The Stochastic-Quantum Theorem

Jacob A. Barandes1, ∗
1
Jefferson Physical Laboratory, Harvard University, Cambridge, MA 02138
(Dated: September 7, 2023)
This paper introduces several new classes of mathematical structures that have close connections
with physics and with the theory of dynamical systems. The most general of these structures,
called generalized stochastic systems, collectively encompass many important kinds of stochastic
processes, including Markov chains and random dynamical systems. This paper then states and
proves a new theorem that establishes a precise correspondence between any generalized stochastic
system and a unitarily evolving quantum system. This theorem therefore leads to a new formulation
of quantum theory, alongside the Hilbert-space, path-integral, and quasiprobability formulations.
The theorem also provides a first-principles explanation for why quantum systems are based on the
complex numbers, Hilbert spaces, linear-unitary time evolution, and the Born rule. In addition,
arXiv:2309.03085v1 [quant-ph] 3 Sep 2023

the theorem suggests that by selecting a suitable Hilbert space, together with an appropriate choice
of unitary evolution, one can simulate any generalized stochastic system on a quantum computer,
thereby potentially opening up an extensive set of novel applications for quantum computing.

cluding Hilbert spaces over the complex numbers, linear-

unitary evolution, and the Born rule.
Seen from another point of view, this stochastic-
I. INTRODUCTION quantum correspondence yields an alternative way to for-
mulate quantum theory, in the language of trajectories
In the development of physical theories, it sometimes unfolding stochastically in configuration spaces. This
turns out that existing definitions are too conceptually alternative formulation is distinct from the traditional
limiting, and that more flexible definitions are needed. Hilbert-space formulation [4, 5], the path-integral formu-
Working with more flexible definitions at a higher level of lation [6–8], and the quasi-probability formulation [9, 10].
abstraction or generality may make it easier to discover From a more practical perspective, turning this
new connections or prove new theorems, which would stochastic-quantum correspondence around suggests that
then also apply down at the lower level of the original unitarily evolving quantum systems can be put to work
definitions. simulating a very broad class of non-Markovian stochas-
This paper will argue that by appropriately generaliz- tic processes, thereby potentially opening up an extensive
ing standard definitions of dynamical systems, one can suite of new applications for quantum computers.
obtain novel classes of mathematical structures that en- Section II begins by defining deterministic generaliza-
compass an extensive array of physically important mod- tions of dynamical systems, followed by the introduction
els. As with a traditionally defined dynamical system, of important distinctions between indivisible, divisible,
each such mathematical structure describes a physical and Markovian dynamics. Section III provides a gener-
system moving deterministically or stochastically along alized definition of a system with stochastic laws, shows
some trajectory in a configuration space, albeit with a how to represent such a generalized stochastic system
more general set of laws than according to standard def- in the formalism of linear algebra, describes connections
initions. (For pedagogical treatments of the standard between this work and the existing research literature,
theory of dynamical systems, see, for instance, [1–3].) defines the relationship between a composite system and
This paper also states and proves a new theorem show- its subsystems, and introduces the crucial notion of a
ing that despite being based on trajectories in configu- unistochastic system. Section IV states the stochastic-
ration spaces, the newly introduced class of generalized quantum theorem, whose proof is this paper’s primary
stochastic systems actually includes all quantum systems. goal, and then discusses some important corollaries and
As a consequence, this stochastic-quantum theorem of- provides a simple example of the theorem in practice.
fers a more conceptually transparent way to understand Section V lays out the theorem’s proof, which entails ex-
quantum systems, with superpositions no longer regarded plicitly constructing the claimed correspondence between
as literal blends of physical states. The theorem also pro- stochastic systems and quantum systems along the way.
vides a first-principles explanation for features of quan- Section VI concludes the paper with a brief discussion of
tum theory that are usually taken to be axiomatic, in- future work.

∗ jacob [email protected]
2

II. DETERMINISTIC SYSTEMS that will be called the system’s dynamical map.
This dynamical map f takes as inputs any state
A. Generalized Dynamical Systems i and any time t, and outputs a state f (i, t) ∈ X :

i, t 7→ f (i, t) ∈ X [for all i ∈ X , t ∈ T ]. (3)

Dynamical systems are abstract mathematical struc-
tures that usefully model many deterministic physical
processes. According to the standard definition [1–3], • Fixing the time t turns f into a time-dependent
a dynamical system consists of a map representing some dynamical map
kind of evolution law that can be repeatedly applied to
the elements of some set of states. A dynamical system ft : X → X (4)
is usually assumed to be divisible, in the sense that one
can ‘divide up’ its evolution law over any time duration defined by
into well-defined evolution laws that describe intermedi-
i 7→ ft (i) ≡ f (i, t) [for all i ∈ X ]. (5)
ate time durations. The more general case would be an
indivisible dynamical system that might lack this feature.
The terms ‘divisible’ and ‘indivisible’ for dynamical • Without any important loss of generality, the set
laws are remarkably new. The terminology appears to of times T will be assumed to include an element
be due to Wolf and Cirac, who introduced it in a 2008 denoted by 0 and called the initial time. It will
paper [11] on quantum channels.1 be further assumed that at the initial time 0, the
To accommodate the eventual possibility of indivisible time-dependent dynamical map ft trivializes to the
evolution, this paper will define a generalized dynamical identity map idX on X :
system to mean a tuple of the form
f0 = idX , or f0 (i) = i [for all i ∈ X ]. (6)
(X , T , f ) (1)
• One can regard the argument i appearing in the ex-
that consists of the following data. pression ft (i) as an initial state of the system at the
initial time 0, with the time-dependent dynamical
• The symbol X denotes a set that will be called map ft then describing the evolution of that state
the generalized dynamical system’s state space (or i from the initial time 0 to the time t.
phase space), and whose individual elements i ∈ X
denote the system’s (allowed) states. • Given a fixed state i, the set of states

• Note that X may or may not be a finite set, and it {ft (i) | t ∈ T } ⊂ X (7)
may or may not involve additional structure in its
definition, such as measure-theoretic structure or a describes the orbit, or trajectory, of the initial state
vector-space structure. For the purposes of this pa- i through the system’s state space X according to
per, no such additional structure will be specified the dynamical map f .
or assumed. More broadly, for reasons of brevity
and simplicity, this paper will entirely set aside In many applications, one takes the set of times T to
measure-theoretic considerations that arise for the be a semigroup, meaning that the definition of T includes
case of uncountable sets. an associative binary operation ⋆ (which is often denoted
instead by + in the commutative case):
• The symbol T denotes a set of times t ∈ T , where
T may or may not be isomorphic to a subset of the t, t′ 7→ t ⋆ t′ ∈ T [for all t, t′ ∈ T ], (8)
real line R under addition.

• The symbol f denotes a map (t ⋆ t′ ) ⋆ t′′ = t ⋆ (t′ ⋆ t′′ ) [for all t, t′ , t′′ ∈ T ]. (9)

f :X ×T →X (2) One usually also takes the initial time 0 to be the identity
element under this binary operation,

0⋆t=t⋆0=t [for all t ∈ T ], (10)

1 Note that this terminology is unrelated to the much older concept
of infinite divisibility, which refers to a probability distribution
in which case T becomes a monoid, meaning a semigroup
that can be expressed as the probability distribution of a sum with an identity element. If, furthermore, every time t
of any integer number of independent and identically distributed has an inverse t′ such that t ⋆ t′ = t′ ⋆ t = 0, then T
random variables. becomes a group.
 3

