See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/226717715

On intrinsic randomness of dynamical systems

Article in Journal of Statistical Physics · January 1981

DOI: 10.1007/BF01008481

CITATIONS READS

89 123

3 authors, including:

Sheldon Goldstein Bibhu R. Misra

Rutgers, The State University of New Jersey Virginia Polytechnic Institute and State University
179 PUBLICATIONS 5,821 CITATIONS 36 PUBLICATIONS 1,222 CITATIONS

SEE PROFILE SEE PROFILE

All content following this page was uploaded by Sheldon Goldstein on 09 September 2014.

The user has requested enhancement of the downloaded file.

Journal of Statistical Physics, Vol. 25, No. 1, 1981

On Intrinsic Randomness of Dynamical Systems

S. Goldstein, ~ B. Misra, 2 and M. Courbage 2

Received August 27, 1979

We discuss the problem of nonunitary equivalence, via positivity-preserving

similarity transformations, between the unitary groups associated with determin-
istic dynamical evolution and semigroups associated with stochastic processes.
Dynamical systems admitting such nonunitary equivalence with stochastic Mar-
kov processes are said to be intrinsically random. In a previous work, it was
found that the so-called Bernoulli systems (discrete time) are intrinsically ran-
dom in this sense. This result is extended here by showing that a more general
class of dynamical systems---the so-called K systems and K flows---are intrinsi-
cally random. The connection of intrinsic randomness with local instability of
motion is briefly discussed. We also show that Markov processes associated
through nonunitary equivalence to nonisomorphic K flows are necessarily non-
isomorphic.

KEY WORDS: Dynamical systems; Markov processes; K flows; H theorem;

time operator; irreversibility; instability.

1. INTRODUCTION
The study of the possible connections that may exist between deterministic
dynamics and probabilistic processes is of obvious importance for the
foundation of nonequilibrium statistical mechanics. As is well known,
stochastic Markov processes provide the best possible models to represent
irreversible evolution admitting a Lyapounov functional or % function. The
important question, thus, is how is the passage from deterministic dynamics
to probabilistic Markov processes to be achieved?
It is generally believed that probabilistic processes can come from

Dr. Goldstein's research was supported in part by NSF Grant No. PHY78-03816.
I Department of Mathematics, Rutgers University, New Brunswick, New Jersey.
2Facult6 des Sciences, Universit6 Libre de Bruxelles, Campus Plaine, C.P. 231, 1050 Brussels,
Belgium.
111
0022-4715/81/0500-0111$03.00/0 9 1981 Plenum Publishing Corporation
112 Goldsteln, Mlsra, and Courbage

deterministic dynamics only as a result of some form of "coarse graining"

and approximations which necessarily involve a loss of information. In this
connection let us recall that the generally adopted procedure for obtaining,
say, a Master equation from dynamics involves an initial "contraction of
description" brought about by a projection operator followed by suitable
limit procedure designed to restore the Markovian character of the evolu-
tion which is generally lost in the process of "contraction of description."
Recently a totally different possibility for obtaining stochastic Markov-
ian evolution from deterministic dynamics has been proposed (see Ref. 1
and the earlier discussion in Ref. 2). The basic idea of this approach is to
obtain a stochastic Markov process from a deterministic dynamics simply
through a "change of representation" brought about by a (necessarily)
nonunitary but invertible similarity transformation. More explicitly, let Ut
denote the unitary group (acting on the space of distribution functions on
the phase space) which is induced from the dynamical group T t describing
the (deterministic) motion of the phase points. One considers now the
possibility of defining an invertible similarity transformation A mapping
states (i.e., positive and normalized distribution functions) to states such
that AUrA - 1 - - W* represents the semigroups of a stochastic Markov
process. The existence of such a A means that under the "change of
representation" p ~ A p the originally given deterministic dynamics de-
scribed by the unitary group Ut is transformed into the stochastic Markov-
ian dynamics associated with Wt*. The important point to note is that this
scheme for the passage from deterministic to stochastic dynamics involves
no approximation and (as the required invertibility of A assures) no "con-
traction of description" or coarse graining.
For this reason dynamical systems admitting the existence of a similar-
ity transformation A with the above-stated properties have been termed
intrinsically random dynamical systems in Ref. 1.
How can it be possible that a "change of representation" can lead
from deterministic to stochastic evolution? Should not a change of repre-
sentation in classical mechanics be given simply by a canonical transforma-
tion of the phase space into itself? As suggested in ReL 1 the answer to this
is to be found in the concept of instability of motion. If the dynamical
motion is sufficiently unstable so that each open region of the phase space,
no matter how small, rapidly spreads out in time to far-separated regions in
phase space then, obviously, one cannot arrive at the concept of the phase
space trajectories by considering the motion of smaller and smaller initial
regions of the phase space. Now, a physical measurement is necessarily of
finite accuracy and hence can determine the initial conditions of the system
to lie in only an open region of the phase space, which may be taken to be
arbitrarily small but which can never be reduced to an individual phase
On Intrinsic Randomness of Dynamical Systems 113

