Discovering Quantum Causal Models

Sally Shrapnel

December 28, 2015

I present an Interventionist account of quantum causation, based on the process matrix

formalism of Oreshkov, Costa and Brukner (2012), and more recent work by Costa and

Shrapnel (2015). The formalism generalises the classical methods of Pearl (2000), and

allows for the discovery of quantum causal structure. I show that classical causal

structure emerges in certain situations as a special case. I emphasise the crucial role

causal discovery plays, in order to distinguish this approach from other recent

alternatives.

Contents

1 Introduction. 3

2 Motivating quantum causal models. 6

3 What makes a classical causal model causal? 7

1



4 Quantum causal models: a new formalism. 15

5 Discovery and explanation: Bell inequalities. 26

6 The emergence of classical causal structure. 27

7 Conclusions. 29

2



1 Introduction.

It is notoriously di�cult to apply classical causal modelling techniques to systems

involving quantum phenomena. Entanglement correlations seem to defy causal

explanation, and it seems near impossible to produce an account that avoids the

di�culties posed by non-locality, contextuality, and the measurement problem. In the

last thirty years or so, philosophers have advocated a number of fixes, arguing for

non-local common causes (Suárez and San Pedro 2011; Egg and Esfeld 2014),

non-screening o↵ common causes (Butterfield 1992) and “uncommon” common causes

(Hofer-Szabó et al. (2013), Naeger (2015). Others have argued for more exotic solutions

such as retrocausation (Evans, Price and Wharton (2013); Evans (2014)) and

superdeterminism (’t Hooft 2009).

All of these approaches share a common theme. Roughly speaking, one starts with

quantum correlations, applies classical causal modelling methods, identifies a

contradiction and then decides what has to go. For interventionist accounts, the

dilemma is presented as a choice between relinquishing one of two assumptions: the

Causal Markov Condition or Faithfulness (no-fine-tuning)1. Whilst these accounts have

much to commend them, unfortunately, they have not led to any kind of consensus. It

seems reasonable to advocate a new direction.

The primary motivation for this new direction comes from quantum engineering.

Physicists have been producing technologies that harness quantum e↵ects for decades

now. Control and manipulation of quantum systems is commonplace in the laboratory,

and physicists express the relata and relations that comprise each specific design via the

use of quantum circuit diagrams. I think it entirely reasonable to suggest that these

1See in particular Glymore (2006) and Naeger (2015).

3



diagrams may represent causal structure. They are acyclic and align with temporal

direction. They allow physicists to predict the e↵ect of possible interventions and also to

di↵erentiate between e↵ective and ine↵ective strategies. On balance then, there is much

to suggest that if one were looking for an interventionist notion of quantum causation,

quantum circuit diagrams of engineered systems would be a good place to start.

A number of physicists have recently taken this approach towards causation and

produced a generalised notion of quantum circuits.2 Such “circuit models” now pervade

the physics literature on quantum causation and the framework I introduce in this

paper follows in this same tradition. The starting point for these accounts is not the

classical causal modelling formalism we are used to from the philosophical literature.

Rather, the starting point is simply the formalism of quantum mechanics, defined in a

suitably operational manner. We will take a closer look at these circuit models below.

I think this approach to quantum causation represents an important alternative to

traditional methods. As we shall see, it provides a new way of analysing the causal

structure of the Bell experiments. As mentioned above, the search for causal

explanations for Bell violating correlations has inspired much hair pulling on the part of

both philosophers and physicists in the past: here I will present a fresh perspective on

this old problem.

One of the founding fathers of classical causal modelling, Judea Pearl, is well aware

that his formalism cannot be applied to situations involving quantum mechanics:

... the Laplacian [deterministic] conception is more in tune with human
intuitions. The few esoteric quantum experiments that conflict with the

2See, for example, (Gutoski and Watrous (2006), Chiribella, D’Ariano and Perinotti

(2009), Chaves, Majenz and Gross (2015)).

4



predictions of the Laplacian conception evoke surprise and disbelief, and they
demand scientists give up deeply entrenched intuitions about locality and
causality. Our objective is to preserve, explicate and satisfy - not destroy - those
intuitions.(2000), [26]

He chooses to defend his account against various quantum qualms by claiming his

causal structure applies only to macroscopic relata:

Only quantum mechanical phenomena exhibit associations that cannot be
attributed to latent variables, and it would be considered a scientific miracle if
anyone were able to discover such peculiar associations in the macroscopic world
(2000), [62]

Whilst one can quibble over what ought to count as “macroscopic”,

experimentalists have pushed hard against the claim that there is some fundamental

upper limit at which quantum mechanics simply ceases to apply (see, for example,

Eibenberger et al., (2013)). One can no longer sensibly claim that quantum mechanics

applies just to the microscopic world. Nor can one sensibly hold that descriptions,

explanations and models involving quantum systems are simply bereft of causal

content3.

A flurry of recent papers capture something of this new perspective of quantum

causation. Bell experiments have been analysed using the causal modelling framework

(Wood and Spekkens (2014), Naeger (2015)), dualist versions of causal models have

been developed to capture both quantum and classical correlations (Laskey (2007),

Henson, Lal and Pusey (2014), Pienaar and Brukner (2014)) and new generalised

frameworks suggest quantum causal structure may be fundamental, with classical

causal structure emerging as a special case (Oreshkov and Giarmatzi (2015), Costa and

3See, for example, Shrapnel (2014) and Naeger (2015).

5



Shrapnel (2015)). It is the latter framework that I shall choose to explicate here.

Whilst all accounts are worthy of exploration, I believe the latter account comes closest

to the new direction I am advocating. The aim is to provide a consistent formalism

that allows for representation of quantum causal structure, with classical causal

structure emerging under certain (decoherent) situations.

The paper is structured as follows: first I motivate the project. Second, I examine

exactly what makes Pearl’s classical causal models causal. Third, I introduce the

quantum causal modelling framework of Costa and Shrapnel. Fourth, I explain how

causal discovery and explanation work in this new framework. Finally, I show how one

can recover classical causal structures as a special case. I do not wish to suggest that

this is the only possible formulation of a quantum causal model. This is a very new

field, and undoubtedly there is much more to be said.

2 Motivating quantum causal models.

I see three distinct reasons to be interested in the proposal I present here. Firstly, the

interventionist account of causation was largely inspired by casual modelling in

engineering and the special sciences. As Jenann Ismael puts it, in her usual succinct

manner:

Intuitions play almost no role in this [interventionist] literature. The emphasis
there is on providing a framework for representing causal relations in science, i.e.
a formal apparatus for rendering the deep causal structure of situations...and
provides normative solutions to causal inference and judgement problems.

