VOL. 18, 1932 MA THEMA TICS: KOOPMA N AND v.

NE UMANN 255
gives an easy method of characterizing the unit point of the coordinate
system. If we impose the conditions, for example, that the coefficients of
the first, second and third degree terms in (4) shall all be unity, then the
unit point lies on Q and its projection from (0, 0, 0, 1) upon X4 = 0 lies
on a tangent of Segre. This point, the two points of intersection of the
tangent with the above cubic surface, and the point of intersection of the
tangent and the line xi = X4= 0 have a cross ratio equal to -2. The three
choices of the tangent of Segre correspond to the cube roots of unity which
appear when the above conditions on the coefficients are imposed.
1 Wilczynski, Trans. Amer. Math. Soc., 9, 79-120 (1908).
2 Green, Ibid., 20, 79-153 (1919).
3 Stouffer and Lane, Bull. Amer. Math. Soc., 34, 460 (1928).
4 Bompiani, Rend. Acc. Lincei, [6] 6, 187-190 (1927), and Math. Zeitschrift, 29,
678-683 (1929).
'Fubini e Cech, Geometria Proiettiva Differenziale, 1, 148 (1926).
6 Stouffer, Bull. Amer. Math. Soc., 34, 301 (1928).
Downloaded from [Link] by [Link] on July 31, 2025 from IP address [Link].

DYNAMICAL SYSTEMS OF CONTINUOUS SPECTRA

By B. 0. KOOPMAN AND J. v. NEUMANN
DEPARTMENTS OF MATHEMATICS, COLUMBIA UNIVERSITY AND PRINCETON UNIVERSITY
Communicated January 21, 1932
1. In a recent paper by B. 0. Koopman,I classical Hamiltonian mechan-
ics is considered in connection with certain self-adjoint and unitary opera-
tors in Hilbert space t (= 22). The corresponding canonical resolution
of the identity E (X), or "spectrum of the dynamical system," is introduced,
together with the conception of the spectrum revealing in its structure
the mechanical properties of the system.2 In general, E(X) will consist
of a discontinuous part (the "point spectrum") and of a continuous part.
The case of a pure point spectrum, and the other extreme, that in which
the inner product (E(X)f, g) is, for every f and g in ID, the Lebesgue in-
tegral of one of its derivatives, may readily be treated by known analytical
tools.3 The present paper is devoted to the case where E(X) is continuous
(X # 0), but without (E(X)f, g) being necessarily equal to the integral
of its derivative. It will further be assumed that the system is non-
integrable in the sense that any f in t such that U1f = f almost every-
where on Q must be almost everywhere constant. In other words, we
are assuming the following hypothesis:
C. E(X + 0) - E(X -0) = 0, for X # 0:
If Eo = E(+ 0) -E(- 0), then Eof = almost everywhere constant.
 256 MA THEMA TICS: KOOPMA N A ND v. NE UMA NN PROC. N. A. S.

Suppose that 4 is a characteristic function of a non-integrable system:

Upo = e' 4. Then UJ O I = 4 !, so that 4 must be almost every-
where constant, and hence we may take 4 = eio, 0 being defined almost
everywhere (mod 27r). From U14 = e$" 4 follows that Ut= 0 + Xt
(mod 27r); thus 0 is an "angle variable." Since 4 is in p, ) 11 =
fne' e-edv = ,u must be finite, so that a non-integrable system of
infinite ,uQ can have no characteristic function. These considerations
permit the following restatement of our hypothesis:
C'. The system possesses no invariant subset of Q of positive finite
measure, and in the case where ,uQ is finite, no angle variables.
We will show that under this hypothesis all the initially observed
properties of the system are obliterated by the lapse of time: the method
of elementary mechanics of computing the final from the initial state
must be replaced by the methods of the theory of probability. Contrary
to the case of classical statistical mechanics, this situation is not de-
pendent upon the system's having an enormous number of degrees of
Downloaded from [Link] by [Link] on July 31, 2025 from IP address [Link].

freedom: this number may. perfectly well reduce to two.

