Subadditive Ergodic Theory

Author(s): J. F. C. Kingman
Source: The Annals of Probability , Dec., 1973, Vol. 1, No. 6 (Dec., 1973), pp. 883-899
Published by: Institute of Mathematical Statistics

Stable URL: https://www.jstor.org/stable/2959077

REFERENCES
Linked references are available on JSTOR for this article:
https://www.jstor.org/stable/2959077?seq=1&cid=pdf-
reference#references_tab_contents
You may need to log in to JSTOR to access the linked references.

JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide
range of content in a trusted digital archive. We use information technology and tools to increase productivity and
facilitate new forms of scholarship. For more information about JSTOR, please contact support@jstor.org.

Your use of the JSTOR archive indicates your acceptance of the Terms & Conditions of Use, available at
https://about.jstor.org/terms

Institute of Mathematical Statistics is collaborating with JSTOR to digitize, preserve and

extend access to The Annals of Probability

This content downloaded from

35.8.11.2 on Mon, 28 Jun 2021 14:25:04 UTC
All use subject to https://about.jstor.org/terms
The Annals of Probability
1973, Vol. 1, No. 6, 883-909

SPECIAL INVITED PAPER

SUBADDITIVE ERGODIC THEORY'

BY J. F. C. KINGMAN

University of Oxford

It is now ten years since Hammersley and Welsh discovered (or

invented) subadditive stochastic processes. Since then the theory has
developed and deepened, new fields of application have been explored, and
further challenging problems have arisen. This paper is a progress report
on the last decade.

1. Theory.

1.1. Subadditive processes. An interesting class of stochastic processes was

isolated by Hammersley and Welsh in their wide-ranging survey [4] of the prob-
lems of percolation theory. They observed that certain of these problems could
be formulated in terms of families of random variables x8t (s < t), where the
indices s and t run over the set T of nonnegative integers, which satisfy three
conditions:

S1. Whenever s < t < u,

(1.1.1) x U< x t+ xtu

S2'. The distribution of x8t depends o
S3. The expectation

(1.1.2) gt = E(xot)
exists, and satisfies

(1.1.3) gt > -At

for some constant A and all t > 1.

From S2', E(x8t) = gt-_ and taking expectations in (1.1.1) shows that gu-8
g,8 + g,-t, so that
gin+n < gm + gn (m,n 1) .
The familiar theory of subadditive functions [5]
(1.1.4) limt 0g/t = r
where

(1.1.5) r = inft?,lg/t
is finite because of (1.1.3).

Received February 19, 1973.

1 Professor Kingman's paper was presented as a Special Invited Paper at the Central Regional
Meeting of the Institute of Mathematical Statistics in Iowa City, Iowa, October 8-11, 1973.
AMS 1970 subject classifications. Primary 60F15; Secondary 28A65, 60G10.
Key words and phrases. Subadditive processes, ergodic theory, percolation processes, products
of random matrices, random permutations.
883

This content downloaded from

35.8.11.2 on Mon, 28 Jun 2021 14:25:04 UTC
All use subject to https://about.jstor.org/terms
 884 J. F. C. KINGMAN

Although Hammersley and Welsh were able to deduce some consequences

from Sp S2' and S3, these were fragmentary and incomplete, partly b
stationarity requirement S2' is not strong enough to bring into play the powerful
tools of ergodic theory. A more complete theory is possible if S2' is strength-
ened to

S2. The joint distributions of the process (x8+lt+l) are the same as those of
(X80.)

This is strictly stronger than S2', in the sense that it is possible to construct
processes satisfying S2' but not S2. But these examples are highly artificial, an
it seems that, for all the applications so far envisaged, there is no advantage in
using the weaker condition S2'.
A family x = (x8t; s, t e T, s < t) of random variables x8t satisfying S1, S2 and
S3 is called a subadditive process, and the constant r associated with it will be
denoted where necessary by T(x).
If the random variables of a subadditive process x are all degenerate, then S2
shows that x8t = xOt_8 = gt-8, so that (1.1.4) implies that
limt_- Xot/t = r

On the other hand, suppose that (1.1.1) is strengthened to

(1.1.6) xSU = x8t + Xtu

(in which case x is called an additive process). Then the process (yt; t > 1)
defined by

(1.1.7) Yt =Xt-t
is stationary, and

(1.1.8) Xot = Z=iYr

Hence Birkhoff's ergodic theorem shows that the limit

(1.1.9) limt_-C x~t/t

exists with probability one.
These two special cases led Hammersley and Welsh to conjecture that the
limit (1 1.9) might exist for all subadditive processes x, an assertion wh
generalise simultaneously the limit tjheorem for subadditive functions
Birkhoff ergodic theorem. Such a result was proved in [8], and will be described
in Sub-section 1.2. In Sub-sections 1.3-1.4 various ramifications and reformu-
lations are discussed, and in Sub-sections 2.1-2.4 some typical applications of
the theory are described. The paper ends with a list of unsolved problems
which stand as a challenge to future research.

