Canad. Math. Bull. Vol. 39 (2), 1996 pp.

138-150

THE DOOB-MEYER DECOMPOSITION REVISITED

RICHARD F. BASS

ABSTRACT. A new proof is given of the Doob-Meyer decomposition of a super-

martingale into martingale and decreasing parts. Although not the most concise proof,
the proof is elementary in the sense that nothing more sophisticated than Doob's in-
equality is used. If the supermartingale is bounded and the jump times are totally inac-
cessible, then it is shown that discrete time approximations converge to the decreasing
part in I?. The general case is handled by reduction to the above special case.

1. Introduction. The Doob-Meyer decomposition says that under mild integrabil-

ity conditions, a supermartingale can be decomposed into the difference of a martingale
and an increasing process. This wasfirstproved by Doob [D] in the discrete time case and
Meyer [Ml, M2] in the much harder continuous time case. There are a number of other
proofs, including those of Doléans-Dade [DD] and Rao [R]. Doléans-Dade uses the no-
tions of predictable projections and dual predictable projections. Rao's proof is probably
the simplest; he uses a discrete time approximation, Doob's decomposition for discrete
time supermartingales, and a limit procedure. Despite this, the Doob-Meyer result is still
considered a hard theorem, most likely because the limit procedure uses convergence in
the topology a(Ll,L°°), which in turn uses the Dunford-Pettis compactness criterion.
In this paper we give a new proof of the Doob-Meyer decomposition. Our proof,
although not the most concise, is completely elementary in the sense that the most so-
phisticated technique we use is Doob's inequality. We also start with a discrete time
approximation, but now the convergences are in probability. In fact, when the super-
martingale is bounded with totally inaccessible jump times, we show directly that the
discrete approximations converge in L2.
In Section 2 we look at the case where the jump times are totally inaccessible. The
general case, which involves a reduction to the case of Section 2, is done in Section 3.
When the supermartingale is continuous, our proof can be made much more straight-
forward; this case has been presented in Bass [B]. For more information on predictable
and totally inaccessible stopping times, see [DM1, DM2].
After this paper was submitted, we learned that T. Brown had earlier found an ele-
mentary proof of this theorem, but his proof was never published.

Supported in part by NSF grant DMS-9100244.

Received by the editors January 25, 1995.
AMS subject classification: Primary: 60G07; secondary: 60G05, 60G44.
Key words and phrases: supermartingales, Doob-Meyer decomposition, predictable, totally inaccessible,
martingales.
© Canadian Mathematical Society 1996.
138

https://doi.org/10.4153/CMB-1996-018-8 Published online by Cambridge University Press

 DOOB-MEYER DECOMPOSITION 139

2. The totally inaccessible case. In this section we do the case when the jumps of
the supermartingale are totally inaccessible. This is the heart of the matter and also the
most interesting and useful situation. Recall that a stopping time S is predictable if there
exist stopping times Sn strictly less than S which increase up to S, a.s., on (S < oo).
A stopping time S is totally inaccessible if P(5 = T < oo) = 0 for each predictable
stopping time T.
Let (Q, f, P) be a probability space, let Jt be a right continuous filtration, and let Zt
be a supermartingale. Without loss of generality we may suppose Z has paths that are
right continuous with left limits and that Z0 = 0. If Zs- denotes the left hand limit of Z at
s, the jump at time s is AZS — Zs — Zs-. In this section we assume that the jumps of Zt are
totally inaccessible. That is, if ^ ^ is the «-th time that Zt jumps more than e in absolute
value, then S„i£ is totally inaccessible for each n and e. Another way of phrasing this is
to say that whenever S„ are stopping times increasing up to S, then Z$n —* Zs, a.s. The
Markov theory literature calls this property quasi-left continuity. A supermartingale is
said to be of class D if the set of random variables {ZT : T a stopping time} is uniformly
integrable.
We prove
THEOREM 2.1. Let Zt be a supermartingale of class D with Zo = 0 where the paths
are right continuous with left limits. Suppose the jumps ofZt are totally inaccessible.
Then there exists a continuous increasingprocess At such thatMt = Zt+At is a uniformly
integrable martingale. The decomposition Zt—Mt— At is unique.
REMARK 2.2. The proof of the uniqueness is easy (see, for example, [P], [IW], or
[B]) and we have nothing to add here. In the remainder of the section we concentrate on
the existence.
LEMMA 2.3. Suppose {Q, k = 0, I,...} is an increasing sequence of random vari-
ables and Jk is an increasing sequence of cr-fields such that CQ = 0, Q is 5k-\ measur-
able, and there exists N E (0, oo) such that for all k,
E[Coo-Q|ft]<tf, a.s.
2
Then E ( 4 < IN .
PROOF. We have
ECoo = E [ E [ C o o - C 0 | ^ o ] ] < ^
Let c/c — Cjfc+i — Q > 0. Some algebra shows that
00 OO

