Threshold Behavior of Random Structures -Advances and Challanges
Gil KalaiHebrew University of Jerusalem
Random Structures and Algorithm number 16!,Poznan, August 2013
Gil Kalai Thresholds
Erdos and Renyi
The critical probability for connectivity for G (n, p) is log n/n thethreshold interval for connectivity is of magnitude 1/n.
Gil Kalai Thresholds
Erdos and Renyi
Gil Kalai Thresholds
Bollobas and Thomason
For every ε > 0 there exists k such that for every monotone(graph) property A, if the probability that G ∈ G (n, p) satisfies Ais at least ε then the probability that G ∈ G (n, kp) satisfies A is atleast 1− ε.
In other words, for fixed ε > 0 the length of the threshold intervalis bounded by a constant times the critical probability.
Gil Kalai Thresholds
Heppy birthday dear Bela!!!
Gil Kalai Thresholds
Heppy birthday dear Bela!!!
Gil Kalai Thresholds
Four threshold men
Gil Kalai Thresholds
This lectureThis lecture is about advances and challenges in the abstract studyof threshold behavior of stochasic systems, and relations withdiscrete isoperimetry and discrete harmonic analysis. The abstractstudy of threshold behavior is a small but interesting fragment ofthe wide theory of threshold behavior for random graphs andrelated stochastic properties.
I will start by mentioning one major theorem (Friedgut 1999) andone conjecture (Kahn and Kalai 2006).
Gil Kalai Thresholds
Introduction: A theorem and a conjecture
Gil Kalai Thresholds
Friedgut’s theorem
Definition: A B-local monotone property of graphs is the propertyof containing a graph among a finite family of graphs of boundedsize B.Example: To contain a K5 is a local property.
Notation: For a monotone property of graphs P denote by µp(P)the probability that a random graph in G (n, p) satisfies P.
Friedgut’s theorem (1999): For every ε and C , there is B suchthat for every monotone property P of graphs, if
dµp(P)/dp < pC ,
then P is ε-close to a B-local monotone property Q.
Gil Kalai Thresholds
A striking consequence of Friedgut’s theorem
Coarse threshold occurs only at threshold functions of the form nβ
where β is rational.
Corollary: Connectivity has a sharp thrshold. (Because thethreshold function is log n/n.)
Gil Kalai Thresholds
Threshold and Expectation threshold
Consider a random graph G in G (n, p) and the graph property: Gcontains a copy of a specific graph H. (Note: H depends on n; amotivating example: H is a Hamiltonian cycle.) Let q be theminimal value for which the expected number of copies of H ′ in Gis at least 1/2 for every subgraph H ′ of H. Let p be the value forwhich the probability that G contains a copy of H is 1/2.
Conjecture: [Kahn, Kalai 2006]
p/q = O(log n).
The conjecture can be vastly extended to general Booleanfunctions, and we will hint on possible connection with harmonicanalysis and discrete isoperimetry. (Sneak preview: it will require afar-reaching extension of results by Friedgut, Bourgain andHatami.)
Gil Kalai Thresholds
Part II: Boolean functions and influences
Gil Kalai Thresholds
The discrete n-dimensional cube and Boolean functions
The discrete n-dimensional cube Ωn is the set of 0-1 vectors oflength n.
A Boolean function f is a map from Ωn to 0, 1.
A boolean function f is monotone if f cannot decrease when youswitch a coordinate from 0 to 1.
Gil Kalai Thresholds
The Bernoulli measure
Let p, 0 < p < 1, be a real number. The probability measure µp isthe product probability distribution whose marginals are given byµp(xk = 1) = p. Let f : Ωn → 0, 1 be a Boolean function.
µp(f ) =∑x∈Ωn
µp(x)f (x) = µpx : f (x) = 1.
Gil Kalai Thresholds
The total influence
Two vectors in Ωn are neighbors if they differ in one coordinate.
For x ∈ Ωn let h(x) be the number of neighbors y of x such thatf (y) 6= f (x).The total influence of f is defined by
I p(f ) =∑x∈Ωn
µp(x)h(x).
If p = 1/2 we will omit p as a subscript or superscript.
Gil Kalai Thresholds
Russo’s lemma
Russo’s lemma: For a monotone Boolean function f ,
dµp(f )/dp = I p(f ).
Very useful in percolation theory and other areas.
The threshold interval for a monotone Boolean function f isthose values of p so that µp(f ) is bounded away from 0 and 1.(Say 0.01 ≤ µp(f ) ≤ 0.99.)
A typical application of Russo’s lemma: If for every value p in thethreshold interval I p(f ) is large, then the threshold interval itself isshort. This is called a sharp threshold phenomenon.
Gil Kalai Thresholds
(A version of) Harper’s theorem
Harper’s theorem: If µp(f ) = t then
I p(f ) ≥ 2t · logp t.
There is a 3 line proof by induction.Harmonic analysis proof: without the log(1/t) factor it followsfrom Parseval.
