Lectures on Quantum Tensor Networks

a pathway to modern diagrammatic reasoning

Jacob Biamonte
Deep Quantum Laboratory
Skolkovo Institute of Science and Technology
3 Nobel Street
Moscow Russia 143026

e-mail: [email protected]
web: http://quantum.skoltech.ru

Version current as of December 21, 2019.

I would be grateful if you email me about any mistakes and typos that you find.

Preprint available at arXiv:XXX

Contents

Foreword 1

I From Tensors to Networks 2

1 Penrose Graphical Notation for Tensor Networks 3

2 Penrose Wire Bending Duality 12

3 Bending Density Operator Wires 15

3.1 Bipartite Matrix Operations 19
3.2 Vectorization of Matrices 20

4 Worked Examples 23

5 Problems 25

6 Further Reading 26

II Matrix Product States 28

7 The Diagrammatic SVD 28

8 Matrix Product Factorization of States 32

9 Numerical Tensor Network Algorithms and Packages 38

9.1 Online Resources Describing Tensor Network Software Implementations 39
9.2 Open Source Tensor Network Software Packages 39

10 Problems 40

III Boolean Tensor Networks 42

11 Introduction 42

12 Overview of the Chapter 42

–i–
13 Quantum Legos: a tensor tool box 50
13.1 COPY-tensors: the “diagonal” 50
13.2 XOR-tensors: the “addition” 51
13.3 Quantum AND-state tensors: Boolean universality 53
13.3.1 Summary of the XOR-algebra on tensors 55
13.3.2 Merging COPY-tensors by node equivalence 58
13.3.3 Algebras on valence-3 tensors 60

14 Examples of Boolean Tensor Network States 62

15 Discussion 65

16 Problems 65

IV Symmetries and Stabilizer Tensor Theory 70

17 Introduction 70

18 Symmetries in Tensor Network States 70

19 The Interaction of Networks Comprised of δ ijk and ⊕ijk 75

20 Stabilizer Tensor Theory 81

21 The Clifford Group 83

22 Stabiliser Tensor Theory 84

23 Boolean Stabilizer States 87

24 Problems 89

V Tensor Networks and Entanglement Invariants 91

25 Generating Invariants for n-party Density Operators 103

26 Some Symmetries of ρ 107

27 Polynomial Invariants of Symmetric States 109

28 From Binary Forms to Qubit States 111

29 Problems 113

– ii –
VI Tensor Networks for Open Quantum System 115

30 Kraus / Operator-Sum Representation 115

31 System-Environment / Stinespring Representation 116

32 Louiville-Superoperator Representation 118

33 Choi-Matrix Representation 119

34 Process Matrix Representation 122

35 Transforming between representations 124

35.1 Transformations between the Choi-matrix and superoperator representations 124
35.2 Transformations to the superoperator representation 125
35.3 Transformations to the Choi-matrix representation 126
35.4 Transformations to the Kraus Representation 127
35.5 Transformations to the system-environment representation 129

36 Applications 131
36.1 Vectorization of composite systems 131
36.2 Composing superopators 134
36.3 Matrix operations as superoperators 135
36.4 Reduced superoperators 136
36.5 Ancilla Assisted Process Tomography 137
36.6 Average Gate Fidelity 139
36.7 Entanglement Fidelity 142

VII Counting Solutions by Tensor Contraction 144

37 Introduction 144

38 Returning to Boolean Quantum States 144

39 Returning to Stabilizer Tensor Theory 146

40 Elementary Theorems of Tensor Contraction 147

41 A 3-fold way 149

VIII Appendix 155

A Algebra on Quantum States 155

– iii –
B XOR-algebra 157

C The Minimization Method of Karnaugh 158

D Tensors and Tensor Products 160

– iv –
Foreword
Tensor network methods represent a collection of techniques to understand and
reason about multi-linear maps which have found particular use in applications
to quantum information processing. These methods form the backbone of tensor
network contraction algorithms to model physical systems and are used in the ab-
stract tensor languages to represent channels, maps, states and processes appearing
across quantum information science.

In these chapters—which were complied based on years of teaching—we out-

line the building blocks needed to understand the elementary properties of tensor
networks, the associated mathematical techniques and the diagrammatic reasoning
language. This book is intended to be self-contained, and accessible to graduate
students studying quantum information and quantum computation. It is hoped
that advanced readers will let this book serve as an active research reference in the
many areas that this multi-disciplinary topic has its roots.

Indeed, the topic of tensor networks touches on a number of topics yet the
vast majority of writing on the topic is much more specific and often limited to be
accessible by a narrow community. This book attempts to broadly cover the foun-
dations of tensor network theory as it applies generally across quantum information.

The aesthetically appealing development of tensor networks as a unifying lan-

guage across quantum information science has long been close to my own research
interests. I have conducted research on quantum and classical circuits, as it applies
to quantum computing. This research included developing methods to embed logic
functions into spin Hamiltonian ground states and the realization of quantum al-
gorithms by quantum circuits. The settings of both classical and quantum circuits
comes with a well known graphical language.

This gave me a base to spend time merging ideas from (i) modern tensor net-
works as they appear in condensed matter physics; (ii) quantum circuits and their
graphical language; (iii) aspects of categorical quantum mechanics as well as (iv)
the graphical language of digital circuits to create a common notation and to de-
velop and use rewrite rules that intersect these topics. The text should be accessible
to graduate students and of value to practicing researchers in the field of quantum
information science.

Sincerely and happy reading,

Jacob Biamonte—Moscow December 21, 2019

1
 PART

From Tensors to Networks I

Tensors are a mathematical concept that encapsulates and generalizes the idea of
multilinear maps, i.e. functions of multiple parameters that are linear with respect
to every parameter. A tensor network is simply a countable collection of tensors
connected by contractions. Tensor network methods is the term sometimes
given to the entire collection of associated tools, which are regularly employed in
modern quantum information science, condensed matter physics, mathematics and
computer science.
Tensor networks come with an intuitive graphical language that can be used
in formal reasoning and in proofs. This diagrammatic language found applications
in physics at least as early as the 1970s by Roger Penrose [1]. Tensor network
theory has recently seen many advancements and adaptations to different domains
of physics, mathematics and computer science. An important milestone was David
Deutsch’s use of the diagrammatic notation in quantum computing, developing the
quantum circuit (a.k.a. quantum computational networks as Deutsch would call
them) model [2]. Quantum circuits are a special class of tensor networks, in which
the arrangement of the tensors and their types are restricted. A related diagram-
matic language slightly before that is due to Richard Feynman [3]. The quantum
circuit model—now well over two decades old—is widely used to describe quan-
tum algorithms and their experimental implementations, to quantify the resources
they use (by e.g. counting the quantum gates required), to classify the entangling
properties and computational power of specific gate families, and more.
There is growing excitement concerning numerical algorithms that preform
tensor contractions. These algorithms are important in condensed matter physics
and beyond. There are many reviews and surveys devoted to this important
direction—see [4–17], as well as, Tensor Networks in a Nutshell, which I wrote
with Ville Bergholm [18]. Some of the best known applications of tensor networks
are 1D Matrix Product States (MPS), Tensor Trains (TT), Tree Tensor Networks
(TTN), the Multi-scale Entanglement Renormalization Ansatz (MERA), Projected
Entangled Pair States (PEPS)—which generalize matrix product states to higher
dimensions—and various other renormalization methods [5–8, 12, 15, 19]. The ex-
citement is based on the fact that certain classes of quantum systems can now be
simulated more efficiently, studied in greater detail, and this has opened new av-
enues for a greater understanding of certain physical systems. The concept is to
factor a quantum state ψ into various network structures, as follows.

TTN PEPS ...

... ...
MPS
...
... ...

These methods approximate a complicated quantum state using a tensor net-

work with a simplistic, regular structure—essentially applying lossy data compres-

2
sion that preserves the most salient properties of the quantum state. The method
is known to efficiently work for certain classes of ground and thermal states.
We assume that most readers will have a basic understanding of some quantum
theory, linear algebra and tensors.

1 Penrose Graphical Notation for Tensor Networks

We will present a variant of the graphical notation used by Penrose [1, 20, 21]
that we have tailored to describe quantum states and circuits by tensor networks
[22, 23]. We have merged these old ideas with three ingredients where and if they
apply: (i) modern tensor networks as they appear in condensed matter physics;
(ii) quantum circuits and their graphical language [24]; (iii) aspects of categorical
quantum mechanics [25] as well as (iv) the graphical language of digital circuits.
The output of this merger is an increasingly popular collection of ideas re-
lated to the application of tensor networks to quantum information and quantum
computation. The notation matches quantum circuit notation and the presenta-
tion should hopefully be approachable for a wide audience of modern quantum
information scientists. Indeed, the techniques do differ from any of the respective
ingredients we have mixed together; so any of the above listed communities—(i),
(ii), (iii), (iv)—should go away after reading these notes with new techniques.
Tensor Network Tensors can be thought of as indexed multi-dimensional arrays
of complex numbers with respect to a fixed standard basis.

Definition 1.1. Let X , Y, Z be finite-dimensional complex Hilbert spaces,

L(X , Y) is the space of bounded linear operators A : X → Y with L(X ) ≡
L(X , X ). Figure 1. “It now ceases
to be important to maintain a
For example, consider the Hilbert space X ∼
= Cd , where as is typical in modern distinction between upper and
quantum theory, we choose our standard basis to be the computational basis lower indices.” – Roger Pen-
rose, 1971 [1].
{|ii : i = 0, ..., d − 1}.

Then in Dirac notation a vector |vi ∈ X is a 1st-order tensor which can be expressed
in terms of its tenor components vi := hi|vi with respect to the standard basis as
|vi = i=0 vi |ii. Similarly one can represent linear operators on this Hilbert
Pd−1

space, A ∈ L(X ), as 2nd-order tensors with components Aij := hi| A |ji as A =

Pd−1
i,j=0 Aij |iihj|.
Hence, in Dirac notation the number of indices of a tensors components more
define what we called a tensors order. Vectors |vi ∈ X refer to tensors which only
have ket “|ii” basis elements, vectors in the dual vector space hu| ∈ X † (or more
typically denoted X ? or X ∗ ) refer to those with only bras “hi|”, and linear operators
on A ∈ L(X ) refer to tensors with a mixture of kets and bras in their component
decomposition.
Remark. Like Penrose, we use the word valence or order instead of rank when
referring to the number of indices on a tensor, since rank is used elsewhere.
A tensor with n indices up and m down is called a valence-(n, m) tensor and
sometimes a valence-k tensor for k = n + m.
Remark. The concurrent evaluation of all indices returns a complex number.
This is called total or full contraction.
–3–
 Remark. The idea of representing quantum states, operators and maps (etc.)
diagrammatically is credited to works by Penrose and is sometimes referred
to as Penrose graphical notation or string diagrams. Though Penrose unques-
tionably pioneered many of the applications and uses of the language and
deserves credit, Arthur Cayley developed much earlier variants of graphical
languages.
We mostly adhere to Penrose’s notation of representing states (vectors) and ef-
fects (dual-vectors) as triangles, linear operators as boxes, and scalars as diamonds,
as found in Illustration 1.1. Here each index corresponds to an open wire on the
diagram and so we may define higher order tensors with increasingly more wires.
The number of wires is then the order of the tensor, with each wire acting on a
separate vector space Xj .

Illustration 1.1 (Graphical depiction of elementary tensors) Non-zero scalars (d) are also repre-
sented as ‘blank’ on the page. We represent vectors (states) and dual-vectors (effects) as triangles, linear
operators as boxes, and scalars as diamonds, with each index of the tensor depicted as an open wire on the
diagram. The orientation of the wires determines the type of tensor, in our convention the open end of the
wires point to the left for vectors, right for dual-vectors, and both left and right for linear operators.

v v A Λ
(a) Vector |vi ∈ X (b) Dual-vector hv| ∈ X † (c) Linear Operator A ∈ L(X ) (d) Scalar λ ∈ C

v v A
Nn Nn
(e) Vector |vi ∈ i=1
Xi (f) Dual-vector hv| ∈ i=1
Xi† (g) Linear operator A :
Nn Nm
i=1
Xi → j=1
Xj

It is typical in quantum physics to think of a tensor as an indexed multi-array

of numbers.1 For instance, 1 Abstract index notation is a
mathematical notation for tensors
(a) (b) that uses indices as place holders
identifying space(s), rather than
their components in a particular
basis.

represent the tensor (a) ψi in the space Hi and (b) T ijk in the space

H i ⊗ Hj ⊗ Hk

respectively.
Remark (Diagram convention—top to bottom, or right to left). Open wires
pointing towards the top of the page, correspond to upper indices (bras),
open wires pointing towards the bottom of a page correspond to a lower in-
dices (kets). For ease of presentation, we will often rotate this convention 90
degrees counterclockwise.

–4–
 There are three specific tensors that (essentially) play the role of Kronecker’s
delta. These tensors allow for (i) tensor index contraction by diagrammatic connec-
tion, (ii) raising and lowering indices, and (iii) they give rise to a duality between
maps, states and linear maps in general. The bijection induced by bending wires,
is sometimes called Penrose Duality, after its inventor [1]. As in [1], these three
tensors are given diagrammatically as

By thinking of these tensors now in terms of components, e.g. δij = 1 − (i − j)2

for i, j = 0, 1, we note that

1= δji |ii hj| = (1.1)

X X
|ki hk| ,
ij k

h00| + h11| + · · · hnn| = δ ij hij| = (1.2)

X X
hkk| ,
ij k

|00i + |11i + · · · |nni = δ ij |iji = (1.3)

X X
|kki ,
ij k

where the identity map (a) corresponds to Equation (1.1), the cup (b) to (1.2) and
the cap (c) to (1.3). The relation between these three equations is given by bending
wires. In a basis, bending a wire corresponds to changing a bra to a ket, and vise
versa.
The contraction of two tensor indices diagrammatically amounts to joining
those indices with a single wire. Given tensors T ijk , Al n and B qm we form a
contraction by multiplying by δlj δqk resulting in the tensor

T ijk Ajn B km = Γi nm , (1.4)

def

where the tensor Γi nm is introduced per definition to simplify notation. As a linear

map, in quantum physics notation, this is typically expressed in equational form as

Γ= Γi nm |nmi hi| = (1.5)

X X
T ijk Ajn B km |nmi hi| .
inm ijknm

Connection to quantum computing notation

As mentioned, tensors are multilinear maps. They can be expanded in any given
basis, and expressed in terms of their components. In quantum information sci-
ence one often introduces a computational basis {|ki}k for each Hilbert space and
expands the tensors in it, using kets (| i) for vectors and bras (h |) for dual vectors:

T = (1.6)
X
T ijk |ii hjk| .
ijk

Here T ijk is understood not as abstract index notation but as the actual compo-
nents of the tensor in the computational basis. In practice there is little room for
confusion. The Einstein summation convention is rarely used in quantum informa-
tion science, hence we write the sum sign explicitly.

–5–
 So far we have explained how tensors are represented in tensor diagrams, and
what happens when wires are connected. The ideas are concluded by four exam-
ples; we urge the reader to work through the examples and check the results for
themselves.
The first example introduces a familiar structure from linear algebra in ten-
sor form. The next two examples come from quantum entanglement theory—see
connecting tensor networks with invariants [26, 27]. The fourth one showcases
quantum circuits, a subclass of tensor networks widely used in the field of quantum
information. The examples are chosen to illustrate properties of tensor networks
and should be self-contained.
Remark. We occasionally will work with equality up to a scalar when manip-
ulating tensor diagrams by hand. This is common and typically amounts to a
loss of (unit) normalization. In quantum theory, a global phase is undetectable
and call this working in the unit gauge: in tensor networks we sometimes work
in the scalar gauge, C/{0}. This amounts to mapping numbers picked up dur-
ing calculation as
C/{0} → 1
and representing the unit 1 as a blank on the page.
Example (The  tensor). A tensor is said to be fully antisymmetric if swapping
any pair of indices will change its sign: Aij = −Aji . The  tensor is used to
represent the fully antisymmetric Levi-Civita symbol, which in two dimensions
can be expressed as

00 = 11 = 0, 01 = −10 = 1. (1.7)

The  tensor can be used to compute the determinant of a matrix. In two

dimensions we have
det(S) = ij S i0 S j1 . (1.8)
Using this we obtain

S
= ·detS
S

as can be seen by labeling the wires in the diagram. In equational form this
is
ij S im S jn = det(S) mn . (1.9)
In terms of quantum mechanics,  corresponds to the two-qubit singlet state:
1 1
√ |i = √ (|01i − |10i). (1.10)
2 2
This quantum state is invariant under any transformation of the form U ⊗
U , where U is a 2 × 2 unitary, as it only gains an unphysical global phase
factor det(U ).
Example (Concurrence and entanglement). Given a two-qubit pure quantum
state |ψi, its concurrence C(ψ) = |C 0 (ψ)| is the absolute value of the following

–6–
tensor network expression [28]:

C'( ) =

Here ψ is the complex conjugate of ψ in the computational basis. The con-

currence is an entanglement monotone, a function from states to nonnegative
real numbers that measures how entangled the state is. |ψi is entangled if and
only if the concurrence is greater than zero.
Consider now what happens when we act on |ψi by an arbitrary local unitary
operation, i.e.
ψ = (U1 ⊗ U2 ) |ψi .
 0

Using the result of Example 2 we obtain

C ((U1 ⊗ U2 ) |ψi) = C(ψ)| det(U1 ) det(U2 )|. (1.11)

Due to the unitarity

| det U1 | = | det U2 | = 1,
which means that the value of the concurrence is invariant (i.e. does not
change) under local unitary transformations. This is to be expected, as local
unitaries cannot change the amount of entanglement in a quantum state.
More complicated invariants can also be expressed as tensor networks [26].
We will leave it to the reader to write the following network as an algebraic
expression:

τ'( ) =

If |ψi is a 3-qubit quantum state, τ (ψ) = 2|τ 0 (ψ)| represents the entanglement
invariant known as the 3-tangle [29]. It is possible to form invariants also with-
out using the epsilon tensor. For example, the following expression represents
the 3-qubit entanglement invariant known as the Kempe invariant [30]:

K(ψ) = ψ ijk ψ ilm ψ nlo ψ pjo ψ pqm ψ nqk . (1.12)

The studious reader would draw the equivalent tensor network.

Example (Quantum circuits). Quantum circuits are a restricted subclass of
tensor networks that is widely used in the field of quantum information. In
a quantum circuit diagram each horizontal wire represents the Hilbert space
associated with a quantum subsystem, typically a single qubit. “I learned very early the
The tensors attached to the wires represent unitary propagators acting on difference between knowing
those subsystems, and are called quantum gates. Additional symbols may be the name of something
used to denote measurements. The standard notation is described in [31]. The and knowing something.”
graphical language of quantum circuits was explored in detail in [23]. — Richard P. Feynman,
Here we will consider a simple quantum circuit that can generate entangled co-discover (with Norman
Margolus) of the CNOT-
–7– a.k.a. Feynman-gate.
Bell states. It consists of two tensors, a Hadamard gate (H) and a controlled
NOT gate (CNOT, denoted by the symbol inside the dashed region):

k j i
H
m l

The CNOT and Hadamard gates are defined as

CNOT = (1.13)
X
|a, a ⊕ biha, b| ,
ab
1 X
H= √ (−1)ab |aihb| , (1.14)
2 ab

where the addition in the CNOT is modulo 2. The reader should verify Addition modulo 2: 1 ⊕ 1 =
that acting on the quantum state |00i the above circuit yields the Bell state 0 ⊕ 0 = 0, 1 ⊕ 0 = 0 ⊕ 1 = 1.
√1 (|00i + |11i), and acting on |11i it yields the singlet state √1 (|01i − |10i).
2 2

Example (COPY and XOR tensors). One can view the CNOT gate itself as a
contraction of two order-three tensors [22, 23]:

The top tensor (• with three legs) is called the COPY tensor. It equals unity
when all the indices are assigned the same value (0 or 1), and vanishes other-
wise:
0 1
= = 1
0
0

1
1

Hence, COPY acts to copy the binary inputs 0 and 1:

COPY |0i = |0i |0i , (1.15a)

COPY |1i = |1i |1i . (1.15b)

The bottom tensor (⊕ with three legs) is called the parity or XOR tensor. It
equals unity when the index assignment contains an even number of 1s, and
vanishes otherwise:
0 1 1 0
= = = = 1
0

1
0

0
1

1
1
0

The XOR and COPY tensors are related via the Hadamard gate as

–8–
 H
1
−
2 =
H
H
(1.16)
Thus one can think of XOR as being a (scaled) copy operation in another
basis:
1
√ XOR |+i = |+i |+i , (1.17a)
2
1
√ XOR |−i = |−i |−i , (1.17b)
2
where |+i := H |0i and |−i := H |1i. In terms of components,

COPYij k = (1 − i)(1 − j)(1 − k) + ijk, (1.18a)

XOR qr
s = 1 − (q + r + s) + 2(qr + qs + sr) − 4qrs. (1.18b)

The CNOT gate is now obtained as the tensor contraction

COPYqmi XORrmj = CNOTqrij . (1.19)

X

The COPY and XOR tensors will be explored further in later examples and
have many convenient properties [23, 32, 33].
.tex
Example (Quantum circuits for cups and epsilon states). The quantum circuit
from Example 2 is typically used to generate entangled qubit pairs. For
instance, acting on the state |00i yields the familiar Bell state—as a ten-
sor network, this is equal to a normalized cup. Here we also show the
mathematical relationship the XOR and COPY tensors have with the cup
(|+i := H |0i = √12 (|0i + |1i)):

0 H +
1 1
=−
2
=−
2
0 0

+ = 0 =
(1.20)

Similarly, one can use the circuit (2) to generate the epsilon state. Let us
denote the Pauli matrices by X := |0i h1| + |1i h0|, Y := −i |0i h1| + i |1i h0|
and Z := |0i h0| − |1i h1|. The Z gate commutes with the COPY tensor, and
the X or NOT gate commutes with XOR. Commuting those tensors to the
right hand side, allows us to apply Eq. (1.20). Making use of the Pauli algebra

–9–
 identity ZX = iY , one recovers the epsilon state:

1 H 1 +
1 1
=−
2
=−
2
1 0 0

1 1
=− =−
2
2
(1.21)
Remark (Graphical tensor calculus [1]). In practice, tensor networks contain
an increasing number of tensors, making it difficult to form expressions using
(inherently one-dimensional) equations. The two-dimensional diagrammatic
depiction of tensor networks can simplify such expressions, reduce calcula-
tions and often depict internal structure that can lend insight into physical
phenomena. The applications of these methods to contract and reason about
quantum circuits appears in [23].
Equational identities will also be cast into diagrammatic form. For example,
if Γ is totally symmetric in any arm or leg exchange, then we could adopt the
convention to draw it as a circle (b). The tensor in (c) illustrates the equation
Abdcg B acf .

(a) (b) (c)

Remark (Graphical Trace). The trace in the graphical calculus is given by

connecting wires to close loops [1]. Diagram (a) below represents the trace Aii .
Diagram (b) represents the trace B iqiq .

i j
(a) (b) (c)
A B C
i q p k
i

Partial trace means contracting only some of the outputs with their corre-
sponding inputs, such as with the tensor C ijkpk shown in diagram (c).

Example (Partial trace). The following is an early rewrite representing entan-

gled pairs due to Penrose [20].

The diagram on the left represents the partial trace of |ψihψ| over the second

– 10 –
 subsystem. Readers can prove that this equality follows by interpreting the
bent wires as cups and caps, and the crossing wires as SWAPs.
Example (Partial trace of Bell states). Continuing on from Example 2, if we
choose |ψi = |∪i, i.e. |ψi is an unnormalized Bell state, we obtain

= =

In general, a quantum state on n-spins is a single tensor (such as a box or

triangle) with n protruding legs. Several methods exist to factor such states into
elementary building blocks. For example, it has been shown that the tensor network
for each quantum state in the Boolean class of states arises in turn from the classical
decomposition of f into fundamental gates. This is made precise by Theorem 38.1.
Remark (Notation). We use B to denote the Boolean field, given by an element
of the set {0, 1}. Numbers in Bn are then n-long Boolean bit strings. Here
and elsewhere, bold font x is shorthand for bit strings x1 , x2 , . . . , xn .

Definition 1.2 (The class of Boolean quantum states — covered in detail in § III).
Let
f : Bn → B, (1.22)
be any switching function. Then

ψB = f (x) |xi , (1.23)

X

is an arbitrary representative in the class of Boolean states. In this fashion,

every Boolean function save the constant zero function, gives rise to a quantum
state. Conversely, every quantum state written in a local basis with amplitude
coefficients taking binary values in {0, 1} gives rise to a Boolean function. This
defines the so called, class of Boolean quantum states [22].

We will state and prove the following as Theorem 38.1:

A tensor network representing a Boolean quantum state is determined from the

classical network description of the corresponding function.

This was proven in [22], where the quantum tensor networks are found by letting
each classical gate act on a linear space and from changing the composition of
functions, to the contraction of tensors.
Remark (Quantum Legos: Boolean tensor Networks [22]). An example of a
Boolean tensor is the AND-tensor introduced in [22]. This tensor stores the
truth table for the local AND function in its superposition.

∧ijk = |x, yi |x ∧ yi = (|00i + |01i + |10i) |0i + |11i |1i . (1.24)

X

x,y=0,1

Under Penrose duality, if we bend a wire to raise the index labeled k we arrive

– 11 –
 at
∧ij k = (|00i + |01i + |10i) h0| + |11i h1| . (1.25)

Remark (Tensor juxtaposition). When two or more disconnected tensors appear

in the same diagram they are multiplied together using the tensor product. In
quantum physics notation, they would have a tensor product sign ⊗ between
them. In the abstract index notation the tensor product sign is omitted.
Tensors can be freely moved past each other. This is sometimes called planar
deformation.

B A
= = A B
A B

From the diagram above, using equations we have

(1 ⊗ B)(A ⊗ 1) = A ⊗ B = (A ⊗ 1)(1 ⊗ B), (1.26)

where we make use of the wire also playing the role of the identity tensor 1—
detailed in Section 2. As we shall soon see, wires are allowed to cross tensor
symbols and other wires, as long as the wire endpoints are not changed. This
is one reason why tensor diagrams are often simpler to deal with than their
algebraic counterparts.

R Q = Q R

In the diagram above we did not label the wires, since it is an arbitrary
assignment. If we did, we could for example denote it as Qdegb Rf ac .

2 Penrose Wire Bending Duality

Bending and exchanging wires. Let us consider n Hilbert spaces H⊗n . These
spaces are essentially equivalent to each other. We will then consider the SWAP
operator which exchanges the position of two Hilbert spaces in a composite system,
moreover it will exchange the ith space with the i+1th space, for 1 ≤ i ≤ n−1. This
generates the permutation group with order that divides n!, with the generators
given diagrammatically as in (a)
(a) (b)
=

where (b) shows that (a) is self inverse. The operator is unitary, and (a) may be
written as the tensor
SWAPij kl = δ i l δ jk ,

– 12 –
or expanded in the computational basis as

SWAP =
X
|ijihji| .
ij

It also has a well-known implementation in terms of three CNOT gates as

Example (SWAP on two Hilbert spaces). Let X and Y be complex Hilbert

spaces of dimensions d1 and d2 respectively, then the SWAP operation is the
map
SWAP : X ⊗ Y → Y ⊗ X
(2.1)
SWAP : |xi ⊗ |yi 7→ |yi ⊗ |xi ,
for all |xi ∈ X , |yi ∈ Y.
Given any two orthonormal basis

{|xi i : i = 0, . . . , d1 − 1}

and
{|yj i : j = 0, . . . , d2 − 1}
for X and Y respectively, we can give an explicit construction for the SWAP
operation as Repeated indices to be
1 −1 dX
dX 2 −1 summed can share the same
SWAP = |yj ihxi | ⊗ |xi ihyj | . (2.2) color in wire diagrams [34].
i1 =0 j2 =0

The SWAP operation is represented graphically by two crossing wires as shown:

:=

The basis decomposition in (2.2) is then an application of the resolution of

the identity to each wire.
We will now consider the transformation of raising or lowering an index. One
can raise an index and then lower this index or vice versa, which amounts essentially
to the net effect of doing nothing at all. This is captured diagrammatically by the
so called, snake or zig-zag equation, as

= =
together with its vertical reflection across the page. The snake or zig-zag equation
in diagrammatic form dates back at least to Penrose [1]. Given a basis makes a
duality between flipping a bra to a ket, that is, raising or lowering an index, precise.
In tensor index notation, it is given simply by δji δ ik = δj k .

– 13 –
 The mathematical rules of tensor network theory assert that the wires of tensors
may be manipulated, with each manipulation corresponding to a specific contrac-
tion or transformation.

Definition 2.1. Transposition of 1st-order vectors and dual-vectors, and 2nd-order linear operators is repre-
sented by a bending of a tensors wires as follows:

v v AT  A  A
 v  v

(2.3)
(a) Vector transposition: (b) Dual-vector transposition: (c) Linear operator transposition

|viT = hv| hv|T = |vi

Complex conjugation of a tensor’s coefficients however is depicted by a bar over the tensor label in the
diagram:

v v A (2.4)
(a) Complex conjugation of |vi (b) Complex conjugation of hv| (c) Complex conjugation of A

We stress that under this convention a vector |vi = i vi |ii and its hermitian conjugate dual-vector hv| =
P

i v i hi| are represented as shown in Figure 1.1a and 1.1b respectively.

P

For explicit examples of writing down the equational form of a tensor diagram refer
to the proofs in Section 4.
Equivalence class. In further detail, we will consider the class of operations
formed from bending tensor wires forwards or backwards using cups and caps,
as well as exchanging wires using SWAP. We can conceptualize this class of trans-
forms acting on a tensor as, amounting essentially, to matrix reshapes. From the
snake equation, action with a cup or cap is invertible and SWAP is self inverse.
This means that, all possible configurations of a tensors legs using these operations
are equivalent, when the equivalence is taken up to Penrose duality.
Lemma 2.1 (Cardinality of index manipulations). Given a tensor T ij with fixed
labels i, j we can use cups and caps to arrive at

T ij , T ij , Tij , Ti j , (2.5)

the SWAP operation reorders i and j and then the cups and caps yield

T ji , T ji , Tji , Tj i . (2.6)

In general, for a tensor with a total of n indices, each index can be up or down,
yielding 2n possibilities. The symmetry group formed by SWAP is of order n!
and acts to arrange the n legs of a tensor, yielding

n! · 2n (2.7)

– 14 –
 different ways to reorder the indices of a tensor, provided we distinguish Ti j
and T ji etc.
Remark (Ordering operators by numbers of inputs and outputs). In the previous
remark, we considered Ti j (b) and T ji (a) etc., as distinct. For all practical
purposes, they are not however. This is shown as follows.

(a) (b)

This shows an awkward property of standard Dirac notation. Both (a) and
(b) represent the same map, but when we write this in a basis, one of them
will require us to write hi| ⊗ |ji.
With this in mind, we note that the tensor Tij which was considered in the
last section actually has 6 unique reshapes, as two of the reshapes are dia-
grammatically equivalent.
(a) (b) (c)

(d) (e) (f)

The duality is well know, but we have not seen mention of the order. We
call this the natural tensor symmetry class. In Theorem 2.2 are going to count (i)
the number of possible ways a tensor can have its wires bent, either forward or
backwards using the cups and caps, in conjunction with (ii) the number of ways a
tensor can have its arms and/or legs exchanged.

Theorem 2.2 (Natural tensor symmetry class). The arms and legs of a tensor
Γij···k
qr···s with m input arms (ij · · · k) and n output legs (qr · · · s), can be rear-
ranged in
(n + m + 1)! (2.8)
different ways.
Proof. Exercise.

3 Bending Density Operator Wires

We will now apply the natural tensor symmetry class counted in Theorem 2.2 to
the analysis of the quantum states arising from bending wires on density operators.
These states are found by bending all the wires of a tensor representing a density
operator to the same side, as follows.
(a) (b) (c)

– 15 –
For the case of a density map, m = n.
We will first compare the idea of a matrix basis with that of a vector space basis
both for an inner product space. We will use these concepts to study symmetries
of density operators. In fact,
Remark (Injection from Density Operators to States — Theorem 3.1). We will
soon prove that the existence of an injective map sending each n-party density
matrix ρ to a quantum state. The map is found by bending wires. The
resulting state is naturally equivalent under SWAP to 2n! states, by Theorem
2.2.
The process is invertible. However, every quantum state does not always give
rise to a ρ under wire duality. The purpose of the present section is to make these
statements precise.

Matrix basis
We expand d-dimensional operators using a matrix basis {Ai }, which is orthonormal
with respect to the Hilbert-Schmidt inner product, defined as
1  † 
hAi , Aj i := Tr Ai Aj = δij . (3.1)
d
The product Ai Aj is given in (a)
(a) (b)

and the trace inner product in (b). Here and elsewhere in this work, all scale factors
in the diagrams are omitted graphically, but we note that care must be taken when
one is summing over diagrams.
Example (Pauli matrix basis). A nascent example of a matrix basis with the
described properties is the Pauli matrices on qubits. Any operator of type
C2 → C2 can be written in terms of the Pauli matrices as

a1 + p~ · ~σ = a1 + bX + cY + dZ, (3.2)

for a, b, c, d ∈ C. Note that this provides a decomposition into a symmetric

subspace
S = span{1, X, Z}, (3.3)
and an antisymmetric subspace A = span{Y }. This symmetry is exhibited
by ρ = ρ> ∈ S and antisymmetry as ρ = −ρ> ∈ A. Diagrammatically,
transposition is done by twisting a map (see (a) or (b) in Remark 2) [35].
Vector basis
For a vector space basis {ai } in C2 ⊗ C2 we use the typical inner product where

hai , aj i = δij , (3.4)

which is given diagrammatically as in (b). (a) Follows from the diagrammatic

SVD, and the black square is intended to depict a not necessarily unitary, or for
that matter invertible, map.

– 16 –
 (a) (b)

= =

Example (Bell vector basis). An example of a vector basis is the Bell basis on
qubits. Any vector in C2 ⊗ C2 can be written in terms of this basis as (see
Table 1)
ψ(a, b, c, d) = aΦ+ + bΦ− + cΨ+ + dΨ− , (3.5)
for a, b, c, d ∈ C. This partitions the space into a symmetric subspace

s = span{Φ+ , Φ− , Ψ+ }, (3.6)

and an antisymmetric subspace a = span{Ψ− }. This symmetry is illustrated

by ψ(i, j) = ψ(j, i) ∈ s and antisymmetry as ψ(i, j) = −ψ(j, i) ∈ a. Diagram-
matically, this amounts to letting the swap gate act on both output wires,
which serves to exchange them.
Comparison
We note that in the diagrammatic language, the Bell vector basis and the Pauli
matrix basis become essentially equivalent as they are related by bending wires. In
particular, we note that they have identical form as the operator basis norm and
the inner product of vectors. These are both, up to a scale factor, identical in the
graphical language. To illustrate our point, we present the following table.

C2 → C2 C2 ⊗ C2
1 =
∼ Φ+ = |00i + |11i
σx =
∼ Ψ+ = |01i + |10i
σy =
∼ Ψ− = i(|01i − |10i)
σz =
∼ Φ− = |00i − |11i

Table 1. Mapping between Pauli matrices and Bell states.

Table 1 illustrates the specific mapping between Pauli matrices and Bell states.
For example, the second row represents

(a) (b) (c)

=:

where (a) is the σ x matrix. By bending a wire, we arrive at (b) which is identically
equal to the bell state Ψ+ in (c).

Density operators vs pure states.

We are in a position to carry out an analysis of the class of states found by bending
the wires of a density operator all to the same side. We first consider the expansion
of a density operator into the so called Hilbert-Schmidt basis of Pauli operators.
We let
Pn = {1, X, Y, Z}⊗n , (3.7)
be the set of all n letter words, formed from the alphabet of Pauli matrices, with
⊗ as the concatenation operator. The span of Pn forms a Hermitian operator basis
– 17 –
as each element is invariant under the †. We can expand any density operator ρ in
terms of this basis as
ρ= (3.8)
X
aijk···l σi σj σk · · · σl ,
for an n-long index ijk · · · l, where each i, j, k etc indexes an operator in P1 . We
can now study the natural embedding of this operator ρ on C2 into
⊗n ⊗n
→ C2
.
2 ⊗2n
ψρ ∈ C

Theorem 3.1 (Injection from density operators to states). Every density oper-
ator ρ on H⊗n → H⊗n gives rise to a state ψρ in H⊗2n . This state has 2n!
natural symmetries induced by swap.
Proof. Each density operator is dual to a state by bending wires. We
will express this starting with the density operator

ρ= (3.9)
X
aijk···l σi σj σk · · · σl ,

and writing it as

ψρ = aijk···l Ψi ⊗ Ψj ⊗ Ψk ⊗ · · · ⊗ Ψl , (3.10)
X

where Ψm ∼ = σm per bending a wire on each σm , and arriving at the

state ψρ . This results in the quantum state ψρ . Note that there are a
number of choices of ordering when bending the wires. In addition, these
wires can be arbitrarily ordered after bending to still form an essentially
equivalent state. In fact, one can act with the symmetry group on the
open wires, to arrive at the natural symmetry class with the same order
as the permutation group. This scenario is depicted below.

Example (Injection from density operators to states). We can illustrate the idea
behind Theorem 3.1 in the following figure. (a) depicts density operator ρi j
and (b) ρij its state dual.

(a) (b)

Remark (From states to operators by Penrose wire duality). While every density
operator gives rise to a quantum state, the converse is not necessarily true.
The condition ρ = ρ† corresponds to aijk···l ∈ R which limits the possible
states. For the case of qubits, a quantum state has 2(2n − 1) real degrees of
freedom in general, whereas the states arising under Theorem 3.1 have 2n − 1
real degrees of freedom.
Remark (In general wire bending is not purification). Let
1 1
ρ= 1 + σ.A = 1 + aX + bY + cZ, (3.11)
2 2
be the arbitrary state of a qubit. Then by Theorem 3.1,

ψρ = (1 + c) |00i + (a + bi) |01i + (a − bi) |10i + (1 − c) |11i , (3.12)

– 18 –
 is the natural state-dual to ρ with norm hψρ |ψρ i = 2(a2 + b2 + c2 ). However,
we note that Theorem 3.1 does not in general provide a purification of ρ.
Remark (Symmetric density operators vs symmetric states). A symmetric single
qubit density operator necessarily has bY = 0 and so gives rise to a symmetric
two qubit state with real valued coefficients parametrized by two real degrees
of freedom, ψρ (a, 0, c). Diagrammatically, a symmetric two-party state is in-
variant under exchange of its legs (a) whereas a symmetric density operator
is invariant under exchange of its arms and legs (e.g. transpose), (b).
(a) (b)
= =

3.1 Bipartite Matrix Operations

Bipartite matrices are used in several representations of CP-maps, and manipula-
tions of these matrices will be important in the following discussion.

Definition 3.1. Consider two complex Hilbert spaces X , and Y with dimen-
sions dx and dy respectively. The bipartite matrices we are interested in are
then d2x × d2y matrices M ∈ L(X ⊗ Y) which we can represent as 4th-order
tensors with tensor components

Mmµ,nν := hm, µ| M |n, νi (3.13)

where m, n ∈ {0, ..., dx −1}, µ, ν ∈ {0, ..., dy −1} and |n, νi := |ni⊗|νi ∈ X ⊗Y
is the tensor product of the standard bases for X and Y.

Graphically this is given by

m n
MmΜ,nΝ
Μ
M Ν

We can also express the matrix M as a 2nd-order tensor in terms of the stan-
dard basis {|αi : α = 0, . . . , D − 1} for X ⊗ Y where D = dx dy . In this case M has
tensor components
Mαβ = hα| M |βi (3.14)
This is represented graphically as

MΑΒ Α M Β Α M Β

We can specify the equivalence between the tensor components Mαβ and Mmµ,nν
by making the assignment

α = dy m + µ (3.15)
β = dy n + ν, (3.16)

– 19 –
where dy is the dimension of the Hilbert space Y.
The bipartite matrix operations which are the most relevant for open quantum systems (see Fig. VI) are the
partial trace over X (TrX ) (and TrY over Y), transposition (T ), bipartite-SWAP (S), col-reshuffling (Rc ), and
row-reshuffling (Rr ). The corresponding graphical manipulations are:

M
M
M M
M M
(a) Partial Trace (b) Partial Trace (c) Transpose (d) Bipartite-Swap (e) Row-Reshuffle (f) Col-Reshuffle
T S Rr
TrX [M ] TrY [M ] M M M M Rc

In terms of the tensor components of M these operations are respectively given by:
Partial trace over X TrX : L(X ⊗ Y) → L(Y), 7→ m Mmµ,mν
P
Mmµ,nν
Partial trace over Y TrY : L(X ⊗ Y) → L(X ) 7→ µ Mmµ,nµ
P
Mmµ,nν
Tranpose T : L(X ⊗ Y) → L(X ⊗ Y), Mmµ,nν 7→ Mnν,mµ
Bipartite-SWAP S : L(X ⊗ Y) → L(Y ⊗ X ), Mmµ,nν 7→ Mµm,νn
Row-reshuffling Rr : L(X ⊗ Y) → L(Y ⊗ Y, X ⊗ X ), Mmµ,nν 7→ Mm,n,µ,ν
Col-reshuffling Rc : L(X ⊗ Y) → L(X ⊗ X , Y ⊗ Y), Mmµ,nν 7→ Mνµ,nm
Note that we will generally use reshuffling R to refer to col-reshuffling Rc . Similarly we can represent the
partial transpose operation by only transposing the wires for X (or Y), and the partial-SWAP operations by only
swapping the left (or right) wires of M .

