ASPECTS OF PART VS WHOLE RELATIONSHIPS IN QUANTUM INFORMATION PROCESSING A Thesis Submitted to the Faculty in partial fulfillment of the requirements for the degree of Doctor of Philosophy in Physics and Astronomy by Peter Douglas Johnson DARTMOUTH COLLEGE Hanover, New Hampshire October 20, 2016 Examining Committee:
Lorenza Viola, Chair
Chandrasekhar Ramanathan
Miles Blencowe
F. Jon Kull, Ph.D. Fernando G.S.L. Brandão Dean of Graduate and Advanced Studies
Abstract
Since its inception in quantum theory, the phenomenon of quantum entanglement has evolved from Einstein’s enigmatic “spooky action at a distance” to a crucial resource for quantum information processing. Recent technological advances geared towards controlling quantum systems and harnessing quantum entanglement have borne new perspectives and challenges. One major challenge is the development of a complete theory of multi-partite entanglement. Quantum theory places highly non-trivial con- straints on how entanglement may be distributed among the parts of a whole com- posite quantum system. In the simplest example, the more entangled system A is with system B, the less entangled system B can be with system C. This principle, known as the “monogamy of entanglement”, is a uniquely quantum feature enabling, in particular, secure quantum key distribution protocols and having ramifications for control of many-body quantum systems. In the first half of this thesis, I describe our contributions toward understanding the principles governing the distribution of multipartite entanglement. In particular, we elucidate surprising connections between the underlying kinematic constraints and the dynamical constraints stemming from the “no-cloning” principle and the uncer- tainty principle for incompatible quantum observables. In the second half of the thesis, I describe our contributions towards developing methods to create and control multipartite entanglement under realistic resource con- straints. Thanks to a number of recent experimental realizations, dissipative control of quantum systems is garnering increasing attention, alongside traditional unitary approaches. We investigate the use of dissipative control for driving a quantum sys- tem towards a target entangled state independently of initialization, a task known as “stabilization” – subject to the constraint that control resources be “quasi-local”. In particular, we develop mathematical tools for discovering hidden structures among the parts of a multi-partite entangled state which enable their stabilization.
ii
Preface
I cannot imagine a more ideal setting for balancing work and play than Dartmouth College in Hanover, New Hampshire. I am grateful to have spent such formative years in an environment that combines intense intellectual stimulation with striking natural beauty. I will truly miss the Norwich hills cycling, Connecticut river swimming, and granite peaks hiking. Certainly, this sense of connection to place is rooted in connections to friends. I am sincerely thankful to all of you who have helped me write such a vibrant chapter of my life. First and foremost, I thank my PhD adviser and mentor Lorenza Viola. You have shaped my development as a scientist in ways that I have yet to fully appreciate and have taught me lessons that extend far beyond physics. For the remainder of my career I will be drawing inspiration from your tireless demand of quality. I thank my local thesis committee Chandrasekhar Ramanathan and Miles Blencowe for guidance and encouragement throughout my PhD. I also thank my outside ex- aminer Fernando Brandão for participating in my defense and encouraging me as a scientist. I owe thanks to two other members of my Italian academic family. Thanks to my academic older brother, Francesco Ticozzi, for friendship, for expanding my mathe- matical tool set, and for supporting me in many ways. Thanks to my academic uncle, Roberto Onofrio, for thoughtful, timely guidance over the past six years. My working days were made brighter by the companionship of two wonderful friends Abhijeet Alase and Salini Karuvade. I look forward to the evolution of our friendship and collaboration for years to come. A number of physicists have been responsible for steering my course at various points over the past six years. Thanks to: Sandu Popescu for a conversation over billiards that blossomed into an undying pursuit; Stephon Alexander for helping me to see some some beautiful connections between my passions of physics and music; Carlton Caves for some memorable anecdotes, your encouragement, and for being a role model. Finally, I want to give special thanks to Ben Schumacher for setting my course and to Bill Wootters for sustaining it. Outside of academia, I thank my Hanover friends Dan Reeves, Ian Adelstein, Billy Braasch, and Mana Francisquez for pulling me out of the office for adventures and
iv then still talking shop along the way. Thanks to Sam, Russ, Andrew, and Ben for making deep thinking a fond passtime. Lastly, I want to especially thank my parents, brothers, and Ariana for your love and for putting up with (and even encouraging!) my pursuit of physics.
v Contents
Abstract ...... ii Preface ...... iv
1 Introduction 1
2 Quantum marginals: sharability and joinability 9 2.1 Introduction ...... 10 2.2 Joining and sharing classical vs. quantum states ...... 12 2.2.1 Joinability ...... 13 2.2.2 Sharability ...... 17 2.3 Joining and sharing Werner and isotropic states ...... 20 2.3.1 Werner and isotropic qudit states, and their classical analogues 20 2.3.2 Classical joinability limitations ...... 23 2.3.3 Joinability of Werner and isotropic qudit states ...... 24 2.3.4 Isotropic joinability results from quantum cloning ...... 30 2.3.5 Sharability of Werner and isotropic qudit states ...... 31 2.4 Further remarks ...... 33 2.4.1 Joinability beyond the three-party scenario ...... 33 2.4.2 Sharability of general bipartite qubit states ...... 35
3 Joinability of causal and acausal relationships 37 3.1 Introduction ...... 38 3.2 General quantum joinability framework ...... 40 3.2.1 Homocorrelation map and positive cones ...... 41 3.2.2 Generalization of joinability ...... 45 3.3 Three-party joinability settings with collective invariance ...... 48 3.3.1 Joinability limitations from state-positivity and channel-positivity 48 3.3.2 Joinability limitations from local-positivity ...... 53 3.4 Agreement bounds for quantum states and channels ...... 57
vi 4 Towards an alternative approach to joinability: enforcing positivity through purification 63 4.1 Introduction ...... 64 4.2 Joinability limitations from hard constraints ...... 66 4.3 Joinability limitations from soft constraints ...... 70
5 Asymptotic stabilization of quantum states with continuous-time quasi-local dynamics 77 5.1 Introduction ...... 78 5.2 Preliminaries ...... 79 5.2.1 Notation and background ...... 79 5.2.2 Fixed points of quantum dynamical semigroups ...... 82 5.2.3 Quasi-local semigroup dynamics ...... 86 5.2.4 Quasi-local stabilizability: prior pure-state results and frustration- free semigroup dynamics ...... 88 5.3 Frustration-Free Stabilizable States: Necessary Conditions ...... 90 5.3.1 Linear-algebraic tools ...... 90 5.3.2 Invariance conditions for quasi-local generators ...... 92 5.3.3 From invariance to necessary conditions for stabilizability . . . 94 5.4 Frustration-Free Stabilizable States: Sufficient Conditions ...... 96 5.4.1 A key result on frustration-free Markovian evolutions . . . . . 96 5.4.2 Sufficient conditions for full-rank target states ...... 98 5.4.3 Sufficient conditions for general target states ...... 100 5.5 Illustrative Applications ...... 102 5.5.1 Some notable failures of quasi-local stabilizability ...... 102 5.5.2 Quasi-local stabilization of graph product states ...... 103 5.5.3 Quasi-local stabilization of commuting Gibbs states ...... 107 5.5.4 Quasi-local stabilization beyond commuting Hamiltonians . . 111 5.5.5 Approximate FFQLS ...... 120
6 Finite-time stabilization of quantum states with discrete-time quasi- local dynamics 125 6.1 Introduction ...... 126 6.2 Preliminaries ...... 127 6.2.1 Quasi-local discrete-time dynamical semigroups ...... 127 6.2.2 Convergence notions ...... 129 6.2.3 Quasi-local parent Hamiltonians ...... 130 6.3 Finite-time stabilization ...... 132 6.3.1 Conditions for finite-time stabilization ...... 132 6.3.2 Efficiency of finite-time stabilization ...... 137 vii 6.4 Robust finite-time stabilization: necessary conditions ...... 140 6.5 Robust finite-time stabilization: sufficient conditions ...... 144 6.5.1 Non-operational sufficiency criteria ...... 144 6.5.2 Operational sufficiency criteria ...... 152 6.6 Efficiency of robust finite-time stabilization ...... 158 6.6.1 Circuit complexity ...... 158 6.6.2 Connection to rapid mixing ...... 161 6.7 Extension of results to mixed target states ...... 163 6.8 Proofs ...... 165
7 Towards finite-time dissipative quasi-local quantum encoders 187 7.1 Preliminaries ...... 188 7.2 Repetition code ...... 191 7.3 Toric code ...... 195 7.4 General stabilizer codes ...... 200 7.5 Further questions ...... 203
8 Summary and outlook 205
A List of abbreviations 213
References 215
viii
List of Tables
2.1 n-m sharability of Werner states ...... 34
x
List of Figures
1.1 Overlapping neighborhoods ...... 4
2.1 Werner and isotropic state joinability ...... 25 2.2 Werner and isotropic state projected joinability region ...... 29 2.3 Qubit Werner and isotropic state 1-n sharability ...... 33
3.1 State-channel duality commutativity diagram ...... 42 3.2 State-positive and channel-positive cone example ...... 44 3.3 Joinability problem homomorphism ...... 48 3.4 Brauer operator positivity ...... 51 3.5 Comparison of Werner state and depolarizing channel joinability . . . 53 3.6 Local-positive joining of Werner operators ...... 54 3.7 Comparison of separable joining and state-channel intersection joining 57
6.1 1-D nearest-neighbor neighborhood structure ...... 128 6.2 Cubic graph AKLT state: not finite-time stabilizable ...... 133 6.3 Finite-time stabilization scheme for AKLT state ...... 138 6.4 Non-factorized robust finite-time stabilizable state ...... 149 6.5 Generalized Bravyi-Vyalyi state ...... 153 6.6 Robust finite-time stabilizable state with non-commuting canonical frustration-free parent Hamiltonian ...... 157 6.7 Low-depth dissipative circuit for robust finite-time stabilization on lat- tice system ...... 158 6.8 CCZ-state on Kagome lattice ...... 160
7.1 Stabilizers and labeling scheme for 2-D toric code ...... 196 7.2 Logical operators for toric code ...... 197 7.3 Sequence of correction maps for finite-time encoding ...... 198
xii
Chapter 1
Introduction
1 Introduction
Quantum information processing provides a radically different means of comput- ing that shows promise for solving previously intractable problems. From a physical standpoint, one of the most compelling applications is to the simulation of quan- tum systems [1, 2]. A very practical and promising use for quantum simulation is, for instance, the calculation of molecular energies [3]. With a classical computer, such calculations are feasible for small systems, but they soon become too resource- intensive as larger molecules are considered. Quantum algorithms have been devel- oped which are believed to solve this problem by requiring exponentially fewer steps than the best known classical algorithms. A few months before the time of writing, Google Inc. and collaborators published work that demonstrated the first example of a scalable quantum simulation calculating molecular energies [4]. Besides its technological applications, quantum information science also offers a diverse set of mathematical tools and concepts. Many of these are actively being developed or, even, created. These tools and concepts are proving useful in areas of physics as diverse as condensed matter physics [5, 6], quantum field theory [7, 8], and quantum gravity [9, 10, 11]. Even with its technological potential and powerful mathematical toolbox, quan- tum information science has captivated many researchers, including the author, for a different reason. Some of the seminal contributions to the field were initiated by physicists asking foundational questions about quantum mechanics [12]. Quantum information science has given us deep insights into the workings of Nature. Certainly, more insights lie just beyond the horizon. In the most common setting, quantum information is processed with quantum bits, or qubits as opposed to bits. In principle, any two-level quantum system may serve as a qubit. Examples include the spin of an electron, the polarization of a photon, or the two lowest energy levels in a superconducting Josephson junction [13]. No single system has yet to triumph as the universally used qubit. In practice, certain systems (e.g. trapped ions or superconducting qubits) are more accurately controlled than others, and therefore, show more promise for becoming the standard qubit. Quantum bits obey more subtle rules than classical bits do. In particular, multiple quantum systems can exhibit a uniquely quantum type of correlation known as quan- tum entanglement. Although we do not yet fully understand its ramifications, it is thought that quantum entanglement plays a crucial role in providing the advantages of quantum computing over classical computing. Also, not all classical information processes translate to the quantum case. For instance, there is no quantum process which can perfectly copy, or clone, an unknown quantum bit [14]. This mentioning of quantum entanglement and the no-cloning principle anticipate some of the nuances of quantum information processing that we address throughout the thesis. The field of quantum information science, which gained traction in the 1990s, is now a firmly established research field [15]. Many of the most pressing challenges in
2 Introduction
quantum information science demand improvements in hardware engineering. Much of the theoretical groundwork is, by now, relatively well-established. The important concepts from classical information theory have been appropriately repurposed for use in quantum information theory [16]. There are a number of well-established models of quantum computing such as the circuit model [17], measurement-based quantum computing [18], and, towards solving certain optimization problems, quantum an- nealing [19]. The theory of quantum error correction, necessary for building scalable, fault-tolerant quantum computers, is relatively well-understood [20]. Despite these advances, many theoretical challenges remain and are being actively pursued. We still lack a complete understanding of the principles ensuring “quantum supremacy” over classical information processing. And, furthermore, we cannot antic- ipate the breakthroughs in understanding and unforeseen applications that are likely to result from a deeper inquiry into the quantum. The theory of bipartite entanglement, quantum correlations between two systems, is well-understood, including mixed-state entanglement. However, even in the sim- plest setting where subsystems are distinguishable, multipartite entanglement de- mands a much richer tool-kit of mathematical concepts, many of which are currently being developed. Some of these tools are already finding diverse applications, such as towards exploring space-time as an emergent phenomenon based on quantum entan- glement [11]. Nevertheless, questions remain as for how best to describe multipartite entanglement, to characterize its features, and to use it as a resource. In the circuit model approach to quantum information processing, one applies a sequence of unitary quantum gates to some fiducial input state, seeking to maintain sufficient coherent control of the relevant parts of the system. However, even within the quantum circuit model, access to suitable non-unitary control is crucial for proper initialization or “entropy removal” in fault-tolerant architectures, as well as for uni- versal “digital” simulation of open-system dynamics [21]. Control theoretic advances in open quantum systems have further revealed new avenues for utilizing incoherent control by means of engineered dissipation. Such non-unitary control resources have been shown to be beneficial for the tasks of robust preparation of resource states [22], rapid quantum state purification [23], and engineered dissipative quantum memories [24]. The last few years, in particular, have seen a surge of interest in proposals for dissipatively preparing strongly-correlated and topological phases of matter. Recent experimental advances have used engineered dissipation to autonomously drive a two- qubit system towards an entangled resource state. This feat has been achieved with superconducting qubits [25] as well as with ion traps [26]. With these advances comes a demand for developing mathematical tools which can fully describe the capabilities and limitations of controlled dissipation. This thesis contributes to the problems identified above. The unifying theme tying this work together is the quantum part vs whole relationship. Although the content
3 Introduction
Figure 1.1: Three quantum systems ABC with delineated subsystems AB and BC.
is varied in motivation and mathematical techniques, nearly all of the problems we consider admit a simple example on three subsystems as pictured in Fig. 1.1. Many questions we consider involve confronting the interaction, or “overlap”, of subsystems (e.g. subsystem AB overlaps with subsystem BC). Like discussing the weather, the ubiquitous difficulty of “the interacting case” gives even physicists with differing backgrounds a topic with which to mutually relate. It may be useful for the reader to keep this picture in mind throughout. The thesis naturally divides into two parts. The first half explores a concept that we developed known as quantum joinability. The notion of joinability addresses the question of whether or not there exists a description of the whole (e.g. ABC) which is consistent with constraints on the parts (e.g AB and BC); equivalently, we ask if the constrained parts can be “joined” into some physically allowed whole. Due to the phenomenon of entanglement, quantum theory places non-trivial constraints on the ways that the parts of a quantum system may be correlated among one another. As an example, the more entangled qubit A is with B, the less entangled qubit B can be with C. This phenomenon, known as the monogamy of entanglement, is uniquely quantum in that classical correlations are not limited by such a constraint. As the number of systems increases, the complexity of such constraints grows drastically. The study of such constraints is often referred to as the quantum marginal problem [27]. This problem originated from the field of quantum chemistry [28], motivated by the attempt to simplify calculations of atomic and molecular ground state energies. More recently, this problem has gained attention from the quantum information com- munity since it addresses the nature of entanglement distribution in a multipartite quantum system. Researchers have drawn on and developed diverse mathematical tools towards tackling this very difficult problem. We have contributed to this effort by analyzing, in detail, a number of tractable cases which bear insight on the more general problem. Furthermore, we have established rigorous connections among the concepts of monogamy of entanglement, the no-cloning principle, and incompatible measurements. These concepts are ubiquitous in quantum information and are find- ing application in other areas of physics such as black-hole thermodynamics [29]. We develop the framework of quantum joinability in order to put these notions on equal footing and elucidate their common origin.
4 Introduction
• Chapter 2 defines the problems of quantum sharability and quantum joinability for bipartite quantum states. We investigate some basic scenarios in which a given set of bipartite quantum states may consistently arise as the set of re- duced states of a global N-partite quantum state. We restrict the discussion to bipartite reduced states that belong to the paradigmatic classes of Werner and isotropic states in d dimensions, and focus on two specific versions of the quantum marginal problem which we find to be tractable. The first is Alice- Bob, Alice-Charlie joining, with both pairs being in a Werner or isotropic state. The second is m-n sharability of a Werner state across N subsystems, which may be seen as a variant of the N-representability problem to the case where subsystems are partitioned into two groupings of m and n parties, respectively. By exploiting the symmetry properties that each class of states enjoys, we de- termine necessary and sufficient conditions for three-party joinability and 1-n sharability for arbitrary d. Our results explicitly show that although entangle- ment is required for sharing limitations to emerge, correlations beyond entan- glement generally suffice to restrict joinability, and not all unentangled states necessarily obey the same limitations. The relationship between joinability and quantum cloning as well as implications for the joinability of arbitrary bipartite states are discussed. In particular, the observations regarding quantum cloning lead naturally into the investigation of the subsequent chapter. • Chapter 3 develops the framework of quantum joinability which unifies seem- ingly different joinability problems for bipartite quantum states and channels. This includes well known problems such as optimal quantum cloning and quan- tum marginal problems as special instances. Central to our generalization is a variant of the Choi-Jamiolkowski isomorphism between bipartite states and dynamical maps which we term the “homocorrelation map”: while the former emphasizes the preservation of the positivity constraint, the latter is designed to preserve statistical correlations, allowing more direct contact with entan- glement. In particular, we define and analyze state-joining, channel-joining, and local-positive joining problems in three-party settings exhibiting collective U ⊗ U ⊗ U symmetry, obtaining exact analytical characterizations in low di- mension. Suggestively, we find that bipartite quantum states are limited in the degree to which their measurement outcomes may agree, while quantum channels are limited in the degree to which their measurement outcomes may disagree. Loosely speaking, quantum mechanics enforces an upper bound on the extent of positive correlation across two subsystems at a single time, as well as on the extent of negative correlation between the state of a single system across two instants of time. We argue that these general statistical bounds inform the quantum joinability limitations, and show that they are in fact sufficient for the three-party U ⊗ U ⊗ U-invariant setting.
5 Introduction
• Chapter 4 delves deeper into underpinnings of quantum joinability. We adopt a different approach to the joinability problem, ensuring positive-semidefiniteness or complete positivity from the outset. With this, we formalize the “composition law” of correlations from the previous two chapters. We incorporate the notion of incompatibility of observables as another example of quantum joinability. Furthermore, we draw parallels between the quantum and classical problems, finding a intuitive common cause for their respective joinability constraints. The mathematical techniques developed here are intended to elucidate some quantum peculiarities such as the distinction between causal and acausal quan- tum relationships and the origin of monogamy of entanglement, no-cloning, and measurement incompatibility. The second half of this thesis explores engineering dissipation to stabilize quantum states both asymptotically and in finite time. Towards preparing the quantum re- source of many-body entanglement in a realistic setting, one must address the issue of constrained control capabilities. This becomes increasingly relevant as the number of parts in the systems is scaled up, as is needed for quantum information processing to deliver its full potential. In particular, a possible realistic constraint on one’s control of the system is that only parts of the system can be addressed at a time. Much like a logical circuit, we may assume that manipulations of the whole are achieved by a series of manipulations of the parts. Such an implementation is said to be quasi-local. Thus, in quasi-local stabilization, the part-whole relationship features, in that we seek a preparation of the whole (e.g. ABC) by addressing only its parts (e.g. AB and BC). The task of engineering quasi-local dissipation to drive a quantum system, all-to- one, into a target quantum state has been initiated and explored in [30, 31, 32, 33, 34]. These works consider a continuous-time dynamics generated by an engineered quasi-local Markovian master equation, giving examples and exploring conditions for preparing, or “stabilizing”, a target pure state. In practice, pure quantum states are never available and natural dissipative dynamics exhibit mixed steady states. Accordingly, we contribute to this line of research by exploring quasi-local stabilization in the case of mixed target states. Some implementations of engineered dissipation are best modeled by discrete-time dynamics [35]. Constraining the dynamical maps to act quasi-locally, we can view the sequence of maps as a dissipative quantum circuit. An advantage in this case, compared to that of continuous-time, is that a target state may be exactly stabilized in a finite time. We contribute to the field of dissipative quantum control by determining conditions for finite-time stabilization of a target state. We describe how well-known resource states, such as graph states for measurement based quantum computing, may be stabilized with these schemes. To conclude, we present some preliminary work geared towards achieving quantum encoding using quasi-local resources.
6 Introduction
• Chapter 5 builds off of previous work by L. Viola and F. Ticozzi to investi- gate the engineering of dissipative continuous-time dynamics to render a target mixed quantum state as the unique global attractor of the dynamics. In par- ticular, we determine necessary and sufficient conditions for whether or not a given target state can be the unique steady state of frustration-free quasi-local continuous-time Markovian dynamics. We investigate under which conditions a mixed state on a finite-dimensional multipartite quantum system may be the unique, globally stable fixed point of frustration-free semigroup dynamics subject to specified quasi-locality constraints. Our central result is a linear- algebraic necessary and sufficient condition for a generic (full-rank) target state to be frustration-free quasi-locally stabilizable, along with an explicit procedure for constructing Markovian dynamics that achieve stabilization. If the target state is not full-rank, we establish sufficiency under an additional condition, which is naturally motivated by consistency with pure-state stabilization re- sults yet provably not necessary in general. Several applications are discussed, of relevance to both dissipative quantum engineering and information process- ing, and non-equilibrium quantum statistical mechanics. In particular, we show that a large class of graph product states (including arbitrary thermal graph states) as well as Gibbs states of commuting Hamiltonians are frustration-free stabilizable relative to natural quasi-locality constraints. Likewise, we provide explicit examples of non-commuting Gibbs states and non-trivially entangled mixed states that are stabilizable despite the lack of an underlying commuting structure, albeit scalability to arbitrary system size remains in this case an open question.
• Chapter 6 complements the work of Chapter 5 by investigating the discrete- time analog of quantum state stabilization with quasi-local dynamics. While continuous-time Markovian dynamics cannot exactly stabilize a target state in finite time, discrete-time dynamics can, in principle. We develop necessary and sufficient conditions for establishing if a given target state can be stabilized by a finite sequence of quasi-local dynamical maps. Then we investigate the efficient scheme of robust finite-time stabilization, whereby the target state is stabilized regardless of the implementation order of the dynamical maps. A main theme in this chapter is the role that certain “commuting structures” play in facilitating robust stabilization.
• Chapter 7 turns to the task of preparing quantum information in a quantum error correcting code. As in the previous chapter, we explore several examples whereby this task can be achieved exactly, in principle, by a finite sequence of quasi-local dynamical maps. We develop a number of principles which aid the construction of such finite-time dissipative encoders utilizing the quantum
7 Introduction
stabilizer formalism.
8 Chapter 2
Quantum marginals: sharability and joinability
9 Quantum marginals: sharability and joinability
This chapter presents material that appeared in Physical Review A, 88:032323 (2013), in an article titled “Compatible quantum correlations: Extension problems for Werner and isotropic states”, which is joint work with Lorenza Viola.
2.1 Introduction
Understanding the nature of quantum correlations in multiparty systems and the distinguishing features they exhibit relative to classical correlations is a central goal across quantum information processing (QIP) science [17], with implications rang- ing from condensed-matter and statistical physics to quantum chemistry, and the quantum-to-classical transition. From a foundational perspective, exploring what different kinds of correlations are, in principle, allowed by probabilistic theories more general than quantum mechanics further helps to identify under which set of physical constraints the standard quantum framework may be uniquely recovered [36, 37]. In this context, entanglement provides a distinctively quantum type of correla- tion, that has no analogue in classical statistical mechanics. A striking feature of entanglement is that it cannot be freely distributed among different parties: if a bi- partite system, say, A(lice) and B(ob), is in a maximally entangled pure state, then no other system, C(harlie), may be correlated with it. In other words, the entan- glement between A and B is monogamous and cannot be shared [38, 39, 40, 41, 42]. This simple tripartite setting motivates two simple questions about bipartite quan- tum states: given a bipartite state, we ask whether it can arise as the reduced state of A-B and of A-C simultaneously; or, more generally, given two bipartite states, we ask if one can arise as the reduced state of A-B while the other arises as the reduced state of A-C. It should be emphasized that both of these are questions about the existence of tripartite states with given reduction properties. While formal (and more general) definitions will be provided later in the chapter, these examples serve to in- troduce the notions of sharing (1-2 sharing) and joining (1-2 joining), respectively. In its most general formulation, the joinability problem is also known as the quantum marginal problem (or local consistency problem), which has been heavily investigated both from a mathematical-physics [27, 43, 44] and a quantum-chemistry perspective [45, 46] and is known to be QMA-hard [47]. Our choice of terminology, however, facilitates a uniform language for describing the joinability/sharability scenarios. For instance, we say that the joinable correlations of A-B and A-C are joined by a joining state on A-B-C. The limited sharability/joinability of entanglement was first quantified in the sem- inal work by Coffman, Kundu, and Wootters, in terms of an exact (CKW) inequality obeyed by the entanglement across the A-B, A-C and A-(BC) bipartitions, as mea- sured by concurrence [38]. In a similar venue, several subsequent investigations at- tempted to determine how different entanglement measures can be used to diagnose
10 2.1 Introduction
failures of joinability, see e.g. [48, 49, 41]. More recently, significant progress has been made in characterizing quantum correlations more general than entanglement [50, 51], in particular as captured by quantum discord [52]. While it is now established that quantum discord does not obey a monogamy inequality [53], different kinds of limitations exist on the extent to which it can be freely shared and/or communicated [54, 55]. Despite these important advances, a complete picture is far from being reached. What kind of limitations do strictly mark the quantum-classical correlation boundary? What different quantum features are responsible for enforcing different aspects of such limitations, and how does this relate to the degree of resourcefulness that these correlations can have for QIP? While the above are some of the broad questions motivating this work, our spe- cific focus here is to make progress on joinability and sharability properties in low- dimensional multipartite settings. In this context, reference [56] has obtained a nec- essary condition for three-party joining in finite dimension in terms of the subsystem entropies, and additionally established a sufficient condition in terms of the trace- norm distances between the states in question and known joinable states. For the specific case of qubit Werner states [57], Werner himself established necessary and sufficient conditions for the 1-2 joining scenario [58]. With regards to sharability, necessary and sufficient conditions have been found for 1-2 sharing of generic bipar- tite qubit states [59], as well as for specific classes of qudit states [60]. To the best of our knowledge, no conditions that are both necessary and sufficient for the joinability of generic states are available as yet. In this chapter, we obtain necessary and suffi- cient conditions for both the three-party joinability and the 1-n sharability problems, in the case that the reduced bipartite states are either Werner or isotropic states on d-dimensional subsystems (qudits). Though our results are restricted in scope of applicability, they provide key in- sights as to the sources of joinability limitations. Most importantly, we find that standard measures of quantum correlations, such as concurrence and quantum dis- cord, do not suffice to determine the limitations in joining quantum correlations. Specifically, we find that the joined states need not be entangled or even discordant in order not to be joinable. Further to that, although separable states may have join- ability limitations, they are, nonetheless, freely (arbitrarily) sharable. By introducing a one-parameter class of probability distributions, we provide a natural classical ana- logue to qudit Werner and isotropic quantum states. This allows us to illustrate how classical joinability restrictions carry over to the quantum case and, more inter- estingly, to demonstrate that the quantum case demands limitations which are not present classically. Ultimately, this feature may be traced back to complementarity of observables, which clearly plays no role in the classical case. It is suggestive to note that the uncertainty principle was also shown to be instrumental in constraining the sharability of quantum discord [54]. It is our hope that further pursuits of more
11 Quantum marginals: sharability and joinability
general necessary and sufficient conditions may be aided by the methods and findings herein. The content is organized as follows. In Sec. 2.2 we present the relevant mathemat- ical framework for defining the joinability and sharability notions and the extension problems of interest, along with some preliminary results contrasting the classical and quantum cases. Sec. 2.3 contains the core results of our analysis. In particular, after reviewing the defining properties of Werner and isotropic states on qudits, in Sec. 2.3.1 we motivate the appropriate choice of probability distributions that serve as a classical analogue, and determine the resulting classical joinability limitations in Sec. 2.3.2. Necessary and sufficient conditions for three-party joinability of quantum Werner and isotropic states are established in Sec. 2.3.3, and contrasted to the clas- sical scenario. Sec. 2.3.4 shows how the results on isotropic state joinability are in fact related to known results on quantum cloning, whereas in Sec. 2.3.5 we establish simple analytic expressions for the 1-n sharability of both Werner and isotropic states, along with discussing constructive procedures to determine m-n sharability properties for m > 1. In Sec. 2.4, we present additional remarks on joinability and sharability scenarios beyond those of Sec. 2.3. In particular, we outline generalizations of our analysis to N-party joinability, and show how bounds on the sharability of arbitrary bipartite states follow from the Werner and isotropic results. For ease and clarity of presentation, we have omitted the technical proofs of the results in Sec. 2.3 from this thesis. These proofs can be found in the appendix of [61].
