This preview shows pages 1–10. Sign up to view the full content.
This preview has intentionally blurred sections. Sign up to view the full version.
View Full DocumentThis preview has intentionally blurred sections. Sign up to view the full version.
View Full DocumentThis preview has intentionally blurred sections. Sign up to view the full version.
View Full DocumentThis preview has intentionally blurred sections. Sign up to view the full version.
View Full DocumentThis preview has intentionally blurred sections. Sign up to view the full version.
View Full Document
Unformatted text preview: 43(4 1:) 45m hataﬂom W M )zéwvs? [aH‘EV [P MA Amte aﬁawba 62% W5» ear» We 6W6: («ﬁfkmwmf ' hm (cow’s (WWWMJ 0“) ‘éwm g‘WsV ‘0‘ col/«>ke' XAC‘J [6% me #30? x3 61v? ‘ “ok a” (/7, (ALAQ 5} Hot,» {mad he >(CWONVXQASQ 4 he“ Qad’x 37va W («41 a‘ggug‘e') Mvg (5 0‘ V1]
'C’N‘WW: 6% a 5.. “Mfg j; l \‘WYbOMMﬁQjY \‘S
f 4 indeponwa (91$ evm*§. DEE. 1:; & (wk Canaan» AHAQHIALQZ‘
wz La»; $1M, MM. we! mam?
MMW “1
I W(‘A¢3\A41\w aAgg): lg: P(A~) fmwmﬁ “3:31;: '6? “We (5
HAMMW, a}: 3M6} {Ly/Lire wig! MMMM‘I ¥¥’”AgmNé ' PfLﬂfi;ﬁﬂiﬂﬂﬂfL
(“5&2 ‘
WW “N
EYWMQWQ Tame, jaw ECL/dclwzm Q/aCV‘ 9—f_
WWC‘A c9 QMW AKth 5m be “53% or R
ﬁuc‘ \miePMaU/wh’) as, lbw. o M
‘ v Clﬂl‘AVQ/v‘ .4312 MSW”?
7;; ,SL «M a? 2L 7:
Q 3 EL MM [9 Ma$r\" We, C F/V‘g’ m gawk.“ Chc‘wxw 4’00»! 3&4;
L15 skow Aﬂg \WOQZP', M (g) Cam/“14%)
I 1. Aﬁj’ﬂin‘r (“00.2013 /p A .0 p ‘L A n“. C’Wb‘aé) I'd W ’PEV‘H S’=I4M,Mb,ccc/ «EC/Each;
/ ‘ 56“be («5,645 , H
Ai:§,zh‘1’ex‘f.evig aj)§(,\ow SA ; “WW
Ejﬁfbgl® W A] arm} : IN Amgk) M n W Wk) \gA—Drr
L\ A A7/ / A“ £39. emu/J1 in 10M
«6’ l } 62v A' WA W2 W AC
CQWWKM 6°? M V C 5 &\CA ‘Aok CDCQ‘N(>' QC‘AC‘Z AL :@I if ‘21 (WAN 4\ / W (PM R370. ‘TarV \P((\K\>_ K" K‘?(K£JAL)
5/ \eitP(Ac> 70,b~gk:j:f' I ﬂ) [ avﬁleaa/cﬁv“ W‘bn’“ ? I4()’47)A?>):/P{Al:). Lem AWN/3‘“ new/W43 WI}L M6“ “‘2
jvéw G M “J (A A 4/
c M #6 t 4, f L
/ ghG
am“ MUN—I rélPM‘L)>o
PM  gag “Wow? UM WWW)
”” r /.—~———/' Mum) 4??? WM. U, M EX?“ “u 69 / M all
ice
the
zed
l is
are [so ige
)er lﬂS H Ml
.r tic
we
en :nt
II'S S a
161' .nd whs LECTURE 8 The Lovész Local Lemma The Lemma. Let A1,    , An be events in a probability space. In combinatorial
applications the ‘A, are “bad” events. We wish to show Pr [All] > 0 so that there is a point (coloring, tournament, conﬁguration) x which is good. The basic
probabilistic method of Lecture 1 may be written: Counting sieve. If X Pr [A‘] < 1, then Pr [AA] > O _
There are other simple conditions that ensure Pr [AAi] > 0. Independence sieve. If A1 , ‘  v , A" are mutually independent and all Pr [Ai] < 1,
then Pr [/\A,] > O. The Lovész Local Lemma is a sieve method which allows for some dependence
among the A;. A graph G on vertices [n] (the indices for the Ag) is called a
dependency graph for A1,   ‘ , A,1 if for all i A is mutually independent of all
Aj with {i,j} E G. (That is, A. is independent of any Boolean function ofthese Aj.) LovAsz LOCAL LEMMA (Symmetric case). LetA; ,  ‘ ~ , A" be events with depen
dency graph G such that Pr [A]: p for all i, de'g (i): d for all i and 4dp < 1.
