Professional Documents
Culture Documents
Paolo Baldi, Laurent Mazliak, Pierre Priouret - Martingales and Markov Chains - Solved Exercises and Elements of Theory (2002, Chapman and Hall - CRC)
Paolo Baldi, Laurent Mazliak, Pierre Priouret - Martingales and Markov Chains - Solved Exercises and Elements of Theory (2002, Chapman and Hall - CRC)
2) <
< = sup E((M; ~ M,)*)
‘This implies P(V > 0) = 0 and
0.eal CHAPTER 3: MARTINGALES
‘Theorem 3.7 Let (My )nzu be 4 martingale such that sup,»4E(M;) < +00. Then
(Miy)y29 converges a.s, and in L?,
Convergence Theorems
#3.10 One of the reasons for the importance of martingales is the convergence the-
orems; they guarantec that a martingale converges a.s., under a set of assumptions
that are often easy to check. Already Theorem 3.7 states that a bounded martingale
converges a.s.
Let (Xq)nz0 be a submartingale bounded in absolute value by some number K
and let X, = My + A, be its Doob decomposition (Theorem 3.2). Then E(A,) =
E(X1,) — E(Mo) < 2K so that E(Ago) < 2K, which gives Ano < +00 as. We sct
tp = ifm; Angi > pl: tp is a stopping time and |Mz,anl £|Xeypan— Aepanl S
K + p. so that Mis a bounded martingafe and converges a.s. by Theorem 3.7. We
get then the a.s, convergence of (X;")n20 and therefore that (X,)n20 converges on
[tp = +00). Finally (X;,)n20 converges a.s. Since, Ago being as, finite, U% {tp =
+00] = U% {Ace < p) = {Avo < +00) = Rass,
If (Xp)az0 is u positive supermartingale, sinice the exponential function is convex
and increasing, ¥, = e~** is a submartingale (3.3) satisfying 0 < ¥, < 1, Thus
(Y¥n)n20 and (X,)q,29 (with values in R*+) both converge a.s. Moreover by Futou’s
Jemma (21.5)
F(Xeo| Fi) = E( lim Xn| Fe) < lim EX_| Fi) sXe as.
‘This implies that, for every M > 0, E(Xzol poem) = E(XolxocMy)+80 that Xoo <
+00 on [Xo < +00}. We have proved
‘Theorem 3.8 Let (X,,)n20 be a positive supermartingale, Then it converges a.8, 10
anv Xoo and X_, > E(Xo| %,), Moreover, if P(Xq < +00) = 1, thenP(Xx <
+00) = 1,
Of course, a positive martingale isa positive supermartingale and the previous theorem
gives the convergence also in this ease.
Theorem 3.9 Let (X,,)a>0 be a submartingale such that sup, 29 E(|X»l) < +00.
Then (Xu)n20 converges i :
‘The proof simply amounts to showing that (X,)20 can be written as the difference
of a positive integrable martingale and a positive integrable supermartingale and then
applying Theorem 3.8, Let Xt = My + Ay be the Doob decomposition of the
submartingale (X})q20- Then E(Aze) = supysg E(An) < ElMol +sup,»o E|Xnl <
+00. If ¥y = Mn + E(Aco| F,), since Yr > Xf > Xns (¥q nad is a positive
mnartingale and Z, = Yq —X, isa positive supermartingale, Obviously X, = ¥,—Z,,
Note that Theorem 3.8 states that E(¥ao) < +00 and E(Zeo) < +00, 80 that Feo <
+00 and Zao < +90 4.8.
Let us point out that, if (X.)nxo is a submartingale, since |x] = 2x+ — x, it holds
E|Xn| = 2E(X+) — E(X,) 5 2E(X$) — E(Xo). Thus supyso E|X,| < +00 ifCONVERGENCE THEOREMS Fy
and only if sup,syE(X,}) < +00. Similarly, tor a (integrable) supermartingule,
Sup,20 E[X,| < +00 if and only if sup, > E(X,) < +00.
11 We have already seen that a martingale that is bounded in L? converges a.s. anu
in L?, More generally o martingale that is bounded in L?. p > |, converges a.s. an
in LP, Actually itis even bounded in L' and thus, thanks to Theorem 3,9, there exists
ZV. Xoo such that Xq —>n-+09 Xao @8, Moreover, by Doob’s inequality, the rv.
X* = sup,20 [Xn betongs to LP. Thanks to the inequality |X, — Xool? < 2PX*",
we are allowed to apply Lebesgue’s theorem and get
lim, BUX — Xool”) = 0.
Contrary to the cuse L?, p > 1, the condition supyso E|X»| < +00 does not imply
the convergence in L', as we shall see in the exercises (Exercise 3.5, for instance).
#3.12 Let us go back to our first example of a martingale, i.e., X, = E-**(X),X € L!
(see @3,2). This martingale is bounded in L',
E(X pl) = EUE%*(X))) < EE (1X|)) = EUXD.
‘Therefore (Theorem 3.9) Xn —*n-+00 Xoo 4.8. Let us study the convergence in L'
‘Thanks to the decomposition X = X+ — X™, we can assume X > 0. Leta > 0.
‘The Lebesgue theorem gives ||X — X A all; -» Gas. > +00, We just proved that
K(X A a| Fy) converges a.s. but, being bounded, E(X A a |.%y) also converges in
L', It suffices then to write
VEX Fy) ~ BOX | Fwy
S WE(X | Fn) — E(X Aa| Fq), + WE(X Aa | F,) — E(X Aa | Fu) +
HHE(X Aa | Fy) — E(X| Fini =
<2I1X — X vail, + WE(X A.4| Fn) — EX Ac | Fndhy
in order to derive that (E**(X)),>0 is a Cauchy sequence in L' and thus that it con-
verges in L'. Let us identity its limit Xoo. Cleatly Xco is Foo-measurable. Moreover,
by the definition of X,,
[xara [xan torvey eh
a A
from which, since (Xq)q20 converges in L',
[XwaP =f XaP, forevery AG Fy. as)
A A
It is easy to see, by a monotone class argument (#24), that (3.5) holds for every
A Feo = 6 (q20 Fn)- This shows that Xoo = E(X | Foo). Summarizing,
Theorem 3,10 Assume X € L', The martingale X = E(X | Fy) converges a.s. and
in L' t0 E(X | Foo).
43.13 We just mention the following
Theorem 3.11 Let X € L! and (n)nz0 be a decreasing sequence of o-algebras
with Goo = My>0Gn- The sequence X, = E(X |G) then converges a.s. and in L!
10 E(X | Go).30 CHAPTER 3: MARTINGALES
‘The sequence (Xn)nz0 of Theorem 3.11 is called an inverse martingale,
Regular Martingales
£3.14 A martingale of the form X, = E(X | F,). X € L' is said to be regular. We
have just proved that such a martingale converges a. andin L' to Xq = E(X | Foo)
Conversely, let (Xj)n20 be @ martingale converging in L' to a cv. X, Since |Xnl <
IX] + IX — Xl SuPpso ECXnl) < +00. Thus (Theorem 3,9), (X,)n20 converges
as, to X, We then have, for every p > Oand for every A € Fn,
ff Xesvars f xnaP
F i
und, letting p goto +00. since X,, 4 converges in L',
f xaP -f XndP, forevery Ae F,
or equivalently, X,, = E(X | ¥)). In conclusion,
Theorem 3.12 A martingale is regular if and only if it converges in L'.
93.15 Let (X_)q20 be a regular martingale having the &.v. Xao as its limit, Then it
is closed meaning that (X,),,¢4 is a martingale, i.e, verifies the definition for every
m