2.2 Joining and sharing classical vs. quantum states
Although our main focus will be to quantitatively characterize simple low-dimensional settings, we introduce the relevant concepts with a higher degree of generality, in order to better highlight the underlying mathematical structure and to ease connections with existing related notions in the literature. We are interested in the correlations among the subsystems of a N-partite composite system S. In the quantum case, we thus require a Hilbert space with a tensor product structure:
N (N) O (1) (1) H ' Hi , dim(Hi ) ≡ di, i=1
(1) where Hi represents the individual “single-particle” state spaces and, for our pur- poses, each di is finite. In the classical scenario, to each subsystem we associate a sample space Ωi consisting of di possible outcomes, with the joint sample space being
12 2.2 Joining and sharing classical vs. quantum states
given by the Cartesian product:
(N) Ω ' Ω1 × ... × ΩN .
Probability distributions on Ω(N) are the classical counterpart of quantum density operators on H(N).
2.2.1 Joinability The input to a joinability problem is a set of subsystem states which, in full generality, may be specified relative to a “neighborhood structure” on H(N) (or Ω(N)) [31, 34]. That is, let neighborhoods {Nj} be given as subsets of the set of indexes labeling individual subsystems, Nk ( ZN . We can then give the following:
Definition 2.2.1. [Quantum Joinability] Given a neighborhood structure {N1, N2, (N) ..., N`} on H , a list of density operators
(ρ1, . . . , ρ`) ∈ (D(HN1 ),..., D(HN` ))
is joinable if there exists an N-partite density operator w ∈ D(H(N)), called a joining state, that reduces according to the neighborhood structure, that is,
Tr ˆ (w) = ρ , ∀k = 1, . . . , `, (2.1) Nk k ˆ where Nk ≡ ZN \Nk is the tensor complement of Nk.
The analagous definition for classical joinability is obtained by substituting corre- ˆ sponding terms, in particular, by replacing the partial trace over Nk with the corre- sponding marginal probability distribution. As remarked, the question of joinability has been extensively investigated in the context of the classical [62] and quantum [27, 63, 64, 56] marginal problem. A joining state is equivalenty referred to as an extension or an element of the pre-image of the list under the reduction map, while the members of a list of joinable states are also said to be compatible or consistent. Clearly, a necessary condition for a list of states to be joinable is that they “agree” on any overlapping reduced states. That is, given any two states from the list whose neighborhoods are intersecting, the reduced states of the subsystems in the intersec- tion must coincide. From this point of view, any failure of joinability due to a dis- agreement of overlapping reduced states is a trivial case of non-compatible N-party correlations. We are interested in cases where joinability fails despite the agreement on overlapping marginals. This consistency requirement will be satisfied by construc- tion for the Werner and isotropic quantum states we shall consider in Sec. 2.3.
13 Quantum marginals: sharability and joinability
One important feature of joinability, which has recently been investigated in [65], is the convex structure that both joinable states lists and joining states enjoy. The set of lists of density operators satisfying a given joinability scenario is convex under component-wise combination; this is because the same convex combination of their joining states is a valid joining state for the convex combined list of states. Similarly, the set of joining states for a given list of joined states is convex by the linearity of the partial trace. As mentioned, one of our goals is to shed light on limitations of quantum vs. classical joinability and the extent to which entanglement may play a role in that respect. That quantum states are subject to stricter joinability limitations than classical probability distributions are, can be immediately appreciated by considering two density operators ρAB = |ΨBihΨB| = ρAC , where |ΨBi is any maximally entangled Bell pair on two qubits: no three-qubit joining state wABC exists, despite the reduced state on A being manifestly consistent. In contrast, as shown in [56, 62], as long as two classical distributions have equal marginal distributions over A, p(A, B) and p(A, C) can always be joined. This is evidenced by the construction of the joining state: w(A, B, C) = p(A, B) p(A, C)/p(A). As pointed out in [56], although the above choice is not unique, it is the joining state with maximal entropy and represents an even mixture of all valid joining distributions. Although any two consistently-overlapped classical probability distributions may be joined, limitations on joining classical probability distributions do typically arise in more general joining scenarios. This follows from the fact that any classical probability assignment must be consistent with some convex combination of pure states. Consider, for example, a pairwise neighboorhood structure, with an associated list of states p(A, B), p(B,C), and p(A, C), which have consistent single-subsystem marginals. Clearly, if each subsystem corresponds to a bit, no convex combination of pure states gives rise to a probability distribution w(A, B, C) in which each pair is completely anticorrelated; in other words, “bits of three can’t all disagree”. In Sec. 2.3.3, we explicitly compare this particular classical joining scenario to analogous quantum scenarios. While all the classical joining limitations may be expressed by linear inequalities, the quantum joining limitations are significantly more complicated. The limitations arise from demanding that the joining operator be a valid density operator, namely, trace-one and non-negative (which clearly implies Hermiticity). This fact is demon- strated by the following proposition, which may be readily generalized to any joining scenario:
Proposition 2.2.2. For any two trace-one Hermitian operators QAB and QAC which obey the consistency condition TrB (QAB) = TrC (QAC ), there exists a trace-one Her- mitian joining operator QABC . Proof. Consider an orthogonal Hermitian product basis which includes the identity
14 2.2 Joining and sharing classical vs. quantum states
for each subsystem, that is, {Ai ⊗ Bj ⊗ Ck}, where A0 = B0 = C0 = I. Then we can construct the space of all valid joining operators QABC as follows. Let dABC be the dimension of the composite system. The component along A0 ⊗ B0 ⊗ C0 is fixed as 1/dABC , satisfying the trace-one requirement. The components along the two- body operators of the form Ai ⊗ Bj ⊗ I are fixed by the required reduction to QAB, and similarly the components along the two-body operators of the form Ai ⊗ I ⊗ Ck are determined by QAC . The components along the one-body operators of the form Ai ⊗ I ⊗ I, I ⊗ Bi ⊗ I, and I ⊗ I ⊗ Ci are determined from the reductions of QAB and QAC . This leaves the coefficients of all remaining basis operators unconstrained, since their corresponding basis operators are zero after a partial trace over systems B or C. Thus, requiring the joining operator to be Hermitian and normalized is not a limiting constraint with respect to joinability: any limitations are due to the non- negativity constraint. Understanding how non-negativity manifests itself is extremely difficult in general and far beyond our scope here. We can nevertheless give an ex- ample in which the role of non-negativity is clear. Part of the job of non-negativity is to enforce constraints that are also obeyed by classical probability distributions. For example, in the case of a two-qubit state ρ, if hX ⊗ Iiρ = 1 and hI ⊗ Xiρ = 1, then hX ⊗ Xiρ must equal 1. More generally, consider a set of mutually commuting ob- k servables {Mi}i=1 and any basis {|mi} in which all Mi are diagonal. Any valid state must lead to a list of expectation values (Tr (ρM1) ,..., Tr (ρMk)), whose values are element-wise convex combinations of the vertexes {(hm|M1|mi,..., hm|Mk|mi)|∀m}. The interpretation of this constraint is that since commuting observables have simul- taneously definable values, just as classical observables do, probability distributions on them must obey the rules of classical probability distributions. We call on this fact when we compare the quantum joining limitations to the classical analogue ones in Sec. 2.3.3. Non-negativity constraints that do not arise from classical limitations on compat- ible observables may be labeled as inherently quantum constraints, the most familiar being provided by uncertainty relations for conjugate observables [66, 67]. Although complementarity constraints are most evident for observables acting on the same system, complementarity can also give rise to a trade-off in the information about a subsystem observable vs. a joint observable. This fact is essentially what allows Bell’s inequality to be violated. For our purposes, the complementarity that comes ~ ~ into play is that between “overlapping” joint observables (e.g., between S1 · S2 and ~ ~ S1 · S3 for three qubits). We are thus generally interested in understanding the inter- play between purely classical and quantum joining limitations, and in the correlation trade-offs that may possibly emerge. Historically, as already mentioned, a pioneering exploration of the extent to which quantum correlations can be shared among three parties was carried out in [38],
15 Quantum marginals: sharability and joinability yielding a characterization of the monogamy of entanglement in terms of the well- known CKW inequality: 2 2 2 min CAB + CAC ≤ (C )A(BC), where C denotes the concurrence and the right hand-side is minimized over all pure- state decompositions. Thus, with the entanglement across the bipartition A and (BC) held fixed, an increase in the upper bound of the A-B entanglement can only come at the cost of a decrease in the upper bound of the A-C entanglement. One may wonder whether the CKW inequality may help in diagnosing joinability of reduced states. If a joining state wABC is not a priori determined (in fact, the existence of such a state is the entire question of joinability), the CKW inequality may be used to obtain a necessary condition for joinability, namely, if ρAB and ρAC are joinable, then
2 2 CAB + CAC ≤ 1. (2.2)
However, there exist pairs of bipartite states – both unentangled (as the following Proposition shows) and non-trivially entangled (as we shall determine in Sec. III.B, see in particular Fig. 2.2a) – that obey the “weak” CKW inequality in Eq. (2.2), yet are not joinable. The key point is that while the limitations that the CKW captures are to be ascribed to entanglement, entanglement is not required to prevent two states from being joinable. In fact, weaker forms of quantum correlations, as quantified by quantum discord [52], are likewise not required for joinability limitations. Consider, specifically, so-called “classical-quantum” bipartite states, of the form
X i X ρ = pi|iihi|A ⊗ σB, pi = 1, i i
i where {|iiA} is some local orthogonal basis on A and σB is, for each i, an arbitrary state on B. Such states are known to have zero discord [68]. Yet, the following holds:
Proposition 2.2.3. Classical-quantum correlated states need not be joinable.
Proof. Consider the two quantum states
ρAB = (|↑X ↑X ih↑X ↑X | + |↓X ↓X ih↓X ↓X |)/2,
ρAC = (|↑Z ↑Z ih↑Z ↑Z | + |↓Z ↓Z ih↓Z ↓Z |)/2, on the pairs A-B and A-C, respectively. Both have a completely mixed reduced state over A and thus it is meaningful to consider their joinability. Let wABC be a joining state. Then the outcome of Bob’s X measurement would correctly lead him to predict Alice to be in the state | ↑X i or | ↓X i, while at the same time the outcome of Charlie’s Z measurement would correctly lead him to predict Alice to be in the state | ↑Z i or
16 2.2 Joining and sharing classical vs. quantum states
| ↓Z i. Since this violates the uncertainty principle, wABC cannot be a valid joining state. The existence of separable but not joinable states has been independently reported in [56]. While formally our example is subsumed under the more general one presented in Thm. 4.2 therein (strictly satisfying the necessary condition for joinability given by their Eq. (2.2)), it has the advantage of offering both a transparent physical interpretation of the underlying correlation properties, and an intuitive proof of the joinability failure.
2.2.2 Sharability As mentioned, the second joinability structure we analyze is motivated by the concept of sharability. In our context, we can think of sharability as a restricted joining (2) (1) (1) scenario in which a bipartite state is joined with copies of itself. If H 'H1 ⊗H2 , (1) consider a N-partite space that consists of m “left” copies of H1 and n “right” (1) copies of H2 , with each neighborhood consisting of one right and one left subsystem, respectively (hence a total of mn neighborhoods). We then have the following:
Definition 2.2.4. [Quantum Sharability] A bipartite density operator ρ ∈ D(HL⊗ ⊗m ⊗n HR) is m-n sharable if there exists an N-partite density operator w ∈ D(HL ⊗HR ), called a sharing state, that reduces left-right-pairwise to ρ, that is,
Tr (w) = ρ, ∀i = 1, . . . , m, j = 1, . . . , n, (2.3) LˆiRˆj where the partial trace is taken over the tensor complement of neighborhood ij. Each m-n sharability scenario may be viewed as a specific joining structure with the additional constraint that each of the joining states be equal to one another, the list being (ρ, ρ, . . . , ρ). In what follows, we shall take arbitrarily sharable to mean ∞-∞ sharable, whereas finitely sharable means that ρ is not m-n sharable for some m, n. Also, each property “m-n sharable” (sometimes also referred to as a “m-n extendible”) is taken to define a sharability criterion, which a state may or may not satisfy. It is worth noting the relationship between sharability and N-representability. The N-representability problem asks if, for a given (symmetric) p-partite density operator (1) ⊗p ρ on (H1 ) , there exists an N-partite pre-image state for which ρ is the p-particle reduced state. N-representability has been extensively studied for indistinguishable bosonic and fermionic subsystems [69, 45, 46] and is a very important problem in quantum chemistry [70]. We can view N-representability as a variant on the shara- bility problem, whereby the distinction between the left and right subsystems is lifted, and m + n = N. Given the p-partite state ρ as the shared state, we ask if there exists a sharing N-partite state which shares ρ among all possible p-partite subsystems. In
17 Quantum marginals: sharability and joinability
the setting of indistinguishable particles, the associated symmetry further constrains the space of the valid N-partite sharing states. Just as with 1-2 joinability, any classical probability distribution is arbitrarily sharable [37]. Likewise, similar to the joinability case, convexity properties play (1) an important role towards characterizing sharability. If dim(H1 ) = d1 ≡ dL and (1) dim(H2 ) = d2 ≡ dR, then it follows from the convexity of the set of joinable states lists that m-n sharable states form a convex set, for fixed subsystem dimensions dL and dR. This implies that if ρ satisfies a particular sharability criterion, then any mixture of ρ with the completely mixed state also satisfies that criterion, since the completely mixed state is arbitrarily (∞-∞) sharable. Besides mixing with the identity, the degree of sharability may be unchanged under more general transformations on the input state. Consider, specifically, completely- positive trace-preserving bipartite maps M(ρ) that can be written as a mixture of local unitary operations, that is,
X i i i† i† X M(ρ) = λiU1 ⊗ V2 ρU1 ⊗ V2 , λi = 1, (2.4) i i
i i where U1 and V2 are arbitrary unitary transformations on HL and HR, respectively. These (unital) maps form a proper subset of general Local Operations and Classical Communication (LOCC) [17]. We establish the following:
Theorem 2.2.5. If ρ is m-n sharable, then M(ρ) is m-n sharable for any map M that is a convex mixture of unitaries.
Proof. Let M(ρ) be expressed as in Eq. (2.4). By virtue of the convexity of the set of m-n sharable states (for fixed subsystem dimensions), it suffices to show that each term, UV ρU †V †, in M(ρ) is m-n sharable. Let w be a sharing state for ρ, and define
0 † † † † w = U1 ...UmVm+1 ...Vm+n w U1 ...UmVm+1 ...Vm+n .
Then, for any left-right pair of subsystems i and j, it follows that
0 † † † † Tri,j (w ) = UiVjTri,j (w) Ui Vj = U ⊗ V ρU ⊗ V = ρUV .
0 Hence, w is an m-n-sharing state for ρUV , as desired.
This result suggests a connection between the degree of sharability and the en- tanglement of a given state. In both cases, there exist classes of states for which these properties cannot be “further degraded” by locally acting maps (or any map for that matter). Obviously, LOCC cannot decrease the entanglement of states with no entanglement, and convex unitary mixtures as above cannot increase the sharability
18 2.2 Joining and sharing classical vs. quantum states
of states with ∞-∞ sharability (because they are already as sharable as possible). These two classes of states can in fact be shown to coincide as a consequence of the fact that arbitrary sharability is equivalent to (bipartite) separability. This result has been appreciated in the literature [40, 36, 71, 37] and is credited to both [72] and [73]. We reproduce it here in view of its relevance to our work:
Theorem 2.2.6. A bipartite quantum state ρ on HL ⊗ HR is unentangled (or sepa- rable) if and only if it is arbitrarily sharable. L R Proof. (⇐) Let ρ be separable. Then for some set of density operators {ρi , ρi }, it P L R P can be written as ρ = i λiρi ⊗ ρi , with i λi = 1. Let n and m be arbitrary, and let the N-partite state w, be defined as follows:
X L ⊗m R ⊗n w = λi(ρi ) ⊗ (ρi ) , i with N = m + n. By construction, the state of each L-R pair is ρ, since it follows straighforwardly that Eq. (2.3) is obeyed for each i, j. Thus, w is a valid sharing state. (⇒) Since ρ is arbitrarily sharable, there exists a sharing state w for arbitrary values of m, n. In particular, we need only make use of a sharing state w for m = 1 and arbitrarily large n, whence we let n → ∞. Given w, let us construct another sharing state w˜, which is invariant under permutations of the right subsystems, that is, let 1 X † w˜ = Vπ wVπ, |Sn| π∈Sn ⊗n where Sn ≡ {π} is the permutation group of n objects, acting on H via the natural Q R n-fold representation, Vπ( i |ψii) = ⊗i|ψπ(i)i, i = 1, . . . , n. It then follows that w˜ shares ρ:
1 X † TrL,ˆ Rˆ (w ˜) = TrL,ˆ Rˆ Vπ wVπ |Sn| π∈Sn 1 X 1 X = Tr (w) = ρ = ρ. L,πˆ (Rˆi) |Sn| |Sn| π∈Sn π∈Sn
⊗∞ Having established the existence of a symmetric sharing state w˜ ∈ D(HL ⊗ HR ), Fannes’ Theorem (see section 2 of [73]) implies the existence of a unique representation P i i i of w˜ as a sum of product states, w˜ = i λiρL ⊗ ρR ⊗ ρR ⊗ .... Reducing w˜ to any L-R pair leaves a separable state. Thus, if ρ is 1-n sharable it must be separable. As we alluded to before, a Corollary of this result is that in fact 1-∞ sharability implies ∞-∞ sharability. In closing this section, we also briefly mention the concept
19 Quantum marginals: sharability and joinability
(1) ⊗p of exchangeability [74, 75]. A density operator ρ on (H1 ) is said to be exchangeable if it is symmetric under permutation of its p subsystems and if there exists a symmetric (1) ⊗(p+q) state w on (H1 ) such that the reduced states of any subset of p subsystems is ρ for all q ∈ N. Similar to sharability, exchangeability implies separability. However, the converse only holds in general for sharability: clearly, there exist states which are separable but not exchangeable, because of the extra symmetry requirement. Thus, the notion of sharability is more directly related to entanglement than exchangeability is.
2.3 Joining and sharing Werner and isotropic states
Even for the simplest case of two bipartite states with an overlapping marginal, a general characterization of joinability is extremely non-trivial. As remarked, no con- ditions yet exist which are both necessary and sufficient for two arbitrary density op- erators to be joinable; although, conditions that are separately necessary or sufficient have been recently derived [56]. In this Section, we present a complete characteriza- tion of the three-party joining scenario and the 1-n sharability problem for Werner and isotropic states on arbitrary subsystem dimension d. We begin by introducing the relevant families of quantum and classical states to be considered.
2.3.1 Werner and isotropic qudit states, and their classical analogues
The usefulness of bipartite Werner and isotropic states is derived from their simple analytic properties and range of mixed state entanglement. For a given subsystem dimension d, Werner states are defined as the one-parameter family that is invariant under collective unitary transformations [57] (see also [75]), that is, transformations of the form U ⊗ U, for arbitrary U ∈ U(d). The parameterization which we employ is given by d I 1V ρ(Ψ−) = (d − Ψ−) + Ψ− − , d2 − 1 d2 d d where V is the swap operator, defined by its action on any product ket, V |ψφi ≡ |φψi. This parameterization is chosen because Ψ− is a Werner state’s expectation value with respect to V , Ψ− = Tr[V ρ(Ψ−)]. Non-negativity is ensured by −1 ≤ Ψ− ≤ 1, and the completely mixed state corresponds to Ψ− = 1/d. Furthermore, the concurrence of Werner states is simply given by [76]
C(ρ(Ψ−)) = −Tr V ρ(Ψ−) = −Ψ−, Ψ− ≤ 0. (2.5)
20 2.3 Joining and sharing Werner and isotropic states
For Ψ− > 0, the concurrence is defined to be zero, indicating separability. Werner states have been experimentally characterized for photonic qubits, see e.g. [77]. In- terestingly, they can be dissipatively prepared as the steady state of coherently driven atoms subject to collective spontaneous decay [78]. Isotropic states are defined, similarly, as the one-parameter family that is invariant under transformations of the form U ∗ ⊗ U [79]. We parameterize these states as
d I 1 ρ(Φ+) = (d − Φ+) + Φ+ − |Φ+ihΦ+| , d2 − 1 d2 d where |Φ+i = p1/d P |iii. The value of the parameter is given by the expectation i h i + TA + value with respect to the partially transposed swap operator, Φ = Tr V(AB)ρ(Φ ) , and is related to the so-called “singlet fraction” [80] by Φ+ = dF . Non-negativity is now ensured by 0 ≤ Φ+ ≤ d, whereas the concurrence is given by [81], s 2 C(ρ(Φ+)) = (Φ+ − 1), Φ+ ≥ 1, (2.6) d(d − 1)
and is defined to be zero for Φ+ ≤ 1. Before introducing probability distributions that will serve as the analogue classi- cal states, we present an alternative way to think of Werner states, which will prove useful later. First, the highest purity, attained for the Ψ− = −1 state, is 2/[d(d − 1)], with the absolute maximum of 1 corresponding to the pure singlet state for d = 2. Second, collective projective measurements on a most-entangled Werner state return only disagreeing outcomes (e.g., corresponding to |1i ⊗ |3i, but not |1i ⊗ |1i). The following construction of bipartite Werner states demonstrates the origin of both of these essential features. For generic d, the analogue to the singlet state is the following d-partite fully anti-symmetric state:
− 1 X |ψ i = √ sign(π)Vπ|1i|2i ... |di, (2.7) d d! π∈Sd
where, as before, Sd ≡ {π} denotes the permutation group and {|`i} is an orthonor- mal basis on H(1) ' Cd. The above state has the property of being “completely disagreeing”, in the sense that a collective measurement returns outcomes that dif- fer on each qudit with certainty. The most-entangled bipartite qudit Werner state − is nothing but the two-party reduced state of |ψd i. Thus, we can think of general bipartite qudit Werner states as mixtures of the completely mixed state with the − two-party-reduction of |ψd i. The inverse of 2/[d(d − 1)] (the purity) is precisely the number of ways two “dits” can disagree. Understanding bipartite Werner states to
21 Quantum marginals: sharability and joinability
− arise from reduced states of |ψd i will inform our construction of the classical analogue states, and also help us understand some of the results of Sec. 2.3.3 and 2.3.5.
For Werner states, increased entanglement corresponds to increased “disagree- ment” for collective measurement outcomes. For isotropic states, increased entangle- ment corresponds to increased “agreement” of collective measurements, but only with respect to the computational basis {|ii} relative to which such states are defined. It is this expression of agreement vs. disagreement of outcomes which carries over to the classical analogue states, which we are now ready to introduce. The relevant probabil- ity distributions are defined on the outcome space Ωd × Ωd = {1, . . . , d} × {1, . . . , d}. To resemble Werner and isotropic quantum states, these probability distributions should have completely mixed marginal distributions and range from maximal dis- agreement to maximal agreement. This is achieved by an interpolation between an even mixture of “agreeing pure states”, namely, (1, 1), (2, 2),..., (d, d), and an even mixture of all possible “disagreeing pure states”, namely, (1, 2),..., (1, d), (2, 1),..., (d, d − 1). That is: α 1 − α p(A = i, B = j) = δ + (1 − δ ), (2.8) α d i,j d(d − 1) i,j
where α is the probability that the two outcomes agree.
To make the analogy complete, it is desirable to relate α to both Ψ− and Φ+. We define α in the quantum cases to be the probability of obtaining |ki on system A, conditional to outcome |ki on system B for the projective measurement {|ijihij|}. For Werner states, this probability is related to Ψ− by
Ψ− + 1 p(|ki | |ki ) = ≡ α , (2.9) A B W d + 1 W and, similarly for isotropic states, we have
Φ+ + 1 p(|ki | |ki ) = ≡ α . (2.10) A B I d + 1 I We may thus re-parameterize both the Werner and isotropic states in terms of their respective above-defined “probabilities of agreement”, namely:
d I 1V ρ(α ) = (1 − α ) + α − , (2.11) W d − 1 W d2 W d d d I 1 ρ(α ) = (1 − α ) + α − |Φ+ihΦ+| , (2.12) I d − 1 I d2 I d
22 2.3 Joining and sharing Werner and isotropic states
subject to the conditions 2 1 0 ≤ α ≤ , ≤ α ≤ 1. W d + 1 d + 1 I
For Werner states, αW can rightly be considered a probability of agreement be- cause it is independent of the choice of local basis vectors in the projective measure- † † ment {U ⊗ U|ijihij|U ⊗ U }. For isotropic states, αI does not have as direct an interpretation. We may nevertheless interpret α as a probability of basis-independent agreement if we pair local basis vectors on A with their complex conjugates on B. In other words, αI can be thought of as the probability of agreement for local projective measurements of the form {U ∗ ⊗ U|ijihij|U ∗† ⊗ U †} 1.