(Ismael 2015)[2]

For the interventionist then, the existence of quantum engineered systems should

immediately alert one to the possibility of causal structure. To design such systems it

6



must be possible to distinguish between e↵ective and ine↵ective strategies.

The second motivation for the proposal comes from other recent philosophical work

characterising possible quantum causal models. Naeger (2015) and Evans (2015) both

advance quantum causal models that take Pearl’s characterisation to be correct, but

relax one of the key assumptions, allowing for models with fine-tuned variables. Whilst

I think both these accounts are worthy contributions, there is still a concern that they

may be, to some extent, throwing the baby out with the bathwater. Causal structures

are discovered rather than merely specified. They don’t arise in a conceptual vacuum,

but rather evolve in a co-evolutionary manner. As a scientific domain develops and

matures, causal explanation and causal discovery typically become two sides of the

same coin. Allowing for fine-tuned models, even though the fine-tuning is of a

particular kind, disturbs the fine balance between discovery and explanation.

I make this point clearer in the following section. For now, just think of it as

reason to look to alternative quantum causal models. The alternative I present here will

not try to shoe-horn quantum causal relations into classical ones. Instead, quantum

causal structure is seen as fundamental, with classical structure being recovered in an

appropriate limit.

Finally, one can also motivate the task of defining quantum causal models as

simply providing a new method for gaining some traction on another well-known

question: how does the quantum world di↵er from the classical?

3 What makes a classical causal model causal?

In virtue of what, exactly, is a causal model a representation of causal structure? For

many interventionists it is primarily a pragmatic matter. Causal models enable one to

identify e↵ective strategies by distinguishing between probabilistic correlations that are

7



due to causes and those that are merely accidental. On this view, causal models are

vehicles for learning about the manipulable elements of the world. It is therefore

relevant to ask whether causal models are purely of epistemological significance. Many

think not. The concept of interventionist causation (at the hands of Woodward and

others) has matured beyond an agent-centric approach to one that can be articulated

without explicit reference to human action or knowledge. The formal structures of

causal modelling are increasingly viewed as devices that give us a handle on the

objective, causal structure of the world.4 For those who prefer a modest, less

metaphysically loaded interpretation of causal modelling (see, for example, Woodward

(2003) ,Woodward (2007) and Frisch (2014)) the models I discuss here will still be of

interest.

A number of authors have contributed in important ways to the development of

causal modelling, for example Spirtes et al. (2000) and Woodward (2003), but we will

follow the physicists here and use Pearl’s account to focus our discussion. I will spend a

little time here explaining the detail of Pearl’s account. This will help the reader

appreciate the problems one encounters when using these methods to characterise

quantum causal structure. It will also pave the way for a deeper understanding of the

framework presented to account for quantum causal structure.

The causal relata of Pearl’s account are classical random variables X1, .., Xn. It is

assumed that each variable can be associated with a range of “values”: properties that

we can unambiguously reveal by measurement or direct observation. Such variables can

be binary and used to represent the occurrence or otherwise of an event, can take on a

4See, for example, Ismael (2015). See Price (2007) for an alternative, agent based

view.

8



finite range of values or have values that are continuous. It is generally assumed that

the properties that these values represent are non-contextual (in the sense of quantum

contextuality) and exist prior to, and independently of the act of measurement or

observation. Ultimately, these values represent the point of contact between the model

and the world.

The causal model is formed using an ordered triple hV, G, Pri. Here V is the set of
variables, G a directed graph and Pr a joint probability distribution over the variables

V . The graph captures the qualitative relationships between the variables, with the

nodes of the graph being the variables in V and the arrows between them representing

the causal dependencies. Graphs are usually acyclic, in the sense that no path is closed

to form a loop; causes cannot be their own e↵ects. Some basic terminology will be

useful: a variable A is a ‘parent’ of B when there is a single arrow from A to B. In such

a situation B is a ‘child’ of A. A is an ‘ancestor’ of B when there is a ‘directed path’ of

several linked arrows from A to B, in such a case B is a ‘descendant’ of A. The relation

of parent to child node, characterised by an arrow, is assumed to represent direct

causation.

The probability measure P is defined over propositions such as A = a, where A is a

variable in V with possible value a. P is also defined over conjunctions, disjunctions,

and negations of such propositions. Conditional probabilities over such propositions are

also consequently well-defined whenever the event being conditioned on is well-defined

(Hitchcock 2012).

Pearl was initially motivated by causal discovery rather than explanation. His

algorithms were designed to discover causal structure from within probabilistic

empirical data sets; in many cases a dauntingly complex task. Imagine one has a table

of statistical data pertaining to a set containing n variables. If each variable has k

9



possible values, then to exactly specify a probability distribution over all possible

combinations of values in the model, one needs kn parameters. Pearl identified that a

Bayesian Network provides a possibility for representing such a joint distribution in a

more compact form, by identifying conditional independencies and dependencies.

Pearl’s causal models are essentially Bayesian Networks with some additional

assumptions, known as the Causal Markov Condition and Faithfulness. A Bayesian

Network is a directed acyclic graph (DAG), G, where each node Xi has an associated

conditional probability distribution that denotes the dependence of the values of Xi on

its parents in the graph, P(Xi|PaG(Xi)). This network represents a joint probability
distribution via the rule P(X1, ..Xn) =

Q
i P(Xi|PaG(Xi)). In words, the joint

distribution (P), taken over all the variables in the model (V ), decomposes over the

graph (G) into a collection of smaller conditional probability distributions. This

factorisation is crucial to the formal manipulations of the graph that make various

causal and statistical inferences possible.

For such probabilistic BN models, directed edges need not convey causal meaning,

which is why one needs to assume the Causal Markov Condition and Faithfulness. In

Pearl’s own words:

. . . behind every causal conclusion there must lie some causal assumption that is
not discernible from the original distribution.” (loc 10189).

An easy way to see this is to consider that there are certain aspects of DAGs that

are unidentifiable from observational data alone. Several graph structures can return

the same probabilistic conditional independencies. For example:

X Y Z

10



X Y Z

X Y Z

all capture X?Z|Y . So if we start with purely observational data, and identify the
statistical relationship X?Z|Y , we can only construct an equivalence set of graphs.
Such graphs are said to be “independence equivalent”, I-equivalent or “Markov

equivalent”. To di↵erentiate between such graphs we need to feed in other assumptions,

for example temporal or interventionist information. The important point is that for a

probabilistic graph (Bayesian Network) the three structures are equivalent, but for a

causal graph (Causal Bayesian network) we must have a means for disambiguating

between these possible structures. One can either perform interventions to get to the

correct factorisation, or assume statistical constraints.

