Location via proxy:   [ UP ]  
[Report a bug]   [Manage cookies]                

Chatterji1971 Differentiation Along Algebra

Download as pdf or txt
Download as pdf or txt
You are on page 1of 12

manuscripta math.

4, 213 - 224 (1971)


~ by Springer-Verlag 1971

DIFFERENTIATION ALONG ALGEBRAS

S.D. Chatterji

The convergence of the generalized derivatives of a sequence


of finitely additive set functions along a sequence of alge-
bras is proved very simply by using only the maximal inequal-
ity. This gives a purely measure-theoretical proof and exten-
sion of the semi-martingale convergence theorems.

i. Introduction

The purpose of this paper is to show that very general


theorems (see e.g. Ths. 1,2,4) asserting convergence a.e. of
generalized derivatives of finitely additive set-functions
on an increasing sequence of algebras with respect to a count-
ably additive or even finitely additive measure follow from
one simple inequality namely the usual maximal inequality of
the subject (lemma 2). The usual martingale and semi-martin-
gale convergence theorems are immediate corollaries of our
theorems and thus one has a simple purely measure-theoretic
treatment of these important theorems without recourse to
any probabilistic inequalities involving upcrossings. These
latter are of c%urse important and indispensable in the the-
ory of stochastic processes. But since the convergence
theorems in question are of wide interest outside probability
theory, our approach may be worth while in its own right.
Besides, the method has the added virtue of being directly
applicable to vector-valued situations (see Chatterji [2]).
Also as we shall show in another work, our method of proof
centering solely around the maximal inequality clarifies the
theorems of Krickeberg and others (see [8] for an account)
concerning differentiation along a directed set of algebras
and the various Vitali conditions which are needed therein.

213
2 CHATTERJI

H o w e v e r in the last section of this p a p e r we show that using


classical ideas of Lebesgue, De la Vall~e Poussin and Riesz-
Nagy (see in p a r t i c u l a r [10] pp. 87-89) it is p o s s i b l e to
give a short proof of the classical differentiation theorem
of Lebesgue along r e g u l a r sequences of sets using just the
simplest case of our Th. I. This proof assumes only the
global d i f f e r e n t i a t i o n t h e o r e m w h i c h goes under the name of
Radon-Nikodym and avoids all covering theorems.

We are of course w h o l l y aware of the fact that our


theorems can also be deduced from the standard Doob version
of m a r t i n g a l e theory ([3] pp. 319) despite their g e n e r a l i t y
involving algebras instead of o-algebras and finitely addi-
tive set-functions. It is the method w h i c h we want to stress
here.

Given a measure ,, w h i c h will always be o-finite, we


call a set function 8 a u-integral if 8(A) = fA f du for
some function f. We define u-singularity of 8 by using the
e-~ d e f i n i t i o n on sets of finite u-measure. In one t h e o r e m
(Th. 3), we shall need the derivative of a ~-additive but
possibly n o n - o - f i n i t e 8. We define this to be a function f
such that (f+m) + = D(e+mu) + for m = 1,2,... and where the
right-hand side is the u - R a d o n - N i k o d y m derivative of the
absolutely continuous part of the o-finite positive m e a s u r e
(8+mu) +. This d e f i n i t i o n is in a c c o r d a n c e with all others
used in such cases.

2. C o n v e r g e n c e theorems: u o-additive

Throughout this section, (S~Z,u) will be a fixed


(totally) o-finite m e a s u r e space. If G c ~ is an algebra of
sets, M(G) will stand for the space of all real-valued,
bounded and finitely a d d i t i v e set-functlons on G. Given a
8 g M(G), let 8 + and e- be the positive set-functions (i.e.
in M+(G)) w h i c h arise from the Jordan d e c o m p o s i t i o n of e
+
and let 181 = e +8 be the absolute variation function

214
CHATTERJI 3

of e. We shall say that e is u-continuous if ~§


If p were finite, u-continuity of e w o u l d imply its g-addi-
tivity. We shall say that e is u - s i n g u l a r if for any posi-
tive E, ~ and A e G of finite u-measure there exists BeG
such that ~ ( A n B ) <e and [el(A-B) < ~.

