Arxiv:2012.01441V2 [Quant-Ph] 10 Jun 2021

A no-go theorem on the nature of the gravitational ﬁeld beyond quantum theory

Thomas D. Galley§,1, ∗ Flaminia Giacomini§,1, † and John H. Selby§2, ‡ 1Perimeter Institute for Theoretical Physics, 31 Caroline St. N, Waterloo, Ontario, N2L 2Y5, Canada 2ICTQT, University of Gdańsk,Wita Stwosza 63, 80-308 Gdańsk,Poland Recently, table-top experiments involving massive quantum systems have been proposed to test the interface of quantum theory and gravity. In particular, the crucial point of the debate is whether it is possible to conclude anything on the quantum nature of the gravitational field, provided that two quantum systems become entangled due to solely the gravitational interaction. Typically, this question has been addressed by assuming an underlying physical theory to describe the gravitational interaction, but no systematic approach to characterise the set of possible gravitational theories which are compatible with the observation of entanglement has been proposed. Here, we introduce the framework of Generalised Probabilistic Theories (GPTs) to the study of the nature of the gravitational field. This framework has the advantage that it only relies on the set of operationally accessible states, transformations, and measurements, without presupposing an underlying theory. Hence, it provides a framework to systematically study all theories compatible with the detection of entanglement generated via the gravitational interaction between two non-classical systems. Assuming that such entanglement is observed we prove a no-go theorem stating that the following statements are incompatible: i) the two non-classical systems are independent subsystems, ii) the gravitational field is a physical degree of freedom which mediates the interaction and iii) the gravitational field is classical. Moreover we argue that conditions i) and ii) should be met, and hence that the gravitational field is non-classical. Non-classicality does not imply that the gravitational field is quantum, and to illustrate this we provide examples of non-classical and non-quantum theories which are logically consistent with the other conditions.

I. INTRODUCTION

The fundamental description of spacetime and, in particular, the nature of the gravitational field are among the deepest puzzles that fundamental physics is facing nowadays. The twofold role of the gravitational field —as a dynamical quantity and as the metric of the spacetime— precludes us the use of standard quantisation techniques. In general, no agreement has been reached on whether gravity should admit a quantum description at all: some arguments deem it impossible to find a quantum description of the gravitational field [1, 2], others call for a radically new perspective on its quantisation [3], while the majority of results are in favour of its quantum description [4–9]. arXiv:2012.01441v2 [quant-ph] 10 Jun 2021 The lack of an explanation on the nature of the gravitational field is not a matter of abstract speculation, but poses important challenges for experiments which may soon be within the reach of technology. For instance, there are several open questions which challenge the conclusions which could be drawn from experiments at low-energies, in which massive bodies prepared in a quantum state interact solely via the gravitational field. The nature of the gravitational field associated to a quantum source has been first discussed by Feynman [10, 11]. Since then, many experiments involving massive bodies in superposition have been proposed [12–26]. Recently, the debate around these topics has been revived by two papers [27, 28], in which two masses A and B become entangled solely by the gravitational interaction,

∗ [email protected] † [email protected] ‡ [email protected] § All authors contributed equally to this work. 2 and in particular the Newtonian potential. The critical point of the debate originating from these results [29, 30] is whether any quantum feature of the gravitational field can be deduced from this experiment, provided that entanglement between the masses is measured. One of the crucial points of the debate is that experiments only involving Newtonian gravity only test the gauge component of the gravitational field, which is not quantised in the standard formalism [30]. This criticism is valid, however there is an important caveat: it is possible to devise a thought experiment where, with analogous arguments to those employed in the electromagnetic version of the same experiment [31, 32], one runs into faster-than-light signalling unless some quantum-field elements of the gravitational field are included in the description [21]. This hints at the fact that the Newtonian field itself has some information content, and thus the discussion cannot be simply dismissed as inconclusive on the nature of the gravitational field [22]. In part, the disagreement around these topics arises from the fact that different definitions of what it means for the gravitational field to be quantum have been used in the literature. Some authors [10, 11, 23] argue that the existence of a superposition of spacetimes is a sufficient feature to deem the gravitational field as quantum; others invoke the result that Local Operations and Classical Communication (LOCC) do not generate entanglement [33] as a criterion to establish the quantum nature of the gravitational field [15, 16, 20, 27, 28]. Another possibility is that gravity can be considered quantum only if quantised radiation (i.e., gravitons) is emitted [30]. Finally, there has been an attempt to understand whether gravity can be compatible with some so-called post- quantum theories, i.e., theories that are neither classical nor quantum [34, 35], within the framework of constructor theory [36] (see Appendix F for a comparison with the approach presented here). All these approaches, with the exception of the one in Ref. [34, 35], make statements about the nature of the gravitational field by considering the field by itself, without specifically referring to the set of all possible measurements. However, in order to have a full characterisation of a physical theory, one needs to consider, together with the set of states (in this case, of the two systems and the gravitational field) and their transformations, also the set of all possible measurements. If the latter is neglected, it might happen that a specific state could seem quantum, but that its quantum behaviour never manifests itself in any possible experiment. An additional difficulty in experiments in which gravity is sourced by quantum systems is that the full set of measurements is not accessible to us: we cannot directly measure the (possibly) quantised gravitational radiation to establish whether gravity is quantum. Hence, we need to resort to probing the nature of the gravitational field indirectly through the matter systems. The need to verify via indirect measurements whether some degree of freedom is quantum is not special to the gravitational field but happens, for instance, in the Stern-Gerlach experiment: we infer the quantum nature of the spin degree of freedom not by directly probing it, but by measuring the deflection of a quantum particle traversing a magnetic field. The question we here ask is then: “Provided that we observe gravitationally-induced entanglement in the laboratory, which conclusions can we draw on the nature of the gravitational field?” The answer to this question is nontrivial, because, in order to give a fully consistent answer, we cannot assume an underlying theory for the gravitational field. We therefore have to check the consistency of our experimental results in a systematic way, taking into account all logically-possible theories. This task is precisely what Generalised Probabilistic Theories (GPTs) have been useful for in the context of non-relativistic quantum theory. For example, GPTs have been used in the study of thermodynamics [37–40], interference [41–46], decoherence [47–50], computation [51–57], cryptography [58–63], contextuality [64–67], information processing [68–75], and correlations [76– 82] as well as deepening our understanding of the general structure of physical theories [83–98] in a way that does not presuppose the quantum formalism but instead studies these phenomena in a theory-agnostic way. The basic idea of the formalism, is that any theory of physics should be able 3 to make operational, potentially probabilistic, predictions about the outcomes of experiments. This does not mean that a theory cannot, and should not, do more than this by providing an ontological level of description, but, that the operational level of description is a necessary prerequisite of any good physical theory. It is this operational level of description which is captured by the framework of GPTs. One imposes simple mathematical axioms on the structure of the theory, which faithfully capture the conceptual operational underpinnings, and, by doing so, one can derive a rich mathematical structure which any GPT must have. For example, this means that we are no longer free to propose arbitrary modifications of quantum theory, but, having proposed a particular modification, we must then check that it is indeed compatible with the basic mathematical axioms – for example, that probabilistic predictions must be valued in the interval [0, 1]. In this paper, we introduce the theoretical tools of GPTs to determine whether gravity can be classical, quantum, or a different type of theory when gravity-induced entanglement between two massive bodies is observed in the laboratory. We prove a no-go theorem stating that it is impossible to satisfy at the same time the following conditions: i) gravity is a no-signalling theory; ii) gravity is a physical degree of freedom which mediates the gravitational interaction; iii) gravity is classical§. Assuming that such gravitationally induced entanglement is observed, we argue that conditions i) and ii) should be met, and hence that the gravitational field is non-classical. The violation of iii) does not necessarily mean that gravity is quantum. There are, in fact, other possibilities, which are neither classical nor quantum, which we explore. The paper is organised as follows. In Section II we briefly review the framework of GPTs and and put it in the context of the rest of this work. This Section provides the reader who is not familiar with GPTs the preliminaries to understand the no-go theorem. Section III contains our main result: we state the no-go theorem and show the implications of the failure of each condition. Section IV contains a discussion on different notions of classicality, and their ontological status. This last Section is not strictly necessary to understand our result, but it comments on alternatives which are consistent with our no-go theorem, while still allowing gravity to be, in a sense that we specify, classical.

II. THEORY INDEPENDENT FRAMEWORK

In order to explore these questions we need a framework which can describe arbitrarily conceiv- able physical systems together with their interactions, rather than just simple quantum or classical systems. The framework of GPTs is a well established formalism designed for doing just that [70, 99]. In Appendix A we provide an introduction to GPTs, focusing on the particular aspects of the formalism which we use to prove our no-go theorem. See [70, 99, 100], however, for a more in depth introduction to the framework. The basic idea of the formalism is that, ultimately, any physical theory has an operational representation which can be characterised in terms of probabilities which can be, in principle, observed in experiments. This operational representation, in the case of quantum theory, is isomorphic to the standard representation of quantum theory in terms of Completely Positive and Trace-Preserving (CPTP) maps, density matrices and so on. Indeed, another way to view the GPT formalism is that it starts with this representation and then removes all of the structure speciﬁc to quantum theory, whilst maintaining only the structure necessary to ensure that the theory makes sensible probabilistic predictions. For example, in quantum theory a system has an associated set of states described by density matrices which have a rich geometric structure. The GPT formalism identiﬁes the fact that this set of states is

§ In the sense of the standard notion for GPTs which we will formally define later. 4 a convex set as being the essential ingredient needed to ensure that the theory makes sensible predictions, hence, state spaces for a GPT can correspond to arbitrary convex sets. Similarly, in quantum theory, systems are composed by the tensor product which has a precise definition which is characteristic of quantum theory. The GPT formalism identifies, for example, associativity of the tensor product as being the essential feature which must be preserved by an arbitrary GPT. Despite the minimalistic nature of the assumptions underpinning the framework of GPTs, they have surprisingly strong consequences. For example, implying nontrivial bounds on the correlations that can be observed [79] and the computational power of the theories [51]. We will formally define classical and quantum theory as GPTs in Appendix A. However, for the proofs in the following section we only require one particular property of these two theories that we now discuss. Ref. [50] introduced the notion of a leak for a system within some general physical theory, and noted that classical theory was a theory that had maximal leaks whilst quantum theory had only trivial leaks. A corollary of this, is that classical systems can, in principle at least, be perfectly characterised by local non-disturbing measurements, whilst non-disturbing measurements for quantum theory are all trivial, in that no information can be gained about the quantum system in a non-disturbing manner. Ref. [101] demonstrates that this is equivalent to another condition, namely, that in classical theory, the identity transformation is maximally reducible to a coarse graining of effect-state processes, whilst in quantum theory the identity is irreducible, that is, it cannot be thought of as the coarse graining of any other processes. It is this decomposition of the identity process in classical theory which will be vital to proving our main theorem. One of the great upshots to the GPT formalism is that, whilst it is based on very basic and well motivated principles, the mathematical structure which is derived from these is extremely rich. This has two dividends, on the one hand, it allows for extremely general results to be proven based on this mathematical structure, and, on the other hand, it shows that any physical theory which is lacking this structure must necessarily violate one of these basic principles. Indeed, this is the case for many of the modifications of quantum theory which have been considered in the literature. For example, any modification to the measurement postulates of quantum theory does not lead to an operationally consistent physical theory [83]. More generally the GPT framework shows that when modifying certain parts of quantum theory, or when putting forward a new physical theory there are many consistency constraints which must be met in order for the theory to be operationally well defined. This is an important consideration when proposing novel interactions between a gravitational field and quantum systems, since it shows that the interaction used in conjunction with other existing interactions and operational procedures must give a consistent operational procedure in the theory.

