Quantum Theory and Beyond: Is Entanglement Special?

Borivoje Dakić1 and Časlav Brukner1, 2

1
Faculty of Physics, University of Vienna, Boltzmanngasse 5, A-1090 Vienna, Austria
2
Institute of Quantum Optics and Quantum Information,
Austrian Academy of Sciences, Boltzmanngasse 3, A-1090 Vienna, Austria
Quantum theory makes the most accurate empirical predictions and yet it lacks simple, comprehensible phys-
ical principles from which the theory can be uniquely derived. A broad class of probabilistic theories exist
which all share some features with quantum theory, such as probabilistic predictions for individual outcomes
(indeterminism), the impossibility of information transfer faster than speed of light (no-signaling) or the im-
possibility of copying of unknown states (no-cloning). A vast majority of attempts to find physical principles
behind quantum theory either fall short of deriving the theory uniquely from the principles or are based on ab-
arXiv:0911.0695v1 [quant-ph] 3 Nov 2009

stract mathematical assumptions that require themselves a more conclusive physical motivation. Here, we show
that classical probability theory and quantum theory can be reconstructed from three reasonable axioms: (1) (In-
formation capacity) All systems with information carrying capacity of one bit are equivalent. (2) (Locality) The
state of a composite system is completely determined by measurements on its subsystems. (3) (Reversibility)
Between any two pure states there exists a reversible transformation. If one requires the transformation from the
last axiom to be continuous, one separates quantum theory from the classical probabilistic one. A remarkable
result following from our reconstruction is that no probability theory other than quantum theory can exhibit
entanglement without contradicting one or more axioms.

I. INTRODUCTION disturbance and measurement [16] – generally thought of as

specifically quantum, yet being shown to be present in all
The historical development of scientific progress teaches us except classical theory. These generalized probabilistic the-
that every theory that was established and broadly accepted at ories can allow for stronger than quantum correlations in the
a certain time was later inevitably replaced by a deeper and sense that they can violate Bell’s inequalities stronger than the
more fundamental theory of which the old one remains a spe- quantum Cirel’son bound (as it is the case for the celebrated
cial case. One celebrated example is Newtonian (classical) “non-local boxes” of Popescu and Rohrlich [17]), though they
mechanics which was superseded by quantum mechanics at all respect the “non-signaling” constraint according to which
the beginning of the last century. It is natural to ask whether correlations cannot be used to send information faster than the
in a similar manner there could be logically consistent theo- speed of light.
ries that are more generic than quantum theory itself. It could Since the majority of the features that have been highlighted
then turn out that quantum mechanics is an effective descrip- as “typically quantum” are actually quite generic for all non-
tion of such a theory, only valid within our current restricted classical probabilistic theories, one could conclude that addi-
domain of experience. tional principles must be adopted to single out quantum the-
At present, quantum theory has been tested against very ory uniquely. Alternatively, these probabilistic theories indeed
specific alternative theories that, both mathematically and in can be constructed in a logically consistent way, and might
their concepts, are distinctly different. Instances of such al- even be realized in nature in a domain that is still beyond
ternative theories are non-contextual hidden-variable theo- our observations. The vast majority of attempts to find phys-
ries [1], local hidden-variable theories [2], crypto-nonlocal ical principles behind quantum theory either fail to single out
hidden-variable theories [3, 4], or some nonlinear variants the theory uniquely or are based on highly abstract mathe-
of the Schrödinger equation [5, 6, 7, 8]. Currently, many matical assumptions without an immediate physical meaning
groups are working on improving experimental conditions to (e.g. [18]).
be able to test alternative theories based on various collapse On the way to reconstructions of quantum theory from
models [9, 10, 11, 12, 13]. The common trait of all these pro- foundational physical principles rather than purely mathemat-
posals is to suppresses one or the other counter-intuitive fea- ical axioms, one finds interesting examples coming from an
ture of quantum mechanics and thus keep some of the basic instrumentalist approach [19, 20, 21], where the focus is pri-
notions of a classical world view intact. Specifically, hidden- marily on primitive laboratory operations such as prepara-
variable models would allow to preassign definite values to tions, transformations and measurements. While these recon-
outcomes of all measurements, collapse models are mecha- structions are based on a short set of simple axioms, they still
nisms for restraining superpositions between macroscopically partially use mathematical language in their formulation.
distinct states and nonlinear extensions of the Schrödinger Evidentally, added value of reconstructions for better un-
equation may admit more localized solutions for wave-packet derstanding quantum theory originates from its power of ex-
dynamics, thereby resembling localized classical particles. planation where the structure of the theory comes from. Can-
In the last years the new field of quantum information has didates for foundational principles were proposed giving a ba-
initialized interest in generalized probabilistic theories which sis for an understanding of quantum theory as a general theory
share certain features – such as the no-cloning and the no- of information supplemented by several information-theoretic
broadcasting theorems [14, 15] or the trade-off between state constraints [22, 23, 24, 25, 26]. In a wider context these ap-
 2

proaches belong to attempts to find an explanation for quan-

tum theory by putting primacy on the concept of information
or on the concept of probability which again can be seen as
a way of quantifying information [27, 28, 29, 30, 31, 32, 33,
34, 35, 36]. Other principles were proposed for separation of
quantum correlations from general non-signaling correlations,
such as that communication complexity is not trivial [37, 38],
that communication of m classical bits causes information
gain of at most m bits (“information causality”) [39], or that
any theory should recover classical physics in the macroscopic
limit [40].
In his seminal paper, Hardy [19] derives quantum the- FIG. 1: State spaces of a two-dimensional system in the generalized
ory from “five reasonable axioms” within the instrumentalist probabilistic theories analyzed here. d is the minimal number of real
framework. He sets up a link between two natural numbers, parameters necessary to determine the (generally mixed) state com-
d and N, characteristics of any theory. d is the number of de- pletely. From left to right: A classical bit with one parameter (the
grees of freedom of the system and is defined as the minimum weight p in the mixture of two bit values), a real bit with two real pa-
number of real parameters needed to determine the state com- rameters (state ρ ∈ D(R2 ) is represented by 2×2 real density matrix),
pletely. The dimension N is defined as the maximum number a qubit (quantum bit) with three real parameters (state ρ ∈ D(C2 ) is
represented by 2 × 2 complex density matrix) and a generalized bit
of states that can be reliably distinguished from one another in
for which d real parameters are needed to specify the state. Note
a single shot experiment. A closely related notion is the infor- that, when one moves continuously from one pure state (represented
mation carrying capacity of the system, which is the maximal by a point on the surface of a sphere) to another, only in the classical
number of bits encoded in the system, and is equal to log N probabilistic theory one must go trough the set of mixed states. Can
bits for a system of dimension N. probability theories that are more generic than quantum theory be ex-
Examples of theories with an explicit functional depen- tended in a logically consistent way to higher-dimensional and com-
dence d(N) are classical probability theory with the linear de- posite systems? Can entanglement exist in these theories? Where
pendence d = N − 1, and quantum theory with the quadratic should we look in nature for potential empirical evidences of the the-
dependence for which it is necessary to use d = N 2 − 1 real ories?
parameters to completely characterize the quantum state [65].
Higher-order theories with more general dependencies d(N) Both experiments put an upper bound on the extent of the ob-
might exist as illustrated in Figure 1. Hardy’s reconstruc- servational effects the two alternative theories may produce.
tion resorts to a “simplicity axiom” that discards a large class
of higher-order theories by requiring that for each given N,
d(N) takes the minimum value consistent with the other ax- II. BASIC IDEAS AND THE AXIOMS
ioms. However, without making such an ad hoc assumption
the higher-order theories might be possible to be constructed
Here we reconstruct quantum theory from three reasonable
in agreement with the rest of the axioms. In fact, an explicit
axioms. Following the general structure of any reconstruc-
quartic theory for which d = N 4 − 1 [41], and theories for
tion we first give a set of physical principles, then formulate
generalized bit (N = 2) for which d = 2r − 1 and r ∈ N [42],
their mathematical representation, and finally rigorously de-
were recently developed, though all of them are restricted to
rive the formalism of the theory. We will only consider the
the description of individual systems only.
case where the number of distinguishable states is finite. The
It is clear from the previous discussion that the question on three axioms which separate classical probability theory and
basis of which physical principles quantum theory can be sep- quantum theory from all other probabilistic theories are:
arated from the multitude of possible generalized probability
theories is still open. A particulary interesting unsolved prob- Axiom 1. (Information capacity) An elementary system has
lem is whether the higher-order theories of Refs. [19, 41, 42] the information carrying capacity of at most one bit. All sys-
can be extended to describe non-trivial, i.e. entangled, states tems of the same information carrying capacity are equiva-
of composite systems. Any progress in theoretical under- lent.
standing of these issues would be very desirable, in particular
because experimental research efforts in this direction have Axiom 2. (Locality) The state of a composite system is com-
been very sporadic. Although the majority of experiments in- pletely determined by local measurements on its subsystems
directly verify also the number of the degrees of freedom of and their correlations.
quantum systems [66], there are only few dedicated attempts Axiom 3. (Reversibility) Between any two pure states there
at such a direct experimental verification. Quaternionic quan- exists a reversible transformation.
tum mechanics (for which d = 2N 2 − N − 1) was tested in
a suboptimal setting [45] in a single neutron experiment in A few comments on these axioms are appropriate here. The
1984 [43, 44], and more recently, the generalized measure the- most elementary system in the theory is a two-dimensional
ory of Sorkin [46] in which higher order interferences are pre- system. All higher-dimensional systems will be built out of
dicted was tested in a three-slit experiment with photons [47]. two-dimensional ones. Recall that the dimension is defined
 3