The Causal Markov Condition states that for a graph G, each variable Xi is

independent of all its non-descendants, given its parents Pa(Xi) in G. One can think of

it as a generalisation of Reichenbach’s screening o↵ criteria. That is, it allows the

inference from connectedness in the graph to causal dependence. Faithfulness

(alternatively stability or no fine-tuning) states that the only conditional

independencies in the distribution P are the independencies that hold for any set of

causal parameters D. Another way to put this is that all the independence relations in

the probability distribution over the variables in V must be a consequence of the

Markov condition. The idea here is that one does not wish to allow for “accidental”

independencies that are created when causal paths cancel. Faithfulness licences the

inference from unconnectedness in the graph to causal independence.

One good reason for disallowing fine-tuned models, is that otherwise it becomes

possible to trivially associate any probability distribution with a Markovian graph by

11



ensuring the graph is complete. This means each node in the graph is connected to all

other nodes in the graph. In such a situation, causal discovery becomes impossible.

Many have argued against this constraint, claiming it is too strong. There are

many cases of models of physical systems that appear fine-tuned, despite the fact that

intuitively we still wish to call them causal (Cartwright (2001), Andersen (2013)).

There is a subtlety that somewhat mitigates this concern however. Recall that causal

models are relative to a number of pragmatic choices: the model scope, the range of

invariants (background variables we exclude from the model) and the level of detail.5

For any fine-tuned model it will be possible to recover a faithful (stable) model by

either changing scope, increasing the level of detail or altering the range of invariance

(assuming Pearl’s interpretation below is correct). This is also the case with

Markovianicity: one can fine-grain variable values, increase the scope of the model and

look for correlated noise to restore an otherwise recalcitrant causal model to obey the

Causal Markov condition. Thus, in classical modelling situations, the Markov condition

and Faithfulness guide us toward the most e�cacious level to express the causal

structure of a given system and play an indispensable part in causal discovery.6

Pearl suggests two physical facts underpin his claim that BNs that are Markovian

5Invariants is Ismael’s term (2015). The range of invariants is simply the range of

values that the background variables can assume such that the causal relations of the

model still hold. Woodward uses the term stability for this concept.

6In some sense, this is part of the real power of Bell’s theorem. Even if one tries all

these options, the statistics one arrives at will nonetheless still be able to violate the

Bell inequality. That is, adding local hidden variables can not alleviate the need for

fine-tuning to account for the correlations.

12



and Faithful can be interpreted to represent causal structure. Firstly, he assumes the

local conditional probability distributions (between parent nodes and their respective

child node) encode information about local (in time and space) stochastic mechanisms,

where the stochasticity is due to our ignorance about other variables that we do not

include in the model. This means that the relationships between the values of parent

and child nodes can be expressed as a function Xi = fi(paXi, Ui), where the Ui

represent all the unmodelled influences on the variable values (noise). Crucially, Pearl

assumes the Ui to be independent, thus guaranteeing the Causal Markov Condition will

hold. On this interpretation the model is probabilistic by virtue of our ignorance about

the values of the Ui: the distribution over the Ui generates the probability distribution

over the model.

The second physical fact that Pearl appeals to states that the act of setting a

variable to a particular value (an intervention) can deterministically override the

natural causal mechanisms of the model, providing us with new information by

disrupting only the local mechanism associated with that node. The intervention is

assumed to replace the original causal mechanism with one that determines the child

variable value X = x with probability 1. Arrows into the intervened variable are broken,

and a new probability distribution is associated with the altered graph. Remember, the

joint distribution of the undisturbed graph is P(X1, ..Xn) =
Q

i P(Xi|PaG(Xi)), where
the product on the right is over all the nodes in the graph. The joint distribution for

the altered graph, for example when X3 is set to x, is

P(X1, ..Xn|do(X3 = x) = p(X1,...Xn)p(X3|Pa(X3). Note, it is the factorisation of the joint probability
into local conditional probabilities between parent and child nodes that ensures this

manipulation of the graph accurately represents the e↵ect of a localised intervention.

Pearl claims his framework rests on a hierarchy of assumptions, statistical then

13



interventionist. The basic axioms of probability theory (positivity, normalisation and

additivity) and the causal assumptions of Markovianicity and Faithfulness provide the

possibility of testing the model via local interventions.

Pearl has developed two well known algorithms that take as input a list of

conditional independencies (found in a joint distribution over a given set of variables)

and return a set of DAG’s as output. The IC algorithm will return a DAG, under the

assumptions of Causal Markovianicity, Faithfulness and Causal Su�ciency (no

unmeasured common causes). The IC* algorithm does not require the assumption of

causal su�ciency, but will in general only return a partial ancestral graph (PAG): a

DAG with any number of undirected edges. These algorithms use observational data

only, so the underdetermination of the causal structure can be further reduced if one

has access to interventionist data7.

Let us take a step back, and look at two general features of this formalism. Firstly,

what are the points of contact between such causal models and the world? Roughly

speaking there are two: the data that underlies the variable “values”, and the data that

we use to characterise the local interventions. It is worth noting here that both kinds of

data are not explicitly included in the final model: it is the axioms of probability theory

that get us from the data to the final model.

Secondly, and more importantly (for our purposes at least), is the fact that the

application of causal modelling techniques mirrors the iterative manner in which science

develops. One does not arrive at a causal model in a conceptual vacuum: typically one

does not either just “start with the data”, so to speak, and produce a fully fledged

causal model. Nor does one start with a plausible causal structure, according to domain

7Although see Eberhardt (2013) for some interesting problem cases.

14



knowledge and theory, and expect it to perfectly match an empirically derived joint

distribution. Rather, the process is one of mutual refinement, where model and theory

are developed in a co-evolutionary manner. Pearl’s techniques capture this feature of

causal modelling in a mathematically consistent manner. This is also the case for the

modelling strategies utilised by Scheines, Glymore and Spirtes (2000).

Recall, a fine-tuned, or non-Markovian model will often imply we have not settled

on the best representational strategy for the phenomena in question. Rather than signal

a lack of causation, such failures can instead be seen as an indication that more work

needs to be done. Searching for hidden variables, fine-graining measurement values,

analysing noise models and improving intervention control are all viable options for

arriving at an improved model. When the set of variables is expanded, causal relations

can come into existence. It is the fit between causal discovery and causal explanation

that drives the possibility of such an iterative approach. Valuable work is being done in

the field of causal modelling to refine and extend the reach of such techniques.8

For all these reasons, classical causal modelling techniques now take centre stage in

much of the current philosophical work involving causation. There is just one problem.

