Noise-Insensitive Noise-Insensitive Boolean-Functions Boolean-Functions
are Juntas are Juntas
Guy Kindler & Muli Safra Guy Kindler & Muli Safra
Slides prepared with help of: Adi Akavia
Slides prepared with help of: Adi Akavia
Influential
Influential People People
The theory of the The theory of the influence influence of variables of variables on Boolean functions
on Boolean functions [BL, KKL] [BL, KKL] and and
related issues, has been introduced to related issues, has been introduced to
tackle
tackle social choice social choice problems, problems,
furthermore has motivated a magnificent furthermore has motivated a magnificent
sequence of works, related to economics sequence of works, related to economics
[K], percolation [BKS], Hardness of [K], percolation [BKS], Hardness of
approximation [DS]
approximation [DS]
Revolving around the
Revolving around the Fourier/Walsh Fourier/Walsh analysis of Boolean functions
analysis of Boolean functions … …
Where to go for Dinner?
Where to go for Dinner?
Who has suggestions:
Who has suggestions:
Each cast their vote in an Each cast their vote in an
(electronic) envelope, and (electronic) envelope, and
have the system decided, have the system decided,
not necessarily according to not necessarily according to
majority…
majority…
It turns out someone –in the It turns out someone –in the
Florida wing- has the power Florida wing- has the power
to flip some votes to flip some votes
Power
Power
Voting Systems Voting Systems
n n agents, each voting either “for” ( agents, each voting either “for” ( T T ) or ) or
“against” (
“against” ( F F ) – a Boolean function over ) – a Boolean function over n n variables
variables f f is the outcome is the outcome
The values of the agents (variables) may The values of the agents (variables) may each, independently, flip with probability each, independently, flip with probability
It turns out It turns out : one cannot design an : one cannot design an f f that that would be robust to such noise -that is,
would be robust to such noise -that is, would, on average, change value w.p.
would, on average, change value w.p. < < O(1) O(1) - - unless taking into account only very few of unless taking into account only very few of
the votes
the votes
Dictatorship Dictatorship
Def Def : a Boolean function : a Boolean function P([n]) P([n]) {-1,1} {-1,1} is a is a monotone
monotone e e - - dictatorships dictatorships --denoted --denoted f f e e -- -- if: if:
e
T e x
f x F e x
e
T e x
f x F e x
Juntas Juntas
Def Def : a Boolean function : a Boolean function f:P([n]) f:P([n]) {-1,1} {-1,1} is a is a j j - - Junta Junta if if J J [n] [n] where where |J|≤ j |J|≤ j , ,
s.t. for every
s.t. for every x x [n] [n] : : f(x) = f(x f(x) = f(x J) J) Def Def : : f f is an is an [ [ , j] , j] - - Junta Junta if if
j- j- Junta Junta f’ f’ s.t. s.t.
Def Def : : f f is an is an [ [ , j, p] , j, p] - - Junta Junta if if
j- j- Junta Junta f’ f’ s.t. s.t.
x~U
nf x f' x
Pr
x~U
nf x f' x
Pr
x~
pf x f' x
Pr
x~
pf x f' x
Pr
We would tend to omit p
Long-Code Long-Code
In the long-code In the long-code L:[n] L:[n] {0,1} {0,1} 2 2 n n each element is each element is encoded by an
encoded by an 2 2 n n -bits -bits
This is the most extensive binary code, having This is the most extensive binary code, having one bit for every subset in
one bit for every subset in P([n]) P([n])
Long-Code Long-Code
Encoding an element Encoding an element e e [n] [n] : :
E E e e legally-encodes legally-encodes an element an element e e if if E E e e = f = f e e
Long-Code
Long-Code Monotone-Dictatorship Monotone-Dictatorship
The truth-table of a Boolean function The truth-table of a Boolean function over
over n n elements, can be considered as a elements, can be considered as a 2 2 n n bits long string (each corresponding bits long string (each corresponding
to one input setting – or a subset of
to one input setting – or a subset of [n] [n] ) ) For a long-code, the legal code-words For a long-code, the legal code-words
are all monotone dictatorships are all monotone dictatorships
How about the Hadamard code?
How about the Hadamard code?