1.2. The ergodic theorem. The fundamental theorem proved in [8] is a com-
plete generalization to subadditive processes of Birkhoff's theorem, when the
latter is formulated in terms of the additive process (1.1.8).

This content downloaded from

35.8.11.2 on Mon, 28 Jun 2021 14:25:04 UTC
All use subject to https://about.jstor.org/terms
 SUBADDITIVE ERGODIC THEORY 885

THEOREM 1. If x is a subadditive process, then the finite limit

(1.2.1) e = 1imt x0t/t

exists with probability one and in mean, and

(1.2.2) E(e) = r.
In general, the random variable e may be non-degenerate. If Y denotes the
a-field of events defined in terms of x and invariant under the shift x -x'
where

(1.2.3) xi = x
then e is an >-measurable random variable, with the explicit representation

(1.2.4) e = limt As t-'E(xot l A)

Thus, if X7 is trivial (as may in many particular ca
law), we have e = almost surely, so that e may be replaced by the constant
r in (1.2.1).
The proofs of these results can be found in [8], as can that of a "maximal
ergodic theorem":

(1.2.5) E(x01jx, > 0O forsome t> 1)>O.

Unfortunately, the asymmetry of (1.1.1) means that the maxima
rem does not at once imply (as it does in the additive case) the more important
Theorem 1. The proof of that theorem depends rather on a decomposition

(1.2.6) x8t = Yst + ZSt

in which y is an additive process with E(y01) -, and z is a nonnegative sub-

additive process with T(z) = 0.
It is useful to note that the condition S3 can for some purposes be weakened to:

S3'. E(x +) < oo

Note that by (1.1.1) this implies that

(1.2.7) E(x+t) < 00

for all s < t. Weakening S3 to S3' adm
is finite for all t but that

(1.2.8) gt/t > - o (t --a oo)

and secondly that

(1.2.9)

for some (and then for all sufficiently large) t.

THEOREM 2. If x satisfies S1, S2 and S3' but not S3, then the limit (1.2.1) exists
with probability one in - oo < e < oo, and

(1.2.10) E(e) = -00

This content downloaded from

35.8.11.2 on Mon, 28 Jun 2021 14:25:04 UTC
All use subject to https://about.jstor.org/terms
886 J. F. C. KINGMAN

PROOF. It is trivial to check th

X(N = max (x8t, -N(t - s))

and N is any positive integer. Hence Theorem 1 shows that the finite limit

i(N) = limt-oo X(N /t

exists for all N, with probability one. Since

X(N)lt = max (xo0lt, - N)

this means that the limit (1.2.1) exists, and that
(N) = max ($,-N) .
Hence
E(e) < E( N) = T(X(N))
- infe? t-'E(xoN))
_ t'E( O N))
for any N, t. Letting N -+o,
E(t) < t-'E(xot) = gt/t,
and letting t -* oo,
E(e) ? -oo,
which completes the proof.

1.3. Other formulations. Theorem 1 can, of course, be formulated in the

language of ergodic theory in an obvious way. Thus, without loss of generality
(cf. [1]), we may suppose that the probability space (Q. , P) on which the
subadditive process x is defined is sufficiently rich to admit a measure-preserving
transformation 0: Q -k Q such that

(1.3.1) X8+lt+l(wO) = X8,t(O)

for all w e Q. s, t e T, s < t. Then S2 is automatically sa
can be expressed by the assumption:

SE. The functions f xno,,,: Q R belong to L1(Q) and satisfy

(1.3.2) fm+((O) < fm((0)) + f.(Omw)
for all m, n > 1 and all wt e Q, and

(1.3.3) Sfn dP > -An

for some constant A and all n > 1.

Then according to Theorem 1, SE implies that

(1.3.4) ,(W) = lim.,.f.((o)/n

exists for almost all w, and that

(1.3.5) so f ( -(w)
n
- -dP-+ 0 (n oo)

This content downloaded from

35.8.11.2 on Mon, 28 Jun 2021 14:25:04 UTC
All use subject to https://about.jstor.org/terms
 SUBADDITIVE ERGODIC THEORY 887

Now let I be the set of all functions

95: {(s, t); s, t = 0, 1, 2, ;s < t} -+ R

which satisfy

(1.3.6) s0(s, u) ? (s, t) + 0(t, u)

whenever s < t < u, and make I a measurable space by means of the smallest
a-field with respect to which the coordinate maps o -> -o(s, t) are measurable.
Then, if Q, I, P, 0, fA satisfy SI; there is a measurable function f: Q ->
defined by

(1.3.7) [f(o)](s, t) = X8t((O) = ft-8(08o)

and f induces a probability measure

(1.3.8) II = Pf

on S.
If a: I S- is the shift

(1.3.9) (a )(s, t) = (0 s + 1, t + 1).

then HI is invariant under a. Moreover, for some constant A,

(1.3.10) -An < S , (0, n)ll(do) < o

for all n. Hence the theory of subadditve processes can be regarded as the theory
of probability measures on I, invariant under a, which satisfy (1.3.10). Theo-
rem 1 (whose proof in [8] depends on taking just this point of view) implies that
any such measure 11 is concentrated on the subset

(1.3.11) 0 = {in e I; limbo. n-1ts(0, n) exists} .