vidual system, such as (1) “The outcome of measurement A is

+1” and (2) “The outcome of measurement A is -1”. An alter-
native choice for the pair of propositions can be propositions
about joint properties of two systems, such as (1’) “The out-
comes of measurement A on the first system and of B on the
second system are correlated” (i.e. either both +1 or both -1)
and (2’) “The outcomes of measurement A on the first system
and of B on the second system are anticorrelated”. The two
choices for the pair of propositions correspond to two choices
of basis states which each can be used to span the full state
FIG. 2: Illustration of the assumption stated in axiom 1. Consider space of an abstract elementary system (also called “general-
a toy-world of a two-dimensional system in which the set of pure ized bit”). As we will see later, taking the latter choice, it will
states consists of only x1 and x2 and their orthogonal states x⊥1 and follow from axiom 1 alone that the state space must contain
x⊥2 respectively, and where only two measurements exist, which dis- entangled states.
tinguish {x1 , x⊥1 } and {x2 , x⊥2 }. The convex set (represented by the grey Axiom 2 assumes that a specification of the probabilities for
area within the circle) whose vertices are the four states contains all a complete set of local measurements for each of the subsys-
physical (pure or mixed) states in the toy-world. Now, choose a point tems plus the joint probabilities for correlations between these
in the set, say y = λx1 + (1 − λ)x2 . Axiom 1 states that any phys- measurements is sufficient to determine completely the global
ical state can be represented as a mixture of two orthogonal states
state. Note that this property does hold in both quantum the-
(i.e. states perfectly distinguishable in a single shot experiment), e.g.
y = ηx+(1−η)x⊥ . This is not fulfilled in the toy world, but is satisfied ory and classical probability theory, but not in quantum theory
in a theory in which the entire circle represents the pure states and formulated on the basis of real or quaternionic amplitudes in-
where measurements can distinguish all pairs of orthogonal states. stead of complex. A closely related formulation of the axiom
was given by Barrett [16].
Finally, axiom 3 requires that transformations are re-
as the maximal number of states that can be reliably distin- versible. This is assumed alone for the purposes that the set
guished from one another in a single shot experiment. Under of transformations builds a group structure. It is natural to
the phrase “an elementary system has an information capac- assume that a composition of two physical transformations is
ity of at most one bit” we precisely assume that for any state again a physical transformation. It should be noted that this
(pure or mixed) of a two-dimensional system there is a mea- axiom could be used to exclude the theories in which “non-
surement such that the state is a mixture of two states which local boxes” occur, because there the dynamical group is triv-
are distinguished reliable in the measurement. An alternative ial, in the sense that it is generated solely by local operations
formulation could be that any state of a two dimensional sys- and permutations of systems with no entangling reversible
tem can be prepared by mixing at most two basis (i.e. per- transformations (that is, non-local boxes cannot be prepared
fectly distinguishable in a measurement) states (see Figure 2). from product states) [49].
Roughly speaking, axiom 1 assumes that a state of an elemen- If one requires the reversible transformation from our ax-
tary system can always be represented as a mixture of two iom 3 to be continuous:
classical bits. This part of the axiom is inspired by Zeilinger’s
Axiom 3’. (Continuity) Between any two pure states there ex-
proposal for a foundation principle for quantum theory [23]. ists a continuous reversible transformation,
The second statement in axiom 1 is motivated by the in-
tuition that at the fundamental level there should be no dif- which separates quantum theory from classical probability
ference between systems of the same information carrying theory. The same axiom is also present in Hardy’s reconstruc-
capacity. All elementary systems – be they part of higher tion. By a continuous transformation is here meant that every
dimensional systems or not – should have equivalent state transformation can be made up from a sequence of transfor-
spaces and equivalent sets of transformations and measure- mations only infinitesimally different from the identity.
ments. This seems to be a natural assumption if one makes A remarkable result following from our reconstruction is
no prior restrictions to the theory and preserves the full sym- that quantum theory is the only probabilistic theory in which
metry between all possible elementary systems. This is why one can construct entangled states and fulfill the three axioms.
we have decided to put the statement as a part of axiom 1, In particular, in the higher-order theories of Refs. [19, 41, 42]
rather than as a separate axiom. The particular formulation composite systems can only enjoy trivial separable states. On
used here is from Grinbaum [48] who suggested to rephrase the other hand, we will see that axiom 1 alone requires en-
the “subspace axiom” of Hardy’s reconstruction using physi- tangled states to exist in all non-classical theories. This will
cal language rather than mathematical. The subspace axiom allow us to discard the higher-order theories in our reconstruc-
states that a system whose state is constrained to belong to an tion scheme without invoking the simplicity argument.
M dimensional subspace (i.e. have support on only M of a set As a by product of our reconstruction we will be able to
of N possible distinguishable states) behaves like a system of answer why in nature only “odd” correlations (i.e. (1, 1, −1),
dimension M. (1, −1, 1), (−1, 1, 1) and (−1, −1, −1)) are observed when two
In logical terms axiom 1 means the following. We can think maximally entangled qubits (spin-1/2 particles) are both mea-
of two basis states as two binary propositions about an indi- sured along direction x, y and z, respectively. The most famil-
 4

iar example is of the singlet state |ψ− i = 12 (|0i1 |1i2 − |0i1 |1i2 ) We can decompose the state into a mixture of pure states, each
with anticorrelated results for arbitrarily but the same choice of which has to return probability one, and thus we can use
of measurement directions for two qubits. We will show that one of them to be a basis state. We will show later that each
the “mirror quantum mechanics” in which only “even” corre- pure state corresponds to a unique measurement outcome.
lations appear cannot be extended consistently to composite If the system in state p is incident on a transformation de-
systems of three bits. vice, its state will be transformed to some new state U(p). The
Our reconstruction will be given in the framework of typi- transformation U is a linear function of the state p as it needs
cal experimental situation an observer faces in the laboratory. to preserve the linear structure of mixtures. For example, con-
While this instrumentalist approach is a useful paradigm to sider the mixed state p which is generated by preparing state
work with, it might not be necessary. One could think about p1 with probability λ and p2 with probability 1 − λ. Then, in
axioms 1 and 3 as referring to objective features of elemen- each single run, either p1 or p2 is transformed and thus one
tary constituents of the world which need not necessarily be has:
related to laboratory actions. In contrast, axiom 2 seems to
acquire a meaning only within the instrumentalist approach as U(λp1 + (1 − λ)p2 ) = λU(p1 ) + (1 − λ)U(p2 ). (1)
it involves the word “measurement”. Even here one could fol- It is natural to assume that a composition of two or more
low a suggestion of Grinbaum [48] and rephrase the axiom to transformations is again from a set of (reversible) transforma-
the assumption of “multiplicability of the information carry- tions. This set forms some abstract group. Axiom 3 states that
ing capacity of subsystems.” the transformations are reversible, i.e. for every U there is an
Concluding this section, we note that the conceptual inverse group element U −1 . Here we assume that every trans-
groundwork for the ideas presented here has been pre- formation has its matrix representation U and that there is an
pared most notably by Weizsäcker [50], Wheeler [51] and orthogonal representation of the group: there exists an invert-
Zeilinger [23] who proposed that the notion of the elemen- ible matrix S such that O = S US −1 is an orthogonal matrix,
tary yes-no alternative, or the “Ur”, should play a pivotal role i.e. OT O = 11, for every U (We use the same notation both for
when reconstructing quantum physics. the group element and for its matrix representation). This does
not put severe restrictions to the group of transformations, as
it is known that all compact groups have such a representa-
III. BASIC NOTIONS tion (the Schur-Auerbach lemma) [55]. Since the transforma-
tion keeps the probabilities in the range [0, 1], it has to be a
Following Hardy [19] we distinguish three types of devices compact group [19]. All finite groups and all continuous Lie
in a typical laboratory. The preparation device prepares sys- groups are therefore included in our consideration.
tems in some state. It has a set of switches on it for varying Given a measurement setting, the outcome probability
the state produced. After state preparation the system passes Pmeas can be computed by some function f of the state p,
through a transformation device. It also has a set of switches
Pmeas = f (p). (2)
on it for varying the transformation applied on the state. Fi-
nally, the system is measured in a measurement apparatus. It Like a transformation, the measurement cannot change the
again has switches on it with which help an experimenter can mixing coefficients in a mixture, and therefore the measured
choose different measurement settings. This device outputs probability is a linear function of the state p:
classical data, e.g. a click in a detector or a spot on a observa-
tion screen. f (λp1 + (1 − λ)p2 ) = λ f (p1 ) + (1 − λ) f (p2 ). (3)
We define the state of a system as that mathematical object
from which one can determine the probability for any conceiv-
IV. ELEMENTARY SYSTEM: SYSTEM OF
able measurement. Physical theories can have enough struc-
INFORMATION CAPACITY OF 1 BIT
ture that it is not necessary to give an exhaustive list of all
probabilities for all possible measurements, but only a list of
probabilities for some minimal subset of them. We refer to A two-dimensional system has two distinguishable out-
this subset as fiducial set. Therefore, the state is specified by comes which can be identified by a pair of basis states {p, p⊥ }.
a list of d (where d depends on dimension N) probabilities for The state is specified by d probabilities p = (p1 , ..., pd ) for d
a set of fiducial measurements: p = (p1 , . . . , pd ). The state is fiducial measurements, where pi is probability for a particu-
pure if it is not a (convex) mixture of other states. The state lar outcome of the i-th fiducial measurement (the dependent
is mixed if it is not pure. For example, the mixed state p gen- probabilities 1 − pi for the opposite outcomes are omitted in
erated by preparing state p1 with probability λ and p2 with the state description). Instead of using the probability vector p
probability 1 − λ, is p = λp1 + (1 − λ)p2 . we will specify the state by its Bloch representation x defined
When we refer to an N-dimensional system, we assume that as a vector with d components:
there are N states each of which identifies a different outcome xi = 2pi − 1. (4)
of some measurement setting, in the sense that they return
probability one for the outcome. We call this set a set of basis The mapping between the two different representations is an
or orthogonal states. Basis states can be chosen to be pure. To invertible linear map and therefore preserves the structure of
see this assume that some mixed state identifies one outcome. the mixture λp1 + (1 − λ)p2 7→ λx1 + (1 − λ)x2 .
 5

