Hidden Nonlocality
Flavien Hirsch
September 2012 - June 2013
University of Geneva
Supervisor : Prof. Nicolas Brunner Abstract
In this essay we study the hidden nonlocality which can be revealed by local filtering. We present the idea and the physical motivations, and we recapitulate all the known cases. We then construct some new examples in lower dimension and discuss the local models of the Werner states. For the bipartite qubits case, we propose some improvements on these models that allow us to simulate a larger class of states, and we construct a protocol that, from a projective-local state, gives a POVM-local one. Finally, we prove the existence of genuine hidden nonlocality with a few concrete examples, including maximal and genuine hidden nonlocality.
2 Contents
Contents 3
1 Introduction 5
2 Concepts and motivations 6 2.1 Nonlocality and Bell’s game ...... 6 2.1.1 The scenario ...... 6 2.1.2 Classification of the probability distribution ...... 7 2.1.3 Bell’s game ...... 9 2.2 Entanglement ...... 11 2.2.1 Quantum formalism ...... 11 2.2.2 Characterization of quantum states ...... 12 2.3 SLOCC and hidden nonlocality ...... 14 2.4 Notations ...... 16 2.5 Classes of measurement ...... 17 2.5.1 Quantum observable ...... 18
3 Local states 19 3.1 Local models ...... 19 3.1.1 Simulate POVMs ...... 20 3.2 Werner state ...... 20 3.2.1 Werner’s model ...... 21 3.2.2 Barrett’s model ...... 22 3.2.3 Interpretation of Werner’s and Barrett’s model ..... 22 3.3 New classes of local states ...... 23 s 3.3.1 ra ...... 23 3.3.2 The model ...... 24 s 3.3.3 ra,b ...... 26 s 3.3.4 ra,b,g ...... 28
4 Examples of filterable states 30 4.1 General filtering and qubits filtering ...... 30 4.2 Werner state ...... 32 4.3 Gisin’s state ...... 33 4.4 Collins-Gisin’s state ...... 34 4.5 Hidden nonlocality for a two-qubits state ...... 35
3 CONTENTS 4
5 Maximal and genuine hidden nonlocality 36 5.1 From projective measurements to POVMs ...... 36 5.2 Genuine hidden nonlocality ...... 38 5.3 Maximal hidden nonlocality ...... 39 5.4 Maximal and genuine hidden nonlocality ...... 42 5.5 Generalization of rGHNL and of the POVM protocol ...... 44 5.5.1 The protocol for a qudit ...... 44 a,b,g,d 5.5.2 rGHNL ...... 44 6 Open questions 46 6.1 Sequential filtering ...... 46 6.2 POVM versus projectors ...... 48 6.3 Genuine locality ...... 48
7 Conclusion 49
8 Acknowledgments 50
Bibliography 51 Chapter 1
Introduction
For more than a century Quantum Mechanic has been studied and tested every day by the physicists. Thanks to the progress of the science and in particular those of the experimental apparatus, the precision with which one is able to verify the theories has become better everyday. So, for more than a century Quantum Mechanic has been put to the proof more and more rigor- ously, and to this day has never been faulted. The precision of the quantum theory is spectacular and it therefore seems to describe perfectly the observed phenomena; in all cases better than any other theory.
Yet, however good are its predictions, Quantum Mechanic continues to defy the human intuition. Indeed, three properties of the theory are very difficult to accept: intrinsic randomness, collapse of the wavefunction and nonlocality. While the two first ones are fundamental properties of the theory and have given rise to fruitful discussions about independence of Nature and about determinism, the third one was discovered later and was considered first as an anomaly of the theory. Nowadays, we have tested and accepted nonlocality as a characteristic of Nature. It means that two systems can be well-separated in space (and in particular space-like separated) but act as if they were one single system. We strongly think that no realistic theories are able to predict the future.
Knowing this, it is not astonishing that nonlocality has been vastly studied in the past thirty years. What seems less normal is the actual theoretical disorder about nonlocality. Multipartite scenarios, Bell’s inequalities, link between entanglement and local models: despite a lot of nice results and the work of brilliant people, these tasks are very partially understood, and we still lack an overview, a general formalism, an analytical method to solve the problems we have.