1.4. Continuous-parameter processes. There is no reason why the parameter
set T should always consist of the nonnegative integers. If T is any subset of
the real line which is closed under addition, we can define a subadditive process
on T to be a family x = (x8t; s, t e T, s < t) of random variables x8t satisfyin

S1. The inequality (1.1.1) holds whenever s, t, u e T, s < t < u.

S2. For all z e T, the joint distributions of x'-' = (x8+?,tT+) are the same as
those of x.
S3. For all positive t e T, the expectation (1. 1.2) exists and satisfies (1. 1.3).

Exactly as before, the finite limit

(1.4.1) 7 = Iim too t C 7r 9tt - inft I T t>Og t/t

will exist.
Apart from T -{0, 1, 2, * * }, the most interesting case is that in which T
consists of all nonnegative real numbers, and we therefore assume in this section
that T = [0, oo). In this case x is called a continuous-parameter subadditive process.

This content downloaded from

35.8.11.2 on Mon, 28 Jun 2021 14:25:04 UTC
All use subject to https://about.jstor.org/terms
 888 J. F. C. KINGMAN

If (1.1.1) is replaced by the corresponding equality (1.1.6), so that x is a

continuous-parameter additive process, we have

x8t = Xot - X0X

Then the process (xo,; t > 0) is, by S2, a process with stationary increm
Reversing this argument, we can construct large classes of additive processes.
For example, if (yt; t ? 0) is any stationary process with finite expectat
is immediate that

(1.4.2) X8t = Yt-Ys

defines a process satisfying S1 (with equality), S2 and S3. This example would
seem to rule out the possibility of any general theorems about the local behavior
of continuous-parameter subadditive processes, and interest therefore centres on
the behavior of x0t for large values of t.
It would be nice if Theorem 1 could be generalized to continuous-parameter
subadditive processes, to yield the assertion that

(1.4.3) e = limtO. x0t/t

exists for all such processes. Such an assertion would obviously require an
assumption of separability, but might perhaps be expected to be true without
any further conditions. If true, it would have as a special case (when x is addi-
tive) the corollary that, for any separable process (yt; t > 0) with stationary
increments and finite expectations, the finite limit lim, S yt/t exists with prob-
ability one. Such a theorem does not (as far as I know) appear in the literature,
for the excellent reason that the conclusion is false, and in a very strong sense.

THEOREM 3. Let F(t) (t > 0) be any positive increasing function. Then there
exists a process (yt; t > 0) with stationary increments and finite expectations, whose
sample functions have derivatives of all orders, such that

(1 .4.4) P{jytj < F(t) for all sufficiently large t} = 0

PROOF. Choose a C- function S6 on R which vanishes outside the interval
(Q, 3) and satisfies 0 < Ob(x) < 0Q(') = 1. Let (, >, > .2 be indep
random variables, Y2 having a uniform distribution on (0, 1), and the
the same distribution, which attaches mass [n(n + 1)]-' to the integer next above
F(n+ 1)forn= 1, 2, * Define

(1.4.5) Yt= [ -n)] (n < t < n + 1; n = 1. l, 2,

yt =yt+ .
Then (yt) is clearly stationary, and therefore has stationary increments, and its
sample functions are of class C-o. Moreover,

E(Jyti) = E(yo) = S OE(Y8) ds

= E L ,0 b(vos) ds
= E So"o 0(u) du = SI 0(u) du < oo

This content downloaded from

35.8.11.2 on Mon, 28 Jun 2021 14:25:04 UTC
All use subject to https://about.jstor.org/terms
 SUBADDITIVE ERGODIC THEORY 889

However,

P{fytf< r(t) for all sufficiently large t}

< P{ly(n + vj' - i)f < r(n + 2 - i) for all sufficiently large n}
< Pfvn < r(n + 2) for all sufficiently large n}
=0

by the second Borel-Cantelli lemma, since the vn are independent and

E ln
P{"= >2~D
r(n --
+ 2)}
r=i> E- I E- I~ _ = 00.
r(r?1)
Hence (1.4.4) is established, and the proof is complete.
To state a condition under which Theorem 1 can be generalized, define the
oscillation of the subadditive process x on an interval I c T as

(1.4.6) Q = SUp8<t,8,tE I Ix8t1t

It is easy to see that, if

(1.4.7) E(Q1) < oo

is true for any non-degenerate interval I, it is true for every bounded interval.