Gil Kalai Thresholds
Influence of variables on Boolean functions
Let
σk(x1, . . . , xk−1, xk , xk+1, . . . , xn) = (x1, . . . , xk−1, 1−xk , xk+1, . . . , xn).
The influence of the kth variable on a Boolean function f isdefined by:
I pk (f ) = µp(x ∈ Ωn, f (x) 6= f (σk(x))).
Gil Kalai Thresholds
KKL’s theorem
Theorem (Kahn, K, Linial, 1988; Bourgain Katznelson KKL 1992;Talagrand 1994 Friedgut Kalai 1996) There exist a variable k suchthat
Ik(f ) ≥ Cµ(f )(1− µ(f )) log n/n.
A sharp version (due to Talagrand)∑I pk (f )/ log(1/I p
k (f )) ≥ C (p)µp(f )(1− µp(f )).
Gil Kalai Thresholds
Discrete Fourier analysis
We assume now p = 1/2. Let f : Ωn → R be a real function.
Let f =∑
f (S)WS be the Fourier-Walsh expansion of f .
HereWS(x1, x2, . . . , xn) = (−1)
Pxi :i∈S.
Gil Kalai Thresholds
Hypercontructivity and Harper’s theorem:
We assume now p = 1/2. f =∑
f (S)WS is the Fourier-Walshexpansion of f . Key ideas:
0 Parseval gives I (f ) = 4∑
f 2(S)|S |.1 Bonami-Gross-Beckner hypercontractive inequality.
||∑
f (S)(1/2)|S | || 2 ≤ || f || 5/4.
2 For Boolean functions the qth power of the q norm is themeasure of the support and does not depend on q. If thesupport is small this means that the q-norm is very differentfrom the r -norm if r 6= q.
(See also : Ledoux’ book on concentration of measure phenomena)
Gil Kalai Thresholds
Part III: Thresholds: Baker Dozen’s challenges
1. Threshold and symmetry. The threshold window for“containing a clique of size log n” is of size 1/ log2 n. Thisseems to be the largest threshold for graph properties. Wheredoes the extra log n come from?
2. Connectivity; small ps. Understand the sharp thresholdbehavior of connectivity as part of a general framework. Moregenerally, understand threshold intervals when p is small.
3. Hunting sharp thresholds. Prove sharp-threshold behaviorfor various monotone properties of random graphs.
4. Uniform vs. non-uniform models. Understand the thresholdbehavior for the 3-SAT problem, 3-colorability of graphs, etc..Relate to 0-1 laws.
Gil Kalai Thresholds
5 More isoperimetry. Margulis and Talagrand extended theErdos-Renyi result on connectivity in a different direction,relying on a different isoperimetric result. Combine this studywith the study of influence and discrete Fourier analysis.
6 General conditions for sharp threshold. What is anecessary and sufficient condition for a sequence of Booleanfunctions to have the sharp threshold property ?
7 Locating the threshold. Can one sometimes find and provethe location of the threshold based on sharp thresholdbehavior ?
8 Larger alphabets, etc.. Study larger alphabets, products ofother groups, products of graphs, analogous settings for thesymmetric group.
9 Positional games and other games. Extend theinfluence/sharp-threshold framework to positional games ongraphs and other games.
10 Models of statistical physics. Apply the influence/shresholdtheory to percolation and other models of statistical physics.
Gil Kalai Thresholds
11 Polynomially-small threshold intervals. Are there generalconditions that will guarantee that the total influence is atleast nα for some α > 0? At most n1/2−α for some α > 0?
12 Phase transition phenomena. Relations between influence,isoperimetry and discrete Fourier analysis and other phasetransition phenomena for random graph such as theemergence of giant component.
13 Non-product distributions. Such as the equal-slicedistribution, distributions arising in the Potts model, FKGdistributions.
14 Graph parameters rather than properties Studydistributions, concentration phenomena, etc. for parametersof random graphs.
15 Continuous settings. Extend the notions and the results tofunctions with continuous domain and range. Importantexamples: Gaussian spaces, Gelfand pairs,...
Gil Kalai Thresholds
Part IV: Threshold behavior and symmetry
Gil Kalai Thresholds
Invariance under transitive group
Theorem: If a monotone Boolean function f with n variables isinvariant under a transitive group of permutation of the variables,then
I p(f ) ≥ Cµp(f )(1− µp(f )) log n.
Proof: Follows from KKL’s theorem (extended to general Bernoulimeasures) since all individual influences are the same.
Gil Kalai Thresholds
Total influence under symmetry of primitive groups
For a transitive group of permutations Γ ⊂ Sn, let I (Γ) be theminimum influence for a Γ-invariant function Boolean functionwith n variables.Theorem: [Bourgain and Kalai 1998] If Γ is primitive then one ofthe following possibilities hold.
I
I (Γ) = θ(√
n),
I
(logn)(k+1)/k−o(1) ≤ I (Γ) ≤ C (log n)(k+1)/k ,
I I (Γ) behaves like (log n)µ(n), where µ(n) ≤ log log n isgrowing in an arbitrary way.
