The Importance of Being
Biased
Irit Dinur
S. Safra
(some slides borrowed from Dana Moshkovitz)
VERTEX-COVER
Instance: an undirected graph G=(V,E).
Problem: find a set CV of minimal size
s.t. for any (u,v)E, either uC or vC.
Example:
©S.Safra
Minimum VC NP-hard
Observation: Let G=(V,E) be an undirected
graph. The complement V\C of a vertexcover C is an independent-set of G.
Proof: Two vertices outside a vertex-cover
cannot be connected by an edge.
©S.Safra
VC Approximation Algorithm
©S.Safra
C
E’ E
while E’
do let (u,v) be an arbitrary edge of E’
C C {u,v}
remove from E’ every edge incident
to either u or v.
return C.
Demo
©S.Safra
Polynomial Time
C
O(n2)
E’ E
while E’ do
let (u,v) be an arbitrary edge of E’
C C {u,v}
O(n2)
remove from E’ every edge incident to
either u or v
return C
©S.Safra
O(1)
O(n)
Correctness
The set of vertices our algorithm returns is
clearly a vertex-cover, since we iterate until
every edge is covered.
©S.Safra
How Good an Approximation is it?
Observe the set of edges our algorithm chooses
no common vertices! any VC contains 1 in each
©S.Safra
our VC contains both, hence at most twice as large
How well can VC be
Approximated?
Upper bound
A little better (w/hard work) : 2-o(1)
Hardness results
Previously: 7/6
Thm: NP-hard to approximate to
within 105-21 1.36 (> 4/3)
Conjecture: NP-hard to within 2-
>0
©S.Safra
(m,r)-co-partite Graph G=(MR, E)
Comprise m=|M| cliques of size r=|R|:
E {(<i,j1>, <i,j2>) | iM, j1≠j2 R}
m
©S.Safra
h-Clique
Gap Independent-Set
Instance: an (m,r)-co-partite
graph G=(MR, E)
Problem: distinguish between
Good: IS(G) = m
Bad: every set I V s.t. |I|>
m
m contains an edge
h-Clique
m
Thm: hIS(r,
IS( r,h,)) is NP-hard as long as r ( h1 / )c
for some constant c
, r and h constant!
©S.Safra
Hardness of Vertex-Cover
Problem: the size of G’s Vertex-Cover is
Good: (1-1/r) |G|
Bad: (1- /r) |G|
Resulting in a factor smaller than 1+1/r
We show: A reduction from hIS(G) to a graph H
Good:
1
IS(H) 3 H o(1) H
hIS(G) m IS(H) 1 H o(1) H
9
IS(G) m
Bad:
implying NP-hardness of 4/3 factor for
Vertex-Cover
©S.Safra
Encode I.S.’s Representatives
supposedly encoding IS’s representative
jR
IS
assignment:
Edges:
two vertices
can’t
both
if inof
the
ISbe 1 in
Replace clique iM bythat
a 1set
vertices,
the
any Apply
encoding
of an IS
0
if
out
1 for each bit of some binary-code
of R
of
G
long-code
m
©S.Safra
Long-Code of R
©S.Safra
One bit (vertex) for every subset of R
Long-Code of R
One bit (vertex) for every subset of R
to encode an element eR
0
©S.Safra
0
1
1
1
Long-Code to Co-partite’s
what edges do weI.S.
have within a part?
