The law of the iterated logarithm in game-theoretic...
Transcript of The law of the iterated logarithm in game-theoretic...
![Page 1: The law of the iterated logarithm in game-theoretic ...kenshi.miyabe.name/wordpress/wp-content/uploads/2012/10/gtp-lil.pdf · The law of the iterated logarithm in game-theoretic ...](https://reader031.fdocuments.net/reader031/viewer/2022022712/5c04663809d3f2093a8ba9d7/html5/thumbnails/1.jpg)
The law of the iterated logarithm in game-theoretic
probabilityFourth Workshop on Game-Theoretic Probability
and Related Topics12 -14 Nov 2012 The University of Tokyo
Kenshi MiyabeResearch Institute for Mathematical Sciences,
Kyoto University
This is a joint work with A. Takemura.
1
![Page 2: The law of the iterated logarithm in game-theoretic ...kenshi.miyabe.name/wordpress/wp-content/uploads/2012/10/gtp-lil.pdf · The law of the iterated logarithm in game-theoretic ...](https://reader031.fdocuments.net/reader031/viewer/2022022712/5c04663809d3f2093a8ba9d7/html5/thumbnails/2.jpg)
Overview
Brief history of the LIL
The main result and its related results
The proof
2
![Page 3: The law of the iterated logarithm in game-theoretic ...kenshi.miyabe.name/wordpress/wp-content/uploads/2012/10/gtp-lil.pdf · The law of the iterated logarithm in game-theoretic ...](https://reader031.fdocuments.net/reader031/viewer/2022022712/5c04663809d3f2093a8ba9d7/html5/thumbnails/3.jpg)
Brief history of the LIL
3
![Page 4: The law of the iterated logarithm in game-theoretic ...kenshi.miyabe.name/wordpress/wp-content/uploads/2012/10/gtp-lil.pdf · The law of the iterated logarithm in game-theoretic ...](https://reader031.fdocuments.net/reader031/viewer/2022022712/5c04663809d3f2093a8ba9d7/html5/thumbnails/4.jpg)
The law of the iterated logarithm
Theorem (Khintchine 1924)
Let {Xn} be the sequence of i.i.d. random variables such
that P (X1 = 0) = P (X1 = 1) =
12 . Then
lim sup
n!1
Pnk=1 Xk � n
2p2n ln lnn
=
1
2
almost surely.
Actually, this equation holds for each Schnorr random set.
4
![Page 5: The law of the iterated logarithm in game-theoretic ...kenshi.miyabe.name/wordpress/wp-content/uploads/2012/10/gtp-lil.pdf · The law of the iterated logarithm in game-theoretic ...](https://reader031.fdocuments.net/reader031/viewer/2022022712/5c04663809d3f2093a8ba9d7/html5/thumbnails/5.jpg)
The Kolmogorov LILTheorem (Kolmogorov 1929)
Let {Xn} be the sequence of independent random variables
with EXn = 0 and EX
2n < 1. Put An =
Pnk=1 EX
2k and
Sn =
Pnk=1 Xk. If An ! 1 and there exists a sequence
{cn} such that
|Xn| cn = o(
rAn
ln lnAn) a.s.,
then
lim sup
n!1
Snp2An ln lnAn
= 1 a.s.
5
![Page 6: The law of the iterated logarithm in game-theoretic ...kenshi.miyabe.name/wordpress/wp-content/uploads/2012/10/gtp-lil.pdf · The law of the iterated logarithm in game-theoretic ...](https://reader031.fdocuments.net/reader031/viewer/2022022712/5c04663809d3f2093a8ba9d7/html5/thumbnails/6.jpg)
The Hartman-Wintner LIL
Theorem (Hartman-Wintner 1941)
Let {Xn} be the sequence of i.i.d. random variables with
EX1 = 0 and EX21 < 1. Then
lim supn!1
Snp2An ln lnAn
= 1 a.s.
Strassen (1966) showed that the converse also holds.