It
P is convenient to define a totaly mixed state E =
1
N x∈Spure x, where Spure denotes the set of pure states and
N is the normalization constant. In the case of a continu-
ous set of pure states the summation has to be replaced by a
proper integral. It is easy to verify that E is a totally invariant
state. This implies that every measurement and in particular
the fiducial ones will return the same probability for all out-
comes. In the case of a two-dimensional system this proba-
bility is 1/2. Therefore, the Bloch vector of the totally mixed
state is the zero-vector E = ~0.
The transformation U does not change the totaly mixed FIG. 3: (Left) Illustration to the proof that the entire d-dimensional
state, hence U(~0) = ~0. The last condition together with the ellipsoid (here represented by a circle; d = 2) contains physical
linearity condition (1) implies that any transformation is rep- states. Consider a line x(t) = tx through the origin. A point on
resented by some d × d real invertible matrix U. The same the line can be expanded into a set of linearly independent vectors xi
reasoning holds for measurements. Therefore, the measured (here x1 and x2 ). For sufficiently small t (i.e. when the line is within
the gray square) the point x(t) can be represented as a convex mixture
probability is given by the formula:
over xi and their orthogonal vectors x⊥i and thus is a physical state.
1 According to axiom 1, x(t) can be represented as a convex mixture of
Pmeas = (1 + rT x). (5) two orthogonal pure states x0 and x⊥0 : x(t) = tx = αx0 + (1 − α)(−x0 ),
2 where x = x0 (see text for details). This implies that every point in
The vector r represents the outcome for the given measure- the ellipsoid is a physical state. (Right) Illustration to the proof that in
ment setting. For example, the vector (1, 0, 0, . . . ) represents the orthogonal representation the measurement vector m that identi-
one of the outcomes for the first fiducial measurement. fies the state x, i.e. for which the probability Pmeas = 21 (1 + mT y) = 1,
is identical to the state vector, m = y. Suppose that m , y, then
According to axiom 1 any state is a classical mixture of
||m|| > 1, since the state vector is normalized. But then the same
some pair of orthogonal states. For example, the totally mixed measurement for state y′ parallel to m would return a probability
state is an equally weighted mixture of some orthogonal states larger than 1, which is nonphysical. Thus m = y.
~0 = 1 x + 1 x⊥ . Take x to be the reference state. According to
2 2
axiom 3 we can generate the full set of states by applying all
possible transformations to the reference state. Since the to-
tally mixed state is invariant under the transformations, the For every pure state x, there exists at least one measurement
pair of orthogonal states is represented by a pair of antipar- setting with the outcome r such that the outcome probability
allel vectors x⊥ = −x. Consider the set Spure = { Ux | ∀U} is one, hence rT x = 1. Let us define new coordinates y = 1c S x
of all pure states generated by applying all transformations to and m = cS −1T r in the orthogonal representation. The set of
the reference state. If one uses the orthogonal representation pure states in the new coordinates is a (d − 1)-sphere Sd−1 =
of the transformations, U = S −1 OS , which was introduced {y | ||y|| = 1} of the radius. The probability rule (5) remains
above, one maps x 7→ S x and U 7→ O. Hence, the transfor- unchanged in the new coordinates:
mation Ux 7→ S Ux = OS x is norm preserving. We conclude
that all pure states are points on a d-dimensional ellipsoid de- 1
scribed by ||S x|| = c with c > 0. Pmeas = (1 + mT y). (7)
2
Now, we want to show that any vector x satisfying ||S x|| = c
is a physical state and therefore the set of states has to be the
Thus, one has mT y = 1. Now, assume that m , y. Then
whole ellipsoid. Let x be some vector satisfying ||S x|| = c and
||m|| > 1 and the vectors m and y span a two-dimensional
x(t) = tx a line trough the origin (totaly mixed state) as given
plane as illustrated in Figure 3 (right). The set of pure states
in Figure 3 (left). Within the set of pure states we can always
within this plane is a unit circle. Choose the pure state y′ to
find d linearly independent vectors {x1 , . . . , xd }. For each state
be parallel to m. Then the outcome probability is Pmeasur =
xi there is a corresponding orthogonal state x⊥i = −xi in a set 1 ′
of states. We can expand a point on 2 (1 + ||m||||y ||) > 1 which is non-physical, hence m = y.
P the line into a linearly Therefore, to each pure state y, we associate a measurement
independent set of vectors: x(t) = t di=1 ci xi . For sufficiently
vector m = y which identifies it. Equivalently, in the original
small t we can define a pair of non-negative P numbers λi (t) =
1 1 ⊥ 1 1 ⊥ coordinates, to each x we associate a measurement vector r =
( + tc i ) and λ (t) = ( − tc i ) with i (λi (t) + λ (t)) =1
2 d i 2 d i Dx, where D = c12 S T S is a positive, symmetric matrix. A
such that x(t) is a mixture x(t) = di=1 λi (t)xi + λ⊥i (t)x⊥i and
P
proof of this relation for the restricted case of d = 3 can be
therefore is a physical state. Then, according to axiom 1 there found in Ref. [19].
exists a pair of basis states {x0 , −x0 } such that x(t) is a mixture
of them From now one, instead of the measurement vector r we will
use the pure state x which identifies it. When we say that the
x(t) = tx = αx0 + (1 − α)(−x0 ), (6) measurement along the state x is performed we mean the mea-
surement given by r = Dx. The measurement setting is given
where α = 1+t
2 and x = x0 . This implies that x is a pure state by a pair of measurement vectors r and −r. The measured
and therefore all points of the ellipsoid are physical states. probability when the state x1 is measured along the state x2
 6

follows from formula (5): ity one predicts at the Planck scale the dimension of space-
time to be different from 3 + 1 [54]. If one considers direc-
1 tional degrees of freedom (spin), then the d − 1-sphere (Bloch
P(x1 , x2 ) = (1 + xT1 Dx2 ). (8)
2 sphere) might be interpreted as the state space of a spin system
We can choose orthogonal eigenvectors of the matrix D as embedded in real (ordinary) space of dimension d, in general
different than 3 which is the special case of quantum theory.
the fiducial set of states (measurements):
The reversible transformation R preserves the purity of state
Dxi = ai xi , (9) ||Rx|| = ||x|| and therefore R is an orthogonal matrix. We have
shown that the state space is the full (d − 1)-sphere. Accord-
where ai are eigenvalues of D. Since xi are pure states, they ing to axiom 3 the set of transformations must be rich enough
satisfy xTi Dx j = δi j . The set of pure states becomes a unit to generate the full sphere. If d = 1 (classical bit), the group
sphere Sd−1 = {x | ||x|| = 1} and the probability formula is of transformations is discrete and contains only the identity
reduced to and the bit-flip. If d > 1, the group is continuous and is some
subgroup of the orthogonal group O(d). Every orthogonal ma-
1 trix has determinant either 1 or -1. The orthogonal matrices
P(x1 , x2 ) = (1 + xT1 x2 ). (10)
2 with determinant 1 form a normal subgroup of O(d), known
as the special orthogonal group SO(d). The group O(d) has
This corresponds to a choice of a complete set of mutually
two connected components: the identity component which
complementary measurements (i.e. mutually unbiased basis
is the SO(d) group, and the component formed by orthogo-
sets) for the fiducial measurements. The states identifying out-
nal matrices with determinant -1. Since every two points on
comes of complementary measurements satisfy P(xi , x j ) = 21
the (d − 1)-sphere are connected by some transformation, the
for i , j. Two observables are said to be mutually comple-
group of transformations is at least the SO(d) group. If we in-
mentary if complete certainty about one of the observables
clude even a single transformation with determinant -1, the set
(one of two outcomes occurs with probability one) precludes
of transformations becomes the entire O(d) group. (Later we
any knowledge about the others (the probability for both out-
will show that only some d are in agreement with our three
comes is 1/2). Given some state x, the i-th fiducial measure-
axioms and for these d’s the set of physical transformations
ment returns probability pi = 12 (1+xi ). Therefore, xi is a mean
will be shown to be the SO(d) group).
value of a dichotomic observable bi = +1xi − 1x⊥i with two
possible outcomes bi = ±1.
A theory in which the state space of the generalized bit V. COMPOSITE SYSTEM AND THE NOTION OF
is represented by a (d − 1)-sphere has d mutually comple- LOCALITY
mentary observables. This is a characteristic feature of the
theories and they can be ordered according to their number.
For example, classical physics has no complementary observ- We now introduce a description of composite systems. We
ables, real quantum mechanics has two, complex (standard) assume that when one combines two systems of dimension L1
quantum mechanics has three (e.g. the spin projections of and L2 into a composite one, one obtains a system of dimen-
a spin-1/2 system along three orthogonal directions) and the sion L1 L2 . Consider a composite system consisting of two
one based on quaternions has five mutually complementary geneneralized bits and choose a set of d complementary mea-
observables. Note that higher-order theories of a single gen- surements on each subsystem as fiducial measurements. Ac-
eralized bit are such that the qubit theory can be embedded in cording to axiom 2 the state of the composite system is com-
them in the same way in which classical theory of a bit can be pletely determined by a set of real parameters obtainable from
embedded in qubit theory itself. local measurements on the two generalized bits and their cor-
Higher-order theories can have even better information pro- relations. We obtain 2d independent real parameters from the
cessing capacity than quantum theory. For example, the com- set of local fiducial measurements and additional d2 param-
putational abilities of the theories with d = 2r and r ∈ N eters from correlations between them. This gives altogether
in solving the Deutsch-Josza type of problems increases with d2 + 2d = (d + 1)2 − 1 parameters. They are the components
the number of mutually complementary measurements [42]. xi , yi , i ∈ {1, ..., d}, of the local Bloch vectors and T i j of the
It is likely that the larger this number is the larger the error correlation tensor:
rate would be in secret key distribution in these theories, in a
xi = p(i) (A = 1) − p(i) (A = −1), (11)
similar manner in which the 6-state is advantageous over the
4-state protocol in (standard) quantum mechanics. In the first y j = p( j) (B = 1) − p( j) (B = −1), (12)
case one uses all three mutually complementary observables T i j = p(i j) (AB = 1) − p(i j) (AB = −1). (13)
and in the second one only two of them. (See Ref. [52] for a
review on characterizing generalized probabilistic theories in Here, for example, p(i) (A = 1) is the probability to obtain
terms of their information-processing power and Ref. [53] for outcome A = 1 when the i-th measurement is performed on
investigating the same question in much more general frame- the first subsystem and p(i j) (AB = 1) is the joint probability
work of compact closed categories.) to obtain correlated results (i.e. either A = B = +1 or A =
A final remark on higher-order theories is of more specula- B = −1) when the i-th measurement is performed on the first
tive nature. In various approaches to quantum theory of grav- subsystem and the j-th measurement on the second one.
 7

