Signature Redacted Certified By

Home , Gilles Brassard, PBR theorem

Quantum Correlations, Quantum Resource Theories and Exclusion Game ARCHNES MASSACHUSETTS INSTITUTE by OF TECHNOLOLGY Zi-Wen Liu JUL 3 0 2015 Submitted to the Department of Mechanical Engineering LIBRARIES in partial fulfillment of the requirements for the degree of

Master of Science in Mechanical Engineering

at the

MASSACHUSETTS INSTITUTE OF TECHNOLOGY

Author ..... Signature ...... Department of Mechanical Engineering May 8, 2015 Z- / 7/ Signature redacted Certified by...... 0...... Seth Lloyd Professor of Mechanical Engineering Thesis Supervisor

Accepted by. Signature redacted ...... David E. Hardt Ralph E. and Eloise F. Cross Professor of Mechanical Engineering Graduate Officer, Department of Mechanical Engineering 2 Quantum Correlations, Quantum Resource Theories and Exclusion Game by Zi-Wen Liu

Submitted to the Department of Mechanical Engineering on May 8, 2015, in partial fulfillment of the requirements for the degree of Master of Science in Mechanical Engineering

Abstract

This thesis addresses two topics in quantum information theory. The first topic is quantum correlations and quantum resource theory. The second is quantum communication theory. The first part summarizes an ongoing work about quantum correlations beyond entanglement and quantum resource theories. We systematically explain the concept quantum correlations beyond entanglement, and introduce a unified framework of measuring such correlations with entropic quantities. In particular, a new measure called Diagonal Discord (DD), which is simpler to compute than discord but still possesses several nice properties, is proposed. As an application to real physical scenarios, we study the scaling behaviors of quantum correlations in spin lattices with these measures. On its own, however, the theory of quantum correlations is not yet a satisfactory quantum resource theory. Some partial results towards this goal are introduced. Furthermore, a unified abstract structure of general quantum resource theories and its duality is formalized. The second part shows that there exist (one-way) communication tasks with an infinite gap between quantum communication complexity and quantum information complexity. We consider the exclusion game, recently introduced by Perry, Jain and Oppenheim [80], which exhibits the property that for appropriately chosen parameters of the game, there exists an winning quantum strategy that reveals vanishingly small amount of information as the size of the problem n increases, i.e., the quantum (internal) information cost vanishes in the large n limit. For those parameters, we prove the quantum communication cost (the size of quantum communication to succeed) is lower bounded by Q (log n), thereby proving an infinite gap between quantum information and communication costs. This infinite gap is further shown to be robust against sufficiently small error. Some other interesting features of the exclusion game are also discovered as byproducts.

Thesis Supervisor: Seth Lloyd

3 Title: Professor of Mechanical Engineering

4 Acknowledgments

First of all, I would like to thank my academic and research advisor, Professor Seth

Lloyd, for leading me into the magnificent world of quantum information that I knew little of prior to coming to MIT, and for all his generous support and insightful guidance over the last two years. I can never know how blessed I am to have such an opportunity.

I would like to thank Professors Scott Aaronson, Harry Asada, Gang Chen, Isaac

Chuang, Eddie Farhi, Aram Harrow, Mehran Kardar, Seth Lloyd, Hong Liu, Peter

Shor, Wati Taylor, Salil Vadhan, Evelyn Wang, Xiao-Gang Wen and many more, from whom I learned priceless knowledge through courses and/or discussions.

I would like to thank my fellow graduate students and close collaborators, especially Can Gokler, Dax Koh, Kevin Thompson, Elton Zhu and Quntao Zhuang, for all the happy time we spent together talking about everything from quarks to universe(s).

I would also like to thank all my friends, for sharing my happiness and sorrow, and for allowing me to do the same for them.

Thanks to all the people above, for making me believe I am doing the right thing, at the right place.

And at last, I would like to thank my parents and my girlfriend Xiaoyu for en- couraging and trusting me, unswervingly. Miraculously, I feel that you are all right here with me, when I write down these words.

5 6 Contents

I Quantum correlations and resource theories 14

1 Introduction 15

2 Quantum correlations beyond entanglement 19

2.1 Purely classical correlations ...... 19

2.1.1 Classically correlated states ...... 20

2.1.2 Creating nonclassical correlations ...... 22

2.2 Entropic measures ...... 25 2.2.1 Candidates ...... 26 2.2.2 Hierarchy ...... 34

2.2.3 Multipartite generalization ...... 39

2.3 Scaling behaviors in spin lattices ...... 42

2.3.1 Quantum correlation between two spins ...... 43 2.3.2 Example: 1D Heisenberg XXZ chain ...... 46 2.3.3 Area laws ...... 48 2.3.4 O utlook ...... 52 2.4 Diagonal Discord (DD) ...... 52 2.4.1 M otivation ...... 52

2.4.2 Properties ...... 54

2.4.3 Physical interpretation ...... 58

3 Quantum Resource Theories (QRTs) 61 3.1 Unified framework ...... 62

7 3.1.1 Elements ...... 62

3.1.2 Perfect QRT ...... 68

3.1.3 Hierarchical structure ...... 72

3.1.4 Combining QRTs ...... 73

3.2 Quantum correlations as a resource ...... 76

3.2.1 Free states ...... 78

3.2.2 So ...... 79

3.2.3 Promotion to S ...... 83

3.2.4 Map zoo ...... 85

3.2.5 Remarks ...... 87

3.3 Dual QRTs ...... 87

3.3.1 General structure ...... 88

3.3.2 Examples ...... 91

4 Summary and outlook 95

II Exclusion game 97

5 Introduction 99

6 Preliminaries 103

6.1 General formulation of communication tasks ...... 103

6.1.1 Mathematical structure ...... 103

6.1.2 Exclusion game ...... 104

6.2 Information and communication ...... 105

6.3 Classical communication complexity ...... 106

6.4 PJO strategy ...... 106

6.4.1 Protocol ...... 107

6.4.2 Quantum information cost ...... 108

8 7 Quantum communication complexity 109 7.1 Zero error ...... 109

7.1.1 Classical encodings of quantum states ...... 110 7.1.2 Lower bound of Qcc ...... 111 7.1.3 G aps ...... 113

7.2 Robustness against error ...... 114 7.2.1 Lem m as ...... 114

7.2.2 Quantum-classical separation of communication ...... 118

7.2.3 Maximum error ...... 119

8 Concluding remarks 121

Appendix 123

A QD and DD of real X states 123 A.1 Preparations ...... 123

A.2 Pairwise quantum correlation ...... 125 A.2.1 Optimal basis ...... 125

A.2.2 Explicit calculation of pairwise QD ...... 126 A.2.3 Pairwise DD ...... 130

9 10 List of Figures

2-1 (Adapted from [104]) Area law. Intuitively speaking, under locally

interacting Hamiltonians, the correlation length in noncritical phases

is finite so that sites in A and B that are separated by a distance further than the correlation length (the shaded stripe) should not contribute

to the mutual information or correlation measures between A and B, hence bounded by the number of sites at the boundary and therefore

scales as the boundary area...... 49

3-1 Intuitive illustration of the basic content and structure of a QRT. .. 63

3-2 A hierarchical structure of maps. Columns represent correspondences, and rows represent strict hierarchies...... 73

3-3 A sketch of a strategy for determining a qubit state that queries the QRTs of coherence and purity. The dashed circle represents the states that has the same entropy (connected by unitary transformations), and

the solid line represents the states that are diagonal in the appointed

basis (incoherent states). For an arbitrary state p, the unitary U brings

it to one of the incoherent states, while preserving purity/entropy. .. 76

3-4 Geometrical (Bloch sphere) demonstration of the effect of a unital map

on a qubit. The unital channel keeps the two basis vectors symmetric

with respect to the center of the Bloch sphere, and the output state

can be diagonalized in the orthonormal basis corresponding to the in-

tersections of the connecting line and the surface of the Bloch sphere. 82

11 3-5 (Adapted from [91]) The detailed hierarchy in between mixture of unitaries and unital maps. AQBP denotes "Asymptotic Quantum Birkhoff

P roperty"...... 86

3-6 Illustration of the dual QRT. The parts with grey fill means "free".

Dashed lines represent the pair of elements in each theory, and arrow

represent partial orders, determined by quantifiers. Note that this is

only a sketch of the idea of duality. The geometries of sets and the

partial orders in each space do not necessarily resemble this figure. . . 91

7-1 Exclusion game EXC,.. where m E (n), 1/2 < a < I in the large n

limit. Solid arrows indicate established separations (pointing towards

the smaller one), while the dashed one indicates an unknown separation

(at most exponential)...... 114

12 List of Tables

2.1 An LOCC state preparation protocol that generates quantum correla-

tions. According to the outcomes of flipping their private coins, Alice

and Bob respectively apply the local operation indicated in the table. The rightmost column shows the corresponding output states. .... 23

3.1 Level 1 descriptions of some typical QRTs. *Since NO is maximal, this

quantifier is unique. **With respect to a particular basis...... 68

13 Part I

Quantum correlations and resource theories

14 Chapter 1

Introduction

Since its origination, quantum physics has been bewildering people in all sorts of

aspects. As famously said by Niels Bohr, "If quantum mechanics hasn't profoundly

shocked you, you haven't understood it yet." Even to this day, more than 100 years af-

ter its debut, people are still trying to understand more fully how quantum mechanics

works, and how to take advantage of it.

First recognized by Einstein, Podolsky and Rosen (EPR) in 1935 [35], entangle-

ment, which Einstein termed, the "spooky action at a distance", lies at the heart of

quantum weirdness. In essence, this concept captures a peculiar form of nonlocal

correlation that has no classical counterpart - in principle, the output of measuring

the spin of an electron on the earth (which can be completely random) will directly

tell you what your friend will see if he (even if immediately afterwards) measures

another electron on Mars, or arbitrarily further away, as long as the two electrons are

entangled.

Due to its apparent weirdness, entanglement is one of the most investigated con-

cepts in quantum information theory. One may naturally ask: is there any other form

of correlations besides classical correlations and entanglement? The answer turns out

to be no. In mixed states, there exist correlations that are not entanglement, yet have

no classical counterparts as well. Ollivier and Zurek first proposed the quantity Quan-

tum Discord (QD) to quantify this type of correlations in their seminal 2001 work 177]. and started a boom of the studies along this direction for more than 10 years.

15 Since then, many papers on the topic of quantum correlations beyond entanglement have appeared. Despite considerable efforts, however, no unified and universally accepted method for understanding quantum correlations has emerged. Two messages can be taken home from this situation:

1. Quantum correlations are difficult to measure.

2. We probably haven't found the right way to understand quantum correlations.

These are exactly the problems that this part of the thesis is trying to tackle.

The first message can be interpreted from two perspectives. First, there are many possible measures of quantum correlations to choose between. To simplify this choice, I shall introduce a unified framework of entropic measures, and systematically discuss the motivation for defining them. Second, most measures are difficult to compute.

For example, the calculation of the earliest and best-known measure, QD, is shown to be NP-complete. Indeed, such convex optimization problems are always hard. In a word, the field is a mess. To address the issue of computational complexity, a new measure which is originated very naturally from both the mathematical and physical perspectives, the Diagonal Discord (DD), is proposed. This measure is significantly easier to compute and study. More importantly, diagonal discord can still serve its job as a quantifier of quantum correlations beyond entanglement pretty well, as supported by its various features.

The second message is closely related to the first one. To unify our understanding of quantum correlations, we look at Quantum Resource Theories (QRTs). The QRT of some quantum property x is essentially the theory of x as a physical resource. A resource theory should be able to reveal multiple aspects of the quantity x, including how to create it, how to quantify it, how to convert it, how to make use of it and so on. To gain better understanding of a quantum object, establish the resource theory of it. On its own, the theory of quantum correlations is far from a satisfactory QRT, since the set of operations that generates/does not generate quantum correlations is not fully characterized. One of the primary targets of this thesis work is to construct a theory of quantum correlations that allows the natural construction of quantum

16 resource theories. Furthermore, different QRTs frequently exhibit very similar mathematical structure, which leads to a abstract framework that unifies many QRTs. I shall discuss the details of this framework in this part, and point out an interesting dual of it. Note that these two topics are seemingly quite self-contained, but as you will see from the content, they are closely related everywhere.

This part is organized as follows. In Chapter 2, I systematically discuss the interpretation and creation of purely classical correlations, and motivate the concept of quantum correlations beyond entanglement as the complement. Several candidate entropic measures are defined in a consistent way. Their relationships and generalizations are then studied. The proposed measures are then used to study the scaling behaviors of quantum correlations in spin models. Specifically, the new measure DD is studied in the last section. Surprisingly, quantum correlations beyond entanglement turn out to be the necessary condition for heat transfer, and DD measures its rate in the infinitesimal limit. In Chapter 3, I present an abstract mathematical framework that unifies QRTs, and discuss several interesting products. Using this framework, partial results towards a full QRT of quantum correlations are introduced. In the final section, a dual framework of QRT is established, with examples given.

17 18 Chapter 2

Quantum correlations beyond entanglement

As already mentioned in Chapter 1, entanglement is not the only form of correlation present in quantum systems that cannot be described classically - there is a gap between purely classical correlations and entanglement. In this chapter, I systematically explain the concept quantum correlations beyond entanglement, and introduce a unified framework of measuring such correlations with entropic quantities. In particular, a new measure called Diagonal Discord (DD), which is simpler to compute than discord but still possesses several nice properties, is proposed. As an application to real physical scenarios, we study the scaling behaviors of quantum correlations in spin lattices with these measures.

2.1 Purely classical correlations

To understand quantum correlations, one must first understand classical correlations.

Two questions arise:

1. What kinds of correlations are purely classical?

2. What quantum states are classically correlated?

19 In this Section, I shall first answer the above questions, and then discuss ways to prepare quantum states that are separable (unentangled), but that possess quantum correlations, thereby illustrating the gap between purely classical correlations and entanglement. Note that the latter part (creation of quantum correlations) is also closely related to the resource theory of quantum correlations, which will be discussed in Section 3.2.

2.1.1 Classically correlated states

For the purpose of explaining the basic ideas, it is sufficient to consider bipartite correlations for the moment. According to classical information theory, the state of a classical system consisting of two parties respectively represented by random variables A and B can be described by a joint probability distribution {Pab}, while the respective marginal probability distributions are denoted as {p"} and {Pb}. Consider the case where the possible values a and b are assumed to be discrete. To determine whether these two probability distributions are correlated or not from these probability distributions, one can apply Shannon's theory to calculate the mutual information

I(A : B) = H(A) + H(B) - H(AB), (2.1) where H(-) denotes the Shannon entropy of a probability distribution, e.g., H(AB) =

- Ea bPab log Pab. Nonzero I(A; B) means that A and B are not independent of each other, and the value of I(A : B) quantifies the amount of information they share in common.

Then comes the second question: Which quantum states are completely classically correlated? Or more specifically, how does one write down a quantum state that registers classical correlation as introduced in the above paragraph? (Note that such a quantum state lives in the joint Hilbert space 'RAB = RAO RB.) The answer is that this quantum state represents the probabilistic mixture (note: different from coherent

superposition) of a set of quantum states according to {Pab}, where any pair of states

chosen from this set are perfectly distinguishable,since all events (labeled by ab) are by

20 definition distinguishable in the classical setting. As is well known, the necessary and sufficient condition for two quantum states being reliably distinguishable is that they are orthogonal [761. Therefore, the set of perfectly distinguishable states spans some

Hilbert space ?AB whose dimension is the number of possible ab's, i.e., forms the orthonormal basis of NAB. Denoting the corresponding elements of this orthonormal set as lab), a completely classically correlated state actually represents the statistical ensemble {Pab, lab)}.

Definition 1 (Completely classically correlated state; CC state). A quantum state is classically correlated, iff it takes the form

PAB = ZPaba)A(a 0 lb)B(b= Z Pablab)(abl, (2.2) ab ab where EabPab = 1, (abla'b') = S5aalbb'. For reasons that will become clear later, members of this class of states are also called classical-classical (CC) states [781.

Furthermore, one may also argue that if all possible local measurements on one subsystem disturb the joint quantum state, that means its correlation with the other subsystem exhibits nonclassical nature. In other words, if one subsystem is classically correlated with the other, then there should exist a local measurement that does not alter the joint state. The only possible local measurement basis is the Schmidt basisi

(not necessarily unique), since other choices do not preserve the local state (nonzero off-diagonal entries will be erased). This intuition not only provides another nice way of understanding of the "quantumness" of correlations, but also naturally leads to the definition of "one-way" classically correlated states by restricting the local measurement to one side (say, A):

Definition 2 (One-way classically correlated state; CQ state). A bipartite quantum state is called a classical-quantum (CQ) state, iff there exists a local measurement on A that does not disturb the joint state. Equivalently, it can be written in the

'Schmidt basis is defined as the basis with respect to which the local density matrix is diagonal (local spectral basis).

21 following form

pA = Lpala)A(al &Pa (2.3) a where Ia) is the Schmidt (local spectral) basis for A (block diagonal in A's eigenbasis), and p' is some valid density matrix for B.

As can be directly seen, the name "CQ" comes from the form of the state: it looks classical on A's side and quantum on B's. Likewise, we can define QC states by placing the local measurement criterion on B. And if we require the joint state to be undisturbed for local measurements on both sides, the definition of CC states is easily recovered.

Note that CC is a stronger constraint than CQ or QC: CC states definitely qualify as both CQ and QC states. I shall call the union of these three sets "-y states" from now. Lastly, I introduce two important properties of the set of 7 states:

1. It has Lebesgue measure zero and is nowhere dense in the parameter space of

all quantum states [38], i.e., topologically negligible. That is, almost every state

possesses quantum correlations;

2. It is not convex: mixtures of -y states may possess quantum correlations.

I shall leave the explanation of the second property to Section 2.1.2. These two features will show great importance in later discussions.

2.1.2 Creating nonclassical correlations

Equipped by the in-depth understanding of the classical-type correlations, we shall

now explore the interesting features of those correlations that are not classical, i.e., quantum. A nice first step is to think about what kind of operations are able to

generate quantum correlations between two systems.

LOCC

As has been extensively studied, a key feature of entanglement is that it cannot be

created by Local Operations and Classical Communication (LOCC) [28, 50]. (The

22 Coin flip LO Output Alice Bob Alice Bob Alice Bob + + I H 10) +) + - X HX 11) 1-) - + HX I --) 10) - - H X 1+) 11)

Table 2.1: An LOCC state preparation protocol that generates quantum correlations. According to the outcomes of flipping their private coins, Alice and Bob respectively apply the local operation indicated in the table. The rightmost column shows the corresponding output states. study of LOCC is actually the foundation of the resource theory of entanglement, which will be further discussed in Chapter 3.) That is, one can never prepare an entangled state by operations within the scope of LOCC.

Now I present a state preparation protocol that creates quantum correlations with

the help of LOCC only. As usual, we name the two parties Alice and Bob respectively

with a classical communication channel in between, both having an unbiased private

coin, and |0) as the initial state: the initial bipartite state is simply the direct product

of two local states, i.e., I0)A 0 I0)B (completely uncorrelated). They respectively

flip their private coins, and then communicate the (independent) outcomes via the

classical channel (say, they send each other "+" or "-", respectively represents heads or

tails, via text message). They then follow the instruction in Table. 2.1 to apply the

corresponding local unitary their initial states, e.g., if Alice gets heads and Bob gets

tails, Alice acts Pauli X on her state and Bob acts Pauli X followed by Hadamard on

his state. The final bipartite state will be

1 PAB = -(IO)A(0 ® +)B(+ + 1)A(1I 0 I-)B(- 4 +|-)A(- 0 I)B(0| + +)A(+I 0 |1)B(1I) 1 = 4-(I0+)(0 + I + 1-)(1 - I + I - O)(-0 + I + 1)(+1). (2.4)

This state is definitely not entangled because the initial state is an uncorrelated

product state (possess no entanglement), and the preparation procedure only in-

volves LOCC. However, it is impossible use the forms of -y states to write PAB. More

23 elaborately, I will show in the Section 2.2 that the measures of quantum correlations, which may only vanish on 'y states, are positive for PAB. Generally, one can always follow the similar procedure to prepare the following state

Sep PAR = piUi|0)(0|U 0 |0) B0O~

= piluivi)(Uivil, (2.5) where Ui and V are unitaries acting on A and B corresponding to the classical flag i, and I denote Ui|O) = Jui), V10) = jvi). Or equivalently,

pR = ipp, (2.6) where E pi = 1, pi and pi are arbitrary local states. All states that can be written in the form of Eqs. (2.5) or (2.6) are separable/unentangled. Or in other words, they are just probabilistic mixtures of product (uncorrelated) states. Note that the problem of determining separability is computationally hard in general [57]. One can also use the Perez-Horodecki Positive Partial Transpose (PPT) criterion [791 to test the separability of this state, though it is not always a sufficient condition.

Throwing away information

Alternatively, PAB can be viewed as the uniform probabilistic mixture of four product states, which exemplifies the nonconvexity of the set of y states, as mentioned in

Section 2.1.1. Nonconvexity also leads to a way to create quantum correlations by simply throwing away information: Suppose one has two y states in hand, and he flips a coin to determine which one to choose. By simply forgetting the outcome of the

(public) coin flip, one obtains the state which is just a mixture of these two y states, and by nonconvexity it can be and frequently is quantumly correlated. As will be discussed in Chapter 3, such a scenario is not good for a reasonable quantum resource theory since one can seemingly create quantum resources from classical correlations

24 arbitrarily - Further assumptions need to be made.

Discussions

The above example indicates that there exist unentangled states that possess quantum correlations, which live in the gap between purely classical correlations and entanglement. The notion of entanglement is not sufficient to capture all correlations that have no classical counterpart. In fact, people have found that unentangled states in many cases may exhibit nonclassical physical behaviors [731. Comparing the forms of classically correlated states Eq. (2.2) and unentangled states Eqs. (2.5) or (2.6), it is easy to tell that the restriction on the latter class is much weaker. In fact, the connection between separable states and -y states is very similar to the purification of mixed states. It is shown in [61] that a bipartite state PAB is separable iff there exists a CC state PAABB, (partition: AA'/BB') on an extended Hilbert space where

PAB ~ trA'B'PAA'BB'. That is, the "separable-CC" relation can be thought of as the bipartite analogy of the "mixed-pure" relation for quantum states.

