
• • •
0 eiα
Controlled U =
U
C B A

Here, the single-qubit gates A, B, and C are defined such that

U = eiα AXBXC ABC = 1

Universal gate set

In computer science, the term universal computer is used to refer to a computing ma-
chine that is in principle capable of running any algorithm. This concept extends to
quantum computing, where people talk about universal quantum computers, which
are theoretical devices capable of running any quantum algorithm. In quantum
computing, an algorithm can be thought of as a particular sequence of gates and
measurements such that when it is applied to a set of qubits that are appropriately
initialized, the desired outcome of the algorithm is contained in the final state of the
qubits.

The concept of universal quantum computers naturally leads to the question:

What kinds of gates do we need in order to be able to realize any algorithm? A set
of gates that achieves this is called a universal gate set. Because quantum algorithms
are essentially just quantum circuits, and quantum circuits represent multi-qubit uni-
tary operators, this question boils down to asking what sets of gates do we need in
order to be able to make any multi-qubit unitary operator by combining them? As
you might expect, this question has been studied quite extensively over the past 40
years since people first started thinking about quantum computers. As you might
also expect, there are infinitely many universal gate sets. People are usually inter-
ested in the smallest such sets though, because if we have a smaller gate set, then
there are fewer gates that we need a quantum computer to implement reliably, and
getting a single gate to work well in a real quantum device can be quite hard. Here
are the two most common universal gate sets that people talk about:

100
 Set 1: CNOT, all single-qubit gates

Set 2: {CNOT, Hadamard (H), Phase (S), π/8 gate (T)}

The second set is particularly remarkable due to the fact that it contains only
four gates. From these four, we can make any multi-qubit unitary to arbitrary ac-
curacy. We will not prove this here. The proof only requires linear algebra, but it is
somewhat tedious. It can be found in e.g., the book by Nielsen and Chuang.

5.3 Quantifying entanglement

5.3.1 Cartan/KAK decomposition of two-qubit gates
So
√ far, we have mentioned four types of two-qubit gates: CNOT, CZ, SWAP, and
SWAP. Some of these (CNOT, CZ) can produce maximally entangled states from
product states, while others (SWAP) cannot. A generic two-qubit gate will produce
some amount of entanglement when it acts on certain product states. We can classify
two-qubit gates according to how much entanglement they are capable of creating.
The Cartan or KAK decomposition provides a natural way to do this classification.
A long time ago (long before quantum computing was a thing), Cartan showed that
a general two-qubit unitary U can always be written as the product of three different
two-qubit unitaries:
U = K (1) AK (2)
where K (1) and K (2) are each tensor products of single-qubit unitaries:
(1) (1)
K (1) = u1 ⊗ u2
(2) (2)
K (2) = u1 ⊗ u2

while A has the form

A = exp[i(αXX + βY Y + γZZ)]

where α, β, γ are three real parameters that depend on U . A cannot be expressed

as a tensor product of single-qubit unitaries. Clearly, K (1) and K (2) are not capable
of creating any entanglement since they rotate each qubit separately. Therefore, any
entanglement created by U must come from the entanglement created by A. Since
A is fully determined by α, β, γ, we see that these three parameters can be used to

101
quantify the entangling power of U . If two different unitaries U1 and U2 have the
same α, β, γ, then we say that they are locally equivalent, meaning that they are
the same up to single-qubit (i.e., local) gates. In this case, they will have the same
entangling power.

Let’s start with the simplest case where α = β = γ = 0. In this case, A is the
identity matrix, which is obviously not capable of creating entanglement. In this
case, U = K (1) K (2) is a tensor product of single-qubit unitaries, and so it too is inca-
pable of generating entanglement. We conclude that any U for which α = β = γ = 0
has zero entangling power.

Now consider the case α = β = 0, γ 6= 0, for which we have

 iγ 
e 0 0 0
 0 e−iγ 0 0
A = eiγZZ = 
0

0 e−iγ 0 
0 0 0 eiγ

If we choose γ = π/4, then if we stare at this long enough, we see that we can write
A as

A = e−iπ/4 eiπZ/4 ⊗ eiπZ/4 CZ
| {z }
single-qubit

Inverting this, we obtain a decompositionn for the CZ gate:

 
iπ/4 (1) (1)
CZ = e u1 ⊗ u2 eiπZZ/4
u1 = u2 = 1
(1) (1) (2) (2)
u1 = u2 = e−iπZ/4 ,
π
α = β = 0, γ = 4

What about CNOT? We can figure out the KAK decomposition for CNOT by
using the fact that it can be obtained from the CZ gate by applying Hadamard gates
to the target:
• •
CNOT = =
H • H
| {z }
Based on identity HZH = X

Combining this identity with the decomposition for CZ found above, we obtain the

102
following KAK decomposition for CNOT:


CNOT = eiπ/4 e−iπZ/4 ⊗ He−iπZ/4 eiπZZ/4 (1 ⊗ H)
π


We see that CNOT has the same entangling power α = β = 0, γ = 4
as and is
locally equivalent to CZ.

What about SWAP? It turns out we can express SWAP in terms of the following
matrix exponentiation:

SWAP = e−iπ/4 eiπ(XX+Y Y +ZZ)/4 (5.2)

This is already in the KAK form with α = β = γ = π/4 and K (1) = K (2) = 1 ⊗ 1.
However, we already know that SWAP has no entangling power, so we see that just
because the parameters are nonzero, α 6= 0, β 6= 0, γ 6= 0, this does not mean the
gate is entangling.
√
Now consider the case α = β = γ = π/8. This gives the entangling SWAP
gate:  
1 0 0 0
√ 0 λ iλ 0 1
SWAP =  0 iλ λ 0
 λ = √ e−iπ/4
2
0 0 0 1
√
SWAP |01i = λ |01i + iλ |10i = maximally entangled

√
We see that SWAP can produce maximally entangled states from product states,
but it’s not in the same class as CNOT and CZ since the values of α, β, γ are different.

In addition to classifying two-qubit gates, the KAK decomposition also provides

some information about what kinds of interactions are needed to produce different
kinds of entangling gates. For example, suppose that the two-qubit Hamiltonian is
H = JZ ⊗ Z. This is called an Ising interaction. In this case, usingn the above
decompositionn of the CZ gate, we see that the evolution operator U (t) = e−iHt
π
at time t = 4J is locally equivalent to a CZ gate. On the other hand, if we have
a Heisenberg interaction,
√ H = J(XX + Y Y + ZZ), then the evolution operator is
π
equivalent to a SWAP at time t = 8J .

103
5.3.2 Concurrence
Consider a general pure two-qubit state

|Ψgen i = c00 |00i + c01 |01i + c10 |10i + c11 |11i , (5.3)
P
with |cj |2 = 1. How can we tell if such a state is entangled versus separable? It is
actually simple to work this out. Consider the most general separable pure state of
two qubits:

|Ψsep i = (α0 |0i + α1 |1i) ⊗ (β0 |0i + β1 |1i) , (5.4)

P P
with |αj |2 = |βj |2 = 1. Expanding out the terms we have

|Ψsep i = α0 β0 |00i + α0 β1 |01i + α1 β0 |10i + α1 β1 |11i . (5.5)

Notice that for a product state, the coefficients of the four basis states are related.
Specifically, we see that iff a state is separable, the following holds:

c00 c11 − c01 c10 = 0. (5.6)

For a Bell state, this quantity is ±1/2. Of course, we know that all the Bell states have
the same amount of entanglement. We would also like the measure of entanglement
to range from 0 to 1. Given that this measure increases monotonically as we go from
a separable to a maximally entangled state, we define the concurrence of a pure state
as

C = 2|c00 c11 − c01 c10 |. (5.7)

It is worth making a connection between the concurrence and the coefficients of

a two-qubit state when it is expressed in the so-called ‘magic basis’. The latter is
defined as

|Φ1 i = Φ+
|Φ2 i = −i Φ−
|Φ3 i = Ψ−
|Φ4 i = −i Ψ+

104
 By expressing the general state (Eq. (5.3)) in the magic basis
X
|Ψgen i = µj Φ j , (5.8)
j

we can easily prove that the concurrence is

X
C= µ2j . (5.9)
j

Note that the definition of the concurrence for a mixed state ρ of two qubits
is much
p√more complicated. We define the matrices ρ̃ = (Y ⊗ Y ) ρ∗ (Y ⊗ Y ) and
√
R= ρρ̃ ρ. The matrix R is Hermitian, which means that it has real eigenvalues.
The concurrence is defined as C(ρ) = max(0, λ1 − λ2 − λ3 − λ4 ), where λ1 , ..., λ4 are
the eigenvalues of R in decreasing order. One can check that for a pure state this
expression is equivalent to what we defined above.

5.3.3 Makhlin invariants

An efficient way to test if two different 2-qubit gates have the same entangling power
is the use of the so-called Makhlin invariants. First we define the matrix
 
1 0 0 i
1 0 i 1 0
Q≡ √  
2 0 i −1 0 
1 0 0 −i

. This is the transformation matrix from the computational basis to the basis of magic
states defined above. Given any two-qubit unitary U , we can use Q to compute the
following 4 × 4 matrix:
M = Q| U | Q∗ Q† U Q.
Somewhat surprisingly (and very nontrivially), the eigenvalues of this matrix depend
only on the A part of the KAK decomposition of U . The eigenvalues of M are
completely independent of the angles inside the single-qubit “K” parts of the KAK
decomposition. They can therefore be used to quantify the entangling power of a
two-qubit gate. Specifically, we can define the following Makhlin invariants to classify
two-qubit gates into sets where all the gates in each set are related to each other by

105
single-qubit gates. There are two Makhlin invariants called G1 and G2 :

(TrM )2
G1 = ,
16 det(U )
(TrM )2 − Tr(M 2 )
G2 = .
4 det(U )

If two unitaries have the same G1 , G2 , they have the same entangling power. How
do G1 , G2 work? You can show by direct computation two important properties:
(i) If U = u1 ⊗ u2 is a tensor product of single-qubit gates, then M = 1. (ii) If
U = ei(αXX+βY Y +γZZ) , then the corresponding M is
 
M = diag e2i(α−β+γ) , e2i(α+β−γ) , e−2i(α+β+γ) , e−2i(α−β−γ) .

These two results can be used to show explicitly that M only depends on the A part
of the KAK decomposition. To see this, take U = K1 AK2 . Plugging this into the
general formula for M , we have

MU = Q| K2| A| K1| Q∗ Q† K1 AK2 Q

= Q| K2| Q∗ Q| A| Q∗ Q| K1| Q∗ Q† K1 QQ† AQQ† K2 Q
= Q| K2| Q∗ Q| A| Q∗ Q† AQQ† K2 Q
= Q| K2| Q∗ MA Q† K2 Q.

In the second equality, we used that Q is unitary so that QQ† = 1 = Q∗ Q| . In the

third equality, we used result (i) to get rid of K1 :

Q| K1| Q∗ Q† K1 Q = MK1 = 1.

Now if we take the trace, we find that all the K2 dependence also drops out:


Tr(MU ) = Tr Q† K2 QQ| K2| Q∗ MA = Tr(MA ). (5.10)
| {z }
M † † =1
K2

106
Similarly, Tr(MU2 ) = Tr(MA2 ). As a consequence, the Makhlin invariants G1 and G2
are purely functions of α, β, γ:
1 −4i(α+β+γ)  2
G1 = e 1 + e4i(α+β) + e4i(α+γ) + e4i(β+γ) ,
16
G2 = cos(4α) + cos(4β) + cos(4γ).

Let’s see what the Makhlin invariants are for the common two-qubit gates we in-
troduced earlier:

CNOT or CZ: α = β = 0, γ = π/4

1
⇒ G1 = (−1)[1 + 1 − 1 − 1]2 = 0
16
G2 = 1 + 1 − 1 = 1

Identity gate: α = β = γ = 0
1
G1 = [1 + 1 + 1 + 1]2 = 1
16
G2 = 3

SWAP: α = β = γ = π/4
1
G1 = (−1)[1 + 1 + 1 + 1]2 = −1
16
G2 = −1 − 1 − 1 = −3

107
6 Qubit interacting with quantized field
When we discussed the Rabi problem and related problems, we examined what hap-
pens when we couple a qubit to a classical electric or magnetic field. Now we want
to treat the field quantum mechanically, at the same level as the qubit. We will first
see how to quantize electric and magnetic fields, and then we will re-examine the
problem of a qubit coupled to a (now quantum) field. Treating the field itself as
a quantum mechanical object is important for understanding at a microscopic level
phenomena like spontaneous emission that are caused by the fluctuating quantum
mechanical vacuum. It is also important in cases where the field is strongly confined
in one or more dimensions, such as in a cavity or waveguide.

6.1 Field quantization

To quantize the electric and magnetic fields, we first start from Maxwell’s equations
in free space (no charge or current sources):

~ = 0, ~ = 0, ~˙
~ = −B, ~ = 1 ~˙
∇·E ∇·B ∇×E ∇×B E
c2
Recall that we can derive wave equations for the fields. For example, by taking the
curl of the third equation and using the fourth, we find:

⇒ ∇×∇×E ~˙ = − 1 E
~ = −∇ × B ~¨
| {z  } c2
∇ ∇·E ~ −∇2 E~
| {z }
=0

~ = 1 ~¨
⇒ ∇2 E E
c2
This wave equation admits solutions that can be separated into spatial and temporal
parts. In general, the temporal part can be expanded in a Fourier series. Here for
simplicity, we will focus on the special class of solutions where this series contains
only a single term corresponding to frequency ω:
~ r, t) = f~(~r) α0 e−iωt +c.c.
E(~ (6.1)
| {z }
α(t)

108
The spatial part satisfies the Helmholtz equation:

ω2 ~
∇2 f~ = − f (6.2)
c2

The first Maxwell equation also requires that the spatial part satisfies ∇ · f~ = 0.
The function f~ is called a mode function. The word mode just refers to a Fourier
component. Here, we have just one mode. If we had more than one mode, then we’d
have a different f~ for each, and we’d need to label them accordingly. This is obvious
from the Helmholtz equation 6.2, which depends on ω. The function f~ is therefore
characteristic of a particular mode, which is why it is called a mode function. The
particular form of this function depends on the boundary conditions. For example,
if we are interested in the electromagnetic field inside a three-dimensional, perfectly
conducting box, then the mode function would be

conducting 3D box: f~(~r) = f0,x cos(Kx x) sin(Ky y) sin(Kz z)x̂

+ f0,y sin(Kx x) cos(Ky y) sin(Kz z)ŷ
+ f0,z sin(Kx x) sin(Ky y) cos(Kz z)ẑ,

where
πnx πny πnz
Kx = , Ky = , Kz = .
Lx Ly Lz
This solution follows from the fact that for a perfect conductor, the electric field
is perpendicular to the surface (because the free charges in the conductor perfectly
screen fields). For the conducting box, this means that e.g., at the x = 0 wall,
we must have f~ ∼ x̂. The fact that the wave vector is quantized means that the
frequency is also quantized:

~
2 ω2
K = ⇒ ω quantized
c2
Once we have a solution for the electric field, we can obtain the magnetic field
by computing the curl and then integrating with respect to time:
 
B~˙ = −∇ × E ~ = −∇ × f~ α0 e−iωt + c.c.
 
⇒B ~ = −i ∇ × f~ α + c.c.
ω
So far we have just reminded ourselves how to solve the electric and magnetic fields

109
in free space, possibly with nontrivial boundary conditions. To figure out how to
quantize these fields, we will examine the Hamiltonian. Recall that this is the integral
of the energy density of the fields, which depends on E~ 2 and B
~ 2:
Z   Z  
1 ~2 1 ~2 0 ~ 2 + c2 B
~2
H= dV 0 E + B = dV E (6.3)
2 µ0 2

Again sticking with the case of a single mode and assuming that the mode function
R 2
is normalized, dV f~(~r) = 1, the Hamiltonian can be computed explicitly:
Z Z   2 
2
~2 = 2 ~2 ~∗ ∗ 2 2 ~
dV E dV α f + (α ) f + 2|α| f
Z 2 Z
  2  2 
2 ~ =
2 c 2 ~ ∗ 2 ~∗ 2 ~
2
c dV B dV −α ∇ × f − (α ) ∇ × f + 2|α| ∇ × f
ω2
Z  2 Z h  i Z  
~ ~
dV ∇ × f = dV ∇ · f × ∇ × f + ~ dV f~ · ∇ × ∇ × f~
| {z } | {z }
~ [f~×∇×f~]=0 2 R
= dV f~·(−∇2 f~)= ω2 dV f~2
R R
= dS· (Helmholtz)
c

Here, the underlined terms cancel each other, and we assumed that the mode function
vanishes at the boundary of the space. The remaining terms then give
Z  
0 2 2 c2
2 ~ 2 ~
⇒H = dV 2|α| f + 2|α| ∇ × f
2 ω2
Z 2
Z h  i Z  
~ ~ ∗ ~
dV ∇ × f = dV ∇ · f × ∇ × f + dV f · ∇ × ∇ × f ~∗ ~
| {z } | {z }
=0 ω2
| ~|2
R
dV f
c2
Z
0

⇒H = dV 4|α|2 = 20 V |α|2 = 20 V αr2 + αi2
2
In the last step, we defined the real and imaginary parts of α to be αr and αi ,
respectively. Since α(t) = α0 e−iωt , these are given explicitly by

αr (t) = α0 cos(ωt), αi (t) = −α0 sin(ωt)

Notice that these two functions obey the following simple differential equation:
−1
αr = α̇i .
ω

110
This is in fact the defining equation of a simple harmonic oscillator. This is easier
to see if we redefine these two functions in terms of a generalized coordinate q and
its time-derivative q̇:
r r
1 mω 2 −1 m
αi = √ q ⇒ αr = √ q̇.
20 2 20 2
Here, we included some dimensionful constants so that when we insert these expres-
sions back into the Hamiltonian,

m 2 mω 2 2 p2 mω 2 2
H= q̇ + q = + q , p ≡ mq̇
2 2 2m 2
we obtain precisely the usual Hamiltonian for a harmonic oscillator. What we have
shown is that each mode of an electromagnetic field in free space behaves like a
harmonic oscillator. This is the key to quantizing the fields: We can re-interpret q
and p as quantum mechanical operators instead of ordinary functions of time. This
will mean that E ~ and B~ become quantum mechanical operators also. Before we
explore the repercussions of this, let’s recall some basic properties of the quantum
harmonic oscillator. The so-called algebraic solution in terms of ladder operators
will be particularly important.