III. THE NO-GO THEOREM ON THE NATURE OF THE GRAVITATIONAL FIELD

Consider the following experimental set up, illustrated in Fig. 1. We have two distant physical systems A and B which we assume to be non-classical, and which interact via some mediating system G. We assume that it is possible to prove experimentally that the only way in which they interact is gravitationally. Hence, we take the system G to represent the gravitational field. We prepare the two systems A and B in a separable state, and after some time the experiment detects gravitationally-induced entanglement. We then ask whether this measurement allows us to infer anything on the nature of the gravitational field. The systems A, B, and G are all described within the framework of Generalised Probabilistic Theories (GPTs) and, as such, are called GPT systems. The definition of a GPT system only relies on operational concepts, such as the preparation, transformation, and measurement of the system. 5

A G B

FIG. 1: Illustration of the experimental situation. Two masses A and B are initially prepared in a separable state (where at least one of them, A in the ﬁgure, is in a superposition state in position basis). The masses interact via the gravitational ﬁeld G. After some time, the full state becomes entangled.

We can then give the following deﬁnition of a GPT system.

Definition 1 (GPT system). A GPT system S consists of a set of states, ΩS, effects, ES, and transformations, TS, which can be embedded in a finite dimensional real vector space VS. States are mapped to vectors, effects are mapped to dual vectors and transformations induce linear maps VS → VS.

By states, we mean the set of all operationally equivalent (i.e., not distinguishable experimentally) preparation procedures. By effects, we mean the set of operationally equivalent measurement outcomes. Notice that the definition of the GPT system is agnostic as to whether the system is compatible with a classical or quantum description. If we want to consider only classical systems, further conditions need to be specified, and in our case these are, notably, conditions on the measurements. We here take a strong notion of classicality, which corresponds to the standard notion used in classical physics textbooks.

Example 1 (Finite classical systems). Consider a classical system with a finite configuration space, Λ. The states for the GPT then correspond to probability distributions over Λ, effects to response functions over Λ, and transformations to stochastic maps from Λ to itself. The vector Λ space structure comes by noting that the function space R (that is, the space of functions from Λ to R) is a real vector space where addition and scalar multiplication are defined point-wise.

Any GPT system which is not of the above type is a non-classical system.

Deﬁnition 2 (Classical theory). A theory is said to be classical if every system can be characterised by local non-disturbing measurements.

We can also describe quantum systems as a particular kind of GPT system.

Example 2 (Quantum systems). The real vector space associated to a quantum system S described by a Hilbert space H, is the real vector space of Hermitian operators on H, which we denote by B(H). Density operators on H can then be viewed as particular vectors in B(H), Positive Operator- Valued Measure (POVM) elements can be viewed as particular covectors in B(H)∗, and CPTP — or Completely Positive Trace non-Increasing (CPTNI) — maps can be viewed as linear maps from B(H) to itself. 6

Quantum systems are a specific type of non-classical systems. Quantum theory as an operational theory is defined in terms of systems as well as how they compose. Definition 3 (Quantum theory). Quantum theory has primitive systems which are quantum systems. The composite of two systems S1 and S2 with associated Hilbert spaces H1 and H2 is given by a system S1,2 with associated Hilbert space H1,2 where dim (H1) · dim (H2) = dim (H1,2). The composite of a pair of procedures is given by the tensor product of these. Further examples of non-quantum, non-classical GPT systems are given in App. A. With these tools in mind, we can formulate the following no-go theorem. Theorem III.1. We consider two non-classical systems A and B, initially in a separable state, and some unknown field G. If, after some time t, entanglement between the systems A and B is observed, then the following statements are incompatible: 1. Subsystem independence of A and B; 2. A and B interact locally via the mediator G; 3. G is classical. Proof. See App. B. Discussion of condition 1.— Condition 1 is always taken to be valid in any GPT theory §; subsystem independence states that for any two systems A and B the state of one system is independent of local actions on the other. In quantum theory this manifests as ρA = TrB (ρAB) = † TrB (IA ⊗ UB)ρAB(IA ⊗ UB) for all local unitaries UB on B (and all CPTP maps on B more generally). Subsystem independence is a necessary condition for the existence of subsystems: a subsystem must have degrees of freedom which can be isolated and probed independently of the rest of the universe. When two subsystems are space-like separated subsystem independence is equivalent to no-signalling, which is a purely operational condition of locality. It is therefore somewhat surprising that a theory which does not meet Condition 1 can still have a mathematically well defined notion of separable states and entanglement. In such a theory given two systems A and B one could have an inclusion map I : A × B 7→ C mapping the two systems A and B to their composite system C. However, the absence of subsystem independence would mean that there were no reduction maps (generalisations of the partial trace) RA : C → A and RB : C → B. In such a theory any states of C which are in the image of I would be product states, those in the convex hull of the image would be separable, and any states which lay outside this would be entangled (see Appendix A for more details). There exists other notions of locality, such as Bell non-locality which is an ontological concept and is stronger (see Section IV B for a more detailed discussion). Discussion of condition 2.— In condition 2, by mediator we mean that the gravitational field G is described as a GPT system, on equal footing with A and B. In addition, A and B do not interact directly, but only via G. The violation of condition 2 implies that gravity is not a physical degree of freedom in its own right. In Section III A we will show that if this is the case we can incur a violation of the no-faster-than-light-signalling condition. In the following we argue that the gravitational field (specifically the Newtonian field) should be treated as a physical degree of freedom. This argument, coupled with a view that A and B are independent subsystems leads us to argue that conditions 1 and 2 of the no-go theorem should be met. This leads us to study violations of condition 3. Discussion of condition 3.— The most interesting cases arise when condition 3 is violated, i.e., when system G is non-classical. Notice that in quantum theory a field can be treated as: a)

§ In the Operational Probabilistic Theory (OPT) framework [87] this is an assumption known as causality [102]. 7 a non-dynamical entity, in which case it enters the equations of motion as a potential; or b) a dynamical entity, in which case it is assigned some Hilbert space. Similarly, in the GPT setting if a field is modelled as a system then it is a dynamical entity, whereas if it is not modelled as a system (i.e., it enters as a parameter in some transformation of other systems) then it is modelled non-dynamically. Crucially, for a field the term ‘classical’ is typically used to describe something which has not been quantized; however this can be the case for two different reasons a) the field is not treated dynamically or b) it is dynamical (i.e. a system) but classical. A failure of condition 3 should not be interpreted as a demonstration that the gravitational field is quantum. Of course, we know that entanglement can be generated if G is quantum, but we also show that entanglement can be generated by alternative GPTs that are neither quantum nor classical. This is one of the main novel contributions of this work, and we will study explicit examples of this violation of condition 3.

A. The gravitational ﬁeld as a physical degree of freedom

δx TB A 0 B d L x FIG. 2: Illustration of the protocol. Long before the experiment starts, a massive particle A with mass mA is put in a spatial coherent superposition in such a way that the distance of the position of its centre of mass in the two amplitudes is d. At time t = 0 an observer B, located at a distance L from A, can decide whether to release a probe particle B with mass mB, initially kept in a trap. If B does not release the trap, the state of the full system stays separable. If B releases the trap, and assuming that gravity does not have a physical degree of freedom associated to it, the state of A’s and B’s particles gets entangled in a time TB. A can then recombine her state, ending the experiment at time TA. With an appropriate choice of the parameters L, d, TA,TB, mA, mB, and by performing an interference experiment, A can tell whether B released the trap or not before a light-crossing time (in the ﬁgure, the light cone is illustrated as an orange-shaded area). Hence, if gravity is not treated as a physical degree of freedom, we run into a violation of the no-faster- than-light signaling.

In this Section, we explore the implications of the failure of the condition that system G (the gravitational ﬁeld) is a physical degree of freedom mediating the interaction between systems A and B. At this level, we are not making any assumption on the other two conditions of the no-go theorem. We only assume that entanglement between two particles A and B can be induced by gravitational interaction. 8

In order to explore the implications of the violation of condition 2, we revise a protocol first introduced in Ref. [31] in the electromagnetic case, and then studied in Refs. [21, 22] in the gravitational case. The protocol is illustrated in Fig. 2. We consider two observers, A and B, who are located at distance L from each other. Long before the experiment starts, 1) A prepares a massive particle of mass mA in a spatial coherent superposition state, such that the separation between the centre of mass of A’s particle in the two amplitudes is d L, and 2) B localises a probe particle with mass mB mA in a trap. At time t = 0, B decides whether to release the trap or not. It is a corollary of the results in Refs. [21, 22] that if we do not assign some physical degree of freedom to the gravitational field we run into faster-than-light signalling. This is easily shown with the following argument. If gravity does not possess its own degrees of freedom, this means that there is no physical state associated to its left (L) or right (R) configuration. Firstly, let us consider the situation in which B does not release its trap. The full state of A and B is in this case |Ψi = √1 (|Li + |Ri ) |0i , where |Li and |Ri stand for the left and right 2 A A B A A amplitude of the state of the centre of mass of A’s particle, and |0iB stands for the initial position of B’s particle in the trap. This state is unchanged for the whole duration of the experiment. If B instead decides to release the trap, the state of B’s particle starts to become entangled with the state of A’s particle. Calling TB the time it takes for the state of A and B to become entangled, at T the full state of A and B is |Φi = √1 (|Li |Li + |Ri |Ri ). Similarly to the case when B B 2 A B A B does not release the trap, A can now recombine her superposition, ending her experiment at time TA. It is possible to find a combination of the parameters L, d, TA,TB, mA, mB such that A can recombine her state before a light-crossing time. It is now clear that the outcome of A’s experiment will be very different according to whether B decided to release the trap or not. If B didn’t release the trap, A sees an interference pattern when recombining the state; if B does not release the trap, the state decoheres and A does not see any interference. Hence, A is able to know whether B decided to release the trap or not before a light-crossing time, thus violating the no-faster-than-light signaling principle. In Refs. [21, 22] it was shown that, if gravity is treated as a degree of freedom which can get entangled with the states of A and B, then the violation of faster-than-light signalling does not arise for any value of the parameters. It is relevant to notice that the fact that the Newtonian potential is responsible of the entanglement is no approximation here, and that the violation of the no-faster-than-light signaling principle would be the same if, for instance, we considered a retarded potential. Analogously, no propagating degrees of freedom of gravity are to be considered before A starts recombining her superposition because, when the experiment starts, particle A and the gravitational field are already in a static configuration. It was then argued in Ref. [22] that a mass in a spatial superposition should be considered entangled with its own Newtonian field, and that the Newtonian field can carry (quantum) information.