point. This limitation on the possible physical measurements is of no

consequence for stable dynamical evolution as one can arrive at the
concept of phase space trajectory as the idealized limit of the motion of
smaller and smaller phase space regions. But as we have seen, this is no
longer possible in the case of sufficiently unstable dynamical motion. Thus
for such systems the very concept of deterministic motion along phase
space trajectories becomes an idealization which is beyond the possibility of
physical realization even as a limiting case. As a consequence the fundamen-
tal objects of dynamical theory must not, now, be the individual phase
points or trajectories but open regions or more generally, the distribution
functions on the phase space: A change of representation of dynamics may
therefore correspond now to more general transformations acting on the
distribution function than those given by the canonical transformations of
the phase space into itself.
According to the viewpoint expressed above the appearance of strong
forms of instability of motion marks the breakdown of the deterministic
description of dynamics given in terms of the phase space trajectories. In
this situation it is both possible and natural to pass to a new representation
of dynamics (provided by necessarily "nonlocal" transformations on the
phase space) which eliminates from the theory the physically unrealizable
idealization of determiministic motion along phase space trajectories and
results in a stochastic Markov process. The notion of intrinsic randomness
formalized this idea.
The consistency of this conception of the close link between intrinsic
randomness and instability is supported by the result that the instability
leading to mixing on phase space (in the sense of ergodic theory) is a
necessary condition for the system to be intrinsically random. (1) As regards
sufficiency, it was shown in Ref. 1 that the so-called Bernoulli systems
(discrete time) are intrinsically random. (Actually only the simplest system
of the Bernoulli systems--the so-called baker's transformation--is explicitly
considered in Ref. 1. But as emphasized there it is not difficult to see that
the same considerations can be generalized to an arbitrary Bernoulli
system.(16~) The question as to whether Bernoulli flow (continuous time)
and more generally the K systems and K flows are also intrinsically
random, however, remained unsettled. The main purpose of this paper is to
answer this question in the affirmative.
Let us mention that a great variety of systems of physical interest--
infinite ideal gas and hard rods system,(3'4) motion of a billiard ball on a
table with convex obstacle, (5'6) a hard sphere gas, (7) and geodesic flow on a
manifold of negative curvature, (8'9) etc.--are known to be K flows and even
Bernoulli. In view of the intrinsic randomness of K flows, one can associate
with these systems stochastic Markov processes that are obtained from the
114 Goldstein, Misra, and Courbage

underlying deterministic motion through nonunitary but invertible similarity

transformations. It would be interesting to examine the physical meaning of
these Markov processes and discuss questions such as the conditions under
which they are diffusion processes.