2.3.2 Classical joinability limitations In order to determine the joinability limitations in the classical case, we begin by not- ing that any (finite-dimensional) classical probability distribution is a unique convex combination of the pure states of the system. In our case, there are five extremal three-party states, for which the two-party marginals are classical analogue states, as defined in Eq. (2.8). These are
1 X p(A, B, C agree) = (i, i, i), d i 1 X p(A, B agree) = (i, i, j), d(d − 1) i6=j 1 X p(A, C agree) = (i, j, i), d(d − 1) i6=j 1 X p(B,C agree) = (j, i, i), d(d − 1) i6=j 1 X p(all disagree) = (i, j, k), d(d − 1)(d − 2) i6=j6=k
1In principle, one might question the validity of our analogy on account of the quantum states being more pure than the mixed classical states in Eq. (2.8), e.g., the classical analogue with α = 1 may seem closer to the separable quantum state (|11ih11|+...+|ddihdd|)/d. We note, however, that any POVM gives a mapping from a quantum state to a probability distribution; we have chosen here two families of quantum states for which the resulting probability distribution is minimally dependent on the choice of measurement, as argued. Furthermore, we aim to understand which features of joinability limitations are of a quantum origin and which are simply due to classical P 2 limitations. In contrast, the family of “classical-like” quantum states p i |iiihii|/d + (1 − p)I/d would be un-interesting, as manifestly only constrained by classical joining limitations.
23 Quantum marginals: sharability and joinability
where (i, j, k) stands for the pure probability distribution p(A, B, C) = δA,iδB,jδC,k. The first four of these states are valid for all d ≥ 2 and each corresponds to a vertex of a tetrahedron, as depicted in Fig. 2.1(left). The fifth state is only valid for d ≥ 3 and corresponds to the point (αAB, αAC , αBC ) = (0, 0, 0) in Fig. 2.1(right). Any valid three-party state for which the two-party marginals are classical analogue states must be a convex combination of the above states. Therefore, the joinable-unjoinable boundary is delimited by the boundary of their convex hull. For the d = 2 case, the inequalities describing these boundaries are explicitly given by the following:
p(A, B, C agree) ≥ 0 ⇒ αAB + αAC + αBC ≥ 1,
p(C disagrees) ≥ 0 ⇒ −αAB + αAC + αBC ≤ 1,
p(B disagrees) ≥ 0 ⇒ αAB − αAC + αBC ≤ 1,
p(A disagrees) ≥ 0 ⇒ αAB + αAC − αBC ≤ 1,
where each inequality arises from requiring that the corresponding extremal state has a non-negative likelihood. In the d ≥ 3 case, the inequality p(A, B, C agree) ≥ 0 is replaced by αAB, αAC , αBC ≥ 0.
2.3.3 Joinability of Werner and isotropic qudit states We now present our results on the three-party joinability of Werner and isotropic states and then compare them to the classical limitations just found in the previous section. While, as mentioned, the proofs have been relegated to the published version of the paper [61], the basic idea is to exploit the high degree of symmetry that these classes of states enjoy. Consider Werner states first. Our starting point is to observe that if a tripartite state wABC joins two reduced Werner states ρAB and ρAC , then the “twirled state” w˜ABC , given by Z † w˜ABC = (U ⊗ U ⊗ U) wABC (U ⊗ U ⊗ U) dµ(U), (2.13)
is also a valid joining state. In Eq. (2.13), µ denotes the invariant Haar measure on U(d), and the twirling super-operator effects a projection into the subspace of operators with collective unitary invariance [82]. By invoking the Schur-Weyl duality [83], the guaranteed existence of joining states with these symmetries allows one to narrow the search for valid joining states to the Hermitian subspace spanned by representations of subsystem permutations, that is, density operators of the form X w = µπVπ, w ∈ WW , (2.14)
π∈S3
24 2.3 Joining and sharing Werner and isotropic states
(a) Werner joinability limitations (d=2) (b) Werner joinability limitations (d=5)
Figure 2.1: Three-party quantum and classical joinability limitations for Werner and isotropic states, and their classical analogue, as parameterized by Eqs. (2.11), (2.12), (2.8), respectively. a) Qubit case, d = 2. The Werner state boundary is the surface of the darker cone with its vertex at (2/3, 2/3, 2/3), whereas the isotropic state boundary is the surface of the lighter cone with its vertex at (1/3, 1/3, 2/3). The classical boundary is the surface of the tetrahedron. b) Higher-dimensional case, d = 5. The Werner state boundary is the surface of the bi-cone with vertices at (0, 0, 0) and (1/3, 1/3, 1/3), whereas the isotropic state boundary is the flattened cone with its vertex at (1/6, 1/6, 1/3). The classical boundary is the surface of the two joined tetrahedra. In both panels the grey line resting on top of the cones indicates the colinearity of the cone surfaces along this line segment.
25 Quantum marginals: sharability and joinability
∗ where Hermiticity demands that µπ = µπ−1 . Given wABC which joins Werner states, each subsystem pair is characterized by the expectation value with the respective − swap operator, Ψij = Tr[wABC (Vij ⊗ Iij)], where i, j ∈ {A, B, C} with i 6= j. Hence, − − − the task is to determine for which (ΨAB, ΨBC , ΨAC ) there exists a density operator wABC consistent with the above expectations. Our main result is the following:
− − − Theorem 2.3.1. Three Werner qudit states with parameters ΨAB, ΨBC , ΨAC are join- − − − able if and only if (ΨAB, ΨBC , ΨAC ) lies within the bi-cone described by
2 − − 2 − 1 ± Ψ− ≥ Ψ + ωΨ + ω Ψ , (2.15) 3 BC AC AB for d ≥ 3, or within the cone described by
2 − − 2 − 1 − Ψ− ≥ Ψ + ωΨ + ω Ψ , Ψ− ≥ 0, (2.16) 3 BC AC AB for d = 2, where 1 − − − i 2π Ψ− = (Ψ + Ψ + Ψ ), ω = e 3 . (2.17) 3 AB BC AC
Similarly, if a tripartite state wABC joins isotropic states ρAB and ρAC , then the “isotropic-twirled state” w˜ABC , given by Z ∗ ∗ † w˜ABC = (U ⊗ U ⊗ U) wABC (U ⊗ U ⊗ U) dµ(U), (2.18)
is also a valid joining state. A clarification is, however, in order at this point: although we have been referring to the isotropic joinability scenario of interest as three-party isotropic state joining, this is somewhat of a misnomer because we effectively consider the pair B-C to be in a Werner state, as evident from Eq. (2.18). Compared to Eq. (2.14), the relevant search space is now partially transposed relative to subsystem A, that is, consisting of density operators of the form
X TA w = µπVπ , w ∈ Wiso. (2.19) π∈S3 Our main result for three-party joinability of isotropic states is then contained in the following:
Theorem 2.3.2. Two isotropic qudit states ρAB and ρAC and qudit Werner state + + − + + − ρBC with parameters ΦAB, ΦAC , ΨBC are joinable if and only if (ΦAB, ΦAC , ΨBC ) lies within the cone described by
+ + − ΦAB + ΦAC − ΨBC ≤ d , (2.20)
26 2.3 Joining and sharing Werner and isotropic states
+ + − 1 + ΦAB + ΦAC − ΨBC ≥ (2.21) r − 2d iθ + −iθ + d(Ψ − 1) + (e Φ + e Φ ) , BC d − 1 AB AC e±iθ = ±ip(d + 1)/(2d) + p(d − 1)/(2d), or, for d ≥ 3, within the convex hull of the above cone and the point (0, 0, −1). The results of Theorems 2.3.1 and 2.3.2 as well as of Sec. 2.3.2 are pictorially sum- marized in Fig. 2.1. We now compare these quantum joinability limitations to the joinability limita- tions in place for classical analogue states. As described in Sec. 2.3.2, the non- negativity of p(A, B, C agree) and p(A disagrees) is enforced by the two inequalities αAB + αAC + αBC ≥ 1 and −αAB − αAC + αBC ≥ 1, respectively. We expect the same requirement to be enforced by the analogue quantum-measurement statistics. For d = 2, the bases of the Werner and isotropic joinability-limitation cones are − − − + + − determined by ΨAB + ΨAC + ΨBC ≥ 0 and ΦAB + ΦAC − ΨBC ≤ 2, respectively. Writing down each of these parameters in terms of the appropriate probability of agreement α, as defined in Eqs. (2.9) and (2.10), we obtain αAB + αAC + αBC ≥ 1 and −αAB − αAC + αBC ≥ 1. Hence, for qubits, part of the quantum joining limita- tions are indeed derived from the classical joining limitations. This is also illustrated in Fig. 2.1(left). Of course, one would not expect the quantum scenario to exhibit violations of the classical joinability restrictions; still, it is interesting that states which exhibit manifestly non-classical correlations may nonetheless saturate bounds obtained from purely classical joining limitations. For d ≥ 3, the only classical boundary which plays a role is the one which bounds + + − the base of the isotropic joinability-limitation cone: ΦAB +ΦAC −ΨBC ≤ d. Again, in terms of the agreement parameters, this is (just as for qubits) −αAB −αAC +αBC ≥ 1. In the Werner case, the quantum joinability boundary is not clearly delineated by the classical joining limitations. We can nevertheless make the following observation. By the non-negativity of Werner states, the three-party joinability region in Fig. 2.1(right) is required to lie within a cube of side-length 2/(d + 1) with one corner at (0, 0, 0). Consider the set of cubes obtained by rotating from this initial cube about an axis through (0, 0, 0) and (2/(d + 1), 2/(d + 1), 2/(d + 1). It is a curious fact that the exact quantum Werner joinability region (the bi-cone) is precisely the intersection of all such cubes. Another interesting feature is that there exist trios of unentangled Werner states − − − which are not joinable. For example, the point (ΨAB, ΨAC , ΨBC ) = (1, 1, 0) cor- responds to three separable Werner states that are not joinable. This point is of particular interest because its classical analogue is joinable. Translating (1, 1, 0) into the agreement-probability coordinates, (αAB, αAC , αBC ) = (2/3, 2/3, 1/3), we see that
27 Quantum marginals: sharability and joinability this point is actually on the classical joining limitation border. Thus, these three sepa- rable, correlated states are not joinable for purely quantum mechanical reasons. Note that the point (αAB, αAC , αBC ) = (2/3, 2/3, 1/3) does correspond to a joinable trio of pairs in the isotropic three-party joining scenario: this point lies at the center of the face of the isotropic joinability cone, as seen in Fig. 2.1(left). The same fact holds for (2/3, 1/3, 2/3) or (1/3, 2/3, 2/3) when the Werner state pair in the isotropic joining scenario describes A-C or A-B, respectively; in both cases, we would have obtained yet another cone in Fig. 2.1 that sits on a face of the classical tetrahedron boundary. Having determined the joinable trios of both Werner and isotropic states, we are now in a position to also answer the question of what pairs A-B and A-C of states are joinable with one another. In the Werner state case, this is obtained by projecting − − the Werner joinability bicone down to the ΨAB-ΨAC plane, resulting in the following:
− − Corollary 2.3.3. Two pairs of qudit Werner states with parameters ΨAB and ΨAC − − 1 are joinable if and only if ΨAB, ΨAC ≥ − 2 , or if the parameters satisfy 1 (Ψ− + Ψ− )2 + (Ψ− − Ψ− )2 ≤ 1, (2.22) AB AC 3 AB AC
− − 1 or additionally, in the case d ≥ 3, if ΨAB, ΨAC ≤ 2 . For isotropic states, we may similarly project the cone of Eq. (2.21) onto the + + ΦAB-ΦAC plane to obtain the 1-2 joining boundary. This yields the following:
+ + Corollary 2.3.4. Two pairs of qudit isotropic states with parameters ΦAB and ΦAC are joinable if and only if they lie within the convex hull of the ellipse
(Φ+ /d + Φ+ /d − 1)2 (Φ+ /d − Φ+ /d)2 AB AC + AB AC = 1, (2.23) (1/d2) (d2 − 1)/d2
+ + and the point (ΦAB, ΦAC ) = (0, 0). Lastly, by a similar projection of the isotropic cone given by Eqs. (2.20)-(2.21), we may explicitly characterize the Werner-isotropic hybrid 1-2 joining boundary:
+ Corollary 2.3.5. An isotropic state with parameter ΦAB and a Werner state with − parameter ΨBC are joinable if and only if they lie within the convex hull of the ellipse (Φ+ /d + Ψ− /d − 1)2 (Φ+ /d − Ψ+ /d)2 AB BC + AB BC = 1, (2.24) (1/d2) (d2 − 1)/d2
+ − and the point (ΦAB, ΨBC ) = (0, 1), and, for d ≥ 3, within the additional convex hull + − introduced by the point (ΦAB, ΨBC ) = (0, 1).
28 2.3 Joining and sharing Werner and isotropic states
(a) Werner state joinability (b) Isotropic state joinability
Figure 2.2: Two-party joinability limitations for Werner and isotropic qudit states. (a) Werner states. The shaded region corresponds to joinable Werner pairs, with the lighter region being valid only for d ≥ 3. The rounded boundary is the ellipse determined by Eq. (2.22). This explicitly shows the existence of pairs of entangled Werner states that are within the circular boundary determined by the weak CKW inequality, Eq. (2.2), yet are not joinable. (b) Isotropic states. The three regions correspond to the joinable pairs of isotropic states for d = 2, d = 3 and d = 1000. This shows how, in the limit of large d, the trade-off in isotropic state parameters becomes linear, consistent with known results on d-dimensional quantum cloning [84].
29 Quantum marginals: sharability and joinability
The above results give the exact quantum-mechanical rules for the two-pair join- ability of Werner and isotropic states, as pictorially summarized in see Figs. 2.2a and 2.2b. A number of interesting features are worth noticing. First, by restricting to the − − line where ΨAB = ΨAC , we can conclude that qubit Werner states are 1-2 sharable if and only if Ψ− ≥ −1/2, whereas for d ≥ 3, all qudit Werner states are 1-2 sharable. As we shall see, this agrees with the more general analysis of Sec. 2.3.5. Second, some insight into the role of entanglement in limiting joinability may be gained. In the first quadrant of Fig. 2.2a, where neither pair is entangled, it is no surprise that no joinability restrictions apply. Likewise, it is not surprising to see that, in the third quadrant where both pairs are entangled, there is a trade-off between the amount of entanglement allowed between one pair and that of the other. But, in the second and fourth quadrants we observe a more interesting behavior. Namely, these quadrants show that there is also a trade-off between the amount of classical correlation in one pair and the amount of entanglement in the other pair. In fact, the smoothness of the boundary curve as it crosses from one of the pairs being entangled to unentangled suggests that, at least in this case, entanglement is not the correct figure of merit in diagnosing joinability limitations.
2.3.4 Isotropic joinability results from quantum cloning Interestingly, the above results for 1-2 joinability of isotropic states can also be ob- tained by drawing upon existing results for asymmetric quantum cloning, see e.g. [84, 85] for 1-2 and 1-3 asymmetric cloning and [86, 87, 88] for 1-n asymmetric cloning. One approach to obtaining the optimal asymmetric cloning machine is to exploit the Choi isomorphism [89] to translate the construction of the optimal cloning map to the construction of an optimal operator (or a “telemapping state”). This connection is made fairly clear in [86, 90]; in particular, “singlet monogamy” refers to the trade-off in fidelities of the optimal 1-n asymmetric cloning machine or, equivalently, to the trade-off in singlet fractions for a (1 + n) qudit state. We describe how the approach to solving the optimal 1-n asymmetric cloning problem may be rephrased to solve the 1-n joinability problem for isotropic states. The state |Ψi described in Eq. (4) of [86] is a 1-n joining state for n isotropic states characterized by singlet fractions F0,j (related to the isotropic state parameter + by F0,j = Φ0,j/d, as noted). The bounds on the singlet fractions are determined by the normalization condition of |Ψi, together with the requirement that |Ψi be an eigenstate of a certain operator R defined in Eq. (3) of [86]. That |Ψi is an isotropic joining state is readily seen from its construction, and that it may optimize the singlet fractions (hence delineate the boundary in the {F0,j} space) is proven in [90]. Our contribution here is the observation that this result provides the solution to the 1-n joinability of isotropic states. The equivalence is established by the fact
30 2.3 Joining and sharing Werner and isotropic states that optimality is preserved in either direction by the Choi isomorphism. Quantitatively, the boundary for 1-n optimal asymmetric cloning, is given by Eq. (6) in [86] in terms of singlet fractions. Specializing to the 1-2 joining case and rewriting in terms of Φ+, we have
1 q q 2 Φ+ + Φ+ ≤ (d − 1) + Φ+ + Φ+ . AB AC n + d − 1 AB AC As one may verify, this is equivalent to the result of Corollary 2.3.4. In light of this connection, the fact that, as d increases, the isotropic-joinability cone of Fig. 2.1(right) becomes flattened down to the αAB-αAC plane is directly related to the linear trade-off in the isotropic state parameters for the semi-classical limit d → ∞, as discussed in [84]. Within our three-party joining picture, we can give a partial explanation of this fact: namely, it is a consequence of the classical joining boundary in tandem with the upper limit on the agreement parameter αBC for the Werner state on B-C: αBC ≤ 2/(d + 1). In the limit of d → ∞, these two boundaries conspire to limit the (A-B)-(A-C) isotropic state joining boundary to a triangle, as explicitly seen in Fig. 2.2b(right). For the general 1-n isotropic joining scenario, the quantum-cloning results addi- tionally imply the following: + + Theorem 2.3.6. A list of n isotropic states characterized by parameters Φ0,1,..., Φ0,j is 1-n joinable if and only if the (positive-valued) parameters satisfy
n n 2 X 1 X q Φ+ ≤ (d − 1) + Φ+ . (2.25) 0,j n + d − 1 0,j j=1 j=1
Interestingly, similar to our discussion surrounding Eq. (2.2), the authors of [86] argue how the “singlet monogamy” bound can lead to stricter predictions (e.g., on ground-state energies in many-body spin systems) than the standard monogamy of entanglement bounds based on CKW inequalities [38, 41].
2.3.5 Sharability of Werner and isotropic qudit states We next turn to sharability of Werner and isotropic states in d dimension, beginning from the important case of 1-n sharing. For Werner states, a proof based on a representation-theoretic approach is given in [61]. Although we expect a similar proof to exist for isotropic states, we obtain the desired 1-n sharability result by building on the relationship with quantum cloning problems highlighted above. We then outline a constructive procedure for determining the more general m-n sharability of Werner states.
31 Quantum marginals: sharability and joinability
Our main results are contained in the following: Theorem 2.3.7. A qudit Werner state with parameter Ψ− is 1-n sharable if and only if d − 1 Ψ− ≥ − . (2.26) n Theorem 2.3.8. A qudit isotropic state with parameter Φ+ is 1-n sharable if and only if d − 1 Φ+ ≤ 1 + . (2.27) n Proof. Specializing Eq. (2.25) to the case of equal parameters for all n isotropic states, the above result immediately follows. As stated in [86], this is consistent with the well known result for optimal 1-n symmetric cloning. We depict the qubit case (d = 2) of the above result in Fig. 2.3. In the case of Werner state sharing, Eq. (2.26) implies that a finite parameter range exists where the corresponding Werner states are not sharable. In contrast, for d ≥ 3, every − Werner state is at least 1-2 sharable. This simply reflects the fact that |ψd i (recall Eq. (2.7)) provides a 1-(d − 1) sharing state for a most-entangled qudit Werner state. With isotropic state sharing, for all d there is, again, a finite range of isotropic states which are not sharable. The simplicity of the results in Eqs. (2.26)-(2.27) is intriguing and begs for intu- itive interpretations. Consider a central qudit surrounded by n outer qudits. If the central qudit is in the same Werner or isotropic state with each outer qudit, then Theorems 2.3.7 and 2.3.8 can be reinterpreted as providing a bound on the sums of concurrences. For Werner states, we have that the sum of all the central-to-outer concurrences cannot exceed the number of modes by which the systems may disagree (i.e., d−1). In the isotropic state case, the sum of the n pairwise concurrences cannot p exceed the maximal concurrence value given by Cmax,d = 2(d − 1)/d. These rules do not hold in more general joining scenarios, as we already know from Sec. 2.3.3. There, we found that the trade-off between A-B concurrence and A-C concurrence is not a linear one, as such a simple “sum rule” would predict; instead, it traces out an ellipse (recall Fig. 2.2a). Starting from the proof of Thm. 2.3.7 found in [61] in conjunction with similar representation-theoretic tools, it is possible to devise a constructive algorithm for determining the m-n sharability of Werner states. The basic observation is to re- alize that the most-entangled m-n sharable Werner state corresponds to the largest eigenvalue of a certain Hamiltonian operator Hm,n, which is in turn expressible in terms of Casimir operators. Calculation of these eigenvalues may be obtained using Young diagrams. Although we lack a general closed-form expression for max(Hm,n), the required calculation can nevertheless be performed numerically. Representative results for n-m sharability of low-dimensional Werner states are shown in Table 2.1.
32 2.4 Further remarks
Figure 2.3: Pictorial summary of sharability properties of qubit Werner and isotropic states, according to Eqs. (2.26) and (2.27). The arrow-headed lines depict the pa- rameter range for which states satisfy each of the sharability properties displayed to the right and left, respectively. The vertical ticks between end points of these ranges indicate the points at which subsequent 1-n sharability properties begin to be satisfied.
2.4 Further remarks
2.4.1 Joinability beyond the three-party scenario
In Sec. 2.3, we focused on considering joinability of three bipartite (Werner or isotropic) states in a “triangular fashion”, namely, relatively to the simplest over- (3) lapping neighborhood structure N1 = {A, B}, N2 = {A, C} on H . In a more general N-partite scenario, other neighborhood structures and associated joinability problems may naturally emerge. For instance, we may want to answer the follow- ing question: Which sets of N(N − 1)/2 Werner-state (or isotropic-state) pairs are joinable? The approach to solving this more general problem parallels the specific three-party case we discussed. If each pair is in a Werner state, then if a joining state exists, there must exist a joining state with collective invariant symmetry (that is, invariant under arbitrary collective unitaries U ⊗N ). Thus, we need only look in the set of states respecting this symmetry. Any such operator may be decomposed into a sum of operators, which each have support on just a single irreducible subspace. This is useful because posi-
33 Quantum marginals: sharability and joinability
Table 2.1: Exact results for n-m sharability of Werner states for different subsystem dimension, with m and n increasing from left to right and from top to bottom in each table, respectively. For each sharability setting, the value −Φ is given. Asterisks correponds to entries whose values have not been explicitly computed.
n, m 1 2 3 4 5 1 1 1/2 1/3 1/4 1/5 2 1/2 1/2 1/3 1/4 1/5 d = 2 3 1/3 1/3 1/3 1/4 1/5 4 1/4 1/4 1/4 1/4 1/5 5 1/5 1/5 1/5 1/5 1/5
n, m 1 2 3 4 5 1 1 1 2/3 1/2 2/5 2 1 1/2 1/2 1/2 2/5 d = 3 3 2/3 1/2 1/3 1/3 1/3 4 1/2 1/2 1/3 1/4 1/4 5 2/5 1/3 1/3 1/4 ∗
n, m 1 2 3 4 5 1 1 1 1 3/4 3/5 2 1 1 2/3 1/2 1/2 d = 4 3 1 2/3 5/9 1/2 ∗ 4 3/4 1/2 1/2 ∗ ∗ 5 3/5 1/2 ∗ ∗ ∗ tivity of the joining operator when restricted to each irreducible subspace is sufficient for positivity of the overall operator. The joining operators may then be decomposed into the projectors on each irreducible subspace and corresponding bases of traceless operators on the projectors. The basis elements will be combinations of permuta- tion operators and the dimension of each such operator subspaces is given by the square of the hook length of the corresponding Young diagram [91]. The remaining task is to obtain a characterization of the positivity of the operators on each irre- ducible subspace. In [92], for example, a method for characterizing the positivity of low-dimensional operator spaces is presented. As long as the number of subsystems remains small, this approach grants us a computationally friendly characterization of positivity of the joining states. The bounds on the joinable Werner pairs may then be obtained by projecting the positivity characterization boundary onto the space of Werner pairs, analogous to the space of Fig. 2.2a. While a complete analysis is beyond our scope, a similar method may in principle
34 2.4 Further remarks
be followed to determine more general joinability bounds for isotropic states. How- ever, a twirling operation that preserves the joining property only exists for certain isotropic joining scenarios. For instance, we took this issue into consideration when we required the B-C system to be in a Werner state while A-B and A-C were isotropic states; it would not have been possible to take the same approach if all three pairs were isotropic states.
2.4.2 Sharability of general bipartite qubit states For qubit Werner states, one can use the methods of the proof of Thm. III.6 to show that 1-n sharability does imply n-n sharability [cf. Table I.(a)]. This property neither holds for Werner qudit states nor bipartite qubit states in general. The simplest example of a Werner state which disobeys this property is the most-entangled qutrit − Werner state ρ(Ψ = −1)d=3. This state is 1-2 sharable, as evidenced by the point (−1, −1, −1) lying within the bi-cone described by Eq. (2.15). The corresponding sharing state is the totally antisymmetric state on three qutrits as given by Eq. (2.7). This is the unique sharing state because the collective disagreement between the subsystems of each joined bipartite Werner state forces collective disagreement among the subsystems of the tripartite joining state; the totally antisymmetric state is the only quantum state satisfying this property. Since the only 1-2 sharing state − for ρ(Ψ = −1)d=3 is pure and entangled, clearly there can exist no 2-2 sharing. Additionally, we present below a counter-example that involves qubit states off the Werner line:
Proposition 2.4.1. For a generic bipartite qubit state ρ, 1-n sharability does not imply n-n sharability.
Proof. We claim that the following bipartite state on two qubits, 1 i ρ = |00i + |11i h00| + h11| + |10ih10| ≡ ρ , 3 L1R1 is 1-2 sharable but not 2-2 sharable. To show that ρ is 1-2 sharable, direct calculation shows that the two relevant partial-trace constraints uniquely identify w3 ≡ |ψihψ| as the only valid sharing state, with 1 |ψi ≡ √ (|000i + |101i + |110i). 3 The above state may in turn be equivalently written as
1 r2 1 |ψi = √ |0i ⊗ |00i + |1i ⊗ √ (|01i + |10i) . 3 3 2
35 Quantum marginals: sharability and joinability
In order for ρ to be 2-2 sharable, a four-partite state w4 must exist, such that Tr (w ) = ρ, for i, j = 1, 2. Any state which 2-2 shares ρ must then 1-2 share LˆiLˆj 4 the pure entangled state w3. That is, in constructing the 2-2 sharing state for ρ, we bring in a fourth system L2 which must reduce (by tracing over L1 or L2) to w3. But, since w3 is a pure entangled state, it is not sharable. Thus, there cannot exist a 2-2 sharing state for ρ. We conclude by stressing that our Werner and isotropic state sharability results allow in fact to put bounds (though not necessarily tight ones) on the sharability of an arbitrary bipartite qudit state. It suffices to observe that any bipartite state can be transformed into a Werner or isotropic state by the action of the respective twirling map (either Eq. (2.13) or (2.18)). Theorem 2.2.5 proves that the sharability of a state cannot be decreased by a unitary mixture map, and hence twirling cannot decrease sharability. This thus establishes the following:
Corollary 2.4.2. A bipartite qudit state ρ is no more sharable than the Werner state Z † † ρ˜ ≡ U ⊗ UρV U ⊗ U dµ(U), and the isotropic state Z ∗ T † ρ¯ ≡ U ⊗ UρV U ⊗ U dµ(U),
† for any ρV = I ⊗ V ρ I ⊗ V , with V ∈ U(d). In the qubit case, for instance, any maximally entangled pure state can be trans- formed into |Ψ−i or |Φ+i by the action of some local unitary I ⊗ V . Thus, all maximally entangled pure qubit states and their “pseudo-pure” versions, obtained as mixtures with the fully mixed states, have the same sharability properties as the Werner/isotropic states.