B. Examples of non-classical systems mediating entanglement between quantum systems

In order to find non-quantum systems which mediate entanglement between quantum systems, we must first have some GPT which includes quantum theory as a subtheory, that is, the GPT must be capable of describing both the quantum systems which we want to entangle, as well as the non-quantum systems which represent the gravitational field. Indeed, there are very few such post-quantum theories, and strong constraints on them were found in Ref. [48]. However, there are three such GPTs (to our knowledge) which all allow for entanglement to be generated via non-quantum mediators, that is Refs. [93], [103] and [104]. Hence, it is clear that whilst showing 9 that gravity can entangle two quantum systems means that it cannot be classical, it does not mean that it is necessarily quantum. It could be one of these hypothetical post-quantum systems. The first of these, Ref. [93], is a GPT which has standard, complex, quantum systems as well as real and quaternionic systems§. To see that these non-complex systems can be used to generate entanglement one can simply note that there is a way to embed a standard qubit within a quaternionic system. Hence, we can use the standard protocol for generating entanglement within quantum theory, but with the mediating qubit embedded within a quaternionic system instead. The second of these, Ref. [103], is a GPT which has a hyperdecoherence mechanism [48] whereby the post-quantum theory decoheres to quantum theory. It is therefore easy to see that these post- quantum systems can be used to generate entanglement, as the definition of hyperdecoherence means that anything that can be achieved in quantum theory can instead be achieved in the post-quantum theory. We elaborate on this example in Appendix E. Finally, in Ref. [104], a GPT is constructed which allows arbitrary no-signalling quantum channels to be realised via interactions mediated by non-quantum systems. An example of a no- signalling quantum channel is a channel which deterministically prepares the maximally entangled state, regardless of what the input state is. To realise this within quantum theory, for example, by discarding the inputs and preparing a maximally entangled state, would require a non-local interaction. However, within the GPT of Ref. [104] this non-local quantum interaction can instead be realised via local interactions with a mediating non-quantum system, hence, these non-quantum systems are capable of generating entanglement between quantum systems. We have shown three example GPTs in which non-quantum systems can mediate entanglement between quantum systems. However, it is clear that not every non-classical system is capable of mediating entanglement. In fact, the possibility of non-trivial interactions between a quantum system and a non-quantum system is severely constrained by operational requirements (such as no-signalling). We leave a full exploration of this topic to future work, but pose two open problems to suggest future directions of study.

Open Problem 1: Find a well motivated class of GPTs such that the only one that can generate entanglement between quantum systems is quantum theory itself.

Open Problem 2: Develop a classiﬁcation of the GPTs which are capable of generating entanglement between quantum systems.

C. Comparison to existing work

Theorem III.1 can be viewed as an extension of the arguments of Refs. [27, 28] to theories beyond quantum theory. Moreover the explicit framing as a no-go theorem shows that the conclusion that gravitationally mediated generation of entanglement of quantum particles entails that the gravitational ﬁeld relies on extra assumptions (conditions 1 and 2). In particular, condition 1 is not explicitly stated in [27, 28]. This assumption is non-trivial since certain analyses, such as constrained systems, would not necessarily make them. Another related work is [35] which is further discussed in Appendix F. In [35] the authors also consider the case of a non-classical gravitational ﬁeld using the framework of constructor theory. This framework has ontological commitments beyond those of the GPT framework used in this work which uses a strictly operational methodology and so avoids any ontological commitments (see Sec. 18 of Ref. [105] for a detailed discussion of this point). Moreover the theorem of Ref. [35]

§ That is, a GPT systems in which pure states are taken to have real or quaternionic coefficients rather than the standard complex coefficients. 10 make an assumption of local tomography (which is called locality in their framework). This is a highly non-trivial assumption, since many GPTs (such as real vector space quantum theory [106], general theories based on Jordan algebras [93], general composite systems with modified Born rules [84], and many others [107, 108]) violate this assumption.

IV. A DISCUSSION ON CLASSICALITY

A. The eﬀective versus fundamental nature of systems

In the present work we derive results concerning the nature of the gravitational field. The nature of a system could be classical, quantum or non-classical more generally. It is important to emphasise what we mean by the nature of a system. Here we are concerned only with the operational features of a system, i.e. its effective behaviour and we do not make claim about the underlying ontology. A system is classical if it generates classical statistics, and similarly a system is quantum if it generates quantum statistics. The statistical description of a system is given by its convex state and effect spaces. These encode only the statistical behaviour of the system and make no assumptions about ontology. As such when we talk about the nature of the gravitational field we mean its effective behaviour. In the following we distinguish two levels of description of a system: the operational level (also known as effective or apparent) and the ontic level (also known as the fundamental or realist level). A system may have an apparent classical behaviour, but have a fundamentally quantum nature. An example of this would be a qubit with a superselection rule such that only σZ eigenstates can be prepared. The state space of such a system would be a classical bit (it has only two perfectly distinguishable operational states and no other extremal states) even it is fundamentally quantum. Alternatively consider a qubit such that all states can be prepared but only the σZ measurement can be carried out. This system is effectively classical also, since all pure ‘ontic’ states with the same σZ expectation value are indistinguishable with the available measurements and thus correspond to the same operational state. The effective state space is just that of a bit. A major question in the foundations of quantum theory is whether an operational quantum system admits of an underlying classical explanation, and which features such classical explanations have. Classical explanations of quantum systems are known as classical ontological models.

B. Classical ontological models

Any quantum system can be modelled using an underlying classical ontological model (also known as a hidden variable model). A classical ontological model of a quantum (or more generally non-classical) system accounts for the effective non-classical behaviour of the system in terms of an underlying (fundamental) classical system. The underlying classical system consists of a measurable space Ω (the ontic space), and an ontological model describes the states of the non-classical system as probability measures on Ω and the measurements of the non-classical system using response functions on Ω. An example of a classical model of a single quantum system is the Holevo-Beltrametti-Bugasjski model [109]. This model accounts for all non-classical behaviours of a quantum system in terms of an underlying classical model. This model has multiple drawbacks, such as being contextual and requiring an infinite dimensional classical system to model a finite dimensional quantum system. By construction it is highly fine-tuned: it only appears quantum because there are many classical degrees of freedom which can only be probed indirectly. Given a quantum system, or indeed, any GPT system, one cannot rule out that it is in fact a highly fine tuned classical system. 11

As such one cannot definitively prove the fundamental quantum nature of a system exhibiting effective quantum behaviour. However there exists various theorems showing that classical models of quantum systems have undesirable features such as contextuality and non-locality. The most famous of the theorems is Bell’s theorem, which shows that any classical model which reproduces the operational statistics of space-like separated entangled quantum systems must be Bell non- local. As such one can rule out classical ontological models of quantum systems by postulating additional requirements that a ‘good’ model should possess. For instance if one requires a classical explanation which is Bell local then one can use Bell’s theorem to rule out classical models of quantum systems. We stress that Bell non-locality is a feature of the ontological model, and that a Bell non-local model of a pair of quantum systems still yields an operational description of these which is non-signalling (i.e., operationally local). The main theorem of this work makes no claim about the ontological nature of the gravitational field, just about its operational nature. Namely our theorem states that if entanglement is observed between two quantum systems A and B, and conditions 1 and 2 are met then the gravitational field behaves non-classically. This does not rule out the existence of some fine-tuned classical ontogical 1 model for the gravitational field, just as the quantum behaviour of spin 2 particles does not rule out the existence of fine-tuned classical explanations.

C. Diﬀering notions of classicality

The notion of classicality of Definitions 1 and 2 states that a theory is classical if every system can be characterised by local non-disturbing measurements. This is the common notion of classicality in GPTs, but there exists other notions of classicality. There might be some systems mediating entanglement between qubits, which are considered non-classical according to our Def- initions 1 and 2, but classical according to a weaker notion of classicality. This point is made in Ref. [29], where the authors argue that the proposals of Refs. [27, 28] only tests classical models of gravitational fields of a specific type (i.e. those which are classical according to Definitions 1 and 2). An example of such a different notion of classicality is used in the Reginatto-Hall model [110, 111], which redefines classicality to allow for conjugate variables. The authors then give an example of a local interaction between a ‘classical’ system and two quantum systems which generates entanglement. However this model is signalling in certain situations [112], thus violating both conditions 1 and 3 of our theorem. One can define alternative notions of classicality in terms of features of classical ontological models for systems. For instance one could define a system to be classical* (where we use the * to differentiate from our pre-existing definition, see Definition. 5 in Appendix. A based on [65] for a formal definition.) if it has a classical ontological model with certain desirable features. An example of such a property is non-contextuality [113]. The property of being non-contextual is strictly weaker than being classical: all classical systems are non-contextual but the converse direction does not hold. We make the following conjecture:

Conjecture 1: the mediating system must be contextual in order to generate entanglement between quantum systems.

V. CONCLUSIONS

In this paper, we have introduced the apparatus of Generalised Probabilistic Theories to the study of the nature of the gravitational field. While previous results mostly adopted a specific 12 model of gravity, Generalised Probabilistic Theories allow us to take a model-independent approach and derive general results on the nature of the gravitational field. The long-term goal of this type of study is to be able to constrain the type of physical theories which are compatible with experimentally testable scenarios, and to systematically check the internal consistency of proposed theories of gravity. Here, we paved the way for reaching such long-term goal by proving a no-go theorem on the nature of the gravitational field, stating that, if gravity-induced entanglement between two masses is detected in an interferometric experiment, then either gravity violates the no faster-than-light signaling principle (because it has some internal mechanism which allows for superluminal communication or because it is not a physical degree of freedom) or it has to be non-classical. We have also shown that non-classicality of gravity does not necessarily imply that it is quantum, and we have provided alternative models of systems which are neither classical nor quantum but, that can nonetheless entangle quantum systems. Our description of the nature of the gravitational field is only concerned with the operational, or effective, behaviour of gravity, and we do not make claims about how gravity should be described at the ontological level. Our result does not exclude the possibility that there is an intrinsic decoherence or collapse mechanism which prevents gravity from entangling two quantum systems. Here, we address a different question, namely which conclusions can be drawn on the nature of the gravitational field if experiments show that two quantum systems can be entangled solely via the gravitational interaction. It would also be interesting to consider such theories using the GPT framework, in particular, to ascertain whether or not they satisfy the basic operational principles that it imposes. In future work we will address the two open problems given in Section III B as well as exploring Conjecture 1 further. Examples of well-motivated classes of GPTs to consider for the first open problem may be transitive GPTs, tomographically local GPTs or GPTs obeying the purification principle. In preliminary work we showed that although an interaction between a non-contextual system (a stabilizer qubit) and two qubits could generate entanglement, this interaction led to operational inconsistencies when applied in different scenarios. This gives some credence to our conjecture but is still a way off constituting a general proof. The main aim of future work is to provide a full characterisation of all non-classical gravitational theories which can mediate entanglement between quantum systems. This would shed light on the possible features of a unified account of general relativity and quantum theory.

ACKNOWLEDGMENTS

The authors would like to thank Markus Aspelmeyer and Markus M¨ullerfor helpful comments on the draft. T.D.G. and F.G. acknowledge support from Perimeter Institute for Theoretical Physics. Research at Perimeter Institute is supported in part by the Government of Canada through the Department of Innovation, Science and Economic Development and by the Province of Ontario through the Ministry of Colleges and Universities. J.H.S. was supported by the Foundation for Polish Science through IRAP project co-ﬁnanced by EU within Smart Growth Operational Programme (contract no. 2018/MAB/5). Most of the diagrams were prepared using TikZit.