THEOREM 4. If x is a separable continuous-parameter subadditive process satisfying

(1.4.7), then the limit (1.4.3) exists with probability one and in mean, and E(e) = .

PROOF. Without loss of generality take I- [0, 1], and consider the discrete-
parameter subadditive process

xDXD =_.2
= (x8t; s, t - O 1 2, . s < * t).

Then y(XD) =r limnO n-'E(x,,) = y, and Theorem 1 shows that the limit

(1.4.8) = limn-o, n x0On

(as n increases through the integers) exists with probability one and in mean,
and that E(e) r. For any t, let n be the integer part of t, and note that

XOn+1 - Xt,n+l ? xot ?_ Xo + Xnt

so that
XOn+ l Q [nn+l1] < xot < Xon + Q[nn+l]

Then (1.4.3) will follow from (1.4.8) if we can show that

(1.4.9) limnoo nQ [nn+[1] 0

with probability one and in mean.
Using S2 and separability, the distribution of Q[nn+l] is the same as that of
Q, so that
E~n-'rn~nll)= E(n-'Q1) = --> 0
and (1.4.9) holds in mean. For any e > 0,

n=l P(n~n+13 > en) = P(Q, > en)

?s-E(Qj) < ??

This content downloaded from

35.8.11.2 on Mon, 28 Jun 2021 14:25:04 UTC
All use subject to https://about.jstor.org/terms
 890 J. F. C. KINGMAN

and the Borel-Cantelli lemma proves (1.4.9) with probability one. Hence the
proof is complete.

Comparing the proof with the counterexample of Theorem 3, we see that th

process (1.4.5), divided by t, is an instance of the phenomenon discovered in
[7]; a process with continuous sample functions whose discrete skeletons all con-
verge to zero with probability one, but which does not itself converge to zero.
The continuous-parameter ergodic theorem is usually stated (for instance in
[1]) for additive processes of the form

(1.4.10) X8t = S u(r) dT,

where u is a measurable stationary proc

process (1.4.7) is immediate, since

QI < 5 Iu(z)l dz,

so that the ergodic theorem for such processes is a corollary of Theorem 4.
Notice however that (1.4.7) does not necessarily imply continuity of xs, in s
t (since Q, is not required to be small) and the theorem applies to many d
continuous subadditive and additive processes.

2. Applications.

2.1. Percolation problems. The situation which led Hammersley and

to formulate the notion of a subadditive process was of the followin
.v be a connected graph, and suppose that with each edge e = (v, v') of the
graph there is associated a positive random variable ue = u(v, v'), with finite
mean, the ue for distinct edges e being independent. If v and v' are any two ver-
tices, and p = (v. = v, V1, V2, ..., Vk = v') is a path in ? from v to v', define
(2.1.1) Up= -E=l U(V71 Vr)
and denote by U(v, v') the infimum of Up o
properties of the random variable U(v, v') are of great importance in various
problems of percolation theory (for details of which we refer to [4]), and their
calculation is usually of great difficulty.
The graphs of interest usually have a certain regularity of structure, at least
to the extent of possessing interesting isomorphisms. By an isomorphism, we
mean a function so from the set of vertices of ?s onto itself, such that o(e) =
(D(v), sD(v')) is an edge of <( if and only if e = (v, v') is. If so has this property,
and if v. is any vertex, define a sequence (v,) of vertices recursively by
(2.1.2) vn =((Vn-1) (n = 1, 2 ...).
Then the random variables

(2.1.3) X S = U(VS Vt) (s, t - 0 1, 2, ... ; s < t

clearly satisfy (1.1.1). If the random variables are such that
distribution as ue, then
(X8+1,t+l) = (U(9(V8) p(vt)))

This content downloaded from

35.8.11.2 on Mon, 28 Jun 2021 14:25:04 UTC
All use subject to https://about.jstor.org/terms
 SUBADDITIVE ERGODIC THEORY 891

has the same distributions as (x,,), so that S1

S3 holds, since for any path p from v, to vn

0 < E(x0n) ? E(U) < Leep E(Ue)

Hence the theory of Sub-section 1 .1 is directly applicable, and we can con-
clude that the limit

(2.1.4) g = lima,, n 1U(v0, Vn)

exists with probability one and in mean. The fact that E(e)-T is of rather
little use, since there seems in general to be no way of calculating the constant
7 exactly. However, as was pointed out in a special case in [8], it is often pos-
sible to replace i by 7 in (2.1.4) by invoking a zero-one law. For example, if s0
is such that, for every edge e, and any finite set F of edges of <, the sequenc
(en) defined by
eo = e , en = (en-1)
remains ultimately outside F, then the invariant a-field <7 will be trivial, so
that P(- y) 1.
2.2. Products of random matrices. A problem attacked by a number of authors
in different contexts is the following. Let Y, Y2, * be random (k x k) matrices,
and define

(2 .2. 1 Xn - Yi ) 2 ..Yn

What can be said about the product Xn

tribution was made by Furstenberg and Kesten [2] who proved among other
results (a slight variant of) the following theorem. As will be seen from the
proof given here, the result is a simple corollary of Theorem 1.