3.2 Vectorization of Matrices

We now recall the concept of vectorization which is a reshaping operation, trans-
forming a (m × n)-matrix into a (1 × mn)-vector [36]. This is necessary for the
description of open quantum systems in the superoperator formalism, which we
will consider in Section 32.

Definition 3.2. Vectorization can be done with using one of two standard
conventions: column-stacking (col-vec) or row-stacking (row-vec). Consider
two complex Hilbert spaces X ∼ = Cm , Y ∼= Cn , and linear operators A ∈
L(X , Y) from X to Y. Column and row vectorization are the mappings

col-vec: L(X , Y) → X ⊗ Y : A 7→ | Aiic (3.17)

row-vec: L(X , Y) → Y ⊗ X : A 7→ | Aiir (3.18)

respectively, where the operation col(row)-vec when applied to a matrix, out-

puts a vector with the columns (rows) of the matrix stacked on top of each
other.

– 20 –
Illustration 3.2 Graphical representations for the row-vec and col-vec operations are found from bending
a wire to the left either clockwise or counterclockwise respectively:

A
Ar := Ac := A
(a) Row-vec (b) Col-vec

Vectorized matrices in the col-vec and row-vec conventions are naturally equivalent under wire exchange
(the SWAP operation)

A
Ac    Ar
A

In particular we can see that the unnormalized Bell-state Φ+ ∈ X ⊗ X is in fact the vectorized identity
 

operator 1 ∈ L(X )
Φ = | 1iir = | 1iic . (3.19)
 E
 +

Definition 3.3. We may also define a vectorization operation with respect to

an arbitrary operator basis for L(X , Y). Let X ∼ = Cdy , and Z ∼
= Cdx , Y ∼ = CD
where D = dx dy . Vectorization with respect to an orthonormal operator basis
{σα : α = 0, ..., D − 1} for L(X , Y) is given by

σ-vec: L(X , Y) → Z : A 7→ | Aiiσ . (3.20)

This operation extracts the coefficients of the basis elements returning the
vector
D−1
| Aiiσ := Tr σα† A |αi (3.21)
X h i

α=0

where {|αi : α = 0, ..., D − 1} is the standard basis for Z ∼

= CD . This is
depicted in our graphical calculus as

AΣ
Σ† A

Remark. To distinguish between these different conventions we use the

notation | Aiix to denote the vectorization of a matrix A, were the sub-
script x = c, r, σ labels which convention we use; either c for col-vec, r
for row-vec, or σ for an arbitrary operator basis.
For the case X ∼=Y ∼ = Cd , we can define row-vec and col-vec in terms terms
of (3.21) by taking our basis to be the elementary matrix basis {Ei,j = |iihj| :

– 21 –
 i, j = 0, ..., d2 − 1}, and making the assignment α = di + j and α = i + dj
respectively. Hence we have
d−1
| Aiir := (3.22)
X
Aij |ii ⊗ |ji
i,j=0
d−1
| Aiic := (3.23)
X
Aij |ji ⊗ |ii .
i,j=0

Example. Using the definition of the unnormalized Bell-state Φ+ and sum-

 

ming over i and j one can rewrite Eq. (3.22) and (3.23) as

| Aiir = (A ⊗ 1) Φ+ (3.24)
 E


| Aiic = (1 ⊗ A) Φ+ (3.25)
 E


which are the equational versions of our graphical definition of row and col
vectorization shown in (3.2).
When working in the superoperator formalism for open quantum systems, it is
sometimes convenient to transform between vectorization conventions in different
bases. Given two orthonormal operator bases {σα } and {ωα } for L(X , Y), the basis
transformation operator

Tσ→ω : Z → Z : | Aiiσ 7→ | Aiiω (3.26)

transforms vectorized operators in the σ-vec convention to the ω-vec convention.

Graphically this is given by

AΩ = TΣ®Ω AΣ

The basis transformation operator Tσ→ω is given by the equivalent expressions

Tσ→ω = |αi hhωα |σ = (3.27)

X X
| σα iiω hα| ,
α α

and the corresponding graphical representations are:

For the remainder of this chapter we will use the col-vec convention by default,
and drop the vectorization label subscripts unless referring to a general σ-basis.
The main transformation we will be interested in is then from col-vec to another
arbitrary orthononormal operator basis {σα }. Tensor networks for the change of
basis Tc→σ and its inverse Tσ→c are

– 22 –
 Tc Σ TΣ c
Σ† Σ

(a) Col-vec to σ-basis (b) Row-vec to σ-basis

In the case where one wants to convert to row-vec convention, as previously shown
the transformation is given by

Tc→r = Tr→c = SWAP. (3.28)

One final important result that often arises when dealing with vectorized ma-
trices is Roth’s Lemma for the vectorization of the matrix product ABC [36]. Given
matrices A, B, C ∈ L(X ) we have

| ABCii = (C T ⊗ A)| Bii (3.29)

The graphical tensor network proof of this lemma is as follows:2 2 The theory of tensor net-
works leverages one to study the
mathematical structure formed by
the composition of processes and
states on the same footing.

4 Worked Examples
We will now prove the consistency of several of the basic tensor networks introduced
in Part I, and in doing so illustrate how one may use our graphical calculus for
diagrammatic reasoning.
The color summation convention we have presented represents diagrammatic
summation over a tensor index by coloring the appropriate tensors in the diagram.
In this convention summation over a Kronecker delta, i,j δij = i,j hi|ji, is as
P P

shown:
  ij∆ij
(4.1)
This expression is used in several of the following proofs.
We begin with the proof of the trace of an operator A:

A  A

 A

 i Aii
(4.2)

For illustrative purposes, to prove this algebraically we

note that the
 +tensor
networks for trace correspond to the algebraic expressions Φ A ⊗ 1 Φ
+ and



– 23 –
 Φ  1 ⊗ A Φ+ , and that

Φ+  1 ⊗ A Φ+ = hi|ji hi| A |ji =

D   E X X
δij Aij
 
i,j i,j

= (4.3)
X
Aii
i
= Tr[A].

Similarly we get Φ+  A ⊗ 1 Φ+ = Tr[A].

  

To prove the snake equation we must first make the following equivalence for
tensor products of the elements |ii and hj|:

hj| ⊗ |ii ≡ |ii ⊗ hj| ≡ |iihj| (4.4)

This is illustrated diagrammatically as

i j
  i j
j i
(4.5)

With this equivalence made, the proof of the snake-equation for the “S” bend
is given by

 (4.6)
The proof for the reflected “S” snake-equation follows naturally from the equiva-
lence defined in (4.5).
The proof of our tensor network for the transposition of a linear operator A is
as follows:

A  A

 A

 AT

AT
(4.7)


To prove this algebraically we note that the corresponding algebraic equation for

– 24 –
the transposition tensor network is

A
= 1 ⊗ Φ+  (1 ⊗ A ⊗ 1) Φ+ ⊗ 1 (4.8)
D   E

= (4.9)
X
hj| A |ii |ii ⊗ hj|
i,j

= (4.10)
X
hj| A |ii |iihj|
i,j

= (4.11)
X
hi| AT |ji |iihj|
i,j

= (4.12)
X
|iihi| AT |jihj|
i,j

= AT . (4.13)

The proof for transposition by counter-clockwise wire bending follows from the
equivalence relation in (4.4) and (4.5).
With the tensor network for transposition ofan operator proven, the proof of
transposition by contracting through a Bell-state Φ+ is then an application of the
snake equation as shown:

AT 
A


A
(4.14)

5 Problems

Exercise 5.1. (Bell basis). Using the right hand side equations,

σ0 = σ1 =
X X
|aiha| , |1 − aiha| ,
σ2 = i (−1)a+1 |aih1 − a| , σ3 = (−1)a |aiha| .
X X

for a ∈ {0, 1}. Show that

 
1 1 X
hl| ⊗ hl| σi ⊗ 1 = √ (hl| σi ) ⊗ hl| , (5.1)
X
√ 
2 l∈{0,1} 2 l∈{0,1}

defines the Bell effects (which defines an orthonormal basis in C2 ⊗C2 ). Here
σi indexes the Pauli matrices.3 3 Effects are dual to states. Also
called, costates.

Remark. Complete the following problems using standard techniques and then
compare this to graphical tensor network approach.

– 25 –
 Exercise 5.2. Suppose ψ ∈ C⊗2
2 and E > 0 with E ∈ L(C2 ). Show that
⊗2

hψ| E ⊗ 1 |ψi , (5.2)

takes the same values when ψ is any of the four Bell states.

Exercise 5.3 (Transpose or Ricochet Trick). Show that

(M A ⊗ 1B ) |ΦiAB = (1A ⊗ (M > )B ) |ΦiAB , (5.3)

for maximally entangled |ΦiAB and any matrix M .

Exercise 5.4. Show that the purity P (ρA ) is equal to

P (ρA ) = Tr{(ρA ⊗ ρA )F AA }, (5.4)

0 0

where Hilbert space A is isomorphic to A0 and F AA is the swap operator

0

defined on a basis (indexed by x, y) as

0 0
F AA |xiA |yiA = |yiA |xiA . (5.5)
0

Hint. Establish that

Tr{f (ρA )} = Tr{(f (ρA ) ⊗ 1A )F AA }, (5.6)

0 0

for function f of the operators on A.

6 Further Reading
Readers should be aware of the high number of quality tutorials covering various
aspects of tensor networks available for free download from the arXiv.org preprint
server. Many but not all of these are also published in journals. For those inter-
ested in applications to condensed matter, we particularly recommend.

Hand-waving and Interpretive Dance: An Introductory Course on Ten-

sor Networks [14]
Jacob Bridgeman and Christopher Chubb
J. Phys. A: Math. Theor. 50 223001 (2017) arXiv:1603.03039

For those interested in the mathematics of string diagrams (category theory),

the most accessible introduction covering the foundations of the building blocks
presented in this chapter can be found in.

A Prehistory of n-Categorical Physics [37]

John Baez and Aaron Lauda
Deep Beauty 13–128, Cambridge University Press (2011) arXiv:0908.2469

– 26 –
 In addition to this work on the categories of tensor networks [37], readers might
also find the survey [38] of interest.

– 27 –
 PART

Matrix Product States II

One of the most common uses of tensor networks in quantum information is rep-
resenting states which belong to small but physically relevant subset of a must
larger Hilbert space. This often includes low-entanglement states. The backbone
of this idea rests on low rank matrix approximations which we will consider in this
chapter.
We will see that if one partitions a network, by cutting it in two, the number
of wires that were cut in this process provides an upper bound on the maximum
amount of possible entanglement between spins. This can be made more precise
by considering the (unitarily invariant) entanglement entropy of a bipartite split.

E = − Tr (ρlnρ) = − λi lnλi . (6.1)

X

Here λi are the singular values of the reduced density operator of either subsystem. Figure 2. “You should call it
The quantity is maximized for all λi equal to the inverse of the dimension of a entropy, for two reasons. In
the first place your uncertainty
reduced density matrix. The value of (6.1) provides a quantitative measure of
function has been used in
correlations. This will be elaborated as a central concept in what follows. statistical mechanics under that
name, so it already has a name.
In the second place, and more
7 The Diagrammatic SVD important, no one really knows
what entropy really is, so in a
In this section, we will introduce a diagrammatic form of the singular value decom- debate you will always have the
position (SVD). It is assumed that the reader has solved Exercise 10.1 (see also advantage.”
Exercise 7.1).
There are several utilities to our approach. The first stems from the fact that John von Neumann sug-
the known invariants we have studied can be simplified by network contraction gesting to Claude Shannon a
using the diagrammatic SVD. The method factors tensor into well defined building name for his new uncertainty
blocks with simplistic interaction properties: black COPY-tensors and white unitary function, as quoted in Scientific
American 225(3) 180, (1971).
boxes. We will also consider the iteration of this process, allowing one to arrive at
matrix product states (MPS) in terms of our network building blocks.
An aim of the present work is to consider how graphical depictions of tensors
can leverage a better understanding of how certain properties evident in a network
are reflected in properties related to a quantum state. Intuitively one thinks of a
connected network as representing a correlated or entangled state. This chapter
will push this idea further.
COPY-tensors have been studied in the setting of the Penrose tensor calculus,
in work dating back at least to Lafont — see also [22, 23]. Here we apply them
in the diagrammatic SVD factorization of quantum states. The key diagrammatic
properties are illustrated as follows.

Definition 7.1 (Properties of the COPY-tensor). The copy property is illus-

trated in (b). Here a basis state, or copy point |xi is contracted with the
COPY-tensor, which then breaks into two copies of |xi. These tensors are de-
fined in d dimensions. If the tensor was written in the standard basis, it would
copy |0i as well as |1i and for the case of qubits be written as |0i h00| + |1i h11|.

28
We use the plus symbol |+i to represent an equal sum over all so called, copy
points of a COPY-tensor. When contracting a COPY-tensor with such a basis
state, the effect is to prune an arm or leg, as shown in (c) and (d). This is
also called a unit. COPY-tensors can be composed. (e) illustrates that appro-
priate composition of two COPY-tensors is equal to the identity, as would be
expected.

(a) (b) (c) (d) (e)

= = = = =

=
Remark (COPY-tensors on as generalized delta functions). The valence-three
COPY-tensor in terms of components for the case of qubits can be expressed
as
δijk = (1 − i)(1 − j)(1 − k) + ijk, (7.1)
which can be thought of a generalized delta function, where i, j, k = 0, 1. This
can be extended to valence-n COPY-tensors as

δijk...l = (1 − i)(1 − j)(1 − k)...(1 − l) + ijk...l. (7.2)

In figure (e) from Definition 7.1, the composition of two COPY-tensors becomes

δ i jk δ jkm = δ i m . (7.3)

We note that the valence-two delta tensor in components is given as

δ i j = 1 − (i − j)2 = (1 − i)(1 − j) − ij, (7.4)

where i, j = 0, 1.

Exercise 7.1. (Rank-1 projectors). Show that a non-trivial operator P? is

a Schmidt rank-1 projector if and only if it can be written as |ψihψ|.

Theorem 7.2 (Diagrammatic SVD). Every operator A : H1 → H2 can be

factored into a nonnegative order-one tensor Σ (unique), an order-three COPY-
tensor and unitary order-two tensors U and V :

A = V Σ U = V U

– 29 –
 Proof. Using the SVD, every valence-two tensor Ail can be written as

Ail = Uji Σjk Vlk , (7.5)

where U and V are unitary and Σ is diagonal and nonnegative in the

standard bases of H1 and H2 . Σ can be written as
min(d1 ,d2 )−1
Σ=
X
σj |ji2 hj|1 ,
j=0
min(d1 ,d2 )−1
= (where σk ≥ 0).
X X X
|ii2 hi|1 |ji1 hjj|1 σk |ki1 ,
i=0 j k
| {z }| {z }| {z }
Q12 COPY σ
(7.6)
We have then expressed the tensor Σ as a contraction of an order-one
tensor σ with the COPY-tensor. The tensor Q12 is only necessary if H1
and H2 have different dimension.

Remark (Eckart-Young-Mirsky Theorem). The rank of a matrix A is the number

of non-zero singular values it has. To determine its optimal rank-r approxi-
mation (with r < rank(A)), we can turn to a classic theorem by Eckart and
Young which was generalized by Mirsky.
Given the SVD, A = U ΣV † , we will discard rank(A) − r smallest singular
values in Σ by setting them to zero, obtaining Σ0 . This process is often called
trimming.
This gives rise to A0 = U Σ0 V † , an approximation of A.

Theorem 7.3 (Eckart-Young-Mirsky). For m × m matrices A, A0

kA − A0 k = min kA − Âk, (7.7)

rank(Â)≤r

for any unitarily invariant matrix norm k·k with r < rank(A).

Here  is any approximation to A of the same or lesser rank as A0 . This implies Interested readers can try to
that truncating or trimming Σ in this way yields as good of an approximation get their hands on copies of
as one can expect. In the following section, we will specifically consider the C. Eckart and G. Young,
induced error for such an approximation. “The approximation of
one matrix by another of
Corollary 7.4 (Diagrammatic Schmidt decomposition). Given a bipartite state
lower rank,” Psychometrika
|ψi, we use the snake equation to convert it into a linear map (inside of the dashed
1, (1936) and L. Mirsky,
region). Now we apply the SVD as in Theorem 7.2, resulting in unitary maps U
“Symmetric gauge functions
and V , a COPY-tensor and the order-one tensor σ representing the singular values.
and unitarily invariant
Diagram reorganization leads to the diagrammatic Schmidt decomposition of |ψi:
norms,” The Quarterly
Journal of Mathematics
11:1, 50–59 (1960).

– 30 –
 σ

V U U
ψ = ψ = = σ
T
V

(Sliding V around the cup takes the transpose but the resulting map is still unitary).
The singular values σ0 , . . . , σd−1 in σ correspond to the Schmidt coefficiets.
Example (Graphical map-state duality). In the figure from Corollary 7.4, we
arrive at an example of map-state duality from [39] as

(a) (b)
ψ = ψ

In (a) we start with a state |ψi. We can think of this (vacuously, it would
seem) as a state being acted on by the identity operation. Application of the
snake-equation to one of the outgoing wires allows one to transform this into
the diagram in (b). We can think of (b) a Bell state (left) being acted on by
a map found from coefficients of the state |ψi. We have illustrated this map
acting on the bell state by a light dashed line around |ψi.

Definition 7.2 (Partition χ). Given a many-body quantum state, partition the
state into two halves and perform the diagrammatic SVD on this system,
with respect to this partition. The term “χ” stands for the number of non-
zero singular values across a bipartition of a state. These values are used to
compute the entanglement entropy (6.1) of either reduced subsystem.

Example (Entanglement topology). The most significant topology change oc-

curs when the input state to the black COPY-tensor is a copy point — this
causes the diagram to break into two. When the input state is a unit for the
COPY-tensor, the tensor structure is converted to a smooth wire (this is the
maximally entangled case).
(a) (b) (c)

(a) is the general form of a state ψ = αi |ϕi i |φi i with χ > 1 and at least
P

two singular values taking different values, that is ∃λi 6= λj for some i 6= j.
The state takes the form in (b) iff the singular values in the triangle from (a)
all take the same values, so λi = λj ∀i, j with clearly χ = d . In this case, the
state is LU equivalent to a generalized Bell state, in dimension d. The state

– 31 –
 takes the form in (c) iff the first singular value equals one, which necessarily
implies that the remaining singular values are zero. In such a case, the state
is separable and χ = 1.
Corollary 7.5 (Diagrammatic state purification). The diagrammatic SVD from
Theorem 7.2 gives rise to a diagrammatic representation of state purification. To
begin with consider
(a) (b)

In (a) we have a two-party pure density state |ψi hψ| and in (b) we trace out one
subsystem ρ0 = Tr2 (|ψi hψ|). Now consider
def

(a) (b) (c)

= =

(a) is found from applying the diagrammatic SVD to the reduced state ρ0 . (b) fol-
lows from applications for simple diagrammatic rewrite rules, allowing the bottom
unitaries to cancel. This follows from pulling both boxes around the bends, which
takes the transpose of each map. We arrive at U †> U > = (U U † )> = 1. In (c) the
COPY-tensors merge, resulting in multiplication of the singular values stored in the
valence-one triangular tensors.
Alternatively, |ψi can be seen as a purification of ρ1 (the square roots of the sin-
gular values multiply (c) resulting again in ρ1 .) These diagrams translate between
a purification of a density operator and the density operator itself.

8 Matrix Product Factorization of States

As a key application of the diagrammatic SVD, we will consider Matrix Product
States (MPS), an iterative method to factor quantum states into a linear chain of
tensors (see [6, 7] for a recent review and [23] for work considering the category
theory behind MPS). We will express this factorization in terms of the diagram-
matic SVD 7.2, and our focus will be on exposing the degrees of freedom which
are invariant under local groups acting on open tensor legs. A basis to expand any
local unitary invariant of an MPS will be explored in Section 8.

Definition 8.1. Given an n-party quantum state |ψi, fully describing this state
generally requires an amount of information (or computer memory) that grows
exponentially with n. If |ψi represents the state of n qubits,

|ψi = (8.1)
X
ψij···k |ij · · · ki ,
ij···k

the number of independent coefficients ψij···k in the basis expansion in general

would be 2n which quickly grows into a computationally unmanageable number
as n increases. The goal is to find an alternative representation of |ψi which

– 32 –
 is less data-intensive. We wish to write |ψi as

|ψi = tr(Ai Aj · · · Ak ) |ij · · · ki , (8.2)

X [1] [2] [n]

ij···k

where Ai , Aj , . . . , Ak are indexed sets of matrices and trace (tr) means

[1] [2] [n]

matrix trace and could be omitted (e.g. .Ai and Ak are row and column
[1] [n]

vectors respectively). Calculating the components of |ψi then becomes a mat-

ter of calculating the products of matrices, hence the name matrix product
state.
If the matrices are bounded in size, the representation becomes efficient in
the sense that the amount of information required to describe them is only
linear in n. The point of the method is to choose these matrices such that
they provide a good (and compact) approximation to |ψi. For instance, if the
matrices are at most χ by χ, the size of the representation scales as ndχ2 ,
where d is the dimension of each subsystem.

Without loss of generality, we will apply the MPS method to a four-party state,
and explain the procedure in terms of three distinct steps. Consider a quantum
state, expressed as a triangle in the Penrose graphical notation with a label 1 inside
and open legs labeled i, j, k, m.

1
i j k m

(Step I). We will now create a partition of the legs of this state, into a first
collection containing only leg i and a second collection containing legs j, k, m. We
will then apply the diagrammatic SVD across this partition. The partition is
illustrated with the dashed cut below in (a). Figure (b) results from applying the
diagrammatic SVD across this partition, factoring the original state with label 1 in
(a) into a valence-two unitary box with label 2, a valence-one triangle containing the
singular values with label 3, and a valence-four triangle with label 4, all contracted
with a COPY-tensor, as illustrated. A new internal label (d) for the wire connecting
the COPY-tensor to the valence-four triangle (4) was introduced for clarity. (see
also Figure 3 (a) and (b)).
cut
(a) (b) d
2
1 =
33
4
i j k m i j k m

Remark (Isometric internal tensors). The valence-four triangle tensor in (b)

above is actually a unitary map. The only input leg shown is labeled d. The
other legs are contracted with a fixed basis state |0i, from the SVD (a). We
then depict this as the triangle 4 as in (b). From the unitarity property, the
isometry property follows, as illustrated graphically in (c).

– 33 –
 (a) d
d (c)
(b)
U = 4 =

j k m
j k m

Remark (Contraction of Unitaries). In tensor network diagrams, two unitaries

compose to form unitaries. In (a) below, we factor a tensor with three legs into
an order-two unitary (white box), a black order three COPY-tensor contracted
with an order-one triangle of singular values and an order-four unitary with
label U . In (b) we remove U and act on it with an arbitrary order-two tensor.
These compose to form U e which is still unitary. In a general MPS, unitaries
acting on the open legs do not alter the singular values found in a factorization.
(a) (b)

=
=

(Step II). To illustrate the next step in the factorization, we will remove the
tensor labeled 4 by breaking the wire connecting it to the COPY-tensor (a). We
will then partition this separate tensor into two halves, one containing wires d, j
the other half wires k, m. This partition is illustrated by placing a dashed line
(labeled cut) in (a). We arrive at the the structure in (b), which we have explained
in the first step. (see also Figure 3 (b) and (c)).
cut
(a) d (b) d e
5
4 = 6 7

j k m j k m

Remark (An elementary property of tensor network manipulation). It is a fun-

damental property of tensor network theory that one can remove a portion
of a network, alter this removed portion of the network without changing its
function, and replace it back into the original network, leaving the function of
the original network intact.
(Step III). In the third and final step of the MPS factorization applied to this
four-party example, following remark 8 we first place the tensor we have factored
in the second step, back into the original network from the first step, see (a) below.
We then repeat the second step, applied to the triangular isometry tensor, labeled
internally with a 7. This results in the factorization appearing in (b). (see also
Figure 3 (c) and (d)).
(a) cut (b)
d e d e
2 5 2 5 8 10
3 6 = 3 6 9
7
i j k m i j k m

Remark (Step n). The iterative method continues in the same fashion as the
first three steps, resulting in a factorization of an n-party state. A summary
of the MPS factorisation applied to a four-party state is shown in Figure 3.

– 34 –
 cut cut
(a) (b) (c) cut (d)
d
2 d e d e
2 5 2 5 8 10
1 =
33
4 = 3 6 = 3 6 9
7
i j k m i j k m i j k m i j k m

Figure 3. (Diagrammatic summary of steps I, II and III). The quantum state (a) is
iteratively factored into the 1D Matrix Product State (d). This procedure readily extends
to n-body states.

(Summary). We will now consider Figure 3, which summarizes the factorization

scheme. In the steps we have outline, we have factored the Figure 3 original state
(a) into the MPS in Figure 3 (d), in terms of the components listed below.

(i) States (labeled 3, 6 and 9; denoted φ3 , φ6 and φ9 , respectively): φ3 =

(λ0 , λ1 )> , φ6 = (λ2 , λ3 , λ4 , λ5 )> and φ9 = (λ6 , λ7 )> . The λi ’s are the sin-
gular values across each partition. The number of non-zero singular values
(χ) is given by the minimum dimension of the two parties in the cut. For the
case of qubits, the first outside partition has at most two non-zero entries,
and the next inside partition has at most 4. One might also consider the
singular values as the eigenvalues of either member of the pair of reduced
density matrices found from tracing out either half of a partition.

(ii) Unitary gates (labeled 2 and 10; denoted U2 and U10 , respectively).

(iii) Isometries (labeled 5 and 8; denoted I5 and I8 respectively). The isometry

condition describes the tensor relation Ijq
d I jq = δ d . It is a consequence of
r r
the fact that tensors 5 and 8 arise from unitary gates, as explained in Step
II. The isometry condition plays a more relevant role in structures other than
1D tensor chains.

We note that by appropriately combining neighboring tensors as in Figure 4

(a), one recovers the familiar matrix product representation of quantum states 4
(b). Matrix product states are written in equational form as

ψ= (8.3)
X [1] [2] [3]
Ai Aj Ak A[4]
m |ijkmi .
i,j,k,m

Here A[1] becomes a new tensor formed from the contraction of tensors labeled 2,
3, and A[2] is a contraction of tensors labeled 5 and 6, etc. The exact grouping has
some ambiguity.
A utility of our approach Figure 4 (a) is that the COPY-tensor is well defined
in terms of purely graphical rewrite identities (as seen in Definition 7.1). These
graphical relations allow one to gain insights (into e.g. polynomial invariants as will
be seen), and to contract portions of tensor networks by hand. The factorization
we present however, allows one to preform many diagrammatic manipulations with
ease, and exposes more structure inherent in a MPS.
Remark (Data compression). The compact representation of a MPS is recov-
ered by picking a cutoff value for the singular values across each partition, or
a minimum number of allowed singular values. This allows one to compress
data by truncating the Hilbert space and is at the heart of MPS computer

– 35 –
 algorithms in current use.

(a) (b)
2 5 8 10
3 6 9 =

i j k m i j k m

Figure 4. Conversion from our notation (a), to conventional MPS notation (b). The
factorization methods we have presented here and elsewhere [22, 23] allows one to “zoom
in” and expose internal degree of freedom (a) or “zoom out” and expose high-level structure
(b). The equational representation of the MPS in (b) is given in (8.3).

The singular values found from the MPS factorization can be used to form a
complete basis to express any quantity related to an MPS that is invariant under
local unitary operations. This includes providing a complete basis to express any
entanglement monotone.
Example (MPS for the GHZ state). The standard MPS representation of the
Greenberger-Horne-Zeilinger (GHZ) state is given as Daniel Greenberger, Michael
!n Horne and Anton Zeilinger
1 |0i 0 1 first studied what is now
|GHZi = √ tr = √ (|00 . . . 0i + |11 . . . 1i). (8.4)
2 0 |1i 2 named the GHZ-state in 1989
[40].
Alternatively, we may use a quantum circuit made of CNOT gates to construct
the GHZ state, and then use the rewrite rules employed in Examples 2 and 2
to recover the familiar MPS comb-like structure consisting of COPY tensors:

Diagrammatically, any tensor network formed from connected COPY-tensors

reduces to a single dot with the appropriate number of input and output legs.
Hence one might write the n-party GHZ-state as
1 X
|GHZi = √ COPYijk...l |ijk . . . li . (8.5)
2 ijk...l

Example (MPS for the W state). Like the GHZ state from Example 8, the
n-qubit W state (n ≥ 3) has the following MPS representation:
!n−2
1   |0i 0
!
|0i
|W i = √ |1i |0i ,
n |1i |0i |1i
(8.6)
1
= √ (|10 . . . 0i + |010 . . . 0i + . . . + |0 . . . 01i) .
n

– 36 –
Invariant Basis for Matrix Product States
We will consider generating a full monomial basis in terms of the singular values
found in the factorization of Matrix Product States. The monomial basis is gener-
ated as a matter of convenience and can be used to define a basis for entanglement
monotones. In fact, we will see that the diagrammatic factorization can be used to
prove that certain tensor contractions give rise to certain invariants, which allow
one to calculate quantities of interest, such as the concurrence or Reyni entropy.
These tensor contractions are not dependent on the factorization method used, and
are general. The diagrammatic SVD will be used to prove that the contraction of
certain tensors, results in an expression that is in terms of the singular values.
Our objective will be to develop tensor contractions that evaluate to specific
quantities of interest. These quantities of interest will be invariants of the local
unitary group. Such invariants have expansions in terms of the singular values of
reduced density operators. We will use tensor network contractions to evaluate a
full basis that can be used to expand any function of the singular values. This
includes quantities of interest such as concurrence and Reyni Entropy.
For the general case, one can, for instance, form a polynomial basis using the
elementary symmetric polynomials by combining the λ as

S1 = (8.7)
X
λi ,
i

S2 = (8.8)
X
λi λj ,
i6=j

S3 = (8.9)
X
λi λj λk ,
i6=j6=k

and so on. Such polynomials are used to calculate the d-concurrence, see Definition
8.3. Any polynomial in the S’s is necessarily a local unitary invariant.
Another basis of interest, is the basis formed by summed powers of the singular
values
Bn = (8.10)
X
λni ,
i

which is of great interest to evaluate Reyni’s entropy, see Definition 8.4. The
polynomial B2 is related to the concurrence measure of entanglement in Definition
8.2. For the case of qubits, B2 = λ20 + λ21 which is greater than zero iff the state
is entangled. Increase n, pattern continues up to Bd . Bd being greater than zero
implies that Bn ≥ 0 for all 1 ≤ n ≤ d. In addition, Bn = 0 implies that Bk = 0 for
all n ≤ k ≤ d.
We will first recall the definition of the concurrence (see [28]) and then the
definition of Reyni’s Entropy (see for instance [41]). These quantities are of physical
interest. As will be shown, they can be calculated by contracting specific tensor
networks.

Definition 8.2 (The concurrence). The concurrence of a pure bipartite normal-

ized state |ψi is defined as
s
d
C(|ψi) := (1 − Trρ2 ), (8.11)
d−1

– 37 –
 where ρ is obtained by tracing over one subsystem. The factor d/(d − 1)
p

ensures that 0 ≤ C(|ψi) ≤ 1.

Remark (Tensor contractions for the concurrence). We will contract tensor net-
works that evaluate to B2 = Tr(ρ2 ) and these hence can be used to evaluate
the concurrence.

Definition 8.3 (The d-concurrence). Consider a d x d-dimensional bipartite

pure state |ψi with Schmidt numbers λ := (λ0 , λ1 , ..., λd−1 ) the d concurrence
monotones, Ck (|ψi), k = 1, 2, ..., d, of the state |ψi are defined as follows

Sk (λ0 , λ1 , ..., λd−1 )

 1/k
Ck (|ψi) := , (8.12)
Sk (1/d, 1/d, ...1/d)

where Sk is the kth elementary symmetric polynomial.

Definition 8.4 (The Reyni Entropy). The Reyni entropy of order α is defined
to be
1
Sα := ln (8.13)
X
λαi ,
1−α
and in the limit α → 1

E = lim Sα = − λi ln λi . (8.14)
X
α→1
i

Remark (Lower bounds on Entropy). Consider a tensor network representing a

quantum spin state. Partition this state into a block of k-spins and a block
of m-spins. Let the number of wires connecting the two blocks be given as l.
Then
E ≤ min{k, m, l}. (8.15)
In the remaining sections, we will consider tensor networks that enable the
evaluation of the quantities of interest we have mentioned here. At the heart of
the MPS method, is the factorization of a state into two halves. Our diagrammatic
SVD allows one to prove that certain networks contract to quantities that can
readily be related to the concurrence or Reyni’s entropy. In what follows, we build
examples for biaprtite states. These generally apply to Matrix Product States by
considering partitions. In fact, in the diagrammatic language, it is often useful to
group k-wires into one wire, for the purpose of manipulation [23].

9 Numerical Tensor Network Algorithms and Packages

This book is focused on graphical reasoning and the applications of tensor networks
to quantum information science. A primary driving force behind tensor network
applications is numerical algorithms for condensed matter physics application. This
is not our main focus, however.
However, no book would be complete without touching on these important
numerical applications. Indeed, we have described several tensor contractions which
can be utilized to solve counting problems (see § VII). And to simulate quantum
systems using matrix product states (Part 8).

– 38 –
 Those reading this text that want to explore the numerical application of these
ideas should be aware of both open source software packages and papers that
describe in detail the most effective tensor contraction algorithms. That is precisely
the objective of this appendix.

9.1 Online Resources Describing Tensor Network Software Implemen-

tations
Several webpages are devoted to listing papers and tensor contraction algorithms.
Here we provide a short listing of those papers which go into more detail and are
of a primary software centric focus. We have done our best to be as inclusive
as possible, however this list is merely an editors pick and is not designed to be
comprehensive.

1. Resource. TensorNetwork.ORG
Brief description. An open-source ‘living’ article containing many tensor
network resources, applications, and software.
Link. https://tensornetwork.org

2. Resource. Tensors.NET
Brief description. A collection of resources, including links to software,
tutorials and code in several languages.
Link. https://www.tensors.net

9.2 Open Source Tensor Network Software Packages

Software packages and programs to do various tensor contractions and related tasks
is increasingly available in a variety of languages. This includes TEBD programs
as well as others. Here we have done our best to include an active and up to date
listing of the main packages. It was last updated December 21, 2019. As software
packages can appear and also become inactive at any time, a more updated listing
of available packages can be found on this books homepage.

1. Title of package. ITensor

Language. C++
Brief description. ITensor or Intelligent Tensor is well featured tensor
network library implementing a bit of everything out of the box.
Status. Active (version 2.0.11)
Link. http://itensor.org

2. Title of package. TNT

Language. Various
Brief description. The TNT library contains highly optimised routines for
manipulating tensors and routines that can be used to build the most common
tensor network algorithms [42].
Status. Active
Link. http://www.tensornetworktheory.org

3. Title of package. TensorNetwork

Language. Phython
Brief description. A Google/X package for tensor networks described in
[43] and related to Google’s popular tensor flow package (https://www.
tensorflow.org).
– 39 –
 Status. Active
Github. https://github.com/google/tensornetwork

4. Title of package. Quantomatic

Language. Phython
Brief description. A diagrammatic proof assistant supporting reasoning
with diagrammatic languages with applications to tensor networks [44, 45].
Status. Active
Github. https://quantomatic.github.io

10 Problems

Exercise 10.1 (Tensor products, Density operators, Singular values and

purification). Let ψ = |001i + |010i + |100i be a state in H1 ⊗ H2 with
dim(H1 ) = 2 and dim(H2 ) = 4, and let

0 1 0 0
 
 √1 0 0 − √12 
V † =  √12 . (10.1)
 

2
0 0 √12 
0 0 1 0

(i) By writing φ = ij Cij |ii |ji, where |ii ∈ H1 , |ji ∈ H2 are both
P

orthonormal basis, state the values of i ∈ {0, 1}, j ∈ {0, 1, 2, 3} that

correspond to non-zero coefficients of Cij and hence express ψ in the
basis |ii , |ji.

(ii) Write the 2 × 4 matrix C = (Cij )ij and show that CC† and C† C
are (non-normalized) density operators, equivalent to ρ1 , ρ2 found by
tracing over the systems H1 and H2 respectively. (Here the adjoint †
means matrix conjugate transpose.)

(iii) From the singular value decomposition, one can write C = U ΣV † . For
V given above, and U the 2 × 2 identity matrix, find the 2 × 4 matrix
of singular values Σ.

(iv) Find purifications for ρ1 and ρ2 (other than ψ).

Exercise 10.2. (Lie product formula—a.k.a. Trotter-Suzuki decomposi-

tion). Show that
k
eA+B = lim (10.2)

eA/k eB/k
k→∞

for Hermitian matrices A and B.

Exercise 10.3. The Schmidt number of |ψiAB —denoted Sch(ψ)—is the

rank of the reduced density matrix ρA = TrB (|ψihψ|) where rank of a Her-
mitian operator is defined as the dimension of its support.

– 40 –
Exercise 10.4. Let |ψi = α |φi + β |γi and prove that

Sch(ψ) ≥ kSch(φ) − Sch(γ)k1 (10.3)

Exercise 10.5. Let non-negative A, B ∈ L(Cn ). Show equality of the sets

as Spec{AB} = Spec{BA}.

Exercise 10.6. A matrix A ∈ L(Cn ) is non-negative (A ≥ 0) if

hψ| A |ψi ≥ 0 (10.4)

∀ψ ∈ Cn . Show that A ≥ 0 implies the existence of a unique B such that

B 2 = A.

Exercise 10.7. Let Hermitian A, B ∈ L(Cn ). Now suppose that A2 = A

and B 2 = B and further that {A, B} = 0. Let kψk2 = 1 and show that

hψ| A |ψi2 + hψ| B |ψi2 ≤ 1 (10.5)

Exercise 10.8. Consider

1
ρ = (1 + X ⊗ X − Y ⊗ Y ) (10.6)
4
and use Lagrange multipliers to find kρk∞ .

Exercise 10.9. Let density operator ρAB represent the state of a qubit pair
(A, B) and define the spin-flipped density matrix as

ρ̃AB = (Y ⊗ Y )ρ̄AB (Y ⊗ Y ) (10.7)

where ρ̄ is complex conjugate of ρ in the standard basis. As ρAB , ρ̃AB ≥ 0,

so is their non-Hermitian product (Problem (10.5)). Consider ψ as the two-
qubit Bell state and ρ = |ψihψ|. Find the eigenvalues of ρρ̃.

– 41 –
 PART

Boolean Tensor Networks III

11 Introduction
Now we will explain part of a tool set and framework largely following [22, 23].
Hence, we will approach tensor networks by focusing on familiar components, “It is not of the essence of
namely Boolean logic gates (and multi-valued logic gates in the case of qudits), mathematics to be conver-
applied to the tensor network context. See Appendix B for background on Boolean sant with the ideas of num-
algebra. The concept of Boolean linearity and non-linearity was used to form a ber and quantity.” — George
dichotomy between fundamental Boolean tensor building blocks in [22]. Boole
This subject has an increasingly long history. The first categorical model of
Boolean circuits (as well as some progress on the quantum case) can be found
in the seminal work by Lafont [46]. Lafont’s pioneering ideas were subsequently
developed by the categorical quantum mechanics community and also influenced
our work in [22]. In the condensed matter community, Boolean tensors, such as
COPY-tensors, are very commonly used. Closely related to Boolean tensors are
stabilizer tensors (§ 20), sometimes called Clifford tensor networks or sometimes
even the ZX-calculus [25].

12 Overview of the Chapter

Tensor network representations of quantum states
A qudit is a d-level generalization of a qubit.4 As has been seen in the last section, 4 A q-dit generalization of the ZX-

a quantum state of n-qudits has an exact representation as a order-n tensor with calculus appears in [23]. The ZX-
each of the open legs corresponding to a physical degree of freedom, such as a calculus is equivalent to Clifford
circuits plus cups and caps to bend
spin with (d − 1)/2 energy levels. Such a representation, shown in Figure 12.1(a) wires and compose states. It was
is manifestly inefficient since it will have a number of complex components which first proposed in terms of interact-
grows exponentially with n. The purpose of tensor network states is to decompose ing quantum observables in [47].
this type of structureless order-n tensor into a network of tensors whose order is
bounded. 5 5 Lafont was the first to devel-

There are now a number of ways to describe strongly-correlated quantum lattice oped a categorical theory of quan-
systems as tensor-networks. These include tum circuits [46]. Subsequent work
built on these pioneering ideas.
(i) Matrix Product States, MPS [48–50]

(ii) Projected Entangled Pair States, PEPS [7, 51]

(iii) Multiscale Entanglement Renormalisation Ansatz, MERA [52, 53]

(iv) Tree Tensor Networks, TNN [54, 55]

(v) Categorical Tensor Network States, CTNS [22, 23].