cL = 2£(Coo-Q) Q -£<i
*=0 Jc=0
Then
oo
ECl < 2E £ E[Coo - Ck | %}ck\ < 2NE J2 ck = 27VECoo < 2N\
*=0 k=0
as required.

https://doi.org/10.4153/CMB-1996-018-8 Published online by Cambridge University Press

 140 RICHARD F. BASS

LEMMA 2.4. Suppose C^\ C^2) are two increasing sequences, each satisfying the
hypotheses of Lemma 23. Let Dk = Ô^—Ô^.Suppose there exists Y> OwithEY2 < oo
such that for all k,
|E[Doo-A| J i ] | < E [ y | ^ ] , a.s.

Then
EsupD2k < 8EF2 +32v / 2ME7 2 ) 1 / 2 .
k

PROOF. Let dk = Dk+l - Dk and cf = C ^ - C}. As above

D2OQ = 2YJ(Doo-Dk)dk-YJdl
k=0 k=0

Then

+ c
EDl < 2 E E E[£>oo - A | J * ] * < 2E £ £[Y I ftK^ ?)
k=0 k=0

= 2E
L
^ y ( 4 ' > + 42>)l = 2 E t F ( ^ + 0].
k=0

The Cauchy-Schwarz inequality and the bounds for E(CS)2 show ED2^ <
4v/2Ar(EF2)1/2.
TogetthesupremumoverA:,letM^ = E[Doo | !fk],Nk = E[Y | ^Fk],andXk — Mk—Dk.
Since \Xk\ = |E[Ax> ~ Dk | 7k\\ < Nk, by Doob's inequality

E s u p ^ | <EsupN2k < 4EA^ = 4EY2.

k k

Another use of Doob's inequality shows that

EsupM2<4EM^-4ED^.
k

Since sup^ \Dk\ < sup^ \Xk\ + sup^ \Mk\, the result follows. •

REMARK 2.4. We will use the following observation several times.

If Yn -+ 0 in I 2 , then supE[7„ | J , ] -> 0 in L2.

This follows from the fact that Mn{t) = E[Yn \ !ft] is a martingale, so by Doob's inequal-
ity,
EsupM„(0 2 < 4EM„(oo)2 = 4E72 -» 0.
t

The following lemma is of interest in its own right. Let v be a positive integer and let
En = {k/2n : 0 < k/2n < v}.

https://doi.org/10.4153/CMB-1996-018-8 Published online by Cambridge University Press

 DOOB-MEYER DECOMPOSITION 141

LEMMA 2.5. Suppose T is a totally inaccessible stopping time. For 8 > 0 let
R(8) = supP(t<T<t + 8\ 7t\

Then R(8) —> 0 in probability as 8 —> 0.

PROOF. Let a > 0 and let

S„(S) = inf{t e En : P(r < T < t + 8 | ft) > a} A v.