THEOREM 5. Suppose that the elements of the matrices Yn are strictly positive,
and that their logarithms have finite expectations. Suppose also that the sequence
(r ) is stationary. Then the finite limit

(2.2.2) i limno n-1 log [Xnj1j

exists with probability one and in mean, and does not dep

(The notation [.],j stands for the (i, j)th' element of a

PROOF. If
Z81t =[Y+1 '1+2 . . . yt]1I

then, for s < t < u,

Zsu - k=J1 [ yS+ + '1[+1 y Yt2 YU]jI >~ Zt Zt.

so that
xst =-log St
satisfies S, of Sub-section 1 .1. Since (Yj) is stationary, S8 holds, and by assump-
tion x8t has finite expectation gt-,.

This content downloaded from

35.8.11.2 on Mon, 28 Jun 2021 14:25:04 UTC
All use subject to https://about.jstor.org/terms
 892 J. F. C. KINGMAN

For any matrix A, write

HAIl = maxi EX I[A]iil

for the i1-norm. Then

Eflog I I Y11l} = E{maxi log Z j [ Y1]ij}

< E{maxij log k[ Y1]ij}
< E[i j(log[Yl]ij)+} + logk < o0,
and hence
-gn= E{logzO} < E{log IIY1 Y2 ... Yell}
< 1] n= Eflog 11Yll1},
so that
inf gn/n > -E{log 11Y11} > -00
and S3 is satisfied.
Theorem 1 may therefore be applied to x to show that the limit (2.2.2) exists
when i j = 1. To deal with the other values of (i, ]), note first that

[Xnlij > [ yllilZl,n-d[ynblj

so that with probability one

lim infn,0 n-1 log [Xn]2j > lim infn_. n-1 log [ Y1]j1 + i
+ lim infn_.c n-1 log [ Y]1j

The finiteness of E{log [ Yn]1jl shows (as in the proof of Theorem 4) that the fina
term is zero, so that

(2.2.3) P{lim inf n-1 log [Xn]ij > 2 1 .

Similarly, the inequality

ZO,n+2 >- [Y1]1i[Xn']ij[ Yn+2]j1

where Xn' = Y2 Y3 * Yn+1, shows that

P{lim sup n-1 log [Xn']ij < 4 1 i

and stationarity allows us to deduce that

P{lim sup n-1 log [Xn]ij < 41 = 1

Comparing this with (2.2.3) shows that (2.2.2) holds with probability one. The
proof of convergence in mean is exactly similar, and the theorem is therefore
proved.
It is not difficult to extend the analysis to cover various cases in which some
of the elements of the matrices Yn may vanish, or (using Theorem 2) wher
some of the expectations of logarithms are infinite.
If the matrices Yn are of infinite order, the argument carries through except
that the manipulations of norms may not be available to show that 2' - 00.
But in many problems this can be proved in other ways. For example, if the
Y,, are infinite stochastic matrices, z8t < 1, so that r > 0.

This content downloaded from

35.8.11.2 on Mon, 28 Jun 2021 14:25:04 UTC
All use subject to https://about.jstor.org/terms
 SUBADDITIVE ERGODIC THEORY 893

2.3. Random products in Banach algebras. At the core of the argument of the
last section is the supermultiplicative property of the diagonal elements of posi-
tive matrices:

(2.3.1) [AB]11 > [A]11[B]11

A dual analysis can be based on inequalities expressing subadditivity of matrix
norms:

(2.3.2) IIABHI < IIA[ *IBHJ .

Such an analysis carries through in much more general contexts, as the next
theorem illustrates.

THEOREM 6. Let (Ye) be a stationary sequence taking values in a (real or complex)

Banach algebra X, and suppose that

(2.3.3) E{(log IIY11H)+} < ?

Then

(2.3.4) - lim___ nw log | |Y1 Y2 ... Y I

exists in -0oo < < oo with probability one, and

(2.3.5) E(e) = limn_ n-'E{log I [Y1 Y2 ... YnIl}

PROOF. If

Xst = log II[ Ys+ Ys+2 ... Yt l

then (2.3.2) implies that x = (xst) satisfies S1, S2 and S3'. The result then
from Theorems 1 and 2.
If the Y. were all equal to a nonrandom element y of A, then ee would be
the spectral radius of y. Thus ee could be regarded as a stochastic spectral radius
for the sequence (Y.). In particular, if the Y, are independent, and if w denote
their common probability distribution, then ee is a constant p(w) depending only
on r. Hence we have a way of defining a "spectral radius" p(w) for a probability
measure w on A
If X is the algebra of (k x k) matrices, endowed with any matrix norm, then
Theorem 6 reduces to another result of Furstenberg and Kesten [2]. In particu-

lar, if the elements of Y. are positive, the limit e can be identified with th
i of Theorem 5. This identification depends critically, however, on the finite
order of the matrices. For example, if X- is the algebra of infinite matrices
with finite B1-norm, then it will be true that i < $, but strict inequality can
occur (cf. [6]).