Review of Simple Harmonic Oscillator

The Hamiltonian of a simple harmonic oscillator is
  r
p2 mω 2 2 ~ω p2 q 2 √ ~
H= + q = + p0 ≡ ~mω, q0 ≡ .
2m 2 2 p20 q02 mω

We are particularly interested in the algebraic solution using the so-called ladder
operators a, a† , which are defined as
   
1 q p † 1 q p
a≡ √ +i a ≡√ −i (6.4)
2 q0 p0 2 q0 p0

Using that [q, p] = i~, we can easily find that commutator of a with a† :
 
 † 1 q p q p i i ~
a, a = +i , −i = [p, q] − [q, p] = =1
2 q0 p0 q0 p0 2p0 q0 2p0 q0 p0 q 0

111
This commutation relation is the key to solving for the eigenvalues and eigenstates
of H. Solving Eq. (6.4) for p and q in terms of a and a† , we find
p 1
q 1

= √ a − a† = √ a + a†
p0 2i q0 2

We can use these expressions to express H in terms of a and a† :

 
~ω 1

† 2 1

† 2 ~ω  †
⇒H= − a−a + a+a = aa + a† a + aa† + a† a
2 2 2 4
 
~ω  †  † † ~ω  † † 1
= 2a a + 2 a, a + 2a a = 4a a + 2 = ~ω a a + (6.5)
4 4 2

Our goal is to find the energy spectrum and eigenstates of H:

H |ni = En |ni

From Eq. (6.5), we see that this is tantamount to finding the spectrum of a† a ≡ n̂,
which is referred to as the “number operator” for reasons that will become clear
shortly. The states |ni are eigenstates of n̂ with eigenvalues n:

n̂ |ni = n |ni assume hn|ni = 1 normalized (6.6)

Because the expectation value of the number operator can be interpreted as the norm
of a state, it follows that the eigenvalues n must be nonnegative:

n = hn|n̂|ni = hn|a† a|ni = ka |nik2 ≥ 0 (6.7)

Therefore, we see that the lowest possible eigenvalue is n = 0 with a |0i = 0. The
ground state energy is therefore E0 = ~ω 2
.
Now let’s focus on the excited states. The first step is to note the following
commutation relations involving the number operator and each ladder operator:
   
[n̂, a] = a† a, a = a† , a a = −a, (6.8)

and similarly  
n̂, a† = a† . (6.9)
With these commutators in hand, now consider what happens when we apply the

112
number operator to a |ni and a† |ni:

n̂(a |ni) = (an̂ − a) |ni = (n − 1)a |ni (6.10)

n̂ a† |ni = a† n̂ + a† |ni = (n + 1)a |ni (6.11)

These results show that a |ni and a† |ni are both eigenstates of n̂ = a† a with eigen-
values n ∓ 1. This means that a |ni ∝ |n − 1i and a† |ni ∝ |n + 1i. We can find the
proportionality constants by again using the fact that the norms of these states can
be interpreted as expectation values of n̂:
√
ka |nik2 = n → a |ni = n |n − 1i
2
√
a† |ni = hn|aa† |ni = hn| a† a + 1 |ni = n + 1 → a† |ni = n + 1 |n + 1i

We see that a, a† decrease/increase the eigenvalue n by 1. This is why they are

called ladder operators—they can be used to descend/ascend the ladder of energy
eigenstates. a is called the lowering operator, while a† is called the raising operator.
They are also called the annihilation/creation operators. This is because from the
point of view of a quantized electromagnetic field, we can interpret n as the number
of photons comprising the field, and so a (a† ) reduces (increases) the number of
photons.
Note that n can only assume integer values because non-integer eigenvalues of n̂
would lead to negative-norm states if we applied a enough times. Starting from the
ground state |0i, we can generate all the excited states by repeatedly applying the
raising operator: √
a† |0i = |1i a† |1i = 2 |2i , . . . (6.12)
1 1
n
|ni = √ a† |n − 1i = . . . = √ a† |0i (6.13)
n n!
The energies of these states are
 
1
En = ~ω n + , n = 0, 1, 2, . . . (6.14)
2

We see that the energies are perfectly equispaced with spacing ~ω. See the sketch of
the spectrum below.

113
 The last bit of information about the harmonic oscillator we need is what the
ladder operators look like in the Heisenberg picture. Recall that an operator in the
Schrödinger picture, AS , transforms to a Heisenberg operator AH (t) by taking the
frame transformation operator to be the full evolution operator u(t):

AH (t) = u† (t)AS u(t) (6.15)

The (generally time-dependent) operator AH (t) obeys the Heisenberg equation of

motion:  
dAH i ∂AS
= [H, AH ] + (6.16)
dt ~ ∂t H
Let’s solve this equation to obtain the lowering operator for the harmonic oscillator:
daH i i  
= [H, aH ] = [H, a]H = iω a† a, a H = −iωaH
dt ~ ~
⇒ aH (t) = ae−iωt (6.17)
Note that the constant of integration must be an operator. From the general defini-
tion of AH (t) above, we see that AH (0) = AS since u(t) = 1. This is how we knew to
set the integration constant equal to the Schrödinger operator a. Now that we know
what the lowering operator looks like in the Heisenberg picture, we can immediately
write down the raising operator in this picture:

a†H (t) = a† eiωt (6.18)

114
 Let’s return to the task of quantizing the electromagnetic fields. The key step
is that we upgrade α(t) (the time-dependent part of the fields) from an ordinary
function of time to a quantum mechanical operator:
" r #
1 −1 mω 2
α(t) = αr + iαi = √ √ p+i q (6.19)
20 2m 2
" √ √ #
1 ~ω a − a† ~ω

=√ − +i a + a† (6.20)
20 2 i 2
r
~ω
=i a(t) (6.21)
20

Because α(t) is time-dependent, it follows that the operators p, q, a, a† that we

rewrite it in terms of are in fact Heisenberg-picture operators. Here, we are sup-
pressing the H subscript on a and a† , e.g., a(t) = ae−iωt . We then arrive at the
following quantized electric and magnetic fields:
r h i
~ r, t) = i ~ω f~(~r)a(t) + f~∗ (~r)a† (t)
E(~ (6.22)
20
r
~ ~  ~

B(~r, t) = ∇ × f a(t) + H.c. (6.23)
2ω0

6.2 Jaynes-Cummings model

We are now ready to revisit the problem of a qubit coupled to an electric field. This
time we will treat the field quantum mechanically. We will continue to focus on the
case where the field contains only a single mode with frequency ω. We can think of
the qubit as two levels of an atom |gi and |ei, and we’ll shift the overall energy so
that the lower level |gi has zero energy. The total Hamiltonian is then

H = HA + HC + HAC , (6.24)

where HA is the Hamiltonian for the atom,

HA = ~ω0 |eihe| , (6.25)

115
with transition frequency ω0 , and HC is the Hamiltonian of the field, which we now
know can be written like a harmonic oscillator Hamiltonian:
 
† 1
HC = ~ω a a + (6.26)
2

Here, we are thinking of the atom as sitting inside a cavity, and so HC describes the
energy of the field inside this cavity. The atom couples to this field through a dipolar
coupling like we considered in the case of a qubit coupled to a classical field (as in
the Rabi problem):
HAC = −d~ · E
~ Schrod (6.27)
~ Schrod is the electric field in the Schrödinger picture, which does not have
Here, E
time-dependence. It is given by
r h i
~ω ~
~
ESchrod = i f (~r)a + f~∗ (~r)a† , (6.28)
20

where a and a† are (time-independent) Schrödinger ladder operators. The interaction

depends on the strength of the field at the location ~r of the atom. The vector d~ in
HAC is the dipole moment operator of that atom. In terms of the two atomic levels,
we can write it as


d~ = atomic dipole operator = hg|d|ei
~ (|gihe| + |eihg|) ≡ d~ge σ + σ † (6.29)

where in the last equality we introduced raising and lowering operators for the atom:
σ = |gi he| and σ † = |ei hg|. We may then write the interaction term in the Hamil-
tonian as r
~ω
h i
HAC = −i σ + σ † d~ge · f~(~r)a + f~∗ (~r)a† (6.30)
20
To simplify the notation a bit, we can define the atom-field coupling energy:
r
~ω ~
~g(~r) = −i dge · f~(~r) (6.31)
20
in terms of which, the interaction is now

 
HAC = ~ σ + σ † g(~r)a + g ∗ (~r)a† (6.32)

116
Note that the coupling strength depends on the position ~r of the atom. We can
always choose the phases of the atomic states such that g(~r) is real and positive,
yielding


HAC = ~g σ + σ † a + a† (6.33)
The last step is to make use of the rotating-wave approximation (just like we did for
the Rabi problem we studied earlier). This step is summarized in the following box:

RWA: First go to the interaction picture:

† −iω0 tσ †σ   1  
e|iω{z
0 tσ σ
} σe = σ + iω0 tσ † σ, σ + iω0 tσ † σ, iω0 tσ † σ, σ + . . .
2
u†A

 † 
σ σ, σ = σ † σσ − σσ † σ = −σσ † σ = −σ
1
u†A σuA = σ − iω0 tσ − (ω0 t)2 σ + . . . = σe−iω0 t
2
Similarly, u†C auC = ae−iωt . Assuming that |ω − ω0 |  ω + ω0 , this means that
two of the terms in HAC are rapidly rotating and can therefore be discarded:

⇒ σa ∼ e−i(ω0 +ω)t , σ † a† ∼ ei(ω0 +ω)t → discard (rapidly oscillating)

Going back to the lab frame, we have

HAC = ~g σa† + σ † a

We are then left with the following Hamiltonian describing the interaction be-
tween a single mode of an electric field and a two-level system:


Jaynes-Cummings: H = ~ω0 σ † σ + ~ωa† a + ~g σa† + σ † a (6.34)

This is a famous Hamiltonian called the Jaynes-Cummings model. Note that we have
dropped the vacuum field energy ~ω/2 because it just amounts to an overall energy
shift of the whole system. Let’s spend some time understanding the physics of this
model. First we’ll study the free evolution of an initial state under this Hamiltonian.
We will then come back and analyze the energy spectrum and eigenstates.

117
6.2.1 Free evolution in Jaynes-Cummings model
In this section, we are interested in seeing how an initial state evolves under the
Jaynes-Cummings Hamiltonian. We will see that the behavior is quite simple and also
very familiar. Consider a general state describing the combined atom-field system:
∞
X
|ψi = [cg,n |g, ni + ce,n |e, ni] (6.35)
n=0

We can think of n as the number of photons in the cavity. From now on, let’s set
~ = 1. To see how this state evolves, we need to solve the time-dependent Schrödinger
equation:
∂
i |ψi = (HA + HC + HAC ) |ψi (6.36)
∂t
To evaluate the action of the Hamiltonian on |ψi, it helps to use the Dirac form of
the ladder operators:
∞
X ∞
X
√ √
a= n + 1 |nihn + 1| , a† = n + 1 |n + 1ihn| (6.37)
n=0 n=0

Plugging |ψi into the Schrödinger equation gives

∞ h
X ∞
X
iċg,n |g, ni + iċe,n |e, ni] = [ω0 ce,n |e, ni + nω(cg,n |g, ni + ce,n |e, ni)
n=0 n=0
√ √ i
+ g n + 1ce,n |g, n + 1i + g ncg,n |e, n − 1i (6.38)

Equating the coefficients of each ket on both sides of the equation results in the
following coupled differential equations:
√
iċg,n+1 = (n + 1)ωcg,n+1 + g n + 1ce,n (6.39)
√
iċe,n = (ω0 + nω)ce,n + g n + 1cg,n+1 (6.40)

We see that |e, ni only couples to |g, n + 1i and vice-versa; this means that we
can separately solve the Schrödinger equation in each of these 2 × 2 subspaces. To do
this more easily, we can rewrite these two coupled equations as a single 2-component

118
vector equation:
   √  
ċg,n+1 (n√+ 1)ω g n + 1 cg,n+1
i = (6.41)
ċe,n g n + 1 ω0 + nω ce,n
| {z }
Hn

In terms of Pauli matrices, Hn is given by

∆ ∆ √ ∆ λ
Hn = (nω + ω0 + )1 + σz + g n + 1σx ≡ (nω + ω0 + )1 + σθ (6.42)
| {z 2 } 2 2 2
ignore in dynamics

Here, ∆ = ω − ω0 . As usual, we can ignore the first, identity term since it only
contributes a time-dependent global phase on the state. The other terms in Hn
look very similar to what we had in the Rabi problem in the rotating frame of the
(classical) driving field. Let’s recall what the Hamiltonian and evolution operator
looked like in that case:
Recall Rabi problem:
  p
1 ∆ Ω0 λ λ = Ω20 + ∆2
Hef f = = σθ
2 Ω0 −∆ 2 tan(θ) = Ω∆0
   
λt λt
⇒ uef f (t) = e −iλtσθ /2
= cos 1 − i sin σθ
2 2
√
If we make the replacement Ω0 → 2g n + 1, then Hef f becomes exactly Hn
(after dropping the identity term in the latter). Therefore, we can borrow our Rabi
problem expression for the evolution operator to write down the evolution operator
for the Jaynes-Cummings Hamiltonian:
   
λn t λn t
⇒ uJC = e −iλn tσθ /2
= cos 1 − i sin σθ
2 2
√
p 2g n+1
λn = 4g 2 (n + 1) + ∆2 tan(θn ) =
∆
σθ = cos(θn )σz + sin(θn )σx

Suppose we start in the state |e, ni. What happens? To answer this question, we

119
just need to apply the evolution operator to this initial state:
      
λn t λn t 0
uJC (t) |e, ni = cos 1 − i sin [cos(θn )σz + sin(θn )σx ]
2 2 1
          
λn t 0 λn t 0 λn t 1
= cos + i sin cos(θn ) − i sin sin(θn )
2 1 2 1 2 0
       
λn t λn t λn t
= cos + i sin cos(θn ) |e, ni − i sin sin(θn ) |g, n + 1i
2 2 2
(6.43)

What is the probability that the atom will end up in its ground state at a later time?
This is easily computed:
 
(n) 2 2 λn t
Pe→g (t) = | hg, n + 1|uJC (t)|e, ni| = sin sin2 (θn ) (6.44)
2
4g 2 (n + 1) 2
p
= 2 sin ( 4g 2 (n + 1) + ∆2 t/2) (6.45)
4g (n + 1) + ∆2 | {z }
Quantized Rabi frequency $n

√
We thus find what looks like the usual Rabi oscillations, but where 2g n + 1 plays
the role of the driving amplitude. The atom emits a photon and relaxes to the
ground state. It then re-absorbs a photon and goes into the exited state again, and
so on. The frequency of this cyclic process increases with the number of photons
n. It makes sense that Ω0 has been replaced by an expression that increases with
n since if we have more photons, then this should correspond to a field with higher
intensity. Notice that these Rabi oscillations occur even if the cavity initially has no
photons. In this case, the generalized Rabi frequency is
p
$0 = “vacuum Rabi frequency” = 4g 2 + ∆2 (6.46)

6.2.2 Energy spectrum of the Jaynes-Cummings model

Now let’s examine the energy spectrum and eigenstates of the Jaynes-Cummings
model. First of all, we can identify the ground state by inspection. Notice that
HJC annihilates the state |g, 0i corresponding to the atom in its ground state and
no cavity photons: HJC |g, 0i = 0. This tells us that |g, 0i is the ground state, with
zero energy.
Now consider the excited states. We can again borrow our previous results for
the classical field Rabi problem to write down the energies and eigenstates. The

120
Figure 5: Energy spectrum of the Jaynes-Cummings Hamiltonian for g = 0 (left)
and g > 0 (right).

eigenvalues of Hn are
∆ λn ∆ 1p 2
nω + ω0 + ± = nω + ω0 + ± 4g (n + 1) + ∆2 (6.47)
2 2 2 2
while the corresponding eigenstates are
   
θn θn
cos |g, n + 1i + sin |e, ni ≡ |+, (n)i (6.48)
2 2
   
θn θn
− sin |g, n + 1i + cos |e, ni ≡ |−, (n)i (6.49)
2 2

Interestingly, these states describe hybrid light-matter excitations which are com-
monly referred to as polaritons. They describe something which is neither light nor
matter, but something in between. The energy spectrum of these states are sketched
below for the case of zero detuning: ω = ω0 ⇒ ∆ = 0. For comparison, we show
both the zero-coupling case and the case of nonzero coupling g.
For both zero and nonzero coupling, there is a unique ground state |g, 0i. When
g = 0, each excited state is doubly degenerate, and the energy levels are uniformly
spaced with spacing ω = ω0 . On the other hand, when g > 0, each degenerate pair
of excited states couple together to form two new states that are non-degenerate.
Moreover, the splitting between these states increases with photon number n. The
structure of this interacting spectrum gives rise to interesting phenomena, including
something called photon blockade, which will be discussed later in Sec. 6.3.1.

121
6.2.3 Open cavity and lossy qubit
So far, we considered the JC Hamiltonian in an idealized setting, i.e., without any
sources of non-unitary dynamics. In reality, there is an environment that can interact
with the cavity and atom. This is the electromagnetic field outside the cavity, which
in the idealized situation is completely decoupled. In reality, however, the modes of
the cavity can couple to modes of free space, and decay and dissipation can occur.
We will see later how this is done in a more formal way, but for now we will consider
a decay rate for the cavity, usually denoted by κ and a decay rate for the qubit,
usually denoted by γ, as we saw in Sec. 2.6.
Then, we can write down an equation for the joint density operator ρ of the
qubit-cavity system using Lindblad terms:
κ †
γ †

ρ̇ = i[ρ, H] − a aρ + ρa† a − 2aρa† − σ σρ + ρσ † σ − 2σρσ † , (6.50)
2 2
where H is the JC Hamiltonian and σ (σ † ) is the lowering (raising) operator for the
qubit. There are then three types of transitions that can occur in this system:
(1) Free space emission couples |e, ni to |g, ni
(2) Coherent evolution couples |e, ni to |g, n + 1i
(3) Cavity decay couples |e/g, n + 1i to |e/g, ni
…

|#3⟩
…

#
|%2⟩

* 3)
2)

# |#2⟩
|%1⟩
*
) 2)

# |#1⟩
|%0⟩
*
)

|#0⟩

Let’s consider the implications of the dissipative terms in the physics of the

122
coupled system. Clearly, the timescales (or, equivalently, energy scales) we need to
consider are the coupling strength g and the two decay rates κ and γ. When g  κ, γ,
i.e., in the so-called strong coupling regime, we essentially have the dynamics we saw
above in Sec. 6.2.1, but with decay at longer times. When, on the other hand, the
dissipation rates are comparable to g, then the atom and cavity don’t have time to
interact enough (i.e., complete a full Rabi cycle) before decay sets in. This is how
we can understand the different regimes from the point of view of timescales. We
can also think about it in the frequency (energy) domain, where dissipation broadens
the transitions, and if this broadening is large then the splitting between the states
(which is set by g) cannot be resolved.