Appendix A: Beyond quantum theory: an introduction to GPTs

In this section we provide a brief introduction to the formalism of GPTs. In particular, we take a diagrammatic approach to GPTs based on process theories. This approach has two key benefits, first, the proofs are much more intuitive and straightforward than using alternative methods, second, the results can be much more general, for instance, applying to infinite dimensional systems 13

[114] and non-convex theories [115]. We will illustrate this abstract formalism by means of examples towards the end of this section. There are two key components to a GPT, systems, denoted by labeled wires, for example:

A (A1) and processes, denoted by labeled boxes, for example:

C A A f (A2) A B

Such processes have input systems at the bottom, in this example systems A and B, and output systems at the top, in this example C, A and A. These processes can be wired together to form diagrams, such as:

C C A A f A B A g (A3) C A C A B s C a key feature of a GPT is that it is closed under forming such diagrams, that is, that these diagrams must themselves be processes within the theory, in this case with input C and output C, C, A, and A. This wiring together of processes is not completely free, it must be acyclic and connect outputs to inputs of the same type. Two diagrams are the same if they have the same connectivity, that is, if one can be deformed into the other simply by sliding around the processes on the page while preserving the connectivity of the diagram, for example:

C C A A C C A A f f A B A A g = B g (A4) C A A C A B s s C C

Whilst it might be tempting to imagine these diagrams as being embedded within some space-time diagram, it is important to note that this is not always possible. In particular, the systems that we consider and represent by 1D wires do not necessarily have to correspond to localised degrees of freedom, and hence, may not simply trace out world-lines if drawn on a space-time background. For example, one could consider the electromagnetic field as a system, or, as we will do in this paper, the gravitational field will be represented by a system but these clearly should not be thought of as being localised. There are three special types of processes within a GPT which are important: those with no inputs, known as states; those with no outputs, known as effects; and, those with neither inputs 14 nor outputs, known as scalars. These are denoted as:

A B s , e and n respectively. (A5) A C B

We work with so-called causal GPTs, these have a unique eﬀect for each system A, the discarding map, denoted by:

A (A6)

This uniqueness has important consequences, one the one hand it means that discarding two components of a system is the same as discarding the full system:

A B = A B (A7) and, on the other hand, that if we discard the output of any process it is the same as discarding its input:

C A A f = (A8) A B A B

This condition is what ultimately means that the systems in a GPT satisfy subsystem independence – that is, we can uniquely deﬁne reduction maps from a composite system to its components. This in turn ensures no-signalling between space-like separated parties [102, 116] and, more generally, compatibility with the causal structure of space-time [117]. Finally, it also implies that there is a unique scalar, which we can denote by the empty diagram:

(A9)

Note that in many approaches to GPTs the scalar encode probabilities, but here we have seen that there is just a single scalar. From this point of view we instead ﬁnd that the probabilistic component of the GPT comes from the fact that it is assumed that there is a classical subtheory described by stochastic maps. That is, we have particular systems which are labeled by sets X which we will draw as:

X (A10) to distinguish them from the general systems. Processes between these classical systems, such as:

Y σ (A11) X correspond to stochastic maps between these sets. In particular, this means that the states of the classical systems are simply probability distributions over the set:

X p (A12) 15 and moreover, that there is a unique eﬀect which discards systems which corresponds to marginal- ising over a variable:

(A13) X

Note also that the stochasticity condition ensures that the causality condition is satisﬁed:

Y σ = . (A14) X X

With the classical subtheory in place we can then deﬁne measurements and controlled processes with the theory. A destructive measurement is simply a process which has a generic input and a classical output:

X M (A15) A whilst a more general non-destructive measurement would also have a non-classical output:

B X M (A16) A

Now suppose that we have some state for the input system, how do we compute the probability that a particular outcome occurs? It is simply via the following diagram:

X B M (A17) A s this is a state of the classical system, and hence, a probability distribution over the outcome space X. Classical systems can moreover be used as control systems, and so, for instance, we can control which process occurs conditioned on the outcome of some measurement. This would be described by the diagram:

A C (A18) B X C M D

Perhaps the deﬁning feature of a GPT is that it satisﬁes the principle of tomography, that is, if two processes are not equal then there is some experimental way to distinguish them. Within our 16 formalism this is captured by the condition:

X X M M C A A C A A C A A C A A f = g ⇐⇒ ∀s, M, E f E = g E (A19) A B A B A B A B s s An important concept for this project is known as a leak [50, 118, 119], l, that is, a process denoted:

A L l (A20) A which satisﬁes the following constraint:

A L l = A (A21) A Note, in particular, that if the leaked system, L, is classical, i.e.:

A X m (A22) A then this captures the notion of a non-disturbing measurement. A key feature of classical theory is that it satisﬁes a stronger notion of tomography. Namely, classical systems can (in principle at least) be characterised by performing non-disturbing measurements which factorise over subsystems:

B M N L B M N L Y Z Y Z m m n n B X Z B Y Z Y Z Y Z f = g ⇐⇒ ∀l, L, m, M, n, N f = g A X A X X X l l X X A µ A µ (A23) where µ is the uniform probability distribution over X. A key consequence of this characterisation of classical theory, which will be useful for this project, is that we can decompose a classical system as a sum of projectors [101]:

X X x X = (A24) x∈X x X 17

Note that the eﬀects in this decomposition are not included as processes within the GPT as they do not satisfy condition A8. To keep track of this fact we invert the colours of the process. Nonetheless they are a useful computational tool, so we typically work with the minimal extension of the GPT which includes them as processes. In this case, there is no longer a unique scalar, and instead, scalars correspond to probabilities, for example, the probability of obtaining outcome x for a given measurement corresponds to the diagram:

x X B M = p ∈ [0, 1] . (A25) A s

We can then deﬁne the eﬀect space for a general system via non-causal processes:

ex|M X = M (A26) A A when we compose these with a state, s, we obtain the probability that we would observe outcome x ∈ X having performed measurement M on the system A prepared in state s. With these in place we are now in the position that we can retrieve the geometric data which (in the literature) has typically been associated with a GPT system:

Definition 4 (GPT system). A generic GPT system, S, consists of a set of states, ΩS, effects, ES and transformations, TS, which can be embedded in a finite dimensional real vectors space VS. States are mapped to vectors, effects are mapped to dual vectors and transformations induce linear maps VS → VS. In the case of classical systems, the fact that they can be characterised by non-disturbing measurements is equivalent to the following geometric characterisation:

Example 3 (Finite dimensional classical system). A finite dimensional classical system Cd associated to a set {1, ..., d} has states which are the probability measures on X and the measurements consists of all finite tuples {fi} of positive measurable functions on X which sum to 1 for all x ∈ X. The state space of a system Cd is a simplex with d-vertices. A classical system Cd can be embedded d in a vector space R . A defining property of finite dimensional classical systems is that every mixed state can be uniquely decomposed into a convex combination of extremal states. Recent work, Refs. [108, 120] however has demonstrated that the assumption that all systems are classical does not imply that the theory as a whole is classical. Indeed, they surprisingly show that they may even still admit entangled states. In order to obtain classical theory as a whole we must also impose how these classical systems compose: Example 4 (Finite dimensional classical theory). Systems correspond to finite dimensional classical systems (Def. 3) which compose via Cd ⊗ Cd0 = Cdd0 . This follows from our definition from the assumption that the non-disturbing measurements, which characterise the classical systems, factorise over subsystems. This is the notion of local tomography [105, 106] which was shown in [120] to suffice (together with classicality of systems) to pick out classical theory. 18

(a) Boxworld state space. This (b) Qubit state space. The sur- (c) Rebit state space. The rebit is a state space for which there face of the sphere corresponds can be obtained from the qubit exist two extremal measurement to the pure states whilst the in- by restricting to convex combi- (equivalent to the X and Z mea- terior of the sphere consists of nations of pure states of the form surements on a qubit) for which mixed states. Antipodal states a |0i+b |1i with a, b ∈ R. Equiv- the extremal states (vertices of on the sphere (corresponding to alently it is the state space ob- the square) give deﬁnite values orthgonal pure states) can be tained if one removes the Y mea- for both measurements (unlike perfectly distinguished. surement from the qubit. states of a qubit). There is however no simultaneous measurement of both X and Z.

Example 5 (Quantum system). A ﬁnite dimensional quantum system Qd associated to a complex d d Hilbert space C has states corresponding to trace one self-adjoint operators on C , measurements correspond to POVMs and transformations correspond to CPTP maps. A quantum system Qd d2 can be embedded in a vector space R .

Similarly to classical theory, we can characterise quantum theory as a tomographically local theory where systems are quantum systems:

Example 6 (Quantum theory). Systems correspond to quantum systems (Def. 5) which compose via Qd ⊗ Qd0 = Qdd0 .

We are also interested in systems which are not classical in the strict sense deﬁned above, but which, nonetheless, be understood as having a satisfying underlying classical explanation. It is well known [121, 122] that any generalised probabilistic system can be viewed as a restriction on some underlying classical system. These, however, typically end up violating the desiderata for classical explanations – based on Leibniz’s principle of the identity of indiscernibles [123]– known as generalised noncontextuality [113]. Within the language of GPTs this is equivalent to a very natural simulability criteria known as the simplex embeddability condition [65].

Definition 5 (Simplex embeddability condition and classical* systems). We say that some GPT system, S, with associated state space ΩS, effect space ES and vector space VS is simplex embed- d dable if and only if there exists some classical system C with associated state space ∆d and effect ∗ d ∗ d∗ space δd, and a pair of linear maps ι : VS → R and κ : VS → R such that:

ι(ΩS) ⊆ ∆d, (A27) ∗ κ(ES) ⊆ ∆d, and (A28) ∀s ∈ ΩS, e ∈ ES κ(e)[ι(s)] = e[s]. (A29) 19

A classic example of a simplex embeddable, that is, a classical*, system is given by the epistemically restricted theories of Spekkens [121, 124]. For example:

Example 7 (Epistemically restricted bit). The state and effect spaces are explicitly constructed ∗ as subsets of ∆4 (a tetrahedron) and ∆4 respectively. In particular, the states are restricted to an octahedron defined by the convex hull of the midpoints of the edges of the tetrahedron, and the ∗ effects are similarly restricted to the midpoints of the edges of the dual hypercube ∆4.

Whilst a classic example of a GPT which is not simplex embeddable is colloquially known as Boxworld [70].

Example 8 (Boxworld bit). The statespace of a boxworld bit is given a square embedded in R3 in a hyperplane that does not intersect the origin. The eﬀect space is then the dual convex set (i.e., the space of linear functionals that evaluate to [0, 1] on the square state space) living in (R3)∗, this is an octahedron with one vertex the origin.

Finally, we will introduce some basic notions from quantum entanglement theory within the GPT framework.

Deﬁnition 6 (Entangled and separable states). A bipartite state, S, is separable if it can be decomposed in the following way:

A B A B L R S = PS PS (A30)

X pS where the white dot copies the classical input. A state that is non-separable (that is, for which L R there does not exist, X, pS, PS and PS such that the above equality holds) is deﬁned to be entangled.

More geometrically, separability is equivalent to the condition that:

A B X A B = px L R (A31) S sx sx x∈X which can be seen by using the decomposition of the classical system of Eq. (A24).

Deﬁnition 7 (LOCC transformations). An LOCC transformation is a bipartite transformation, 20

T , which can be decomposed in the following way:

C D R tn

Xn L tn Ln−1 Yn−1 Rn−1 C D ...... T = ...... (A32) A B L2 Y2 R2 R t2

X2 L t2 L1 R1 Y1 R t1

X1 L t1

A B where n labels the number of rounds of classical communication of the protocol. Like with separable states, we could gain a more geometric picture of this notion by decomposing the mediating classical systems using Eq. (A24).