2.4. Random permutations. A quite different application of subadditive ergodic

theory has been made by Hammersley [3] to a problem of Ulam. Let Cn be the
group of permutations of {1, 2, * X ., n}, and define 1(Z) for w e J/n to be
length of the longest ascending sequence in w; the largest integer k for wh

This content downloaded from

35.8.11.2 on Mon, 28 Jun 2021 14:25:04 UTC
All use subject to https://about.jstor.org/terms
 894 J. F. C. KINGMAN

there exist integers i1, i2,

1 <_ i1 < i2 < . . .< ik < n, 17(1) < 17(i2) < . . .< 17(4).
It is required to discuss the distribution of 1(Z) when w is drawn at random from
a uniform distribution on Age and to do this especially when n is large.
Hammersley attacks this problem by a most ingenious device. He first con-
structs a Poisson process II of unit rate in the plane, and defines l, for 0 < s <
to be the length of the longest ascending path with vertices in II lying in the
square R8, with vertices (s, s), (s, t), (t, s), (t, t). More explicitly, 4st is the larges
integer k for which there exist points (xi, yi) (i = 1, 2, * * *, k) of II with

S < x1 < x2 < . .. < Xk _< t, 5 < Y1 < Y2 < < Yk <_ t.
Then it is immediate that

lSU >- 1st + ltu (s < t < l)+

so that -1= (-1st) satisfies (1.1.1). It is in fact easy to check that -I is a

continuous-parameter subadditive process, that it satisfies (1.4.7), and that its
invariant a-field 7 is trivial (since the shift of II along the diagonal is ergod
Hence Theorem 4 implies the existence of an absolute constant c such that

(2.4.1) limt-0f lott = c 9

with probability one.
Now let r(n) be the smallest value of t for which the square Rot contains n
points of II. By the strong law for the Poisson process,

limn-w nfz(n)2 = 1

with probability one, so that (2.4.1) implies that

(2.4.2) limn-00 nQ lo(%) = c

with probability one.
In terms of II, we can define a random element rn of 9? as follows. Let th
points of II in Ro:n) be written as (xi, yi), where O < xl < x2 < . . . < xn <_ (n
Then rn is to be the unique w e J/,9 such that

Yr () < Yr (2) < < Yzr(n)

It is clear that
1070n) = I07n),

and hence (2.4.2) shows that

(2.4. 3) limn-00 n-il(lr) =c

with probability one.
The properties of the Poisson process imply that, for each n, the distribution
of 17n is uniform over 9?. Since convergence with probability one implies
convergence in probability, and since this latter property is an attribute of the

This content downloaded from

35.8.11.2 on Mon, 28 Jun 2021 14:25:04 UTC
All use subject to https://about.jstor.org/terms
 SUBADDITIVE ERGODIC THEORY 895

marginal distributions of the in alone, it follows that (2.4.3) holds in probability

whenever a sequence (an) with wn e S< has the property that each in is uniformly
distributed.

THEOREM 7 (Hammersley). Let wn be a random permutation uniformly distrib-

uted over 9n. Then, as n o->, n- il(rn) converges in probability to the absolute
constant c.

The obvious question is: what is the value of c? Hammersley in [3] shows
that c must satisfy ir < c < e, and his arguments may be refined to give the
following bounds. (Dr. Hammersley tells me that he now has a quite different
technique which lends support to the hypothesis that c = 2.)

THEOREM 8. The constant c in (2.4.3) satisfies

(2.4.4) (8/w)' ? C ?< f,

where J3 = 6 + b-t, and 8 is the unique positive root of

log(1 +8) = 2 8
1+
Thus

(2.4.5) 1.59 < c < 2.49.

PROOF. Choose a sequence (xr,

point of H with the smallest value of x + y subject to x > 0, y > 0 and for
r > 2, (xr, yr) is the point of II with the smallest value of x + y subject to
X > Xr-1i Y > yr-l Then
O < X1 < X2 < * * < yl < y2 < ...

so that, if t(n) = max (xn, yn),

lot(n) > n

Now the differences xr- Xr1 are independent and identically distributed, with
mean
s xe- JL+y2 dx dy = (i/8)1
Hence by the strong law,

limbo, Xn/n - (i/8 I

and likewise

limno, Ynjn = (w/8)1

so that
limn-w t(n)/n = (i/8)&,
with probability one. Therefore

c = limbo lOt(,)!t(n) > limn-- n/t(n) = (8/w)2 > 1 .59 .