6.3 Physics and uses of the JC Hamiltonian

6.3.1 Photon Blockade
Consider an ideal cavity (strong coupling regime). Suppose the cavity is initially
empty, n = 0, and the atom is in |gi. If we inject a photon with frequency ωp = ω −g,
the system gets excited to |−, (0)i (see Fig. 5). If we want to add a second photon,
the system
√ must undergo the transition
√ |−, (0)i → |−, (1)i, but this has frequency
(2ω − 2g) − (ω − g) = ω − ( 2 − 1)g 6= ωp . Thus, when g is large enough, if we
send in two photons with frequency ωp , only the first one gets trapped in the cavity.
This is a way to create effective interactions between photons, which otherwise don’t
interact with each other.

6.3.2 Dispersive readout

Let’s revisit the situation of the JC Hamiltonian with large detuning, i.e., when the
qubit and photon frequencies differ by more than the coupling strength, |ω − ω0 | =
|∆| > g. Physically, we expect that at the limit of a very large detuning between
the two systems, exchange of energy will be suppressed, i.e., the qubit cannot absorb
a cavity photon and be excited from the ground to the excited state, and it also
cannot emit a photon into the cavity mode and transition from the excited to the
ground state (strictly speaking, it can, but these are very low probability events). In
that case, we can transform the Hamiltonian by taking into account the smallness
of g/∆ so that we obtain a form that does not involve the energy exchange terms
σa† and σ † a. We can do this via a canonical transformation (sometimes also called
a Schrieffer-Wolff transformation) using the operator
hg
i
† †
T = exp σa − σ a , (6.51)
∆

123
taking

H −→ T HT † = Hef f , (6.52)

and keeping the lowest order. This results in an effective Hamiltonian

 
g2 g2
Hef f = ~ ω0 − σ † σ + ~ωa† a + ~ σz a† a. (6.53)
∆ ∆

This approximate Hamiltonian, which is valid in the dispersive limit, manifestly sup-
presses joint transitions between qubit and photon states. However, the interaction
is still nontrivial and, in fact, useful as we will see in a moment. The atom ‘sees’
the cavity through the last term, which manifests as a shift of the qubit transition
frequency. This shift is photon-number dependent, and it is present even in the case
of an empty cavity (this is analogous to the so-called Lamb shift from atomic physics,
where the interaction of an atom with the vacuum electromagnetic field leads not
only to spontaneous emission but also to a shift in the atomic frequency). The fact
that the shift is photon-number dependent is interesting: it means that by probing
the qubit we can deduce the photon number in the cavity. Now, there is symmetry
between the two systems: we can also focus on the cavity, and view the last term
as the qubit shifting the cavity frequency (this is clear if the second and third term
are combined). In that case, by probing the cavity we can tell whether the qubit
is in the ground or excited state. This is commonly used in superconducting qubit
platforms for qubit readout (so-called dispersive readout).

6.3.3 Cavity-mediated two-qubit interaction

Another crucial capability the JC Hamiltonian offers is an effective coupling of two
qubits. Consider two qubits, which are separated enough such that they do not
interact with each other directly. We can consider the case where each interacts
with the same cavity mode. It is straightforward to write down the Hamiltonian
for the total system: it will include the free Hamiltonians for each qubit, the free
Hamiltonian for the cavity, and the interaction terms between the cavity and qubit
1 and the cavity and qubit 2:
   
† † † † † † †
H = ~ω1 σ1 σ1 + ~ω2 σ2 σ2 + ~ωa a + ~g1 σ1 a + σ1 a + ~g2 σ2 a + σ2 a . (6.54)

If we were to solve the eigenvalue equation, we would of course find a generalization

of the JC spectrum for one qubit discussed above in Sec. 6.2.2. Now, the types of

124
states that would couple in each manifold would conserve excitation number over the
three subsystems. For example, introducing the convention that the first (second)
entry corresponds to the first (second) qubit and the third to the cavity, the set of
states {|g, e, 1i , |e, g, 1i , |e, e, 0i , |g, g, 2i} would couple and form a manifold (the two-
excitation manifold). For most quantum information applications, we are interested
in storing information in the qubits. The cavity serves as an auxiliary system, e.g.,
for readout as shown above in Sec. 6.3.2 and for coupling as discussed here. In these
cases, we do not want to populate the cavity with photons. We can therefore imagine
a regime where we detune the cavity from the qubits (the qubits themselves need not
be in exact resonance, but should be closer to each other than to the cavity) such
that there are no real excitations transferred to the cavity from the qubits. Then,
we can follow a similar procedure as we did above for dispersive readout to eliminate
the energy exchanging terms with the cavity. The transformation in this case is
" #
X  †

†
S = exp λj a σj − aσj , (6.55)
j

where
gj
λj = − . (6.56)
ωj − ω

Then the effective Hamiltonian is

 
Hef f = SHS † = ~ω̃1 σ1† σ1 + ~ω̃2 σ2† σ2 + ~g̃ σ1 σ2† + σ1† σ2 , (6.57)

where ω̃j and g̃ depend on the original couplings gj and on the detunings ∆j = ωj −ω.

6.3.4 Purcell effect

Returning to a single qubit coupled to a cavity, and taking dissipative terms into
account, we now consider the weak coupling limit, where κ  g, γ. In this case, it
can be shown that the decay rate of the qubit is modified by the interaction with
the cavity. The effective decay rate of the qubit becomes γ + γκ , where
κ
γκ = g 2 . (6.58)
∆2 + (κ/2)2

125
Note the distinct behavior of different limits of the detuning. For ∆  κ, this
expression becomes

g2
γκ ' κ, (6.59)
∆2
while for ∆  κ, the enhancement becomes independent of the detuning:

4g 2
γκ = . (6.60)
κ
This enhancement of the spontaneous emission rate is useful for developing efficient
single-photon sources, where it is desirable to be able to produce photons at a high
rate.

6.4 Spontaneous emission in free space

So far, we have focused entirely on a single mode. If we consider the electromagnetic
field inside a very large box of linear dimension L, then there will be a large number
of modes that are closely spaced in frequency since the wave vector components
are integer multiples of 2π/L, and ω = c|~k|. This is the regime of spontaneous
emission. We can describe this situation with a Hamiltonian that is similar to the
Jaynes-Cummings model, but where we now have a dense set of modes:
X  
† 1
H0 = ~ω0 |eihe| + ~ωk ak,ζ ak,ζ + . (6.61)
k,ζ
2

Here, k labels the different field modes, while ζ labels the two independent polariza-
tions, so that the quantized electric field has the form
r
~ωk
Ek,ζ (r) = fk,ζ (r)ak,ζ + H.c. (6.62)
20
The atom-field dipolar interaction then becomes
r  
X ~ωk †
† ∗
HAF = − (σ + σ )dge · fk,ζ (r)ak,ζ + fk,ζ (r)ak,ζ , (6.63)
k,ζ
20

126
where dge = hg| d |ei is the dipole matrix element. Again invoking the rotating-wave
approximation, we can write this as
X  
HAF = ~ gk,ζ σ † ak,ζ + gk,ζ
∗
σa†k,ζ , (6.64)
k,ζ

where the coupling is defined as

r
ωk
gk,ζ (r) = − dge · fk,ζ (r). (6.65)
20 ~

We can use this Hamiltonian to derive the spontaneous emission rate. Sponta-
neous emission is the process by which an atom initially in its excited state decays
to its ground state and emits a photon. Therefore, we should consider the following
state, which consists of all one-excitation basis states:
X
|ψi = ce |ei + ck,ζ |g, 1k,ζ i . (6.66)
k,ζ

The first term corresponds to the case in which the atom is in its excited state and
there are no photons. The second term includes all states in which the atom is in its
ground state, and there is one photon. Spontaneous emission corresponds to starting
in state |ei and transitioning to a superposition of one-photon states. To compute
the rate, we must solve the time-dependent Schrödinger equation:
i
∂t |ψi = − (H0 + HAF ) |ψi . (6.67)
~
Plugging Eq. (6.66) into this equation and matching coefficients, we obtain
X
∂t c̃e = −i gk,ζ c̃k,ζ e−i(ωk −ω0 )t (6.68)
k,ζ
∗
∂t c̃k,ζ = −igk,ζ c̃e ei(ωk −ω0 )t . (6.69)

Here, we defined c̃e = ce eiω0 t and c̃k,ζ = ck,ζ eiωk t . Integrating Eq. (6.69) and plugging
the result into Eq. (6.68) gives
X Z t
0
∂t c̃e = − |gk,ζ |2
dt0 c̃e (t0 )e−i(ωk −ω0 )(t−t ) . (6.70)
k,ζ 0

127
Inserting the expression from Eq. (6.65), using that in a free space of volume V the
mode function is a plane wave,
1
fk,ζ (r) = √ ε̂k,ζ eik·r , (6.71)
V
and assuming the atom is spherically symmetric so that |ε̂k,ζ · dge |2 = d2ge /3, we
obtain Z t
d2ge X 0
∂t c̃e = − ωk dt0 c̃e (t0 )e−i(ωk −ω0 )(t−t ) . (6.72)
30 ~V k 0

To deal with the sum over k, we consider the limit where V → ∞, in which case the
sum can be replaced by an integral:
X Z Z ∞
V 3 V
→ 3
dk→ 2 3 dωω 2 . (6.73)
k
(2π) 2π c 0

The last step here makes sense so long as the integrand only depends on ω, which is
true in the present calculation. We then have
Z ∞ Z t
d2ge 0
∂t c̃e = − 2 3
dωω 3
dt0 c̃e (t0 )e−i(ω−ω0 )(t−t ) . (6.74)
6π 0 ~c 0 0

0
Now consider the frequency dependence of the integrand. The function e−i(ω−ω0 )(t−t )
oscillates very quickly and causes the time integral to average out to zero, except
when ω ≈ ω0 . Therefore, the time integral is strongly peaked at ω = ω0 . In
comparison, the ω 3 factor varies slowly, and so we can set ω 3 → ω03 and bring it
outside the integral. The remaining frequency integration gives a delta function:
Z ∞
0
e−i(ω−ω0 )(t−t ) ≈ 2πδ(t − t0 ). (6.75)
0

This is approximate because the lower integration limit is 0 instead of −∞. With
these approximations, we are then left with

ω03 d2ge c̃e

∂t c̃e = − . (6.76)
3π0 ~c3 2

128
Thus, the excited state decays at a rate

ω03 d2ge
γ= . (6.77)
3π0 ~c3
This is the spontaneous emission rate. We see that it depends on the transition
frequency of the atom as well as on the dipole moment of this transition.

7 Quantum computing platforms

7.1 Semiconductor spin qubits
The electron spin is conceptually the simplest possible qubit. It is naturally a two-
level system, and its control has been demonstrated in experiments since the early
days of quantum mechanics (although, granted, the type of control required for
quantum information processing is at an entirely different level of sophistication).
Semiconductors are ideal hosts of spin qubits, as they constitute a mature technology,
notably as chips for classical computers.

7.1.1 Brief note on semiconductors

Solids involve a huge number of atoms. As such, the discrete energy levels of atoms
become continuous energy bands in solids. Semiconductors have an energy gap be-
tween the highest occupied band (valence band) and the lowest unoccupied band
(conduction band), and carriers (electrons or holes) can be excited to the conduc-
tion band. Semiconductors can be grown into nanostructures such as quantum wells
(with confinement in one direction), quantum wires (with confinement in two direc-
tions), and quantum dots (with confinement in all three directions). Here we will
exclusively focus on quantum dots and donors, both of which effectively feature con-
finement in all directions, but we note that quantum wires are also investigated for
quantum computing, notably for topological (Majorana) qubits. The two dominant
semiconductor materials that are explored for quantum information processing are
silicon and GaAs (which is a type of III-V semiconductor). One crucial advantage
of silicon is that it can be isotopically purified, meaning that the ions with spinful
nuclei (∼ 4.6% natural abundance) can be eliminated. On the other hand, all of the
nuclear spins in GaAs are spinful, which leads to decoherence of the electron spin
qubit through the hyperfine interaction, and isotopic purification is not possible.
The energy En (k) of the nth electronic band in a semiconductor is expressed as
functions of the wave vector k (also known as crystal momentum when multiplied

129
by ~). For our purposes, it suffices to think of this as analogous to momentum.
At the bottom of the conduction band and the top of the valence band we can
approximate the energy as a parabolic function of k. As such, we can define an
effective mass m∗ , which is determined by the curvature of the energy at these
extrema: m∗ = ~2 (∂ki ∂kj En |k=k∗ )−1 . Here, k∗ is the location of the band extremum,
and we are assuming that the various second-order derivatives are all equal; this does
not have to be the case, and more generally the effective mass is described by a tensor
rather than a single number. Defining the effective mass is extremely useful, because
then for many practical purposes we can think of electrons near the band extrema in
the semiconductor as being just like free particles, and the only consequence of the
surrounding semiconductor crystal is that the mass differs from that of free electrons.
Thinking of semiconductor electrons as free particles with mass m∗ allows us to use
simple models to describe them. For example, the particle in a box description is
surprisingly good in estimating properties such as energy level spacings in a quantum
dot.

7.1.2 Quantum dots

A quantum dot can be defined as an abstract object that can trap single electrons
and that is connected to a source and a drain, which can be biased to move electrons
into and out of the dot. Here we will exclusively focus on semiconductor quantum
dots. These have a size of 30-100 nm in the lateral direction and about roughly
10 nm thickness. Quantum dots are created by first putting together two different
types of semiconductor, for example silicon with SiGe or with SiO2 , or GaAs with
AlGaAs. This creates an electrostatic confinement potential at the interface of the
two materials, trapping a two-dimensional gas of electrons there. Researchers then
place carefully shaped metallic “gate electrodes” on the surface of one of the materi-
als using electron beam lithography techniques. By adjusting the voltages on these
gate electrodes, it is possible to push most of the electrons in the 2D gas away, while
leaving a small number of electrons trapped in the vicinity of the electrodes. The
precise number of trapped electrons can be chosen at will by carefully adjusting the
gate electrode voltages. The dot is also connected to a source and a drain, which
are reservoirs made up of the remaining 2D electron gas. These are used to either
provide additional electrons or take away electrons from the dots.

Typically, there is a static magnetic field, which splits the two spin states of an
electron (Zeeman splitting). This allows both single-qubit gates and readout. An
important concept here is the Pauli exclusion principle, which prohibits two electrons

130
from occupying the same state. As a result, an electron can be added to the dot only
if it has enough energy to occupy the next highest available energy state (see figure).

I. Single electron spin qubit

Measurement: spin-to-charge conversion

Single-shot measurement is very important in quantum information processing

(meaning, we only have one copy of the system and need to measure it with high
fidelity instead of averaging over different runs). Measuring single spins is extremely
challenging. Spin-to-charge conversion allows us to convert the spin information to
charge information, and this is hugely important because single charge measurements
are much easier. Spin-to-charge measurements work through a combination of the
Pauli exclusion principle, Coulomb blockade, and something called a quantum point
contact. See figure.
The basic idea is that a gate electrode is used to adjust the energy levels in the
dot such that an electron is trapped only if it is in the lower-energy spin state, while
if it is in the other spin state, it will tunnel out of the dot and into the reservoir.
In this way, information about the spin of the electron is mapped to a statement
about how much electrical charge is in the dot. For concreteness, let’s assume that
|↑i has lower energy than |↓i, which is the case in GaAs. If the electron is in |↑i,
then it stays in the dot, and the dot remains (negatively) charged at all times. On
the other hand, if the electron is in |↓i, then after some time it will tunnel into the
reservoir, leaving the dot uncharged. However, this leaves open a state on the dot
that is lower in energy than the states in the reservoir. Therefore, an electron from
the reservoir in state |↑i will eventually tunnel into the dot, so that the dot again
becomes negatively charged. Thus, if we can tell that the dot was empty for some
amount of time, then we can infer that the original electron in the dot had spin ↓,
while if the dot was charged throughout, then it must have had spin ↑.
Now that we know how to convert spin information to charge information, the
question becomes: How can we measure the charge in the dot? This is done with
something called a quantum point contact (QPC). This is a small constriction in a
semiconductor that only allows electrons to pass through one at a time. We can
think of this constriction as a small dip in an otherwise high electrostatic barrier
that allows electrons to tunnel through the barrier. The rate at which the electrons
tunnel through it depends sensitively on the shape and size of the dip. The current
generated by electrons tunneling through the barrier can be measured with very high
precision. The key idea is to put such a constriction very close to the quantum dot so

131
that the barrier shape depends sensitively on whether there is an electron in the dot
or not. This allows us to determine whether or not the dot is charged by measuring
the current through the QPC. If the current remains steady throughout, then the
electron must have been in |↑i, while if the current increases momentarily (when the
dot is empty, the QPC barrier lowers a little), then the electron must have been in
|↓i. This is the most commonly used method for measuring an electron spin in a
quantum dot.

Decoherence

The most important decoherence mechanisms in semiconductor quantum dot spin

qubits are charge noise and hyperfine interactions with the nuclear spins of the sur-
rounding ions in the semiconductor lattice. Charge noise is a ubiquitous effect in
solid state devices; the precise origins are unclear, and it is likely that there are
several sources depending on the details of the device. The bottom line is that some-
thing causes fluctuations in the electrostatic potential that confines the electrons,
and this in turn can give rise to decoherence of the spin. The decoherence caused
by nuclear hyperfine interactions is particularly significant in GaAs because in this
material, all the ions have spinful nuclei. In silicon quantum dots, the situation is
better because only a small fraction of the ions have nonzero spin, and this fraction
can be reduced substantially using isotopic purification. Tremendous effort has been
devoted to understanding and overcoming these sources of noise; they remain one
of the central challenges to realizing quantum information technologies with these
systems.