2. By a P-set we shall mean a set of points of Q, by a t-set, a set of
points on the time axis. The P-set for which f(P) > a, etc., shall be de-
noted as usual by [f(P) > a], etc., and similarly for t-sets and functions
of t. Along with the ,A-measure AM of a P-set M defined with respect to
the volume element dv = pdw, we shall define the T-measure 4I of a
t-set I as follows:
TI = lim m(I-[jtj. T])
T=+c 2T
Here m denotes the ordinary Lebesgue measure, and the customary set-
multiplication is referred to.5 By a zero P-set or a zero t-set we shall mean
one for which the IA-measure or the r-measure, respectively, is zero. The
"characteristic function" of a P-set of a t-set 0 shall be denoted, as usual,
by Xe: xe(P) = 1 or 0 according as P is or is not on 0, etc. In these
terms we are able to state the fundamental theorem of this paper:
THEOREM I. Under the hypothesis C, there exists a zero t-set I such that,
for any two P-sets M and N of finite ,I-measure, as t - + co (or t
- co) through values not on I,

A(Mt -N) >IAM,uN,(1)

Ml being the image of M after the lapse of time t. When ,4 = co, the
right-hand member is to be replaced by zero.
Proof.-It has been shown6 that in non-integrable systems,
 VOL. 18, 1932 MA THEMA TICS: KOOPMA N A ND v. NE UMA NN 257

(EoxM, XN) = _____;

hence (1) is equivalent to

-
lim (UtXM, XN) = (EOXM, XN),
t co
i not on I

which, in turn, may be replaced by

lim (Utf,g)= (EJ,g), (1)
t - co
t not on I

since the aggregate of x-functions and their linear combinations is every-

where dense in ,,. That is, we must prove that as t > h o through
values not on I, Utf converges "weakly" to Eof.7 Now (Utf, g) ->
(Eof, g) for all f, g if and only if (Ulf, f) o- (Eof, f) for all f.8 Using,
now, the known properties:9 U1EO = EoU1 = U1, E' = Eo, etc., we have
((Ut - Eo)f,f) = ((Ut- U1Eo)f, f) = (U (I - Eo)f,f) = (U1(I - Eo) 2f,f)
= (U1(I - Eo)f, (I - Eo)f), so that our problem is to show that
Downloaded from [Link] by [Link] on July 31, 2025 from IP address [Link].

(U(I - Eo)f, (I - Eo)f) - 0. That is, we are to show that a zero

t-set I exists such that, as t -> 4 co, t not on I, (Utg, g) > 0 when-
ever Eog = 0. Since U1 is a bounded operator, it is sufficient to show this
only for an everywhere dense sequence gi, g2, g3, ... of functions g.
We now introduce the expression
1 fT
2T (Utg, g) 2 | dt.
It has the value
rT
1 fr+T +o

2TJ.T J o e'ldx 11E(X)gj112 2dt

`
2T J-T co- dx
f/ dt e/ II E(X)g 112. I eI dA I1 E(j)g 112

cjo f o (2T f e'('")' dt) d, II E(X)g 112. 11 E(j.)g 112

+ T
f + o ei(X-ju)T -i(X-p))T
-~~
J
c

J o
e
2i(X
-~~e
- I)T
*dx I I E()J2 .d 1EAg
,+XJ+sin(XM)T. dx 1 E(X)gjj2 * d,, 1 E(,a)g 112.
Since E(X) has no point-spectrum except possibly for X = 0, for which
value precisely we have (E(+ 0) - E(- 0))g = Eog = 0, 11 E(X) 112 is a
 *~ t> co
258 MA THEMA TICS: KOOPMAN A ND v. NE UMANN PROC. N. A. S.

continuous function of X, which, as X goes from- co to + co, goes in a

non-decreasing manner from 0 to II g 112. If )X = 4(x) is the inverse of
X = 11 E(X) 112, 4 (x) increases monotonically from - o to + o as x
goes from 0 to 11 g 112, it is constant on no interval, although it may have
finite jumps. Hence the above expression may be written as
f11 fI 11jg 1I1 sin{4+(x) - +(y)}T
J0J0 {~~qS(x) -4(y)}IT dx&dy.
Since the region of integration is finite and the integrand in absolute value
less than 1, and since (except for the zero set x = y) it approaches zero,
the expression approaches zero as T -> aC. (This would not be
true if +(x) were constant on certain intervals, i.e., if 1I E(X)g 112 had dis-
continuities, and thus, E(X) had a point spectrum.)
Thus,
lim 1 T( Ug,g) 12dt =
0. (2)
Downloaded from [Link] by [Link] on July 31, 2025 from IP address [Link].