36 Chapter 3
Joinability of causal and acausal relationships
37 Joinability of causal and acausal relationships
This chapter presents material that appeared in Journal of Physics A: Mathemat- ical and Theoretical, 48:035307 (2015), in an article titled “On state versus channel quantum extension problems: exact results for U ⊗ U ⊗ U symmetry”, which is joint work with Lorenza Viola.
3.1 Introduction
It has long been appreciated that many of the intuitive features of classical probability theory do not translate to quantum theory. For instance, every classical probability distribution has a unique decomposition into extremal distributions, whereas a gen- eral density operator does not admit a unique decomposition in terms of extremal operators (pure states). Entanglement is responsible for another distinctive trait of quantum theory: as vividly expressed by Schrödinger back in 1935 [93], “the best possible knowledge of a total system does not necessarily include total knowledge of all its parts,” in striking contrast to the classical case. Certain features of classi- cal probability theory do, nonetheless, carry over to the quantum domain. While it is natural to view these distinguishing features as a consequence of quantum theory being a non-commutative generalization of classical probability theory in an appropri- ate sense, thoroughly understanding how and the extent to which the purely quantum features of the theory arise from its mathematical structure remains a longstanding central question across quantum foundations, mathematical physics, and quantum information processing (QIP), see e.g. Refs. [94, 95, 96, 97]. In this chapter, we investigate a QIP-motivated setting which allows us to directly compare and contrast features of quantum theory with classical probability theory, namely, the relationship between the parts (subsystems) of a composite quantum sys- tem and the system as a whole. Specifically, building on our earlier work [98], we develop and investigate a general framework for what we refer to as quantum join- ability, which addresses the compatibility of different statistical correlations among quantum measurements on different systems. Arguably, the most familiar case of joinability is provided by the “quantum marginal” (aka “local consistency”) problem [27, 99]. In this case, we ask whether there exists a joint quantum state compatible with a given set of reduced states on (typically non-disjoint) groupings of subsys- tems. The quintessential example of a failure of joinability is the fact that two pairs of two-level systems (qubits), say, Alice-Bob (A-B) and Alice-Charlie (A-C), cannot simultaneously be described by the singlet state, |ψ−i = p1/2(|↑↓i − |↓↑i). A semi- nal exploration of this observation was carried out by Coffman, Kundu, and Wootters [38] and later dubbed the “monogamy of entanglement” [40]. In classical probability theory, a necessary and sufficient condition for marginal probability distributions on A-B and A-C to admit a joint probability distribution (or “extension") on A-B-C is that the marginals over A be equal [27, 62]. The analogous compatibility condition
38 3.1 Introduction remains necessary in quantum theory, but, as demonstrated by the above example, is clearly no longer sufficient. The identification of necessary and sufficient conditions in general settings with overlapping marginals remains an actively investigated open problem as yet [61, 56, 100]. Physically, standard state-joinability problems as formulated above for density op- erators, may be regarded as characterizing the compatibility of statistical correlations of two (or more) different subsystems at a given time. However, correlations between a single system before and after the action of a quantum channel – a completely posi- tive trace-preserving (CPTP) dynamical map – may also be considered, for example, in order to characterize the “location” of quantum information that one subsystem may carry about another [101] and/or the causal structure of the events on which probabilities are defined [96, 102]. The work in [102] thoroughly explores, in par- ticular, the idea of placing kinematic and dynamic correlations on equal footing, by introducing a formalism of “quantum conditional states” to represent the correlations of either bipartite quantum states or quantum channels as bipartite operators. With these ideas in mind, one may want to formulate a quantum marginal problem for quantum channels (see also Ref. [103]). For example, given two quantum channels MAB : B(HA) → B(HB) and MAC : B(HA) → B(HC ) (with B(H) denoting the space of bounded linear operators on H), one may ask whether there exists a quan- tum channel MABC : B(HA) → B(HB ⊗ HC ), whose reduced channels are MAB and MAC , respectively. A motivation for considering such channel-joinability problems is that questions regarding the optimality of paradigmatic QIP tasks such as quantum cloning [84, 85] or broadcasting [104] may be naturally recast as such. A fundamental tool here is the Choi-Jamiolkowski isomorphism [105, 89], which may been used to translate optimal cloning problems into quantum marginal problems [86, 106], and vice-versa [61]. Both monogamy of entanglement and the no-cloning theorem [14] have significant impli- cations for the behavior of quantum systems: the former effectively constrains the kinematics of a multipartite quantum system, while the latter constrains the dynam- ics of a quantum system (composite or not). As both of these fundamental concepts are closely related to respective quantum joinability problems, we are prompted to explore in more depth their similarities and differences. Identifying a general join- ability framework, able to encompass all such quantum marginal problems, is one of our main aims here. The content is organized as follows. In Section 3.2, we introduce and motivate the use of what we term the homocorrelation map as our main tool for representing quantum channels as bipartite operators. Despite the different motivation, this repre- sentation will share suggestive points of contact with the conditional-state formalism of [102]. Formally, we show how it enables a notion of quantum joinability that in- corporates all joinability problems of interest, and discuss ways in which different
39 Joinability of causal and acausal relationships
joinability problems may be (homomorphically) mapped into one another. In Section 3.3, we obtain a complete analytical characterization of some archetypal examples of low-dimensional quantum joinability problems. Namely, we address three-party join- ability of quantum states, quantum channels, and block-positive (or “local-positive”) operators, in the case that the relevant operators are invariant under the group of collective unitary transformations, that is, under the action of arbitrary transforma- tions of the form U ⊗ U ⊗ U. These examples allow us to distinguish the joinability limitations stemming from classical probability theory from those due to quantum theory and, furthermore, to contrast the joinability properties of quantum channels vs. states. In Section 3.4, we investigate a possible source for the stricter joinability bounds in quantum theory, as compared to classical probability theory. We introduce the notion of degree of agreement (disagreement), that is, the probability that a ran- dom local collective measurement yields same (different) outcomes, as given by an appropriate two-value POVM. We find that quantum theory places different bounds on the degree of agreement arising from quantum states than it does on that of quantum channels: while quantum states are limited in their degree of agreement, quantum channels are limited in their degree of disagreement. The differences in these bounds point to a crucial distinction between quantum channels and states. At least in the examples of Section 3.3 and a few others, these limitations suffice in fact to determine the bounds of joinability exactly. Possible implications of such bounds with regards to joinability properties of general quantum states and channels are also discussed.
3.2 General quantum joinability framework
We begin by reviewing the standard state-joinability (quantum marginal) problem, framing it in a language suitable for generalization. Given a composite Hilbert space (N) NN H = i=1 Hi, a joinability scenario is defined by a list of partial traces {Tr`k }, with each `k ⊆ [1,...,N], along with a set of allowed “joining operators,” W , which in this case is the set of positive trace-one operators acting on H(N); accordingly, we may
associate a joinability scenario with a 2-tuple (W, {Tr`k }). For a given joinability
scenario, the images of W under the Tr`k define a set of reduced states {Rk} = {Trk(W )}. For any list of states {ρk} ∈ {Rk}, the following definition then applies:
Definition 3.2.1. [State-Joinability] Given a joinability scenario described by the
pair (W ≡ {w| w ≥ 0}, {Tr`k }), the reduced states {ρk} ∈ {Rk} are joinable if there
exists a joining state w ∈ W such that Tr`k (w) = ρk for all k.
The first step toward achieving the intended generalization of the above definition to quantum channels is to represent the latter as bipartite operators. In the following subsection, we establish a tool to achieve this and highlight its broader utility.
40 3.2 General quantum joinability framework
3.2.1 Homocorrelation map and positive cones
One way to identify channels with bipartite operators is by use of the Choi-Jamiolkowski (CJ) isomorphism [89, 107]. This isomorphism, denoted J , identifies each map M ∈ L(HA, HB) with the state resulting from the map’s action on one member of a maximally entangled state:
1 X J (M) ≡ [I ⊗ M](|Φ+ihΦ+|) = |iihj| ⊗ M(|iihj|), (3.1) A d A ij √ + P where IA is the identity map on B(HA), |Φ i = i |iii/ dA and dA = dim(HA). + + TA We note that dA|Φ ihΦ | = V , where V is the swap operator on HA ⊗ HA and TA denotes partial transposition on subsystem A. The transformation is an isomorphism in that it preserves the positivity of the objects it maps to and from; namely, quantum channels (CPTP maps) are mapped to quantum states (positive trace-one operators). Consequently, J is a useful diagnostic tool for determining whether a map is CP 1. Here, we employ an alternative means of identifying quantum channels with bi- partite operators, building on an identification that was introduced for the special case of qubits in [108]. In this approach, basis-dependence is avoided by replacing the reference state with the normalized swap operator V/d. Since the latter is not a den- sity operator, this correspondence lacks an interpretation as a physical process. But, for our purposes, this comes at the greater benefit of yielding a resulting bipartite operator that bears the statistical properties of the corresponding channel. Formally, we define a homocorrelation map, H, which takes any map M ∈ L(HA, HB) (with L(HA, HB) being the set of linear maps, or “superoperators”, from B(HA) to B(HB)), to a “channel operator” MH ∈ B(HA ⊗ HB) according to
1 X H(M) ≡ [I ⊗ M](V/d ) = |iihj| ⊗ M(|jihi|), (3.2) A A d A ij
P 2 where V = i,j |ijihji| with respect to any orthonormal basis {|ii} . While the CJ isomorphism is a handy diagnostic tool, the homocorrelation map serves a different purpose. It does not take CP maps to positive operators. Instead, it takes each map to an operator which exhibits the same statistical correlations as that map:
1 + Note that J depends on a choice of local basis, needed to define |Φ i and TA. In order for the isomorphism to hold, the reference state (|Φ+ihΦ+| above) must be maximally entangled; again for d > 2, any such state reflects a choice of local bases. 2The homocorrelation map H is closely related to the causal conditional states defined in [102]. Namely, a channel operator resulting from the homocorrelation map is precisely the causal condi- tional state conditioned on ρA being the completely mixed state. Note that Eq. (3.2) above differs from Eq. (29) in [102] by a factor of dA.
41 Joinability of causal and acausal relationships
(a) (b)
Figure 3.1: (a) Commutativity diagram summarizing the relationship between the Choi-Jamiolkowski isomorphism and the homocorrelation map defined in Eqs. (3.2)- (3.1). In (b), the corresponding actions are given in terms of tensor network diagram notation [109]. Proposition 3.2.2 may be straightforwardly proved using this notation.
Proposition 3.2.2. A bipartite state ρ ∈ B(HA ⊗ HB) and a quantum channel M : B(HA) → B(HB) exhibit the same correlations, that is, 1 Tr[ρA ⊗ B] = Tr[M(A)B], ∀A ∈ B(HA),B ∈ B(HB). (3.3) dA if and only if the equality H(M) = ρ holds.
Proof. The two operators ρ and H(M) are equal if and only if their expectations Tr[ρA⊗B] = Tr[H(M)A⊗B] for all A, B. Thus, it suffices to show that Tr[H(M)A⊗ B] = 1 Tr[M(A)B] for all A, B. This equality may be established as follows: dA
Tr[H(M)A ⊗ B] = 1 P Tr[|iihj| ⊗ M(|jihi|)A ⊗ B] dA i,j = 1 P Tr[(|iihj|A) ⊗ (M(|jihi|)B)] dA i,j = 1 P hj|A|iiTr[M(|jihi|)B] = 1 Tr[M(A)B]. 2 dA i,j dA
Equation (3.3) may be taken as the defining property of the homocorrelation map. An example may explicitly demonstrate the utility of this representation. Consider the one-parameter family of qudit depolarizing channels [17], defined as
I D (ρ) = (1 − η)Tr(ρ) + ηρ. (3.4) η d
† The action of this channel commutes with all unitary channels in that Dη(UρU ) = † UDη(ρ)U . Under the homocorrelation map, the depolarizing channels are taken to
42 3.2 General quantum joinability framework operators with U ⊗ U symmetry, namely, I ⊗ I V H(D ) = (1 − η) + η , (3.5) η d2 d where V is, again, the swap operator. Trace-one, positive operators of this form are the well-known Werner states [57] (see also Sec. 3.3.1). Imagine that an observer does not know a priori whether her two measurements are made on distinct systems in a Werner state or if they are made on the same system before and after a depolarizing channel has been applied. If presented with a Werner state or depolarizing channel 1 1 having η = − d2−1 to d+1 , the observer will not be able to distinguish between the two cases. The homocorrelation map makes this operational identification explicit. To contrast, the CJ map takes the depolarizing channels to so-called isotropic states [80], I ⊗ I J (D ) = (1 − η) + η|Φ+ihΦ+|, (3.6) η d2 where as before |Φ+i is the maximally entangled state. The isotropic states are defined by their symmetry with respect to U ⊗ U T transformations. An observer in the scenario above would certainly be able to distinguish between the correlations of the depolarizing channel and the isotropic states, as long as η 6= 0. The distinction between the CJ isomorphism and the homocorrelation map can be further appreciated by contrasting the sets of operators they produce. The set of CP maps forms a cone in the set of superoperators L(HA, HB). Both the CJ isomorphism and the homocorrelation map are cone-preserving maps (by linearity) from L(HA, HB) to B(HA ⊗HB). While in the case of the CJ isomorphism, the resulting cone is exactly the cone of bipartite states, in the case of the homocorrelation map, the cone is distinct from the cone of states. One of the main findings of this chapter is that the correlations exhibited by bipartite states and the ones exhibited by quantum channels need not be equivalent. Furthermore, we find that this difference plays a role in their distinct joinability properties. The homocorrelation representation of channels provides us with a natural framework for exploring this difference: a channel and a state with differing correlations will be represented as distinct operators in the same operator space; as an example, the classes of bipartite Werner states and depolarizing channels are depicted in Fig. 3.2. These notions and their use in joinability are fleshed out in what follows. The cone of positive operators plays a central role in defining joinability of quan- tum states. Analogously, the cone of homocorrelation-mapped channels (or “channel- positive operators”) will play a central role in defining joinability of quantum channels.
Definition 3.2.3. [State-positivity] An operator M ∈ B(H) is state-positive if Tr(MP ) ≥ 0 for all Hermitian projectors P = P † = P 2 ∈ B(H). We notate this
43 Joinability of causal and acausal relationships
condition as M ≥st 0 and emphasize that the resulting set is a self-dual cone. Recall that a map M is a valid quantum channel if Tr[J (M)P ] ≥ 0 for all P = P 2 ∈ B(HA ⊗ HB) [89]. Using the relationships of Fig. 3.1, we translate this condition to one on the homocorrelation-mapped operator M = H(M). Specifically, we define:
Definition 3.2.4. [Channel-positivity] An operator M ∈ B(HA ⊗HB) is channel- positive with respect to the A-B bipartition if Tr(MP TA ) ≥ 0 for all Hermitian projec- † 2 tors P = P = P ∈ B(HA ⊗ HB). We notate this condition M ≥ch 0, and emphasize that the resulting set is, again, a self-dual cone.
V I
I – V/d
θ θ
0 T r M = 1 Figure 3.2: State- and channel-positive cones for two qudit Werner operators. The re- gion of the solid arc (blue) corresponds to state-positive operators, while the region of the dashed arc (pink) corresponds to channel-positive operators. The overlapping re- gion, seen as purple, corresponds to PPT operators; of these, the normalized operators are also unentangled state-positive operators. The self-dual nature of the state- and channel-positive cones is consistent with the right angles of each cone’s vertex. The 1 Young diagrams represent the corresponding projectors into the symmetric 2 (I + V ) 1 and antisymmetric 2 (I − V ) subspaces, respectively. For qudit dimension d, the angle θ is calculated to be cos θ = Tr[V (I + V )]/pTr[V 2]Tr[(I + V )2] = p(d + 1)/2d.
In the general case, we can give a characterization of the intersection of the two cones and their complements. This is aided by the fact that the CJ isomorphism and the homocorrelation map are related to one another by partial transpose. A commutivity diagram of these relationships is given in Fig. 3.1, where the tensor network diagram calculus [109] may be used to concisely demonstrate that, up to normalization, −1 −1 J ◦ H = H ◦ J = TA.
44 3.2 General quantum joinability framework
Proposition 3.2.5. A bipartite state ρ ∈ B(HA ⊗ HB) and a quantum channel M : B(HA) → B(HB) exhibit the same correlations if and only if the density operator (or equivalently, channel operator) has a positive partial transpose (PPT). Proof. By Prop. 3.2.2, if a bipartite state and a quantum channel exhibit the same correlations, then ρ = H(M). Since J is related to H by a partial trace, we also TA have H(M) = J (M)/dA. By the positivity preservation of J , M being CPTP TA implies that J (M) is a positive operator. Thus, we have that ρ = J (M)/dA is positive. This result may be used to directly connect quantum channels to entanglement:
Corollary 3.2.6. If the correlations of a bipartite state ρ ∈ B(HA ⊗ HB) cannot be exhibited by a quantum channel, then the state is entangled. Proof. Since the correlations cannot be exhibited by a quantum channel, the operator is not PPT, by Prop. 3.2.5. Then, by the Peres-Horodecki criterion [110], the state is necessarily entangled.
3.2.2 Generalization of joinability We are now poised to use the homocorrelation representation to define the joinability of channels. The channel-positive operators provide an alternative set with which to define the allowed joining operators W . As a warm-up, we rephrase the channel- joinability problem that was posed in the Introduction. Consider quantum channels from HA to HB ⊗HC . Under the homocorrelation map, these correspond to tripartite operators lying in the channel-positive cone, notated WA|BC . The partial traces TrC and TrB take channel-positive operators in WA|BC to channel-positive operators in WA|B and WA|C , respectively; that is, operators in WA|B and WA|C correspond to valid quantum channels via the homocorrelation map. The corresponding channel- joining scenario is then defined as (WA|BC , {TrC , TrB}). A channel-joinability problem presents two channel operators MAB ∈ WA|B and MAC ∈ WA|C and seeks to determine the existence of a channel operator MABC ∈ WA|BC which reduces to the two channel operators in question. In general, we thus have the following: Definition 3.2.7. [Channel-Joinability] Given a joinability scenario described by the pair (W ≥ch 0}, {Tr`k }), the reduced operators {Mk} ∈ {Rk} are joinable if there exists a joint operator M ∈ W such that Tr`k (M) = Mk for all k. We note that a channel joinability (or extension) problem can be stated using the CJ isomorphism instead of the homocorrelation map, as done in [103]. However, as we argued, the homocorrelation map provides a platform to directly compare the joinability of states and channels of equivalent correlations. For instance, it will allow
45 Joinability of causal and acausal relationships
us to simultaneously compare the joinability of local-unitary-invariant quantum states and channels, and consequently to compare these both to the joinability of analogous classical probability distributions (c.f. Fig. 3.5). Before proceeding to the general notion of joinability, we also remark that allowed joining operators in W have thus far been considered to be either state-positive or channel-positive. However, from a mathematical standpoint, a sensible joinability problem only needs W to be a convex cone. To investigate this generalization and (as motivated later) to meld state and channel joining, we consider a third type of positivity that we call local-positivity. This notion is equivalent to both block- positivity [111] and to map-positivity (not necessarily CP) [112, 113], in that by representing linear maps using the homocorrelation map, the cone of (transformed) positive maps is equal to the cone of bipartite block-positive operators. Formally:
Definition 3.2.8. [Local-positivity] An operator M ∈ B(HA⊗HB) is local-positive with respect to the A-B factorization if Tr(MPA ⊗ PB) ≥ 0 for all pure states PA = 2 2 PA ∈ B(HA) and PB = PB ∈ B(HB). We notate this condition M ≥loc 0. The set of channel-positive operators and state-positive operators are each sub-cones of the local-positive operators, as local-positivity clearly is a weaker condition. Local- positive operators are directly relevant to QIP, in particular because they may serve as an entanglement witnesses [114]. Moreover, in comparing quantum joinability limita- tions to analogous limitations stemming from classical probability theory, joinability scenarios defined with respect to W ≥loc 0 may allow the identification of quantum limitations in a “minimally constrained" setting, closer to the (less strict) classical boundaries. In Sec. 3.3.2, we find that local-positivity does nevertheless provide stricter-than-classical limitations on joinability. Another way of viewing the various definitions of positivity is to understand the subscript on the inequality to indicate the dual cone from which inner products with M must be positive. For M ≥st 0, M ≥ch 0, and M ≥loc 0, the respective dual cones are the positive span of rank-one projectors, the positive span of partially- transposed projectors, and the positive span of product projectors (from which the trace-one condition confines to the set of separable states). We note that the first two cones are self-dual (and, furthermore, symmetric [115]), while the local-positive cone is not. With several important examples of positivity established, each being a different convex set with which to define W , we are in a position to give the following: Definition 3.2.9. [General Quantum Joinability] Let W be a convex cone in (N) B(H ), and {Tr`k } be partial traces with `k ⊂ ZN . Given the joinability scenario
(W, {Tr`k }), the operators {Mk} ∈ {Rk} are joinable if there exists a joining operator
w ∈ W such that Tr`k (w) = Mk for all k. This general definition naturally encompasses the various joinability problems ref- erenced in the Introduction. Specifically, in the case where W is the set of quan-
46 3.2 General quantum joinability framework
tum states on a multipartite system, the joinability problem reduces to the quan- tum marginal problem, while if W consists of channel-positive operators describing quantum channels from one multipartite system to another, one recovers the channel- joining problem instead. Specific instances of this problem are the optimal asymmet- ric cloning problem [84, 85, 88], the symmetric cloning problem [116, 117], and the k-extendibility problem for quantum maps [118]. In addition to providing a unified perspective, our approach has the important advantage that different classes of join- ability problems may be mapped into one another, in such a way that a solution to one provides a solution to another. This is made formal in the following:
Proposition 3.2.10. Let W and W 0 be two positive cones of operators acting on (N) the space H , let {Tr`k } be a set of partial traces that apply to both cones, and let φ : W → W 0 be a positivity-preserving (homo)morphism, which permits reduced actions φk satisfying φk ◦ Tr`k = Tr`k ◦ φ. If {Mk} ∈ {Tr`k (W )} is joinable with 0 0 respect to W , then {φk(Mk)} ∈ {Tr`k (W )} is joinable with respect to W .
Proof. Assume that w is a valid joining operator for the operators {Mk} ∈ {Tr`k (W )}. 0 Then, the set of operators {φk(Mk)} ∈ {Tr`k (W )} is joined by the operator φ(w), 0 since Tr`k [φ(w)] = φk(Tr`k [w]) = φk(Mk) and φ(w) ∈ W .
This is shown in the commutative diagram of Fig. 3.3. We shall use a stronger corollary of this result in the remaining sections:
Corollary 3.2.11. Let φ be a one-to-one positivity-preserving map from W to W 0,
with invertible reduced actions φk satisfying φk ◦ Tr`k = Tr`k ◦ φ (and similarly for
their inverses). Then a set of operators {Mk} ∈ {Tr`k (W )} is joinable if and only if the set of operators φk(Mk) is joinable.
Proof. The forward implication follows from Proposition 3.2.10, while the backwards implication follows from the fact that φ and the φk are invertible, along with the contrapositive of Proposition 3.2.10.
The joinability-problem isomorphism we make use of is the partial transpose map. The latter permits a natural reduced action, namely, partial transpose on the remain- ing of the previously transposed subsystems. As explained, the partial transpose is a positivity-preserving bijection between states and channel operators (via H). Thus, if we determine the joinable-unjoinable demarcation for a class of states, we will determine the joinable-unjoinable demarcation for a corresponding class of channel- operators.
47 Joinability of causal and acausal relationships
Figure 3.3: Commutivity diagram showing a homomorphism of joinability problems.
3.3 Three-party joinability settings with collective invariance
In this Section, we obtain an exact analytical characterization of the state-joining, channel-joining, and local-positive joining problems in the three-party scenario, by taking advantage of collective unitary invariance. That is, we determine what trio of bipartite operators MAB, MAC , and MBC may be joined by a valid joining operator wABC , subject to the appropriate symmetry constraints. As noted, the most familiar case is state joinability, whereby the bipartite operators along with the joining tri- partite operator are state-positive. The next case considered is referred to as “1-2 channel joinability”: here, we specify a bipartition of the systems (say, A|BC) and consider the bipartite operators which cross the bipartition (MAB and MAC ), along with the joining operator, to be channel-positive with respect to the bipartition, while the remaining bipartite operator (MBC ) is state-positive. Since each of the three possible bipartition choices (A|BC, B|AC, and C|AB) constitutes a different channel joinability scenario, a total of four possibilities arise for three-party state/channel joinability. Lastly, motivated by the suggestive symmetry arising from these results and their relation to classical joining, we consider the weaker notion of local-positive joining, in which all operators involved are only required to be local-positive.
3.3.1 Joinability limitations from state-positivity and channel- positivity
We begin by describing the operators which are to be joined. The bipartite reduced operators inherit the collective unitary invariance from the tripartite operators from which they are obtained. Therefore, by a standard result of representation theory [119], the operators which are to be joined are of the following form:
I V ρ(η) = (1 − η) + η , (3.7) d2 d 48 3.3 Three-party joinability settings with collective invariance
where V is the swap operator defined earlier. The above operators are known to 1 1 be state-positive for the range − d−1 ≤ η ≤ d+1 , corresponding to the d-dimensional (qudit) Werner states we already mentioned. The parameterization is chosen so that η is a “correlation” measure: if d = 2, η = −1 corresponds to the singlet state, η = 0 to the maximally mixed state, while η = 1 is not a valid quantum state, but expresses perfect correlation for all possible collective measurements. Note that a value η = 1, for instance, does correspond to a valid quantum channel. Intuitively, channel-positive operators with U ⊗ U-invariance correspond to depolarizing channels. It is known that complete positivity (or channel-positivity) of the depolarizing channel is ensured 1 provided that − d2−1 ≤ η ≤ 1 [120]. However, we find it instructive to independently establish state- and channel- positivity bounds using the CJ isomorphism. To this end, we enlarge the above class of U ⊗ U-invariant operators to the class of operators with collective orthogonal invariance, namely, invariance under transfor- mations of the more general form O ⊗ O, belonging to the so-called Brauer algebra [121, 122]3. In addition to U ⊗ U-invariant operators, the Brauer algebra also con- tains U ∗ ⊗ U-invariant operators. The latter class of operators, which includes the well-known isotropic states, are spanned by the operators I and V TA . Thus, the set of O ⊗ O-invariant operators are of the form
I V V TA ρ(η, β) = (1 − η − β) + η + β . (3.8) d2 d d In particular, the operator ρ(0, 1) is a generic Bell state on two qudits, ρ(0, −1/(d−1)) is the maximally entangled Werner state (namely, the singlet state for d = 2), ρ(1, 0) is the identity channel, and ρ(0, 0) is the completely mixed state (or the completely depolarizing channel). We can then establish the following:
Proposition 3.3.1. A bipartite operator ρ(η) with collective unitary invariance is 1 channel-positive if and only if − d2−1 ≤ η ≤ 1. Proof. The Brauer algebra includes all the state-positive operators which are mapped, via the CJ isomorphism J , to the U ⊗ U-invariant channel-positive operators; under J , η and β in Eq. (3.8) are swapped with one another. Since J takes state-positive operators to channel-positive operators, we need only obtain the set of state-positive operators. State-positivity of these operators is enforced by the inner products with
3Operators in this algebra have been extensively analyzed in [123, 124], and recent work char- acterizing their irreducible representations may be found in [125, 126]. The Brauer algebra act- ing on N d-dimensional Hilbert spaces is spanned by representations of subsystem permutations {Vπ|π ∈ SN }, along with their partial transpositions with respect to groupings of the subsystems Tl {Vπ |π ∈ SN , l ⊆ {1,...,N}}. In terms of tensor network diagrams, each element of this basis is represented by a set of disjoint pairings of 2N vertices, with the vertices arranged in two rows, both containing N of them.