Long-code Tests Long-code Tests
Def Def (a (a long-code test long-code test ): given a code- ): given a code- word
word w w , probe it in a constant number of , probe it in a constant number of entries, and
entries, and
accept w.h.p if accept w.h.p if w w is a monotone dictatorship is a monotone dictatorship
reject w.h.p if reject w.h.p if w w is not close to any is not close to any monotone dictatorship
monotone dictatorship
Efficient Long-code Tests Efficient Long-code Tests
For some applications, it suffices if the test may For some applications, it suffices if the test may
accept illegal code-words, nevertheless, ones accept illegal code-words, nevertheless, ones
which have short list-decoding:
which have short list-decoding:
Def Def (a (a long-code list-test long-code list-test ): given a code-word ): given a code-word w w , , probe it in 2/3 places, and
probe it in 2/3 places, and
accept w.h.p if accept w.h.p if w w is a monotone dictatorship, is a monotone dictatorship,
reject w.h.p if reject w.h.p if w w is not even is not even approximately approximately
determined by a short list of domain elements
determined by a short list of domain elements , that , that is, if
is, if a a Junta Junta J J [n] [n] s.t. s.t. f f is close to is close to f’ f’ and and f’(x)=f’(x
f’(x)=f’(x J) J) for all for all x x
Note Note : a long-code list-test, distinguishes : a long-code list-test, distinguishes between the case
between the case w w is a is a dictatorship dictatorship , to the , to the
Background Background
Thm (Friedgut) Thm (Friedgut) : a Boolean function : a Boolean function f f with small with small average-sensitivity
average-sensitivity is an is an [ [ ,j]- ,j]- junta junta
Thm (Bourgain) Thm (Bourgain) : a Boolean function : a Boolean function f f with small with small high- high- frequency weight
frequency weight is an is an [ [ ,j]- ,j]- junta junta
Thm (Kindler&Safra) Thm (Kindler&Safra) : a Boolean function : a Boolean function f f with small with small high-frequency weight
high-frequency weight in a in a p p - - biased biased measure is an measure is an [ [ ,j]- ,j]- junta junta
Corollary Corollary : a Boolean function : a Boolean function f f with with small small noise- noise- sensitivity
sensitivity is an is an [ [ ,j]- ,j]- junta junta
Parameters Parameters : : average-sensitivity average-sensitivity [BL,KKL,F] [BL,KKL,F]
high-frequency weight
high-frequency weight [KKL,B] [KKL,B]
[n]
x [n]
z II
[n]
[n]
Noise-Sensitivity Noise-Sensitivity
How often does the value of
How often does the value of f f changes changes when the input is perturbed?
when the input is perturbed?
x
z II
Def Def ( ( ,p,x ,p,x [n] [n] ): Let ): Let 0< 0< <1 <1 , and , and x x P([n]) P([n]) . . Then
Then y~ y~ ,p,x ,p,x , if , if y = (x\I) y = (x\I) z z where where
I~ I~ [n] [n] is a is a noise subset noise subset , and , and
z~ z~ p p I I is a is a replacement replacement . .
Def Def ( ( -noise-sensitivity -noise-sensitivity ): let ): let 0< 0< <1 <1 , then , then
[ When
[ When p=½ p=½ equivalent to flipping each equivalent to flipping each
[n] [n]
p ,p,x
x~ ,y~
ns f = Pr f x f y
[n]
[n]
p ,p,x
x~ ,y~
ns f = Pr f x f y
[n] x [n]
z II
Noise-Sensitivity
Noise-Sensitivity
Fourier/Walsh Transform Fourier/Walsh Transform
Write
Write f:{-1, 1} f:{-1, 1} n n {-1, 1} {-1, 1} as a polynomial as a polynomial What would be the monomials?
What would be the monomials?
For every set For every set S S [n] [n] we have a monomial which is the we have a monomial which is the product of all variables in
product of all variables in S S (the only relevant (the only relevant powers are either
powers are either 0 0 or or 1 1 ) )
?????
?????
Make sense now to consider the degree of
Make sense now to consider the degree of f f or or to break it according to the various degrees of to break it according to the various degrees of
the monomials..
the monomials..
