Order Tutorial

Journal of Mathematical Psychology ( ) –
Contents lists available at ScienceDirect
Journal of Mathematical Psychology

journal homepage: www.elsevier.com/locate/jmp
Quantum cognition and decision theories: A tutorial

James M. Yearsley a,b,∗ , Jerome R. Busemeyer c
a
Department of Psychology, Vanderbilt University, Nashville, TN 37240, USA
b
Department of Psychology, School of Arts and Social Sciences, City University London, Whiskin Street, London EC1R 0JD, UK
c
Department of Psychological and Brain Sciences, Indiana University, Bloomington, IN 47405, USA
highlights
• We present a tutorial on quantum models of cognition and decision aimed at those with little or no prior experience of such models.
• We focus on the question of how to build quantum models in practice.
• We give examples from the study of order effect, including a new derivation of the QQ Equality.
article info abstract

Article history: Models of cognition and decision making based on quantum theory have been the subject of much
Available online xxxx interest recently. Quantum theory provides an alternative probabilistic framework for modelling decision
making compared with classical probability theory, and has been successfully used to address behaviour
Keywords: considered paradoxical or irrational from a classical point of view.
Quantum theory
The purpose of this tutorial is to give an introduction to quantum models, with a particular emphasis
Order effects
on how to build these models in practice. Examples are provided by the study of order effects on
Similarity
judgements, and we will show how order effects arise from the structure of the theory. In particular,
we show how to derive the recent discovery of a particular constraint on order effects implied by
quantum models, called the Quantum Question (QQ) Equality, which does not appear to be derivable from
alternative accounts, and which has been experimentally verified to high precision. However the general
theory and methods of model construction we will describe are applicable to any quantum cognitive
model. Our hope is that this tutorial will give researchers the confidence to construct simple quantum
models of their own, particularly with a view to testing these against existing cognitive theories.
© 2015 Elsevier Inc. All rights reserved.
1. Introduction Tversky & Kahneman, 1974.) These quantum models posit that, at
least in some circumstances, human behaviour does not align well
Models of decision making based on the mathematics of with classical probability theory or expected utility maximisation.
quantum theory have attracted a large amount of interest re- However unlike, for example, the fast and frugal heuristics pro-
cently (Aerts, 2009; Busemeyer & Bruza, 2014; Khrennikov, 2010; gramme (see, e.g. Gigerenzer, Hertwig, & Pachur, 2011), quantum
Mogiliansky, Zamir, & Zwirn, 2009; Pothos & Busemeyer, 2013; cognition aims not to do away with the idea of a formal structure
Wang, Busemeyer, Atmanspacher, & Pothos, 2013; Yukalov & Sor- underlying decision-making, but simply to replace the structure
of classical probability theory with an alternative theory of proba-
nette, 2011). These models have arisen in part as a response to the
bilities. This new probability theory has features, such as context
empirical challenges faced by ‘rational’ decision-making models,
effects, interference effects and constructive judgements, which
such as those based on Bayesian probability theory. (Such exam-
align well with psychological intuition about human decision-
ples are mostly associated with the famous Tversky–Kahneman
making. Initial research involving quantum models tended to focus
research tradition. See e.g. Chater, Tenenbaum, & Yuille, 2006;
mainly on explaining results previously seen as paradoxical from
the point of view of classical probability theory, and there have
been a number of successes in this area (Aerts, Gabora, & Sozzo,
∗ Corresponding author at: Department of Psychology, Vanderbilt University, 2013; Blutner, Pothos, & Bruza, 2013; Bruza, Kitto, Ramm, & Sitbon,
Nashville, TN 37240, USA. 2015; Pothos & Busemeyer, 2009, 2013; Trueblood & Busemeyer,
E-mail address: [email protected] (J.M. Yearsley). 2011; Wang et al., 2013; White, Pothos, & Busemeyer, 2014). More
http://dx.doi.org/10.1016/j.jmp.2015.11.005
0022-2496/© 2015 Elsevier Inc. All rights reserved.
2 J.M. Yearsley, J.R. Busemeyer / Journal of Mathematical Psychology ( ) –
recently, the focus has switched to some extent to testing new material in this tutorial. Second, we will not touch upon ‘entan-
predictions arising from quantum models, and designing better glement’ or issues around quantum information. Finally, some ad-
tests of quantum vs. classical decision theories (Atmanspacher & vanced topics, such as CP-Maps and the full theory of POVMs will
Filk, 2010; Yearsley & Pothos, 2014, in preparation). not be covered, as they are best learnt about once one is famil-
One key success of quantum models of cognition has been the iar with the basics. They will be covered in a subsequent tutorial
treatment of question order effects (Moore, 2002). ‘Order effects’ (Yearsley, in preparation).
here describes a phenomenon where, for example, given two We will assume the reader has a good familiarity with linear
particular questions, each with a number of possible responses, algebra in the usual form of vectors, matrices etc., but for reference,
the expected distribution of responses to a particular question and to set notation, we give a brief summary of some important
depends on whether it was asked first or second in the series. In ideas in the Appendix.
other words, asking a prior question can influence the outcome of The rest of this tutorial is structured as follows; in Section 2 we
a subsequent one. We will explain in more detail below exactly introduce the basic elements of quantum cognition. In Section 3
how to characterise this effect. we then expand upon some points, with the aim of guiding readers
As we shall see, order effects arise naturally in quantum the- through the process of constructing a quantum model in practice
rather than in theory. In Section 4 we give a brief introduction to
ory, and thus they can be accounted for by quantum cognitive
order effects in quantum theory, and in Section 5 we expand on
models. However what is more remarkable is that quantum the-
this to include a derivation of the QQ Equality. In Section 6 we
ory also predicts particular constraints on the probabilities that
give a brief introduction to POVMs, which can be used to represent
can be generated by these models, most notably in the form of
noisy or imperfect measurements, and in Section 7 we apply these
the Quantum Question (QQ) Equality (Wang & Busemeyer, 2013).
in the setting of order effects, our goal being to see to what
These constraints seem to be extremely well satisfied in the
extent the QQ Equality generalises to the case of more realistic
data from real world experiments and surveys (Wang, Solloway,
noisy measurements. In Section 8 we briefly talk about another
Shiffrin, & Busemeyer, 2014). Thus as well as being a natural appli-
application of quantum theory to modelling similarity judgements.
cation of quantum theory, question order effects also represent a
We summarise in Section 9. A number of mathematical details are
striking empirical confirmation of the idea of using quantum the- contained in the Appendix.
ory to model decisions.
Although the mathematical machinery of quantum theory is
2. The basics of quantum cognition
not inherently more complex than that required by many other
cognitive models, essentially linear algebra and a small amount of
The aim of this section is to present the basic formalism of quan-
calculus, it is rather unfamiliar to most cognitive scientists. Our aim
tum cognition, including information about the state, the dynam-
in this tutorial paper is therefore to introduce readers to the ideas ics, and the description of measurements. Our goal here is to give
and machinery of basic quantum theory, such that after working a reasonably concise account of the essentials; in the next sec-
their way through this tutorial readers will feel more confident tion we will return to each element in turn and ask in more detail
making use of quantum models in their research. what it means and how it may be specified for a particular model.
As well as existing cognitive scientists, we hope this tutorial We hope this format will make it easy for readers to grasp the
may find a secondary audience in those researchers who already essential structure of quantum models. All of the material in this
have a background in quantum theory gained from studying the section is standard, and we will not give references for individual
physical sciences, who are interested in the application of these results/definitions. For a more compete account see Isham (1995)
ideas in social science.1 To help these readers we have structured or for an alternative description with a more cognitive focus see
our discussion of the basic formalism of quantum theory in a way Busemeyer and Bruza (2014).
which should feel familiar to anyone who first encountered it in
the context of the physical sciences (see for example Isham, 1995 2.1. What is quantum cognition?
or the notes by Plenio, 2002 available online). Hopefully this should
enable those already familiar with quantum theory to quickly Quantum cognition is a framework for constructing cognitive
grasp how to apply their existing knowledge to the construction models based on the mathematics of quantum probability
of cognitive models. theory, which is itself a mathematical framework for assigning
The material we will cover in this tutorial is essential back- probabilities to events, much like classical probability theory (for
ground to any application of quantum theory in judgement and a full account see Busemeyer & Bruza, 2014). For a given event,
decision making. We will pay particular attention to two impor- usually thought of as the outcome of some judgement process,
tant but sometimes overlooked issues; first how exactly does one and specification of the decision maker by means of a cognitive
choose a particular framework of Hilbert space, basis vectors, ini- state, quantum cognition gives a real number between 0 and 1
tial state etc. to suit the problem at hand, and what do these choices which is to be interpreted as the probability that the decision
mean? Second, how are the various calculations actually carried maker will make that particular choice. Quantum cognition also
out? Grasping both of these issues is essential for any student of the includes information about the set of possible dynamics, state
field, and we hope this tutorial will help researchers bridge the gap transformations and measurements that can be performed on a
between reading about quantum models and actually constructing system, although to a large extent this follows directly from the
them for themselves. basic probabilistic structure.
There are a number of things we will not cover in this tutorial, In its most conservative form, quantum cognition is simply an
which may be worth stating now. First, although we will mention algorithm for computing probabilities, without any claim to reflect
it, we will not cover the dynamics of quantum systems in any de- the underlying way decisions are made in the brain. In this way of
tailed way; this is mainly a tutorial on quantum statics. Quantum thinking, the success or otherwise of the approach is to be judged
dynamics are relatively simple to grasp once one understands the purely by the empirical success of its predictions. However steps
are being taken towards viewing quantum cognitive theories as
process models, that do reflect in some way the process of arriving
at a given decision (Kvam, Pleskac, YU, & Busemeyer, 2015). The
1 Indeed one of the present authors (JMY) has such a background. attraction of quantum models in this case stems in part from
J.M. Yearsley, J.R. Busemeyer / Journal of Mathematical Psychology ( ) – 3
the fact they have features, such as contextuality, interference equivalent representations of |ψ⟩ (one is the Fourier transform of
effects, order effects and constructive judgements, to name a few, the other, and they carry the same information), but they live in
that seem to align well with the way we think human decision different spaces and have different units, so they are certainly not
makers process information. However an important feature of equal, and neither are they equal to the abstract vector |ψ⟩.
quantum models of decision is that they generally do not assume An important concept is the idea of an operator on a Hilbert
the underlying processing that happens in the brain is represented space. An operator on a space is a map which sends an element
by quantum physics. In this sense that are distinct from so of the space to another element of the same space. e.g.,
called quantum brain models (Hagan, Hameroff, & Tuszynski, 2002;
Hameroff & Penrose, 1996; McKemmish, Reimers, McKenzie, Mark, A:H →H
(4)
& Hush, 2009; Tegmark, 2000). |u⟩ → A |u⟩ .
The ingredients of any quantum cognitive model are a space
of possible thoughts/judgement outcomes etc. with subspaces In a particular basis, the operator A will be represented by a matrix.
corresponding to particular beliefs/opinions/choices, together Note though that in a similar way to states, an operator is not equal
with a cognitive state that keeps track of a decision maker’s to a matrix, since the particular matrix depends on the choice of
current state of mind, and a dynamics that specifies how this basis.
state of mind changes with time/evidence presentation etc. From An important property of operators in Hilbert space is that
these ingredients we can compute probabilities for any relevant they need not commute. Let A and B be two operators on H . The
judgement outcome. commutator [A, B] is defined as,
In the following subsections we will introduce each element in
[·, ·] : H × H → H
turn. (5)
[A, B] → AB − BA
2.2. The space i.e. the commutator represents the difference between the two
possible orders in which A and B could be applied to a state. A and B
The most basic ingredient in quantum cognition is a space of are said to commute if [A, B] = 0. Otherwise they are said to be non-
possible thoughts/beliefs/etc. This space is a finite dimensional commuting. That many important operators do not commute with
Hilbert space,2 which is essentially just a vector space equipped each other, such as those representing position and momentum,
with an inner product ⟨·|·⟩. In quantum theory the Hilbert spaces is what lies behind many interesting features of quantum theory,
we consider are complex spaces, which means they can be spanned such as the uncertainty principle.
by vectors whose entries are complex numbers. It is usual to denote
A basic overview of operators on Hilbert spaces is given in the
the Hilbert space by the symbol H .
Appendix.
We will write the elements of H as |u⟩. |u⟩ is known as a ‘ket’,
and is roughly analogous to a column vector in standard linear
Aside on infinite dimensional spaces
algebra. The inner product of two vectors in H is denoted as
In this tutorial we will be working mainly with finite dimen-
⟨|v⟩ , |u⟩⟩ ≡ ⟨v|u⟩ (1) sional vector spaces, that are appropriate for modelling finite sets
which is roughly analogous to the dot product of two vectors in of possible question answers. However we may sometimes need to
linear algebra. The inner product induces a norm | · | on H via, model situations where the variables can take any value in a con-
tinuous range, such as the position of a slider, or situations where
|·|:H →R the set of possible outcomes is so large it may be more convenient
 (2) to think of them as continuous, for example confidence judgements
|u⟩ → ⟨u|u⟩. on a 0%–100% scale, where the corresponding finite dimensional
Technically our Hilbert space is an abstract space, i.e. it can be representation has 101 dimensions. In these cases we can extend
spanned by many different possible sets of basis states. One the formalism to a infinite dimensional Hilbert space. Instead of
important point about this is that the object, |u⟩, while a ‘vector’ vectors our space is now spanned by complex valued functions.
in the technical sense, is not equal to a column vector, i.e. Most of what we say in this tutorial still holds, but the maths be-
  comes slightly more complex. Interested readers should consult
a any good book on quantum theory (e.g. Isham, 1995), or for the
|u⟩ ̸= . (3)
b full mathematical glory, see Reed and Simon (1980).
The reason for this is that the right hand side of Eq. (3) is defined
relative to a particular basis, whereas the left hand side is basis 2.3. The knowledge state
independent. We will often abuse notation to write a ket as equal
to a column vector etc., but when we do so we will always take care The second ingredient in any quantum model is the specifica-
to specify the basis in which we are computing the vector. tion of the initial knowledge state of a participant, or group of par-
One particularly striking example of the difference between ticipants. This is given in the simplest case by a vector |ψ⟩ in H .
vectors in H and vectors/functions in a particular basis occurs This state vector encodes information about the cognitive state of
when dealing with states of a particle moving along a line (Isham, participants prior to any computation. This state can also be writ-
1995). Let |ψ⟩ be the state of a particle in 1D. Then ψ(x) = ⟨x|ψ⟩ ten as an operator on the Hilbert space by taking the outer product
is the wave function of the particle in position space, and ψ̃(p) = of |ψ⟩ with itself, to give the density operator,
⟨p|ψ⟩ is the wave function in momentum space. ψ(x) and ψ̃(p) are ρ = |ψ⟩ ⟨ψ| . (6)
The main advantage of using density matrices to represent
knowledge states, rather than just vectors, is that it allows us to
2 The restriction to finite dimensions is not strictly necessary. The extension to
deal with situations where we have some (classical) uncertainty
infinite dimensional spaces such as those required to model motion in a continuous
space requires only some extra mathematical precision. However the spaces most
about the state. Suppose we have an ensemble of different groups
often used in quantum cognition are finite, so we will mostly restrict to this case in of participants and each group can be characterised by the state
this tutorial. See below or Isham (1995) and Reed and Simon (1980) for more. vector |ψi ⟩ for i = 1, 2 . . . N. Let the proportion of participants
in each group be ωi . Then the state of the whole ensemble can be The expected value of the observable A for a participant with a
written as, cognitive state given by ρ is equal to,
⟨A⟩ = Tr(Aρ) = ⟨ψ| A |ψ⟩