Genuine hidden nonlocality is one of these big questions: can a fully local state becomes nonlocal after local operations? Even if some partial results are known (for projective measurements), no definitive answer has been yet given. In this work we will prove that the answer to the above question is yes.
5 Chapter 2
Concepts and motivations
The aim of this first chapter is to introduce all the concepts that will be used later in this work and to motivate them physically.
2.1 Nonlocality and Bell’s game
Since quantum mechanics gives probability to events, i.e. is not deterministic, we will focus ourselves on probability theory, define nonlocality and charac- terize the quantum probability distributions. Finally we will explain what we call a Bell’s game.
2.1.1 The scenario The standard situation we will focus on is the one with two parties, sharing some resources. Let’s call Alice the "left" one and Bob the "right" one for convenience.
Figure 1: the general situation we consider
Now Alice and Bob can perform a measurement on their respective re- sources, and obtain a result. We do not care about the exact mechanism of what they do, but our assumptions are the following ones: they can freely choose between a finite set of possible measurements and the set of the results they can obtain after the measurement is finite as well.
If Alice has NA possible measurements we symbolically define MA = A , A , ..., A as the set of measurements and we call the integers num- { 1 2 NA } bers that label the measurements the inputs of Alice. We define RA = RA , RA , ..., RA as the set of the n results she can obtain and we call the { 1 2 na } A integers numbers that label the results the outputs. Similarly, the Bk represent
6 CHAPTER 2. CONCEPTS AND MOTIVATIONS 7
Bob’s measurements and the RBk his results (with set of size NB and nB, respectively).
Note that it is a very general way to study two physical systems; the assumption about the finite number of inputs and outputs is used just for convenience (one can extend to countable and uncountable infinite number of possibility, in particular the quantum formalism allow such case) and we did not mention anything about the physical system nor what the measurement is.
Since we do not focus on what really happens when a measurement is applied on a part, we can represent our scenario in the following way: both parties have a box, with a finite number of buttons on the top which can be pushed individually and a finite number of lights on the bottom which can light individually. What Alice and Bob can do is push one of the buttons -which therefore represent the inputs- and note which light is lit -i.e. which output they obtain. This approach is often called the device-independent approach.
Figure 2: Each party has a box with possible inputs and outputs
All the information about this bipartite system is contained in the prob- ability of obtaining a pair of outputs given a pair of measurements. In other terms we want to know
p(a, b A , B ) (2.1) | x y where a 1, 2, ..., N , b 1, 2, ..., N That is (2.1) is the so-called joint 2{ A} 2{ B} probability that Alice obtains the result RAa given the measurement Ax and Bob obtains the result RBb given the measurement By.
2.1.2 Classification of the probability distribution From here we will rewrite (2.1) in the following form: p(a, b A, B), for con- | venience. Please keep in mind that A, respectively B, can be any measurement chosen in the set MA, respectively MB. Definition 1 (Determinism). A joint probability distribution p(a, b A, B) is de- | terministic if
p(a, b A, B)=d d (2.2) | a,a0 b,b0 CHAPTER 2. CONCEPTS AND MOTIVATIONS 8
where a0 (b0) is the result that Alice (Bob) obtains after measurement A (B), and di,j = 1 if i = j, 0 otherwise. That is, a measurement determines precisely its result. Obviously this will never be the case with a quantum system.
Definition 2 (Independence). A joint probability distribution p(a, b A, B) is inde- | pendent if it can be written as
p(a, b A, B)=p(a A)p(b B) (2.3) | | | i.e. if the probability distribution factorizes.
This corresponds to the trivial case where Alice and Bob are completely uncorrelated; there is no particular interest to study them together rather than separately in this situation.
Definition 3 (Locality). A joint probability distribution p(a, b A, B) is local if it | can be written as
p(a, b A, B)= p(a A, l)p(b B, l)r(l)dl (2.4) | ZL | | where L is a set, l L and r : L [0, 1] is a probability density. 2 ! The interpretation of this decomposition of the probability distribution is that Alice and Bob share some local resources, described by the variable l. The variable is said to be local because we assume that physically it is created somewhere (according to the probability distribution r(l)) and sent to Alice and Bob. The origin of the randomness on this variable can be quantum or classical but the point is that the correlations of Alice and Bob are completely explained by this shared variable and what they do locally.