Single-qubit gates

Single-qubit gates are conceptually simple in these systems. Since we have a two-
level spin, we can use a time-dependent AC magnetic field, just as we discussed in
Sec. ??, to induce Rabi oscillations between the two states. As we saw before, this
allows us to generate arbitrary single-qubit gates. This technique is called electron
spin resonance, because the most efficient approach is to use an AC magnetic field
that oscillates at a frequency that is close to or equal to the Zeeman splitting of the
electron.
Note that in practice it is often more convenient to generate time-dependent elec-
tric fields, since this can be done through the gate electrodes on the device. There
exist ways to convert this electric field into an effective magnetic field through the
use of a micromagnet that creates a magnetic field gradient, so there is an effective

132
coupling between the spin and the position of the electron. Since the spatial part of
the wavefunction (which can be thought of as a charge distribution) is sensitive to
the electric field, and that couples to the spin, we can use an electric field to drive
the spin. This technique is called electric dipole spin resonance.

Two-qubit gates

There are several ways to implement two-qubit gates between two quantum dot
spin qubits. Most of these methods rely on the exchange interaction, which we now
describe. Afterward, we briefly describe how each of these types of gates are imple-
mented.

Exchange interaction—Suppose we have two electrons trapped in two neighboring

quantum dots (one in each). Also suppose that each electron is in the lowest orbital
state of its dot. Let’s call the lowest orbital in the left dot |Li and the lowest orbital
in the right dot |Ri. There are four possible states of the two electrons:
1
|Si = (|RLi + |LRi) ⊗ (|↑↓i − |↓↑i) , (7.1)
2
1
|T0 i = (|RLi − |LRi) ⊗ (|↑↓i + |↓↑i) , (7.2)
2
1
|T+ i = √ (|RLi − |LRi) ⊗ |↑↑i , (7.3)
2
1
|T− i = √ (|RLi − |LRi) ⊗ |↓↓i . (7.4)
2
Each of these states is antisymmetric under the exchange of the two electrons, as is
required for indistinguishable fermions. The first state |Si, which is called the singlet,
is antisymmetric in the spin degrees of freedom and symmetric in the orbital part,
so that the entire state is antisymmetric. On the other hand, each of the remaining
triplet states is antisymmetric in the spatial part and symmetric in the spin part.
The fact that the triplet states have a different orbital part compared to the
singlet leads to a difference in the Coulomb energy between the two electrons in the
two cases. This difference in Coulomb energy essentially comes from the fact that
the electron charge is distributed differently in the two types of states. However,
because the antisymmetry of the entire state ties the (anti)symmetry of the spatial
part to the (anti)symmetry of the spin part, we can say that effectively the Coulomb
energy depends on the spin state of the two electrons. This is what is known as the
exchange interaction. Let’s see it in more detail.

133
 The Coulomb interaction is VC = e2 /|r1 − r2 |, where r1 , r2 are the positions of
the two electrons. The Coulomb energy of the singlet state is
1
hS| VC |Si = (hRL| + hLR|) VC (|RLi + |LRi)
2
1
= (hRL| VC |RLi + hLR| VC |LRi + hRL| VC |LRi + hLR| VC |RLi) .
2
(7.5)

On the other hand, the Coulomb energy of a triplet state is

1
hT | VC |T i =(hRL| − hLR|) VC (|RLi − |LRi)
2
1
= (hRL| VC |RLi + hLR| VC |LRi − hRL| VC |LRi − hLR| VC |RLi) .
2
(7.6)

Note that this expression holds for all three triplet states. Taking the difference of
these two energies gives the exchange energy:

J = hT | VC |T i − hS| VC |Si = − hRL| VC |LRi − hLR| VC |RLi . (7.7)

Notice that the exchange energy involves matrix elements in which the single-electron
states are switched (or exchanged) between the bra and the ket. This is the key signa-
ture of an exchange energy. It means that the exchange energy will only be nonzero
if there is overlap between the spatial wavefunctions ψL (r) and ψR (r), because oth-
erwise the integrand in e.g.,
Z Z
e2
hRL| VC |LRi = d r1 d3 r2 ψR∗ (r1 )ψL∗ (r2 )
3
ψL (r1 )ψR (r2 ) (7.8)
|r1 − r2 |

will always be zero if ψL (r) is zero whenever ψR (r) is nonzero.

To understand better why the exchange energy can be treated as arising from an
effective spin-spin interaction, consider the operator JS1 ·S2 , where S1 and S2 are the
spin operators for the first and second electron, respectively. This operator mimics
VC for the states in Eqs. (7.1)-(7.4) in the sense that it has the same expectation
value for all three triplet states, while its expectation value for the singlet is smaller

134
by J. This is easy to check if we use the identity
1 
S1 · S2 = (S1 + S2 )2 − S21 − S22 (7.9)
2
and the fact that |Si and |T i are eigenstates of total spin (S1 + S2 )2 with eigenvalues
0 and 2, respectively. These states are also eigenstates of S21 and S22 with eigenvalues
3/4. Therefore

J
hS| JS1 · S2 |Si = [0 − 3/4 − 3/4] = −3J/4, (7.10)
2
J
hT | JS1 · S2 |T i = [2 − 3/4 − 3/4] = J/4, (7.11)
2
and so the difference of the two expectation values is J. This effective spin-spin
interaction is convenient to work with because it allows us to ignore the spatial
degrees of freedom and only worry about the spins. However, you should always keep
in mind that its fundamental origin is the Coulomb interaction and not some sort of
magnetic spin-spin coupling such as a dipole interaction. The fact that the exchange
interaction ultimately comes from the Coulomb force is important because it means
that J can be quite large, much larger than a magnetic dipole-dipole interaction for
instance. This means it can be used to create fast entangling two-qubit gates. It’s
also worth noting a bit of terminology that is widely used: An interaction of the form
S1 · S2 is called a Heisenberg interaction, as we noted in Sec. 5.3.1. Consequently,
the effective interaction JS1 · S2 is often called the Heisenberg exchange interaction.
The full two-qubit Hamiltonian is

H = gµB B1 · S1 + gµB B2 · S2 + JS1 · S2 , (7.12)

where g is the Landé factor, and µB is the Bohr magneton. If we assume that the
magnetic fields in the two dots both point in the z direction, then this becomes
ω1 ω2 J
H= σ1z + σ2z + σ1 · σ2 , (7.13)
2 2 4
where ω1 , ω2 are the Zeeman splittings of the two spins.
√
SWAP two-qubit gate— √ The exchange interaction JS1 · S2 can be used to gen-
erate the entangling gate SWAP. In a quantum dot, J can be tuned by raising
and lowering the barrier between the two dots, because this changes the overlap of
the two orbitals ψL (r) and ψR (r). In the limit of infinite barrier, the two dots are

135
isolated, so hL|Ri → 0, and therefore J → 0. By√lowering the barrier, we switch on
J, enabling us to implement a SWAP gate or a SWAP entangling gate depending
on how long J is left on for and assuming that J  ω1 , ω2 . We can see that these
gates result by recalling from Sec. 5.3.1 that they are generated from Heisenberg
interactions. This approach requires very large J to get good results.
√
CNOT gate—One way to realize a CNOT gate is to combine two SWAP gates
together with single-qubit rotations using the identity:
√ √
CNOT = i(e−i 4 σ1z ⊗ Hei 4 σ2z ) SWAP(1 ⊗ e−i 2 σ2z ) SWAP(1 ⊗ H),
π π π
(7.14)

where H is the Hadamard gate. Alternatively, a CNOT gate can be achieved through
a combination of tuning J and applying a time-dependent AC field. To see how this
works, consider the effect that J has on the spectrum of Eq. (7.13):
> wz
Assuming w ,

J -

-
O J 70

III. In
" ITT)
wit
# f*"÷ ,
497

HA
www..ro#IIw..otIM
n

HA
We see that the transition frequency of each qubit now depends on the state of the
other qubit. For example, to first order in J, the transition frequency of the first
qubit is ω1 + J/2 when the second qubit is in state |↑i, while the transition frequency
is ω1 − J/2 when the second spin is in |↓i. Therefore, we can generate a conditional
operation by driving the first qubit using an (effective) AC magnetic field with fre-
quency ω1 + J/2. If we time the magnetic pulse appropriately, this will perform a
g
NOT gate on qubit 1 only if qubit 2 is in state |↑i. Note that the “dressed” states |↑↓i
g are linear combinations of |↑↓i and |↓↑i. Ultimately, we want to perform an
and |↓↑i
operation on the computational states |↑↑i, |↑↓i, |↓↑i, |↓↓i. We can accomplish this
by adiabatically turning J on, applying the magnetic pulse, and then adiabatically
turning J off. During the adiabatic parts of the evolution, the computational states
evolve into the dressed states (and vice versa), so that the CNOT gate implemented

136
on the dressed states becomes a CNOT gate on the computational states. See Zajac
et al, Science 359, 439 (2018) for a nice experimental demonstration of this approach
in silicon quantum dot spin qubits. This approach has the benefit of not requiring
J to be very large.

Cavity-mediated gates–A single qubit can also be defined in a double quantum

dot. The motivation for doing this is that, in combination with a magnetic field
gradient, the spin state can become sensitive to position, which in turn means an
increased sensitivity to electric fields. The motivation for doing this is that it can
allow coupling to a resonator (cavity). This is desirable for scalability, as it allows
us to couple remote qubits to each other and provides more complex architectures
compared to a 2D grid of quantum dots with nearest-neighbor couplings.

Let’s take a closer look at how the spin can be sensitive to an electric field. This
is the same mechanism that underlies the technique called electric dipole spin reso-
nance (EDSR) that we mentioned earlier. Suppose that the two dots are separated
by some distance along the x direction. Also suppose that there is an external mag-
netic field of the form B = B0 ẑ + βxŷ, where B0 and β are constants. That is,
the component in the z direction is uniform in space, while the component in the
y direction increases linearly in x, such that it is larger in the right dot compared
to the left. We can quickly shuttle the electron back and forth between the two
dots by oscillating voltages on the gate electrodes. Because the y component of the
magnetic field changes from one dot to the other, the electron then effectively sees a
time-dependent magnetic field that oscillates at the frequency of the shuttling. This
field is in the y direction, and so it will drive transitions between the spin states,
provided the shuttling frequency is close to the Zeeman splitting caused by B0 . This
is EDSR. The same mechanism can also be used to couple the spin to the electric
field of a cavity, because the cavity field can also shuttle the electron back and forth
in a manner similar to oscillating gate voltages.

II. Singlet-triplet qubit

In addition to using a single electron spin as a qubit, it is also possible to define

a qubit using the spin states of multiple electrons. The simplest and most common
multi-electron spin qubit is called the singlet-triplet spin qubit, or ST qubit for short.
In this case, two electrons are used, with one electron in each of two neighboring dots.
The idea is to use the singlet and one of the triplet states as the two logical states.
In particular, the triplet with sz = 0 is used. These states were already discussed

137
above (see Eqs. (7.1) and (7.2)). There we kept both the orbital and spin parts of
the states, but it is common practice to ignore the spatial parts when talking about
singlet-triplet qubits, in which case we simply write
1 1
|Si = √ (|↑↓i − |↓↑i) , |T0 i = √ (|↑↓i + |↓↑i) . (7.15)
2 2
The motivation for using these states as a qubit as opposed to the spin states of a
single electron is that these states do not couple to uniform magnetic fields (both
states have ms = 0). Thus, if the environment produces magnetic field fluctuations
(and both dots see the same fluctuations), then the ST qubit will be immune to this,
whereas a qubit based on a single electron spin would not be.

Another advantage of the ST qubit is that it can be controlled using only DC

electric fields, without the need for EDSR. To see how this works, consider again the
two-qubit Hamiltonian with exchange coupling, Eq. (7.13). Taking matrix elements
of this in the ST qubit subspace yields the ST Hamiltonian:
J ω1 − ω2
HST = σz + σx . (7.16)
2 2

Although the Zeeman frequencies ω1 , ω2 are typically fixed in experiments, J is easily

tuned by changing the voltages on gate electrodes. Therefore, we can treat J(t) as a
function of time that we are free to shape. We can then recognize HST as being just
like the Rabi Hamiltonian in the rotating frame, only that we’ve swapped the two
Pauli matrices, which just amounts to a simple change of basis. Therefore, we have
the same amount of controllability as in the Rabi problem and, in particular, we can
generate any single-qubit gate by choosing J(t) appropriately. Of course, the phys-
ical origin of this Hamiltonian is very different. One extra complication compared
to the Rabi problem, is that in order to design J(t), we need to first understand
how J depends on the voltages of the gate electrodes. Mapping out this dependence
requires some extra experimental effort. Another downside to using ST qubits is that
now we need twice as many quantum dots to create the same number of qubits. If
we want to demonstrate a two-qubit gate for instance, we need a 4-dot device.

Initialization of an ST qubit is done by raising the ground state energy of one dot
relative to the other until both electrons end up in the same dot. The lowest-energy
state corresponds to both electrons occupying the lowest spatial orbital of the dot,
which can only happen if the two electrons are in a singlet state. Therefore, if we

138
tilt the dots and wait long enough, the two electrons will end up in the singlet state.
We can then reverse the tilt so that both dots become symmetric, and the electrons
separate into the two dots. If this is done quickly enough, then the electrons will
still be in a singlet at the end of this procedure. Operations can then be performed
by raising and lowering the barrier between the dots to change J while maintaining
symmetry between the dots.

Readout of an ST qubit also utilizes spin-to-charge conversion, only here it looks

a bit different compared to the one-dot case. The process involves tilting the dots
like we do for initialization, only now we pay attention to how much we need to tilt
the double dot before both electrons end up in the lower dot. If the electrons are
in the singlet state, then not too much tilting is needed to achieve this. However, if
the electrons are in the |T0 i state, then there is a higher energy penalty for putting
both electrons in the same dot because they cannot occupy the same spatial orbital
(due to the Pauli exclusion principle). A very large tilt is needed to put the second
electron in the first excited spatial orbital of the dot. Thus, if we tilt the dots enough
so that a singlet in one dot can form, but a triplet cannot, then we can use a QPC to
measure the charge distribution (i.e., one electron in each dot versus two electrons
in the same dot) and infer the spin state.

7.1.3 Donors in silicon

In addition to quantum dots, an alternative way to create spin qubits in semicon-
ductors is to use electronic or nuclear spins localized on acceptor or donor atoms.
Acceptor or donor atoms are just additional atoms residing in a semiconductor that
are not a part of the normal lattice structure of the material. When an acceptor is
placed in a semiconductor, it accumulates some extra charge from the electrons in
the semiconductor. In contrast, donors “donate” some electrons to the semiconduc-
tor. The intentional doping of semiconductors with acceptors and/or donors is the
basic technique used to create transistors, and thus it has been fundamental to the
semiconductor electronics industry for many decades.

In the context of spin qubits, the most popular approach is to dope silicon (Si)
with phosphorous (P) donors. Once implanted, the P donor carries one excess elec-
tron and has a spin 1/2 nucleus. It thus offers two types of spin qubit, each with
advantages and disadvantages. A big advantage of the nuclear spin qubit is that it
has very long coherence times because it interacts very weakly with its environment.

139
Recent experiments have demonstrated nuclear spin coherence times on the order of
minutes. This is the record for any spin qubit. Electron spins are more sensitive
to their environment, and so their coherence times are typically about two orders of
magnitude shorter. However, the advantage of electron spins is that they are easier
to manipulate and couple to one another. Thus, both qubits are being investigated
as possible building blocks of a quantum computer.

In 1998, Bruce Kane proposed to use both types of spin qubit in what is now
known as the Kane quantum computer. The idea is to use the nuclear spins as
long-lived quantum memories, while the electron spins are used to store and read
information on these memories and to create entanglement between multiple spins.
The information storage and readout processes can be performed by exploiting the
hyperfine interaction between the electron and nuclear spins. This interaction is of
Heisenberg form: Hhf = AS · I, where S and I are the spin operators of the electron
and nucleus, respectively. A is called the hyperfine coupling constant. Because this
interaction is of Heisenberg form, it naturally generates SWAP operations, which
is precisely what we need in order to swap information between the electronic and
nuclear spins. Electron spins are selectively rotated by using local gate electrodes to
change the Landé g factors (which depend on applied electric fields), which in turn
brings the spins into or out of resonance with an AC magnetic field. Interactions are
mediated by the electron spins. These can be either exchange interactions, dipolar
interactions, or microwave cavity-mediated interactions. Recent experiments have
demonstrated very precise single-qubit control for both electron and nuclear spin
qubits. The first two-qubit gate between donor spins was only realized in the last
two years using exchange interactions. The donor approach to spin qubits offers very
long coherence times and high-fidelity gates, but the fabrication of devices is very
challenging.

7.2 Superconducting qubits

7.2.1 LC circuits
LC circuits are crucial concepts of superconducting quantum information processing.
Here we will review classical LC circuits and then proceed to quantize them.

An LC circuit consists of a capacitor (with capacitance C) and an inductor (with

inductance L), with the former (latter) storing energy in the form of an electric
(magnetic) field. In the LC circuit, as the current oscillates, the energy moves from

140
the capacitor to the inductor, i.e., it transforms from electric to magnetic and back.
When the capacitor is charged, an electric field is created between the capacitor
plates. This creates a current that flows from the positively to the negatively charged
plate. When the current passes through the inductor, a magnetic field is generated.
The time-varying magnetic flux in the inductor causes a voltage with the opposite
sign, recharging the capacitor, but with the opposite charge. The physics can be
described by the combination of Ampere’s law and Faraday’s law:
Z
B~ · d~` = µ0 N I (7.17)
Z
E~ · d~` = − dΦ , (7.18)
dt
where Φ is the magnetic flux through the inductor. The total energy can be expressed
as a function of the squares of the electric and magnetic fields.
Z   Z  
1 ~ 2 1 ~2 0 ~ 2 + c2 B
~2
H= dv 0 E + B = dv E (7.19)
2 µ0 2

As we saw in Section X, we can promote the fields into operators and find the
commutation relation between them using the standard canonical quantization pro-
cedure. Expressing the E (B) field in terms of the charge and capacitance (flux and
inductance),

V Q
E = = , (7.20)
d Cd
Φ
B = , (7.21)
N AL
where d is the distance between the capacitance plates, AL is the area of the solenoid
cross section and N is the number of its turns, we automatically have the quantization
condition for those quantities as

[Q̂, Φ̂] = i~ (7.22)

Using the relations

0 AC
C = (7.23)
d
µ0 AL N 2
L = , (7.24)
`

141
where AC is their area of the capacitor plates, we obtain the Hamiltonian

Q̂2 Φ̂2
H= + . (7.25)
2C 2L
Now that this Hamiltonian has been written as that of a simple harmonic os-
cillator, we can introduce creation and annihilation operators, as we did for the
electromagnetic field in Section Y:
s r
~ C
Q̂ = (a + a† ), (7.26)
2 L
s r
~ L
Φ̂ = −i (a − a† ). (7.27)
2 C

(Note that sometimes people choose a different convention in which Φ̂ is proportional

to a + a† and Q̂ is proportional to a − a† .) This then gives
 
† 1
H = ~ω a a + , (7.28)
2
√
where ω = 1/ LC is the frequecy of the oscillator. Below we will discuss how this
relates to the Hamiltonian of a superconducting circuit. First, we will give some brief
information on superconductivity.