We first show Sn(8) < T, a.s. Note 5W(<5) takes on only the values k/2n;T cannot take
on any of the values k/2n with positive probability, or else part of T could be predicted
by the stopping times k/2n - 1 jm. Hence P(Sn(8) = f) = 0. If A Q(T<t) and A is <ft
measurable, then
E[P(f < T < t + 8 | fi);A] = ¥(t<T< t + 8; A) = 0.
If 7 were less than Sn(8) with positive probability, then for some t E En, we have
P ( r < f, 5„(S) = f) > 0. Let A = (T < t, S„(S) = f). Observe that
P(* < T < t + 6 | fi) > a on the set (S„(6) = i)9 hence on the set A, which is a
contradiction. We conclude that Sn(8) < T.
We next define a stopping time 5. Let
50) = inf Sn(8% 5 = sup 5(1 In).
n
n

Since 5W(<5) < T, then 50) < 7, a.s. Since 71 is totally inaccessible, we must have
P(S=T) = 0. This implies
(2.1) P(S=T\?s-) = 09 a.s.
(If U is a predictable stopping time predicted by the stopping times U„9 then ^iy_ is the
a-field generated by the sets in U„!fun.)
We now complete the proof of the lemma. Suppose there exists e > 0 such that
(2.2) P(R(8)>a)>e,
no matter how small 8. Let j5 > 0 and take 5 < /}. For n sufficiently large,
P(p{Sn(8) <T< Sn(8)+8 | ?SM) >a)>e.
T cannot equal S{8) + 8 with positive probability, or else part of T could be predicted by
the stopping times S(S) + 8 — 1 jm. So the probability of the symmetric difference of the
set (Sn(8) <T< S„(S) + 8) and the set (S(8) <T< S(8) + 8) tends to 0 as n - * oo.
Using Remark 2.4,
P(P(50) < 7 < S{8) + 8 | ^ ) > a) > e9

hence ?{S(8) < T < S(8) + /? | f^ô)) *s greater than or equal to a with probability at
least e. Next we let 5 = 1 jn and let « —» oo. Repeating the argument we just gave, we
get that
P(P(5 < T < 5 + /3 | fs-) >a)>£.
Letting (3 —• 0, P(5 = 7 | ^ _ ) > a with positive probability, which contradicts (2.1).
Hence (2.2) fails and the lemma is proved. •

https://doi.org/10.4153/CMB-1996-018-8 Published online by Cambridge University Press

 142 RICHARD F. BASS

LEMMA 2.6. Suppose Z satisfies the hypotheses of Theorem 2.1 and in addition \Z\
is bounded by N and the paths ofZ are constant after time v. Let

(2.3) W(6)= sup E[Zt-Zu\Tt].

t<u<t+8

Then W(5) -> 0 in L2 as 8 -> 0.

PROOF. Since \Z\ is bounded by N, then W{8) is bounded by 2Af, so it suffices to

show W{8) —> 0 in probability.
Let e, a > 0 and b — a^fz. Let

=
^t =X)^l(fe<AZ s ) 5 Zt ^2&Zsl(b<-AZs),
s<t s<t
Zf = Z,-(Zf+Z?).

Let
WS(S)= sup \E[Z?-Zsu\f<]\,
t<u<t+6

with ^(tf) and WM(8) defined analogously.

Note
Ws{8) < supE[ sup \Zsr - Zf | | ?t\
t r<s<r+8

By Doob's inequality,

P(Ws(8)>a)<E[ sup \Zsr - Zss\2]l a2.

r<s<r+6

Since Z 5 is right continuous with left limits and the jumps of Zs are bounded by b in
absolute value, the lim sup of the right hand side as 8 —> 0 is less than or equal to
b21 a2 = e by Fatou's lemma.
Since Z is right continuous with left limits, there are only finitely many jumps of size
larger than b. Let T\ = inf{f : AZt > b}, and for / > 1, Ti+] = inf{t > 7} : AZ, > 6}.
These are the times when Z has a jump of size larger than b. Since \Z\ is bounded by N,
then | AZT;. | is bounded by 2N. Choose K such that P(7* < v) < e. Then

?{WP{8) >a)< P(TK < v) + ^P(supE[AZr i l ( r<r / <i^ I 7A > a/K)

K
SU
(2.4) < e + J2 H P P (' <Ti<t + 6\yrt)> ajlKN).