Note that axiom 2 “The state of a composite system is com- The last expression is called Schmidt decomposition of the
pletely determined by local measurements on its subsystems state.
and their correlations” is formulated in a way that the non- The local Bloch vectors satisfy ||x||, ||y|| ≤ 1 which implies
signaling condition is implicitly assumed to hold. This is a bound on the correlation ||T || ≥ 1 for all pure states. The
because it is sufficient to speak about “local measurements” following lemma identifies a simple entanglement witness for
alone without specifying the choice of measurement setting pure states. The proof of this and all subsequent lemmas is
on the other, potentially distant, subsystem. Therefore, xi does given in the Appendix.
not depend on j, and y j does not depend on i.
We represent a state by the triple ψ = (x, y, T ), where x Lemma 1. The lower bound ||T || = 1 is saturated, if and only
and y are the local Bloch vectors and T is a d × d real matrix if the state is a product state T = xyT.
representing the correlation tensor. The product (separable) Recall that for every transformation U we can find its or-
state is represented by ψ p = (x, y, T ), where T = xyT is of thogonal representation U = S OS −1 (the Schur-Auerbach
product form, because the correlations are just products of the lemma), where S is an invertible matrix and OT O = 11. The
components of the local Bloch vectors. We call the pure state matrix S is characteristic of the representation and should be
entangled if it is not a product state. the same for all transformations U. If we choose some local
The measured probability is a linear function of the state ψ. transformation U = (R1 , R2 ), U will be orthogonal and thus
If we prepare totaly mixed states of the subsystems (0, 0, 0), we can choose to set S = 11. The representation of transfor-
the probability for any outcome of an arbitrary measurement mations is orthogonal, therefore they are norm preserving. By
will be 1/4. Therefore, the outcome probability can be written applying simultaneously all (local and non-local) transforma-
as: tions U to some product state (the reference state) ψ and to the
1 measurement vector which identifies it, r = ψ, we generate the
Pmeasur = (1 + (r, ψ)), (14) set of all pure states and corresponding measurement vectors.
4
Since we have 1 = P(r = ψ, ψ) = P(Ur, Uψ), correspondence
where r = (r1 , r2 , K) is a measurement vector associated to r = ψ holds for any pure state ψ. Instead of the measurement
the observed outcome and (..., ...) denotes the scalar product: vector r in formula (14) we use the pure state which identifies
it. If the state ψ1 = (x1 , y1 , T 1 ) is prepared and measurement
(r, ψ) = rT1 x + rT2 y + Tr(K T T ). (15) along the state ψ2 = (x2 , y2 , T 2 ) is performed, the measured
probability is given by
Now, assume that r = (r1 , r2 , K) is associated to the
outcome which is identified by some product state ψ p = 1
P12 (ψ1 , ψ2 ) = (1 + xT1 x2 + yT1 y2 + Tr(T 1T T 2 )). (22)
(x0 , y0 , T 0 ). If we preform a measurement on the arbitrary 4
product state ψ = (x, y, T ), the outcome probability has to fac-
The set of pure states obeys P12 (ψ, ψ) = 1. We can define
torize into the product of the local outcome probabilities of
the normalization condition for pure states P12 (ψ, ψ) = 14 (1 +
the form (10):
||x||2 + ||y||2 + ||T ||2) = 1 where ||T ||2 = Tr(T T T ). Therefore we
1 have:
Pmeasur = (1 + rT1 x + rT2 y + xT Ky) (16)
4
= P1 (x0 , x)P2 (y0 , y) (17) ||x||2 + ||y||2 + ||T ||2 = 3, (23)
1 1 for all pure states.
= (1 + xT0 x) (1 + yT0 y) (18)
2 2 An interesting observation can be made here. Although
1 seemingly axiom 2 does not imply any strong prior restric-
= (1 + xT0 x + yT0 y + xT x0 yT0 y), (19)
4 tions to d, we surprisingly have obtained the explicit number
3 in the normalization condition (23). As we will see soon this
which holds for all x, y. Therefore we have r = ψ p . For each relation will play an important role in deriving d = 3 as the
product state ψ p there is a unique outcome r = ψ p which iden- only non-classical solution consistent with the axioms.
tifies it. We will later show that correspondence r = ψ holds
for all pure states ψ.
If we preform local transformations R1 and R2 on the sub- VI. THE MAIN PROOFS
systems, the global state ψ = (x, y, T ) is transformed to

(R1 , R2 )ψ = (R1 x, R2 y, R1 T RT2 ). (20) We will now show that only classical probability theory and
quantum theory are in agreement with the three axioms.
T is a real matrix and we can find its singular value decompo-
sition diag[t1 , . . . , td ] = R1 T RT2 , where R1 , R2 are orthogonal
matrices which can be chosen to have determinant 1. There- A. Ruling out the d even case
fore, we can choose the local bases such that correlation tensor
T is a diagonal matrix: Let us assume the total inversion Ex = −x being a physical
transformation. Let ψ = (x, y, T ) be a pure state of composite
(R1 , R2 )(x, y, T ) = (R1 x, R2 y, diag[t1 , . . . , td ]). (21) system. We apply total inversion to one of the subsystems
 8

and obtain the state ψ′ = (E, 11)(x, y, T ) = (−x, y, −T ). The (x1 , x2 , . . . , xd )T and y = (y1 , y2 , . . . , yd )T . Consider the one-
probability bit transformation R with the property Re1 = −e1 . The local
transformation of this type maps the state from S 12 to S 34 as
1 shown by the following lemma:
P12 (ψ, ψ′ ) = (1 − ||x||2 + ||y||2 − ||T ||2) (24)
4
1 Lemma 3. If the state ψ ∈ S 12 , then ψ′ = (R, 11)ψ ∈ S 34 and
= (||y||2 − 1) (25) ψ′′ = (11, R)ψ ∈ S 34 .
2
Let us define Ti(x) = (T i1 , . . . , T id ) and T(y) T
i = (T 1i , . . . , T di ) .
has to be nonnegative and therefore we have ||y|| = 1. Simi-
The correlation tensor can be rewritten in two different ways:
larly, we apply (11, E) to ψ and obtain ||x|| = 1. Since the local
vectors are of the unit norm we have ||T || = 1 and thus, accord-  (x) 
 T1 
ing to lemma 1, the state ψ is a product state. We conclude  (x) 
 T2   
that no entangled states can exist if E is a physical transfor- T =  .  or T = T(y) T(y)
. . . T(y)
. (32)
mation. As we will soon see, according to axiom 1 entangled  ..  1 2 d
 (x) 
states must exist. Thus, E cannot represent a physical trans- Td
formation. We will now show that this implies that d has to be
odd. Recall that the set of transformations is at least the SO(d) Consider now the case d > 3. We define local transforma-
group. d cannot be even since E would have unit determinant tions Ri flipping the first and i-th coordinate and R jkl flipping
and would belong to SO(d). d has to be odd in which case E the first and j-th, k-th, and l-th coordinate with j , k , l , 1.
has determinant -1. The set of physical transformations is the Let ψ = (x, y, (T1(x) , . . . , Td(x) )T ) belong to S 12 . According to
SO(d) group. Lemma 2, the states ψi = (Ri , 11)ψ and ψ jkl = (R jkl , 11)ψ be-
long to S 34 , therefore P12 (ψ, ψi ) = 0 and P12 (ψ, ψ jkl ) = 0. We
have:
B. Ruling out the d > 3 case.
0 = P12 (ψ, ψi ) (33)
1 − x21 + x22 + · · · − x2i + · · · + x2d + ||y||2 (34)
Let us define one basis set of two generalized bit product
states: −||T1(x) ||2 + ||T2(x) ||2 + · · · − ||Ti(x) ||2 + · · · + ||Td(x) ||2 (35)
= 1 − 2x21 − 2x2i − 2||T1(x) ||2 − 2||Ti(x) ||2 + ||x|| + ||y|| + ||T ||2
2 2
ψ1 = (e1 , e1 , T 0 = e1 eT1 ) (26)
ψ2 = (−e1 , −e1 , T 0 ) (27) = 2(2 − x21 − x2i − ||T1(x) ||2 − ||Ti(x) ||2 ). (36)
ψ3 = (−e1 , e1 , −T 0 ) (28) Similarly, we expand P12 (ψ, ψ jkl ) = 0 and together with the
ψ4 = (e1 , −e1 , −T 0 ) (29) last equation we obtain:

with e1 = (1, 0, . . . , 0)T . Now, we define two subspaces S 12 x21 + x2i + ||T1(x) ||2 + ||Ti(x) ||2 = 2 (37)
and S 34 spanned by the states ψ1 , ψ2 and ψ3 , ψ4 , respectively. x21 + x2j + x2k + x2l + ||T1(x) ||2 + ||T(x)
j ||
2
+ ||Tk(x) ||2 + ||Tl(x) ||2 = 2.
Axiom 1 states that these two subspaces behave like one-bit
spaces, therefore they are isomorphic to the (d − 1)-sphere Since this has to hold for all i, j, k, l we have:
S 12  S 34  Sd−1 . The state ψ belongs to S 12 if and only if
x2 = x3 = · · · = xd = 0 (38)
the following holds:
T2(x) = T3(x) = · · · = Td(x) = 0. (39)
P12 (ψ, ψ1 ) + P12 (ψ, ψ2 ) = 1. (30)
We repeat this kind of reasoning for the transformations
Since the ψ1 , . . . , ψ4 form a complete basis set, we have (11, Ri ) and (11, R jkl ) and obtain:

y21 + y2i + ||T(y) 2 (y) 2

1 || + ||Ti || = 2 (40)
P12 (ψ, ψ3 ) = 0, P12 (ψ, ψ4 ) = 0. (31)
y21 + y2j + y2k + y2l + ||T(y)
1 ||
2
+ ||T(y)
j ||
2
+ ||T(y)
k ||
2
+ ||T(y)
l ||
2
= 2.
A similar reasoning holds for states belonging to the S 34 sub-
space. Since the states ψ ∈ S 12 and ψ′ ∈ S 34 are per- Therefore, we have
fectly distinguishable in a single shot experiment, we have
y2 = y3 = · · · = yd = 0 (41)
P12 (ψ, ψ′ ) = 0. Therefore, S 12 and S 34 are orthogonal sub-
spaces. T(y)
2 = T(y)
3 = ··· = T(y)
d = 0. (42)
Axiom 1 requires the existence of entangled states as it is
The only non-zero element of the correlation tensor is T 11 and
apparent from the following Lemma 2.
it has to be exactly 1, since ||T || ≥ 1. This implies that ψ is a
Lemma 2. The only product states belonging to S 12 are ψ1 product state, furthermore ψ = ψ1 or ψ = ψ2 .
and ψ2 . This concludes our proof that only the cases d = 1 and
d = 3 are in agreement with our three axioms. To distinguish
We define a local mapping between orthogonal subspaces between the two cases, one can invoke the continuity axiom
S 12 and S 34 . Let the state ψ = (x, y, T ) ∈ S 12 , with x = (3’) and proceed as in the reconstruction given by Hardy [19].
 9

VII. “TWO” QUANTUM MECHANICS

We now obtain two solutions for the theory of a composite

system consisting of two bits in the case when d = 3. One
of them corresponds to the standard quantum theory of two
qubits, the other one to its “mirror” version in which the states
are obtained from the ones from the standard theory by partial
transposition. Both solutions are regular as far as one consid- FIG. 4: Correlations between results obtained in measurements of
ers composite systems of two bits, but the “mirror” one cannot two bits in a maximal entangled (Bell’s) state in standard quantum
be consistently constructed already for systems of three bits. mechanics (Left) and “mirror quantum mechanics” (Right) along x,
Two conditions (30) and (31) put the constraint to the form y and z directions. Why do we never see correlations as given in the
of ψ: table on the right? The opposite sign of correlations on the right and
on the left is not a matter of convention or labeling of outcomes. If
x1 = −y1 , T 11 = 1. (43) one can transport the two bits parallel to the same detector, one can
distinguish operationally between the two types of correlations [57].
2
The subspace S 12 is isomorphic to the sphere S . Let us
choose ψ complementary to the one bit basis {ψ1 , ψ2 } in S 12 . √
We have P12 (ψ, ψ1 ) = P12 (ψ, ψ2 ) = 1/2 and thus x1 = y1 = 0. The transformed correlation tensor has a component 2
For simplicity we write ψ in the form: which is non-physical. Therefore, the transformation
! ! !! (11, R(φ))ψ draws a full circle of pure states in a plane orthog-
0 0 1 TTy onal to ψ1 within the subspace S 12 . Similarly, the transforma-
ψ= , , , (44)
x y Tx T tion (R(φ), 11) draws the same set of pure states when applied
with x = (x2 , x3 )T , y = (y2 , y!3 )T , Ty = (T 12 , T 13 )T , T x = to ψ. Hence, for every transformation (11, R(φ1 )) there exists a
T 22 T 23 transformation (R(φ2), 11) such that (11, R(φ1))ψ = (R(φ2 ), 11)ψ.
(T 21 , T 31 )T and T = . This gives us a set of conditions:
T 32 T 33
Let R(φ) be a rotation around the e1 axis. This transforma- R(φ2 )x = x (51)
tion keeps S 12 invariant. Now, we show that the state ψ as
R(φ1 )y = y (52)
given by equation (44) cannot be invariant under local trans-
formation (11, R(φ)). To prove this by reductio ad absurdum R(φ2 )T x = Tx (53)
suppose the opposite, i.e. that (11, R(φ))ψ = ψ. We have three TTy RT (φ1 ) = TTy (54)
conditions R(φ2 )T = T RT (φ1 ), (55)
R(φ)y = y, TTy RT (φ) = TTy , T
T R (φ) = T, (45)
which are fulfilled if x = y = T x = Ty = 0 and T =
which implies y = 0, TTy
= 0 and T = 0 thus diag[T 1 , T 2 ]. Equation (37) gives T 22 = T 32 = 1 and we finally
     end up with two different solutions:
 0   0   1 0 0 

ψ =  x2  ,  0  ,  T 2 0 0  . (46) ψQM = (0, 0, diag[1, −1, 1]) ∨ ψMQM = (0, 0, diag[1, 1, 1]).
x3 0 T3 0 0 (56)
The first “M” in ψMQM stands for “mirror”. The two so-
According to equations (37) and (40) we can easily check that lutions are incompatible and cannot coexist within the same
||x|| = 1, and thus ψ is locally equivalent to the state: theory. The first solution corresponds to the triplet state φ+ of
 0   0   1 0 0 
      ordinary quantum mechanics. The second solution is a totally
′       ′
ψ =  0  ,  0  ,  T 2 0 0  . (47) invariant state and has a negative overlap with, for example,
1 0 T 3′ 0 0 the singlet state ψ− for which T = diag[−1, −1, −1]. That is,
if the system were prepared in one of the two states and the
Let χ1 = (−e3 , e1 , −e3 eT1 ) and χ2 = (−e3 , −e1 , e3 eT1 ). The two other one were measured, the probability would be negative.
conditions P(ψ′ , χ1 ) ≥ 0 and P(ψ′ , χ2 ) ≥ 0 become Nevertheless, both solutions are regular at the level of two
bits. The first belongs to ordinary quantum mechanics with
1 1
(1 − 1 − T 3′ ) = − T 3′ ≥ 0 (48) the singlet in the “antiparallel” subspace S 34 and the second
4 4 solution is “the singlet state in the parallel subspace” S 12 . We
1 1
(1 − 1 + T 3′ ) = T 3′ ≥ 0 (49) will show that one can build the full state space, transforma-
4 4 tions and measurements in both cases. The states from one
and thus T 3′ = 0. The normalization condition (23) gives T 2′ = quantum mechanics can be obtained from the other by partial
±1. The state ψ′ is not physical. This can be seen when one transposition ψPTQM = ψMQM . In particular, the four maximal
performs the rotation (R, 11) where entangled states (Bell states) from “mirror quantum mechan-
 1 ics” have correlations of the opposite sign of those from the
 √2 − √12 0 

standard quantum mechanics (see Figure 4).
R =  √1 √1 0  .
 