7.2.2 A brief note on superconductivity

Superconductors are certain materials which below a certain (low!) temperature ex-
hibit the following properties: their resistance drops to zero, which means that cur-
rent flows without dissipation, and magnetic fields are expelled from them (the latter
is called the Meissner effect). Typical superconductors are the elements aluminum
(Al), bismuth (Bi), niobium (Nb), tin (Sn), etc. There exist also superconducting
compound materials. Superconductivity is a quantum mechanical, collective effect,
meaning that it can be explained by considering the many-body wavefunction of the
conduction electrons.

A key aspect of the theory of conventional superconductors is an effective at-

tractive potential between electrons (mediated by the positively charged ions). This
leads to a many-body wavefunction, which consists of so-called Cooper pairs, i.e.,

142
pairs of electrons with opposite wavevectors and opposite spin. These Cooper pairs,
which have charge −2e, are the relevant quasiparticles and carriers of current. In the
ground state of the superconductor, the electrons are paired up into Cooper pairs.
The first excited state is obtained from the ground state by breaking a Cooper pair, a
process that comes with an energy cost 2∆, which is called the superconducting gap.
To remain in the ground state, and maintain the superconductive behavior, clearly
the temperature has to be much lower than the energy gap. (We note in passing that
the sensitivity of superconductivity to excitations that overcome the energy gap is
used in superconducting single photon detectors, a technology that is used in very
different platforms of quantum information processing.)

7.2.3 The Josephson junction

Since here we are considering the low temperature limit, kB T  2∆, we can focus
on the ground state only. We can therefore think of the effect of superconductivity,
for our purposes, as reducing the huge Hilbert space of all the electrons to a single
state. The idea of reducing a large Hilbert space to a finite number of sharp, atomic-
like states, is critical for using solid-state systems as qubits. Here we have only one
state, so to make this a proper qubit we need to double the size of the Hilbert space.
To do this, we will consider two small superconducting electrodes connected by a
tunnel junction, i.e., a small piece of an insulator through which Cooper pairs can
tunnel coherently. The fact that Cooper pairs, rather than single electrons, can tun-
nel quantum mechanically through such barriers is called the Josephson effect, and
it is a key component of superconducting qubits.

We will now introduce a simple model to describe the complicated system given
by two superconducting islands coupled by a tunnel junction. We consider the total
number of Cooper pairs to be fixed over the two islands N = NL + NR , where NL
(NR ) is the number of pairs in the left (right) island. We can describe the state using
the basis

|mi = |NL − m, NR + mi , (7.29)

where m represents the number of Cooper pairs that have transferred through the
Josephson junction from the left to the right island. We will describe tunneling using
the phenomenological Hamiltonian
1 X
HT = − EJ (|mi hm + 1| + |m + 1i hm|) , (7.30)
2 m

143
where we introduced the parameter EJ , which is called the Josephson coupling energy
and quantifies the ability of Cooper pairs to tunnel through the Josephson junction.
The eigenstates of this Hamiltonian are superpositions of the {|mi} states with the
following (unnormalized) form:
∞
X
|φi = eimφ |mi . (7.31)
m=−∞

To see that these are indeed eigenstates and to find the eigenvalues, we apply the
Hamiltonian from Eq. (7.30):

EJ X X imφ 0

HT |φi = − e |m i hm0 + 1| |mi + eimφ |m0 + 1i hm0 | |mi
2 m m0
EJ X imφ
= − e (|m − 1i + |m + 1i) (7.32)
2 m

Redefining the variables that are summed over, this becomes

EJ X i(m+1)φ

HT |φi = − e |mi + ei(m−1)φ |mi
2 m
EJ iφ
X imφ
= − e + e−iφ e |mi
2 m
= −EJ cos φ |φi (7.33)

Next let’s define a Cooper pair number operator

X
n̂ = m |mi hm| . (7.34)
m

It is straightforward to show that

X 0
hφ| n̂ |φ0 i = meim(φ −φ) . (7.35)
m

To see what the form of the operator n̂ is in the φ representation, we calculate

Z
1 π 0
hφ| n̂ |Ψi = dφ hφ| n̂ |φ0 i hφ0 | |Ψi , (7.36)
N −π

144
where we have inserted the identity in the form of an integral over the states {|φi}.
Using Eq. (7.35) we have
Z
1 π 0 X im(φ0 −φ) 0
hφ| n̂ |Ψi = dφ me hφ | |Ψi ,
N −π m
Z
1 X −imφ π 0 imφ0
= e dφ me Ψ(φ0 )
N m −π
X Z π  
1 −imφ 0 ∂ imφ0
= e dφ −i 0 e Ψ(φ0 ). (7.37)
N m −π ∂φ

Doing an integration by parts and using the fact that everything is periodic in φ, we
get
Z !
1 π 0 X im(φ0 −φ) ∂
hφ| n̂ |Ψi = dφ e i 0 Ψ(φ0 ). (7.38)
N −π m
∂φ

The quantity in parenthesis equals N δ(φ − φ0 ). Substituting this into the equation
above gives
∂
hφ| n̂ |Ψi = i Ψ(φ). (7.39)
∂φ
We can therefore identify
∂
n̂ = i , (7.40)
∂φ

which implies that φ̂ and n are conjugate operators, i.e.,

[n̂, φ̂] = −i (7.41)

From the eigenvalue equation of HT , and given that this Hamiltonian shares
eigenstates with φ̂, we can express it as

HT = −EJ cos φ̂. (7.42)

Next we will understand why the Josephson junction can be thought of as an

inductor. If in Eq. (7.41) we replace n̂ with −Q̂/2e, then to recover the LC circuit

145
commutator, Eq. (7.22), we need to replace φ̂ with
2e
φ̂ = Φ̂, (7.43)
~
Note however that here we do not have a magnetic field (due to the Meissner effect),
and the origin of the flux is the kinetic energy of the Cooper pairs.

7.2.4 Cooper pair box

So far we have focused on the tunneling aspect of the Josephson junction. However,
the discussion has been incomplete because it ignored the fact that as current flows
through the junction, the current that builds up on the superconducting islands and
the resulting Coulomb energy become important. To include the Coulomb energy
we add a term

U = 4EC (n̂ − ng )2 , (7.44)

where the factor of 4 is because the Cooper pairs have double the charge of the
electron, and ng is an offset charge due to an applied field or inherent junction
asymmetry. The total Hamiltonian for this two-superconducting-island system is
then

HCP B = 4EC (n̂ − ng )2 − EJ cos φ̂. (7.45)

Here, the subscript CPB stands for Cooper pair box, which is how this system is
called.
To connect back to the LC circuit discussion above (which we have already in
part motivated by showing the correspondence between φ̂ and the flux), we can think
of the first term as originating from the capacitor of the circuit and the second from
the inductor. This can be made more obvious by considering the limit EJ  EC ,
in which case the potential term −EJ cos φ̂ forces hφ̂i to stay close to zero. We can
then expand the cosine up to second order (and ignore ng for now) to get

1
HCP B ' 4EC n̂2 + EJ φ̂2 . (7.46)
2

Note that strictly speaking, φ̂ is not a proper operator, as it describes a compact

variable, so only periodic functions of φ̂ are properly defined. Nevertheless, in the

146
limit we are considering, the wave function has significant support only near φ = 0.
Therefore ignoring the periodicity of φ does not really introduce a large error.

We should stress that the LC Hamiltonian, Eq. (7.46), cannot be used as a qubit.
This is because the spectrum is that of a simple harmonic oscillator, which means
that the energy levels are equispaced and, as a result, a qubit cannot be isolated: A
field that addresses the lowest transition would also drive all the other ones as well,
causing severe leakage out of the ‘qubit’ subspace. The cosine term however avoids
this issue, as the spectrum becomes now one of an anharmonic oscillator.

We can use the analogy between the Cooper pair box and the LC circuit to write
down the Cooper pair box Hamiltonian in terms of creation and annihilation oper-
ators b† and b by borrowing results from the LC circuit analysis above. Comparing
the quadratic terms of HCP B to those of the quantum LC circuit Hamiltonian, we
can read off an effective capacitance and inductance:

e2 ~2
CJ = , LJ = . (7.47)
2EC 4e2 EJ
Using Eq. (7.26), we find
s r
 1/4
Q̂ 1 ~ CJ † −5/4 EJ
n̂ = − = − (b + b ) = −2 (b + b† ), (7.48)
2e 2e 2 LJ EC
s r
 1/4
2e 2e ~ LJ † EC
φ̂ = Φ̂ = −i (b − b ) = −i21/4
(b − b† ), (7.49)
~ ~ 2 CJ EJ

and so to fourth order in φ̂, we obtain

1 EJ 4
HCP B ≈ 4EC n̂2 + EJ φ̂2 − φ̂
2
r 24 r
E J E J
√ EC EJ EC
= 4EC 2−5/2 (b + b† )2 − 2 (b − b† )2 − 2 (b − b† )4
EC 2 EJ 24 EJ
p  
† 1 EC
= 8EC EJ b b + − (b − b† )4 . (7.50)
2 12
√
The first term is a simple harmonic oscillator with frequency ω01 = 8EC EJ /~.
The second term creates an anharmonicity, meaning that the energy levels are not
perfectly equispaced. For example, if the transition frequency between the lowest

147
two energy levels |0i and |1i is ω01 , then the transition frequency between |1i and
|2i is ω12 = ω01 − η, where η  ω01 is the anharmonicity. We can see from Eq. (7.50)
that the anharmonic term at 4th order causes a negative shift of the energy, so that
η < 0. Thus, the level spacing shrinks as we go to higher energy. We can also see
from Eq. (7.50) that the size of the anharmonicity is controlled by the size of the
charging energy.

7.2.5 Superconducting transmon qubits

The Cooper pair box is the simplest example of a class of superconducting qubits
called charge qubits. They are referred to this way because the system is described
by the number of Cooper pairs residing on a superconducting island, and hence by
the charge on this island. A disadvantage of the Cooper pair box that hindered early
experiments is that it is very sensitive to charge noise from the environment. From
the point of view of the Cooper pair box Hamiltonian, charge noise causes the charge
offset, ng , to fluctuate. A plot of the energy spectrum of HCP B as a function of ng
reveals that the energies are very sensitive to the value of ng when EJ ∼ EC , and
thus large dephasing errors result if ng fluctuates. One way around this problem is
to increase the ratio EJ /EC . This reduces the relative importance of the charging
energy term in HCP B , and consequently, the energy spectrum becomes less sensitive
to ng . The ratio EJ /EC can be increased by adding an additional capacitor with
capacitance C 0 in parallel to the Josephson junction. Because the net capacitance
of two capacitors in parallel is the sum of the capacitances, this means that the
total capacitance will become C + C 0 . And since EC ∼ 1/C, this means that EC is
reduced, and thus EJ /EC is enhanced. The new qubit that results from adding the
capacitor C 0 is called a transmon qubit. As expected, it has proven to be much less
sensitive to charge noise, and as a result, it exhibits much longer coherence times (a
few orders of magnitude longer) compared to the Cooper pair box qubit.

However, according to the no-free-lunch theorem, something must be sacrificed

in order to gain this improvement in coherence times. This is indeed the case. As
we saw in the previous section above, when EJ  EC , hφ̂i stays close to zero, and
cos φ̂ can be replaced by its second-order Taylor series expansion. This leaves us
with a harmonic oscillator. But as we noted above, harmonic oscillators cannot be
used as qubits because it is impossible to drive transitions between the lowest two
energy levels without also driving transitions to higher levels. This is a consequence
of the fact that all the energies are equispaced. Thus, when EJ  EC , it becomes
hard to isolate two qubit levels from the infinite harmonic oscillator spectrum. Of

148
course, in the case of the transmon, we only approximately have a harmonic oscillator
spectrum. The higher-order terms in the Taylor series of cos φ̂ will create some an-
harmonicity η as shown in Eq. (7.50). The good news is that as EJ /EC is increased,
the energy levels become insensitive to ng at a rate that is exponential in EJ /EC ,
while the anharmonicity grows polynomially in EJ /EC . Thus, there is a range of
EJ /EC values where it is possible to make η reasonably large while still maintaining
the robustness to charge noise.

Tunable transmons–It is also possible to make a transmon whose spectrum can

be tuned by changing the strength of an applied magnetic field. This is done using
the so-called split transmon in which the single Josephson junction of the regular
transmon is replaced by two Josephson junctions placed in two parallel branches of
a circuit loop. The energies of the two junctions add together, but one has a relative
phase compared to the other if there is magnetic flux threading the loop:
 
e 0 .
HJ,split = −EJ1 cos φ̂ − EJ2 cos φ̂ + 2π Φ/Φ (7.51)

e is the magnetic flux threading the circuit loop, and Φ0 = h/2e is the magnetic
Here, Φ
flux quantum. Using trigonometric identities, this can be rewritten as
!v
u !
πΦe u (E − E )2 π e
Φ  
HJ,split = −(EJ1 + EJ2 ) cos t 1+
J1 J2
tan2
cos φ̂ + φ0 ,
Φ0 (EJ1 + EJ2 )2 Φ0
(7.52)
where φ0 is a constant phase shift defined by the equation
   
e 0 = EJ1 − EJ2 tan π Φ/Φ
tan φ0 + π Φ/Φ e 0 . (7.53)
EJ1 + EJ2
The important point here is that Eq. (7.52) looks like the Josephson energy of a single
junction, but where now the Josephson energy is a function of the applied magnetic
flux. Thus, the split transmon device behaves like a regular transmon, but where the
energy spectrum can be tuned with an applied magnetic field. This tunability plays
an important role in many two-qubit gate designs, as we discuss below.

7.2.6 Superconducting transmission line resonators

Qubit readout and control are typically done by coupling the qubit to a transmission
line resonator. Such a resonator is usually in the form of a coplanar waveguide,

149
which is the two-dimensional equivalent of a coaxial cable. It consists of three long,
narrow “rails” of superconductor lying parallel to one another on the surface of a
substrate, separated from one another by about 10 mm. The outer two rails are
connected to ground and are called ground planes. The middle rail is broken in two
places separated by a distance of about d = 1 cm. This effectively creates an open
lumped-element circuit between the breaks, with a linear inductance per unit length
` along the rails, and a capacitance per unit length c between the middle rail and the
outer ground plane rails. The coplanar waveguide can thus be modeled as a series of
infinitesimal LC circuits. The classical Hamiltonian is given by
Z d  2 
q (∂x Φ)2
H= dx + . (7.54)
0 2c 2`

Here, q is the charge density along the waveguide, while ∂x Φ is the flux density.
Aftering performing an integration by parts on the second term in H, Hamilton’s
equations give
1 1
∂t q = ∂x2 Φ, ∂t Φ = q. (7.55)
` c
Differentiating the second equation by t and plugging the result into the first equa-
tion, we see that Φ satisfies a wave equation:

∂t2 Φ = v 2 ∂x2 Φ, (7.56)

√
where the effective velocity is v = 1/ c`. Since we have an open circuit, we want
to focus on solutions with open boundary conditions: ∂x Φ x=0 = ∂x Φ x=d = 0. The
general solution is
∞
√ X
Φ(x, t) = 2 ξn (t) cos(kn x), (7.57)
n=0

where kn = nπ/d, and ξn (t) obeys the harmonic oscillator equation ξ¨n = −ωn2 ξn with
ωn = vkn . Plugging this solution into H and performing the x integration yields
cd X ˙2
H= (ξ + ωn2 ξn2 ). (7.58)
2 n n

Thus, we see that each mode of frequency ωn behaves like a simple harmonic oscilla-
tor. We can then follow the same quantization procedure we used for the electromag-
netic field, where we upgrade ξn and ξ˙n to quantum operators that can be expressed

150
in terms of raising and lowering operators a†n and an :
r r
~ ˙ ~ωn
ξˆn = −i (an − a†n ), ξˆn = (an + a†n ). (7.59)
2ωn cd 2cd
In fact, we can think of this as quantizing the actual electric and magnetic fields that
live between the rails of the transmission line. Recall from Faraday’s law that the
time derivative of the flux is the voltage (both of which are now quantum operators):
˙
V̂ = ∂t Φ̂, so that the quantized electric field associated with mode n is Ên ∼ ξˆn ∼
an + a†n . Similarly, the quantized magnetic field is B̂n ∼ ξˆn ∼ an − a†n . The ladder
operators obey the commutation relations

[an , a†m ] = δnm . (7.60)

In terms of these operators, the Hamiltonian becomes

X  1

†
H=~ ωn an an + . (7.61)
n
2

The transmission line resonator can thus be described as a collection of quantum

harmonic oscillators with the discrete frequencies ωn . Note that these frequencies
depend on the capacitance and inductance per unit length, and on the length of the
middle rail of the coplanar waveguide:
nπ
ωn = √ . (7.62)
c`d
For d on the order of 1 cm and for typical values of c and `, these frequencies are on
the order of GHz, i.e., these modes are in the microwave regime.