They don’t work for systems involving quantum phenomena.

4 Quantum causal models: a new formalism.

It is clear that there are substantial di�culties to overcome in order to clarify a possible

candidate for a quantum causal model. The relata of quantum causal models are not

going to be variables in the usual classical sense. One can not associate measurement

outcomes with properties that exist prior to, and independently of the act of

8See, for example, Zhang and Spirtes (2015).

15



measurement or observation. The Kochen-Spekker and Bell theorems both confirm the

impossibility of ascribing the usual local, non-contextual, classical hidden variables to

represent quantum causal structure. This means one can not simply enlarge the

variable set, or fine-grain the values, and recover a Markovian, Faithful causal DAG9.

The first change we shall consider then, is the nature of the causal variables.

For classical causal models, one can think of a variable as capturing a possibility

space for a collection of spatio-temporally located events. For the quantum case, we can

generalise this notion to include events characterised by possible operations inside

specific space-time regions.10 Intuitively, one can think of a quantum causal model as a

structure that represents modal information, where the information tells us how what

can happen in one space-time region depends on what can happen in another. Such

9See Wood and Spekkens (2014) for a lengthy analysis of Bell experiments using

Pearl’s classical causal methods. The upshot is that one must allow for fine-tuning to

produce a causal model that explains such correlations. For Wood and Spekkens (as for

myself) the cost is too great: one loses the ability to discover causal structure.

10The process matrix framework that follows was developed by Oreshkov, Costa and

Brukner (2012), although they did not identify quantum causal models in the sense

reproduced here. Rather they were interested in developing a formalism that could

describe the possibility of indefinite causal structure. This framework has been adopted

and modified by Costa and Shrapnel (2015) as a possible candidate for producing

quantum causal models. The issue of indefinite causal structure, whilst interesting, will

not be addressed here. In the present work I assume that operations take place in a

fixed space-time background with definite causal order.

16



modal information is accessed by considering the possibility of (i) local choices of

operations associated with each space-time region, and (ii) physical systems passing

information between each space-time region. The nodes correspond to the space-time

regions where we can perform our interventions, and the edges correspond to the

transferring physical systems that carry information from one region to another. As

such, the connecting physical systems play a similar role as the local mechanisms of

Pearl’s causal models, that determine the functional relationship between parent and

child nodes.11

Circuit diagrams are a representation of just this kind of structure. There are two

distinct varieties of diagram we need to distinguish here. The first is the familiar kind,

as depicted in Neilsen and Chuang’s famous textbook (Nielsen and Chuang 2000).

Wires are associated with quantum systems. The state of these systems can be

manipulated into superposition states: logic gates represent a physical interaction that

changes the state of the quantum system in a predictable manner.

X | iout| iin

As the system passes between the gates, it is assumed that the state does not

change. Thus the wires in these diagrams represent the identity map, 1. The gates are

assumed to e↵ect a unitary evolution of the state, and can be switched on and o↵ at

various times (in the diagram, the box marked “X”). Measurements are typically

pushed to the far right of the circuit, and represent the outcomes of the computation.

The second kind of circuit is a generalisation of this basic structure, to include a

variety of di↵erent possible experimental situations in a more abstract manner. The

11In the language of quantum information, an edge corresponds to a quantum channel

from the output space of the parent node to the input space of the child node.

17



wires still represent the quantum state of a physical system, but the state can now

change between the gates. This is reflected by associating the more general completely

positive map with the wire. Intermediate nodes no longer represent single possible

gates, but rather are place holders for a variety of possible interactions, including

non-deterministic ones (to allow for the possibility of measurements). It is this latter,

more general, kind of circuit that will form the basis for the quantum causal model

formalism. The details of these circuits will become clearer below as we examine the

mathematical structure in more detail and see some specific examples.

For now, the picture to keep in mind is simply one where nodes correspond to the

space-time regions where we perform our interventions, and the edges correspond to the

transferring physical systems that carry information from one region to another.

How does one go about attaching probabilities to such a structure? The

assignment of probabilities to quantum data is usually via the familiar Born Rule:

P(j) = Tr(Oj⇢), where Oj here is the relevant measurement operator and ⇢ the density

matrix. We shall need to generalise this rule, but for now it is enough to recall that this

will shape how we attach probabilities to objects in the model.

Typically, we associate measurements with operators, or more generally POVM’s.12

However, in many cases we wish to encode not just the outcome of a measurement, but

also the transformation to the state that occurs during the measurement process.

Mathematically, the most general object one can associate with such a quantum

operation is a trace-non-increasing completely positive map (CP map)13. We can

12Recall a POVM is a Positive Operator Valued Measure: a set of positive,

semi-definite matrices that sum to the identity.

13See Timpson (2013) pp. 255-257 for a good introduction to CP maps.

18



associate each region with sets of such CP maps. Ultimately, these CP maps will

perform the same formal function as the variable values in the classical case, they

provide us with the possibility space associated with each node. A particular set of CP

maps together can characterise a quantum instrument. Essentially an instrument

represents one out of a number of ways we can interact with the system. For example it

may represent a choice of measurement setting, basis, preparation etc. We shall see

that such instruments ultimately represent the possible interventions we can perform to

test the causal structure of the model.

The mathematical arena of the CP map is the Hilbert space we associate with the

space-time region. One can decompose this space into the tensor product of two

subsystems: one associated with the incoming system, HAI , and one with the outgoing
system, HAO. The CP map, denoted MAIAOj , is a linear map that sends states in the
input Hilbert space HAI to states in the output Hilbert space HAO. Each use of an
operation in a given space time region will be associated with one out of a possible set

of outcomes, and we label these j = 1, ...., n. We can thus consider that each outcome

induces a particular transformation in the state from input to output, and this is

captured by the CP map associated with that outcome. For example, for region A and

outcome j we denote the associated CP map as MAj : L(HAI ) ! L(HAO).

A

j k

HAO HBI
B

HAI HBO

In the diagram above, the nodes A and B represent distinct spatio-temporal

regions. The Hilbert space associated with these regions is defined by the tensor

product of the spaces associated with the incoming and outgoing systems, for example,

HAi ⌦ HAo for region A. The outcome of the operation occurring at A is represented as

19



j, and the outcome of the operation at B is represented as k. One can generalise to

cases that include multiple incoming and outgoing edges, as we shall see in the

examples below.