It is worth mentioning that in contrast to -y states, the set of separable states exhibits the following features:

1. It possesses a finite volume in the set of all quantum states 155], i.e., has nonzero

measure;

2. It is convex: the mixtures of separable states are still separable.

It has been discussed that these two properties are not satisfied by -y states.

2.2 Entropic measures

In previous sections, I have discussed about the boundaries in the classical-quantum-

entanglement hierarchy of correlations in quantum states, in order to motivate readers

to pay attention to quantum correlations beyond entanglement. However, this char-

acterization of quantum correlations is far from the purpose of fully understanding

and utilizing such correlations as a resource for quantum information processing tasks

25 (which is the motivation and mission for Quantum Resource Theories (QRT), as will be discussed in detail in Chapter 3). In general, a good resource theory should both quantify the amount of the resource associated with a quantum state, and also provide a partial order of quantum states.

Originating from the perspectives of thermodynamics (physical) and information theory (mathematical) quite independently, the concept of entropy, which quantifies the uncertainty of information content, has been playing a central role in all kinds of scientific studies. As Landauer's principle and Szilard's engine has taught us, these two fields are closely bond together - information is physical. Indeed, correlations between physical systems can be understood as shared information: entropy is the most natural and powerful tool to study correlations. In this section, I shall introduce and analyze some candidate entropic quantifiers for quantum correlations in a unified manner. Note that some measures other than entropic quantites can be suitable for certain scenarios, including the geometric measure [311 etc. 2 , but I shall focus on those measures that are defined in terms of entropic quantities here.

2.2.1 Candidates

Quantum Discord (QD)

As discussed earlier, quantum correlation is simply the part of total correlation that is not classical. This intuition leads to the original and most famous quantification of quantum correlations, namely, Quantum Discord (QD), which was originally proposed by Ollivier and Zurek in 2001 [77]. By definition, QD is the discrepancy between total mutual information and the classical part, interpreted as the maximal amount of mutual information that can be accessed locally (by measurements). I shall start by introducing the underlying ideas of this quantity, and then proceed to other reasonable measures. Note that without loss of generality, I assume that A is the party that does local measurements when talking about one-way scenarios.

As introduced in Section 2.1.1, the amount of correlation between classical random

2 R.efer to 1731 for a review.

26 variables A and B is captured by the mutual information I(A: B) = H(A) + H(B) -

H(AB) where H(X) = - E_ plogp, denotes the Shannon entropy, where X is a classical variable with values x occuring with probability px. On the other hand, Bayes' rule allows us to define an equivalent form for the classical mutual information

as I'(A: B) = H(B) - H(BIA) = Ii(A: B) with the conditional entropy H(BIA) =

EZpaH(Bla), which can be understood as the information of B that can be obtained by gaining knowledge on A. Note that for classical systems the mutual information

yielded by these two definitions are exactly the same.

Now we "quantize" the above notions by promoting random variables A and B to

quantum systems, and Shannon entropies to von Neumann entropies. Furthermore, the conditioning on A is now realized by making measurements on the local subspace.

Then one may immediately notice that the entropic quantities will depend on the

measurement, which is indeed the case, and we should treat all knowledge of the total

correlation that can be accessed locally as classical, leading to the optimization over

all POVMs or von Neumann measurements. It can be shown that the optimal POVM

is always rank-1 and extremal [71, 431, but not necessarily orthogonal projective (by

Davies' theorem, the optimal POVM may have up to d2 elements, where d is the

dimension of the local Hilbert space), indicating that optimizing over von Neumann

measurements is not always optimal. However for certain purposes, restricting to von

Neumann measurements is good enough (indeed, simpler at least), so I shall leave it

here as a possibility. Gathering all the above considerations together, we define the

amount of classical correlation as

CA(A : B) = S(B) - min Z pkS(Bk), (2.7)

where {Hek} denotes a set of measurement operators, and PB,k = trA [(lk0IA)pAB]/pk is the reduced state of subsystem S corresponding to outcome k. On the other hand

the total mutual information is simply given by

I(A: B) = S(A) + S(B) - S(AB), (2.8)

27 where S(X) = -px log(px) denotes the von Neumann entropy of a quantum state px. Therefore we have the following definition of QD:

Definition 3 (Quantum Discord (QD)). QD is defined as the minimum amount of mutual information that cannot be accessed via local measurements on A, or in other words, the discrepancy between total mutual information and the maximum amount of mutual information that can be accessed via local measurements on A:

D(A -+ B) I(A: B) - CA(A: B)

= S(A) + S(B) - S(AB) - S(B) + min I:pkS(Bk) {nI} k = min E pkS(Bk) + S(A) - S(AB) {nk} k min S(BIA) - S(BIA), (2.9) {nI} where the classical correlation CA(A : B) can also be interpreted as the maximum amount of correlation in the postmeasurement state.

I emphasize again that due to Bayes' rule, D(A -+ B) always vanishes for classical probability distributions since the classical analogs of the two conditional entropies in the last line are exactly equivalent, but it can be positive for some separable states

(that are not CQ). Also note that QD is not necessarily a symmetric measure as its value depends on the party chosen to carry out the measurement: it is not an ordinary distance measure, or say, a metric.

Measurement and entropy

Now let's consider a general question: Without postselection, how do measurements affect the entropy of a quantum system? For von Neumann measurements, it is well known that the entropy of the state being measured cannot decrease, and it remains unchanged when the spectral basis (in which the density matrix is diagonal) is chosen as the measurement basis since the state itself is unchanged (Theorem 11.9 in (761).

An equivalent interpretation is that von Neumann measurements are closed operations

(uncoupled to the outside), which should never decrease the entropy of the system by

28 the second law of thermodynamics. However, POVMs (generalized measurements), which can always be thought of as reductions of projective measurements on a larger

Hilbert space (Naimark's dilation theorem). indicating that they are essentially open system operations, which may cause information to flow from the bath to the system, thus decreasing the entropy. What if one does von Neumann measurements on a subspace? The answer is given by the following statement3 :

Theorem 1. Consider an arbitraryquantum state p living in the d-dimensional space

Rd. Projective measurements on a d-dimensional subspace (di <; d, Rd, 9 Rd,

1= I-td\l-d,, when d = dl: empty) cannot decrease the entropy of the state (the postmeasurement state reads p' = Z(HUi 0 In )p(Ui 0 I )), i.e., S(p') > S(p), with equality if p can be written as a CQ state with the part in Rd, being classical, and {H} being the Schmidt basis.

Proof. By definition, Ei H = IHd1 , thus ZE Hs 0 IjJ = 'Rd; and (li 0 IR-)2 di d Hi 0 Ih . Therefore d1

S(p') = -tr(p log p')

= -tr [Z(i 0 Is )p(He 0 In ) logp'

= -tr Z(fi0 Isi))PlogpI(Hi0I.t)

-tr Z(i0M IR)plogp']

-tr(p log p'), (2.10) where the third equality follows from the fact that (fli 0 IH ) commutes with p', and thus also its logarithm. Then by definition of the relative entropy distance:

S(p') - S(p) = S(pjIp') > 0, (2.11) with equality iff p = p', by Klein's inequality. For d, = d case, the statement simply

3This fact is known. Variants of the same result can be found in literatures, e.g., 1631.

29 reduces to Theorem 11.9 in [76] mentioned earlier. When d, < d, there exists local basis that doesn't disturb the state (Schmidt basis) iff it's CQ (classical over 'Nd), and p' is the same CQ state.

Indeed, since dephasing (measurement) only happens inside the system, this kind of operations can also be considered closed, which never decrease entropy.

Note that whenever p' $ p, S(p') > S(p), which directly indicates that for non-CQ states (i.e., states that possess positive one-way quantum correlation), min{rl-L} [S(p')-

S(p)] > 0. Treating '-t as the space of the bath, this minimum joint entropy production appears to be a very natural measure of the quantum part of the coupling strength, i.e., the one-way quantum correlation.

Definition 4 (Minimum Joint Entropy Production (minJEP)). As implied by the name, minJEP is defined as the minimum amount of entropy generation over all possible local projective measurements on the appointed subsystem A:

M(AB) = min S(AB) - S(AB), (2.12) {InA} where S(AB) and S(AB) correspond to S(p') and S(p) in Theorem 1 respectively.

Note that the local measurement is equivalent to local dephasing with respect to the measurement basis, so this quantity can also be interpreted as the minimum amount of information lost due to the constraint of classical communication. For pure states, QD trivially reduces to the entanglement entropy S(A) = S(B).

This measure actually coincides with several quantities that has been defined in different contexts, e.g., thermodynamics [109]. And by Eq. (2.11), it is easy to argue that minJEP is also exactly equivalent to the minimum relative entropy distance to CQ states, namely "relative entropy of discord", which has also been proposed independently as a measure [48].

Combined with the above conclusions, for a fixed quantum state, the lower bound of entropy change upon measurements seems to increase as the "effective" space being

30 measured.4 Independent of the current purpose of analyzing quantum correlations, it might be interesting to study the details of this observation, e.g., bounds depending on dimensions. (A trivial case is the maximally mixed state, whose entropy cannot be changed by measurements.) As a side remark, the maximal set of operations that never decreases the entropy of a quantum system is called Noisy Operations (NO), which will be introduced in Chapter 3 as the core element of the resource theory of purity.

In the above I only considered the entropy production of the whole system upon local projective measurements, which may also generates entropy locally. Taking this fact into consideration, we arrive at another important observation:

Theorem 2. Consider projective measurements on a subsystem (which do not decrease entropy both jointly and locally): the minimum difference between local and joint entropy production reduces to QD.

Proof. As introduced in Definition 3, QD can be understood as the minimum discrepancy between pre- and postmeasurement mutual information of the two parties, thus can be rewritten in the following form:

D(A -+ B) = I(A : B) - max (A : B) {nt} = S(A) + S(B) - S(AB) - max[S(A) - S(B) + S(AB)] {ni}

= min [S(AB) - S(AB)] - [S(A) - S(A)]}, (2.13) where the two brackets are respectively the joint and local entropy production. El

Combining the above theorem with the nonnegativity of QD [73], we directly obtain the following

Corollary. Upon local projective measurements, the joint and local entropy production

are both zero if the state is CQ. Otherwise, the former is always greater than the

latter.

4 For local projective measurements, it simply refers to 7d,, while for POVMs, it means the extended Hilbert space by Naimark's dilation.

31 Diagonal Discord (DD): a natural simplification

QD and minJEP both involve the optimization over all possible measurements, which is computationally very difficult in general, i.e., unfavorable for practical use, especially when dealing with large systems. Is there a way to define a reasonable measure that is easier to calculate? An immediate idea is to see if we can put restrictions on the measurement. Obviously, the most natural and promising choice is the Schmidt basis measurement due to its very special feature: does not disturb the reduced state, which leads to the definition of Diagonal Discord (DD):

Definition 5 (Diagonal Discord (DD)). Instead of carrying out the minimization over all local measurement bases for computing QD, we directly apply local measurement in the Schmidt basis, and name the resulting discord value as Diagonal Discord (DD). By Eq. (2.13), it takes the form

DD(A -+ B) = S(AB) - S(AB), (2.14) where the measurement is made in Schmidt basis, which does not perturb the reduced state, thereby generating no entropy locally - It is completely equivalent to substitute the optimization over measurements in the definition of minJEP with Schmidt basis.

Note that the diagonal basis is unique iff the decomposition is non-degenerate.

The basis choice within the degenerate subspace doesn't matter, e.g. for maximally mixed state all bases are equivalent. As a specific example, Now recall the separable bipartite state in Eq. (2.4). One can directly observe that the reduced state of either party (the marginals) are maximally mixed, implying that all local measurement bases are equivalent: QD reduces to DD. Here I calculate the exact value for illustration:

S(Ao) = S(A 1 ) = -1 log 1 - log ~ 0.81 bits, and therefore D(A -+ B) = D(B -+

A) = DD(A -+ B) = DD(B -+ A) = log i 0.31 bits, which supports our earlier claim that the correlation exhibited in this state is not purely classical, i.e., quantum correlations can be created by LOCC.

I shall discuss this measure in more detail in Section 2.4.

32 Criteria for candidacy

In the above, several entropic measures of quantum correlations are defined. Generally speaking, a reasonable one-way measure must satisfy the following basic criteria:

1. It vanishes for, and only for CQ states; otherwise it's positive;

2. It is invariant under local unitaries;

All measures defined earlier do have these properties. There are some other features that one may expect a good measure should possess, especially for certain purposes. But these are the nonnegotiable criteria that any acceptable quantification for quantum correlations should satisfy. I shall explore some interesting common properties in Section 2.4.

Two-way generalization

By Theorem 2, the one-way quantifiers of quantum correlations can be easily generalized to two-way, or even multipartite cases (Section 2.2.3). Specifically for bipartite QD, the two-way version can be defined as follows:

Definition 6 (Two-way bipartite QD). The two-way QD can be defined as the minimum difference between the joint entropy production and the sum of local entropy

productions, when local measurement are made by both parties, i.e.,

D(A ++ B) = min [S(Ab) - S(AB)] - [S(A) - S(A)] - [S(B) - S(B)]I, (2.15) where {Ik}, {III} denote local projective measurements on A and B respectively. The same quantity is known as WPM discord [1061.

Two-way generalizations of minJEP and DD can be similarly defined. Since local entropy productions are not involved in these two measures: the only difference is that local measurements are made on both sides.

33 2.2.2 Hierarchy

Now I analyze the magnitudes of these different entropic quantifiers, and compare them to the entropic measures of entanglement and total correlations.

Measures for quantum correlations

In the previous subsection, I defined QD, minJEP and DD as candidate measures for quantum correlations beyond entanglement. It is easy to show that for a certain state, QD can never be larger than minJEP: Suppose there exists a state p such that

QD is larger than minJEP. This directly indicates that the local entropy production is negative upon projective measurements, which contradicts Theorem 11.9 in [76j

Then one can immediately tell that DD is always larger than or equal to the other two since it is essentially JEP without minimization. Summarizing, we have

QD < minJEP < DD. (2.16)

As discussed earlier, this hierarchy collapses at least for states with maximally mixed marginals, and pure states [60]. The maximal set of states that takes the equalities is yet to be specified, and is important especially for studying DD.

However, despite that these entropic measures strictly obey the above order, it can be shown that they determine different partial orders of states with positive quantum correlations:

Theorem 3. For any pair of different entropic measures of quantum correlations 61

and 52, there exist two different states p and o- such that S1(p) > 62 (-), but 62(P) <

61(a), i.e., are ordered differently under these two measures.

Proof. The proof goes similarly as the proof of different state orderings under "good asymptotic entanglement measures" [102]. It is known that for any pure state IV),

61(10)) = 62(10)) =-(-4)) (property of entropic measures of quantum correlations). Assume that 61 and 62 place the same ordering on all states. Since 6(1,0)) obviously covers the whole range, one can always find two pure states 1@) and 10) such that

34 6 61 (1V))) = Si(p) + e and 61(|#)) i(p) -- e for any mixed state p (e > 0). Therefore

By the equivalence of 61 and 62 on pure states,

6 6i()) 2(P) > 61(I#)), (2.18) i.e.,

61(P) + f >_ 62(P) > 61(P) - .(2.19)

Taking e to zero, we see that 6 i(p) = 62(p) for all p: they are the same thing. For them to be different measures, the partial orders have to be different. 0

This theorem indicates that any entropic measure is not a monotone of another.

From the perspective of resource theory, these measures serve to quantify the use-

fulness of a state for some certain tasks. In this sense, they are not universal: it

only makes sense to talk about a certain measure in relation to specific information

processing tasks (giving it an operational interpretation), since more resourceful state

for this task might be less resourceful for another.

Comparison to entanglement

For general mixed states, many different measures of entanglement, including entan-

glement cost [121, distillable entanglement [12, 84] etc. have been proposed 5 Here

I pick one of the most widely used entropic measures of entanglement, namely En-

tanglement of Formation (EoF) [121, for comparison with the measures of quantum correlations:

Definition 7 (Entanglement of Formation (EoF)). For a bipartite statae PAB, EoF

'See 1501 for a review.

35 is an entanglement measure defined as

EF(PAB) = mm piS(p ), (2.20)

where the minimization is over all ensembles of pure states {pi, 1I0)} such that PAB =

E pil10)(Vi1, and p' = trBI4~)4.

The following Koashi-Winter relation [591 establishes a quantitative dual relation involving EoF and QD distributed in a tripartite system, due to the monogamy of entanglement measures:

EF(A: B) = D(C -4 B) + S(A) - S(AB), (2.21) where S(-) denotes von Neumann entropy. Already knowing that the set of 7 states

(classically correlated states) is strictly contained in that of separable states (on which any entanglement measure should vanish), and that LOCC can create quantum correlations but not entanglement, one may speculate that entanglement should always be considered as only a portion of quantum correlations since measures of quantum correlations beyond entanglement should aim at capturing all correlations that do not exhibit classicality. However, both as entropic measures that share the same dimension (bits), EoF can be larger than QD for some states [25]: they do not obey a strict ordering. This observation is directly supported by the so-called "quantum conser- vation law" [37], a further implication of the above Koashi-Winter relation, which states that for a tripartite pure state, one can pick a particular subsystem, and the total amount of EoF that the other two subsystems share with this one cannot be increased without increasing the total amount of QD, by the same amount:

EF(A : B) + EF(A : C) = D(B -- A)+ D(C -+ A), (2.22) when we pick A. One can immediately tell that EoF and (one-way) QD for the two pairs of subsystems (AB and AC) are either both equal, or ordered differently. Similar dual relations for other measures of entanglement have also been studied in literature,

36 e.g., distillable entanglement [97]. In summary, further considerations are needed in fully explaining the relationship between entanglement and quantum correlations in mixed states.

Comparison to the total amount of correlations

Analogous to the classical case, entropy (information) shared by two quantum systems captures the total amount of correlations between them. A formal definition is as follows:

Definition 8 (Mutual Information (MI)). The total amount of correlations between two subsystems of a bipartite quantum state is given by quantum Mutual Information (MI): I(A: B) = S(A) + S(B) - S(AB), (2.23) where S(-) denotes von Neumann entropy. Equivalently, we can write MI as

I(A: B) = S(pABIIpA 0 PB), (2.24) where S(-1.-) is the relative entropy.

It is obvious that QD < MI since the classical correlation CA(A: B) > 0; For the other two measures of quantum correlations, S(A) + S(B) S(AB) under any local measurement: minJEP, DD < MI.

For entanglement measures of mixed states, it can be shown that the amount of purely classical correlations (denoted PCC in the following) lower bounds some entanglement meausures:

CAB( A : B) ED(A : B), (2.25) where ED(.) denotes Distillable Entanglement (DE) 1100], and

CA'(A : B) < Ec(A: B), (2.26)

37 where COB(A: B)= limT(o[CAB(pAOB)/n] is the regularized PCC, and Ec(.) denotes

Entanglement Cost (EC) [991, which is equivalent to the regularized version of EoF, i.e., Ec(p) = EP (p) [82]. There are also indications (without general proof though, as claimed in 1821) that EoF is additive, meaning that EF(p) = EF(p). Therefore

I treat EC and EoF as a whole for the moment. Combining with the known result that EoF is lower bounded by DE, which reflects the irreversibility of the process of formation, we have DE < EC, EoF for the entanglement measures that I mentioned.

Note that CA(A : B) CAB(A : B) always holds, so I do not distinguish them in

PCC for now. Summarizing, we have the following hierarchy:

PCC < DE < EC, EoF. (2.27)

However, it turns out that the above hierarchy is not strictly upper bounded by MI, which is discussed in the following:

Lemma 4. While D(A -+ B) < S(A) is always true [32] (note that A is the subsystem being measured), D(A -+ B) S(B) does not hold generally, though it is true in most cases [621.

Theorem 5. EF(A : B) I(A : B) (EoF < MI) does not always hold. It fails for the same set of states such that D(A -+ B) S(B) fails.

Proof. Combining the definition of MI and Koashi-Winter relation in Eq. (2.21), we have

EF(A: B) - I(A: B) = D(C -+ B) - S(B). (2.28)

By Lemma 4, the right hand side does not have a strict relation compared to zero.

Consequently, EoF < MI fails generally, but for the same set of states such that D(A -+ B) S(B) does not hold.

Not being able to find any related discussions in literature about this result, I may investigate into it in more detail as future work.

38 Summary

I conclude all the above results in the following hierarchy:

QD < minJEP < DD < MI, (2.29)

PCC < DE < EC, EoF, (2.30) where the bold ones are measures of quantum correlations beyond entanglement, and the italic ones are entanglement measures. Note that the second line does not fit into the first line since no strict relations can been established. I argue that QD and DD are respectively the lower and upper bound. Importantly, this hierarchy collapses for pure states: all quantum correlations can be identified as entanglement, which is simply measured by the marginal entropy.

2.2.3 Multipartite generalization

In the previous section, the entropic measures of quantum correlations were originally defined for correlations in bipartite states, and I presented a unified way to view all of them as quantities depending on joint and local entropy productions upon local measurements. This directly leads to natural generalizations of multipartite versions for all measures, as will be introduced in succession:

Definition 9 (Multipartite QD). Consider a quantum system consisting of n subsystems A 1 through An. We can quantify the amount of quantum correlations among these parties by the minimum difference between joint and the sum of local entropy productions upon local measurements on some specified subsystems labeled by j E [n]

(call the whole set of chosen indices J: j E J), i.e.,

Dj(PA,...An)= min [SOA ...A.) - S(pAi...AJ - Z[S(OAj) - S(PA,)]1, (2.31)

where the minimization is over all local von Neumann measurements on Aj, and

PiA ...An and PA, are postmeasurement joint and reduced states respectively.

39 The multipartite generalization QD is previously studied in a different form in

[87] (named as global QD, or GQD). The main idea is to rewrite the bipartite QD in terms of relative entropies and use this form to make the generalization:

DGQD(PA 1... An)-= min S(PA 1 -...-A. A1 --A - S(PA IIA)1 (2.32)

Note that in the original definition the local measurements are made on all n subsystems. For the more general case of making local measurements on Aj where jEj

(denote this GQD as DGQD,J), we simply need to substitute E>"= by Ej,. It is easy to show that our generalization of QD is completely equivalent to GQD:

Theorem 6. DJ(PA1 ... An) = DGQD,J(PA 1 ...A,). Elements (labels of the subsystems being measured) in J are denoted by J1 up to jj).

Proof. It is well-known that S(PAi IIfiAi) = S(fiA) - S(pA) for global projectors [761.