7.2.7 Superconducting qubit coupled to resonator

Next let’s see how to couple a superconducting qubit like the Cooper pair box or
transmon to a transmission line resonator. This is done by placing the qubit (i.e.,
Josephson junction and capacitor in the case of the transmon) between one of the
outer ground rails and the middle rail of the transmission line. The qubit should
be placed at a position where the electric field of the resonator has an antinode to
maximize the coupling. The locations of the antinodes depend on ωn , so first we
have to decide which cavity mode we want to couple to. Typically the device is set
up so that the qubit couples strongly to only one of the modes. Let’s call this the

151
frequency of this mode ωr , where the subscript r stands for resonator. The coupling
comes from the standard coupling between an electric dipole moment and an electric
field:
U (r) = −p · E(r), (7.63)
where r is the position of the qubit. Note that this is the same interaction we used to
derive the Jaynes-Cummings model of an atom coupled to an electric field. Since the
typical size of the resonator is 1 cm, if we assume that ωr is one of the lowest modes
in the cavity, then the wavelength of this mode will also be of order 1 cm. This is
much larger than the size of the qubit, which is on the order of microns. Thus, the
electric field is approximately constant across the qubit when the latter is located at
an antinode. The dipole moment p is proportional to the charge, Q̂, which in turn is
proportional to the Cooper pair number operator n̂ ∼ b + b† . Thus, the interaction
can be written as
Hint = ~g(b + b† )(a + a† ), (7.64)
where the coupling constant g depends on the strength of the electric field at r
and on the relative orientation of the field polarization to the dipole. As we did
in the Jaynes-Cummings model, we can apply the rotating-wave approximation and
simplify the interaction to an excitation number-conserving interaction:

Hint = ~g(ba† + b† a). (7.65)

We can then write the total qubit-cavity Hamiltonian as

Xh p
i
H = ~ωr a† a + j |jihj| + ~g j + 1 |jihj + 1| a† + |j + 1ihj| a . (7.66)
j

This is a Jaynes-Cummings Hamiltonian. The only difference compared to our earlier

discussion of the JC model is that here, the qubit system has more than two levels.
Aside from this, the physics we previously uncovered for the JC model applies here
too. In particular, the spectrum divides up into pairs of states labeled by excitation
number, photons can be repeatedly absorbed and emitted by the qubit in a manner
that closely resembles Rabi oscillations, and the qubit state can be measured using
dispersive readout. For the latter, recall that the dispersive regime is when the
coupling is much smaller than the qubit-cavity detuning: ~g  1 − 0 − ~ωr . In this
case, a canonical transformation can be performed that diagonalizes the Hamiltonian,
yielding an effective qubit-cavity interaction in which the cavity frequency depends
on the state of the qubit (see Sec. 6.3.2). The state of the qubit can then be measured
by reflecting microwaves off the cavity to determine its resonance frequency.

152
7.2.8 Driving a qubit in a cavity
We can perform quantum operations on the qubit by applying a time-dependent
voltage to the cavity:
V̂d (t) = Ω(t) cos(ω01 t)(a + a† ), (7.67)
where ω01 = (1 − 0 )/~ is the transition frequency of the qubit, and Ω(t) is an
arbitrary pulse envelope. If we are in the dispersive regime where the resonator
frequency ωr is far detuned from ω01 , then V̂d (t) will (approximately) not create
photons in the cavity. However, it will cause transitions between the qubit states.
To see how this works, let’s simplify things as much as possible and just focus on the
lowest two levels of the qubit system (the logical states). In this case, the qubit-cavity
Hamiltonian is just the usual two-level JC model that we are familiar with:

H = ~ωr a† a + ~ω01 σ † σ + ~g(σa† + σ † a), (7.68)

where σ = |0ih1|. Recall that to derive the effective Hamiltonian  g in the dispersive

regime, we need to apply the canonical transformation T = exp ∆ (σa − σ † a) :
†

 
† † g2 g2
Hef f = T HT = ~ωr a a + ~ ω01 − σ † σ + ~ σz a† a, (7.69)
∆ ∆

where ∆ = ωr − ω01 . We can apply the same transformation to the applied voltage:
n hg io
† † † † †
T V̂d (t)T = Ω(t) cos(ω01 t) a + a + (σa − σ a), a + a + ...
n ∆ o
g g
= Ω(t) cos(ω01 t) a + a† − σ − σ † + . . .
n ∆ o∆
g
= Ω(t) cos(ω01 t) a + a† − σx + . . . (7.70)
∆
We see that we end up with a driving term for the qubit:
e cos(ω01 t)σx ,
Hd (t) = Ω(t) (7.71)

e
with an effective driving envelope given by Ω(t) = − ∆g Ω(t). This is just like the
driving terms we considered when we talked about controlling arbitrary two-level
systems to generate single-qubit gates. Thus, we see that we can use transmission
line resonators to both readout and control superconducting qubits.

153
7.2.9 Two-qubit coupling and gates
The two most common physical mechanisms for coupling a pair of superconducting
qubits is to use either a direct, capacitive coupling (from the point of view of an
effective circuit, the two qubits are connected to opposite sides of a capacitor) or a
coupling mediated by a transmission line resonator. Each approach has merits and
drawbacks, and each can be used to generate a variety of different two-qubit entan-
gling gates. Many of these approaches require tunable qubits like the split transmon
and/or tunable resonators. Here, we briefly discuss some of the most popular ap-
proaches/gates, focusing on the case of transmon-transmon coupling.

Qubit-qubit “exchange” gates: Exchange-like interactions between two qubits can

be obtained either through direct capacitive coupling or with a resonator-mediated
coupling. In the case of the direct capacitive coupling, the charge degree of freedom of
one qubit couples to the voltage operator of the other. Assuming that the two qubit
frequencies are close to each other, we can employ the rotating-wave approximation
to arrive at the Hamiltonian

H = Hq1 + Hq2 + ~J(b†1 b2 + b1 b†2 ), (7.72)

where Hqi = ~ωqi b†i bi − E12Ci (bi − b†i )4 is the Hamiltonian of the ith transmon. The
effective coupling constant is given by
 1/4
2EC1 EC2 EJ1 EJ2
~J = , (7.73)
ECc 4EC1 EC2

where ECc is the charging energy of the coupling capacitor, and EJi , ECi are the
Josephson and charging energies of the individual transmons. Keeping only the
lowest two levels of each transmon, setting the frequencies equal (ωq1 = ωq2 ), and
going to the interaction picture, we find
~J
H 0 = ~J(σ1† σ2 + σ1 σ2† ) = (σ1x σ2x + σ1y σ2y ) . (7.74)
2
We recognize this as the operator that swaps the states of the two qubits. This is why
this is referred to as an “exchange” interaction. However, it should be kept in mind
that this is not a real exchange interaction in the
√ sense that it does not originate
from Pauli’s exclusion principle. An entangling iSWAP gate can be generated by
tuning the qubits into resonance, waiting a time t = π/(4J), and then tuning them

154
 √
out of resonance. The iSWAP gate has the form
 
1 0 0 0
√  0 √12 √i 0
iSWAP = eiπ/8(XX+Y Y ) = 
0 √i
2 . (7.75)
2
√1
2
0
0 0 0 1
√
Note that this gate is very different from the SWAP gate, which is also entangling.
We can see this by computing√ the Makhlin invariants: G1 = 1/4, G2 = 1, which
are distinct from those of SWAP. They are also distinct from the invariants for
the CNOT and identity gates. Also notice that we can multiply two of these gates
together to form the gate iSWAP,
 
1 0 0 0
0 0 i 0
iSWAP =  
0 i 0 0  , (7.76)
0 0 0 1

which, unlike the SWAP gate, is also entangling. It has Makhlin invariants G1 = 0,
G2 = −1. The extra factor of i in the off-diagonal entries of iSWAP compared to
those
√ of SWAP make all the difference. Finally, we point out that by combining two
iSWAP gates together with a certain set of single-qubit gates, one can create a
CNOT gate.

We can also get an exchange-like interaction in the case where the two transmons
are coupled to a common transmission line resonator. Now the Hamiltonian in the
lab frame is
X 2
†
H = Hq1 + Hq2 + ~ωr a a + ~gi (a† bi + ab†i ). (7.77)
i=1

This is like the Jaynes-Cummings Hamiltonian, except now we have more than
one qubit coupled to the cavity. In this case, it is called the Tavis-Cummings
model. Like we did for the Jaynes-Cummings model, in the dispersive regime
(gi  ωhr − ωqi ) and for ωiq1 ≈ ωq2 , we can perform a canonical transformation
P gi † †
T = exp i ∆i (a bi − abi ) and the rotating-wave approximation to obtain the ef-

155
fective Hamiltonian

Hef f = T HT † ≈ H e q1 + H
e q2 + ~J(b†1 b2 + b1 b†2 ) + ~ωr a† a
X ~ X ECi gi gj †  † 
2
ECi gi2 † † †
− 2~ 2
a ab b
i i + b b
i i b b
j i + b i j ,
b (7.78)
i=1
∆ i 2 i6=j
∆ i ∆j

where ∆i = ωqi −ωr . The tildes on the single-qubit terms He qi reflect the fact that the
qubit resonance frequencies are modified by the interaction. The effective “exchange”
coupling is now given by  
g1 g2 1 1
J= + . (7.79)
2 ∆1 ∆2
The first term in the second line of Eq. (7.78) is similar to the dispersive interaction
we found in the JC model, only now there is one such interaction per qubit. The
last term in Eq. (7.78) is a higher-order type of exchange interaction
√ that is much
smaller than J and thus usually neglected. We can then generate iSWAP gates as
we discussed in the context of a direct capacitive coupling.

Flux-tuned phase gates:

Another class of gates that can be realized either with a direct capacitive coupling
or with a resonator-mediated coupling are flux-tuned phase gates. Here, the idea is
to take advantage of the coupling between the |11i and |02i two-qubit states. Note
that this coupling is allowed because the excitation number is conserved, that is, both
states live in the two-excitation subspace of Hilbert space. The strength of the effec-
tive qubit-qubit interaction depends on the qubit frequencies. When the qubit-qubit
interaction is turned on, the two states hybridize such that the two-qubit spectrum
has two energy eigenstates that are linear combinations of |11i and |02i. For certain
values of the qubit frequencies ωqi , the interaction is strongest, and the hybridization
is most pronounced. On the other hand, as we tune the qubit frequencies away from
these special values, the hybridization goes away, and |11i and |02i become energy
eigenstates. Thus, we can effectively turn on the interaction by tuning one or both
qubits using an applied magnetic field of appropriate strength. If the interaction is
turned on for the right amount of time, the |11i state acquires a -1 phase, generating
a two-qubit CZ gate. Compared to other approaches, this method generates fast
(10s of nanoseconds), high-fidelity entangling gates. One of the challenges with this
approach though is that some technical issues can arise when trying to create fast
flux pulses.

156
 All-microwave gates:

Entangling gates can also be generated without having to tune any device el-
ements. This has the advantage of introducing less noise to the qubits, because
generally the more tunability a system has, the more sensitive it is to environmental
noise. For example, the flux-tunable split transmon is more prone to errors caused by
magnetic field fluctuations compared to the regular, fixed-frequency transmon. The
way to generate entanglement without tuning device components is to use shaped
microwave pulses that drive one or both qubits while a fixed qubit-qubit interaction
is on. Several methods for doing this have been proposed and demonstrated. Here,
we’ll focus on one of the most popular ones, namely the cross-resonance gate.

We start with two transmons interacting under the Hamiltonian

H = ~ωq1 σ1† σ1 + ~ωq2 σ2† σ2 + ~J(σ1† σ2 + σ1 σ2† ). (7.80)

We are working in the two-level approximation for simplicity. As shown above, the
interaction J can come from either a direct capacitive interaction or from a resonator-
mediated one. Diagonalizing Eq. (7.80), we find that the qubit frequencies effectively
become
eq1 ≈ ωq1 + J 2 /∆12 ,
ω eq2 ≈ ωq2 − J 2 /∆12 ,
ω (7.81)
where ∆12 = ωq1 − ωq2 . Now suppose that we drive the first qubit (for instance
through its own readout resonator). In the new basis, the driving term in the Hamil-
tonian becomes  
J
~Ω(t) cos(ωd t) σ1x + σ1z σ2x , (7.82)
∆12
to leading order in J/∆12 . The second term performs an x rotation on the second
eq2 , then the
qubit that depends on the state of the first qubit. If we choose ωd = ω
second qubit will be resonantly driven, and a conditional entangling gate results.
This gate is locally equivalent to CNOT. This type of entangling gate is commonly
used by IBM and other groups working on transmon systems.

7.3 Trapped ions

Trapped ion quantum processors are an advanced platform. This is the first example
of atomic qubits that we encounter. Atomic qubits are distinct from the systems
we discussed so far (spins in quantum dots and superconducting qubits) in a few

157
ways: they are naturally occurring systems, meaning that for a fixed species they
are identical to each other; they are naturally occurring; and they have long coherence
times, as they do not live in a ‘dirty’ solid-state environment. On the other hand,
they are harder to confine and trap, and device integration is not as straightforward.

7.3.1 Trapping and cooling

Unlike neutral atoms, ions are electrically charged, which means that we can use
electric fields to trap them. This, however, is not trivial, due to Earnshaw’s theorem,
which states that a charged particle cannot be confined in all three dimensions by an
electrostatic potential. To overcome this, we can use time-dependent electric fields
combined with electrostatic potentials. One way of doing this is through a Paul trap,
which was developed by Wolfgang Paul, who won the Nobel prize (1/2 of it together
with Dehmelt) in Physics in 1989. The particular version of the Paul trap we will
focus on consists of four parallel cylinders arranged in a way that their cross-section
forms a square (each cylinder lies at a vertex of the square). There is an alternating
rf voltage between two of the cylinders located diagonally from each other, while the
other two are grounded. There is a static voltage at the ends of the trap, which
keeps the ions confined in the direction along the cylinder axes, which we will call
z. To obtain an effective static Hamiltonian, we time-average the time-dependent
rf potential. For carefully chosen parameters, such that there is a regime of stable
points in the trap, the resulting Hamiltonian is given by a 3D harmonic oscillator
added to the Coulomb repulsion between the ions:

XN   XN X
m 2 2 2 2 2 2 p~2i e2
H= ωx Xi + ωy Yi + ωz Zi + 2 + (7.83)
2 m ~ ~
i=1 i=1 j<i 4π0 |Ri − Rj |

Here, R~ i = (Xi , Yi , Zi ) is the position of the ith ion, and p~i is its momentum.
We will first consider a single ion in the trap, which means that the potential is
that of a 3D harmonic oscillator. The ion has an internal (electronic) structure, such
that the state can be written as |Ψi = |internali ⊗ |motionali. When the ions are
loaded into the trap—a process that happens by ionizing atoms and heating them
near the trap—they are in an excited motional state. It is therefore important to
cool them down prior to attempting to use them for quantum information processing.

The cooling happens in two stages, by combining Doppler cooling and sideband
cooling. We will use a simplified two-level model for the internal structure of the
ion to give a qualitative idea of the cooling processes. We denote the transition

158
frequency between those two states as ωeg . In Doppler cooling, there is a cooling
laser beam of frequency ωL and momentum pL . The ions that move in the opposite
direction (toward) the laser beam see a larger laser frequency due to the Doppler
effect. Assuming that the Doppler-shifted frequency is ωeg , the photon can be ab-
sorbed by the ion. Assuming for simplicity that the initial state of the ion is a plane
wave with momentum p~I while the internal state is the ground state, we can write
the initial energy of the ion as p2I /2m, where m is the ion mass. After absorption
of the laser photon, the momentum of the ion is p~0 = p~I − p~L and its energy is
p02
2m
+ ωeg . At some point, the ion will spontaneously emit a photon in a random
direction and drop to its lower internal state. Let’s assume for concreteness that the
original direction of the ion’s motion was x and that the photon was emitted along
y. Then the momentum
and energy of the ion
will be, respectively, (p~0 − p~y , p~y , 0)
and (p0 − py )2 + p2y /2m = p02 + 2p2y − 2py p0 /2m. Plugging in p0 and making the
reasonable assumption that pL , py  pI , we see that the kinetic energy of the ion is
reduced.

Doppler cooling will get rid of significant kinetic energy, but when the ion’s mo-
mentum drops to low values it will not be able to cool it further. Cooling to the
motional ground state is achieved through sideband cooling, which is a special case of
optical pumping. Optical pumping is a technique that uses states outside the relevant
Hilbert subspace (in our case the qubit subspace) in order to purify a state that is
mixed within that subspace. We will explain it here with a simple example before
discussing it in the context of trapped ions. For simplicity, we will focus on a three-
level system, where the two lower levels are the qubit states and the upper level has
a nonzero dipole matrix element with both of the qubit states. We will assume there
is some difference in the two transitions, either the dipole orientation is different or
the frequency is different (or both). This will allow us to use a time-dependent field
to selectively drive only one transition. The excited state can relax to either state
(let’s assume with equal probability for simplicity, even though that is not a require-
ment) through spontaneous emission of photons. We take the density operator to
have the form diag(1/2, 1/2, 0), where the first two entries correspond to the qubit
states and the last to the excited state. Moving the population of the second state
to the excited state with a π pulse, the density operator becomes diag(1/2, 0, 1/2),
while after spontaneous emission we have diag(1/2 + 1/4, 1/4, 0). Clearly, repeating
this process will asymptotically move all the population from the second qubit state
to the first, using the excited level as a mediator. We will now see how this principle
applies to the specific case of a single trapped ion with a two-level internal structure
coupled to its motional degree of freedom. To do that, we will need to introduce the

159
Hamiltonian of the system in the presence of a driving laser field.

To see how sideband cooling works, let us first consider the electric field of a
driving laser taken to propagate along the z axis with electric field along x and
magnitude

E(z, t) = E0 cos(ωt − kz). (7.84)

We assume dipole coupling −d~ · E ~ and consider for simplicity an ion with a two-
level structure (|gi and |ei), with Hamiltonian H0 = ωeg |ei he| and dipole d~ =
x̂d0 (|gi he| + |ei hg|). Defining the Rabi frequency Ω0 = −d0 E0 /2, we can express
the interaction Hamiltonian as


Hint = Ω0 ei(ωt−kz) + e−i(ωt−kz) (|gi he| + |ei hg|) . (7.85)

Going to the interaction picture and performing the rotating wave approximation
(i.e., dropping rapidly rotating terms e±i(ω+ωeg )t ) we get


H̃int ' Ω0 ei[(ω−ωeg )t−kz] |ei hg| + e−i[(ω−ωeg )t−kz] |gi he| , (7.86)

where the tilde indicates the rotating frame. From this we see that we could have just
kept two terms out of the four appearing in Eq.(7.85). This is what we will do, and
we’ll write the operators in the lab frame. Expanding the kz term in the exponent,
we will only keep the first-order term, as we are considering the Lamb-Dicke regime
kz0  1, where z0 is the characteristic length of the SHO potential. Then we can
express
(0) (1)
Hint = Hint + Hint , (7.87)

where
(0)
Hint ' Ωeiωt |ei hg| + Ωe−iωt |gi he| (7.88)

drives purely internal dynamics (a tensor product with the identity is understood in
this equation) and

(1) 2πz0 iωt
2πz0 −iωt

Hint ' −iΩ e |ei hg| a + a† + iΩ e |gi he| a + a† (7.89)
λ λ
drives transitions between different internal and motional states simultaneously.