2. INTRINSIC RANDOMNESS, OPERATOR TIME, AND INSTABILITY

OF MOTION
In this section we recall the precise definition of intrinsic randomness
and the "canonical" procedure for constructing the transformation A which
leads from deterministic to stochastic dynamics. For more details see Refs.
1 and 10.
Let us consider an abstract dynamical system {F, ~ , #, Tt}. Here F
denotes the phase space of the system equipped with the o-algebra ~ of
subsets of F, Tt a group of measurable motions mapping F onto itself and
preserving the measure/z. For example, F could be the energy surface of a
classical dynamical system, Tt the group of dynamical evolution, and/~ the
invariant measure whose existence is assured by Liouville's theorem. For
convenience, we shall assume the measure/~ to be normalized:/~(F) = 1. As
is well known, the evolution of density functions O in L~ under the given
deterministic dynamics is described by the unitary group Ut induced by Tt:
(Utp)(~o) = p ( T ,co), ~0 E F
Every deterministic evolution thus defines a unitary group. Conversely
(under the mild assumption that (F, ~ , #) is a standard measure space)
every unitary group which preserves positivity (i.e., maps nonnegative
functions to nonnegative functions and leaves the constant functions un-
changed) is induced by a group Tt of measure-preserving transformation on
F (see, for example, Ref. 11).
On the other hand, stochastic Markov processes on the state space F,
preserving/z, are associated with contraction semigroups of L~.(12) In fact,
let P(t, ~o,A) denote the probability of transition from the point ~o E F to the
region A E ~ in time t. Then the operators Wt defined by
(Wtf)(~0) = f f(,o')e(t,,o,d,o'), f E L~
form a semigroup for t/> 0 having the following properties:
(i) Wt preserves positivity (i.e., f(w) >10 implies W,f >10 for t I> 0);
(ii) Wtl = 1.
The evolution of the distribution functions ~ under the Markov process
is described now by the adjoint semigroup W* which also preserves
positivity since W t does: tS-~ Ot = Wt*tS. Since the measure/z is an invariant
On Intrinsic Randomness of Dynamical Systems 115

measure for the process (or equivalently the microcanonical distribution

function 1 is the equilibrium state of the process) we also have the follow-
ing:
(iii) W*I = 1.
Every Markov process on F with stationary distribution /z is thus
associated with a contraction semigroup satisfying the conditions (i)-(iii).
Conversely every contraction semigroup Wt on L~ satisfying the above
conditions comes from a stochastic Markov process, the transition proba-
bilities P(t, o~, A) being given by
e ( t, a) = (
Here rpa denotes the characteristic (or indicator) function of the set A.
In the following we are interested in a special class of Markov
processes whose semigroups W t satisfy [in addition to the conditions
(i)-(iii)] the following condition:
(iv) IIWT(p - 1)ll decreases strictly monotonically to 0 as t ~ oo [f(t) is
strictly monotonically decreasing if t < s ~ f ( t ) > f(s)] for all states p ~ 1
(i.e., for all nonnegative distribution functions p ~ 1 with frpdt~ = 1). This
condition expresses the requirement that any initial distribution p tends
strictly monotonically in time to the equilibrium distribution 1. For such
processes the functional

I1~,112= frPt2d/x, (Pt = Wt*P)

and indeed any other convex functional including the usual expression for
negative entropy
~ptlnptdt t

is an ~ function. Such Markov processes thus provide the best possible

model of irreversible evolution obeying the law of monotonic increase of
entropy.
Semigroups satisfying the conditions (i)-(iv) above have been called
strong Markov semigroups in Ref. 1. The term strong Markov process is,
however, generally used in probability theory in an entirely different sense.
To avoid this confusion, we shall henceforward refer to semigroups satisfy-
ing conditions (i)-(iv) as monotonic Markov semigroups.
The intrinsic randomness of dynamical systems may now be defined in
terms of the existence of a "change of representation" under which the
unitary group associated with the deterministic dynamics is transformed
into a monotonic Markov semigroup. More explicitly, the deterministic
dynamics with induced unitary group Ut on L~ is said to be intrinsically
random if there exists a (bounded) linear operator A defined on L~ such
116 Goldsteln, Mlsra, and Courbage

that
(a) A preserves positivity: Ap > 0 if p > O;
(b) frodg = frAodg;
(c) A 1 = 1;
(d) A has a densely defined inverse A-l;
(e) AUtA- 1 VI/', (for t > 0) is a contraction semigroup satisfying the
____.

conditions (i)-(iv) formulated above.