ρ= ωi |ψi ⟩ ⟨ψi | . (7) (10)
i
where the final equality holds if the cognitive state is pure.
It is cumbersome to represent this type of ensemble using state However if we perform a measurement of A we do not get the result
vectors alone. ⟨A⟩ in general. An important feature of quantum theory is that the
Let us introduce some simple nomenclature. States that can be possible results of a measurement of any observable are one of its
written in the form ρ = |ψ⟩ ⟨ψ| for a single |ψ⟩, i.e. that have eigenvalues.
ωi = 1 for some i and ωj̸=i = 0 are called pure states. States that Because the operators representing observables are self-
cannot be written in this form, i.e. that have ωi ̸= 0 for more than adjoint, the spectral theorem for self adjoint operators (see
one i, are called mixed. Appendix) tells us that A may be written as,
The knowledge state plays a role roughly analogous to a prior 
probability distribution in classical theories. If the states |ψi ⟩ are A= ai PiA (11)
orthonormal then the ωi play the role of the probabilities for the i
initial state to be one of the |ψi ⟩. The following properties of ρ
where the ai are the eigenvalues of A and PiA are the projection
guarantee that the ωi can be interpreted as probabilities in this
operators onto the subspaces corresponding to these eigenvalues.
way;
The probability of measuring A and getting the result ai is given by,
• ρ = ρ Ď . This guarantees the
ωi are real.
• Tr(ρ) = 1. This guarantees i ωi = 1. p(ai ) = Tr(PiA ρ) = ⟨ψ| PiA |ψ⟩ (12)
• ⟨ψ| ρ |ψ⟩ ≥ 0, ∀ |ψ⟩ ∈ H . This property, known as positivity,
where again the last equality holds if the cognitive state is pure.
guarantees the ωi are non-negative.
A crucial feature of quantum theory is what happens to the state
A particularly important example of a knowledge state is the when this measurement is carried out. Suppose we measure the
case where the {|ψi ⟩} form an orthonormal basis for H and ω1 = observable A and we get the answer ai . Then the state collapses to
ω2 . . . = 1/D, where D is the dimension of the space. The the new state,
knowledge state can be written in this case as,
PiA ρ PiA
1 ρ→ (13)
ρ= 1 (8) Tr(PiA ρ)
D
or
where 1 is the identity operator, i.e. the operator that has the
property that, for any A, 1A = A1 = A. This state corresponds P A |ψ⟩
to a state of maximum (classical) uncertainty, and it plays a similar |ψ⟩ →  iA  (14)
P |ψ⟩
i
role to a uniform prior in classical models. Note that it is possible
to define the notion of the ‘entropy’ of a quantum state, in a similar if the state is pure. This is known as the collapse postulate in
way to the entropy of a classical probability distribution. This is quantum theory and is responsible for much of the strange non-
beyond the scope of this tutorial, see Neilsen and Chuang (2000) classical behaviour of the theory. Note that if we take the collapsed
or Yearsley (in preparation) for details, however it can be shown state and immediately perform another measurement of A then we
that ρ = D1 1 does indeed give the maximum value of the entropy, will get the answer ai again with certainty, since,
equal to ln(D).
PiA ρ PiA
 
We note that it is usual in quantum cognitive models to see
Tr PjA = δij . (15)
the initial knowledge state represented as a pure state (e.g. in Tr(PiA ρ)
Busemeyer & Bruza, 2014 density matrices are introduced only
at the very end of the book). In the interests of completeness we This means that measurements represented by projection opera-
will mainly work in this tutorial with density matrices, so that the tors can also be used to prepare systems to be in a known state.
initial state may be either pure or mixed. Where appropriate we These are often known as ‘ideal first kind’ measurements. That such
will show how our expressions simplify in the case of pure states. operations can actually be carried out in practice is a critical as-
sumption in quantum cognition.3
It is worth saying a little more about the role of the collapse pos-
2.4. Observables and measurements
tulate, since anecdotally it seems to cause some concern amongst
cognitive scientists. On the surface it looks a lot like Bayesian up-
The term observable in quantum theory refers to any quantity
dating. Imagine that the initial state is a density operator ρ which
whose value can be measured by a suitable experiment. In quan-
can be written as a weighted sum of projection operators onto
tum cognition observables are generally preferences or choices,
eigenvectors of a particular observable A, i.e.,
such as whether a suspect if guilty or innocent, which of two gam-
bles are preferred, or how confident a decision maker feels about

ρ= p(ai ) |ai ⟩ ⟨ai | (16)
a particular choice. The observables in a quantum cognitive theory
i
are ultimately the same as they would be in a classical theory.
Observables in quantum theory are represented by Hermitian with {p(ai )} a probability distribution over the possible ai . Such a
operators on H . An operator is Hermitian if decomposition is not guaranteed to be possible, because although
the |ai ⟩ form an orthonormal basis ρ need not be diagonal in this
A = AĎ ≡ (AT )∗ . (9) basis. However suppose it is possible for the particular state under
The operation Ď is known as Hermitian conjugation. Hermitian op-
erators are an important class because they have real eigenvalues,
which makes sense, since any measurement must have a result that 3 Although one can argue that it is a critical assumption in any theory of cognition,
is a real number. quantum or otherwise.
consideration. Then if we measure the observable A and get the re- The solution is to model the cognitive state not as a closed
sult aj then the state collapses to, system, but as an open one, where information can flow between
the cognitive system and some external system or environment.
ρ →  aj aj  .
  
(17) This allows for the possibility of irreversible evolutions not
represented by unitary operators. It is also beyond the scope of
In other words, if one thinks of the ωi in Eq. (16) as defining a proba-
this tutorial (Yearsley, in preparation). Concerns about reversibility
bility distribution over the set of states |ai ⟩, then the collapse looks
notwithstanding, unitary evolutions are the most widely used in
exactly like a Bayesian updating of this probability distribution.
quantum models of decision, and have indeed proven to be very
The reason this interpretation does not work in general is that
successful. However one should recognise that such a choice may
a Bayesian update happens when one gets some new information
not always be appropriate.
about the state as a result of performing a measurement. Because
There are two further points we want to make about evolution.
of this, a Bayesian update may, at least in some cases, decrease
First we want to quickly discuss how to write expressions
the uncertainty, or entropy, associated with our description of the
for the probabilities associated with measurements at different
system. In contrast, at least for pure quantum states, increasing
times, second we want to give two simple examples of useful
our knowledge of an observable A will decrease our knowledge
Hamiltonians.
of any observable that does not commute with A. For pure states
If we start with an initial cognitive state ρ0 , evolve for a time t
this means state collapse does not reduce our uncertainty about
and then perform a measurement of A, the probability we will get
the state, in contrast to a classical Bayesian update. That being
the outcome ai is given by any of the following expressions,
said, there is an approach to understanding the interpretation
of quantum theory which emphasises the connection between p(ai , t ) = Tr(PiA U (t )ρ0 U Ď (t )) (22)
collapse and Bayesian updating, called Quantum Bayesianism. (See,
for example, Caves, Fuchs, & Schack, 2002.) We will not discuss = Tr( ρ(t ))
PiA (23)
this further here since the interpretation of quantum theory is
= Tr( (t )ρ0 )
PiA (24)
beyond the scope of this tutorial. The important point to take away
from this discussion is that although the collapse postulate is a where,
fundamentally quantum feature, it does have some similarities
with more familiar ideas in classical probability theory. ρ(t ) = U (t )ρ0 U Ď (t )
(25)
PiA (t ) = U Ď (t )PiA U (t ).
2.5. Evolution
Eqs. (23) and (24) refer to the Schrödinger and Heisenberg pictures
of evolution respectively. They differ in terms of whether one
We will not be especially concerned in this tutorial with the
regards the evolution as changing the cognitive state, or the
evolution of quantum states. Nevertheless, for the sake of com-
observables. For the most part we tend to work in the Schrödinger
pleteness we note here the basics of quantum evolution.
picture in quantum cognition, but both representations are equally
For a closed quantum system, evolution is generated by an
valid.
Hermitian operator H known as the Hamiltonian.4 The evolution
Now suppose we measure A at time t1 , then we continue to
of a quantum state is governed by,
evolve our state, and now we also measure B at time t2 . The
ρ(t ) = U (t )ρ(0)U Ď (t ) (18) probability that we get the outcome ai at t1 and bj at t2 is,
or p(ai , t1 ; bj , t2 ) = Tr(PjB U (t2 − t1 )PiA U (t1 )ρ0 U Ď (t1 )
|ψ(t )⟩ = U (t ) |ψ(0)⟩ (19) × PiA U Ď (t2 − t1 ))