160
 |e, 0i
|e, 1i
|e, 0i
|e, 2i
|e, 1i
|e, 0i |e, 3i …
|e, 2i
|e, 1i
|e, 0i |e, 3i
|e, 2i
|e, 1i
|e, 3i
|e, 2i
|e, 3i |g, 0i
|g, 1i
|g, 0i |g, 2i
…
|g, 1i |g, 3i
|g, 0i
|g, 2i
|g, 1i
|g, 0i |g, 3i
|g, 2i
|g, 1i
|g, 3iof a trapped ion. Levels corresponding to
Figure 6: Energy levels and transitions
|g, 2i
two internal states |gi and |ei and one normal mode with frequency ωn are shown.
The sideband transitions (orange
|g, 3i and blue) are used to cool the ion, as explained in
the text.

These transitions are called sidebands, and are shown in Fig. 6. Notice that the
sideband transitions have different frequencies ωeg ± ωn from the internal transitions,
and they can thus be resolved. Therefore, if we drive resonantly with transition
|gi ⊗ |1i → |ei ⊗ |0i, which will also drive all |gi ⊗ |ni → |ei ⊗ |n − 1i, with n > 0,
spontaneous emission will eventually take all the population to |gi ⊗ |0i, which is
not driven. This is of course an example of optical pumping.

7.3.2 Qubit: definition, readout, single-qubit gates

The internal structure of atoms (and thus ions) is very rich: there are many different
pairs of energy levels, characterized by their energy splittings, coherence times, selec-
tion rules, couplings to other states, etc. All these features determine their suitability
as qubits. A common feature of all trapped ion qubits that are presently explored is
the availability of a cycling transition. This means that there exists a transition from
one qubit level to a level outside the qubit subspace which does not decay to any
other state. This implies that the transition can be repeatedly driven without the

161
driving leading to optical pumping to any other states. As a result, by continuously
driving the transition, multiple emitted photons can be measured to enable robust
readout. This, in addition to the requirement of a simple level structure, leads to the
choice of atoms with two electrons in the outer shell, such that when ionized they
will have one electron in the outer shell. These are the alkaline earths (Be, Ca, Ba,
etc) and the lanthanides (e.g., Yb).

Examples of qubits in trapped ions are spin qubits, fine-structure qubits, hyper-
fine qubits, and optical qubits. Each of these choices has its own advantages and
drawbacks. Hyperfine qubits are hybrid states of electron and nuclear spin. One can
choose one state from each of two different F manifolds as qubits in a way that they
are insensitive to magnetic field noise. These are sometimes called clock qubits, as
they are also used in atomic clocks due to their noise insensitivity. Hyperfine qubits
have a splitting on the order of GHz. They can be controlled directly using mi-
crowave fields or optically via intermediate excited levels in a Λ system configuration
(sometimes called Raman control). The coherence times are long (seconds to many
hundreds of seconds with dynamical decoupling). Optical qubits consist of one state
from the ground state manifold (that is s type) and one from an excited state man-
ifold (p, d, f ). Typically a d state is used because it is longer-lived than p. This is
because the dipole moment with s is zero, and the quadrupolar moment provides the
lowest nonzero matrix element. However, the coupling is still sufficient so that the
transition can be driven efficiently with a stronger laser. The large splitting of optical
qubits leads to faster direct control (gates). Another advantage of these qubits is
that typical frequencies are in the infrared, where lasers are stable and compatible
with on-chip integration. An obvious disadvantage is that these qubits suffer from
spontaneous emission, such that T2 is limited by T1 (on the order of 1-10 sec).

In trapped ion quantum information processing (and in atomic physics more

broadly) unitary gates between two states are often implemented through lasers
that couple both these states to a common excited state. Labeling the two lower
(qubit) states by |0i and |1i and the excited state by |ei, the resulting Λ system is
characterized by an interaction Hamiltonian

Hc = Ωa eiωa t |0i he| + Ωb eiωb t |1i he| + H.c., (7.90)

where we have performed RWA, and a and b label the lasers driving transitions
|0i ↔ |ei and |1i ↔ |ei, respectively. We define the detuning of each transition
as ∆a = ωa − e + 0 and ∆b = ωb − e + 1 , where j is the energy of state |ji.

162
Taking ∆a = ∆b is called the two-photon resonance. When the ∆k are taken to be
much larger than the Ωk , the excited state is only virtually excited, and we can use a
canonical transformation (or, equivalently, adiabatic elimination of the excited state)
to obtain an effective Hamiltonian in the qubit subspace. This will give us terms
proportional to Ω2k /∆k for the diagonal terms and Ωa Ωb /∆k for the off-diagonals.

7.3.3 Normal modes of multiple ions in a trap

Let’s now consider how we can couple two or more ions together and create entangling
gates. To understand the coupling mechanism, we first need to derive the normal
modes that describe the motion of the ions in the trap. The normal modes can
be obtained by starting from the potential term in the multi-ion Hamiltonian from
Eq. (7.83):

X
N
m
XN X
β
~ i) =
V (R ωx2 Xi2 + ωy2 Yi2 + ωz2 Zi2 + , (7.91)
2 ~i − R
|R ~ j|
i=1 i=1 j<i

where β = e2 /4π0 , and R~ i = (Xi , Yi , Zi ) is the position of the ith ion. We assume
that we have cooled the ions down enough using Doppler and sideband cooling that
they do not stray far from their equilibrium positions. This allows us to perform a
Taylor series expansion of V (R~ i ) around the equilibrium positions. The second-order
terms in this expansion carry information about the normal modes.

Let’s illustrate this procedure for N = 2 ions. First of all, we need to find the
equilibrium positions. These correspond to the values of R~ i that minimize V (R
~ i ). If
we focus on a linear Paul trap, then we can assume that the ions are displaced from
each other by some distance 2` along the z direction, which we take to be the long
dimension of the trap. Taking the origin of the coordinate system to be the halfway
point between the two ions, the equilibrium positions are then (0, 0, ±`). What is `?
We find this by minimizing V (`):

β
V (`) = mωz2 `2 + . (7.92)
2`
Taking the derivative, setting the result to zero, and solving for `, we find
 
d 2 2 β β β
mωz ` + = 2mωz2 ` − 2 = 0 ⇒ `3 = . (7.93)
d` 2` 2` 4mωz2

163
We see that the equilibrium positions of the ions move toward each other as ωz in-
creases. This makes sense since the potential becomes more confining in this case,
pushing the ions together. The ions stay separated of course, because of their strong
Coulomb repulsion, which is captured by β.

Now that we know the equilibrium positions of the ions, we need to perform a
Taylor series expansion. Consider small displacements ~ri away from equilibrium:
~ 1 = (0, 0, −`) + ~r1 ,
R ~ 2 = (0, 0, `) + ~r2 .
R (7.94)

Let’s expand the Coulomb part of the potential first since that’s harder:
β β β
= =p
~1 − R
|R ~ 2| | − 2`ẑ + ~r1 − ~r2 | (x1 − x2 )2 + (y1 − y2 )2 + (−2` + z1 − z2 )2
 
β (x1 − x2 )2 (y1 − y2 )2 z1 − z2 3 (z1 − z2 )2 (z1 − z2 )2
≈ 1− − + + −
2` 8`2 8`2 2` 8 `2 8`2
β β(x1 − x2 )2 β(y1 − y2 )2 β(z1 − z2 ) β(z1 − z2 )2
= − − + + . (7.95)
2` 16`3 16`3 4`2 8`3
Here, we assumed that |~r1 − ~r2 |/`  1, and we kept all terms up through second
order. The expansion of the trap potential to second order is exact:
m 2 2
[ω (x + x22 ) + ωy2 (y12 + y22 ) + ωz2 (z12 + z22 ) − 2ωz2 `(z1 − z2 ) + 2ωz2 `2 ]. (7.96)
2 x 1
Combining Eqs. (7.95) and (7.96), we have

3β
V (~ri ) = + Vx (x1 , x2 ) + Vy (y1 , y2 ) + Vz (z1 , z2 ), (7.97)
4`
where
mωx2 2 β
Vx (x1 , x2 ) = (x1 + x22 ) − (x1 − x2 )2 ,
2 16`3
mωy2 2 β
Vy (y1 , y2 ) = (y1 + y22 ) − 3
(y1 − y2 )2 ,
2 16`
mωz2 2 β
Vz (z1 , z2 ) = (z1 + z22 ) + 3 (z1 − z2 )2 . (7.98)
2 8`
We see that the potential breaks up into three pieces, each of which depends on only
a pair of coordinates—one from each ion. This allows us to separately examine each

164
piece and extract normal modes from it. Note that the term linear in z1 − z2 from
Eq. (7.95) cancelled with the linear term from Eq. (7.96).

Let’s look at the Vx part first. The goal is to find two different linear combinations
of x1 and x2 , call them xa and xb , such that Vx can be rewritten as Vx = m2 [ωxa 2 2
xa +
2 2
ωxb xb ] for some constants ωxa and ωxb . xa and xb would then correspond to two
normal modes with frequencies ωxa and ωxb . To facilitate this rewriting of Vx , let’s
first express this part of the potential as
  
2 2

A B x1
Vx (x1 , x2 ) = A(x1 + x2 ) + 2Bx1 x2 = x1 x2 , (7.99)
B A x2

where
mωx2 β β
A= − 3
, B= . (7.100)
2 16` 16`3
We see that, essentially, all we need to do is diagonalize the matrix A1 + Bσx . Since
the first term is proportional to the identity, this matrix
√ has the same eigenvectors as
σx , and we already know that these are (1, ±1)/ 2. The corresponding eigenvalues
are A ± B. Therefore, we have
     

1 1 1 A+B 0 1 1 1 x1
Vx (x1 , x2 ) = x1 x2 √ √
2 1 −1 0 A − B 2 1 −1 x2
   
1
A+B 0 1 x1 + x2
= √ x1 + x2 x1 − x2 √
2 0 A−B 2 x1 − x2
 2  2
x1 + x 2 x1 − x2
= (A + B) √ + (A − B) √ . (7.101)
2 2
√
We see that the first mode is xa = (x1 + x2 )/ 2, which describes the center-of-mass
motion in the x direction. This mode has frequency
r r r
2√ 2 mωx2
ωxa = A+B = = ωx . (7.102)
m m 2
We see that when the two ions are in this mode, they oscillate synchronously back in
forth in the x direction with a frequency equal to the trap frequency. This is called
√
a transverse center-of-mass mode. The second mode we find is xb = (x1 − x2 )/ 2,

165
with frequency
r r r
2√ 2 m 2 p
ωxb = A−B = (ωx − ωz2 ) = ωx2 − ωz2 . (7.103)
m m 2

In the second-to-last step, we used A − B = mωx2 /2 − β/8`3 together with `3 =

β/4mωz2 to simplify the result. When the two ions are in this mode, they oscillate
in the x direction antisynchronously, i.e., they are perfectly out of phase. This is
called a transverse tilt mode. These oscillations occur with a frequency that is less
than that of the transverse center-of-mass mode by an amount that depends on the
confinement energy in the z direction (long axis of the trap).

Now consider the Vy part of the potential. From Eq. (7.98), we see that this part
has the same form as Vx , which is not surprising since there is complete symmetry
between these two directions. We can therefore borrow the same analysis from above
to immediately write down the two normal modes and their frequencies:
√
ya = (y1 + y2 )/ 2, ωya = ωy ,
√ q
yb = (y1 − y2 )/ 2, ωyb = ωy2 − ωz2 . (7.104)

These are again called transverse center-of-mass and tilt modes, respectively.

The normal modes from Vz are different owing to the asymmetry of the trap. This
is reflected in Eq. (7.98), where we see that the coefficient of the contribution from
the Coulomb term is different compared to those in Vx and Vy . We again re-arrange
the terms to make Vz look like two simple harmonic oscillators:

mωz2 2 β
Vz (z1 , z2 ) = (z1 + z22 ) + 3 (z1 − z2 )2
2 8`
2 2
= A(z1 + z2 ) + 2Bz1 z2
 2  2
z1 + z2 z1 − z2
= (A + B) √ + (A − B) √ , (7.105)
2 2
2
where now A = mω 2
z
+ 8`β3 , B = − 8`β3 . The mode za = z1 + z2 is the axial center-of-
mass mode, while zb = z1 − z2 is called an axial breathing mode, since it describes
a motion in which the two ions periodically move toward each other and away from

166
each other. The frequencies of these two modes are
r
2√
ωza = A + B = ωz (7.106)
m
r
2√ √
ωzb = A − B = 3ωz . (7.107)
m
A similar procedure can be followed for N > 2 ions to obtain 3N normal modes.
We first solve for the equilibrium positions of the ions. If we assume the ions line
up along the z axis, then we need to solve the system of equations ∂V (R~ i )∂Zi = 0.
This is a set of N equations for the N unknowns Zi . We then expand the potential
around the equilibrium positions to second order and re-organize the terms into
simple harmonic oscillators to obtain the normal modes and their frequencies. We
can then quantize each normal mode in the usual way to arrive at the following
multi-ion quantum Hamiltonian:

X
3N
ωeg X
N
H0 = ~ ωn (a†n an + 1/2) + ~ σjz , (7.108)
n=1
2 j=1

where the first term describes the motional degrees of freedom, and the second term
describes the internal degrees of freedom of each ion. Here, we only include two
internal energy levels |gi and |ei, with energy splitting ~ωeg . Typically, the internal
energy splitting is large compared to the normal mode energies ~ωn , and so we can
think of each internal energy level for a given ion as being split into 3N closely spaced
levels.

7.3.4 Coupling trapped ions and generating entangling gates

How can we couple multiple ions together using the normal modes? The way this is
usually done is to apply lasers that are tuned close to a sideband transition so that
only one normal mode is (potentially) excited. The laser driving is described by the
Hamiltonian
X ~Ωj h † i[ηj /√N (a+a† )−ωj t] i
N
Hint = σj e + H.c. . (7.109)
j=1
2

Here, Ωj is the Rabi frequency associated with the jth laser acting on the jth ion,
ωj is the laser frequency, and ηj is the Lamb-Dicke parameter discussed above. a†
and a are the raising and lowering operators for the normal mode associated with
the sideband transition that is being driven. Here, we assume that this is the axial

167
center-of-mass mode, since this is the one that is usually used for coupling ions
√ due
to the fact that it is the lowest-frequency mode (its frequency is ωz ). The N in
the exponent arises because the mass p associated with† this mode is the total mass of
all the ions, N m, and so zc.o.m. = ~/N mωz (a + a ). The other q factors in zc.o.m.
√
aside from N enter into the Lamb-Dicke parameter: ηj = 2π λj
~
2mωz
. We called
the above Hamiltonian Hint because it describes an interaction between each ion and
the normal mode. Physically, we can understand this coupling mechanism between
internal and motional degrees of freedom by noting that when an ion absorbs pho-
tons from the laser, its electrons get excited to a higher internal state, while at the
same time the ion as a whole picks up some momentum from the absorbed photons,
changing its motional state. Thus, changes to the internal and motional states are
intrinsically linked.

The fact that each ion that is driven by a laser talks to the same normal mode
provides a mechanism to couple different ions together in a manner similar to the
Tavis-Cummings Hamiltonian. One ion can exchange energy between its internal
state and the normal mode in the trap. A second ion can then pick up this energy
from the same normal mode and transfer it to its internal degrees of freedom. In this
way, the two ions effectively talk to each other through the normal mode. Notice that
we can choose which ions participate in the interaction by choosing whether or not
to drive them. If we don’t want to couple some of the ions to the rest, then we just
switch off their lasers, which corresponds to setting the respective Rabi frequencies
to zero: Ωj = 0. This ability to select how many ions interact with each other at the
same time is a powerful feature of the trapped ion system.

Cirac-Zoller gate—Let’s see how we can use the normal mode-mediated coupling
described above to implement a two-qubit entangling gate. Starting from Eq. (7.109),
we allow for additional internal states, and we expand the exponent to first order
in the Lamb-Dicke parameter. Focusing the laser to a particular sideband (as in
Fig. 6) and
P going to the interaction picture, we arrive at the effective Hamiltonian
Hef f = jq Hj,q , where

ηj Ω h i
Hj,q = √ |eq ij hg| a + |gij heq | a† . (7.110)
2 N
The index j labels the different ions, and q labels different internal excited states.
Here, we have assumed that the Rabi frequencies are the same for both ions. Each of
these Hamiltonians is just like a Jaynes-Cummings Hamiltonian. Note however that

168
there is an important physical difference here compared to the Jaynes-Cummings
problem. The latter normally refers to the case of a qubit-cavity system in which
the qubit is nearly resonant with a cavity mode. In contrast, here the normal mode
frequency is orders of magnitude smaller than the optical frequency of the ion. This
large energy mismatch is compensated by the applied laser, enabling the exchange
of energy between the internal degrees of freedom and the motional ones.

The Cirac-Zoller gate works by assuming that there are two excited states |e0 i and
|e1 i such that transitions between |gi and |e0 i are driven for one choice of the laser’s
polarization, while transitions between |gi and |e1 i require the opposite polarization.
Thus, only one of these two possible transitions is driven at any given time. The
computational states are |gi and |e0 i, while |e1 i is an auxiliary state that is helpful
for creating entanglement between the computational√states, as we will see. Now
suppose that laser j is turned on for a time t = kπ N /Ωηj . The corresponding
evolution operator for ion j is
 i
k,q kπ h
Uj = exp −i |eq ij hg| a + H.c. . (7.111)
2

This evolution operator preserves excitation number. It drives Rabi oscillations

between states within subspaces of fixed excitation number. In the subspace with
excitation number 1, the states evolve according to

|gij |1i → cos(kπ/2) |gij |1i − i sin(kπ/2) |eq ij |0i

|eq i |0i → cos(kπ/2) |eq ij |0i − i sin(kπ/2) |gij |1i . (7.112)

In each term, the first ket is the internal state of the ion, while the second is the mo-
tional state. Thus, we have Rabi flopping between the internal and motional states,
analogous to what we saw in the atom-cavity Jaynes-Cummings problem.