The central problem faced by all types of tensor networks is that the resulting
tensor network for the quantity hψ| (O |ψi, where O is some product operator, needs
to be efficiently contractible (efficient is taken to mean polynomial in the problem
size) if any physically meaningful calculations, e.g., expectation values, correlations

42
or probabilities, are to be computed. For MPS and TTN efficient contractibility
follows from the 1D chain or tree-like geometry, while for MERA it follows from
its interesting causal cone structure [53]. For PEPS and CTNS, however, exact
contraction is not proven to be efficient in general, but can often be rendered
efficient if approximations are made [7, 51].
For MPS and PEPS, shown in Figures 12.1(b) and (c), the resulting network
of tensors follows the geometry of the underlying physical system, e.g., a 1D chain
and 2D grid, respectively.
Alternatively a Tensor Tree Network (TTN) can be employed which has a
hierarchical structure where only the bottom layer has open physical legs, as shown
in Figure 12.1(d) for a 1D system and Figure 12.1(e) for a 2D one.
For MERA the network is similar to a TTN, as seen in Figure 12.1(f) for 1D,
but is instead comprised of alternating layers of order-four unitary and order-three
isometric tensors.
A categorical tensor network state (CTNS) contains some algebraically con-
strained tensors obeying some clearly defined diagrammatic laws, along with pos-
sible generic tensors. Indeed, when recast, certain widely used classes of tensor
network states can be readily exposed as examples of CTNS [22]. Specifically, vari-
ants of PEPS have been proposed called string-bond states [56]. Although these
string-bond states, like PEPS in general, are not efficiently contractible, they are
efficient to sample.

Illustration 12.1

(a) A generic quantum state |ψi for n degrees of freedom represented as a tensor with n open legs. (b) A
comb-like MPS tensor network for a 1D chain system [48, 49]. (b) A grid-like PEPS tensor network for a 2D
lattice system [7, 51]. (d) A TTN for a 1D chain system where only the bottom layer of tensors possess open
physical legs [54, 55]. (e) A TTN for a 2D lattice system. (f) A hierarchically structured MERA network
for a 1D chain system possessing unitaries (order-4 tensors) and isometries (order-3 tensors) [52, 53]. This
tensor network can also be generalized to a 2D lattice (not shown).

Remark (Generalisation of String Bond States). When recast, certain widely

used classes of tensor network states can be readily exposed as examples of
CTNS. Specifically, variants of PEPS have been proposed called string-bond
states [56]. Although these string-bond states, like PEPS in general, are not
efficiently contractible, they are efficient to sample. By this we mean that for
these special cases of PEPS, any given amplitude of the resulting state (for

– 43 –
a fixed computational basis state) can be extracted exactly and efficiently, in
contrast to generic PEPS. This permits variational quantum Monte-Carlo cal-
culations to be performed on string-bond states where the energy of the state
is stochastically minimized [56]. This remarkable property follows directly
from the use of a tensor, called the COPY-tensor, which forms one of several
tensors in the fixed toolbox considered in great detail later in this lecture.
As its name suggests, the COPY-tensor duplicates inputs states in the compu-
tational basis, and thus with these inputs breaks up into disconnected compo-
nents, as depicted in Figure 12.2(a). By using the COPY-tensor as the “glue"
for connecting up a TNS, the ability to sample the state efficiently is guar-
anteed so long as the individual parts connected are themselves contractible.
The generality and applicability of this trick can be seen by examining the
structure of string-bond states, as well as other types of similar states like
entangled-plaquette-states [57] and correlator-product states [58], shown in
Figure 12.2(c)-(e). A long-term aim of this work is that by presenting our
toolbox of tensors, entirely new classes of CTNS with similarly desirable con-
tractibility properties can be devised.

Illustration 12.2
(a) (b) (c)
...

... ...
...
=

(e) (d)

(a) One of the simplest tensors, the COPY-gate or the COPY-dot in classical boolean circuits, copies compu-
tational basis states |xi where x = 0, 1 for qubits and x = 0, 1, ..., d − 1 for qudits. The tensor subsequently
breaks up into disconnected states. (b) A generic PEPS in which we expose a single generic order-5 tensor.
This tensor network can neither be contracted nor sampled exactly and efficiently. However, if the tensor
has internal structure exploiting the COPY-tensor, then efficient sampling becomes possible. (c) The tensor
breaks up into a vertical and a horizontal order-3 tensor joined by the COPY-tensor. Upon sampling com-
putational basis states the resulting contraction reduces to many isolated MPS, each of which are exactly
contractible, for each row and column of the lattice. This type of state is known as a string-bond state and
can be readily generalized [56]. (d) An even simpler case is to break the tensor up into four order-2 tensors
joined by a COPY-tensor forming a co-called correlator-product state [58]. (e) Finally, outside the PEPS
class, there are entangled plaquette states [57] which join up overlapping tensors (in this case order-4 ones
describing a 2×2 plaquette) for each plaquette. Efficient sampling is again possible due to the COPY-tensor.

Remark (From qubits, to qtrits, ..., qdits). By invoking known theorems assert-
ing the universality of multi-valued logic [59] (also called d-state switching),
our methods can be readily applied to tensors of any finite dimension. This

– 44 –
 was considered explicitly in [23].

Tensor network components defined by diagrammatic laws

Here we will review the collection of tensors that form a universal tensor tool box.
In mathematical logic and computer science, formal semantics is an important
field of study. Throughout this lecture series, we adopt the semantics developed
in [22, 23] which aimed to offer a natural extension of the graphical language of
quantum circuits in wide spread use in quantum physics. This was done by merging
the overlapping concepts in various fields into a common language that deviates as
little as possible from the standard language of quantum circuits.
To get an idea of how the tensor calculus will work, consider Figure 12.3, which
forms a presentation of the linear fragment of the Boolean calculus [46]): that is,
the calculus of Boolean algebra we represent on quantum states, restricted to the
building blocks that can be used to generate linear Boolean functions—as described
in [22].
Remark (Quantum linear states are non-trivial). In the setting of tensors, the
linear fragment of the calculus is already non-trivial. In fact, these are the
building blocks that appear in my important quantum information protocols
and are the backbone of the widely studied class of stabilizer states. For
instance, in [32] the authors construct exactly contractible 2D networks for
use in topological states.
To recover the full Boolean-calculus, we must append a non-linear Boolean gate
as done in [22]: we use the AND-gate. Figure 12.3 together with Figure 12.4 form a
full presentation of the calculus [46]. The origin and consequences of these relations
will be considered in full detail in Section 13. The presentations in Figure 12.3
together with Figure 12.4 represent a complete set of defining equations.

Illustration 12.3 A summary of the linear fragment of the Boolean calculus on tensors (reproduced from
Lafont [46]). The plus (⊕) tensors are XOR and the black (•) tensors represent COPY. The details of (a)-(g)
will be given in Sections 13. For instance, (d) represents the bialgebra law and (g) the Hopf-law (in the
case of qubits x ⊕ x = 0, in higher dimensions the units h+| becomes h0| + h1| + · · · + hd − 1|). (Read top
to bottom.)

=
= (f) =
= =

(a) (b) (c) (d) (e)

=
= = (g) =
= =

Remark (Alternative approaches to the linear fragment). The structures in Fig-

ure 12.3 which we have found by casting classical circuits into tensor networks
are used to form the building blocks needed to represent CNOT gates and are

– 45 –
 related to other approaches [60–62] which have been used as a graphical lan-
guage for measurement based quantum computation [61] and for graph states
[62]. Our method of arriving at this collection of tensors (Figure 12.3) affords
more general options and our presentation of the linear fragment here offers (i)
improved semantics and (ii) a better theoretical understanding by pinpointing
precisely that these networks correspond to the so called linear fragment of the
XOR or mod sum algebra carries with it new proof techniques. These results
were found by casting the theory of classical networks into a theory of tensors
[22], which carried with it all of the known and desirable graphical rewrite
properties from classical networks, and from this and some other methods, in
[22] we assert that we have subsumed the existing graphical languages present
in quantum information science by considering the the graphical system ap-
pearing in Figure 12.4 together with the linear fragment from Figure 12.3.

Illustration 12.4

=
(a) (b) (c) (d) = = (f) = (h)
= = =

= (e) = (g) =

A summary of the quantum AND-tensor calculus we introduce to quantum theory (reproduced from Lafont
[46]). This figure with Figure 12.3 is a summary of the Boolean-calculus. The details of (a)-(g) will be given
in Sections 13. For instance, (h) represents distributivity of AND(∧) over XOR (⊕), and (d) shows that
x ∧ x = x. (Diagrams read top to bottom.)

Bending wires. Proceeding axiomatically we need to add additional tensors to

represent operators and quantum states. Our network model of quantum states
requires that we are able to bend wires. As is well known in modern algebra, we
can hence define transposition graphically (see Figure 11 (d)). Cups and caps (wire
bending) was also used in the categorical model of teleportation [38].
The way forward is to add what mathematicians refer to as compact struc-
tures (see Section 13.3.3 for further details). These compact structures are given
diagrammatically as
(a) (b)

and as will be explored in Section 13.3.3 these two structures allow us to formally
bend wires and to define the transpose of a linear map/state, and provide a formal
way to reshape a matrix. We understand (a) above as a cup, given as the generalized
Bell-state
d−1
|iii = |00i + |11i + ... (12.1)
X

i=0

– 46 –
and (b) above as the so-called cap, Bell-costate
d−1
hii| = h00| + h11| + ... (12.2)
X

i=0

or effect.
Remark (Normalization factors omitted). As we have mentioned before, we will
often omit global scale factors (contracted tensor networks with no open wires
are sent to blank on the page). This is done for ease of presentation. We note
that for Hilbert space H there is a natural isomorphism

=H∼
C⊗H∼ = H ⊗ C,

and care must be taken when summing over diagrams where a relative scale
factor could exist.
As readers will recall from § I, compact structures provide a formal way to
bend wires — indeed, we can now connect a diagram represented with an operator
with spectral decomposition X
βi |ii hi| ,
i

bend all the open wires (or legs) towards the same direction and it then can be
thought of as representing a state
X 
βi |ii i ,
i

where overbar is complex conjugation), bend them the other way and it then can
be thought of as representing a measurement outcome
X

βi i hi| ,
i

that is an effect. One can also connect inputs to outputs, contracting indices and
creating larger and larger networks. With these ingredients in place, let us now
consider the class of Boolean quantum states.
Remark (Overbar notation). The isomorphism

=
βi i hi| ∼ =
βi |ii hi| ∼ (12.3)
X
 X X 
βi |ii i ,
i i i

for a real valued basis becomes

=
βi hi| hi| ∼ =
βi |ii hi| ∼ (12.4)
X X X
βi |ii |ii ,
i i i

which amounts to flipping a bra to a ket and vise versa. Here we will always
assume a real valued basis so will always omit the overbar on kets.
Defining the class of Boolean tensor network states
Figure 5 which depicts a simple but key network building block: the use of the
so-called “quantum AND-tensor” [22] which we consider in detail in Section 13.3.
This is a representation of the familiar Boolean operation in the bit pattern of a

– 47 –
 (a) (b)

= =

Figure 5. Example of the Boolean quantum AND-state or tensor. In (a) the tensors output
is contracted with h1| resulting in the tensor splitting to the product state |11i. In (b) the
tensors output is contracted with h0| resulting in the entangled state |00i + |01i + |10i.

three-qubit quantum state as

|ψAND i = |x1 i ⊗ |x2 i ⊗ |x1 ∧ x2 i = |000i + |010i + |100i + |111i ,

def X

x1 ,x2 ∈{0,1}

and hence the truth table of a function is encoded in the bit pattern of the super-
position state. This utilizes a linear representation of Boolean gates on quantum
states as opposed to the typical direct sum representation common in Boolean
algebra.
In this lecture, we will construct tensor networks with components that take
binary values 0 or 1.
This is done by contracting the output of a switching tensor network with |1i.
The function realized in the tensor network is constructed in such a way that any
time the input qubits states represent a desired term in a quantum state (e.g. create
a function that outputs logical-one on designated inputs |00i, |01i and |10i and zero
otherwise as shown in Figure 5). We then insert a |1i at the network output. This
procedure recovers the desired Boolean state as illustrated in Figure 6(a) with the
resulting state appearing in Equation (12.5).

h1|f (x1 , x2 , ..., xn )i |x1 , x2 , ..., xn i =

X

x1 ,x2 ,...,xn ∈{0,1}

(12.5)
f (x1 , x2 , ..., xn ) |x1 , x2 , ..., xn i .
X

x1 ,x2 ,...,xn ∈{0,1}

The network representing the circuit is read backwards from output to input. Al-
ternatively the full class of Boolean states is defined as:

Definition 12.1 (The Class of Boolean Quantum States). We define the class
of Boolean states as those states which can be expressed up to a global scalar
factor in the form (12.6)

|x1 , x2 , ..., xn i |f (x1 , x2 , ..., xn )i , (12.6)

X

x1 ,x2 ,...,xn ∈{0,1,...,d−1}

where
f : {0, 1}n → {0, 1},
is a switching function and the sum is taken over all variables xj taking 0 and
1 for qubits (see Figure 6 (a)).

– 48 –
 Remark (Better notation). In practice it is often simpler to express equations
such as
f (x1 , x2 , ..., xn ) |x1 , x2 , ..., xn i , (12.7)
X

x1 ,x2 ,...,xn ∈{0,1}

as
f (x) |xi , (12.8)
X

where the sum is over all assignments of x := x1 , x2 , ..., xn .

(a) (b)

... ...

Figure 6. A general boolean quantum state arising from function f can either be formed
as (a) by network contraction with a logical-one at the output of the circuit as described
by Equation (12.5) or (b) by bending the output of the tensor network around, as in
Equation (12.6).

Example (GHZ-states and W-states). Examples of Boolean states include the familiar GHZ-state |00 · · · 0i +
|11 · · · 1i which on qudits in dimension d becomes
d−1
|GHZd i = |ii |ii |ii = |0i |0i |0i + |1i |1i |1i + · · · + |d − 1i |d − 1i |d − 1i , (12.9)
X

i=0

as well as the W-state |00 · · · 1i + |01 · · · 0i + · · · + |10 · · · 0i which again on qudits becomes
d−1 3
|Wd i := (Xj )i |0i |0i |0i = |0i |0i |1i + |0i |1i |0i + |1i |0i |0i + |0i |0i |2i +
XX

i=1 j=1 (12.10)

|0i |2i |0i + |2i |0i |0i + · · · · · · + |0i |0i |d − 1i + |0i |d − 1i |0i + |d − 1i |0i |0i .

In Equation (12.10) the operator X |mi = |m + 1(mod d)i is one way to define negation in higher dimensions.
The subscript labels the ket (labeled 1,2 or 3 from left to right) the operator acts i times.

Remark (Extensions to arbitrary quantum states). What is clear from this def-
inition is that Boolean states are always composed of equal superpositions
of sets of computational basis states, as the allowed scalars take binary val-
ues, 0,1. Despite this apparent limitation, tensor networks composed only of
Boolean components can nonetheless describe any quantum state. To do this
we require a minor extension to include superposition input/output states,
e.g. order-1 tensors of the form |0i + β1 |1i + · · · + βd−1 |d − 1i. This gives
a universal class of generalized Boolean tensor networks which subsumes the
important subclass of Boolean states. This class is then shown to form a
nascent example of the exhaustiveness of CTNS and to give rise to a wide
class of quantum states that we show are exactly and efficiently sampled [22].
Remark (Comparison to other approaches). The theory of tensor network states
has received recent interest fueled by developments that have been made re-

– 49 –
 lated to the important problem of quantum simulation, using tensor con-
traction algorithms [6, 7]. There is also an established language of quantum
circuits, appearing in most text books on quantum computing and quantum
information. These circuits are effectively tensor networks and efforts have
been made to form an extension and unite the two [23].

13 Quantum Legos: a tensor tool box

A key point to this is that the introduction of Boolean logic gate tensors into
the tensor network context allows the seminal logic gate universality results from
classical network theory to be applied in the setting of tensor network states.
Remark (Dual spaces). Any vector space V has a dual V ∗ : this is the space
of linear functions f from V to the ground field C, that is f : V → C. This
defines the dual uniquely. We must however fix a basis to identify the vector
space V with its dual. Given a basis, any basis vector |ii in V gives rise
to a basis vector hj| in V ∗ defined by hj|ii = δij (Kronecker’s delta). This
defines an isomorphism V → V ∗ sending |ii to hi| and allowing us to identify
V with V ∗ . In what follows, we will fix a particular arbitrarily chosen basis
(called the computational basis in quantum information science). We will now
concentrate on Boolean building blocks that are used in our construction.
As we have mentioned in Remark 6, the graphical language appearing in the
linear fragment (Figure 12.3) was considered first by Lafont [46] as well as in other
work [22, 47, 60–63]. The AND-tensor and the connection to Toffoli gates appeared
first in [22].

Review of basic boolean algebra

We arrived at these structures by casting the theory of Boolean circuits into a
theory of tensor networks, here we encourage the readers to review Boolean algebra
to better understand the presented structures. We advise readers to quickly review
Appendix 13.2 on XOR-algebra as well as Appendix C on the method of Karnaugh
map equation reduction. The following sections will assume these methods are
known to the reader.

13.1 COPY-tensors: the “diagonal”

The copy operation arises in digital circuits [64, 65] and more generally, in the
context of category theory and Algebra, where it is called a diagonal. The operation
is readily defined in any finite dimension as
d−1
4 = (13.1)
def X
|iii hi| .
i=0

As |0i and |1i are eigenstates of σ z , we might give 4 the alternative name of Z-
copy. In the case of qubits COPY is succinctly presented by considering the map
4 that copies σ z -eigenstates:
(
|0i 7→ |00i
4 : C2 → C2 ⊗ C2 :
|1i 7→ |11i

This map can be written in operator form as 4 : |00i h0| + |11i h1| and under
cup/cap induced duality (on the right bra) this state becomes a GHZ-state as
– 50 –
|ψGHZ i = |000i + |111i ∼= |00i h0| + |11i h1|. The standard properties of COPY
are given diagrammatically in Figure 13.1 and a list of its relevant mathematical
properties are found in Table 2.

Illustration 13.1
(a) (b) (c) (d)
= = = = =

Some diagrammatic properties of the COPY-tensor. (a) Full-symmetry. (b)

Copy points, e.g. |xi 7→ |xxi for x = 0, 1 for qubits. (c) The unit — in this
case the unit corresponds to deletion, or a map to the terminal object which
is given as h+| = h0| + h1| for qubits and h+| = h0| + h1| + · · · + hd − 1| for
def def

d dimensional qudits. (d) Co-interaction with the unit creates a Bell state.

Remark (The COPY-gate from CNOT). The CNOT-gate is defined as |0i h0|1 ⊗
12 + |1i h1|1 ⊗ σ2x . We will set the input that the target acts on to |0i then
calculate CNOT(11 ⊗ |0i2 ) = |0i h0|1 ⊗ |0i2 + |1i h1|1 ⊗ |1i2 . We have hence
defined the desired COPY map copying states from the Hilbert space with
label 1 (subscript) to the joint Hilbert space labeled 1 and 2. Note that this
was used in reverse in the optimal circuit example from Lecture I.
Remark (The types of possible states built from COPY). An alternative defini-
tion of the COPY-tensor would be to define the operation by raising or lowering
indices on δ ij k , a Kronecker delta function. In that regard, one might write
the n-party GHZ-state as

ψGHZ = (13.2)
X
δ ijk...l |ijk...li .

Tensor products of state of this form are precisely the only types of states
constructible with the COPY-tensor alone.
13.2 XOR-tensors: the “addition”
The classical XOR-gate implements exclusive disjunction or addition (mod 2 for
qubits) and is denoted by the symbol ⊕ [66, 67]. We note that for multi-valued
logic a modulo subtraction gate can also be defined as in [23].
Remark (relation to COPY). As is a well known fact in algebra, the XOR-gate
is simply a Hadamard transform of the COPY-gate, appropriately applied to
all of the tensors legs. This can be captured diagrammatically in the slightly
different form:

Remark (Symmetry). The XOR-tensor is symmetric under leg exchange. In

components, if we write Xijk then Xijk = 1 for i = j = k = 0 (mod 2) and 0
otherwise.
To define the gate on the computational basis, we consider f (x1 , x2 ) = x1 ⊕ x2
then f = 0 corresponds to (x1 , x2 ) ∈ {(0, 0), (1, 1)} and f = 1 corresponds to
(x1 , x2 ) ∈ {(1, 0), (0, 1)}, where the truth table for XOR follows.

– 51 –
 x1 x2 f (x1 , x2 ) = x1 ⊕ x2
0 0 0
0 1 1
1 0 1
1 1 0

Under cap/cap induced duality, the state defined by XOR is given as

|ψ⊕ i = |x1 i |x2 i |f (x1 , x2 )i = |000i + |110i + |011i + |101i , (13.3)

def X

x1 ,x2 ∈{0,1}

which is in the GHZ-class by LOCC equivalence viz. |ψ⊕ i = H⊗H⊗H(|000i+|111i).

The operation of XOR is summarized in Table 3. Since the XOR-gate is related to
the COPY-gate by a change of basis, its diagrammatic laws have the same structure
as those illustrated in Figure 13.1. The gate acting backwards (co-XOR) is defined
on a basis as follows:

|0i 7→ |00i + |11i

( (
|+i 7→ |++i
⊕:C →C ⊗C :
2 2 2
or equivalently
|1i 7→ |10i + |01i |−i 7→ |−−i

Generating the affine class of networks

Thus far we have presented the XOR- and COPY- gates. This system allows us to
create the linear class of Boolean functions. As explained in the present subsection,
this class can be extended to to the affine class by introducing either a gate that
acts like an inverter, or by appending a constant |1i into our system. This constant
will allow us to use the XOR-gate to create an inverter.

Definition 13.1 (Complemented vs Uncomplimneted Boolean Variables). A com-

plemented Boolean variable is a Boolean variable that appears in negated form,
that is ¬x or written equivalently as x. Negation of a Boolean variable x can
be expressed as the XOR of the variable with constant 1 as x = 1 ⊕ x. Whereas
uncomplimneted Boolean variables are Boolean variables that do not appear
in negated form (e.g. negation is not allowed). Linear Boolean functions con-
tain terms with uncomplemented Boolean variables that appear individually
(e.g. variable products are not allowed such as x1 x2 and higher orders etc., see
Section B).

Definition 13.2 (Linear boolean functions). Linear Boolean functions take the
general form
f (x1 , x2 , ..., xn ) = c1 x1 ⊕ c2 x2 ⊕ ... ⊕ cn xn , (13.4)
where the vector (c1 , c2 , ..., cn ) uniquely determines the function.

Definition 13.3 (Affine boolean functions). The affine Boolean functions take
the same general form as linear functions. However, functions in the affine
class allows variables to appear in both complemented and uncomplemented
form. Affine Boolean functions take the general form

f (x1 , x2 , ..., xn ) = c0 ⊕ c1 x1 ⊕ c2 x2 ⊕ ... ⊕ cn xn , (13.5)

– 52 –
 where c0 = 1 gives functions outside the linear class. From the identities,
1 ⊕ 1 = 0 and 0 ⊕ x = x we require the introduction of only one constant (c0 ),
see Appendix B.

Together, XOR and COPY are not universal for classical circuits. When used
together, XOR- and COPY-gates compose to create networks representing the class
of linear circuits. The affine circuits are generated by considering the constant |1i.
The state |1i is indeed copied by the black tensor. However, our axiomatization
(Figure 12.3) proceeds through considering the XOR- and COPY-gates together
with |+i, the unit for COPY and |0i the unit for XOR. It is by appending the
constant |1i into the formal system (Figure 12.3) that the affine class of circuits
can be realized.
Remark (Affine functions correspond to a basis). Each affine function is labeled
by a corresponding bit pattern. This can be thought of as labeling the com-
putational basis, as states of the form |{0, 1}n i are in correspondence with
polynomials in algebraic normal form (see Appendix B).

13.3 Quantum AND-state tensors: Boolean universality

The proceeding sections have introduced enough machinery to generate the linear
and affine classes of classical circuits. These classes are not universal. To recover
a universal system introduce a the AND gate as a tensor. The unit for this gate is
h1| and so can be used to elevate the linear fragment to the affine class.
The AND gate (that is, ∧) implements logical conjunction [64, 65]. The AND-
gate relates to the OR-gate via De Morgan’s law. This can be captured diagram-
matically as

To define the gate on the computational basis, we consider f (x1 , x2 ) = x1 ∧ x2

which we write in short hand as x1 x2 . Here f = 0 corresponds to (x1 , x2 ) ∈
{(0, 0), (0, 1), (1, 0)} and f = 1 corresponds to (x1 , x2 ) = (1, 1).
Under cap/cap induced duality, the state defined by AND is given as

|ψ∧ i = |x1 i |x2 i |f (x1 , x2 )i = |000i + |010i + |010i + |111i . (13.6)

def X

x1 ,x2 ∈{0,1}

The key diagrammatic properties of AND are presented in Figure 13.2 and the gate
is summarized in Table 4.
The gate acting backwards (co-AND) is defined on a basis as follows:

|0i 7→ |00i + |01i + |10i

( (
|+i 7→ |++i
∧ : C2 → C2 ⊗C2 : or
|1i 7→ |11i |−i 7→ |00i + |01i + |10i − |11i

Illustration 13.2 Salient diagrammatic properties of the AND-tensor.

– 53 –
 (a) (b) (c) (d)
= =
= =

(a) Input-symmetry. (b) Existence of a zero or fixed-point. (c) The unit |1i.
(d) Co-interaction with the unit creates a product-state. Note that the gate
forms a valid quantum operation when run backwards as in (d).

Example (AND-states from Toffoli-gates). The AND-state is readily con-

structed from the Toffoli gate [22] as illustrated in Figure 13.3. This al-
lows some interesting states to be created experimentally, for instance, post-
selection of the output to |0i would yield the state |00i + |01i + |10i.
Remark (Quantum universality and universal states). The problem of determin-
ing universal quantum gate families has received significant research interest
resulting in the simplistic universal gate sets appearing in [68–70] and else-
where. It is even known that Toffoli and Hadamard are universal for quantum
computation [68]. Toffoli can be generated by combining one AND-state and
two COPY-states (see also Figure 13.4). 6 6 In [22] the ZX calculus (Clifford
gates plus cups and caps to bend
Theorem 13.3 (Toffoli contracts to give the AND-state). The following rewrites
wires) by considering the addition
of AND-states which can repre-
exhibit the use units to prepare the AND-state [22]. Using this state together sent Toffoli gates was shown to
with single qubit NOT-gates, one can construct tensor networks which any be quantum computationally uni-
Boolean qubit state as well as any of the states appearing in Table 13.3.1. versal [22]. With the addition
We note that the box around the Toffoli gate (left) is meant to illustrate a of scalars, the AND+ZX calculus
presented in [22] was proven to be
difference between our notation and that of quantum circuits. In our notation, approximately universal for linear
those dots inside the box would merge which of course is not a valid unitary maps between qubits.
gate.

Theorem 13.4. Hadamard follows from contracting the AND-state together

with |−i = √12 (|0i − |1i) [22].
def

Proof.

= H

Follows by direct calculation.

– 54 –
 Definition 13.4. Hadamard states are defined as

|ψH i = |00i + |01i + |10i − |11i .

Lemma 13.5. Tensor contractions formed from Hadamard states, COPY-and

AND-states are universal for quantum computation [22].
Proof. The proof [22] follows from wire bending duality and the proof
that Hadamard and Toffoli are universal for quantum circuits [68].
13.3.1 Summary of the XOR-algebra on tensors
We will now present the three previously referenced Tables (2, 3 and 4) which
summarize the quantum logic tensors we introduced in the previous subsections
(13.1, 13.2 and 13.3). The tables contain entries listing properties that describe
how the introduced network components interact. These interactions are defined
diagrammatically and explained in Section 13.

Gate Type Co-copy point(s) Unit Co-unit Interaction

COPY |0i,|1i |+i Bell state: |00i + |11i
Symmetry Associative Commutative Frobenius Algebra
Full Yes Yes Yes (Fusion Law)

Table 2. Summary of the COPY-gate from Section 13.1.

Gate Type Co-copy point(s) Unit Co-unit Interaction

XOR |+i,|−i |0i Bell state: |00i + |11i
Symmetry Associative Commutative Frobenius Algebra
Full Yes Yes Yes (Fusion Law)

Table 3. Summary of the XOR-gate from Section 13.2.

Gate Type Co-copy point(s) Unit Co-unit Interaction

AND |1i |1i Product state: |11i
Symmetry Associative Commutative Bialgebra Law
Inputs Yes Yes Yes (with GHZ)

Table 4. Summary of the AND-gate from Section 13.3.

co-COPY: the co-diagonal

What is evident from our subsequent discussions on logic gates is that in the context
of tensors, the bending of wires implies that gates can be used both forwards in
backwards. We can therefore form tensor networks from Boolean gates in a very
different way from classical circuits. Indeed, it becomes possible to flip a COPY
– 55 –
operation upside down, that is, instead of having a single leg split into two legs,
have two legs merge into one. In terms of tensor networks, co-COPY is simply
thought of as being a dual (transpose) to the familiar COPY operation. This is
common in algebra: to consider the dual notation to algebra, that is co-algebra.
In general, while a product is a joining or pairing (e.g. taking two vectors and
producing a third) a co-product is a co-pairing taking a single vector in the space
A and producing a vector in the space A ⊗ A.
Remark (co-algebras [71]). co-algebras are structures that are dual (in the
sense of reversing arrows) to unital associative algebras such as COPY and
AND the axioms of which we formulated in terms of picture calculi (Sections
13.1 and 13.3). Every co-algebra, by (vector space) duality, gives rise to an
algebra, and in finite dimensions, this duality goes in both directions.
Co-COPY can be thought of as applying a delta function in the transition from
input to output. That is, given a copy point x = 0, 1, ..., d − 1 for qudits on dim d.
Depicting COPY as the map 4

4 (|xi) = |xi ⊗ |xi , (13.7)

we define co-COPY by the map 5 such that

5 (|ii , |ji) = δij |ii , (13.8)

that is, the diagram is mapped to zero (or empty) if the inputs |ii, |ji do not agree.
This is succinctly expressed in terms of a delta-function dependent on inputs |ii,
|ji where i, j = 0, 1, ..., d − 1 for qudits of dim d.
Example (Simple co-pairing). Measurement effects on tripartite quantum sys-
tems can be thought of as co-products. This is given as a map from one
system (measuring the first) into two systems (the effect this has on the other
two). GHZ-states are prototypical examples of co-pairings. In this case, the
measurement outcome of |0i (|1i) on a single subsystem sends the other qubits
to |00i (|11i) and by linearity this sends |+i to |00i + |11i.

The remaining Boolean tensors: NAND-states etc.

We have represented a logical system on tensors — this enables us to represent any
Boolean function as a connected network of tensors and hence any Boolean state.
We chose as our generators, constant |1i, COPY, XOR, AND. Other generators
could have also been chosen such as NAND-tensors. Our choice however, was made
as a matter of convenience. If we had considered other generators, we could have
ended up considering the following cases: weak-units (Definition 13.5) and fixed
point pairs (Definition 13.6).

Definition 13.5 (Weak units). An algebra (or product see Appendix A) on a

tripartite state |ψi has a unit (equivalently, one has that the state is unital) if
there exists an effect hφ| which the product acts on to produce an invertible
map B, where B = 1 (see Example 13.3.1). If no such hφ| exists to make
B = 1, and B has an inverse, we call hφ| a weak unit, and say the state |ψi
is weak unital and if B 6= 1 and B 2 = 1 we call the algebra on |ψi unital-
involutive. This scenario is given diagrammatically as:

– 56 –
 = =

Example (NAND and NOR). NAND and NOR have weak units, respectively
given by |1i and |0i. These weak units are unital-involutive.

|ψNAND i = |001i + |011i + |101i + |110i , (13.9)

|ψNOR i = |001i + |010i + |100i + |110i . (13.10)

For |ψNAND i to have a unit, there must exist a |φi such that

φ0 |01i + φ0 |11i + φ1 |01i + φ1 |10i = |00i + |11i , (13.11)

D  E D  E D  E D  E
   

and hence no choice of |φi makes this possible, thereby confirming the claim.

Definition 13.6 (Fixed Point Pair). An algebra (see Appendix A) on a tripartite

state |ψi has a fixed point if there exists an effect hφ| (the fixed point) which
the product acts on to produce a constant output, independent of the other
input value. For instance, in Figure 8(c) on the left hand side the effect h1|
induces a map (read bottom to top) that sends |+i 7→ |1i. Up to a scalar, this
map expands linearly sending both basis effects h0|, h1| to to the constant state
|1i. If the resulting output is the same as the fixed point, we say hφ| has a zero
(|1i is the zero for the OR-gate in Figure 8(c)). A fixed point pair consists of
two algebras with fixed points, such that the fixed point of one algebra is the
unit of the other, and vise versa (see Figure 8). Diagrammatically this is given
in Figure 7.

= =

Figure 7. Diagrammatic equations satisfied by a fixed point pair (see Definition 13.6).

(a) (b) (c)

= = = =

Figure 8. AND and OR tensors form a fixed point pair. The unit for AND (|1i see a) is
the zero for OR (c) and vise versa: the unit of OR (|0i see a) is the zero for AND (b).

Summarizing: network composition of quantum logic tensors

We have considered sets of universal classical structures in our tensor network
model. In classical computer science, a universal set of gates is able to express any
n-bit Boolean function

f : Bn → B : (x1 , ..., xn ) 7→ f (x1 , ..., xn ), (13.12)

– 57 –
where we note that Z2 ∼= B allowing us to use the alternative notation for f as
f : Zd → Zd with d = 2 for the binary case. Universal sets include
n

1. {COPY, NAND},

2. {COPY, AND, NOT},

3. {COPY, AND, XOR, |1i},

4. {OR, XNOR, |1i} and others.

One can also consider the states |ψi formed by the bit patterns of these functions
f (x1 , x2 ) as
|ψf i = |x1 i |x2 i |f (x1 , x2 )i . (13.13)
X

x1 ,x2 ∈{0,1}

This allows a wide class of states to be constructed effectively. In the following

Table (13.3.1) we illustrate the quantum states representing the classical function
of two-inputs.
The bit pattern of the following quantum states represents a Boolean function
(given by the subscript) such that the right most bit is the Boolean functions
output, and the two left bits are the functions inputs, and the non-linear Boolean
functions are on the left side of the table and the linear functions on the right.
Consider the state |ψAND i, and Boolean variables x1 and x2 , then the superposition
|ψAND i encodes the function |x1 , x2 , x1 ∧ x2 i in each term in the superposition, and

|ψAND i =
X
|x1 , x2 , x1 ∧ x2 i .
x1 ,x2 ∈{0,1}

As outlined in the text, cup/cap induced-duality allows us (for instance) to

express this state as the operator

|0i h00| + |0i h01| + |0i h01| + |1i h11| : |x1 , x2 i 7→ |x1 ∧ x2 i

which projects qubit states to the AND of their bit value.

non-linear linear (Frobenius Algebras)

|ψAND i = |000i + |010i + |100i + |111i
|ψOR i = |001i + |011i + |101i + |111i |ψXOR i = |000i + |011i + |101i + |110i
|ψNAND i = |001i + |011i + |101i + |110i |ψXNOR i = |001i + |010i + |100i + |111i
|ψNOR i = |001i + |010i + |100i + |110i

13.3.2 Merging COPY-tensors by node equivalence

COPY-tensors are readily generalized to an arbitrary number of input and output
legs. As one would rightly suspect, a COPY-tensor with n inputs and m outputs
corresponds to an n + m-partite GHZ-state. Neighboring tensors of the same type
can be merged into a single tensor: this is called node equivalence in digital circuits.
COPY-tensors represent Frobenius algebras7 [71, 74]. 7 In the work [72], the COPY-
tensor, through its properties as a
Frobenius algebra [73], was shown
to be equivalently characterised by
an orthogonal basis for a finite-
– 58 – dimensional Hilbert space.
 Theorem 13.6 (Node equivalence or fusion law). Given a connected graph with
m inputs and n outputs comprised solely of COPY-tensors of equal dimension,
this map can be equivalently expressed as a single m-to-n tensor, as shown as

Node equivalence or fusion law. Connected black-tensors (•) as well as

connected plus-tensors (⊕) can be merged and also split apart at will. The
intuition for digital or qudit circuits follows by connecting a state |φi to one
of the legs and iterating over a complete basis |0i, |1i,...,|d − 1i.

This rule goes by many different names, depending on the community. For
example, in classical circuits this is called node equivalence whereas researchers
in categorical quantum mechanics, credit node equivalence as their own spider
law [47].

Associativity, distributivity and commutativity

The products we have considered are all associative and commutative. As algebras,
AND, XOR and COPY are associative, unital commutative algebras. This was
already expressed diagrammatically in Figures 12.3(a) and Figure 12.4(c). The
diagrammatic laws relevant for this subsection represent the following Equations

(x1 ∧ x2 ) ∧ x3 = x1 ∧ (x2 ∧ x3 ), (13.14)

(x1 ⊕ x2 ) ⊕ x3 = x1 ⊕ (x2 ⊕ x3 ). (13.15)

Distributivity of AND over XOR then becomes (see (h) in Figure 12.4)

(x1 ⊕ x2 ) ∧ x3 = (x1 ∧ x2 ) ⊕ (x1 ∧ x2 ). (13.16)

We have commutativity for any product symmetric in its inputs: this is the case
for AND and XOR.

Bialgebras on tensors
There is a powerful type of algebra that arises in our setting: a bialgebra defined
graphically on tensors in Figure 9 (see Kassel, Chapter III [35], [71]).
Such an algebra is simultaneously a unital associative algebra and co-algebra
(for the associativity condition see (b) in Figure 9). Specifically, we consider the
following two ingredients:

(i) A product (black tensor) with a unit (black triangle) see the right hand side of
Figure 9(a).
(ii) A co-product (white tensor) with a co-unit (white triangle) see the left hand
side of Figure 9(a).
– 59 –
To form a bialgebra, these two ingredients above must be characterized by the
following four compatibility conditions:

(i) The unit of the black tensor is a copy-point of the white tensor as in (e) from
Figure 9.
(ii) The (co)unit of the white tensor is a copy-point of the black tensor as in (d)
from Figure 9.
(iii) The bialgebra-law is satisfied given in (c) from Figure 9.
(iv) The inner product of the unit (black triangle) and the co-unit (white triangle)
is non-zero (not shown in Figure 9).

(a) (b) (c) (d) (e)

= = = =
=
=

Figure 9. Bialgebra axioms [71] (scalars are omitted). (a) unit laws (these are of course
left and right units); (b) associativity; (c) bialgebra; (d,e) co-COPY points.

Example (GHZ, AND form a bialgebra). We are in a position to study the

interaction of GHZ-AND. This interaction satisfies the equations in Figure
9: (a) the bialgebra law; (b) the co-copy point of AND is |1i; and (c) the
co-interaction with the unit for GHZ creates a compact structure. In addition,
(a) and (b) show the copy points for the black GHZ-tensor; in (c) we have the
unit and fixed point laws.
Even if a given product and co-product do not satisfy all of the compatibility
conditions (given in (a), (b), (c), (d), (e) in Figure 9), and hence do not form bialge-
bras, they can still satisfy the bialgebra law which is given in Figure 9(c). Examples
of states that satisfy the bialgebra law in Figure 9(c), but are not bialgebras are
given in Definition 13.7. Notice that bialgebra provides a highly constraining char-
acterization of the tensors involved and is tantamount to defining a commutation
relation between them.

Definition 13.7 (Bialgebra Law [71]). A pair of quantum states (black, white
tensors) satisfy the bialgebra law if (c) in Figure 9 holds. The Boolean states,
AND, OR, XOR, XNOR, NAND, NOR all satisfy the bialgebra law with COPY.

13.3.3 Algebras on valence-3 tensors

A particularly important class of bialgebras are known as Hopf-algebras [71]. This
is characterized by the way in which algebras and co-algebras can interact. This is
captured by the Hopf-law, where the linear map A is known as the antipode.

Definition 13.8 (Hopf-Law [71]). A pair of quantum states satisfy the Hopf-
Law if an A can be found such that the following equations hold:

= =

– 60 –
 Example (XOR and COPY are Hopf-algebras on Boolean States [46]). It is well
known (see e.g. [46]) that the Boolean state XOR, satisfies the Hopf-algebra
law with trivial antipode (A = 1) with COPY. Recall Figure 12.3(g).

Bending wires: compact structures

As mentioned in the preliminary section (12), we make use of what’s called a
compact structure in category theory which amounts to introducing cups and caps,
to provide a formal way to bend wires and define transposition. See Figures 10 and
11.
A compact structure on an object H consists of another object H∗ together
with a pair of morphisms (note that we use the equation H∗ = H in Hilbert space
making objects self dual which simplifies what follows).