By Lemma 2.5, the right hand side can be made less than 2e if 8 is small enough.
WM(8) is treated similarly. Since W{8) < WS(8)+WP(8)+WM(8\ the result follows. •

https://doi.org/10.4153/CMB-1996-018-8 Published online by Cambridge University Press

 DOOB-MEYER DECOMPOSITION 143

PROPOSITION 2.7. Suppose Zt satisfies the assumptions of Theorem 2.1 and in addi-
tion \Z\ is bounded by N and the paths of Z are constant after time v. Then the conclusion
of Theorem 2.1 holds.
PROOF. Fix n for the moment and let ^ f = 7k/2n- Let
a
\ = ^iZ(k-\)/2" -Z/c/2» I 7k-\\
Since Zt is a supermartingale, the a\ are nonnegative. Note the a\ are J£_x measurable.
Let A\ — Y!JZ\ <*]• It is trivial to check that Zk/2n + A\ is a discrete time martingale with
respect to ?». Let 5? = Ank if (A:- \)/2n < t < k/2n.
We will show the B" converge in L2 as n —* oo, uniformly over /, by showing
(2.5) E[sup|^-^|2]-^0 as/z,m-+oo.

Suppose m > n. Since B™ and B" are constant over intervals (k/2m, (k + 1)/2 W ], the
supremum of the difference will take place at some k/2m. We will apply Lemma 2.3
(with Ôkl) = A%, Ô2} = ££/2„,, and y equal to W(2~n), and with respect to the a-fields

Fix t — kl2m and let w be the smallest element of En bigger than or equal to /.
(2.6) £[(&> - 6? | J f ] - ELAZ - A? | ^ " ] = E[Z, - Z^ | Jt\
which is bounded by IN. On the other hand,
(2.7) E[C£> - C f | Jf] = E[<, - 5? | £ ] = E[E[<, - 5^ | ft] | Ft]
= E[E[Z„ - ZM | ft] | J,] - E[Z„ - Zoo | Ft],
which is also bounded by 2N.
Taking the difference of (2.6) and (2.7),
E[(ZC - / £ > ) - (A? - 5?) | ft] = E[Z( - Z„ | ft].
The right hand side is nonnegative and bounded by W{2~n). Since the right hand side is
% measurable, it is also bounded by E[W(2~n) \ 5t\- So by Lemmas 2.3 and 2.6, we get
(2.5). Let us denote the limit of the Bnt by At.
Next we want to show that At is continuous. The jumps of Bnt are

AB" = E[Z(k_iy2* - zk/2" I 7{k-\)li»\ t

= V2"'
which are bounded by W(2~n). Hence sup, \AB"\ —> 0 in L2. By looking at a suitable
subsequence «/, sup, \ABÏJJ\ —» 0, a.s., and so the limit is continuous.
Finally we show that Zt + At is a uniformly integrable martingale. Since Z, is right
continuous and^4, is continuous and both are square integrable, it suffice to show that for
s, teEn,s <t,mdB e 7,
E[Zt+At;B] = E[Zs+As;B].
This follows readily by a passage to the limit from the corresponding equation for Zt+B".
The uniform integrability follows since \Z\ is bounded and A is square integrable. •

https://doi.org/10.4153/CMB-1996-018-8 Published online by Cambridge University Press

 144 RICHARD F. BASS

COROLLARY 2.8. Suppose Zt satisfies the hypotheses of Theorem 2.1, but in addition
the jumps ofZt are bounded. Then the conclusion of Theorem 2.1 holds.
PROOF. Let TN = M{t : \Zt\ > N} A Wand let Zf = ZtATN- Since the jumps of Zt are
bounded, ZN will be bounded, and by Proposition 2.7, there exist a continuous increasing
process A?f and a martingale Aff such that Z^ — M? — A?.
Suppose L >N. Then Zf = ZLtKT^ — MLtKr^ — A^ATN is another decomposition of ZN.
By the uniqueness result (Remark 2.2), A^ATN = A?. Thus if we define At to be A^ for
t < TV, the definition of At is unambiguous.
By monotone convergence and the fact that Z is of class D,

EAoo < l i m L ^ = — limEZj^ < oo.

