some topics in analysis of boolean functions

109
Ryan O’Donnell Carnegie Mellon University

Upload: guest756c74

Post on 13-Jul-2015

572 views

Category:

Education


5 download

TRANSCRIPT

Page 1: Some topics in analysis of boolean functions

R ya n O’Donne llC a rne g ie Me llon Unive rs ity

Page 2: Some topics in analysis of boolean functions

P a rt 1:

A. F ourie r e xpa ns ion ba s ic s

B . Conc e pts :

B ia s , In flue nc e s , Nois e S e ns itiv ity

C. Ka la i’s proof o f Arrow’s Th e ore m

Page 3: Some topics in analysis of boolean functions

10 Minute B re a k

Page 4: Some topics in analysis of boolean functions

P a rt 2:

A. Th e Hype rc ontra c tive Ine qua lity

B . Alg orith m ic G a ps

Page 5: Some topics in analysis of boolean functions

S a dly no tim e for:

Le a rn ing th e ory

P s e udora ndomne s s

Arith m e tic c ombina toric s

R a ndom g ra ph s / pe rc ola tion

Communic a tion c omple xity

Me tric / B a na c h s pa c e s

Coding th e ory

e tc .

Page 6: Some topics in analysis of boolean functions

1A. F ourie r e xpa ns ion ba s ic s

Page 7: Some topics in analysis of boolean functions

f : {0,1}n {0,1}

Page 8: Some topics in analysis of boolean functions

f : {−1,+1}n {−1,+1}

Page 9: Some topics in analysis of boolean functions

ℝ3(+1,+1,+1)

(−1,−1,−1)

(+1,+1,−1)

(+1,−1,+1)

(−1,+1,+1)

−1

−1

−1

+1+1

+1

+1

−1

Page 10: Some topics in analysis of boolean functions

ℝ3(+1,+1,+1)

(−1,−1,−1)

+1

+1

+1

+1−1

−1

−1

−1

Page 11: Some topics in analysis of boolean functions

ℝ3(+1,+1,+1)

(−1,−1,−1)

−1

−1

−1

+1−1

−1

−1

−1

Page 12: Some topics in analysis of boolean functions

ℝ3(+1,+1,+1)

(−1,−1,−1)

+1

+1

+1

+1+1

+1

+1

−1

Page 13: Some topics in analysis of boolean functions

ℝ3(+1,+1,+1)

(−1,−1,−1)

+1

+1

+1

+1+1

+1

+1

+1

Page 14: Some topics in analysis of boolean functions

ℝ3(+1,+1,+1)

(−1,−1,−1)

−1

−1

−1

−1−1

−1

−1

−1

Page 15: Some topics in analysis of boolean functions

ℝ3(+1,+1,+1)

(−1,−1,−1)

−1

−1

+1

+1−1

+1

+1

−1

Page 16: Some topics in analysis of boolean functions

ℝ3(+1,+1,+1)

(−1,−1,−1)

−1

+1

−1

+1+1

−1

+1

−1

Page 17: Some topics in analysis of boolean functions

ℝ3(+1,+1,+1)

(−1,−1,−1)

+1

−1

−1

+1+1

+1

−1

−1

Page 18: Some topics in analysis of boolean functions

ℝ3(+1,+1,+1)

(−1,−1,−1)

−1

−1

−1

+1+1

+1

+1

−1

Page 19: Some topics in analysis of boolean functions

(+1,+1,+1)

+1

+1

+1+1

−1

−1

−1−1

+1

+1

+1

+1

−1

−1 −1

−1

(+1,+1,−1)

(+1,−1,−1)

Page 20: Some topics in analysis of boolean functions

=

Page 21: Some topics in analysis of boolean functions
Page 22: Some topics in analysis of boolean functions

=

Page 23: Some topics in analysis of boolean functions

P ropos ition:

E ve ry f : {−1,+1}n {−1,+1} c a n be

e xpre s s e d a s a multilinear po lynom ia l,

Th a t’s it. Th a t’s th e “F ourie r e xpa ns ion” o f f.

(un ique ly)

(inde e d, )→ ℝ

Page 24: Some topics in analysis of boolean functions

P ropos ition:

E ve ry f : {−1,+1}n {−1,+1} c a n be

e xpre s s e d a s a multilinear po lynom ia l,

Th a t’s it. Th a t’s th e “F ourie r e xpa ns ion” o f f.

(un ique ly)

(inde e d, )→ ℝ

Page 25: Some topics in analysis of boolean functions

R e s t: 0

Page 26: Some topics in analysis of boolean functions

Why?

Coe ffic ie nts e nc ode us e fu l in form a tion.

When?

1. Uniform proba b ility involve d

2. Ha m m ing d is ta nc e s re le va nt

Page 27: Some topics in analysis of boolean functions

P a rs e va l’s Th e ore m :

Le t f : {−1,+1}n {−1,+1}.