High/Low Frequencies and High/Low Frequencies and
their Weights their Weights
Def Def : the : the high-frequency high-frequency portion of portion of f f : :
Def Def : the : the low-frequency low-frequency portion of portion of f f : : Def Def : the : the high-frequency-weight high-frequency-weight is: is:
Def Def : the : the low-frequency-weight low-frequency-weight is: is:
k S
S k
f f S
k S
S k
f f S
k S
S k
f f S
k S
S k
f f S
2
k 2
2 S k
f
f S
2
k 2
2 S k
f
f S
2
k 2
2 S k
f
f S
2
k 2
2 S k
f
f S
Low High-Frequency Weight Low High-Frequency Weight
Prop Prop : the : the -noise-sensitivity can be expressed in Fourier -noise-sensitivity can be expressed in Fourier transform terms as
transform terms as
Prop Prop : Low : Low ns ns Low Low high-freq weight high-freq weight Proof
Proof : By the above proposition, low noise-sensitivity : By the above proposition, low noise-sensitivity implies
implies
nevertheless,
nevertheless, f f being being {-1, 1} {-1, 1} function, by Parseval function, by Parseval formula (that the
formula (that the norm 2 norm 2 of the function and its of the function and its Fourier transform are equal) implies
Fourier transform are equal) implies
( )
S
2( )
S
1 - l f S ~ 1
å ( )
S
2( )
S
1 - l f S ~ 1
å
2
f S 1
f S 1
2
S 2
S
2 ns f =1
1
Sf S
2
S
2 ns f =1
1 f S
Average and Restriction Average and Restriction
Def Def : Let : Let I I [n], [n], x x P([n]\I) P([n]\I) , , the the restriction function restriction function is is
Def Def : the : the average function average function is is
Note Note : :
I
I y P I
A f : P I
A f x E f x y
I
I y P I
A f : P I
A f x E f x y
I I
f x : P I 1,1
f x y f x y
I I
f x : P I 1,1
f x y f x y
I y P I I
A f x E f x y
I y P I I
A f x E f x y
I [n]
x y
I [n]
x y y
y y y
Fourier Expansion Fourier Expansion
Prop Prop : :
Prop???? Prop???? : :
Corollary Corollary : : I S
S I
A f f(S)
I S
S I
A f f(S)
I T S
S I T I S
f x f T x
I T S
S I T I S
f x f T x
x P I S I
E g x g S
x P I S I
E g x g S
Variation Variation
Def Def : the : the variation variation of of f f : :
Prop Prop : the following are equivalent : the following are equivalent definitions to the
definitions to the variation variation of of f f : :
2 2
I I 2
S I
f f A f f S
variation I I 2 2 2
S I
f f A f f S
variation
I f x P I E var f x y y P I I
variation I y P I I
x P I
f E var f x y
variation
Low-freq Variation and Low-freq Variation and
Low-freq Average-Sensitivity Low-freq Average-Sensitivity
Def Def : the : the low-frequency variation low-frequency variation is: is:
Def Def : : the the average sensitivity average sensitivity is is And in Fourier representation:
And in Fourier representation:
Def Def : the : the low-frequency average sensitivity low-frequency average sensitivity is: is:
i
i [n]
f f
as variation i
i [n]
f f
as variation
2
S
f f (S) S as 2
S
f f (S) S as
2
k k
I I
S I S k
f f f S
variation I k variation I k 2
S I S k
f f f S
variation variation
f f f (S) S 2
k k
as k f variation k f f (S) S 2
as variation
Main Result Main Result
Theorem Theorem : :
constant constant >0 >0 s.t. any Boolean function s.t. any Boolean function f:P([n])
f:P([n]) {-1,1} {-1,1} satisfying satisfying is an
is an [ [ ,j]-junta ,j]-junta for for j=O( j=O(
-2-2k k
33
2k2k) ) . . Corollary
Corollary : : fix a
fix a p p -biased distribution -biased distribution
ppover over P([n]) P([n]) . . Let Let >0 >0 be any parameter. be any parameter.
Set Set k=log k=log
1-1-(1/2) (1/2) . . Then
Then constant constant >0 >0 s.t. any Boolean function s.t. any Boolean function f:P([n])
f:P([n]) {-1,1} {-1,1} satisfying satisfying is an
is an [ [ ,j]-junta ,j]-junta for for j=O( j=O(
-2-2k k
33
2k2k) ) . .
k 2 2 2
f
k 22 O k
2f
O k
2ns f
O k
2ns f
O k
Where to go for Dinner?
Where to go for Dinner?
Who has suggestions:
Who has suggestions:
Each cast their vote in an Each cast their vote in an
(electronic) envelope, and (electronic) envelope, and
have the system decided, have the system decided,
not necessarily according to not necessarily according to
majority…
majority…
It turns out someone –in the It turns out someone –in the
Florida wing- has the power Florida wing- has the power
to flip some votes to flip some votes
Power Power
Of course they’ll have to discuss it over
dinner….