non-intersecting:
F1F2 =
VLC = M P[R]
m
©S.Safra
ELC = {(F1,F2) | F1 F2 E}
Problem: all F, |F| >½r are IS
In
Between
each part:
parts:
intersecting
assume a co-matching
m
©S.Safra
Weighted Graphs
Assign weights to V - hence G = (V, E, )
Consider a probability distribution :V[0,1]
and let the size of a set of vertices be
(S) Pr v S
hence
v
(v)
v S
IS(G) max (I)
i.s.IV
©S.Safra
Easily reducible to graphs with no weights
Biased Long-Code
Consider the p-biased product
distribution p:
Def: The probability of a subset F
pR F p F (1 p) R\F
and for a family of subsets
pR Pr F pR F
R
F p
©S.Safra
F
p <½- F‘s of size >½r Vanish
discriminating against large subsets
solves the >½ problem,
however…
m
©S.Safra
Problem: consistent large subsets
almost all subsets
have a representative in
those subsets
m
Si
Sj
©S.Safra
what if any pair of
cliques i & j have a pair
of large subsets Si & Sj
that are all-wise
consistent
The (m’,r’)-co-partite Graph GB
Fix a large lT and l=r·2lT
V
B ,
m' B
l
R' { a : [l] {T, F } | a 1 (T ) lT }
m
©S.Safra
m’
The (m’,r’)-co-partite Graph GB
Fix a large lT and
l=r·2l
V T
B ,
m' B
l
R' { a : B {T, F } | a 1 (T ) lT }
m
©S.Safra
m’
The (m’,r’)-co-partite Graph GB
Vertices:
Fix a large lT and l=r·2lT
let B=V(l), m’ =|B|
For every BB
Prop: IS(G) = m
IS(GB) > m’ (1-2–(lT))
RB a : B T,F | a 1 (T) lT
Edges:
Let B’ = V(l-1): B1=B’{v1}, B2=B’{v2}
(a1, a2) EB for a1RB1, a2RB2 if
©S.Safra
a1|B’ a2|B’
or
(v1, v2)E and a1(v1) = a2(v2) = T
Now Apply Long-Code to GB
The final graph H = (B P[ RB ], EBLC, )
Vertices: one B B and a subset F P[RB]
Edges: EBLC (F1, F2) for FProof:
P[R
] if
1 P[Rgiven
B1], F2an
ISB2in
F1 F2 EGB , I, consider the
B
corresponding set of
Weights: (F) = p(F) / |B|
singletons in H; take
monotone extension
Prop (Completeness):
IS(H) p · IS(GB) / m’
Thm (Soundness): For p≤(3-5)/2,
hIS(G) < m IS(H) < P + ’
©S.Safra
[for p 1/3: P=p2]
IS of size P even in Bad Case
Partition V into V1 and V2
For every block B, let
©S.Safra
a1 assign T to V1 and F to V2
a2 assign T to V2 and F to V1
and let
B = { F {a1, a2} }
These B‘s form an IS of weight p2 in H
Erdös-Ko-Rado
Def: A family of subsets
P[R] is t-intersecting if for every
F1, F2 , |F1 F2| t
p (P) = max p (Ai,2 )
i
Thm[Wilson,Frankl,Ahlswede-Khachatrian]:
For a t-intersecting ,
where
©S.Safra
p ( ) max p (Ai,t )
i
Ai,t F | F 1,...,2i t i t
Corollary: p() > P is not 2-intersecting
Soundness Proof
Important Observation:
Assume I is a maximal IS in H
I’s intersection with any block
I[B] I P[ RB ]
is monotone and intersecting
It follows:
©S.Safra
q(I[B]) is a non-decreasing function of q
Soundness Proof
We prove: If H has an IS I s.t. (I) > P + 500 then
hIS(G) > m
Let
B[I] = { B | p(I[B]) > P + 250 }
Prop:
|B[I]| > 250 |B|
Observation:
©S.Safra
V
B'
l
1
B B' V \ B'
Soundness Proof
(Naïve) Plan:
Find, for every B B [I], a
distinguished block-assignment aB
Let
VB’ ={ v | B’{v} B [I] and
aB’{v}(v)=T}
©S.Safra
There must be B’ V(l-1) s.t.
|VB’| > 124m
Are I[B]’s juntas?
Long-Code’s Junta
Def: A family of subsets P[R] is Cdecided if membership of F in is
decided according to FC
©S.Safra
P[R] is C-decided to within if
there exists a C-decided ’ so that
( ’)
We refer to C as the (q, )-core of
Influence and Sensitivity
The influence of an element e R on a family
P[R], according to q is
influenceeq () PrR F {e} F \ {e}
Fq
The average-sensitivity of is the sum of
element’s influences:
asqR ( ) influence eq ( )
eR
©S.Safra
Friedgut’s Lemma
Thm[Friedgut]: A Family of subsets
P[R] of average-sensitivity
k = asq() is C-decided to within , where |C|
2kO(k/)
Namely,
has a (q, )-core C R of size
|C| 2O(k/)
©S.Safra
Thm [Margulis-Russo]:
For monotone
asq ( )
d q ( )
dq
Hence
Lemma:
For monotone
> 0, q[p, p+] s.t. asq() 1/
Proof: Otherwise p+() > 1
©S.Safra
Now Comes the Hard Part
Hence I[B] has low, 1/, average-sensitivity
with regards to q
Which, for any , implies a small (q, )-core CB
Let the core-family
CF B
©S.Safra
F P CB | PrF'RB\CB F F' I B 3
4
q
Thus CF[B] is of size > P
hence there exist aB and Fь, F# CF[B] s.t.