Definition 4 (Nonlocality). A joint probability distribution p(a, b A, B) is nonlocal | if it cannot be written as (2.4).
Definition 5 (Non-signalling). A joint probability distribution p(a, b A, B) is | nonsignalling if
 p(a, b A, B)= p(a, b A, B0) (2.5) b | b |
 p(a, b A, B)= p(a, b A, B0) (2.6) a | a |
A, A , B, B 8 0 0 It means that the measurement of Bob does not influence the result of Alice and vice versa. We notice that a local probability distribution is always non-signalling. CHAPTER 2. CONCEPTS AND MOTIVATIONS 9
Definition 6 (Marginals). We call p(a A, B)=Â p(a, b A, B) the marginals of | b | Alice. Bob’s marginals are defined similarly. Note that if a probability distribution is non-signalling we have p(a A, B)= | Â p(a, b A, B)=Â p(a, b A, B )=p(a A, B ), B, B . Therefore p(a AB)= b | b | 0 | 0 8 0 | p(a A) in the non-signalling case. | To summarise, non-signalling is the weakest condition we can put on a joint probability distribution, it strictly includes the local probability distributions, which strictly includes the independent ones, which strictly includes the deterministic. The two relevant sets for this essay will be the local and non-signalling ones and the nonlocal and non-signalling ones.
2.1.3 Bell’s game From now on, we will add the condition that Alice and Bob are space-like separated during the time between the choice of their measurements and their appliance. So, according to relativity, without faster-than-light communication what one party does locally cannot give any information to the other one. Accordingly we are in the non-signalling scenario, since if Alice’s marginals depended on the measurement that Bob would choose, it would be possible, in principal, to extract locally some informations on Bob’s choice whatever the distance between the two parties is, and that would violates the relativity. However, there still is an interesting property that the probability distribution may have: the nonlocality. As we defined above, a probability distribution is nonlocal if it cannot be written in the form (2.4), means it cannot be deduced from a common variable that Alice and Bob have, and of what they do locally.
Now an important question is the following one: if we have a certain probability distribution in hand, how can we check whether it is nonlocal or not? To show that it is local, one can give explicitly the functions p(a A, l), | p(b B, l) and r(l) which satisfy the condition (2.4). But if it is not, one has to | show that for all p(a A, l), p(b B, l) and r(l) this condition is never fulfilled, | | that is, there exists no such decomposition of the probability distribution. This seems to be a very tedious task. Fortunately, we have a powerful tool: the Bell’s inequalities. Definition 7 (Bell’s inequality). Let M = A , A , ..., A and M = B , B , ..., B A { 1 2 NA } B { 1 2 NB } be two sets of measurements with nA and nB their number of outputs, respectively. Let p(a, b A, B) be a probability distribution define for all A M and B M . | 2 A 2 B Let f be a function on the following form:
f (MA, MB, p)= Â aa,b,x,y p(a, b Ax, By) (2.7) a,b,x,y |
where the aa,b,x,y are some real numbers.
We call a Bell’s inequality all pair of f , L with f on the form (2.7) and L a real { } number such that the condition f (MA, MB, p) > L is a sufficient condition for the nonlocality of the probability distribution p. CHAPTER 2. CONCEPTS AND MOTIVATIONS 10
The first example of such an inequality was discovered by Clauser, Horne, Shimony and Holt [1]. Let’s consider a set of two possible measurements on Alice’s side M = MA , MA and similarly for Bob M = MB , MB . A { 1 2} B { 1 2} The so-called CHSH inequality is define in the following way:
CHSH(MA, MB, p)= p(00 MA , MB )+p(11 MA , MB ) p(01 MA , MB ) p(10 MA , MB ) | 1 1 | 1 1 | 1 1 | 1 1 + p(00 MA , MB )+p(11 MA , MB ) p(01 MA , MB ) p(10 MA , MB ) | 2 1 | 2 1 | 2 1 | 2 1 + p(00 MA , MB )+p(11 MA , MB ) p(01 MA , MB ) p(10 MA , MB ) | 1 2 | 1 2 | 1 2 | 1 2 p(00 MA , MB ) p(11 MA , MB )+p(01 MA , MB )+p(10 MA , MB ) | 2 2 | 2 2 | 2 2 | 2 2 (2.8)
Defining the correlator of MAi and MAj as
MA MB = p(i = j MA , MB ) p(i = j MA , MB ) (2.9) i j | i j 6 | i j we can rewriting⌦ CHSH↵ in a more convenient form:
CHSH(M , M , p)= MA MB + MA MB + MA MB MA MB A B h 1 1i h 1 2i h 2 1i h 2 2i (2.10)
For any local probability distribution p we have 2 CHSH(M , M , p) A B 2. With this tool, answering the question of nonlocality of a probability distribution p is quite simple: one has to find the existence of two sets MA and M such that CHSH(M , M , p) > 2. That is, optimise the CHSH B | A B | function over all possible MA, MB. At this point, we have everything we need to give the definition of a Bell’s game. The goal of this kind of game is to check concretely the violation of a Bells inequality. If one has the analytical expression of a probability distribution, the problem is entirely mathematical and can be solved, at least in principal. Consequently the aim of the game is to obtain the probability distribution from the actual situation.