Gil Kalai Thresholds
Jumps in the behavior of I (Γ) for primitive groups Γ
If Γ is not An and Sn then I (Γ) ≤ (log n)2.
If I (Γ) ≤ (log n)1.99 then I (Γ) ≤ (log n)3/2
If I (Γ) ≤ (log n)3/2−ε then I (Γ) ≤ (log n)4/3
If I (Γ) ≤ (log n)4/3−ε then I (Γ) ≤ (log n)5/4
...
If I (Γ) ≤ (log n)1+o(1) then I (Γ) ≤ log n · log log n
Gil Kalai Thresholds
Threshold behavior for random graphs
The case that Γ is Sn acting on unordered pairs from[n] = 1, 2, . . . , n describes graph properties. The conclusion isthat the threshold interval for graph properties is at most
1/ log2−o(1) n.
Gil Kalai Thresholds
Hypercontractivity and the lower bounds
Both the upper bounds and the lower bounds depend on findinginvariants of the group which causes the threshold to go abovelog n. Giving constructions for the upper bounds requires a detailedunderstanding of primitive permutation groups based on theclassification theorem and O’Nan-Scott theorem.
The lower bounds are based on delicate and complicated harmonicanalysis.
Step I: hypercontractivity + random restriction argument + cleverinequalities takes you in the graph case from log n to log n3/2.
Step II: Extremely subtle ”bootstrap” to amplify the outcome.
Gil Kalai Thresholds
The Entropy Influence conjecture (Friedgut + Kalai 1996)
If the Fourier-Walsh expansion of f is f =∑
f (S)WS define
E (f ) =∑
f 2(S) log(1/f 2(S)).
Conjecture: For some absolute constant C ,
I (f ) ≥ C · E (f ).
Gil Kalai Thresholds
Scaling-limit symmetry, critical exponents, spectraldistribution,...
Gil Kalai Thresholds
Prelude: A necessary and sufficient condition for o(1)threshold window.
The Shapley value of the kth variable is defined by
ψk(f ) =
∫ 1
0I pk (f )dp.
Theorem: (Kalai 2005) A necessary and sufficient condition fordiminishing threshold window is that the maximum of the Shapleyvalues tends to 0.
Problem 1 : Close the exponential gap in this theorem.
Problem 2 : Find a necessary and sufficient condition for sharpthreshold (when the critical probability is small).
Gil Kalai Thresholds
Part V: Stability for edge-isoperimetry: From Friedgut toHatami and beyond
Gil Kalai Thresholds
Low Influence and Juntas
A dictatorship is a Boolean function depending on one variable. AK -junta is a Boolean function depending on K variables.
Theorem: (Friedgut) If p is bounded away from 0 and 1 andI p(f ) < C then f is close to a K (C )-Junta.
This works if log p/ log n = o(1) the most interesting applicationswould be when p is a power of n. There the theorem is not true.
Gil Kalai Thresholds
The works of Friedgut and Bourgain (1999)
Suppose that f is a Boolean function and
I p(f ) < pC ,
thenFriedgut’s theorem (1999) (already mentioned): If f represent amonotone graph property then f is close to a a “locally defined”monotone function g .
Bourgain’s theorem (1999): Unconditionally, f has a substantial“locally defined” ingredient.
Gil Kalai Thresholds
The hunt for sharp thresholds
These results are very useful for proving shaprp threshold forspecific examples. However implementing them often requiresdifficult analysis. Here are a few examples where sharp thresholdswere established.
I K-SAT (Friedgut)
I Colorability (Achlioptas, Friedgut)
I Ramsey properties for graphs (Fridgut Krivelevich; Friedgut,Rodl, Rucinski, Tetali)
I Random sets with a monochromatic arithmetic progression inevery 2-coloring (Friedgut, Han, Person, Schacht)
I k-cores (Shoham Letzter).
Gil Kalai Thresholds
Hatami’s theorem: Pseudo-juntas
Suppose that for every subset of variables S , we have a functionJS : 0, 1S− > 0, 1 which can be viewed as a constraint overthe variables with indices in S . Now there are two conditions:A Boolean function is a K -psudo-junta if(1) the expected number of variables in satisfied constraints isbounded by a constant K .(2) f (x) = f (y) if the variables in satisfied constraints and alsotheir values are the same for x and y .
Hatami’s theorem: For every C there is K (C ), such that if
I p(f ) < pC ,
then f is close to a K (C )-pseudo-junta.
Gil Kalai Thresholds
A conjectural extension of Hatami’s theorem
Conjecture: Suppose that µp(f ) = t and
I (f ) ≤ C log(1/t)t
then f is εt-close to a O(log(1/t))-pseudo-junta.
If true, may apply toward the conjecture on expectation thresholds,and may enable us to use the abstract threshold theory for huntingthresholds’ locations.
Gil Kalai Thresholds
Thank you very much!
Gil Kalai Thresholds