ηH : C −→ H ⊗ H, H : H ⊗ H −→ C,

where the standard representation in Hilbert space with dimension d and basis {|ii}
is given by
d−1 d−1
ηH = H =
X X
|ii ⊗ |ii , hi| ⊗ hi| ,
i=0 i=0

and in string diagrams (read from the top to the bottom of the page) as
(a) (b)

These cups and caps give rise to cup/cap-induced duality: this amounts to being
able to create a linear map that “flips” a bra to a ket (and vise versa) and at
the same time taking an (anti-linear) complex conjugate. In other words, the cap
i=0 hii| sends quantum state |ψi = α |0i + β |1i to α h0| + β h1| which is equal to
P1

the complex conjugate of |ψi† = hψ| = α h0| + β h1|. Diagrammatically, the dagger
is given by mirroring operators
D  across the page, whereas transposition is given by
bending wire(s). Clearly, ψ  = α h0| + β h1|.


In the case of relating the Bell-states and effects to the identity operator, under
cup/cap-induced duality, we flip the second ket on ηH and the first bra on H . This
relates these maps and the identity 1H of the Hilbert space: that is, we can fix a
basis and construct invertible maps sending ηH ∼ = 1H ∼ = H . More generally, the
maps ηH and H satisfy the equations given in Figure 10 and their duals under the
dagger.
A second way to introduce cups and caps is to consider a Frobenius form [71]
on either of the structures in the linear fragment from Figure 12.3 (COPY and
XOR). This is simply a functional that turns a product/co-product into a cup/cap.
This allows one to recover the above compact structures (that is, the cups and caps
given above) as
+

(a) = (b) =
0

Again, we will use these cups and caps as a formal way to bend wires in tensor
networks: this can be thought of simply as a reshape of a matrix.

– 61 –
Figure 10. Cup identities. (a) Symmetry. (b) Conjugate state. (c) the snake equation.
(d) Sliding an operator around a cup transposes it.

(a) (b)
= =

Figure 11. Diagrammatic adjoints. Cups and caps allow us to take the transpose of a
linear map. Note that care must be taken, as flipping a ket |ψi to a bra hψ| is conjugate
transpose, and bending a wire is simply
P transposition,
so the conjugate must be taken: e.g.
acting on |ψi with a cap given as i hii| results in ψ .

14 Examples of Boolean Tensor Network States

Constructing Boolean states
Since the fixed building blocks of our tensor networks are the logic tensors AND,
OR, XOR and COPY, along with ancilla bits, we can immediately apply the uni-
versality of these elements for classical circuit construction to guarantee that any
Boolean state has a categorical tensor network decomposition. However our con-
struction goes beyond this because as we have seen, categorical tensor networks can
be deformed and rewired in ways which are not ordinarily permitted in the stan-
dard acyclic-temporal definition of classical circuits. The W-state will be shown to
provide a non-trivial example of this.
Example (Functions on W- and GHZ-states). We consider the function fW
which outputs logical-one given input bit string 001, 010 and 100 and logical-
zero otherwise. Likewise the function fGHZ is defined to output logical-one on
input bit strings 000 and 111 and logical-zero otherwise. See Examples 14 and
14 which consider representation of these functions as polynomials. We will
continue to work with a linear representation of functions on quantum states;
here bit string 000 7→ |000i (etc.).
Example (MPS form for W-state). Like the GHZ state, the W-state has a simple
MPS representation
!n
|0i 0
|Wn i = h0| |1i = |10...0i + |01...0i + ... + |00...1i . (14.1)
|1i |0i

This description (14.1) is succinct. All MPS-states have essentially this same
topological or network structure. In contrast, our categorical construction
described below breaks this network up further.
Remark (Exact-value functions). The function fW takes value logical-one on
input vectors with k ones for a fixed integer k. Such functions are known in

– 62 –
 the literature as Exact-value symmetric Boolean functions. When cast into our
framework, exact-value functions give rise to tensor networks which represent
what are known as Dicke states [75].
Example (Function realization of fW and fGHZ : the Boolean case). One can
express (using x to mean Boolean variable negation)

fW (x1 , x2 , x3 ) = x1 x2 x3 ⊕ x1 x2 x3 ⊕ x1 x2 x3 (14.2)

by noting that each term in the disjunctive normal form of fW are disjoint,
and hence OR maps to XOR as ∨ 7→ ⊕. The algebraic normal form (see
Appendix B) becomes

fW (x1 , x2 , x3 ) = x1 ⊕ x2 ⊕ x3 ⊕ x1 x2 x3 (14.3)

fGHZ (x1 , x2 , x3 ) = 1 ⊕ x1 ⊕ x2 ⊕ x3 ⊕ x1 x2 ⊕ x1 x3 ⊕ x2 x3 (14.4)

Example (Function realization of fW and fGHZ : the set function case). Set func-
tions are mappings from the family of subsets of a finite ground set (e.g.
Booleans) to the real or complex numbers. In the circuit theory literature,
functions from the Booleans to the reals are known as pseudo-Boolean func-
tions and more commonly as multi-linear polynomials or forms (see [76] where
these functions are used to embed a co-algebraic theory of logic gates in the
ground state energy configuration of spin models). There exists an algebraic
normal form and hence a unique multi-linear polynomial representation for
each pseudo-Boolean function (see Appendix B). This is found by mapping
the negated Boolean variable as x 7→ (1 − x). For the GHZ- and W-functions
defined in Example 14 we arrive at the unique polynomials (14.5) and (14.6).

fGHZ (x1 , x2 , x3 ) = 1 − x1 − x2 + x1 x2 − x3 + x1 x3 + x2 x3 (14.5)

fW (x1 , x2 , x3 ) = x1 + x2 + x3 − 2x1 x2 − 2x1 x3 − 2x2 x3 + 3x1 x2 x3 (14.6)

These polynomials (14.5) and (14.6) are readily translated into categorical
tensor networks.
Example (Network realisation of W- and GHZ-states). A network realization of
W- and GHZ-states in our framework then follows by post-selecting the rele-
vant network to |1i on the output bit — leaving the input qubits to represent
a W- or GHZ-state respectively. An example of this is shown in Figure 12.

(a) = (b)

=
Figure 12. Left (a) the circuit realization (internal to the triangle) of the function fW
of e.g. (14.3) which outputs logical-one given input |x1 x2 x3 i = |001i, |010i and |100i and
logical-zero otherwise. Right (b) reversing time and setting the output to |1i (e.g. post-
selection) gives a network representing the W-state. The naïve realization of fW is given
in Figure 14 with an optimized co-algebraic construction shown in Figure 14.

– 63 –
Figure 13. Naïve CTNS realization of the familiar W-state |001i + |010i + |100i. A
standard (temporal) acyclic classical circuit decomposition in terms of the XOR-algebra
realizes the function fW of three bits. This function is given a representation on tensors.
As illustrated, the networks input is post selected to |1i to realize the desired W-state.

(a) (b)
= =

Figure 14. W-class states in the categorical tensor network state formalism. (a) is the
standard W-state. (b) is found from applying De Morgan’s law (see Section 13.3) to (a)
and rearranging after inserting inverters on the output legs. Notice the atemporal nature
of the circuits, as one gate is used forwards, and the other backwards.

Two different categorical constructions for the building blocks of the W-state
are shown in Figure 13 and Figure 14. Notice that in Figure 14 the resulting
tensor network forms an atemporal classical circuit and is much more efficient than
the naïve construction in Figure 13. Moreover by appropriately daisy-chaining the
networks in Figure 14 we construct a categorical tensor network for an n-party W-
state as shown in Figure 15. The resulting form of this tensor network is entirely
equivalent (up to regauging) to the MPS description given earlier, but now reveals
internal structure of the state in terms of CTNS building blocks.

Figure 15. W-state (n-party) in the categorical tensor network state formalism. The
comb-like feature of efficient network contraction remains, with the internal structure of
the network components exposed in terms of well understood algebraic structures.

– 64 –
15 Discussion
We have introduced a class of quantum states, known as Boolean quantum states.
This class is of interest, since it allows one to study quantum states using the well
understood Boolean algebra. In addition, states in this class have an evident tensor
network.

Theorem 15.1 (The Class of Boolean Quantum States [22]). Every switching
function f (x) gives rise to a quantum state with binary coefficients in {0, 1}.
Moreover, a tensor network representing this state is determined from the
classical network description of f (x).

Remark (From composition to contraction). The quantum tensor network is

found by letting each classical gate act on a linear space and from changing
the composition of functions, to the contraction of tensors.
Remark (Shannon and Davio decomposition). There are several elegant meth-
ods that allow one to factor a classical function into networks of different
structures. We should only have time to briefly mention these. Those inter-
ested can consult wikipedia for Shannon and Davio decompositions.
We have examined in some detail a Boolean tensor tool box. This is perhaps
the most accessible part of tensor network states, due to its strong relation to
Boolean circuits and widespread awareness of techniques to manipulate standard
Boolean networks. This tool box forms the glue or building blocks behind the more
complicated applications.

16 Problems

Exercise 16.1 (LOCC equivalence). Show that the AND-tensor is locally

bit-flip equivalent to NAND-, NOR-, and OR-tensors.

Example (Two-site reduced density operator of n-party GHZ-states). GHZ-states

on n-parties have a well known matrix product expression given as

|0i 0 · · · 0
 n 

 0 |1i · · · 0 
 
|GHZn i = Tr  . .. .. 
 
.
 . · · · . (16.1)
 
. 
 

0 · · · 0 |d − 1i
= |0i |0i |0i + |1i |1i |1i + · · · + |d − 1i |d − 1i |d − 1i .

Such MPS networks are known to be efficiently contactable. We note that the
networks in Figure 13.6 do not appear a priori to be contractible due to the
number of open legs. What makes them contractible (in their present from)
is that the tensors obey the fusion law allowing them to be deformed into a
contractible MPS network. The reduced density matrix of an n-party GHZ-
state then becomes (a) in Figure 16 and the expectation value of an observable
is shown in (b) where we included the normalisation constant.

– 65 –
 =

The GHZ-state tensor is simply a order-n COPY-tensor. Node equivalence

implies that this tensor can be deformed into any network geometry including
a MPS comb-like structure (right).

(a) (b)
= =

Reduced density operator. Left (a) reduced density operator ρ0GHZ found from
applying the fusion law to a n-qubit GHZ-state. Right (b) the expectation
value of observable O1 ⊗ O2 found from connecting the observable and con-
necting the open legs (i.e. taking the trace).

Exercise 16.2. Write down the matrix product state for the n-party GHZ
and n-party W-states. Let us define the two-point correlation as

Ci,j (ψ, Ai , Aj ) = hψ| Ai Aj |ψi − hψ| Ai |ψi hψ| Aj |ψi (16.2)

Find C(j ≤ n) := C1,j (ψ, X1 , Xj ) for both GHZ- and W- where X is the
familiar Pauli matrix. Using Mathematica, make a publication quality plot
for some fixed n (e.g. label everything and create a caption explaining the
plot).

Exercise 16.3 (Basic properties of Boolean quantum states). This exercise

considers some elementary properties of Boolean quantum states.

(i) Count the number of Boolean quantum states on n-qubits. Write truth
tables for all two-input boolean functions and label appropriately those
columns corresponding to NOT, AND, XOR, XNOR, NAND, NOR, and
OR.

(ii) Let ψB2 = c0 |00i+c1 |01i+c2 |10i+c3 |11i represent a boolean quantum
state, and hence ∀i, ci = 0, 1. Also let the output of all possible two-
bit functions be given by a truth vector c = (c0 , c1 , c2 , c3 ) and classify
entangled vs non-entangled boolean states on two qubits.

(iii) Hence, using the result in (ii) or otherwise, show that the possible
values of entanglement (using the quantity K1 or J2 ) is course grained
and give the possible values.

Example (valence-(0,3) tensor factorization). Every tensor in A ⊗ A ⊗ A has

a factorization as: (i) two valence-(1,0) tensors in A (e.g. the triangles); (ii)
two valence-(2,1) COPY-tensors; (iii) two unitary morphisms of type A → A
(white boxes). This factorization is given diagrammatically as follows.

– 66 –
 = =

Exercise 16.4 (MPS factorization of the AND-state). Recall the MPS fac-
torization of quantum states covered in lecture I. Present an MPS factoriza-
tion of the AND-state

ψ∧ = |000i + |010i + |100i + |111i (16.3)

into five elementary tensors, as illustrated in Example 16. Write this as a

matrix product. (Note: it is acceptable to use a computer to calculate the
SVD. Another method follows from considering the eigenvalues of reduced
density states.)

Exercise 16.5 (The class of linear quantum states). We define the linear
class of quantum states as quantum states of the form

ψ⊕L = (16.4)
X
c1 x1 ⊕ c2 x2 ⊕ ... ⊕ cn xn |x1 , x2 , ..., xn i .

Here we will consider a 1D system. What is the maximum possible χ (as

defined in lecture I) in a partition in one of these states?

Exercise 16.6 (Correlations in polarity states (Optional)). Here let us con-

sider a state defined as

ψ := (−1)f (x) |xi , (16.5)

X

where the sum is over all x. These states are often considered in quantum
algorithm theory. Prove that ψ is separable iff the function implements a
linear function.

Definition 16.1 (Boolean Density Operators). In general, a Boolean density

matrix takes the following form.

ρB = f (x)f (y) |xi hy| =

X

= f (x1 , ..., xk , ..., xn )f (y1 , ..., yk , ...., yn ) |x1 , ..., xk , ..., xn i hy1 , ..., yk , ..., yn |
X

(16.6)

Exercise 16.7 (The Boolean trace theorem). Show that performing the
partial trace over the kth subsystem of a Boolean density state results in

Trk ρB = δxk ,yk ρB =

= f (x1 , ..., xk , ..., xn )f (y1 , ..., xk , ..., yn ) |x1 , ..., xn i hy1 , ..., yn |
X

(16.7)

– 67 –
Exercise 16.8 (Two-qubit entanglement). To prepare for the next problem,
here we will continue our study of two-qubit entanglement, by considering
again the results in Lecture I.

(i) By considering a general quantum state

α |00i + β |01i + γ |10i + δ |11i , (16.8)

and by using wire bending duality, construct the induced matrix M .

(ii) Show that Det(M ) vanishes identically for α = ac, β = ad, γ = bc and
δ = bd with a, b, c, d ∈ C and factor states of this type into a product
of local states φ1 (a, b)φ2 (c, d).

(ii) Using this result from (ii) or otherwise, show that Det(M ) vanishes
iff the the state giving rise to M under wire duality takes the form
φ1 (a, b)φ2 (c, d).

Exercise 16.9 (Quantum AND-tensors). With the contractions in Figure 5

(a) and (b) in mind, consider instead the contraction with the state

α |0i + β |1i , (16.9)

and hence, form an order-two tensor T with components take values in α

and β.

(i) Write down the quantum state resulting from this contraction. By
bending a wire, write down the resulting two by two matrix, and label
this M (α, β).

(ii) By considering the matrix determinant or otherwise, find the values

that α and β must take for the state to be separable.

(iii) Consider now products of M (α, β). Give a one sentence proof or dis-
proof of the following statements: The products of M for (a) a magma
(or groupoid); (b) a semigroup; (c) a monoid.

(iv) Are there values of α and β that make a state local unitary equivalent
to a bell-state? Are there values of α and β that can break leg exchange
symmetry in the state?

Exercise 16.10.
(i) Write down the functions for a two-party bell state, a GHZ and a W
state. Use the trace theorem to provide analytical closed formula for
the reduced density states found from tracing over the last bit. Expand
this as a matrix and compare to the result obtained using standard
methods.
(ii) Using standard properties of Boolean algebra, show that ρ2B = ρB and
hence that ρB is a projector.

– 68 –
(iii) Show that U = 1 − 2ρB is self-adjoint and unitary. Give values for the
matrix trace and determinant of U .

(iv) Show that (1 − ρB ) and (ρB ) project onto eigenspaces of U and relate
the dimension of the subspaces to properties of the boolean function.

Exercise 16.11. Define the Werner states acting on C2 ⊗ C2 as

1−r
ρr = r φ+ φ+  + 1 (16.10)
 ED 
 
4
√  
where 2 φ+ = |00i + |11i is the standard Bell state and r takes only
values in the real interval [0, 1]. Find (i) the matrix trace of ρr and (ii) the
eigenvalues of ρr . The concurrence is

C(ρ) = max{λ1 − λ2 − λ3 − λ4 , 0} (16.11)

where λ1 ≥ λ2 ≥ λ3 ≥ λ4 Using part (ii) or otherwise, find C(ρr ).

Exercise 16.12 (The Shannon effect — Research). For many fundamental

functions one can design efficient circuits. Is it possible to compute each
function by an efficient circuit? In 1949 Shannon proved that almost all
functions are hard functions, optimal circuits for almost all functions have
exponential size and linear depth. This was proved quite easily using a
counting argument [65]. The number of circuits with small circuit size or
small depth grows much slower than the number of different Boolean func-
tions implying that almost all functions are hard. This means that a random
Boolean function is hard with very large probability.

(i) The circuits we consider here don’t have the same temporal structure
enforced by classical gates. In that regard, we have seen that the
function for the W-state can be realized using two COPY-gates, one
AND, and one OR.

(ii) With this observation in mind, what can we say about the Shannon
effect in the present case?

– 69 –
 PART

Symmetries and Stabilizer Tensor IV

Theory
17 Introduction
In quantum theory, one predicts experiments by performing calculations involving
the processes that transform the mathematical representative of a quantum system,
such as operators on a state space or an observable. This mathematical structure
has proven very rich, and consequently well studied. It is often the case that
mathematical structures, such as symmetries or other algebraic properties, allow
us to understand much more about a situation at hand.
The theory of Penrose tensor networks leverages one to study the mathematical
structure formed by the composition of processes themselves. Tensor networks are
not just a new conceptual tool, but in fact are becoming applicable across a range of
disciplines. Primarily they have practical applications in the context of condensed
matter and many-body physics. There are still fundamental questions related to
these tools, left almost entirely unexplored.
In this lecture, we will examine some basic aspects of symmetry in tensor
network states.8 8 When describing known results
in quantum information, tensor
network practitioners should hope
18 Symmetries in Tensor Network States to find simple and elegant rewrites
between structurally intuitive and
Let ψ ijk be a tensor with valence-(3,0). In the graphical tensor notation of Penrose aesthetically pleasing representa-
tions.
[1] (who also often used triangles to depict quantum states) we represent this as

When we refer to tensor symmetry, we typically mean under index exchange. For-
mally,

Definition 18.1 (Exchange or index symmetric). We say a valance-(n,m) tensor

is wire exchange symmetric if one can exchange any of the arms amongst
themselves, or any of the legs amongst themselves, with the effect leaving the
tensor unchanged.

Definition 18.2 (Full symmetry). We say a valence-(n,m) tensor is fully sym-

metric if in addition to being exchange symmetric per Definition 18.1, the
tensor is left invariant under arm and leg exchange.

Remark (The trivial representation). We say that an exchange symmetric tensor

of valence-(n,m) carries the trivial representation of the symmetric group of
order n! (arm exchange symmetric) and order m! leg exchange symmetric.
The diagrammatic generator of the symmetry group is expressed in (a).

70
 (a) (b)
=

Here (a) represents the tensor δji δlk and (b) represents the contraction

(δji δlk )(δji δlk ) = δii δkk . (18.1)

In diagrammatic form, this exchange operator is used to generate the group of

operators S1 in (a), S2 in (b) and S3 of order 3! in (c).
(a) (b) (c)

Returning now to ψ ijk . We say ψ ijk is index symmetric provided it is left

invariant when acted on with any operator in (c) above.
Remark (Symmetry breaking). Contraction of symmetric tensors breaks sym-
metry. Given two tensors Γ1 and Γ2 satisfying Definition 18.2 the contraction
of Γ1 and Γ2 over one or more indices is not necessarily symmetric. Exam-
ples abound. The sum of tensors (defined for tensors of the same valance) is
symmetric.
Symmetrizers. Let us consider methods to perform the symmetrization of tensors.
In the definition we avoid clutter by letting the reader figure out how the group
should act on the arms and legs of the tensor.

Definition 18.3 (Symmetrizer). Let us define the operator, acting on appropri-

ate types
1 X
RG : Γ 7→ g{Γ} (18.2)
|G| g∈G

then RG {Γ} is necessarily symmetric under the group G.

Example (Example of symmetrization). Consider φijk which is not necessarily

symmetric. We will symmetrize the first two indices labeled i and j under S2 .
In this case we find
1 1
RG {φijk } = φijk + φjik . (18.3)
2 2
The shorthand notation for this procedure and expansion is to write round
brackets over the indices we are symmetrizing as φ(ij)k .
Remark (Symmetrization can vanish). It is possible to have a tensor vanish
under RG . Examples include the antisymmetric state ψ = |01i − |10i. Here
1 1 1 1
RG {ψ ij } = ψ ij + ψ ji = ψ ij − ψ ij . (18.4)
2 2 2 2
Example (Symmetric over other groups). We have defined RG in a way where
we can pick groups G other than the symmetry group acting on the tensor
indices. Consider say G = {H ⊗ H ⊗ H, 1} where H 2 = 1 is the standard
Hadamard transformation. We consider the valence-(1,2) COPY-tensor δ i jk
and let RG act as
1 1
RG {δ i ij } = δ i jk + ⊕i jk (18.5)
2 2

– 71 –
 here ⊕i jk and from now on, ⊕i jk denotes the XOR-tensor. The resulting tensor
still has three indices. This operation is given diagrammatically as

Antisymmetrizers. One can also consider antisymmetrizers. These are used reg-
ularly in the study of fermionic particles for instance.

Definition 18.4 (Antisymmetrizer). We define the operator of antisymmetriza-

tion as
1 X
A= (−1)π g (18.6)
N ! g∈S
n

where π = 0 for g even and π = 1 for g odd.

Example (Antisymmetrization). Consider the tensor Γij of valance-(2,0). Then

1 1
A{Γij } = Γij − Γji (18.7)
2 2
and provided the tensor A{Γij } is non-zero, it is said to carry the sign rep-
resentation of S2 . It is standard to write antisymmetrization over indicies,
by including them in square brackets. The example here would then be
Γ[ij] = A{Γij }.
Remark (Antisymmetrization can map to zero). Examples abound of tensors
which map to zero under A. For example, there is no fully antisymmetric
tensor on (C2 )⊗3 .
Example (Could consider other groups). We will let

T ijk = δ i jk − ⊕i jk (18.8)

then if we consider action under

G = {H ⊗ H ⊗ H, 1}, (18.9)

we have that the action

H ⊗ H ⊗ H{T ijk } = −T ijk (18.10)

and
1{T ijk } = T ijk . (18.11)
Let us now consider some of the symmetries present in the tensors we have
defined in the first two lectures. We will start with the COPY-tensor.

– 72 –
This is written using standard quantum theory notation as

(18.12)
X
δ i jk |jki hi| .
ijk

As we recall, copy is fully symmetric. In addition,

Remark (The types of possible states built from COPY). A concise definition of
the COPY-tensor is given as any combination of raised or lowered indices on
δ ijk , a Kronecker delta function on three indices. In step, one might write the
n-party GHZ-state as

ψGHZ = (18.13)
X
δ ijk...l |ijk...li .

Tensor products of states of this form are precisely the only types of states
constructible with the COPY-tensor alone.

(a) (b) (c) (d) (e)

= = = = =
=

Figure 16. Diagrammatic properties of the COPY-tensor [46]. (a) Full-symmetry. (b)
Copy points, e.g. |xi 7→ |xxi for x = 0, 1 for qubits. (c) The unit given as h+| = h0| + h1|.
def

(d) Applying the unit to the top leg creates a cap. (e) Copy then delete is the same as
identity.

We are going to study these diagrammatic equations from Figure 16 under

local unitary change of basis. We will first define the rotation operator

Definition 18.5 (Tensor rotation operator). Let us define a rotation operator

which acts on a valence-(n,m) tensor as follows. On the n input arms, U is
applied, on the m output legs, U † is applied. We will define this map as UG
which is given diagrammatically as follows

Rotations of COPY-tensors. It is of course desirable to extend the definitions of

the valence-(1,2) COPY-tensors to act on arbitrary bases. We would then change
to a basis B as

Example (Engineering a mapE to copy a particular basis). Say you have an or-
thonormal basis {|ψi , ψ }, which we will call B and you wish to define a
 ⊥

COPY-tensor on this basis as |ψi 7→ |ψ, ψi and ψ ⊥ 7→ ψ ⊥ , ψ ⊥ . To repre-

 E  E
 

sent such a tensor diagrammatically we draw

– 73 –
To transform the COPY-tensor to copy this new basis, define the unitary

U= (18.14)
X
|ii hφi |
i

where {φi }i and {i}i are bases for the same space such that hφi |φj i = δij and
hi|ji = δij . The map U is unitary as

U †U = ( |ji hφj |) = |φi i δij hφj | = 1. (18.15)

X X X
|φi i hi|)(
i j j

Remark. There are important cases were we only consider transforming

COPY-tensors using self adjoint unitary maps. Hadamard is an example with
1
H = √ (X + Z) = H † . (18.16)
2
Every projector gives rise to such a map as U = 1 − 2P is unitary when
P 2 = P . The map is invertible, and hence every self adjoint unitary gives rise
to a projector, illustrating the bijection between self adjoint unitary maps and
projects.
Example (Transforming COPY-tensors into XOR-tensors). It is well known in
the theory of algebra that Hadamard transforms (the Fourier transform over
Z2 ) relate COPY-tensors and XOR-tensors. This is also known in quantum
circuits. For the higher dimensional case see [23]. The transformation of a
COPY-tensor into an XOR-tensor is given diagrammatically as follows.

Remark (Arbitrary rotations). We will now mention exactly how the general
form of the diagrammatic laws can transform properly under rotation. The
general form is straightforward. States rotate under U † and tensors rotate as
we have already defined.

(a) (b)
= =

Exercise 18.1 (General rotation). Prove that if the COPY-tensor is rotated

as by U above, then the diagrammatic identities hold.

– 74 –
19 The Interaction of Networks Comprised of δ ijk and ⊕ijk
We will now consider networks comprised of COPY-tensors and XOR-tensors. Let
us first recalls DeMorgan’s law.

Definition 19.1 (DeMorgan’s Law). DeMorgan’s law is a relationship between

logical AND, (∧) and logical OR, (∨). The relationship is found by considering
what in quantum circuit theory would be called bit flips (X), induced by
logical negation (¬). The equation reads

¬(a ∧ b) = (¬a) ∨ (¬b) (19.1)

In diagrammatic form, this equation becomes

we note that X 2 = 1.

Remark (Relating COPY and XOR-tensors). A similar structure to the De-

Morgan’s law relating AND and OR-tensors also holds between COPY- and
XOR-tensors. In this case, instead of using negation, we use the Hadamard
gate H. In diagrammatic form, this becomes

and we note that H 2 = 1.

To continue, let us recall the definitions of the tensor at hand. In constraint
equation form, together with graphical representations, these read as

from the constraint equations, we can readily construct, a sort of “truth table”,
which gives the value of the tensor contraction, provided we contract the wires
(labeled i, j and k) with states |0i or |1i. We illustrated these possibilities in the
contraction table below.
i j k ⊕ijk δijk
|0i |0i |0i 1 1
|0i |0i |1i 0 0
|0i |1i |0i 0 0
|0i |1i |1i 1 0
|1i |0i |0i 0 0
|1i |0i |1i 1 0
|1i |1i |0i 1 0
|1i |1i |1i 0 1
– 75 –
 Exercise 19.1 (Linearity of tensor contraction). Tensor contraction is lin-
ear. Given the table above, use linearity to determine the contraction found
from other states, such as say |+i, |−i, etc.

To consider further properties relating the interaction of COPY-tensors and

XOR-tensors we will recall Lafont’s presentation of the Boolean calculus [22, 46],
given in Figure 19.2.

Illustration 19.2

=
= (f) =
= =

(a) (b) (c) (d) (e)

=
= = (g) =
= =

Lafont’s presentation of the linear fragment of the Boolean calculus [46]. (a) associativity; (b) unit laws; (c)
input symmetry (as mentioned, these tensors are in fact fully symmetric by Definition 18.2); (d) bialgebra
law; (e) illustrates that the unit for XOR is a copy point for the COPY-tensor and vice versa; (f) the inner
product h0|+i ∈ C. Global scale factors are represented as blank space on the page; (g) is the Hopf law.

These rules, and also the algebraic properties of XOR-algebra result in the
following class of functions (the only type possible to construct using the operations
at hand.

Definition 19.2 (Linear and affine boolean functions). Linear Boolean functions
take the general form

f (x1 , x2 , ..., xn ) = c1 x1 ⊕ c2 x2 ⊕ ... ⊕ cn xn (19.2)

where the vector (c1 , c2 , ..., cn ) uniquely determines the function. The affine
Boolean functions take the same general form as linear functions. However,
for functions in the affine class, variables can appear in both complemented
and uncomplemented form. Affine Boolean functions take the general form

f (x1 , x2 , ..., xn ) = c0 ⊕ c1 x1 ⊕ c2 x2 ⊕ ... ⊕ cn xn (19.3)

where c0 = 1 gives functions outside the linear class. From the identities,
1 ⊕ 1 = 0 and 0 ⊕ x = x we require the introduction of only one constant (c0 ).

When contracted, XOR- and COPY-tensors compose to create tensor networks

representing the class of linear quantum networks. This gives rise to the following
class of quantum states.

Definition 19.3 (The class of linear quantum states). We define the linear class

– 76 –
 of quantum states as quantum states of the form

ψ⊕L = (19.4)
X
c1 x1 ⊕ c2 x2 ⊕ ... ⊕ cn xn |x1 , x2 , ..., xn i

Exercise 19.3. Here we will consider a 1D system. What is the maximum

possible χ (as defined in lecture I) in a bi-partition of a linear quantum state?

Exercise 19.4 (Correlations in polarity states (Optional)). Here let us con-

sider a state defined as
ψ := (−1)f (x) |xi (19.5)
X

where the sum is over all x. These states are often considered in quantum
algorithm theory. Prove that ψ is separable iff the function implements a
linear function.

Example (Examples of linear quantum states).

Φ+ = a ⊕ b |a, bi = |01i + |10i (19.6)

X

Ψ+ = a ⊕ ¬b |a, bi = |00i + |11i (19.7)

X

H ⊗ H ⊗ H |GHZi = a ⊕ b ⊕ c |a, b, ci = |010i + |100i + |001i + |111i (19.8)

X

The connection to quantum circuits. Lafont’s 2003 paper [46] considered an

algebraic theory of classical switching networks, but also considered certain aspects
of quantum networks. We also mention the results in [23]. These ideas are readily
applied to quantum circuits, and combined with the known gate identities appearing
in text books on quantum information science. The development was in part,
influenced by quantum circuits [24], developments in the use of tensor networks in
condensed matter, development of the ZX-calculus [47] as well as Lafont’s influential
work [46].
The structures in Figure 19.2 are also related to other approaches [60–62, 77].
which have been used as a graphical language for measurement based quantum
computation [61] and for graph states [62]. Our method of arriving at this collection
of tensors (Figure 19.2) affords more general options and our presentation of the
linear fragment here offers (i) improved semantics and (ii) a better theoretical
understanding by pinpointing precisely that these networks correspond to the so
called linear fragment of the XOR-algebra [22].
Remark (From CNOT to COPY-tensors or XOR-tensors). The following figure
illustrates the contractions needed to transform a CNOT-gate into either a
COPY-tensor (top) or an XOR-tensor (bottom).

– 77 –
 = =

= =

These types of identities are common when considering interacting quantum

observables [47] as well as quantum circuits [23].
Example (GHZ-class circuits). We can realize the GHZ-sate by the following
circuit:

The simplification from left to right is a sequence of contractions [23], which

recover the familiar form of the COPY-tensor. On the other hand, one could
also realize GHZ by bending a wire as follows.

These circuits scale to create n-qubit GHZ-states in the evident way.

Example (Transformations of the controlled Z-gate). The following illustrates
the well known identity that there is a symmetry between the control and
target on a controlled Z-gate. This sequence of rewrites is an application of
the DeMorgan’s type law relating COPY- and XOR-tensors.

Such a transformation is common in applications of graphical rewrite systems

to measurement based quantum computation [61].
Now we will recall Lafont’s diagrammatic presentation of the bialgebra law,
between contractions of COPY- and XOR-tensors.

Definition 19.4 (COPY and XOR-tensors form a bialgebra [46]). As mentioned,

we work up to a global scalar in many of the diagrams in this work. One
can think of the bialgebra law as a type of commutation relationship amongst
tensor pairs, satisfying the law. Turning it side ways we have

– 78 –
 Exercise 19.5 (COPY and XOR-tensors form a bialgebra [46]). Verify that
COPY and XOR-tensors form a bialgebra [46], up to a global scale factor,
which should be determined.

We will now relate this abstract definition to well known quantum circuit iden-
tities. Let us first recall the well known factorization of the swap gate into a triple
product of CNOT-gates.
Example (Factorization of the swap gate). The CN-gate together with its hor-
izontally mirrored pair allows one to construct the swap gate as follows.

= =

Theorem 19.6 (Relating swap and the bialgebra law). In the past, we have
provided a diagrammatic proof that the square of the controlled not gate is
equal to the identity. This in turn allows one to relate the swap gate (using
its factorization above) and the bialgebra law.

So the factorization of swap together with CN2 = 1 is enough to have the

bialgebra law on tensors.

We will now introduce the gate-copy rewrite rule. Gate-copy allows one to
pull controls and targets through each other. When this happens, they are copied,
along with the attaching wires, leaving the attaching tensor intact.

Theorem 19.7 (Gate-copy). The following graphical rewrites in (a) and (b) hold.

(a)
(b)

Proof. The proof of gate-copy follows from application of the bialgebra and fusion laws, as follows.

– 79 –
 Example (Circuit simplification using gate-copy). Here we apply gate-copy to simplify the circuit from [24]
designed to simulate time evolution under the σ z σ z σ z Hamiltonian.

Starting from the circuit from Figure 4.19 on page 210 of [24], we apply a sequence of transformations including
the Gate-copy reduction rule introduced in Theorem 19.7. The network resulting from the simplification appears
in the bottom right.

Hopf law. We recall Lafont’s diagrammatic form of the Hopf-law [46].

Theorem 19.8 (COPY- and XOR-tensors satisfy the Hopf-law [46]). The follow-
ing diagrammatic equations, reproduced from [46], depict the Hopf-law, which
is satisfied by COPY- and XOR.

Exploration of the consequences of this identity in terms of quantum circuit

– 80 –
 manipulation can be found in the ZX-calculus [47] as well as other works on
categorical models of quantum circuits [23].

Using the Hopf law, we can justify a well known gate identity. This identity
was generalized to arbitrary finite dimensions in [23].

= = = =

We have considered the key properties and defining equations of the XOR- and
COPY-tensors. We will use these results as building blocks for the sections that
follow on from here.

20 Stabilizer Tensor Theory

Let us recall the notation of a stabilizer. “We may always depend on
it that algebra, which can-
Definition 20.1 (General stabilizer). A unitary map U stabilizes a quantum not be translated into good
state ψ iff English and sound common
U ψ = (+1)ψ. (20.1) sense, is bad algebra.” —
William Kingdon Clifford
Stabilizers of ψ from a group represented trivially by ψ.

We are concerned here with a subclass of the above definition. We are con-
cerned with what are commonly known as “stabilizer states”.

Definition 20.2 (Stabilizer state). An n-qubit stabilizer state ψ is defined by

n commuting and different operators S formed from the Pauli algebra with
presentation
{1, X, Y, Z, ±i, ·, ⊗}. (20.2)
These n operators generate the stabilizer group of order 2n for ψ. Clearly we
have that sψ = ψ, ∀s ∈ S.

Example (Single qubit stabilizer states). Here are the single qubit states stabi-
lized by the Pauli-group

(i) Pauli-X: σ x stabilizes |+i = |0i + |1i and −σ x stabilizes |−i = |0i − |1i
(ii) Pauli-Y: σ y stabilizes |y+ i = |0i + i |1i and −σ y stabilizes |y− i = |0i −
i |1i
(iii) Pauli-Z: σ z stabilizes |0i and −σ z stabilizes |1i
The following is simply a reminder of the essential properties of the Pauli
operators. These are useful for general knowledge and for working through the
details of later calculations, but could be skimmed on a first read.

Definition 20.3. Recall from angular momentum theory, the familiar Pauli
Matrices. We let σ1 ≡ σx , σ2 ≡ σ y and σ3 ≡ σ z which satisfy

(i) [σi , σj ] = 2iijk σk

– 81 –
 (ii) complex conjugation is generated by σj ∀w ∈ {i, j, k} as σj σw
∗ σ = −σ
j w
for w 6= j

(iii) Tr(σi σj ) = 2δij

Note that it is common to change subscripts to superscripts σx ≡ σ x to distin-

guish powers of operators σx2 from operators acting on specific indices σ3x —
that is, σx acting on the third qubit and not σx cubed. This should be evident
from context.

These familiar operators from quantum mechanics form what is called a Ge-
ometric Algebra (a.k.a. Clifford Algebra). Consider {σil } as a basis for a real left
and right distributive vector space, such that

σx2 = σy2 = σz2 = 1 = −iσx σy σz . (20.3)

Then consider the product

σi σj = δij 1 + iijk σk (geometric product of vectors). (20.4)

It is clear that for i 6= j

{σil , σjl } = 0 (20.5)
meaning that the vectors anti-commute viz.,

σil σjl = −σjl σil . (20.6)

We also note that

{σ i , σ j } = 2δij 1. (20.7)

Exercise 20.1. Verify Equations (20.4), (20.5), (20.6) and (20.7).

Single qubit density operators can be expanded in terms of a dot product of a

polarisation vector P := (p1 , p2 , p3 ), ∀i, pi ∈ R, and a sigma vector σ := (σ1 , σ2 , σ3 )
as
1 1
ρ = 1 + σ.P = 1 + pi σ i = p1 σ 1 + p2 σ 2 + p3 σ 3 . (20.8)
2 2
Clearly the vectors P are in the vector space R3 . We can elevate this vector
space to a Hilbert space by defining an inner product between vectors — that is a
map (−, −) taking two elements from the vector space (in this case Hamiltonians
on C2 ) and producing a scalar in the underlining field C.

(−, −) : (C2 → C2 ) × (C2 → C2 ) → C

(σ.A, σ.B) 7→ T r[(σ.A).(σ.B)] = A.B (20.9)

Exercise 20.2 (Product of vectors). Verify that straight forward calculation

also yields

(σ.A, σ.B) = Ai Bi + iijk σ k = A.B + iσ(A ∧ B). (20.10)

– 82 –
 Exercise 20.3 (Eigenvalues). Verify, by using the fact that for A = B the
cross product vanishes or otherwise, that the eigenvalues of (A.σ) are ±|A|,
where both roots necessarily appear as A.σ is traceless.

We can relate symmetry in density operators and symmetry in states as follows.

Before continuing on to define the Clifford group of quantum circuits, we will
consider an example of a stabilizer state.
Example (The Bell state is a stabilizer state). The bell state

Φ+ = a ⊕ ¬b |a, bi = |00i + |11i (20.11)

X

is a stabilizer state with stabilizer group

S = {1, XX, −Y Y, ZZ} (20.12)

Note that this matches the following claims in the definition (i) the stabilizer
group is of order 2n , here n = 2, with generators given by e.g. XX, ZZ, etc.
and (ii) the group is abelian.

Exercise 20.4. Verify that XX, −Y Y and ZZ are in fact stabilizers of the
Bell state Φ+ graphically. Using the rules of the Pauli algebra, verify that
these operators commute.

Exercise 20.5. Find the stabilizers for the following states.

Φ− = |00i − |11i (20.13)

Ψ− = |01i − |10i (20.14)

Hint. One method would be to use the Pauli algebra and the stabilizes for
Ψ+ . First cancel the local operator separating the given states from Ψ− ,
apply the known stabilizer and then reapply the operator to return to the
starting state. Other methods exist.

21 The Clifford Group

There is another useful way of generating stabilizer states. We will consider a The following are equivalent.
quantum circuit acting on the state |0i⊗n . The circuit is comprised solely of Clifford
gates. 1. The ZX-calculus [25].

2. Clifford gates plus cups

Definition 21.1 (Clifford gates). The collection of Clifford gates is as follows.
and caps.
(a) CNOT; (b) the Hadamard gate H = √12 (X + Z); (c) The phase gate
P = |0i h0| + i |1i h1|, and (d) the Pauli gates generated by the Pauli algebra 3. Ability to bend wires
on a single qubits, with 16 elements (the group generated by {X, Y, Z}). and compose stabilizer
states.
(a) (b) (c) (d)

– 83 –
 Remark (Single qubit Clifford group). The standard properties of single qubit
gates follow.

(i) HXH = Z; and HZH = X

(ii) P XP † = Y ; and P Y P † = Z = P 2

These gates above generate the single qubit Clifford group.

