BOOK REVIEWS 113
BULLETIN (New Series) OF THE
AMERICAN MATHEMATICAL SOCIETY
Volume 21, Number 1, July 1989
©1989 American Mathematical Society
0273-0979/89 $1.00 + $.25 per page
Ergodic theory of random transformations, by Yuri Kifer. Progress in Prob-
ability and Statistics, vol. 10, Birkhàuser, Boston, Basel, Stuttgart, 1986,
210 pp., $34.00. ISBN 0-8176-3319-7
Traditionally ergodic theory has been the qualitative study of iterates of
an individual transformation, of a one-parameter flow of transformations
(such as that obtained from the solutions of an autonomous ordinary dif-
ferential equation), and more generally of a group of transformations of
some state space. Usually ergodic theory denotes that part of the theory
obtained by considering a measure on the state space which is invariant
or quasi-invariant under the group of transformations. However in 1945
Ulam and von Neumann pointed out the need to consider a more gen-
eral situation when one applies in turn different transformations chosen at
random from some space of transformations. Considerations along these
lines have applications in the theory of products of random matrices [2,
3], random Schrödinger operators [2], stochastic flows on manifolds [6],
and differentiable dynamical systems.
Mathematically the set up is as follows. Let M be a space, £% a a -algebra
of subsets of M and let ST be a collection of measurable transformations
of M into M. For example, if M is a topological space we could choose y
to be the space, C(M, M), of all continuous transformations of M into M,
and if M is a smooth manifold we could take ST to the space, D(M, M)
of all smooth transformations of M into M. Suppose &~ is equipped with
a (T-algebra so that the map (f
9
x) —• f(x) of J^ x M —> M is measurable.
Let m be a probability measure on ET. We want to study the action on M
of compositions of elements of £7~ chosen independently with distribution
m. So consider the direct product space Q = ^
N
equipped with the
direct product measure p = m
N
, where N denotes the natural members.
The elements of Q are sequences w = (w\, w
2
, w^,...) of members of ST.
There is a natural transformation, S: Q, —• £2, of Q called the shift map
and defined by S((w\
9
w
2
, w^,... )) = (w
2
, w^,... ). The shift preserves the
probability p (i.e. p(S~
l
A) = p(A) for every measurable subset A of Q) and
p is ergodic for S (i.e. if A is a measurable subset of Q and S~
l
A = A then
p(A) = 0 or 1). Consider the skew-product transformation T:Q x M ^
flxM defined by T(w,x) = (Sw,w\(x)) where w = (w\,w
2
,...) E Q,
and x G M. Iterating gives T
n
(w,x) = (S
n
w,w
n
o w
n
-\ o • • • o W\(x)) for
n > 1, and the second coordinate gives the action of the randomly chosen
maps on M.
This induces on M a discrete-time Markov Process with the probability,
P(x, B), of moving from the point i G M t o a point of the measurable sub-
set B c M in one unit of time given by P(x
9
B) - m({f e ^\f(x) e B}).
For some applications, such as stochastic stability of diffeomorphisms [5],
it seems more natural to consider certain Markov processes on M rather
than actions by random maps, so one should consider which Markov