B. Divisible Dynamical Systems where ◦ denotes function composition. The di-

visibility condition means that the time-dependent
In the most general case, a generalized dynamical sys- transition map gt←t′′ factorizes or ‘divides up’ into
tem will not provide a way to evolve a system from a a part gt′ ←t′′ that carries out the evolution from t′′
non-initial time t′ ̸= 0 to another time t. A generalized to t′ , followed by a part gt←t′ that carries out the
dynamical system will also generically lack any means evolution from t′ to t.
of ‘dividing up’ the evolution from 0 to t ̸= 0 into well- • Choosing the time t′ in the time-dependent tran-
defined forms of evolution over intermediate time dura- sition map gt←t′ to be the initial time 0 naturally
tions between 0 and t (even assuming that the set of defines a time-dependent dynamical map (4),
times T has a notion of ordering). To make contact with
the kinds of dynamical systems considered more widely ft ≡ gt←0 [for all t ∈ T ], (18)
in the research literature, it will therefore be necessary to
introduce a somewhat less general class of mathematical which then also defines an overall dynamical map
structures. f : X × T → X according to f (i, t) ≡ ft (i). The
This paper will define a divisible dynamical system to trivialization condition g0←0 = idX from (16) en-
be a tuple of the form sures that ft satisfies the corresponding trivializa-
tion condition f0 = idX in (6). It follows that every
(X , T , g). (11) divisible dynamical system is, in particular, a gen-
eralized dynamical system, one that includes the
• Here X is a state space and T is a set of times additional structure that corresponds to having a
forming a monoid, whose identity element 0 plays transition map g.
the role of an initial time, as usual.
• Meanwhile, setting t′′ = 0 in the divisibility con-
• The symbol g denotes a map dition (17) on the transition map g yields the sub-
g :X ×T2 →X (12) sidiary divisibility condition

that will be called the divisible dynamical system’s ft = gt←t′ ◦ ft′ [for all t, t′ ∈ T ], (19)
transition map. This transition map g takes as in- which means that the time-dependent dynamical
puts any state i and any pair of times t, t′ , and map ft ‘divides up’ into a part ft′ that carries out
outputs a state g(i, (t, t′ )) ∈ X : the evolution from the initial time 0 to t′ , followed
by a part gt←t′ that carries out the evolution from
i, t, t′ 7→ g(i, (t, t′ )) ∈ X (13)
′
t′ to t.
[for all i ∈ X , t, t ∈ T ].
• It also follows from the subsidiary divisibility condi-
The transition map g here should be understood as tion (19), together with the trivialization condition
describing the evolution or transition of the state i f0 = idX in (6), that the time-dependent dynamical
at the time t′ to the state g(i, (t, t′ )) at the time t. map ft has a well-defined inverse ft−1 given by
• Fixing two times t, t′ turns g into a time-dependent ft−1 = g0←t [for all t ∈ T ]. (20)
transition map
That is, a divisible dynamical system’s time-
gt←t′ : X → X (14) dependent dynamical maps ft are invertible.2
defined by
C. Markovian Dynamical Systems
i 7→ gt←t′ (i) ≡ g(i, (t, t′ )) (15)
[for all i ∈ X ].
A divisible dynamical system will be called memory-
This time-dependent transition map will be re- less, or Markovian, if it has the special property
quired to trivialize to the identity map idX on X
gt⋆t′ ←t′ = gt←0 [for all t, t′ ∈ T ]. (21)
when t′ = t,
gt←t = idX [for all t ∈ T ], (16)
2 If the set of times T has a well-defined ordering relation, and
or gt←t (i) = i [for all i ∈ X , t ∈ T ],
if one modifies the definition of a divisible dynamical system
as well as satisfy the divisibility condition (X , T , g) by restricting the transition map (12) so that gt←t′ is
only defined when t′ comes before t according to that ordering
gt←t′′ = gt←t′ ◦ gt′ ←t′′ (17) relation, then the arguments leading to (20) will break down.
In that case, the time-dependent dynamical maps ft will not
′ ′′
[for all t, t , t ∈ T ], necessarily have inverses.
4

Equivalently, using ft ≡ gt←0 from (18), one can write III. STOCHASTIC SYSTEMS
this special property as
A. Generalized Stochastic Systems
′
gt⋆t′ ←t′ = ft [for all t, t ∈ T ]. (22)
This paper will be primarily concerned with a mathe-
matical structure that replaces the deterministic behav-
It then follows immediately from the subsidiary divisibil- ior of a generalized dynamical system (1) with proba-
ity condition (19) that the dynamical map f : X ×T → X bilistic, or stochastic, behavior. A generalized stochastic
gives a semigroup action of the set of times T on the state system will be defined to mean a tuple of the form
space X , in the sense that the semigroup operation ⋆ on
T is mapped to function composition: (C, T , Γ, p, A) (24)

ft⋆t′ = ft ◦ ft′ [for all t, t′ ∈ T ]. (23) that consists of the following data.

• The symbol C will denote a set called the system’s

This equation is an example of Markovianity, or a Markov configuration space, and the elements of C will be
property. Roughly speaking, a Markov property implies called the system’s (allowed) configurations. Con-
that time evolution depends only on duration, and not figurations and configuration spaces will play an
on history. analogous role for generalized stochastic systems
Notice that the Markov property (23) is phrased di- that states and state spaces play for generalized
rectly in terms of a dynamical map f , so it can be im- dynamical systems.
posed on a generalized dynamical system (X , T , f ), with-
out any need to invoke a transition map g. In that case, it • The reason for switching the terminology from
also follows that there is no longer a meaningful distinc- ‘states’ to ‘configurations’ is conceptual. In ap-
tion between a generalized dynamical system (X , T , f ) plications of the theory of dynamical systems to
and a divisible dynamical system (X , T , g). As such, if physical situations, such as in classical Hamilto-
the Markov property (23) is imposed, then the resulting nian mechanics, the notion of a ‘state’ is often taken
mathematical structure will simply be called a Markovian to include rates of change or momenta in addition
dynamical system. to configurations, because defining a ‘state’ in that
way can make it possible to obtain deterministic
What most references call a ‘dynamical system’ corre- laws in the form of first-order differential equations.
sponds to what this paper would call a Markovian dy- For a generalized stochastic system, by contrast,
namical system. That is, a dynamical system, without the system’s probabilistic laws may mean that rates
any further qualifiers, is a tuple (X , T , f ) for which X is of change and momenta are not well-defined in the
a state space, T is a set of times forming a monoid, and absence of specifying a trajectory. In that case, the
f : X ×T → X is a dynamical map satisfying the Markov only available notion of a ‘state’ is limited to the
property (23). It follows that a generalized dynamical more rudimentary notion of a ‘configuration.’
system (1), as befits its name, really is a generalization
of the kinds of dynamical systems that are usually con- • For the purposes of this paper, C will be assumed to
sidered in textbooks, in the research literature, and in be a finite set, with some (possibly very large) num-
many applications. ber N of elements denoted by 1, 2, 3, . . . , N , and
In the most general cases, however, generalized dynam- labeled abstractly by Latin letters i, j, . . . . By an
ical systems will not satisfy the Markov property (23). abuse of notation, the symbol C will sometimes be
Nor will it necessarily be possible to obtain a given gen- used to refer to the generalized stochastic system
eralized dynamical system (X , T , f ) by starting with a as a whole.
divisible dynamical system (X , T , g) and then defining a
dynamical map f according to ft ≡ gt←0 for all times • The symbol T will again denote a set of times, in-
t, as in (18). Indeed, in light of the inversion formula cluding a time 0 that will be called the system’s
(20), the absence of any inverse time-dependent dynam- initial time.
ical maps ft−1 would present an immediate obstruction
• The symbol Γ will denote a map
to deriving a dynamical map f from a transition map g.
A generalized dynamical system (X , T , f ) that cannot Γ : C 2 × T → [0, 1] ⊂ R (25)
be derived from a divisible dynamical system (X , T , g)
will lack the structure needed to divide up its time evo- that will be called the system’s stochastic map,
lution into intermediate durations, and will therefore be where [0, 1] denotes the closed unit interval 0 ≤
called an indivisible dynamical system. x ≤ 1. Each value of this stochastic map will be
 5

labeled as The normalization condition (28) on the stochastic

map Γ and the normalization condition (32) on the
Γij (t) ≡ Γ((i, j), t) (26) standalone probabilities at the initial time 0 then
[for all i, j ∈ C, t ∈ T ], together ensure that the probability distribution p
satisfies the standard normalization condition more
and will be called the conditional probability or
generally:
transition probability p(i, t|j, 0) for the system to
be in its ith configuration at the time t, given that N
X
the system is in its jth configuration at the initial pi (t) = 1 [for all t ∈ T ]. (34)
time 0: i=1