49 Joinability of causal and acausal relationships
respect to their (operator) eigenspaces being non-negative. Such eigenspaces are PA, P+, and PY , independent of η and β: the first is the anti-symmetric subspace, the second is the one-dimensional space spanned by |Φ+i, and the third is the space spanned by vectors |yi satisfying hy|(|yi)∗ = 04. The eigenvalues are as follows:
2 ρ(η, β)PA = [(1 − η − β)/d − η/d]PA, 2 ρ(η, β)P+ = [(1 − η − β)/d + η/d + β]P+, 2 ρ(η, β)PY = [(1 − η − β)/d + η/d]PY .
Hence, state-positivity of the bipartite Brauer operators is ensured by
1 ≥ (d + 1)η + β, 1 ≥ −(d − 1)η − (d2 − 1)β, 1 ≥ −(d − 1)η + β.
The inequalities bounding channel-positivity are obtained by swapping the ηs and βs. In particular, we recover that the state-positive range for U ⊗ U-invariant operators 1 1 1 is − d−1 ≤ η ≤ d+1 , whereas the channel-positive range is − d2−1 ≤ η ≤ 1. In a similar manner, we can also obtain the ranges of local-positivity: Proposition 3.3.2. A bipartite operator ρ(η) with collective unitary invariance is 1 local-positive if and only if − d−1 ≤ η ≤ 1. Proof. Local positivity is ensured by the non-negativity of expectation values with respect to the product vectors {|xxi, |xx¯i, |yyi, |yy¯i}, satisfying |xi∗ = |xi and |yi∗ = |y¯i, where the bar indicates a vector orthogonal to the original vector. In terms of η and β, these constraints read
2 0 ≤ hρ(η, β)ixx = (1 − η − β)/d + η/d + β/d, 2 0 ≤ hρ(η, β)ixx¯ = (1 − η − β)/d , 2 0 ≤ hρ(η, β)iyy = (1 − η − β)/d + η/d, 2 0 ≤ hρ(η, β)iyy¯ = (1 − η − β)/d + β/d.
More compactly, these boundaries are given by 1 − ≤ η + β ≤ 1, −(d − 1)η + β ≤ 1, η − (d − 1)β ≤ 1. d − 1 Thus, for bipartite Brauer operators, local-positive operators are equivalent to convex combinations of state- and channel-positive operators5. In particular, for the local-
4As noted, both the definition of |Φ+i and the use of complex conjugation are basis-dependent notions. It is understood that all usages of either refer to the same (arbitrary) choice of basis. 5This property is known as decomposability [112]. Interestingly, such an equivalence also holds for arbitrary bipartite qubit states [127].
50 3.3 Three-party joinability settings with collective invariance
(a) (b)
Figure 3.4: Positivity regions for bipartite Brauer operators: (a) d = 2. (b) d = 5. The solid triangle (blue) encloses the state-positive region, the dashed triangle (pink) encloses the the channel-positive region, and the outer boundary encloses the local- positive region.
1 positive range of U ⊗ U-invariant operators, it follows that − d−1 ≤ η ≤ 1, as stated.
A pictorial summary of the three positivity bounds is presented in Fig. 3.4. Having characterized all types of positivity for the (bipartite) operators to be joined, we now turn to characterize the positivity for the (tripartite) joining opera- tors. For each positive tripartite set (W ≥st 0, W ≥ch 0, and W ≥loc 0), we obtain the trios of joinable bipartite operators by simply applying the three partial traces (TrA, TrB, TrC ) to each positive operator. In more detail, our approach is to obtain an expression for the positivity boundary of the tripartite operators in terms of operator space coordinates, and then re-express this boundary in terms of reduced-state param- eters (the three Werner parameters in this case). For state- and channel-positivity, the desired characterization follows directly from the analysis reported in our previous work [61].
Corollary 3.3.3. With reference to the parameterization of Eq. (3.7), we have that: (i) Three Werner states with parameters (ηAB, ηAC , ηBC ) are joinable with respect to the (WABC ≥st 0, {TrA, TrB, TrC }) scenario if and only if
1 2 i2π/3 2 (1 − ηAB − ηAC − ηBC ) ≥ |ηAB + ωηAC + ω ηBC |, ω ≡ e , ηAB + ηAC + ηBC ≥ −1,
for d = 2, while for d ≥ 3 they need only satisfy
2 |ηAB + ωηAC + ω ηBC |.
51 Joinability of causal and acausal relationships
(ii) Three U ⊗ U-invariant operators with parameters (ηAB, ηAC , ηBC ) are channel- joinable with respect to the (WA|BC ≥ch 0, {TrA, TrB, TrC }) scenario if and only if r 1 2 2d iθ −iθ + ηAB + ηAC − ηBC ≥ + dηBC + (e ηAB + e ηAC ) , d − 1 d − 1 d − 1 eiθ ≡ p(d − 1)/2d ± ip(d + 1)/2d.
The channel-joinability limitations in the other two scenarios B|AC and C|AB may be obtained by permuting the ηs accordingly.
Proof. Result (i) corresponds to Theorem 3 in [61], re-expressed in terms of the parametrization of Eq. (3.7) (with reference to the notation of Eqs. (15)–(17) in [61], 2 − one has η` = (d/(d − 1))(Ψ` − 1/2), ` = AB, AC, BC). In order to establish (ii), note that J may be used to translate any U ∗⊗U-invariant state-positive joinability problem into a U ⊗ U-invariant channel-positive joinability problem, drawing on Corollary 3.2.11. Explicitly, under J (partial transpose in the ∗ case of operators), the U ⊗ U ⊗ U-invariant state-positive operators WA∗BC are in one-to-one correspondence with the U ⊗ U ⊗ U-invariant channel-positive operators WA|BC . Hence, by the joinability isomorphism induced by the partial transpose, the solution to a joinability problem of the scenario (WA∗BC , {TrA, TrB, TrC }) gives a solution to a corresponding joinability problem of (WA|BC , {TrA, TrB, TrC }). Thus, to obtain the depolarizing channel-joinability boundaries, we simply translate the isotropic state parameters of Eqs. (20)-(21) in [61] into η parameters.
The joinability limitations of all four scenarios are depicted in Fig. 3.5. As stressed in [61], the quantum joinability limitations must adhere to the analogous classical joinability limitations (seen as the tetrahedra in Fig. 3.5). In the qubit case, we find it intriguing that the inclusion of the quantum channel-joinability limitations allows us to regain the tetrahedral symmetry imposed by the classical limitations; whereas each scenario on its own expresses a continuous rotational symmetry that is not reflected classically. In other words, if we consider the joinability scenario defined by (span{WABC ,WA|BC ,WB|AC ,WC|AB}, {TrA, TrB, TrC }), the joinable bipartite operators respect the tetrahedral symmetry suggested by the classical joinability bounds. This amounts to asking the question: what trios of bipartite correlations – as derivable from either quantum states or channels, or from probabilistic combinations of the two – may be obtained from the measurements on three systems? Though the result expresses the tetrahedral symmetry of the classical joinability limitations, these classical joinability limitations do not suffice to enforce the stricter quantum joinability limitations, as manifest in the fact that the corners
52 3.3 Three-party joinability settings with collective invariance
(a) (b)
Figure 3.5: Joinability of operators on the β = 0 line for (a) d = 2 and (b) d = 5 (each from a different perspective). State-positivity, along with channel-positivity with respect to each of the three bipartitions, obtains the four cones depicted here. The joinability limitations for classical probability distributions are given by (a) the tetrahedron with black edges and (b) the union of the two tetrahedra with black edges. of the classical joinability tetrahedron are not reached by the quantum boundaries. We diagnose such limitations as strictly quantum features that do not have classical analogues – as we will discuss later in this work.
3.3.2 Joinability limitations from local-positivity
We now explore how local-positive joinability (a strictly weaker restriction, as noted) relates to the state/channel-joinability limitations above, as well as to the underlying classical limitations. As of yet, we only know that the local-positive limitations will lie between the classical and the quantum boundaries. Since obtaining a simple analytical characterization for arbitrary subsystem dimension d appears challenging in the local-positive setting, and useful insight may already be gained in the lowest- dimensional (qubit) setting, we focus on d = 2 in this section. Our main result is contained in the following:
Theorem 3.3.4. With reference to Eq. (3.7), three qubit Werner operators (con- strained by local-positivity) with parameters (ηAB, ηAC , ηBC ) are joinable by a local-
53 Joinability of causal and acausal relationships
Figure 3.6: Boundary of local-positive Werner operators which are joinable via local- positive operators, as described by the Roman surface, see Theorem 3.3.4.
positive tripartite Werner operator w if an only if the following conditions hold:
1 + ηAB + ηAC + ηBC ≥ 0, 1 + ηAB − ηAC − ηBC ≥ 0,
1 − ηAB + ηAC − ηBC ≥ 0, 1 − ηAB − ηAC + ηBC ≥ 0,
and
2 2 2 2 2 2 2ηABηAC ηBC − ηABηAC − ηABηBC − ηAC ηBC ≥ 0.
The proof is rather lengthy and can be found in the appendix of [98]. The resulting boundary is depicted in Fig. 3.6; the shape and its determining equation is recognized as the convex hull of the Roman surface (aka Steiner surface) [112, 128]. Comparing with Fig. 3.5(a), we see that, still, the quantum joinability limitation arising from from local-positivity is stricter than the corresponding classical one. However, it is closer to the classical limitations than the state/channel-positive limitations obtained in the previous section for d = 2. To shed light on the cause of the quantum bound- ary here, we can explicitly construct a product-state projector, whose probability would be negative if joinability outside of this shape were allowed. The family of joining states w that we need to consider (see appendix of [98] for details) may be parameterized in terms of the bipartite reduced state Werner parameters as 1 η η η w(η , η , η ) = I + AB (V − I/2) + AC (V − I/2) + BC (V − I/2). AB AC BC 8 4 AB 4 AC 4 BC 54 3.3 Three-party joinability settings with collective invariance
Consider, in particular, the following state on A-B-C:
1 cos 2π/3 cos 4π/3 |ψi = ⊗ ⊗ , (3.9) 0 sin 2π/3 sin 4π/3
which corresponds to the pure product state with the local Bloch vectors as anti- parallel with one another as possible. Computing its expectation with respect to w(ηAB = ηAC = ηBC ≡ η), the largest value of η that admits a non-negative value is η = 2/3. Hence, local-positivity limits the simultaneous joining of these Werner operators to a maximum of η = 2/3. The operational interpretation of this result deserves attention. Consider a local projective measurement made on each of three qubit systems. Furthermore, consider the three systems to have a collective unitary symmetry, in the sense that there are no preferred local bases. In our general picture, where local positivity is considered, the systems need not be three distinct systems – they may also be the same system at two different points in time. Local positivity enforces the rule that “all probabilities arising from such measurements must be non- negative”. In the example above (i.e. ηAB = ηAC = ηBC ), this rule implies that the three equal correlations (as measured by the ηs) can never exceed 2/3. As this example and Fig. 3.6 show, local-positivity enforces joinability limitations more strict than those of classical probability theory. Notwithstanding, these limitations reflect the same symmetry as the classical limitations do, namely, symmetry with respect to individually inverting two axes. The state-joining and channel-joining scenarios reflected a preference towards the negative axis (anticorrelation) and the positive axis (correlation) of the ηs, respectively.
Before concluding this section, we connect the above discussion to the relationship between local-positivity and separability. As mentioned earlier, the cone of local positive operators and the cone of separable operators are dual to one another. The operator subspace we are dealing with is spanned by the orthonormal operators √1 I, 8 q √1 (V − I/2), √1 (V − I/2), and √1 (V − I/2) with coordinates √1 , 3 η , 6 AB 6 AC 6 BC 8 8 AB q 3 q 3 8 ηAC , and 8 ηBC , respectively. In Theorem 3.3.4, we determined the algebraic surface bounding the local positive operators; hence, the dual to this surface will bound the separable operators within this space. The dual to the Roman surface is known as the Cayley’s cubic surface [129], which, for a given scale parameter w is characterized by
w x y
x w z = 0.
y z w
q 3 q 3 q 3 We first set x = 8 ηAB, y = 8 ηAC , and z = 8 ηBC . Then we must set w so that
55 Joinability of causal and acausal relationships the Cayley surface delimits the separable states. For each extremal separable state in our space, there is a corresponding local-positive operator acting as an entanglement witness; a state is separable if the inner product with its entanglement witness is nonnegative.
Consider the extremal local-positive operator ηAB = ηAC = ηBC = 2/3 that we made use of previously. This operator will act as an entanglement witness for another operator with ηAB = ηAC = ηBC = σ. We obtain σ by solving
√1 √1 q 8 q 8 3 2 3 8 3 8 σ q · q = 0, 3 2 3 8 3 8 σ q 3 2 q 3 8 3 8 σ
1 to arrive at σ = − 6 . With this, the only value of w allowing the Cayley surface to be solved by σ = − 1 is w = √1 . Setting the scaling value and evaluating the 6 24 determinant, we find that the separable states are bound by the surface
2 1 + 54ηABηAC ηBC − 9(ηAB + ηAC + ηBC )
+18(ηABηAC + ηABηBC + ηAC ηBC ) ≥ 0.
This inequality may also be obtained using Theorem 1 in [82]. The shape of the separable states is depicted in Figure 3.7. Several remarks may be made. First, the set of separable states exhibits the tetrahedral symmetry shared by the classical joinability boundary and the local-positive joinability boundary. Thus, among the various boundaries we have considered in this three dimensional Euclidean space, the state- and channel-positive boundaries are the only ones not obeying tetrahedral symmetry. However, both the convex hull and the intersection of the state- and channel-positive cones bound regions which recover this symmetry. It is a curious observation that the convex hull of these cones is “nearly” the local-positive region, while the intersection is “nearly” the set of separable states. Earlier we found, in the two-qudit case, that local-positivity coincides with the union of the state- and channel-positive regions, as well as that the separable region was their intersection. Here we consider the analog for three qubits. The result is that i) the convex hull of state- and channel- positive operators is strictly contained in the set of local-positive operators; and ii) the intersection of the state- and channel-positive operators is strictly contained in the set of separable states. We may further interpret the latter result in terms of PPT considerations. The operators which result from a homocorrelation-mapped channel necessarily have PPT. Corollary 1 in Ref. [82] states that the PPT and bi-separable Werner operators co-
56 3.4 Agreement bounds for quantum states and channels
(a) (b)
Figure 3.7: (a) Set of separable operators within the set of local-positive operators. (b) Intersection of the state- and channel-positive cones within the set of local-positive operators. While the separable operators are a subset of the intersection set, the two objects coincide (only) at their vertices. In panel (a) the closest point in the separable set to the (−1, −1, −1) corner of the figure is (−1/6, −1/6, −1/6), whereas in panel (b), the closest point in the intersection set is (−1/5, −1/5, −1/5). incide. Thus, any state-positive operator which is also a homocorrelation mapped channel is necessarily bi-separable. Hence, the intersection of the four cones will be the set of states which are bi-separable with respect to any of the three partitions. This set is clearly contained in the set of tri-separable states. These observations il- luminate the relationships among entanglement, quantum states, and quantum chan- nels. Specifically, the homocorrelation map allows us to place quantum channels in the same arena as quantum states, and hence to directly compare and contrast them. Finding that the tri-separable operators are a proper subset of the bi-separable ones, we wonder what features these strictly bi-separable operators possess, and what does bi-separability imply for the states or channels supporting such correlations.
3.4 Agreement bounds for quantum states and chan- nels
In what remains, we illustrate some crucial differences between channel- and state- positive operators. These differences inform the nature of their respective joinability limitations. In order to directly compare states to channels we restrict our consid- erations here to operators in B(Hd ⊗ Hd). Qualitatively, state-positive operators are restricted in the degree to which they can support agreeing outcomes, whereas
57 Joinability of causal and acausal relationships
channel-positive operators are restricted in the degree to which they can support disagreeing outcomes. We define the degree of agreement to be the likelihood of a certain POVM element. Specifically, consider a local projective measurement M = {|ijihij|}. We can coarse-grain this into a two-element projective measure- P ment with the bipartition into “agreeing” outcomes, EA = |iiihii|, and “disagree- P i ing” outcomes, ED = i6=j |ijihij|, respectively. Lastly, so that these outcomes are basis-independent, we can “twirl” EA and ED as follows: Z X † † EA = dµ(U)U ⊗ U |iiihii| U ⊗ U i Z X † † ED = dµ(U)U ⊗ U |ijihij| U ⊗ U , i6=j
where dµ(U) denotes integration with respect to the invariant (Haar) measure. It is simple to see that these two operators yield a resolution of identity and hence form a POVM. We can compute these two operators explicitly as follows. By the invariance of the Haar measure, we can rewrite EA as Z ⊗2 EA = d dµ(ψ)|ψihψ| ,
for which the above integral is proportional to the projector onto (or identity operator + ⊗2 ⊗2 I2 in) the totally symmetric subspace H+ ⊂ H [130]. Explicitly, we can write d + d I + V + + 2 + d − 1 EA = + I2 = + , d2 ≡ dim(H2 ) = , (3.10) d2 d2 2 2
ED = I − EA. (3.11)
We define the degree of agreement to be the likelihood of EA and, similarly, the degree of disagreement to be the likelihood of ED. Operationally, these values are the prob- ability that, for a randomly chosen local projective measurement made collectively, the local outcomes will agree or, respectively, disagree.
We now proceed to show how quantum channels differ from quantum states in their allowed range of agreement likelihood. In the case of a bipartite operator ρ ∈ B(H ⊗ H), we are familiar with computing this agreement probability as Tr(EAρ). To carry out the same computation for a channel operator, the homocorrelation map becomes expedient. Given a quantum channel M : B(H) → B(H), we wish to determine the probability that the outcome of a randomly chosen projective measurement (made on the completely mixed state) will agree with the outcome of the same measurement after the application of M. Assume the outcome was |ii from an orthogonal basis
58 3.4 Agreement bounds for quantum states and channels
{|ii}. Then the post-channel state is M(|iihi|), and the likelihood that the post- channel measurement will also be |ii is hi|M(|iihi|)|ii. Lastly, if we want to average this likelihood of agreement over all choices of basis we integrate, p(agree) = R dµ(U)Tr M(U|iihi|U †)U|iihi|U † = R dµ(ψ)Tr M(|ψihψ|)|ψihψ| .
If we wish to find the bounds on this value, the above form does not make transparent the fact that we are performing an optimization problem in a convex cone. But, recalling the namesake property of the homocorrelation map, Eq. (3.3), the above expression may be rewritten as Z p(agree) = Tr H(M)d dµ(ψ)|ψihψ| ⊗ |ψihψ| = Tr[H(M)EA].
Accordingly, the likelihood of agreement is calculated for channel operators in the homocorrelation representation in just the same way as it is for bipartite density operators. With the stage set, the desired bounds are described in the following theorem:
Theorem 3.4.1. Let w be an operator in B(Hd ⊗ Hd), and consider a POVM with operation elements {EA, ED} as in Eqs. (3.10)-(3.11). Then the degree of agreement for w ≥st 0 as calculated by the likelihood of EA is bounded by 2 0 ≤ Tr(wE ) ≤ , (3.12) A 2 + d − 1
while the degree of agreement for w ≥ch 0 is bounded by 1 ≤ Tr(wE ) ≤ 1. (3.13) 2 + d − 1 A
Proof. In the case of state-positive operators, the maximal value of Tr(wEA) is achieved by setting w = EA/Tr(EA), which results in Tr(wEA) = 2/(d + 1). For the lower bound, it is simple to see that choosing w to lie in the complement of the projector yields a value of zero. Hence, we have obtained the bound of Eq. (3.12).
In the case of channel-positive operators, the value of Tr(wEA), where w ≥ch 0, is unchanged by a partial transposition of both operators. Thus, we may seek bounds TA TA TA on the value of Tr(w EA ), where w is a density operator. By using Eq. (3.10),
59 Joinability of causal and acausal relationships
the partial transposition of EA is
TA TA d I + V EA = + . d2 2
TA TA Thus, the upper and lower bounds on Tr(wEA) are achieved by setting w = V /d and wTA = (I − V TA )/(d2 − d), respectively. Accordingly, the resulting bounds are d 1 d 1+d + 2 ≤ Tr(wEA) ≤ + 2 , which simplify to those of Eq. (3.13). d2 d2
By virtue of the homocorrelation map, the above result may be understood geo- metrically. The objects involved are the agreement/disagreement POVM operators EA and ED, and the state- and channel- positive cones Wst and Wch, respectively. Theorem 3.4.1 places an upper bound on the inner product between vectors in Wst and EA, and, similarly, on the inner product between vectors in Wch and ED. This geometric understanding is aided by the example of Werner operators shown in Fig. 3.2. Lastly, we proceed to show that general joinability limitations (though not strict ones) can be derived based solely on i) the above agreement bounds of channels and states; ii) joinability bounds of classical probabilities; and iii) the fact that the agreement likelihoods must obey rules of classical joinability. Ultimately, the reduced states must satisfy certain limitations arising from joining limitations of classical probability distributions. In the three-party joining scenario, the bipartite marginal distributions of three classical d-nary random variables must have probabilities of agreement αAB, αAC , and αBC satisfying the following inequalities [61]:
−αAB + αAC + αBC ≤ 1, (3.14)
αAB − αAC + αBC ≤ 1, (3.15)
αAB + αAC − αBC ≤ 1, (3.16)
and, in the case of d = 2, also
αAB + αAC + αBC ≥ 1. (3.17)
Since Tr(wEA) is a probability of agreement, it too is subject to the above con- straints. Hence, we identify Tr(ρiEA) ≡ α`, where ` = AB, AC, or BC. Consider the case where systems B-C are state-positive. Theorem 3.4.1 then sets the bound 2 Tr(ρBC EA) ≤ d+1 . Setting the parameter αBC = Tr(ρBC EA) to this upper limit of 2 d+1 , Eq. (3.16) becomes d + 3 α + α ≤ . AB AC d + 1
60 3.4 Agreement bounds for quantum states and channels
In the case of αAB = αAC ≡ α, this yields d + 3 α ≤ , 2(d + 1)
which corresponds precisely to the optimal bound for qudit cloning [131] (cf. Eq. (21) therein, where their F coincides with our α). We can similarly recover the exact bound for the 1-2 sharability of qubit Werner states determined in [61]. Again, we 2 set the B-C agreement to its extremal value Tr(ρBC EA) = d+1 , as given by Theorem 3.4.1. For d = 2, Eq. (3.17) applies, and substituting in the extremal value of αBC 1 1 we obtain αAB + αAC ≥ 3 . Again, in the case of αAB = αAC ≡ α, this yields α ≤ 6 , which is the exact condition for 1-2 sharability of Werner qubits. While obtaining a full generalization of Theorem 3.4.1 to multiparty systems would entail a detailed understanding of representation theory for Brauer algebras which is beyond our current purpose, we can nevertheless establish the following:
⊗N Theorem 3.4.2. Let w ∈ B(Hd ), and consider a POVM with operation elements d + {EA = + IN , ED = I − EA} (analogous to Eqs.(3.10)-(3.11)). Then the degree of dN agreement for w ≥st 0 as calculated by the likelihood of EA is bounded by d 0 ≤ Tr(wE ) ≤ . (3.18) A d−1+N N
Proof. The maximal and minimal values of Tr(wEA) are achieved by setting w = + + EA/Tr(EA) and w = (dI/dN − EA)/Tr(dI/dN − EA), respectively, which yields the desired bounds of Eq. (3.18). From the above multiparty bound, one may attempt to recover, for instance, the known bounds on 1-n sharability of Werner states [61]. However, thus far we have not been successful in this endeavor. In the tripartite qudit setting, such bounds were found to be sufficient, but this might be a special feature of this particular case. Therefore, it remains an open question to determine whether there exists a simple principle (or simple principles) which govern joinability limitations beyond the tripartite setting.
61
Chapter 4
Towards an alternative approach to joinability: enforcing positivity through purification
63 Towards an alternative approach to joinability: enforcing positivity through purification 4.1 Introduction
The aim of this chapter is to follow up on some of the ideas brought up in the previous two chapters and also to draw further connections using the concept of joinability. We develop a new approach to the joinability problem which allows us to formalize the “composition law” that was touched upon in the previous chapter, to easily incorporate the concept of measurement incompatibility into the joinability framework, and to find a simple principle that governs joinability failures in both the classical and quantum cases. The techniques used in this chapter are rather elementary. Yet, to our knowledge, they have not been applied to the problem of joinability. Surprisingly, with basic linear algebra, we unveil important connections among principles that govern the structure of quantum relationships. We emphasize that the ideas presented in this chapter are just sprouting and that the purpose of presenting them is to provide an environment in which they can grow. In the previous chapters, failures of joinability were due to a failure of positive- semidefiniteness or complete positivity. In this chapter we make positivity manifest by purifying the quantum state with an ancillary system or by “lifting” the quantum channel to an isometric description in a larger Hilbert space. In this way, failures of joinability are only due to failures of trace-normalization, such as the conclusion that the likelihood of every possible event is zero. We emphasize that the motivation for pursuing the ideas in this chapter are mostly theoretical, urged by the desire to better understand the fundamental concepts of monogamy of entanglement, no-cloning, and measurement incompatibility. Previously, we presented the following explanation for the fact that qubit pairs A-B and A-C could not both be in the singlet state. The singlet state exhibits per- fectly disagreeing measurement outcomes for the same spin measurement made on both qubits. Assuming a singlet state were shared between A-B and A-C, we can “compose” the disagreements to conclude that the same spin measurement made on B-C must produce perfectly agreeing outcomes. However, no valid quantum state (i.e. no positive-semidefinite trace-one operator) can produce such perfectly agreeing outcomes for all collective spin measurements. Thus, we have that the A-B disagree- ment and the A-C disagreement force the contradictory B-C agreement. A similar argument can explain the no-cloning result. The initial motivation for the ideas in this chapter had been to further explore the idea of “composing” two conditions, such as the A-B disagreement and the A-C disagreement, to obtain a third condition. Such an argument was made to work in the singlet state sharing example and to obtain the no-cloning principle. But, it is not clear how it could be properly generalized. Furthermore, could such an approach be extended beyond the three-body joinability scenario? In the spirit of generalized en- tanglement [132], could such an approach be applied to general consistency problems
64 4.1 Introduction
which do not admit a natural tensor product structure as in fermionic systems? Before discussing our new approach to joinability we establish the nature of the constraints that are considered. We consider constraints to be on the manifestly positive objects such as the purification of the quantum state or the isometry of the quantum channel. We briefly review the construction of these objects. Given a density operator ρ ∈ B(H ), there always exists a quantum state |ψi ∈ A √ H1 ⊗ H2 such that Tr (|ψihψ|) = ρ. This is easily verified. ρ ≥ 0 implies ρ = √A A 2 √ ρ† is well-defined. Then |ψi = ( ρ ⊗ I)|Ωi satisfies the partial trace condition, P where |Ωi = i |iii is the unnormalized maximally entangled state. The state |ψi is unique up to a norm-preserving linear map, or isometry, acting on the ancillary system. A similar argument can be used to show that, given a quantum channel 1 1 2 3 E : B(HA) → B(HB), there exists an isometry V : HA → HB ⊗ HB ⊗ HA such † that Tr23 V · V = E(·). Here, trace-preservation of E corresponds to the isometry condition on V . The isometry is defined up to the action of another isometry acting on the traced-out ancillary systems. We consider two types of constraints that we refer to as hard constraints and soft constraints. Hard constraints express knowledge of statistical impossibilities, while soft constraints bound the likelihoods of various outcomes. We represent a hard constraint on a quantum state as a set of linear equations satisfied by its purification:
(hφ|A ⊗ IB)|ψiAB = 0, ∀ |φi ∈ SA ⊆ HA, (4.1)
where SA is the set of states involved in the hard constraint. A hard constraint on a quantum channel is, similarly, represented by a set of linear equations satisfied by its isometry:
(hφ|AB ⊗ IC )(IA ⊗ VA→BC )|ΩiAA = 0, ∀ |φiAB ∈ SAB ⊆ HA ⊗ HB, (4.2) P where |ΩiAA = i |iii is the unnormalized maximally entangled state and SAB is, again, the set of vectors used to subject |ψi to the hard constraint.