First Attempt:
First Attempt:
Following Freidgut’s Proof Following Freidgut’s Proof
Thm Thm : any Boolean function : any Boolean function f f is an is an [ [ ,j]- ,j]- junta for junta for
Proof Proof : :
1. 1. Specify the junta Specify the junta where, let
where, let k=O(as(f)/ k=O(as(f)/ ) ) and fix and fix =2 =2 -O(k) -O(k)
2. 2. Show the complement of Show the complement of J J has small variation has small variation
f /
j = 2 O as O as f /
j = 2
i
J i| variation
i f
J i| variation f
P([n])
Following Freidgut - Cont Following Freidgut - Cont
Lemma Lemma : :
Proof Proof : :
Now, lets bound each argument:
Now, lets bound each argument:
Prop Prop : : Proof
Proof : characters of size : characters of size k k contribute to the contribute to the average-sensitivity
average-sensitivity at least at least (since
(since ) )
k 2 2
f k 2 as f k
2
f as f k
J f 2 variation J f 2
variation
k k 2
J f J f f 2
variation J f variation J k f f k 2 2
variation variation
P([n])
J
k 2
k f
k 22k f
2
2
as f f S S
2
S
as f f S S
k i J
2 2
S O(k) S
i S, S k i S, S k
i J 2 i J r
2 4/r
O(k) O(k)
S S
i S i S
i J r i J 2
2/r 2 O(k)
k J
O(k)
f
f(S) 2 f(S)
2 f(S) 2 f(S)
2 f 2 as f
f
variation vari on ati i
k i J
2 2
S O(k) S
i S, S k i S, S k
i J 2 i J r
2 4/r
O(k) O(k)
S S
i S i S
i J r i J 2
2/r 2 O(k)
k J
O(k)
f
f(S) 2 f(S)
2 f(S) 2 f(S)
2 f 2 as f
f
variation vari on ati i
Following Freidgut - Cont Following Freidgut - Cont
Prop Prop : : Proof Proof : :
k
J f 4 variation J k f 4
variation
we do not know
whether as(f) is small!
True only since this is a {-1,0,1} function.
So we cannot proceed this way with only
this way with only as as
kk! !
If If k k were 1 were 1
Easy case
Easy case (!?!): If we’d have a bound on the non- (!?!): If we’d have a bound on the non- linear weight, we should be done.
linear weight, we should be done.
The linear part is a set of independent The linear part is a set of independent
characters (the singletons) characters (the singletons)
In order for those to hit close to 1 or -1 most of In order for those to hit close to 1 or -1 most of
the time, they must avoid the law of large the time, they must avoid the law of large
numbers, namely be almost entirely placed on numbers, namely be almost entirely placed on
one singleton [by Chernoff like bound]
one singleton [by Chernoff like bound]
Thm[FKN, ext.]: Assume
Thm[FKN, ext.]: Assume f f is close to is close to linear, linear, then
then f f is is close to close to shallow shallow ( ( a constant a constant
How to Deal with Dependency How to Deal with Dependency
between Characters between Characters
Recall Recall
(theorem’s premise) (theorem’s premise)
Idea Idea : Let : Let
Partition Partition [n]\J [n]\J into into I I 1 1 ,…,I ,…,I r r , for , for r >> k r >> k
w.h.p w.h.p f f I I [x] [x] is close to is close to linear linear (low freq (low freq characters intersect
characters intersect I I expectedly by expectedly by 1 1
k 2 k
J f f 2 J f
variation J f f k 2 2 +variation J k f
variation +variation
k 2
2
1
2f
k 22 O k 1
2f
O k
J i| variation
i k f
J i| variation
i kf
P([n])
J I
1I
2I
rI
So what?
So what?
f f I I [x] [x] is close to is close to linear linear
By By FKN FKN f f I I [x] [x] is either a constant-function is either a constant-function or a dictatorship, for any
or a dictatorship, for any x x Still,
Still, f f I I [x] [x] could be a different could be a different dictatorship for every
dictatorship for every x x , hence the , hence the
P([n])
J I
1I
2I
rI
almost linear almost linear
almost shallow almost shallow
Theorem(
Theorem( [FKN] [FKN] ) ) : : global constant global constant M M , , s.t.
s.t. Boolean function Boolean function f f , ,
shallow shallow Boolean function Boolean function g g , s.t. , s.t.