6
![Page 7: The law of the iterated logarithm in game-theoretic ...kenshi.miyabe.name/wordpress/wp-content/uploads/2012/10/gtp-lil.pdf · The law of the iterated logarithm in game-theoretic ...](https://reader031.fdocuments.net/reader031/viewer/2022022712/5c04663809d3f2093a8ba9d7/html5/thumbnails/7.jpg)
Predictably Unbounded Forecasting
Players: Forecaster, Skeptic, Reality
Protocol:
K0 := 1.
FOR n = 1, 2, . . .:
Forecaster announces mn 2 R, cn � 0, and vn � 0.
Skeptic announces Mn 2 R and Vn 2 R.Reality announces xn 2 R such that |xn �mn| cn.
Kn := Kn�1 +Mn(xn �mn) + Vn((xn �mn)2 � vn).
Collateral Duties: Skeptic must keep Kn non-negative.
Reality must keep Kn from tending to infinity.
7
![Page 8: The law of the iterated logarithm in game-theoretic ...kenshi.miyabe.name/wordpress/wp-content/uploads/2012/10/gtp-lil.pdf · The law of the iterated logarithm in game-theoretic ...](https://reader031.fdocuments.net/reader031/viewer/2022022712/5c04663809d3f2093a8ba9d7/html5/thumbnails/8.jpg)
The GTP counterpart
Theorem (Shafer and Vovk 2001)
In the predictably unbounded forecasting protocol, Skeptic
can force
An ! 1 & cn = o
rAn
ln lnAn
!!
=) lim sup
n!1
Pni=1(xi �mi)p2An ln lnAn
= 1.
8
![Page 9: The law of the iterated logarithm in game-theoretic ...kenshi.miyabe.name/wordpress/wp-content/uploads/2012/10/gtp-lil.pdf · The law of the iterated logarithm in game-theoretic ...](https://reader031.fdocuments.net/reader031/viewer/2022022712/5c04663809d3f2093a8ba9d7/html5/thumbnails/9.jpg)
Question
Q. How to express the HW LIL in GTP?
The importance:(1) How to express i.i.d. in GTP?(2) How much can we weaken the condition of the LIL?
9
![Page 10: The law of the iterated logarithm in game-theoretic ...kenshi.miyabe.name/wordpress/wp-content/uploads/2012/10/gtp-lil.pdf · The law of the iterated logarithm in game-theoretic ...](https://reader031.fdocuments.net/reader031/viewer/2022022712/5c04663809d3f2093a8ba9d7/html5/thumbnails/10.jpg)
The main result
10
![Page 11: The law of the iterated logarithm in game-theoretic ...kenshi.miyabe.name/wordpress/wp-content/uploads/2012/10/gtp-lil.pdf · The law of the iterated logarithm in game-theoretic ...](https://reader031.fdocuments.net/reader031/viewer/2022022712/5c04663809d3f2093a8ba9d7/html5/thumbnails/11.jpg)
THE UNBOUNDED FORECASTING GAME WITH
QUADRATIC AND STRONGER HEDGES (UFQSH)
Players: Forecaster, Skeptic, Reality
Protocol:
K0 := 1.
FOR n = 1, 2, . . .:
Forecaster announces mn 2 R, vn � and wn � 0.
Skeptic announces Mn 2 R, Vn 2 R and Wn � 0.
Reality announces xn 2 R.Kn := Kn�1 +Mn(xn �mn) + Vn((xn �mn)2 � vn)
+Wn(h(xn �mn)� wn).
With the usual collateral duties.
11
![Page 12: The law of the iterated logarithm in game-theoretic ...kenshi.miyabe.name/wordpress/wp-content/uploads/2012/10/gtp-lil.pdf · The law of the iterated logarithm in game-theoretic ...](https://reader031.fdocuments.net/reader031/viewer/2022022712/5c04663809d3f2093a8ba9d7/html5/thumbnails/12.jpg)
Assumption
1. h is an even function.
2. h 2 C
2and h(0) = h
0(0) = h
00(0) = 0.
3. h
00(x) is strictly increasing, unbounded and concave
(upward convex) for x � 0.
Example
The function h(x) = |x|↵ for 2 < ↵ 3 satisfies the assump-
tion above.