Since At is increasing in /, this implies the uniform integrability of Mt. m

PROOF OF THEOREM 2.1. By the proof of Corollary 2.8, it suffices to obtain a de-
composition of ZtAr, where N>0 and T = inf{t : \Zt\ > N} A N. We may thus suppose
that \Zt-\ is bounded and that Z has at most a single jump larger than 2N, occurring at
time T.
Let
Zt — AZrl(r<f)l(Azr>i)5 Z* = AZrl(r<ol(-AZr>i)-
Z/ and —Zp are both supermartingales since they are decreasing processes. Suppose we
can find a Doob-Meyer decomposition for each: ZJt = MJt — AJt and —Zpt — MP — Ap.
Then note that

Zf = Zt~(Zp - Ap) - (Z-j+A-j) = Zt+Mp-Afj

will be a supermartingale with jumps bounded by IN + 1, and by Corollary 2.8 will

have a decompositionMf — Af. Zt = (Mf — Mp + Aff) — Af will then be our desired
decomposition for Zt.
We proceed to decompose Z/, the decomposition of—Zf being similar. Note \AZT\ <
\ZT-\ + \ZT\ < N+ \ZT\, SO \AZT\ is integrable. Let a, e > 0. Chooser > 1 large enough
so that E[|AZr|; \AZT\ > R] < ea. Let
=
Zt = AZ7-l(r<r)l(-AZ r >/?)î %t %t ~ Zt •

Define #/'", #f'M, and #fn in terms of Z 7 , ZL, and Z 5 in exactly the same way Bnt was
defined in terms of Z in the proof of Proposition 2.7.
We show #/,w converges in probability, uniformly in t, by showing it is a Cauchy
sequence. We have

(2.8) P(sup |^' w - £f'm| > a) < P(sup |#f" - #fm| > a/3)
* t

+ P(sup|^'"| > a / 3 )

+ P(sup|5f'm| > a / 3 ) .

https://doi.org/10.4153/CMB-1996-018-8 Published online by Cambridge University Press

 DOOB-MEYER DECOMPOSITION 145

The second term is small since

P(sup |5f'"| > a/3) = P(Z&" > a/3) < (3/a)Ei&"
t

< (3/a)E|Z^| < (3/A)E[|AZH; |AZr| >tf]

and similarly for the third term. \Zlt\ is bounded by 7?, so the first term on the right of
(2.8) can be made small by taking m and n large as in the proof of Proposition 2.7.
Therefore Bf'n converges, uniformly in /, as n —* oo. Let the limit be denoted by Aj.
The continuity of AJt is exactly as in Proposition 2.7. For each n, EB^ = — EZ^, so by
Fatou's lemma, EAJœ is integrable. With this fact it is not hard to see that Z/ + AJt is a
martingale, and that this martingale is uniformly integrable. •

3. The general case. In this section we prove the general case of the Doob-Meyer
decomposition. IfR and S are stopping times, let [ i ? , 5 ] C f i x [0, oo] denote {(UJ,S) :
R(u) < s < S(UJ)}. The graph of a stopping time S is the set [S, S]; the finite part of
the graph of S will be defined to be {(a;, S(ujfj : S(u) < oo}. A process is predictable
if, considered as a map from Q x [0, oo], it is measurable with respect to the <j-field
generated by the sets {[0, S] : S a predictable stopping time}.
THEOREM 3.1. Suppose Zt is a supermartingale of class D with ZQ = 0 and with
paths that are right continuous with left limits. Then there exists a predictable increasing
process At such thatMt = Zt+At is a uniformly integrable martingale. The decomposition
Zt — Mt — At is unique.
REMARK 3.2. Again, the uniqueness is not difficult—see [IW].
LEMMA 3.3. Suppose R and S are predictable stopping times. Let

K
' I oo ifS(u) = R(u).
Then S' is a predictable stopping time.
PROOF. Let Ri9 Si be stopping times predicting R and S, respectively. Define

Sf = lSi {
™i>R , Sf = mfSf, SA = supSf.
I oo otherwise j>i J t

It is easy to see that Sf is a stopping time, hence so is SA. If S > R, then for all / sufficiently
large, 5/ > R. Hence for / large, Sf = St, so Sf = Sh and thus SA = S. If S < R, then
Sf < R for all z, or Sf — oo for all /', hence SA — oo. Thus SA — SifS> R and equals oo
otherwise. If SA < oo, then S > R, hence Sf = Si < S = SA for / large. Therefore SA is
predictable.
On (S = R) for each / we have supy Sj > Rj. Picky'/ >y'/-i so that