The conditions (a) and Co) are necessary conditions for p--> Ap =- t~ to
be interpreted as a change of representation. The condition (d) assures that
the transformation does not involve "contraction of description" or coarse
graining; the latter being given by projection operators. The condition (e)
expresses the fact that under the change of representation p ~ A o the
deterministic evolution p ~ Utp is transformed into a stochastic Markovian
evolution: Ap-~ A Utp = W*Ap described by IV,*.
It may be noted that we have not required that A -1 should also
preserve positivity. Indeed the following proposition shows that if one
requires this then one is necessarily confined to the change of representa-
tions provided by (canonical) point transformations of the phase space onto
itself which cannot, obviously, lead from deterministic dynamics to stochas-
tic processes.
Proposition 1 [cf. Refs. 11 and 15]. Let A, in addition to satisfying
conditions (a), (b), and (c) given above, also satisfy the following: (a') A-1
preserves positivity, and (b') the domain of A - l contains all the characteris-
tic functions ~0a (A ~ ~). Then, there exists a measure-preserving point
transformation T of F onto itself such that
(Ap)(~) = p(T~)
Thus for intrinsically random systems the reverse passage ~-~A-lfi,
which leads from the stochastic description of the time evolution provided
by Wt* to the deterministic dynamics Ut, cannot physically be interpreted
as a change of representation. One might say that the "reverse operation"
leading to the deterministic description, though mathematically well de-
fined, is a physically inadmissible operation. This is in conformity with the
view that intrinsic randomness is a property only of highly unstable
dynamical motion and in this situation the deterministic description of
dynamic development constitutes an unphysical idealization. The above
proposition shows also that A U,A-l = W*, although defined for all t, can
be required to be positivity preserving only either for positive t or for
negative t but not for both at the same time.
Thus the passage to the stochastic description as contemplated in
defining intrinsic randomness naturally breaks the symmetry between the
On Intrinsic Randomness of Dynamical Systems 117

"forward" and "backward" directions of time, causing the temporal evolu-

tion to be described by a semigroup rather than a group.
Let us now turn to the method of constructing the A transformation. A
first remark to this end is that if AUtA- 1 is a monotonic Markov semigroup
then the operator
M =--A * A
is necessarily a Lyapounov variable for Ur In this connection, by a
Lyapounov variable for the group Ut one means a positive operator M on L~z
such that the functional
~(Pt) ~ ( Utp, MUtp)
decreases strictly monotonically with t to its minimum value f~(1) for all
initial states p ~ l, the single exception being the microcanonical ensemble
for which the functional f~ is, obviously, constant (cf. Ref. 10).
Intrinsic randomness thus implies the existence of Lyapounov vari-
ables which expresses the inherent irreversibility of the dynamical evolu-
tion. The existence of Lyapounov variables, on the other hand, is known (1~
to imply that the system must be at least mixing in the sense of ergodic
theory. In this way, one sees once again the close connection between
intrinsic randomness and instability of motion: Instability leading to mix-
ing on the phase space is a necessary condition for the manifestation of
intrinsic randomness.
The foregoing remark shows that the A transformation (when it exists)
may be constructed as essentially the "square root" of a Lyapounov
variable. Now as described in Ref. 10, there is a general procedure for
constructing the Lyapounov variables of a class of (abstract) dynamical
systems, the so-called K flows. Briefly, for such systems one can construct
an operator of "internal time" T which is defined to be a self-adjoint
operator on % x satisfying the relation
U~'TUt = T + tI (2.1)
(Here %-o~ denotes the one-dimensional subspace spanned by the constant
functions and %_x~ denotes the orthogonal complement of % o~.)
An operator M of the form
M = f(T) + P_~ (2.2)
is then easily verified to be a Lyapounov variable if f ( T ) is taken to be an
operator function of T (in the sense of functional calculus) corresponding
to a bounded positive function f decreasing strictly monotonically to 0, and
P_ ~r denotes the projection onto %_ oo-
The operator A defined to be the positive square root of an operator M
118 Goldsleln, Mlsra, and Courbage

of the form (2.2) is, obviously, again of the same form:

A = h(T) + e_~ (2.2 3
where h = f l / 2 is again a positive function decreasing strictly monotonically
to 0. A possible strategy for establishing the intrinsic randomness of K
flows would, therefore, be to show that for suitable choice of operator time
T and of a monotonically decreasing positive function h the operator given
by (2.2') as well as the semigroup W* = AUrA -I (for t >1 0) preserve
positivity. We do this in the following section.

3. NONUNITARY EQUIVALENCE BETWEEN THE UNITARY GROUP

INDUCED BY K FLOWS AND MARKOV SEMIGROUPS
Let (I', ~ , ~, Tt) be an abstract dynamical system. It is called a K flow
(cf. Ref. 13) if there exists a o-subalgebra 80 C_~ with the following
properties:
(i) TtCs = 6-3t C T , ~ o = ~,, t < s
(ii) The o-algebra generated by all 6fx (with - o o < X < + oo) coin-
cides with the entire o-algebra of measurable sets of the system:
+oo
v ~x --- 63
~ = --oo