We can now consider how to associate probabilities with such operations. If a CP

map is performed on a quantum state ⇢, then MAj (⇢) describes the updated,
non-normalised state in the case that outcome j is observed. The probability to observe

this particular outcome is given by the Born Rule: P(MAj ) = Tr[MAj (⇢)]. If we assume
our list of possible outcomes (j = 1, ..., n) is complete, then the sum of all CP maps will

be trace preserving and completely positive. Another way to put this is that we are

guaranteed to have a legitimate probability distribution: with certainty at least one

outcome j will occur. Such a complete set of CP maps constitutes a particular

instrument.14

We now have a description that can link a probability distribution to a single node.

The next step is to consider how to extend this to provide a valid distribution over all

the nodes in the graph. First let us consider the simple case of two nodes, labelled A

and B, as in the diagram above. We assume that the joint probability, P(Mj, Mk), for
a pair of maps to be realised should be independent of the particular set of possible CP

14Whilst the definition of a quantum instrument is relatively intuitive in the case of

finite, discrete systems, it can become more complicated for the continuous case. See

(Davies and Lewis 1970) for a more formal definition of a quantum instrument that

captures both situations. The generalisation to the continuous case does not threaten

any of the arguments made here. Roughly, one can think of an instrument as a

generalisation of a POVM that also captures transformations to the sysytem: for a one

dimensional output space an instrument will reduce to a POVM.

20



maps associated with either A or B: the joint probabilities must reflect a kind of

non-contextuality of the CP maps. That is, what kind of instrument we choose to use

at A should not e↵ect the kind of instrument we choose to use at B.15 Characterising a

probability distribution over a pair of outcomes, j and k, that correspond to maps

P(Mj), P(Mk) is not entirely straightforward. We will need to make use of a little
mathematical trick. I give a simplified account in the footnote below, but see Oreshkov,

Costa and Brukner (2012) for the full story.16

The details of why we represent the CP maps in this form are not crucial for our

purposes, we just note that as a result, we can represent the probabilities for two

measurement outcomes in di↵erent space-time regions A and B, as a bilinear function

of the corresponding CJ operators:

P(MAj , MBk ) = Tr[W AIAOBIBO(MAIAOj ⌦ MBIBOk )] (1)

15This is essentially the no-signalling condition.

16The trick involves representing a CP map as particular kind of matrix using an

isomorphism called the Choi-jamiolkowski isomorphism (CJ). First consider that a CP

map can also be represented as a matrix on the tensor product Hilbert space,

Mj 2 (HAI ⌦ HAO), where I� Mj � 0. We denote the complete set of outcomes
j = (1....m) for a given operation as {MAIA0j }mj=1. Each element of this set is positive
and, by virtue of the tensor product structure, if we trace over the output space, we will

recover the identity on the input space. The CJ matrix corresponds to a particular

form of linear map, defined as MAIAOj :=
⇥
I ⌦ M(|�+ih�+|)

⇤T
, where |�+i is a

maximally entangled state whose basis states form an orthonormal basis for HAI , I is
the identity map and T indicates the transpose.

21



Where W AIAOBIBO is a matrix in L(HAI ⌦ HAO ⌦ HBI ⌦ HBO), known as the process
matrix. This matrix can be generated by taking the tensor product of all the maps

representing connections between the nodes, written in their CJ form. One can think of

equation 1 as a further generalisation of the Born Rule, with the process matrix

analogous to a quantum state and the object (MAIAOj ⌦ MBIBOk ) analogous to an
operator.

With respect to our causal model, the process matrix W encodes possible

connections between the two space time regions A, B. This is easiest to see by

considering a less trivial concrete example, involving three nodes and two wires.

Imagine one is given three boxes, labelled A, B and S, connected by wires, labelled (i)

and (ii). Each box comes equipped with a set of levers to represent possible

instruments, and a readout associated with possible outcomes. By gathering statistics

that relate outcome results with respective instrument settings, one can produce a W

matrix. The aim is to use this information to di↵erentiate between the two possible

situations depicted below.

A B

S

(i) (ii)

(1) Bipartite entangled state: common cause.

A B

S

(i) (ii)

22



(2) Direct cause structure.

Figure (1) depicts the familiar Bell scenario: the production of an entangled

bipartite state. We shall refer to this as the common cause structure. Figure (2) depicts

the direct transfer of a quantum physical system between three distinct locations: a

direct causal structure. The process matrix ought to be able to tell us which of the two

circuits has been implemented.

In order to depict these two structures as quantum causal models, we first need to

associate an input and output space for all three nodes. In this simple case we need an

input and output space for each of A and B (labelled HAI , HAO, HBI , HBO). For S we
need one input space (HSI ) and two output spaces (HSO1 , HSO2 )17.

First let us use these example networks to compose a process matrix for each case

from the terms in each example. For a known causal structure, one can compose the

process matrix from the tensor product of terms corresponding to the various

components of the network. For example, the process matrix for the common cause

structure will compose as (1) Wc = ⇢
SI ⌦ T SO1AI1 ⌦ T SO2BI2 ⌦ 1AOBO. Where ⇢SI is the

input state at S, T SO1AI1 is the CJ matrix representing the channel from S to A, T
SO2BI
2

is the CJ matrix representing the channel from S to B, and 1AOBO is the identity map

on the output states at A and B. In a similar manner, one can compose the process

matrix for the direct cause structure as (2) Wd = ⇢
AI ⌦ T AOSI1 ⌦ T SO1BI2 ⌦ ⇢SO2 ⌦ 1BO.

Here ⇢AI represents the input state at A, T SO1AI1 represents the connection from A to S,

T
SOIBI
2 represents the connection between S and B, ⇢

SO2 represents that component of

the output space of S that in this case we trace out, and 1BO is the identity on the

output of BO.

17See the diagram in Appendix A

23



These examples show that, generally speaking, it is easy enough to decompose the

process matrix into tensor products of various circuit components if we already know

the structure being described. The crucial question, of course, is whether one can go in

the other direction. Can one reconstruct the correct W matrix from the observed

statistics of outcomes at A, B and S? Furthermore, is it possible to di↵erentiate

between the two cases depicted and thus discover the causal structure of the network

from that data alone?