Appendix B: Proof of Theorem III.1 (no-go theorem)

Let us observe that the statement of Theorem III.1 is equivalent to saying that a classical system cannot mediate entanglement via local interactions . Condition 1 (subsystem independence) is part of the definition of a GPT. Let us observe that the statement of Theorem III.1 is equivalent to saying that a field G which is modelled either as a classical system or non-dynamically cannot mediate entanglement via local interactions. The fact that a non-dynamical entity cannot mediate entanglement is immediate, since mediation is defined as requiring a system to do the mediating. Therefore we can rephrase our no-go theorem in the following manner: Theorem B.1. Interactions mediated by classical systems cannot generate entanglement for any GPT as any interaction mediated by a classical gravitational field describes an LOCC protocol.

Proof. Let us denote a classical gravitational ﬁeld as G. Then, given two quantum systems A and B, we can consider an example of an interaction between these mediated by the gravitational ﬁeld:

A G B

IB G I A (B1) G s A B 21

That is, we have the initial state of the ﬁeld s, which then ﬁrst interacts with the quantum system A via the interaction IA and then with the system B via the interaction IB.

To see that this is incapable of generating entanglement, we insert a resolution of the identity for the system G (which is possible following Eq. (A24) as G is assume to be classical) in between the two interactions, and post select on some ﬁnal ﬁeld state s0:

A B s0 G IB G X f (B2) f f G IA G s A B

This can be rewritten as follows:

A B

A B f s0 G G X X f f IA IB = EA EB (B3) f G G f s f A B

A B

f f where EA and EB are completely positive trace nonincreasing maps for the respective quantum systems. This transformation is therefore manifestly within Sep and hence is known to not generate entanglement.

This, however, is just a simple example of an interaction between two quantum systems mediated by a classical gravitational ﬁeld. One may therefore be tempted to speculate that some more complicated interaction could generate entanglement.

As discussed in detail in Appendix C the most general form that such an interaction can take 22 can be written as a sequence of pairwise interactions of the above form, i.e.:

A G B (n) IB G (n) IA A G B

......

A G B (2) IB G (2) IA A G B (1) IB G I(1) A (B4) G s A B Here, we could follow a similar argument to before, inserting resolutions of identities of G between each interaction. However, perhaps a more illuminating argument is as follows. Namely, under a trivial rewriting we can see that this is nothing but an n-round LOCC protocol:

A G B (n) IB G (n) IA A G B

......

A G B (2) IB G (2) IA A G B (1) IB G I(1) A (B5) G s A B Alice Bob We can then just use the known fact that LOCC is incapable of generating entanglement to obtain the desired result for quantum systems A and B. This fact moreover holds when A and B are arbitrary GPT systems, that is, it is GPT-LOCC or GPT-SEP transformations are incapable of generating entanglement for GPTs. 23

Appendix C: Linearized quantum gravity model

The goal of this section is to give a concrete model of the quantized gravitational field to justify the decomposition present, e.g., in Theorem B.1, where each single system interacts with the gravitational field alone. In the case in which there is only one source particle for the gravitational field this condition is easily satisfied, as the interaction is only determined by the contribution of the single source on the gravitational field. The non-trivial case is when two particles, say P1 and P2, are the source of the gravitational field. In such case, it is not straightforward that the interaction is, at least at first order, factorisable into two commuting terms Uˆ1GUˆ2G = Uˆ2GUˆ1G, where i = 1, 2 labels the Hilbert space of particle Pi. In order to show that this decomposition arises naturally in a concrete model, we consider linearized quantum gravity. In this model, gravity is described as a quantum field on some fixed background. Specifically, the (quantum) gravitational fieldg ˆµν is split into a classical term, serving as a background (usually Minkowski background), and a quantized part, described as a spin-2 field on such fixed background. In particular, it is assumed that the quantized part of the gravitational field is a small perturbation. Loosely speaking, this means that the metric field is decomposed asg ˆµν = ηµν + hˆµν, where ηµν is the Minkowski metric and hˆµν is a small quantum correction, corresponding to the quantized part of the gravitational field. All the results only hold to first √ q 16πG order in κ = c4 . Let us start by considering the Hamiltonian of linearized gravity interacting with matter, which can be written as Hˆ = Hˆ0 + κVˆ + HˆG + HÎ , where Hˆ0 is the free Hamiltonian of the matter field, Vˆ is a potential which is a function of the matter degrees of freedom, HˆG is the free Hamiltonian of the linearized gravitational field, and HÎ is the interaction Hamiltonian between matter and the gravitational degrees of freedom. For later convenience, we can write Hˆ = Hˆfree + HÎ , where Hˆfree = Hˆ0 + κVˆ + HˆG contains all the terms but the interaction between gravity and matter. We focus on the gravitational part of the Hamiltonian, which can be expressed more explicitly as

X Z d3k Hˆ = ω ˆb†ˆb , G (2π)3 k k k r r (C1) κ X Z d3k h i Z Hˆ = − ~ √ Jˆ (k)ˆb† + Jˆ†(k)ˆb = d3xhˆ (x)Tˆij(x), I 2 (2π)3 2ω r k r k ij r k

ˆ† ˆ where the sum over r is the sum over the polarisations of the gravitational field, bk and bk are respectively the creation and annihilation operators of the gravitational field, and Jˆr(k) is the matter current sourcing the gravitational field. Notice that the interaction Hamiltonian can be expressed also as the coupling between the stress energy tensor Tîj(x) and the quantum corrections to the metric. This is analogous to considering the minimal coupling in electromagnetism. The matter current has an explicit expression as Z ˆ r 3 −ik·x îj Jr(k) = Lij(k) d xe T (x), (C2)

r where Lij(k) is the polarisation matrix (and not vector as in the electromagnetic case) with the P r r 1 kikj property that r LijLlm = 2 (PilPjm + PimPjl − PijPlm), with Pij = δij − |k|2 being the projector on the transverse direction of the wave vector. We have now all the elements to proceed to the calculation. We first move to the interaction picture. Concretely, this maps the conserved current of the matter degrees of freedom and the creation and annihilation operators of the gravitational field into their free-field time-evolved 24 equivalent, i.e.,

ˆ ˆ iωkt bk(t) = bke , (C3) i ˆ ˆ i ˆ ˆ ˆ (H0+κV )t ˆ − (H0+κV )t Jr(k, t) = e ~ Jr(k)e ~ . (C4) The state then evolves according to the interaction Hamiltonian r κ X Z d3k h i Hˆ (t) = − ~ √ Jˆ (k, t)e−iωktˆb† + Jˆ†(k, t)eiωktˆb . (C5) I 2 (2π)3 2ω r k r k r k We are now going to use the assumption that the matter field is in a superposition of classical configurations. Hence, in every amplitude, the source can be treated as classical. Since the Hamil- tonian is time-dependent, we have to calculate the time evolution from time t = 0 to time t = T as the time-ordering of infinitesimal terms, i.e., ˆ ←− n T/δt ˆ o U(T ) = lim T Πn=0 U((n + 1)δt, nδt) , (C6) δt→0 which gives, apart from an overall phase

P R d3k h † ∗ i − fr(k)ˆb −f (k)ˆb Uˆ(T ) = e r (2π)3 k r k , (C7) where r i κ 1 Z T ~ −iωks fr(k) = − √ dsJr(k, s)e . (C8) ~ 2 2ωk 0 This result shows that the evolution operator is a displacement operator on the gravitational degrees of freedom, with a parameter of the displacement that is different for each classical configuration of the matter field. This means that the gravitational field gets entangled with the two masses§ and the vacuum state of the field, which is our initial state, is mapped to a coherent state. From the definition of fr(k) and Jr(k, s) we obtain r i κ 1 Z T Z ~ r 3 −ik·x −iωks ij fr(k) = − √ Lij ds d xe e T (x) = ~ 2 2ωk 0 r (C9) i κ 1 Z T Z ~ r 3 −iωks ij = − √ Lij ds d xe T (x), ~ 2 2ωk 0 because the oscillating term e−ik·x is negligible for a nonrelativistic particle. We assume that, whatever the free Hamiltonian of the matter is, at the leading order it does not contain direct interactions between the two particles P1 and P2. This means that we can approximate the stress- energy tensor of the two particles as the sum of the two stress-energy tensors

ij ij ij T (x) = T1 (x) + T2 (x). (C10)

A+B A B − 1 [A,B]+··· With this assumption, and using the Baker-Campbell-Haussdorﬀ formula e = e e e 2 for A and B two general operators, we can factorise the total evolution into two (commuting) factors, each of which is an interacting term between the classical conﬁguration of the matter and the gravitational degrees of freedom. Concretely, the unitary operator can be written as

3 h i 3 h i − P R d k f 1(k)ˆb† −f 1∗(k)ˆb − P R d k f 2(k)ˆb† −f 2∗(k)ˆb Uˆ(T ) = e r (2π)3 r k r k e r (2π)3 r k r k , (C11)

§ Notice that fr(k) vanishes if the particle sourcing the gravitational ﬁeld is not accelerated. 25 where all other terms are higher order in κ. This can be easily seen by computing the commutator

Z 3 3 0 X d k d k h 1 † 1∗ 2 0 † 2∗ 0 i f (k)ˆb − f (k)ˆb , f 0 (k )ˆb − f 0 (k )ˆb 0 = (2π)3 (2π)3 r k r k r k0 r k r,r0 Z 3 3 0 X d k d k n 1 2∗ 0 hˆ† ˆ i 1∗ 2 0 hˆ ˆ† io = − f (k)f 0 (k ) b , bk0 + f (k)f 0 (k ) bk, b 0 = (C12) (2π)3 (2π)3 r r k r r k r,r0 Z 3 X d k 1 2∗ 0 1∗ 2 0 = f (k)f 0 (k ) + f (k)f 0 (k ) = O(κ). (2π)3 r r r r r

In particular, this result implies that

Uˆ(T ) = Uˆ1G(T )Uˆ2G(T ) = Uˆ2G(T )Uˆ1G(T ), (C13)

P R d3k h i ˆ† i∗ ˆ i − r 3 fr(k)bk−fr (k)bk where UîG(T ) = e (2π) and i = 1, 2. This result implies that the interaction of particles A and B with the gravitational field can be factorised, at the first order in perturbation theory, into two commuting terms describing the interaction of each system A and B with the gravitational field. This interaction is compatible with the structure of interactions we consider in the proof of Theorem B.1, and is verified for linearized quantum gravity—the only widely accepted model for describing quantum aspects of the gravitational field. Since linearized quantum gravity generates entanglement, it violates condition 3 of our no-go theorem, namely that gravity is classical.

Appendix D: Methods: Circuit based approach

In this Appendix, we show how to cast the thought experiments in Refs. [27, 28] and in Ref. [21] in the language used in this paper. In particular, this step is needed to apply our no-go theorem to these previous results and show that they are consistent with it. As we are considering the gravitational field as a physical degree of freedom in its own right, then we can describe its interactions with other physical degrees of freedom using a circuit based approach. In this section we will explore various different circuits which could describe a fundamental interaction between various degrees of freedom which could generate entanglement, depending on the nature of the physical degree of freedom representing the gravitation field. That is, whether it is a quantum, a classical, or some other sort of physical system.