With the definitions of the the Clifford group in place, one can state the alter-
native definition of stabilizer states.

Definition 21.2 (Stabilizer states). If ψ can be produced from the all-|0i state
by Clifford gates, then ψ is stabilized by 2n tensor products of Pauli matrices
or their sign opposites (where n is the number of qubits). This means that the
stabilizer group is generated by log(2n ) = n such tensor products. The state
ψ is then the stabilizer state uniquely determined by these generators.

Remark (Properties of Clifford circuits). We now list elementary properties of

Clifford circuits.

(i) Clifford circuits generate the Clifford group.

(ii) Let Pn be the collection of 4n n-letter words with ⊗ as concatenation

generated from the alphabet

{±1, ±X, ±Y, ±Z, ±i1, ±iX, ±iY, ±iZ}. (21.1)

All operators g in the Clifford group acts as an involution when Pn is

conjugated by g, that is gPn g † = Pn .

(iii) Note, alternative notation to write P1 could be

P1 = {±1, ±i}{1, X, Y, Z}. (21.2)

(iv) Defining properties of the Pauli matrices:

X 2 = Y 2 = Z 2 = 1 = −iXY Z. (21.3)

Exercise 21.1. Show that conjugation by H lifts to an involution on P1 by

considering (iv) with H = √12 (Z + X).

22 Stabiliser Tensor Theory

Let us cast the stabilizer theory into a theory of tensors, in the Penrose graphical
calculus. We will first define the notation of an abstract stabilizer for a tensor.

Definition 22.1 (Abstract stabilizer). Let Γ be a valence-(n,m) tensor. A sta-

bilizer for Γ is given by m + n local invertible operators, satisfying

– 84 –
 We then turn to a definition of tensors, which as we will soon show, generate
the Clifford circuits as a subclass.

Theorem 22.1 (Stabilizer tensors). Contraction of tensors taken from the fol-
lowing are sufficient to generate the stabilizer group.

(a) (b) (c) (d) (e)

Where (a) is the COPY-tensor, (b) is the y-plus-state |y+ i = |0i+i |1i, the cup
def

(c) and cap (d) are to bend wires and hence reshape maps and take transposes,
(e) is the Hadamard gate.
Proof. [Stabilizer tensors] Let us first consider generating the states |+i,
|−i, |y− i. These are found from the following tensor contractions.

(a) (b) (c)

=
= =

From these states, the following contractions generate our elementary

gates.
(a) (b)

= =

We then note that

HZH = X (22.1)
P =P P
† 2
(22.2)
and that
P XP † = P XP 3 = Y (22.3)
We recover the COPY-tensor by contracting with Hadamard gates and
CNOT is found from reshapes of the following contraction

We then recover the generators of the Clifford group.

Now we will consider stabilizers of the COPY-tensor.

– 85 –
 Example (Stabilizers of COPY). The COPY-tensor has stabilizer generators
σ1x ⊗ σ2x ⊗ σ3x and σiz ⊗ σjz which uniquely determine ψGHZ = |000i + |111i and
result in the following stabilizer group of order 23 = 8.

{σ x σ x σ x , −σ x σ y σ y , −σ y σ x σ y , −σ y σ y σ x , 1 ⊗σ z ⊗σ z , σ z ⊗ 1 ⊗σ z , σ z ⊗σ z ⊗ 1, 1}
(22.4)
Diagrammatically these relations are given in Figure 17.

= =

Figure 17. (Top) Diagrammatic depiction of the stabilizer equation σiz σjz (|000i + |111i) =
|000i + |111i. (Bottom) Uses the stabilizer identity together with σz2 = 1 to show that the
σ z commutes with the COPY-tensor. (Middle) Diagrammatic depiction of the stabilizer
equation σ1x ⊗ σ2x ⊗ σ3x (|000i + |111i) = |000i + |111i. (Bottom) Diagrammatic depiction
(up to a sign) of the stabilizer equation −σix ⊗ σjy ⊗ σky (|000i + |111i) = |000i + |111i.

Exercise 22.2 (Gottesman-Knill Theorem). A graphical rewrite proof (by

bounding the number of rewrites) of the Gottesman-Knill theorem follows
by considering the action of the COPY- and XOR-tensors on σ z and σ x .
Derive the gate identities in paper [78] and prove the main theorem using
the methods from this lecture.

Exercise 22.3 (Stabilizers for δijkl ). Write down the 24 stabilizers of the
state
δ ijkl |ijkli (22.5)
Now consider stabilizers in (a). We can expand this arriving at (b). What are
the conditions on C and D such that A, B, E, F are stabilizers? Compare
this with the stabilizers given for COPY, prior to contraction.
(a) (b) (c)
= =

We can now consider how stabilizers of a tensor transform, when the tensor
undergoes a local change of basis.

– 86 –
 Theorem 22.4 (Transformation properties of stabilizers). Let Γ be a tensor with
stabilizer A ⊗ B ⊗ · · · ⊗ C. Then if we rotate Γ as Γ0 = UG (Γ), then

UG (A ⊗ B ⊗ · · · ⊗ C) = UG (A) ⊗ UG (B) ⊗ · · · ⊗ UG (C) (22.6)

is a stabilizer for Γ0 .

23 Boolean Stabilizer States

Now one can consider the intersection found from the class of Boolean states which
are in addition to being Boolean states, also stabilizer states. We will construct
them explicitly for the case of a single qubit, and leave the two qubit case as
homework.

Single qubit Boolean Stabilizer States

There is one stabilizer state for each single qubit boolean state. This correspon-
dence is a special case of the dim = 2 state space and does not hold in any higher
dim. “To know all about anything
is to know how to deal with it
Single bit functions under all circumstances.” —
Consider William Kingdon Clifford
f : B2 → B (23.1)
then a general function f is expanded as a sum-of-products (LHS) and as a pseudo
Boolean form (RHS)
c0 x ∨ c1 x = c0 + (c1 − c0 )x (23.2)
for c0 , c1 , x ∈ B. There are exactly 22 boolean functions. Evaluated at n = 1 gives
n

four possible single qubit Boolean states. One of these however, corresponds to
ψ = 0 and so is trivial. The others are |0i, |1i and |0i + |1i.

Single qubit stabilizer operators

We will consider ±Z and X. These are the only possible stabilizer operators that
stabilize single qubit boolean states.

X(c0 |0i + c1 |1i) = c1 |1i + c0 |0i (23.3)

and hence, Xψ = +1ψ iff c1 = c0 . Over B this has non-trivial solutions (that is,
kψk > 0) for c0 = c1 = 1 and hence, we recover the only boolean state stabilized
by X as |0i + |1i. We will then consider

± Zψ = ±c0 |0i ∓ c1 |1i (23.4)

and ±Zψ = +1ψ iff ±c0 = ∓c1 . It follows that

+ Z ⇒ c0 = 1, c1 = 0 (23.5)

− Z ⇒ c0 = 0, c1 = 1 (23.6)

– 87 –
Correspondence between stabilizer states and boolean states
In this section we consider the correspondence between single qubit stabilizer states
and single qubit boolean states. Let us introduce two boolean variables, b0 and b1 .
We then write
(−1)b1 (1 − b0 )Z + b0 X (23.7)
We find that b0 ⇒ c0 = c1 = 1. The case that b0 = 0 implies that b1 decides c0 , c1
as
b1 = 1 ⇒ c1 = 1, c0 = 0 (23.8)
b1 = 0 ⇒ c1 = 0, c0 = 1 (23.9)
We will then parameterize the single qubit boolean state in terms of b0 and b1 as

ψB = (b0 ∨ b0 ∧ b1 ) |0i + (b0 ∨ b0 ∧ b1 ) |1i = (b0 + b1 − b0 b1 ) |0i + (1 + b0 − b1 + b0 b1 ) |1i

(23.10)
and hence the choice c0 = b0 + b1 − b0 b1 and c1 = 1 + b0 − b1 + b0 b1 provides the
correspondence.
Remark. In or around circa. 2010, I discussed development of classification
scheme for Boolean stabilizer states with Oscar Dahlsten. I believed Oscar
solved that problem, specifying exactly the class of Boolean stabilizer states.

Exercise 23.1 (Two qubit Boolean stabilizer states). For two qubits, there
are 22 − 1 = 15 boolean states. Let q and r take boolean values. Then we
2

arrive at the following truth table.

q r f0 f1 f∧
0 0 0 1 1 1 1 0 1 1 0 0 1 0 1 0 0 0
0 1 0 1 1 1 0 1 1 0 1 0 0 1 0 1 0 0
1 0 0 1 1 0 1 1 0 1 1 1 0 0 0 0 1 0
1 1 0 1 0 1 1 1 0 0 0 1 1 1 0 0 0 1

(i) Label OR, NOR, NAND, XOR in the table above.

(ii) Using the definition of K1 , determine which states in this table are
separable, and classify them based on the resulting values of K1 .

(iii) Which states in the table are not stabilizer states?

Exercise 23.2 (Invariants of symmetric Boolean states).

(i) Write the general form of a symmetric three qubit boolean state. How
many possible symmetric boolean states are there?

(ii) Consider the discriminant of the cubic.

∆ = a20 a23 − 6a0 a1 a2 a3 + 4a0 a32 − 3a21 a22 + 4a31 a3 (23.11)

and find maximum and minimum values when ∀i, ai ∈ {0, 1}.

– 88 –
24 Problems
Theorem 24.1 (Sufficient expression stabilizer states). Let

f, g, k : Bn → B (24.1)

then the quantum state

ψB = (−1)f (x) (i)g(x) k(x) |xi (24.2)

X

is sufficient to express any stabilizer state.

Proof. [Normal forms on stabilizer states] Each stabilizer state is an
equally weighted superposition with coefficients ±1, ±i, 0. The functions
f, g, k determine these for each basis vector |xi.

Remark (Normal forms). Using a PPRM from lecture II, we can expand f, g, k
to a normal form. The functions then become uniquely determined by a
coefficient vector.

Exercise 24.2 (Stabilizer states as pseudo Boolean forms). Find f , g, and

k to express the following states (note, the second is not a stabilizer state).

(i) ψ1 = |000i + i |111i

(ii) ψ2 = |001i + i |010i − i |100i

Exercise 24.3 (Stabilizer generators). Consider U as an arbitrary Clifford “An expert is someone who
circuit. Then, knows some of the worst
ψ = U |0i⊗n (24.3) mistakes that can be made
in their subject and how
is an arbitrary stabilizer state. Show that evolution of Zi under U in the
to avoid them.”— Werner
Heisenberg picture
Heisenberg
U Zi U † (24.4)
is necessarily a stabilizer for ψ.

Exercise 24.4 (Further Exercises on Pauli Matricies).

(i) Let P .σ := p0 X +p1 Y +p2 Z where |P | = 1. Show that exp −i 2θ P .σ =

 

1 cos(θ/2) − i(P .σ) sin(θ/2) and find the values of θ, P to recover the
Hadamard gate, up to a phase factor.

(ii) Find the time of the evolution of the Hamiltonian |11i h11| to create
a CZ-gate, then write down a quantum circuit in terms of H and CZ
to create a CNOT-gate. What are the input states needed to use the
CNOT-gate to prepare the singlet state |Ψ− i = |01i − |10i?

(iii) Show that the SWAP operator 12 (1 + σ A · σ B ) permutes the values of

bits A and B as
SWAP |iA i |iB i = |iB i |iA i (24.5)

– 89 –
 where the notation σ A · σ B is typically said to stand for the scalar and
tensor product: σ A · σ B = 3i=1 σiA ⊗ σiB .
P

(iv) In the computational basis, express the general form of a two-qubit

symmetric eigenstate of the SWAP operator and count the real de-
grees of freedom. Repeat this for anti-symmetric eigenstates (e.g.
SWAP|ψi = − |ψi).

(v) Using the notation from (iii) above, find a value for q to show that
the two-site quantum Heisenberg model Jσ 1 · σ 2 can be written as
J
(σ + )2 − q 1 and show that |Ψ± i = |01i ± |10i are energy eigen-


2 1 σ 2
states.

– 90 –
 PART

Tensor Networks and Entanglement V

Invariants
It is typical to consider a symmetry of a density operator ρ or state ψ as an operator
satisfying
ρ = V ρV † (24.6)
ψ = Mψ (24.7)
and so one could say that ρ is invariant under V and ψ is invariant under M — see
Example V. If the relations holds for all elements of a matrix group, then we say
that ρ or ψ are invariant under the left multiplicative action of the group. There
are more subtle symmetries however. These are formed by considering polynomials
in the coefficients of ρ or ψ that are left invariant under the action of a group. The
study of such polynomials is known as Invariant Theory [79]. David Hilbert made
notable progress on the topic of invariant theory, which he perused throughout his
life. There has been past work on considering these invariants in the context of
quantum information science. Some of our personal favorites include [80? –82] as
well as some work more closely related to this chapter [26]. Algebraic geometry has
also been applied to tensor networks in [27].
Example (Group Symmetry of ρ). We will consider a general density operator
ρ and look for V ∈ U (d) such that

ρ = V ρV † (24.8)

this implies that [ρ, V ] = 0 and we arrive at a basis for G by noting the unitary
operators that commute with ρ. That is, {|λi i}i such that ρ = i pi |λi i hλi |.
P

It then follows from V V † = 1 that every V ∈ G can be written as

V = (24.9)
X
eiθi |λi i hλi |
i

Introduction to polynomial invariants

One can form polynomials out of the coefficients of a state or an operator. These
algebraic invariants, are called polynomial. For example, given a state with coeffi-
cients αij ,
ψ= (24.10)
X
αij |iji ,
ij

we could form a real valued polynomial function out of these variables αij and their
complex conjugates αij

f (α00 , α01 , α10 , α11 , α00 , α01 , α10 , α11 ) (24.11)

Acting on the state ψ with some linear transformation induces in turn an action of
this linear transformation on the polynomial f . We will be concerned with the case
that the linear transformation can be any element of a group G. If the polynomial

91
f in the coefficients of the state remains unchanged under the induced action of all
g ∈ G, then the polynomial is said to be a polynomial invariant under G.
For example, the polynomial J1 (24.12) corresponds to the norm of the state,
and is invariant under unitary transformations of ψ.

J1 := (24.12)
X
αij αij
ij

Remark (Basis independence). To form a polynomial out of the coefficients of

a state, one first chooses a basis to express the state in. The coefficients of
the state will change depending on the basis chosen. A polynomial invariant
that is invariant under any group that contains the local unitary group as a
subgroup is however inherently basis independent. The basis chosen to express
the polynomial depends on the local unitary group, however the polynomial
is invariant under the local unitary group, by construction.
Remark (Polynomial invariants are not state specific). As typical, symmetry is
state specific. We consider symmetries of a given state (Example V) or con-
sider a family of states that satisfy a symmetry property (so called, symmetric
states). In contrast to this, polynomial invariants, as typically considered, are
state independent invariants. They give rise to an invariant value for all states,
though this numerical value can be different when evaluating the polynomial
itself for different states. A polynomial invariant is invariant for all states
acted on by some group, but the numerical value can differ from state to
state. The numerical value of the invariant illuminates important properties
about the specific state in question.
Example (Example of Z2 invariance). Here we will illustrate properties of form-
ing polynomial invariants out of the coefficients of a state by considering a toy
example. Let
ψ= (24.13)
X
cx |xi
x

be a quantum state of two qubits. For the purpose of this example, we will
explore what a polynomial invariant is by considering a state-specific example
(in contract to remark V). That is, we will pick specific values of the cx ’s to
illustrate our point.
We must pick a matrix group G acting on states in C2 ⊗ C2 . Each g ∈ G in
turn induces an action on polynomials in cx as follows

cx 7→ hx, gψi, g ∈ G (24.14)

As an illustrative example, we pick G = {σx σx , 1} as the group G and

ψ2 = c00 |00i + c11 |11i (24.15)

for the state. We then consider polynomials

f (c00 , c11 , c00 , c11 ) (24.16)

that are invariant under G.

As stated, action of the group G on the state, induces an action on any
polynomial in the coefficients of the state via (24.14). For g ∈ G it is standard

– 92 –
 to write this action using the notation of putting g ∈ G in the superscript
(24.17).
f g (c00 , c11 , c00 , c11 ) := f ((c00 , c11 , c00 , c11 )g > ) (24.17)
then f 1 = f and

f σx σx (c00 , c11 , c00 , c11 ) = f (c11 , c00 , c11 , c00 ) (24.18)

It becomes clear that one can write certain polynomials f that are invariant,
e.g.
f = (c00 + c00 )(c11 + c11 ) (24.19)
under this group and others that are not, e.g.

f = (c00 + c11 )(c11 + c00 ) (24.20)

This example served to illustrate several key aspects of polynomial invariants,

but as this toy example is state specific and basis dependent, it does not
capture the philosophy and key aspects present in the more general setting.
A Graphical Language for Polynomial Entanglement Invariants
The invariants we have studied have all taken a remarkably simplistic form when
cast into our framework. The key to the simplification we have found was though
our introduction of the diagrammatic SVD in Theorem 7.2. In fact, the study of
invariants was a motivating factor which lead us to introduce this factorization,
which we soon found to have other applications. Through this factorization and
through other methods we are currently exploring, we have found that Penrose
graphical tensor notation and the invariant theory of Hilbert et al. connect very
well together. We are aware of a more cumbersome graphical language appearing
in the classic literature and reviewed in the book by Oliver [83], which he credits
to Cayley. We feel the approach taken here, facilitated by the diagrammatic SVD
and developed particularly to study entanglement invariants is significantly more
natural.

Invariants of pure qubit states

Tensor contractions for LOCC 9 invariants 9 LOCC: Local Operations and
Classical Communication.
We have seen in Example V that when we pick a finite group and a basis for
our state, one can form polynomials that are invariant under the group action.
The standard case, however, is to consider invariance under the local action of a
continuous group, and to consider this action on arbitrary states, of appropriate
dimension.

Definition 24.1 (LOCC Equivalence). Two states are LOCC equivalent iff they
are equivalent under local unitary transformations, that is, action of the group

GLOCC := U (1) × SU (d1 ) × SU (d2 ) × . . . × SU (dn ) (24.21)

Equivalence under LOCC yields a partitioning of states. Two states are in the
same equivalence class iff they are related by a local unitary transformation.
LOCC equivalence gives rise to a partitioning of the set of entanglement values.
An entanglement measure should remain constant on the equivalence classes.

– 93 –
 Remark (Single qubit invariants). When considering the local unitary group
acting on a single pure qubit, the only polynomial invariant is the norm of the
state
J1 = (24.22)
X
ψiψi
invariant under U (1) × SU (2). This is because the local orbit moves through
the angles on the Bloch sphere, and so the norm is the only possible ambiguity.
By fixing the norm, we fix the only invariant.
We have made adaptations to the diagrammatic language the utility of which
we will first illustrate by considering the case of two-qubits. These adaptations
allow for the confluent graphical contraction to evaluate invariants to equate them
in terms of singular values. This is illustrated by the following Theorem.
Lemma 24.5 (Graphical contraction of the norm J1 ). The graphical language
enables a sequence of tensor contractions to relate the polynomial invariant
J1 to the singular values of the state.
Proof. For pure states of two qubits, starting from Equation (24.12)

J1 := (24.23)
X
αij αij
ij

we arrive at the contracted network in (a) below. To show that this

contracts to a value determined by the singular values of the state, we
apply the diagrammatic SVD which exposes the internal network building
blocks shown in (b).

(a) (b) (c) (d)

= = =

The white boxes (valence-two tensors) represent unitary maps. They

therefore cancel resulting in the network illustrated in (c). The COPY-
tensor contracted as in (c) results in the identity wire (recall Figure (e) in
Definition 7.1). We then find the inner product of the triangular tensors
containing singular values,

⇒ J1 = λ2i = λ20 + λ21 (= 1 for a normalized state) (24.24)

X

Which is the desired result.

In addition to the norm J1 , for pure two-qubit states one finds another inde-
pendent invariant given as10 10 A way to understand why we
can expect only two invariants for
J2 = (24.25)
X
αij αkl αil αkj two pure qubits is by considering
the reduced density operator of a
The invariant is at quartic order and its tensor diagram is given by contracting two single qubit. This can be done
because there is only one possi-
copies of the state, and two copies of its dual under the dagger. ble partition for a two-qubit state.
This density operator has two sin-
gular values and we should be able
to find (algebraically independent)
invariants that separate the orbit,
based on these two singular values.

– 94 –
 Lemma 24.6 (Graphical contraction of J2 ). The graphical language enables a
sequence of tensor contractions to relate the polynomial invariant J1 to the
singular values of the state.
Proof. From the graphical expression of J2 (a) we arrive at (b) which
is found from applying the diagrammatic SVD to (a). The two pairs of
unitary order-two tensors cancel (c) and after contraction we are left with
a product of four valence-one triangles (d). The center portion, attached
to the four triangles (c) contracts to a single valence-four COPY-tensor
as11 11 This elementary property of

δ i jk δ qrs δ jsl δ rk m = δ iqlm (24.26) contraction, can also be seen as the

special case of a more general nor-
J2 thus becomes a product of singular values and hence (b) provides a mal form, see page 64 of [71]. It
graphical variant of (24.25) given as can also be seen as a tensor ver-
sion of node equivalence in digital
circuits [22].
J2 = λ4i = λ40 + λ41 = 1 − 2λ21 + 2λ41 (24.27)
X

where the last step follows from the constraint on the states norm.

(a) (b) (c) (d)

= = =

These graphical contractions inspired the following derivations.

Lemma 24.7 (General formulas equating state coefficients to singular values).
Given a general two-qubit state

|Ψi = α |00i + β |01i + γ |10i + δ |11i (24.28)

it can be shown that the Schmidt coefficients are given as

1
q 
λ21,2 = 1 ± 1 − 4|βγ − αδ|2 (24.29)
2
where α, β, γ, and δ are complex numbers.
Corollary 24.8. The invariance of J2 implies the invariance of the Schmidt coef-
ficients and vise versa.
Proof. It can be shown that the invariant J2 can be expressed as

J2 = 1 − 2|βγ − αδ|2 (24.30)

which implies that

1 − J2
|βγ − αδ|2 = (24.31)
2
and therefore
1
λ21,2 = 1 ± 2J2 − 1 (24.32)
p 
2

– 95 –
 Remark (Relating J2 to entanglement). Parameterizing λ0 := cos θ and λ1 :=
sin θ, with 0 ≤ θ ≤ π/4 (which gives λ0 ≥ λ1 ) the invariant J2 becomes
1
J2 (θ) = cos4 θ + sin4 θ = (3 + cos(4θ)) (24.33)
4
where the last line again follows from the norm. The angle θ now becomes
a characterization of the entanglement. θ = 0 iff the state is separable and
θ = π/4 iff the state is locally equivalent to a maximally entangled Bell state.
We can see this from the following plot of J2 (θ). For small angles, the value
of the plot is ≈ 1 as the angle increases, the entanglement increases and the
value of j2 goes to its maximum value of one half (at θ/π = 1/4).

1.0

0.9

0.8
J2

0.7

0.6

0.5
0.00 0.05 0.10 0.15 0.20 0.25
θ/π

Remark (Algebraic independence of J1 and J2 : the fundamental ring). J1 and

J2 are the only algebraically independent invariants under local action of the
unitary group. Any polynomial function of such invariants is also a polynomial
invariant. In this fashion, it is a remarkable feature that functions of J1 and
J2 are all that is needed to express any local unitary invariant of two-qubit
pure states. This elementary result follows from a much more powerful and
general result in classical invariant theory. That is a proof by Hilbert that the
ring of polynomial invariants is finitely generated [79]. This corresponds to
freely generated linear sums and products of J1 , J2 (the ring

{J1 , J2 , (R, +, ·)}. (24.34)

Any minimal complete set of invariants that can freely generate the full ring
are called fundamental invariants.
n-qubit LOCC invariants. The method we have described in detail for two qubits
is readily extended to n-qubits. Consider the general expression for an n-qubit
pure state
ψ= (24.35)
X
ψijk... |ijk...i
A general polynomial of the state coefficients together with their complex conju-
gates is expressed as

(24.36)
X ijk... qrs... lmn...
cqrs...lmn ψ ψ · · · ψijk... · · ·

If the polynomial (24.36) has an equal numbers of ψ’s and ψ’s and all the

– 96 –
indices of the ψ are contracted using the invariant tensor δ with those of the ψ,
each index being contracted with an index corresponding to the same party then
the polynomial is manifestly invariant under LOCC transformations.
Remark (Connection to the diagrammatic language). Invariant polynomials
(24.36) can be written in terms of permutations on the indices, and given
diagrammatically as contraction with a permutation operator.
Remark (Fundamental ring of invariants). Although we can generate a full basis
of invariants (24.36), except in rare cases, a minimal set of invariants is not
known.
Tensor contraction for SLOCC12 invariants 12 SLOCC: Stochastic Local Op-
erations and Classical Communi-
SLOCC equivalence amounts to equivalence under local invertible transformations. cation.
This results in a coarser partitioning of the set of states as compared to LOCC
equivalence, since SLOCC shows which states are accessible to different parties
with non-zero probability. The following result was proven in [84].

Theorem 24.9 (SLOCC). Two states are SLOCC-equivalent iff there exists an
invertible local operator relating them given by the action of the local general
linear group [84]

GL(d1 , C) × GL(d2 , C) × . . . × GL(dn , C). (24.37)

Remark (SLOCC under the special linear group). Up to a scale factor, and
without loss of generality, we will instead consider action of the special linear
group, SL(2, C) on qubits.

SL(2, C) × SL(2, C) × . . . × SL(2, C) (24.38)

The precise relation between GL and SL invariants is mentioned in Remark 11.

We then consider transforming an n-party state ψ to S |ψi with S := ni=0 Si
N

where each Si has unit determinant. Action of either group is known to

correspond precisely to quantum operations that preserve true entanglement
under a general quantum operation.
Graphical properties of the Levi-Civita symbol. In order to study SLOCC in-
variants, and their tensor networks, we will need to recall the Levi-Civita symbol
and consider a tensor network representation.

Definition 24.2 (Levi-Civita symbol). The order-n Levi-Civita symbol εij···k is

the fully antisymmetric tensor with coefficients in {−1, 0, 1}. All its legs have
the same dimension d = n.

1 when (i, j, . . . , k) is an even permutation of the index values,



εij···k = −1 when (i, j, . . . , k) is an odd permutation, and


 0 otherwise.


(24.39)

Before proceeding to an analysis of the SLOCC invariant tensor network struc-

ture, we must explore some properties of the epsilon tensor.

– 97 –
Remark (Matrix determinant). The determinant for an n-by-n matrix A can
be expressed in terms of the Levi-Civita symbol as follows:

det(A) = εi0 ···in−1 A0i0 · · · An−1in−1 . (24.40)

Diagrammatically, the determinant takes the form

or equivalently, the determinant is directly proportional to the expected value

of the operator
n
(24.41)
O
h| Ai |i
i=1

with respect to the epsilon state in that dimension n.

Lemma 24.10 (Invariance of epsilon under SL(n, C)). We will consider the
representation L of the group SL(n, C) on Cn⊗n , defined as

L(S) |ψi := S ⊗n |ψi , (24.42)

where S ∈ SL(n, C). The order-n epsilon state is invariant under this repre-
sentation:
0
L(S) |εi = S i i S j j · · · S k k εij···k i0 j 0 · · · k 0 = εi j ···k i0 j 0 · · · k 0 = |εi .
0 0   0 0 0  

Proof. If an index value is repeated in the set {i0 , j 0 , . . . , k 0 }, the antisym-

metry of epsilon dictates that the expression must vanish. Hence the only
nonvanishing index combinations are permutations of the n allowed index val-
ues. Using (24.40), we can see that the permutation

(i0 , j 0 , . . . , k 0 ) = (0, 1, . . . , n − 1)

corresponds to det(S), which equals 1 since S ∈ SL(n, C). Invoking the

antisymmetry of epsilon again we conclude that all even permutations must
also yield 1, and all odd permutations −1, thus recovering the definition of
epsilon.

However, we will then note that under wire duality we arrive at the relation

which says that SεS > = ε. So under wire duality, we find that (S ⊗ S)ψε ∼
=
SεS > .

– 98 –
 Remark (Defining equations relating δ and ). It is a classical result of invariant
theory that every identical relation satisfied by contractions of δ’s and ’s can
be built up from the following relations (see e.g. [20]).

AB = −BA ; AB = −BA (24.43)

δA δB = δA
B C C
;  ;
B AC
δA B AC
δA  = BC ; AB AC = δB
C
(24.44)
A
δA = 2 = AB  AB
(24.45)
Also, the identity,

AB CD + AD BC + AC DB = 0 (24.46)

holds, together with the equivalent identities obtained by raising indices of

the above, with Levi-Civita symbols, e.g.
B D
δA δC = AC BD
D B
δC − δA (24.47)

We also note that when we assign values to A, B = 0, 1 we arrive at the

following relations, between the indices and the tensor components as

AB = A − B (24.48)
A
δB = 1 − (A − B)2 (24.49)
Remark (Matrix determinant in dim = 2). The case of the determinant in two
dimensions becomes

εij Ai0 Aj1 = ε01 A00 A11 + ε10 A10 A01 (24.50)
X

where ε10 = −1 and ε01 = 1. In two dimensions, when all i, j, m, n are in

{0, 1}, the following identities hold.

εij εmn = δim δjn − δin δjm (24.51)

εij εin = δjn (24.52)

εij εij = 2 (24.53)
Remark (Epsilon symbol in two dimensions). In two dimensions the Levi-Civita
symbol is an element of the Pauli group: ε = iY = −XZ = ZX, that is,a

= =

where the last step of bending output to inputs and inputs to outputs takes
the transpose of a linear map.
a
Note that the diagram is not written using the convention from quantum circuits, where
the order of composition of operations in an equation is reversed in the circuit.

SLOCC invariants. For pure two-qubit states, the single SL invariant is the de-
terminant of the coefficient matrix α as

K1 = εij εkl αik αjl = 2 det ψ i j = 2(α00 α11 − α01 α10 ) (24.54)
X  

– 99 –
here ε is the fully antisymmetric tensor on two indices defined as ε00 = ε11 = 0 and
ε01 = −ε10 .
Remark. The norm is not a SLOCC invariant.
Remark (Algebraic independence of K1 : the fundamental ring). Any other poly-
nomial invariant of the induced action of the special linear group is necessarily
a polynomial in K1 . Since there is only one invariant in this special case, each
member of this class could be considered a fundamental invariant. Finding the
fundamental invariants becomes complicated for higher dimensional systems.
With the epsilon identities in place, we are now able to consider the tensor
structure of K1 :
K1 = (24.55)
X
εij εkl αik αjl

– 100 –
 Lemma 24.11 (Diagrammatic contraction of K1 ). The diagrammatic language contracts the tensor network
for K1 (24.55) to be twice the determinate of the matrix of state coefficients ( Φ+  ⊗ 1) |ψi found from bending



a wire on ψ.
Proof. [Diagrammatic contraction of K1 ] In the following figure, (a) represents K1 which simplifies to
(d) using our previous results.
(a) T
T
T =
= (b) = T
(c) (d)

=
In (c) we have added > symbols in the unitary boxes, to denote transpose as found from sliding a box
around a bent wire. Figure (d) shows the SU (2, C) invariant of ε as in Lemma 24.10.
To continue the analysis of this network, we will need to introduce a few more identities that we have
proven in detail elsewhere [22, 23]. We need to then consider the stabilizer group of the COPY-tensor.
This is an eight element group

{1, X ⊗ X ⊗ X, −X ⊗ Y ⊗ Y, −Y ⊗ X ⊗ Y, −Y ⊗ Y ⊗ X, Z ⊗ Z ⊗ 1, Z ⊗ 1 ⊗ Z, 1 ⊗ Z ⊗ Z} (24.56)

and then we consider the diagrammatic laws on a choice of generators

(a ) (b )

From these identities we then return to our equation for the determinant.

(a) (b) (d)

= = =

In (a) we express epsilon in using the Pauli algebra. In (b) we apply the fact that the COPY-tensor is
stabilized by the tensor product of three bit flip operators X ⊗ X ⊗ X, as well as pair products of Z’s.
We combine this with the fact that X 2 = 1 = Z 2 to show that the internal order-two tensors all cancel
out in (b). This results in the network (d). What remains is one bit flip operator. It can be contracted
with an valence-one triangular tensor, performing the map λi 7→ λj for i 6= j. This leads to the inner
product
(λ0 , λ1 ).(λ1 , λ0 ) = 2λ1 λ0 = 2 det(ψ) (24.57)
which matches precisely what we expect.

In Theorem 24.9 we recalled the definition of SLOCC invariance, with respect

to GL, the general linear group. We then stated that there is a connection to the
special linear group in Remark 11. We make this connection precise in the following
remark.
Remark (Relating GL and SL in SLOCC). Let G ∈ GL. We have that det(G) =
k, for some constant k and that S ∈ SL has det(S) = 1. Let S = wG, then

– 101 –
 letting
1
w= 1 , ⇒ det(S) = det{wG} = 1 (24.58)
det(G) n
From this it follows that
1
G |ψi = det(G) n S |ψi (24.59)

and so the transformations are related by constant factors (global scale factor).
Each polynomial SLOCC invariant is a homogeneous function with respect to
this global scale factor. The invariant changes only by a factor that does not
depend on the state, but only on the transformation G, S.

Example (Calculated Values of the Invariants for Typical States). Here we calculate the values of each invariant
for several common states of interest. The reduced AND-state is found from recalling Equation (1.25).

∧ij k = (|00i + |01i + |10i) h0| + |11i h1| (24.60)

We evaluate this Boolean tensor at k = 0 as

∧ij 0 = (|00i + |01i + |10i) h0| + |11i h1|) |0i = |00i + |01i + |10i (24.61)

Type State J2 K1 (λ0 , λ1 )

Product |0i |0i 1 0 (1, 0)
Bell √1 (|00i + |11i)
2
1
4
1
2
√1 (1, 1)
2
√ 
r 
AND-type √1 (|01i + |10i + |00i) 7 1
λ± = 16 3 ± 5
3 9 3
√ √ √ √
General λ |00i + 1 − λ |11i 2λ(λ − 1) + 1 2 λ(1 − λ) ( λ, 1 − λ)
p

Invariant composition law

Consider two initially non-interacting quantum wavefunctions |ψi and |φi. If these
states interact by a coupling unitary propagator U , the polynomial invariants of
each state individually are not necessarily polynomial invariants, of the join system
U |φi ⊗ |ψi.
It is desirable to devise methods to compose invariants. By this we mean that if
we know the polynomial invariants of |ψi, |φi and U , together with a tensor network
defining their composition, from this information alone, how can we determine
the resulting invariants of the joint system? In the present section, we solve this
problem for the case of bipartite systems of equal dimension.
Composition of the invariant K1 We will consider a quantum state ψ ∈ H ⊗ H.
We will act on this state locally on each leg, with invertible linear maps. A possible
scenario is depicted as follows. On the left, we have a state ∈ H ⊗ H factored using
the diagrammatic SVD. We then act on this state with invertible maps S1 and S2 .

– 102 –
 The invariant K1 is equal to twice the determinate of the coefficient matrix,
found from bending either wire back, on the tensor representing the state. The
invariant K1 is equal to zero iff det(ψ) = 0. This is true iff ψ is separable, thereby
partitioning the SLOCC class into a disjoint union: entangled vs. not. We can
calculate K1 for not only the state, but also for the maps S1 and S2 . Given the
quantities
K1 (ψ), K1 (S1 ), K1 (S2 ) (24.62)
we can deduce from the standard properties of the determinant, the value of the
invariant found from acting on ψ with the map S1 ⊗ S2 as

K1 (S1 ⊗ S2 ψ) = K1 (ψ)K1 (S1 )K1 (S2 ) (24.63)

Such a scenario extends to the case of qudits, and the value of the invariants
evaluates to the product of singular values of all parties in the contracted network.

25 Generating Invariants for n-party Density Operators

A method to systematically generate polynomial invariants of density operators
acted on by the local unitary group exists and was shown to me by Markus Grassl
[80]. Here we will cast this method into the Penrose tensor calculus. The method
generates a complete basis of monomials that are necessarily invariants of the lo-
cal unitary group acting on a density matrix. A polynomial invariant could be
expressed in this complete basis. Although we can generate a complete basis of
invariants, except in rare cases, finding the minimal collection of polynomial in-
variants is computationally difficult. A utility of generating this basis, comes from
the fact that the tensor networks we consider can be used to calculate any proper-
ties that are invariant under the action of the local unitary group. This includes
Reyni entropies and the entanglement spectrum.
The invariant basis. Consider a density operator ρ ∈ H → H. We can write this
in a basis and consider acting on ρ by a given a group G. A polynomial in

f (x0 , x1 , x2 , x3 , x0 , x1 , x2 , x3 ) =: f (x) (25.1)

is invariant under G iff ∀g ∈ G

f g (x) = f (x) (25.2)

where the notation f g (x) means

 > 
g(x)> = f (x)g > (25.3)
 
f

so we let g act on x and then let f act — f is constant under G iff f is invariant.
The question we are considering is how to generate a complete monomial basis
where each basis element is invariant under G. This would then imply that any
holomorphic function, can be written in this basis. In fact, such a function would
(i) necessarily be invariant under G and (ii) be a function of the density operator
ρ.
First consider the network

– 103 –
representing the equation F ij ρjk then Tr(F ρ) = δik F ij ρjk that is

= F ij ρji If we were considering

ρ = x0 |0i h0| + x1 |0i h1| + x2 |1i h0| + x3 |1i h1| (25.4)

then this trace would evaluate to

F00 x0 + F01 x1 + F10 x2 + F11 x3 (25.5)

and we can pick an valence-four F such to expand the second order monomials as
Tr(F ρ⊗2 ) = Fjk
il ρj ρl which translates to the tensor network as
i k

This procedure carries on in this fashion, arriving at a monomial of order n as

Tr F ρ⊗n = Fjk···r
il···q j l
ρi ρk · · · ρqr (25.6)

The question then changes. We have generated a complete basis in the coefficients
of the density operator, with coefficients in Fjk···r
il···q
. We then will act on the density
operators by elements g ∈ G
ρ0 = gρg −1 (25.7)
and search for Fjk···r
il···q
that satisfy

Tr F g ⊗n ρ⊗n g ⊗n−1 = Tr F ρ⊗n (25.8)

 

Equation (25.8) is satisfied by monomial invariants. From this it follows that

Tr F g ⊗n ρ⊗n g −1⊗n = Tr ρ−1⊗n F g ⊗n g ⊗n (25.9)

   

∀ρ and so
[F, g ⊗n ] = 0, ∀g ∈ G (25.10)
We are then faced with finding matrices F that commute with g ⊗n for each g ∈ G.
This problem was solved in a different setting around 1937, and the solution is
roughly stated in the following theorem.

Theorem 25.1 (R. Brauer, 1937). The algebra of matrices that commute with
each U ⊗n for U ∈ U (n) is generated by a certain representation of the permu-
tation group.

The permutation group has a well known and evident diagrammatic form. In
(a) we show the elements of the permutation group on one system. In (b) we

– 104 –
show the elements on two systems. (c) Illustrates the permutation group on three
elements, S3 of order 6.13 13 Of possible related interest, is
the diagrammatic presentation of
(a) (b) (c) the Temperly-Lieb algebra [? ].

We then carry on to evaluate

Tr Sk g ⊗n ρ⊗n ρ−1⊗n (25.11)

 

to form invariant monomials. We consider the case where we wish to generate an

order one invariant in (a) below. Order two invariants are found below in (b).
We note that the first invariant is simply the square of the invariant found in
(a). The second however, is independent. An example of an invariant of order
three is given in (c). The situation continues. The basis is finitely generated form
the Cayley-Hamilton Theorem. We have that every density operator satisfies its
own characteristic polynomial, giving an nth order polynomial equation in ρ that
vanishes.
(a) (b) (c)

The general form. We consider the monomial invariant generated by tracing over
the contraction with the SWAP-operator in (a) below. Here we have acted on ρ
with some unitary operation U , and give a diagrammatic proof that the network
contracts to a quantity that is invariant, under unitary transformations of ρ.
To see this, we slide U, U † around the bends, resulting
in (b). The diagram
reduces to (c), showing that the invariant evaluates to Tr ρ2 .

(a) (b)
(c)
= = =

Remark (Generating a complete basis of invariants by tensor contraction). To

generate a complete basis, we have to consider operations like SWAP that
permute all elements to different elements. These are called complete permu-
tations.
Lemma 25.2 (Diagrammatic proof of the basis). Every complete permutation
on n wires contracted with n tensor copies of ρ and traced over, contracts to
the elementary diagram giving the trace of ρn .
The unitary invariants of any contraction with a complete permutation op-
erator is readily shown diagrammatically. Any complete permutation is can be
shown to contract to (a) below. It then possible to understand a complete basis of
invariants by evaluating the quantity as shown in (c) below.