Γij (t) ≡ p(i, t|j, 0) ∈ [0, 1] (27) • Note that the definition of the stochastic map Γ
[for all i, j ∈ C, t ∈ T ]. is independent of the choice of standalone proba-
bilities p1 (0), . . . , pN (0) at the initial time 0. That
• The stochastic map Γ will be required to satisfy the is, Γ can be freely adjusted independently of those
standard normalization condition initial standalone probabilities. Importantly, the
N N Bayesian marginalization condition (33) therefore
defines a linear relationship between the standalone
X X
Γij (t) = p(i, t|j, 0) = 1 (28)
i=1 i=1 probabilities p1 (0), . . . , pN (0) at the initial time 0
[for all j ∈ C, t ∈ T ], and the standalone probabilities p1 (t), . . . , pN (t) at
any other time t. In the work ahead, it will be
as well as the initial condition argued that this linear relationship is ultimately
( responsible for the linearity of time evolution in
1 for i = j,
Γij (0) ≡ δij ≡ (29) quantum theory.
̸ j,
0 for i =
• The stochastic map Γ will not be assumed to have
where δij is the usual Kronecker delta. anything like a Markov or divisibility property.
• The symbol p will denote a map More precisely, no assumption will be made that
given any pair of times t, t′ , there will exist a time
p : C × T → [0, 1] ⊂ R (30) t′′ such that the following Markov property holds:
that will be called the system’s probability distribu- N
X
tion. Each value of this map will be labeled as Γij (t) = Γik (t′′ )Γkj (t′ ) (35)
k=1
pi (t) ≡ p(i, t) ∈ [0, 1] (31) [for all i, j ∈ C].
[for all i ∈ C, t ∈ T ],
Nor will it be assumed that Γ satisfies even the
and will be called the standalone probability for the weaker divisibility condition that for any pair of
system to be in its ith configuration at the time t. times t, t′ , there exists a set of real-valued, non-
• Only the standalone probabilities p1 (0), . . . , pN (0) negative quantities Xik ≥ 0 for which3
at the initial time 0 will be taken to be freely N
X N
X
adjustable, subject to the standard normalization Xij = 1 and Γij (t) = Xik Γkj (t′ ) (36)
condition i=1 k=1
N
X [for all i, j ∈ C].
pj (0) = 1. (32)
j=1
That is, a generalized stochastic system will gener-
ically be non-Markovian and indivisible, in a sense
The standalone probabilities at every other time that parallels the notion of indivisibility for a gen-
t will be assumed to be defined by the following eralized dynamical system, as introduced earlier in
Bayesian marginalization condition: this paper.
N
X
pi (t) ≡ Γij (t)pj (0) (33)
j=1 3 Note that if the quantities Γkj (t′ ) are regarded as forming an
N
X N × N stochastic matrix, a term to be defined shortly, then the
= p(i, t|j, 0)pj (0) inverse matrix, if it exists, will always have negative entries un-
j=1 less both matrices are permutation matrices. It follows that one
cannot safely define the non-negative quantities Xij by invoking
[for all i ∈ C, t ∈ T ]. inverse matrices.
6

• This notion of non-Markovianity represents a dis- • Fixing the time t, let p(t), called the system’s (time-
tinct way to generalize the Markovian case, as com- dependent) probability vector, denote the N × 1 col-
pared with most forms of non-Markovianity de- umn vector whose ith entry is the ith standalone
scribed in textbooks and in the research literature. probability pi (t):
According to those more traditional forms of non-  
Markovianity, one assumes that the set of times T p1 (t)
has an ordering relation, and one further assumes p(t) ≡  ... . (40)
 
the existence of higher-order conditional probabil-
pN (t)
ities p(i, t|j1 , t1 ; j2 , t2 ; . . . ) that are conditioned on
arbitrarily many times t1 , t2 , . . . . From that more
traditional standpoint, the system is Markovian if • Again fixing the time t, let Γ(t), called the sys-
and only if the latest conditioning time t1 always tem’s (time-dependent) transition matrix, denote
screens off all the earlier conditioning times t2 , . . . , the N × N matrix for which the entry in the
so that p(i, t|j1 , t1 ; j2 , t2 ; . . . ) = p(i, t|j1 , t1 ). The ith row, jth column is the transition probability
definition of a generalized stochastic system pre- Γij (t) ≡ p(i, t|j, 0):
sented in this paper does not assume the existence  
of such higher-order conditional probabilities in the Γ11 (t) Γ12 (t)
first place. Γ(t) ≡ Γ (t) . . .
 
21

• The symbol A will denote a commutative algebra ΓN N (t)
 
of maps of the form p(1, t|1, 0) p(1, t|2, 0)
= p(2, t|1, 0)
 .. .

(41)
.
A:C×T →R (37)
p(N, t|N, 0)
under the usual rules of function arithmetic, and
will be called the system’s algebra of random vari- It follows that for any time t, the transition matrix
ables. The individual values of each random vari- Γ(t) is a (column) stochastic matrix in the mathematical
able A ∈ A will be labeled as sense, meaning that its entries are all non-negative real
numbers,
ai (t) ≡ A(i, t) ∈ R [for all i ∈ C, t ∈ T ]. (38)
Γij (t) ≥ 0 [for all i, j ∈ C, t ∈ T ], (42)
Each such value ai (t) will be called the magnitude
of the random variable A when the system is in its and that its columns each sum to 1, as required in (28).
ith configuration at the time t. No assumption will The initial condition (29) on the stochastic map Γ then
be made here that these magnitudes are all distinct, becomes the statement that the transition matrix Γ(0)
even at any fixed time t. at the initial time 0 is just the N × N identity matrix 1:
• For the purposes of this paper, the algebra of ran- 
1 0

dom variables A will always be taken to be max-
Γ(0) = 1 ≡ 0 . . . . (43)
 
imal, in the sense of containing every well-defined
map of the form (37). 1

• Fixing the time t, the expectation value ⟨A(t)⟩ of a Observe that the Bayesian marginalization condition (33)
random variable A will denote its statistical average naturally takes the form of matrix multiplication:
according to the probability distribution p at the
time t: p(t) = Γ(t)p(0) [for all t ∈ T ]. (44)
N
X
⟨A(t)⟩ ≡ ai (t)pi (t). (39)
i=1
C. Connections with Other Constructions

There is a definite sense in which a generalized stochas-

B. Ingredients from Linear Algebra tic system (C, T , Γ, p, A), as introduced in (24), pro-
vides a probabilistic extension of a generalized dynam-
Given a generalized stochastic system (C, T , Γ, p, A), ical system (X , T , f ), as introduced in (1). Indeed, if the
with C a configuration space of finite integer size N , it will stochastic map Γ : C 2 × T → [0, 1] from (25) outputs
be convenient to introduce some formalism from linear only the trivial probabilities 1 and 0, then Γ is effectively
algebra. deterministic. In that case, one can naturally set X ≡ C
 7

and define a dynamical map f : X × T → X according • Notice that a random dynamical system assigns
to a well-defined probability to any set of dynamical
maps {fω | ω ∈ Ω′ } that corresponds to a measur-
f (j, t) = i if and only if Γ((i, j), t) = 1 (45) able subset Ω′ of the sample space Ω. Thus, after
[for all i, j ∈ X , t ∈ T ]. fixing a state j at the initial time 0, a random dy-
namical system assigns a well-defined probability
A generalized stochastic system also naturally general- to any set of trajectories of the form {{fω,t (j) | t ∈
izes a mathematical structure called a stochastic dynam- T } | ω ∈ Ω′ }. (The nested subset notation here
ical system, or random dynamical system [12, 13]. is just intended to make clear that this expres-
• The definition of a random dynamical system starts sion refers to an indexed set of indexed sets.) In
with what this paper would call a generalized dy- particular, by generalizing the right-hand side of
namical system (X , T , f ), with T a monoid, and (50) to impose more conditions on the system’s dy-
replaces the single dynamical map f with a prob- namical maps, one can define conditional probabil-
abilistic family or ensemble of dynamical maps ities p(i1 , t1 ; i2 , t2 ; . . . |j, 0) that involve arbitrarily
{fω | ω ∈ Ω} that are indexed by a point ω in some many times. Random dynamical systems therefore
sample space Ω. That is, a randomly sampled point contain far more information than the generalized
ω in Ω picks out an entire dynamical map fω , as a stochastic systems defined in this paper, meaning
whole. that a generalized stochastic system indeed repre-
sents a less constrained, more general kind of math-
• For any randomly sampled point ω and any fixed ematical structure.
time t, one can define a map