(50) Now we show that the theory with “mirror states” is physi-
 2 2 
0 0 1 cally inconsistent when applied to composite system of three
 10

bits. Let us first derive the full set of states and transforma- Now, we show that “mirror mechanics” cannot be consis-
tions for two qubits in standard quantum mechanics. We have tently extended to composite systems consisting of three bits.
seen that the state ψQM belongs to the subspace S 12 , and fur- Let ψ p = (x, y, z, T 12, T 13 , T 23 , T 123 ) be some product state of
thermore, that it is complementary (within S 12 ) to the product three bits, where x, y and z are local Bloch vectors, T 12 , T 13 ,
states ψ1 and ψ2 . The totally mixed state within the S 12 sub- T 23 and T 123 are two- and three-body correlation tensors, re-
space is E12 = 21 ψ1 + 12 ψ2 . The states ψ1 and ψQM span one spectively. We can apply the transformations PTUi j PT to a
two-dimensional plane, and the set of pure states within this composite system of i and j, and we are free to choose with
plane is a circle: respect to which subsystem (i or j) to take the partial trans-
position. Furthermore, we can combine transformations in
ψ(x) = E12 + cos x (ψ1 − E12 ) + sin x (ψQM − E12 ) (57) 12 and 13 subsystems such that the resulting state is genuine
= (cos x e1 , cos x e1 , diag[1, − sin x, sin x]). (58) three-partite entangled, and we can choose to partially trans-
pose subsystem 2 in both cases. We obtain the transformation
We can apply a complete set of local transformations to the
set ψ(x) to obtain the set of all pure two-qubit states. Let us U123 = PT2 U12 PT2 PT2 U23 PT2 (62)
represent a pure state ψ = (x, y, T ) by the 4 × 4 Hermitian = PT2 U12 U23 PT2 . (63)
matrix ρ:
3 3 3 When we apply U123 to ψ p we obtain the state
1 X X X
PT2 U12 U23 φ p , where φ p = PT2 ψ p is again some product state.
ρ = (11⊗ 11+ xi σi ⊗ 11+ yi 11⊗σi + T i j σi ⊗σ j ), (59)
4 i=1 i=1 i, j=1
The state U12 U23 φ p is a quantum three qubit state. Since states
ψ p and φ p are product states and do belong to standard quan-
where σi , i ∈ {1, 2, 3}, are the three Pauli matrices. It is tum states, we can use the formalism of quantum mechanics
easy to show that the set of states (57) corresponds to the and denote them as |ψ p i and |φ p i. Furthermore, since the state
set of one-dimensional projectors |ψ(x)ihψ(x)|, where |ψ(x)i = |ψ p i is an arbitrary product state, without loss of generality we
cos 2x |00i + sin 2x |11i. The action of local transformations set |φ p i = |0i|0i|0i. We can choose U12 and U23 such that:
(R1 , R2 )ψ corresponds to local unitary transformation U1 ⊗
U2 |ψihψ|U1† ⊗ U2† , where the correspondence between U and U12 |0i|0i = |0i|0i (64)
R is given by the isomorphism between the groups SU(2) and 1
U12 |0i|1i = √ (|0i|1i + |1i|0i) (65)
SO(3): 2
r
1 2
  3
3 X
 
† 1  X   
U23 |0i|0i = √ |0i|1i + |1i|0i. (66)
UρU = 11 +  Ri j x j  σi  . (60) 3
2 3
i=1 j=1
This way we can generate the W-state
Here Ri j = Tr(σi Uσ j U † ) and xi = Trσi ρ. When we apply a
complete set of local transformations to the states |ψ(x)i we |Wi = U12 U23 |0i|0i|0i (67)
obtain the whole set of pure states for two qubits. The group 1
of transformations is the set of unitary transformations SU(4). = √ (|0i|0i|1i + |0i|1i|0i + |1i|0i|0i). (68)
3
The set of states from “mirror quantum mechanics” can be
obtained by applying partial transposition to the set of quan- When we apply partial transposition with respect to subsystem
tum states. Formally, partial transposition with respect to sub- 2, we obtain the corresponding “mirror W-state” which we
system 1 is defined by action on a set of product operators: denote as WM -state, WM = PT2 W. The local Bloch vectors
and two-body correlation tensors for the W state are
PT1 (ρ1 ⊗ ρ2 ) = ρT1 ⊗ ρ2 . (61)
x = y = z = (0, 0, 31 )T , (69)
where ρ1 and ρ2 are arbitrary operators. Similarly, we can
define the partial transposition with respect to subsystem 2, T 12 = T 13 = T 23 = diag[ 32 , 32 , − 31 ], (70)
PT2 . To each unitary transformation U in quantum mechan-
ics we define the corresponding transformation in “mirror where |0i corresponds to result +1. Consequently, the local
mechanics”, e.g. with respect to subsystem 1: PT1 UPT1 . Bloch vectors and the correlation tensor for WM -state are
Therefore, the set of transformations is a conjugate group x = y = z = (0, 0, 31 )T , (71)
PT1 SU(4)PT1 := {PT1 UPT1 | U ∈ SU(4)}. Note that we could
equally have chosen to apply partial transposition with respect T 12 = T 23 = diag[ 32 , − 32 , − 13 ], (72)
to subsystem 2, and would obtain the same set of states. In T 13 = diag[ 32 , 32 , − 31 ]. (73)
fact, one can show that PT1 UPT1 = PT2 U ∗ PT2 , where U ∗ is
a conjugate unitary transformation (see Lemma 4 in the Ap- The asymmetry in the signs of correlations in the tensors
pendix). Therefore, the two conjugate groups are the same T 12 , T 23 and T 13 leads to inconsistencies because they de-
PT1 SU(4)PT1 = PT2 SU(4)PT2 . We can generate the set of fine three different reduced states ψi j = (xi , x j , T i j ), i j ∈
“mirror states” by applying all the transformations PTUPT {12, 23, 13}, which cannot coexist within a single theory. The
to some product state, regardless of which particular partial states ψ12 and ψ23 belong to “mirror quantum mechanics”,
transposition is used. while the state ψ13 belongs to ordinary quantum mechanics.
 11

To see this, take the state ψ = (0, 0, diag[−1, −1, 1]) which IX. WHAT THE PRESENT RECONSTRUCTION TELLS
is locally equivalent to state ψMQM = (0, 0, 11). The overlap US ABOUT QUANTUM MECHANICS
(measured probability) between the states ψ13 and ψ is nega-
tive It is often said that reconstructions of quantum theory
within an operational approach are devoid of ontological com-
1 2 2 1 1 mitments, and that nothing can be generally said about the on-
P(ψ, ψ13 ) = (1 − − − ) = − . (74)
4 3 3 3 6 tological content that arises from the first principles or about
the status of the notion of realism. As a supporting argument
We conclude that “mirror quantum mechanics” – while be- one usually notes that within a realistic world view one would
ing a perfectly regular solution for a theory of two bits – can- anyway expect quantum theory at the operational level to be
not be consistently extended to also describe systems consist- deducible from some underlying theory of “deeper reality”.
ing of many bits. This also answers the question why we find After all, we have the Broglie-Bohm theory [58] which is a
in nature only four types of correlations as given in the table nonlocal realistic theory in full agreement with the predictions
(Figure 4) on the left, rather than all eight logically possible of (non-relativistic) quantum theory. Having said this, we can-
ones. not but emphasize that realism does stay “orthogonal” to the
basic idea behind our reconstruction.
Be it local or nonlocal, realism asserts that outcomes cor-
respond to actualities objectively existing prior to and inde-
VIII. HIGHER-DIMENSIONAL SYSTEMS AND STATE
UP-DATE RULE IN MEASUREMENT
pendent of measurements. On the other hand, we have shown
that the finiteness of information carrying capacity of quantum
systems is an important ingredient in deriving quantum theory.
Having obtained d = 3 for a two-dimensional system we This capacity is not enough to allow assignment of definite
have derived quantum theory of this system. We have also values to outcomes of all possible measurements. The ele-
reconstructed quantum mechanics of a composite system con- mentary system has the information carrying capacity of one
sisting of two qubits. Further reconstruction of quantum me- bit. This is signified by the possibility to decompose any state
chanics can be proceeded as in Hardy’s work [19]. In particu- of an elementary system (qubit) in quantum mechanics in two
lar, the reconstruction of higher-dimensional systems from the orthogonal states. In a realistic theory based on hidden vari-
two-dimensional ones and the general transformations of the ables and an “epistemic constraint” on an observer’s knowl-
state after measurement are explicitly given there. We only edge of the variables’ values one can reproduce this feature at
briefly comment on them here. the level of the entire distribution of the hidden variables [59].
In order to derive the state space, measurements and trans- That this is possible is not surprising if one bears in mind that
formations for a higher-dimensional system, we can use quan- hidden-variable theories were at the first place introduced to
tum theory of a two-dimensional system in conjunction with reproduce quantum mechanics and yet give a more complete
axiom 1. The axiom requires that upon any two linearly inde- description [67]. But any realism of that kind at the same
pendent states one can construct a two-dimensional subspace time assumes an infinite information capacity at the level of
that is isomorphic to the state space of a qubit (2-sphere). The hidden variables. Even to reproduce measurements on a sin-
state space of a higher dimensional system can be character- gle qubit requires infinitely many orthogonal hidden-variable
ized such that if the state is restricted to any given two dimen- states [60, 61, 62]. It might be a matter of taste whether or
sional subspace, then it behaves like a qubit. The fact that all not one is ready to work with this “ontological access bag-
other (higher-dimensional) systems can be built out of two- gage“ [60] not doing any explanatory work at the operational
dimensional ones suggests that the latter can be considered as level. But it is certainly conceptually distinctly different from
fundamental constituents of the world and gives a justification the theory analyzed here, in which the information capacity of
for the usage of the term “elementary system” in the formula- the most elementary systems – those which are by definition
tions of the axioms. not reducible further – is fundamentally limited.
When a measurement is performed and an outcome is ob- To further clarify our position consider the Mach-Zehnder
tain, our knowledge about the state of the system changes and interferometer in which both the path information and inter-
its representation in form of the probabilities must be updated ference observable are dichotomic, i.e. two-valued observ-
to be in agreement with the new knowledge acquired in the ables. It is meaningless to speak about “the path the parti-
measurement. This is the most natural update rule present in cle took in the interferometer in the interference experiment”
any probability theory. Only if one views this change as a because this would already require to assign 2 bits of infor-
real physical process conceptual problems arise related to dis- mation to the system, which would exceed its information
continuous and abrupt “collapse of the wave function”. There capacity of 1 bit [63]. The information capacity of the sys-
is no basis for any such assumption. Associated with each tem is simply not enough to provide definite outcomes to all
outcome is the measurement vector p. When the outcome is possible measurements. Then, by necessity the outcome in
observed the state after the measurement is updated to p and some experiments must contain an element of randomness and
the measurement will be a certain transformation on the ini- there must be observables that are complementarity to each
tial state. Update rules for more general measurements can other. Entanglement and consequently the violation of Bell’s
accordingly be given. inequality (and thus of local realism) arise from the possibility
 12