1. Interactions in the Hamiltonian description and circuit description

a. Classical gravitational ﬁeld acting locally

Let us consider a scalar ﬁeld φ(x) corresponding to our classical system G. Its interaction with a quantum system A is given by a coupling term of the form gφ(x) in the Hamiltonian HA. The scalar ﬁeld G acts locally on two quantum systems A and B when the interaction Hamil- tonian is of the form H = HA + HB. The unitary generated by this Hamiltonian is of the form i(H +H )t UAB(t) = e A B = UA(t)⊗UB(t) where the unitaries UA(t) and UB(t) depend on the state φ(x) 26 of G. As such in circuit notation this is a circuit of the form:

A UA . (D1) G • •

B UB

One typically leaves out the system G in the circuit.

b. Classical gravity acting non-locally

First consider the case where the gravitational ﬁeld acting on two quantum systems A and B such that H 6= HA + HB. This entails that the unitary is no longer product, which is represented as:

C • . (D2) Q1 U Q2

The above is in general a non-local action of the classical ﬁeld C on Q1 and Q2 (which could be space like separated). It is in eﬀect joint unitary on Q1Q2 controlled on the classical system C.

c. Gravitational ﬁeld as a quantum system mediating interaction

Whatever its nature, we wish to treat the gravitational field as a system G. Namely it has dynamical degrees of freedom and can interact with quantum systems. We assume in addition that the gravitational field acts locally on each quantum system (thus ruling out the coupling constant modelling of G) and that quantum systems do not interact directly with each other: all interactions between systems are mediated by G. We consider two qubits A and B interacting pairwise with a third system G corresponding to the gravitational field. A general circuit is of the form:

A ... . (D3) U1 Un G ... V1 Vn B ... If the system G is taken to be a quantum system, then we can consider a Hamiltonian description of an interaction between AB which is mediated by G. This has the following form:

H = HAG ⊗ IB + IA ⊗ HGB , (D4)

Using the Lie formula:

HAG+HBG HAG/n HBG/n n e = limn→∞(e e )

We see that this is just an inﬁnite sequence of local unitaries between AG and BG of inﬁnitesimally iH /n small time steps. This is the n → ∞ limit of the circuit of Equation D3 when Ui = e AG and iH /n Vi = e BG for all i.. 27

| + ⟩A A | + ⟩A A UAB UAB | + ⟩B B | + ⟩B B

|0⟩S SA |0⟩S SA A = A SB SB |0⟩SB |0⟩SB

|00⟩G G |00⟩G G

FIG. 4: Circuit for the BMV experiment. The CNOTX,Y operation is represented as a black dot (indicating the system X on which the conditioning depends) linked by a solid black line to the ⊕ symbol (representing the NOT operation on system Y ). The two × symbols connected by a solid black line indicates a SWAP operation between two systems.

2. Circuit description of Bose-Marletto-Vedral

In the Bose-Marletto-Vedral experiment, illustrated in Fig. 4, two masses are initially put in a spatial superposition. The state corresponding to the mass i = A, B being on the left is labelled as |0ii, and the state corresponding to the mass i being on the right is labelled as |1ii. Because of the Newtonian interaction between the masses, the phase accumulated in the interferometer is different according to the relative distance between the masses. In our treatment, we explicitly include the state of the gravitational degrees of freedom, and consider the two masses entangled with it. For each configuration of the two masses, we have a different state of the gravitational 2 2 field. This is naturally described by associating to the gravitational field a C ⊗ C Hilbert space, with the corresponding basis {|00iG , |01iG , |10iG , |11iG}. The state of the gravitational field at the beginning of the experiment can be obtained by applying two CNOT gates controlled on the systems A and B, as shown in Fig. 4 |0i + |1i |0i + |1i 1 X CNOT CNOT A√ A B√ B |00i = |ii |ji |iji (D5) A,G B,G G 2 A B G 2 2 i,j=0,1 While traversing the interferometer, the systems A and B acquire a phase, which depends on the Newtonian interaction, and thus on the distance, between them. In the circuit model, we can represent the evolution of the full state in the interferometer as a unitary operator controlled on the state of the gravitational field, which acts as

ˆ iφij UAB |iiA |jiB |ijiG = e |iiA |jiB |ijiG . (D6) In order to test the entanglement, we need to swap it from the systems A and B to some spin states embedded in each mass, which we label as SA and SB. The state becomes 1 X iφij CNOTA,S CNOTB,S e |ii |ji |iji |0i |0i = A B 2 A B G SA SB i,j=0,1 (D7) 1 X = eiφij |ii |ji |iji |ii |ji . 2 A B G SA SB i,j=0,1 At the end of the experiment, the masses are sent through a reverse Stern-Gerlach and recom- bined. The state after the recombination is 1 X 1 X eiφij |ii |ji |iji |ii |ji → |0i |0i |00i eiφij |ii |ji . (D8) 2 A B G SA SB A B G 2 SA SB i,j=0,1 i,j=0,1 28

| + ⟩A A | + ⟩A A | + ⟩A A

|0⟩G G |0⟩G G |0⟩G G

|0⟩S S = |0⟩S S = |0⟩S S

|0⟩B B |0⟩B B |0⟩B B

FIG. 5: Circuit of the protocol of Refs. [21, 22]. The CNOTX,Y operation is represented as a black dot (indicating the system X on which the conditioning depends) linked by a solid black line to the ⊕ symbol (representing the NOT operation on system Y ). The two × symbols connected by a solid black line indicates a SWAP operation between two systems.

These operations can formally be written as a series of four CNOT operations controlled on the state of the spin SA and SB and acting on the Hilbert space of A and G, i.e., CNOTSA,ACNOTSA,G, and B and G, i.e., CNOTSB ,BCNOTSB ,G respectively. The result of the protocol is equivalent to a swap of the state of the systems A and B with the state of the respective spin systems SA and SB.

3. Circuit description of the protocol of Refs. [21, 22]

The protocol of Refs. [21, 22] is illustrated, in its circuit equivalent, in Fig. 5. In this protocol, we consider a massive particle A, which is the source of the gravitational field, the gravitational field G, an auxiliary spin degree of freedom S, and a probe particle B, which does not backreact on the gravitational field. Initially, A is put in a spatial superposition of two locations. We here denote as the qubit

|0iA the state corresponding to the left amplitude of the superposition, and with qubit |1iA the state corresponding to the right amplitude of the superposition. The initial state is then |+iA = √1 (|0i + |1i ). As the superposition is created, the state of the gravitational field gets entangled 2 A A with the state of A. The full quantum state of A, G, S, and B is 1 |Ψi = √ (|0i |0i + |1i |1i ) |0i |0i , (D9) 2 A G A G S B where |0iG , |1iG denote the quantum states of the gravitational field respectively corresponding to the left and right configuration of the quantum system A. This operation corresponds to a CNOT gate on the Hilbert spaces of A and G, where A acts as the control qubit, and G behaves as the target qubit. When B releases the trap, the gravitational interaction (e.g., Newtonian potential) between A and B maps the state to 1 |Ψi = √ (|0i |0i |0i + |1i |1i |1i ) |0i , (D10) 2 A G B A G B S which also corresponds to a CNOT gate on B controlled on the Hilbert space of G. Since it is not relevant to our argument, we here neglect the relative phase due to the gravitational interaction. It would be possible to take it into account by adding a unitary operator like in Appendix D 2. If we want to certify that the systems A and B become entangled via the gravitational field, we have two possibilities:

• We can measure system G directly. In this case, we could for instance perform a measurement in the basis |±i = √1 (|0i ± |1i ) of the gravitational ﬁeld. This corresponds to being G 2 G G 29

able to measure a quantum state of the gravitational ﬁeld, which is out of reach of current technologies. We do not consider this possibility here; • We infer features of G via an indirect measurement of the two systems A and B. One possibility is to recombine the state of A adiabatically, in such a way that it does not decohere [125]. In this case, the full state becomes a product state of the state of G and the state of A and B. In order to show that the state of A and B is entangled, we then need to swap the state of A and G with an auxiliary quantum system S, for instance a spin state. We can then measure the state of S and B in order to show that the state of A, G, and B was initially entangled.

If we follow the latter procedure, the protocol we obtain is illustrated in Fig. 5. The ﬁnal result is that the protocol is equivalent to swapping the state of A with the state of S, and the state of G with the state of B. The remaining CNOT gates after the preparation in the third circuit in Fig. 5, when acting on the speciﬁc initial state that we choose, reduce to an identity operator.

Appendix E: Entanglement via density hypercubes (DHC)

Density hypercubes [103] deﬁne a GPT, DHC, such that quantum theory, Quant, is a full subtheory of the Karoubi envelope [126], K, of the theory. That is, we have that Quant ⊂ K[DHC]. In particular, the systems, Q, in this full subtheory are given by hyperdecohered density hypercube systems, H. In particular, this means that there is some idempotent hyper : H → H which split through a quantum system Q:

H Q H e d hyper = Q and Q = H (E1) H d e H Q hence, in particular, we have a reversible encoding of Q into H. Suppose then, that we have some simple quantum circuit which can generate entanglement between quantum systems A and B:

A Q B , (E2) U

then, we know, we can always factorise the mediating quantum system through a hypercube system: 30

B V Q d A H B , (E3) e Q U A hence, by identifying the dashed boxes with processes U and V we obtain a protocol for generating entanglement mediated by a hypercube system, H:

A H B . (E4) U

Clearly this method can be straightforwardly generalised to more complicated scenarios in which there is some more complicated circuit which generates the entanglement between the two quantum systems.

Appendix F: Comparison to constructor theory

In [35] the authors use the framework of constructor theory to show that any system M generating entanglement between two qubits must be non-classical, where non-classicality is defined as the existence of incompatible observables. The main difference to our result is the use of the constructor theory framework (as opposed to our use of the GPT framework). One of the main strengths of the GPT framework is that it allows us to provide explicit examples of a non-quantum system which can mediate entanglement between two qubits. Indeed, we provide explicit examples of such GPTs in section III B. To our knowledge there are no explicit examples of constructor theories beyond classical and quantum theory. Moreover, GPTs are a framework for describing the effective nature (or behaviour) of systems, whilst the constructor theory approach is a framework for describing the ontological nature of systems. As such the result of [35] assumes that the gravitational field is fundamentally an information medium and then makes statements about its fundamental nature. As such their result applies only modulo assumptions about ontology. This is akin to a Bell like theorem which states that under the assumption of a specific ontology (namely classical ontological models) then quantum systems are non-local. On the other hand our result makes no assumption about underlying ontology, and is solely about effective behaviour of the gravitational system. That is, as discussed in Sec. 18 of [105], GPTs apply an operational methodology, and, as such, do not require one to make any ontological commitments. 31

Finally, the constructor theory framework makes an assumption of locality, which, when trans- lated to the GPT formalism is known as tomographic locality [99]. This is a substantial restriction on the scope of GPTs, since many GPTs (such as real vector space quantum theory [106], general theories based on Jordan algebras [93], general composite systems with modiﬁed Born rules [84], and many others [107, 108]) violate this assumption.