– 105 –
 (a) ...
... =
(c)
(b) ... ...
... = ... =

Note that in (b), we have pulled both of the unitaries around bends. We should
have written U > and U †> but we have omitted the transpose symbol > for ease of
producing the figures. These unitary maps still contract to identity as is quickly
verified.
Example (Trace invariants for qubit density operators). We will consider gen-
erating the fundamental invariants for a qubit density operator. The first
invariant follows from tracing over the identity operator (the only permuta-
tion group element on a single system)

I1 = Tr(ρ) (25.12)

is the constant norm. The second is

I2 = Tr ρ2 (25.13)
 

which turns out to be precisely what is known in other areas of the literature
as the purity. In terms of the singular values we have

I1 = λ + + λ − = 1 (25.14)

which is the norm and

I2 = λ2+ + λ2− (25.15)
For the purpose of generating polynomial invariants, it is enough to stop with
I1 and I2 . In fact, from the Cayley-Hamilton Theorem, we have that there is
a polynomial in second order in ρ, that vanishes identically. In other words,
constants a, b, c exist such that

aρ2 + bρ + c1 = 0 (25.16)

so higher powers of ρ can be expressed in terms of I1 and I2 through this

relation. These invariants are indeed algebraically independent and complete,
meaning any other polynomial invariant can be expressed in {R, +, ·, I1 , I2 }.
For instance,
1 1
Det(ρ) = Tr(ρ)2 − Tr ρ2 = (I1 − I2 ) = 1 − (c2 + a2 + b2 ) = λ0 λ1
 
2 2
(25.17)

Topological equivalence of pure and mixed state invariants

We have considered pure state invariants as well as mixed state invariants, under
the action of the local unitary group. We will consider the relation between these
two approaches, when cast into our diagrammatic framework. “Any effect, constant, theo-
Let us return to the expression for J2 rem or equation named after
Professor X was first discov-
J2 = (25.18)
X
αij αkl αil αkj ered by Professor Y, for some
value of Y not equal to X.” —
– 106 – John C. Baez
We will then write
αij αil =: B jl (25.19)
which is given graphically below.

:=

We also fix B l j := αkl αkj . It now follows that the expression for J2 becomes

J2 = B l j B jl = Tr B 2 (25.20)
X  

In (b) we see that this is found graphically when acting on two reduced states with
a SWAP operation, and then tracing over the result.

(a) (b) (c)

= =

We will go on to show that this invariant is in fact identical to the pure state
invariant. From application of the graphical identity (Penrose’s graphical represen-
tation of a density state is on the right)

ψ
ψ*
=
ψ
ψ*

it now follows by applying this identity to (a) we arrive at the expression for J2 we
considered in Section V.

(a) (b)

26 Some Symmetries of ρ
We have expressed two Theorems (3.1 and 2.2) which relate density operators
to quantum states precisely. This might be thought of as a type of symmetry.
The other type of symmetry we have considered are invariant polynomials in the
coefficients of a density operator. These symmetries are general, as they are found
for arbitrary density operators. In the present section, we will remind the reader

– 107 –
of other types of symmetry that are sometimes considered in quantum theory.
Symmetry plays a key role in the lectures that follow, so here is just an introduction.
In particular, symmetry will be considered in detail in Lecture III. To begin, let us
recall the typical symmetry that is often considered in quantum physics.
Example (Group Symmetry of ρ). We will consider a general density operator
ρ and look for V ∈ U (d) such that

ρ = V ρV † (26.1)

this implies that [ρ, V ] = 0 and we arrive at a basis for G by noting the unitary
operators that commute with ρ. That is, {|λi i}i such that ρ = i pi |λi i hλi |.
P

It then follows from V V † = 1 that every V ∈ G can be written as

V = (26.2)
X
eiθi |λi i hλi |
i

In quantum computer science another symmetry is often considered. This is

the symmetry found by so called, stabilizer states. We will discuss this in detail in
Lecture III and also cover a bit of it in Lecture IV. A quantum gate U which is a
tensor product of Pauli operators is said to stabilize the state |ψi iff U |ψi = |ψi.
All the gates stabilizing a given state form a group.
Example (Pauli stabilizers). Here are some standard examples of states stabi-
lized by the Pauli-group

(i) Pauli-X: σ x stabilizes |+i = |0i + |1i and −σ x stabilizes |−i = |0i − |1i
(ii) Pauli-Y: σ y stabilizes |y+ i = |0i + i |1i and −σ y stabilizes |y− i = |0i −
i |1i
(iii) Pauli-Z: σ z stabilizes |0i and −σ z stabilizes |1i
Remark (Gottesman-Knill Theorem). A graphical rewrite proof (by bounding
the number of rewrites) of the Gottesman-Knill theorem follows by considering
the action of the black and plus dots on σ z and σ x . We will set proving this
as an exercise in a later Lecture.
In addition to these two, one also finds the class of so called, symmetric states
(or operators for that matter). These are characterized as follows.
Example (Invariance under Sk ). The other form of symmetry that is considered,
is invariance under the symmetric group. Diagrammatically this amounts to
braiding wires (where here the order of the braids is not relevant).
We can relate symmetry in density operators and symmetry in states as follows.

Further symmetries of ρ.
Now using these ideas, we can consider other symmetries of ρ. We want to find
solutions of
ρ = V ρV † (26.3)
We will then consider
ρ0ε = a1 + bε (26.4)

– 108 –
We then note that b ∈ R and that ε = iY . So we arrive at the density operator

ρε = a1 + bY (26.5)

and will search for solutions to the equation

ρε = V ρε V † = aV V † − biV εV † (26.6)

and note that under wire duality,

ε = SεS > (26.7)

(see Lemma 24.10) and so ρε is invariant for

V ∈ SU (n, R) (26.8)

as those matrices satisfy SS > = 1. We can go further however. In fact, the

following is easy to verify.
Lemma 26.1 (Isomorphism between projectors and self-adjoint unitary maps). It
is readily verified that
U = 1 − 2P (26.9)
is unitary with P 2 = P . The equation is then solved for P in terms of U .
We then see that each projector gives rise to a symmetry of ρε , as does the real
orthogonal subgroup of the special unitary group.

27 Polynomial Invariants of Symmetric States

A symmetric multipartite state or fully symmetric tensor carries the trivial repre-
sentation of the symmetric group and hence is invariant under any permutation of
the parties. Examples are the three qubit GHZ-state |GHZi = |000i + |111i and the
three qubit W-state |Wi = |001i + |010i + |100i.
Remark (Symmetric basis). A general n qubit quantum system has the 2n
orthonormal basis vectors {|00 . . . 00i, |00 . . . 01i, . . . , |11 . . . 11i}. For the sub-
space of symmetric n qubit states, an orthonormal basis is given by the n + 1
symmetric basis states {|S0 i, |S1 i, . . . , |Sn i}. They are defined as
!− 1
2
n
|Sk i = (27.1)
X
|0i|0i · · · |0i |1i|1i · · · |1i
k perm
| {z }| {z }
n−k k

and called Dicke states. We can therefore write |Wi = |S1 i and |GHZi =
√1 (|S0 i + |S3 i).
2

Exercise 27.1 (Boolean symmetric states). Provide the number of symmet-

ric Boolean quantum states.

A polynomial basis. Symmetric qubit states and symmetric polynomials in two

variables are isomorphic as vector spaces. The known mapping is accomplished as

– 109 –
follows.
|0i ↔ x, |1i ↔ y (27.2)
Remark (Creation and annihilation operators). If we consider a symmetric poly-
nomial basis in xn , we can define two operators

a+ := x (27.3)

∂
a− := := ∂x (27.4)
∂x
From this we can define an inner product on our space. For instance, we can
calculate the norm of xn as

hxn |xn i = xn a+ xn−1 = a− xn xn−1 = n! (27.5)

D  E D  E
 

This concept is readily extended to the case of binary forms in x and y when
considering creation and annihilation operators for each variable x and y sep-
arately.

Example (Polynomials for the GHZ-class). The n-partite GHZ state |0i⊗n +
|1i⊗n has homogeneous polynomial f (x, y) = xn + y n .
Example (Polynomials corresponding to common quantum states). The three-
qubit W state |W3 i = |001i + |010i + |100i is isomorphic to the monomial
3x2 y. Furthermore, two appropriately braided copies of |W3 i read |W3,3 i =
(|003i + |030i + |300i) + (|012i + |021i + |102i + |120i + |201i + |210i) which
is given diagrammatically as

:=

and becomes a sum of two Dicke states having corresponding homogeneous

polynomials x20 x3 and x0 x1 x2 . It is isomorphic to the homogeneous polynomial
3x20 x3 + 6x0 x1 x2 .
It turns out that the invariant theory of binary forms is well studied in the
literature. An invariant of a binary form is a polynomial in the coefficients of
a binary form in two variables x and y that remains invariant under invertible
transformations of the variables x and y.

Invariants of forms
Example. Consider
ax2 + 2bx + c = 0 (27.6)
where the discriminant is
∆ = b2 − ac (27.7)
If ∆ = 0 we have a double root, and if ∆ < 0 the roots are complex conjugate.

– 110 –
 Now if we consider the affine change of variables, where α 6= 0

x0 = αx + β (27.8)

Upon calculating the discriminant for the polynomial in the new variables we
arrive at
1
∆0 = 2 ∆ (27.9)
α
which is the same as ∆ in the original polynomial, up to a multiplicative factor
that depends only on the transformation. The properties of the roots remain
unchanged under such transformations.

28 From Binary Forms to Qubit States

We will now consider our subject of interest. Forms in two variables, called
quadratic forms, or binary forms. For example, the general degree two binary
form is given as
Q(x, y) = a2 x2 + 2a1 xy + a0 y 2 (28.1)
which gives rise to the quantum state

ψQ = a2 |00i + a1 (|01i + |10i) + a0 |11i (28.2)

Definition 28.1 (General linear invertible transformation). A general linear in-

vertible transformation of two variables takes the form

x0 = αx + βy (28.3)

y 0 = γx + δy (28.4)
where
αδ − βγ 6= 0 (28.5)
When a variable is transformed, we get an induced transformation of the
coefficients. It is important to note that this transformation does not change
the degree of a given polynomial.

Definition 28.2 (Invariant of a binary form [83]). An invariant of a binary form

Q(x, y) is a function I(a) = I(a0 , ..., a1 ) depending on the coefficients a =
(a0 , ..., an ) of the form, which, up to a determinantal factor, does not change
under the general linear transformation

I(a) = (αδ − βγ)k I(a0 ) (28.6)

where a0 = (a00 , ..., a0n ) are the transformed coefficients.

An integer k is called the weight of the invariant and we are concerned with
the case that I(a) is a polynomial. It is then called, a polynomial invariant.
Remark. Classical invariant theory has provided us elegant solutions to the
following questions [79, 83].

(i) How many independent polynomial invariants are there of a given de-
gree?

– 111 –
 (ii) What do they tell us about the form?

(iii) Can we find a basis for the polynomial invariants?

We are missing an important piece to our puzzle. These are the covariants,
which are polynomials.

Definition 28.3 (Covariant [83]). A covariant of weight k of a binary from Q

of degree n is a function
J(a, x, y) (28.7)
depending both on the coefficients ai and on the independent variables (x, y)
which up to a determinantal factor, in unchanged under general linear trans-
formations
J(a, x, y) = (αδ − βγ)k J 0 (a0 , x0 , y 0 ) (28.8)

Three qubit example

The cubic binary form is given as

Q(x, y) = a3 x3 + 3a2 x2 y + 3a1 xy 2 + a0 y 3 (28.9)

It is known that there is just one fundamental invariant, the discriminant of the
cubic.
∆ = a20 a23 − 6a0 a1 a2 a3 + 4a0 a32 − 3a21 a22 + 4a31 a3 (28.10)
Remark. ∆ = 0 iff Q has a double or triple root.

Exercise 28.1. Consider the discriminant of the cubic.

∆ = a20 a23 − 6a0 a1 a2 a3 + 4a0 a32 − 3a21 a22 + 4a31 a3 (28.11)

and find maximum and minimum values when ∀i, ai ∈ {0, 1}.

The obvious covariant is the form itself. Another important covariant is the
Hessian
H = Qxx Qyy − Q2xy (28.12)
where we use subscripts to denote partial derivatives. For the binary cubic, the
Hessian becomes
1
H = (a1 a3 − a2 )x2 + (a0 a3 − a1 a2 )xy + (a0 a2 − a21 )y 2 (28.13)
36

Theorem 28.2 (vanishing Hessian [83]). A binary form Q(x, y) has vanishing
Hessian, H = 0 iff Q(x, y) = (cx + dy)n , that is iff Q(x, y) is the nth power of
a linear form.

Remark (Applications to quantum entanglement). It is clear that a linear form

Q(x, y) = (cx + dy)n (28.14)

– 112 –
 gives rise to a factorisable quantum state as

ψQ = (c |0i + d |1i)n (28.15)

So clearly, the Hessian corresponding to a quantum state, determines if the

state can be written as a local product of operators. Other covariants give
additional information.
Remark (Sylvester’s tables). In the mid 1800s Sylvester produced tables
predicting numbers of invariants and covariants of a given degree. The ta- “The object of pure physics is
ble should only be trusted for covariants up to degree 6 and invariants up to the unfolding of the laws of
degree 8. the intelligible world; the ob-
ject of pure mathematics that
degree 2 3 4 5 6 7 8 9 10 12 of unfolding the laws of hu-
invariants 1 1 2 4 5 26 (30) 9 89 104 109 man intelligence.” — James
covariants 2 4 5 23 26 124 (130) 69 415 475 947 Joseph Sylvester.

29 Problems
Anti-linear operators
Remark. Tensor contraction is linear. Tensor networks encapsulate multilinear
algebra. What about antilinear operators or non-linear operators (such as
those found in tensor flow software)? As a prelude, let us first consider some
properties of antilinear operators. Readers should first show that complex
conjugate is antilinear.

Definition 29.1 (Qubit density operator). Consider a density operator of a single

spin (a.k.a. qubit) as
1
ρ = (1 + a.σ) (29.1)
2
where a = (a1 , a2 , a3 ) ∈ [−1, 1]⊗3 ⊂ R3 , the Pauli vector σ = (σ1 , σ2 , σ3 ) =
(X, Y, Z) and ‘.’ signifies a sum over the entry wise product, viz. a.σ = i ai σi .
P

Here Tr ρ = 1, ρ ≥ 0, ρ = ρ† , i∈{1,2,3} |ai | ≤ 1 (the vector a also satisfies

P

||a||2 ≤ 1).

Definition 29.2 (T-action). For the polarization vector σ, we define T through

the action T σT † = −σ.

Exercise 29.1. Prove that T as defined in 29.2 is not unitary. (Hint con-
sider preservation of the definition of the Pauli-algebra under unitary group
homomorphism).

Exercise 29.2. Show that i=1,2,3 Tr σi ρ = a1 a2 a3 .

Q

We will consider the form of T = U K where U is unitary and K is complex

conjugation.

– 113 –
Exercise 29.3. For a single qubit, let T = e−ı 2 Y K and show that this
π

definition satisfies the definition of the T-reversal operator from 29.2.

Exercise 29.4. Show that [ρ, T ] = 0 for T = e−ı 2 Y K iff ρ = 1.

π
1
2

Exercise 29.5. Show that [ρ, K] = 0 iff a2 = 0, in which case T ∼ Y .a

a
∼ denotes equality up to an equivalency class defined by unit norm complex numbers.

Remark (Dagger of qubit maps). Let us write the general local qubit map,
using the following parameterization which can express any single-qubit map.

U (2, C) = U (1, C) × SU (2, C) (29.2)

so up to a negligible global phase, every local unitary map on qubits is effec-

tively given by a special unitary map. Consider then,

S = α |0ih0| − β |0ih1| + β |1ih0| + α |1ih1| . (29.3)

It is a straight forward calculation to verify that

[> S]> = S † . (29.4)

Which has the diagrammatic form as:

– 114 –
 PART

Tensor Networks for Open Quan- VI

tum System
In this section we recall several common mathematical descriptions for completely-
positive trace-preserving maps, and show how several key properties may be cap-
tured graphically using the diagrammatic notation we have introduced. The rep-
resentations we will consider are the Kraus (or operator-sum) representation, the
system-environment (or Stinespring) representation, the Liouville superoperator
description based on vectorization of matrices, and the Choi-matrix or dynamical
matrix description based on the Choi-Jamiołkowski isomorphism. We will also de-
scribe the often used process matrix (or χ-matrix) representation and show how
this can be considered as a change of basis of the Choi-matrix. Following this, in
Section 35 we will show how our framework enables one to freely transform between
these representations. The work on open systems we present here in the language
of tensor networks, follows primarily [34].

Spectral Decomposition
Choi-Matrix Kraus /
(Χ-Matrix) Jamio kowski Isomorphism Operator-Sum

Jam
Stinespring Dilation

io
ko
Reshuffling

ws
ki
Iso
mo
a t ion rp
riz his
cto m
Ve

Liouville Vectorization System-Environment

Superoperator (Stinespring)

Remark. In this chapter we will use the notation that X , Y, Z are finite-
dimensional complex Hilbert spaces, L(X , Y) is the space of bounded linear
operators A : X → Y (with L(X ) ≡ L(X , X )), T (X , Y) is the space of opera-
tor maps E : L(X ) → L(Y) (with T (X ) ≡ T (X , X )), and C(X , Y) is the space
of operator maps E which are CP.

30 Kraus / Operator-Sum Representation

The first representation of CPTP-maps we cast into our framework is the Kraus [85]
or operator-sum [24] representation. This representation is particularly useful in
phenomenological models of noise in quantum systems.

Theorem 30.1. Kraus’s theorem states that a linear map E ∈ T (X , Y) is

115
 CPTP if and only if it may be written in the form
D
E(ρ) = Kα ρKα† (30.1)
X

α=1

where the Kraus operators {Kα : α = 1, ..., D}, Kα ∈ L(X , Y), satisfy the
completeness relation
D
Kα† Kα = 1X . (30.2)
X

α=0

The Kraus representation of E in (30.1) has the graphical representation

Ρ K Ρ K†

The maximum number of Kraus operators needed for a Kraus description of E

is equal to the dimension of L(X , Y). For the case where X ∼
=Y∼ = Cd the maximum
number of Kraus operators is d2 , and the minimum number case corresponds to
unitary evolution where there is only a single Kraus operator.
It is important to note that the Kraus representation of E is not unique as there
is unitary freedom in choosing the Kraus operators. We can give preference to a
particular representation called the Canonical Kraus Representation [86] h whichi is
the unique set of Kraus operators satisfying the orthogonality relation Tr Kα† Kβ =
λα δαβ . The canonical Kraus representation will be important when transforming
between representations in Section 35.

31 System-Environment / Stinespring Representation

The second representation of CPTP-maps we consider is the system-environment
model [24], which is typically considered the most physically intuitive description
of open system evolution. This representation is closely related to (and sometimes
referred to as) the Stinespring representation as it can be thought of as an ap-
plication of the Stinespring dilation theorem [87], which we also describe in this
section.

Theorem 31.1. In this model, we consider a system of interest X , called the

principle system, coupled to an additional system Z called the environment.
The composite system of the principle system and environment is then treated
as a closed quantum system which evolves unitarily. We recover the reduced
dynamics on the principle system by performing a partial trace over the en-
vironment. Suppose the initial state of our composite system is given by
ρ ⊗ τ ∈ L(X ⊗ Z), where τ ∈ L(Z) is the initial state of the environment.
The joint evolution is described by a unitary operator U ∈ L(X ⊗ Z) and the
reduced evolution of the principle system’s state ρ is given by

E(ρ) = TrZ [U (ρ ⊗ τ )U † ] (31.1)

For convenience we can assume that the environment starts in a pure state
τ = |v0 ihv0 |, and in practice one only need consider the case where the Hilbert space

– 116 –
describing the environment has at most dimension d2 for X ∼
= Cd [24]. The system-
environment representation of the CP-map E may then be represented graphically
as

Ρ
Ρ  U U†
Ν Ν

The system-environment model is advantageous when modelling the environ-

ment as a physical system. However, care must be taken when ascribing physical
reality to any particular model as the system-environment description is not unique.
This is not surprising as many different physical interactions could give rise to the
same reduced dynamics on the principle system. This freedom manifests in an
ability to choose the initial state of the environment in the representation and then
adjust the unitary operator accordingly. In practice, the system-environment model
can be cumbersome for performing many calculations where the explicit dynamics
of the environment system are irrelevant. The remaining descriptions, which we
cast into diagrammatic form, may be more convenient in these contexts.
Note that the system-environment evolution for the most general case will be
an isometry and this is captured in Stinespring’s representation [87].

Theorem 31.2. Stinespring’s dilation theorem states that a CP-map E ∈

C(X , Y) can be written in the form

E(ρ) = TrZ AρA† (31.2)

h i

where A ∈ L(X , Y ⊗ Z) and the Hilbert space Z has dimension at most equal
to L(X , Y). Further, the map E is trace preserving if and only if A† A = 1X
[87].

In the case where Y = ∼ X , the Hilbert space X ⊗ Z mapped into by the Stine-
spring operator A is equivalent to the joint system-environment space in the system-
environment representation. Hence one may move from the system-environment
description to the Stinespring representation as follows:

Ρ
Ρ  U U†
Ν Ν

Ρ
 A A†

where |v0 i ∈ Z is the initial state of the environment, and we have defined the
Stinespring operator
A = U · (1X ⊗ |v0 i), . (31.3)
– 117 –
 This close relationship is why these two representations are often referred to
by the same name, and as we will show in Section 35.5, it is straight forward to
construct a Stinespring representation from the Kraus representation. However,
generating a full description of the joint system-environment unitary operator U
from a Stinespring operator A is cumbersome. It involves an algorithmic completion
of the matrix elements in the unitary U not contained within the subspace of the
initial state of the environment [86]. Since it usually suffices to define the action
of U when restricted to the initial state of the environment, which by (31.3) is the
Stinepsring representation, this is often the only transformation one need consider.
Remark. A further important point is that the evolution of the principle sys-
tem E(ρ) is guaranteed to be CP if and only if the initial state of the system
and environment is separable; ρX Z = ρX ⊗ ρZ . In the case where the physi-
cal system is initially correlated with the environment, it is possible to have
reduced dynamics which are non-completely positive [88, 89], however such
situations are beyond the scope of this chapter.

32 Louiville-Superoperator Representation
We now move to the linear superoperator or Liouville representation of a CP-map
E ∈ C(X , Y).

Definition 32.1. The superoperator representation is based on the vectoriza-

tion of the density matrix ρ 7→ | ρiiσ with respect to some orthonormal operator
basis {σα : α = 0, ..., d2 −1} as introduced in Section 3.2. Once we have chosen
a vectorization basis (col-vec in our case) we define the superoperator for a
map E ∈ T (X , Y) to be the linear map

S : X ⊗ X → Y ⊗ Y : | ρii 7→ | E(ρ)ii (32.1)

This is depicted graphically as

EHΡL = S Ρ

In the col-vec basis we can express the evolution of a state ρ in terms of tensor
components of S as

E(ρ)mn = (32.2)
X
Snm,νµ ρµν .
µν

For the case where E ∈ T (X ), it is sometimes useful to change the basis of

our superoperators from the col-vec basis to an orthonormal operator basis {σα }
for L(X ). This is done using the basis transformation operator Tc→σ introduced in
Section 3.2. We have

Sσ = Tc→σ · S · Tc→σ
†
(32.3)
= (32.4)
X
Sαβ | σα iihhσβ |.
αβ

where the subscript σ indicates that Sσ is the superoperator in the σ-vec convention.

– 118 –
The tensor networks for this transformation is given by

S †
SΣ = Tc®Σ Tc®Σ

Remark. For a general map E ∈ T (X , Y) we could do a similar construction

but would need different bases for the initial and final Hilbert spaces L(X )
and L(Y).
The structural properties the superoperator S must satisfy for the linear map
E to be hermitian-preserving (HP), trace-preserving (TP), and completely positive
(CP) are [86]:

E is HP ⇐⇒ S = S S (32.5)
⇐⇒ S = S (32.6)
E is TP ⇐⇒ Smm,nν = δnν (32.7)
⇐⇒ S = (32.8)
E is CP ⇐⇒ SI⊗E | ρAB ii ≥ 0 ∀ρAB ≥ 0 (32.9)
Remark. There is not a convenient structural criteria on the superoperator S
which specifies if E is a CP-map. To test for positivity or complete positivity
one generally uses the closely related Choi-matrix representation.
Superoperators are convenient to use for many practical calculations. Unlike
the system-environment model the superoperator S is unique with respect to the
choice of vectorization basis. Choosing an appropriate basis to express the super-
operator in can often expose certain information about a quantum system. For
example, if we want to model correlated noise for a mutli-partite system we can
vectorize with respect to the mutli-qubit Pauli basis. Correlated noise would then
manifest as non-zero entries in the superoperator corresponding to terms such as
σx ⊗ σx . We discus in more detail how this may be done in Section 36.2.

33 Choi-Matrix Representation
The final representation shown in Fig. VI is the Choi matrix [90], or dynamical
matrix [86]. This is an application of the Choi-Jamiołkowski isomorphism which
gives a bijection between linear maps and linear operators [91].

Definition 33.1. Similarly to how vectorization mapped linear operators in

L(X , Y) to vectors in X ⊗ Y or Y ⊗ X , the Choi-Jamiołkowski isomorphism
maps linear operators in T (X , Y) to linear operators in L(X ⊗ Y) or L(Y ⊗ X ).
The two conventions are

col-Λ : T (X , Y) → L(X ⊗ Y) : E 7→ Λc (33.1)

row-Λ : T (X , Y) → L(Y ⊗ X ) : E 7→ Λr . (33.2)

– 119 –
 For X ∼
= Cd , the explicit construction of the Choi-matrix is given by
d−1
Λc = (33.3)
X
|iihj| ⊗ E(|iihj|)
i,j=0
d−1
Λr = (33.4)
X
E(|iihj|) ⊗ |iihj|
i,j=0

where {|ii : i = 0, . . . , d − 1} is an orthonormal basis for X .

We call the two conventions col-Λ and row-Λ due to their relationship with the
vectorization conventions introduced in Section 3.2.

Definition 33.2. The Choi-Jamiołkowski isomorphism can also be thought of

 +having
as a map E ∈ T (X , Y) act on one half of an unnormalized Bell-state
Φ = i |ii ⊗ |ii ∈ X ⊗ X , and hence these conventions corresponding to
 P

which half of the Bell state it acts on:

Λc = (I ⊗ E) Φ+ Φ+  (33.5)
 ED 
 

Λr = (E ⊗ I) Φ+ Φ+  (33.6)
 ED 
 

where I ∈ T (X ) is the identity map.

In what follows we will use the col-Λ convention and drop the subscript from Λc .
We note that the alternative row-Λ Choi-matrix is naturally obtained by applying
the bipartite-SWAP operation to Λc .
As will be considered in Section 35.3, if the evolution of the CP map E is de-
scribed by a Kraus representation {Ki }, then the Choi-Jamiołkowski isomorphism
states that we construct the Choi-matrix by acting on one half of a bell state with
the Kraus map as shown:

Remark. Note that in general any tensor network describing a linear map
E, not just the Kraus description,
 + 
+ may be contracted with one-half of the
maximally entangled state Φ Φ  to construct the Choi-matrix.
With the Choi-Jamiołkowski isomorphism defined, the evolution of a quantum
state in terms of the Choi-matrix is then given by

E(ρ) = TrX (ρT ⊗ 1Y )Λ (33.7)

h i

or in terms of tensor components

E(ρ)mn = Λµm,νn ρµν . (33.8)

X

n,m

The tensor network for (33.7) is given by

– 120 –
 Ρ
EHΡL = L
ΡT = L

The graphical proof of (33) for the case where E is described by a Kraus represen-
tation is as follows:

Ρ Ρ

% $ ! !!

$ ! Ρ !!

$ !!Ρ#

Remark. The structural properties the Choi-matrix Λ must satisfy for the
linear map E to be hermitian-preserving (HP), trace-preserving (TP), and
completely positive (CP) are [86]:

E is HP ⇐⇒ Λ† = Λ (33.9)

⇐⇒ L = L (33.10)

E is TP ⇐⇒ TrY [Λ] = 1X (33.11)

=
⇐⇒ L (33.12)

E is CP ⇐⇒ E is CP ⇐⇒ Λ ≥ 0. (33.13)
The Choi-matrix for a given map E is unique with respect to the isomorphism
convention chosen. We will provide tensor networks to illustrate a close relationship
to the superoperator formed with the corresponding vectorization convention in
Section 35.1. The Choi-matrix finds practical utility as one can check the complete-
positivity of the map E by computing the eigenvalues of Λ. It is also necessary to
construct the Choi-matrix for a given superoperator to transform to the other
representations.
Due to the similarity of vectorization and the Choi-Jamiołkowski isomorphism,
one could then ask what happens if we vectorize in a different basis. This change
of basis of the Choi-matrix is more commonly known as the χ-matrix which we will
discuss next. However, such a change of basis does not change the eigen-spectrum
of a matrix, so the positivity criteria in (33.13) holds for any basis.
Another desirable property of Choi matrices is that they can be directly de-
termined for a given system experimentally by ancilla assisted process tomography
(AAPT) [92, 93]. This is an experimental realization of the Choi-Jamiołkowski
isomorphism which we discuss in detail in Section 36.5.

– 121 –
34 Process Matrix Representation
As previously mentioned, one could consider a change of basis of the Choi-matrix
analogous to that for the superoperator. The resulting operator is more commonly
known as the χ-matrix or process matrix [24].

Definition 34.1. Consider Hilbert spaces X ∼

= Cd x , Y ∼
= Cdy and let D = dx dy ,
and Z ∼= CD . If one chooses an orthonormal operator basis {σα : α = 0, ..., D−
1} for L(X , Y), then a CPTP map E ∈ C(X , Y) may be expressed in terms of
a matrix χ ∈ L(Z) as
D−1
E(ρ) = χαβ σα ρσβ† (34.1)
X

α,β=0

where the process matrix χ is unique with respect to the choice of basis {σα }.

The process matrix with respect to an orthonormal operator basis {σα } is

related to the Choi matrix by the change of basis

χ = Tc→σ · Λ · Tc→σ
†
(34.2)
⇒Λ = (34.3)
X
χαβ |σα iihhσβ |
α,β

where Tc→σ is the vectorization change of basis operator introduced in Section 3.2.
Thus evolution in terms of the χ-matrix is analogous to our Choi evolution as shown
below:

†
Tc%Σ Χ Tc%Σ
!!Ρ# '

Starting with the expression for process matrix evolution in (34.1), the graph-
ical proof asserting the validity of (34.2) is as follows

– 122 –
 Ρ  Σ Ρ Σ†

 Σ Χ Σ†


Σ Χ Σ

†
TcΣ Χ TcΣ





We also see that if one forms the process matrix with respect to the col-vec
basis σα = Ej,i where α = i + dj and d is the dimension of H, then we have χ = Λ.
Remark. Since the process matrix is a unitary transformation of the Choi-
matrix, it shares the same structural conditions for hermitian preservation
and complete-positivity as for the Choi-matrix given in (33.9) and (33.13)
respectively. The condition for it to be trace preserving may be written in
terms of the matrix elements and basis however. These conditions are

E is TP ⇐⇒ TrY Tc→σ χTc→σ = 1X (34.4)

h i
†

χα,β σαT σ β = 1X (34.5)

X
⇐⇒
α,β

E is HP ⇐⇒ χ† = χ (34.6)
E is CP ⇐⇒ χ ≥ 0. (34.7)
To convert a process-matrix χ in a basis {σα } to another orthonormal operator
basis {ωα }, we may use the same change of basis transformation as used for the
superoperator change of basis in Section 32. That is

χω = Tσ→ω · χσ · Tσ→ω
†
(34.8)
= (34.9)
X
χσαβ | σα iiω hhσβ |ω
αβ

where the superscripts σ, ω denote the basis of the χ-matries. This is illustrated as

ΧΩ T Ω ΧΣ T† Ω

– 123 –
35 Transforming between representations
35.1 Transformations between the Choi-matrix and superoperator rep-
resentations
The Choi-matrix and superoperator are naturally equivalent under the reshuffling
wire bending duality introduced in Section 3.1. In the col (row) convention we
may transform between the two by applying the bipartite col (row)-reshuffling
operation R introduced in Section 3.1. Let Λ ∈ L(X ⊗ Y) be the Choi-matrix, and
S ∈ L(X ⊗ X , Y ⊗ Y) be the superoperator, for a map E ∈ T (X , Y). Then we have

Λ = SR S = ΛR (35.1)

The tensor networks for these transformations using the col convention are

  S S  

In terms of tensor components we have

Λmn,µν = Sνn,µm (35.2)

where m, n and µ, ν index the standard bases for X and Y respectively. Graphical
proofs of the relations ΛRc = S and S Rc = Λ are given below

Ρ
Ρ
Ρ S  S
 

 S Ρ
Ρ


 S Ρ

Ρ

Ρ
 
Ρ

 Ρ
 Ρ
Ρ

  Ρ
 

Remark. To transfer between a χ-matrix with respect to an arbitrary operator

basis, and a superoperator with respect to an arbitrary vectorization basis, we
must first convert both to col-vec (or row-vec) convention and then proceed
by reshuffling.

– 124 –
 Note that reshuffling is its own inverse, i.e (ΛR )R = Λ, hence the solid bi-
directional arrow connecting the Choi-matrix and superoperator representations in
Fig. VI. This is the only transformation between the representations we consider
which is linear, bijective, and self-inverse.

35.2 Transformations to the superoperator representation

Transformations to the superoperator from the Kraus and system-environment rep-
resentations of a CP-map are also accomplished by a wire-bending duality, in this
case vectorization. However, unlike the bijective equivalence of the Choi-matrix
and superoperator under the reshuffling duality, the vectorization duality is only
surjective.
If we start with a Kraus representation for a CPTP map E ∈ C(X , Y) given by
{Kα : α = 0, ..., D − 1}, with Kα ∈ L(X , Y), we can construct the superoperator
S ∈ L(X ⊗ X , Y ⊗ Y) by
D−1
S = (35.3)
X
K α ⊗ Kα .
α=0

The corresponding tensor network is

K
S K

and the graphical proof of this relationship follows directly from Roth’s lemma:

K
!!Ρ# $ †
$ $ " Ρ
K Ρ K K Ρ

Starting with a system-environment (or Stinespring) representation of a map

E ∈ C(X , Y) with input and output system Hilbert spaces X ∼= Cdx and Y ∼ = Cd y
respectively, and environment Hilbert space Z = C with 1 ≤ D ≤ dx dy , we
∼ D

may construct the superoperator for this map from the joint system-environment
unitary U and initial environment state |v0 i by

S = (35.4)
X
hα| U |v0 i ⊗ hα| U |v0 i ,
α

where {|αi : α = 0, ..., D − 1} is an orthonormal basis for Z. The corresponding

tensor network is

U
v0
S 
U
v0

As with the Kraus to superoperator transformation, the proof of (35.4) follows

from Roth’s lemma.
– 125 –
 Note that while the vectorization wire bending duality is invertible, these trans-
formations to the superoperator from the Kraus and system-environment represen-
tations are single directional. In both cases injectivity fails as the superoperator
is unique, while both the Kraus and system-environment representations are not.
Hence we have solid single directional arrows in Fig. VI connecting both the Kraus
and system-environment representations to the superoperator. The inverse trans-
formation from a superoperator to the Kraus or system-environment representation
requires a canonical decomposition of the operator S (via first reshuffling to the
Choi-matrix), which is detailed in Sections 35.4 and 35.5.

35.3 Transformations to the Choi-matrix representation

Transforming to the Choi-matrix from the Kraus and system-environment repre-
sentations is accomplished via a wire-bending duality which captures the Choi-
Jamiołkowski isomorphism. As with the case of transforming to the superoperator,
this duality transformation is surjective but not injective.
Given a set of Kraus matrices {Kα : α = 0, ..., D − 1} where Kα ∈ L(X , Y)
for a CPTP-map E ∈ C(X , Y), one may form the Choi-Matrix Λ as was previ-
ously illustrated in (33) in Section 33. In terms of both Dirac notation and tensor
components we have:

!
Λ= Kα |iihj| Kα† (35.5)
X X
|iihj| ⊗
i,j α

= (35.6)
X
|Kα iihhKα |
α
Λmn,µν = (Kα )µm (K α )νn . (35.7)
X

where {|ii} is an orthonormal basis for X , m, n index the standard basis for X ,
and µ, ν index the standard basis for Y.
Given a system-environment representation with joint unitary U ∈ L(X ⊗ Z)
and initial environment state |v0 i ∈ Z we have

Λ= |iihj| ⊗ TrZ U |iihj| ⊗ |v0 ihv0 | U † (35.8)

X h i

i,j

Graphically this is given by

U U†
v0 v0

The proof of these transformations follow directly from the definition of the Choi-
matrix in (33.3), and the tensor networks for the evolution via the Kraus or system-
environment representations given in (30) and (31) respectively. As with the vector-
ization transformation to the superoperator discussed in Section 35.2, even though
the Choi-Jamiołkowski isomorphism is linear these transformations are single direc-
tional as injectivity fails due to the non-uniqueness of both the Kraus and system-
environment representations. Hence we have the solid single-directional arrows

– 126 –
connecting both the Kraus and system-environment representations to the Choi-
matrix in Fig. VI.
This completes our description of the linear transformations between the repre-
sentations of CP-maps in Fig. VI. We will now detail the non-linear transformations
to the Kraus and system environment representations.

35.4 Transformations to the Kraus Representation

We may construct a Kraus representations from the Choi matrix or system envi-
ronment representation by the non-linear operations of spectral-decomposition and
partial trace decomposition respectively. To construct a Kraus representation from
the Superoperator however, we must first reshuffle to the Choi matrix.
To construct Kraus matrices from a Choi matrix we first recall the graphical
Spectral decomposition we introduced as an example of our color summation con-
vention. If E is CP, by (33.13) we have Λ ≥ 0 and hence the spectral decomposition
of the Cho -matrix is given by

Λ= (35.9)
X
µα |φα ihφα | ,
α

where µα ≥ 0 are the eigenvalues, and |φα i the eigenvectors of Λ. Hence we can
√
define Kraus operators Kα = λα Aα where λα = µα and Aα is the unique operator
satisfying | Aα ii = |φα i as illustrated:

KΑ = ΦΑ ΛΑ

The number of Kraus operators will be equal to the rank r of the Choi matrix,
where 1 ≤ r ≤ dim(L(X , Y)). The graphical proof of this transformation is as
follows:

Ρ  

Φ Λ Φ


 A Ρ A†

Λ Λ

 K Ρ K†

The proof that Kraus operators satisfy the completeness relation follows from the
trace preserving property of Λ in (33.11):

– 127 –
 K† K  Λ Φ Φ Λ

 Φ Λ Φ

 

Note that since Λ, and the χ-matrix are related by a unitary change of basis,
the Kraus representations constructed from their respective spectral decomposi-
tions will also be related by the same transformation. Each will give a unitarily
equivalent Canonical Kraus representation of E since the eigen-vectors are orthog-
onal. Thus we have described the arrow in Fig. VI connecting the Choi matrix
to the Kraus representation. It is represented as a dashed arrow as it involves a
non-linear decomposition, and is single directional as this representation transfor-
mation is injective, but not surjective. Surjectivity fails as we can only construct
the canonical Kraus representations for E. The reverse transformation is given by
the Jamiołkowski isomorphism described in Section 35.3.
Starting with a system-environment representation with joint unitary U ∈
L(X ⊗ Z) and initial environment state |v0 i ∈ Z, we first choose an orthonormal
basis {|αi : α = 0, ..., D − 1} for Z. We then construct the Kraus representation
by decomposing the partial trace in this basis as follows

E(ρ) = TrE U (ρ ⊗ |vihv|) U † (35.10)

h i

D−1
= hα| U |v0 i ρ hv0 | U † |αi (35.11)
X

α=0
D−1
= Kα ρKα† . (35.12)
X

α=0

Hence we may define Kraus matrices

Kα = hα| U |v0 i (35.13)

leading to the tensor network

KΑ 
U
Α v


The graphical proof of (35.13) and (35.4) is as follows

– 128 –
 Ρ

Ρ
 U U†
Ν Ν

 Ρ
U U†
Ν Ν

 Ρ
U U†
Ν Ν

 K Ρ K†

Though the Kraus and system-environment representations are both non-unique,

for a fixed environment basis this partial trace decomposition is an injective trans-
formation between the Kraus and Stinespring representations (or equivalently be-
tween the Kraus and system-environment representations when the joint unitary
is restricted to a fixed initial state of the environment). To see this let {Kα } and
{Jα } be two Kraus representations for a CPTP-map E ∈ C(X , Y), constructed
from Stinespring representations A and B respectively. We have that

Kα = Jα (35.14)
⇔ (Kα )ij = (Jα )ij (35.15)
⇔ Aiα,j = Biα,j (35.16)
⇔ A = B. (35.17)

Since the Stinespring operators satisfy A = U |v0 i and B = V |v0 i for some joint
unitaries U and V , we must have that U0 = V0 where U0 and V0 are the joint
unitaries restricted to the subspace of the environment spanned by |v0 i.
This transformation can be thought of as the reverse application of the Stine-
spring dilation theorem, and hence for a fixed choice of basis (and initial state of
the environment) it is invertible. The inverse transformation is the Stinespring
dilation, and as we will show in Section 35.5, since the inverse transformation is
also injective this transformation is a bijection. However, since the partial trace
decomposition involves a choice of basis for the environment it is non-linear —
hence we use a dashed bi-directional arrow to represent the transformation from
the system-environment representation to the Kraus representation in Fig. VI.