P(5 = R, Sj. < Ri) < I'1.

https://doi.org/10.4153/CMB-1996-018-8 Published online by Cambridge University Press

 146 RICHARD F. BASS

Define
SB = f Sj, if Sj, <Ri 9 SB = M§B SB = mpSB
y oo otherwise *>/ t

\ÎS <R, then for / large, Sy. < S < Rh or Sf = SJr So for / large, Sf = SJi9 and thus
SB = S. If S > R, then for / large, Sj. >R> Ri, so Sf = oo, hence SB = oo. By our choice
of/,-, we have P(S = £, 5f 7^ 00) < 2"*, hence P(S = R, Sf ^ 00) < E£,-2-* - 2~i+\
hence P(S = /Î, SB ^ 00) = 0. Thus SB equals 5 if S < R and equals 00 otherwise. That
SB is a predictable stopping time is proved just as for SA.
Since S' = SA ASB and the minimum of two predictable stopping times is predictable,
the assertion is proved. •
REMARK 3.4. Given predictable stopping times R\, R2,...,Ri, and S, iterating
Lemma 3.3 shows that if Sf is defined to be 00 if S = Rt for some i and equal to S
otherwise, then S' is predictable.
LEMMA 3.5. Suppose T is a stopping time. There exist stopping times U and V\,
Vi,... such that U is totally inaccessible, each V[ is predictable, the finite parts of the
graphs of the V\ are disjoint, and thefinite part of the graph ofT is contained in the union
of the graphs of U and the Vt.
PROOF. Let T{ = T and

fx = sup{P(S = T\ < 00) : S predictable}.

Choose Si predictable so that P(5i = T{ < 00) > f/2 and let V\ = S\. Define T2 to
equal T\ if V\ ^ T\ and to equal 00 otherwise.
We continue by induction. We let^J = sup{P(S = 7} < 00) : S predictable}. If
f = 0, we stop; if not, we choose St predictable so that P(5/ = Tt < 00) >f/2. We use
Remark 3.4 to let Vt equal St on the set where St is not equal to any of V\, V2,..., Vt-i
and equal to 00 otherwise. We let Ti+\ equal 7/ on the set where Vt ^ T\ and otherwise
set it equal to 00.
Note the finite parts of the graphs of the V[ are disjoint by construction. Also, the sets
(7}+i 7^ 7/, Ti < 00) are disjoint. Then

J^f/2 < E P ( i ; + i ^ Ti9 Tt < 00) < 1,

oxf —+ 0. The Ti increase, and we let U = lim, 7}. U must be totally inaccessible: if
not, there exists S predictable such that P(C7 = S) > 0. Then P(U = S)>f for some /,
contradicting the fact that we chose Si at the i-th stage, not S. u
REMARK 3.6. Suppose S\, S2,... are predictable stopping times. Set V\ — S\ and
use Remark 3.4 to let Vi equal Si on the set where 5/ is not equal to any of V\, F2,..., K/_ 1
and equal to 00 otherwise. Then the Vt are predictable stopping times, the union of the
finite parts of their graphs is the same as the union of the finite parts of the graphs of the
Si, and the finite parts of the graphs of the Vi are disjoint.

https://doi.org/10.4153/CMB-1996-018-8 Published online by Cambridge University Press

 DOOB-MEYER DECOMPOSITION 147

LEMMA 3.7. Suppose Zt is a supermartingale such that ZQ = 0, the paths ofZt

are right continuous with left limits, Zt is bounded by N and constant after time v, and
furthermore

(3.1) E[AZ5 | ft-] = 0, a.s.

whenever S is a predictable stopping time. Let W(8) be defined as in (2.3). Then W(6) —> 0
in probability as 6 —> 0.
PROOF. The proof is a modification of the proof of Lemma 2.6. If 7} is the z'-th time
that Zt jumps more than b, then provided we can show that

P( sup EfAZ^. l(t<Ti<t+è) | %] > a/K) -» 0

as S —» 0, the remainder of the proof is exactly as in Lemma 2.6.