To prove the reverse inequality, let k be any positive integer, let w be d

This content downloaded from

35.8.11.2 on Mon, 28 Jun 2021 14:25:04 UTC
All use subject to https://about.jstor.org/terms
 896 J. F. C. KINGMAN

from a uniform distribution on 9ge and let v be the number of sequences

il < i2 < ..< < ik< n with r(il) < Wr(i2) < ... < (ik). Then

E(>) = E%1<i2< ... <k P{wf(il) < (i2) < .* .< w(ik)}

= k(n)(k !)-1.

However, if 1(w) > k, there is an ascending sequence of length 1(wr), and each
subsequence of length k is ascending, so that

2.-? (Or))
(k)

In particular, since (k) is non-decreasing,

(2.4.7) P{l(w) > rJ < (n)[k! (r)]- , (r > k) .

Now let n --> o, k -> oo, r -> oo in such

k , ant , r~ An0
where 0 < a < / < oo.
Then Stirling's formula shows easily that

P{l(w) > 9n} - 0

as n -> oo if

(2.4.8) 2a + ( a-) log -a) -a log a - /log: < .

Hence c < /3 if a can be chosen to satisfy (2.4.8). It is easy t
infimum of the values of /3 with this property is the numb
statement of the theorem, and the proof is complete.
If the random permutations wn happen to be defined on the same probability
space, the question arises whether Theorem 7 can be strengthened to yield con-
vergence with probability one. Equation (2.4.3) shows that this can be done
when the wn are constructed from the Poisson process II, but this is by no
the only way in which they might arise,
To take one example, let zn (n = 1, 2, 3, ***) be independent random variables
with a common continuous distribution (whose form does not affect the prob-
lem). Let wn be the unique permutation w e -9 such that

(2.4.9) Zr(1) < Z (2) < *. . < Zr(n) '

Then wn is obviously uniformly distributed over

ability; it is not known whether convergence takes place with probability one.
The obvious ways of defining the 7rn on a common probability space all have
the property that L(n) = 1(w) is non-decreasing, and when this is true one wa
of arguing might be as follows. Let K be any positive constant, and let r = r(n)
be the largest integer with rK < n. Then

L(re) < L(n) < L((r + 1)K),

where L is defined for nonintegral arguments by linear interpolation, so that

lim sup L(r1)r-ir < lim sup L(n)n-i < lim sup L((r + 1)K)(r + 1)-i,

This content downloaded from

35.8.11.2 on Mon, 28 Jun 2021 14:25:04 UTC
All use subject to https://about.jstor.org/terms
 SUBADDITIVE ERGODIC THEORY 897

and the same with "rim sup" replaced by "rim inf." Since

(r + 1)-K r-ir
this shows that

(2.4.10) P{lim.O. L(n)n-= c} 1

will follow from
P{limr-oo L(rr)r-il c} 1

Hence, by the Borel-Cantelli lemma, (2.4.10) will follow if we can prove that

tr=1 P{)L(r) -crill > srik} < oc

In particular, if

(2.4.11) P{IL(n) - cn'l > Eni} < An-'

for positive constants A, ( (perhaps depending on E), then almost sure conver-
gence will follow (taking K > 4).
Exactly similar arguments yield one-sided inequalities. For example, the
explicit bounds which follow from (2.4.7) by Stirling's formula show that, so

long as the sequence (wr) is defined so that 1(7w.) < l(zrn+1) then

(2.4.12) P{lim sup,,. nAl(^)) ? / < 2.49} =1

3. Problems.

3.1. The constant T. Pride of place among the unsolved problems o

tive ergodic theory must go to the calculation of the constant r (which in the
presence of a zero-one law is the same as the limit i). In none of the applica-
tions described here is there an obvious mechanism for obtaining an exact
numerical value, and indeed this usually seems to be a problem of some depth.
Of course, it may well be possible (as in the argument of Sub-section 2.4) to
derive bounds for a. For example, (1.1.5) shows that

(3.1.1) ;o <g.n/
for any n, and for small values of n it may be possible explicitly to compute
this bound. In the other direction, it should be noted that, if x is a given sub-
additive process, and if we can construct an additive process with

(3.1.2) XSt>Y
then

(3.1.3) i(x) > 1(Y) E(yj1).

The representation (1.2.6) shows that this lower bound is potentially sharp,
though it may not be possible explicitly to compute an additive process y attain-
ing equality in (3.1.3).

3.2. Rate of convergence. How fast does the convergence of xt/tl to e

This content downloaded from

35.8.11.2 on Mon, 28 Jun 2021 14:25:04 UTC:56 UTC
All use subject to https://about.jstor.org/terms
 898 J. F. C. KINGMAN

place? In the additive case it is of course possible to assert, under suitable

conditions, that
rqxot - et)
has an asymptotically normal distribution and is of order (log log t)i as t -
Presumably such conclusions hold also for subadditive processes which do not
deviate too far from additivity. No significant general results in this direction
are known, though there some first steps in [4].