[1] R. Penrose, “On the gravitization of quantum mechanics 1: Quantum state reduction,” Foundations of Physics, vol. 44, no. 5, pp. 557–575, 2014. [2] L. Diosi, “A universal master equation for the gravitational violation of quantum mechanics,” Physics Letters A, vol. 120, no. 8, pp. 377–381, 1987. [3] S. Hossenfelder, “Experimental search for quantum gravity,” Classical and quantum gravity: Theory, Analysis and Applications, vol. 5, 2010. [4] M. Bronstein, “Republication of: Quantum theory of weak gravitational fields,” General Relativity and Gravitation, vol. 44, no. 1, pp. 267–283, 2012. [5] D. N. Page and C. D. Geilker, “Indirect evidence for quantum gravity,” Phys. Rev. Lett., vol. 47, pp. 979–982, 1981. [6] K. Eppley and E. Hannah, “The necessity of quantizing the gravitational field,” Foundations of Physics, vol. 7, no. 1, pp. 51–68, 1977. [7] J. Mattingly, “Why Eppley and Hannah’s thought experiment fails,” Phys. Rev. D, vol. 73, p. 064025, 2006. [8] S. Carlip, “Is quantum gravity necessary?,” Class. Quant. Grav., vol. 25, no. 15, p. 154010, 2008. [9] S. M. Giampaolo and T. Macr`ı,“Entanglement, holonomic constraints, and the quantization of fundamental interactions,” arXiv preprint arXiv:1806.08383, 2018. [10] C. M. DeWitt and D. Rickles, The role of gravitation in physics: Report from the 1957 Chapel Hill Conference, vol. 5. epubli, 2011. [11] H. D. Zeh, “Feynman’s interpretation of quantum theory,” The European Physical Journal H, vol. 36, no. 1, pp. 63–74, 2011. [12] L. Ford, “Gravitational radiation by quantum systems,” Annals of Physics, vol. 144, no. 2, pp. 238– 248, 1982. [13] N. H. Lindner and A. Peres, “Testing quantum superpositions of the gravitational field with Bose- Einstein condensates,” Physical Review A, vol. 71, no. 2, p. 024101, 2005. [14] D. Kafri and J. Taylor, “A noise inequality for classical forces,” arXiv preprint arXiv:1311.4558, 2013. [15] D. Kafri, J. Taylor, and G. Milburn, “A classical channel model for gravitational decoherence,” New Journal of Physics, vol. 16, no. 6, p. 065020, 2014. [16] N. Altamirano, P. Corona-Ugalde, R. B. Mann, and M. Zych, “Gravity is not a pairwise local classical channel,” Classical and Quantum Gravity, vol. 35, no. 14, p. 145005, 2018. [17] C. Anastopoulos and B.-L. Hu, “Probing a gravitational cat state,” Classical and Quantum Gravity, vol. 32, no. 16, p. 165022, 2015. [18] C. Anastopoulos and B.-L. Hu, “Quantum superposition of two gravitational cat states,” Classical and Quantum Gravity, vol. 37, no. 23, p. 235012, 2020. [19] M. Carlesso, M. Paternostro, H. Ulbricht, and A. Bassi, “When Cavendish meets Feynman: A quantum torsion balance for testing the quantumness of gravity,” arXiv preprint arXiv:1710.08695, 2017. [20] M. Bahrami, A. Bassi, S. McMillen, M. Paternostro, and H. Ulbricht, “Is gravity quantum?,” arXiv preprint arXiv:1507.05733, 2015. [21] A. Belenchia, R. M. Wald, F. Giacomini, E. Castro-Ruiz, C.ˇ Brukner, and M. Aspelmeyer, “Quantum superposition of massive objects and the quantization of gravity,” Physical Review D, vol. 98, no. 12, p. 126009, 2018. [22] A. Belenchia, R. M. Wald, F. Giacomini, E. Castro-Ruiz, C.ˇ Brukner, and M. Aspelmeyer, “Informa- tion content of the gravitational field of a quantum superposition,” International Journal of Modern Physics D, vol. 28, no. 14, p. 1943001, 2019. [23] M. Christodoulou and C. Rovelli, “On the possibility of laboratory evidence for quantum superposition of geometries,” Physics Letters B, vol. 792, pp. 64–68, 2019. 32

[24] R. Howl, V. Vedral, M. Christodoulou, C. Rovelli, D. Naik, and A. Iyer, “Testing quantum gravity with a single quantum system,” arXiv preprint arXiv:2004.01189, 2020. [25] R. J. Marshman, A. Mazumdar, and S. Bose, “Locality and entanglement in table-top testing of the quantum nature of linearized gravity,” Physical Review A, vol. 101, no. 5, p. 052110, 2020. [26] T. Krisnanda, G. Y. Tham, M. Paternostro, and T. Paterek, “Observable quantum entanglement due to gravity,” npj Quantum Information, vol. 6, no. 1, pp. 1–6, 2020. [27] S. Bose, A. Mazumdar, G. W. Morley, H. Ulbricht, M. Toroˇs, M. Paternostro, A. A. Geraci, P. F. Barker, M. Kim, and G. Milburn, “Spin entanglement witness for quantum gravity,” Physical Review Letters, vol. 119, no. 24, p. 240401, 2017. [28] C. Marletto and V. Vedral, “Gravitationally induced entanglement between two massive particles is sufficient evidence of quantum effects in gravity,” Physical Review Letters, vol. 119, no. 24, p. 240402, 2017. [29] M. J. Hall and M. Reginatto, “On two recent proposals for witnessing nonclassical gravity,” Journal of Physics A: Mathematical and Theoretical, vol. 51, no. 8, p. 085303, 2018. [30] C. Anastopoulos and B.-L. Hu, “Comment on “a spin entanglement witness for quantum gravity” and on “gravitationally induced entanglement between two massive particles is sufficient evidence of quantum effects in gravity”,” arXiv preprint arXiv:1804.11315, 2018. [31] A. Mari, G. De Palma, and V. Giovannetti, “Experiments testing macroscopic quantum superpositions must be slow,” Scientific reports, vol. 6, p. 22777, 2016. [32] G. Baym and T. Ozawa, “Two-slit diffraction with highly charged particles: Niels Bohr’s consistency argument that the electromagnetic field must be quantized,” Proceedings of the National Academy of Sciences, vol. 106, no. 9, pp. 3035–3040, 2009. [33] C. H. Bennett, G. Brassard, C. Crépeau, R. Jozsa, A. Peres, and W. K. Wootters, “Teleporting an unknown quantum state via dual classical and Einstein-Podolsky-Rosen channels,” Physical Review Letters, vol. 70, no. 13, p. 1895, 1993. [34] C. Marletto and V. Vedral, “Why we need to quantise everything, including gravity,” npj Quantum Information, vol. 3, no. 1, pp. 1–5, 2017. [35] C. Marletto and V. Vedral, “Witnessing non-classicality beyond quantum theory,” arXiv preprint arXiv:2003.07974, 2020. [36] D. Deutsch, “Constructor theory,” Synthese, vol. 190, no. 18, pp. 4331–4359, 2013. [37] H. Barnum, J. Barrett, L. O. Clark, M. Leifer, R. Spekkens, N. Stepanik, A. Wilce, and R. Wilke, “Entropy and information causality in general probabilistic theories,” New Journal of Physics, vol. 12, no. 3, p. 033024, 2010. [38] G. Chiribella and C. M. Scandolo, “Entanglement and thermodynamics in general probabilistic theories,” New Journal of Physics, vol. 17, no. 10, p. 103027, 2015. [39] G. Chiribella and C. M. Scandolo, “Microcanonical thermodynamics in general physical theories,” New Journal of Physics, vol. 19, no. 12, p. 123043, 2017. [40] H. Barnum, J. Barrett, M. Krumm, and M. P. Müller,“Entropy, majorization and thermodynamics in general probabilistic theories,” arXiv preprint arXiv:1508.03107, 2015. [41] C. M. Lee and J. H. Selby, “Higher-order interference in extensions of quantum theory,” Foundations of Physics, vol. 47, no. 1, pp. 89–112, 2017. [42] A. J. Garner, “Interferometric computation beyond quantum theory,” Foundations of Physics, vol. 48, no. 8, pp. 886–909, 2018. [43] H. Barnum, M. P. Müller,and C. Ududec, “Higher-order interference and single-system postulates characterizing quantum theory,” New Journal of Physics, vol. 16, no. 12, p. 123029, 2014. [44] B. Dakić,T. Paterek, and C.ˇ Brukner, “Density cubes and higher-order interference theories,” New Journal of Physics, vol. 16, no. 2, p. 023028, 2014. [45] H. Barnum, C. M. Lee, C. M. Scandolo, and J. H. Selby, “Ruling out higher-order interference from purity principles,” Entropy, vol. 19, no. 6, p. 253, 2017. [46] S. Horvat and B. Dakić, “Interference as an information-theoretic game,” arXiv preprint arXiv:2003.12114, 2020. [47] J. G. Richens, J. H. Selby, and S. W. Al-Safi, “Entanglement is necessary for emergent classicality in all physical theories,” Physical Review Letters, vol. 119, no. 8, p. 080503, 2017. [48] C. M. Lee and J. H. Selby, “A no-go theorem for theories that decohere to quantum mechanics,” Pro- ceedings of the Royal Society A: Mathematical, Physical and Engineering Sciences, vol. 474, no. 2214, 33

p. 20170732, 2018. [49] C. M. Scandolo, R. Salazar, J. K. Korbicz, and P. Horodecki, “Is it possible to be objective in every physical theory?,” arXiv preprint arXiv:1805.12126, 2018. [50] J. Selby and B. Coecke, “Leaks: quantum, classical, intermediate and more,” Entropy, vol. 19, no. 4, p. 174, 2017. [51] C. M. Lee and J. Barrett, “Computation in generalised probabilisitic theories,” New Journal of Physics, vol. 17, no. 8, p. 083001, 2015. [52] J. Barrett, N. de Beaudrap, M. J. Hoban, and C. M. Lee, “The computational landscape of general physical theories,” arXiv preprint arXiv:1702.08483, 2017. [53] M. Krumm and M. P. Müller,“Quantum computation is an island in theoryspace,” arXiv preprint arXiv:1804.05736, 2018. [54] C. M. Lee and J. H. Selby, “Generalised phase kick-back: the structure of computational algorithms from physical principles,” New Journal of Physics, vol. 18, no. 3, p. 033023, 2016. [55] C. M. Lee and J. H. Selby, “Deriving Grover’s lower bound from simple physical principles,” New Journal of Physics, vol. 18, no. 9, p. 093047, 2016. [56] H. Barnum, C. M. Lee, and J. H. Selby, “Oracles and query lower bounds in generalised probabilistic theories,” Foundations of physics, vol. 48, no. 8, pp. 954–981, 2018. [57] M. P. Müllerand C. Ududec, “Structure of reversible computation determines the self-duality of quantum theory,” Physical Review Letters, vol. 108, no. 13, p. 130401, 2012. [58] J. Sikora and J. Selby, “Simple proof of the impossibility of bit commitment in generalized probabilistic theories using cone programming,” Physical Review A, vol. 97, no. 4, p. 042302, 2018. [59] J. H. Selby and J. Sikora, “How to make unforgeable money in generalised probabilistic theories,” Quantum, vol. 2, p. 103, 2018. [60] J. Sikora and J. H. Selby, “On the impossibility of coin-flipping in generalized probabilistic theories via discretizations of semi-infinite programs,” arXiv preprint arXiv:1901.04876, 2019. [61] H. Barnum, O. C. Dahlsten, M. Leifer, and B. Toner, “Nonclassicality without entanglement enables bit commitment,” in Information Theory Workshop, 2008. ITW’08. IEEE, pp. 386–390, IEEE, 2008. [62] L. Lami, C. Palazuelos, and A. Winter, “Ultimate data hiding in quantum mechanics and beyond,” Communications in Mathematical Physics, vol. 361, no. 2, pp. 661–708, 2018. [63] J. Barrett, L. Hardy, and A. Kent, “No signaling and quantum key distribution,” Physical Review Letters, vol. 95, no. 1, p. 010503, 2005. [64] D. Schmid, J. H. Selby, M. F. Pusey, and R. W. Spekkens, “A structure theorem for generalized- noncontextual ontological models,” arXiv preprint arXiv:2005.07161, 2020. [65] D. Schmid, J. Selby, E. Wolfe, R. Kunjwal, and R. W. Spekkens, “The characterization of noncontextuality in the framework of generalized probabilistic theories,” arXiv preprint arXiv:1911.10386, 2019. [66] F. Shahandeh, “Contextuality of general probabilistic theories and the power of a single resource,” arXiv preprint arXiv:1911.11059, 2019. [67] G. Chiribella and X. Yuan, “Measurement sharpness cuts nonlocality and contextuality in every physical theory,” arXiv preprint arXiv:1404.3348, 2014. [68] J. Bae, D.-G. Kim, and L.-C. Kwek, “Structure of optimal state discrimination in generalized probabilistic theories,” Entropy, vol. 18, no. 2, p. 39, 2016. [69] H. Barnum and A. Wilce, “Information processing in convex operational theories,” Electronic Notes in Theoretical Computer Science, vol. 270, no. 1, pp. 3–15, 2011. [70] J. Barrett, “Information processing in generalized probabilistic theories,” Physical Review A, vol. 75, no. 3, p. 032304, 2007. [71] A. Jenˇcováand M. Plávala, “Conditions on the existence of maximally incompatible two-outcome measurements in general probabilistic theory,” Physical Review A, vol. 96, p. 022113, 2017. [72] H. Barnum, J. Barrett, M. Leifer, and A. Wilce, “Generalized no-broadcasting theorem,” Physical Review letters, vol. 99, no. 24, p. 240501, 2007. [73] H. Barnum, J. Barrett, M. Leifer, and A. Wilce, “Teleportation in general probabilistic theories,” in Proceedings of Symposia in Applied Mathematics, vol. 71, pp. 25–48, 2012. [74] H. Barnum, C. P. Gaebler, and A. Wilce, “Ensemble steering, weak self-duality, and the structure of probabilistic theories,” Foundations of Physics, vol. 43, no. 12, pp. 1411–1427, 2013. [75] T. Heinosaari, L. Leppäjärvi,and M. Plávala, “No-free-information principle in general probabilistic 34