For a e in M+(G), let

Co(A) = inf{5-e(Aj) = AC 0 Aj} .


j:l

It is clear that e ~ e and that e is u - a d d i t i v e on G. Indeed


O g
e ~ is c h a r a c t e r i z e d by the fact that it is the largest g-addi-
tive, n o n - n e g a t i v e set function less than or equal to 8. For
an a r b i t r a r y e in M(G), let e
= e + - e - and e = 8 - 8 . e is
o o p o o
called the u - a d d i t i v e part and e
the purely finitely a d d i t i v e
P
part of e. ep has the r e m a r k a b l e p r o p e r t y of being singular
with r e s p e c t to any u - a d d i t i v e set-function on G and hence is
in p a r t i c u l a r u-singular. The notion of pure f i n i t e - a d d i t i v i t y
and the d e c o m p o s i t i o n of a f i n i t e l y additive e into a o-addi-
tive e and a p u r e l y f i n i t e l y additive e are due to Hewitt
o p
and Y o s i d a [5] where a d e t a i l e d d i s c u s s i o n may be found. Our
construction of e is different from that in [5] and follows
O
a well-known classical m e t h o d due to C a r a t h ~ o d o r y [i] pp.
245-255. Let us now apply the L e b e s g u e - d e c o m p o s i t i o n theorem
to e g to obtain (assuming that the r e s t r i c t i o n of u to G is
again u-finite)

e~ = I f du + e'(A)
A
where e' is a u - s i n g u l a r set-function and f is a u - i n t e g r a b l e
function m e a s u r a b l e w i t h respect to the o - a l g e b r a g e n e r a t e d
by G. Finally then any e in M(G) can be w r i t t e n as

f
e(A) | f du + e (A) (8 : e + e')
JA s s p

where e s is u-singular. In the following we shall denote the


function f so o b t a i n e d by DO. Of course the o p e r a t o r D de-
pends on G and on u too. Our c o n v e n t i o n will be to take G
always as the domain of d e f i n i t i o n of e. Du can be c o n s i d e r e d

215
4 CHATTERJI

as a g e n e r a l i z e d derivative of e w i t h respect to ~ and G


such that "IG is e-finite.

Let G n c Gn+ 1 c Z be a sequence of algebras and let


G = U G n (n 91). Clearly G is an algebra. Given 8 gM(G) we
shall w r i t e Wn e for the r e s t r i c t i o n of e to G n. We shall
also assume that Wl ~ and hence Wn ~' n > l, is e-finite. Our
first theorem is

THEOREM i. D~ e ~D8 a.e. (,).


n

We shall deduce theorem 1 from a standard special case


of it which we call

LEMMA I. If e e M ( G ) is of the form de : f-d~ then f : D~ e


-- n n

converges a.e. and in LI(~) t__qof = E f = the Z -conditional


expectatipn o f f where z i__ssth___~ee - a l g e b r a generated by G.

We note that if Z n : o-algebra generated by G n then


f = D~ e = E f = the Z - c o n d i t i o n a l expectation of f.
n n n n

We recall that lemma 1 is p r o v e d by using the following


two remarks w h i c h we f o r m u l a t e as lemmas

LEMMA 2.
~{s : m a x Ifk(s)l ~t} ~ t-l.ljfILl , t > 0 (i)
14k4n

LEMMA 3. I f T n are linear operators o__nnLI(~) such that

(i) U{ m a x ITkf I ~ t } ~ t-l.llflll , t > 0


l~k~n
and
(ii) lim
T f exists a.e. (u) for a set of f densein LI(u)
n
then lim T f exists for all f e LI(~).
n

Lemma 3 is g e n e r a l l y considered so trivial (and r i g h t l y


so) that authors use it w i t h o u t mention. Condition (i) of
lemma 3 can be varied of course in many ways. If we r e p l a c e
it by the w e a k e r condition suplTnfl < | a.e. (u) we get a
much deeper theorem of Banach cf. [4] pp. 332-333. C u r i o u s l y
however, it seems that in all concrete cases one proves the

216
CHATTERJI 5

a.e. boundedness of Tnf only via an i n e q u a l i t y like (i) and


so the t r i v i a l lemma 3 is almost always a good enough sub-
stitute for the d e e p e r theorem of Banach.

As r e g a r d s lemma 2, called a maximal inequality, we


refer to [3] pp.314. The simple argument used to prove this
inequality will be used and shown explicitly in the proof of
lemma 4.

PROOF OF T H E O R E M i. Writing 0 = ~ + 6, ~,8 c M ( G ) where a is


a u-integral and ~ is ~ - s i n g u l a r (possibly finitely additive)
we see that De = D~ and D~ne = D~n~ + D~n6. By lemma i,
D~ a >D~ = De a.e. Hence all we need to prove is
n

LEMMA 4. If 6 a M ( G ) i__ssu - s i n g u l a r then D~n6 .0 a.e. (n).