where, if H is time independent, U (t ) is given by, = Tr(PjB (t2 )PiA (t1 )ρ0 PiĎA (t1 )). (26)
U (t ) = e−iHt . (20) This can easily be extended to further evolutions and measure-
ments. The Heisenberg picture of evolution comes into its own
The operator U (t ) is unitary which turns out to be an extremely here.
important property. Unitary operators preserve the length of the We want to conclude this section on evolution with a couple
state vector, or more generally the trace of the density operator, of examples of useful Hamiltonians. The first is the Hamiltonian
which means they conserve total probability. They also have an for a single particle evolving in some potential V (x), such as might
inverse, U −1 (t ) = U Ď (t ) = eiHt , so that, be appropriate for modelling a set of beliefs which could take any
value in a given range. In the x basis we have,
U Ď (t )U (t ) = U (t )U Ď (t ) = 1. (21)
1 ∂2
The reason this is important is the following: In practice we H =− + V (x). (27)
generate evolution of a cognitive state by the presentation of 2m ∂ x2
some stimuli. If we assume this evolution can be modelled by Here m is some constant (the mass in physical terms) which gives
a Hamiltonian operator leading to a unitary evolution, then we rise to a sort of ‘inertia’ of our cognitive state, in other words it
also assume that this evolution is reversible. That means we are determines how quickly the state changes as a result of applying
implicitly assuming there is some second set of stimuli that, if some force determined by V (x). This sort of Hamiltonian has been
presented to a participant, could undo the change in cognitive used in studies of confidence judgements (Kvam et al., 2015). Good
state caused by the first set of stimuli (assuming no measurement choices for the potential might be linear (V (x) = ax + b) which
intervenes). Whether this is reasonable or not depends on the generates a constant force in a particular direction, or a quadratic
experimental setup, but it is frequently not realistic. (V (x) = ax2 ) which produces a state which oscillates around x = 0.
Second, the Hamiltonian for a simple two state system can be
written (like any other Hermitian operator on this space) as a
weighted sum of the identity operator and the three Pauli matrices,
4 Since H is Hermitian, it is an observable. In physics H corresponds to the energy
of a system, however its significance in cognition is less clear. H = a1 + bσx + c σy + dσz (28)
where 1. The space of states. In this case they are probability distributions
      (functions) in the interval (0, 100).
0 1 0 i 1 0
σx = , σy = , σz = (29) 2. The initial state. Or at least what that allowable class of initial
1 0 −i 0 0 −1
states looks like.
and a, b, c , d are real numbers. Adding an overall scale to the 3. The dynamics. Which we will assume can be written in terms
numbers just speeds up or slows down the evolution, which of a classical Hamiltonian function.
is equivalent to rescaling the time. Thus only the ratios of the 4. The observables. In this case the probabilities of lying in various
different coefficient are important. Now, the term involving the intervals of the form [a, b].
identity matrix just adds an overall phase to the state. Since the
absolute phase is unobservable we can ignore this term. The terms Given these four features, we can now construct a fairly unique
involving σx , σy , σz generate rotations about the x, y, z directions quantum analogue of this model. Let us explain how to do this,5
respectively, depending on the setup one might be able to further • The Hilbert space is given by the set of complex valued functions
simplify this expression by considering rotation around only a on the interval (0, 100), together with the condition that these
single axis. Thus, for example, in Yearsley and Pothos (2014) it was functions are normalisable. Strictly H = L2 ([0, 100]). These
argued that the Hamiltonian, are just the infinite dimensional analogues of column vectors.
H = bσx (30) • The initial state is a normalised state |ψ⟩ ∈ H .
• The dynamics follow from the classical Hamiltonian. Suppose
gives a reasonable general dynamics if one is interested only in the
we have H = f (x, p) where x and p are the classical position and
observable σz .
momentum of the state. A typical example might be a particle
evolving in according to some potential V (x) (representing the
3. What does it all mean?
effect of evidence presented maybe) H = p2 /2m + V (x). To get
3.1. Introduction the equivalent quantum dynamics one makes the replacement6
x → x̂, p → p̂, so H → Ĥ = p̂2 /2m + V (x̂). Here p̂ is the
In the previous section we introduced the basics of quantum operator in Hilbert spaces corresponding to momentum, in the
theory, as applicable to constructing theories of decision making. x basis it is given by −i ∂∂x . The dynamics are then given by the
In theory this gives all the information you need to go and study Schrödinger equation,
existing quantum models, and even in principle to build models of
∂
your own. −i |ψ⟩ = Ĥ |ψ⟩ . (31)
However much was left unsaid about the meaning of the various ∂t
objects introduced. As a consequence, while it should now be • Finally, the observables are given by self adjoint operators
obvious that, for example, to build a quantum model of order on H , but the most important one is simply x̂, which is the
effects one first needs to identify the Hilbert space, it is probably operator corresponding to the confidence judgement. Typical
less clear exactly how to do this in practice. What we want to do judgements might be whether a participant’s confidence was
in this section is to fill in some of the gaps regarding the meaning between a% and b%, which corresponds to the projection
of the various pieces of the quantum formalism. Our aim will be operator,
to give some clues about how to build a quantum model for a
b
particular set of judgements.

There are two possible positions one might find oneself in when P̂[a,b] = dx |x⟩ ⟨x| (32)
a
attempting to construct a quantum model of a particular decision
making process. First we might have in mind some existing so that,
classical model, and want to explore a quantum analogue, either  b
because the classical model is already empirically inadequate, or p(a, b) = ⟨ψ| P̂[a,b] |ψ⟩ = dx|ψ(x)|2 . (33)
simply because one is interested in whether a quantum model a
might produce novel predictions. The second position is that we
might have little idea about any existing treatment of the problem, So much for this ideal situation. In practice we are unlikely to
and we have to start modelling from scratch. In practice we tend to have a concrete classical model to work from, and we have to be
encounter the latter more often than the former, but we will look considerably more inventive to set up our quantum models. We
at both scenarios here. will explain how to do this below.
3.2. Quantisation? 3.3. Building quantum models from scratch
Suppose we already have a classical model of a decision making

In the last section we saw that if one has a well defined classical
problem, and we are interested in creating a quantum analogue.
model of decision making, setting up a quantum analogue is a
This is very similar to the typical situation in the physics literature,
fairly straightforward process. However it is much more common
and physicists already have a reasonably straightforward process
in practice to approach a modelling problem with little or no prior
for constructing quantum models from classical ones. This process
idea about how to model it. In this case one cannot start with an
is known as ‘quantisation’ (The classic reference is Dirac, 1982), and
existing classical model, and we need to find a different way to
we will describe it briefly.
It is easy to take a specific example, so we will look at confi- proceed.
dence judgements than can be expressed on a 0%–100% scale. We The difficulty we face is not so much that it is hard to build
will assume the range of possible values is continuous, even if the a quantum model, but that it is hard to constrain the possible
set of possible responses is discrete. Assume we have some initial
distribution of confidence levels, dynamics that might be gener-
ated by some evidence presentation, and then a judgement pro- 5 To avoid confusion, in this subsection we use hats to denote operators.
cess that corresponds to splitting the scale into some intervals, 6 This hides a multitude of sins. If the classical Hamiltonian contains terms like
e.g. 0% ≤ x < 10%, or 20% ≤ x < 30% etc. The precise details are xn pm there is no unique quantisation, since classically this term does not depend on
unimportant, but the key is the classical model gives us four things, the order of x and p, whereas of course in quantum theory it does.
choices. For simple models this is rarely a problem, but as the 3.4.1. Identifying H
number of, say, responses grows so does the number of possible For the sets of simultaneously definable observables, find the
ways of encoding these responses as vectors in a Hilbert space. For set with the largest set of associated states. If two or more sets of
this reason what we will describe below is a reasonably algorithmic commuting observables tie for the largest number of states then
way to construct quantum models, but it may often be the case pick either, call this number D. Your Hilbert space then must have
that the modeller has some extra information that suggests a D complex dimensions, or in other words H = CD .
particular structure to use. Remember that, provided a few simple To go back to our example, it turns out our two sets of com-
rules are followed, it is hard to produce a model that is ‘wrong’ muting observables {A, α} and {B} have equal numbers of distinct
in the sense of not being a valid quantum system. Some degree of states, four each, so we can work with either. Since the set {B} has
experimentation when building models is therefore fine, and may four distinct states, we will need a 4D space for them to live in. This
indeed be necessary to get a model which works well. means we need a Hilbert space of four complex dimensions, so we
Here is a rough outline of how you might go about building a have determined that H = C4 .
What we have done here is used the spectral theorem for
quantum model of a particular decision making process.
Hermitian operators in reverse (see Appendix). We know that
the eigenstates of an Hermitian operator span the Hilbert space
3.4. Identify the observables in which it lives. Thus if we know that we have an observable
with D distinct eigenvalues, that means it must live in a complex
space with at least dimension D. The trick is then to look at
Identify all the observables/measurements/judgements you
our observables and find the one with the largest number of
wish to include in your model. An observable must be defined for eigenvalues, this gives the minimum dimension of our Hilbert
any state. Decide which of them you expect to be able to define space. Care needs to be taken if we have commuting subsets of
simultaneously, and for each set of simultaneously well defined observables, like A and α above, since the number of distinct states
observables decide how many distinct states there are. In other now depends on the degeneracy (e.g. if A and α can each take one
words, identify the commuting subsets of your observables, and of three possible values, the number of vectors may vary between
the associated eigenstates. If all observables commute you have a 3 and 9).
So in our example H = C4 and the states |b1 ⟩ , |b2 ⟩ , |b3 ⟩ ,|b4 ⟩
 
classical model.
As an example, suppose we have three variables, A, α and B. and {|a1 , α1 ⟩ , |a1 , α2 ⟩ , |a2 , α1 ⟩ , |a2 , α2 ⟩} form two orthonormal
Suppose A can take values a1 , a2 , α can take values α1 , α2 and B can bases for H .
take values b1 , b2 , b3 , b4 . Suppose we expect to be able to define A
and α simultaneously, but we do not expect to be able to define 3.4.2. The relationship between the observables
either together with B. So our two sets of commuting observables Now by construction the eigenstates of B and those of the pair
are {A, α} and {B}. {A, α} span H , so we already know two different bases for our
Now an important question is whether states with the same Hilbert space. The question is, what is the relationship between
value of A might have different values of α or vice versa. If this is the sets of bases? Or in other words, we started by assuming that A
not the case then A and α are simple functions of one another and and B could not be defined simultaneously, which means that they
that is not very interesting. Suppose instead that states with the will not commute as operators. So the question is, what is their
same value of A can indeed have different values of α . Then there commutator?
are four possible states defined by the set {A, α}, let us denote them Well the short answer is that we cannot tell you! This is
{|a1 , α1 ⟩ , |a1 , α2 ⟩ , |a2 , α1 ⟩ , |a2 , α2 ⟩}, where e.g. state |a1 , α1 ⟩ has essentially where the maths ends and the psychologist’s intuition
value a1 for observable A and value α1 for observable α . has to take over. Quantum cognitive theories, like any other
theory, require certain inputs from the modeller, and one of these
Once we have identified the sets of observables we can use
is the relationship between the observable quantities. However
these to determine H .
although we cannot tell you exactly how to proceed, we can
give you some clues about how to implement different choices.
Aside on degeneracy One comes from the fact that since the two sets of vectors
One possibility we are ignoring here is that all the commuting {|a1 , α1 ⟩ , |a1 , α2 ⟩ , |a2 , α1 ⟩ , |α2 , α2 ⟩}, and {|b1 ⟩ , |b2 ⟩ , |b3 ⟩ , |b4 ⟩}
each form an orthonormal basis for H , they must be related to
subsets we generate may have degenerate eigenspaces.
each other by a unitary transform. A unitary transformation in
It might happen that all eigenstates are degenerate because of
the space CD has D2 parameters, so that gives us the maximum
some observable we have failed to account for. For example, as well
number of parameters to be fixed. However this number will often
as the operators A, α, B there may be some fourth operator C which
be much lower in practice, because of simplifying assumptions we
commutes with all the other operators and gives multiple possible can make.
flavours for all existing states, e.g. |b1 , c1 ⟩ , |b1 , c2 ⟩ . . . . This could Another is that we may sometimes be able to make simplifying
happen if there is a mismatch between the way a judgement is assumptions about the relationship between the states. Suppose
elicited and the way that observable is represented cognitively, e.g. for example we are dealing with emotion states from two different
we may ask participants to give preference judgements on a 1–9 points of view, perhaps two different people, so that our states
scale when their preferences are actually encoded on a 0–100 scale are {|Happy⟩ , |Sad
 ⟩ , |Excited ⟩ , |Scared⟩} 
from one point of view,
and Happy′ , Sad′ , Excited′ , Scared′ from a different point
 
in the brain.