fω,t : X → X (46)
In a sense, the absolutely most general kind of stochas-
by tic mathematical structure is a stochastic process, which
essentially requires only the specification of a state space
i 7→ fω,t (i) ≡ fω (i, t) [for all i ∈ X ]. (47) X , a set of times T , an initial probability distribution p,
and a set of one or more time-dependent random vari-
A random dynamical system is then assumed to ables A. (See [14–17] for textbook treatments.) Im-
satisfy an initial condition that generalizes (6), portantly, the definition of stochastic process lacks the
fω,0 = idX [for all ω ∈ Ω], (48) specification of a dynamical law. By requiring a dynam-
ical law in the form of a stochastic map Γ, a generalized
as well as a Markov property that generalizes (23), stochastic system (C, T , Γ, p, A) is not quite as general
as a stochastic process, but will still be general enough
fω′ ,t⋆t′ = fθ(ω′ ,t′ ),t ◦ fω′ ,t′ (49) to encompass a large class of physical and mathematical
[for all ω ′ ∈ Ω, t, t′ ∈ T ]. models.
Here θ : Ω × T → Ω is a map that is part of the def- In particular, one can regard any Markov chain as a
inition of the random dynamical system, and spec- special case of a generalized stochastic system. The start-
ifies a deterministic rule for updating the sample ing place is to assume that the set of times T is isomor-
point ω ′ and the dynamical map fω′ to accommo- phic to the integers, T ∼ = Z, with each time t = n δt an
date replacing the original initial time 0 with the integer number n ∈ Z of steps of some fixed, elementary
effectively new initial time t′ . That is, θ is neces- time scale δt. One then further assumes that for each
sary to account for any (purportedly deterministic) integer n, the time-dependent transition matrix Γ(n δt)
evolution in the underlying source of randomness originally defined in (41) can be expressed as the nth
itself. power of the transition matrix Γ(δt) that implements the
evolution for just the first time step δt:
• It follows from the foregoing definitions that the
conditional probability p(i, t|j, 0) for the system to n
be in its ith state at the time t, given that the Γ(n δt) = [Γ(δt)] [for all n ∈ Z]. (51)
system is in its jth state at the initial time 0, is
obtained by adding up the probabilities for all the
More broadly, a generalized stochastic system can there-
points ω in the sample space Ω whose correspond-
fore be understood as a kind of non-Markovian general-
ing dynamical maps fω take the state j at the initial
ization of a Markov chain.
time 0 and yield the state i at the final time t:
As this paper will also show, the class of generalized
p(i, t|j, 0) = probability({ω ∈ Ω | fω,t (j) = i}) (50) stochastic systems essentially includes all quantum sys-
[for all i, j ∈ X , t ∈ T ]. tems as well.
8

D. Composite Systems and Subsystems Given a composite system (C, ˜ T̃ , Γ̃, p̃, Ã) with configu-
˜ ′
ration space C = C×C , as in (52), the composite system’s
Introducing a notation of primes and tildes now to set of times T̃ trivially defines two sets of times T and
distinguish between specific generalized stochastic
sys- T ′ according to
tems, a generalized stochastic system C, ˜ T̃ , Γ̃, p̃, Ã will
T ≡ T ′ ≡ T̃ . (58)
be called a composite system if its configuration space C˜
is naturally expressible as a nontrivial Cartesian product Meanwhile, the composite system’s probability distribu-
of two sets C and C ′ : tion p̃ defines two probability distributions p : C × T →
[0, 1] and p′ : C ′ × T ′ → [0, 1] according to the respective
C˜ = C × C ′ . (52) marginalization formulas
′
A composite system has the following salient features. N
X
pi (t) ≡ p̃ii′ (t) [for all i ∈ C, t ∈ T ], (59)
• Letting N denote the size of C, and letting N ′ de- i′ =1
note the size of C ′ , the composite system’s configu- N
X
ration space C˜ has size Ñ ≡ N N ′ . p′i′ (t) ≡ p̃ii′ (t) [for all i′ ∈ C ′ , t ∈ T ′ ]. (60)
i=1
• Labeling the elements of C by unprimed Latin let-
ters i, j, . . . , and labeling the elements of C ′ by These two probability distributions p and p′ do not nec-
primed Latin letters i′ , j ′ , . . . , the composite sys- essarily turn C and C ′ into generalized stochastic systems
tem’s stochastic map Γ̃ : C˜2 × T̃ → [0, 1] has indi- of their own, in the sense of (24), due to the generic lack
vidual values that will be denoted by of well-defined stochastic maps Γ : C 2 × T → [0, 1] and
Γ′ : C ′2 × T ′ → [0, 1]. Indeed, in place of the Bayesian
Γ̃ii′ ,jj ′ (t) ≡ Γ̃(((i, i′ ), (j, j ′ )), t) marginalization condition (33), one instead has the rela-
tions
≡ p̃((i, i′ ), t|(j, j ′ ), 0) (53)
′ ′
′ ′ ′ N X N
N X
[for all i, j ∈ C, i , j ∈ C , t ∈ T̃ ]. X
pi (t) ≡ Γ̃ii′ ,jj ′ (t)p̃jj ′ (0) (61)
i′ =1 j=1 j ′ =1
The normalization condition (28) on the stochastic
map Γ̃ takes the form [for all i ∈ C, t ∈ T ],
′
N X
N X N
N′
N X
X
X p′i′ (t) ≡ Γ̃ii′ ,jj ′ (t)p̃jj ′ (0) (62)
Γ̃ii′ ,jj ′ (t) = 1 (54) i=1 j=1 j ′ =1
i=1 i′ =1
[for all i′ ∈ C ′ , t ∈ T ′ ].
[for all j ∈ C, j ′ ∈ C ′ , t ∈ T̃ ].
The two sets C and C ′ will be called the configu-
• The composite system’s probability distribution p̃ : ration spaces of subsystems of the composite system
C˜ × T̃ → [0, 1] has individual values at any fixed ˜ T̃ , Γ̃, p̃, Ã).
(C,
time t that will be denoted by

p̃ii′ (t) ≡ p((i, i′ ), t) (55) E. Unistochastic Systems

[for all i ∈ C, i′ ∈ C ′ , t ∈ T̃ ].
Recall that an N × N matrix U with complex-valued
entries is called unitary if it satisfies
The normalization condition (34) on the probabil-
ity distribution p̃ takes the form U † U = U U † = 1. (63)
′
N X
N
X Here † denotes the adjoint operation, meaning complex
p̃ii′ (t) = 1 [for all t ∈ T̃ ], (56) conjugation combined with the transpose operation.
i=1 i′ =1 A stochastic matrix X is called unistochastic if each
of its entries Xij is the modulus-squared |Uij |2 of the
and the Bayesian marginalization condition (33)
corresponding entry Uij of a unitary matrix U :
becomes the definition
|U11 |2 |U12 |2
   
N X
N ′ X11 X12
..  = |U |2 . . .
X
p̃ii′ (t) ≡ Γ̃ii′ ,jj ′ (t)p̃jj ′ (0) (57) X = X .
   