The game will imply a third party that we will call the "referee", whose role is to verify that Alice and Bob do not "cheat". The game is the following: Alice and Bob share some resources. In the first round, each of them receives a measurement from the referee, from their respective measurements set, and they have to give their results to the referee. The referee fixes a time during which they have to give their results, and he ensures that they are space-like separated during the interval that separates the receipt of the measurement and the answer they give. The second round is played according to same rules with the possibility of others measurement. At each round, the referee collect Alice’s and Bob’s outputs. After a certain number of rounds, the game is stopped by the referee and he can estimate the probability distribution for the measurements he gave. CHAPTER 2. CONCEPTS AND MOTIVATIONS 11
Since the number of rounds is finite, this will not be the exact probabilities, but one can have arbitrarily close values by increasing the number of rounds. Then the referee can compute the value of a suitable Bell’s inequality (note that the number of measurement involves in a Bell’s inequality is fixed, but for each number of measurements there exists in principal a Bell’s inequality) and determine if the probability distribution is nonlocal or not. Why do we need a external person to calculate these quantities? If Alice and Bob simply choose some measurements that they then apply to their system to evaluate the probability distribution, where is the problem? The answer is that we did not make any assumption on Alice and Bob and we do not trust them. If we did, it would be indeed true that the above game is useless, Alice and Bob would be able to play without any referee and to obtain the right answer. But it could be the case that they fool us, willingly or not. Although we call them with human names, in general they can represent any entity. And for instance, if these entity are causally connected, they can tune their outputs to optimise the value of a Bell’s inequality. Even if they are space-like separated during the whole game, they can choose in advance the inputs that they will use at each round and the related outputs -which is totally equivalent to choose completely the probability distribution. In other terms, we have to require the no-knowledge on the other part’s inputs, and that’s what this game does.
2.2 Entanglement
Bipartite quantum states are introduced in this section as well as the notion of entanglement and its link with nonlocality.
2.2.1 Quantum formalism Let d be an n-dimensional Hilbert space on C. We will define the quantum H quantities we need to treat according to what we have already defined above ; we present everything for the bipartite case -since it is the one in which we are interested- but one can trivially extend all these definitions to the n-partite case.