Consider the series

co 1
(t) = Ei 2nIIgX 112 W(Uggngm)j 2.
It is uniformly convergent for t < + oo, since it is dominated by the
1 II U1gfl lgj[ 11
series E non account of the inequality| (Ug gn)|2_ g|
11 g 112. Hence, in virtue of (2),
1 rT
lim I 7ir(t) dt = 0. (3)
)1> + X J-TT
Now if we can show that a zero t-set I can be found such that
lim 7r(t) = 0, (4)
I not on I

then, on account of the inequality

0 < j(UIgg, gX)12 < 2' || g 1127r(t),
we shall have the consequence that
lim (U1gn, gn) = 0,
I not onI

that is, the theorem will be proved.

Let Im = [7r(t) > -] Evidently I, C 12C I3 C ..., and, in virtue
of (3),
 VOL. 18, 1932 MA THEMA TICS: KOOPMAN AND v. NEUMANN 259

lim m(Im [it IT]) 0

o
T- > + co 2T
We now choose T1 < T2 < ... such that the left-hand member of the
1
above is, for all m = 2, 3, ..., and T > Tm.., less than -. Then we set
I = Ih[jI t < T1] + I2 [Tl < lt T2] +
+ Im.[Tm-, < It<_ Tm] +
For Tm.' < T < Tm we have
I [I t . T] CIm [it < Tmi1] + Im [itt < T],
hence,
m(Ii t T]) < m(Im.*[it I< Tm-J) + m(Im *[I tI < TD)
2T = 2T
< m(m_1 * [I t < TmI- m(Im . [I t _T])
= 2Tm-i 2T
Downloaded from [Link] by [Link] on July 31, 2025 from IP address [Link].

1 1
m-1 m
As m - + ', that is, as T -* + a', this approaches zero; hence
I is a zero t-set. Since, further, when t > Tm -1, a t on Im is on I, it
1
follows that outside I, ir(t) <-. Hence as m -> + a, or t >- =,= co
m
ir(t) - 0, so that (4) is established.
CoroUary.-The set I of Theorem I can be taken as a set of intervals
such that there are only a finite number of intervals of the set in any
arbitrarily chosen finite interval.
Proof.-Since the expressions (U1g, g) are continuous function of t,
7r(t) is likewise, so that every I, is a closed set, and, likewise, I. Hence
the Lebesgue and that Jordan measure of I are equal. In each of the
regions ti < 1, 11 < 2, ...,m-1 < Itt < m, ...wecanreplaceIby
a finite set of overlapping intervals without increasing its measure by
more than, for example, 2' 4' ' 2, ... ,-thus, in all, by a quantity < 1,
so that it remains a zero t-set.
THEOREM IL. When the hypothesis C is not realized, the conclusion
of Theorem I is invalid.
Proof.-To begin with, suppose the system to be integrable. Then there
will exist an invariant A C Q: At = A, with gA > 0 and finite, and
 260 MA THEMA TICS: KOOPMAN AND v. NE UMANN PROC. N. A. S.

,u(Q-A) > 0. If ,u is finite, take M = A, N = - A in (1), when

A (Ml * N) = M(A (. Q - A)) = M(A (2- A)) = 0, whereas
A A( -A) O. If a), take M =N =A, in which
uQ=
case I(Mt * N) = I(At A) = MA > 0, whereas
-
0 . Thus,
in either case, (1) is violated.
Suppose, on the other hand, that the system is non-integrable, but that
it has a characteristic function p of characteristic number X > 0: Ut1p =
e(p. In this case, as we have seen, we may take sp = eO, and ,uQ. will be
finite. In (1') takef = g = so; then (Ulf, g) = (ei'tp,o) = e>(,p, op)
= elp1g, so that (1'), which in the non-integrable case is a consequence
of (1), is impossible.
Theorem II is thus proved.
Now let us consider to what extent the presence of the exceptional t-set
I appears necessary in Theorem I. We have:
THEOREM III. There exist spectra E(X) (not necessarily belonging to a
Downloaded from [Link] by [Link] on July 31, 2025 from IP address [Link].

dynamical system'0) such that, under the hypothesis C, the conclusion (1')
of Theorem I fails to hold when I is empty.
Proof. -Take f such that Eof = 0; then it is sufficient to show that
lim (Utf, f) does not exist, or else that it is not zero. We have"
r+
(UJ,f)=J etd11E(X)f 112,
and since x = I E(X)f 112 can evidently be taken to be an arbitrarily given
continuous non-decreasing function, increasing from zero to a finite
a (= If j2) as t goes from - co to + o, its inverse X = so(x) can ob-
viously be taken to be equal to an arbitrary monotonically increasing
function, possibly with finite jumps, in the interval of definition 0 < x < a,
and going from - co to + co. Hence

(U1f f = f "(x)t dx.