So I just need to show the equivalence of the first term. Similar to Theorem 1, we have

= Z .. ~ (flk ® ~ 17. ® )l.Af(Ilk, ®... ,) (2.33) (nki } { Aii A where H's are local projectors. Each k goes up to the dimension of its subspace.

40 Likewise,

S(fiAl...A.)

--tr il A... A. 109g~ . A.--,)

= -tr -- 0 -0-- T@ )PA 1... A -(r . .(, i ) logP Ai...A.]

Ir~kj } (HU An>31 I = -tr o ... H kj )PA...A. logpA...A.(lk, D 0 Hkis)

{njr }

= -tr --- (11k 0Aj 1 (2.34) ni ..{r.} {

= -tr(pA .. ,log pA...A.), (2.34)

where the third equality follows from the fact that (). 0... 0f - -- j' ) commutes A1 jI7|

with PA 1...A, and thus its logarithm, the fourth equality uses the cyclic property of

trace and the fact that ([P1 (D . (D0 - k ) is idempotent, and the last equality uses A3, Au5 moetadte ateuaiyue the fact that they sum to identity. Therefore

S(fi ... A.) - S(PA1 ... An)

- -tr(A 1 ... A. log pAi...A A) +tr(pA,...An log pA 1... A.)

= S(PA 1...AI||VAi... A). (2.35)

Since corresponding terms with respect to the subsystems are equal, the overall quan-

tity is also minimized by the same set of local measurements. Therefore the two

definitions are equivalent. El

For minJEP and DD, since local entropy productions are not involved in their

definitions, it is even easier to make multipartite generalizations:

41 Definition 10 (Multipartite minJEP). Using the same assumptions and notations for defining multipartite QD, the multipartite generalization of minJEP is defined as

M-J(PA..A) = min[S(PA 1...A) - S(pAl...A)] . (2.36)

Definition 11 (Multipartite DD). Similarly, the multipartite generalization of DD is defined as

DDj = S( ,A...A) - S(PA...A.), (2.37) where all local measurements are done in the local spectral basis (with respect to which the reduced state is diagonal).

Using the above ideas, several properties of multipartite quantum correlations can be shown. Moreover, interesting connections may possibly be drawn to the study of genuine multipartite correlations [13].

2.3 Scaling behaviors in spin lattices

I have already discussed how to understand and measure different classes of correlations living in quantum systems from an entropic point of view in previous sections of this thesis, and now we can apply the knowledge to analyze real physical systems. Since its origination, the concept of correlation has always played an essential role in studying many-body systems, e.g., spin lattices, since it indicates how numerous elements of the complex system may behave collectively. For quantum many-body physics, quantum entanglement may prove crucial for studying anomalous nonclassical phenomena, such as quantum phase transitions [88]. Moreover, the scaling behaviors of entanglement, especially in ground (zero temperature) or thermal states, have been shown to exhibit counterintuitive properties, e.g., the entanglement entropy frequently scales as the area enclosing a subsystem, instead of its volume, in ground states [36], i.e., obeying the so-called area law. This type of scaling behavior connected to other aspects and topics in fundamental physics, such as black holes and holographic principles. However, as mentioned previously, people found that nontriv-

42 ial quantum correlations also exist in certain unentangled quantum systems, leading to nonclassical behaviors, which directly raises lots of questions, e.g., how do they behave in certain many-body systems? And how do we characterize them?

In this section, I study the behaviors of quantum correlations in a typical model of quantum many-body systems - spin lattices, with the aid of the results provided by

Section 2.2. First I shall introduce the calculation of measures of quantum correlations between pairs of spins in spin-1/2 lattice models with Z2 symmetry (no Z2 symmetry breaking terms in the Hamiltonian, e.g., magnetic field). As the two-site reduced states are generally mixed, as we will see, the analysis of quantum correlations beyond entanglement may provide new physical insights. Heisenberg XXZ chain will be discussed as a specific example. Then for general dimensions, I shall discuss the scaling behaviors of quantum correlations between a chosen part of the system with

the environment, in analogy to area law. Note that I shall use QD and DD as

the typical measures of quantum correlations in this chapter, since they respectively

represent the tighter and looser reasonable quantifiers, as shown previously.

2.3.1 Quantum correlation between two spins

In order to calculate the correlation between two sites in a lattice, we need the joint

state of these two spins. First I shall discuss some general properties of the reduced

state of two sites in quantum spin models obeying Z2 symmetry as preliminaries for further results.

It has been shown in [541 that in the computational basis {l00), 101), 110), 111)},

the reduced density matrix of two sites i and j (tracing out other sites) in the Z 2 - symmetric quantum spin models takes the form

oo 0 0 Lo3

pg 0 91 Q12 0 , (2.38) 0 Q12 922 0

\03 0 0 L33

43 with only the diagonal and anti-diagonal entries being non-zero, therefore bears the name of "X state", which is mixed in general: measures of quantum correlations do not simply reduce to forms of entanglement entropy. Here we only need to consider real X state, i.e., 912 = *12 and 0 0 3 Qs, since they can always be transformed into Eq. (2.38) via local unitary transformations, which do not disturb the correlation.

Considering the trace constraint, we see that this density matrix actually has only five degrees of freedom.

For some of these states, QD and DD can be explicitly evaluated. Since the detailed calculation is lengthy and technical, I shall place all intermediate steps in

Appendix A, and directly present the final results here.

For the following two classes of real X states, one-way QD can be explicitly calculated:

1. IV/0oo33 - P1Q221 19121 + 10031:

D(j -+ i)

-A+, log A+, - A+, log A+,-

4 1 +GT 1-G + E A, log A. - 2 (I+ G') - 2log2 log (I G') + 1, a=1 (2.39)

2. (10121 + IL03) 2 (GoO - Lu)(L33 - 022):

D(j- i)

=- S Ak,I log Ak,I {k=0,1} It=+,-} 41 +Gz 1G + A, log A, - 2 log (1 + G') - 2 log (1 - G') + 1, a=1 (2.40)

44 where G's denote correlation functions

Gz = (of) = tr(ou pij) = Poo Pu - P22 - Q33, (2.41)

G- = (or) = tr(ojp~i) = Poo - P1 + P22 - 233, (2.42)

G- = (ouff) = tr[(ou 0 o-)pij] = 2 (912 + 003), (2.43)

G Y = (uroj) = tr[(o 9 o0j')pij] = 2 (912 - 003), (2.44)

G = (oafuj) = tr[(ou 0 o )pij) = oo - L1 - 022 + 933, (2.45)

and

A = (1 + Gx + GYF - G ), (2.46)

S = (1 - Gf - Gg - Gff), (2.47)

2 A 3 = (1+Gij /+4(Gf)2+ (G: - GYF) ), (2.48)

4 = (1 + Gzz - V4(Gz)2 + (GX - GY)2), (2.49)

A+,i = A-, = ( "t +(Gj)2 (2.50)

Ao,+ = 1(1 + Gj t Gz Gz), (2.51)

Ai= (1 - G : G t Gf). (2.52)

On the other hand, since the computational basis is automatically the Schmidt basis, DD of these states are given by

DD(j -- i)

= - Ak,1 log Ak,1 {k=0,1} {'=+,-} 1 +oz 1 G. + A, log Aa- log (1 + G') - - 2 log (1 - Gf) + 1, a=1 (2.53) using the above notations, and can be extended to all real X states that are outside

45 these two regimes.

2.3.2 Example: 1D Heisenberg XXZ chain

In this subsection, the two-site scaling behaviors of correlation measures is preliminar- ily illustrated via a specific example: the 1D spin-1/2 anisotropic Heisenberg XXZ spin model, whose Hamiltonian reads

Hxxz(A) = Z(ui + (70ol + Ao4i ), (2.54) where the anisotropy parameter A controls the quantum phases. Note that this model can be solved by the Bethe ansatz [261. For A > 1, the system is in the antiferromagnetic N6el phase which breaks the lattice translation symmetry, and for

A < -1, the ferromagnetic Ising phase, which breaks the spin reflection symmetry.

Both of the above phases are gapped and have two-fold degenerate groung states.

A -+ +oo and A -4 -oo are respectively the antiferromagnetic and classical Ising limit. The model is in the critical XY phase (i.e., gapless) when A E (-1, 1], which is known to be described by a c = 1 conformal field theory (CFT), as the correlation length diverges and the system becomes scale invariant [361.

Note that the XXZ chain exhibits U(1) invariance [89], namely, [H, & = 0, which is even a stronger constraint over the elements of the density matrix than the

Z 2 symmetry, i.e., the two-site reduced state is an X state and 903 also vanishes, hence we are safe to use results previous results.

Quantum entanglement in many-body systems, especially of ground states, are important in studying the behavior of the systems, e.g., quantum phase transitions

[88]. For a pure bipartite quantum state, the entanglement entropy corresponding to a certain partition is uniquely defined, which is very useful for indicating quantum criticality. The volumetric entanglement entropy scaling in different regimes of the

XXZ model is discussed in [26, 2]. In the 1D critical regimes (in this model, A E

46 (-1,11), CFT yields that the (subsystem) entanglement entropy scales as

SA(1) = log 1 + k, (2.55) 6 where k is a model-dependent constant and c, i are holomorphic and antiholomorphic central charges respectively [211, indicating different universality classes.

Since the pairwise correlation in gapped phases is naturally expected to decay exponentially, I now analyze the two-site scaling of QD and DD in the critical phase where A E (-1, 1) at zero temperature. In this gapless regime, using our notations, the pairwise spin-correlation functions scales as [67]

Gx = G' ~, li- j,-0, (2.56)

Gf ~ i j-2+e2ikFliiI j -~ j-, (2.57) with critical exponent given by

1 arcsin(-A) (2.58) 2 7r and e2ikFIa'I a phase factor. Notice that although the leading order term in Gf is not fixed, Gf is always less significant than G-, G'.

Plugging these spin correlation functions into the parametrization Eq. (2.41)-

(2.45), we see that this state falls into the first class: the exact QD is given by Eq.

(2.39). Note that G should vanish. By Taylor expanding Eq. (2.39), up to leading order, QD scales as

2 D(j -+ i) ~ (Gf) - 2, (2.59) and it turns out that DD obeys the same scaling:

DD(j -+ i) ~ -ijl-20, (2.60) by expanding Eq. (2.40). Therefore in the critical XY phase at zero temperature, measures of quantum correlations beyond entanglement decay polynomially, which

47 resembles the behavior of spin correlation function, but with different exponents.

It is also interesting to study the behavior of quantum correlations depending on the temperature of the bath, e.g., in ID XYZ chains, QD increases as the bath temperature grows, while entanglement is expected to decay [1031.

2.3.3 Area laws

When the interactions in quantum many-body systems are local, the ground state entanglement entropy typically grows linearly with respect to the boundary area of the subregion instead of the volume, in contrast with the expected extensive behavior. This kind of scaling behavior is said to obey an "area law" [36]. The general mathematical statement if a physical quantity -1 of region A obeys

4D(A) = O(b9AI), (2.61) where OA denotes the boundary area of A, we say that the area law is satisfied. The intuitive picture here is shown in Fig. 2-1. In locally interacting noncritical systems, the correlation length is finite so that sites in A and B that are separated by a distance further than the correlation length (the shaded stripe) should not contribute to the mutual information or correlation measures between A and B, hence bounded by the number of sites at the boundary and therefore scales as the boundary area.

It is worth mentioning that the area laws for entanglement entropy has deep connections [94, 14] with the famous area dependence of the Bekenstein-Hawking black hole entropy [9, 45], which states that the entropy of a black hole is proportional to its horizon area A: kA 3 kc A (2.62) SBH - 42 ~ h' 4lp 4Gh' (' where lp = VGh/c3 is the Planck length, as they share similar scaling behaviors.

These discoveries of black hole entropy scaling laws were the driving force for several studies of entanglement entropy scaling in quantum fields later on [46, 22]. It has been argued that the information contained in a volume of space can be represented

48 000000004000000000

V00 0000000 .000 00 .0002 000 c00 00 00oo 00o 000 00 00000000 000

000 00 00 0 0000 000 000 o 0000

00000000000000000

Figure 2-1: (Adapted from [1041) Area law. Intuitively speaking, under locally interacting Hamiltonians, the correlation length in noncritical phases is finite so that sites in A and B that are separated by a distance further than the correlation length (the shaded stripe) should not contribute to the mutual information or correlation measures between A and B, hence bounded by the number of sites at the boundary and therefore scales as the boundary area. by a theory which lives on the boundary of that region [98, 15], which is widely known as the holographic principle - the information contained by a region depends on its surface area, rather than on its volume. This insight remains one of the central topics in theoretical physics to this day. The famous AdS/CFT correspondence [681 is a beautiful realization of the holographic principle. At a fundamental level, all these area laws might be related. In this section I shall briefly discuss the area laws for entanglement entropy and total mutual information for general spin systems, and in turn present the area law for measures of quantum correlations in locally interacting noncritical spin systems.

Entanglement entropy

The area laws of entanglement entropy in various contexts have been been extensively studied for years'. For quantum many-body systems on lattice (the entire system) W where A is a subregion and B = W \ A its complement, S(A) = O(IAI) implies that

6 Refer to 1361 for a review.

49 the area law is satisfied in the system. We emphasize that, in fact, it is truly unusual for a quantum state to satisfy an area law 1361 as it has been shown that the typical entropy of a subsystem is nearly maximal [90], indicating that it should scale as the volume instead of boundary area. For the purpose in this section it suffices to know that for general ground states of quantum spin systems in gapped, i.e., noncritical phases the area law is obeyed. The area law statement was first made rigorous by

Hastings in 1D [44], and recently the idea that exponential decay of correlations leads to area law scaling of entanglement was also formally shown in ID [171 (quantum data hiding states being obstacles of this argument in general dimensions [341). Notable violations take place at quantum criticalities, models with Fermi surfaces etc., and the non-trivial topological order will result in a negative term in the ground state entanglement entropy, namely topological entanglement entropy. The analysis on 1D

Heisenberg XXZ chain earlier can serve as a nice example: in noncritical regimes, quantum correlation between two spins decays exponentially as the distance between them grows, indicating that the entanglement entropy fulfills the area law. While for critical regimes, QD and DD are both shown to decay polynomially at zero temperature. Indeed, area law should fail in this case - quantum correlation is not local

(finite-ranged) anymore.

Mutual Information

As has been mentioned earlier, the total correlation in a bipartite quantum state

AB is given by quantum Mutual Information (MI), i.e., I(A: B) = S(A) + S(B) -

S(AB), where S(.) denotes the von Neumann entropy. Now consider states in thermal equilibrium, i.e., thermal states, which take the form PAB - e H /tre-H with inverse temperature 3. Thermal states minimizes the free energy F(p) = tr(Hp) - S(p)/, and F(PAB) < F(PA 0 PB), from which one can obtain the following

Lemma 7 (Area law for MI [104]). We denote the total Hamiltonian as H HA + HB + 1&Hwhere HO collects interactions crossing the boundary. The MI of a thermal

50 state is bounded by a first order function of Ha, i.e.,

I(A : B) /3tr[Ha(pA 0 PB - PAB)] , (2-63) the right hand side of which only depends on the size of the boundary. Therefore the area law is satisfied.

More specifically if we only consider two-site interactions, we will have

I(A : B) < 20A max I|hijiI, (2.64) i,jcaA where ||hjj|| are eigenvalues (strengths) of the two-site interaction between i and j

(across the boundary).

Notably, it is also shown that finite correlation length generally implies an area law scaling of MI [1041.

Quantum correlations

Therefore for thermal states, the following scaling theorem for QD and DD can be directly proven by the hierarchy in Eq. (2.29):

Theorem 8 (Area law for quantum correlations). For general quantum spin systems in noncritical regime with local interactions, the amount of quantum correlations between a subgraph A and its complement in thermal states scales as the boundary area.

Proof. Denote the entire spin system as W and B = W \ A. As shown in Section

2.2.2, we have the hierarchy QD < DD < MI for the same state, i.e.

D(A -4 B) DD(A -+ B) < I(A : B)

51 Specifically for two-site interactions [54], by Eq. (2.64), we immediately see that

D(A -+ B) < DD(A -+ B) < I( A: B) < 2/310A max I|hjjI|, (2.66) ij EOA

Conclusively, D(A -+ B), DD(A - B) |OAI,I i.e., scales with the boundary area. The specific argument about QD can be found in [54]. El

2.3.4 Outlook

As indicated by the holographic principle, the information of a region should depend on its surface area instead of volume, at the fundamental level. Therefore studies on scaling behaviors of correlation measures in many-body systems may prove fruitful for both the fields of quantum information theory and many-body statistical physics, and may even provide some unique insights into other aspects of physics. New findings may emerge at this interface.

2.4 Diagonal Discord (DD)

In Section 2.2, DD, which can be thought of as the simplified version of QD by fixing the local measurement to Schmidt basis measurement, or the joint entropy production upon such a local measurement, was proposed as a candidate measure of quantum correlations beyond entanglement. (The formal definitions of DD and its multipartite generalization were given in Definitions 5 and 11).

Although this quantity has showed up in other contexts [60, 65], its properties and physical interpretations have not been well explored yet. In this section, I shall systematically discuss the motivations for using this quantification for quantum correlations, and some of its important features, as well as possible future directions.

2.4.1 Motivation

The quantification of quantum correlations is based on local measurements. Since there are infinitely many ways to measure a quantum system, a perfect quantification

52 of quantum correlations has to involve an optimization over all possible measurements, which seems extremely difficult for evaluation. Indeed, computing QD is shown to be an NP-complete problem [53], and this proof can be directly extended to other entropic measures involving optimization over all possible measurements, such as minJEP. The rigorous statement goes as follows:

Theorem 9. For some measure of quantum correlations6 that involves optimization of an entropic quantity over all possible measurements (e.g., QD or minJEP), given a bipartite quantum state PAB of dimension m x n and a real number b with the promise that either (Y) S(A -+ B) < b or (N) 6(A -+ B) > b+c where e is inverse polynomial in m, n (E = 1/poly(m, n)), it is NP-complete to decide which is the case.

Proof. The NP-completeness of QD along with several entanglement measures including EoF and EC, is shown in [531, by reducing the problems of computing these measures to the problem of separability (polynomially), which is known to be NP- complete [56]. Notably, QD and EoF is connected by Koashi-Winter relation, which indicate the polynomial reduction. Similarly, computing other entropic measures of quantum correlations involving optimization terms can be polynomially reduced to computing QD, thus is NP-complete. El

Therefore, the primary advantage of DD is that it is much easier to evaluate, since the complexity of diagonalizing a matrix is generally polynomial. Based on Koashi-

Winter relation (Eq. (2.21)), we can define the following entanglement measure dual to DD:

Definition 12 (Diagonal Entanglement of Formation (DEoF)). Koashi-Winter relation establishes a duality between EoF and QD (between different parties). By substituting the QD term in Koashi-Winter relation by DD, an analogous dual relation can be obtained:

EF(A: B) = DD(C -+ B) + S(A) - S(AB), (2.67) the left hand side of which is named DEoF. Similar as DD compared to QD, DEoF is a

53 simplified version of EoF without the optimization, which is much easier to compute.

2.4.2 Properties

Due to the significant simplification, DD is not expected to capture all aspects of quantum correlations. However, since it is easy to calculate, this quantity will be of much more practical value if we can argue that DD satisfies the basic criteria that a reasonable measure should possess, and still works well in various scenarios. Now I investigate some of the properties of DD.

I shall first explicitly argue that DD satisfies the basic criteria of an acceptable measure of quantum correlations discussed in Section 2.2.1:

1. DD is nonnegative. Moreover, the null set of one-way DD (without loss of

generality, assume the local measurement is done on on A) is exactly the set of

CQ states, (for the two-way case, the set of CC states), i.e., the same as QD

and any other reasonable measure of quantum correlations.

Theorem 10. For any bipartite state PAB, DD is nonnegative, i.e., DD(A -+ B) > 0.

Proof. Since QD is nonnegative [32], and QD < DD, DD is nonegative. Or directly by Theorem 1 which states that local projective measurements can never

decrease entropy, DD(A -+ B) > 0 since DD is by definition the joint entropy

production upon local measurement in the Schmidt basis (projective). El

Theorem 11. For a bipartite state PAB, DD(A -+ B) - 0 iff PAB is CQ, which

is exactly the null set of QD. Likewise, DD(A ++ B) 0 iff PAB is CC, also

same as QD. The above null sets are also shared by minJEP.

Proof. Necessity: As has been discussed, the states with zero one-way QD is

called CQ states, which take the form PAB = EkpkIk)A(ki & pB, where {Ik)A} is automatically the Schmidt basis, implying that DD vanishes.

54 Sufficiency: DD(A -+ B) = 0 implies that the diagonal basis gives zero discord

value, which is minimal, i.e., exactly the value of QD, since QD is nonnegative

by Theorem 10.

The above arguments can be directly generalized to the two-way case. Since QD and DD are respectively lower and upper bounds of minJEP, the null sets of these quantities must be the same. l

As mentioned in Section 2.1.1, the set of 7 states is measure zero and nowhere

dense in the space of al quantum states: almost all quantum states have positive QD and DD [381.

2. DD is invariant under local unitary transformations. This argument is included

in the following theorem that applies for more general cases:

Theorem 12. Suppose U and V are respectively arbitraryunitaries on the sub-

spaces A and B of a bipartite quantum state pAB, then any entropic measures

for quantum correlations, including any DD (namely, one-way from any sub-

system, two-way) of the state P'AB = (U 0 V)pAB(Ut 0 Vt), is the same as that

of the original state PAB-

Proof. The essence is that local unitaries can be thought of pure rotations of

basis vectors of the local Hilbert space since they preserve distances. Without

loss of generality, use the computational basis { ij)} to write the matrix rep-

resentation of PAB, and use {Uji) 0 Vjj)} for P'AB. One can directly tell that

these two matrices are completely identical. Therefore all entropic quantities of

PAB and P'B, and consequently all entropic measures of quantum correlations, are the same. l

From the above, DD survives the basic criteria as an acceptable measure of quantum correlations beyond entanglement. In fact, DD also exhibits several other important properties. Compared to QD and other entropic measures:

3. QD < minJEP < DD (Section 2.2.2).

55 4. The partial order of quantum states determined by DD is different from those

determined by other entropic measures, e.g., QD and minJEP, given that they

reduce to the same quantity for pure states (Theorem 3 in Section 2.2.2). There

are several equivalent statements, e.g., there must exist some states whose order

of resourcefulness given by different measures are different, or these measures

are not monotones of one another.