Although it seems like this might be a problem, it is actually of view. It might be that we have reason to believe that e.g. Happy′

unlikely to cause difficulty as long as the dynamics do not contain

is a mixture of |Happy⟩ and |Sad⟩ and Excited′ is a mixture of

interactions between C and other observables. However a good |Excited⟩ and |Scared⟩, but that the change of perspective does not
general rule is to choose your observables to match the way you mix e.g. |Happy⟩ and |Excited⟩. The unitary transformation can then
think the quantities are encoded in the brain, and then if necessary be split into two separate transformations, each operating on a 2
represent the measurements in a coarse grained way as projection dimensional subspace, and the number of parameters drops from
operators onto large dimensional subspaces. D2 = 16 to 2 ∗ 22 = 8. Another simplification occurs when we
have a small number of observables compared with the dimension observable A on the initial cognitive state and get the result a1 , then
of H . We may then be able to pick all our basis vectors to be real. we know that the cognitive state has now collapsed to |a1 ⟩. Thus
A further special case is when we think the different bases are we can use an initial measurement as a priming, to fix the cognitive
maximally mixed, in the sense that a basis state from one basis is a state. Assuming not all participants give the same answer to this
mixture of all basis states from the other basis, so that there is equal initial question, they will not all be in the same state afterwards.
probability to measure this state to be in any of the basis states for We can deal with this in one of two ways; if there is an extremely
the other basis. In other words, if we have two bases {|ei ⟩} and {|fj ⟩} high probability of participants giving one answer, say a1 , then we
then, can simply screen out any participant who answers otherwise. Our
cognitive state is then ρ = |a1 ⟩ ⟨a1 |.
 ei |fj  = 1 ,
 2
∀i, j ∈ 1, 2 . . . D. (34) If, on the other hand, there is a spread of likely answers, we
D can either model each participant separately with a cognitive state
If our two bases have this property then they are called mutually corresponding to his/her initial answer, or we can use the density
unbiased (mutually unbiased bases were introduced in, Schwinger, operator,
1960. For a more recent review see Bengtsson, Unpublished). There 
are only a fixed number of mutually unbiased bases in a given ρ= ωi |ai ⟩ ⟨ai | (35)
Hilbert space, which means only a fixed number of options for i
choosing {|fj ⟩} given {|ei ⟩}. If the dimension of the Hilbert space, where the {ωi } are equal to the probabilities that participants give
D, is a power of a prime number, then the number of mutually each of the possible answers {ai } to the initial question. A special
unbiased bases is D + 1. It is not known how many mutually case of this is when there is equal likelihood of participants being in
unbiased bases there are for general D, but it is always less than any of the |ai ⟩, in which case the most appropriate initial state is the
or equal to D + 1 (Durt, Englert, Bengtsson, & Zyczkowski, 2010). completely mixed state, corresponding to maximum ignorance,
Thus in this case there is only a finite set of possible choices for the ρ = D1 1.
basis vectors. We can then take this idea further. Suppose we have good rea-
What exactly does the choice of relation between observables son to believe participants would, with extremely high probability,
mean for a model? In a classical model all observable are compati- answer the initial question in a particular way. One example might
ble, which means they can be defined at the same time. In quan- be if our experimental setup was something like a jury trial, and
tum models observables are generally not compatible, and they the initial question was prior to the presentation of any evidence,
cannot be defined simultaneously. Assuming one wants to use a since participants should overwhelmingly prefer innocent to guilty
quantum model in the first place, then presumably one is looking in the absence of any evidence. In this case since we are close to
for incompatibility. The relationship between observables in a par- certain that participants would respond innocent if questioned, we
ticular model therefore determines two things; how much incom- can say their initial state must be very close to the ‘innocent’ eigen-
patibility one has, and whether the incompatibility occurs equally vector. However this rapidly becomes unreliable as our confidence
across all states of those observables. in what the initial judgement would be drops from 100%. This is, in
As we will show for a specific example in Section 4, the degree its simplest form, because it is impossible to
√ √tell the difference be-
of incompatibility, measured through non-commutation or order tween a superposition of the form a |0⟩+ 1 − a |1⟩ and a mixed
effects, depends on the overall size of the unitary rotation between state of the form a |0⟩ ⟨0| + (1 − a) |1⟩ ⟨1| by only measuring in the
bases for A and B. However if the unitary transformation does not {|0⟩ , |1⟩} basis.
effect all basis states equally, it might be the case that, for example, Nevertheless, the key is that the application of some quantum
some subset of the possible outcomes of A commutes with a subset know how together with some psychological intuition can usually
of the possible outcomes of B, so that although the observables are help us to identify the initial cognitive state in a straightforward
technically incompatible this may not be the case when acting on manner.
all possible states. In general deciding on the relationship between
a given set of observables is a hard problem, that needs a lot of 3.4.4. Identifying the dynamics
psychological insight to solve successfully. We shall briefly talk though how the dynamics might be spec-
ified. In the same way as for the initial state above we shall begin
3.4.3. Identifying the initial state with the technical story, and then we shall show how the problem
The initial state is straightforward to identify, at least in the- can be simplified with some psychological intuition.
ory. The initial state can be fixed from the probabilities for the out- Evolution is generated by a Hermitian operator, the Hamilto-
comes of measurements. So if one knows the initial measurement nian. For finite dimensional systems this is an D × D Hermitian
probabilities, one can identify the state. matrix, and thus has D2 parameters. We can fix some of these pa-
The problem is that there are 2D − 2 free parameters in a pure rameters by using the transition probabilities,
quantum state, and D2 − 1 in a general density operator. By con-
trast, a complete set of measurements can fix at most D−1 parame- p(i → j; t ) = |⟨j|U (t ) |i⟩|2 . (36)
ters (see Appendix for information on how to compute these). Thus There are D2 /2 of these transition probabilities, so we either need
to fix all the parameters of a pure state we need probabilities for them for two different bases, or for the same basis at two different
at least two non-commuting observables, and we will need many times. In theory we know this information if we know the expected
more than that to fix a density operator. dynamics.
However, this is almost never the way things are done in In practice, in much the same way as we could assume a simpli-
practice. Usually we set the initial state to be a particular eigenstate fied form for the initial state, we can probably assume a simplified
of one observable. How can we get away with this? The answer lies form for the dynamics. A standard assumption is that the dynamics
in recognising that because measurements change the state via the of a quantum system functions like a transformation between two
collapse postulate, measurements can be thought of as equivalent known bases. e.g. it might rotate the states {|0⟩ , |1⟩} to the states
to state preparation (Peres, 1998). {|+⟩ , |−⟩} after some total time T . This means,
Suppose we have a D dimensional Hilbert space with some
unknown initial cognitive state |ψ⟩ and an observable A with 1
U (T ) |0⟩ = |+⟩ = √ (|0⟩ + |1⟩) (37)
eigenvectors |a1 ⟩ , |a2 ⟩ . . . . If we perform a measurement of the 2
so we can take, Most operators in quantum theory do not commute with each
 π  other, and thus the order in which they are evaluated matters.
U (T ) = exp −i σy (38) If the two operators represent two possible survey questions, for
4 example, then a lack of commutation means that the expected
thus we conclude, answers to these questions depend on the order in which they
are asked. Thus quantum theory can naturally incorporate order
πt
 
U (t ) = exp −i σy . (39) effects via non-commuting operators (Atmanspacher & Römer,
4T 2012).
Again, we have seen that with some psychology intuition about As an example, let us consider a two dimensional Hilbert space
the expected dynamics, we can fix the unitary operator controlling spanned by {|0⟩ , |1⟩}, and two projection operators given by
evolution fairly simply. 1
P0 = |0⟩ ⟨0| , P+ =(|0⟩ + |1⟩)(⟨0| + ⟨1|). (41)
2
3.5. The interpretation of quantum theory
A striking example of order effects in this case occurs when the
initial state is |ψ⟩ = |1⟩. We have
Despite all that we have said above, it is reasonable to feel like
we have not really scratched the surface of what quantum theory 
1
2
1
p(+, 0) = |P0 P+ |1⟩ | =  (|0⟩ ⟨0| + |0⟩ ⟨1|) |1⟩ =
2

means as a theory of cognition. What does superposition really  (42)
mean? Why are certain variables incompatible and what does that 2 4
mean for the way they are actually encoded in the brain? If wave but,
function collapse is not exactly Bayesian updating then what is it?  2
These questions about the meaning of the theory are in some ways 1
p(0, +) = |P+ P0 |1⟩ |2 =  (|0⟩ ⟨0| + |1⟩ ⟨0|) |1⟩ = 0.

(43)
the most interesting and important. 2
Unfortunately we cannot cover these questions in this tutorial.
If we examine the commutator, we see,
The reason is essentially twofold; first, the primary purpose of
this tutorial is to explain how to use quantum theory to model 1
certain effects. Furthermore, addressing the meaning of the theory
P0 P+ = (|0⟩ ⟨0| + |0⟩ ⟨1|)
2
is generally best done once one has a good grasp of the way the (44)
1
theory works in practice, and so it is more profitable for the reader P+ P0 = (|0⟩ ⟨0| + |1⟩ ⟨0|)
2
to return to this question once they have more experience with
so
how the theory is used.
Second, there is in fact no general agreement on how to in- 1
[P0 , P+ ] = (|0⟩ ⟨1| − |1⟩ ⟨0|) ̸= 0. (45)
terpret quantum theory as a theory of cognition. The confusion is 2
partly inherited from physics, since the interpretation of quantum However this masks a more complicated structure. The com-
theory in physics is hardly without controversy, but it is also wors- mutator of two operators is itself an operator. There may be some
ened by the fact that quantum theory is generally regarded as an states for which the expectation value of this operator is 0, and thus
effective theory of cognition, where the actually neuroprocessing for which the order of operations does not matter, even though the
may take place in an essentially classical way. Thus, in cognition, commutator does not vanish identically. Indeed, it is easy to see
quantum theory is regarded as an effective theory possibly aris- that,
ing from an underlying classical process. For various reasons this
cannot be the case in physics (Bell, 2004; Isham, 1995), and thus ⟨0| [P0 , P+ ] |0⟩ = ⟨1| [P0 , P+ ] |1⟩ = 0. (46)
it is not obvious whether the interpretative framework associated On the other hand, we can show that the largest order effects come
with quantum theory can be directly imported to quantum cogni- from the two eigenstates of the commutator, |±i⟩, where,
tion in the same way as the mathematics (see beim Graben & At-
manspacher, 2006; Blutner & beim Graben, 2014 for some attempts 1 1
|+i⟩ = √ (|0⟩ + i |1⟩), |−i⟩ = √ (|0⟩ − i |1⟩) (47)
to interpret quantum cognition as arising directly from classical 2 2
neurophysics). We hope that further research will help us to un-
derstand why quantum theory works, and what it tells us about and we have,7
the underlying process of cognition. i
⟨±i| [P0 , P+ ] |±i⟩ = ± . (48)
2
4. Order effects I
In addition we have the following very important property: Con-
sider any two operators A and B, and look at the average value of
The basic framework we have outlined so far can be used
their commutator across a complete basis of H . This is equivalent
to build quantum models for any particular setup one wishes.
to computing the expectation value of the commutator in a state
However it is useful to look at specific examples to see how the
where ρ = D1 1. Such a state corresponds to an ensemble where
theory works in practice. For the rest of this tutorial we will
participants are equally likely to be in any initial knowledge state,
therefore focus on the particular example of order effects, but bear
i.e. there are no privileged states. We see,
in mind this is just one possible application of these ideas.
Order effects are a central feature of quantum theory because Tr([A, B]1) = Tr(AB − BA) = 0. (49)
of the non-commutation of operators. Recall; if A and B are two In other words, this type of order effect averages out across Hilbert
operators on H then the commutator [A, B] is defined as, space. The presence of a particular order effect is therefore as much
[·, ·] : H × H → H
(40)
[A, B] → AB − BA.
7 It is easy to see that for two Hermitian operators A and B, [A, B] is anti-Hermitian,
A and B are said to commute if [A, B] = 0. Otherwise they are said i.e. [A, B]Ď = −[A, B]. Anti-Hermitian operators have pure imaginary eigenvalues,
to be non-commuting. a fact that follows from a simple modification of Theorem 1 of Appendix.
about the initial state as it is about the operators in question. (We The eigenvectors can be computed if required, but they are not
will see in a later section an example of a different type of order necessary here. The key is that the maximum possible order effects
effect which does not average out in this way.) that we can observe are given by the eigenvalues of this operator.
The discussion so far might lead you to conclude that commu- It is worth looking in more detail at how these eigenvalues
tation is a binary property, two operators either commute or they behave with different values of θ . Firstly they are zero for θ =
do not. This is technically true, but it is more helpful to think of 0, π /2 as expected. The maximum value is,
commutation as a continuum. To explore this let us consider the
max(cos2 (θ ) sin(θ ))
projection operator Pθ , defined by
√ √
     
Ď −1 2 −1
Pθ = U (θ )P0 U (θ ) (50) = sin 2 tan 5−2 6 cos 2 tan 5−2 6
where, in the basis {|0⟩ , |1⟩}, ≈ 0.385. (60)

cos(θ ) − sin(θ )
 
Which occurs at
U (θ) = , (51)
sin(θ ) cos(θ )
√
 
so Pθ is a unitary rotation of P0 . In the basis {|0⟩ , |1⟩} we have, θ = 2 tan−1 5−2 6 ≈ 35.2◦ . (61)
cos2 (θ ) sin(θ ) cos(θ )