21 . 21
j=1 j ′ =1
XN N |UN N |2
′ ′
[for all i ∈ C, i ∈ C , t ∈ T̃ ]. (64)
 9

The study of unistochastic matrices was initiated in a B. Corollaries of the Theorem

1954 paper [18] by Horn, who originally called them
‘ortho-stochastic matrices.’ Unistochastic matrices were The proof of the stochastic-quantum theorem (65) will
given their modern name in 1989 by Thompson [19, 20], involve the construction of a representation of the given
and today orthostochastic matrices refer to the special generalized stochastic system in the formalism of Hilbert
case in which U can be taken to be a real orthogonal spaces, and will show that every generalized stochastic
matrix, meaning a unitary matrix with entries that are system corresponds to a unitarily evolving quantum sys-
all real-valued. tem on a Hilbert space. This paper will therefore es-
Every unistochastic matrix is doubly stochastic, or bis- tablish an important new correspondence between gen-
tochastic, meaning that its columns and rows each sum eralized stochastic systems and quantum systems, and
to 1.4 All 2 × 2 doubly stochastic matrices are unis- thereby turn some of the puzzling axiomatic ingredi-
tochastic,5 but this equivalence does not extend beyond ents of quantum theory—the complex numbers,8 Hilbert
the 2 × 2 case.6 Again beyond the 2 × 2 case, not every spaces, linear-unitary time evolution, and the Born rule
unistochastic matrix is orthostochastic.7 in particular—into the output of a theorem.
A generalized stochastic system (C, T , Γ, p, A) whose One can also read this stochastic-quantum correspon-
stochastic map Γ defines a unistochastic matrix Γ(t) at dence in the other direction, as the statement that all
every time t will be called a unistochastic system. One generalized stochastic systems can be modeled in terms
of the main goals of this paper will be to establish that of unitarily evolving quantum systems. From this per-
the study of generalized stochastic systems essentially re- spective, unitarily evolving quantum systems actually
duces to the study of unistochastic systems, which will represent the most general way to model a system with
also turn out to correspond to unitarily evolving quan- stochastic dynamical laws.
tum systems.

C. A Simple Example
IV. THE STOCHASTIC-QUANTUM THEOREM
The notion of embedding a generalized stochastic sys-
A. Statement of the Theorem tem into a unistochastic system does not always require
an elaborate construction. As a simple example, con-
With all these preliminaries now in place, this paper’s sider a discrete Markovian dynamical system (X , T , f )
next major goal will be to prove the following theorem, whose state space X has some finite size N , whose set of
which is a new result. times T is isomorphic to the integers Z under addition,
and whose dynamical map f : X × T → X satisfies the
appropriate version of the Markov property (23):
The Stochastic-Quantum Theorem
Every generalized stochastic system ft+t′ = ft ◦ ft′ [for all t, t′ ∈ T ∼
= Z]. (66)
(65)
can be regarded as a subsystem of a
unistochastic system. Based on these definitions and assumptions, it follows
from the Markov property and the trivialization condi-
Remarkably, to study the class of generalized stochastic tion f0 = idX from (6) that the time-dependent dynam-
systems, this theorem implies that it suffices to restrict ical map ft is invertible for any time t ∈ T , with inverse
one’s attention to the subclass of unistochastic systems. given by
ft−1 = f−t [for all t ∈ T ∼
= Z]. (67)

 †  Markovian dynamical systems of this kind provide a dis-

4 2 2
P P P
Proof: i |Vij | = i Vij Vij = V V jj = 1 and j |Vij | = cretization of many of the kinds of deterministic, time-
†
P  
j Vij Vij = V V ii
= 1, where bars denote complex conjuga- reversible physical systems that show up in classical
tion. QED
5 Proof: Every 2 × 2 doubly stochastic matrixis √of the form physics.
  √ 
x 1−x x − 1−x
1−x x , where 0 ≤ x ≤ 1, and the matrix √1−x √x
is easily seen to be unitary. QED
011
 
8
6 For example, the 3 × 3 doubly stochastic matrix 12 1 0 1 is Time-reversal transformations for quantum systems are carried
110
provably not unistochastic. out by anti-unitary operators, which take the general form V K,

1 1 1
 where V is unitary and K is an abstract operator that carries
7 For example, the 3 × 3 matrix √1 1 z z2 is unitary for z = out complex-conjugation. By definition, K 2 = 1 and Ki = −iK,
3 1 z2 z √
1 111
  where i ≡ −1 is the imaginary unit. One can therefore
exp(2πi/3), so the 3 × 3 matrix 3 1 1 1 is unistochastic. But show that i, K, and iK generate a Clifford algebra called the
111
this unistochastic matrix is not orthostochastic, because there pseudo-quaternions [21]. As a consequence, there is a sense in
does not exist a set of three mutually orthogonal 3 × 1 column which quantum systems are actually defined over the pseudo-
vectors that feature only 1s and −1s as entries. quaternions, rather than merely over the complex numbers.
10

Letting δt be the elementary duration of a single time For each fixed time t, the complex numbers Θij (t) collec-
step, and letting n ∈ Z denote an integer number of time tively form their own N × N matrix, which will be called
steps, the system’s specific state i at the time t = n δt the system’s time-evolution operator Θ(t):
can be identified as an N × 1 column vector with a 1 in  
its ith entry and 0s in all its other entries. Moreover, the Θ11 (t) Θ12 (t)
Θ(t) ≡ Θ (t) . . . . (69)
 
time-dependent dynamical map fn δt can be expressed as 21
the nth power of a fixed permutation matrix Σ, meaning ΘN N (t)
a matrix consisting of only 1s and 0s, with a single 1 in
each row and in each column. (As a concession to terminological conventions, the terms
Every permutation matrix is unitary, and is also unis- ‘matrix’ and ‘operator’ will be used more-or-less inter-
tochastic, because computing the modulus-squares of 1s changeably in what follows.)
and 0s gives back 1s and 0s. Hence, a discrete Marko- The normalization condition (28) on the system’s tran-
vian dynamical system defined in this way is already a sition matrix Γ(t) then becomes the summation condition
unistochastic system, and therefore trivially satisfies the N
X
stochastic-quantum theorem (65). 2
|Θij (t)| = 1 [for all j ∈ C, t ∈ T ], (70)
Interestingly, there exists an analytic interpolation of i=1
this discrete-time unistochastic system that yields a cor-
which can roughly be regarded as a generalization of a
responding continuous-time unistochastic system, at the
unitarity constraint. In keeping with the initial condition
cost of introducing nontrivial stochasticity into the in-
Γ(0) = 1 from (43), the time-evolution operator Θ(0) at
tervals between the discrete time steps. This new unis-
the initial time 0 will be taken to be the N × N identity
tochastic system is based on a time-dependent unis-
matrix 1:
tochastic matrix Γ(t) whose entries are the modulus-  
squares of the corresponding entries of the N ×N unitary 1 0
time-evolution operator defined by U (t) ≡ Σt/δt , and Θ(0) ≡ 1 ≡ 0 . . . . (71)
 
therefore satisfies Γ(n δt) = Σn for any integer n. One
1
can even go on to define an N × N self-adjoint Hamil-
tonian H as the infinitesimal generator of U (t), so there
is ultimately a Schrödinger equation underlying this sys- B. The Hilbert Space
tem.9
The time-evolution operator Θ(t) acts on an N -
V. THE HILBERT-SPACE REPRESENTATION dimensional Hilbert space H defined as the space CN
of N × 1 column vectors with complex-valued entries,
A. The Time-Evolution Operator
together with the standard inner product v † w for all
v, w ∈ CN :
To commence the proof of the stochastic-quantum the- H ≡ CN . (72)
orem (65), one starts with a given generalized stochas-
tic system (C, T , Γ, p, A). The non-negativity (42) of The standard orthonormal basis e1 , . . . , eN is defined by
the system’s transition probabilities, Γij (t) ≥ 0, means 1
0
0
0
that each transition probability can be written as the e1 ≡  .. , ..., eN ≡  .. . (73)
modulus-squared of a (non-unique) complex number . .
0 0
Θij (t): 0 1