As a simple example, consider a system of three qubits HA⊗HB⊗HC and constrain the density matrix of this system to have reduced state AB described by the (anti- symmetric) singlet state |ψ−i = √1 (|01i − |10i). Denote the space of symmetric 2 states on AB as S2(H). This is the space orthogonal to the singlet state. Hence, the condition that AB is described by the singlet state is expressed by the hard constraint
2 (hφ|AB ⊗ ICD)|ψiABCD = 0, ∀ |φi ∈ S (H) (4.3)
Standing alone, this means of expressing the AB-singlet condition is not very illu- minating. The value of this rephrasing is revealed when we consider multiple hard constraints in tandem. As we show in the following section, we can recover a number
65 Towards an alternative approach to joinability: enforcing positivity through purification of basic quantum “no-go” principles as stemming from an over-constraining of hard constraints.
Soft constraints bound the likelihoods of certain projective measurement out- comes. We represent a soft constraint on a quantum state as a set of inequalities satisfied by its purification:
2 2 2 k(hφ|A ⊗ IB)|ψiABk ≤ || |φi|| || |ψi|| , ∀ |φi ∈ SA ⊆ HA, (4.4) where represents the likelihood bound associated to the set of states SA. A soft constraint on a quantum channel is, similarly, represented by a set of inequalities satisfied by its isometry:
2 2 k(hφ|AB ⊗ IC )(IA ⊗ VA→BC )|Ωik ≤ || |φi|| || (IA ⊗ VA→BC )|Ωi|| ,
∀ |φiAB ∈ SAB ⊆ HA ⊗ HB. (4.5)
4.2 Joinability limitations from hard constraints
In this section we explore joinability where the partial descriptions of the system are given by hard constraints. We begin with a trivial example from classical probability theory. Consider the simplest classical system, the probabilistic bit. The probability distribution is given by the likelihoods p(0) and p(1). An example of inconsistent hard constraints is p(0) = 0 and p(1) = 0. No distribution on outcomes 0 and 1 can satisfy these constraints while also satisfying a non-trivial normalization condition.
This example seems too trivial to bear insight. However, we will argue that all inconsistent sets of hard constraints fail to be consistent for the same reason as the above example fails: the likelihood of any outcome is zero. Surprisingly, this same reasoning explains the non-sharability of maximally entangled states, the no-cloning principle, the incompatibility of certain observables, as well as the simple classical principle that three bits cannot disagree.
Consider three classical probabilistic bits A, B, and C with the constraints, A and B disagree, B and C disagree, and A and C disagree. These conditions are equivalent to the three hard constraints, A and B cannot agree, B and C cannot agree, and A
66 4.2 Joinability limitations from hard constraints and C cannot agree. These three conditions exclude all possible configurations:
0 0 0 0 0 0 × 0 0 0 × 0 0 0 × 0 0 1 0 0 1 × 0 0 1 × 0 0 1 × 0 1 0 0 1 0 0 1 0 0 1 0 × 0 1 1 0 1 1 0 1 1 × 0 1 1 × AB → BC → AC → 1 0 0 1 0 0 1 0 0 × 1 0 0 × 1 0 1 1 0 1 1 0 1 1 0 1 × 1 1 0 1 1 0 × 1 1 0 × 1 1 0 × 1 1 1 1 1 1 × 1 1 1 × 1 1 1 × (4.6) The constraints imply that the likelihood of any outcome is zero. Therefore, no distribution can satisfy these three conditions. Consider a single qubit S with the constraints “|0i is certain upon making a Pauli- Z measurement” and “|+i is certain upon making a Pauli-X measurement”. We can translate each into a hard constraint on the purification of the density operator to an ancillary system A: h1|S ⊗ IA|ψSAi = 0 and h−|S ⊗ IA|ψSAi = 0. Any state vector |φi ∈ HS can be written as some linear combination |φi = α|1i + β|−i. Hence, for any |φi ∈ HS , there exists α and β such that
∗ ∗ hφ| ⊗ IA|ψSAi = α h1| ⊗ IA|ψSAi + β h−| ⊗ IA|ψSAi = 0. (4.7)
Only |ψSAi = 0 satisfies hφ| ⊗ IA|ψSAi = 0 for all |φi. Thus, no non-trivial state can satisfy both constraints. It seems that we have taken a simple fact of quantum mechanics (i.e. uncertainty of Z and X measurements cannot both be zero) and given an unfamiliar or cumbersome argument for reaching it. However, a failure of the trace-one condition seems to be more direct or basic than one involving algebraic properties of observables, especially when the example is more complicated than the one above. Next, we use the concept of hard constraints to easily prove a fact used in the previous chapter: there is no bipartite quantum state that exhibits perfectly agreeing outcomes for all collective projective measurements. Considering the system to be described by the Hilbert space H = Cd ⊗ Cd we can rephrase the condition in terms of hard constraints as,
⊥ ⊥ ⊥ (hφ|A ⊗ hφ |B ⊗ IC )|ψiABC = 0, ∀ |φi, |φ i, s.t. hφ |φi = 0. (4.8)
Since the set of vectors |φi ⊗ |φ⊥i span Cd ⊗ Cd (this can be taken as an exercise for the enthusiastic reader), there is no non-trivial solution to the set of hard constraints, verifying the statement. Note that we did not have to make any reference to a positive-semidefinite constraint as we did while making this argument in the previous
67 Towards an alternative approach to joinability: enforcing positivity through purification
chapter. The key insight was to phrase the condition in terms of a set of negative statements, or impossibilities, as opposed to a number of positive expectation values. With a parallel argument we can show that no quantum channel can map every input pure state to an orthogonal pure state. Considering the quantum channel to d d be described by an isometry V : C → C ⊗ HA we rephrase the condition in terms of hard constraints as follows. The condition states that the output of the channel, † TrA V |φihφ|V , should be orthogonal to |φi, or, equivalently, hφ|V |φi = 0 for all |φi ∈ Cd. Defining hφT | ≡ T (|φi), and noting that |φi = (hφT | ⊗ I)|Ωi, we write the hard constraints as
T d (hφ | ⊗ hφ| ⊗ IA)(I ⊗ V )|Ωi = 0, ∀ |φi ∈ C . (4.9)
Then, because span{hφT | ⊗ hφ|} = (Cd ⊗ Cd)† (this, too, can be taken as an exercise), there is no non-trivial solution for V . These two previous observations address the quantum distinction between causal and acausal relationships. There is certainly a one-to-one correspondence between the vectors |φi ⊗ |φi and matrices |φi ⊗ hφ|. The former linear objects inhabit the space H ⊗ H, which corresponds to acausal quantum relationships, while the latter inhabit the space B(H) ∼= H⊗H∗, which corresponds to causal quantum relationships. Despite the one-to-one correspondence, the span of the |φi ⊗ |φi is a strict subspace of H ⊗ H, while the span of the |φi ⊗ hφ| is equal to H ⊗ H. In particular the vectors |φi ⊗ |φi ∈ Cd ⊗ Cd only span a d(d + 1)/2-dimensional space, while the set of |φihφ| ∈ Cd ⊗ Cd† spans Cd ⊗ Cd†. Next we consider the impossibility of having entanglement between both A-B and B-C. We return to the three-qubit system with the constraint that A-B and B-C are each described by the singlet state |ψ−i = √1 (|01i − |10i). We can rephrase this 2 in terms of a set of hard constraints on the purified state |ψiABCD,
2 (hφφ|AB ⊗ hτ|C ⊗ ID)|ψiABCD = 0 ∀ |φi, |τi ∈ C , (4.10)
and similarly for the BC system. Since span{|φφτi, |τφφi} = (C2)⊗3, again, there is no non-trivial solution. Thus, we can view the impossibility of singlet sharing to be due to the constraints having ensured zero likelihood for any measurement outcome. In this sense, the singlet sharing is impossible for the same reason that three bits disagreeing is impossible. For completeness, we prove the no-cloning theorem using the hard constraint approach. Consider a quantum channel from system A to BC, where HA = HB = HC ≡ H, represented by the isometry V : HA → HB ⊗ HC ⊗ HD. We can phrase the cloning condition as a negative statement by requiring that the output of the ⊥ ⊥ channel applied to |φi be orthogonal to span(|φ i)⊗HC and HB ⊗span(|φ i), where
68 4.2 Joinability limitations from hard constraints
hφ|φ⊥i = 0. As a set of hard constraints, we have
T (hφ φτ| ⊗ ID)(I ⊗ V )|Ωi = 0 ∀ |φi, |τi ∈ H, (4.11)
along with the analogous statement from switching the roles of B and C. Since span{hφT φτ|, hφT τφ|} = H† ⊗ H† ⊗ H†, no non-trivial solution exists for V . As with the no singlet sharing example, we can explain the impossibility of cloning as due to the set of constraints ruling out all possibilities. As a final example, we consider the incompatibility of measuring devices. A quantum measurement can be described by a quantum channel from the measured system A to a classical measuring device D. The measuring device is made “classical” by requiring that operators appearing on B(HD) are restricted to being diagonal in a particular basis. Equivalently, operators on D must be in the commutative algebra C⊕d, where d is the number of outcomes. As with any quantum channel, such measurement channels admit an isometric representation V : HA → HD ⊗ HE for d some ancillary system E. For a complete projective measurement on HA = C with outcomes corresponding to the basis of vectors |ji ∈ HA, an isometric representation P of the channel is V = j |jjihj|, where we have let the |ji also correspond to an orthonormal basis of vectors in HD and in HE. With this formalism in place, we consider the case of having a single system be measured by two different measurement devices simultaneously. Following [133], we say that the two measurements are compatible if there exists a single measurement device from which the two measurements can be obtained via partial trace of the outputs. Note that this is equivalent to the definition of channel joinability. Re- stricting to the qubit case, consider an X and a Z measurement device, from the quantum system A to the devices DX and DZ . We label the basis vectors of the output systems as {|+i, |−i} and {|0i, |1i}, respectively, such that the labels corre- spond to the states of the input system. We can ensure the X and Z measurement conditions by requiring that, with input |±i, the output on DX must be orthogonal to |∓i and that, with input |0/1i, the output on DZ must be orthogonal to |1/0i
(the “/” indicates “or”). Letting V : HA → HDX ⊗ HDZ ⊗ HE be the isometry of the quantum measurement, the hard constraints are (h∓| ⊗ hφ| ⊗ IE)V |±i = 0
for all |φi ∈ HDZ and (hφ| ⊗ h1/0| ⊗ IE)V |0/1i = 0 for all |φi ∈ HDX . Since span{h∓| ⊗ hφ| ⊗ |±i, hφ| ⊗ h1/0| ⊗ |0/1i} = H† ⊗ H† ⊗ H , no non-trivial solution DX DZ A exists. Thus, we have shown that many of the important quantum no-go principles, in- cluding the impossibility of singlet sharing, no-cloning, and measurement incompati- bility can be explained with the same argument used to show p(0) = 0 and p(1) = 0 are incompatible. The approach used is not restricted to needing a tensor product structure. Therefore, it would be interesting to explore these ideas for fermionic or
69 Towards an alternative approach to joinability: enforcing positivity through purification
bosonic systems, making better connection, for instance, with the N-representability problem [134, 28, 27].
4.3 Joinability limitations from soft constraints
We now turn to the case where, instead of ruling out any possibilities, we simply place bounds on their likelihoods. As mentioned, the motivation for the approach to joinability, developed in this chapter, was to elucidate the composition rule that we previously used to show the impossibility of singlet sharing. The first example returns to the singlet sharing scenario. Consider a tripartite system of qubits ABC. We aim to show that the correla- tions of AB in tandem with the correlations of BC ensure certain correlations on AC. For instance, we used the “perfect disagreement” of the singlet state to ensure that the singlet describing AB and BC ensures the correlations on AC to be perfectly agreeing. After much effort towards trying to develop and generalize this “compo- sition observation”, we have arrived at a simple idea which makes this possible. A key insight, which was conveyed in the previous section, is the utility of expressing constraints in terms of negative statements. There, instead of formalizing the com- position (disagree)AB + (disagree)BC ⇒ (agree)AC , we found it simpler to formalize the composition (not agree)AB + (not agree)BC ⇒ (not disagree)AC as represented by hard constraints. Then, we expect the corresponding composition rule in the soft con- straint case to be (low agreement)AB + (low agreement)BC ⇒ (low disagreement)AC . We enforce some degree of singlet correlation on a pair of subsystems by upper bounding the inner product of the state with the set of |φφi, which span the symmetric subspace. Assume that, for the purified state |ψi on ABCD, the soft constraint on AB is given by 2 2 k(hφφτ| ⊗ ID)|ψik ≤ αAB, ∀|φi, |τi ∈ C (4.12) while the soft constraint on BC is given by
2 2 k(hτφφ| ⊗ ID)|ψik ≤ αBC , ∀|φi, |τi ∈ C . (4.13)
We expect to be able to bound the likelihood of “disagreeing” outcomes on AC, such as |φi ⊗ |φ⊥i, where the two states are orthogonal. Consider, then, the expression for the likelihoods on AC,
⊥ 2 (hφτφ | ⊗ ID)|ψi . (4.14)
Since the Hilbert space of system B is two-dimensional, an arbitrary vector |τiB can be written as a linear combination |τi = λ|φi + µ|φ⊥i. Making this replacement in
70 4.3 Joinability limitations from soft constraints the likelihood expression, we obtain
⊥ ⊥ ⊥ 2 (λhφφφ | ⊗ ID + µhφφ φ | ⊗ ID)|ψi . (4.15)
We can bound the value of this expression using the triangle inequality,
⊥ 2 ⊥ 2 ⊥ ⊥ 2 (hφτφ | ⊗ ID)|ψi ≤ (λhφφφ | ⊗ ID)|ψi + (µhφφ φ | ⊗ ID)|ψi . (4.16)
Each term of the right-hand side of the inequality is the form of the constraints we ⊥ 2 place on AB and BC. Letting the maximum value of (hφτφ | ⊗ ID)|ψi be δAC and noting that |λ|, |µ| ≤ 1, we then obtain the monogamy-like inequality
δAC ≤ αAB + αBC . (4.17)
This inequality captures how an “agreement-bound” on AB and on BC (i.e. αAB and αBC ) enforce a “disagreement-bound” on AC (i.e. δAC ). Note that the hard-constraint case is recovered by setting αAB = αBC = 0, which forces δAC = 0 (an impossibility). Also, using the fact that span{|φφ⊥i} = C2 ⊗ C2, there is no vector which can be ⊥ orthogonal to all |φφ i. Thus, there is a lower bound to the value of δAC , which can 1 be computed to be 2 (show as an exercise). This expresses a “trade-off” in the values that αAB and αBC may take, 1 ≤ α + α . (4.18) 2 AB BC This inequality gives a necessary condition for determining if the soft constraints are consistent. It does not diagnose all inconsistent sets of constraints. This is, in part, due to the fact that the inequality is weakened by upper-bounding λ and µ. In future work, we hope to properly compare these observations to our previous Werner joinability findings. We only briefly make a few observations. First, applying the inequality in Eq. (4.18) to Werner states gives a linear trade-off for the Werner parameters (see Chapters 2 and 3 for various parameterizations). We found, however, that the exact trade-off between the Werner parameters for AB and BC is quadratic (see Eq. (2.22)). We can now understand another opportunity afforded by phrasing the constraints as negative statements (or upper bounds on likelihoods): the upper-bounds of the constraints allow us to use the triangle inequality to upper bound the derived con- straint. If instead, we had used lower-bounds, the direction of these inequalities would have opposed the direction of the triangle inequality. The singlet state is not particular with respect to monogamy constraints. We expect that other Bell states should lend themselves to similar composition laws. Sticking to the qubit case, each maximally entangled state is obtained by applying a particular unitary transformation to, without loss of generality, system B: |ψU i =
71 Towards an alternative approach to joinability: enforcing positivity through purification
(I ⊗ U)|ψ−i. With this, the hard constraint that ensures two qubits are described † 2 by |ψU i is (hφφ| ⊗ IC )(IA ⊗ UB ⊗ IC )|ψiABC = 0 for all |φi ∈ C . This is a simple modification of the singlet state hard constraints.
Accordingly, the soft constraints which ensure a state is “close to” the state |ψU i are stated as 2 † U 2 (hφφ| ⊗ IC )(IA ⊗ UB ⊗ IC )|ψiABC ≤ αAB, ∀ |φi ∈ C . (4.19) Consider the three-party joining scenario, let the above soft constraint apply to AB (where we purify the quantum state using an ancilla system D). Let the following soft constraint apply to BC,
2 † V 2 (hτφφ| ⊗ ID)(IAB ⊗ VC ⊗ ID)|ψiABCD ≤ αBC , ∀ |φi ∈ C . (4.20) We follow the reasoning of the singlet state example. Consider the following expression for certain likelihoods on AC
⊥ † † 2 (hφτφ | ⊗ ID)(IAB ⊗ (U V )C ⊗ ID)|ψi . (4.21)
Now, we can write |τi = λU|φi + µU|φ⊥i for any |τi, giving
2 ⊥ ⊥ ⊥ † † † (λhφφφ | ⊗ ID + µhφφ φ | ⊗ ID)(IA ⊗ UB ⊗ (U V )C ⊗ ID)|ψi . (4.22) We can bound the value of this expression using the triangle inequality,
⊥ † † 2 (hφτφ | ⊗ ID)(IAB ⊗ (U V )C ⊗ ID)|ψi (4.23) 2 ⊥ † † † ≤ (λhφφφ | ⊗ ID)(IA ⊗ UB ⊗ (U V )C ⊗ ID)|ψi 2 ⊥ ⊥ † † † + (µhφφ φ | ⊗ ID)(IA ⊗ UB ⊗ (U V )C ⊗ ID)|ψi . (4.24)
By setting |τ 0i = VU|φ⊥i and |φ0i = U|φ⊥i, we simplify the inequality
⊥ † † 2 (hφτφ | ⊗ ID)(IAB ⊗ (U V )C ⊗ ID)|ψi (4.25) 2 † ≤ (λhφφτ| ⊗ ID)(IA ⊗ UB ⊗ ICD)|ψi 2 0 0 † + (µhφφ φ | ⊗ ID)(IAB ⊗ VC ⊗ ID)|ψi . (4.26) Each term on the right-hand side of the inequality can be recognized as the expressions
72 4.3 Joinability limitations from soft constraints in the soft constraints on AB and BC, respectively. Define
VU ⊥ † † 2 δAC ≡ max (hφτφ | ⊗ ID)(IAB ⊗ (U V )C ⊗ ID)|ψi . (4.27) |φi,|τi
Using |λ|, |µ| ≤ 1, we obtain the inequality
VU U V δAC ≤ αAB + αBC . (4.28)
This generalized composition rule simplifies to the singlet state case by setting U = V = I. This inequality demonstrates how, by ensuring certain quantum correlations for AB and for BC, certain correlations are forced upon AC. We parameterized the maximally entangled qubit states with a unitary transformation. This shows that, if AB is close to the maximally entangled state |ψU i and BC is close to the maximally ⊥ entangled state |ψV i, then the likelihood of any measurement outcome |φi ⊗ VU|φ i is upper bounded. It would be interesting to explore further generalizations of these observations to qudits and to more than three systems. The above “composition rule” inequality does not, on its own, tell us when the initial soft constraints, themselves, are inconsistent with one another. As with the singlet example, we must further evaluate or place a lower bound on the disagreement parameter δAC , in order to obtain necessary conditions on the consistency of the soft constraints. We have yet to provide a general intuition behind joinability failures in the soft-constraint case. For instance, it would be useful to directly obtain an inequal- ity involving αAB and αBC , or involving a general set of soft constraint parameters i. We outline an approach to obtaining such inequalities and show that joinability fail- ures, in this case too, are on account of failures of the trace-normalization condition. Thus, we can view this cause of joinability failures as being a more general version of the “zero-total-probability” explanation in the hard-constraint case. As with the quantum case, for a classical probability distribution the sum of all likelihoods must be 1. Considering the classical soft constraints of p(0) ≤ and p(1) ≤ δ, the constraints are inconsistent unless 1 ≤ +δ. A simple quantum example obtains analogous inconsistency bounds. Consider a qubit system A (with ancilla B) 2 2 subject to the soft constraints k(h+| ⊗ I)|ψiABk ≤ δ √and k(h0| ⊗ I)|ψiABk ≤ . Writing the trace as h0| · |0i + h1| · |1i, we replace |1i = 2|+i − |0i to give √ Tr (ρ) = h0|ρ|0i + 2h+|ρ|+i − 2(h+|ρ|0i + h0|ρ|+i) + h0|ρ|0i. 2 2 ≤ 2k(h0| ⊗ I)|ψiABk + 2k(h+| ⊗ I)|ψiABk √ + 2 2k(h0| ⊗ I)|ψiABk · k(h+| ⊗ I)|ψiABk. (4.29) √ ≤ 2(δ + ) + 2 2δ, (4.30) where ρ = TrB (|ψihψ|AB). Thus, the trace-normalization condition is impossible
73 Towards an alternative approach to joinability: enforcing positivity through purification
√ 1 unless 2 ≤ δ + + 2δ. The inequality expresses the fact if δ and are both too small, the total probabilistic weight cannot amount to the proper normalization value. Intuitively, the constraints have “ruled out too much”. We can apply this same technique to the soft constraints in the singlet sharing example. Consider the following soft constraint that ensures proximity to the singlet state on AB
2 2 2 2 k(hφ|AB ⊗ hτ|C ⊗ ID)|ψik ≤ αAB, ∀ |φi ∈ S (C ), ∀ |τi ∈ C , (4.31)
2 2 2 2 2 where S (C ) denotes the symmetric subspace. Defining SAB ≡ S (C ) ⊗ C , we can equivalently express this soft constraint as
2 k(hν|ABC ⊗ ID)|ψik ≤ αAB, ∀ |νi ∈ SAB. (4.32)
2 2 2 Defining SBC ≡ C ⊗ S (C ), we can give the analogous soft constraint on BC as
2 k(hω|ABC ⊗ ID)|ψik ≤ αBC , ∀ |ωi ∈ SBC . (4.33)
2 ⊗3 Since span{SAB, SBC } = (C ) , there are no states |ψi which admit likelihood zero for all states in SAB and SBC . Furthermore, this ensures that, just like the previous qubit example, we can express the trace operation in terms of vectors chosen from SAB 2 ⊗3 and SBC . Let {|ji} be an arbitrary basis for (C ) . Each basis vector can be written as a linear combination |ji = xj|νji + yj|ωji, where |νji ∈ SAB and |ωji ∈ SBC . Letting ρ = TrD (|ψihψ|ABCD), we can write the trace of ρ as
8 X Tr (ρ) = hj|ρ|ji j=1 8 X 2 ∗ ∗ 2 = |xj| hνj|ρ|νji + xj yjhνj|ρ|ωji + xjyj hωj|ρ|νji + |yj| hωj|ρ|ωji. j=1 8 X 2 ∗ ∗ √ 2 ≤ |xj| αAB + (xj yj + xjyj ) αABαBC + |yj| αBC . (4.34) j=1 √ Thus, the agreement parameters are bound by 1 ≤ λαAB + η αABαBC + µαBC , where, the expressions for λ, η, and µ are given in the last line above. Crucially, the values of λ, η, and µ depend on the choice of basis {|ji} and the choice of decomposition |ji = xj|νji + yj|ωji. Some choices will lead to less-strict joinability limitations than other choices. Unfortunately, there is not a single choice of basis and decomposition which diagnoses all joinability failures. Rather, we expect there to be family of “optimal” decompositions. Nevertheless, any choice of decomposition
74 4.3 Joinability limitations from soft constraints gives finite values for λ, η, and µ and, therefore, leads to non-trivial constraints on √ the agreement parameters 1 ≤ λαAB + η αABαBC + µαBC . This approach can certainly be extended to more general settings. Furthermore, it would be valuable to investigate the structure of choosing the decompositions |ji = xj|νji+yj|ωji, and to understand the features of the optimal family of decompositions. We anticipate that convex geometry might play a role in determining such optimal families. If so, we will have returned to an undesirable vantage point, in that, one of the motivations of the work in this chapter was to avoid the use of convex geometry for understanding joinability. We concede that, it is possible (and probably likely), that any complete exploration of the concept of joinability must resolve to using the tools of convex geometry. Regardless, we have shown how the failure of the trace- normalization condition is responsible for certain failures of joinability. As we emphasized in the introduction to this chapter, the ideas presented here are very preliminary and represent a starting point for further investigation. While in the previous two chapters we have emphasized the role that positivity plays in limiting joinability, here, we have argued that an alternative explanation exists. In particular, we showed that failures of the consistency of hard or soft constraints can be explained by over-constraints on the total probabilistic weight of the quantum state or channel. We achieved this by enforcing the quantum state or channel to be manifestly positive-semidefinite or completely positive by means of purification or an isometric extension, respectively. When positive semi-definiteness is made manifest, joinability failures can be diagnosed with failures of the trace-normalization condition.
75
Chapter 5
Asymptotic stabilization of quantum states with continuous-time quasi-local dynamics
77 Asymptotic stabilization of quantum states with continuous-time quasi-local dynamics
This chapter presents material that appeared in Quantum Information and Compu- tation, 7:0657, (2016), in an article titled “General fixed points of quasi-local frustration- free quantum semigroups: from invariance to stabilization”, which is joint work with Francesco Ticozzi and Lorenza Viola.