 
Somewhat surprisingly then, the angle between the projectors P0
Pθ = . (52)
sin(θ ) cos(θ ) sin2 (θ ) and Pθ at which the measured order effect is maximised is not the
same as the angle at which the commutator attains its maximum
We can show,
value, which recall as π /4 rad or 45°. Also, the maximum possible
 
1 0 1 order effect, in the sense of the difference between the probabilities
[P 0 , P θ ] = sin(2θ ) . (53) p(A and then B) and p(B and then A) is limited to about 0.385. This
2 −1 0
is significant because logically, since the probabilities themselves
The
 eigenvectors
 of [P0 , Pθ ] are equal to the eigenvectors of are bounded by 0, 1, one might have expected to be able to produce
0 1
−1 0
, i.e. they are the same as |±i⟩ above, but the eigenvalues order effects of any size from 0 to 1.
i sin(2θ) So far in this section, we have glimpsed how order effects may
are now given by λ± = ± 2 . The maximum order effects that be modelled, and we have seen that there are some non-trivial
we can observe are therefore given by, restrictions on the type of effects that can be accounted for. In
±i the next section we will introduce a precise test of the quantum
⟨±i| [P0 , Pθ ] |±i⟩ = sin(2θ ). (54) account of order effects which allows the quantum account to be
2
compared with other models.
Therefore although strictly P0 and Pθ do not commute unless θ is a
multiple of π /2, there is a sense in which they approximately com-
5. Order effects 2
mute for values of θ close to these. In a realistic experiment with
some low level of noise in the experimental data, for these values of
Designing a theory to fit existing data is relatively easy. What
θ the maximum possible order effects may be too small to observe
is much harder is to develop an a priori or parameter free test of
against the noise.
a theory which would allow the theory to be falsified. As we will
Now thus far we have been interested in the expected value
now show, quantum theory is very constrained with respect to the
of commutators as a measure of order effects. However when we
types of order effects it can account for. This leads to a remarkable
perform a sequence of measurements of two observables, A and B
equality between various probabilities which must be satisfied if
this is not in fact what we obtain. Rather the probability of getting
the probabilities are to be explained by quantum theory (Wang &
a positive result for the measurement of A followed by a positive
Busemeyer, 2013; Wang et al., 2014).
one for the measurement of B is given by,
Our aim in this section is to show in detail how this constraint,
p(A and then B) = Tr(PB PA ρ PA ) = |PB PA |ψ⟩|2 (55) which is known as the Quantum Question, or QQ Equality, is
derived. The derivation in the original papers was rather long, here
where the second equality holds when the initial knowledge state we will make use of the machinery of quantum theory we have
is pure. The probability for the same measurements in the opposite discussed in this tutorial to give a simpler and hopefully more
order is given by, transparent derivation.
p(B and then A) = Tr(PA PB ρ PB ) = |PA PB |ψ⟩|2 . (56) Suppose we have two different measurements we can per-
form, A or B, and crucially these measurements have two pos-
We are therefore interested in the quantity, sible outcomes, A, Ā and B, B̄. Denote the projection operators
p(A and then B) − p(B and then A) corresponding to these measurement outcomes as PA , PĀ , PB , PB̄ ,
with PA + PĀ = PB + PB̄ = 1. Let us write the probabilities
= Tr(PB PA ρ PA ) − Tr(PA PB ρ PB )
associated with these measurements as p(A, B), p(Ā, B) . . . where
= Tr({PA PB PA − PB PA PB }ρ). (57) p(A, B) = Tr (PB PA ρ PA ) etc.
The order effect we will observe in a measurement therefore Now since these measurements do not necessarily commute,
depends on the operator PA PB PA − PB PA PB . Note that this operator p(A, B) ̸= p(B, A). (62)
is Hermitian. Using the same 2D Hilbert space as above, and taking
PA = P0 and PB = Pθ we can show, in the {|0⟩ , |1⟩} basis, However, slightly unexpectedly, the following is true,
sin(θ ) − cos(θ ) p(A, B) + p(Ā, B̄) − p(B, A) − p(B̄, Ā) = 0. (63)

 
P0 Pθ P0 − Pθ P0 Pθ = cos2 (θ ) sin(θ ) . (58)
− cos(θ ) − sin(θ ) This is the QQ Equality (Wang & Busemeyer, 2013). (This equality
The eigenvalues of this operator can be easily, if somewhat also holds if we relabel A ↔ Ā, or B ↔ B̄, and the original form of
tediously, shown to be, the equality given in Wang & Busemeyer, 2013; Wang et al., 2014
had B ↔ B̄, but the form given above is better to see the essential
λ± = ± cos2 (θ ) sin(θ ). (59) symmetry of the equality.)
The QQ Equality holds because the following operator is 6. A brief introduction to POVMs
identically zero,
The measurements we have been dealing with so far are ideal
PA PB PA + PĀ PB̄ PĀ − PB PA PB − PB̄ PĀ PB̄ = 0. (64)
measurements, in the sense that they are perfectly idempotent
We give a proof of this below. and perfectly orthogonal. This means if the cognitive state of a
participant happens to be an eigenstate |ai ⟩ of an operator A, if we
This derivation makes use of the properties of commutators and
perform a measurement of the observable represented by A we will
projection operators, which together imply,
get the result ai with probability 1.
[PA , PB ] = [PA , (1 − PB̄ )] = −[PA , PB̄ ] = [PB̄ , PA ] Realistic experiments are never this straightforward, and
there is always some irreducible noise that means that real
= −[PB̄ , PĀ ] = [PĀ , PB̄ ] etc. (65)
measurements are not exactly ideal. The reason for this could
Note however that the use of commutators is just a mathematical be as simple as the presence of response errors; for example
convenience. if an experiment is performed under conditions where speed
is emphasised. Alternatively noise could enter because different
Now for the derivation. We begin with,
response options are not viewed as orthogonal, or because it is
[PA , PB ] − [PA , PB ] = 0. (66) impossible to separate out the cognitive variable one is interested
in from others which are also effectively ‘measured’ when a
Inserting two copies of the identity gives, decision is made. In order to account for this we need to generalise
the ideal measurements we have been using so far. The correct way
[PA , PB ]PA + [PA , PB ]PĀ − PB [PA , PB ] − PB̄ [PA , PB ] = 0. (67) to do this in quantum theory is via the use of what are known as
Positive Operator Valued Measures or POVMs for short.
Now we use the property of the commutator noted above, to get,
This tutorial is not the place to give a full introduction to
[PA , PB ]PA + [PĀ , PB̄ ]PĀ − PB [PA , PB ] − PB̄ [PĀ , PB̄ ] = 0. (68) POVM measurements.8 However the noise inherent in realistic
measurements can play an important part in determining whether
Expanding out the commutators gives, we observe order effects in experiments, and it is therefore worth
presenting a bare bones account that will enable us to see how
PA PB PA + PĀ PB̄ PĀ − PB PA PB − PB̄ PĀ PB̄ = 0 (69) noise in the measurements interacts with order effects.
Suppose we wish to model an experiment where we have
which is Eq. (64). Since this operator is identically zero, it follows
participants express a preference for one of two alternatives, A or B,
that for any density operator ρ ,
and that these are exhaustive and exclusive alternatives. In an ideal
Tr ({PA PB PA + PĀ PB̄ PĀ − PB PA PB − PB̄ PĀ PB̄ }ρ) = 0. (70) measurement these would be represented by projection operators
P0 = |0⟩ ⟨0|, P1 = |1⟩ ⟨1|. Suppose instead our measurement is not
By the linearity and cyclic property of the trace this gives, ideal but, intentionally or otherwise, is subject to some noise. This
means some participants who really prefer 0 will select option 1,
p(A, B) + p(Ā, B̄) − p(B, A) − p(B̄, Ā) = 0 (71) and vice versa.
Let us see how we might model this. What we want is an
where p(A, B) = Tr (PB PA ρ PA ) etc. operator E0 , whose expectation value in the state |0⟩ is close to one,
This QQ Equality was first introduced by Wang and Busemeyer but which also has a non-zero expectation value in the state |1⟩,
(2013), and it was strongly supported experimentally in a very and likewise for E1 . That is,
compelling paper by Wang et al. (2014). That the QQ Equality is
obeyed by real world data is surprising in view of the fact that ⟨0| E0 |0⟩ = 1 − ϵ, ⟨1| E0 |1⟩ = ϵ,
(72)
it does not follow from general contextual probability theories, ⟨0| E1 |0⟩ = ϵ, ⟨1| E1 |1⟩ = 1 − ϵ
but can only (apparently) be derived assuming the particular
where 0 ≤ ϵ ≤ 1 is some small error probability.9 Let us also
structures of quantum theory. This means observing that the QQ assume,
Equality is obeyed in the wild provides very good evidence that
human decision-making can be described by quantum theory ⟨0| E0 |1⟩ = 0, etc. (73)
in particular, rather than just a contextual probability theory in
general. In the basis {|0⟩ , |1⟩} these operators can therefore be written as,
It is worth asking exactly what properties of quantum theory
1−ϵ ϵ
   
0 0
the derivation relies on. That is, exactly what properties does the E0 = , E1 = . (74)
0 ϵ 0 1−ϵ
QQ Equality test? The key ones appear to be the properties of
the projection operators representing measurements, specifically Can we use these operators to describe a measurement process? It
the fact that they are orthogonal and idempotent. This raises the is easily seen that they are not projection operators, nevertheless
question of whether the QQ Equality still holds if the measurement they satisfy the following properties,
operators do not have these ‘ideal’ properties, such as one might
expect to be appropriate for more realistic measurements. We will
address this question in a later section. 8 The best book JMY has read on POVMs is Busch, Grabowski, and Lahti (1995),
Another property that this derivation relied upon was that the however this is currently out of print. A popular alternative is Neilsen and Chuang
measurements had binary outcomes. Despite extensive work, we (2000). See also Yearsley (in preparation).
have not been able to prove a generalisation of this result for 9 An important point here is that we are assuming the errors are unbiased, i.e. the
measurements with more than two possible outcomes. It remains probability of incorrectly answering 1 when the state is |0⟩ is the same as the
probability of incorrectly answering 0 when the state is |1⟩. This need not hold in
unclear whether any such analogue exists, but it would obviously
general, but this would describe a something other than pure noise, so we will ignore
be of huge interest to resolve this question. this possibility here.
• They are positive operators, which means they have positive

eigenvalues.
• They are complete, in the sense that E0 + E1 = 1, where, recall,
1 is the identity operator over H .
These properties mean that for any density operator,
0 ≤ Tr(Ei ρ) ≤ 1 (75)
and

Tr(Ei ρ) = 1. (76)
i=0,1
The quantities Tr(Ei ρ) can thus be interpreted as probabilities,

and so E0 and E1 are good candidates to describe a measurement
process.
But what measurement process do they describe? Well there
are many ways to think about this, but probably the easiest is to
note that we can write,
Fig. 1. pϵ (B and then +), pϵ (+ and then B), and their difference, plotted against ϵ .
E0 = (1 − ϵ)P0 + ϵ P1 , E1 = ϵ P0 + (1 − ϵ)P1 . (77)
In other words, we can write these operators like, We will replace the projection operators PA , PB , with the
 following POVMs,10
E0 = p0 (i)Pi (78)
ϵ
 
i 0
PB → EB =
0 1−ϵ
where p0 (i) have (loosely) the interpretation of probabilities. So
one way to think about these measurements is that instead of 1 1 − 2ϵ
 
(83)
performing a measurement P0 , we instead perform one of the 2 2
P+ → E+ =  .
 
possible measurements Pi with some probabilities p0 (i). So these 1 − 2ϵ 1
measurements look like noisy versions of ideal ones.
2 2
E0 and E1 are specific examples of elements of POVMs. The
These have associated measurement operators,
probability that a measurement described by Ei gives a positive
answer is then given by, √
ϵ 0

MB = √
p(i) = Tr(Ei ρ). (79) 0 1−ϵ
√ √ √ √ 
Suppose we perform a measurement of Ei , what is the resulting 1−ϵ+ ϵ 1−ϵ− e (84)
state after the measurement? Writing, 2 2 .
 
M+ = 
 √1 − ϵ − √e √ √ 
Ď 1−ϵ + ϵ
Ei = Mi Mi (80)
2 2
we can show that if we get a positive answer to the measurement Now we can see that,
described by Ei then the state collapses to,
pϵ (+ and then B) = Tr(EB M+ ρ M+ )
Ď
Mi ρ Mi 1 √ √
ρ′ = .