Let e > 0. We decompose 7} into stopping times £/and V\9 Vi,... as in Lemma 3.5.

(3.2) P( s u p E f A Z t / l ^ ^ ) | 7A > a/K)

< P( supPO < U < t + 6 | # ) > a/2^7V) - ^ 0

just as in the proof of Lemma 2.6.

Fix t, 6 for the moment, and let Vj = (tV Vj) A (t + 5). Then

(3.3) E[AZF, 1(,<^<^) | %\ = E[AZ^ | J,]

= E[E[AZ^ | ^ _ ] | 7\ = 0, 7=1,2,...

Combining (3.2) and (3.3) and using the boundedness of |Z|, we get our result. •
PROPOSITION 3.8. Suppose Zt satisfies the conditions of Theorem 3.1 and in addition
(3.1) holds. Then the conclusion of Theorem 3.1 holds.
PROOF. We follow the proof of Theorem 2.1, using Lemma 3.7 in place of
Lemma 2.6. •
REMARK 3.9. Let S\,..., S„ be a sequence of predictable stopping times such that
the finite parts of their graphs are disjoint. Let V\ = S\ A • • • A S„,

S*u j . = min{Si : 1 < i <n,i^ I'I , . . . , i)},

and
Vj = max{S!lt„Jj_l:ii<---<ij-l}.
It is not hard to check that Vj(u) is they'-th smallest of the Si(u), so V\ < • • • < V„.
Since the maximum and minimum of a finite number of predictable stopping times is a
predictable stopping time, then each Vt is a predictable stopping time. And notice that
the finite parts of the graphs of the Vj are still disjoint and their union is the same as the
union of the finite parts of the graphs of the S,-.

https://doi.org/10.4153/CMB-1996-018-8 Published online by Cambridge University Press

 148 RICHARD F. BASS

PROOF OF THEOREM 3.1. Let Tnj be they-th time \AZt\ is in (2~' l ,2"' l+1 ]. Using
Lemma 3.5 we decompose each Tnj- into predictable and totally inaccessible parts. Let
us relabel the collection of such stopping times, n an integer,y a positive integer, by Si,
i — 1, 2,..., so that each S\ is either totally inaccessible or predictable, for each / there
exists a bt such that bt < | AZ^. | < 26/, and the set of jump times of Zt is contained in the
union of the graphs of the 5/. By means of Remark 3.6 we may assume without loss of
generality that the finite parts of the graphs of the 5/ are all disjoint.
Let Z 0 (0 = Zt. MSi is totally inaccessible, let Zi+\(t) — Z((t) and ,4/(0 = 0. If S/ is
predictable, let
Ai(t) = -E[AZSl | ^ - ] l ( s , < o , 2»i(0 = Zi(t) + Ai(t).
We will show (a) each^/(r) is increasing, (b) each Z\(i) is a supermartingale, and (c)
E Ej=i Ajipo) < C, where C is a constant not depending on /. Once we show these three
facts, the proof is quick. In view of (a) and (c), Tfi=\ Ai(t) converges, uniformly in t, a.s.,
as / —> oo. Call the limit Aœ(t). By (c) and Fatou's lemma, Aœ(oo) will be integrable. It
is easy to see that each .4/(f) is predictable, hence so is^oo(0- Let
Zoo = Zt + Aoo(t) = limZ/(0.

It follows from (b) and the uniform convergence of £{=i At{t) that ZQO(0 is a supermartin-
gale. From our construction, each Z/ will have paths that are right continuous with left
limits; using the uniform convergence of £/=i Ai(t), we see that Z ^ will, too. Because of
(c), ZQO will be of class D. By our construction of the St, E[AZOQ(T) \ fj-} — 0 for all pre-
dictable stopping times T. By applying Proposition 3.8 to Zoo, we get Z^t) = Mt —AR(t).
Setting At = Aoo(t) +AR{i) then completes the proof.
We show (a), (b), and (c) by induction. Let us start with (a). There is nothing to prove
when Si is totally inaccessible. If Si is a predictable stopping time, let Sy be stopping
times predicting 5/. Z/ is uniformly integrable by the induction hypothesis (c). Using (b)
and the martingale convergence theorem,
E[AZi(Si) | J 5/ _] = limE[AZ,(S,-) | fal
n
But for each n,
E[AZi(Si) | TsJ = limElZiiSi) - Zt(Sim) \ 7sin]
m

= HmE[E[Z;(Si-) - Z,(5/m) | ïsjïs,»] < 0.