FьF# ={aB}
aB is the distinguished block-assignment of B
The (m’,r’)-co-partite Graph GB
Fix a large lT and
l=r·2l
V T
B ,
m' B
l
R' { a : B {T, F } | a 1 (T ) lT }
m
©S.Safra
m’
Now Comes the Harder Part
©S.Safra
Assuming CB is preserved with respect
to B’
if I[B] were exactly the extensions of
CF[B]
Let’s show that if there is an h-clique Q
in VB’, I would not have been an IS
Apply Sunflower construction, PigeonHole-Principle, to find two blocks with
‘same’ Fь, F#
Sunflower Lemma [Erdös-Rado]
Every family of subsets of a domain U of
large enough size has a subfamily ’ s.t.
each element of U either
©S.Safra
Belongs to no subset F’
Belongs to 1 subset F’
Belongs to all subset F’
G, GB and H
For some q [p, p+]
m
©S.Safra
m’
VB’
Assume VB’ contains an h-clique Q
m
©S.Safra
B’
RB’
m’
VB’
Apply Sunflower lemma and PHP
partial-views
on B’
m’
To obtain a kernel K and two blocks
B1 and B2 of Q whose restriction to
partial-views of B’ is same on K and
disjoint outside K
©S.Safra
Yet Harder
©S.Safra
Given an h-Clique Q in VB’:
Let eCB be the set of partial-views of B of
non-negligible (>2–O(|C|)) influence
Redefine VB’ ={ v | B’{v} B [I] and
aB’{v}(v)=T and
eCB’{v} preserved on B’}
Prop: VB’ still large!
Apply Sunflower construction on eC’s, PigeonHole-Principle on C, Fь, F#, to find two blocks
with ‘same’ Fь, F#
Non-negligible Partial-Views
Extended-Core {a | influencea > 2–O(|C|) }
m
©S.Safra
m’
Non-negligible Partial-Views
m
©S.Safra
B’
m’
Taken Care of Kernel
Fь1 and F#2 disagree on K
partial-views
on B’
m’
Let us redefine
VB’ = { v | B’{v} B [I] and
aB’{v}(v)=T and
eCB preserved on B’}
©S.Safra
Almost There
©S.Safra
Assume an h-clique Q of VB’
Consider the projection of eCB on B’ for all
BQ
Apply the Sunflower lemma to obtain Q’ (a set
of blocks whose eC’s form a Sunflower)
These eC’s are thus disjoint outside the
Sunflower’s kernel K
Q’ being large enough, by PHP it must contain
two blocks B1 and B2 with ‘same’ C, Fь, F#
An Edge between I[B1] and I[B2]
Extend Fь within I[B1] and F# within I[B2] so
as not to agree on any a’ in RB’
Not on C
Not on C’s “spouses”
©S.Safra
Fь disagrees with F# except for the distinguished partialview
which is assigned T in both blocks
Make the extension in each block avoid the other’s
spouses; as all spouses have low influence, this changes
little the size of the extension, leaving it bounded away
from ½
Now show outside C and spouses, there exist two
extensions that disagree
An Edge between I[B1] and I[B2]
©S.Safra
As long as q is so that 1-q(1-q)
Open Problems
©S.Safra
Conj: Vertex-Cover is hard to
approximate to within 2-o(1)
Conj: Coloring a 3-Colorable graph with
>O(1) colors is hard
Free Bit Complexity
Max-Cut
Property-Testing
Max-Bisection
Fourier Transform
Consider all ‘linear’ functions, one for each
character S[n]
S (x) 1
Sx
Given a function
f : 1, 1 1, 1
n
Let the Fourier coefficients of f be
f(S) : E f(x) S (x)
©S.Safra
x
Simple Observations
f E f(x)
Claim:
Let the influence function be
x
fi x
fx f x i
2
the Fourier coefficients of f be
fi Let
x
1
1
f(S) S x f(S) S x S i
2 S
2 S
©S.Safra
Simple Observations
Claim:
Let the influence function beiS
Let the Fourier coefficients of f be
influencei f f(S)
asi f f(S) S
©S.Safra
S
© Copyright 2025 Paperzz