Independently, DD itself also exhibits the following features:

5. DD is upper bounded by the von Neumann entropy of the subsystem being

measured.

Conjecture 1. Given a bipartite quantum state PAB, DD(A -+ B) < S(A), where S(-) denotes von Neumann entropy.

I currently state it as a conjecture due to the lack of rigorous proof. Note

that if S(A) (the subsystem being measured) is replaced by S(B) (the other

subsystem), this inequality does not always hold since it is not even generally

true for QD, as discussed earlier.

6. DD is not monogamous for all states, which is a direct implication of [95].

The statement only excludes the possibility that all quantum states satisfy monogamy of measures of quantum correlations. The complete characterization

of the states that are monogamous for different measures is an interesting open

problem which may provide insights for the nature of quantum correlations.

7. DD is invariant upon attaching pure ancilla to any local subsystem.

Theorem 13. The operation of attaching local pure ancilla on any subsystem

of a bipartite quantum state PAB preserves DD, i.e.,

DD(A -+ BC) DD(A -+ B), (2.68)

DD(AC -+ B) DD(A -+ B), (2.69)

where PABC = PAB 0 1,0)c(VI -

56 Proof. By Theorem 12, local unitaries do not vary DD. One can always find a

unitary U acting on subsystem C such that U14')c = I0)c. Now Eqs. (2.68)

and (2.69) are true by inspection, since the density matrix of PCAB takes the

form PAB 0 ... 0 0 UPCABUt = (2.70) 0 0/

i.e., only the top-left block being nonzero (PAB). It is now easy to tell that all

eigenvalues involved remains invariant (only some more O's), thus the entropic

quantities as well, which completes the proof.

8. For quantum lattice systems in thermal equilibrium (at finite temperature), the

scaling of DD satisfies an area law (Theorem 8 in Section 2.3.3).

Notably, the above features are discussed under the bipartite and one-way setting for simplicity, but they can be easily generalized to two-way or multipartite cases, as I have done several times in this thesis.

This incomplete list includes some of the important properties that one would naturally expect DD to exhibit as a good candidate. In these statements, "DD" can be replaced by "well-behaved measures of quantum correlations". That is, although the quantity is defined with a significant simplification, DD still works well as a candidate measure, making it quite favorable in practice. Note that there does exist a simple feature that is exhibited by QD [1081 but not DD, namely, continuity:

9. DD is not a continous function of density matrices, i.e., there exist states that

are arbitrarily close, but have non-vanishing gap in DD. Examples can be found in [107].

Further discussions are needed for the consequences of this property.

For the moment I omitted some properties that do not seem very meaningful, or have not been rigorously proved. To reveal the complete nature of DD, and in

57 particular, what kind of role it should play among measures of quantum correlations in various physical scenarios, is a fruitful research direction to pursue. Specifically, the list of interesting open questions concerning the mathematical properties of DD may include, but not restricted to:

" Is there a good upper bound for the difference between DD and QD?

" As several measures including QD and geometric discord etc. are known to

exhibit non-contractivity under general local operations (more specifically, non-

unitary evolutions without measurements or attaching product states) while

some others do not, what about DD? (In Section 3.2 I shall elaborate more

on this issue. Locally even tracing out orthogonal flags may generate quantum

correlations from -y states, thus is forbidden for this question.)

Generally, the characterization of the sets of states where different inequalities about

DD reduce to equalities, and the comparison with those sets for QD, may provide useful insights of DD and quantum correlations as well:

" For what states is DD equal to the entropy of the subsystem being measured? Is this set different from that of QD?

" For what states does the monogamy of DD hold? Is this set different from that of QD?

" For what states does the hierarchy QD < minJEP < DD collapse, i.e., the optimal basis is exactly the Schmidt basis?

Answers to these questions may provide significant progress and insights for related

areas. Besides its mathematical nature, the understanding and significance of DD

associated with physical phenomena will be among the topics of the following section and Chapter 3.

2.4.3 Physical interpretation

From the discussions previously, one may already sense from several aspects that the

concept of quantum correlations is in very close relation to thermodynamics. Indeed,

58 the idea of QD was kind of originated from such a context [77, 109J. Remarkably, it can be further shown that positive quantum correlations beyond entanglement is the necessary and sufficient condition of generating heat transfer between quantum systems 1641. I shall only briefly mention the results of this work here without going into details. The rigorous theorem goes as follows:

Theorem 14. Consider two quantum systems A and B evolving continuously from an initial state PAB (0) under a unitary time evolution UAB (t). If the joint state PAB (t) remains CC for all t, then the time evolution for A and B can always be written as

pAB(t) = UA(t) 0 UB(t) pAB(0) Uj4(t) 0 UB(t), (2.71) i.e., they are effectively non-interacting, thus heat transfer cannot be generated.

A step further, a quantitative relation between the rate of heat transfer and generation of quantum correlations (measured by DD) over a sufficiently short period of time At can be established. Suppose A and B respectively starts from uncorrelated thermal states at temperature T, i.e., the initial state reads PAB(0) =

(e-AHA/IZA) 0 (e -,HB/ZB) where 3 = 1/(kT), then under H = HA + HB + HAB, the rate of energy transport turns out to be directly proportional to the rate of DD generation in the infinitesimal time limit: By Eq. (2.14), the infinitesimal generation of DD can be written as

ADD(B -+ A) = St(A) - SAt(AB)

= -trApA3 log PAs

3 = tr AHAAPA + trI3 BHBAPB

= 3AAEA + 3BAEB

= (, 3B - A)AEB, (2.72)

where we discarded high order terms for the second line, plugged PAB(0) in for the third line, and used the fact that AEA = -AEB in the weak coupling limit for the last line. Indeed, since thermal states are diagonal in the energy eigenbasis, DD has

59 a direct physical interpretation in terms of energy flow and entropy increase, and the resulting "energetic discord" can be measured experimentally.

Still, it is natural to ask whether the assumption on the local measurement basis for the measure for quantum correlations we used in establishing the above quantitative relation can be dropped. Therefore we propose the following

Conjecture 2. In the infinitesimal limit of evolution, DD is equivalent to QD, i.e., the Schmidt basis can be considered the optimal basis for QD. In other words, dis- continuities of optimal basis do not take place at y states.

Possible ways to prove or disprove this argument may include brute-force calculations of the behaviors of these quantities in the infinitesimal limit, or looking for counterexamples from the class of states for which the optimal basis is already known, e.g., the two classes of real X states discussed in Appendix A. Unfortunately, these two classes do not share boundaries in the state space [27], so the results about them are not useful for studying the discontinuity of optimal basis, but it might help to keep looking in this direction.

60 Chapter 3

Quantum Resource Theories (QRTs)

A goal of quantum information theory is to make use of peculiar properties of the quantum world to achieve effects that are inaccessible classically. The desire to find a unified and well-structured framework in order to organize these ideas has led to the notion of Quantum Resource Theory (QRT) [49J. The resource x is usually an attribute of quantum states. In this sense, QRT of x essentially the theory of x as a physical resource. A complete theory should be able to tell you basically anything you want to know about x, including how to create it, how to quantify it, how to convert it, how to make use of it and so on. If you want to gain full understanding of a quantum object, a good way to do so is to study or establish the resource theory of it. Quite a few specific QRTs that find use in a wide range of scenarios has been identified, e.g., entanglement [50], stabilizer quantum computation [1011, contextuality [42j, quantum coherence [8]. Surprisingly, the ideas of QRTs have reached far beyond the scope of quantum information theory. Some striking examples are the resource theory of asymmetry [701 which leads to a generalized Noether's theorem, and the resource theory of quantum states out of thermal equilibrium [181 etc.

As can be already seen, the spirit of QRT is quite generic, and all QRTs share a similar kind of structure. Instead of looking at scattered theories for various resources, we ought to depict the whole family of QRTs from a unified point of view.

In this chapter, I shall first discuss the unified framework of traditionalQRTs where the resources are associated with quantum states, so that the theories of specific re-

61 sources can easily fit into this structure, and work as sub-theories. As mentioned, people have established lots of QRTs within this framework, among which the QRT of entanglement, LOCC being the corresponding set of free operations, is the most well-understood and famous one. Entanglement has long been considered as the most important resource for various kinds of quantum advantages. However, people found that quantum correlations beyond entanglement, which is the topic of Chapter 2, are actually responsible for the triumph of "quantum" in several contexts [33, 30, 81].

Nonetheless, a satisfactory QRT for quantum correlations has not been fully characterized. Some discussions on this problem will be carried out in Section 3.2. In fact, analogous to the studies on traditional QRTs, quantum operations can also be treated as resources, thus a dual framework can be formulated. In Section 3.3, I shall introduce the ideas of dual QRTs. As will be seen, this duality is quite natural from a mathematical point of view, and an even more general category-theoretic structure might be identified.

3.1 Unified framework

The purpose of this section is to establish an underlying unified framework such that specific QRTs can fit into this structure as a sub-theory. I shall start from very intuitive analysis on the whole structure, and then proceed to establishing a more rigorous mathematical framework.

3.1.1 Elements

To begin with, let's review the general logic of QRTs by considering the basic ingredients that a well-defined QRT should contain. First of all, a particular resource associated with quantum states serves to label the QRT, along with which there is a quantifier that determines the "resourcefulness" of a particular state. Correspond- ingly, a set of free operations which cannot increase the amount of resource should also be identified. Or in the other way, one can put particular restrictions on the allowed operations, and then those states that can never be prepared with these op-

62 Free Operations * 9 CPTP maps - Cannot increase the amount of resource - -+ Set of free states: closed

+ COMPLEMENTARY Resource - Usually peculiar quantum properties - Cannot be created by free operations - Monotones: quantification 4f Larger set of accessible operations, states etc.: additional power, universality...

Figure 3-1: Intuitive illustration of the basic content and structure of a QRT. erations become resources. With the aid of these operations, it is also impossible to create more resource than before, as measured by the quantifier. The power of free operations (which can only prepare free states) may be very limited, but with resources, we are enabled to achieve a certain kind of universality. One should have already noticed the complementarity between these ingredients. Certainly, these respective components also have to satisfy some basic properties. For example, the allowed operations should be a strict but not empty subset of all possible quantum maps/channels (the theory is trivial if all or none operations are free). Moreover, one would expect that putting garbage together will not create resource, and a series of free states should not converge to a resourceful one. Math- ematically, the set of non-resourceful states should be closed under tensor product, and is a closed set.

In summary, the basic structure and ingredients of QRTs are illustrated in Fig. 3-1. Note that as expected, an abstract framework that unifies the ideas of QRTs can be constructed using category theory [29].

63 With the above basic understanding of how QRTs work, we can now give more rigorous and reliable descriptions of the unified framework for QRTs. Every QRT is consisted of the following three fundamental pieces of elements, respectively satisfying several basic postulates:

1. Free operations.

Denote the set of free operations as A1 . It satisfies:

(a) A1 is the subset of all possible operations (CPTP/CP maps, depending on the context).

(b) A1 can never increase the amount of resource, as measured by a particular quantifier.

2. Resourceful states. Denote the set of resourceful states as R, and its complement, i.e., the set of free states, as F. F needs to satisfy:

(a) F is closed upon doing tensor product.

(b) F is a closed set.

Note that one can also define the "maximally resourceful states" as those states that can be deterministically converted to any other state by free operations.

3. Resource monotone/quantifier. The quantifier is a continuous function of quantum states (denoted by m(p)) that quantify its resourcefulness (ratio determines conversion rate). It determines the partial order of states with nonzero resource. The properties that m has to satisfy are:

(a) m(p) > 0: always nonnegative.

(b) m(p) = 0 iff p E F.

64 We have to be very careful in explaining the relationships between these elements.

Representing different objects, these components do possess a certain extent of inde- pendence, but are indeed all tied together. The set of free operations Af need not be maximal. The quantifier is associated with the state, but as will be discussed in

Section 3.1.2, given that Af is not maximal, the quantifier may not be unique: there can be various measures corresponding to different information processing tasks. For instance, the entropic measures of entanglement/quantum correlations even give different orderings for entangled/non-y states (Theorem 3). In this case, as mentioned in

Chapter 2, it only makes sense to talk about resource quantifiers in relation to specific operational contexts involving conversions between resources. This is the reason why

I suggest the monotone/quantifier as an element instead of the resource, since when the quantifier is not unique, a resource actually corresponds to a class of theories.

Since the second postulate on Af heavily depends on the quantifier, so for this class

of theories the free operation sets and quantifiers are paired up.

Therefore we have the following hierarchy of QRTs:

Level 1 : {., R, .}

Level 2 : {Af, R, m}

The "resource" labels the Level 1 QRT, where {Af, m} may be non-unique. This is

the level where usual notions of QRTs are at, e.g., entanglement theory. For Level 2

QRTs, all elements are fixed.

Note that a very recent work [16] introduced a general structure of QRTs in a

slightly different way. The basic ingredients of QRTs as they suggested are i) the

resources (e.g., entanglement); ii) the non-resources or free states (e.g., separable

states); iii) the restricted set of free (or allowed) operations (e.g., LOCC). Here the

ingredient "resources" are essentially the labels of level 1 QRTs (corresponding to a

class of resource quantifiers) in this framework, and the set of free states is the comple-

ment of the set of resourceful states. As I discussed in the last paragraph, a resource

may correspond to various theories by my definition. Moreover, [16] proposed some

65 more postulates on F including convexity and closeness under tracing out spatially separated systems and party swap (which seems redundant), which I do not regard as basic requirements. But they can be set as restrictions in specific QRTs.

Here are some quintessential examples described by Level 1 language:

1. Entanglement: Af = LOCC.

The resource theory of quantum entanglement has been widely studied and rec-

ognized [51, 101. If two or more parties are only allowed to do local operations

and communicate with one another via classical channels (LOCC), then entan-

glement becomes a corresponding resource in the sense that it can never be

created using LOCC, and furthermore, allows us to perform tasks that are im-

possible with LOCC only, e.g., quantum teleportation [11], and therefore serves

as the essential element of quantum communication. Entanglement is also iden-

tified the indispensable resource for lots of computational and cryptographic

tasks (I shall not delve into details here).

Note that this is a typical example of Level 1 QRT. There are various measures

of mixed state entanglement (note that for pure states, quantum correlations

fully reduce to entanglement, and all entropic measures of entanglement become

equivalent), in relation to different physical scenarios [511, such as EoF, EC, DE as defined in Chapter 2, each determining the conversion rate under some

specific settings (corresponding to a Level 2 QRT), e.g., distillation and dilution

[75, 10]. The fundamental reason for this non-uniqueness is that LOCC is not

the maximal set of operations that cannot create entanglement (LOCC is a

strict subset of all non-entangling operations), as will be further illustrated in

Section 3.1.2. Interestingly, if we restrict to pure state entanglement, LOCC

becomes maximal, and the different measures all reduce to a unique one - the entanglement entropy.

2. Purity: Af = NO.

Purity, which represents knowledge/information of the state, is defined as log d-

S(p), where S(-) denotes von Neumann entropy (which is automatically the

66 quantifier). Indeed, purity is zero for maximally mixed states (zero knowledge),

and maximal for pure states (full knowledge). The corresponding free operation

set is characterized as Noisy Operations (NO) consisting of

(a) attaching maximally mixed (I/d) ancilla;

(b) doing partial trace (PT);

which is the maximal set of operations that can never decrease the entropy: By

Theorem 16, this theory is reversible, thus a perfect QRT (refer to Section 3.1.2).

NO that preserves dimension reduces to Exactly Factorizable maps, which will

be discussed later. Surprisingly, people have discovered connections between

the resource theory of purity and the generalizations of Birkhoff statements, which are essential elements in the hierarchy of operations that I shall discuss

in Sec. 3.1.3. The idea of NO is adopted to disprove the asymptotic Birkhoff

conjecture, which claims that the n-fold tensor product of a doubly stochastic

map can always be approximated by mixtures of unitaries in the limit of large

n 192, 91]. The resource theory of purity will also play an important role in that of quantum correlations beyond entanglement.

3. Magic: Af is stabilizer/Clifford operations.

Natural connections can be drawn between QRTs and universal quantum com-

putation. According to the well-known Gottesman-Knill theorem [41, 75] and

Aaronson's improvement [11, quantum computation based on stabilizer opera-

tions (generated by Clifford group elements) can be efficiently simulated on a

classical computer (in classical complexity class EL [11). This seems to have

closed the gap of computational power between quantum and classical com-

puters for some cases, however the point is that with only Clifford operations

we cannot obtain universal fault-tolerant quantum computation. The resource

needed to achieve this universality is then non-stabilizer states, or so-called

magic states [19, 31. In this sense, the resource is the "magic", or negativity

67 Resource Af F Example m, Entanglement LOCC Separable states o,, infsP R(pI Usep) Purity NO Maximally mixed state I/d R(pj lI/d)= logd -S(p)* Magic Clifford Stabilizer states -stab inf,,,tb R(plI Ustab) Coherence** [81 p s.t. Vi j, Pi= 0 Z piI

Table 3.1: Level 1 descriptions of some typical QRTs. *Since NO is maximal, this quantifier is unique. **With respect to a particular basis.

of Wigner function. It is worth mentioning that one step further, it has been

shown that contextuality is the actual resource of this magic state model [93, 69].

Similar to entanglement, "magic" can also be quantified by'suitable monotones, and distilled using the idea of quantum codes 119].

4. Coherence.

The QRT of coherence is systematically discussed in [8]. The authors estab-

lished a framework for quantifying quantum coherence essentially by identifying

the set of incoherent states (diagonal in the appointed basis) and operations,

with respect to a certain basis. The 11 -norm of coherence, i.e., the sum of abso- lute values of off-diagonal entries in the density matrix, is suggested as a good quantifier of coherence.

The elements of the above example QRTs are summarized in Table 3.1.1.

3.1.2 Perfect QRT

According to the discussions above, the theory is not necessarily unique when we only specify the resource (Level 1), which often makes related topics quite messy, such as mixed state entanglement. Despite of some practical situations where we have to deal with those technicalities, we would still want to see a cleaner theoretical framework.

This is the purpose of this subsection.

Intuitively, legitimate monotones should vanish for the free set, and the farther a state is from free states, the more valuable it should be. In this sense proper distance measures naturally become candidates of a resource monotone. The quantum relative

68 entropy, which is actually the quantum analog of Kullback-Leibler divergence for two classical probability distributions, becomes a natural choice.

Definition 13 (Relative entropy). The relative entropy of two density operators is defined by S(p lo-) = tr(p log p) - tr(p log o-) (3.1)

By Klein's inequality, this quantity is nonnegative. Actually the relative entropy distance is not strictly a distance measure since it is not necessarily symmetric under party swap, whereas it still only vanishes iff two states are exactly the same. Then we define the regularized relative entropy distance of a state p to a set of states Q as

Definition 14 (Regularized relative entropy distance). The relative entropy distance of a state p to a set of states Q is the minimal relative entropy distance between p and all members of Q:

7(p, Q) = inf S(p||j-). (3.2)

Then the regularization of this quantity simply reads

q'(p, Q) = lim i (3.3) n-+oo n

A general theorem introduced in [471 indicates that the regularized relative entropy distance to the free set, which is defined as following, is the unique monotone, as long as the QRT is reversible. The theorem goes as follows:

Theorem 15 ([471). In nontrivial QRTs (some, but not all states are resourceful), for two arbitrary states p and a with some amount of resource, i.e., outside F, the asymptotic conversion rate is given by the ratio of their respective regularized relative entropy distance to the F:

R(p -+ -) = r, (3.4) 7(O-, )o given that this theory is asymptotically reversible.

Qualitatively speaking, "reversible" simply means that the conversion between resources can be two-way, i.e., with the aid of the set of operations specified by the

69 QRT, as long as p*"f can be converted into o-", U" can be transformed back into p "', without loss. The formal definition is as follows:

Definition 15 (Asymptotically reversible QRT [47]). A QRT is asymptotically reversible iff for any two states p and - we have 0 < R(p -+ -) < 00. and

R(p -÷ -)R(ap) = 1, (3.5) where R(- -) denotes the asymptotic conversion rate defined by Definition 3 in [471.

A direct corollary of this theorem is that

Corollary. The regularized relative entropy distance to the free set F, r; (p, F), is the unique measure of the resourcefulness of a quantum state p in reversible QRTs.

Note that concerning the resource theory of quantum correlations beyond entanglement, 77(p, F) where F is the set of CQ states is shown to be equivalent to one-way QD [48], which indicates that the QRT of QD as the chosen measure has special importance. Remarkably, it is recently shown in [16] that under a similar set of postulates, the maximality of free operation set indicates reversibility, for a QRT:

Theorem 16 ([16]). A well-defined (satisfying the set of postulates suggested in [16]) QRT is asymptotically reversible, if the set of free operations is maximal, i.e., includes all operations that cannot increase the amount of resource.

I shall now formally define a special class of QRTs:

Definition 16 (Perfect QRT). The set of free operations of a perfect QRT is maximal for the resource x, i.e., it contains all possible operations that do not increase x.

It is straightforward to tell that among any Level 1 QRTs there is a single perfect Level 2 QRT. Note that among the examples I provided in that last section (summarized in Table 3.1.1), only the theory of purity is a perfect QRT. The theory of entanglement, for example, is not a perfect one, in the sense that LOCC is known to

70 be strictly smaller than the set of all non-entangling operations, leading to various measures in relation to different conversion tasks. As noted earlier though, the QRT for pure state entanglement satisfies the requirements for a perfect QRT.

It can be shown that perfect QRTs are unique given knowledge on any of the basic elements:

Theorem 17. Knowing any of the three basic elements of a QRT gives you a unique perfect one. More specifically, the characterizationof either of the basic elements A1 , R (the resource) or the partial order of quantum states uniquely determines a perfect QRT.

Proof. Note that F is simply the complement of R, thus the knowledge of either one implies the other. Combining Theorem 16 and the above corollary of Theorem 15, we know that the unique resource monotone is (p, F) for a perfect QRT. Therefore full knowledge about this monotone is equivalent to knowing F or R.

If R is specified, then the monotone is directly given by 71'(p, F), and Af of the corresponding perfect QRT is simply the set that contains all operations A such that

T (A(p), F) <; q' (p, F), (3.6) for all p, which is unique.

If Af is specified, then F contains the states that can be prepared by operations in A 1 , and the monotone is given by r'(p, F).