Next we look at (b). To show Z,+i is a supermartingale, it suffices to show that
(3.4) EZ,+1(C/,)>EZ,+1(C/2)
whenever U\ and Ui are stopping times with U\ <Ui. If Sy are stopping times predicting
Sh
EZl(Ux)-EZi{U2)- EZ,(t/,)-EZ,((C/,V Sij) AU2)

EZ/((t/, V Sy) A U2) - EZ,-((l/, V Sd A U2)

EZ,(([/i VSi)AU2)-EZi(U2)],

https://doi.org/10.4153/CMB-1996-018-8 Published online by Cambridge University Press

 DOOB-MEYER DECOMPOSITION 149

and each of the summaries on the right is nonnegative. Lettingy —> oo,

EZi(Ui) - EZ,(f/2) - [EZi(((Ui V 5/) A C/ 2 )-) - EZ^IA V 5/) A U2)] > 0,

which is (3.4).
Lastly we look at (c). We need to get a bound on

E£'E[-AZ5. | 7sr\ = - E £ ' A ^ ,

where £ ' denotes the sum over/s such that Sj is predictable andy < /. Since we have a
finite sum, let us use Remark 3.9 and relabel the stopping times so that S\ < S2 < • • •
on the set where they are finite. Let Sjm be stopping times predicting Sj. Since we have a
finite sum and Z is a supermartingale,

-E£ , AZ$ = l i m £ , E [ Z ^ - ^ ]
< lim^fEtZ^v^, - ZSj] + E[Z^ - Z W l ]
+ E[Z 5/ -Z oo ] + E[Z 0 -Z 5 l j]
= E[Z0 - Zoo],

which is bounded by a constant independent of /. •

A supermartingale is said to be regular if EZs„ —» EZ$ whenever Sn | S.

COROLLARY 3.8. At is continuous if and only ifZ is regular.

PROOF. Clearly, if Zt = Mt — At is the decomposition of Z and At is continuous,

then Z is regular. On the other hand, suppose At has a jump of size b > 0 with positive
probability and let S = inf{/ : AAt > b). Since At is predictable, it is easy to see that S
is predictable. Let Sn be stopping times predicting S. Then by monotone convergence,

-EZSn = EASn -> EAS- <EAS= -EZ 5 ,

or Z is not regular. •

REFERENCES

[B] R. F. Bass, Probabilistic Techniques in Analysis, New York, Springer, 1995.

[DM1] C. Dellacherie and R-A. Meyer, Probabilités et Potentiel, Vol. 7, Paris, Hermann, 1975.
[DM2] , Probabilités et Potentiel: Théorie des Martingales, Vol. 2, Paris, Hermann, 1980.
[D] J .L. Doob, Stochastic Processes, New York, Wiley, 1953.
[DD] C. Doléans-Dade, Existence du processus croissant naturel associé à un potentiel de la classe (D),
Zeit. Wahrschein. 9(1968), 309-314.
[IW] N. Ikeda and S. Watanabe, Stochastic Differential Equations and Diffusion Processes, North Holland
Kodansha, Amsterdam, 1981.
[Ml] P.-A. Meyer, A decomposition theorem for supermartingales, Illinois J. Math. 6(1962), 193-205.
[M2] , Decomposition of supermartingales: the uniqueness theorem, Illinois J. Math. 7(1963), 1—17.

https://doi.org/10.4153/CMB-1996-018-8 Published online by Cambridge University Press

 150 RICHARD F. BASS

[P] P. Protter, Stochastic Integration and Differential Equations, New York, Springer, 1990.
[R] K. M. Rao, On decomposition theorems of Meyer, Math. Scand. 24(1969), 66-78.

Department of Mathematics
University of Washington
Seattle, Washington 98195
U.S.A.

https://doi.org/10.4153/CMB-1996-018-8 Published online by Cambridge University Press