to define an abstract elementary system carrying at most one Proof. If the state is a product state then ||T ||2 = ||x||2||y||2 =
bit such that correlations (“00” and “11” in a joint measure- 1. On the other hand, assume that the state ψ = (x, y, T ) sat-
ment of two subsystems) are basis states. isfies ||T || = 1. Normalization (23) gives ||x|| = ||y|| = 1.
Let φ p = (−x, −y, T 0 = xyT ) be a product state. We have
P(ψ, φ p ) ≥ 0 and therefore
X. CONCLUSIONS
1 − ||x||2 − ||y||2 + Tr(T T T 0 ) = −1 + Tr(T T T 0 ) ≥ 0. (75)
Quantum theory is our most accurate description of nature The last inequality Tr(T T T 0 ) ≥ 1 can be seen as (T, T 0) ≥ 1
and is fundamental to our understanding of, for example, the where (, ) is the scalar product in Hilbert-Schmidt space. Since
stability of matter, the periodic table of chemical elements, the vectors T, T 0 are normalized, ||T || = ||T 0 || = 1, the scalar
and the energy of the sun. It has led to the development product between them is always (T, T 0) ≤ 1. Therefore, we
of great inventions like the electronic transistor, the laser, or have (T, T 0 ) = 1 which is equivalent to T = T 0 = xyT .
quantum cryptography. Given the enormous success of quan- QED
tum theory, can we consider it as our final and ultimate theory?
Quantum theory has caused much controversy in interpreting Lemma 2. The only product states belonging to S 12 are ψ1
what its philosophical and epistemological implications are. and ψ2 .
At the heart of this controversy lies the fact that the theory Proof. Let ψ p = (x, y, xyT) ∈ S 12 . We have
makes only probabilistic predictions. In recent years it was
however shown that some features of quantum theory that one 1 = P12 (ψ p , ψ1 ) + P12 (ψ p , ψ2 ) (76)
might have expected to be uniquely quantum, turned out to be 1
highly generic for generalized probabilistic theories. Is there = (1 + xe1 + ye1 + (xe1 )(ye1 )) (77)
4
any reason why the universe should obey the laws of quantum 1
theory, as opposed to any other possible probabilistic theory? + (1 − xe1 − ye1 + (xe1 )(ye1 )) (78)
4
In this work we have shown that classical probability the- 1
ory and quantum theory – the only two probability theories = (1 + (xe1 )(ye1 )) (79)
2
for which we have empirical evidences — are special in a ⇒ xe1 = ye1 = 1 ∨ xe1 = ye1 = −1 (80)
way that they fulfill three reasonable axioms on the systems’
information carrying capacity, on the notion of locality and ⇔ x = y = e1 ∨ x = y = −e1 . (81)
on the reversibility of transformations. The two theories can QED
be separated if one restricts the transformations between the
pure states to be continuous [19]. An interesting finding is Lemma 3. If the state ψ ∈ S 12 , then ψ′ = (R, 11)ψ ∈ S 34 and
that quantum theory is the only non-classical probability the- ψ′′ = (11, R)ψ ∈ S 34 .
ory that can exhibit entanglement without conflicting one or Proof. If ψ ∈ S 12 we have
more axioms. Therefore – to use Schrödinger’s words [64]
– entanglement is not only “the characteristic trait of quan- 1 = P12 (ψ, ψ1 ) + P12 (ψ, ψ2 ) (82)
tum mechanics, the one that enforces its entire departure from = P12 ((R, 11)ψ, (R, 11)ψ1 ) + P12 ((R, 11)ψ, (R, 11)ψ2 ) (83)
classical lines of thought”, but also the one that enforces the = P12 (ψ′ , ψ3 ) + P12 (ψ′ , ψ4 ). (84)
departure from a broad class of more general probabilistic the-
ories. Similarly, one can show that (11, R)ψ ∈ S 34 .
QED
Lemma 4. Let U be some operator with the following action
Acknowledgments
in the Hilbert-Schmidt space; U(ρ) = UρU † , and PT 1 and
PT 2 are partial transpositions with respect to subsystems 1
We thank M. Aspelmeyer, J. Kofler, T. Paterek and A. and 2, respectively. The following identity holds: PT1 UPT1 =
Zeilinger for discussions. We acknowledge support from PT2 U ∗ PT2 , where U ∗ is the complex-conjugate operator.
the Austrian Science Foundation FWF within Project No.
P19570-N16, SFB and CoQuS No. W1210-N16, the Euro- Proof. We can expand UP into some product basis in the
pean Commission Project QAP (No. 015848) and the Foun- Hilbert-Schmidt space U = i j ui j Ai ⊗ B j . We have
dational Question Institute (FQXi).
PT1 UPT1 (ρ1 ⊗ ρ2 ) = PT1 {UρT1 ⊗ ρ2 U † } (85)
X
= ui j u∗kl (A∗k ρ1 ATi ) ⊗ (B jρ2 B†l )
XI. APPENDIX i jkl
X
= PT2 { ui j u∗kl (A∗k ρ1 ATi ) ⊗ (B∗l ρT2 BTj )}
In this appendix we give the proofs of the lemmas from the i jkl
main text. X
= PT2 { u∗kl ui j (A∗k ⊗ B∗l )(ρ1 ⊗ ρT2 )(ATi ⊗ BTj )}
Lemma 1. The lower bound ||T || = 1 is saturated, if and only i jkl
∗
if the state is a product state T = xyT . = PT2 U PT2 (ρ1 ⊗ ρ2 ),
 13

for arbitrary operators ρ1 and ρ2 .