Definition 8 (Quantum state). A quantum state of dimension d is a positive semi-definite operator r that acts on d and satisfies tr(r)=1. H This is the very general definition of a quantum state, also called a qudit (the "d" holds for the dimension). A bipartite state is a particular case of this formalism:
Definition 9 (Bipartite quantum state). A bipartite quantum state is a positive semi-definite operator r that acts on d1 d2 and satisfies tr(r)=1. H ⌦H This is the formalism we need to describe a quantum resource shared by Alice and Bob, corresponding to the fact that they each receive one qudit. Now we need to define precisely what is a measurement on a quantum state. CHAPTER 2. CONCEPTS AND MOTIVATIONS 12
Definition 10 (Quantum measurement). A measurement on a quantum state r is a set of N linear operators that act on d,M= M , M , ..., M which satisfy H { 1 2 N}
† Â Mk Mk = 1 (2.11) k
r M† After the measurement on the state r the final state is Mk k with probability tr( r M†) Mk k tr( rM†). Mk k One can check that for all r that are quantum states the sum over all the probabilities is indeed 1, due to condition 2.11. From a measurement M one can obtain the related POVM defined by { k} the set of M M† . It is quite convenient to work with POVMs, rather than { k k } measurement:
Definition 11 (POVM). A POVM (Positive Operator Valued Measurement) on a bipartite state r is a set of N linear positive semi-definite operators that acts on d, H E = E , E , ..., E which satisfy { 1 2 N}
 Ek = 1 (2.12) k
After the measurement on the state r the output is k with probability tr(Ekr). Since from the POVM we cannot extract the final states but only the output, we will use that formalism when we do not care about the quantum state after a measurement but only about its statistics. Finally, we will define what is a composite measurement :
Definition 12 (Composite measurement). Let r be a bipartite quantum state. If Alice performs a measurement A and Bob a measurement B on that state the composite measurement M which describes the global measurement on the state is A B. ⌦ Explicitly the measurement is the set M = A B , A B , ..., A B , A { 1 ⌦ 1 1 ⌦ 2 1 ⌦ N 2 ⌦ B , A B , ..., A B , ..., A B , A B , ..., A B 1 2 ⌦ 2 2 ⌦ N N ⌦ 1 N ⌦ 2 N ⌦ N} 2.2.2 Characterization of quantum states At this point we have everything we need to express the probability distri- bution that comes from a bipartite quantum state shared by Alice and Bob. Indeed, if Alice locally performs a POVM A and Bob a POVM B the composite POVM A B describes the global measurement applied on the system and ⌦ according to definition 11 we have
p(a, b A, B)=tr(A B r) (2.13) | a ⌦ b First we remark that this probability distribution is always non-signalling, by using the definition 6 of the marginals: CHAPTER 2. CONCEPTS AND MOTIVATIONS 13
p(a A, B)= p(a, b A, B) | b | =  tr(Aa Bbr) b ⌦ = tr( Aa Bbr) b ⌦ = tr(Aa  Bbr) ⌦ b = tr(A 1r) a ⌦ = p(a A) | Which is the definition (5) of non-signalling. Therefore quantum mechanics is actually non-signalling. For a quantum state an interesting feature is the entanglement. First we need the notion of separability:
Definition 13 (Separability). A bipartite quantum state r of dimension d d is ⇥ separable if it can be written on this form:
k k r = Â pk rA rB (2.14) k ⌦
k k where the rA and rB are quantum states of dimension d and the pk are positive and real numbers which sum to one.
Why do we call such states separable? Let’s assume that we have a k k separable state, which means we can find some rA, rB and pk that decompose our state in the form (2.14). The corresponding probability distribution will be
p(a, b A, B)=tr(A B r) | a ⌦ b k k = tr(Aa Bb  pkrA rB) ⌦ k ⌦ k k =  pk tr(Aa BbrA rB) k ⌦ ⌦ k k =  pk tr(AarA) tr(BbrB) k k k =  pk pA(a A)pB(b B) k | | We obtain a convex combination of separated probabilities, in the sense that they factorise. Which corresponds to classical correlations with local probability distribution pk (a A) and pk (b B) for Alice and Bob. On the A | B | contrary, an entangled state cannot be describe in that way.
Definition 14 (Entanglement). A bipartite quantum state r of dimension d d is ⇥ entangled if it is not separable. CHAPTER 2. CONCEPTS AND MOTIVATIONS 14
We have said that all the information about a physical state is contained on the probability distribution. Due to the degree of freedom involved each complete collection of quantum probability distributions corresponds to a unique quantum state, i.e. there is no ambiguity or redundancy in the quantum description. Knowing this we will define quantum nonlocality in a natural way:
Definition 15 (Quantum nonlocality). A bipartite quantum state r is local if its probability distribution is local, nonlocal otherwise.
As we already noticed above, the separable states have probability distribu- tions which can be written as a convex combination of factorised probabilities, thus they are local since we directly have the decomposition (2.4). All separ- able states are local. But is the reciprocal true: are all entangled states nonlocal? Intuitively we can think that it is actually the case since we said that quantum states and probability distributions are in one-to-one correspondence and since the definition of separability seems to be the equivalence to locality for quantum states.