2 These vectors are labeled as ei , where each value of

Γij (t) = |Θij (t)| [for all i, j ∈ C, t ∈ T ]. (68) i = 1, . . . , N denotes a configuration in the system’s con-
figuration space C, so this basis will be called the system’s
9
configuration basis.
To define the unitary matrix Σt/δt , one starts by using the fact
that every permutation matrix is also unitary to write the permu- There exists an associated set of rank-one projectors
tation matrix Σ as V † DV for some unitary matrix V and some P1 , . . . , PN defined by
diagonal matrix D, where the entries of D are the eigenvalues
of Σ. Because Σ is unitary, its eigenvalues are all phase factors Pi ≡ ei e†i = diag(0, . . . , 0, 1, 0, . . . , 0) [for all i ∈ C].
↑
exp(iθm ), where m = 1, . . . , N , and where each phase θm is a real ith entry
number. Setting all the eigenvalues to the power t/δt, so that
they each take the new form exp(iθm t/δt), one ends up with
(74)
the matrix Σt/δt ≡ V † Dt/δt V , which is still unitary and now These configuration projectors P1 , . . . , PN form a
depends analytically on the time parameter t. The quantum- projection-valued measure (PVM) [22, 23], meaning that
theoretic Hamiltonian matrix H for this system then has energy they satisfy the conditions of mutual exclusivity,
eigenvalues defined by Em ≡ −ℏθm /δt for m = 1, . . . , N , and is
diagonalized by the same unitary matrix V that diagonalizes Σ. Pi Pj = δij Pi [for all i, j ∈ C], (75)
 11

with δij again the usual Kronecker delta, and complete- Here A(t) denotes the N × N diagonal matrix
ness,  
N a1 (t) 0
N
ai (t)Pi ≡  0 . . .
X
A(t) ≡ . (82)
X  
Pi = 1, (76)
i=1 aN (t)
i=1

with 1 again the N × N identity matrix. Observe, in particular, that the magnitudes
a1 (t), . . . , aN (t) of the random variable A become
C. The Dictionary
the eigenvalues of the N × N matrix A(t).
Notice also that the formula for the standalone prob-
ability pi (t) in (78) and the formula for the expectation
It follows from a short calculation that the relationship
value ⟨A(t)⟩ in (81) are both special cases of the Born
Γij (t) = |Θij (t)|2 in (68) can be expressed in terms of a
rule.
matrix trace as

Γij (t) = tr Θ† (t)Pi Θ(t)Pj

(77)
E. An Aside on ‘Classical Wave Functions’
[for all i, j ∈ C, t ∈ T ].
Pausing for a moment, recall the smooth unistochas-
This ‘dictionary’ essentially translates between the the- tic interpolation of a discrete Markovian dynamical sys-
ory of generalized stochastic systems, as expressed by the tem described in Subsection IV C. For that system, the
left-hand side, and a corresponding Hilbert-space rep- unitary time-evolution operator U (t) ≡ Σt/δt trivial-
resentation, as expressed by the right-hand side. This izes to a permutation matrix Σn at every integer time
equation will turn out to form the core of a stochastic- step n δt. It follows that the system’s density matrix
quantum correspondence that will play a crucial role in ρ(t), as defined in terms of its initial value ρ(0) from
the proof of the stochastic-quantum theorem. (79), reduces to a diagonal matrix ρ(n δt) at each integer
time step. Taking the square root of each of its diago-
D. The Density Matrix
nal entries p1 (n δt), . . . , pN (n δt), and allowing for arbi-
trary phase factors, one obtains a ‘classical wave func-
tion’ with components Ψ1 (n δt), . . . , ΨN (n δt) satisfying
Inserting the dictionary (77) into the Bayesian |Ψi (n δt)|2 = pi (n δt) and capturing precisely the same
marginalization condition (33) yields the following equa- information as the diagonal density matrix ρ(n δt).
tion:
This classical wave function is the starting place for
pi (t) = tr(Pi ρ(t)) [for all i ∈ C, t ∈ T ]. (78) a representation of classical deterministic physics known
popularly as the ‘Koopman-von Neumann’ formulation,
Here ρ(t), which will be called the system’s density ma- due to its superficial resemblance to work by Koop-
trix, is an N × N time-dependent, self-adjoint, unit- man [24] and von Neumann [25, 26] in the 1930s. How-
trace, generically non-diagonal matrix defined for any ever, as pointed out explicitly in [27], Koopman and
fixed time t by von Neumann were actually using Hilbert spaces to rep-
resent observables, rather than to represent probabil-
ρ(t) ≡ Θ(t)ρ(0)Θ† (t), (79) ity distributions. The formulation of classical physics
where its value at the initial time 0 is the following N ×N in terms of ‘classical wave functions’ is more properly
diagonal matrix: due to Schönberg [28], Loinger [29], Della Riccia and
  Wiener [30], and Sudarshan [31].
N p1 (0) 0
..
X
ρ(0) ≡ pj (0)Pj ≡  0 . (80)
 
. F. The Kraus Decomposition
j=1 pN (0)
Importantly, one sees from this analysis that the linearity Returning to the proof, and fixing the time t, and let-
of the Bayesian marginalization condition (33) underlies ting β denote an integer from 1 to N that will play a
the linearity of the relationship (79) between the system’s conceptually different role from a configuration index, let
density matrix ρ(0) at the initial time 0 and its density Kβ (t) denote the N ×N matrix whose βth column agrees
matrix ρ(t) at other times t. with the βth column of the time-evolution operator Θ(t),
For any fixed time t, one can similarly express the ex- with 0s in all its other entries. That is, the entry in the
pectation value (39) of a random variable A as ith row, jth column of Kβ (t) is given by

⟨A(t)⟩ = tr(A(t)ρ(t)). (81) Kβ,ij (t) ≡ δβj Θij (t) [for all β, i, j ∈ C, t ∈ T ]. (83)
12

It follows from the summation condition (70) on Θ(t) • The dilated Hilbert space H̃ is defined as the tensor
that these new matrices satisfy the Kraus condition: product
N
H̃ ≡ H ⊗ H′ , (92)
Kβ† (t)Kβ (t) = 1 [for all t ∈ T ].
X
(84)
β=1 where H ≡ CN is the system’s original Hilbert
space (72), and where
Moreover, one can write the dictionary (77) as
′
N H ′ ≡ CN (93)
tr Kβ† (t)Pi Kβ (t)Pj
X

Γij (t) = (85)
is an ancillary Hilbert space whose dimension N ′
β=1
satisfies Ñ = N N ′ ≤ N 3 . That is, N ′ is an integer
[for all i, j ∈ C, t ∈ T ], lying in the bounded interval
and one can express the time-evolution rule (79) for the 1 ≤ N ′ ≤ N 2. (94)
system’s density matrix as
N • The first trace tr(· · · ) in (91) denotes the partial
Kβ (t)ρ(0)Kβ† (t).
X
ρ(t) = (86) trace over just the original Hilbert space H, and the
β=1 second trace tr′ (· · · ) similarly denotes the partial
trace over the ancillary Hilbert space H′ .
The matrices K1 (t), . . . , KN (t) are therefore Kraus oper-
ators [32], and (86) gives a Kraus decomposition of ρ(t). • Importantly, for each fixed time t, Ũ (t) is an Ñ × Ñ
Abstracting these results, one obtains a completely pos- unitary matrix that reduces to the Ñ × Ñ identity
itive trace-preserving map, or quantum channel, matrix Ũ (0) = 1̃ at the initial time 0.
Et : CN ×N → CN ×N , (87) • The symbol 1′ denotes the N ′ × N ′ identity matrix
defined on the algebra CN ×N of N × N matrices over the on the ancillary Hilbert space H′ .
complex numbers according to • Letting e′1 , . . . , e′N ′ denote the standard orthonor-
N mal basis for the ancillary Hilbert space H′ , in
Kβ (t)XKβ† (t).
X
X 7→ Et (X) ≡ (88) analogy with the configuration basis (73) for the
β=1 system’s original Hilbert space H, and letting the
primed Latin letters i′ , j ′ , . . . each denote an ele-
ment of an ancillary configuration space C ′ consist-
G. Dilation ing of the integers from 1 to N ′ , the symbol Pi′′
denotes the rank-one projector
By the Stinespring dilation theorem [33, 34], the quan-
tum channel (87) can be purified, meaning made into a Pi′′ ≡ e′i′ e′†
i′ = diag(0, . . . , 0, 1, 0, . . . , 0) (95)
↑
′
form of unitary time-evolution. Specifically, for some in- i th entry
teger Ñ in the bounded interval [for all i′ ∈ C ′ ],
N ≤ Ñ ≤ N 3 , (89) which is an N ′ × N ′ diagonal matrix with a 1 in its
there exists an Ñ × Ñ unitary matrix Ũ (t) on a poten- i′ th diagonal entry and 0s in all its other entries.
tially enlarged or dilated Hilbert space These projectors form a projection-valued measure
(PVM) on H′ satisfying the conditions of mutual
H̃ ≡ CÑ (90) exclusivity,
such that the dictionary (77) can be written as10 Pi′′ Pj′′ = δi′ j ′ Pi′′ [for all i′ , j ′ ∈ C ′ ], (96)
  