Interestingly, if one has enough statistical data one can construct a unique

(minimal) circuit from a given process matrix W . As with the classical case, if one

doesn’t require the minimality assumption, an equivalence class of models can be

recovered. So how much data is enough? This is related to the notion of

“informationally complete” sets of measurements (POVMS): a single quantum state can

be discovered correctly given such a set of measurements (Prugovecki 1977). One can

generalise the notion of informational completeness to extend to sets of CP maps and

situations that model multiple systems.18 As with the classical case, causal discovery in

this more general setting is intimately related to the characterisation of an ideal

intervention. We now turn to consider a plausible quantum analogue.

Interventions in such a model follow naturally from the fact that we associate nodes

with the space of possible local CP maps. An intervention can thus be characterised by

choosing a particular kind of possible local instrument from within the local space-time

region. For example, this may correspond to a choice of measurement setting, basis

18This is due to the possibility of process and state tomography (Nielsen and Chuang

2000), and can be formally extended to the case of process matrices (Costa and

Shrapnel 2015). See (Reid 2015) for a specific example of such ‘causal’ tomography.

24



choice, or preparation. This will define the specific sets of CP maps associated with the

node. To ensure that such interventions are, in the relevant sense, local, we can demand

that the interior of a region can interact with the rest of the universe only through the

region’s input and output spaces: it is closed to any tra�c not explicitly expressed via

edges in the model. This means particular kinds of interventions on a single space-time

region will arrow-breaking in a similar manner to the classical case. One can conceive of

freely choosing the particular kinds of instrument to use within a given region in a

manner that is statistically independent of any other, unconnected space-time region in

the model. As with the classical case, one can implement a randomisation procedure

here to provide an appropriately agent-independent notion of “free choice”.

Let us look at our specific example to make this connection to the classical idea of

intervention a little clearer. Imagine trying to di↵erentiate between the two cases we

mentioned above. One can distinguish between these two structures by performing an

arrow breaking intervention at A. An obvious example of such an intervention is the set

of CP maps that realises a state preparation at A.

How is it possible to di↵erentiate between these two structures using such an

intervention? Remember, the causal structure of these models, in general, identifies

e↵ective strategies for changing distributions over outcomes, rather than particular

outcomes themselves. Thus, one cannot intervene at A and produce a change in the

distribution over outcomes at B for the common cause structure in figure 2a.

Consequently, there is no direct causal relation between A and B. However, one can

intervene at S and make changes to the distributions over outcomes at both A and B,

validating the common cause structure depicted. Similarly, for the structure in figure

2b, one can intervene at A and e↵ect a change at both S and B. We give a more formal

treatment of this intuitive picture in Appendix A.

25



5 Discovery and explanation: Bell inequalities.

Recall Pearl’s quote from section 1:

... The few esoteric quantum experiments that conflict with the predictions of
the Laplacian conception evoke surprise and disbelief, and they demand
scientists give up deeply entrenched intuitions about locality and causality. Our
objective is to preserve, explicate and satisfy - not destroy - those
intuitions.(Pearl 2000)[26]

It seems that it is possible to preserve the spirit of these intuitions and explicate a

version of quantum causation that observes a well defined notion of both locality and

causality. The causal models presented here are local in the sense that they support

causal inference via local interventions and also in the sense that the physical systems

responsible for the connections are assumed to observe relativistic constraints. They are

causal because they allow us to di↵erentiate between e↵ective and ine↵ective strategies.

On this view, contra to the intuitions that so-called “steering” phenomena a↵ord, one

cannot intervene at A and produce a change at B for the common cause structure in

figure 2a. Thus, in this picture there is no direct causal relation between A and B, in

accordance with Woodward’s intuitions regarding Bell correlations (Woodward 2007)

and also in accordance with the no-signalling principle. However, one can intervene at

S and make changes at both A and B, validating the common cause structure depicted.

Interestingly, for these quantum causal structures there is also an intuitive notion

of fine-tuning. Consider once again the direct cause structure depicted in Figure 2. For

the case where the set of instruments realised at S were to range over a set that only

produced maximally entangled states, for example the four Bell states: |00i+|11ip
2

,

|00i�|11ip
2

, |01i+|10ip
2

, |01i�|10ip
2

. Recall, one output system will be traced over and one output

sent to the node at B (see Appendix). For such states, the input state at B will always

26



be maximally mixed. This means that regardless of the choice of instrument, one will

never see a causal relation between S and B. There is an independence in the data that

is not reflected in the true causal structure, just as in the classical case of fine-tuning.

As with the classical case, one can demand that it is an assumption of the framework

that one does not allow for such fine-tuned models.

6 The emergence of classical causal structure.

I believe one of the key virtues of this approach is that quantum causal structure is seen

as fundamental, with classical structure emerging as a special case. There are three

main reasons why I think this ought to be the case. Generally speaking, the most

prolific users of both classical and quantum theory - physicists - take it as given that

quantum theory is the more fundamental theory of the two (Copenhagenism aside).

Secondly, there is a well described dynamics that takes one from quantum structure to

classical structure under certain circumstances (decoherence)19. Finally, one can

provide examples of how the quantum models described in this paper can also apply to

classical structures as a special case.

A classical model can be characterised by considering sets of instruments that are

restricted to be diagonal in the product basis of the Hilbert space associated with each

node. Recall each node is associated with a total Hilbert space, generated via the

tensor product of the Hilbert spaces associated with each incoming and outgoing edge.

One can ask why we might be restricted to the use of only these particular instruments,

and here we can fall back on the usual arguments from decoherence theory. It is the

19See Zurek (2014) and Schlosshauer (2010) for an introduction to the foundational

implications of decoherence.

27



contingent features of the unknown environment, interacting with the instrument and

determining the pointer basis for that node, that so limits our choice of instrument.

Decoherence theory provides tools to explain how and when quantum probability

distributions change to classical probability distributions. Thus, it seems likely that we

can co-opt these techniques to help explain the relationship between quantum and

classical causal models. As yet, this feature of quantum causal models has not been

studied in any detail, so for now I will just give a flavour of what might be expected.

One of the important lessons of this field is that the quantum-classical boundary is

mobile: it can be shifted smoothly by varying particular experimental parameters. In

principle then, one ought to be able to model a gradual and continuous shift between a

quantum and classical causal model. It would seem that this could be done by

explicitly adding in another quantum node to account for a decohering environment.

On such a view, the structure and statistical constraints that we typically associate

with classical causal models are in fact a special subset of the deeper causal structures

that quantum correlations allow for. Recall Pearl’s picture of the physical story that

underlies causation: deterministic relations between events, characterised by local

mechanisms, with the probabilities entering by virtue of our ignorance of all the facts.