(iii) N~oo~x = ~ - ~ is the trivial o-algebra which consists only of sets

of measure g or complements of such sets.
Another characterization of K flows is that they have completely
positive Kolmogorov entropy. (13) This latter characterization shows that the
observed behavior of K flows contains an essential element of randomness.
In fact, the complete positivity of Kolmogorov entropy may be interpreted
to mean that the knowledge about the evolution of the system in the past
obtained from an infinite repetition of any realistic measurement that
corresponds to a partition of the phase space into a finite number of
disjoint cells is insufficient for predicting the result one would obtain if the
s a m e measurement is performed in the future. Intuitively, then, the ob-
served dynamical evolution of K systems is inherently nondeterministic in
character. In the following we show that K flows are also intrinsically
random in the stronger sense that the originally given deterministic evolu-
tion is equivalent through a suitable change of representation which
involves no coarse graining or contraction of description--to the stochastic
evolution of a Markov process.
Denote by Px the orthogonal projection operator onto L2(~'x, iz), the
subspace of functions in L~ that are measurable with respect to ~x. The
properties (i)-(iii) of ~x then translate, respectively, into the following
On Intrinsic Randomness of Dynamical Systems 119

properties of Px:
(i') Px < Px' if h < ~';
(ii') limx_~ooPx = I;
(iii') limx.~_~P x = P-oo is the projection on the subspace spanned by
the constant functions on F. Moreover, the obvious facts that
(iv) T , ~ = ~-~+, and
(v) T,~-_ o~ = ~
imply, respectively, the following:
(iv') U7 e~ U, = P~_,
(v,) t:*e
Now let F x stand for the projections Px - P - ~ (X real). The family F x
is then a spectral family of projections (or resolution of identity) in %-~o0.
Moreover, owing to properties (iv') and (v'), the spectral family F x is a
system of imprimitivity for the group Ut:
U*F~U, = Fx_ , (3.1)
The operator of "internal time" T mentioned in the preceding section may
now be defined as the self-adjoint operator that has F x as its spectral family
(cf. Ref. 10):
T = fTt dEx (3.2)

In fact, if T is given by (3.2) then the imprimitivity condition (3.1) on F x

and the defining condition (2.1) on T are mutually equivalent.
As is well known, the functional calculus of the self-adjoint operator T
associates with the function h(k) of a real variable an operator function
h (T) of T which is given by
h (T) = f h (h) aFx (3.3)

To establish the intrinsic randomness of K flows let us now study the

positivity-preserving property of operator A of the form
A = h(T) + P-oo (2.2')
as well as of the semigroup A UtA-1 = Wt, (for t ~ 0).

Theorem 1. Let h(h) be a positive and monotonically decreasing

function with h ( - ~ ) < 1. Then the operator A of the form (2.2') is
positivity preserving: f / > 0 a.e. implies Af/> 0 a.e.
To prove this we start with the following lemma.

Lemma 2. The projection operator Px onto L2(~x, /z) is positivity

preserving.
120 Goldsteln, Mlsra, and Courbage

Proof of Lemma 2. Let f > 0 a.e. Define now a measure/,f on the o

algebra fx by
/zf(A) = fa f dt,
for A E fx. It is clear that /,f is a nonnegative measure on fx, and it is
absolutely continuous with respect to the measure/z restricted to fx [i.e.,
if(A) = 0 for A ~ fx implies/zf(A) -- 0].
Thus by Radon-Nykodym theorem, there exists an essentially unique
function, say, X having the following properties:
(i) fis measurable with respect to the a-algebra fx;
(ii) f~fd~--/.tf(A), A E fix. [The essential uniqueness of f means
that if f ' is another function satisfying (i) and (ii) then f = f ' a.e. with
respect to/,.]
This function f is called the Radon-Nykodym derivative of /zf with
respect to #; since both/z and/zf are nolmegative measures f > 0 a.e. I f is,
by definition, the conditional expectation E ( f I fx) of f given fx.]
Now let us consider the function Pff; since it belongs to L2(fx,/z) it is
obviously measurable with respect to fx. Moreover,
f~(P~)(o 0 d/~= (P~0(,WA) = (f, PxWA)