theories,” Quantum, vol. 3, p. 157, 2019. [76] L. Czekaj, M. Horodecki, and T. Tylec, “Bell measurement ruling out supraquantum correlations,” Phys. Rev. A, vol. 98, p. 032117, Sep 2018. [77] H. Barnum, S. Beigi, S. Boixo, M. B. Elliott, and S. Wehner, “Local quantum measurement and no-signaling imply quantum correlations,” Physical Review Letters, vol. 104, no. 14, p. 140401, 2010. [78] L.Czekaj, A. B. Sainz, J. Selby, and M. Horodecki, “Correlations constrained by composite measurements,” arXiv preprint arXiv:2009.04994, 2020. [79] J. Henson, R. Lal, and M. F. Pusey, “Theory-independent limits on correlations from generalized bayesian networks,” New Journal of Physics, vol. 16, no. 11, p. 113043, 2014. [80] M. Weilenmann and R. Colbeck, “Analysing causal structures in generalised probabilistic theories,” Quantum, vol. 4, p. 236, 2020. [81] L. Lami, “Non-classical correlations in quantum mechanics and beyond,” arXiv preprint arXiv:1803.02902, 2018. [82] P. J. Cavalcanti, J. H. Selby, J. Sikora, T. D. Galley, and A. B. Sainz, “Witworld: A generalised probabilistic theory featuring post-quantum steering,” arXiv preprint arXiv:2102.06581, 2021. [83] L. Masanes, T. D. Galley, and M. P. Müller,“The measurement postulates of quantum mechanics are operationally redundant,” Nature Communications, vol. 10, no. 1, pp. 1–6, 2019. [84] T. D. Galley and L. Masanes, “Classification of all alternatives to the Born rule in terms of informational properties,” Quantum, vol. 1, p. 15, 2017. [85] T. D. Galley and L. Masanes, “Any modification of the Born rule leads to a violation of the purification and local tomography principles,” Quantum, vol. 2, p. 104, 2018. [86] T. D. Galley and L. Masanes, “How dynamics constrains probabilities in general probabilistic theories,” arXiv preprint arXiv:2002.05088, 2020. [87] G. Chiribella, G. M. D’Ariano, and P. Perinotti, “Informational derivation of quantum theory,” Phys- ical Review A, vol. 84, no. 1, p. 012311, 2011. [88] G. Chiribella, “Dilation of states and processes in operational-probabilistic theories,” arXiv preprint arXiv:1412.8539, 2014. [89] G. Chiribella, “Distinguishability and copiability of programs in general process theories,” arXiv preprint arXiv:1411.3035, 2014. [90] H. Barnum and A. Wilce, “Local tomography and the jordan structure of quantum theory,” Founda- tions of Physics, vol. 44, no. 2, pp. 192–212, 2014. [91] A. Wilce, “Four and a half axioms for finite dimensional quantum mechanics,” arXiv preprint arXiv:0912.5530, 2009. [92] A. Wilce, “A royal road to quantum theory (or thereabouts),” Entropy, vol. 20, no. 4, p. 227, 2018. [93] H. Barnum, M. A. Graydon, and A. Wilce, “Composites and categories of Euclidean Jordan algebras,” arXiv preprint arXiv:1606.09331, 2016. [94] H. Barnum, R. Duncan, and A. Wilce, “Symmetry, compact closure and dagger compactness for categories of convex operational models,” Journal of Philosophical Logic, vol. 42, no. 3, pp. 501–523, 2013. [95] A. Wilce, “Symmetry and composition in probabilistic theories,” Electronic Notes in Theoretical Com- puter Science, vol. 270, no. 2, pp. 191–207, 2011. [96] L. Masanes and M. P. Müller,“A derivation of quantum theory from physical requirements,” New Journal of Physics, vol. 13, no. 6, p. 063001, 2011. [97] L. Masanes, M. P. Müller,R. Augusiak, and D. Pérez-Garc´ıa,“Existence of an information unit as a postulate of quantum theory,” Proceedings of the National Academy of Sciences, vol. 110, no. 41, pp. 16373–16377, 2013. [98] M. P. Müllerand L. Masanes, “Three-dimensionality of space and the quantum bit: an information- theoretic approach,” New Journal of Physics, vol. 15, no. 5, p. 053040, 2013. [99] L. Hardy, “Quantum theory from five reasonable axioms,” arXiv preprint quant-ph/0101012, 2001. [100] M. P. Müller,“Probabilistic Theories and Reconstructions of Quantum Theory (Les Houches 2019 lecture notes),” arXiv preprint arXiv:2011.01286, 2020. [101] G. M. D’Ariano, P. Perinotti, and A. Tosini, “Information and disturbance in operational probabilistic theories,” arXiv preprint arXiv:1907.07043, 2019. [102] G. Chiribella, G. M. D’Ariano, and P. Perinotti, “Probabilistic theories with purification,” Physical Review A, vol. 81, no. 6, p. 062348, 2010. 35

[103] J. Hefford and S. Gogioso, “Hyper-decoherence in density hypercubes,” arXiv preprint arXiv:2003.08318, 2020. [104] J. H. Selby, P. Cavalcanti, and A. B. Sainz, “Extended boxworld: a generalised probabilistic theory for type independent common cause resources,” forthcoming. [105] L. Hardy, “Reformulating and reconstructing quantum theory,” arXiv preprint arXiv:1104.2066, 2011. [106] L. Hardy and W. K. Wootters, “Limited holism and real-vector-space quantum theory,” Foundations of Physics, vol. 42, no. 3, pp. 454–473, 2012. [107] G. M. D’Ariano, F. Manessi, P. Perinotti, and A. Tosini, “Fermionic computation is non-local tomographic and violates monogamy of entanglement,” EPL (Europhysics Letters), vol. 107, no. 2, p. 20009, 2014. [108] G. M. D’Ariano, M. Erba, and P. Perinotti, “Classical theories with entanglement,” Physical Review A, vol. 101, no. 4, p. 042118, 2020. [109] E. G. Beltrametti and S. Bugajski, “A classical extension of quantum mechanics,” Journal of Physics A: Mathematical and General, vol. 28, pp. 3329–3343, jun 1995. [110] M. Reginatto and M. J. W. Hall, “Quantum-classical interactions and measurement: a consistent description using statistical ensembles on configuration space,” Journal of Physics: Conference Series, vol. 174, p. 012038, jun 2009. [111] M. Hall and M. Reginatto, Ensembles on Configuration Space: Classical, Quantum, and Beyond. Fundamental Theories of Physics, Springer International Publishing, 2016. [112] M. J. W. Hall, M. Reginatto, and C. M. Savage, “Nonlocal signaling in the configuration space model of quantum-classical interactions,” Phys. Rev. A, vol. 86, p. 054101, Nov 2012. [113] R. W. Spekkens, “Contextuality for preparations, transformations, and unsharp measurements,” Phys- ical Review A, vol. 71, no. 5, p. 052108, 2005. [114] S. Gogioso and F. Genovese, “Infinite-dimensional categorical quantum mechanics,” arXiv preprint arXiv:1605.04305, 2016. [115] B. Coecke, B. Edwards, and R. W. Spekkens, “Phase groups and the origin of non-locality for qubits,” Electronic Notes in Theoretical Computer Science, vol. 270, no. 2, pp. 15–36, 2011. [116] B. Coecke, “Terminality implies no-signalling... and much more than that,” New Generation Comput- ing, vol. 34, no. 1-2, pp. 69–85, 2016. [117] A. Kissinger, M. Hoban, and B. Coecke, “Equivalence of relativistic causal structure and process terminality,” arXiv preprint arXiv:1708.04118, 2017. [118] B. Coecke, J. Selby, and S. Tull, “Two roads to classicality,” arXiv preprint arXiv:1701.07400, 2017. [119] J. H. Selby, C. M. Scandolo, and B. Coecke, “Reconstructing quantum theory from diagrammatic postulates,” arXiv preprint arXiv:1802.00367, 2018. [120] G. M. D’Ariano, M. Erba, and P. Perinotti, “Classicality without local discriminability: Decoupling entanglement and complementarity,” Physical Review A, vol. 102, no. 5, p. 052216, 2020. [121] P. Janotta and H. Hinrichsen, “Generalized probability theories: what determines the structure of quantum theory?,” Journal of Physics A: Mathematical and Theoretical, vol. 47, no. 32, p. 323001, 2014. [122] A. S. Holevo, Probabilistic and statistical aspects of quantum theory, vol. 1. Springer Science & Business Media, 2011. [123] R. W. Spekkens, “The ontological identity of empirical indiscernibles: Leibniz’s methodological principle and its significance in the work of Einstein,” arXiv preprint arXiv:1909.04628, 2019. [124] R. W. Spekkens, “Evidence for the epistemic view of quantum states: A toy theory,” Physical Review A, vol. 75, no. 3, p. 032110, 2007. [125] W. G. Unruh, “False loss of coherence,” in Relativistic quantum measurement and decoherence, pp. 125–140, Springer, 2000. [126] F. Borceux and D. Dejean, “Cauchy completion in category theory,” Cahiers de Topologie et Géométrie Différentielle Catégoriques, vol. 27, no. 2, pp. 133–146, 1986.