It would be interesting to find out the relation between the set of operations that cannot create resource out of a free state (denoted as A0 1 ), and the set that cannot generate more resource from any state (A 1 ). Obviously, Af 9 Ao1 since the constraint on Af is stronger. More formally, if any operation A C Af satisfies

Vp, y' (A (p), F) _< r1' (p, F), (3.7) then it must satisfy

Vpo, y' (A(po), F) = 0, (3.8)

71 where po E F. In the context of perfect QRTs, a clean answer for this question wou;d significantly simplify the study of QRTs.

3.1.3 Hierarchical structure

Intuitively speaking, by placing stronger and stronger restrictions on the allowed set of operations, the set of free states (those can be prepared by the free operations) should also become smaller and smaller. Consider two extremes: if almost all operations are allowed, then almost all states are free since they can easily be prepared from anything; if almost no operations are allowed, then almost everything becomes a resource. A more formal statement of perfect QRTs goes as follows:

Theorem 18. Consider two perfect QRTs, whose free sets of states are F and F2 , and the set of allowed operations are Af,1 and AJ,2, respectively. Af, Af,2 implies

F1 C F2 , and vice versa.

Proof. Consider a state p such that p E F1 , i.e., it can be prepared by operations in Af,1. Given that Af,1 g Af,2, p can certainly be prepared by operations in Af,2, indicating that p E F2 . Therefore, there does not exist any state that is free for theory

1 but not free for theory 2, i.e., F1 C F2 . By Theorem 17. the other direction is also true. El

Note that for general QRTs this conclusion does not necessarily hold.

By strengthening the constraints in different directions, hierarchical networks can be established to organize the behaviors of QRTs, in analogy to the "zoo" of computational complexity classes. Fig. (3-2) roughly depicts an example of such a hierarchical structure of operations, coming from the discussion of QRT of quantum correlations in Section 3.2, which I shall come back to later.

Interesting connections can be drawn between map classes and the computational complexity classes. For example, the computational power of Stabilizer (Clifford group) operations is not better than classical Turing machines (corresponding to the complexity GL). But with the aid of magic states as resources (refer to example

72 Local Local unitaries Mixtures of Local unital? C LOCC maps? quantum? (LU) - local unitaries in in in

No Q. Birkhoff Unital CPTP Mixtures of C C CPTP Quantum Unitaries (U) G ua s r-maps maps

Birikoff Doubly C Mixtures of C Stochastic maps Classical Permutations permnutations stochastic maps

Figure 3-2: A hierarchical structure of maps. Columns represent correspondences, and rows represent strict hierarchies.

3 in Section 3.1.1: the QRT of magic), we now have access to universal quantum computation. If we allow less or more operations, analogous connections may be established to other computational complexity classes.

3.1.4 Combining QRTs

To this point, the framework of QRTs has already revealed its essence as a unified mathematical structure to some extent, which guides us to consider the QRTs as abstract mathematical objects. Then a natural question comes into mind: Can we do "algebra" on QRTs? To illustrate the idea, let's ask the following specific question, which is the simplest case:

Suppose we have two perfect QRTs, respectively labeled by free sets F1 and

F2 . The corresponding free sets of operations are respectively Af,1 and Af, 2 .

Now define a new perfect QRT with F3 = F1 U F2. What can be concluded

about the free set of operations Af,3 of this theory?

First of all, by Theorem 17, any element equivalently labels a perfect QRT, so this question is well-defined. We then obtain:

Theorem 19. Consider two perfect QRTs, whose free sets of states are F1 and F2 , and the set of allowed operations are Af,1 and Af, 2, respectively. For a new perfect

QRT with F= F1 U F2 , (Af,1 n Af,2 ) C Af,3 , i.e., the operations that belong to both

73 Af,1 and Af,2 is also in Af,3 .

Proof. Denote the unique resource monotone, the regularized relative entropy distance, as r7' as earlier. A is an operation. The members of Af,1 satisfy

Vp, q' (A (p), F1) < qr (p, F1), (3.9)

while the members of Af, 2 satisfy

Vp, r (A(p), F2) < r (p, F2 ). (3.10)

By inspection,

r (p, F U F2 ) = min{r (p, F1 ), j7 (p, F2 )}, (3.11) 0 0 (A(p), F U F2) = min{r (A(p), F1 ), r (A(p), F2)}. (3.12)

If an operation A satisfies both Eqs. (3.9) and (3.10), i.e., A E (Af,1 n Af, 2 ), the following relation always holds:

Vp, minfr (A(p), F1), ry (A(p), F2)}I min fr/j (p, F1), n' (p, F2)}, (3.13)

By Eqs. (3.11) and (3.12),

Vp, q' (A(p), F U F2) r (p, F U F2 ), (3.14) i.e.,

Vp, r7' (A(p), F3) < r (p, F3 ). (3.15)

That is, A E Af,3 . El

Obviously, this result can be easily extended to the more general case where more

QRTs are involved.

There is another interesting way to think about the combination of QRTs. Con-

74 sider a task where you want to determine a quantum state by querying QRTs, which serve as a black box that can always tell you the resourcefulness of this state, and the state after some known operations. How would you devise the strategy?

Obviously, one single QRT is not enough, simply because generally there can be a family of states possessing the same amount of resource. Then how many QRTs are informationally complete for determining a quantum state? Or even probabilistically?

I am still working on obtaining more general results for this question, but a simple and intuitive example can be given as follows:

Theorem 20. By unitary transformation and the QRTs of coherence and purity as oracles respectively, one can correctly determine an arbitrary qubit p with probability at least 1/2.

Proof. Given a qubit p, one can always find unitaries U (preserves entropy/purity) such that p' = UpUt is incoherent in the appointed basis for the QRT of coherence. If the QRT of purity tell you the purity of p' is 0, then the game is over since p' 1/2, thus the answer is simply p = 1/2. Otherwise if the purity of p' is a (0 < a ; 1), i.e., S(p') = 1 - a, we can obtain a unique solution of the two real eigenvalues m and n such that

-mlogm - nlogn = 1 - a, (3.16)

m + n = 1, (3.17)

m,n > 0. (3.18)

Note that there is no way to distinguish the two possible orderings of m and n in the diagonal matrix p', so p is either

Ut M 0 U (3.19) 0 n or

Ut n 0 U. (3.20) (0 M)

75 U

State space

Figure 3-3: A sketch of a strategy for determining a qubit state that queries the QRTs of coherence and purity. The dashed circle represents the states that has the same entropy (connected by unitary transformations), and the solid line represents the states that are diagonal in the appointed basis (incoherent states). For an arbitrary state p, the unitary U brings it to one of the incoherent states, while preserving purity/entropy.

That is, the QRTs of coherence and purity can "cooperate" to give a strategy for this simple task. A sketch of the idea of this strategy is given in Fig. 3-3. 11

This type of task can be generalized in many ways. For example, the player is only allowed to query a limited number of times, or some error is allowed. This is an interesting topic to explore, since it might provide useful techniques or insights for practical quantum tasks.

3.2 Quantum correlations as a resource

Entanglement has long been recognized as the most important and indispensable quantum resource for computational and cryptographic tasks. Indeed, entanglement

76 seems to play the central role in all sorts of famous quantum protocols, such as quantum teleportation [11] and Shor's factoring algorithm [91]. However, the ultimate resource for all quantum advantages remains a concept that no one can really identify. It is possible to achieve an exponential speedup over classical algorithms using mixed states with vanishingly amount of entanglement [331 in a quantum computational model called DQC1 [581. The resource for this task is recognized as quantum correlations measured by QD. In recent years, quantum correlations have been identified to be the resource responsible for many other useful quantum protocols, including remote state preparation [301, quantum cryptography [811 etc. And the quantity QD has also been bestowed an operational interpretation in terms of entanglement con- sumption in an extended quantum-statemerging protocol [231. However, the theory of quantum correlations is not yet a satisfactory QRT at the moment, since the class of operations that do not increase the amount of quantum correlations (which in general depends on the measure we choose) has not been fully characterized. In this section I introduce some efforts towards the full theory. I denote the whole set of free operations as S, and the set of operations that cannot create quantum correlations from a -y state as So. Obviously, S C So.

Before going into detailed discussions, let's first elucidate the motivation. A well- defined resource theory of quantum correlations will, in the most straightforward

sense, tell us how nonclassicality is created and how to make better use of this special type of correlations beyond entanglement, and therefore provide valuable insights of

the foundations quantum theory. And of course, if such a QRT is developed, it will

be the theory that people refer to for discovering the uses of quantum correlations as

a physical resource in practical quantum information processing tasks.

As the key element of a QRT of quantum correlations, the candidate measures

discussed in Chapter 2 can all be possible resource monotones in relation to different

tasks. However, worth mentioning, the partial order of quantum states determined by

different entropic measures (e.g., QD, minJEP, DD) have to be different, by Theorem

3. For the perfect theory, the measure should be the regularized relative entropy

distance to the free set.

77 3.2.1 Free states

For the theory of quantum correlations, the free states are obviously -y states, as defined in Section 2.1. Note that in this section I do not explicitly distinguish CC and CQ since conclusions about them can be easily generalized to the other.

The properties and interpretations of 7 states were studied in Section 2.1, but here let's review some very important points. First of all, recall the mathematical forms of 'y states. A state with no one-way quantum correlations (without loss of generality, from A) is called a CQ state, which takes the form

PA = Pa&aAaI 0 p, (3.21) a where Ia) is the Schmidt (local spectral) basis for A (block diagonal in A's eigenbasis), and pa is some valid density matrix for B. And analogously we have QC (omitted) and CC states, which takes the form

p = ZBPabja)A(aI 0 |b)B(b = ZPabjab)(ab|, (3.22) ab ab where ZabPab = 1, (abja'b') = 6aa'3 Ikl. The common feature of these states is that there exists some local measurement that doesn't perturb the state at all, or say the post- and premeasurement states are exactly the same.

For constructing a QRT of quantum correlations, a crucial property of the set of 7 states is its nonconvexity, i.e., as shown in Section 2.1.2, this set is not closed under probabilistic mixing. States with positive quantum correlations can be created from 7 states simply by throwing away information. The authors of [161 even list the convexity of the set of free states as one of the postulates that a QRT has to satisfy.

To overcome this difficulty, the operation of throwing away information is deemed to be not free (which makes sense, since it creates entropy). An alternative way to think about this assumption is to attach a classical (orthogonal) flag to each 7 state that we work with. Taking the example from in Section 2.1.2 (Eq. 2.4), which can be thought of as the uniform mixture of four product state: 10+), 11-), 1 - 0) and

78 I + 1). Now we attach an ancilla for each subsystem of these for states. Denote the extra subsystems A' and B' respectively. Let the four ancillae for each side (denoted by 10), 11), 12) and 13)) be mutually orthogonal, and the state after the same mixing reads

1 PAA',BB' = -(100)(001 9 + 0)(+01 + 111)(11 &I - 1)(-1 4 +1 - 2)(-210 102)(021 + I + 3)(+310 113)(131), (3.23) which is now CC, since e.g., for subspace AA', 100), 111), 1 - 2) and I + 3) are mutually orthogonal (same for BB'). That is, the creation of quantum correlations

(the problem of nonconvexity) can be prevented by attaching orthogonal flags.

3.2.2 So

In Section 2.1.2, I showed that quantumly correlated state can be created by LOCC.

A direct implication is that So must be a strict subset of LOCC, i.e.,

So C LOCC, (3.24) since the set of 7 states is a strict subset of separable states, which implies that any operation outside LOCC (entangling operations) can definitely generate quantum correlations. Actually, I have already mentioned a nice observation in Section 2.1 that the "separable-CC" relation is very analogous to the "mixed-pure" relation. By this result, we arrive at the following conclusion:

Theorem 21. Suppose as introduced earlier, we forbid the creation of quantum correlations by mixing (throwing away information). Denote the set So under this assumption as So,g. Then So,g = LOCC ("9" means that mixing is forbidden).

Proof. As discussed, we assume that the probabilistic mixing is prevented by attaching mutually orthogonal ancillae (flags) to all states in the ensemble. An example of the flag-attaching protocol is Eq. (3.23), where it can be seen that the LOCC protocol used to create non-7 states from y states is invalidated. Obviously, LOCC

79 keeps separable states separable, and due to the orthogonal flags (actually analogous to purification), y states are kept -y. Thus by Eq. (3.24),

S0, = LOCC, (3.25) under this flag-attaching protocol. El

Now let's consider the things we can still do if mixing is allowed. More specifically, flags are not attached, and classical communications, which actually allows for global mixing of states using analogous protocols as the coin flip introduced previously, is allowed. (To distinguish with the set of CC states introduced before we denote classical communication as "CComm" from now on.)

In this case, Local Unitaries (LU), which transform local basis, cannot be done arbitrarily because the output states of any operations must be orthonormal to one another. In other words, quantum correlations can simply be created by preparing lots of copies, apply certain LUs on them, and mix them up. I believe this is one of the fundamental reasons why the authors of [16] directly argue that acceptable QRTs should not have a nonconvex free set. In our case, LU certainly should not be able to create any sort of "correlations", but the nonconvexity introduces such a problem.

The set of local channels that can still qualify certainly includes the Semi-Classical

(SC) channel in a fixed basis (LSC), which can be decomposed into local measurements with respect to a fixed basis (local dephasing/decoherence, denoted by LM) plus manipulation on the probability distributions, which can be achieved by generalized depolarizations on certain subspaces (GD).

Definition 17 (Semiclassical (SC) channel Asc [96]). A semiclassical (SC) channel maps arbitrary states onto output states that are diagonal in the same orthonormal basis, i.e., Asc (p) = E PA(P)Ik)(k1. Note that by definition SC channels may alter the uniform probability distribution of a maximally mixed state, i.e., are not necessarily unital. So SC and Unital are essentially two independent sets of maps that have a non-empty intersection.

80 Definition 18 (Generalized Depolarizing (GD) channel AGD; white noise). A quantum depolarizing channel is a model for noise in quantum systems, which maps a quantum state p H-+ Ap + (1 - A)I/d. Note that the CPTP requirement bounds the

2 parameter: A E [-1/(d - 1), 1], but the point is that A can be negative.

Here if Markovianity of the map is considered we can restrict A > 0 (information can only flow outwards) and consider transitions from low to high entropy (high to low purity) states without loss of generality. Then for this case:

So,g - L(NI + D) + CComm = LSC + CComm. (3.26)

But in this case, as mentioned, since CComm allows for mixing, which disqualifies

arbitrary LU due to the nonconvexity of the set of 7 states (which does not really

make sense), we need to be careful in using this result. The basic interpretation is

that the states involved have to be locally classical with respect to the same basis.

For qubit systems the answer is clearer. It is shown in [96] that unital channels

are also not able to create quantum correlations (note that this is not true for higher

dimensions):

Definition 19 (Unital channel AUnitai). A unital channel maps maximally mixed

state to itself, i.e., AUnital(I/d) = I/d.

The rigorous theorem combining results from [96] goes as follows:

Theorem 22. For qubit systems, the maximal set of local channels that cannot create

quantum correlationsis S0 = L(Unital + SC) (a fixed basis is appointedfor SC), while

CComm can be allowed.

Proof. First we show that the set of classically correlated two-qubit states, which

can be written in the form pcc 2 = i=OE=OPijji)A(i 0 Ij)B(ji where (ili') =

6 wi and (jj') = j,, is closed under L(Unital + SC), i.e., the sufficiency. Using Bloch sphere representation, the orthonormal basis of each party can be generically

expressed as |0)(0| = (I + ' - o)/2 and 11)(11 = (I - d - 6)/2, where a E ]R3 is the

81 Unital '1/2

Figure 3-4: Geometrical (Bloch sphere) demonstration of the effect of a unital map on a qubit. The unital channel keeps the two basis vectors symmetric with respect to the center of the Bloch sphere, and the output state can be diagonalized in the orthonormal basis corresponding to the intersections of the connecting line and the surface of the Bloch sphere.

Bloch vector and ' are Pauli group elements. Obviously the local SC channel keeps the reduced density matrix diagonal in some orthonormal basis, i.e., after the action of SC the state is still classically correlated. Now notice that the local unital maps are linear, and by definition they map 1/2 onto itself. So local unital channels map the original basis states onto Aunitai(I0)(0) = (I + d - 6')/2 and Avnitai(I1)(1I) =

(I - S- ')/2 where the elements of a' are mapped from the corresponding elements of

a, which are still symmetric with respect to the center point 1/2 of the Bloch sphere.

Therefore the reduced density matrix of the output state can be diagonalized in the

orthonormal basis, whose two basis vectors are just the pure states corresponding

to the intersections of the line crossing Avnita(10)(0j) and Aunitai(I1)(1), and the

surface of the Bloch sphere (See Fig. 3-4 for a geometrical illustration). Then we

immediately conclude that L(Unital + SC) is the sufficient condition of not creating

quantum correlations. Note that this is not generally true for higher-dimensional

systems.

Next we prove the necessity: any local channel that cannot be decomposed into lo-

cal unital and SC channels (denoted as A,) will create quantum correlations. Nonuni- tality implies A.(I/2) = (I+ b - U)/2 where 3bi $ 0, and the channel being not SC implies ]p, = (I + '. a')/2 such that A.(ps) = (I + d- U)/2 where b and d are lin-

82 early independent, which ensures that An(I/2) and A,(ps) are not diagonal in the same basis. According to the proof of sufficiency, the following state is classically correlated: I + C- I.01 - 67 Pcc 0 10)(01+ - 1)(1I. (3.27) 2 = I 2 2

Now if we put this state through the previously defined channel and define An (C-) = e 6 where ]ei -$0, the output state reads

An (cc 2 ) = I+(bIO)(0+ I+(b e)I 1)(11, (3.28) 2 2 where b - e and b - ' are linearly independent since b and d = b+ ' are linearly independent, implying that An(ficc 2 ) is not classically correlated since 1 + (b+ e) - 5 and 1 + (b-) )- a are not diagonal in the same basis. This completes the proof.

Similar statements can be found in [96]. l

Therefore for two-qubit states

O g = L(Unital + SC) + CComm. (3.29)

As shown in Eq. (3.26), unital channels are not generally allowed for higher dimensions: there exist unital channels that are not in SO,g. Eq. (3.29) is a particular case of So,g that is only valid for qubits.

3.2.3 Promotion to S

The above discussions are about the creation of quantum correlations from -y states, which do not depend on the specific quantifiers. However as the general element for a complete QRT, members of the set S should not be able to increase the amount of resource, as measured by some quantifier m(p), which is generally a stronger requirement than So. Furthermore, the analysis on S will depend on the measure chosen

(denoted as Sm(P) where m is the measure), generally (as the same theory at Level 1,

83 at most one of them can be perfect). Thus the general conclusion is that

Sm(p) C So. (3.30)

For the theory of quantum correlations, it is possible to show that

Theorem 23. For the QRT of quantum correlations, the set S for any non-perfect

theory is strictly contained by that of the perfect theory, i.e., Sm(p) C SO (p,Y).

Proof. The theory with q*(p, 7) as the resource monotone is perfect, i.e., SI(P) C

So (P,-). It can be shown that there must exists operations in SiO (P,) that are not in

Sm(p), as long as m(p) is an entropic measure of quantum correlations different from

7700(p, -Y). By Theorem 3, there must exist two states p1 and P2 such that

m(p1) > m(p 2), (3.31)

'(P, 7) < q* (P2, 7) (3.32)

By Theorem 15, the latter monotone q00(p, -y) has clear practical meanings: the ratio

represents the asymptotic conversion rate (reversible), given that the operation set

is maximal (SNO(P,'Y)) [16]. By Eq. (3.32), there exists a process using operations in

S7 (P,-) that transforms p" to pof2, where ni > n 2 , asymptotically. But for the QRT

with measure m(p), since Eq. (3.31) essentially says that pi is more valuable than P2, the same transition is not possible with operations in Sr(p), which implies that the

above transision process must involve operations that is not in Sm(p). Consequently,

SM(P) C sno(p,,). (3.33)

This theorem indicates that the set of free operations for the QRT of DD (or any

measure that is different from q(p, -)) is not maximal.

84 Now we know that Sm(P) C Snoo (P,) C So. The work left is to identify what operations in So may increase QD for general scenarios and exclude them.

For the case where mixing is still forbidden (by attaching flags to every state), due to the resemblance of to entanglement theory, I conjecture that

Conjecture 3. Under the assumption that mixing is not allowed, Sc = LOCC.

When mixing is not mandatorily banned, the general behavior of quantum correlations under SC channels seems complicated. A simple observation is that the depolarizing (A > 0) the whole space of local subsystems will not increase the amount of quantum correlations under any measure. The complete characterization will be left for future work.

3.2.4 Map zoo

Here I introduce some interesting connections between our results and hierarchies of maps that have been studied before. Under the assumption that mixing is not acceptable, we have the following hierarchy (some detailed results are not completely shown):

LU C Sm(P) C S77(,") C So,c C LUnital C So,g = LOCC. (3.34)

We can immediately notice several interesting analogies of some important classes of maps with quantum and classical map hierarchies, as the structure shown in Fig. 3-2, where the second and third row present the analogy between quantum and classical hierarchies. Here a possible generalization of such hierarchies to local quantum maps shown can be proposed (the first row of Fig. 3-2). According to Eq. (3.34), Sm(P) C

SUto(,-) C So,: lives in between LU and LUnital on this row, and SO,g overlaps with

LOCC. Such intuitive hierarchy structures may lead us to clearer understandings and important insights.

An interesting fact to mention is that in the classical hierarchy, Birkhoff's theorem states that doubly stochastic maps can always be written in terms of mixtures of permutations. However this is not true for their quantum counterparts in the sense that not all unital maps can be decomposed into mixtures of unitaries, i.e., the latter

85 set is a strict subset of the former one [91]. The ideas from the resource theory of purity, which is very closely related to that of quantum correlations, has been used to disprove the asymptotic version of quantum Birkhoff's conjecture, and more deep connections between these topics are expected. To see this, let's expand the hierarchy

in between mixture of unitaries and unital maps shown in Fig. 3-5. The class of maps

Unital

Factonizable

Strongly Factorizable

Exactly Factorizable

U' AQBP

Mixtures of Unitaries

Figure 3-5: (Adapted from [91]) The detailed hierarchy in between mixture of unitaries and unital maps. AQBP denotes "Asymptotic Quantum Birkhoff Property".

called Exactly Factorizable (EF) maps, which is essentially those that can be written

in the following form:

AEF(P) - trB[U(p& )U0], (3.35)

is a special case of NO, which preserves the input space. And obviously, all NOs

are unital. That is, i.e., the free operations of the resource theory of purity locates

between EF maps and unital maps. Certainly, more and more results and insights

86 can be found in this direction.