QED

[1] S. Kochen and E.P. Specker, The Problem of Hidden Variables (arXiv:0907.0909)
in Quantum Mechanics, J. Math. Mech. 17, 59 (1967). [22] C. Rovelli, Relational Quantum Mechanics, Int. J. Theor. Phys.
[2] J.S. Bell, On the Einstein-Podolsky-Rosen paradox, Physics 1, 35, 1637 (1996).
195-200 (1964); reprinted in J.S. Bell, “Speakable and Un- [23] A. Zeilinger, A Foundational Principle for Quantum Mechan-
speakable in Quantum Mechanics” (Cambridge Univ. Press, ics, Found. Phys. 29, 631 (1999).
Cambridge, 1987). [24] Č. Brukner and A. Zeilinger, Information and Fundamental El-
[3] A.J. Leggett, Nonlocal Hidden-Variable Theories and Quan- ements of the Structure of Quantum Theory, in “Time, Quan-
tum Mechanics: An Incompatibility Theorem, Found. Phys. 33, tum, Information”, Eds. L. Castell and O. Ischebeck (Springer,
1469 (2003). 2003) (arXiv:quant-ph/0212084). Č. Brukner and A. Zeilinger,
[4] S. Gröblacher, T. Paterek, R. Kaltenbaek, Č. Brukner, M. Information Invariance and Quantum Probabilities, Found.
Żukowski, M. Aspelmeyer and A. Zeilinger, An experimental Phys. 39, 677 (2009).
test of non-local realism, Nature 446, 871 (2007). [25] R. Clifton, J. Bub, and H. Halvorson, Characterizing Quantum
[5] I. Biaynicki-Birula and J. Mycielski, Nonlinear Wave Mechan- Theory in Terms of Information-Theoretic Constraints, Found.
ics, Ann. Phys. 100, 62 (1976). Phys. 33(11), 1561 (2003).
[6] A. Shimony, Proposed neutron interferometer test of some [26] A. Grinbaum, Elements of information-theoretic derivation of
nonlinear variants of wave mechanics, Phys. Rev. A 20, 394 the formalism of quantum theory, Int. J. Quant. Inf. 1(3), 289
(1979). (2003).
[7] C.G. Shull, D. K. Atwood, J. Arthur, and M. A. Horne, Search [27] W.K. Wootters, Statistical distance and Hilbert space, Phys.
for a Nonlinear Variant of the Schrödinger Equation by Neutron Rev. D 23, 357 (1981).
Interferometry, Phys. Rev. Lett. 44, 765 (1980). [28] D.I. Fivel, How interference effects in mixtures determine the
[8] R. Gähler, A.G. Klein, A. Zeilinger, Neutron Optical Tests of rules of quantum mechanics, Phys. Rev. A 59, 2108 (1994).
Nonlinear Wave Mechanics, Phys. Rev. A 23, 1611 (1981). [29] J. Summhammer, Maximum predictive power and the super-
[9] G. C. Ghirardi, A. Rimini and T. Weber, Unified dynamics for position principle, Int. J. Theor. Phys. 33, 171 (1994). J.
microscopic and macroscopic systems, Phys. Rev. D 34, 470 Summhammer, Quantum Theory as Efficient Representation of
(1986). Probabilistic Information, (2007) (arXiv:quant-ph/0701181).
[10] F. Károlyházy, Nuovo Cimento 42, 390 (1966). F. Károlyházy, [30] A. Bohr and O. Ulfbeck, Primary manifestation of symmetry.
Gravitation and Quantum Mechanics of Macroscopic Bodies Origin of quantal indeterminacy, Rev. Mod. Phys. 67, 1 (1995).
(Thesis, in Hungarian), Magyar Fizikai Folyóirat 22, 23 (1974). [31] A. Caticha, Consistency, amplitudes and probabilities in quan-
[11] L. Diosi, Models for universal reduction of macroscopic quan- tum theory, Phys. Rev. A 57, 1572 (1998).
tum fluctuations, Phys. Rev. A 40, 1165 (1989). [32] C.A. Fuchs, Quantum mechanics as quantum information (and
[12] R. Penrose, On Gravity’s role in Quantum State Reduction, only a little more), in Ed. A. Khrenikov “Quantum The-
Gen. Relativ. Gravit. 28, 581 (1996). ory: Reconstruction of Foundations” (Växjo, Växjo University
[13] P. Pearle, Reduction of the state vector by a nonlinear Press, 2002) (quant-ph/ 0205039). C.A. Fuchs and R. Schack,
Schrödinger equation, Phys. Rev. D 13, 857 (1976). Quantum-Bayesian Coherence, (2009) (arXiv:0906.2187).
[14] H. Barnum, J. Barrett, M. Leifer and A. Wilce, Cloning [33] P. Grangier, Contextual objectivity : a realistic interpreta-
and broadcasting in generic probabilistic models, (2006) tion of quantum mechanics, Eur. J. Phys. 23, 331 (2002)
(arXiv:quant-ph/061129). (arXiv:quant-ph/0012122); P. Grangier, Contextual objectivity
[15] H. Barnum, J. Barrett, M. Leifer and A.Wilce, A general and the quantum formalism, Proc. of the conference “Foun-
no-cloning theorem, Phys. Rev. Lett. 99, 240501 (2007) dations of Quantum Information” (Camerino, Italy, 2004)
(arXiv:0707.0620). (arxiv.org/quant-ph/0407025).
[16] J. Barrett, Information processing in general prob- [34] S. Luo, Maximum Shannon Entropy, Minimum Fisher Informa-
abilistic theories, Phys. Rev. A. 75, 032304 (2007) tion, and an Elementary Game, Found. Phys. 32, 1757 (2002).
(arXiv:quant-ph/0508211). [35] R. Spekkens, Evidence for the epistemic view of quantum
[17] S. Popescu and D. Rohrlich, Quantum nonlocality as an axiom, states: A toy theory, Phys. Rev. A 75, 032110 (2007).
Found. Phys. 24, 379 (1994). [36] P. Goyal, Information-geometric reconstruction of quantum
[18] G. W. Mackey, Quantum Mechanics and Hilbert Space, Amer- theory, Phys. Rev. A 78, 052120 (2008).
ican Mathematical Monthly 64, 45 (1957). [37] W. van Dam, Implausible Consequences of Superstrong Nonlo-
[19] L. Hardy, Quantum Theory From Five Reasonable Axioms cality, (2005) (arXiv:quant-ph/0501159).
(2001) (arXiv.org/quant-ph/0101012). [38] G. Brassard, H. Buhrman, N. Linden, A. Methot, A. Tapp and
[20] G.M. D’Ariano, Probabilistic theories: what is special about F. Unger, A limit on nonlocality in any world in which commu-
Quantum Mechanics?, in “Philosophy of Quantum Informa- nication complexity is not trivial, Phys. Rev. Lett. 96, 250401
tion and Entanglement”, Eds. A. Bokulich and G. Jaeger (Cam- (2006).
bridge University Press, Cambridge UK), (arXiv:0807.4383). [39] M. Pawlowski, T. Paterek, D. Kaszlikowski, V. Scarani, A. Win-
[21] P. Goyal, K.H. Knuth and J. Skilling, Origin of Com- ter, and M. Zukowski, A new physical principle: Information
plex Quantum Amplitudes and Feynman’s Rules (2009) Causality (2009) (arXiv:0905.2292).
 14

[40] M. Navascues and H. Wunderlich, A glance beyond the quan- [54] See, for example, J. Ambjorn, J. Jurkiewicz and R. Loll, Recon-
tum model (arXiv:0907.0372). structing the Universe, Phys. Rev. D 72 064014 (2005).
[41] K. Zyczkowski, Quartic quantum theory: an extension of the [55] H. Boerner, Representations of groups, (North- Holland pub-
standard quantum mechanics, J. Phys. A 41, 355302 (2008). lishing company, Amsterdam 1963).
[42] T. Paterek, B. Dakic, Č. Brukner, Theories of systems with lim- [56] R.A. Horn and C.R. Johnson, Matrix Analysis, (Cambridge
ited information content (2008) (arXiv:0804.1423). University Press, Chapter 8, 1990).
[43] A. Peres, Proposed test for complex versus quaternion quantum [57] R. Spekkens, privite communication.
theory, Phys. Rev. Lett. 42, 683 (1979). [58] D. Bohm, A Suggested Interpretation of the Quantum Theory
[44] H. Kaiser, E.A. George, and S.A. Werner, Neutron interferomet- in Terms of “Hidden Variables” I, Phys. Rev. 85, 166 (1952).
ric search for quaternions in quantum mechanics, Phys. Rev A D. Bohm, A Suggested Interpretation of the Quantum Theory in
29, 2276 (1984). Terms of “Hidden Variables” II, Phys. Rev. 85, 180 (1952).
[45] A. Peres, Quaternionic quantum interferometry, in “Quantum [59] See Ref [35] for a local version of such hidden-variable theory
Interferometry”, Eds. F. De Martini et al., (VCH Publ., 1996), in which quantum mechanical predictions are partially repro-
431-437 (arXiv:quant-ph/9605024). duced.
[46] R. D. Sorkin, Quantum Mechanics as Quantum Measure The- [60] L. Hardy, Quantum Ontological Excess Baggage, Stud. Hist.
ory, Mod. Phys. Lett. A 9, 3119 (1994) (arXiv:gr-qc/9401003). Philos. Mod. Phys. 35, 267 (2004).
[47] U. Sinha, C. Couteau, Z. Medendorp, I. Söllner, R. [61] A. Montina, Exponential growth of the ontological space di-
Laflamme, R. Sorkin and G. Weihs, Testing Born’s Rule in mension with the physical size, Phys. Rev. A 77, 022104 (2008).
Quantum Mechanics with a Triple Slit Experiment, (2008) [62] B. Dakic, M. Suvakov, T. Paterek, and Č. Brukner, Efficient
(arXiv:0811.2068). Submitted to the proceedings of Founda- Hidden-Variable Simulation of Measurements in Quantum Ex-
tions of Probability and Physics-5, Vaxjo, Sweden, August periments, Phys. Rev. Lett. 101, 190402 (2008).
2008. [63] Č. Brukner and A. Zeilinger, Young’s experiment and the finite-
[48] A. Grinbaum, Reconstruction of Quantum Theory, Brit. J. Phil. ness of information, Phil. Trans. R. Soc. Lond. A 360, 1061
Sci. 8, 387 (2007). (2002).
[49] D. Gross, M. Mueller, R. Colbeck, and O.C.O. Dahlsten, All [64] E. Schrödinger, Discussion of Probability Relations Between
reversible dynamics in maximally non-local theories are trivial Separated Systems, Proceedings of the Cambridge Philosophi-
(2009) (arXiv:0910.1840). O.C.O. Dahlstein, privite communi- cal Society 31 (1935) 555-563; 32 (1936): 446-451
cation. [65] Hardy considers unnormalized states and for that reason takes
[50] C.F. von Weizsäcker, 1958, Aufbau der Physik (Carl Hanser, K = d + 1 (in his notation) as the number of degrees of freedom.
München,1958). [66] As noted by Zyczkowski [41] it is thinkable that within
[51] J.A. Wheeler, Law without Law in Quantum Theory and Mea- the time scales of standard experimental conditions “hyper-
surement, Eds. J.A. Wheeler and W.H. Zurek (Princeton Uni- decoherence” may occur which cause a system described in the
versity Press, Princeton, 1983) 182. framework of the higher-order theory to specific properties and
[52] H. Barnum and A. Wilce, Information processing in convex behavior according to predictions of standard (complex) quan-
operational theories, to be published in DCM/QPL (Develop- tum theory.
ments in Computational Models / Quantum Programming Lan- [67] That this cannot be done without allowing nonlocal influences
guages) (Oxford University, 2009) (arXiv:0908.2352). from space-like distant regions is a valid point for itself, which
[53] S. Abramsky and B. Coecke, A categorical semantics of quan- we do not want to follow here further.
tum protocols, Proc. 19th IEEE Conference on Logic in Com-
puter Science, 415–425 (IEEE Computer Science Press, 2004).