PROOF OF LEMMA 4. Let b n = D~ n ~. Since b n is a u - i n t e g r a b l e


function measurable with respect to Z= o-algebra generated
n
by Gn, we can find G n - s i m p l e functions b n' such that
Zfibn-b~Idu_ <~. show that b n' ,0 a.e. This implies
We shall
by virtue of the good a p p r o x i m a t i o n of b' to b that b ~0
n n n
a.e. (standard argument).

We can assume that u is f i n i t e , f o r once this case has


been settled, the g e n e r a l case is a simple deduction from
the a-finiteness of G I. The u-singularity of 8 implies that
given e,6 > 0 there exists a set A in G and hence in some G N
such that
IBI(A) < ~/3 u(S-A) < ~/3 .

Further we can choose II so large that

Slbn-b~ld~ < ~a/3 .


n~N
now if Bn = {A;Ib!l.. > ~, Ibll+3 ~ N~j~n-1} we have

M
{A; max
}1-< n-< M
ibnl, .< -1 I IbnldU (2)
n=n Bn

~< s { b n - bnldU +~--- IbnldP}


N N B
n

217
6 CHATTERJI

M
..< 6 / 3 + E - I ~ - I',,nSl(B n ) (3)
N
M
..< ~ / 3 + - i ~ - - I~M6 l(Bn) (4)
N

4 6/3 + c-l-I~MSi(A; max Ib~l. > E) (5)


N ~ n~ M

26/3.

In (2) and (5) we have used the fact that the disjoint sets
B add up to the set on the left-hand side. In (3) and (4)
n
we use the obvious relations ;Bnlbnldu ~ I~nBl(Bn).< I~MBI(B n)
if n~<M. (Note that B 6 G .) Hence
n n
~{ max Ib'l>~}
N.<n.<M
4 ~{A; max Ibnt > e} + u(S-A)
N.<n.<M

i.e. letting M---~- we have that for any ~,6 > 0 there exists
N such that ~{SUn>~ Ibnl > c} < 6. This clearly proves that
b' ~ 0 a.e. and completes the proof of lemma 4 and t h e o r e m i.
n

The same type of proof gives the more general (as before
G is an increasing sequence of algebras):
n

THEOREM 2. Let e e M ( G
) and let lim 8 (A) = 8(A) exist for
n n n
al__~lA 6 G = U G n. Suppose that

(i) e e M(G)
and
(it) 10 n - ~n81 4 v n e M ( G n ) where

Vn+ 1 ~ v n and vn 20
then
De n . De a.e. (u) .

Note that the condition (it) is satisfied in the follow-


ing cases:
(i) (Submartingale case) ~nSn+l > 8n; here v n can be taken
as ~n 8 - 8 n
218
CHATTERJI 7

(2) (supermartingale case ) Wnen+l "< On' Vn = O n - w en

(3) zIl~nen+l-enl I < | ; here we can take Vn = k)nZ I"kek+l-ek I.


This case has been discussed in an unpublished manu-
script by S. Johansen who used different methods.

PROOF. Let
= ~ eM(G
Wn8 - e n n n) .

We shall show that D~n ~0 a.e. This along with the fact
that DWne--~De a.e. (Th. i) and D8 n = Dw n 8 + D~n will prove
theorem 2.

Let h n = D~n. This is a Zn-measurable LI(u) function


where Z n = ~-algebra formed by Gn. As before let h'n be Gn-
simple functions such that Z llhn-h~lll < -. It will suffice
as before to show that hi---*0 a.e.
n

Again we may assume that u is finite. Calculations


parallel to (2) and (3) above give

u{ max Ih~l > c}


N~n~M

~< -l{>-- Ilhn_hnfll + ~- l~nl(Bn) }


n>~N N.<n~<M

where B n = {lh~l > E, Ih~l 4 ~, N~j4n-l}.

Using I~nl 4 ~n 4 ~N (condition (it) of th. 2) and the


disjointness of Bn, we continue the inequality by

-i{
e 7- Ilhn h'II1 + ~N(S)} .
n>N n

By chosing N large enough we can make the last expression


smaller than any pre-assigned 6 > 0. This implies that
h'--~0 a.e. and completes the proof of theorem 2.
n

The proof indicates also how condition (it) of th. 2


may be further weakened to obtain the same conclusion. How-
ever, that some condition is necessary i.e. the convergence
of De n does not simply follow from the convergence of 8n to
8, even if the latter be in M(G), is shown by the following

219
8 CHATTERJI

example. A much more complicated example occurs in the p a p e r


[7] pp. 694.