As a concluding remark of this subsection, this hierarchical network can be extended in analogy to the "complexity zoo", which I name, the "map zoo". This is the land where QRTs find there own place to live. For example, in Fig. 3-2, the entanglement theory lives on the top-right corner, and some important sets of the theory of quantum correlations live on its left. As discussed in Section 3.1.3, there are indeed interesting connections between these two zoos, and hopefully more and more will be established!

3.2.5 Remarks

Another possible way to consider the QRT for quantum correlations is through the combination of QRTs introduced in Section 3.1.4, i.e., the hybrid theory. The authors of [47] pointed out that the QRT of quantum correlations is very closely related to the theory of local purity, and may possibly be the hybrid theory of entanglement and purity. Furthermore, the theory may also be a similar theory to the QRT of coherence

(consider the nature of SC channels). It will be very interesting and helpful to give rigorous mathematical formulations of such hybrid theories.

3.3 Dual QRTs

Due to the complementarity of the elements of QRTs, as has been extensively studied in previous sections, it is natural to expect that the theoretical framework can be symmetrized. Ordinary QRTs specify the set of allowed operations, the corresponding set of resourceful states, and quantify how resourceful they are. In analogy, we can also quantify how useful an operation is, in terms of the capability of this operation to create resource. Based on the above ideas, I shall introduce a structure of theories which is dual to that of the QRTs in this section, as discuss possible interpretations of this new framework. The results presented here are preliminary.

87 3.3.1 General structure

Now I propose the dual (symmetrized) framework of the ordinary QRTs, namely dual

QRTs. In this class of theories, some operations become resources, and the argument of the quantifier becomes operations instead of states. Analogous to ordinary QRTs, the three basic building-blocks of dual QRTs are:

1. Free states.

Denote the set of free states as F. Similarly, it satisfies:

(a) F is closed upon doing tensor product.

(b) F is a closed set.

2. Resourceful operations.

Denote the set of resourceful operations as A,. It should satisfy

(a) A, is the subset of all possible operations.

(b) Under operations that are not in Ar, F is always closed.

3. Quantification of the resourcefulness of operations.

The quantifier is a continuous function of quantum operations (denoted by K(A)) that quantify its resourcefulness. It determines the partial order of operations

with nonzero resource. The properties that , has to satisfy are:

(a) K(p) > 0: always nonnegative.

(b) K(p) = 0 iff p E Af .

Indeed, the quantification of the value of operations/channels seems tricky, since it is

not natural to generally define the conversion between channels. On the other hand, in ordinary QRTs, the basic mathematical process is mapping a quantum state to

another by quantum operations:

pI+p', (3.36)

88 where a quantum channel is indeed the map, as defined. However in dual QRTs, the dual process becomes: A 4 A', (3.37) where a quantum state maps a CP map to another, which is not as natural to give a physical interpretation. However, the mathematical structure is well-defined. There are also some other subtleties, such as the closeness of several sets. Details remain to be worked out in subsquent work.

All in all, the idea about the quantification is that the more resource an operation can possibly generate, the more powerful it is, and a dual QRT serves to quantify the power. Here I suggest the following function as a possible quantifier of the resourcefulness of operations A, of the dual QRT with free set F:

R(A, F) = sup {rf (A(p), F) - r/ (p, F)}. (3.38) P

Indeed, this quantity satisfies the postulates for a legitimate K.

Let's consider the following intuitive example, the dual QRT of entangling operations. Obviously, LOCC is not a resource, since it cannot create entanglement. But for two-qubit states, the CNOT gate is obviously very powerful since it can directly bring a product state (which is not correlated at all) to a maximally entangled Bell state:

CNOT(l+) 0 0)) = 100),Ii) (3.39) where the control qubit is assumed to be the first one. In any reasonable quantification scheme for dual QRT of entanglement, CNOT should be maximally resourceful for two-qubit states. Using the quantifier suggested in Eq. (3.38), the "entangling power" of CNOT is given by R(CNOT, usep) 1. (3.40)

89 Similarly, we have the following hierarchy for the definition of dual QRTs:

Level 1 dual: {-, Ar, }

Level 2 dual: {F, Ar, K

The resource (of operations) labels a Level 1 dual theory, and the three elements together labels a Level 2 dual theory. Then one may naturally ask: what is the dual of a particular QRT? As showed in Section 3.1.2, for each Level 1 QRT there is a single perfect one that can serve as the representative. It can be directly shown that for such a perfect QRT there exists a perfect dual, which can also be the sole representative of Level 1 dual QRTs.

Theorem 24. There always exists a unique dual for a perfect QRT.

Proof. By Theorem 17, a perfect QRT can be uniquely labeled by F or Af. Let F and Ar (the complement of Af) be the labels of the dual theory. Since by definition of the perfect QRT, Af is maximal, i.e., any operation A E Ar is definitely resourceful, in the sense that -p, n' (A(p), F) > ij'(p, F), (3.41)

i.e., R(A) > 0. (3.42)

Therefore, the dual theory is legitimate. That is, the dual of a certain perfect QRT

can be obtained by taking the complements of R and Af as the elements.

In fact, the elements (state and operation sets) of the original and dual perfect

QRTs are exactly complement of each other in the respective space, as illustrated in Fig. 3-6.

This idea of the dual structure of QRTs is actually the underlying mathemati-

cal framework of a diverse set of studies, and may establish interesting connections

between some quite independent subfields.

90 Dual QRT ......

R. QRT ......

States Operations

Figure 3-6: Illustration of the dual QRT. The parts with grey fill means "free". Dashed lines represent the pair of elements in each theory, and arrow represent partial orders, determined by quantifiers. Note that this is only a sketch of the idea of duality. The geometries of sets and the partial orders in each space do not necessarily resemble this figure.

3.3.2 Examples

Although the generalized framework is new, there are already some studies that actually applied the similar idea: quantifying the strength of some specific properties of quantum operations/maps/channels. Here I shall briefly introduce some examples. As claimed at the end of last subsection, this unified framework may provide connections between some quite independent and scattered topics.

Non-Markovianity

Quantum non-Markovianity has been a central theme in the theory of open quantum systems for quite a long time. It essentially characterizes the memory effect in quantum dynamics. Obviously, all the studies on this topic surround two central questions:

1. Characterization: What kind of processes (dynamics) are non-Markovian?

2. Quantification: How do we measure the extent of non-Markovianity?

91 My point is automatically made clear: the whole theory of quantum non-Markovianity

[85] can be formalized using the language from dual QRTs. There are various measures of non-Markovianity based on different ideas, such as the the size of memory space needed, the distance to a convex Markovian set 11051 etc. However, the point I want to emphasize here is that the majority of results on measuring non-Markovianity is based on the idea of "witness": for systems undergoing Markovian dynamics, the witness monotonically decreases as time goes. Therefore if the growth of the witness in the process signals non-Markovianity, and (roughly speaking) the amplitude of such a growth can serve as a measure of how much the dynamics deviates from Markovianity.

Several quantities can be such a witness, including entanglement which leads to the most famous RHP measure [86], QD [51, and MI [66]. Then these QRTs are dual to the theory of non-Markovianity in the sense that non-Markovian process can increase the amount of these quantities. Via the duality of QRTs, the studies of correlations

(Chapter 2) and non-Markovianity can be naturally connected!

Nonclassicality

The resource theory of the nonclassicality of maps is very closely related to the theory of quantum correlations beyond entanglement discussed in Chapter 2, and the QRT of it in Section 3.2.

A recent work [72] characterizes the nonclassicality of CPTP map Q by the extent of its non-commutativity with a complete dephasing channel F:

W(Q) = supS(Qo F(p)||Fo2(p))

= sup(S(FoQoF(p)IFoQ(p)) +S(QoF(p)IIFoQoF(p))). (3.43)

Since the supremum is over all states, the choice of the dephasing basis can be arbitrary. Actually this quantification resembles Eq. (3.38) in many ways. A possible future work is to figure out the exact relation between them. This relation can be used to specify genuinely classical operations Q, among all CPTP maps by W(Qc) = 0. As noted in [72], there is a natural complementarity between quantumness of operations

92 and quantumness of states. At Level 1, this theory is dual to the QRT of quantum correlations beyond entanglement.

93 94 Chapter 4

Summary and outlook

As proposed in the introduction, the ultimate goal of this work is resolving the following two difficulties:

1. Quantum correlations are indeed very difficult to measure.

2. We probably haven't found the right way to understand quantum correlations.

The first part of the thesis introduces some progress towards this goal. There are still lots of interesting and important open problems waiting to be solved under this huge framework. Besides the conjectures and questions already stated in the text, the following problems or messages are of special importance:

1. Study the properties of DD more comprehensively, in order to find out if it

should be recommended as the best overall measure of quantum correlations.

2. Connect DD to real quantum information processing tasks: endow DD opera-

tional interpretation.

3. Establish a satisfactory resource theory of quantum correlations by studying

the S of DD (easy) or the perfect theory (unique).

4. Obtain more general results on the manipulation of abstract QRTs, such as

combination.

5. Establish a "map zoo" (actually, QRT zoo).

95 These questions will be the major targets of future work in this direction.

96 Part II

Exclusion game

97 98 Chapter 5

Introduction

Quantum communication, the area of study that essentially deals with the transfer and manipulation of quantum information, is one of the central topics in the large field of quantum information science. The basic setting of the theory of communication is very simple: two players, usually called Alice and Bob, try to perform some cooperative tasks by exchanging information. For example, Alice and Bob each has a completely random n-bit string, and they would like to know the sum of their strings.

Obviously, the only way to succeed is that one of them send the whole private string to the other. Roughly speaking, all scenarios involving communication can be reduced to some simple form like this, in principle.

The origin of the mathematical foundations of communication can probably be traced back to the 1948 work "A Mathematical Theory of Communication" by Shan- non. Obviously, the playground of this whole theory can be completely classical.

Players have classical information, can they exchange classical bits. However, it turns out that quantum physics can again provide us with surprising extra power, just as for computation. By exchanging quantum states, which encodes quantum information, some tasks can be achieved far more efficiently, e.g., in terms of the amount of information that has to be transferred for any successful protocol, which is measured by the communication complexity of the task.

The notion of quantum communication complexity, which measures the minimum size of quantum message that has to be sent to achieve a certain task, was first in-

99 troduce by Yao in 1979. From Holevo's theorem, we learned that n quantum bits

(qubits) can represent no more than n bits of classical information without entanglement for communication. Even when the two parties share maximal entanglement, n qubits is still only able to carry at most 2n bits of classical information. However, surprisingly, quantum protocols are indeed much more efficient than classical ones, for many communication tasks. For example, there exists some partial functions f with an exponential quantum-classical separation in communication complexity, i.e., as the problem size n goes to infinity, there exists a winning quantum strategy that only requires O(log n) qubits to be sent, while if restricted to classical strategies, at least Q(n) bits of communication is needed [40].

In this part of the thesis, I shall focus on a communication task called the exclusion game, recently introduced by Perry, Jain and Oppenheim [80]. The game, which involves two players Alice and Bob who have infinite computational power, may be described as follows: Alice and Bob randomly draw an n-bit string x and some subset y C [n], where IyI = m, respectively. They win the game if Bob is able to output a string z that is different from x restricted to the bits specified by y, subject to the constraint that the only allowed communication, whether classical or quantum, is from Alice to Bob. Strikingly, the quantum strategy is infinitely better than any classical one for this task, in terms of the amount of information that has to be revealed, namely, the information cost. It can also be shown that although for appropriately chosen parameters of the game, there exists an winning quantum strategy that reveals vanishingly small amount of information as the size of the problem n increases, i.e., the quantum (internal) information cost vanishes in the large n limit, the quantum communication cost (the size of quantum communication to succeed) is lower bounded by Q(log n) for those parameters. That is, there is an infinite gap between quantum communication complexity and quantum information complexity: although for Alice, almost no information of her string needs to be revealed to Bob to succeed the task, she still has to send a very large message via the channel. This infinite gap is further shown to be robust against sufficiently small probability of error: it holds not only for the original zero-error protocols, but also for the generalized case which

100 admits some tiny error. The significance of this result may be compared with its classical counterpart, where only an exponential separation was shown between classical information and communication costs [391. In conclusion, the relations among four important quantities, the quantum/classical information costs and the quantum/classical communication costs, are carefully studied in different regimes of the exclusion game.

This part is organized as follows. In Chapter 6, I shall first formally introduce the generalized mathematical structure of communication tasks and the exclusion game, and rigorously define the informational quantities that I shall repeatedly refer to as preliminaries of the following discussions, and then present the ideas and detailed procedures of a winning quantum strategy for zero-error exclusion game proposed in

[80] with vanishingly small quantum information cost. In Chapter 7, I shall carefully consider the quantum communication complexity of the exclusion game, and prove lower bounds for it in various regimes, to establish the infinite gap between quantum information and communication costs. At last, Chapter 8 contains some concluding remarks and interesting open problems.

101 102 Chapter 6

Preliminaries

In this chapter, I shall introduce the formal mathematical definition of the exclusion game, and the winning quantum strategy proposed by Perry, Jain and Oppenheim with vanishingly small quantum information cost [801. These results are the premises of the infinite gap between quantum information and communication costs that we shall establish later.

6.1 General formulation of communication tasks

The goal of usual communication tasks is to compute a function (which can be total or partial) of Alice's and Bob's private strings. However, the output of the exclusion game is not a "function" of these two strings, basically because there can be multiple acceptable answers. In this section I shall generalize the traditional mathematical formulation of communication tasks, and use it to formally define the exclusion game.

6.1.1 Mathematical structure

The task of a typical one-way communication is to compute such a function f depending on both players' private strings, by transferring information from Alice to Bob: f : {0, 1}a x {o, 1}b -+ {0, 1}. (6.1)

103 The task may be described as follows: Alice draws some string x E {0, I}a and Bob draws some string y E {0, 1}'. Alice sends a message to Bob and then Bob outputs a string z. They win the game if z = f(x, y).

The exclusion game introduced in [801 requires a more general framework, which we introduce here. We shall describe a general one-way communication task by the function

F : {0, I}a x {0, I}b -+ p({0, }*), (6.2) where P(S) denotes the power set of S. As before, Alice and Bob draw some strings

E {0, I}a and y E {0, 1}b respectively. Then, Alice sends a message to Bob and Bob outputs a string z. The winning condition is made more general though. We say they win the game if z E F(x, y).

It is clear that the more general framework reduces to the typical framework when

F(x, y) contains exactly one element, say f(x, y) . In this case, the only way in which they can win the game is if z = f(x, y).

6.1.2 Exclusion game

We shall now define the exclusion game. Let

M : y~nm X {0, 1}" -+0, 1}m

(y, X) MM" W)

where Y(n,m) is the set of all subsets of {1,... , n} of size m, and My(x) is the m-bit

string formed by restricting the string x to the bits specified by y. We also add another

parameter -y to label the allowed probability of error. For the original exclusion game,

= 0. The exclusion game EXCn,r is then defined as

EXCn,my : {0, 1}s(m) x {0, i} -+ {0, 1} m

(Lyl,x) {zlz # My(x)},

104 where -y represents the allowed probability of error, LyJ is the binary representation of y, and s(m) is the number of bits needed to specify a subset y of size m. The winning condition may then be stated as follows: Alice and Bob win if for given x

and y, Bob outputs a string z such that z = My(x).

6.2 Information and communication

There are two informational quantities that are very important for the communication

task, namely the information cost and the communication cost. The information cost

of a protocol measures the amount of classical or quantum information that is actually

revealed, whereas the communication cost measures the number of bits or qubits that

has to be exchanged in order for the protocol to succeed.

We now formally define the communication complexity and information complex-

ity for communication tasks. For a protocol H that wins the game defined by the

function F, we denote the information cost of a A-protocol (where A = C (clas-

sical) or = Q (quantum)) by ACC(I), and the corresponding communication cost

by R F(H). Then A(C(H) is defined to be the number of bits or qubits exchanged

throughout the protocol, and A'(H) is defined as follows:

AC(1) = I(X : 1lY) + I(Y : 1IX), (6.3)

where I(S : TIU) = H(SU) + H(TU) - H(STU) - H(U) (6.4)

measures the mutual information between S and T given U.

The A-information complexity of a game is then defined to be

A c(F) = inf AFc(f). (6.5) 1 ri im

The A-communication complexity is defined similarly.

105 6.3 Classical communication complexity

For the regime of the exclusion game defined in 6.1.2 that is relevant for later discussions, it can be shown that the communication cost of any winning classical strtategy is lower bounded by Q(n), i.e., the number of bits that Alice has to send to Bob for a classical protocol to work scales linearly as n, asymptotically. The rigorous statement goes as follows:

Theorem 25. Let w(fIn) K m < an, where 0 < a < . is a constant. Any classical

strategy that wins the exclusion game EXCn, in the large n limit with certainty, requires that the number of bits sent from Alice to Bob be of order Q(n), i.e., for all

strategies H, CC n",n (H) G Q(n) where m is within the above regime.

Proof. In Theorem 2 of [801, it was shown that ICn'(H) > n - log 2 (EiL ())- For w(n2) m < an, where 0 < a < -, the lower bound could be simplified

to C C '"'(H) E Q(n) (see Appendix B2 of [80]). Since the amount of information

revealed is bounded above by the communication cost, i.e., CIc 5 Ccc for any

communication protocol, it follows that Ccc E Q(n). As Alice can always send the EXCn~ whole string to Bob in order to win, this bound is tight, and we obtain CE ''')m i)E E Q(n) in the specified regime of m, as desired. El

6.4 PJO strategy

Perry, Jain and Oppenheim (PJO) shows the first example of a communication task

for which there is an infinite separation in information costs between classical and

quantum strategies by devising a winning quantum protocol for a certain regime of

the exclusion game, such that its information cost tends to zero in the large n limit, thus upper bounding the information complexity, and prove a Q(n) lower bound for

its classical counterpart [80]. Now we introduce this PJO strategy.

106 6.4.1 Protocol

The PJO strategy runs as follows. For given strings x and Lyi that Alice and Bob, respectively, draw, we first describe how Alice encodes the string x in her quantum

message: she encodes each classical bit xi using the state

I4'xi(0)) = cos ( 0) + (-1)X sin (0) 1i), (6.6)

where 6 = 2 tan-'(2 1/m - 1). Her n-bit string Y = x.- is then encoded as the joint state r kbi (0)), (6.7) IIy(o)) = 1 i=1 which she sends to Bob via the quantum channel.

Upon receiving the state from Alice, Bob is now able to perform a global measure-

ment whose outcome indicates the string z he is going to output. The measurement

technique used may be described as a conclusive-exclusion measurement, which was

first introduced by [24] and subsequently used to prove the PBR theorem [83], a re-

sult in the field of quantum foundations that rules out a certain class of V)-epistemic

models of quantum mechanics. In [80, 71, it was shown that if Bob performs the

projective measurement {(z)}zE{o,im on the subspace designated by y where

1(Z) 1 (10) - (-1)z' Is) . (6.8)

It is easy to see that

(X)(0) (X)= 0, (6.9)

i.e., the probability that the indicated output z = My(x) is zero, and hence, by

outputting the result corresponding to the projection |(2), they win the game with

certainty.

107 6.4.2 Quantum information cost

As shown in [801, this protocol exhibits a striking property: asymptotically (in the limit of large m), the amount of information it actually reveals (the internal information cost) tends to zero as n increases in the m E w( l/-) regime, while the players are always guaranteed to accomplish the task. Indeed, the quantum information can be calculated to be given by Q1 C "'(H) < 2S(MQ) E o(nm- 2 ), where S(MQ) E o(n-i-2) is the von Neumann entropy of the quantum message MQ that

Alice sends to Bob. Hence, it vanishes in the large n limit, when m E w(fd).

108 Chapter 7

Quantum communication complexity

The PJO protocol, however, requires that n qubits be sent from Alice to Bob. There- fore, even though the information cost becomes vanishingly small with increasing n, the communication cost of the protocol grows with n. An interesting question that then arises is whether it is possible to succeed in the game with a smaller communication cost, i.e. can Alice and Bob still succeed if Alice compresses her quantum message? In particular, is it possible for the message to be compressed so that the communication cost also goes to zero with increasing n? In this chapter, we answer this question in the negative. It can actually be shown that for certain parameters of the exclusion game, the quantum communication complexity is at least Q(log n), i.e., a Q(logn)-size quantum message is always necessary to accomplish the task, when no error is allowed. Then one may naturally wonder whether the results are robust against error. We then present a variant of the zero-error protocol that naturally fits into this picture. As it can be shown, the lower bound and thus the infinite gap is robust against sufficiently small error.

7.1 Zero error

In this section we explicitly present the proof of the lower bound of Qcc for the zero- error exclusion game EXCn,m,o where m E o(n). The main idea of the proof is to devise a classical protocol that approximately simulates the imaginary winning quantum

109 protocol. We show that if the size of quantum communication in the winning quantum protocol is too small, then we can also have a classical strategy with communication cost less that linear, thus contradicting the classical lower bound (Theorem 25 in

Section 6.3).

7.1.1 Classical encodings of quantum states

Before proving the main theorem, we shall introduce the following

Lemma 26. A q-qubit quantum state can be classically described by a set of real numbers encoding the real and imaginary parts of all amplitudes to accuracy e using o 2 qlog(1/c)) bits.

Proof. Generically, a q-qubit pure state IQ) can be written as I'Oq) = Eqi aji), where a E C, and {Ii)} is a complete orthonormal basis set containing 2 q elements. We express all complex amplitudes as ai = bi + ici where bi, c, E R, satisfying

2 ZIil JaiI = f 1 (bi + ci) = 1. Thus, 0 < Ibil, Icil < 1. To approximate each of these real numbers to accuracy e = 2-, we keep the first r bits after the binary point, and use one extra bit to indicate its sign, i.e., we can find an (r+1)-bit classical string that encodes an approximation bi of each bi such that for all i,

Abi = |bi - bil < e,

Aci = Jai - cil < e.

Notice that there are 2 -2 q such numbers in total, thus only 2q+l(r+1) = 0 (2 q log(1/e)) bits are needed to encode I bq) such that we have specified the real and imaginary parts of all amplitudes to accuracy e.

This lemma essentially says that generally, a quantum state can be approximately

encoded by exponentially many classical bits by registering its amplitudes to some

predetermined accuracy. This result play an important role in the classical simulation

protocol, thus the proof of the main theorem as well.