(81) = 1 − 2(1 − 2ϵ) ϵ 1 − ϵ (85)
Tr(Ei ρ) 4
The Mi are often called ‘measurement operators’. The decompo- and
sition of Ei into Mi is not unique; there in fact many different
pϵ (B and then +) = Tr(E+ MB ρ MB )
realisations of a given POVM in terms of a set of measurement
ϵ
operators M√ i . We ignore this issue here, and simply take the sim- = . (86)
plest, Mi = Ei . For more details see Yearsley (in preparation) and 2
Busch et al. (1995). We plot these results against the value of ϵ in Fig. 1. The results
To return to our example above, in the basis {|0⟩ , |1⟩} the are interesting. The key is that the difference in the values of the
measurement operator associated with E0 will be, probabilities (plotted as the dotted line) decreases sharply with
increasing ϵ , i.e. with increasing noise. Note however that the
1−ϵ
√ 
M0 = √0 (82) value of ϵ is interpretable in terms of the ‘error’ probability of the
0 ϵ measurement. Realistic experiments would probably have values
of ϵ in the range 1%–5%, and so order effects are still likely to be
which is nice and simple.
visible in such experiments, although they might appear smaller
than one might have expected.
7. Order effects and POVMs We do not have space here to pursue this further, but it is clear
that small amounts of noise will still allow order effects to be
We are now ready to explore the extent to which the presence
and size of order effects depend on whether the measurements
are perfect. We will start with the simple case, and then move 10 Readers are encouraged to convince themselves E is reasonable. Either start
+
on to consider whether the QQ Equality still holds for imperfect with EB and rotate through π/4, or consider a combination of P+ and P− as in Eq.
measurements. (78).
observed, even though very large amounts of noise rapidly kill off The first term on the right hand side of this equation is proportional
such effects. This has important implications for studies looking for to the original QQ operator Eq. (64) and thus vanishes. The
these effects in the wild. second term is also proportional to the QQ operator, with B and B̄
If the noise inherent in realistic measurements can kill off order interchanged, and thus this also vanishes. What we are left with is,
effects, can it also disrupt the important QQ Equality that we
introduced previously to test the quantum explanation for such MA EB MA + MĀ EB̄ MĀ − MB EA MB − MB̄ EĀ MB̄
effects? It is obviously vital to know the answer to this if we want to = (PA PB PĀ + PĀ PB PA + PA PB̄ PĀ + PĀ PB̄ PA )
understand whether we expect to see this equality obeyed under √ √
× ϵ 1 − ϵ − (PB PA PB̄ + PB̄ PA PB + PB PĀ PB̄
realistic conditions.
√ 
What we want to do is to examine what happens to the operator + PB̄ PĀ PB ) η 1 − η. (93)
in Eq. (64) when we replace the projectors with POVMs. Clearly the
derivation that led to Eq. (64) will no longer hold, but it is unclear By using the properties of projection operators, we see e.g.,
whether this means the equality is no longer valid.
PA PB PĀ = PA (1 − PB̄ )PĀ = −PA PB̄ PĀ (94)
We are going to restrict ourselves in what follows to a very
simple class of POVMs, in particular we will take, in the basis etc. From this we can see that the both the terms on the right hand
{|A⟩ , |Ā⟩} side of Eq. (93) also vanish, and so finally we see,
1−ϵ ϵ
   
0 0 MA EB MA + MĀ EB̄ MĀ − MB EA MB − MB̄ EĀ MB̄ = 0. (95)
EA = , EĀ = (87)
0 ϵ 0 1−ϵ
Therefore for any density operator ρ we have,
and in the basis {|B⟩ , |B̄⟩}
Tr({MA EB MA + MĀ EB̄ MĀ − MB EA MB − MB̄ EĀ MB̄ }ρ) = 0 (96)
1−η η
   
0 0
EB = , EB̄ = . (88) and thus,
0 η 0 1−η
We will also chose the measurement operators to have a simple pϵ,η (A, B) + pϵ,η (Ā, B̄) − pϵ,η (B, A) − pϵ,η (B̄, Ā) = 0 (97)
form. Eg in the {|A⟩ , |Ā⟩} basis
where pϵ,η (A, B) = Tr(EB MA ρ MA ) etc.
1−ϵ
√ 
Therefore we have shown the QQ Equality still holds when the
MA = √0 . (89) measurements are given by these simple types of POVM. This is an
0 ϵ
important result, since it helps to explain why the equality seems
This is useful because it means we can write, to hold so well in real world data sets (Wang et al., 2014).
√ √ Note that this result relied on the POVMs having unbiased
EA = (1 − ϵ)PA + ϵ PĀ , MA = 1 − ϵ PA + ϵ PĀ (90)
errors. If we relax this condition then the QQ equality no longer
etc. This choice of POVM is special because of the fact that the EA holds. This corresponds, for example, to having systematic error
etc. can be written as sums of commuting projection operators. that biases the result towards or away from A, and similarly for B.
This means they are in some sense very close in behaviour to the We will not explore this possibility further here.
more usual projective measurements.11 The significance of this is
In summary, using realistic noisy measurements does effect the
discussed in Busch et al. (1995).
size and presence of order effects. In general adding noise to a
We will now see what happens when we replace the projectors
quantum system tends to make it look more classical, and this case
with POVMs in Eq. (64). We will use the trick of replacing the
provides a good example of that. (For a good introduction to the
POVMs with sums of projection operators, as above. This then lets
quantum to classical transition see Halliwell, 2005. For a classic
us write,
review we recommend Zurek, 1991.) However for realistic levels
MA EB MA + MĀ EB̄ MĀ − MB EA MB − MB̄ EĀ MB̄ of noise order effects should still be visible, and crucially the QQ
√ √ Equality will still hold provided the noisy measurements remain
= ( 1 − ϵ PA + ϵ PĀ )((1 − η)PB + ηPB̄ )
√ √ √ √ unbiased.
× ( 1 − ϵ PA + ϵ PĀ ) + ( ϵ PA + 1 − ϵ PĀ )
√ √
× (ηPB + (1 − η)PB̄ )( ϵ PA + 1 − ϵ PĀ ) 8. Order effects in similarity judgements
√
− ( 1 − ηPB + ηPB̄ )((1 − ϵ)PA + ϵ PĀ )

√ √ In this final section we want to explore the presence of order
× ( 1 − ηPB + ηPB̄ ) − ( ηPB + 1 − ηPB̄ )
 
effects in a slightly different context, namely that of similarity
√ judgements. Similarity judgements are an important part of many
× (ϵ PA + (1 − ϵ)PĀ )( ηPB + 1 − ηPB̄ ).

(91) other more complicated cognitive processes such as categorisation.
Some rather tedious algebra gives, In a classic paper Tversky (1977) demonstrated that human
similarity judgements have a number of properties that one would
MA EB MA + MĀ EB̄ MĀ − MB EA MB − MB̄ EĀ MB̄
not expect from the simplest classical models. The most important
= (PA PB PA + PĀ PB̄ PĀ − PB PA PB − PB̄ PĀ PB̄ ) for us is that similarity can sometimes display asymmetry, in other
× ((1 − ϵ)(1 − η) + ϵη) + (PA PB̄ PA + PĀ PB PĀ words, if Sim(A, B) denotes the similarity of A to B, then,
− PB PĀ PB − PB̄ PA PB̄ )((1 − ϵ)η + ϵ(1 − η)) Sim(A, B) ̸= Sim(B, A). (98)
+ (PA PB PĀ + PĀ PB PA + PA PB̄ PĀ + PĀ PB̄ PA )
√ √ In particular, so the theory goes (Tversky, 1977), Sim(A, B) ≥
× ϵ 1 − ϵ − (PB PA PB̄ + PB̄ PA PB + PB PĀ PB̄ Sim(B, A) if we have more knowledge about B than about A.
√  In an attempt to model this interesting finding, Pothos, Buse-
+ PB̄ PĀ PB ) η 1 − η. (92)
meyer, and Trueblood (2013) came up with a model of similarity
judgements based on quantum theory (see also Pothos et al., 2015).
Their model works as follows, suppose we have a space of concepts
11 We are grateful to an anonymous referee for emphasising this point. H , with subspaces representing the concepts A and B defined by
the projection operators12 PA and PB . Then the judged similarity of As our example, we will take a three dimensional Hilbert space
A to B, given our initial knowledge state is ρ , is given by, of possible thoughts, and PA and PB to be one and two dimensional
respectively. It is useful to choose a basis of eigenstates of B, so that,
Sim(A, B) = Tr(PB PA ρ PA )/Tr(PA ρ) (99)
0 0 0
 
or PB = 0 1 0 . (104)
2 0 0 1
|PB PA |ψ⟩|
Sim(A, B) = (100) The bases formed by the eigenstates of A and B are related by a
|PA |ψ⟩|2
unitary transform U, so in the basis of eigenstates of B we can also
if the state is pure.13 In actual fact Pothos et al. did not include the write,
normalisation factor in the denominator of Eq. (99). However they
1 0 0
 
left the question of normalisation open, so we may explore any rea- Ď
PA = U 0 0 0 U = U Ď PB̄ U . (105)
sonable choice of normalisation.
0 0 0
The initial knowledge state is to be set such that the state is
‘unbiased’. Specifically ‘‘in the absence of priming manipulation The relationship between PA and PB , and therefore the asymmetry,
or contextual influence, we require the state vector to be neutral depends on the unitary transformation U.
between the compared concepts’’ (Pothos et al., 2013). One might There appear to be several possible choices for U. However the
expect that this means the initial state is set such that, only important component of any transformation is that which ro-
tates PA towards the plane defined by PB . Anything else is equiv-
Tr(PA ρ) = Tr(PB ρ) (101) alent to a relabelling of the coordinates. Therefore without loss of
generality we can choose the transformation to simply rotate PA
and this is indeed what Pothos et al. had in mind. However there
towards one of the states in PB . Since the setup is symmetric we
are other ways to interpret this. They go on to say ‘‘Such an
can choose,
assumption is equivalent to that of a uniform prior in a Bayesian
model’’. Now the concept of a uniform prior is rather slippery, cos(θ ) − sin(θ ) 0
 
because it depends on what exactly it is uniform over. In this case U = sin(θ ) cos(θ ) 0 (106)
there are two obvious choices; the prior could be uniform over 0 0 1
the whole knowledge space, so one is equally likely to be thinking
this gives,
about any thought in the space, or it could be uniform over the
concepts under consideration, which, recall, are represented by cos2 (θ ) − cos(θ ) sin(θ )
 
0
subspaces of dimension potentially greater than one. A uniform PA = − cos(θ ) sin(θ )
 sin2 (θ ) 0 . (107)
prior over the whole knowledge space would be given by the 0 0 0
density operator ρ = D1 1, whereas a uniform prior over the
concepts under consideration would be given by a density operator Now we need to ensure we have an allowable initial state. Recall
satisfying Eq. (101). One advantage of the former is that it is we have two choices, either our initial state is proportional to the
guaranteed to exist, whereas the latter is not. At any rate, in the identity, or we must choose it so that ⟨PA ⟩ = ⟨PB ⟩.
absence of further information we will explore both possibilities. Let us start with the simpler case, of ρ = 13 1. Clearly,
This similarity measure looks like a quantum version of a
1
conditional probability. It has two nice properties, it is normalised Tr(PB PA ρ PA ) = Tr(PA PB ) = Tr(PA PB ρ PB ) (108)
in the sense that, 3
so the numerators of the two similarities, Eq. (99), are equal.
Sim(A, A) = 1 (102) However we can also see,
and also in the sense that, 1 1 1 2
Tr(PA ρ) = Tr(PA ) = , Tr(PB ρ) = Tr(PB ) = . (109)
 3 3 3 3
Sim(A, Bi ) = 1 (103)
i
Thus we have,

if the set {Bi } is complete, i.e. i Bi = 1. Sim(A, B) Tr(PB PA ρ PA )Tr(PB ρ)
= = 2. (110)
Now the key factor responsible for asymmetry in this model is Sim(B, A) Tr(PA PB ρ PB )Tr(PA ρ)
that PA and PB may have different dimensions. This corresponds We see therefore that in the case Sim(A, B) is always larger than
in cognitive terms to a participant having greater knowledge Sim(B, A) by a factor of two. In general the ratio of the similarities
about one of the stimuli. Greater knowledge implies more possible is the inverse of the ratio of the dimension of the subspaces, and
thoughts connected with that stimuli, which means a higher thus the inverse of the ratio of the ‘degree of knowledge’ we have
dimensional subspace for that stimuli. What we want to do is about A and B, as desired.14 Turning back to our specific example,
explore how the asymmetry depends on the relationship between Sim(A, B) is easily computed to be sin2 (θ ).
the different subspaces corresponding to PA and PB . A full analysis In summary, this choice of initial state is interesting because
is given in Pothos et al. (2013), what we want to do is explore a it always gives rise to the expected asymmetry. It also produces
simple analytically tractable example. a particularly simple expression for the similarities. However it
is an open question whether the observed asymmetry shares this
behaviour.
12 One might also use smeared projectors or POVMs instead of projection
operators if there is some uncertainty about the concept boundary.
13 This assumes that the similarity between a stimulus and itself is always judged 14 This way of achieving asymmetries can also be implemented classically. Pothos
to be 1. If desired, this assumption can be dropped by using POVMs in place of the et al. (2013) argue that this might not represent a good solution. However we are
projection operators. only interested here in exploring the structure of the theory.
Now let us turn to the more challenging case. It is easiest to start

by looking at the eigenvectors of A − B. Once we have these we can
take combinations that give zero expectation value for A − B, and
which are thus unbiased. In the B basis the eigenvectors are given
by,
(cos(θ ) − 1)
 