3.3. Characterization problems. An interesting feature of Theorem 1 is that it

derives a result about the one-parameter process (x0t) from assumptions abou
the two-parameter process (x8,). This suggests the question: given a proce
(X,; t > 1), under what conditions does there exist a subadditive process (x8
with xt _ =ot for all t?
A similar question is: given a general process (x8t), under what conditions a
there subadditive processes x' and x" with xst = xt- x"4? Likewise, given a
one-parameter process (xt), when do there exist subadditive processes x' and x"
such that xt = x- 4?

3.4. Structure problems. If, for any i in an arbitrary index set I, we have an
additive process a' (defined on a probability space independent of i), then

(3.4.1) xst = supiecaSt

defines a subadditive process so long as E(xo
additive process, does it admit a representation (3.4.1) as a supremum of additive
processes? More concretely, if {at} consists of all the additive processes with
ast < xSt for all s < t, is (3.4.1) true? (Certainly, if xst denotes the right hand
side of (3.4.1), x is a subadditive process with x < x, T(*) = T(x).) Questio
like this are crucial to an understanding, for instance, of the non-uniqueness of
the decomposition (1.2.6) (cf. [8]).
When studying such problems from the point of view of ergodic theory, one
encounters questions like the following. With each point t of the unit interval
is associated (in a reasonably measurable way) a subset Et of [0, 1]. When doe
there exist a measure-preserving transformation T of [0, 1] with Tt C Et fo
almost all t? An obvious necessary condition is that, for any Borel subset B of
[0, 1], the measure of U {Et; t C B} should not be less than that of B. By analo
with Hall's theorem on distinct representatives, one might expect this condition
also to be sufficient; I am indebted to Professor W. Parry for pointing out that
life is more complicated than I had expected.

3.5. Independent subadditive processes. Hammersley and Welsh noted that

some significant subadditive processes have the property that, for 0 to <
tl < t2 < ...< tn, the random variables

(3.5.1) Xtr-1tr (r = 1, 2, ...,n)

are independent. For example, in the problems of Sub-sections 2.2 and 2.3,

This content downloaded from

35.8.11.2 on Mon, 28 Jun 2021 14:25:04 UTC
All use subject to https://about.jstor.org/terms
 SUBADDITIVE ERGODIC THEORY 899

this will be the case when the elements Y, are independent. An immediate
consequence is that X7 is trivial, so that $ = T. But there ought surely
more weighty consequences of such a strong condition.
Let x be a nonnegative subadditive process for which the variables (3.5. 1) are
independent. For any 0 > 0, s < t < u,

E{exp (-Ox8u)} > E{exp (-Ox8t) exp (-Ox,)}

- E{exp(-Oxat)IE{exp(-Oxtu)1.

Moreover, E{exp(-0x8t)} depends only on (t - s), whence it follow

(3.5.2) V(0) = limpet n-' log E{exp(-x0ff)}

exists. What properties has the function T, and to what extend does it enshrine
information about the process x?

These are only a few of the questions which one might ask about the theory
described in this paper. Other problems may be found in [4], and yet others
will occur to the reader.

REFERENCES

[1] DOoB, J. L. (1953). Stochastic Processes. Wiley, New York.

[2] FURSTENBERG, H. and KESTEN, H. (1960). Products of random matrices. Ann. Math. Statist.
31 457-469.
[3] HAMMERSLEY, J. M. (1972). A few seedlings of research. Proc. Sixth Berkeley Symp. Math.
Statist. Prob. Univ. of California Press.
[4] HAMMERSLEY, J. M. and WELSH, J. A. D. (1965). First-passage percolation, subadditive
processes, stochastic networks and generalized renewal theory. Bernoulli-Bayes-
Laplace Anniversary Volume, ed. J. Neyman and L. M. Le Cam. Springer, Berlin.
[5] HILLE, E. and PHILLIPS, R. S. (1957). Functional Analysis and Semi-Groups. American
Mathematical Society, Providence.
[6] KINGMAN, J. F. C. (1963 a). The exponential decay of Markov transition probabilities. Proc.
London Math. Soc. 13 337-358.
[7] KINGMAN, J. F. C. (1963 b). Ergodic properties of continuous-time Markov process and their
discrete skeletons. Proc. London Math. Soc. 13 593-604.
[8] KINGMAN, J. F. C. (1968). The ergodic theory of subadditive stochastic processes. J. Roy.
Statist. Soc. Ser. B 30 499-5 10.
MATHEMATICAL INSTITUTE
24-29 ST. GILES
OXFORD OXI 3LB
ENGLAND

This content downloaded from

35.8.11.2 on Mon, 28 Jun 2021 14:25:04 UTC
All use subject to https://about.jstor.org/terms