35.5 Transformations to the system-environment representation

We now describe the final remaining transformation given in Fig. VI, the bi-
jective non-linear transformation from the Kraus representation to the system-
environment, or Stinespring, representation. The system-environment representa-
tion is the most cumbersome to transform to as it involves the unitary competition
– 129 –
of a Stinespring dilation of a Kraus representation. Thus starting from a super-
operator one must first reshuffle to the Choi-matrix, then from the Choi-matrix
description one must then spectral decompose to the canonical Kraus representa-
tion, before finally constructing the system-environment as follows.
Let {Kα : α = 0, ..., D − 1}, where 1 ≤ D ≤ dim(L(X , Y)), be a Kraus
representation for the CP-map E ∈ T (X , Y). Consider an ancilla Hilbert space
Z ∼= CD , this will model the environment. If we choose an orthonormal basis for
the environment, {|αi : α = 0, ..., D − 1}, then by Stinesprings dilation theorem we
may construct the Stinespring matrix for the CP map E by
D−1
A= (35.18)
X
Kα ⊗ |αi .
α=0

Recall from Section 31 that the Stinespring representation is essentially the

system-environment representation when the joint unitary operator is restricted
to the subspace spanned by the initial state of the environment. Hence if we let
|v0 i ∈ Z be the initial state of the environment system, then this restricted unitary
is given by
U0 = (35.19)
X
Kα ⊗ |αihv0 | , .
α

The tensor networks for (35.18) and (35.19) are:

K K
U  A  U0 
Ν Ν

(a) Stinespring operator (a) Restricted unitary

The graphical proof that this construction gives the required evolution of a state ρ
is as follows

Ρ
U0 U0†
Ν Ν

 K Ρ K†

Ν Ν Ν Ν

 K Ρ K†

 K Ρ K†

 Ρ

In principle, one may complete the remaining entries of this matrix to construct the
full matrix description for the unitary U , however such a process is cumbersome
and is unnecessary to describe the evolution of the CP-map E [86].

– 130 –
 We have now finished characterizing the final transformations depicted in
Fig. VI connecting the Kraus representation to the system-environment represen-
tation by Stinespring dilation. As previously mentioned in Section 35.4, for a fixed
choice of basis and initial state for the environment, the transformation between
Kraus and Stinespring representations is bijective (and hence so is the transfor-
mation between Kraus and system-environment representations when restricted to
the subspace spanned by the initial state of the environment). Though both these
representations are non-unique, by fixing a basis and initial state for the environ-
ment we ensure that this transformation is injective. To see this let U0 and V0 be
unitaries restricted to the state |v0 i constructed from Kraus representations, {Kα }
and {Jα } respectively, for E ∈ C(X , Y). Then

U0 = V0 ⇔ Kα ⊗ |αihv0 | = (35.20)
X X
Jα ⊗ |αihv0 |
α α
Kα hβ|αi = (35.21)
X X
⇔ Jα hβ|αi
α α
⇔ Kβ = Jβ (35.22)

Bijectivity then follows from the injectivity of the inverse transformation — the
previously given construction of a Kraus representation by the partial trace de-
composition of a joint unitary operator in (35.4).

36 Applications
We have now introduced all the basic elements of our graphical calculus for open
quantum systems, and shown how it may be used to graphically depict the vari-
ous representations of CP-maps, and transformations between representations. In
this section we move onto more advanced applications of the graphical calculus.
We will demonstrate how to apply vectorization to composite quantum systems,
and in particular how to compose multiple superoperators together, and construct
effective reduced superoperators from tracing out a subsystem. We also demon-
strate the superoperator representation of various linear transformations of matri-
ces. These constructions will be necessary for the remaining examples where we
derive a succinct condition for a bipartite state to be used for ancilla assisted pro-
cess topography, and where we present arguably simpler derivations of the closed
form expression for the average gate fidelity and entanglement fidelity of a quantum
channel in terms of properties of each of the representations of CP-maps given in
Section VI.

36.1 Vectorization of composite systems

We now describe how to deal with vectorization of the general case of composite
system of N finite dimensional Hilbert spaces. Let Xk ∼ = Cdk be a dk -dimensional
complex Hilbert space, and let {|ik i : ik = 0, ..., dk − 1} be the standard basis for
Xk . We are interested in the composite system of N such Hilbert spaces,
N
X = X1 ⊗ ... ⊗ XN = (36.1)
O
Xk
k=1

which has dimensions D = N k=1 dk . Let {|αi : α = 0, ..., D − 1} be the com-

Q

putational basis for X . We can consider vectors in X and the dual space X † as
– 131 –
either 1st-order tensors where their single wire represents an index running over α,
or as a N th-order tensor where each of the N wire corresponds to an individual
Hilbert space Xk . The correspondence between these two descriptions is made by
the concatenation of the composite indices according to the lexicographical order
N
D
α= where c(k) := Qk (36.2)
X
c(k) ik .
k=1 j=1 dj

Note that one could also consider the object as any order tensor between 1st and
N th by the appropriate concatenation of some subset of the the wires.
We define the unnormalized Bell-state on the composite system X ⊗ X to be
the state formed by the column (or row) vectorization of the identity operator
1X ∈ L(X )
D
| 1X ii =
X
|αi ⊗ |αi
α=0
1 −1
dX N −1
dX
= .... |i1 , ..., iN i ⊗ |i1 , ..., iN i . (36.3)
i1 =0 iN =0

where |i1 , ..., iN i := |i1 i ⊗ ... ⊗ |iN i. The tensor network for this state is

N 



N 


As with the single system case the column vectorization of a composite linear
operator A ∈ L(X , Y), where Y = N k=1 Yk , is given by bending all the system
N

wires upwards, or equivalently by the identity

| Aii ≡ (1 ⊗ A)| 1ii. (36.4)

Graphically this is given by

 




A 
  A






Note that the order of the subsystems for the bent wires is preserved by the vec-
torization operation.
In some situations it may be preferable to consider vectorization of the compos-
ite system in terms of vectorization of the individual component systems. Transfer-
ring between this component vectorization and the joint-system vectorization can
be achieved by an appropriate index permutation of vectorized operators which has
a succinct graphical expression when cast in the tensor network framework.
Suppose the operator A ∈ L(X , Y), where X = N k=1 Xk , Y = k=1 Yk , is
N NN

composed of subsystem operators such that

A = A1 ⊗ ... ⊗ AN (36.5)
– 132 –
where Ak ∈ L(Xk , Yk ) for k = 1, ..., N . As previously stated the vectorized com-
posite operator | Aii is a vector in the Hilbert space X ⊗ Y.
We define an operation VN called the unravelling operation, the action of which
unravels a vectorized matrix | Aii = | A1 ⊗ . . . ⊗ AN ii into the tensor product of
vectorized matrices on each individual subsystem Xk ⊗ Yk

VN | A1 ⊗ . . . ⊗ AN ii = | A1 ii ⊗ . . . ⊗ | AN ii. (36.6)

The inverse operation then undoes the unravelling

−1
| A1 ii ⊗ . . . ⊗ | AN ii = | A1 ⊗ . . . ⊗ An ii. (36.7)


VN

More generally the unravelling operation VN is given by the map

N
VN : |xX i ⊗ |yY i 7−→ (|xk i ⊗ |yk i) (36.8)
O

k=1

where |xX i ≡ |x1 i ⊗ . . . ⊗ |xN i , |yY i ≡ |y1 i ⊗ . . . ⊗ |yN i. Hence we can write VN in
matrix form as

VN = (36.9)
X X
|i1 , j1 , . . . , iN , jN i hiX , jY | .
i1 ,...,iN j1 ,...,jN

where |iX i ≡ |i1 i ⊗ . . . ⊗ |iN i , |jY i ≡ |j1 i ⊗ . . . ⊗ |jN i, and |ik i , |jl i are the standard
bases for Xk and Yl respectively.
We can also express VN as the composition of SWAP operations between two
systems. For the previously considered composite operator A ∈ L(X , Y) we have
that | Aii has 2N subsystems. If we label the SWAP operation between two sub-
system Hilbert spaces indexed by k and l by SWAPk:l , where 1 ≤ k, l ≤ 2N , then
the unravelling operation can be composed as

VN = WN −1 ...W1 (36.10)

where
k−1
Wk = SWAPN −k+2j+1:N −k+2j+2 . (36.11)
Y

j=0

For example

W1 = SWAPN :N +1 (36.12)
W2 = SWAPN −1:N SWAPN +1:N +2
WN −1 = SWAP2:3 SWAP4:5 . . . SWAP2N −2:2N −1 .

While this equation looks complicated, it has a more intuitive construction when
depicted graphically. The tensor networks for the unravelling operation in the
N = 2, 3 and 4 cases are shown below

– 133 –
 (a) V2 (a) V3 (a) V4

We also present a graphical proof of this for the N = 3 case:

A ! !
A1
A !A ! ! ! A A2
A1
A1
A A2 A3
A A3 A2

A ! ! A3 A1 A1
A1 A1 A 1

A ! ! A ! !
A2 A1 2 A
2
A1 A2 A2

! A3 !
A3 A3 A3 A3
A2 A2

A1
A1 A1 A A3 A1
3
! A2

! ! A3
! A2
A2 A2 A1
A3

36.2 Composing
A 3
superopators! A3
A2

We now discuss how to compose superoperators A on individual subsystems to form

3

the correct superoperator on the composite system, and vice-versa. Given two su-
A
peroperators S1 , and S2 , if we construct a joint system superoperator via tensor
1

product (S1 ⊗ S2 ), this composite operator acts on the tensor product of vector-
! A
ized inputs | ρ1 ii ⊗ | ρ2 ii, rather than the the vectorization of the composite input
2

| ρ1 ⊗ ρ2 ii. To construct the correct composite superoperator for input | ρ1 ⊗ ρ2 ii

A
we may use the unravelling operation VN from (36.6) and its inverse.
3

If we have a set of superoperators {Sk : k = 1, ..., N } where Sk ∈ L(Xk ⊗

Xk , Yk ⊗ Yk ), then the joint superoperator S ∈ L(X ⊗ X , Y ⊗ Y), where X =
k=1 Xk , Y = k=1 Yk , is given by
NN NN

S = VN
†
(S1 ⊗ . . . ⊗ SN ) VN . (36.13)

The tensor networks for this transformation in the N = 2 and N = 3 cases are
shown below

!1

1 ! ! !2
 
2 !3

N =2 N =3

Composing superoperators from individual subsystem superoperators is useful

when performing the same computations for multiple identical systems. For an
example we consider vectorization in the Pauli-basis for an N -qubit system. While
it is generally computationally more efficient to perform vectorization calculations
in the col-vec (or row-vec) basis, as these may be implemented using structural
– 134 –
operations on arrays, it is often convenient to express the superoperator in the Pauli
basis, or the Choi-matrix in the χ-matrix representation, when we are interested
in determining the form of correlated errors. However, transforming from the col-
vec to the Pauli-basis for multiple (and possibly arbitrary) number of qubits is
inconvenient. Using our unravelling operation we can instead compute √ the single
qubit change of basis superoperator Tc→σ , where σ = {1, X, Y, Z}/ 2 is the Pauli-
basis for a single qubit, and use this to generate the transformation operator for
multiple qubits. In the case of N -qubits we can construct the basis transformation
matrix as
(N )
Tc→σ = VN
† ⊗N
· Tc→σ · VN . (36.14)
The joint-system superoperator in the Pauli-basis is then given by

S σ = Tc→σ
(N ) (N )†
· S · Tc→σ (36.15)

The same transformation can be used for converting a state ρ = ρ1 ⊗ . . . ⊗ ρN to

the Pauli basis: | ρiiσ = Tc→σ | ρiic . These unravelling techniques are also useful for
(N )

applying operations to a limited number of subsystems in a tensor network as used

in many tensor network algorithms.

36.3 Matrix operations as superoperators

We now show how several common matrix manipulations can be written as super-
operators. These expressions are obtained by simply vectorizing the transformed
operators. We begin with the trace superoperator STr which implements the trace
of a matrix STr | Aii := Tr[A] for a square matrix A ∈ L(X ). This operation is
simply given by the adjoint of the unnormalized Bell-state:

STr := hh1 | (36.16)

where 1 ∈ L(X ) is the identity operator. If X is itself a composite system, we

simply use the definition of the Bell-state for composite systems from Eq. (36.3).
This is illustrated in our graphical calculus as

STr | Aii = ! (36.17)

For a rectangular matrix B ∈ L(X EE the transpose superoperator ST which

 , Y),
implements the transpose ST | Bii =  B is simply a swap superoperator between
 T

X and Y.

ST = SWAP (36.18)
SWAP : X ⊗ Y 7→ Y ⊗ X (36.19)

The tensor network for the swap superoperator is

ST | Bii = ! (36.20)

If X and Y are composite vector spaces we may split the crossed wires into their
respective subsystem wires.

– 135 –
 Next we give the superoperator representations of the bipartite matrix op-
erations in (3.1) acting on vectorized square bipartite matrices M ∈ L(X ⊗ Y).
These are the partial trace over X (STrX ) (and STrY over Y), transposition ST , and
col-reshuffling (SRc ).

STrX : X ⊗ Y ⊗ X ⊗ Y 7→ Y ⊗ Y (36.21)
STrY : X ⊗ Y ⊗ X ⊗ Y 7→ X ⊗ X (36.22)
ST : X ⊗ Y ⊗ X ⊗ Y 7→ X ⊗ Y ⊗ X ⊗ Y (36.23)
SRc : X ⊗ Y ⊗ X ⊗ Y 7→ X ⊗ X ⊗ Y ⊗ Y (36.24)

The graphical representation of the superoperators for these operations are:

STrX | M ii = ! STrY | M ii = ! ST | M ii = ! SRc | M ii = !

Algebraically they are given by

STrX = hh1X | ⊗ 1Y ⊗ 1Y V2 (36.25)

 

STrY = 1X ⊗ 1X ⊗ hh1Y | V2 (36.26)

 

ST = SWAP1:3 SWAP2:4 (36.27)

SRc = V2 (36.28)

where V2 is the unravelling operation in Eq. (36.6).

In the general multipartite case for a composite matrix A ∈ L(X ) where X =
k=1 Xk , we can trace out or transpose a subsystem j by using the unravelling
NN

operation in Eq. (36.6) to insert the appropriate superoperator for that subsystem
with identity superoperators on the remaining subsystems:
   
j−1 N
SOj = VN
−1 
O O
−1 SIk  ⊗ SO ⊗  SIk  VN
k=1 k=j+1

where SO ∈ T (Xj ) is the superoperator acting on system j and SIk ∈ T (Xk ) is the
identity superoperator for subsystem L(Xk ). Similarly by inserting the appropri-
ate operators at multiple subsystem locations we can perform the partial trace or
partial transpose of any number of subsystems.

36.4 Reduced superoperators

We now present a simple but useful example of the presented bipartite operations
in the superoperator representation to show how to construct an effective reduced
superoperator for a a subsystem out of a larger superoperator on a composite
system.
Consider states ρXY ∈ L(X ⊗ Y) which undergo some channel F ∈ C(X ⊗ Y)
with superoperator representation S. Suppose system Y is an ancilla which we
initialize in some state τ0 ∈ L(Y), and we post-select on the output state of system
Y being in a state τ1 . We may construct the effective reduced map F 0 ∈ T (X )

– 136 –
for this process for arbitrary input and output states of system X , given by a
superoperator S 0 , as shown:

Formally, we are defining the superoperator representation S 0 of the effective

channel F 0 as the map

F 0 (ρ) = TrY [(1X ⊗ τ1 )F(ρ ⊗ τ0 )] (36.29)

S =0
hhτ1 |V2 SV2† | τ0 ii (36.30)

where S is the superoperator representation of F and | τj ii is implicitly assumed

to have the identity operation on the vectorization of subsystem X ⊗ X (| τj ii :=
1X ⊗ 1X ⊗ | τj ii).
36.5 Ancilla Assisted Process Tomography
Quantum state tomography is the method of reconstructing an unknown quantum
state from the measurement statistics obtained by performing a topographically
complete set of measurements on many identical copies of the unknown state [24].
Quantum process tomography is an extension of quantum state tomography which
reconstructs an unknown quantum channel E ∈ C(X ) from appropriately gener-
ated measurement statistics. One such procedure, known as standard quantum
process tomography, involves preparing many copies of each of a topographically
complete set of input states, subjecting each to the unknown quantum channel,
and performing state tomography on the output [94].
An alternative approach is to directly measure the Choi-matrix for the channel
via a method known as ancilla assisted process tomography (AAPT) [93]. The
simplest case of AAPT is entanglement assisted process tomography(EAPT) which
is an experimental realization of the Choi-Jamiołkowski isomorphism. Here an
experimenter prepares a a maximally entangled state
1
ρΦ = |1iihh1| (36.31)
d
across the system of interest X and an ancilla Z ∼ = X , and subjects the system to
the unknown channel E, and the ancilla to an identity channel I. The output of
this joint system-ancilla channel is the rescaled Choi-matrix:

Λ
ρ0φ = (I ⊗ E) (ρΦ ) = . (36.32)
d
which can be measured directly by quantum state tomography. The tensor network
for EAPT is
– 137 –
 =

In general AAPT does not require ρAS to be maximally entangled. It has been
demonstrated experimentally that AAPT may be done with a state which does
not have any entanglement at all, at the expense of an increase in the estimation
error of the reconstructed channel [93]. A necessary and sufficient condition for a
general state ρAS to allow recovery of the Choi-matrix of an unknown channel E
via AAPT is that it have a Schmidt number equal to d2 where d is the dimension
of the state space X [93]. This conditions has previously been called faithfulness
of the input state, and one can recover the original Choi-matrix for the unknown
channel E by applying an appropriate inverse map to the output state in post-
processing [92]. We provide an arguably simpler derivation of this condition, and
the explicit construction of the inverse recovery operator. The essence of this proof
is that we can consider the bipartite state ρAS to be Choi-matrix for an effective
channel via the Choi-Jamiołkowski isomorphism (but with trace normalization of
1 instead of d) . We can then apply channel transformations to this initial state
to convert it into an effective channel acting on the true Choi-matrix, and if this
effective channel is invertible we can recover the Choi-matrix for the channel E by
applying the appropriate inverse channel.

Proposition 36.1. (a) A state ρAS ∈ L(X ⊗ X ) may be used for AAPT of an
unknown channel E ∈ C(X ) if and only if the reshuffled density matrix SAS = ρR
AS
c

is invertible.
(b) The channel can be reconstructed from the measured output state by ΛE =
(R ⊗ I)(ρ0AS ) where ρ0AS = (I ⊗ E)(ρAS ) is the output state reconstructed by
quantum state tomography, and R is the recovery channel given by superoperator
SR = (SAS T )−1 .

The graphical proof of Prop. 36.1 is illustrated in Fig. 18. This proof demon-
strates several useful features of the presented graphical calculus. In particular it
applies the vectorized reshuffling transformation to a bipartite density matrix input
state to obtain an effective superoperator representation of a state, and uses the
unravelling operation for composition of superoperators. From this construction we
find that if the initial state ρAS is maximally entangled, then it can be expressed
as ρAS = |V iihhV | for some unitary V . In this case the reshuffled superoperator of
the state corresponds to a unitary channel SAS = V ⊗ V , and hence is invertible
with SAS
−1
= SAS
†
. If the input state is not maximally entangled, then the closer it
is to a singular matrix, the larger the condition number and hence the larger the
amplification in error when inverting the matrix.

– 138 –
 =

Figure 18. Graphical proof of the equivalence of an initial state ρAS used for performing
AAPT of an unknown CPTP map E with superoperator representation S, to a channel
(R ⊗ I) acting on the Choi-matrix Λ for a channel E. The Choi-matrix can be recovered
if and only if the the superoperator SR = SAS
T
= (ρR
AS ) is invertible.
c T

36.6 Average Gate Fidelity

When characterizing the performance of a noisy quantum channels a widely used
measure of the closeness of a CPTP map E ∈ C(X ) to a desired quantum channel
F ∈ C(X ) is the Gate Fidelity. This is defined to be

FE,F (ρ) = F (F(ρ)E(ρ)) (36.33)

where 2
√ √
 q
F (ρ, σ) = Tr ρσ ρ (36.34)

is the fidelity function for quantum states [24].

In general we are interested in comparing a channel E to a unitary map U ∈
C(X ) where U(ρ) = U ρU † . In this case we have
" rq q #2
FE,U (ρ) = Tr U ρU † E(ρ) U ρU † (36.35)
2
√ † √
 q
= Tr ρU E(ρ)U ρ (36.36)
2
√ † √
 q
= Tr ρ U (E(ρ)) ρ (36.37)
= FU † E,I (ρ) (36.38)

– 139 –
where I is the identity channel and U † (ρ) = U † ρU , is the adjoint channel of the
unitary channel U. Thus without loss of generality we may consider the gate fidelity
FE (ρ) ≡ FU † F ,I (ρ) comparing E to the identity channel, where we simply define
E ≡ U † F if we wish to compare F to a target unitary channel U.
The most often used quantity derived from the gate fidelity is the average gate
fidelity taken by averaging FE (ρ) over the the Fubini-Study measure. Explicitly
the average gate fidelity is defined by
Z
FE = d ψ hψ| E(|ψihψ| ψ) |ψi . (36.39)

where due to the concavity of quantum states we need only integrate over pure
states FE (|ψihψ| ψ) = hψ| E(|ψihψ| ψ) |ψi.
Average gate fidelity is a widely used figure of merit in part because it is simple
to compute. The expression in (36.39) reduces to explicit expression for F E in terms
of a single parameter of the channel E itself. This has previously been given in terms
of the Kraus representation [95, 96], superoperator [97] and Choi-matrix in [98].
We now present an equivalent graphical derivation of the average gate fidelity in
terms of the Choi-matrix which we believe is simpler than previous derivations.
We start with the tensor network diagram corresponding to (36.39) and perform
graphical manipulations as follows

For the next step of the proof we use the result that the average over ψ of a tensor
product of states |ψihψ|n is given by

Πsym (n, d)
Z
dψ |ψihψ| ψ ⊗n = (36.40)
Tr[Πsym (n, d)]

where Πsym (n, d) is the projector onto the symmetric subspace of X ⊗n . This project
may be written as [99]
1 X
Πsym (n, d) = Pσ (36.41)
n! σ
where Pσ are operators for the permutation σ of n-indices. These permutations
may be represented as a swap type operator with n tensor wires. For the case of
n = 2 we have the tensor diagram:

=
( + (

– 140 –
Here we can see that Tr[Πsum (2, d)] = (d2 + d)/2, and hence we have that
1
Πsym (2, d) = (1 ⊗ 1 + SWAP) (36.42)
2
d2 + d
Tr[Πsym (2, d)] = (36.43)
2
Z
1 ⊗ 1 + SWAP
⇒ dψ |ψihψ| ψ 2 = (36.44)
d(d + 1)

where X ∼ = Cd , 1 ∈ L(X ) is the identity operator, and SWAP is the SWAP

operation on X ⊗ X . Subsituting (36.6) into (36.6) completes the proof:

Hence we have that the average gate fidelity in terms of the Choi-matrix is given
by
d + hh1 |Λ| 1ii
FE = (36.45)
d(d + 1)
where we have used the fact that the Choi-matrix is normalized such thatTr[Λ] = d.
From this proof one may derive expressions for the other representations using the
channel transformations in Section 35. The resulting expressions are

d + Tr[S]
FE = (36.46)
d(d + 1)
d + hh1 |Λ| 1ii
= (36.47)
d(d + 1)
d + j | Tr[Kj ]|2
P
= (36.48)
d(d + 1)
d + dχ00
= (36.49)
d(d + 1)
d + TrX [A† ] · TrX [A]
= (36.50)
d(d + 1)

where S, Λ, {Kj }, χ, A are the superoperator, Choi-matrix, Kraus, χ-matrix and

Strinespring representations for E respectively. In the case of the χ-matrix
√ repre-
sentation, χ is defined with respect to a basis {σj } satisfying Tr[σj ] = dδj,0 .
Similar techniques can be appliedR for tensor networks that may be graphically
manipulated into containing a term dψ |ψihψ| ψ ⊗n for n > 2. This could prove
useful for computing higher order moments of fidelity functions and other quantities
defined in terms of averages over quantum states |ψi. In this case there are n!

– 141 –
permutations of the tensor wires for the permutation operator Pσ in (36.41), and
these can be decomposed as a series of SWAP gates. For example, in the case of
n = 3 we have
1 ⊗3
Πsym (3, d) = 1 + SWAP1:2 + SWAP1:3 + SWAP2:3 (36.51)
6
+ SWAP1:2 SWAP2:3 + SWAP2:3 SWAP1:2

d3 + 3d2 + 2d
Tr[Πsym (3, d)] = . (36.52)
6
36.7 Entanglement Fidelity
Another useful fidelity quantity is the entanglement fidelity which quantifies how
well a channel preserves entanglement with an ancilla [24, 100]. For a CPTP map
E ∈ C(X ) and density matrix ρ ∈ L(X ) the entanglement fidelity is given by

Fe (E, ρ) = inf F (|ψihψ| , (IZ ⊗ E)(|ψihψ|)) :



TrZ [|ψihψ|] = ρ (36.53)

where |ψi ∈ X ⊗ Z is a purification of ρ over an ancilla Z. Entanglement fi-

delity turns out to be independent of the choice of purification |ψi, and a closed
form expression has been given in terms of the Kraus representation [24] and Choi-
matrix [101]. Here we present a simple equivalent derivation in terms of the Choi-
matrix representation of the channel E using graphical techniques. Then by apply-
ing the channel transformations of Section 35 we obtain expressions in terms of the
other representations. The resulting expressions for entanglement fidelity are:

Fe (E, ρ) = hhρ |Λ| ρii (36.54)

= Tr (ρ ⊗ ρ)S (36.55)
h i
T

= | Tr[ρKj ]|2 (36.56)

X

= χij Tr[ρ σi ] Tr ρσj† (36.57)

X h i

i,j

= TrX [ρA† ] · TrX [Aρ] (36.58)

where S, Λ, {Kj }, χ, A are the superoperator, Choi-matrix, Kraus, χ-matrix and

Strinespring representations for E respectively. In the case of the χ-matrix
√ repre-
sentation, χ is defined with respect to a basis {σj } satisfying Tr[σj ] = dδj,0 .
For the graphical proof in terms of the Choi-representation we start with
(36.53) and perform the following tensor manipulations

– 142 –
 =

Now since the infimum is over all |ψi ∈ Z ⊗ X satisfying TrZ [|ψihψ|] = ρ the result
is independent of the specific purification ψ and we have:

Entanglement fidelity is equivalent to gate fidelity for pure states and hence
average entanglement fidelity is equivalent to average gate fidelity. This can be
shown graphically as follows

Alternatively we can also define the average gate fidelity in terms of the entangle-
ment fidelity with the identity operator

d + Fe (E, 1)
FE = . (36.59)
d(d + 1)
.

– 143 –
 PART

Counting Solutions by Tensor Con- VII

traction
37 Introduction
This chapter presents methods to count via tensor contractions.

38 Returning to Boolean Quantum States

Boolean states were first considered in detail in the second lecture and then were
considered again in lecture three. Here we again recall the key notations, with a
more succinct presentation that utilizes some proof strategies we have developed
for the purpose of reasoning about them [102–104].
Remark. A quantum state is called Boolean iff it can be written in a local
basis with amplitude coefficients taking only binary values 0 or 1. We relate
such states with Boolean functions, allowing for a host of tools from algebra
to be applied to their analysis. The present note derives several relations of
these states, related to the contraction of the corresponding tensor networks.
Remark (Notation). We use B to denote a Boolean bit, given by an element of
the set {0, 1}. A number in Bn then denotes an n-long Boolean bit string. If
x is a bit string, then we use |xi as an index for a basis state. If f : Bn → B
then |f (x)i also indexes a basis state.

Definition 38.1 (The class of Boolean quantum states). Let

f : Bn → B (38.1)

be any switching function. Then

ψB = |xi |f (x)i (38.2)

X

is an arbitrary representative in the class of Boolean states. In this fashion,

every Boolean function gives rise to a quantum state. Conversely, every quan-
tum state written in a local basis with amplitude coefficients taking binary
values in {0, 1} gives rise to a Boolean function. This defines the so called,
class of Boolean quantum states [22].

Theorem 38.1 (Boolean tensor network states [22]). A tensor network repre-
senting a Boolean quantum state is determined from the classical network
description of the corresponding function.

Theorem 38.1 was proven in lecture two, where the quantum tensor networks
are found by letting each classical gate act on a linear space and from changing the
composition of functions, to the contraction of tensors.
144
Contracting networks to solve SAT instances
Theorem 38.2 (Counting 3-SAT solutions). Let f be given to represent a 3-SAT
instance. Then the standard two-norm length squared can be made to give
the number of satisfying assignments of the instance [102].
Proof. The quantum state takes the form

ψf = |xi hf (x)|1i = f (x) |xi (38.3)

X X

x x

We calculate the inner product of this state with itself viz

||ψ||2 = f (x)f (y)hx, yi = f (x) (38.4)

X X

xy x

which gives exactly the number of satisfying inputs. This follows since
f (x)f (y) = δxy . We note that for Boolean states, the square of the two
norm in fact equals the one norm.

Remark (Counting 3-SAT solutions). We note that solving the counting prob-
lem (38.2) for general formula is known to be #P-complete.
Corollary 38.3 (Solving 3-SAT instances). The condition

||ψf || > 0 (38.5)

implies that the SAT instance corresponding to f has a satisfying assignment.

Determining if this condition holds for general Boolean states is an NP-complete
decision problem.14 Note that determining this for 14 Instead of SAT, sometimes
the term UNIQUE-SAT or USAT
ψf = |xi hf (x)|0i = (1 − f (x)) |xi (38.6)
X X
is used to describe the problem
x x of determining whether a formula
known to have either zero or one
in general is a tautology problem. satisfying assignments has zero or
has one.
Graphical depiction and physicality
Remark (Graphical depiction). The algorithm is depicted below. (a) gives a
network realization of the function and determining if the network in (b)
contracts to a value greater than zero solves a SAT instance.

(a) = (b)
...

...

As it happens, some time ago Penrose proved a theorem which applies directly
to the physicality of Boolean satisfiable states. We changed the wording of the
theorem only slightly, changing spin network to tensor network.

Theorem 38.4 (Penrose, 1967). The norm of a tensor network vanishes iff the
physical situation it represents is forbidden by the rules of quantum mechanics
[20].

– 145 –
 The above theorem applies to quantum states. Consider instead a process that
involved the impossibility of measuring a state to be in a certain state. To capture
when such a process is impossible, we modify Penrose’s theorem as follows.

Corollary 38.5 (Forbidden process). The contraction of a tensor network van-

ishes iff the physical situation it represents is forbidden by the rules of quantum
mechanics.
Example (Examples of Penrose’s theorem). Consider a Bell state Φ+ = |00i +
|11i. The amplitude of the first party measuring |0i followed by the second
party measuring |1i is zero. This vanishing tensor network contraction is given
by 01Φ+ . A second example is found by considering the norm of a state |ψi

 

formed by a network of connected tensors, by taking an inner product with

a conjugated copy of itself hψ|ψi. If this inner product vanishes, the network
necessarily represents a non-physical quantum state, by Penrose’s theorem.
Corollary 38.6. All physical Boolean states are satisfiable.
Remark (read-once). A function f is called read-once iff it can be represented
as a Boolean expression using the operations conjunction, disjunction and
negation, in which every variable appears exactly once. We call such a factored
expression a read-once expression for f . These correspond exactly to fan-out
only circuits. From this structure we conclude directly that
Corollary 38.7. All read-once formula are satisfiable [102].
Remark (Quantum read-once). A quantum quantum state is called read-once
if it can be represented by a tensor tree containing only isometries.
Corollary 38.8. All quantum read-once formula are satisfiable, with the evaluation
of k-point functions polynomial in the particle number.

39 Returning to Stabilizer Tensor Theory

In § 20, we talked in detail about the Clifford group and stabilizer theory. In fact,
we proved the following theorem, which connects the theory of stabilizer states to
the theory tensor networks which represent pseudo Boolean forms, which we have
developed in our work, and presented in this book.

Theorem 39.1 (Stabilizer states as pseudo Boolean forms). Let

f, g, k : Bn → B (39.1)

then the quantum state

ψB = (−1)f (x) (i)g(x) k(x) |xi (39.2)

X

is sufficient to express any stabilizer state.

We will now take a step in the other direction. That is, we wish to understand
what Boolean states are stabilizer states. Here we will consider the class of linear
quantum states. We consider the general theory elsewhere.

– 146 –
 Definition 39.1 (The class of linear quantum states). We define the linear class
of quantum states as quantum states of the form

ψ⊕L = (39.3)
X
c0 ⊕ c1 x1 ⊕ c2 x2 ⊕ ... ⊕ cn xn |x1 , x2 , ..., xn i

where ∀i, ci = 0, 1 selects the linear function uniquely. (As we have mentioned,
c0 = 1 results technically in the affine class of classical circuits, but we still
define this full class as, the class of linear quantum states.)

Note that the laws of the algebra enforce the strong constraint, x ⊕ x = 0 and
0 ⊕ y = y. So we find immediately that we need only consider two fully entangled
states in this class, as every other state is found from a direct product of states of
this form. The first is

ψ1 = (39.4)
X
x1 ⊕ x2 ⊕ ... ⊕ xn |x1 , x2 , ..., xn i

and the second is given by

ψ2 = 1 ⊕ x1 ⊕ x2 ⊕ ... ⊕ xn |x1 , x2 , ..., xn i (39.5)

X

The tensor network differs only by contraction with the constant |1i. ψ1 is shown
in (a) and ψ2 is shown in (b).

1
(a) (b)

... ...
We will now consider the stabilizers of each of these cases, (a) and (b).
Remark (Stabilizers of case (a)). The network in (a) is found from a Hadamard
transform on all the legs of a COPY-tensor. The 2n stabilizers of the COPY-
tensor are generated by the n operators

X1 ⊗ X2 ⊗ ... ⊗ Xn (39.6)

Zi ⊗ Zj , 0≤i<j≤n (39.7)
We have considered in lecture III how stabilizers transform. Under the
Hadamard transform, the stabilizer generators transform to

Z1 ⊗ Z2 ⊗ ... ⊗ Zn (39.8)

Xi ⊗ Xj , 0≤i<j≤n (39.9)

40 Elementary Theorems of Tensor Contraction

The following can be used to prove graphical identities and represents a conceptual
tool to aid in the analysis and design of tensor networks as a conceptual framework
(as advocated in this lecture series) as well as a numerical tool for the simulation
of quantum and classical physics.

– 147 –
Remark (Linearity of tensor contraction). Tensor contraction is linear in its
arguments. If A is a tensor in a fully contracted network C{A}, if we let
A 7→ A0 + B and then A 7→ kA we readily find that the contraction becomes
C{A0 } + C{B} and k · C{A} respectively.

Theorem 40.1 (Contraction to sum of products transform). Given a tensor

Γ...ijklmn... in a fully contracted network (e.g. a network without open legs),
the following graphical identity transforms the contraction, to a sum over
products.

The circle is meant as an abstraction depicting a fully contracted but otherwise

unknown network.

Theorem 40.2 (COPY-tensors as a resolution of identity). The following se-

quence of graphical rewrites hold.

Proof. In the above figure, on the left, we abstractly depict a other

wise arbitrary tensor network, by showing only one single wire. The unit
for the COPY-tensor is the plus state |+i. This state is defined as a
sum over basis states |ni. The tensor copies these basis states, and splits
apart.

Theorem 40.3 (A tensor contraction inequality). Given a contracted network

and a partition into two halves x, and y. Writing the contraction as C{x, y}
the following inequality holds.

C{x, y} ≤ C{x, x} · C{y, y} (40.1)

with graphical depiction.

x y x x y y

– 148 –
 Proof. By the linearity of tensor contraction, we arrive at an abstract
form of the Cauchy-Schwarz inequality, with equality in the contraction
iff x = α · y. This leads directly to the concept of an angle between
tensors,
C{x, y}
cos θxy = (40.2)
C{x} · C{y}
where the right side is either real valued, or we take the modulus.

41 A 3-fold way
We will now unify three concepts.
Remark (Pseudo Boolean function). A function is called pseudo Boolean when
it is total with type
f (x) : Bn → C (41.1)

Fold I. Quantum States

A quantum state is a map
ψ:C→H (41.2)
since ψ(c) = c · ψ ∈ H for c ∈ C and ψ(1) = ψ uniquely determines ψ by linearity.
We typically fix kψk = 1. Note that

hψ|ij · · · ki = cij···k (41.3)

where
ψ= (41.4)
X
cx |xi
x

given basis |xi. We arrive at the following

∃!f : Bn → C | f (x) = cx (41.5)

and f is pseudo Boolean and ψ can be expressed as

ψ= f (x) |xi (41.6)

X

Fold II. Linear Operators in H → H

Theorem 41.1 (Isomorphism between states and diagonal maps). There is an
isomorphism sending every state ψ ∈ H to a diagonal map L(ψ) ∈ H → H.
Moreover, the eigenvalues of L(ψ) are the amplitudes of ψ expressed in the
spin basis.

– 149 –
 Proof. We write
ψ= (41.7)
X
αx |xi
and then by constructing a map such that

|xi → |xi |xi (41.8)

one can construct L(ψ) as

L(ψ) = (41.9)
X
αx |xi hx|

We then let
|+i := |0i + |1i + · · · + |ni (41.10)
Assume we are considering n qubits, then

ψ = L(ψ) |+i⊗n (41.11)

establishes the bijection. We also note that

L(ψ) |ki = αk |ki (41.12)

satisfying the eigenvalue condition and hence, proving the result.

Lemma 41.2 (Tensor networks equating states and diagonal maps). The maps
relied on in the above theorem can be given in terms of tensor networks. L(ψ)
is shown in (a). This map is invertible as shown in (b).

(a) (b)

.
.
.. .. .. ..
. .
...
...
...
...
...

=
...

Remark (Proof strategy). Here we sketch what we call the argument by lin-
earity. We contract all wires of an open diagram as follows. On the left we
contract with hx, y, z| and on the right with |q, p, ri.

(a) (b)

.
.. ..
...
...

. =

This is equivalent to evaluating the coefficients of three party delta functions

as
δxqi δypj δzrk ψijk = cijk (= cxyz = cqpr ) (41.13)
by linearity, we recover the map L(ψ).

– 150 –
 Remark. Note that from ψ = f (x) |xi we have
P

L(ψ) = L f (x) |xi = f (x)L(|xi) = f (x) |xi hx| (41.14)

X  X X

Fold III. Classical Spin Hamiltonians

We will consider a generalized Ising spin Hamiltonian. Let

h= hi Z i + J ij Zi Zj + · · · + (41.15)
X X X
k ij...k Zi Zj · · · Zk

and let si be a spin variable taking values ±1 and let xi be a Boolean valued 0, 1.
Use
si = 1 − 2xi (41.16)
then
Zi = 1 − 2 |1i h1| (41.17)
and so we arrive at

hx = hi (1 − 2xi ) + J ij (1 − 2xi )(1 − 2xj ) + · · ·

X X
(41.18)
+ k ij...k (1 − 2xi )(1 − 2xj ) · · · (1 − 2xk )
X

and we arrive at the expression for ψ

ψ= hx (x) |xi (41.19)

X

In [76] we developed general methods to reason about spin Hamiltonians. We

explored an embedding reducing k-body interactions (e.g. k ij...k Zi Zj · · · Zk ) into
two-body interactions by adding additional qubits.

A three fold way

We have shown that three different concepts, are effectively equivalent by con-
structing mathematical dualities that relate them precisely.

Quantum States

pseudo Boolean
forms

Linear operators
Generalised Ising models
with spin energies a function of

Remark (Factorization of quantum states into Tensor Networks). In the lectures,

we also presented a universal a factorization of quantum states, into networks
comprised of the building blocks (Quantum Legos), not included here. A
related factorization appeared in [22].

– 151 –
Counting Graph Colorings
Given a 3-regular planar graph15 , how many possible edge colorings using three 15 A graph is k-regular iff every

colors exist, such that all edges connected to each node have distinct colors? This node has exactly k edges connected
counting problem can be solved in an interesting (if not computationally efficient) to it.

way using the order-3  tensor, which is defined in terms of components as

012 = 120 = 201 = 1,

021 = 210 = 102 = −1, (41.20)

otherwise zero. The counting algorithm is stated as

Theorem 41.3 (Planar graph 3-colorings, Penrose 1971 [1]). The number K of
proper 3-edge-colorings of a planar 3-regular graph is obtained by replacing
each node with an order-3 epsilon tensor, replacing each edge with a wire, and
then contracting the resulting tensor network.