0
 
|v1 ⟩ = 0 , |v2 ⟩ = N2  sin(θ )  ,
 
1
1
0
(111)
(cos(θ ) + 1)
 
|v3 ⟩ = N3  sin(θ ) 
 
1
0
(a) Similarities for the choice a = 0. The two similarities coincide in this
with eigenvalues, −1, − cos(θ ) and cos(θ ) respectively. Here N2 case, so there is no asymmetry.
and N3 are normalisation factors given by,
 
sin2 (θ ) sin2 (θ )
N2 = , N3 = . (112)
2(1 − cos(θ )) 2(1 + cos(θ ))
We will choose a pure initial state for simplicity, and we will also
choose the state to be real. Then the following (unnormalised) set
of states have zero expectation value with respect to the operators
A − B,
 
|ψa ⟩ = a cos(θ ) |v1 ⟩ + 1 − a2 |v2 ⟩ + |v3 ⟩ (113)
where 0 ≤ a ≤ 1. Thus these states have the same expectation
value for the operators A and B and are ‘unbiased’ in the sense of
Pothos et al.. Note that there is not a single state with this property,
but crucially rather a whole family of possible states. (b) Similarities for the choice a = 1. Sim(A, B) is always greater than
Now we can compute the similarities Sim(A, B) etc. according Sim(B, A), apart from the end points θ = 0, π/2.
to Eq. (99). We save the reader the algebra involved in this, but
note it is quite straightforward to run in MatLab or similar. We plot Fig. 2. Plots of Sim(A, B) and Sim(B, A) against θ for two choices of a.
the special cases of a = 0 and a = 1 in Fig. 2, the general case
interpolates between these two extremes. The important point is etc. We hope this tutorial has convinced you that quantum mod-
that we generally have Sim(A, B) ≥ Sim(B, A), with equality in the els represent a useful and potentially powerful way to model these
case a = 0 and for the values θ = 0, π /2. phenomena. The existence of order or asymmetry effects strongly
We see that with this choice of initial state the similarities show suggests that theories beyond classical probability theory are
a rather more complex behaviour. The ratio of the similarities is needed to model this type of decision making. However simply ex-
now a function of θ , as well as of the new parameter a. Thus changing non-contextual for contextual probability theories leaves
in this case the degree of asymmetry predicted depends on how us with too many free parameters, and no underlying sense about
incompatible A and B are. why these effects occur. Quantum models, on the other hand, are
The purpose of this section was to introduce a different type both more constrained than the general case of a contextual prob-
of order effect, which occurs in models of similarity judgements, ability theory, and also offer the possibility of achieving an under-
that behaves in a slightly different way to the order effects we see standing of why such effects occur in terms of incompatibility of
in survey designs. Recall, measured order effects in survey designs mental representations. So far, it has been the success of quantum
result because the operator, models at predicting observed constraints on various phenomena
that have been most persuasive. In the future, we hope that these
∆ = PA PB PA − PB PA PB (114) models will also impress by the story they can tell about the rea-
is non-zero. However, although ∆ ̸= 0 it is traceless, which sons behind the occurrence of other puzzling violations of classical
means that the average of ∆ over the whole Hilbert space is probability theory in human decision making.
zero. In other words, order effects in survey designs are generally As well as having a particular focus on order effects, we hope
this tutorial has also been useful to researchers looking to construct
balanced, for every cognitive state which gives rise to a particular
quantum models for other purposes. We have strived to set out
asymmetry, there is another state which gives the opposite
in as clear a way as possible how and why quantum models are
result. In contrast, the order effects that produce asymmetries in
constructed to address particular questions. While it is impossible
similarity judgements do not have this property, and it can be that
to cover every possible scenario one might encounter when trying
case that Sim(A, B) ≥ Sim(B, A) for all states in the cognitive space.
to set up a model, hopefully this tutorial will have given the reader
the confidence to experiment with constructing quantum models
9. Summary of their own.
Order effects are an important feature of the study of human Acknowledgments

decision making. But order effects are just one of many different
kinds of phenomena addressed by quantum models of cognition. JRB was supported by a grant from the Air Force Office of Scien-
Others include the conjunction fallacy, violations of total probabil- tific Research (FA9550-12-1-0397). JMY was supported by Lever-
ity, over and under extension effects in conceptual combinations, hulme Trust grant RPG-2013-00 and by NSF grant SES-1556415.
Appendix A. Review of linear algebra A useful class of operators in quantum theory are those which
are Hermitian or more generally Self-Adjoint.15 An operator is
In this appendix we give a brief review of some relevant Hermitian if,
concepts from linear algebra. Our aim is not to cover all possible
AĎ = A. (A.10)
topics, but just to recap some key ideas. In particular we will show
how some familiar ideas may be expressed in a basis independent We can now state a number of useful theorems about Hermitian
way, which means they can be computed directly in terms of the operators.
elements of Hilbert space.
A Hilbert space H is essentially a complex vector space which Theorem 1. The eigenvalues of an Hermitian operator are real.
is a complete metric space with respect to the inner product, Proof. Let |ai ⟩ be any non-zero eigenvector of A, with eigenvalue
ai .
⟨·|·⟩ : H × H → C
(A.1) A |ai ⟩ = ai |ai ⟩ . (A.11)
(|u⟩ , |v⟩) → ⟨u|v⟩
Then,
with the usual properties (see Isham, 1995; Reed & Simon, 1980 for
a fuller account). The elements of H are denoted |u⟩ and are known ⟨A |ai ⟩ , |ai ⟩⟩ = ⟨ai |ai ⟩ , |ai ⟩⟩ = a∗i ⟨|ai ⟩ , |ai ⟩⟩
as ‘kets’.
= a∗i ⟨ai |ai ⟩ . (A.12)
An operator on Hilbert space is a map from the space to itself.
But because A is Hermitian,
A:H →H
⟨A |ai ⟩ , |ai ⟩⟩ = |ai ⟩ , AĎ |ai ⟩ = ⟨|ai ⟩ , A |ai ⟩⟩
 
(A.2)
|u⟩ → A |u⟩ .
= ⟨|ai ⟩ , ai |ai ⟩⟩ = ai ⟨ai |ai ⟩ . (A.13)
An important operator is the identity operator denoted 1. It is
defined by, Thus,
1A = A1 = A, ∀A : H → H . (A.3) 0 = (ai − a∗i ) ⟨ai |ai ⟩ (A.14)

and since |ai ⟩ is non-zero, we must have ai = a∗i .
The eigenvectors, {|ai ⟩} and eigenvalues ai of A are defined by,
Note that the converse is also true, i.e. an operator is Hermitian
A |ai ⟩ = ai |ai ⟩ . (A.4) if its eigenvalues are real.
The object ⟨u| is known as a ‘bra’. It is not technically an element of
H , but can be thought of either as an element of the dual space, or Theorem 2. The eigenvectors corresponding to two different eigen-
as a functional that maps elements of H to the complex numbers, values are orthogonal.
Proof. Let
⟨u| : H → C
|v⟩ → ⟨v|u⟩ .
(A.5) A |a1 ⟩ = a1 |a1 ⟩ , A |a2 ⟩ = a2 |a2 ⟩ (A.15)
with a1 ̸= a2
Hermitian conjugation of an operator is a useful concept in
Then,
linear algebra. The Hermitian conjugate of an operator AĎ is defined
by, ⟨|a2 ⟩ , A |a1 ⟩⟩ = ⟨a2 , a1 |a1 ⟩⟩ = a1 ⟨a2 |a1 ⟩ . (A.16)
⟨A |u⟩ , |v⟩⟩ = |u⟩ , AĎ |v⟩ ,

 
∀ |u⟩ , |v⟩ ∈ H . (A.6) But because A is Hermitian the left hand side is,
⟨A |a2 ⟩ , |a1 ⟩⟩ = ⟨a2 |a2 ⟩ , |a1 ⟩⟩ = a∗2 ⟨a2 |a1 ⟩ = a2 ⟨a2 |a1 ⟩ . (A.17)
In a particular basis, where A is represented by some matrix,
Thus,
AĎ = (AT )∗ (A.7)
(a1 − a2 ) ⟨a2 |ai ⟩ = 0 (A.18)
where AT is the transpose of A.
and since a1 ̸= a2 , we must have ⟨a2 |a1 ⟩ = 0.
Hermitian conjugation has the following properties,
Another important class of operators are those which are
1. (AĎ )Ď = A Unitary. An operator U is unitary if,
2. (A + B)Ď = AĎ + BĎ
3. (AB)Ď = BĎ AĎ U Ď U = UU Ď = 1 (A.19)
4. (cA)Ď = c ∗ AĎ for c ∈ C. Ď −1
i.e. U = U . The exponential of (i times) an Hermitian operator
is a Unitary operator,
It is useful to define,
Ď Ď Ď
= e(iH ) = e−iH = e−iH = eiH
 −1
eiH .

⟨u| = |u⟩Ď . (A.8) (A.20)
(We will explain how to define the exponential of an operator
From Eq. (A.4) we can see that, below.)
A |ai ⟩ = ai |ai ⟩
→(⟨ai | A)Ď = (⟨ai | ai )Ď (A.9)
15 The distinction is only important for Hilbert spaces with an infinite number of
→AĎ |ai ⟩ = a∗i |ai ⟩ . dimensions. In these spaces it is not necessarily true that an operator and its adjoint
have the same domain. Thus self-adjoint means that the domains of an operator and
Thus the eigenvectors of AĎ are the same as those of A, but the its adjoint match, and that the operators are equal on this shared domain (Reed &
eigenvalues are the complex conjugates. Simon, 1980).
A final important class of operators are those which are Projec- Since this decomposition of |ψ⟩ can be performed in any basis,
tion operators or projectors. An Hermitian operator P is a projector we can write any state as a superposition of the eigenstates of any
if, observable in our theory.
One important consequence of the spectral theorem is that
P2 = P. (A.21) it lets us compute/define functions of operators. For example
We are now ready to state probably the most important theo- suppose we want to define A2 where A is some Hermitian operator.
rem in quantum theory. Well we can write,
D
 D

Theorem 3. The Spectral Theorem for Hermitian (Self-Adjoint) A2 = ai |ai ⟩ ⟨ai | aj |aj ⟩⟨aj |
Operators. i =1 j =1
The set of all eigenvectors of a Hermitian (self-adjoint) operator A D
defined on a Hilbert space H of dimension D forms an orthonormal

= ai aj ⟨ai |aj ⟩ |ai ⟩ ⟨aj |
basis for H . Thus, any vector |ψ⟩ ∈ H can be expanded as, i,j=1
D
 D D
⟨ai |ψ⟩ |ai ⟩ .
 
|ψ⟩ = (A.22) = ai aj δij |ai ⟩ ⟨aj | = a2i |ai ⟩ ⟨ai | (A.27)
i=1 i,j=1 i=1
The operator A itself can be written in the important form, which is very simple. In general,
D
 D
ai |ai ⟩ ⟨ai | .