Now let a = 1, and p(x) have always such a value that its hexadic de-
velopment contains only the digits 0 and 5. For instance, let ,(0) = - 00,
a.
= + o, and, for x = (a,, = O or 1, the latter infinitely often),
let VW(x) =
coa51
5n.
= For t = * 6', the fractional part of- so(x) t -

n=1 6n 2r
1 5
will thus start with the hexadic digit 0 or 5, i.e., it is > 0, 6', or 2 -
 VOL. 18, 1932 MA THEMA TICS: KOOPMA N A ND v. NE UMA NN 261
1 1 1
. 1-i.e., 2 ((x)t has a value (mod 1) > - - < 6. Hence

aJ e t dx = J cos (27r - p(x)t)dx

X cos (2r *) dx

7r 1
Cos = 2 >0

Thus X., (Ulf, f) = 0 is excluded, and our theorem is proved.

III. Theorems I, II, III, delimit a situation which could scarcely be ren-
dered more precise: the hypotheses and diverse restrictions seem all inherent
in the nature of the case. It would, of course, be extremely interesting if
a proof of Theorem I could be found without the use of the spectrum E(X),
etc., for example, along the lines of the recent paper of Birkhoff's on the
Downloaded from [Link] by [Link] on July 31, 2025 from IP address [Link].

ergodic hypothesis.'2
Theorem I expresses the fact that in a system for which C holds, con-
secutive states at two sufficiently different epochs are almost certainly
statistically independent. For /A(Mt . N) is the probability that, in the
time t, the system go from M to N-and this is = IM -AN the prod-
uct of the probabilities of the system's being in M and of its being in N.
(We say "almost certainly," to correspond with the fact that the excep-
tional zero t-set I is excluded.) Such a system can have, a la longue,
no physical properties. Indeed, the only properties which any system can
have a la longue are those which violate C, namely:
1. Invariant sub-sets: barriers that are never passed.
2. Angle variables: clocks that never change.'3
Theorem I may also be expressed by saying that the states of motion
corresponding to any set M of Q become more and more spread out into
an amorphous everywhere dense chaos. Periodic orbits, and such like,
appear only as very special possibilities of negligible probability.
We have already called attention to the essential difference between this
situation, which may exist in very simple systems, and that envisaged in
the kinetic theory of gases, in which the confused character of the motion
is an intuitively evident consequence of the large number of degrees of
freedom of the system.
'"Hamiltonian Systems and Transformations in Hilbert Space," these PROCEEDINGS,
315-318 (May, 1931); this reference will here be abbreviated to (H). Another reference
of importance for us is the paper by v. Neumann on the proof of the quasi-ergodic hy-
 262 MA THEMA TICS: KOOPMA N A ND v. NE UMA NN PROC. N. A. S.

pothesis (these PROCEEDINGS, pp. 70-82 (Jan., 1932)), which will be referred to under the
abbreviation (Q).
We shall, in the present paper, assume the notation, etc., of (H) and (Q) to be known.
2 Properties, of course, which are true "almost everywhere" on the manifold of
states of motion Q, corresponding with the nature of the functional tools. The allied
conception of properties true "almost everywhere" on the time axis with respect to the
time-average r-measure here defined is an essential notion in the present paper; it is
contained implicitly in J. v. Neumann's recent proof of the quasi-ergodic hypothesis
(these PROCEEDINGS, pp. 70-82 (Jan., 1932), and 263-266 (March, 1932)).
3 These cases are made the subject of as yet unpublished investigations by B. 0.
Koopman. In the case of a pure point spectrum, the theory of almost periodic functions
is available: (Utf, g) is an almost periodic function of t, a fact having an obvious physi-
cal interpretation when f = XM(P) and g = XN(P), characteristic functions of the
sets M and N (C Q; ,uM, MN finite). In the second case here noted, in virtue of the
well-known Fourier integral theorem, (Utf, g) - 0 as t - co, showing (on making
the above choice of f, g) that any such region M will flow, in course of time, "almost
entirely" out of a fixed region N of finite measure. This possibility obviously implies
that AuQ is infinite. There is a similarly obvious interpretation when X = 0 is a simple
unique characteristic number.
Cf. (H), p. 318.
5 It is unnecessary for us to enter upon the discussion of r-measurability or the proper-
Downloaded from [Link] by [Link] on July 31, 2025 from IP address [Link].