5.1 Introduction
Convergence of a dynamical system to a stable equilibrium point is a hallmark of dissipative, irreversible behavior. In particular, rigorously characterizing the na- ture and stability of equilibrium states of irreversible quantum evolutions is a long- standing problem central to both the mathematical theory of open quantum sys- tems and the foundations of quantum statistical mechanics [135, 136]. In recent years, renewed interest in these issues has been fueled by the growing theoreti- cal and experimental significance of techniques for quantum reservoir engineering [137] and dissipative quantum control [21] within Quantum Information Process- ing (QIP). Representative applications that benefit from engineered dissipation in- clude robust quantum state preparation, with implications for steady-state entan- glement [32, 138, 139, 140, 141], non-equilibrium topological phases of matter [142], and ground-state cooling [143, 144, 145]; as well as open-system quantum simulation [146, 147], steady-state dissipation-driven quantum computation [148, 149], dissipa- tive quantum gadgets and autonomous quantum error correction [150, 151], along with quantum-limited sensing and amplification [152, 153]. While applications are often developed by making reference to a specific physi- cal setting, a common theme is the key role played by constraints, that may restrict the allowed dynamical models and the extent of the available manipulations. This motivates seeking a rigorous system-theoretic framework for characterizing controlled open-quantum system dynamics subject to given resource constraints. In this work, we focus on dissipative multipartite quantum systems described by time-independent quasi-local (QL) semigroup dynamics, capturing the fact that, in many physically relevant scenarios, both the coherent (Hamiltonian) and irreversible (Lindblad) con- tributions to the semigroup generator must act non-trivially only on finite subsets of subsystems, determined for instance by spatial lattice geometry. The main question we address is to determine what properties an arbitrary target state of interest must satisfy in order to be the unique stationary (“fixed”) point for a given QL constraint, thereby making the state globally QL-stabilizable in principle, in an asymptotic sense. In previous work [31, 34], this question has been addressed under the assumption that the target state is pure, providing in particular a necessary and sufficient linear- algebraic condition for the latter to be stabilizable without requiring Hamiltonian dynamics. Such pure states are called purely Dissipatively Quasi-Locally Stabilizable (DQLS). While restricting to a pure target state is both a natural and adequate
78 5.2 Preliminaries
first step in the context of dissipatively preparing paradigmatic entangled states of relevance to QIP (such as W or GHZ states), allowing for a general mixed fixed- point is crucial for a number of reasons. On the one hand, since mixed quantum states represent the most general possibility, this is a prerequisite for mathematical completeness. On the other hand, from a practical standpoint, QL stabilization of a mixed state which is sufficiently close to an “unreachable” pure target may still be valuable for QIP purposes, a notable example being provided by thermal graph states at sufficiently low temperature [154]. Furthermore, as physical systems in thermal equilibrium are typically far from pure, characterizing mixed-state QL stabilization might offer insight into thermalization dynamics as occurring in Nature and on a quantum computer [155]. From this point of view, a stability analysis of thermal states of QL Hamiltonians is directly relevant to developing efficient simulation and sampling algorithms for the quantum canonical ensemble, so-called “quantum Gibbs samplers,” as analyzed in [156] for commuting Hamiltonians. In the mixed-state scenario, the problem of QL stabilization involves qualitatively different features and is substantially more complex. This is largely due to the fact that the analysis tools used in the pure-state setting do not lend themselves to a for- mulation where the invariance property of the globally defined target state translates directly at the level of QL generator components. We bypass this difficulty by restrict- ing to the important class of frustration-free (FF) semigroup dynamics [157, 156], for which global invariance of a state also implies its invariance under each QL compo- nent. Physically, the FF property is known to hold within standard derivations of Markovian semigroup dynamics, for instance based on Davies’ weak coupling limit or “heat-bath” approaches generalizing classical Glauber dynamics [136, 156].
5.2 Preliminaries
5.2.1 Notation and background Consider a finite-dimensional Hilbert space H, dim(H) = d, and let B(H) be the set of linear operators on H.X† shall denote the adjoint of X ∈ B(H), with self- adjoint operators X = X† representing physical observables. The adjoint operation corresponds to the transpose conjugate when applied to a matrix representation of X, with the simple transpose being denoted by XT and the entry-wise conjugation by X∗. To avoid confusion, we shall use I to indicate the identity operator on B(H), whereas I will indicate the identity map (or super-operator) from B(H) to itself. We shall use X ≡ Y to say that X is defined as Y . The convex subset D(H) ⊂ B(H) of trace-one, positive-semidefinite operators, called density operators, is associated to physical states. We are concerned with state changes in the Schrödinger picture between two arbitrary points in time, say 0 and
79 Asymptotic stabilization of quantum states with continuous-time quasi-local dynamics
t > 0, which are described by a completely-positive trace-preserving (CPTP) linear map (or quantum channel) on B(H) [136]. A map Tt is CP if and only if it admits an operator-sum representation,
X † ρ(t) = Tt(ρ0) = Mkρ0Mk , ρ0 ∈ D(H), (5.1) k
P † for some {Mk} ⊂ B(H), and is also TP if in addition k Mk Mk = I. The oper- ators Mk are referred to as (Hellwig-)Kraus operators or operation elements [158]. The operator-sum representation of a CPTP map is not unique, and new decompo- 1 sitions may be obtained from unitary changes of the operators Mk. Dual dynamics with respect to the Hilbert-Schmidt inner product on B(H) (Heisenberg picture) are associated to unital CP maps T †, that is, obeying the condition T †(I) = I. A continuous one-parameter semigroup of CPTP maps {Tt}t≥0, with T0 = I, characterized by the Markov composition property Tt ◦ Ts = Tt+s, for all t, s ≥ 0, will be referred to as a Quantum Dynamical Semigroup (QDS) [136]. We shall denote by Lt L the corresponding semigroup generator, Tt = e , with the corresponding dual QDS † † {Tt }t≥0 being described by the generator L . It is well known that L (also referred to as the “Liouvillian”) can be always expressed in Lindblad canonical form [159, 160], that is, in units where ~ = 1: X 1 ρ˙(t) = L (ρ(t)) ≡ −i[H, ρ(t)] + L ρ(t)L† − {L† L , ρ(t)} , t ≥ 0, (5.2) k k 2 k k k
where H = H† is a self-adjoint operator associated with the effective Hamiltonian (generally resulting from the bare system Hamiltonian plus a “Lamb shift” term), and the Lindblad (or noise) operators {Lk} specify the non-Hamiltonian component of the generator, resulting in non-unitary irreversible dynamics. Equivalently, L defines a valid QDS generator if and only if it may be expressed in the form (see e.g. Theorem 7.1 in [161]) 1 L(ρ) ≡ E(ρ) − (κρ + ρκ†), κ ≡ iH + E †(I), (5.3) 2 where E is a CP map and the anti-Hermitian part of κ identifies the Hamiltonian operator. We shall denote by L(H, {Lk}) the QDS generator associated to Hamiltonian H and noise operators {Lk}. Throughout this chapter, both H and all the Lk will be assumed to be time-independent, with (5.2) thus defining a linear time-invariant
1While from a probabilistic and operator-algebra viewpoint it would be more natural to consider the dynamics acting on the states as (pre-)dual, we follow here the standard quantum physics notation as it allows for a more direct connection with existing work as well as a more compact notation in our context.
80 5.2 Preliminaries dynamical system. It is important to recall that, as for CPTP maps, the Lind- blad representation is also not unique, namely, the same generator can be associ- ated to different Hamiltonian and noise operators (see e.g. Proposition 7.4 in [161]), and, further to that, the separation between the Hamiltonian and the noise oper- ators is not univocally defined [31, 161]. Specifically, the Liouvillian is unchanged, 0 0 L(H, {Lk}) = L(H , {Lk}), if the new operators may be obtained as (i) linear combi- 0 0 P ∗ † nations of H, {Lk} and the identity, Lk = Lk + ckI,H = H − (i/2) k(ckLk − ckLk), 0 P 0 with ck ∈ C; or (ii) unitary linear combinations, Lk = l uklLl, H = H, with U ≡ {ukl} a unitary matrix (and the smaller set “padded” with zeros if needed), corresponding to a change of operator-sum representation for E in Eq. (5.3). We will denote a †-closed associative subalgebra A ⊆ B(H) generated by a set of operators X1,...,Xk as A ≡ alg{X1,...,Xk}. If T ≡ T ({Mk}) and L ≡ L(H, {Lk}) are a CP map and a QDS generator, then we shall let alg{T } ≡ alg{Mk} and alg{L} ≡ alg{H,Lk}, respectively. These algebras are invariant with respect to the change of representation in the Kraus or, respectively, Hamiltonian and Lindblad operators since, as remarked, equivalent representations are linearly related to one another. Let ⊕ denote the orthogonal direct sum. It is well known that any †-closed associative subalgebra A of B(H) admits a block-diagonal Wedderburn decomposi- tion [162], namely, H may be decomposed in an orthogonal sum of tensor-product bipartite subspaces, possibly up to a summand: M M (A) (B) H ≡ H` ⊕ HR = H` ⊗ H` ⊕ HR, (5.4) ` ` in such a way that M (A) (B) A = B(H` ) ⊗ I` ⊕ OR, (5.5) ` (B) (B) where I` represents the identity operator on the factor H` and OR the zero operator 0 on HR, respectively. Relative to the same decomposition, the commutant A of A in B(H), given by A0 ≡ {Y | [Y,X] = 0, ∀X ∈ A}, has the dual structure 0 M (A) (B) A = I` ⊗ B(H` ) ⊕ B(HR). (5.6) `
Consider now a density operator ρ ∈ D(H) such that supp(A) ⊆ supp(ρ), where for a generic operator space W the support is henceforth defined as supp(W ) ≡ ∪O∈W supp(O). It then follows that
1 1 1 1 Aρ ≡ ρ 2 A ρ 2 = {Y | Y = ρ 2 Xρ 2 ,X ∈ A} ⊆ B(H) (5.7)
81 Asymptotic stabilization of quantum states with continuous-time quasi-local dynamics
is a †-closed subalgebra with respect to the modified operator product X1 ◦ρ X2 ≡ −1 X1ρ X2, where the inverse is in the sense of Moore-Penrose [163] if ρ is not invertible. In the QIP literature, an associative algebra like in Eq. (5.7), which may be thought of as arising from a standard associative algebra A upon replacing each identity factor (B) in Eq. (5.5) with a fixed matrix τ` in each factor has been termed a distorted algebra [164, 165]. In particular, we shall call Aρ a ρ-distorted algebra, and refer to the map 1 1 Φρ(X) ≡ ρ 2 Xρ 2 as a “distortion map”. The ρ-distorted algebra generated by a set of operators X1,...,Xk will be correspondingly denoted by Aρ ≡ algρ({X1,...,Xk}).
5.2.2 Fixed points of quantum dynamical semigroups States that are invariant (aka stationary or “fixed”) under the dissipative dynamics of interest will play a central role in our analysis. Let fix(T ) indicate the set of fixed Lt points of a CP map T ; when Tt = e for t ≥ 0, then clearly fix(Tt) = ker(L). In this section, we summarize relevant results on the structure of fixed-point sets for CPTP maps, and slightly extend them to continuous-time QDS evolutions. Recall that fixed points of unital CPTP maps form a †-closed algebra: this stems from the fact that alg{T } = alg{T †}, along with the following result (see e.g. Theo- rem 6.12 in [161]):
Lemma 5.2.1. Given a CPTP map T , the commutant alg{T }0 is contained in fix(T †). In particular, if there exists a positive-definite state ρ > 0 in fix(T ), then
alg{T }0 = fix(T †). (5.8)
If T is CPTP and unital, its dual map always admits the identity as a fixed point of full rank. It then follows that fix(T ) = alg{T }0 [166, 165]. A similar result can be established for QDS generators (Theorem 7.2, [161]):
Lemma 5.2.2. Given a QDS generator L, the commutant alg{L}0 is contained in the kernel of L†. In particular, if L(ρ) = 0 for some ρ > 0, then
alg{L}0 = ker(L†). (5.9)
A key result to our aim is that, in general, the set of fixed points of a QDS has the structure of a distorted algebra. The following characterization is known for arbitrary (non-unital) CPTP maps (see e.g. Corollary 6.7 in [161]):
Theorem 5.2.3. Given a CPTP map T and a full-rank fixed point ρ,
1 † 1 fix(T ) = ρ 2 fix(T ) ρ 2 , (5.10)
82 5.2 Preliminaries
† L (A) (B) Moreover, with respect to the decomposition fix(T ) = ` B(H` ) ⊗ I` , we have
M (A) (B) ρ = ρ` ⊗ τ` , (5.11) `
(A) (B) where ρ` and τ` are full-rank density operators of appropriate dimension. Building on the previous results, an analogous statement can be proved for QDS dynamics:
Theorem 5.2.4. (QDS fixed-point sets, full-rank case) Given a QDS generator L and a full-rank fixed point ρ,
1 † 1 ker(L) = ρ 2 ker(L ) ρ 2 . (5.12)
† L (A) (B) Moreover, with respect to the decomposition ker(L ) = ` B(H` ) ⊗ I` , we have
M (A) (B) ρ = ρ` ⊗ τ` , (5.13) `
(A) (B) where ρ` and τ` are full-rank density operators of appropriate dimension.
Lt Proof. In order for {e }t≥0 to be a QDS, and thus a semigroup of trace-norm con- tractions [136], L must have spectrum in the closed left-half of the complex plane and no purely imaginary eigenvalues with multiplicity. It is then easy to show, by resorting to its Jordan decomposition [161], that the following limit exists:
Z t 1 Lτ T∞ ≡ lim e dτ. t→∞ t 0 Being the limit of convex combination of CPTP maps, which form a closed convex set, T∞ it also CPTP. Furthermore, T∞ projects onto ker(L), namely, fix(T∞) = ker(L), and T∞ has only eigenvalues 0, 1 with simple Jordan blocks. Similarly, it follows that † † † the unital CP map T∞ ≡ (T∞) projects onto ker(L ). Using these facts along with Theorem 5.2.3, we then have:
1 1 1 1 2 † 2 2 † 2 ker(L) = fix(T∞) = ρ fix(T∞) ρ = ρ ker(L ) ρ .
The structure of the fixed point, Eq. (5.13), follows from Theorem 5.2.3 applied to T∞.
The above two theorems make it clear that, given discrete- or continuous-time CPTP dynamics admitting a full-rank invariant state ρ, the fixed-point sets fix(T )
83 Asymptotic stabilization of quantum states with continuous-time quasi-local dynamics
and ker(L) are a ρ-distorted algebra with structure
M (A) (B) Aρ = B(H` ) ⊗ τ` , (5.14) `
(B) where the states τ` are the same for every element in fix(T ) or ker(L). In addition, since ρ has a compatible block structure [Eq. (5.13)], it is immediate to see that fix(T ) and ker(L) are invariant with respect to the action of the linear map Mλ(X) ≡ λ −λ ρ Xρ for any λ ∈ C, and in particular for the modular group {Miφ} [167]. The same holds for the fixed points of the dual dynamics. In fact, we can show that modular invariance is also a sufficient condition for a distorted algebra to be a fixed-point set for a CPTP map that fixes ρ, as relevant to the problem of designing stabilizing dynamics for ρ. In order to do this, we need a result by Takesaki [168], which we give in its finite-dimensional formulation (adapted from [167], Theorem 9.2): Theorem 5.2.5. Let A be a †-closed subalgebra of B(H), and ρ a full-rank density operator. Then the following are equivalent: (i) There exists a unital CP map E † such that fix(E †) = A, (E †)2 = E † and E(ρ) = ρ. 1 − 1 (ii) A is invariant with respect to M 1 , that is, for every X ∈ A, ρ 2 Xρ 2 ∈ A. 2 These conditions are equivalent to saying that the map E † is a conditional expec- tation on A that preserves ρ. We can then prove the following: Theorem 5.2.6. (Existence of ρ-preserving dynamics) Let ρ be a full-rank den- sity operator. A distorted algebra Aρ admits a CPTP map T such that fix(T ) = Aρ if and only if it is invariant for M 1 . 2
Proof. First, notice that, if Aρ is a distorted algebra, then it is invariant for M 1 if 2 − 1 − 1 and only if the “undistorted” algebra A ≡ ρ 2 Aρρ 2 is invariant for M 1 . This follows 2 from the fact that M 1 commutes with both the distortion map and its inverse. In 2 particular, if Aρ is invariant for M 1 , we have: 2
− 1 − 1 − 1 − 1 − 1 − 1 M 1 (A) = M 1 (ρ 2 Aρρ 2 ) = ρ 2 M 1 (Aρ)ρ 2 ⊆ ρ 2 Aρρ 2 = A. 2 2 2
Thus, by Theorem 5.2.5, a unital CP projection E † onto A exists whose adjoint preserves ρ. By Theorem 5.2.3, the CPTP dual E ≡ T is such that fix(T ) = Aρ, as desired. To prove the other implication, it is sufficient to notice that Eq. (5.11) implies † that M 1 leaves A = fix(T ) invariant, and thus 2
1 1 1 1 1 1 M 1 (Aρ) = M 1 (ρ 2 Aρρ 2 ) = ρ 2 M 1 (A)ρ 2 ⊆ ρ 2 Aρ 2 = Aρ. 2 2 2
84 5.2 Preliminaries
If the dynamics admit no full-rank fixed state, we may restrict to the support of a given fixed point, which is an invariant subspace for the Schrödinger’s-picture evolution:
Theorem 5.2.7. (QDS fixed-point sets, general case) Given a finite-dimensional QDS generator L, and a maximal-rank fixed point ρ with H˜ ≡ supp(ρ), let L˜ denote the reduction of L to B(H˜). We then have
1 † 1 ker(L) = ρ 2 (ker(L˜ ) ⊕ O) ρ 2 . (5.15)
Proof. For any ρ ∈ ker(L), the subspace H˜ ≡ supp(ρ) is invariant for the dynam- ˜ ˜ ics [138]. Assume that L = L(H, {Lk}) and let Π: H → H denote the partial isometry onto H˜. Define the reduced (projected) operators ρ˜ ≡ Π˜ρΠ˜ †, H˜ ≡ Π˜HΠ˜ †, ˜ ˜ ˜ † ˜ and Lk ≡ ΠLkΠ . The dynamics inside H is then determined by the correspond- ˜ ˜ ˜ ing projected Liouvillian L(H, {Lk}) [169], and ρ˜ is, by construction, a full-rank state for this dynamics. Hence, the fixed-point set ker(L˜) is the distorted algebra 1 † 1 ker(L˜) = ρ 2 ker(L˜ )ρ 2 . Consider now a maximal-rank fixed point, satisfying supp(ρ) = H˜ = supp(ker(L)). It then follows from Theorem 9 in [138] that H˜ is not only invariant but also attractive for the dynamics. This means that ˜ ⊥ Lt lim Tr Π e (ρ0) = 0, ∀ρ0 ∈ D(H). t→∞
With H˜ being attractive, we have that ker(L) can have support only in H˜, and can thus be constructed by appending the zero operator on H˜⊥, so that, using Theorem 5.2.4:
1 † 1 ker(L) = ker(L˜)⊕O = ρ 2 (ker(L˜ )⊕O) ρ 2 .
In the above proof, we made the construction explicit in terms of a representation L = L(H, {Lk}) in order to make it clear that the result does not hold if we consider 1 † 1 † ρ 2 ker(L ) ρ 2 , since H˜ need not be invariant for L . Again, it follows that ρ admits a block decomposition as in Eq. (5.13), compatible with that of ker(L˜†) on its support.
85 Asymptotic stabilization of quantum states with continuous-time quasi-local dynamics
5.2.3 Quasi-local semigroup dynamics Throughout this chapter, the open quantum system of interest will consist of a finite number n of distinguishable subsystems, defined on a tensor-product Hilbert space
n O H = Ha, dim(Ha) = da, dim(H) = D. a=1 As in [31, 34], we shall introduce quasi-locality constraints on the system’s evolution by specifying a list of neighborhoods, namely, groups of subsystems on which operators are allowed to “act simultaneously”. Mathematically, neighborhoods {Nj} may be specified as subsets of the set of indexes labeling the subsystems, that is,
Nj ⊆ {1, . . . , n}, j = 1,...,M.
Each neighborhood induces a bipartite tensor-product structure of H as O O H = H ⊗ H , H ≡ H , H ≡ H . (5.16) Nj N j Nj a N j a a∈Nj a/∈Nj
Likewise, with a neighborhood structure N ≡ {Nj} in place, any state ρ ∈ D(H)
uniquely determines a list of reduced neighborhood states {ρNj }:
ρ ≡ Tr (ρ), ρ ∈ D(H ), j = 1,...,M, (5.17) Nj N j Nj Nj where Tr indicates the partial trace over H . Quasi-local dynamical constraints N j N j may be specified by requiring compatibility with the bipartitions in (5.16), in the following sense:
Definition 5.2.8. (Neighborhood operator) An operator X ∈ B(H) is a neigh- borhood operator relative to a given neighborhood structure N if there exists j such
that the action of X is non-trivial only on HNj , that is:
X = X ⊗ I , Nj N j where I is the identity operator on H . N j N j A similar definition may be given for neighborhood CPTP maps and generators. The relevant quasi-locality notion for QDS dynamics is then the following:
Definition 5.2.9. (QL semigroup) A QDS generator L is Quasi-Local (QL) rela- tive to a given neighborhood structure N if it may be expressed as a sum of neighbor- hood generators:
86 5.2 Preliminaries
X L = L , L ≡ L ⊗ I . (5.18) j j Nj N j j
Quasi-locality of a Liouvillian is well-defined, as the structural property in Eq. (5.18) is defined independently of a particular representation of the generator. In terms of an explicit representation, the above definition is equivalent to requiring that there exists some choice L ≡ L(H, {Lk}), such that each Lindblad operator Lk is a neigh- borhood operator and the Hamiltonian may be expressed as a sum of neighborhood Hamiltonians, namely: X L = L ⊗ I ,H = H ,H ≡ H ⊗ I . k k,Nj N j j j Nj N j j
A Hamiltonian H of the above form is called a QL Hamiltonian (often “few-body,” in the physics literature)2. Mathematically, this denomination is natural given that, for the limiting case of closed-system dynamics, a QL Hamiltonian so defined au- tomatically induces a QL (Lie-)group action consistent with Eq. (5.18), with Lj ≡ i adHj (·) = i [Hj, ·].
Remark 1. The above QL notion is appropriate to describe any locality constraint that may be associated with a spatial lattice geometry and finite interactions range (e.g., spins living on the vertices of a graph, subject to nearest-neighbor couplings). QL semigroup dynamics have also been considered under less restrictive assumptions on the spatial decay of interactions [171], and yet different QL notions may be po- tentially envisioned (e.g. based on locality in “momentum space” or relative to “error weight”). The present choice provides the simplest physically relevant setting that allows for a direct linear-algebraic analysis. We stress that, due to the freedom in the representation of the QDS generator, QL semigroup dynamics may still be induced by Lindblad operators that are not manifestly of neighborhood form. In principle, it is always possible to check the QL property by verifying whether a QDS generator L has components only in the (super-)operator subspace spanned by QL generators. While it may be interesting to determine more operational and efficient QL criteria in specific cases, in most practical scenarios (e.g. open-system simulators [147]) avail- able Lindblad operators are typically specified in a preferred neighborhood form from the outset.
2In particular, the notions of neighborhood Hamiltonian and QL Hamiltonian reduce to the standard uni-local and local ones for non-overlapping neighborhoods, see e.g. [170].
87 Asymptotic stabilization of quantum states with continuous-time quasi-local dynamics
5.2.4 Quasi-local stabilizability: prior pure-state results and frustration-free semigroup dynamics Our main focus will be on determining conditions under which a certain state of interest, ρ, is guaranteed to be invariant and the unique asymptotically stable state for some QL dynamics. Formally, an invariant state ρ ∈ D(H) for a QDS with generator L is said to be Globally Asymptotically Stable (GAS) if
Lt lim e (ρ0) = ρ, ∀ρ0 ∈ D(H). (5.19) t→+∞
For QDS dynamics not subject to QL constraints it is known that a state is GAS if and only if it is the unique fixed point [161, 30]. A definition of stabilizable states relevant to our constrained setting may be given as follows:
Definition 5.2.10. (QLS state) A state ρ ∈ D(H) is Quasi-Locally Stabilizable (QLS) relative to a neighborhood structure N if there exists a QL generator L for which ρ is GAS.
Existing work has so far focused on stabilizability of a pure state, with special emphasis on steady-state entanglement [31, 32, 34]. While even in this case, in general, a careful balancing of Hamiltonian and dissipative action is essential, a simple yet non- trivial stabilization setting arises by further requiring that the target can be made QLS by a generator without a Hamiltonian component, namely, by using dissipation alone. Given the freedom in the representation of a QDS generator, in order to formalize this additional constraint we introduced a standard representation for a generator L(H, {Lk}) that fixes a pure state ρ ≡ |ΨihΨ| ∈ D(H), as in the following result (Corollary 1 in [34]):
Proposition 5.2.11. If a generator L(H, {Lk}) makes ρ = |ΨihΨ| GAS, then the ˜ ˜ same generator can be represented in a standard form L(H, {Lk}), in such a way that ˜ ˜ H|Ψi = h|Ψi, h ∈ R and Lk|Ψi = 0, for all k. ˜ In the standard representation, the target |Ψi ∈ ker(Lk) may thus be seen as a common “dark state” for all the noise operators, borrowing from quantum-optics ter- minology. With this in mind, a pure state ρ = |ΨihΨ| may be defined as Dissipatively Quasi-Locally Stabilizable (DQLS) if it is QLS with H˜ ≡ 0 and QL noise operators ˜ {Lk} in standard form. Notice that such a definition implies that ρ is invariant for the ˜ dynamics relative to each neighborhood, namely, ρ ∈ ker(L({Lk})), for each k. Build- ing on this QL-invariance condition allows for proving the following characterization of DQLS states [31]:
88 5.2 Preliminaries
Theorem 5.2.12. A pure state ρ = |ΨihΨ| ∈ D(H) is DQLS relative to N if and only if \ supp(ρ) = supp(ρ ⊗ I ). (5.20) Nk N k k Remark 2. The proof of the above result includes the construction of a set of ˜ stabilizing Lindblad operators Lk that make ρ DQLS, also implying that one such operator per neighborhood always suffices. It is easy to show that any rescaled ˜0 ˜ version of the same operators, Lk ≡ rkLk, also yield a stabilizing QL generator 0 P 2 P L = k |rk| Lk ≡ k γkLk – incorporating “model (γ-)robustness,” in the terminol- ogy of [169]. However, the reasoning followed for QL stabilization of a pure state does not ex- tend naturally to a general, mixed target state. The main reason is that the standard form, hence the DQLS definition itself, do not have a consistent analogue for mixed states. A major simplification if ρ is pure stems from the fact that it is straightforward to check for invariance, directly in terms of the generator components (see Proposi- tion 1 in [34]); for general ρ, we seek a definition that extends the DQLS notion, and that similarly allows for explicitly studying what the invariance of ρ means at a QL level. A natural choice is to restrict to the class of frustration-free dynamics. That is, in addition to the QL constraint, we demand that each QL term in the generator leave the state of interest invariant. Formally, we define [156]: P Definition 5.2.13. (FF generator) A QL generator L = j Lj is Frustration Free (FF) relative to a neighborhood structure N = {Nj} if any invariant state ρ ∈ ker(L) also satisfies neighborhood-wise invariance, namely, ρ ∈ ker(Lj) for all j. Beside allowing for considerable simplification, FF dynamics are of practical interest because they are, similar to the DQLS setting, robust to certain perturbations. As in P Remark 5.2.4, given a QL generator L = j Lj, define a “neighborhood-perturbed” 0 P + QL generator L = j λjLj, with λj ∈ R . If L is FF, then L(ρ) = 0 implies 0 Lj(ρ) = 0 for each j; therefore, λjLj(ρ) = 0 for each j, and thus L (ρ) = 0 as well. Were L not FF, then the kernel of L would not be robust against such neighborhood- perturbations in general. With these motivations, we introduce the notion of QL stability that we analyze for the remainder of this chapter :
Definition 5.2.14. (FFQLS state) A state ρ ∈ D(H) is Frustration-Free Quasi- Locally Stabilizable (FFQLS) relative to a neighborhood structure N if it is QLS with a stabilizing generator L that is FF.
