Scribd
Upload
13 views12 pages

Lecture 26: Influence and KKL Theorem

Uploaded by

Fahad Arjun
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
13 views12 pages

Lecture 26: Influence and KKL Theorem

Uploaded by

Fahad Arjun
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 12

Lecture 26: Influence and KKL Theorem

Lecture 26: Influence and KKL Theorem


Influence

Let f (x1 , . . . , xn ) be a boolean function


The influence of xi represents the probability (over random
choice of all other variables) that f is sensitive to the value of
xi
Alternately,

Inf i (f ) := E [f (x) 6= f (x + ei )],


$
x[n]\{i} ←{0,1}n−1

where ei is the element with 1 exactly at the i-th position and


0 everywhere else

Lecture 26: Influence and KKL Theorem


Influence

Let J ⊆ [n] be a subset of indices


Influence of J is represented by:

Inf J (f ) := E [f (x) is not constant]


$ J
xJ ←{0,1}

Lecture 26: Influence and KKL Theorem


Examples

ANDn is the function that outputs the AND of its n inputs.


The Inf i (ANDn ), for any i, is the probability that ANDn is
sensitive to xi . That happens exactly when all others input bits
are 1, i.e. with probability 2−(n−1) .
ORn also has identical influence.
MAJn , for odd n, outputs the majority bit of n input bits. The
Inf i (MAJn ) is exactly the probability that the number of 0s
and 1s in the remaining
 inputsbits is equal. This happens with
n −1
probability 2−(n−1) ≈ n−1/2 .
(n − 1)/2
Note that sensitivity of ANDn is very low; but it cannot be
high because the output of ANDn is constant with ≈ 1
probability. But, when the output of the function is nearly
balanced does some variable have high influence?

Lecture 26: Influence and KKL Theorem


Example

Note that MAJn has balanced output and its variables have
≈ n−1/2 influence.
In fact, any J with size ≈ n1/2 has constant influence.
Think: Are there balanced functions with lesser influence?
TRIBESs,w is the ORs of ANDw of n = sw input bits. That is,
interpret the input as s blocks of w bits each. Apply ANDw on
each block and output the ORs of the ANDs.
Think: For what values of s and w is TRIBESs,w balanced?
Think: For these values of s and w , what is the influence of
any variable?

Lecture 26: Influence and KKL Theorem


Target Result

Theorem (Kahn-Kalai-Linial)
For every balanced function f , there exists a variable with influence
at least ≈ log n/n.

We will show a result by Talagrand (presented next slide)


Prove the KKL result using that result
Think: This is asymptotically tight!

Lecture 26: Influence and KKL Theorem


Talagrand’s Result
Lemma
Let g be a function with kg k2 6= kg k3/2 , then:

X gb(S)2 2.5 kg k22


6
|S| log kg k2 / kg k3/2
S6=∅

We apply Hypercontractivity with p = 3/2, q = 2 and


ρ2 = 1/2
X X
kg k23/2 > kTρ (g )k22 = gb(S)2 /2|S| > g (S)2 /2k
S S : |S|=k

That is, for any k > 0, we have:


X g (S)2 2k
6 kg k23/2
|S| k
S : |S|=k

Lecture 26: Influence and KKL Theorem


Proof

Therefore, for any m, we have:


X gb(S)2 X X gb(S)2 X gb(S)2
= +
|S| |S| |S|
S6=∅ 16k6m S : |S|=k S : |S|>m
 
X 2k
 kg k2 +
X gb(S)2
6 3/2
k (m + 1)
16k6m S : |S|>m
 
X 2k 1
6  kg k2 + kg k22
3/2
k (m + 1)
16k6m

Choose largest m such that 2m kg k23/2 6 kg k22 . Using the


maximality property, we have: (m + 1) > 2 log kg k2 / kg k3/2

Lecture 26: Influence and KKL Theorem


Proof

By induction we can prove the following upper bound:


X 2k 2 · 2m+1
6
k (m + 1)
16k6m

So, overall we have:


X gb(S)2 4 · 2m 1
6 kg k23/2 + kg k22
|S| (m + 1) (m + 1)
S6=∅
(4 + 1) 5
6 kg k22 6 kg k22
(m + 1) 2 log kg k2 / kg k3/2

This gives the overall bound of the lemma

Lecture 26: Influence and KKL Theorem


Using Talagrand’s Result to get KKL Theorem

For i ∈ [n], define gi (x) := f (x) − f (x + ei ) (the ‘−’ sign in


the definition is subtraction over R and ‘+’ sign in the
definition is addition over {0, 1}n )
Note that gbi (S) = 2fb(S), if i ∈ S; otherwise, gbi (S) = 0
Note that: E $ [|gi (x)|] = Inf i (f )
x ←{0,1}n
Since f is a boolean function, gi has output in {−1, −1}
Therefore, for p > 1, we have: kgi kpp = kgi k1 = Inf i (f )
So, kgi k2 = Inf i (f )1/2 and kgi k3/2 = Inf i (f )2/3
Using this, we can deduce:

kgi k22 = Inf i (f )


log kg k2 / kg k3/2 = (1/6) log 1/Inf i (f )

Lecture 26: Influence and KKL Theorem


Proof Continued

Use Talagrand’s Result on gi :


X gbi (S)2 15Inf i (f )
6
|S| log 1/Inf i (f )
S6=∅

Now, let us understand the relation between the left-hand-side


using f ’s Fourier coefficient:

X gbi (S)2 X 4fb(S)2


=
|S| |S|
S6=∅ S : i∈S

Previous two inequalities gives:

X fb(S)2 Inf i (f )
6 (15/4)
|S| log 1/Inf i (f )
S : i∈S

Lecture 26: Influence and KKL Theorem


Proof Continued
Summing the previous inequality over all i ∈ [n], we get:
X Inf i (f ) X X fb(S)2 (∗) X (†)
(15/4) > = fb(S)2 = Var[f ]
log 1/Inf i (f ) |S|
i∈[n] i∈[n] S : i∈S S6=∅

The (∗) equality is because the term fb(S)2 / |S| is considered


once for every i ∈ S, i.e. |S|
P times. The (†) equality is because
Var[f ] = E[f ] − E[f ] = S6=∅ f (S)2 .
2 2 b
A “nearly balanced f ” has Var[f ] = Ω(1)
So, we have:
X Inf i (f )
> Ω(1)
log 1/Inf i (f )
i∈[n]

So, there exists i = i ∗ ∈ [n] such that:


Inf i (f )
> Ω(1/n)
log 1/Inf i (f )
That is Inf i ∗ (f ) > Ω(log n/n) (the KKL Result)
Lecture 26: Influence and KKL Theorem

You might also like