A= (A.23) An = ani |ai ⟩⟨ai |. (A.28)
i=1 i=1
That is, any Hermitian (self-adjoint) operator may be written as a sum This suggests the following, suppose we take any function f (·),
of projection operators onto subspaces spanned by the eigenvalues which can be defined in terms of a power series, i.e.
of A, each projection operator being multiplied by the associated
eigenvalue. f (x) = α0 + α1 x + α2 x2 + . . . (A.29)
Proof. The spectral theorem is a classical result in linear algebra. then we can define,
See any good book on linear algebra if you are interested. f (A) = α0 + α1 A + α2 A2 + . . .
Note: This analysis assumes the eigenvalues are not degenerate, D

α0 + α1 ai + α2 a2i + . . . |ai ⟩ ⟨ai |
 
i.e. that each eigenvector has a unique eigenvalue. All the theorems =
in this appendix can be trivially extended to the case where the i=1
spectrum of A is degenerate, see any good textbook on linear D

algebra or quantum theory Isham (1995) and Reed and Simon = f (ai ) |ai ⟩ ⟨ai | . (A.30)
(1980). i=1
This theorem is important because Hermitian operators and
With sufficient care this can then be extended to essentially all
their associated eigenspaces play a key role in quantum theory
functions. So we can take square roots, exponentials, trig functions
because of the association of Hermitian operators with observables.
of operators etc. with abandon.
The spectral theorem says that any observable quantity defines a
One final operation we need to define is trace of an operator.
basis for the knowledge space, so that the space can be partitioned
into subspaces each of which has a well defined value of the

Tr(A) = ⟨φi | A |φi ⟩ (A.31)
observable. i
An important example of a Hermitian operator is the identity
operator, 1. This can be thought of as an operator where all where the {|φi ⟩} form an orthonormal basis for H . The trace of an
the eigenvalues are 1. Since it is Hermitian the spectral theorem operator is independent of the basis in which it is computed. The
applies, and so we can write, trace operation has the following useful properties,
D
1. Tr(ABC ) = Tr(BCA) = Tr(CAB), etc.
2. Tr (A |ψ⟩ ⟨ψ|) = ⟨ψ| A |ψ⟩.

1= |φi ⟩ ⟨φi | (A.24)
i=1 In a particular basis, where A is represented by a matrix, the Trace
where the {|φi ⟩} are any orthonormal basis for the space. The trick of A is simply the sum of the diagonal elements of the matrix. One
of inserting copies of the identity is often used in computations. important consequence is that the trace of a projection operator is
We can use this result to show that any state in the Hilbert space simply the dimension of the subspace onto which it projects.
can be expanded in terms of the eigenstates of any observable.
Consider an observable A, from the above we know, Appendix B. Number of degrees of freedom associated with
states, operators, etc.
D

1= |ai ⟩ ⟨ai | . (A.25) Let us briefly explain how to compute the number of parame-
i=1
ters associated with some objects in quantum theory. This is use-
Now consider this operator acting on any state vector, ful information for understanding the number of measurements
needed to fix a particular state, or indeed how many parameters
D D
  a quantum theory has in total, to compare with a classical theory.
|ψ⟩ = 1 |ψ⟩ = |ai ⟩ ⟨ai |ψ⟩ = ψiA |ai ⟩ . (A.26)
A pure state |ψ⟩ is a complex vector in a D dimensional space.
i=1 i=1
As a column vector in a particular basis it would have D complex
The square norms of the ψiA can be interpreted as the probabilities entries, which means 2D parameters to fix. However we also have
that a measurement of the observable A would give the result ai . two constraints. The first is that the state be normalised, the second
is that the overall phase of a state is unobservable, so we can set it Isham, C. (1995). Lectures on quantum theory. Imperial College Press.
to 0 (Isham, 1995). This means a pure state has a total of 2D − 2 Khrennikov, A. (2010). Ubiquitous quantum structure: from psychology to finance.
Berlin: Springer.
free parameters. Kvam, P. D., Pleskac, T. J., YU, S., & Busemeyer, J. R. (2015). Interference effects
An observable is an Hermitian operator, which in a particular of choice on confidence: Quantum characteristics of evidence accumulation.
basis can be written as a D × D matrix with complex entries, which Proceedings of the National Academy of Sciences, 112, 10645.
McKemmish, L. K., Reimers, J. R., McKenzie, R. H., Mark, A. E., & Hush, N. S.
would require 2D2 parameters to specify. However we have the (2009). Penrose–Hameroff orchestrated objective-reduction proposal for
constraint that an observable must be Hermitian, which in terms human consciousness is not biologically feasible. Physical Review E, 80(2),
of matrix elements means Aij = A∗ji . This reduces the number of 021912.
Mogiliansky, A. L., Zamir, S., & Zwirn, H. (2009). Type indeterminacy: A model of
parameters to be specified to D2 . the KT (Kahneman–Tversky)-man. Journal of Mathematical Psychology, 53(5),
Finally, a density operator is an Hermitian operator, together 349–361.
Moore, D. W. (2002). Measuring new types of question order effects. Public Opinion
with the condition that it must be trace normalised. (Tr(ρ) = 1) Quarterly, 66(1), 80–91.
This is one extra constraint, so reduces the number of parameters Neilsen, M., & Chuang, I. (2000). Quantum computation and quantum information.
by one. The number of free parameters in a density operator is Cambridge University Press.
Peres, A. (1998). Quantum theory: concepts and methods. Dordrecht, The Nether-
therefore D2 − 1. lands: Kluwer Academic.
Plenio, M. 2002. Quantum Mechanics. Lecture notes available at: www3.imperial.
ac.uk/pls/portallive/docs/1/613904.PDF.
References Pothos, E. M., Barque-Duran, A., Yearsley, J. M., Trueblood, J. S., Busemeyer, J.,
& Hampton, J. A. (2015). Progress and current challenges with the quantum
similarity model. Frontiers in Psychology, 6, 205.
Aerts, D. (2009). Quantum structure in cognition. Journal of Mathematical
Pothos, E. M., & Busemeyer, J. R. (2009). A quantum probability explanation for
Psychology, 53(5), 314–348.
violations of ‘rational’ decision theory. Proceedings of the Royal Society B, 276,
Aerts, D., Gabora, L., & Sozzo, S. (2013). Concepts and their dynamics: A quantum-
2171.
theoretic modeling of human thought. Topics in Cognitive Science, 5(4), Pothos, E. M., & Busemeyer, J. R. (2013). Can quantum probability provide a new
737–772. direction for cognitive modeling? Behavioral and Brain Sciences, 36(3), 255–274.
Atmanspacher, H., & Filk, T. (2010). A proposed test of temporal non-locality in
Pothos, E. M., Busemeyer, J. R., & Trueblood, J. S. (2013). A quantum geometric model
bistable perception. Journal of Mathematical Psychology, 54, 314–321.
of similarity. Psychological Review, 120(3), 679.
Atmanspacher, H., & Römer, H. (2012). Order effects in sequential measurements of
Reed, M., & Simon, B. (1980). Methods of modern mathematical physics (Vol. 1).
non-commuting psychological observables. Journal of Mathematical Psychology,
London: Academic Press.
56(4), 274–280. Schwinger, J. (1960). Unitary operator bases. Harvard University.
beim Graben, P., & Atmanspacher, H. (2006). Complementarity in classical Tegmark, M. (2000). Importance of quantum decoherence in brain processes.
dynamical systems. Foundations of Physics, 36(2), 291–306. Physical Review E, 61(4), 4194–4206.
Bell, J. S. (2004). Speakable and unspeakable in quantum mechanics. Cambridge, UK: Trueblood, J. S., & Busemeyer, J. R. (2011). A comparison of the belief-adjustment
Cambridge University Press. model and the quantum inference model as explanations of order effects in
Bengtsson, I. (2006). Three ways to look at mutual unbiases bases. Unpublished, but human inference. Cognitive Science, 35, 1518–1552.
available at: arxiv.org/abs/quant-ph/0610216. Tversky, A. (1977). Features of similarity. Psychological Review, 84, 327–352.
Blutner, B., & beim Graben, P. (2014). Descriptive and foundational aspects of Tversky, A., & Kahneman, D. (1974). Judgment under uncertainty: Heuristics and
quantum cognition. Available at: arxiv.org/abs/1410.3961. biases. Science, 185, 1124–1131.
Blutner, R., Pothos, E. M., & Bruza, P. (2013). A quantum probability perspective on Wang, Z., & Busemeyer, J. R. (2013). A quantum question order model supported by
borderline vagueness. Topics in Cognitive Science, 5(4), 711–736. empirical tests of an a priori and precise prediction. Topics in Cognitive Science,
Bruza, P., Kitto, K., Ramm, B. J., & Sitbon, L. (2015). A probabilistic framework 5(4), 689–710.
for analysing the compositionality of conceptual combinations. Journal of Wang, Z., Busemeyer, J. R., Atmanspacher, H., & Pothos, E. M. (2013). The potential of
Mathematical Psychology, 67, 26–38. using quantum theory to build models of cognition. Topics in Cognitive Science,
Busch, P., Grabowski, M., & Lahti, P. J. (1995). Operational quantum physics. Berlin, 5(4), 672–688.
Heidelberg: Springer. Wang, Z., Solloway, T., Shiffrin, R. M., & Busemeyer, J. R. (2014). Context effects
Busemeyer, J. R., & Bruza, P. D. (2014). Quantum models of cognition and decision. produced by question orders reveal the quantum nature of human judgments.
Cambridge University Press. Proceedings of the National Academy of Sciences, 111, 9431.
Caves, C. M., Fuchs, C. A., & Schack, R. (2002). Quantum probabilities as Bayesian White, L. C., Pothos, E. M., & Busemeyer, J. R. (2014). Sometimes it does hurt to ask:
probabilities. Physical Review A, 65, 022305. the constructive role of articulating impressions. Cognition, 133(1), 48.
Chater, N., Tenenbaum, J. B., & Yuille, A. (2006). Probabilistic models of cognition: Yearsley, J.M. (in preparation). Tutorial on Advanced Tools and Concepts for
Conceptual foundations. Trends in Cognitive Sciences, 10(7), 287–291. Quantum Cognitive Theories.
Dirac, P. A. M. (1982). Principles of quantum mechanics. Oxford University Press. Yearsley, J. M., & Pothos, E. M. (2014). Challenging the classical notion of time in
Durt, T., Englert, B. G., Bengtsson, I., & Życzkowski, K. (2010). On mutually unbiased cognition: a quantum perspective. Proceedings of the Royal Society B, 381(1781),
bases. International Journal of Quantum Information, 8, 535–640. 20133056.
Gigerenzer, G., Hertwig, R., & Pachur, T. (Eds.) (2011). Heuristics: the foundations of Yearsley, J.M., & Pothos, E.M. (in preparation). Zeno’s paradox in human decision
adaptive behaviour. Oxford University Press. making. In this paper we ask some participants the same question up to 13 times!
Hagan, S., Hameroff, S. R., & Tuszynski, J. A. (2002). Quantum computation in brain (With other things happening in between.) In this set up it matters hugely whether
microtubules: Decoherence and biological feasibility. Physical Review E, 65(6), the measurements are exactly, or only approximately perfect.
061901. Yukalov, V. I., & Sornette, D. (2011). Decision theory with prospect interference and
Halliwell, J. J. (2005). How the quantum universe became classical. Contemporary entanglement. Theory and Decision, 70(3), 283–328.
Physics, 46, 93. Also available at: arxiv.org/abs/quant-ph/0501119. Zurek, W. H. (1991). Decoherence and the transition from quantum to classical.
Hameroff, S. R., & Penrose, R. (1996). Conscious events as orchestrated spacetime Physics Today, 44(10), 36. Also available in an updated form online at: arxiv.
selections. Journal of Consciousness Studies, 3(1), 36–53. org/abs/quant-ph/0306072.

Order Tutorial

Uploaded by

Copyright:

Available Formats

Order Tutorial

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Order Tutorial

Uploaded by

Copyright:

Available Formats

Journal of Mathematical Psychology ( ) –

Contents lists available at ScienceDirect

Journal of Mathematical Psychology

Quantum cognition and decision theories: A tutorial

article info abstract

⟨A⟩ = Tr(Aρ) = ⟨ψ| A |ψ⟩

or p(ai , t1 ; bj , t2 ) = Tr(PjB U (t2 − t1 )PiA U (t1 )ρ0 U Ď (t1 )

|ψ(t )⟩ = U (t ) |ψ(0)⟩ (19) × PiA U Ď (t2 − t1 ))

3.2. Quantisation? 3.3. Building quantum models from scratch

Suppose we already have a classical model of a decision making

where, in the basis {|0⟩ , |1⟩}, ≈ 0.385. (60)

cos2 (θ ) sin(θ ) cos(θ )

sin(θ ) − cos(θ ) p(A, B) + p(Ā, B̄) − p(B, A) − p(B̄, Ā) = 0. (63)

• They are positive operators, which means they have positive

The quantities Tr(Ei ρ) can thus be interpreted as probabilities,

Now let us turn to the more challenging case. It is easiest to start

Order effects are an important feature of the study of human Acknowledgments

1A = A1 = A, ∀A : H → H . (A.3) 0 = (ai − a∗i ) ⟨ai |ai ⟩ (A.14)

⟨A |u⟩ , |v⟩⟩ = |u⟩ , AĎ |v⟩ ,

You might also like