However the answer is negative. Even if it is true for the particular class of pure states [2], there exist some quantum states that are entangled but local [3]. Therefore the situation is a bit more complex for quantum states than for probability distributions: while these are classified in two classes (local or nonlocal), those are classified in three classes (local, entangled or nonlocal).
This point is a crucial point of quantum theory and nonlocality. First we can ask ourselves if it is a "bug" of the formalism: are the definitions good? There is no simple answer to this question and the rest of this essay will add some new information to discuss this question.
2.3 SLOCC and hidden nonlocality
At this point entanglement is only a mathematical property of a quantum state, corresponding to the fact that there is no decomposition like (2.14), since the nonlocal-local classification refers to something physical. The nonlocality is the relevant concept so what has an entangled but local state that a separable state has not? One possible answer is: hidden nonlocality. Indeed one can "reveal" the nonlocality of some entangled and local states.
For instance in [4] Palazuelos show that one can obtain some violation of a Bell’s inequality by tensorizing a local state with itself. That is, many copies of local states can lead to a nonlocal one (in higher dimension). This phenomenon is called supractivation and answer partially the above question. Moreover, in [5] Masanes, Liang, and Doherty show that for each entangled state s, there exists a state r not violating CHSH, such that r s violates ⌦ CHSH. Which proves that entanglement "contains" necessarly some nonlocal- CHAPTER 2. CONCEPTS AND MOTIVATIONS 15 ity but we sometimes need non-trivial protocol to reveal it.
These two cases both use the idea of tensorizing a local state with some- thing to reveal nonlocality. Contrary, we will use some operations which preserve the dimension of the quantum state. In this section we introduce the concept of SLOCC (Stochastic Local Operation with Classical Communica- tion). First, we present a less powerful operation than SLOCC: the local linear operation.
Definition 16 (Local linear operation). A local linear operation is an operation that transforms a bipartite quantum state r in
† † r0 = Â Ak BlrAk Bl (2.15) k,l ⌦ ⌦
d † with the Ak and the Bl some linear operators acting on that respect Âk Ak Ak = 1 † 1 H and Âl Bl Bl = . One can check that it is the most general linear and separable operation over the constraint of preserving the trace and the positivity, which means that the final state is still a bipartite quantum state if the initial state was one. Barret proves that this operation cannot map a local state into a nonlocal one [6]. Hence, we need a more powerful kind of transformation to reveal nonlocality. Note that the transformations on which we focus are local, since with nonlocal linear operation, one can transform a quantum state in any other possible quantum states. But in that case, one needs nonlocal resources and there is nothing astonishing to obtain nonlocality from nonlocality. In fact the trick is to use a sequence of measurements. After the first (local) measurement, the quantum state has changed, so can we use this changing to map a local state into a nonlocal one? This is the idea of local filtering:
Definition 17 (Local filtering). A local filtering is a four-outcomes measurement that comes from the local two-outcomes measurements A = F , F¯ for Alice and { A A} B = F , F¯ for Bob. { B B} Explicitly it corresponds to the global measurement
F = F F , F F¯ , F¯ F , F¯ F¯ (2.16) { A ⌦ B A ⌦ B A ⌦ B A ⌦ B} Note that we need F F† 1 (this condition means that every eigenvalues A A of F F† are smaller than one) since F F† + F¯ F¯ † = 1 and F F† 1 for the A A A A A A B B same reason (the general solution of this matrix equation is F¯ = 1 F F† U A A A where U is a unitary matrix). q Thus after a local filtering we have four possibles state. Now we add the possibility of classical communication for Alice and Bob, that is, they have a channel with which they can send classical bits. This allows them CHAPTER 2. CONCEPTS AND MOTIVATIONS 16 to post-select the state they want: they will keep their quantum state only F F r F† F† when they obtain A⌦ B A⌦ B after the measurement. This happens with tr(F F r F† F†) A⌦ B A⌦ B probability tr(F F r F† F†). And it is the definition of SLOCC. A ⌦ B A ⌦ B Definition 18 (SLOCC). A stochastic local operation with classical communication (SLOCC) transforms a bipartite quantum state r in
† † FA FB r FA FB r0 = ⌦ ⌦ (2.17) tr(F F r F† F†) A ⌦ B A ⌦ B where F and F are linear operators that acts on d. A B H By its structure, this operation preserve the positivity. The difference with the local linear operation (definition 16) is that it does not preserve the trace, but the trace of the final state is equal to one, because we "renormalize". It is really a post-selection: Alice and Bob keep their state only when the initial state "passes" the filters F F . A ⌦ B
Then if the final r0 is nonlocal, while the initial one was local (over a cer- tain class of measurements), we will say that the initial state contains hidden nonlocality.