But if A E fx then P~CFA= ~Va SO that faPxfd# = (f, cp~)=/tf(A) for all
A ~ fx. Thus owing to the essential uniqueness of Radon-Nykodym deriv-
ative,
Pxf=/>O a.e. 9
Remark. The preceding argument is the standard recipe in probabil-
ity theory for defining conditional expectations E ( f I fx) and identifying Pff
with E ( f Ifa).
Proof of Theorem 1. It will suffice to show that (p, Ap') > 0 for
every pair of nonnegative functions p, p' in L2. Now, owing to the assumed
form of A, we obtain
<0, Ap'> = fh(X) d<,, F # > + <p, P_
Partial integration of the first term on the fight then gives
<p, A0'> = h(oo)(p, Foop') - f ( p , Fxp')dh(X) + (P,P-ooP')
Use of the definition F x = P x - P - ~ and rearrangement of terms finally
yields
(p, ap') = - f (p, Pxp')dh(a) + h(oo)(p,p') + (p,P_~op')(1 - h(-oo))
(3.4)
On Intrinsic Randomness of Dynamical Systems 121

For nonnegative O and p' all the terms in the right of (3.4) are individually
nonnegative. This is so for the first term because (o, PAP') >>"0 (owing to the
positivity-preserving property of Pa and the fact that h(h) is chosen to be
monotonically decreasing). Nonnegativity of the second and third terms is
obvious. 9
The preceding argument actually proves the slightly stronger result
that if h(h) is strictly monotonically decreasing then A is "positivity
improving," i.e., for any nonnegative function O'( ~ 0) and any measurable
set A with/t(A) > 0,
( ~ , Ap') = l a P ' dr > 0
In fact, it follows from (3.4) with ~p~ replacing O that
( ~ , At)') >>.- f (cpa, PAP') dh (~)
Since h(h) is strictly monotonically decreasing and (r PAP') >10 the
integral on the right must be strictly positive unless (r PAP') = 0 for at
least some ~ in every interval.
This latter possibility is, however, ruled out since
lim (r PAP') = (~Pa,P - ~P')
.),--~- oo
>o

This result shows that the action of A is highly "delocalizing" in the

following sense: Even if the support of p' is confined to an arbitrarily small
volume of the phase space F the support of the transformed function covers
the entire phase space. Obviously such a transformation cannot be the
multiplication operator by a function on the phase space, not can it be the
induced operator from an underlying point transformation of the phase
space. It is because of such delocalizing action of A that one may expect
the change of representation P ~ AP to lead from deterministic to stochastic
dynamics.
Let us note that the requirements that A1 = 1 and (since A is self-
adjoint) fo dl~ = fAo dl~ are an easy consequence of the construction of A.
The existence of a densely defined inverse A-1 can also be easily assured
by choosing h(~) to be strictly positive a.e. Thus to prove the intrinsic
randomness of K flows it remains only to show that for a suitable choice of
the function h(h) the operators W,* = A UtA-1 (for t/> 0) form a monotonic
Markov semigroup (see Section 2 for definition). This part of this argument
is essentially the same as the corresponding considerations in Ref. 1. But we
supply the details for the sake of completeness.
Now, A UtA-1 _ W7 are positivity-preserving for t/> 0 if and only if
the operators Ut*A UtA-l (for t/> 0) preserve positivity. Using the imprimi-
122 Goldstein, Misra, and Courbage

tivity condition (3.1) on F x one sees that

Ut*AUt = f h(X)d(U*FxUt) + e _ ~

=fh(X + t)dFx + e-oo

Since, on the other hand,

A-a = f h-~) dF~ + P-~o

the operators U*A UtA-1 are of the form
=,. hQ, + t)
U.AUrA -1 h(X) dFx + P-oo
J
According to Theorem 1, the operators U*A UtA-l (and hence W*) pre-
serve positivity for t/> 0 if h(X) is so chosen that the functions
fit(h) = h(X + t)/h(X)
are monotonically decreasing functions of X for all t > 0. Finally, the
requirement that for all states p S I (p>>-O,fpd~= 1) [IWT(p-1)11--->0
strictly monotonically with t can be assured by requiring h(X) to decrease
strictly monoton&ally to 0 as X~ oo.
In fact, denoting p - 1 by p', we find
IIW*(p - 1)112 = (AUtA-lp',AUtA-'p ')
= ( A - lp,, U, A2UtA - ]p,)

= f h=(X)a<A-'o', U*FaUtA-'o') + (A-lp',e_ooA-tp ' )

= fh2(X + t) a < A - 'o', FxA-1O') + (A- lp,, p_ ~A- 10t)

Now, the term (A-lp',P_~A-Ip') vanishes identically since fpdlx= 1
implies that P - 1 = P' (and hence A-Ip ') belongs to % ~ . The first term
decreases strictly monotonically to 0 as t o + m because the integrand
h2(X + t) is chosen to decrease strictly monotonically to 0 as t increases to

Summarizing the proceeding considerations we obtain the following.