Hence, Hence, ||f ||f I I [x] [x] >1 >1 || || 2 2 is small is small f f I I [x] [x] is is close to
close to shallow shallow ! !
2 1 2
2 2
f g 2 2 M f 1 2
f g M f 2
Dictatorship and its Singleton Dictatorship and its Singleton
Prop Prop : if : if f f I I [x] [x] is a dictatorship, then is a dictatorship, then
coordinate coordinate i i s.t. s.t. (where (where p p is the is the bias).
bias).
Corollary (from [FKN]) Corollary (from [FKN]) : : global constant global constant M M , s.t. , s.t.
Boolean function Boolean function h h , either , either
I
f x i
I p f x i p
h i p h i p
22{1} {2} {i} {n} {1,2} {1,3} {n-1,n} S {1,..,n}
weight
Characters Total weight of no more than
Total weight of no more than 1-p 1-p
f f I I [x] [x] Mostly Constant Mostly Constant
Lemma Lemma : : >0 >0 , s.t. for any , s.t. for any and any and any function
function g:P([m]) g:P([m])
Def Def : Let : Let D D I I be the set of be the set of x x P( P( I I ) ) , s.t. , s.t.
f f I I [x] [x] is a dictatorship is a dictatorship
Next we show, that Next we show, that |D |D I I | | must be small, must be small,
I I
D I x P I : i I,s.t. f x i I p
D x P I : i I,s.t. f x i p
m 1 4k k 2 4 2 k 2 2
x~ Pr g x M g M g
m 1 4k k 2 4 2 k 2 2
x~ Pr g x M g M g
Lemma Lemma : :
Proof Proof : let : let , then , then
|D |D I I | | must be small must be small
[n]
4k k 2
I 1 2 2
x~ Pr x D M M f
[n] I 1 4k 2 k 2 2
x~ Pr x D M M f
i I
g x i f x i I
g x f x i
[n]
i i
I i
x~ i I x P I
4 2
4k k k
1 i I 2 2 i I 2
4 2
1 4k S 2 S
i I S [n], S k,S I i 2 i I S [n], S k,S I i 2
2 2 2
4k k
1 2 2
i I S [n], S k,S I i i I
Pr x D Pr g x p
M g M g
M f S M f S
M f S M f
[n]
i i
I i
x~ i I x P I
4 2
4k k k
1 2 2 2
i I i I
4 2
1 4k S 2 S
i I S [n], S k,S I i 2 i I S [n], S k,S I i 2
2 2 2
4k k
1 2 2
i I S [n], S k,S I i i I
Pr x D Pr g x p
M g M g
M f S M f S
M f S M f
Prev lemma
I
T T [n],T I S
f x S f T
I
TT [n], T I S
f x S f T
Each S is counted only for one index iI. (Otherwise, if S was counted for both i
and j in I, then |SI|>1!)
Parseval
Simple Prop Simple Prop
Prop Prop : let : let {a {a i i } } i i I I be be sub-distribution sub-distribution , that , that is, is, i i I I a a i i 1 1 , , 0 0 a a i i , then , then i i I I a a i i 2 2 max max i i I I {a {a i i } } . .
Proof Proof : : i 2 max max 2 max
i I
a 1 a a a
i 2 max max 2 max
i I
a 1 a a a
1 2 3 max n
a
ino more than no more than 1 1
1
a
i1/a 1/a
maxmax1
|D |D I I | | must be small - Cont must be small - Cont
Therefore Therefore
(since
(since ), ),
Hence Hence
2
2 2 k
i I i I
i I S [n], S k, S [n], S k,
S I i S I i
f S max f S max
f
variation
i
2
2 2 k
i I i I
i I S [n], S k, S [n], S k,
S I i S I i
f S max f S max
f
variation
i
2
S [n], S k,S I i
f S 1
2
S [n], S k,S I i
f S 1
[n]
4k k 2
I 1 2 2
x~ Pr x D M M f
[n] I 1 4k 2 k 2 2
x~ Pr x D M M f
Obtaining the Lemma Obtaining the Lemma
It remains to show that indeed: It remains to show that indeed:
Prop1 Prop1 : :
Prop2 Prop2 : :
I
I f E x~ I f x I variation I f E x~
I variation I f x I variation variation
J1 r
4k k 2
I~ I 2
E
1Jr variation f O f
4k k 2
I~ I 2
E variation f O f
I
2 2
x~ I
S I S I
E f x S f S
I
2 2
x~ I
S I S I
E f x S f S
• Recall Recall
• However However
I
f
x P IE
y P Ivar f x y
I
variation
I
f
x P IE
y P Ivar f x y
I
variation
I
f x
I 0 y P Ivar f x y
I 0
variation
I f x
I
0
y P Ivar f x y
I
0
variation
Obtaining the Lemma – Cont.