12
![Page 13: The law of the iterated logarithm in game-theoretic ...kenshi.miyabe.name/wordpress/wp-content/uploads/2012/10/gtp-lil.pdf · The law of the iterated logarithm in game-theoretic ...](https://reader031.fdocuments.net/reader031/viewer/2022022712/5c04663809d3f2093a8ba9d7/html5/thumbnails/13.jpg)
The main theorem
Theorem (Miyabe and Takemura)
In UFQSH with h satisfying the assumption, Skeptic can
force An ! 1 and
X
n
wn
h(bn)< 1
!
) lim supn!1
Sn �Pn
i=1 mip2An ln lnAn
= 1
where bn =q
Anln lnAn
.
13
![Page 14: The law of the iterated logarithm in game-theoretic ...kenshi.miyabe.name/wordpress/wp-content/uploads/2012/10/gtp-lil.pdf · The law of the iterated logarithm in game-theoretic ...](https://reader031.fdocuments.net/reader031/viewer/2022022712/5c04663809d3f2093a8ba9d7/html5/thumbnails/14.jpg)
Corollary
Let h be an extra hedge satisfying the assumption and
X
n
1
h(pn/ ln lnn)
< 1.
In UFQSH with this h and mn ⌘ m, vn ⌘ v and wn ⌘ w,
the following are equivalent for m0 2 R and v0 � 0.
1. m0 = m and v0 = v.
2. Skeptic can force
lim supn!1
Sn �m0np2n ln lnn
=pv0. (1)
3. Reality can comply with (1).
14
![Page 15: The law of the iterated logarithm in game-theoretic ...kenshi.miyabe.name/wordpress/wp-content/uploads/2012/10/gtp-lil.pdf · The law of the iterated logarithm in game-theoretic ...](https://reader031.fdocuments.net/reader031/viewer/2022022712/5c04663809d3f2093a8ba9d7/html5/thumbnails/15.jpg)
• Marcinkiewicz and Zygmund (1937) constructed a se-
quence of independent random variables for which
An ! 1 and |Xn| = O(
pAn/ ln lnAn) and which
does not obey the LIL.
• There are many su�cient conditions for the LIL.
• Egorov’s su�cient condition (1984) is
Pnk=1 X
2k
An! 1 a.s. and
nX
k=1
EX2kI(|Xk| >
✏An
ln lnAn) = o(An)
for any ✏ > 0.
15
![Page 16: The law of the iterated logarithm in game-theoretic ...kenshi.miyabe.name/wordpress/wp-content/uploads/2012/10/gtp-lil.pdf · The law of the iterated logarithm in game-theoretic ...](https://reader031.fdocuments.net/reader031/viewer/2022022712/5c04663809d3f2093a8ba9d7/html5/thumbnails/16.jpg)
The related results
Takazawa (2012,201?) showed a weaker upper bound with double hedges.
Miyabe and Takemura (2012) showed a necessary and sufficient condition for the SLLN.
16
![Page 17: The law of the iterated logarithm in game-theoretic ...kenshi.miyabe.name/wordpress/wp-content/uploads/2012/10/gtp-lil.pdf · The law of the iterated logarithm in game-theoretic ...](https://reader031.fdocuments.net/reader031/viewer/2022022712/5c04663809d3f2093a8ba9d7/html5/thumbnails/17.jpg)
The proof
17
![Page 18: The law of the iterated logarithm in game-theoretic ...kenshi.miyabe.name/wordpress/wp-content/uploads/2012/10/gtp-lil.pdf · The law of the iterated logarithm in game-theoretic ...](https://reader031.fdocuments.net/reader031/viewer/2022022712/5c04663809d3f2093a8ba9d7/html5/thumbnails/18.jpg)
Proposition
In UFQSH with h satisfying the assumption, Skeptic can
force An ! 1 and
X
n
wn
h(bn)< 1
!
=) lim supn!1
Snp2An ln lnAn
1.