Theorem 3. Let the abstract dynamical system (F, ~,/~, T,} be a K
flow, Ut the unitary group induced by Tt. Further, let F x and P_ ~o be the
projections defined in the beginning of this section and h(?0 a function
satisfying the following conditions:
(i) h(X) is strictly monotonically decreasing with limx_~h(X ) = 0 and
h ( - oo) < 1.
On Intrinsic Randomness of Dynamical Systems 123

(ii) h(~ + s)/h(~) -= h](~) is a monotonically decreasing function of

for every s/> 0. Suppose A is defined by

A =f_+ffh(XldF~+ e_~
Then A is positivity improving, has a densely defined inverse A-1, and the
operators AUtA- l __ Wt. form a monotonic Markov semigroup for t/> 0.
The intrinsic randomness of K flows is therefore established once the
class of functions h(~) satisfying the conditions (i) and (ii) is shown to be
nonempty.
Now condition (ii) means that [lnh(~ + s ) - lnh(~)] is decreasing for
s/> 0. In other terms, condition (ii) means that h(~) is logarithmically
concave, i.e., h(~) is of the form
h(~) = e -•(x)
with 4~(~) convex. Thus conditions (i) and (ii) can be satisfied by choosing
q~(~) to be a convex, positive function which increases to + oo as ~ ~ oo.
For example, h(~) = e -(e~) satisfies conditions (i) and (ii). Finally it may be
remarked that Theorem 3 and its proof remain valid also, essentially as
stated, for K systems (discrete time).

4. ISOMORPHISMS OF MARKOV PROCESSES ASSOCIATED WITH

K FLOWS
We have seen in the preceding sections how stochastic Markov pro-
cesses can be obtained from K flows through nonunitary similarity transfor-
mations A. We show now that this procedure associates nonisomorphic
Markov processes with nonisomorphic K flows. In this connection, two
Markov processes with transition probabilities Pl(t,w,A) and P2(t,w',h')
[on state spaces ( F ] , ~ l , / q ) and (F2,~2, ~2), respectively] are said to be
isomorphic if there exists a one to one map S from all of F l (except possibly
a set of measure 0) onto the entire space F 2 (modulo again a set of measure
0) such that (i) both S and S - ] are measurable and measure-preserving
and (ii) Pl(t, w, 4) = P2(t, Sw, SA).
Let V: L~ ~ L~z be the unitary transformation induced by S:

(Vp)(~) = 0(S -'0~)

for p E L~Z, w ~ F 2.
It is easy to verify that the condition (ii) on transition probabilities Pi
of isomorphic processes is equivalent to the condition VW,O)V* -- Wt(2)
where W,(i) (i = 1, 2) is the semigroup associated with the Markov process
having transition probability Pi(t,., .) (cf. Section 2).
124 Goldsteln, Mlsra, and Courbage

L e t Ut (1) and Ut(z) be, respectively, the unitary groups induced from K
flows TtO)and Tt(2). We suppose that the Markov processes A1Ut(OAi-~
= Wt(1)* and AEUt(2)Af l = W(2)* obtained from the two K flows in
question through A transformations discussed before are isomorphic. We
show now that if this holds then the two K flows in question must
themselves be isomorphic. Now, as explained in the preceding paragraph,
the assumed isomorphism between the Markov processes means that
V W 0)* V -1 = Wt(2)* (4.1)
where V is the unitary operator induced by a measure-preserving transfor-
mation S from the phase space of the flow Tt0) to that of T,(2). Using the
fact that A,. (i = 1,2) are functions of "operator time" of the respective K
flows we easily find that
w(i)* = AiU,(OAT'= U,(')[ ut(i)*AiUt'i)A7 -' ]

Here Ai may be easily computed to be

( h,(X+ t)
hi(X) dF(i)+ P_=

with F (i) denoting the spectral family of the "operator time" of K flow U(i)
and