Obtaining the Lemma – Cont.
Prop3 Prop3 : :
Proof Proof : separate by freq: : separate by freq:
Small freq: Small freq:
Large freq: Large freq:
Corollary Corollary (from props 2,3): (from props 2,3):
J I
1r
2 k 2
I~ ,x~ 2
S I 1
E f S O f
J I
1r
2 k 2
I~ ,x~ 2
S I 1
E f S O f
J I
1r
2 2 2
I~ ,x~
S k S I 1 I
S k
E f S Pr S I 1 | k r O 1
r
k 1
J I
1r
2 2 2
I~ ,x~
S k S I 1 I
S k
E f S Pr S I 1 | k r O 1
r k 1
J I
1r
2 k 2
I~ ,x~ 2
S I 1 S k
E
f S f
J I
1r
2 k 2
I~ ,x~ 2
S I 1 S k
E
f S f
2 2
2 2
Obtaining the Lemma – Cont.
Obtaining the Lemma – Cont.
Recall: by corollary from [FKN], Recall: by corollary from [FKN], Either
Either or or
Hence Hence
By By Corollary Corollary
Combined with Combined with Prop1 Prop1 we obtain: we obtain:
f i p f i p
n f M f 1 2 2
variation n f M f 1 2 2
variation
J I J I J I
1r 1r 1r
1 2
I I I I
I~ ,x~ I~ ,x~ I~ ,x~ 2
E
f x Pr x D E
M f x
variation
J I J I J I
1r 1r 1r
1 2
I I I I
I~ ,x~ I~ ,x~ I~ ,x~ 2
E
f x Pr x D E
M f x
variation
|D |D
II| | is small is small
J I
1r
2 2
1 k
I~ ,x~ I 2 2
E
J1 I f x
O f
r
2 2
1 k
I~ ,x~ I 2 2
E
f x
O f
J1 r
4k k 2
I~ I 2
E I~
1Jr variation I f O 4k f k 2 2
E variation f O f
Important Lemma Important Lemma
Lemma Lemma : : >0 >0 , s.t. for any , s.t. for any and any and any function
function g:P([m]) g:P([m]) , the following , the following holds:
holds:
m 4k k 2 4 k 2 2
x~ Pr g x O g g
m 4k k 2 4 k 2 2
x~ Pr g x O g g
Low-freq high-freq
Beckner/Nelson/Bonami Inequality Beckner/Nelson/Bonami Inequality
Def Def : let : let T T be the following operator on be the following operator on f f
Thm Thm : for any : for any p≥r p≥r and and ≤((r-1)/(p-1)) ≤((r-1)/(p-1)) ½ ½
Corollary
Corollary : for : for f f s.t. s.t. f f >k >k =0 =0
1 ,p,x
y
f x E f y
T
1 ,p,x
y
f x E f y
T
p r
f f
T r
f p f
T
4 4k 4
4 2
g 4 4 4k g 2 4
g g
Probability Concentration Probability Concentration
Simple Bound Simple Bound : :
Proof Proof : :
Low-freq Bound Low-freq Bound : Let : Let g:P([m]) g:P([m]) be of be of degree
degree k k and and >0 >0 , then , then >0 >0 s.t. s.t.
m t
t
x~ Pr g x g t
t
m t t
x~ Pr g x g
m 4 4k 2 4
x~ Pr g x g
m 4 4k 2 4
x~ Pr g x g
4 4k 4
g 4 4k g 4
g g
Lemma’s Proof Lemma’s Proof
Now, let’s prove the lemma: Now, let’s prove the lemma:
Bounding low and high freq separately: Bounding low and high freq separately:
, ,
simple bound
4 2 2
4 4k k k
2 2
g g
4 4k g k 2 4 2 g k 2 2
m
m m
x~
k k
x~ x~
Pr g x
Pr g x Pr g x
m
m m
x~
k k
x~ x~
Pr g x
Pr g x Pr g x
Low-freq bound
Shallow Function Shallow Function
Def Def : a function : a function f f is is linear linear , if only singletons , if only singletons have non-zero weight
have non-zero weight
Def Def : a function : a function f f is is shallow shallow , if , if f f is either a is either a constant or a dictatorship.
constant or a dictatorship.