Lemma
In UFQSH with h satisfying the assumption, Skeptic can
force
An ! 1 andX
n
wn
h(bn)< 1 =) |xn| = o(bn)
18
![Page 19: The law of the iterated logarithm in game-theoretic ...kenshi.miyabe.name/wordpress/wp-content/uploads/2012/10/gtp-lil.pdf · The law of the iterated logarithm in game-theoretic ...](https://reader031.fdocuments.net/reader031/viewer/2022022712/5c04663809d3f2093a8ba9d7/html5/thumbnails/19.jpg)
5.3: THE SHARPNESS OF THE ITERATED-LOGARITHM BOUND 111
Value of martingale
Region A
Region B
Region C
Fig. 5.1 Proving the large-deviation inequality by reducing an initial payoff. Here, , and .
To implement our plan, we shall need the following simple auxiliary result: Suppose weare given some value of . Then the value of for which (5.35) attains its maximum is
(5.36)
and the corresponding value of the payoff (5.35) is
(5.37)
In this proof, we call this value of optimal.Let be a constant, small even compared to . Our basic payoff will be optimal with
respect to the value
and will therefore correspond to
(5.38)
Let us check that the bottom of the redundant region C lies below the thick line. For this andfor
taken from Shafer and Vovk 200119
![Page 20: The law of the iterated logarithm in game-theoretic ...kenshi.miyabe.name/wordpress/wp-content/uploads/2012/10/gtp-lil.pdf · The law of the iterated logarithm in game-theoretic ...](https://reader031.fdocuments.net/reader031/viewer/2022022712/5c04663809d3f2093a8ba9d7/html5/thumbnails/20.jpg)
In Shafer and Vovk (2001), they showed that, as long as
Reality is required to ensure that
|xn| �
rC
ln lnC
,
for all n,
L(s)exp(S(s) � 2
2 C)
(lnC)
8�
for every situation s 2 ⌧ where ⌧ := min{n | An � C} by
considering the strategy
Li+1 = Li1 + xi + (1� �)
2x
2i /2
1 + (1� �)
2vi/2
.
20
![Page 21: The law of the iterated logarithm in game-theoretic ...kenshi.miyabe.name/wordpress/wp-content/uploads/2012/10/gtp-lil.pdf · The law of the iterated logarithm in game-theoretic ...](https://reader031.fdocuments.net/reader031/viewer/2022022712/5c04663809d3f2093a8ba9d7/html5/thumbnails/21.jpg)
Define stopping time ⌧1, ⌧2, ⌧3 by
⌧1 = min {n | · · · } ,⌧2 = min{n | An � C},
⌧3 = min
(n | |xn| > �
rC
ln lnC
).
Lemma
In UFQSH with h satisfying the assumption, there exists a
martingale L such that L(⇤) = 1
Ln
exp(Sn �
2C/2)
(lnC)4�
for n such that n = ⌧2 < ⌧1, ⌧3.
21
![Page 22: The law of the iterated logarithm in game-theoretic ...kenshi.miyabe.name/wordpress/wp-content/uploads/2012/10/gtp-lil.pdf · The law of the iterated logarithm in game-theoretic ...](https://reader031.fdocuments.net/reader031/viewer/2022022712/5c04663809d3f2093a8ba9d7/html5/thumbnails/22.jpg)
A martingale that satisfies the property is defined by
Li
= Li�1
1 + x
i
+
2x
2i
2 � h(xi)h(�1)
1 +
2vi2 � wi
h(�1)
for all i. .
22
![Page 23: The law of the iterated logarithm in game-theoretic ...kenshi.miyabe.name/wordpress/wp-content/uploads/2012/10/gtp-lil.pdf · The law of the iterated logarithm in game-theoretic ...](https://reader031.fdocuments.net/reader031/viewer/2022022712/5c04663809d3f2093a8ba9d7/html5/thumbnails/23.jpg)
Summary
23
![Page 24: The law of the iterated logarithm in game-theoretic ...kenshi.miyabe.name/wordpress/wp-content/uploads/2012/10/gtp-lil.pdf · The law of the iterated logarithm in game-theoretic ...](https://reader031.fdocuments.net/reader031/viewer/2022022712/5c04663809d3f2093a8ba9d7/html5/thumbnails/24.jpg)
The question again
How do we express i.i.d. in GTP?
An idea is to add stronger hedges.
This idea seems to work well to some extentbut not completely.
24