A i = fh,()t) dF(ai) + P
Thus ~ki is a positive and self-adjoint operator and Wt(0. = Ut(~ with U,(0
unitary and Ai >t 0 represents the "'polar decomposition" of the operator
Wt(0.. The important point is that the unitary part in the polar decomposi-
tion of l,Vt(0. coincides with the group Ut(i), this result being a consequence
of the fact that Ai are functions of operator time. Condition (4.1) may be
rewritten now as
(vu,(')v*)(vLv*)= w,(2)*

Since VUt(I)v* is unitary and V A 1V* >10 both sides of the above equation
provide a polar decomposition of Wt(2)*. Uniqueness of polar decomposi-
tion then entails
VUt(l)v * = Ut (2)

which means that the transformation S from which V is induced satisfies

STt(1)S - 1 = Tt(2)
On Intrinsic Randomness of Dynamical Systems 125

In other terms, the two K flows are isomorphic. We have thus established
the following proposition.
Proposition 1.3 Markov processes obtained, through nonunitary A
transformation described in Section 3, from nonisomorphic K flows, are
necessarily nonisomorphic.
The preceding discussion shows also that the symmetry group of the
Markov process IV* = A U t A - 1 associated with the K flow T t is necessarily
a subgroup of the symmetry group of the K flow in question.
In fact, a symmetry operation of the Markov process would corre-
spond to a (positivity-preserving) unitary operator V such that VI,Vt*V-1
= Wt*. But this implies, as before, that VUt V-1 = Ut ' which means that V
is a symmetry operation for the K flow too. One thus finds that the passage
to the stochastic description through nonunitarity transformation as de-
scribed in Ref. 1 and here is expected to lead also to symmetry breaking.
We have already remarked that this passage necessarily involves breaking
the symmetry between positive and negative directions of time. It would be
interesting to examine whether additional symmetries, such as space inver-
sion symmetry, can be broken by this procedure. (16)

ACKNOWLEDGMENTS
It is a pleasure to thank Professor Ilya Prigogine for his encouraging
interest in this work and for discussions on the physical applications of
nonunitary equivalence. We also acknowledge support from Instituts In-
ternationaux de Physique et de Chimie fondrs par E. Solvay and Actions
de Recherche Concertres of the Belgian Government.

REFERENCES
1. B. Misra, I. Prigogine, and M. Courbage, Physica (Utrecht) 98A:1-26 (1979); see also an
earlier shorter version in Proc. Natl. AcacL Sci. USA 76:3607-3611 (1979).
2. I. Prigogine, C. George, F. Hennin, and L. Rosenfeld, Chem. Scr. 4:5-32 (1973).
3. Y.G. Sinai, Funct. Anal. Apgl. 6:35 (1972).
4. M. Aizenman, S. Goldstein, and J. L. Lebowitz, Commun. Math. Phys. 39:289-301
(1975).
5. Y.G. Sinai, Usp. Mat. Nauk 27:137 (1972).
6. G. Gallavotti and D. Ornstein, Commun. Math. Phys. 38:83-101 (1974).
7. Y.G. Sinai, Soy. Math. Dokl. 4:1818-1822 (1963).
8. D. Anosov, Proc. Steklov Inst. No. 90 (1967).
9. D. Ornstein, Ergodic Theory, Randomness and Dynamical Systems (Yale University Press,
New Haven, Connecticut, 1974).

3In the special case of Bernoulli systems this proposition has been previously obtained by
explicity calculation; see Ref. 14.
 126 Goldsteln, Mlsra, and Courbage

10. B. Misra, Proc. Natl. Acad. Sci. USA 75:1627-1631 (1978).

11. K. Goodrich, K. Gustafson, and B. Misra, On a converse to Koopman's lemma, Physica
(UtrechO 102A:379-388 (1980).
12. E.B. Dynkin, Markov Processes (Springer, New York, 1965).
13. V. I. Arnold and A. Avez, Ergodic Problems of Classical Mechanics (Benjamin, New
York, 1968).
14. M. Courbage, C. Coutsomitros, and B. Misra, On the isomorphisms of Markov processes
associated with Bernoulli systems, submitted to Ann. Inst. Henri Poincar~.
15. B. Simon, The P(~')2 Euclidean (Quantum) Fluid Theory (Princeton University Press,
Princeton, New Jersey, 1974), proof of Theorem 1.13.
16. M. Courbage and B. Misra, On the equivalence between Bernoulli dynamical systems
and stochastic Markov processes, to appear in Physiea.

View publication stats