We can use the evolution in Eq. (7.111) to create an entangling gate between ions
a and b in three steps: Ua,b = Ua1,0 Ub2,1 Ua1,0 . In the first step,
√ we apply a laser to
qubit a to drive the |gia ↔ |e0 ia transition for a time t = π N /Ωηa . As we can see
from Eq. (7.112), this causes the following transitions:

|gia |1i → −i |e0 ia |0i , |e0 ia |0i → −i |gia |1i , (7.113)

which exchange an excitation between the internal and motional degrees of freedom.
In the second step, we drive the |gib ↔ |e1 ib transition of qubit b for time t =

169
 √
2π N /Ωηb . This puts a phase on the state |gib |1i:

1 |gib |1i → − |gib |1i . (7.114)

The third step is a repeat of the first, and so it again exchanges an excitation between
internal and motional degrees of freedom. To understand why these three steps
produce an entangling gate, it helps to look at the evolution of each computational
basis state under these steps:

Ua1,0 Ub2,1 Ua1,0

|gia |gib |0i → |gia |gib |0i → |gia |gib |0i → |gia |gib |0i
|gia |e0 ib |0i → |gia |e0 ib |0i → |gia |e0 ib |0i → |gia |e0 ib |0i
|e0 ia |gib |0i → −i |gia |gib |1i → i |gia |gib |1i → |e0 ia |gib |0i
|e0 ia |e0 ib |0i → −i |gia |e0 ib |1i → −i |gia |e0 ib |1i → − |e0 ia |e0 ib |0i
(7.115)

We see that the net result in the computational subspace is a CZ gate. Notice that
although the normal mode is populated during the gate, this is only temporary, as
the normal mode is restored to its ground state at the end of the gate in all cases.

The Cirac-Zoller gate can be generalized to a multi-qubit gate involving more

than two ions. For example, the five-step sequence Ua1,0 Ub1,1 Uc2,1 Ub1,1 Ua1,0 implements
a controlled-controlled-Z (CCZ) gate on qubits a, b, and c. This gate performs a Z
gate on one qubit only if the other two qubits are both in the |e0 i state.

Molmer-Sorenson gate—A disadvantage of the Cirac-Zoller gate is that the nor-

mal mode is populated during the gate. Ideally the normal mode is completely
restored to its ground state at the end of the gate. However, in reality, vibrational
noise and control imperfections will cause some population to remain in the normal
mode, leading to gate errors. Molmer and Sorenson proposed another type of entan-
gling gate for trapped ions that only excites normal modes virtually, i.e., population
never actually moves onto the normal modes. Virtual excitations can happen if the
laser is detuned slightly away from the sideband transition. This is just like we saw in
the Rabi problem, where the probability to end up in the excited state is reduced as
the detuning ∆ is increased. Even though the excited state is not populated (much),
the presence of the excited state can still impart phases on the qubit states. This is
the basic idea behind the Molmer-Sorenson gate. This is similar to the Λ structure
based single-qubit gates described above, where the excited state is only virtually
populated.

170
 To understand this gate in more detail, focus on the case of two ions, both of
which are driven by their respective lasers. Both lasers are tuned close to the same
sideband transition in each ion. However, there is a small positive detuning δ for ion
a, and a small negative detuning −δ for ion b.

The system is still described by the Hamiltonian Hint in Eq. (7.109). However,
this time, we perform perturbation theory up to second order assuming ηΩ is a small
parameter. By keeping terms up to this order, one can show that the Hamiltonian
drives transitions between |ggni and |eeni with an effective Rabi frequency given by

X heen| Hint |mi hm| Hint |ggni (ηΩ)2

e=2
Ω ≈− , (7.116)
m
Eggn + ~ωα(m) − Em 2N (ν − δ)

where ωα(m) is the frequency of the laser that drives transitions to and from the
intermediate state m, i.e., α(m) = a or b. ν is the frequency of the normal mode.
We see that Ωe does not depend on the excitation level n of the normal mode. It
is thus insensitive to vibrational noise, unlike the Cirac-Zoller gate. By
√ rotating
partway between |ggni and |eeni, a gate that is locally equivalent to a SWAP is
implemented.

An important difficulty with both the Cirac-Zoller and Molmer-Sorenson gates

(and in general with coupling trapped ions through vibrational modes) is that the
coupling strength (and hence the entangling rate) decreases with the number of ions
N . This is evident in both Eq. (7.110) and Eq. (7.116). This is due to the fact that
the axial center-of-mass mode becomes heavier with increasing N , making it harder
to excite the mode. Thus, entangling gates become difficult when there are too many
ions in the trap. This is one of the reasons that scaling up trapped ion systems is
non-trivial.

171
8 Quantum networks
8.1 Quantum communications
In addition to quantum computers, quantum technologies include communication
networks. These can be thought of as a quantum version of the internet. The
motivation for developing such networks lies in two key principles of quantum me-
chanics, which provide inherent security to quantum information. These are the
fact that measurement disturbs a quantum state, which means that an eavesdropper
can be detected, and the fact that unknown quantum states cannot be copied (the
‘no-cloning theorem’). As a result, information transmitted over quantum networks
is secure. This can be used for various cryptographic protocols, such as quantum
key distribution, multi-partite protocols such as secret sharing, as well as for se-
cure ‘cloud’ access of a remote quantum computer (also known as blind quantum
computing).

8.1.1 Quantum key distribution

An important result in cryptography is that the encryption of a message is secure
if a key (which is as large as the message to be encrypted) is used only once. The
key is a random string of bits (zeros and ones), which is added bitwise mod 2 to the
message, which has been converted to a bit string as well. The encrypted message
can then be sent over a public unsecure channel. Only those in possession of the key
can decrypt the message. This protocol is called a one-time pad.

Quantum states of single photons can be used to transmit information from a

sender to a receiver, who coordinate in a specific way to generate a secret key. To
do so, first we have to understand that we can perform measurements in bases other
than the computational basis. This is done by doing a unitary gate that maps the
desired basis to the computational basis and then measuring in the computational
basis. We assume that the sender has the ability to prepare states in different bases
and the receiver has the capability of measuring in different bases.

Here we will present the simplest QKD protocol, known as BB84. The main idea
is that the sender encodes the classical bit 0 in either of two non-orthogonal states
(e.g., |0i and |+i) and bit 1 in one of their orthogonal states (|1i and |−i respec-
tively). Thus, if the sender starts with a large number of photons and prepares each
in one of these four states, the resulting collection of photons can be thought of as
encoding a bit string (assuming all the photons are ordered in some way). All the

172
photons are then sent to the receiver. The receiver measures each photon in either
of the two bases (|0i / |1i or |+i / |−i) and records the outcomes. Once the receiver
is done measuring, both the sender and receiver publicly announce which basis they
used for each qubit. For each qubit for which they chose the same basis, their mea-
surement outcomes agree, and these bits can be used to form the shared key. On the
other hand, the measurement outcomes for those qubits for which the sender and
receiver chose different bases are random with respect to each other, and so these
outcomes will not be used as part of the key.

Can someone intercept the information about the key while the photons are be-
ing sent? Suppose an eavesdropper measures a photon and then resends it (or sends
a copy of what they measured). In the process, they disturb the state whenever
the measurement basis they choose does not coincide with that of the sender. As a
result, when this happens, 50% of the time the sender and receiver will not agree on
the state they prepared and measured, respectively, even in cases where they chose
the same basis. In these cases, they know something is wrong, because they know
they should get perfect agreement when the bases are the same. In this way, if they
compare results, they can tell there is an eavesdropper and abort the QKD proce-
dure. During this process of checking for an eavesdropper, the sender and receiver
publicly announce the states they measured/prepared in. Therefore these bits have
to be excluded from the key. As a result, they sacrifice some bits, and thus the length
of the key is shorter. If they find there is no eavesdropper, the remaining qubits,
where the bases of preparation and measurement agree, are then used to generate
the shared secret key (called the sifted key).

Another QKD protocol is based on shared Bell pairs of the two parties. In
general, there are various cryptography protocols where sharing of entanglement
among parties at remote locations is important. Without going into details here, we
will consider this one of the goals of quantum networks and discuss physical systems
that are currently being explored to achieve this goal.

8.2 Quantum repeaters

All the above protocols require sending photons from one location to another, re-
mote location. Many protocols also require sending entangled photons to different
parties and sharing of entanglement. Photons can be sent in free space, but it is
more practical to send them using optical fibers, which is what is used in classical
communications (e.g., for the internet). Light travels in specially designed cables.

173
These are designed to guide the light and minimize absorption, and the frequencies
are chosen to have low attenuation. Inevitably, however, light is eventually absorbed,
in which case the encoded information is lost. For today’s fibers, this happens with
probability 4% after 1 km. To characterize the loss rate of a fiber, people often refer
to the attenuation length, which is the distance a photon can travel successfully with
probability 1/e ≈ 0.368. For a fiber with 4% loss per kilometer, the attenuation
length is about 24.5 km. In classical communications, repeaters are used to amplify
the signal (e.g., measure and resend). In quantum networks, this is not possible
because measurement disturbs the quantum state. Instead, a more clever approach
is required.

8.2.1 Entanglement swapping

Just as we can measure single qubits in different bases, we can also measure two
qubits in a two-qubit basis other than the computational basis. In particular, we are
interested in measuring in the Bell basis. It is not hard to see how this can be done,
once we realize that a Hadamard one on qubit and a CNOT map each computational
basis state to a distinct Bell state. As a result, we can use this transformation (rather
its inverse) along with measurement in the computational basis, to measure in the
Bell basis.

Measuring in the Bell basis allows us to spread entanglement between qubits

that have not interacted directly. To see how this can be done through an ex-
ample, consider two Bell pairs. For concreteness, let us take them both to be
|Φ+ i = √12 (|00i + |11i). The total four-qubit state is |Ψi = |Φ+ i12 ⊗|Φ+ i34 . Imagine
that qubits 1 and 4 are located far from each other. Qubits 2 and 3 are sent to a
common location and measured in the Bell basis. It is not hard to see that for any
of the four possible outcomes, the post-measurement state is two Bell pairs, only
that now the pairs are between qubits 2 and 3 (obviously, since they were measured
in the Bell basis) and between qubits 1 and 4! This means that entanglement was
swapped and the two qubits that never came close to each other (qubits 1 and 4)
became entangled! For example, if qubits 2 and 3 were found to be in state |Ψ+ i,
then the post-measurement state of qubits 1 and 4 is also |Ψ+ i14 . This shows how
we can overcome the difficulty of fiber attenuation and spread entanglement over
longer distances: We chop the distance into smaller segments and perform a series
of entanglement swapping operations.

174
8.3 Quantum networks based on defect qubits
The most common paradigm for implementing quantum networks is having stations
with matter qubits that can be entangled with photons, and sending the photons
to intermediate stations to undergo Bell measurements, which effect entanglement
swapping. The matter qubits of course need to be coherent while the photons travel
and undergo Bell measurements. There are several physical platforms that are ex-
plored for quantum networks. One of the most promising is based on optically active
defects in solids. Out of these, most of the work to date has been done on the
nitrogen-vacancy (NV) center in diamond. Below we focus exclusively on this de-
fect, but note that alternative qubits are also being explored. Popular alternatives
include the silicon-vacancy in diamond, the silicon-carbon divacancy in silicon car-
bide, and the silicon vacancy in silicon carbide. Each system has its advantages and
drawbacks in terms of coherence time, optical efficiency, emission frequency, etc. A
detailed comparison is outside the scope of these notes.

8.3.1 The NV center in diamond

As we discussed in the context of spins in quantum dots, electrons in solids occupy
extended states, and the allowed energies form continuous bands. When an electron
is localized in a solid such that the energy of the state falls within a bandgap, the
state has atomic-level features: it is localized in space and sharp in energy. In addi-
tion to forming quantum dots, other ways to have localized, sharp levels in a solid is
through defects. A simple example is a missing atom in a crystal. Another example
is a substitutional atom of a different element from the host crystal, or a combination
of the two. Such a defect can trap one or more electrons and confine them in a small
region around the defect on the order of angstroms or nanometers.

The most well known example of a deep defect (deep means that it is deep into
the bandgap) is the negatively charged NV center in diamond. This consists of a
substitutional nitrogen in a carbon site with the nearest neighboring carbon also
missing and nothing taking its place (i.e., a vacancy). Carbon has four valence
electrons, so there are three electrons associated with the vacancy. Nitrogen has
five valence electrons, two of which are unpaired. This is a total of 5 electrons. An
additional electron (originating from a different defect in the lattice) is commonly
trapped at the defect site as well, so that 6 electrons are associated with the negatively
charged NV center (which for simplicity we’ll be simply referring to as the NV center
in what follows). Out of these electrons, four pair up, and two are important in
forming the relevant states. These two electrons pair up in a total spin S = 1 state

175
in the ground state. Since the ground state has S = 1, there are three sublevels
(m = 0, ±1) in the ground state manifold. The m ± 1 states are split from the
m = 0 state due to spin-spin interactions in the crystal. Their splitting, also known
as zero-field splitting, is 2.88 GHz. An attractive property of the NV center is its
very long coherence time, even at room temperature. Specifically, T2∗ is a few µs,
while dynamical decoupling gives a T2 of a few ms. For isotopically purified diamond
T2∗ exceeds 100 µs, while at low temperatures with dynamical decoupling T2 can get
close to seconds in such samples.
Another attractive feature of the NV center is that it is optically active, and its
optical transition can be used for the generation of spin-photon entanglement, an
important ingredient in quantum networks. Because lasers (electric fields) do not
interact with the spin, all optically active states also have S = 1. The transition
frequency is approximately 1.9 eV. This transition is also used for readout (it is
almost a cycling transition, similarly to trapped ions). There are also metastable
singlet states (S = 0) in the defect, and they are important in polarizing (initializing)
the qubit. This is done through a combination of spin-orbit interaction, which couples
the S = 1 and S = 0 states, and vibrational modes, which take away the excess
energy. This process initializes the system into the m = 0 state. The other states
can then be populated through the use of microwave fields.

8.3.2 Nuclear spin qubits as a memory register

An additional attractive feature of NV centers in diamond is the availability of nu-
clear spin qubits in the crystal. As we’ve seen before, a dense bath of nuclear spins is
generally undesirable, as its primary consequence is the decoherence of the electronic
qubit. However, diamond only has ∼ 1% nonzero nuclear spin abundance from 13 C
atoms. These nuclear spins, which have I = 1/2, interact with the electronic spin
of the NV center via hyperfine dipolar interactions. The relatively small number of
these nuclear spins, along with their long coherence times, make them interesting
candidates for memory qubits in quantum repeaters.

The interaction between the NV center and a given nuclear spin has the general
form I · A · S, where A is a tensor (i.e., matrix), and I and S are the spin operators
for the nucleus and electron, respectively. We define a qubit out of the three NV
center levels by limiting the NV center population to only states |0i and |−1i (or
|1i). Notice that in this case, the nuclear spin only interacts with the NV center
when the latter is in the state with |m| = 1. We also use the fact that the Zeeman
splittings of the two systems differ by orders of magnitude, which allows us to use

176
a transformation that leads to an effective Hamiltonian that is diagonal in the NV
spin, but allows rotations of the nuclear spins. Under this approximation, the general
form of the hyperfine interaction between the NV spin and a given 13 C is

Hj,int = |0i h0| ⊗ H0 + |1i h1| ⊗ Hj1 , (8.1)

where H0 = ωn Iz and Hj1 = ωn Iz + Aj Ix . Note that the hyperfine coupling depends

on the particular nuclear spin through its distance from the NV center and the angle
between this separation vector and the NV symmetry axis. The total interaction
Hamiltonian with all the nuclear spins will of course be the sum of all such Hamilto-
nians (sum over j). At first sight, one may think that the fact that these interactions
are always on (not switchable) is a deal-breaker for using them in quantum informa-
tion protocols. Fortunately, there is a way to use them despite this. The key idea is
to use dynamical decoupling sequences. As we discussed in Chapter X, these gener-
ically decouple the qubit from its interactions with other systems. However, this is
not always the case. When the inter-pulse time delay of the dynamical decoupling
sequence τ is chosen to satisfy a resonance condition determined by A, the interac-
tion is not affected by the pulses, such that a given nuclear spin can be targeted to
undergo a conditional-X gate controlled by the NV center. Given these conditional
gates, which can also be combined to create SWAP type gates, the nuclear spins can
be used as a quantum memory.

8.3.3 Quantum repeaters based on NV centers

The NV spin–photon entanglement is not a deterministic process due to emission into
unwanted states (specifically, photonic states that are accompanied by the excitation
of vibrational modes). However, we can attempt to create the target Bell state
between the NV spin and the photon multiple times until success. When this is
achieved, the photon will be sent to the intermediate node to be measured in the
Bell basis jointly with a photon arriving from a different node. Note that the Bell
measurement itself is probabilistic for photons. (This is a general statement that has
nothing to do with the NV or how the photon was generated. It is a statement about
linear optics.) Once this Bell measurement succeeds, a SWAP gate is implemented
between the NV and a nearby 13 C following the procedure described above. Then, the
NV center is available again for another spin–photon entanglement generation. This
allows entanglement to be created with another system that is located in a different
location, in the opposite direction from where the first photon was sent. When an
entanglement link is generated by repeating the procedure described above, a Bell
measurement between the NV center and the nuclear spin can be implemented (using

177
 (a) Bell state measurement (b)

Communication qubit
(NV e- spin)

Memory qubit (13C

Location A nuclear spin) Location B Location A Location B

Bell state measurement

(c) (d)

Location A Location B Location C Location A Location B Location C

Figure 7: Repeater protocol based on NV centers. (a) Entanglement is established

between two remote NV centers, one in location A and one in location B. (b) The
state is swapped to the nuclear spins, so that the NVs are available to become
entangled with another photon and, in turn, a different NV. (c) The protocol is
repeated between the NV center in location B and one in location C. (d) Once the
NVs in B and C are entangled, a measurement in the Bell basis between the NV
and the nuclear spin in location B implements entanglement swapping so that the
nuclear spin in location A and the NV center in location C become entangled.

the CNOT gates described above). This will achieve entanglement swapping and
allow the creation of an entangled state between the two (further separated) nodes
from where the photons originated. This whole procedure is shown schematically in
Fig. 7.

178