Several interpretations of hidden nonlocality are possible. One can high- light that well-chosen local filtering allows preparing nonlocal states from local ones, i.e. a selection on many copies of local states can lead to states that all violate a Bell’s inequality. If Alice and Bob can communicate, they can win a Bell’s game with a local state that has hidden nonlocality: when the filter does not pass they saturate the local bound, and when it does, they reach a larger value. Even if they cannot communicate local filtering increases their probability of winning the game. But as demonstrated in [7], hidden nonlocality mainly shows the failure of the local model for sequence of measurements. Are these kinds of states really local, if a local model predicts everything correctly for one measurement, but not for more than one?
2.4 Notations
Let d be a d-dimensional Hilbert space. We will note y a general vector H | i in this space and y his dual. We define 0 , 1 , 2 , ..., d the canonical h | {| i | i | i | i} basis. Hence a general vector y is defined by its decomposition in this basis: | i y = Âd c k = Âd k y k . | i i=k k| i i=kh | i| i We will sometimes use the pure state notation:
Definition 19 (Pure states). A quantum state r that acts in d is pure if it is a H rank-one projector, i.e. r = y y . | ih | CHAPTER 2. CONCEPTS AND MOTIVATIONS 17
When a state is pure we will then denote it by y instead of y y . | i | ih | Finally we define the three Pauli matrices:
01 0 i 10 s = s = s = x 10 y i 0 z 0 1 ✓ ◆ ✓ ◆ ✓ ◆ And ~s =(sx, sy, sz)
2.5 Classes of measurement
We will distinguish two classes of measurement: projective measurements and POVMs. The second one has already been defined (definition11) and is just a rewriting of the general measurement formalism. The first one is a smaller class:
Definition 20 (Projective measurements). A projective measurement (also called Von Neumann measurement or ideal measurement) on a quantum state of dimension d is a measurement P = P , P , .., P where the elements are orthogonal rank-one { 1 2 d} projectors. That is
PiPj = di,jPi (2.18)
i, j = 1...d 8 This class of measurement is in a sense the canonical one since every posit- ive matrices (i.e. POVM elements) can be decomposed in a linear combination of projectors. Moreover it has a natural quantum mechanical interpretation: each element projects on a pure state, P = y y , and all together they k | kih k| span entirely d. After the projective measurement the final state is the pure H state y with probability tr( y y r)= y r y . In particular if the initial | ki | kih k| h k| | ki state is f (i.e. if it is pure) this probability becomes y f 2, that is the | i |h k| i| norm squared of the Hilbertian scalar product.
In the 2-dimensional case a rank-one projector can be written as:
1 +~v ~s P = · (2.19) 2 where ~v is a vector on the Bloch sphere (unit sphere where the three direc- tions are the "spin" directions: sx, sy, sZ). And therefore a rank-one projector in dimension two is uniquely defined by a vector on the Bloch sphere.
We will speak about "hidden nonlocality" when a state is local for projective measurements but nonlocal after filtering, while we will reserve the term "genuine hidden nonlocality" for a state fully local (i.e. local for all POVMs) and nonlocal after filtering. CHAPTER 2. CONCEPTS AND MOTIVATIONS 18
2.5.1 Quantum observable It is sometimes more convenient to work with observables rather than meas- urements. A quantum observable is a hermitian matrix constructed from a POVM E = E , E , ..., E as follows : { 1 2 N}
O = Â akEk (2.20) k
where the real number ak represents the value we assign to the POVM element Ek. In the projective case the ak are the eigenvalues of the observable since the projectors are orthogonal. If the quantum state is a qubit it is common to choose +1 and 1 as eigenvalues and in that case we can write the observables in terms of the Pauli matrices:
O = ~a ~s (2.21) · where ~a is a vector on the two-dimensional sphere. One can check that this is a hermitian matrix with eigenvalues +1 and 1. Chapter 3
Local states
In this chapter, we will present some entangled states that are local over certain classes of measurement. We will then prove that some of them exhibit hidden nonlocality.