Γij (t) = tr tr′ Ũ † (t)[Pi ⊗ 1′ ]Ũ (t)[Pj ⊗ Pj′′ ] (91)
[for all i, j ∈ C, t ∈ T ]. that this matrix satisfies Ṽ † (t)Ṽ (t) = 1N 2 ×N 2 , so it defines
a partial isometry, which can always be extended to a uni-
The formula (91) involves a number of ingredients. tary N 3 × N 3 matrix Ũ(iβm)(ja) (t) by adding N 3 − N 2 addi-
tional columns that are mutually orthogonal with each other
and with the previous N 2 columns already in Ṽ (t), where the
new index a runs through N 2 possible values. These additional
10 From the starting assumptions presented here, one can sketch columns can always be chosen so that at the initial time 0, where
the following proof: Given N × N Kraus matrices Kβ (t), with V (0) = 1 is the N × N identity matrix, they make Ũ (0) co-
β = 1, . . . , N , define an N 3 × N 2 matrix Ṽ (t) according to incide with the N 3 × N 3 identity matrix. The last step is to
Ṽ(iβm)(jl) (t) ≡ Kβ,ij (t)δlm , treating (iβm) as the first index show that Ũ (t) satisfies (91),
P whose right-hand side reduces to
2 2
P
of Ṽ (t) and treating (jl) as its second index. One can show β,m |Ũ(iβm)(jj ′ ) (t)| = β |Kβ,ij (t)| . QED
 13

and completeness, is unistochastic. Moreover, by construction, Γ̃(t)

′
satisfies the marginalization condition
N
X
Pi′′ = 1′ . (97) N
X
′

i′ =1 Γ̃ii′ ,jj ′ (t) = Γij (t) (103)

i′ =1
• Note that the left-hand side of (91) is insensitive to [for all i, j ∈ C, j ′ ∈ C ′ , t ∈ T ],
the specific choice of j ′ ∈ C ′ on the right-hand side.
Extending the foregoing construction, and fixing the where, as in (91), the value of j ′ ∈ C ′ is irrelevant.
time t, one obtains an Ñ × Ñ transition matrix given by • Let the map p̃ : C˜ → [0, 1] be the probability distri-
the new dictionary bution defined by
˜ Ũ † (t)P̃ii′ Ũ (t)P̃jj ′


Γ̃ii′ ,jj ′ (t) ≡ tr (98) N
X
[for all i, j ∈ C, i′ , j ′ ∈ C ′ , t ∈ T ]. p̃ii′ (t) ≡ Γ̃ii′ ,jj ′ (t)pj (0) (104)
j=1
Here the trace is now over the dilated Hilbert space H̃, [for all i ∈ C, i′ , j ′ ∈ C ′ , t ∈ T ].
and
It follows from the marginalization condition (103),
P̃ii′ ≡ Pi ⊗ P̃i′ [for all i ∈ C, i′ ∈ C ′ ] (99) together with the Bayesian marginalization condi-
defines a rank-one projector on H̃. tion (33), that
Unfolding the notation, the dilated dictionary (98) re- N ′

duces to the statement that X

p̃ii′ (t) = pi (t) [for all i ∈ C, t ∈ T ], (105)
2 i′ =1
Γ̃ii′ ,jj ′ (t) = |Ũii′ ,jj ′ (t)| (100)
[for all i, j ∈ C, i′ , j ′ ∈ C ′ , t ∈ T ]. as in (59).

That is, each entry Γ̃ii′ ,jj ′ (t) of the Ñ × Ñ transition • Let the algebra à of random variables be the set of
matrix Γ̃(t) is the modulus-squared of the corresponding all maps of the form à : C˜ × T̃ → R.
entry Ũii′ ,jj ′ (t) of an Ñ × Ñ unitary matrix Ũ (t). Again, ˜ T̃ , Γ̃, p̃, Ã) is a composite
These results establish that (C,
a stochastic matrix with this special feature is called a
unistochastic system, and that the original generalized
unistochastic matrix.
stochastic system (C, T , Γ, p, A) can be regarded as one
of its subsystems. This conclusion completes the proof
H. The Dilated Generalized Stochastic System of the stochastic-quantum theorem (65). QED

One can now define a dilated generalized stochastic I. The Corresponding Quantum System
˜ T̃ , Γ̃, p̃, Ã) in the following way.
system (C,
• Let the dilated system’s configuration space C˜ be Corresponding to the dilated unistochastic system
defined as the Cartesian product ˜ T̃ , Γ̃, p̃, Ã) is a quantum system based on a Hilbert
(C,
space H̃ of dimension Ñ ≤ N 3 , with linear-unitary time
C˜ ≡ C × C ′ , (101) evolution encoded in the unitary time-evolution operator
Ũ (t).
where C is the original system’s configuration space,
Unistochastic matrices are not generally orthostochas-
with N elements labeled by unprimed Latin letters
tic, meaning that they are not guaranteed to be express-
i, j, . . . , and where C ′ is the configuration space of
ible in terms of real orthogonal matrices. As a conse-
an ancillary subsystem, with N ′ elements labeled
quence, if one is given a generalized stochastic system
by primed Latin letters i′ , j ′ , . . . .
whose N ×N transition matrix Γ(t) is already unistochas-
• Let the dilated system’s set of times T̃ be the orig- tic, then there is no guarantee that the corresponding
inal system’s set of times T : N × N unitary time-evolution operator U (t) can be as-
sumed to be a real orthogonal matrix. The stochastic-
T̃ ≡ T . (102) quantum correspondence therefore implies that in order
to provide Hilbert-space representations for the most gen-
• Let Γ̃ : C˜2 × T̃ → [0, 1] be the stochastic map eral kinds of generalized stochastic systems, the complex
defined according to the dilated dictionary (100). numbers C will be an important feature of quantum the-
Then for each fixed time t, the Ñ × Ñ matrix Γ̃(t) ory.
14

Whether in that case or more generally, of course, one of unitary time-evolution operator employed in standard
is always free to start with an N × N time-evolution op- textbook treatments of the measurement process, one in-
erator Θ(t) in (68) whose individual entries are all real. evitably finds that the measuring device ends up in one of
With that choice, the Ñ × Ñ unitary time-evolution oper- its measurement-reading configurations with a stochastic
ator Ũ (t) obtained from the Stinespring dilation theorem probability given by the general form of the Born rule.
will likewise be real, and will therefore be an orthogonal Hence, in principle, one has access to a quantum system’s
matrix. entire noncommutative algebra of observables.
However, it is important to keep in mind that from the With the stochastic-quantum correspondence in hand,
point of view of the stochastic-quantum correspondence, one can refer other exotic features of quantum systems
a generalized stochastic system’s Hilbert-space represen- back to their associated generalized stochastic systems
tations are convenient fictions, and so one is entirely free to give those features a more physically transparent in-
to assume that they involve the complex numbers any- terpretation. For example, as shown in [35], interference
way. Assuming that a given choice of Hilbert-space rep- and entanglement can be understood as artifacts of the
resentation is defined over the complex numbers, rather generic indivisibility of a generalized stochastic system’s
than merely over the real numbers, allows one to take dynamics. (It is also shown in [35] how this interpre-
advantage of the spectral theorem, eigenvectors of the tative framework navigates the various no-go theorems
time-evolution operator Ũ (t), and anti-unitary operators. that have appeared in the research literature over the
Assuming appropriate smoothness conditions in time, past century.)
one can further make use of a self-adjoint Hamiltonian
This overall approach to quantum foundations there-
H̃(t) with real-valued energy eigenvalues, as well as the
fore sheds new light on some of the strangest features of
Schrödinger equation. To the extent that these math-
quantum theory, in addition to suggesting novel applica-
ematical constructs are often taken by textbooks to be
tions of quantum computers. This approach might even
indisputable features of quantum systems, the complex
provide a helpful stepping stone for the development of
numbers become an avoidable part of quantum theory.11
self-consistent generalizations of quantum theory itself.