Remarkably, studying FFQLS states will allow us to recover the results for DQLS pure states as a special case. In fact, a pure state is FFQLS if and only if it is DQLS. This claim is proved in Appendix A of [172]. To summarize, the above definition consists of
89 Asymptotic stabilization of quantum states with continuous-time quasi-local dynamics
four distinct mathematical conditions that the generator L of the dissipative dynamics must obey for a given target invariant state ρ: • (QDS): L is a generator of a CPTP continuous semigroup; • (QL): L is QL, that is, L = P L , with L = L ⊗ I ; j j j Nj N j • (GAS): ρ is GAS, or equivalently ker(L) = span(ρ);
• (FF): L is FF, namely ker(Lj) ≥ span(ρ) for all Nj. The problem we are interested in is to determine necessary and sufficient conditions for a given state to be FFQLS and, if so, to design QL FF dynamics that achieves the task.
5.3 Frustration-Free Stabilizable States: Necessary Conditions
In this section, we derive necessary conditions for a target state to be FFQLS. Frustration-freeness requires such a state to be in the kernel of each neighborhood generator. We show that, if a neighborhood generator is to leave a global state in- variant, the size and structure of its kernel are constrained; in general, the kernel will be larger than the span of the reduced neighborhood state (as a vector in Hilbert- Schmidt space). However, if the target state is to be the unique fixed point of the QDS dynamics, then the intersection of all the neighborhood-generator kernels must coincide with the span of the target state. We shall show in Section 5.4.2 that this condition is also sufficient for a generic (full-rank) state to be FFQLS.
5.3.1 Linear-algebraic tools
Recall that, given a tensor product of two inner-product spaces V = VA ⊗ VB and a vector v ∈ V , a Schmidt decomposition of v is any decomposition X v = λiai ⊗ bi, i
where ai ∈ VA, bi ∈ VB, λi > 0, and {ai}, {bi} are each orthonormal sets of vectors. There are two instances of Schmidt decomposition which are relevant in our context, both well known within QIP [173]. The first is the Schmidt decomposition of a bipartite pure state |ψi ∈ HA ⊗ HB, namely, X |ψi = λi|aii ⊗ |bii. i
90 5.3 Frustration-Free Stabilizable States: Necessary Conditions
The second is the so-called operator-Schmidt decomposition, whereby a bipartite operator M ∈ B(HA ⊗ HB) = B(HA) ⊗ B(HB) is factorized in terms of elements in the vector spaces B(HA) and B(HB), relative to the Hilbert-Schmidt inner product. Specifically, X M = λiAi ⊗ Bi, i † † where Ai ∈ B(HA), Bi ∈ B(HB), λi > 0, and Tr(Ai Aj) = Tr(Bi Bj) = δij. Building on the concept of Schmidt decomposition, we introduce the Schmidt span:
Definition 5.3.1. (Schmidt span) Given a tensor product of two inner product P spaces V = VA ⊗ VB and a vector v ∈ V with Schmidt decomposition v = i λiai ⊗ bi, the Schmidt span of v relative to VA is the subspace n X o ΣA(v) = span ai ∈ VA | v = ai ⊗ bi, bi ∈ VB . (5.21) i
Without referring to a particular tensor-product decomposition, it is possible to show that the Schmidt span is the image of v under partial inner product:
† ΣA(v) = {a ∈ VA | a = (IA ⊗ b )v for some b ∈ VB}. (5.22)
dA dB † One example is when VA ⊗ VB is a matrix space, such as C ⊗ C (where the latter factor is meant as a space of row vectors). In this case, the Schmidt span of a matrix W ∈ CdA ⊗ CdB †, relative to the first factor CdA , is simply the range of the matrix W (namely, the set of all linear combinations of its column vectors), namely ΣA(W ) = range(W ). Similarly, the Schmidt span of W relative to the second factor CdB † is the orthogonal complement of the kernel of W or, in other words, the support: ΣB(W ) = supp(W ). Another example is when VA ⊗ VB is a bipartite operator space, such as B(HA)⊗B(HB). The Schmidt span of M ∈ B(HA)⊗B(HB) relative to B(HA), i.e. “on A”, is the operator subspace ΣA(M) = {TrB [(IA ⊗ B)M] ,B ∈ B(HB)}. The Schmidt span is a useful tool because conditions on how a neighborhood operator is to affect a global state constrains how such an operator must act on the entire operator-Schmidt span of that state. This intuition is formalized in the following Lemma:
Lemma 5.3.2. (Invariance of Schmidt span) Given a vector v ∈ VA ⊗ VB and 0 0 0 MA ∈ B(VA), if (MA ⊗ IB)v = λv, then (MA ⊗ IB)v = λv for all v ∈ ΣA(v) ⊗ VB. In particular:
span(v) ≤ ker(MA ⊗ IB) ⇒ ΣA(v) ⊗ VB ≤ ker(MA ⊗ IB), (5.23)
91 Asymptotic stabilization of quantum states with continuous-time quasi-local dynamics
and span(v) ≤ fix(MA ⊗ IB) ⇒ ΣA(v) ⊗ VB ≤ fix(MA ⊗ IB). (5.24)
P Proof. Consider the Schmidt decomposition of v, v = i γiai ⊗ bi, where ai ∈ VA, bi ∈ VB, γi > 0, and {ai}, {bi} are each orthonormal sets of vectors. Applying the eigenvalue equation for MA to this yields X X γiMAai ⊗ bi = λ γiai ⊗ bi. i i
† † Multiplying both sides by IA ⊗ bj, where bj is the dual vector of bj, selects out the VA-factor of the ith term, i.e., MAai = λai. This holds for each i and any linear combination of the ais. By definition, the Schmidt span of v is ΣA(v) = span{ai}. 0 0 P Denoting {βi} a basis for VB, we may write any v ∈ ΣA(v)⊗VB as v = ij µijai ⊗βj. Applying (MA ⊗ IB) to this we obtain
0 X 0 (MA ⊗ IB)v = µijMAai ⊗ βj = λv . ij
Thus, all elements in ΣA(v) ⊗ VB have eigenvalue λ with respect to MA ⊗ IB, as claimed. Eqs. (5.23) and (5.24) follow by specializing the above result to λ = 0 and λ = 1, respectively.
5.3.2 Invariance conditions for quasi-local generators As remarked, we require the global dynamics to be FF. This simplifies considerably the analysis, as global invariance of the target state is possible only if the latter is invariant for each neighborhood generator. Therefore, we examine the properties of a neighborhood generator that ensure the target state ρ to be in its kernel. Note that if ρ is factorizable relative to the neighborhood structure (i.e., a pure or mixed product state), ρ is invariant as long as each factor of ρ if fixed. Each such reduced neighborhood state can then be made not only invariant but also attractive by a neighborhood generator, if the reduced states are the only elements in the kernels of the corresponding LNj . This automatically makes the global factorized state also GAS. In other words, if ρ is factorizable, then QL stabilizability is guaranteed. If the target state is non-factorizable (in particular, entangled), the above scheme need not work; a non-factorizable state will have some operator Schmidt spans with dimension greater than one. The following Corollary, which follows from Lemma 5.3.2, illustrates the implication of quasi-locally fixing a state with non-trivial operator-Schmidt spans:
92 5.3 Frustration-Free Stabilizable States: Necessary Conditions
Corollary 5.3.3. Let L = L ⊗ I be a neighborhood Liouvillian. If ρ ∈ ker(L ), j Nj N j j then it must also be that
Σ (ρ) ⊗ B(H ) ≤ ker(L ). Nj N j j
Accordingly, if each neighborhood generator Lj is to fix a non-factorizable ρ (as is necessary for global invariance with FF dynamics), then each neighborhood generator must be constructed to leave invariant, in general, a larger space of operators – specifically, the corresponding neighborhood operator-Schmidt span of ρ. However, leaving only the Schmidt spans invariant is, in general, not possible if the dynamics are to be CPTP, since a Schmidt span need not be a distorted algebra (as required by Theorem 5.2.7). We show that, in order for ρ to be in the kernel of a valid QL generator, it is necessary that the dynamics leave certain “minimal fixed- point sets” generated by the Schmidt spans invariant as well. We give the following:
Definition 5.3.4. (Minimal modular-invariant distorted algebra) Let ρ ∈ D(H) be a density operator, and W ⊆ B(H). The minimal modular-invariant dis- torted algebra generated by W is the smallest ρ-distorted algebra generated by W 1 − 1 which is invariant with respect to M 1 (X) = ρ 2 Xρ 2 , where the inverse is in the 2 sense of Moore-Penrose if ρ is not full-rank.
In the finite-dimensional case that we consider, Fρ(W ) can be constructed by the 0 following iterative procedure: define F ≡ algρ(W ), and compute
k+1 k F = alg (M 1 (F )), ρ 2
k+1 k until F = F ≡ Fρ(W ). This particular distorted algebra is the smallest structure whose invariance is required if the dynamics are to be CPTP:
Lemma 5.3.5. (Minimal fixed-point sets) Let W ≤ B(H) be an operator subspace containing a positive-semidefinite operator ρ such that supp(ρ) = supp(W ). If W ≤ fix(T ) for a CPTP map T : B(H) → B(H), then
Fρ(W ) ≤ fix(T ). (5.25)
Proof. Given the iterative construction of Fρ(W ), it suffices to show that if some set W ⊆ fix(T ) includes a density operator with supp(ρ) = supp(W ), then both alg (W ) ⊆ fix(T ) and M 1 (W ) ⊆ fix(T ), and their support is still equal to supp(ρ). ρ 2 Since T is a CP linear map, fix(T ) is closed with respect to linear combinations and
93 Asymptotic stabilization of quantum states with continuous-time quasi-local dynamics
†-adjoint. We are left to show that fix(T ) is closed with respect to the ρ-modified product. Consider a partial isometry V : supp(ρ) → H, and define the reduced map
T˜ : B(supp(ρ)) → B(supp(ρ)), T˜(X) ≡ V †T (VXV †)V.
Since ρ is invariant, the set of operators with support contained in supp(ρ) = supp(W ) is an invariant subspace, and thus T (VV †XVV †) = VV †T (VV †XVV †)VV †. By construction, T˜ is CP, TP, and has a full-rank fixed point ρ˜ ≡ V †ρV . It follows from Theorems 5.2.3 and 5.2.6 that fix(T˜) is a ρ-distorted algebra; hence, it is closed with respect to the modified product, as well as modular-invariant. Now, if X,Y ∈ W are fixed points for T , so are X˜ = V †XV, Y˜ = V †YV † for T˜. Since their adjoint, linear combinations and ρ-distorted products are in fix(T˜), we have:
T (Xρ−1Y ) = T (VV †XVV †ρ−1VV †YVV †) = VV †T (V X˜ρ˜−1YV˜ †)VV † = V T˜(X˜ρ˜−1Y˜ )V † = V X˜ρ˜−1YV˜ † = V XV˜ †V ρ˜−1V †V YV˜ † = Xρ−1Y.
Hence, it must be algρ(W ) ≤ fix(T ), as desired, and we still have supp(algρ(W )) = supp(ρ). On the other hand, if X ∈ alg (W ), then supp(M 1 (X)) ∈ supp(W ), and we have: ρ 2
† 1 − 1 † † 1 − 1 † † T (M 1 (X)) = T (VV ρ 2 Xρ 2 VV ) = VV T (V ρ˜2 X˜ρ˜ 2 V )VV ρ 2 1 − 1 † 1 − 1 † = V T˜(˜ρ 2 X˜ρ˜ 2 )V = V ρ˜2 X˜ρ˜ 2 V 1 − 1 = ρ 2 Xρ 2 .
Accordingly, M 1 (W ) ∈ fix(T ) and supp(M 1 (X)) ⊆ supp(ρ) as well, as desired. 2 2
5.3.3 From invariance to necessary conditions for stabilizabil- ity In order to apply the above lemma to our case of interest, namely, finding necessary conditions for FFQLS, the first step is to show that the reduced neighborhood states of ρ may be used to generate the minimal ρ-distorted algebra containing the Schmidt span:
Proposition 5.3.6. Given a neighborhood Nj ∈ N , the support of the corresponding reduced state, ρ = Tr (ρ), is equal to the support of the operator-Schmidt span Nj N j
ΣNj (ρ).
94 5.3 Frustration-Free Stabilizable States: Necessary Conditions
Proof. Since ρNj ∈ ΣNj (ρ), supp(ρNj ) ≤ supp(ΣNj (ρ)). It remains to show the oppo- site inclusion, that is, by equivalently considering the complements, that ker(ρN ) ≤ j ker(ΣNj (ρ)). Let |ψi ∈ ker(ρNj ). Since ρNj ≥ 0, we then have Tr ρNj |ψihψ| = Tr (ρ(|ψihψ| ⊗ I)) = 0. Let {Ei} be a positive-operator valued measure (POVM) on HN which is informationally complete (that is, span{Ei} = B(HN )). The j P j POVM elements sum to I, giving i Tr (ρ(|ψihψ| ⊗ Ei)) = 0. Since each term is non-negative, Tr (ρ(|ψihψ| ⊗ E )) = 0 for all i. Letting ρ ≡ Tr (ρ(I ⊗ E )), we can i i N j i write 0 = Tr (ρ(|ψihψ| ⊗ Ei)) = hψ|ρi|ψi. Then, ρi ≥ 0 implies ρi|ψi = 0 for all i. Since the E span the operator space B(H ), by using Eq. (5.22), we have that the i N j
corresponding ρi span ΣNj (ρ). Hence, |ψi ∈ ker(ΣNj (ρ)).
The above Proposition, together with Lemma 5.3.2 and Lemma 5.3.5, imply the following:
Corollary 5.3.7. If a state ρ is in the kernel of a neighborhood generator Lj = L ⊗ I , then the minimal fixed-point set generated by the neighborhood Schmidt Nj N j span obeys F (Σ (ρ)) ⊗ B(H ) ≤ ker(L ). (5.26) ρNj Nj N j j
Proof. Assume that ρ ∈ ker(L ). By Lemma 5.3.2, we have Σ (ρ) ⊗ B(H ) ≤ j Nj N j
ker(Lj). By Proposition 5.3.6, we also know that the support of ρNj is equal to that of Σ (ρ), and hence supp(ρ ⊗ I ) = supp(Σ (ρ) ⊗ B(H )). With this and the Nj Nj N j Nj N j fact that ρ ⊗ I ∈ Σ (ρ) ⊗ B(H ) ≤ ker(L ), Nj N j Nj N j j Lemma 5.3.5 implies that FρN ⊗I ΣNj (ρ) ⊗ B(H ) ≤ ker(Lj), j N j N j
or, equivalently, F (Σ (ρ)) ⊗ B(H ) ≤ ker(L ), as desired. ρNj Nj N j j Summing up the results obtained on invariance so far, and recalling that unique- ness of the equilibrium state is necessary for GAS, we have the following necessary condition:
Theorem 5.3.8. (Necessary condition for FFQLS) A state ρ is FFQLS relative to the neighborhood structure N only if \ span(ρ) = F (Σ (ρ)) ⊗ B(H ). (5.27) ρNj Nj N j j
95 Asymptotic stabilization of quantum states with continuous-time quasi-local dynamics
Proof. Let ρ be FFQLS relative to N . Frustration-freeness of L implies that ker(L) = T ker(L ). From QLS, we have ker(L) = span(ρ). Thus, FFQLS implies Nj j \ ker(Lj) = span(ρ). j
Corollary 5.3.7 implies that for each neighborhood, F (Σ (ρ))⊗B(H ) ≤ ker(L ). ρNj Nj N j j Hence, \ \ F (Σ (ρ)) ⊗ B(H ) ≤ ker(L ). ρNj Nj N j j j j By construction, we also have \ span(ρ) ≤ F (Σ (ρ)) ⊗ B(H ). ρNj Nj N j j
Stringing together the three relationships above, we arrive at the desired result, Eq. (5.27).
5.4 Frustration-Free Stabilizable States: Sufficient Conditions
In this section, we move from necessary conditions for FFQL stabilization to sufficient ones, by providing in the process a constructive procedure to design stabilizing semi- group generators. A key step will be to establish a property that arbitrary (convex) sums of Liouvillians enjoy, namely the fact that, as long as the algebras associated with individual components of the generator are contained in the algebra associated to the full generator, the existence of a common full-rank fixed point suffices to prove frustration-freeness. Drawing on this result, we will prove that the necessary condi- tion of the previous section is also sufficient in the generic case where the target state is full-rank, and then separately address general target states.
5.4.1 A key result on frustration-free Markovian evolutions P Consider a QDS of the form L = k Lk, where individual terms need not, at this stage, correspond to neighborhood generators. The following general result holds:
Theorem 5.4.1. (Common fixed points of sums of Liouvillians) Let L = P k Lk be a sum of QDS generators, and assume that the following conditions hold: (i) alg{Lk} ≤ alg{L} for each k; (ii) there exists a positive definite ρ ∈ ker(L) such that ρ ∈ ker(Lk) for all k.
96 5.4 Frustration-Free Stabilizable States: Sufficient Conditions
0 Then ρ is invariant under L only if it is invariant under all Lk, that is:
0 0 ρ ∈ ker(L) =⇒ ρ ∈ ker(Lk) ∀ k.
T Proof. By linearity of L, we clearly have that ker(L) ≥ ker(Lk). We show that T k under the hypotheses, ker(L) ≤ ker(Lk), therefore effectively implying ker(L) = T k k ker(Lk). By (ii), ρ is a full-rank state in ker(L) and ρ ∈ ker(Lk) for all k. Theorem 5.2.4 implies that
1 1 1 † 1 2 † 2 2 2 ker(L) = ρ ker(L )ρ and ker(Lk) = ρ ker(Lk)ρ , ∀k. Then, by Lemma 5.2.2, we also have that
1 1 1 1 1 † 1 1 1 2 † 2 2 0 2 2 2 2 0 2 ρ ker(L )ρ = ρ alg{L} ρ and ρ ker(Lk)ρ = ρ alg{Lk} ρ . In view of condition (i), the relevant commutants satisfy
0 0 alg{L} ≤ alg{Lk} , ∀k.
The above inequality may then be used to bridge the previous equalities, yielding:
1 † 1 1 0 1 ker(L) = ρ 2 ker(L )ρ 2 = ρ 2 alg{L} ρ 2 ≤
1 † 1 1 1 2 2 2 0 2 ker(Lk) = ρ ker(Lk)ρ = ρ alg{Lk} ρ , T for all k. From this we obtain ker(L) ≤ k ker(Lk), which completes the proof.
Remark 3. We note that condition (i) above, namely alg{Lk} ≤ alg{L}, is only ever not satisfied due to the presence of Hamiltonian contributions in Lk. In fact, P if Lk ≡ Lk({Lj,k}) for each k in a given representation, then L = Lk also has a S k purely dissipative representation L( k{Lj,k}), and thus alg{Lk} ≤ alg{L}. On the other hand, suppose that Lk = Lk(Hk, {Lj,k}), with Hk 6= 0 in some representation. S P This implies that L = L(H, k{Lj,k}), with H = k Hk. In this case, since alg(H) need not contain alg(Hk), condition (i) does not hold in general. As a trivial example, consider two generators associated to H1 = M and H2 = −M, with M 6= I. Clearly, {O} = alg(H) does not contain alg(M). Likewise, if H = (C2)⊗3 and Z is a single- qubit Pauli operator, consider QL Hamiltonians H1 = ZZI and H2 = IZZ. Then alg(H1) alg(H1+H2). Intuitively, this stems from the fact that since noise operators enter “quadratically” (bilinearly) in the QDS, they cannot cancel each other’s action – unlike Hamiltonians, which by linearity may “interfere” with one another.
97 Asymptotic stabilization of quantum states with continuous-time quasi-local dynamics
Interestingly, the reasoning leading to Theorem 5.4.1 also applies to CPTP maps, with the simplification that, since no Hamiltonian is present, condition (i) is always satisfied. Formally:
Corollary 5.4.2. (Common fixed points of sums of CPTP maps) Let T = P P k pkTk be a sum of CPTP maps, with pk > 0 and k pk = 1. If there exists a 0 positive definite ρ ∈ fix(T ) such that ρ ∈ fix(Tk) for all k, then ρ is invariant under T only if it is invariant under all Tk, that is:
0 0 ρ ∈ fix(T ) =⇒ ρ ∈ fix(Tk).
The proof is a straightforward adaptation of the one above, and it can actually be extended to a positive semi-definite fixed state ρ, provided that some extra hypotheses on the support of ρ are satisfied. The precise statement and proof of this extended result are given in Appendix B of [172]. Another direct corollary of Theorem 5.4.1, which now specializes to locality- constrained dynamics, provides us with a useful tool to ensure that a QL generator be FF: the generator itself and all of its QL components must share a full-rank fixed state.
Corollary 5.4.3. (Frustration-freeness from full-rank fixed point) Let L = P j Lj be a QL generator, and assume that the following conditions hold: (i) alg{Lj} ≤ alg{L} for each j; (ii) there exists a positive-definite ρ ∈ ker(L) such that ρ ∈ ker(Lj) for all j. Then the QL generator L is FF.
5.4.2 Sufficient conditions for full-rank target states Theorem 5.4.4. (Sufficient condition for full-rank FFQLS) A full-rank state ρ is FFQLS relative to the neighborhood structure N if \ span(ρ) = F (Σ (ρ)) ⊗ B(H ). (5.28) ρNj Nj N j j
Proof. To show that this condition suffices for FFQLS, we must show that there exists some QL FF Liouvillian L for which span(ρ) = ker(L). Our strategy is to first construct a QL generator for which ρ is the unique state in the intersection of the QL-components’ kernels. Then, we use Thm. 5.4.3 to show that this generator is FF, yielding the desired equality.
Fix an arbitrary neighborhood Nj ∈ N , with associated bipartition H = HNj ⊗ H . We shall construct a neighborhood CPTP map E ≡ E ⊗ I , where E N j j Nj N j Nj
98 5.4 Frustration-Free Stabilizable States: Sufficient Conditions
projects onto the minimal fixed-point set containing the neighborhood-Schmidt span (that is, such projection maps are duals to a conditional expectation). Since, by construction, F (Σ (ρ)) is a modular-invariant distorted subalgebra of B(H ), ρNj Nj Nj
Theorem 5.2.6 ensures that there exists a CPTP map ENj such that
fix(E ) = F (Σ (ρ)). Nj ρNj Nj
In particular, we take E 2 = E , so that it projects onto its fixed points. Explicitly, Nj Nj its structure follows from the decomposition in Eq. (5.14):
M (A) (B) F (Σ (ρ)) = B(H ) ⊗ τ , ρNj Nj `,j `,j `
L L (A) (B) with a corresponding Hilbert space decomposition HNj ≡ ` H`,j = ` H`,j ⊗H`,j , (B) (B) and τ`,j a full-rank state on H`,j . Introducing partial isometries Π`,j : H`,j → HNj ,
the sought-after maps ENj can be constructed as:
M † (B) ENj (ρ) ≡ Tr (B) (Π`,jρΠ`,j) ⊗ τ`,j . (5.29) H`,j `
It is straightforward to verify that ENj (ρ) is CPTP. Recalling Eq. (5.3), we may then define a neighborhood QDS generator by taking κ = E † (I)/2 = I/2 and letting Nj
LNj ≡ ENj − INj , ∀j. (5.30)
Let now L ≡ P L = P L ⊗ I define the QL generator of the overall j j j Nj N j dynamics. We constructed each Lj in such a way that
ker(L ) = F (Σ (ρ)) ⊗ B(H ), ∀j. j ρNj Nj N j
Hence, by invoking the hypothesis (Eq. (5.28)), it follows that ρ is the unique state obeying \ span(ρ) = ker(Lj) ≤ ker(L). (5.31) j A priori, it is still possible that span(ρ) < ker(L). However, since we have chosen † κ = κ , the neighborhood generators Lj defined in Eq. (5.30) do not have any Hamiltonian contribution; recalling Remark 3, it follows that the algebra of the global generator contains the algebra of each neighborhood generator,
alg{L} ≥ alg{Lj}, ∀j.
99 Asymptotic stabilization of quantum states with continuous-time quasi-local dynamics
Thus, by Corollary 5.4.3, the generator L is FF. From L being FF, it follows in turn that \ ker(L) ≤ ker(Lj), j which, together with Eq. (5.31), implies span(ρ) = ker(L), as desired.
5.4.3 Sufficient conditions for general target states If the target state ρ is not full-rank, the necessary condition of Theorem 5.3.8 may still be shown to be sufficient for FFQLS if an additional condition (referred to as the “support condition” henceforth) is also obeyed:
Theorem 5.4.5. (Sufficient condition for general FFQLS) An arbitrary state ρ is FFQLS relative to the neighborhood structure N if \ span(ρ) = F (Σ (ρ)) ⊗ B(H ) (5.32) ρNj Nj N j j
and \ supp(ρ) = supp(ρNj ⊗ INj ). (5.33) j
Proof. Our strategy is to use the support condition of Eq. (5.33) to reduce the non-full-rank case to the full-rank one. As in the proof of the previous theorem, fix
an arbitrary neighborhood Nj, and consider the maps ENj , defined in Eq. (5.29). Let P ∈ B(H ) denote the Hermitian projector onto supp(F (Σ )(ρ)), and Nj Nj ρNj Nj P ⊥ = I − P the associated orthogonal projector. In this case, we compose each Nj Nj Nj
ENj with the corresponding map P E 0 (·) ≡ P (·)P + Nj Tr (P ⊥ ·), (5.34) Nj Nj Nj Nj Tr(PNj )
where E 0 is, like E , both CP and TP: Nj Nj Tr (E 0 (M)) = Tr M(P + P ⊥ ) = Tr(M), ∀M ∈ B(H ). Nj Nj Nj Nj
With this, consider new CPTP maps given by E ◦ E 0 , whereby it follows that new Nj Nj neighborhood generators may be constructed as
L ≡ E ◦ E 0 − I , L = L ⊗ I , (5.35) Nj Nj Nj Nj j Nj N j
100 5.4 Frustration-Free Stabilizable States: Sufficient Conditions
P with the global evolution being driven, as before, by the QL generator L = j Lj. Define now Π to be the projector onto supp(ρ), and consider the positive-semidefinite function V (τ) = 1 − Tr (Π τ) , τ ∈ B(H). The derivative of V along the trajectories of the generator we just constructed is
˙ X V (τ) = − Tr (Π Lj(τ)) . j
By LaSalle-Krasowskii theorem [174], the trajectories will converge to the largest in- variant set contained in the set of τ such that the above Lyapunov function V˙ (τ) = 0. We next show that this set must have support only on supp(ρ) = T supp(ρ ⊗I ). Nj Nj N j Since V is defined on global input operators, we first re-express each neighborhood generator Lj in Eq. (5.35) as
L = E ◦ E 0 − I, E ≡ E ⊗ I , E 0 ≡ E 0 ⊗ I , j j j j Nj N j j Nj N j
where we have used the property E ◦ E 0 = (E ◦ E 0 ) ⊗ I . Additionally, let j j Nj Nj N j P ≡ P ⊗ I denote the projector onto supp(F (Σ (ρ)) ⊗ I ). Assume now j Nj N j ρNj Nj N j that supp(τ) supp(ρ ⊗ I ) for some N ∈ N , that is, Tr(τP ⊥) > 0. By using * Nk N k k k the explicit form of the maps E 0 given in Eq. (5.34), we then have: Nj ˙ V (τ) ≤ −Tr (Π Lk(τ)) 0 = −Tr Π(Ek ◦ Ek )τ − Π τ