3.1 Local models
In order to show that a bipartite quantum state is local, we need to prove the existence of the decomposition (2.4) for its probability distribution. A sufficient condition is therefore to exhibit a set L, a probability density r(l) on this set and the functions p(a A, l) and p(b B, l), that we will call the | | response functions. Since the locality means that the probability distribution has only classical correlations, having this three functions allows us to simulate a quantum state, that is having the same probability distribution as the quantum predictions with only classical and local resources. The protocol that explains what Alice and Bob have to do to simulate a quantum state is called a local hidden variable model(LHVM).
Definition 21 (Local hidden variable model). Alice and Bob share a random variable l L (sometimes called the "hidden variable" for historical reasons) with 2 probability density r(l). Then they receive inputs A and B and output fA(A, l) and fB(B, l), respectively.
Figure 3: the scenario of a local model
In the definition above, we have explained the protocol in terms of "outputs functions" rather than response functions but one can always go from one to
19 CHAPTER 3. LOCAL STATES 20 another: the output function is obtained by simulating locally the probabilities given by the response function (on the other side the other way round can be hard).
Although one is sufficient in the local models we will present, we often give both response and output functions because we think it leads to a gain of intuition on the model. More precisely said, the output function is more convenient to understand the model while the response function appears in the proof that the model works (condition (2.4) fulfilled).
3.1.1 Simulate POVMs As pointed out by Barrett[6] to simulate POVMs on a quantum state we just need to focus on the ones with all elements proportional to rank-one projectors. Indeed each element can be decomposed as a summation of rank- one projectors, and hence using this decomposition allow us to simulate in reality something more subtle. Generally we therefore have POVMs on the form P = x P , x P , ..., x P where the real numbers x respect: 0 x 1 { 1 1 2 2 N N} i i and Âi xi = d, and the Pi are rank-one projectors. In the qubit case one can therefore represent such a POVM as three vectors on the Bloch sphere (that respresent the projectors) with norms smaller than one (that represent the coefficients in front of the projectors).
Figure 4: a three-outcomes POVM represented by three vectors inside the Bloch sphere that satisfy the normalization conditions
3.2 Werner state
The Werner state is a qudit-qudit state defined originally by Werner in [3] as follows:
1 W = ((d F)1 +(dF 1)V) (3.1) d3 d where 1 F 1 and V is the "flip" operator: V( y f )= f y . | i ⌦ | i | i ⌦ | i This state is the most general d-dimensional U U invariant state (where U ⌦ CHAPTER 3. LOCAL STATES 21 is any unitary matrix).
Werner showed that this state is entangled if F < 0. However he showed that this state can be simulated for projective measurements for F = 1 + d+1 d2 [3]
3.2.1 Werner’s model Alice and Bob share a complex vector w W, where W is the unit sphere, 2 that is W = w Cd w = 1 . Its probability density is uniform. They { 2 |k k } receive projective measurements A = P , P , ..., P and B = Q , Q , ..., Q , { 1 2 d} { 1 2 d} respectively. Alice’s response function is:
p (k A, w)= w P w (3.2) A | h | k| i Bob’s response function is:
1 if w Qk w < w Ql w for all l = k pB(k B, w)= h | | i h | | i 6 (3.3) | (0 if w Q w > w Q w for some l = k h | k| i h | l| i 6 The interpretation of this response function is that Bob outputs determin- istically the k that satisfies w Q w w Q w k. 0 h | k0 | i h | k| i 8 That is Bob chooses the k that minimizes w Q w . h | k| i Therefore, one can show that
pA(k A, w)pB(k B, w)r(w)dw = Tr(Pk QkW) (3.4) ZW | | ⌦ if F = 1 + d+1 . d2
To link the Werner state with the state rW we will present in 4.2 we will rewrite the Werner state using the fact that the flip operator can be expressed as V = 1 1 2 Â S S . With that an other way to write down a ⌦ i