We will first consider the simplest case, a graph with just two nodes. In this
case we obtain

2 = 6

There are indeed 6 distinct edge colorings for this graph, given as

To understand Theorem 41.3, note first that the contraction K of the epsilon tensor
network is the sum of all possible individual assignments of the index values to the
epsilon tensors comprising the network. Each of the three possible index values can
be understood as a color choice for the corresponding edge. Whenever the index
values for a given epsilon tensor are not all different, the corresponding term in K
is zero. Hence only allowed color assignments result in nonzero contributions to K,
and for a graph that does not admit a proper 3-edge-coloring we will have K = 0.
For instance, for the non-3-colorable Petersen graph we obtain

– 152 –
 ?

= 0
?

However, for K to actually equal the number of allowed colorings, each nonzero
term must have the value 1 (and not −1). This is only guaranteed if the graph is
planar, as can be seen by considering the non-planar graph K3,3 :

= 0

The edges can be colored with three colors—in 12 different ways—yet the contrac-
tion vanishes.
The computational complexity of this problem has been studied in [105]. In-
teresting, by a well known result (Heawood 1897), the 3-colorings as stated above,
are one quarter of the ways of coloring the faces of the graph with four colors, so
that no two like-colored faces have an edge in common.
Example (Physical implementation of abc in quantum computing). In quantum
computing, typically one works with qubits (two level quantum systems) but
implementations using qutrits exist (three level quantum systems, available in
e.g. nitrogen vacancy centers in diamond—see for instance [106]). The epsilon
tensor abc could be realized directly as a locally invariant 3-party state using

– 153 –
qutrits, and can also be embedded into a qubit system. We leave it to the
reader to show that by pairing qubits, abc can be represented with six qubits,
where each leg now represents a qubit pair. (Note that a basis of 3 states can
be isometrically embedded in 4-dimensional space in any number of ways.)
Show further that the construction can be done such that the two qubit pairs
(together representing one leg) are symmetric under exchange.

Exercise 41.4 (Representation of the Epsilon Tensor on Qubit States). The

epsilon tensor is typically considered in C3 ⊗ C3 ⊗ C3 . Show that by pairing
qubits, this can be represented in the space C2⊗6 of six qubits, where each
leg now represents a qubit pair. Show further that the construction can be
done such that the two qubit pairs are symmetric under exchange.

– 154 –
 PART

Appendix VIII
A Algebra on Quantum States
We are concerned with a network theory of quantum states. This on the one hand
can be used as a tool to solve problems about states and operators in quantum
theory, but does have a physical interpretation on the other. This is not foun-
dational per se but instead largely based on what one might call an operational
interpretation of quantum states and processes. A related idea has been used to
study non-locality in quantum physics [77]. This section stems from those ideas
[77] which Bill Edwards introduced me to in Oxford around circa. 2010.
We call an algebra a pairing on a vector space, taking two vectors and producing
a third (you might instead call it a monoid if there is a unit, and then a group if
the set of considered vectors is closed under the product). Let’s now examine how
every tripartite quantum state forms an algebra.
Consider a tripartite quantum state (subsystems labeled 1,2 and 3), and then
ask the question: “how would the state of the third system change after measure-
ment of systems one and two?” Enter Algebras: as stated, an algebra on a vector
space, or on a Hilbert space is formed by a product taking two elements from the
vector space to produce a third element in the vector space. Algebra on states
can then be studied by considering duality of the state, that is considering the
adjunction between the maps of type

1→H⊗H⊗H and H⊗H→H (A.1)

This duality is made evident by using the †-compact structure of the category
(e.g. the cups and caps). It is given vivid physical meaning by considering the
effect measuring (that is two events) two components of a state has on the third
component.
Remark (Overbar notation on Spaces). Given a Hilbert space H, we can con-
sider the Hilbert space H which can be simply thought of as the Hilbert space
H will all basis vectors complex conjugates (overbar). That is, H is a vector
space whose elements are in one-to-one correspondence with the elements of
H:
H = {v | v ∈ H}, (A.2)
with the following rules for addition and scalar multiplication:

v+w = v+w and αv = αv. (A.3)

Remark (Definition of Algebra). We consider an algebra as a vector space A
endowed with a product, taking a pair of elements (e.g. from A ⊗ A) and
producing an element in A. So the product is a map A ⊗ A → A, which
may not be associative or have a unit (that is, a multiplicative identity — see
Example 13.3.1 for an example of an algebra on a quantum state without a
unit).

155
Observation A.1 (Every tripartite Quantum State Forms an Algebra). Let |ψi ∈
H ⊗ H ⊗ H be a quantum state and let Mi , Mj be complete sets of measurement
operators. Then (|ψi , Mi , Mj ) forms an algebra.

= := =
time

The quantum state |Ψi = ijk ψ ijk |ijki is drawn as a triangle, with the iden-
P

tity operator on each subsystem acting as time goes to the right on the page (rep-
resented as a wire). Projective measurements with respect to Mi and Mj are made.
We define these complete measurement operators as
N
M1 = (A.4)
X
i |ψi i hψi |
i=1

N
M2 = (A.5)
X
j |φj i hφj |
j=1

such that we recover the identity operator on the N -level subsystem viz
N N
|φj i hφj | = |ψi i hψi | = 1N (A.6)
X X

j=1 i=1

The measurements result in eigenvalues i, j leaving the state of the unmeasured

system in
|ωi = (A.7)
X D x E D y  E
ψ xyz ψ x φ y |zi
 
xyz

where Q = |Qi> that is, the transpose is factored into: (i) taking the dagger
D 
 def

(diagrammatically this mirrors states across the page) and (ii) taking the complex
conjugate. Hence,
 E†
Q = |Qi = Q = |Qi (A.8)
D 
 >  †

and if we pick a real valued basis for |xi , |yi , |zi = |0i , |1i we recover

|ωi = (A.9)
X
ψ xyz hx|ψx i hy|φy i |zi
xyz

As stated, this physical interpretation is not our main interest. Even in its
absence, we’re able to write down and represent a quantum state purely in terms
of a connected network, where each component is fully defined in terms of algebraic
laws.

– 156 –
B XOR-algebra
Here we review the concept of an algebraic normal form (ANF) for Boolean poly-
nomials, commonly known as PPRMs, (Positive Polarity Reed Muller Forms). See
the reference book [64] and the historical references [66, 67] for further details.

Definition B.1. The XOR-algebra forms a commutative ring with presentation

M = {B, ∧, ⊕} where the following product is called XOR

— ⊕ — : B × B 7→ B :: (a, b) → a + b − ab mod 2 (B.1)

and conjunction is given as

— ∧ — : B × B 7→ B :: (a, b) → a · b, (B.2)

where a · b is regular multiplication over the reals. One defines left negation
¬(—) in terms of ⊕ as ¬(—) ≡

1 ⊕ (—) : B 7→ B :: a → 1 − a. (B.3)

In the XOR-algebra, 1-5 hold. (i) a ⊕ 0 = a, (ii) a ⊕ 1 = ¬a, (iii) a ⊕ a = 0, (iv)

a ⊕ ¬a = 1 and (v) a ∨ b = a ⊕ b ⊕ (a ∧ b). Hence, 0 is the unit of XOR and 1
is the unit of AND. The 5th rule reduces to a ∨ b = a ⊕ b whenever a ∧ b = 0,
which is the case for disjoint (mod 2) sums. The truth table for AND follows

x1 x2 f (x1 , x2 ) = x1 ∧ x2
0 0 0
0 1 0
1 0 0
1 1 1

Definition B.2. Any Boolean equation may be uniquely expanded to the fixed
polarity Reed-Muller form as:

f (x1 , x2 , ..., xk ) = c0 ⊕ c1 xσ1 1 ⊕ c2 xσ2 2 ⊕ · · · ⊕ cn xσnn ⊕

cn+1 xσ1 1 xσnn ⊕ · · · ⊕ c2k−1 xσ1 1 xσ2 2 , ..., xσk k , (B.4)

where selection variable σi ∈ {0, 1}, literal xσi i represents a variable or its
negation and any c term labeled c0 through cj is a binary constant 0 or 1. In
Equation (B.4) only fixed polarity variables appear such that each is in either
un-complemented or complemented form.

Let us now consider derivation of the form from Definition B.2. Because of
the structure of the algebra, without loss of generality, one avoids keeping track of
indices in the N node case, by considering the case where N ≡ 2n = 8.
Example. The vector

c = (c0 , c1 , c2 , c3 , c4 , c5 , c6 , c7 , )| (B.5)

represents all possible outputs of any function f (x1 , x2 , x3 ) over the algebra
formed from linear extension of Z2 × Z2 × Z2 . We wish to construct a normal

– 157 –
 form in terms of the vector c, where each ci ∈ {0, 1}, and therefore c is a
selection vector that simply represents the output of the function

f : B × B × B → B : (x1 , x2 , x3 ) 7→ f (x1 , x2 , x3 ). (B.6)

One may expand f as:

f (x1 , x2 , x3 ) = (c0 · ¬x1 · ¬x2 · ¬x3 ) ∨ (c1 · ¬x1 · ¬x2 · x3 ) ∨ (c2 · ¬x1 · x2 · ¬x3 )
∨(c3 · ¬x1 · x2 · x3 ) ∨ (c4 · x1 · ¬x2 · ¬x3 ) ∨ (c5 · x1 · ¬x2 · x3 )
∨(c6 · x1 · x2 · ¬x3 ) ∨ (c7 · x1 · x2 · x3 ) (B.7)

Since each disjunctive term is disjoint the logical OR operation may be replaced
with the logical XOR operation. By making the substitution ¬a = a ⊕ 1 for
all variables and rearranging terms one arrives at the following normal form: For instance, ¬x1 ·¬x2 ·¬x3 =
(1 ⊕ x1 ) · (1 ⊕ x2 ) · (1 ⊕ x3 ) =
(1⊕x1 ⊕x2 ⊕x2 ·x3 )·(1⊕x3 ) =
f (x1 , x2 , x3 ) = c0 ⊕ (c0 ⊕ c4 ) · x1 ⊕ (c0 ⊕ c2 ) · x2 ⊕ (c0 ⊕ c1 ) · x3 (B.8) 1 ⊕ x1 ⊕ x2 ⊕ x3 ⊕ x1 · x3 ⊕
⊕(c0 ⊕ c2 ⊕ c4 ⊕ c6 ) · x1 · x2 x2 · x3 ⊕ x1 · x2 · x3 .
⊕(c0 ⊕ c1 ⊕ c4 ⊕ c5 ) · x1 · x3 ⊕ (c0 ⊕ c1 ⊕ c2 ⊕ c3 ) · x2 · x3
⊕(c0 ⊕ c1 ⊕ c2 ⊕ c3 ⊕ c4 ⊕ c5 ⊕ c6 ⊕ c7 ) · x1 · x2 · x3 (B.9)

The set of linearly independent vectors, {x1 , x2 , x3 , x1 · x2 , x1 · x3 , x2 · x3 , x1 ·

x2 · x3 } combined with a set of scalars from Equation B.8 spans the eight
dimensional space of the Hypercube representing the Algebra. A similar form
holds for arbitrary N .

f (x1 , x2 , x3 ) = (a1 ) · x1 ⊕ (a2 ) · x2 ⊕ (x3 ) · x3 ⊕ (a1 ⊕ a2 ⊕ a1 ⊕ c2 ) · x1 · x2

⊕(a1 ⊕ a3 ⊕ a1 ⊕ c3 ) · x1 · x3 ⊕ (a2 ⊕ a3 ⊕ a2 ⊕ c3 ) · x2 · x3
⊕(a1 ⊕ a2 ⊕ a3 ⊕ a1 ⊕ a2 ⊕ a3 ) · x1 · x2 · x3 (B.10)

C The Minimization Method of Karnaugh

The Karnaugh map is a tool to facilitate the algebraic reduction of Boolean func-
tions. Many excellent texts and online tutorials cover the use of Karnaugh maps
and should be consulted for more detail.1 This Appendix briefly introduces these
maps to make the lecture notes self contained.
Karnaugh maps (see Table 5 for three examples), or more compactly K-maps,
are organized so that the truth table of a given equation, such as a Boolean equation
(f : Bn → B) or multi-linear form (f : Bn → R), is arranged in a grid form and
between any two adjacent boxes only one domain variable can change value.
This ordering results as the rows and columns are ordered according to Gray
code — a binary numeral system where two successive values differ in only one
digit. For example, the 4-bit Gray code is given as:

{0000, 0001, 0011, 0010, 0110, 0111, 0101, 0100, 1100,

1101, 1111, 1110, 1010, 1011, 1001, 1000}.
1
This includes the wikipedia entry (http://en.wikipedia.org) and the articles linked to
therein as well as the straight forward reference [107].

– 158 –
 x1 x2 x1 x2
00 01 11 10 00 01 11 10

0 0 1 3 2 0 0 1 3 2
z∗ z∗
1 4 5 7 6 1 4 5 7 6

(a) (b)
x1 x2
00 01 11 10

0 0 0 . 0
z∗
1 . . 0 .

(c)

Table 5. Karnaugh maps: (a) 2-local (positive polarity) variable couplings. (b) Linear
(positive polarity) terms. (c) A Karnaugh map illustrating (with ovals) the linear and
quadratic terms needed to an example function.

By arranging the truth table of a given function in this way, a K-map can be used
to derive a minimized function.
To use a K-map to minimize a Boolean function one covers the 1s on the map
by rectangular coverings containing a number of boxes equal to a power of 2. For
example, one could circle a map of size 2n for any constant function f = 1. Table 5
(a) and (b) contain three circles each — all of 2 and 4 boxes respectively. After the
1s are covered, a term in a sum of products expression [65] is produced by finding
the variables that do not change throughout the entire covering, and taking a 1 to
mean that variable (xi ) and a 0 as its negation (xi ). Doing this for every covering
yields a function which matches the truth table.
For instance consider Table 5 (a) and (b). Here the boxes contain simply
labels representing the decimal value of the corresponding Gray code ordering.
The circling in Table 5 (a) would correspond to the truth vector (ordered z? , x1
then x2 )
(0, 0, 0, 1, 0, 1, 1, 1)T . (C.1)
The cubes 3 and 7 circled in Table 5 correspond to the sum of products term x1 x2 .
Likewise (5,7) corresponds to z? x2 and finally (7,6) corresponds to z? x1 . The sum
of products representation of (C.1) is simply

f (z? , x1 , x2 ) = x1 x2 ∨ z? x2 ∨ z? x1 .

Let us repeat the same procedure for Table 5 b.) by again assuming the circled
cubes correspond to 1s in the functions truth table. In this case one finds z? for
the circling of cubes ladled (4,5,7,6), x2 for (1,3,5,7) and x1 for (3,2,7,6) resulting
in the function
f (z? , x1 , x2 ) = x1 ∨ z? ∨ x2 .

– 159 –
 Definition C.1. (Davio Expansion) The Davio expansion is a decomposition
of a boolean function. For a boolean function f (x1 , ..., xn ) we set with respect
to xi :

fxi (x) = f (x1 , ..., xi−1 , 1, xi+1 , ..., xn ) (C.2)

fxi (x) = f (x1 , ..., xi−1 , 0, xi+1 , ..., xn ) (C.3)
∂f
= fxi (x) ⊕ fxi (x) (C.4)
∂xi
as the positive and negative cofactors of f , and the boolean derivation of f .
Then we have for the Reed-Muller or positive Davio expansion:
∂f
f = fxi ⊕ xi (C.5)
∂xi

D Tensors and Tensor Products

The definition of a tensor starts with the tensor product ⊗. There are many equiv-
alent ways to define it, but perhaps the simplest one is through basis vectors. Let
V and W be finite-dimensional vector spaces over the same field of scalars K. In
physics-related applications K is typically either the real numbers R or the complex
numbers C. Now V ⊗ W is also a vector space over K. If V and W have the bases
{ej }j and {fk }k , respectively, the symbols {ej ⊗ fk }jk form a basis for V ⊗ W .
Thus, for finite-dimensional spaces dim(V ⊗ W ) = dim V dim W .
The tensor product of two individual vectors v ∈ V and w ∈ W is denoted
as v ⊗ w. For vectors the tensor product is a bilinear map V × W → V ⊗ W ,
i.e. one that is linear in both input variables. For finite-dimensional spaces one can
obtain the standard basis coordinates of the tensor product of two vectors as the
Kronecker product of the standard basis coordinates of the individual vectors:

(v ⊗ w)jk = v j wk . (D.1)

It is important to notice that due to the bilinearity ⊗ maps many different pairs
of vectors (v, w) to the same product vector: v ⊗ (sw) = (sv) ⊗ w = s(v ⊗ w),
where s ∈ K. For inner product spaces (such as the Hilbert spaces encountered in
quantum mechanics) the tensor product space inherits the inner product from its
constituent spaces:

hv1 ⊗ w1 , v2 ⊗ w2 iV ⊗W = hv1 , v2 iV hw1 , w2 iW . (D.2)

A tensor T is an element of the tensor product of a finite number of vector

spaces over a common field of scalars K. The dual space V ∗ of a vector space V
is defined as the space of linear maps from V to K. It is not hard to show that
V ∗ is a vector space over K on its own. This leads us to define the concept of an
order-(p, q) tensor, an element of the tensor product of p primal spaces and q dual
spaces:
T ∈ W1 ⊗ W2 ⊗ . . . ⊗ Wp ⊗ V1∗ ⊗ V2∗ ⊗ . . . ⊗ Vq∗ . (D.3)
Given a basis {e(i) k }k for each vector space Wi and a dual basis {η (i)k }k for each

– 160 –
dual space Vi∗ , we may expand T in the tensor products of these basis vectors:

T =T (D.4)
i1 ...ip (1)
j1 ...jq e i1 ⊗ . . . ⊗ e(p) ip ⊗ η (1)j1 ⊗ . . . ⊗ η (q)jq .

T 1 pj1 ...jq is simply an array of scalars containing the basis expansion coefficients.
i ...i

Here we have introduced the Einstein summation convention, in which any index
that is repeated exactly twice in a term, once up, once down, is summed over. This
allows us to save a considerable number of sum signs, without compromising on the
readability of the formulas. Traditionally basis vectors carry a lower (covariant)
index and dual basis vectors an upper (contravariant) index.
A tensor is said to be simple if it can be written as the tensor product of some
elements of the underlying vector spaces: T = v (1) ⊗ . . . ⊗ v (q) ⊗ ϕ(1) ⊗ . . . ⊗ ϕ(p) .
This is not true for most tensors; indeed, in addition to the bilinearity, this is one
of the properties that separates tensors from mere Cartesian products of vectors.
However, any tensor can be written as a linear combination of simple tensors, e.g. as
in Eq. (D.4).
For every vector space W there is a unique bilinear map W ⊗ W ∗ → K,
w ⊗ φ 7→ φ(w) called a natural pairing, where the dual vector maps the primal
vector to a scalar. One can apply this map to any pair of matching primal and
dual spaces in a tensor. It is called a contraction of the corresponding upper and
lower indices. For example, if we happen to have W1 = V1 we may contract the
corresponding indices on T :

C1,1 (T ) = T (e i1 ) e(2) i2 ⊗ . . . ⊗ e(p) ip ⊗ η (2)j2 ⊗ . . .

i1 ...ip (1)j1 (1)
j1 ...jq η ⊗ η (q)jq
= (D.5)
k i ...i
T 2 pk i2 ...jq e(2) i2 ⊗ . . . ⊗ e(p) ip ⊗ η (2)j2 ⊗ . . . ⊗ η (q)jq ,

since the defining property of a dual basis is η (1)j1 (e(1) i1 ) = δ j1i1 . Hence the
contraction eliminates the affected indices (k is summed over), lowering the tensor
order by (1, 1).
We can see that an order-(1, 0) tensor is simply a vector, an order-(0, 1) tensor is
a dual vector, and can define an order-(0, 0) tensor to correspond to a plain scalar.
But what about general, order-(p, q) tensors? How should they be understood?
Using contraction, they can be immediately reinterpreted as multilinear maps from
vectors to vectors:

T 0 : V 1 ⊗ . . . ⊗ V q → W1 ⊗ . . . ⊗ Wp ,
T 0 (v (1) ⊗ . . . ⊗ v (q) ) = T ⊗ . . . ⊗ e(p) ip × η (1)j1 (v (1) ) × . . . × η (q)jq (v (q) ),
i1 ...ip (1)
j1 ...jq e i1
(D.6)

where we tensor-multiply T and the vectors to be mapped together, and then

contract the corresponding indices. However, this is not the only possible interpre-
tation. We could just as easily see them as mapping dual vectors to dual vectors:

T 00 : W1∗ ⊗ . . . ⊗ Wp∗ → V1∗ ⊗ . . . ⊗ Vq∗ ,

T 00 (ϕ(1) ⊗ . . . ⊗ ϕ(p) ) = T ϕ(1) (e(1) i1 ) × . . . × ϕ(p) (e(p) ip ) × η (1)j1 ⊗ . . . ⊗ η (q)jq .
i1 ...ip
j1 ...jq
(D.7)

Essentially we may move any of the vector spaces to the other side of the arrow by

– 161 –
taking their dual:

W ⊗V∗ =
∼ K→W ⊗V∗ =
∼ V →W =
∼ V ⊗ W∗ → K =
∼ W ∗ → V ∗,
(D.8)

where all the arrows denote linear maps. Any and all input vectors are mapped
to scalars by the corresponding dual basis vectors in expansion (D.4), whereas all
input dual vectors map the corresponding primal basis vectors to scalars.
If we expand the input vectors v (k) in Eq. (D.6) using the same bases as
when expanding the tensor T, we obtain the following equation for the expansion
coefficients:

T 0 (v (1) ⊗ . . . ⊗ v (q) )i1 ...ip = T (D.9)

i1 ...ip
j1 ...jq v (1)j1 · · · v (q)jq .

This is much less cumbersome than Eq. (D.6), and contains the same information.
This leads us to adopt the abstract index notation for tensors, in which the indices
no longer denote the components of the tensor in a particular basis, but instead
signify the tensor’s order. Tensor products are denoted by simply placing the tensor
symbols next to each other. Within each term, any repeated index symbol must
appear once up and once down, and denotes contraction over those indices. Hence,
xa denotes a vector (with one contravariant index), ωa a dual vector (with one
covariant index), and T abc an order-(2, 1) tensor with two contravariant and one
covariant indices. S abcde xc y d P ea denotes the contraction of an order-(2, 3) tensor S,
an order-(1, 1) tensor P , and two vectors, x and y, resulting in an order-(1, 0) tensor
with one uncontracted index, b.
In many applications, for example in differential geometry, the vector spaces
associated with a tensor are often copies of the same vector space V or its dual V ∗ ,
which means that any pair of upper and lower indices can be contracted, and leads
to the tensor components transforming in a very specific way under basis changes.
This specific type of a tensor is called an order-(p, q) tensor on the vector space V .
However, here we adopt a more general definition, allowing {Vk }k and {Wk }k to
be all different vector spaces.

– 162 –
References
[1] Roger Penrose. Applications of negative dimensional tensors. Combinatorial
Mathematics and its Applications, Academic Press, 1971.
[2] D. Deutsch. Quantum computational networks. Proceedings of the Royal Society of
London A: Mathematical, Physical and Engineering Sciences, 425(1868):73–90,
1989.
[3] Richard P. Feynman. Quantum mechanical computers. Foundations of Phys.,
16:507, 1986.
[4] R. Orús. A practical introduction to tensor networks: Matrix product states and
projected entangled pair states. Annals of Physics, 349:117–158, October 2014.
[5] G. Vidal. Entanglement renormalization: an introduction. In Lincoln D. Carr,
editor, Understanding Quantum Phase Transitions. Taylor & Francis, Boca Raton,
2010.
[6] F. Verstraete, V. Murg, and J. I. Cirac. Matrix product states, projected entangled
pair states, and variational renormalization group methods for quantum spin
systems. Advances in Physics, 57:143–224, 2008.
[7] J. I. Cirac and F. Verstraete. Renormalization and tensor product states in spin
chains and lattices. J. Phys. A Math. Theor., 42(50):504004, 2009.
[8] U. Schollwöck. The density-matrix renormalization group in the age of matrix
product states. Annals of Physics, 326:96–192, January 2011.
[9] S. Sachdev. Viewpoint: Tensor networks—a new tool for old problems. Physics,
2:90, 2009.
[10] Ulrich Schollwöck. The density-matrix renormalization group: a short
introduction. Philosophical Transactions of the Royal Society of London A:
Mathematical, Physical and Engineering Sciences, 369(1946):2643–2661, 2011.
[11] R. Orús. Advances on tensor network theory: symmetries, fermions, entanglement,
and holography. European Physical Journal B, 87:280, November 2014.
[12] J. Eisert. Entanglement and tensor network states. Modeling and Simulation,
3:520, August 2013.
[13] G. Evenbly and G. Vidal. Tensor Network States and Geometry. Journal of
Statistical Physics, 145:891–918, November 2011.
[14] Jacob C Bridgeman and Christopher T Chubb. Hand-waving and interpretive
dance: an introductory course on tensor networks. Journal of Physics A:
Mathematical and Theoretical, 50(22):223001, may 2017.
[15] Andrzej Cichocki, Namgil Lee, Ivan Oseledets, Anh-Huy Phan, Qibin Zhao, and
Danilo P. Mandic. Tensor networks for dimensionality reduction and large-scale
optimization: Part 1 low-rank tensor decompositions. Foundations and Trends in
Machine Learning, 9(4-5):249–429, 2016.
[16] Anastasiia A. Pervishko and Jacob Biamonte. Pushing tensor networks to the
limit. Physics, 12, May 2019.
[17] Andrzej Cichocki, Anh-Huy Phan, Qibin Zhao, Namgil Lee, Ivan Oseledets,
Masashi Sugiyama, and Danilo P. Mandic. Tensor networks for dimensionality
reduction and large-scale optimization: Part 2 applications and future perspectives.
Foundations and Trends in Machine Learning, 9(6):431–673, 2017.
[18] Jacob Biamonte and Ville Bergholm. Tensor networks in a nutshell, 2017.
[19] M. J. Hartmann, J. Prior, S. R. Clark, and M. B. Plenio. Density matrix
renormalization group in the heisenberg picture. Physical Review Letters,
102(5):057202, February 2009.

– 163 –
[20] Roger Penrose. The theory of quantized directions. unpublished, 1967.
[21] Roger Penrose. The road to reality. Alfred A. Knopf, Inc., New York, 2005. A
complete guide to the laws of the universe.
[22] J. D. Biamonte, S. R. Clark, and D. Jaksch. Categorical tensor network states.
2010.
[23] V. Bergholm and J. D. Biamonte. Categorical quantum circuits. Journal of
Physics A Mathematical General, 44(24):245304, June 2011.
[24] Michael Nielsen and Isaac Chuang. Quantum computation and quantum
information. Cambridge University Press, 2000.
[25] Bob Coecke and Aleks Kissinger. Picturing Quantum Processes. Cambridge
University Press, 2017.
[26] Jacob Biamonte, Ville Bergholm, and Marco Lanzagorta. Tensor network methods
for invariant theory. Journal of Physics A Mathematical General, 46(47):475301,
Nov 2013.
[27] A. Critch and J. Morton. Algebraic Geometry of Matrix Product States. SIGMA,
10:095, September 2014.
[28] William K. Wootters. Entanglement of formation of an arbitrary state of two
qubits. Phys. Rev. Lett., 80:2245–2248, Mar 1998.
[29] V. Coffman, J. Kundu, and W. K. Wootters. Distributed entanglement. pra,
61(5):052306, May 2000.
[30] Julia Kempe. Multiparticle entanglement and its applications to cryptography.
Phys. Rev. A, 60:910–916, Aug 1999.
[31] Adriano Barenco, Charles H. Bennett, Richard Cleve, David P. DiVincenzo,
Norman Margolus, Peter Shor, Tycho Sleator, John A. Smolin, and Harald
Weinfurter. Elementary gates for quantum computation. Phys. Rev. A,
52(5):3457–3467, 1995.
[32] S. J. Denny, J. D. Biamonte, D. Jaksch, and S. R. Clark. Algebraically contractible
topological tensor network states. Journal of Physics A Mathematical General,
45(1):015309, Jan 2012.
[33] Jacob Biamonte. Charged string tensor networks. Proceedings of the National
Academy of Sciences, 114(10):2447, 2017.
[34] Christopher J. Wood, Jacob D. Biamonte, and David G. Cory. Tensor networks
and graphical calculus for open quantum systems. Quantum Information &
Computation, 15(9-10):759–811, July 2015.
[35] C. Kassel. Quantum groups. Springer Graduate Texts in Mathematics, 1994.
[36] Roger A. Horn and Charles R. Johnson. Matrix Analysis. Cambridge University
Press, New York, NY, USA, 2nd edition, 2012.
[37] John C. Baez and Aaron D. Lauda. A prehistory of n-categorical physics. Deep
Beauty, page 13–128, 2011.
[38] Samson Abramsky and Bob Coecke. Categorical quantum mechanics. Chapter in
the Handbook of Quantum Logic and Quantum Structures vol II, Elsevier, 2008.
[39] Sebastian Meznaric and Jacob Biamonte. Tensor networks for entanglement
evolution. In Advances in Chemical Physics, pages 567–580. John Wiley & Sons,
Inc., March 2014.
[40] Daniel M. Greenberger, Michael A. Horne, and Anton Zeilinger. Going beyond
bell’s theorem, 1989. in: Bell’s Theorem, Quantum Theory, and Conceptions of the
Universe, M. Kafatos (Ed.), Kluwer, Dordrecht, 69-72.
[41] John C. Baez. Renyi entropy and free energy. 2011.

– 164 –
[42] S. Al-Assam, S. R. Clark, and D. Jaksch. The tensor network theory library.
Journal of Statistical Mechanics: Theory and Experiment, 9(9):093102, Sep 2017.
[43] Chase Roberts, Ashley Milsted, Martin Ganahl, Adam Zalcman, Bruce Fontaine,
Yijian Zou, Jack Hidary, Guifre Vidal, and Stefan Leichenauer. TensorNetwork: A
Library for Physics and Machine Learning. arXiv e-prints, page arXiv:1905.01330,
May 2019.
[44] Lucas Dixon, Ross Duncan, and Aleks Kissinger. Open graphs and computational
reasoning. Electronic Proceedings in Theoretical Computer Science, 26:169–180,
June 2010.
[45] Aleks Kissinger, Alex Merry, and Matvey Soloviev. Pattern graph rewrite systems.
Electronic Proceedings in Theoretical Computer Science, 143:54–66, March 2014.
[46] Yves Lafont. Towards an algebraic theory of boolean circuits. Journal of Pure and
Applied Algebra, 184:2003, 2003.
[47] Bob Coecke and Ross Duncan. Interacting quantum observables: categorical
algebra and diagrammatics. New Journal of Physics, 13(4):043016, Apr 2011.
[48] S. Ostlund and S. Rommer. Thermodynamic limit of density matrix
renormalization. Phys. Rev. Lett., 75:3537, 1995.
[49] M. Fannes, B. Nachtergaele, and R. F. Werner. Finitely correlated states on
quantum spin chains. Lett. Math. Phys., 25:249, 1992.
[50] S. R. Clark, J. Prior, M. J. Hartmann, D. Jaksch, and M. B. Plenio. Exact matrix
product solutions in the heisenberg picture of an open quantum spin chain. New
Journal of Physics, 12(2):025005, February 2010.
[51] F. Verstraete, V. Murg, and J. I. Cirac. Matrix product states, projected entangled
pair states, and variational renormalization group methods for quantum spin
systems. Advances in Physics, 57(2):143–224, 2008.
[52] G. Vidal. Entanglement renormalization. Phys. Rev. Lett., 99:220405, 2007.
[53] G. Vidal. Entanglement renormalization: an introduction, 2010. chapter of the
book “Understanding Quantum Phase Transitions,” edited by Lincoln D. Carr
(Taylor & Francis, Boca Raton).
[54] Y.-Y. Shi, L.-M. Duan, and G. Vidal. Classical simulation of quantum many-body
systems with a tree tensor network. Phys. Rev. A, 74(2):022320, Aug 2006.
[55] L. Tagliacozzo, G. Evenbly, and G. Vidal. Simulation of two-dimensional quantum
systems using a tree tensor network that exploits the entropic area law. Phys. Rev.
B, 80(23):235127, December 2009.
[56] Norbert Schuch, Michael M. Wolf, Frank Verstraete, and J. Ignacio Cirac.
Simulation of quantum many-body systems with strings of operators and monte
carlo tensor contractions. Phys. Rev. Lett., 100(4):040501, Jan 2008.
[57] F. Mezzacapo, N. Schuch, M. Boninsegni, and J. I. Cirac. Ground-state properties
of quantum many-body systems: entangled-plaquette states and variational monte
carlo. New Journal of Physics, 11(8):083026, August 2009.
[58] H. J. Changlani, J. M. Kinder, C. J. Umrigar, and G. K.-L. Chan. Approximating
strongly correlated wave functions with correlator product states. Phys. Rev. B,
80(24):245116, December 2009.
[59] Grzegorz Malinowski. Many-valued logics. Clarendon Press: Oxford University
Press, 1993. Series: Oxford logic guides.
[60] Bob Coecke and Aleks Kissinger. The compositional structure of multipartite
quantum entanglement. pages 297–308, 2010.
[61] Ross Duncan and Simon Perdrix. Rewriting measurement-based quantum
computations with generalised flow. In Samson Abramsky, Cyril Gavoille, Claude
– 165 –
 Kirchner, Friedhelm Meyer auf der Heide, and Paul G. Spirakis, editors, Automata,
Languages and Programming, pages 285–296, Berlin, Heidelberg, 2010. Springer
Berlin Heidelberg.
[62] Ross Duncan and Simon Perdrix. Graphs states and the necessity of euler
decomposition. Lecture Notes in Computer Science, 2009.
[63] Ross Duncan and Simon Perdrix. Rewriting measurement-based quantum
computations with generalised flow. In Automata, Languages and Programming,
pages 285–296. Springer Berlin Heidelberg, 2010.
[64] P. Deschamps M. J. Davio and A. Thayse. Discrete and switching functions.
McGraw-Hill Int. Book Co., 1978.
[65] Ingo Wegener. The Complexity of Boolean Functions. John Wiley & Sons, Inc.,
New York, NY, USA, 1987.
[66] M. Cohn. Inconsistent canonical forms of switching functions. IRE Transactions of
Electronic Computers, 1962.
[67] A. Mukhopadhyay and G. Schmitz. Minimization of exclusive-or and
logical-equivalence switching circuits. IEEE Trans. on Computers, 1970.
[68] D. Aharonov. A simple proof that toffoli and hadamard are quantum universal.
2003.
[69] Y. Shi. Both Toffoli and Controlled-NOT need little help to do universal quantum
computation. 2002.
[70] T. Rudolph and L. Grover. A 2-rebit gate universal for quantum computing. 2002.
[71] Joachim Kock. Frobenius algebras and 2-d topological quantum field theories.
Cambridge University Press, 2003.
[72] Bob Coecke, Dusko Pavlovic, and Jamie Vicary. A new description of orthogonal
bases. Mathematical Structures in Computer Science, 23(3):555–567, Nov 2012.
[73] Dusko Pavlovic. Monoidal computer i: Basic computability by string diagrams.
Information and Computation, 226:94 – 116, 2013. Special Issue: Information
Security as a Resource.
[74] A. Carboni and R.F.C. Walters. Cartesian bicategories i. Journal of Pure and
Applied Algebra, 49:11–32, 1987.
[75] M. Aulbach, D. Markham, and M. Murao. The maximally entangled symmetric
state in terms of the geometric measure. New Journal of Physics, 12(7):073025,
July 2010.
[76] J. D. Biamonte. Nonperturbative k -body to two-body commuting conversion
hamiltonians and embedding problem instances into ising spins. Phys. Rev. A,
77(5):052331, May 2008.
[77] Bob Coecke, Bill Edwards, and Robert Spekkens. Phase groups and the origin of
non-locality for qubits. Electronic Notes in Theoretical Computer Science,
270:15–36, 02 2011.
[78] D. Gottesman. The Heisenberg representation of quantum computers, 1998.
[79] David Hilbert. Theory of algebraic invariants. Cambridge University Press, 1993.
[80] Markus Grassl, Martin Rötteler, and Thomas Beth. Computing local invariants of
quantum-bit systems. Physical Review A, 58(3):1833–1839, Sep 1998.
[81] Yuriy Makhlin. Nonlocal properties of two-qubit gates and mixed states, and the
optimization of quantum computations. Quantum Information Processing,
1(4):243–252, 2002.
[82] E.M. Rains. Polynomial invariants of quantum codes. IEEE Transactions on
Information Theory, 46(1):54–59, 2000.
– 166 –
 [83] Peter Oliver. Classical invariant theory. Cambridge University Press, 1999.
[84] Mark S. Williamson, Marie Ericsson, Markus Johansson, Erik Sjöqvist, Anthony
Sudbery, Vlatko Vedral, and William K. Wootters. Geometric local invariants and
pure three-qubit states. Physical Review A, 83(6), Jun 2011.
[85] Karl Kraus. States, effects and operations: fundamental notions of quantum theory.
Springer, 1983.
[86] Ingemar Bengtson and Karol O Życzkowski. Geometry of quantum states: An
introduction to quantum entanglement. Cambridge University Press, 2006.
[87] WF Stinespring. Proc. Amer. Math. Soc. Positive functions on C*-algebras,
6:211–216, 1955.
[88] Yaakov S Weinstein, Timothy F Havel, Joseph Emerson, Nicolas Boulant, Marcos
Saraceno, Seth Lloyd, and David G Cory. Quantum process tomography of the
quantum fourier transform. The Journal of chemical physics, 121(13):6117–6133,
2004.
[89] Hilary A Carteret, Daniel R Terno, and Karol Życzkowski. Dynamics beyond
completely positive maps: Some properties and applications. Physical Review A,
77(4):042113, 2008.
[90] Man-Duen Choi. Completely positive linear maps on complex matrices. Linear
algebra and its applications, 10(3):285–290, 1975.
[91] Andrzej Jamiołkowski. Linear transformations which preserve trace and positive
semidefiniteness of operators. Reports on Mathematical Physics, 3(4):275–278,
1972.
[92] Giacomo Mauro D’Ariano and Paoloplacido Lo Presti. Imprinting complete
information about a quantum channel on its output state. Physical review letters,
91(4):047902, 2003.
[93] Joseph B Altepeter, David Branning, Evan Jeffrey, TC Wei, Paul G Kwiat,
Robert T Thew, Jeremy L O’Brien, Michael A Nielsen, and Andrew G White.
Ancilla-assisted quantum process tomography. Physical Review Letters,
90(19):193601, 2003.
[94] GM D’Ariano and P Lo Presti. Quantum tomography for measuring
experimentally the matrix elements of an arbitrary quantum operation. Physical
review letters, 86(19):4195, 2001.
[95] Michał Horodecki, Paweł Horodecki, and Ryszard Horodecki. General teleportation
channel, singlet fraction, and quasidistillation. Physical Review A, 60(3):1888, 1999.
[96] Michael A Nielsen. A simple formula for the average gate fidelity of a quantum
dynamical operation. Physics Letters A, 303(4):249–252, 2002.
[97] Joseph Emerson, Robert Alicki, and Karol Życzkowski. Scalable noise estimation
with random unitary operators. Journal of Optics B: Quantum and Semiclassical
Optics, 7(10):S347, 2005.
[98] Nathaniel Johnston and David W Kribs. Quantum gate fidelity in terms of choi
matrices. Journal of Physics A: Mathematical and Theoretical, 44(49):495303, 2011.
[99] Easwar Magesan, Robin Blume-Kohout, and Joseph Emerson. Gate fidelity
fluctuations and quantum process invariants. Physical Review A, 84(1):012309,
2011.
[100] Benjamin Schumacher. Sending entanglement through noisy quantum channels.
Physical Review A, 54(4):2614, 1996.
[101] Andrew S Fletcher, Peter W Shor, and Moe Z Win. Optimum quantum error
recovery using semidefinite programming. Physical Review A, 75(1):012338, 2007.

– 167 –
[102] Jacob D. Biamonte, Jason Morton, and Jacob Turner. Tensor Network
Contractions for #SAT. Journal of Statistical Physics, 160(5):1389–1404, Sep 2015.
[103] T. H. Johnson, J. D. Biamonte, S. R. Clark, and D. Jaksch. Solving search
problems by strongly simulating quantum circuits. Scientific Reports, 3:1235, Feb
2013.
[104] Jason Morton and Jacob Biamonte. Undecidability in tensor network states.
Physical Review A Rapid Communications, 86(3), Sep 2012.
[105] Mingji Xia, Peng Zhang, and Wenbo Zhao. Computational complexity of counting
problems on 3-regular planar graphs. Theoretical Computer Science,
384(1):111–125, 2007.
[106] F. Dolde et al. High-fidelity spin entanglement using optimal control. Nature
Communications, 5:3371, February 2014.
[107] K. Rosen. Discrete mathematics and its applications. McGraw-Hill, 1999.

– 168 –