110 7.1.2 Lower bound of Qcc

Now we prove the main theorem: any winning quantum strategy for a certain regime of the zero-error exclusion game requires at least Q(log n) qubits to be sent.

Theorem 27. QCC(EXCn,m,O) ;> (log n), where m E o(n). That is, the size of quantum communication in any winning protocol for the specified regime of the exclusion game is at least Q(logn).

Proof. Suppose that for the specified regime of the exclusion game, there exists a winning quantum strategy HQ such that Qcc(Hq) = q = o(log n). Then based on

HQ, we can devise a classical strategy Etc with o(n) bits of communication, which would contradict Lemma 1 (classical lower bound), therefore negating the existence of HQ.

Most generally, HQ can be divided into three steps: i) based on her n-bit string x, Alice prepares a quantum message lox) of size o(logn) (defined by log JRI, R being the Hilbert space); ii) Alice sends I4x) to Bob; iii) Bob feeds |ox) into his local quantum computation, and obtains an m-bit string z such that z 7 My(x) according to the output (measurement outcome). (Without loss of generality, we can assume that the quantum message is pure, since any mixed state can always be purified by an ancilla space of the same dimension: the asymptotic scaling of Qcc is not affected. In addition, both players agree on a fixed basis for the matrix representation of operators and amplitudes of state vectors beforehand.

The essence of Hc is to classically simulate all steps of HQ. The basic procedure is the following. First, Alice prepares a classical message C(I'x)) that describes l)=x) = zl7I ajIi) = E (bi + icj)Ij) ({ij)} is the appointed basis.), by registering the real (bj) and imaginary parts (cj) of all amplitudes (as) to some desired accuracy i (we denote the approximations by b and Ej), and then send it to Bob. Note that the size of C(oI)), i.e., the communication cost of 11c, depends on the desired accuracy of the description. (For example, the message needs to be infinitely long to achieve arbitrary precision.) In IIQ, Bob's local strategy can always be modeled as a completely positive trace-preserving (CPTP) map followed by a generalized measurement, which

111 is altogether equivalent to some POVM. Note that although Bob can use an infinite amount of ancillas for his local quantum computation, i.e., the original POVM {Pi} may contain an arbitrary number of elements, there are only 2 "' possible strings that he can eventually output: g(P) = z, where z is an m-bit string. We can therefore combine all Pi's corresponding to the same z as an element P' of a new POVM set {P'} by

P' = (7.2) g(P)=z or in the continuum limit where the elements are labeled by a continuous variable p,

P' = ( d pP([t). (7.3) Jg(PGL))=Z

Due to the convexity of the set of all non-negative Hermitian operators (valid POVM elements), {P'} forms a discrete effective POVM with 2' elements labeled by z.

A subtlety here is that the amplitude vector encoded in C(1,,)) is not necessarily normalized, i.e., the probabilities do not necessarily sum to one. Bob first normalizes the amplitude vector by dividing each component with the 2-norm v

(EyW1 (bj + 2)) 1/2, and then applies Born's rule to compute the approximate probability of obtaining each z:

P 2>3 (bjbk - iajbk + ijk+ ajak)- Jk, (7.4) j,k=l

where p,, is the (j, k)-th entry of P'. The requirement that HQ never fails indicates

that the probability of outputting the POVM elements corresponding to a wrong

answer is exactly zero. As shown in Lemma 29, the whole approximate distribution

{p'} remains very close to the true one (denoted by {pz}) when a is small, so the

classically calculated probability corresponding to the wrong answer My(x) is well

bounded. Therefore Bob simply sets an appropriate threshold value J(i) that p'

cannot exceed, and refuses to output any z with p'

answer above this threshold, the protocol is guaranteed to be successful.

112 Finally, we determine the values of K and 6 in the above protocol 1c. In order to guarantee that there always exists a valid output, it is sufficient that the upper bound of perturbation on all z's, 6, satisfies

6 sup 1p, - p'I < 2-. (7.5) z

Then we can simply set the threshold value

6 = 2-, (7.6) i.e., Bob only outputs a z with p' > 2-' (it always exists). By Lemma 29, 6 < 202/2 Then according to Eq. (7.6), we can set

= 2-(m+q/2) (77) 20 so that 6 < 6. In summary, Hc runs as introduced with the e and 6 respectively specified by Eqs. (7.7) and (7.6).

By Lemma 26, Ccc(Flc) with the above accuracy scales as 0 ((m + q/2)2q) = O(m2q) = Q(n2q) since m = E(na) where 1/2 < a < 1 and q = o(logn). Since

2q = o(nfi) for any positive constant /, we simply set 3 = 1 - a, and it can be directly seen that Ccc(flc) = o(n+O) = o(n). Hence, we have reached a contradiction by constructing a classical protocol with o(n) communication cost for the specified regime of EXC.n,m,O, where the classical communication complexity is Q(n). 0

7.1.3 Gaps

Based on this lower bound on quantum communication cost, we can directly arrive at the following gaps:

Corollary. For EXC,m,o with m C e(na) where 1/2 < a < 1, the gap between

2 2 1. Qic and QcC: O(nm- ) vs. Q(logn). Note that O(nm- ) = o(1), i.e., Qcc is infinitely larger than Qic;

113 Q(n) C cc CC Q

Q(log n) QCC Q/C -+ 0

Figure 7-1: Exclusion game EXC., where m E 0(0), 1/2 < a < 1 in the large n limit. Solid arrows indicate established separations (pointing towards the smaller one), while the dashed one indicates an unknown separation (at most exponential).

2. QcC and Ccc: Q(log n) vs. Q(n), i.e., at most exponential.

The first statement implies the "infinite" separation between quantum information and communication costs. In contrast to the infinite separation of quantum and classical information costs, the second statement upper bounds the quantum-classical separation of communication costs to be exponential. The results about the above separations are illustrated in Fig. 7-1.

7.2 Robustness against error

In this section, we investigate if the above results still hold when some probability of error is allowed, using a slightly different protocol.

7.2.1 Lemmas

Same as the last section, we introduce some useful lemmas before going into details of the main argument. We first rigorously prove a intuitive conclusion: Suppose that the error allowed in the classical encoding of a quantum state (as introduced in the last section) is very small. Then if we use the approximate amplitudes registered in this classical encoding to reconstruct a quantum state, this state is very close to the original one. The rigorous statement goes as follows:

114 Lemma 28. W is a Hilbert space of dimension IR| = 1, with orthonormal basis aIi) =Z_ 1(bj +icj)lj). Suppose{ll), we have --- , l 4') E)}.W with 14') = I {bj,6 } such that Vj, |bj -b|, cj - ajI E < (6V 1)-1 (which is always satisfied in this

paper), and let 1) = E. 1 &ji) = Z_ 1 (b+iaj)|j)/vwhere v (Z 1_(bl +k ) 2 is the normalizationfactor, then D(I|V), 4')) < 10/kE, D(, ) being the trace distance.

Proof. We first deal with the normalization factor:

v2 Z1: Z +

By Cauchy-Schwarz inequality, we have

(b1 + Icjl) 21, (7.9)

and 21c2 < 2l1e = V/le, so V2 < 1 + 3-/21E. (7.10)

Similarly,

1 - 2v/2IE < v 2. (7.11)

Since

1 > 1 - 3v-21 > 1 - 3/2Ie, (7.12) 1 + 3V2,e and 1 < 1 + 3v/2cl < 1 + 3/Wcl, (7.13) 1 - 2v 2lf we have

1 - 3V ik< - < 1 + 3\/21E. (7.14)

Assuming bj > 0, then

(bj - c)(1 - 3v/2ie) < < (b + E)(1 + 3v/ie) if bj-e> 0, (7.15)

(bj - e)(1 + 3v/1e) < < (b + E)(1I + 3v 2e) if bj - e < 0. (7.16) ii

115 For both cases,

(bj + c)(1 + 3v/216) = bj + (1 + 3x/ilb)E + 3V/21E 2 < bj + (2 + 3v-ilb)E. (7.17)

For b. - E > 0,

(bj - E)(1 - 3v/21E) > bj - (I + 3v 2lbj)E. (7.18)

For b3 - E < 0,

(bj - E)(1 + 3v/il) = bj - (1- 3Vi/bj)E - 3V\lE 2 > bj - (2 + 3v/lbj)E. (7.19)

So if b3 > 0,

bj - b < (2 + 3v 2lbj)E. (7.20)

Similarly, if bj < 0, LI

bj - b < (2 - 3v'ilbj)E. (7.21)

So we get

bj - < (2+ 3v'/iibl)E. (7.22)

Similarly, vI Ci - C < (2 + 3v/1cj)E (7.23)

Recall that E) i &1j), where 6d = (b+ ij)/v. Then

|aj - aj|

=bj + ic, - -

; bK - - c -

< (4 + 3V-i(bj| + cjI))E (7.24)

116 Therefore

1 - ( P4)

= (1- () 1+ I( | ) < 2 (1- ({))

< 2 - {@|1 ) - @

|a>j|2 + law 2 - aejd* -

|aj - d |2

< (4+ 3(IbI +Icj1)) 2 62, (7.25)

2 2 where (4 + 3v'21(lbjl + Ici)) = 16 + 24V 1b(I b3 + c) + 18l(bjI + icj ) . With

2 2 2 (Ibjl + cj) < 2(Ibj1 + c3! ) and Eq. (7.9),

12 1 - (#kb) 2 2 2 < E(16 + 242(lbj| + Icj) + 36l(lbjl + Ic31 ))E < (161 + 481 + 361)e 2 100lE 2 . (7.26)

Then

D(@), 4')) = 1 - ('12) < 10VJE. (7.27)

Conclusively, D(jO), 1 )) < 10vlc-. E

Then we can directly argue that the probability distribution of the outcomes of a measurement will be very close to the original one as well:

Lemma 29. Let {Pk} be a POVM, with Pk = (0|PI|'), Pk = (IPkI|). Then

IPk - Pk| < 20v/1i

Proof. By Theorem 9.1 in [76], we directly obtain IP - PkI k IP - Pkj < 2D(10),| 1)) < 2OEV, where the last step came from Lemma 28. 0

117 7.2.2 Quantum-classical separation of communication

Now we show that for exclusion games where the allowed probability of error is non-trivial, the quantum-classical separation of communication cost is at most exponential.

Theorem 30. Suppose that for EXCn,m,, where -y < 2-(+1) and m = Q(poly(n)), there exists a winning quantum strategy H' such that Qcc(fV) = O(s). Then one can construct a classical strategy HlO+ such that Ccc(71?) = 0((m, + s/2)2s), whose probability of error can be made arbitrarilysmall.

Proof. As can be shown, only one bit of classical communication is needed for m >

n,7-y= 2-(m+), so we shall only pay attention to the regime of even smaller probability of error, where the scalings of communication complexities may be interesting.

We revise Bob's local part of the protocol presented in Theorem 27 to devise this

Hl+ as follows. As for the zero error game, Alice prepares an 0 (2' log(1/iy)) classical message that approximately registers the real and imaginary parts of all amplitudes of the quantum message I'P) in H' to accuracy -, and sends it to Bob, who then normalizes the amplitude vector. However, instead of calculating the probability distribution of the output as in Hc, Bob simply prepares a new quantum state P'D according to the normalized state vector (by Lemma 28, this state remains close to the original one when e is small), and then feeds it into his original local quantum computation. By Lemma 29, the probability of outputting the wrong answer

pM(x) < - + 20 28/2. (7.28)

As long as M,(x) is not the winning output, i.e.,

pM ,(x) < 2, (7.29)

Bob can apply amplitude amplification to suppress the error rate: he simply repeats

'It seems true from numerical results that for m = o(poly(n)), an 0(1) size of classical communication cannot guarantee an error rate of smaller than 2 -'. For now I conjecture that for -Y= 2-(m+1) and m = )(poly(n)), Ccc = 0(1).

118 his local protocol for t times (he can use the classical message to prepare as many copies of 1I) as he wants), and outputs the string z that comes out for most times.

We denote the error rate after the whole procedure by -I', then by Chernoff bound, for any r > 0, there exists a i such that as long as t > f, -' < r. That is, -y' can be made arbitrarily small simply by increasing t. Combining Eqs. (7.28) and (7.29), we can set 2- = 7_2-/2 (7.30) 20 in the protocol. Note that the numerator is lower bounded by 2 -(m+1), thus 2 = O(2-(m+1+s/2)) = O(2-(m+s/2)) (we only care about s E w(1)). Therefore the communication cost of I1'+ scales as O((m + s/2)2s). Note that the no-cloning theorem is not violated since Bob does not need to copy quantum states, and we do not care about the scaling of t since local computational resource is not limited. l

This theorem enables us to resolve the following natural question: Is the same infinite Qcc - Qic gap robust against some tiny error perturbation? The answer is yes, as long as the allowed error is sufficiently small.

7.2.3 Maximum error

Note that the above arguments heavily depend on the lower bound of classical communication cost. Intuitively, as the allowed probability of error grows, the necessary size of communication decreases. Then the question is: what is the maximal error scaling that still requires a linear classical communication complexity? As mentioned in the proof of Theorem 30, only one bit of classical communication is needed for m ;> Vn, y = 2 -(0+1), so this maximum error is expected to be even much smaller than 2 -(m+1). Indeed, the answer is given in the following theorem:

Theorem 31. Occ(EXCnm,) = Q(n), where m = 0(n0 ), 1/2 < a < 1, and y < (n + I)~-.

Combining Theorems 30 and 31, we obtain the following

119 Corollary. The same infinite gap (O(nm-2 ) vs. Q(log n)) between quantum information and communication still holds for EXCn,n,y where m E O(na), 1/2 < < 1, and -y < (n + 1)-".

Theorem 30 also indicates that the QcC - Ccc gap is at most Q(s) vs. O((m + s/ 2 ) 2s). In the zero error case, we have ruled out the possibility of super-exponential

Qcc - Ccc gap in a certain regime. When some error is allowed, when does the same conclusion holds? The answer is given in the following corollary:

Corollary. When m < O(2POIY(s)), the gap between Qcc and Ccc is at most exponential.

Note that the complete information of the behaviors of these gaps is encoded in a 3-dimensional (Qic/Qcc/Cic/Cc-m-y) space, but some cross-sections (Fig. 7-

1 being one) may be good enough to illustrate the important information in this diagram.

120 Chapter 8

Concluding remarks

In this work, we considered the exclusion game and showed that an infinite gap exists between its quantum communication complexity and quantum information complexity, whether we allow for zero error or a sufficiently small error. This work actually deals with a very new regime of quantum communication: the error is extremely tiny, and there are infinite separations. As expected, lots of interesting questions around this topic remains open, including

1. Is the Q (log n) lower bound tight? Or in other words, is it possible to compress

the quantum message for a winning protocol?

2. The current proof of the lower bound fails for m E 0(n), i.e., a beyond exponen-

tial gap between Qcc and Ccc is not ruled out for this regime. What makes

this regime so special? Does the lower bound indeed fail for this regime?

For the first question, we note that interesting connections between this game and

Bell nonlocality may be drawn, if there indeed is a gap between Qcc and Ccc [20.

Recently we also realized that a quadratic compression is possible while keeping the error small enough, though it is still much weaker than the desired exponential compression.

The general formulation of communication tasks presented in this work may also open up some new doors. Analogously, the task of quantum state discrimination,

121 where we essentially reduce k possibilities to 1, can be generalized to k -+ (k - m) state exclusion task via semidefinite programming (SDP), which may provide insight for the generalization of communication models. For example, a duality between quantum random access coding (QRAC) 174, 6] and the exclusion game under certain

restrictions may be formalized in this spirit. We expect interesting new problems and results to emerge in this direction.

122 Appendix A

QD and DD of real X states

Here I explicitly show the calculation of QD and DD for a subset of real X states with the aid of results on the optimal measurement basis in [27].

A.1 Preparations

As explained in the main text, I shall only focus on X states with real entries:

goo 0 0 003

Pu 0 Q11 012 0 (A.1) 0 012 Q22 0

03 0 0 033

123 which has five degrees of freedom in total. Thus we can non-redundantly define the five free parameters in terms of spin correlation functions:

G = (of) = tr(ozpij) = ooo + pii - 922 - 233, (A.2)

Gz = (os) tr(ojpij) = oo - On + 922 - 933, (A.3)

G-T = (or j) = tr[(o- 0 ojT)pi] = 2(012 + 003), (A.4)

G = = tr[(ou' 0 u )pij] = 2(012 - 003), (A.5)

Gzj = (ozo,) = tr[(ou D cx)pij] =oo - L1 - 022 + 033, (A.6) with all of which ranging in [-1,11. Here G = (or) (1 i, j) and G = (oc q)

(a, #3 x, y, z) denote the magnetization density at site 1 and two-site spin correlation function of sites i, j, respectively. More importantly, pij can be naturally decomposed as

1 Pij= ( I1 + Gxxoy 0 ,x, + Gioi 0 oTY + Gzcxf 0 aJ + Gza I 0 I+GzIou). (A.7)

This decomposition is very useful in various contexts.

In quantum spin models, the matrix elements of pij can be expressed in terms of spin correlation functions [89] as follows:

goo = 4(1 + Gz + Gz + GZZ), (A.8)

on = (1 + Gf - Gz - Gf), (A.9)

922 = '(1 - Gz + G - GZZ), (A.10) 033 = 1(1 - Gz - Gz + G-7), (A.11)

012=t 2 = (Gxx + GYY), (A.12)

003 0>3 = (Gfx - GY7). (A.13)

124 Then the eigenvalues of p 3 can be expressed as

A = (1 + G-J + Gg - Gff), (A.14)

A2 - Gx - GY - Gf), (A.15)

3= (1 + G + V4(Gf)2 + (G - G)2), (A.16)

A4 = j(+ Gff - 4(Gf)2 +-G G)) (A.17)

A.2 Pairwise quantum correlation

Here we present the calculations of two-site QD and DD in Z2-symmetric quantum spin lattices in terms of pairwise correlation functions. As the scaling laws of these

correlation functions in a number of spin models have already been widely studied, the results in this section can serve as powerful tools to analyse the scaling behaviors

of quantum and generic correlation measures.

A.2.1 Optimal basis

Evaluating QD, even for very simple cases, is extremely hard. Analytical results can

be established for very limited cases. Generally computing the exact value of QD has

been shown to be NP-complete [53]. For X states we have some approximate results

[4, 37], yet the accurate analytical formula is still unknown [27, 521. However, we still have the following useful conclusions that can help us compute the one-way QD of

some regimes of X states of general interest:

Lemma 32 (optimal local measurement for real X states [27]). The local measure-

ment on one subsystem (e.g., without loss of generality, B) that minimizes the quan-

tum conditional entropy of postmeasurement bipartiteX states S(AfB), i.e., gives the

value of QD, is (i) ox, i.e., with respect to local projectors {I+)(+I,|-)(-t} where {I+),I-)} form the eigenbasis of ux if

I PooQ33 - Ve11L221 IL0121 + 10031, (A.18)

125 or (ii) or, i.e., with respect to local projectors {IO)(0|, |1)(1|} where {I0), I1)} (computational basis) is the eigenbasis of o7z if

(0121 + 1,031) < (coo - ell)(933 - 022). (A.19)

The basic idea for the proof is to parametrize the general two-qubit POVM {E'}

4 as { + nk1Ok'B) } , where k k k 1, and k kk 0, or similarly parametrize the von Neumann measurement (as will be shown in the next subsection), and the value of S(Alb) (denoted as SB(!AB) in some literatures) turns out to be a concave function whose minimum is located on the boundary. Note that we shall assume IGTfl ;> IGYY (expressions in terms of matrix elements shown by Eq. (A.4) and (A.5)) without loss of generality, since we can always switch the signs of the involved entries via a local unitary transformation, in which case (i) and (ii) have covered most possibilities, and in addition, even if we adopt uo or oi as the optimal measurement for all X states there is shown to be only a very small error for very few cases numerically. Using o or o- as the optimal measurement suffices for our analysis.

A.2.2 Explicit calculation of pairwise QD

Now I evaluate QD for the above regimes of real X states with the aid of above results. Here I adopt the difference of quantum conditional entropy between pre- and postmeasurement states as the mathematical form of QD and use the conclusions

of optimal measurements to explicitly express pairwise QD in terms of correlation functions.

First from the eigenvalues given in Eq. (A.14)-(A.17), we can easily obtain the von Neumann entropy of the joint state S(ij). Without loss of generality, assume

local measurement is done on j. Taking advantage of the decomposition Eq. (A.7), we simply trace out subsystem i and obtain the reduced density matrix

p3 =(I, + vof) = (Ij + Gio), (A.20)

126 with eigenvalues (1 Gf)/2. Hence the von Neumann entropy of the subsystem is

1 + Gz 1+Gz 1-Gz 1-Gf So) = 2 2 2 2log2 (A.21) 2 2 2 2

So, the quantum conditional entropy of the original two-site state is given as

41 +Gz 1 -Gz S(ilj) = S(ij)-S(j) = - A. log Ac+ 2G log (1 + Gz)+ 2G' log (1 - Gz)-1, a=1 (A.22) with {Aa} given in Eq. (A.14)-(A.17).

Next, according to Lemma 32 and the immediate discussions, there are two classes of postmeasurement states obtained by corresponding optimal measurement bases (i)

{I+)(+I, -)(--I} or (ii) {I0)(0I, 11)(11}. Now we compute the quantum conditional entropies of these two classes, and hence QDs, respectively:

Class (i): {|+)(+I,|-)(-I} as the optimal measurement basis. For pij in this class, the local measurement operation on the original bipartite density matrix is (Ii 0

I )( I), with two possible outcomes, whose corresponding postmeasurement density matrices are given by

pZ3 = (Ii 0& +)(+I)pij

=Ij 91+) (+ I + Gi o +)(+|

1 -1 + Gyyuiy 0 + -G.-C 2 " (1 -1 . 2 ' 1 -1). -1@ (.1 Gzorz 0 1+)(*Ij + 00GjIi --1 , (A.23) 2 1

127 p" (I 0 I-) (-j)pij

+ Gyoy (9 + G.. -1. 2 "' (1 1 ) 1:1 22 1 1i) ,] (A.24) . respectively, with equal probabilities pi = P2 = 1/2. Hence the reduced density matrices after tracing out j read

pi,+ = 2trjp _ = !(I + G oa + Gzu-), (A.25)

pi,_ = 2trjp.:,+ = !(I- G-ox + Gfoz), (A.26) whose spectra are the same:

1 (G-f) 2 + (Gz) 2 A+,= = 2 (A.27) 2

The prefactors 2 come from probabilities on the denominators. So we obtain the quantum conditional entropy of the postmeasurement (optimal) state as

S(ilj) = -A+,+ log A+,+ - A+,- log A+,-, (A.28) which finally gives us the value of QD (for the first class we present the full expression

in terms of eigenvalues):

D(j -+ i)

= -A+,+ log A+,+ - A+,- log A+,-

+ A, log Aa +2G log (I+ Gz) - 12Gi log (1 - Gf) + 1, (A.29) Cf=1

where the A's are given earlier.