VI. DISCUSSION AND FUTURE WORK

The unitarily evolving quantum system that lies on ACKNOWLEDGMENTS

the other side of the stochastic-quantum correspondence
is not limited to a commutative algebra of observables The author would like to thank Scott Aaronson, David
represented by operators that are diagonal in the config- Baker, Craig Callender, Iris Cong, Jordan Cotler, Jenann
uration basis. Ismael, David Kagan, Logan McCarty, Arjun Mirani, and
Indeed, as explained in other work [35], one can model Jill North for helpful conversations.
the quantum measurement process of an observable rep-
resented by an arbitrary self-adjoint operator in terms of
a unistochastic system that contains a subject system, a
measuring device, and an environment, all as explicitly
defined subsystems. By taking the measuring device’s [1] R. Devaney. “An Introduction to Chaotic Dynamical Sys-
allowed configurations to correspond to definite readings tems”. 1989.
[2] S. H. Strogatz. Nonlinear Dynamics and Chaos. West-
of outcomes, and by taking the overall unistochastic sys- view Press, 1994.
tem’s transition matrix to be based on precisely the type [3] A. Katok and B. Hasselblatt. Introduction to the Mod-
ern Theory of Dynamical Systems. Cambridge University
Press, 1995.
11 Even if one assumes that a given choice of Hilbert-space repre- [4] P. A. M. Dirac. The Principles of Quantum Mechanics.
sentation is defined over the complex numbers, one can always Oxford University Press, 1st edition, 1930.
double the dimension of the Hilbert √ space from N to 2N and [5] J. von Neumann. Mathematische Grundlagen der Quan-
represent the imaginary unit i ≡ −1 by the 2 × 2 real ma- tenmechanik. Berlin: Springer, 1932.
trix 01 −1


0 , in which case all N × N unitary matrices become [6] P. A. M. Dirac. “The Lagrangian in Quantum Mechan-
2N × 2N real orthogonal matrices. Interestingly, one can then
ics”. Physikalische Zeitschrift der Sowjetunion, 3(1):64–
also represent the complex-conjugation operation K
needed for
anti-unitary operators as a 2 × 2 real matrix 01 10 . However,
72, 1933.
this approach still ultimately preserves the algebraic structure of [7] R. P. Feynman. The Principle of Least Action in
the complex numbers in the Hilbert space, in the sense that they Quantum Mechanics. PhD thesis, Princeton University,
will correspond to a subalgebra of matrices on the overall Hilbert Princeton, NJ, 1942.
space that commute with the Hamiltonian and all the system’s [8] R. P. Feynman. “Space-Time Approach to Non-
observables. Hence, this approach does not truly eliminate the Relativistic Quantum Mechanics”. Reviews of mod-
complex numbers from the Hilbert-space formalism of quantum ern physics, 20(2):367, April 1948. doi:10.1103/
theory. RevModPhys.20.367.
 15

[9] E. P. Wigner. “On the Quantum Correction For Thermo- [24] B. O. Koopman. “Hamiltonian Systems and Trans-
dynamic Equilibrium”. Physical Review, 40(5):749–759, formations in Hilbert Space”. Proceedings of the Na-
June 1932. doi:10.1103/PhysRev.40.749. tional Academy of Sciences, 17(5):315–318, 1931. doi:
[10] J. E. Moyal. “Quantum Mechanics as a Statistical The- 10.1073/pnas.17.5.315.
ory”. Mathematical Proceedings of the Cambridge Philo- [25] J. von Neumann. “Zur Operatorenmethode In Der Klas-
sophical Society, 45(1):99–124, January 1949. doi:10. sischen Mechanik”. Annals of Mathematics, 33(3):587–
1017/S0305004100000487. 642, 1932. doi:10.2307/1968537.
[11] M. M. Wolf and J. I. Cirac. “Dividing Quantum [26] J. von Neumann. “Zusatze Zur Arbeit ‘Zur Operatoren-
Channels”. Communications in Mathematical Physics, methode...’ ”. Annals of Mathematics, 33(4):789–791,
279:147–168, 2008. arXiv:math-ph/0611057, doi:10. 1932. doi:10.2307/1968225.
1007/s00220-008-0411-y. [27] T. F. Jordan and E. C. G. Sudarshan. “Dynamical
[12] J. Honerkamp. Stochastic Dynamical Systems: Concepts, Mappings of Density Operators in Quantum Mechan-
Numerical Methods, Data Analysis. Wiley-VCH, 1996. ics”. Journal of Mathematical Physics, 2(6):772–775,
[13] L. Arnold. Random Dynamical Systems. Springer, 1998. 1961. URL: http://scitation.aip.org/content/aip/
[14] M. Rosenblatt. Random Processes. Oxford University journal/jmp/2/6/10.1063/1.1724221, doi:10.1063/1.
Press, 1962. 1724221.
[15] E. Parzen. Stochastic Processes. Holden-Day, 1962. [28] M. Schönberg. “Application of Second Quantization
[16] J. L. Doob. Stochastic Processes. Wiley-Interscience, Methods to the Classical Statistical Mechanics (II)”.
1990. Il Nuovo Cimento, 10(4):419–472, April 1953. doi:
[17] S. Ross. Stochastic Processes. John Wiley and Sons, 2nd 10.1007/BF02781980.
edition, 1995. [29] A. Loinger. “Galilei Group and Liouville Equation”.
[18] A. Horn. “Doubly Stochastic Matrices and the Diagonal Annals of Physics, 20(1):132–144, 1962. doi:10.1016/
of a Rotation Matrix”. American Journal of Mathemat- 0003-4916(62)90119-7.
ics, 76(3):620–630, 1954. doi:10.2307/2372705. [30] G. Della Riccia and N. Wiener. “Wave Mechanics in
[19] R. C. Thompson. “Lecture notes from a Johns Hopkins Classical Phase Space, Brownian Motion, and Quantum
University lecture series”. Unpublished lecture notes, Theory”. Journal of Mathematical Physics, 7(8):1372–
1989. 1383, August 1966. doi:10.1063/1.1705047.
[20] P. Nylen, T.-Y. Tam, and F. Uhlig. “On the Eigenval- [31] E. C. G. Sudarshan. “Interaction between classical and
ues of Principal Submatrices of Normal, Hermitian and quantum systems and the measurement of quantum ob-
Symmetric Matrices”. Linear and Multilinear Algebra, servables”. Pramana, 6(3):117–126, March 1976. doi:
36(1):69–78, 1993. doi:10.1080/03081089308818276. 10.1007/BF02847120.
[21] E. C. G. Stueckelberg. “Quantum Theory in Real [32] K. Kraus. “General State Changes in Quan-
Hilbert Space”. Helvetica Physica Acta, 33(4):727– tum Theory”. Annals of Physics, 64(2):311–335,
752, 1960. URL: https://www.e-periodica.ch/digbib/ June 1971. URL: http://www.sciencedirect.
view?pid=hpa-001:1960:33::715#735. com/science/article/pii/0003491671901084,
[22] G. W. Mackey. “Induced Representations of Lo- doi:10.1016/0003-4916(71)90108-4.
cally Compact Groups I”. Annals of Mathematics, [33] W. F. Stinespring. “Positive functions on C*-algebras”.
55(1):101–139, 1952. URL: http://www.jstor.org/ Proceedings of the American Mathematical Society,
stable/1969423, doi:10.2307/1969423. 6(2):211–216, April 1955. doi:10.2307/2032342.
[23] G. W. Mackey. “Quantum Mechanics and Hilbert [34] M. Keyl. “Fundamentals of Quantum Information The-
Space”. The American Mathematical Monthly, 64(8):45– ory”. Physics reports, 369(5):431–548, October 2002.
57, 1957. URL: http://www.jstor.org/stable/ arXiv:0202122, doi:10.1016/S0370-1573(02)00266-1.
2308516, doi:10.2307/2308516. [35] J. A. Barandes. “The Stochastic-Quantum Correspon-
dence”, 2023. URL: https://arxiv.org/abs/2302.
10778, arXiv:2302.10778.