ties of v-measure.
6 Cf. (Q), p. 78.
1 ^T
7 The ergodic theorem (Q) states that -

Utf dt converges "strongly" to Eof,

thus this theorem tells us more than the present one from the point of view of conver-
gence, but less, from the point of view of the function, since f is replaced by its time
average.
-*
8 If we replace in (Utf,f) (Eof,f) f by 2 and by 2 and subtract, the
real part of (Utf, g) > (Eof,g) results. If we replace g by ig, the imaginary part
results. Thus the general formula holds.
9 Cf. (Q), p. 73.
10 The following question is of great interest: When will a canonical resolution of the
identity E(X) be the spectrum of a dynamical system? An (unpublished) formula ob-
tained by Koopman is that, when f, g, and the ordinary product fg are in O, then, for
a dynamical system,

E(x)fg = E(x -s)f-d.E(s)g,

which is a consequence of the fundamental equation obtained in (H):

UtF(f,g,...) = F(Utf,Utg,...),
where F is any single-valued function. Recently J. v. Neumann has shown that the
first-mentioned formula is sufficient and necessary for the Ut being generated by a
suitably chosen group of point-transformations P -* Pt (unpublished).
11 If I IE()X)f 1 2 is differentiable with respect to X (the case for a type of continuous

spectrum defined by Hellinger and Hahn) we have J I E(X) f 12e'tXdX. By a

well-known theorem on Fourier integrals this has lim equal to zero. Cf. reference 3 above.
-* I chc
 VOL. 18, 1932 MA THEMA TICS: J. v. NE UMANN 263
12 These PROCEEDINGS, pp. 650-660 (Dec., 1931).
13 Though it is very probable that the case C or C', in which our theorems hold is the
general one for dynamical systems, it is not easy to construct effective examples. (See
footnote 10.) An example, recently constructed by v. Neumann, will be published soon;
it refers to a two-dimensional flow of the following type: the flow takes place in a rec-
tangle,, oriented parallel to the X and Y axes, the upper side of which has been replaced
by suitable chosen curve Y = F(X). The flow itself is parallel to the positive Y-axis,
and each point X, F(X) has to be identified with the corresponding point X + a, 0.
(The number X + a is to be taken mod. a, where a is the breadth of the parallelogram
in the direction of the X-axis; a is a number incommensurable with a.) If F(X) and
a are suitably chosen this flow can be shown to fulfill C (and C').

PHYSICAL APPLICATIONS OF THE ERGODIC HYPOTHESIS

By J. V. NEUMANN
DEPARTMENT OF MATHEMATICS, PRINCETON UNIVERSITY
Downloaded from [Link] by [Link] on July 31, 2025 from IP address [Link].

Communicated January 21, 1932

I. In a recent issue of these PROCEEDINGS' the author has obtained a
proof of the so-called quasi-ergodic hypothesis. The reader is referred
to that paper for the precise formulation of this hypothesis, which plays
so important a r6le in the foundations of classical statistical mechanics,
and thus in the kinetic theory of gasses; the terminology of that paper
will be used throughout this note. The exact statement of the mathe-
matical result obtained in the previous paper by the author is as follows:
Let Q be either the phase-space 'I of the mechanical system considered,
or a sub space of TX invariant under the transformation (P - P,
P a point of 4i, t the time) induced by the equations of motion.2 Let dv
be the volume element defined in Q invariant3 under the transformation
P -> Pi, ,N the Lebesgue measure (or weight) of N(C Q) defined by
means of dv: ,N = J.'dv. Let the time of sojourn of Pr in N during the
time s < T < t, divided by t - s, be denoted by Zs,,(N; P).
Then there exists a function Z(N; P) such that, as t - s - + c,
the function of P Zs,1(N; P) converges, in the sense of "strong conver-
gence" in the space of functions of P, to the limit Z(N; P); that is
lim f Zs,,(N; P) - Z(N; P) 12dv = 0. (1)

This property determines Z(N; P), which function, in our previous

paper, is studied in more detail and calculated explicitly. The condition
for the validity of the so-called quasi-ergotic hypothesis is that Z(N; P)
be independent of P; we have shown in our earlier paper that this will be
true if and only if there exists in Q no integral of the equations of motion