Then Pr [AA] > 0. Proof We show by induction on s that if [SI 5 s, then for any i Pr [Ai For S =Q this is immediate. Renumber for convenience so that i = n, S '=
{1,   ~ , s} and {i,x}£ G for x>d. Now /\ S 2p.
jeS Pr [An/i1 ‘ ' ’ A—di/A—dkl ‘ ' ‘ As]
PrLAl ' ’ ' AdlAd+l ‘ ' ' As] ' 57 “[AniAl . ‘ I 14—5] 58 LECTURE 8 We bound the numerator
Fri/Mil ‘   All“, ' ‘ ﬁJSPrMnl/LH   A]
= Pr [An] Sp as A" is mutually independent of Ad“, ~   , AS. We bound the denominator d
PITA) ' ' ‘ AdlAd+1 ' ‘ ‘ A3121” 2 Pr[AilAd+l ' ' ' As]
i=1 (1
_>.1— 2 2p (Induction) Hence we have the quotient
Pr [AA/Tl ~   KslSp/%=2p, completing the induction. Finally
Pr[/§1“'5n]=ll Pr[5il51"‘/§s~i]ZH (1—2P)>0~ D
i=1 i=1 The proof is so elementary that it could, and I think» should, be taught in a ﬁrst course in probability. It has had and continues to have a profound eﬁect on
probabilistic methods. The diagonal Ramsey function. A lower bound for R(k, k), our ﬁrst use of the
probabilistic method in Lecture 1, provides a Simple application of the Lovész
Loeal Lemma. Consider a random two—coloring of K,1 with AS the event “S is
monochromatic,” S ranging over the k—setsof vertices. Deﬁne G by placing
{5, T}E G if and only if [SH T122. Then AS is mutually independent of all AT
with T not adjacent to G, since the AT give information only about edges outside
of S. Hence G is a dependency graph. (When Isn TI =2 the events A5, A7 are
independent; note however that mutual independence from a family of AT is far
stronger than pairwise independence with each AT. Recall the old chestnut: I
have two children and at least one is a girl. What is the probability they are both
girls. Conditional on the younger being a girl it is one half. Conditional on the
older being a girl it is one half. Conditional on the disjunction it is one third.
But I digress.) We apply the Lovasz Local Lemma with p = Pr [AS] = 21%;) and d={T: Sﬂ Tl22}s(k>( n 2 k~2 k n k
4 2‘“(2)<
<2)(k—2) 1’
then R(k, k) > n. The asymptotics are somewhat disappointing. COROLLARY. If ttor t in a
:ct on )f the
ovasz
“S is
acing
[11 AT
itside
T are
is far
tut: I
both
n the
third.
and “w... W “WA... a . THE LOVASZ LOCAL LEMMA COROLLARY. R(k, k)>—e@k2k/2(1+o(1)). This improves the lower bound given in Lecture 1 by a factor of 2 and the
improvement via the deletion method in Lecture 2 (which, oddly, was only
published after the better bound) by a factor of x/Z The gap between the upper
and lower bounds has not really been effectively decreased. The lower bound of
Erdos was found in April 1946 (published in 1947) and progress on this diﬂicult
problem has been slow. The van der Waerden function. Here the improvement is more impressive Color
[n] randomly. For each arithmetic progression S of size k let As he the event
that S is monochromatic. Let S, T be adjacent in G if they intersect. (In all our
applications the probability space will be a random Coloring of some set (2. For
Ramsey’s Theorem (I was the edge set of K". Events will be not adjacent if they
deal with the coloring on disjoint sets.) Now 172—2‘4‘ and d5 nk as one pro
gression intersects (exercise) at most nk others. Hence we have the following
theorem. THEOREM. If 4nk21rk< 1, then W(k)> n. That is, W(k)>2k/8lc This greatly improves the bound W( k) > 2”2 of Lecture 1. Still we must in all honesty mention that W(p)2p2”, for p prime, has been shown by completely
constructive means! Algorithm? The Lovasz Local Lemma proves the existence of an x satisfying
/\ 15,» even when Pr [/\ Kijrnay be exponentially small We have seen in Lecture
4 that when X Pr [Ai]<1 there often is an algorithm to ﬁnd a speciﬁc “good” x. Open Problem. Can the Lovasz Local Lemma be implemented by a Good
Algorithm? ' Let us be more speciﬁc. Suppose 51, ~ ‘  , SnC[n] with all 15;]: 10 and all