We face quite a di↵erent picture here. Typically, the quantum circuit model carries

with it the underlying assumption that evolution of states is unitary and probabilities

enter by virtue of fundamental indeterminism.20 One can, in principle, always recover a

quantum causal model from a classical one by expanding the variable set to include a

decohering environment. As such, the deeper physical interpretation of quantum causal

20This assumption is justified by the Stinespring Dilation theorem (Nielsen and

Chuang 2000).

28



models will be hostage to the same concerns as any interpretation of quantum theory.

Choosing to be a Many Worlder, a Collapsican, or a Bohmian will not invalidate

anything I have said here.

7 Conclusions.

Summing up, I have presented an interventionist account of quantum causation based

on the process matrix formalism of Costa and co-workers. I outlined three reasons to

consider this proposal: (i) by interventionist lights, the existence of engineered quantum

systems and the representational resources that accompany their explication and design

flag the possibility of quantum causal structure, (ii) current classical causal modelling

methods fail to produce a solution that retains the possibility of causal discovery, and

(iii) such a project may ultimately cast new light on the question of how the classical

and quantum worlds di↵er.

I then explored Pearl’s classical causal modelling framework in order to illustrate

how local interventions are used to verify putative causal structure and thus play a

crucial role in causal discovery. I suggested that it is the balance between discovery and

explanation that underpins the success of Pearl’s methods. As such, any quantum

causal modelling method ought to a↵ord the possibility of causal discovery. Next, I

introduced the process matrix formalism and gave an example of how it can be used to

discover causal structure. This example was also used to highlight the role that local

interventions play in this framework. Finally, I discussed how it is possible to recover

classical causal structure as a special case, and suggested that this aspect of the

formalism deserves closer attention.

29



Acknowledgements

I wish to thank Fabio Costa for inviting me to collaborate on this project. Thanks also

to Gerard Milburn for multiple enjoyable discussions regarding both classical and

quantum physics. Finally, thanks to Peter Evans for useful conversations regarding

various philosophical aspects of both classical and quantum causation.

30



Appendix A

To distinguish between the two particular causal structures depicted in Figure (1) and

Figure (2), one can limit the necessary instruments to simple state preparation and

measurement options. This will enable one to recover the correct process matrices for

each case. Whilst other possible instruments exist, they do not add any additional

useful information in this particular case. For clarity, we can alter the figure to include

all the required Hilbert spaces:

HAO

A B

S

HSI

(1) Bipartite entangled state: common cause.

A B

S

HAI

HsO2

(2) Direct cause structure.

In each case, we have one input and two output Hilbert spaces associated with

node S, one input and one output Hilbert space associated with node A, and one input

space associated with node B.

31



Let us label the measurement outcomes a, b and s respectively for each node, and

the choice of instrument x, y and r, respectively for each node. The appropriate

instruments to use at each node will be:

MSs|r = E
SI
s ⌦ ⇢

SO1
r ⌦ ⇢SO2r (2)

MAa|x = E
AI
a ⌦ ⇢AOx (3)

MBb|y = E
BI
b (4)

where, for example, ESIs is the POVM element corresponding to measurement outcome

s, and ⇢
SO1
r represents a state preparation at SO1 that depends on r. Recall, a POVM is

a set of positive semi-definite matrices that sum up to identity. The informationally

complete sets of POVM’s and preparations to use in this instance would be the six

projectors associated with the Pauli matrices. The projectors associated with both the

POVMs and preparation procedures would therefore be:

Ea1 = Eb1 = Es1 = 3E1 = |x+ihx+|

Ea2 = Eb2 = Es2 = 3E2 = |x�ihx�|

Ea3 = Eb3 = Es3 = 3E3 = |y+ihy+|

Ea4 = Eb4 = Es4 = 3E4 = |y�ihy�|

Ea5 = Eb5 = Es5 = 3E5 = |z+ihz+|

Ea6 = Eb6 = Es6 = 3E6 = |z�ihz�| (5)

32



⇢x1 = ⇢y1 = ⇢r1 = |x+ihx+|

⇢x2 = ⇢y2 = ⇢r2 = |x�ihx�|

⇢x3 = ⇢y3 = ⇢r3 = |y+ihy+|

⇢x4 = ⇢y4 = ⇢r4 = |y�ihy�|

⇢x5 = ⇢y5 = ⇢r5 = |z+ihz+|

⇢x6 = ⇢y6 = ⇢r6 = |z�ihz�|

(6)

Recall, the W matrix for the common cause and the direct cause structures will

decompose di↵erently as:

Wc = ⇢
SI ⌦ T SOAI1 ⌦ T

SO2BI
2 ⌦ 1AOBO (7)

Wd = ⇢
AI ⌦ T AOSI1 ⌦ T SO1BI2 ⌦ ⇢SO2 ⌦ 1BO (8)

If one has access to the above instruments at each node then one can recover the

terms T SOAI1 and T
SO2BI
2 from the gathered statistics. This means that one can calculate

the conditional probability over all the nodes according to the generalised Born rule:

From equation (1):

P(a, b, s|x, y, r) = Tr
⇥
W(MAI

a|x ⌦ MAIb|y ⌦ MBIs|r )
⇤

(9)

(10)

Using this rule, one can see that the observed statistics will correspond to either

the common cause or direct cause structure, but not both. This fits with our intuitions

regarding how the conditional probability P(a, b, s|x, y, r) ought to decompose in each

33



case. For the common cause P(a, b, s|x, y, r) = P(s)P(a|r)P(b|r). For the direct cause
P(a, b, s|x, y, r) = P(a)P(s|x)P(b|r). The direction of the channels tell us the
appropriate conditionals to use. Each of the conditional probabilities corresponds to the

probability for the child to detect a POVM element given the state of the parent,

transmitted through the corresponding channel.

References

Andersen, Holly. 2013. “When to Expect Violations of Faithfulness and Why it

Matters” Philosophy of Science 80(5): 672-683.

Butterfield, J. 1992 “David Lewis Meets John Bell.” Philosophy of Science 59(1).

Cartwright, N. and Jones, M. 1991. “How to hunt quantum causes.” Erkenntnis 35:

205-231.

Cartwright, N. 2001. “What Is Wrong with Bayes Nets?.” The Monist 84:242-264.

Cavalcanti, Eric G, and Raymond Lal. 2014. “On Modifications of Reichenbach’s

Principle of Common Cause in Light of Bell’s Theorem.” Journal of Physics A:

Mathematical and Theoretical 47(42):424018.

Chaves, R., Majenz, C. and Gross, D. 2015. “Information-theoretic implications of

quantum causal structures.” Nature Communications 6: 5766.