EXAMPLE. Let G = a l g e b r a f o r m e d by the p a r t i t i o n


n
{[m.2 -n, (m+l)2 -n) : 0 ~ m ~ 2 n - l } of [0, i), ~ = L e b e s g u e meas-
ure, 8n(A) = SARnd~, A6 G n R n = nth Rademacher function.
For any A 6 G = U G n it is seen easily that %n(A) )0 i.e.
e ~ 0. But clearly De = R o s c i l l a t e s a.e. between +i and
n n
-i and h e n c e does not c o n v e r g e to anything.

Theorem 2 has the following "unbounded" generalization


in case the en'S form a monotone sequence.

THEOREM 3. Let e be an a d d i t i v e f u n c t i o n on G + G and let


n +n
~ n S n + l > e n. T h e n e = l i m e n exists on G. Let 8 < -. T h e n
De ~De a.e. (~).
n

PROOF. Since 8 + < ~ we can d e f i n e De = D8 + - D e - w h e r e D8 -


n n n n n
is p o s s i b l y +~ s o m e t i m e s . As b e f o r e we can s u p p o s e ~ to be
a finite measure. Using now the inequality (m f i x e d positive

integer) )+ m~nU)+ +
(8 n + m~n~ 4 (~nSn+l + ~ ~n8 + m~n~
and the relation )+
D(8 n + m r n ~ = ( D e n + m)+

we can deduce from th. 2 that


+ m) + (D6
~ (De + m) + a.e.
n
i.e. on the set {lim sup De > -m}, De ~De a.e. On the
n n
rest, lim D8 = -~ i.e. f = lim De e x i s t s a.e. Now
n n
(f+m) + = ( D e + m) + for all m > i h e n c e f = De a.e.

To c o m p l e t e the discussion, we add that the case where


the G ' s form a decreasing sequence is a l i t t l e easier,
n
although not quite symmetrical in the generality in w h i c h
we are discussing here. It is easy to g i v e examples of
8 G M ( G I) so that ~n e 6 M ( G n) m a y be all u-singular although
81G , G = N G n may be a u - i n t e g r a l . So in g e n e r a l lim D~ne
does not equal D(81G). If 8 is a w - i n t e g r a l on G 1 and G is
supposed to be o - f i n i t e then this is the case. Even if G is
not o-finite but 8 is a u - i n t e g r a l on GI, lim D~n8 exists

220
CI~TTERJI 9

and can be i d e n t i f i e d cf. J e r i s o n [6]. The proof again uses


nothing m o r e than the m a x i m a l inequality.

3. p finitely additive

To avoid o v e r b u r d e n i n g the text, we assume that u is


finite i.e. u ( S ) < = and n a t u r a l l y p o s i t i v e and additive. The
proofs of w indicate immediate generalizations to this case.
Let Gn, G be as before and let ~ be defined at least on G.
Given e 6 M ( G n) we can w r i t e e = ~+8 uniquely, e, 8 e M ( G n)
where ~ is p-continuous and 8 is p-singular. This can be
done e.g. by passing to the a p p r o p r i a t e Stone space see e.g.
pp. 312 [4]. A l t h o u g h e need not be a p-integral, a theorem
of B o c h n e r - P h i l l i p s (see [4] pp. 315), provable also by
passing to a Stone space, allows us to find for any ~ > 0 a
function f w h i c h is G n - s i m p l e and such that le(A)-fAfdU I <
for all A e G n. Such a function f we shall call a__nne-approxi-
mate d e r i v a t i v e of e w i t h r e s p e c t t_~o u. The notion of u a.e.
convergence we shall replace by what we call q u a s i - a . e . -
convergence. A sequence fn converges quasi-a.e. (u) to f if
to any e, 6 > 0 there is N such that for all M ~N,

u{ max If - fl > e} < 6 .


N~n~M n
For o - a d d i t i v e u, this notion is clearly equivalent to a.e.
convergence. Similarly we shall say that fn is Cauchy quasi-
a.e. if
u{ max If -fml > E} <
N4n,m(M n
w h e r e the symbols are as before.

W i t h these defintions we can prove the following ana-


logue of t h e o r e m I.

THEOREM. I f e is of the form fAf du where f i__{sG - m e a s u r a b l e


(in the sense of [4] pp. I06) and fn i_~s an C n - a p p r o x i m a t e
derivative of ~n 8 where ZE n < = then f n )f q u a s i - a . e . u .

221
I0 CHATTERJI

l__nngeneral for any e e M(G) if fn is a__nnE n - a p p r o x i m a t e


derivative of the u-continuous part of ~n 8 w h e r e r en
- - < oo

then fn is Cauchy quasi-a.e.

I f E, 6 >0 aye ~iven w_~e can find a p p r o x i m a t e derivatives


f and fn of the ~-continuous parts o f e and ~n e and an N
such that for all M ~N,

~{ max If - fl > e} < ~ .