deg (j) : 10. Two color [21] randomly and let A be the event that S, is monochro
matic. Let i, i’ be adjacent in the dependency graph if and only if their correspond—
ing sets intersect. Each S, is intersected by at most 90 other Sir. We apply the
Lovasz Local Lemma with p = Pr [A] = 2~9 and. d r: 90. As 4dp <1, Pr [A [if] > O
and so there is a two~coloring X for which no S,» is monochromatic. Is there a
polynomial (in n) time algorithm for ﬁnding such a coloring? Notice that the Lovasz Local Lemma here guarantees the existence of a “needle
in a haystack." If, say, S], ‘   , Sn/w are disjoint a random X is good with
probability at most (1 ~2”9)”/1°. Can we actually ﬁnd this exponentially small
needle in polynomial time? Addendum: Joke. Here we give an ironic demonstration of the power of the
Lovasz Local Lemma. THEOREM. Let S, T be ﬁnite sets with T128IS[. Then there exists a function
f: S —> T which is injective. 60 LECTURE 8 Proof. Let f be a random function. For each {x, y}C S let Axy be the event
f(x) =f(y). Then Pr [AW]=1/T=p. Let {x, y} and {u, v} be adjacent in the
dependency graph if and only if they intersect. Then the maximal degree in the
dependency graph is d = 2([Sl — 1). As 4dp < 1, Fr [A/lxy]> 0 and so there exists
an f for which A” for all x, y. That is, f is injective. When IT! 2 365 and I WI = 23 the “birthday problem” says that f has probability
less than % of being injective. When lTI = 8S the probability of a random f being
injective is exponentially small. The Counting Sieve proves the existence of an
injective f only when (I?!) < Cl Anti van der Waerden Here is the original use of the Lovasz Local Lemma. THEOREM. Let k, m satisfy 4m(m —1)(1 —1/k)"'< 1. Let Sc R with Isl 2 m,
Then there exists a kvcoloring x: R —> [k] so that every translate S+t is k—colored.
That is, for all te T and 15 is k there exists s e S with x(s+ t) = 1'. Without use of the Lovasz Local Lemma no proof is known that gives the
existence of an m = m(k) with this property. Notice a fundamental difference
between the translation and homothety groups. Gallai’s Theorem, a consequence
of van der Waerden’s Theorem, states that for all ﬁnite S and all ﬁnite colorings
of R there is a monochromatic S’ = aS+ t. Proof First we let BC R be an arbitrary ﬁnite set and k—color B so that all
S+ t C B have all k colors. Color B randomly. For each I such that S+ t C B let
A, be the event that S+t does not have all k colors. Then i Pr [AJS k(1« l/k)m =p. Let t, t’ be adjacent in the dependency graph if and only if S+t and S+t’
intersect. With given t this occurs only if t': +s—s" for distinct s, s'e S, so
that the dependency graph has degree at most d : m(m — 1). The conditions on
k, m are precisely that 4dp < 1. The Lovasz Local Lemma applies and /\ A, ¢ Q;
there is a kcoloring of B for which all translates of S lying in B have all k
colors. Compactness. To color all of R we need the Compactness Principle. We state
this in a form convenient for us. COMPACTNESS PRINCIPLE. Let Q be an inﬁnite set, k a positive integer, and let
U be afamily ofpairs (B, x) where B C Q is ﬁnite, x: B——> [k] such that (i) U is closed under restriction. That is, if (B, x) e U and B’C B then
(BI,XIB')€ U; (ii) For all B some (B,X) E U.
Then there exists A40“) [k] such that (B,X[B)e U for allﬁnite BC U. Proof Let X be the topological space of all XtQ—9[k]. Here we consider [k]
discrete and X has the usual product topology. That is, a basis for the open sets. is given by the sets {X2 x(bi)= ai, IS is s} over all s, b1, ~ ~  , b5, a1,   , as. For
every ﬁnite B let X3 be the set ofX e X with (B, 143) e U. By (ii) ,XB ¢ @. Splitting
XEX according to xlB gives a ﬁnite (IBlk) partition of X into sets both open am
prc Sin prc
all on
me
inv
be i all We as He ...
View
Full
Document
 Spring '08
 Staff
 Statistics, Probability

Click to edit the document details