Chiribella, G., D’Ariano, G. M. and Perinotti, P. 2009. “Theoretical Framework for

Quantum Networks.” Physical Review A. 80: 022339.

Costa, F and Shrapnel, S. 2015. “Quantum Causal Models”

http://arxiv.org/abs/1512.07106

34



Davies, E. and Lewis, J. “An operational approach to quantum probability.” Comm.

Math. Phys. 17: 239-260.

Eberhardt, Frederick. 2013. “Direct Causes and the Trouble with Soft Interventions.”

Erkenntis 79(4): 755-777.

Egg, Matthias, and Michael Esfeld. 2014. “Non-local Common Cause Explanations for

EPR.” European Journal for Philosophy of Science 4:181–196.

Eibenberger, S., Gerlich, S., Arndt, M., Mayor, M. and Tüxen, J. 2013. “Matter-wave

interference with particles selected from a molecular library with masses exceeding

10,000 amu.” Phys. Chem. Chem. Phys. 15: 14696.

Evans, P. W. and Price, H. and Wharton, K. B. 2013. “New Slant on the EPR-Bell

Experiment” Brit. J. Phil. Sci. 64:297-324.

Evans, P. W. 2014. “Retrocausality at no extra cost”, Synthese.

Evans, P. W. 2015. “Quantum causal models, faithfulness and retrocausality.” arxiv:

1506.08925.

Frisch, Mathias. 2014. “Causal Reasoning in Physics” Cambridge University Press

Cambridge.

Glymore, Clark. 2006. “Markov properties and Quantum Experiments” In Physical

Theory and its Interpretation eds. Demopoulos, W. and Pitowsky, I. Springer.

117:126.

Gutoski, G. and Watrous, J. 2006 “Toward a General Theory of Quantum Games.” In

Proceedings of 39th ACM STOC 565–74.

35



Hausman, Daniel M, and James Woodward. 1999. “Independence, Invariance and the

Causal Markov Condition.” The British Journal for the Philosophy of Science 50:

521–583.

Henson, J. and Lal, R. and Pusey, M. F. 2014. “Theory-independent limits on

correlations from generalized Bayesian networks” New Journal of Physics 16: 113043

Hitchcock, C. 2012. “Probabilistic Causation” In The Stanford Encyclopedia of

Philosophy. ed. Edward N. Zalta.

Hofer-Szabó, Gábor, Miklós Rédei, and László E Szabó. 2013. The Principle of the

Common Cause. Cambridge: Cambridge University Press.

Ismael, J. 2015 “How do causes depend on us? The many faces of perspectivalism”

Synthese: 1-23.

Ismael, J. forthcoming. “Against Globalism about Laws”. In Experimentation and the

Philosophy of Science, eds. Bas van Fraassen and Isabelle Peschard, University of

Chicago Press.

Laskey, K. 2007. “Quantum Causal Networks” Proc. AAAI Spring Symp. on

Quantum142

Leifer, M. S., and Robert W Spekkens. 2013. “Towards a Formulation of Quantum

Theory as a Causally Neutral Theory of Bayesian Inference.” Physical Review A

88(5):052130–052138.

Näger, P. M. 2015. “The causal problem of entanglement” Synthese, 1-29

Nielsen, M. and Chuang, I. 2000. “Quantum Computation and Quantum Information.”

Cambridge University Press.

36



Oreshkov, O., Costa, F. and Brukner, C. 2012. “Quantum correlations with no causal

order.” Nature Communications 3: 1092.

Oreshkov, O. and Giarmatzi, C. 2015. “Causal and causally separable processes.”

http://arxiv.org/abs/1506.05449v2

Pearl, Judea. 2000. Causality: Models, Reasoning and Inference. Cambridge:

Cambridge University Press.

Pienaar, Jacques, and Caslav Brukner. 2014. “A Graph-Separation Theorem for

Quantum Causal Models.” arXiv:1406.0430.

Price, H. 2007. “Causal Perspectivalism” In Causation, Physics, and the Constitution

of Reality: Russell’s Republic Revisited eds. Price, H. and Corry, R. Oxford

University Press. 10: 250-292

Prugovecki, E. 1977. “Information-Theoretical Aspects of Quantum Measurement.”

International Journal of Theoretical Physics 16(5): 321 - 331.

Ried, K., Agnew, M., Vermeyden, L., Janzing, D., Spekkens, R. W., Resch, K J. 2015.

“A quantum advantage for inferring causal structure.” Nature Physics 11: 414 - 420.

Russell, Bertrand. 1912. “On the Notion of Cause.” Proceedings of the Aristotelian

Society 13:1–26.

Schlosshauer, Maximilian A. 2010. Decoherence and the Quantum-To-Classical

Transition. Berlin: Springer.

Shrapnel, Sally. 2014. “Quantum Causal Explanation: Or, Why Birds Fly South.”

European Journal for Philosophy of Science 4:409–423.

37



Shrapnel, Sally and Milburn, Gerard. 2015. Unpublished Manuscript.

Spirtes, Peter, Clark Glymour, and Richard Scheines. 2000. Causation, Prediction, and

Search. Cambridge, Mass.: MIT Press.

Suárez, Mauricio, and Iñaki San Pedro. 2011. “Causal Markov, Robustness and the

Quantum Correlations.” In Probabilities, Causes and Propensities in Physics, ed.

Mauricio Suárez, 173–193. Dordrecht: Springer.

’t Hooft, Gerard. 2009. “Entangled quantum states in a local deterministic theory.”

arxiv: 0908.3408

Timpson, Christopher.2013. “Quantum Information Theory and the Foundations of

Quantum Mechanics.” Cambridge University Press.

van Fraassen, Bas C. 1982. “The Charybdis of Realism: Epistemological Implications of

Bell’s Inequality.” Synthese 52:25–38.

Wood, C. J. and Spekkens, R. W. 2014 “The lesson of causal discovery algorithms for

quantum correlations: Causal explanations of Bell-inequality violations require

fine-tuning” New Journal of Physics 17(3):033002.

Woodward, James. 2003. Making Things Happen. A Theory of Causal Explanation.

Oxford: Oxford University Press.

———. 2007. “Causation with a Human Face.” In Causation, Physics, and the

Constitution of Reality, eds. Huw Price, and Richard Corry, 66–105. Oxford: Oxfo rd

University Press.

Zhang, Jiji and Spirets, Peter. 2015. “The Three Faces of Faithfulness.” Synthese. 1:17.

38



Zurek, W. 2014. “Quantum Darwinism, classical reality, and the randomness of

quantum jumps”, Physics Today 67:44.

39