Claim Claim : Boolean linear functions are shallow. : Boolean linear functions are shallow.
weight
Character
Boolean Linear
Boolean Linear Shallow Shallow
Claim Claim : Boolean linear functions are : Boolean linear functions are shallow.
shallow.
Proof Proof : let : let f f be Boolean linear function, be Boolean linear function, we next show:
we next show:
1. 1. {i {i o o } } s.t. s.t.
( ( i.e. i.e. ) )
2. 2. And conclude, that either And conclude, that either or or i.e. i.e. f f is shallow is shallow
0
S , i ,f S 0
S , i ,f S
0 0
f f f f f i f i 0 0 i i
00
f f
f f f f i f f i 0 0 i i
00Claim 1 Claim 1
Claim 1 Claim 1 : let : let f f be boolean linear function, be boolean linear function, then
then {i {i o o } } s.t. s.t.
Proof Proof : w.l.o.g assume : w.l.o.g assume
for any for any z z {3,…,n} {3,…,n} , consider , consider
x x 00 00 =z =z , , x x 10 10 =z =z {1} {1} , , x x 01 01 =z =z {2} {2} , , x x 11 11 =z =z {1,2} {1,2}
then then . .
Next value must be far from Next value must be far from {-1,1} {-1,1} , ,
A contradiction! (boolean function) A contradiction! (boolean function)
Therefore Therefore
0 i
0f f f i 0 i
0f f f i
f 1 f 2 0 f 1 f 2 0
a,b a',b' : f x
abf x
a'b' min f 1 , f 2
a,b a',b' : f x
ab f x
a'b' min f 1 , f 2
ab a'b'
ab a'b'
1 1
ab a'b'
2 2
f x f x
f 1 x x
f 2 x x
ab a'b'
ab a'b'
1 1
ab a'b'
2 2
f x f x
f 1 x x
f 2 x x
f 2 0
f 2 0 1
?
Claim 2 Claim 2
Claim 2 Claim 2 : let : let f f be boolean function, s.t. be boolean function, s.t.
Then either
Then either or or
Proof Proof : consider : consider f( f( ) ) and and f(i f(i 0 0 ) ) : :
Then Then
but but f f is boolean, hence is boolean, hence
0 i
0f f f i 0 i
0f f f i
f f
f f f f i f f i 0 0 i i
00
0
0 0
f f f i
f i f f i
0
0 0
f f f i
f i f f i
0
0f i
0 f 2 f i
0f i f 2 f i
0
f i
0 f 0,2 f i f 0,2
1
-1 0
f
f f i f i
00
0f i
0f i
Proving FKN:
Proving FKN:
almost-linear
almost-linear close to shallow close to shallow
Theorem Theorem : Let : Let f:P([n]) f:P([n]) be be linear linear , ,
Let Let
let let i i
00be the index s.t. is maximal be the index s.t. is maximal then
then
Note Note : : f f is is linear linear , hence , hence w.l.o.g., assume
w.l.o.g., assume i i 0 0 =1 =1 , then all we need to show , then all we need to show is: is:
We show that in the following claim and lemma.
We show that in the following claim and lemma.
0f i
0f i
2
f 1
2 f 1
22
0
i0 2
f f f i
0
i0
22 1 o 1 f f f i
2 1 o 1
n i i 1
f f f i
n i
i 1
f f f i
n 2
i 2
f i 1 o 1
n 2
i 2
f i 1 o 1
Corollary Corollary
Corollary Corollary : Let : Let f f be linear, and be linear, and then
then a a shallow boolean shallow boolean function function g g s.t. s.t.
Proof Proof : let : let , let , let g g be the be the boolean function closest to
boolean function closest to l l . . Then,
Then,
this is true, as this is true, as
is small (by theorem), is small (by theorem),
and additionally and additionally is small, since is small, since
f g 2 2 3 o 1 f g 3 o 1
0f f i
f f i
0
2
f g 2 2 2 9 o 1 f g 9 o 1
2
f 1
2 f 1
22
l g 2 2
l g f l 2 2
f l
2