128 Class (ii): { 10) (0 1, 11) (1} as the optimal measurement basis. As the calculation pro-

cedures are similar as (i), I directly introduce some important results here. The

postmeasurement density matrices are

0 1 0 -i pig~ - [Is 0 |)(0|i +G of T@ + Gy~or i Pi3'O400 . 3 ( 0 0 +(Gzz + Gz)o 0 10)(01j + GzI 0 10)(0j, (A.30)

11(1I XOX( [I 0 0 0 A3114 i o ~ cj i[hI 1 0 ) i 0

+(-Gzz + Gz)e< 0 I1)(1I| + GIi E 11)(11j, (A.31)

and the corresponding reduced density matrices for i are

1 p,o =1 [(1 + G7)I + (Gz + Gz)o-z, (A.32)

P 1 2~= i[ Pi,1 = [(1 - Gj)I + (-Gzz + Gz)o-z], (A.33)

which are already diagonal in matrix form, but notice that the eigenvalues corre-

sponding to the two measurement outcomes are no longer the same. We omit boring technical steps here, and finally we will obtain

D(j -+ i) = - S Ak,1logAk,1 - S(ilj) (A.34) {k=0,1} {l=+,-}

where S(ilj) is given by Eq. (A.22), and

Ao,+ = ( 1 + Gz t Gz Gz), (A.35) 1 A1, 1 (1 - Gj -F Gz t Gz). (A.36) 2 i

Alternatively, we can apply the following projector parametrization method to

transform the problem of optimization over measurements into optimization over

129

( scalar parameters. For the two-qubit case, we can parametrize the local projectors as {VIO)(OV,VI1)(1IV'} where

cos sin e--iO V= 2 , (A.37) sin Oe'q' - cos and 0 c [0, 7r], # E [0, 27), i.e., V E SU(2). Now QD can be written as an optimization over variables 0 and #, and equivalent results are expected for real X states.

A.2.3 Pairwise DD

Here I consider another candidate measure of quantum correlations, DD. By symmetry, the reduced state is always diagonal in the computational basis. Therefore, it is the Schmidt basis of real X states: o is automatically the local measurement basis in the definition of DD. That is,

DD(j -+ i) = - Ak,1 log Ak,1 - S(ilj), (A.38) {k=0,1} {=+,-}

by Eq. (A.34). Note that Eq. (A.38) can be extended to all real X states that are

not in these two classes.

130 Bibliography

[1] Scott Aaronson and Daniel Gottesman. Improved simulation of stabilizer cir- cuits. Phys. Rev. A, 70:052328, Nov 2004.

[2] Vincenzo Alba, Masudul Haque, and Andreas M Lduchli. Entanglement spec- trum of the heisenberg XXZ chain near the ferromagnetic point. J. Stat. Mech., 2012(08):P08011, 2012.

[3] Colm A. Ryan Alexandre M. Souza, J. Zhang and Raymond Laflamme. Exper- imental magic state distillation for fault-tolerant quantum computing. Nature Comm., 2:169, Jan 2011.

[4] Mazhar Ali, A. R. P. Rau, and G. Alber. Quantum discord for two-qubit X states. Phys. Rev. A, 81:042105, Apr 2010.

[5] S. Alipour, A. Mani, and A. T. Rezakhani. Quantum discord and non- markovianity of quantum dynamics. Phys. Rev. A, 85:052108, May 2012.

[6] A. Ambainis, A. Nayak, A. Ta-Shma, and U. Vazirani. Dense Quantum Coding and a Lower Bound for 1-way Quantum Automata. arXiv:quant-ph/9804043.

[71 Somshubhro Bandyopadhyay, Rahul Jain, Jonathan Oppenheim, and Christo- pher Perry. Conclusive exclusion of quantum states. Phys. Rev. A, 89:022336, Feb 2014.

[8] T. Baumgratz, M. Cramer, and B. Plenio, M.* Quantifying coherence. Phys. Rev. Lett., 113:140401, Sep 2014.

[91 Jacob D. Bekenstein. Black holes and entropy. Phys. Rev. D, 7:2333-2346, Apr 1973.

[101 Charles H. Bennett, Herbert J. Bernstein, Sandu Popescu, and Benjamin Schu- macher. Concentrating partial entanglement by local operations. Phys. Rev. A, 53:2046-2052, Apr 1996.

[11] Charles H. Bennett, Gilles Brassard, Claude Cr6peau, Richard Jozsa, Asher Peres, and William K. Wootters. Teleporting an unknown quantum state via dual classical and einstein-podolsky-rosen channels. Phys. Rev. Lett., 70:1895- 1899, Mar 1993.

131 [12] Charles H. Bennett, David P. DiVincenzo, John A. Smolin, and William K. Wootters. Mixed-state entanglement and quantum error correction. Phys. Rev. A, 54:3824-3851, Nov 1996.

[13] Charles H. Bennett, Andrzej Grudka, Michal Horodecki, Pawel Horodecki, and Ryszard Horodecki. Postulates for measures of genuine multipartite correlations. Phys. Rev. A, 83:012312, Jan 2011.

[14] Luca Bombelli, Rabinder K. Koul, Joohan Lee, and Rafael D. Sorkin. Quantum source of entropy for black holes. Phys. Rev. D, 34:373-383, Jul 1986.

[151 Raphael Bousso. The holographic principle. Rev. Mod. Phys., 74:825-874, Aug 2002.

[16] F. G. S. L. Brandiio and G. Gour. The general structure of quantum resource theories. ArXiv e-prints, February 2015.

[17] F. G. S. L. Branddo and M. Horodecki. An area law for entanglement from exponential decay of correlations. Nature Physics, 9:721-726, November 2013.

[181 Fernando G. S. L. Branddo, Michal Horodecki, Jonathan Oppenheim, Joseph M. Renes, and Robert W. Spekkens. Resource theory of quantum states out of thermal equilibrium. Phys. Rev. Lett., 111:250404, Dec 2013.

[19] Sergey Bravyi and Alexei Kitaev. Universal quantum computation with ideal clifford gates and noisy ancillas. Phys. Rev. A, 71:022316, Feb 2005.

[20] H. Buhrman, L. Czekaj, A. Grudka, M. Horodecki, P. Horodecki, M. Markiewicz, F. Speelman, and S. Strelchuk. Quantum communication complexity advantage implies violation of a Bell inequality. arXiv:1502.01058.

[21] Pasquale Calabrese and John Cardy. Entanglement entropy and quantum field theory. J. Stat. Mech., 2004(06):P06002, 2004.

[22] Curtis Callan and Frank Wilezek. On geometric entropy. Phys. Lett. B, 333(1- 2):55 - 61, 1994.

[23] D. Cavalcanti, L. Aolita, S. Boixo, K. Modi, M. Piani, and A. Winter. Op- erational interpretations of quantum discord. Phys. Rev. A, 83:032324, Mar 2011.

[24] Carlton Caves, Christopher Fuchs, and Ridiger Schack. Conditions for com- patibility of quantum-state assignments. Phys. Rev. A, 66:062111, Dec 2002.

[251 Li-Xiang Cen, Xin-Qi Li, Jiushu Shao, and YiJing Yan. Quantifying quantum discord and entanglement of formation via unified purifications. Phys. Rev. A, 83:054101, May 2011.

132 126] Pochung Chen, Zhilong Xue, I P McCulloch, Ming-Chiang Chung, Miguel Cazalilla, and S-K Yip. Entanglement entropy scaling of the XXZ chain. J. Stat. Mech., 2013(10):P10007, 2013.

[27] Qing Chen, Chengjie Zhang, Sixia Yu, X. X. Yi, and C. H. Oh. Quantum discord of two-qubit X states. Phys. Rev. A, 84:042313, Oct 2011.

[281 E. Chitambar, D. Leung, L. Maneinska, M. Ozols, and A. Winter. Everything You Always Wanted to Know About LOCC (But Were Afraid to Ask). Com- munications in Mathematical Physics, 328:303-326, May 2014.

[29] B. Coecke, T. Fritz, and R. W. Spekkens. A mathematical theory of resources. ArXiv e-prints, September 2014.

[30] B. Daki6, Y. 0. Lipp, X. Ma, M. Ringbauer, S. Kropatschek, S. Barz, T. Pa- terek, V. Vedral, A. Zeilinger, C. Brukner, and P. Walther. Quantum discord as resource for remote state preparation. Nature Physics, 8:666-670, September 2012.

[31] Borivoje Daki6, Vlatko Vedral, and Caslav Brukner. Necessary and sufficient condition for nonzero quantum discord. Phys. Rev. Lett., 105:190502, Nov 2010.

[32] A. Datta. Studies on the Role of Entanglement in Mixed-state Quantum Com- putation. PhD thesis, PhD Thesis, 2008, 2008.

[33] Animesh Datta, Anil Shaji, and Carlton M. Caves. Quantum discord and the power of one qubit. Phys. Rev. Lett., 100:050502, Feb 2008.

[34] D. P. DiVincenzo, D. W. Leung, and B. M. Terhal. Quantum Data Hiding. eprint arXiv:quant-ph/0103098, March 2001.

[35] A. Einstein, B. Podolsky, and N. Rosen. Can quantum-mechanical description of physical reality be considered complete? Phys. Rev., 47:777-780, May 1935.

[36] J. Eisert, M. Cramer, and M. B. Plenio. Colloquium: Area laws for the entanglement entropy. Rev. Mod. Phys., 82:277-306, Feb 2010.

[37] F. F. Fanchini, T. Werlang, C. A. Brasil, L. G. E. Arruda, and A. 0. Caldeira. Non-markovian dynamics of quantum discord. Phys. Rev. A, 81:052107, May 2010.

[38] A. Ferraro, L. Aolita, D. Cavalcanti, F. M. Cucchietti, and A. Acin. Almost all quantum states have nonclassical correlations. Phys. Rev. A, 81:052318, May 2010.

[39] Anat Ganor, Gillat Kol, and Ran Raz. Exponential separation of information and communication. In Proceedings of 55th IEEE Annual Symposium on Foun- dations of Computer Science, FOCS 2014, pages 176-185, Los Alamitos, CA, 2014. IEEE Computer Society.

133 [40] D. Gavinsky, J. Kempe, I. Kerenidis, R. Raz, and R. de Wolf. Exponential separations for one-way quantum communication complexity, with applications to cryptography. eprint arXiv:quant-ph/0611209, November 2006.

[41] D. Gottesman. The Heisenberg Representation of Quantum Computers. eprint arXiv:quant-ph/9807006,July 1998.

[42] A. Grudka, K. Horodecki, M. Horodecki, P. Horodecki, R. Horodecki, P. Joshi, W. Klobus, and A. W6jcik. Quantifying contextuality. Phys. Rev. Lett., 112:120401, Mar 2014.

[43] S. Hamieh, R. Kobes, and H. Zaraket. Positive-operator-valued measure optimization of classical correlations. Phys. Rev. A, 70:052325, Nov 2004.

[44] M B Hastings. An area law for one-dimensional quantum systems. J. Stat. Mech., 2007(08):P08024, 2007.

[451 S. W. Hawking. Nature, 248:30-31, 1974.

[46] Christoph Holzhey, Finn Larsen, and Frank Wilczek. Geometric and renor- malized entropy in conformal field theory. Nucl. Phys. B, 424(3):443 - 467, 1994.

[47] M. Horodecki and J. Oppenheim. (Quantumness in the context of) Resource Theories. Int. J. Mod. Phys. B, 27:45019, January 2013.

[48] Michal Horodecki, Pawel Horodecki, Ryszard Horodecki, Jonathan Oppenheim, Aditi Sen(De), Ujjwal Sen, and Barbara Synak-Radtke. Local versus nonlocal information in quantum-information theory: Formalism and phenomena. Phys. Rev. A, 71:062307, Jun 2005.

[49] Michal Horodecki, Jonathan Oppenheim, and Ryszard Horodecki. Are the laws of entanglement theory thermodynamical? Phys. Rev. Lett., 89:240403, Nov 2002.

[50] Ryszard Horodecki, Pawel Horodecki, Michal Horodecki, and Karol Horodecki. Quantum entanglement. Rev. Mod. Phys., 81:865-942, Jun 2009.

[51] Ryszard Horodecki, Pawel Horodecki, Michal Horodecki, and Karol Horodecki. Quantum entanglement. Rev. Mod. Phys., 81:865-942, Jun 2009.

[52] Yichen Huang. Quantum discord for two-qubit X states: Analytical formula with very small worst-case error. Phys. Rev. A, 88:014302, Jul 2013.

[53] Yichen Huang. Computing quantum discord is NP-complete. New J. Phys., 16:033027, 2014.

[54] Yichen Huang. Scaling of quantum discord in spin models. Phys. Rev. B, 89:054410, Feb 2014.

134 [55] Karol 2yczkowski, Pawel Horodecki, Anna Sanpera, and Maciej Lewenstein. Volume of the set of separable states. Phys. Rev. A, 58:883-892, Aug 1998.

[56] L. M. Ioannou. Computational complexity of the quantum separability problem. eprint arXiv:quant-ph/0603199, March 2006.

[57] Lawrence M. Ioannou. Computational complexity of the quantum separability problem. Quantum Info. Comput., 7(4):335-370, May 2007.

[58] E. Knill and R. Laflamme. Power of one bit of quantum information. Phys. Rev. Lett., 81:5672-5675, Dec 1998.

[591 Masato Koashi and Andreas Winter. Monogamy of quantum entanglement and other correlations. Phys. Rev. A, 69:022309, Feb 2004.

[601 M. D. Lang, C. M. Caves, and A. Shaji. Entropic measures of nonclassical correlations. ArXiv e-prints, May 2011.

[61] Nan Li and Shunlong Luo. Classical states versus separable states. Phys. Rev. A, 78:024303, Aug 2008.

[621 Nan Li and Shunlong Luo. Classical and quantum correlative capacities of quantum systems. Phys. Rev. A, 84:042124, Oct 2011.

[63] Seth Lloyd. Black Holes, Demons and the Loss of Coherence. PhD thesis, Rockfeller University, 1988.

[64] Seth Lloyd, Gang Chen, Vazrik Chiloyan, Yongjie Hu, Samuel Huberman, Zi- Wen Liu, and Poetro Lebdo Sambegoro. No energy transport without discord. In preparation.

[65] Shunlong Luo. Using measurement-induced disturbance to characterize correlations as classical or quantum. Phys. Rev. A, 77:022301, Feb 2008.

[66] Shunlong Luo, Shuangshuang Fu, and Hongting Song. Quantifying non- markovianity via correlations. Phys. Rev. A, 86:044101, Oct 2012.

[67] A. Luther and I. Peschel. Calculation of critical exponents in two dimensions from quantum field theory in one dimension. Phys. Rev. B, 12:3908-3917, Nov 1975.

[68] Juan Maldacena. The large-n limit of superconformal field theories and super- gravity. InternationalJournal of Theoretical Physics, 38(4):1113-1133, 1999.

[69] Victor Veitch Mark Howard, Joel Wallman and Joseph Emerson. Contextuality supplies the 'magic' for quantum computation. Nature, 510:351-355, Jan 2014.

[70] I. Marvian and R. W. Spekkens. Extending Noethers theorem by quantifying the asymmetry of quantum states. Nature Communications, 5:3821, May 2014.

135 [71] G. Mauro D'Ariano, P. Lo Presti, and P. Perinotti. Classical randomness in quantum measurements. Journal of Physics A Mathematical General, 38:5979- 5991, July 2005.

[72] Sebastian Meznaric, Stephen R. Clark, and Animesh Datta. Quantifying the nonclassicality of operations. Phys. Rev. Lett., 110:070502, Feb 2013.

[73] Kavan Modi, Aharon Brodutch, Hugo Cable, Tomasz Paterek, and Vlatko Ve- dral. The classical-quantum boundary for correlations: Discord and related measures. Rev. Mod. Phys., 84:1655-1707, Nov 2012.

[74] A. Nayak. Optimal lower bounds for quantum automata and random access codes. arXiv:quant-ph/9904093.

[75] Michael A. Nielsen and Isaac L. Chuang. Quantum Computation and Quantum Information. Cambridge University Press, New York, 2010.

[76] Michael A. Nielsen and Isaac L. Chuang. Quantum Computation and Quan- tum Information: 10th Anniversary Edition. Cambridge University Press, New York, NY, USA, 10th edition, 2011.

[77] Harold Ollivier and Wojciech H. Zurek. Quantum discord: A measure of the quantumness of correlations. Phys. Rev. Lett., 88:017901, Dec 2001.

[78] Jonathan Oppenheim, Michal Horodecki, Pawel Horodecki, and Ryszard Horodecki. Thermodynamical approach to quantifying quantum correlations. Phys. Rev. Lett., 89:180402, Oct 2002.

[79] Asher Peres. Separability criterion for density matrices. Phys. Rev. Lett., 77:1413-1415, Aug 1996.

[80] C. Perry, R. Jain, and J. Oppenheim. Communication tasks with infinite quantum-classical separation. arXiv:1407.8217.

[81] S. Pirandola. Quantum discord as a resource for quantum cryptography. Sci- entific Reports, 4:6956, November 2014.

[82] M. B. Plenio and S. Virmani. An introduction to entanglement measures. eprint arXiv:quant-ph/0504163, April 2005.

[83] Matthew F. Pusey, Jonathan Barrett, and Terry Rudolph. On the reality of the quantum state. Nat. Phys., 8(6):475-478, 2012.

[84] E. M. Rains. Rigorous treatment of distillable entanglement. Phys. Rev. A, 60:173-178, Jul 1999.

[85] A. Rivas, S. F. Huelga, and M. B. Plenio. Quantum non-Markovianity: characterization, quantification and detection. Reports on Progress in Physics, 77(9):094001, September 2014.

136 [86] Angel Rivas, Susana F. Huelga, and Martin B. Plenio. Entanglement and non- markovianity of quantum evolutions. Phys. Rev. Lett., 105:050403, Jul 2010.

[87] C. C. Rulli and M. S. Sarandy. Global quantum discord in multipartite systems. Phys. Rev. A, 84:042109, Oct 2011.

[88] Subir Sachdev. Quantum Phase Transitions. Cambridge University Press, New York, 2011.

[89] M. S. Sarandy. Classical correlation and quantum discord in critical systems. Phys. Rev. A, 80:022108, Aug 2009.

[90] Siddhartha Sen. Average entropy of a quantum subsystem. Phys. Rev. Lett., 77:1-3, Jul 1996.

[91] Peter Shor. Structure of unital maps and the asymptotic quantum Birkhoff conjecture. Unpublished.

[92] John A. Smolin, Frank Verstraete, and Andreas Winter. Entanglement of as- sistance and multipartite state distillation. Phys. Rev. A, 72:052317, Nov 2005.

[93] Robert W. Spekkens. Negativity and contextuality are equivalent notions of nonclassicality. Phys. Rev. Lett., 101:020401, Jul 2008.

[94] Mark Srednicki. Entropy and area. Phys. Rev. Lett., 71:666-669, Aug 1993.

[95] Alexander Streltsov, Gerardo Adesso, Marco Piani, and Dagmar Brut. Are general quantum correlations monogamous? Phys. Rev. Lett., 109:050503, Aug 2012.

[96] Alexander Streltsov, Hermann Kampermann, and Dagmar Brut. Behavior of quantum correlations under local noise. Phys. Rev. Lett., 107:170502, Oct 2011.

[971 Alexander Streltsov, Hermann Kampermann, and Dagmar BruE. Linking quantum discord to entanglement in a measurement. Phys. Rev. Lett., 106:160401, Apr 2011.

[98] Gerard 't Hooft. Dimensional Reduction in Quantum Gravity. ArXiv e-prints, October 1993.

[991 Barbara M. Terhal, Michal Horodecki, Debbie W. Leung, and David P. Di- Vincenzo. The entanglement of purification. Journal of Mathematical Physics, 43(9), 2002.

[100] Jaegon Um, Hyunggyu Park, and Haye Hinrichsen. Entanglement versus mutual information in quantum spin chains. Journal of Statistical Mechanics: Theory and Experiment, 2012(10):P10026, 2012.

137 [101J Victor Veitch, S A Hamed Mousavian, Daniel Gottesman, and Joseph Emer- son. The resource theory of stabilizer quantum computation. New Journal of Physics, 16(1):013009, 2014.

[1021 S. Virmani and M.B. Plenio. Ordering states with entanglement measures. Physics Letters A, 268(1-2):31 - 34, 2000.

[103] T. Werlang and Gustavo Rigolin. Thermal and magnetic quantum discord in heisenberg models. Phys. Rev. A, 81:044101, Apr 2010.

[104] Michael M. Wolf, Frank Verstraete, Matthew B. Hastings, and J. Ignacio Cirac. Area laws in quantum systems: Mutual information and correlations. Phys. Rev. Lett., 100:070502, Feb 2008.

[105] MichaelM. Wolf and J.Ignacio Cirac. Dividing quantum channels. Communi- cations in Mathematical Physics, 279(1):147-168, 2008.

[106] Shengjun Wu, Uffe V. Poulsen, and Klaus Molmer. Correlations in local measurements on a quantum state, and complementarity as an explanation of nonclassicality. Phys. Rev. A, 80:032319, Sep 2009.

[1071 Shengjun Wu, Uffe V. Poulsen, and Klaus Molmer. Correlations in local measurements on a quantum state, and complementarity as an explanation of nonclassicality. Phys. Rev. A, 80:032319, Sep 2009.

[1081 Zhengjun Xi, Xiao-Ming Lu, Xiaoguang Wang, and Yongming Li. The upper bound and continuity of quantum discord. Journal of Physics A: Mathematical and Theoretical, 44(37):375301, 2011.

[109] Wojciech Hubert Zurek. Quantum discord and Maxwell's demons. Phys. Rev. A, 67:012320, Jan 2003.

138