Nzn~M n

We omit the proof since it would be quite analogous to


that of lemma 4. Analogues of t h e o r e m 2 for finitely addi-
tive u can be formulated just as easily.

4. Lebesgue differentiation theorem

Let the basic m e a s u r e space be the k - d i m e n s i o n a l


Lebesgue m e a s u r e u on the Borel sets Z of R k. For each
j, 0~j~2k-l, we construct in a special way an increasing
sequence EJ of s u b - o - a l g e b r a s such that E = o-algebra
n
spanned by U { E j : n~0}. Let n ~ be the p a r t i t i o n of the
n n 1
line indicated by {[m3 -n (m+l)3 -n) : -~ < m < +=} and
n
that given by { [(m+I/2)3 -n, (m+3/2)3 -n) : -~ < m < +~}. The
1
p a r t i t i o n points of ~ are simply the mid-points of contig-
n
uous p a r t i t i o n points of ~ . If 0~j~2k-l, we can write
j = Eo + El2 + "'" + c k_l 2k-l, E.l = 0 or i, uniquely. Let
~J = ~ o • ~ i • ... • ~ en-I be the product p a r t i t i o n of R k
n n n n
induced in the obvious m a n n e r and let ZnJ be the a-algebra
spanned by the p a r t i t i o n ~] In the t e r m i n o l o g y of De la
n"
Vall~e Poussin, the 2k different nets ~ , for a fixed n, are
conjugate to each other.

If 8 is any o-additive r e a l - v a l u e d set function on Z


and f is the R a d o n - N i k o d y m derivative of its u-continuous
part, then t h e o r e m 1 applied to {E~}n~ o and 18-tul, t
rational, proves the existence of a set N of u-measure zero
such that if x e Rk-N, the equation

222
CHATTERJI Ii

lim le-t~l(An)/u(A n) = If(x)-tl


holds where A is a half-open cube containing x and belong-
1 n
ing to 6 j for some j and A ~x in the obvious sense. A
n n
standard argument shows that the same is true for all real t
simultaneously for the same set N. In particular putting
t = f(x) we get, for x c R k - N
lim 18-f(x)uI(An)/U(A n) = 0

Now we use the crucial geometrical fact that given any cube
C with centre x and side < 1/2 there is a j, n and A e 6 j
n
such that C c A and side of A < 6 (side of C) so that
#(A) < 6ku(c). This follows from the corresponding fact for
k = 1 where it is obvious.

If C is any sequence of cubes with centre x cRk-N and


n
C *x then we can find A ' s converging to x as indicated
n n
above. Hence ( C n C A n , u(An)<6ku(Cn))

le-f(x)wl(Cn)/u(C n) ~ le-f(x)ul(An)/U(C n)

4 6k'le-f(x)ul(An)/u(An)--~0.

The same argument shows that if S n is any sequence of sets


of positive u-measure converging to x regularly (see Munro,
[9J pp. 269) and x e Rk-N then
lim le-f(x)ul(Sn)/U(S n) = 0 .

But this clearly implies that e(Sn)/u(S n) ~f(x) which is


Lebesgue's theorem.

References

[I] CARATHEODORY, C.: Algebraic theory of measure and inte-


gration, New York: Chelsea 1963.

[2] CHATTERJI, S.D.: Martingale convergence and the Radon-


Nikodym theorem in Banach spaces. Math. Scand. 22,
21-41 ( 1 9 6 8 ) .

[3] DOOB, J.L.: Stochastic processes, New York: Wiley 1968.

[4] DUNFORD, N., and SCHWARTZ, J.T.: Linear operators I,


New York: Interscience 1958.

223
12 CHATTERJI

[5] HEWITT, E., and YOSIDA, K.: Finitely additive measures.


Trans. Amer. Math. Soe. 72, 46-86 (1952).

[6] JERISON, M.: Martingale formulation of ergodic theorems.


Proc. Amer. Math. Soc. i0, 531-539 (1959).

[7] JOHANSEN, S., and KARUSH, J.: On the semi-martSngale


convergence theorem. Ann. Math. Statist. 34, 690-694
(1966).

[8] KRICKEBERG, K., and PAUC, C.: Martingales et d~rivation.


Bull. Soc. Math. France 91, 455-544 (1963).

[9] MUNROE, M.E.: Introduction to measure and integration,


Reading: Addison-Wesley 1953.

[lO] RIESZ, F., and Sz.-NAGY, B.: Functional analysis,


New York: Frederick Ungar 1955.

S.D. Chatterji
Battelle Institute
7, route de Drize
CH-1227 Carouse

(Received July 13, 1970)

224

You might also like