Th e n

a vg { f(x)2 }

Page 28: Some topics in analysis of boolean functions

“We ig h t” o f f on S [n ]⊆

=

Page 29: Some topics in analysis of boolean functions

{2}{1}

{3}

{1,3}{1,2} {2,3}

{1,2,3}

Page 30: Some topics in analysis of boolean functions

{2}{1}

{3}

{1,3}{1,2} {2,3}

{1,2,3}

Page 31: Some topics in analysis of boolean functions

{2}{1}

{3}

{1,3}{1,2} {2,3}

{1,2,3}

Page 32: Some topics in analysis of boolean functions

{2}{1}

{3}

{1,3}{1,2} {2,3}

{1,2,3}

Page 33: Some topics in analysis of boolean functions

{2}{1}

{3}

{1,3}{1,2} {2,3}

{1,2,3}

Page 34: Some topics in analysis of boolean functions

{2}{1}

{3}

{1,3}{1,2} {2,3}

{1,2,3}

Page 35: Some topics in analysis of boolean functions

1B . Conc e pts :

B ia s , In flue nc e s , Nois e S e ns itiv ity

Page 36: Some topics in analysis of boolean functions

S oc ia l Ch oic e :

Ca ndida te s ±1

n vote rs

Vote s a re ra ndom

f : {−1,+1}n {−1,+1}

is th e “voting ru le ”

Page 37: Some topics in analysis of boolean functions

B ia s o f f:

a vg f(x) = Pr[+1 wins ] − Pr[−1 wins ]

F a c t:

We ig h t on = m e a s ure s “im ba la nc e ”.∅

Page 38: Some topics in analysis of boolean functions

In flue nc e o f i on f:

Pr[ f(x) ≠ f(x( i)⊕ ) ]

= Pr[vote r i is a s wing vote r]

F a c t:

Page 39: Some topics in analysis of boolean functions

{2}{1}

{3}

{1,3}{1,2} {2,3}

{1,2,3}

Maj(x1,x2,x3)

Page 40: Some topics in analysis of boolean functions
Page 41: Some topics in analysis of boolean functions

+1

+1

+1+1

−1

−1

−1−1

In fi(f) = Pr[ f(x) ≠ f(x( i)⊕ ) ]

Page 42: Some topics in analysis of boolean functions

+1

+1

+1+1

−1

−1

−1−1

In fi(f) = Pr[ f(x) ≠ f(x( i)⊕ ) ]

Page 43: Some topics in analysis of boolean functions

a vg In fi(f) = fra c . o f e dg e s wh ic h

a re c ut e dg e s

Page 44: Some topics in analysis of boolean functions

LMN Th e ore m :

If f is in AC 0

th e n a vg In fi(f)

Page 45: Some topics in analysis of boolean functions

⇒ a vg In fi(P a rityn) = 1

⇒ P a rity AC∉ 0

⇒ a vg In fi(Ma jn) =

⇒ Ma jority AC∉ 0

Page 46: Some topics in analysis of boolean functions

KKL Th e ore m :

If B ia s (f) = 0,

th e n

Corolla ry:

As s uming f m onotone ,

−1 or +1 c a n b ribe o(n) vote rs

a nd win w.p. 1−o(1).

Page 47: Some topics in analysis of boolean functions

Nois e S e ns itivity o f f a t :

NS (f) = Pr[wrong winne r wins ],

wh e n e a c h vote mis re c orde d w/prob

f(

f(

)

)

+ − + + − − + − −

− − + + + + + − −

Page 48: Some topics in analysis of boolean functions
Page 49: Some topics in analysis of boolean functions

Le a rn ing Th e ory princ ip le :

[LMN’93, …, KKMS ’05]

If a ll f ∈ C h a ve s m a ll NS (f)

th e n C is e ffic ie ntly le a rna b le .

Page 50: Some topics in analysis of boolean functions

{2}{1}

{3}

{1,3}{1,2} {2,3}

[3]

Page 51: Some topics in analysis of boolean functions

P ropos ition:

for s ma ll ,

with Electoral College:

10

1

Page 52: Some topics in analysis of boolean functions

1C. Ka la i’s proof o f Arrow’s Th e ore m

Page 53: Some topics in analysis of boolean functions

R a nking 3 c a ndida te s

Condorc e t [1775] E le c tion:

=> (x_i, y_i, z _i) a re Not All E qua l (no 111 -1-1-1)

Condorc e t: Try f = Ma j. Outc om e c a n be “irra tiona l” A

> B > C > A. [e a s y e g ]

Ma ybe s om e oth e r f?

A > B ?

B > C?

C > A?

Page 54: Some topics in analysis of boolean functions

R a nking 3 c a ndida te s

Condorc e t [1775] E le c tion:

=> (x_i, y_i, z _i) a re Not All E qua l (no 111 -1-1-1)

Condorc e t: Try f = Ma j. Outc om e c a n be “irra tiona l” A

> B > C > A. [e a s y e g ]

Ma ybe s om e oth e r f?

• • • • • •

A > B ?

B > C?

C > A?

“C >

A >

B”

“A >

B >

C”

“B >

C >

A”

+

+

+

+

+

+

+

+

+

+

+

+

+

+

Page 55: Some topics in analysis of boolean functions

R a nking 3 c a ndida te s

Condorc e t [1775] E le c tion:

=> (x_i, y_i, z _i) a re Not All E qua l (no 111 -1-1-1)

Condorc e t: Try f = Ma j. Outc om e c a n be “irra tiona l” A

> B > C > A. [e a s y e g ]

Ma ybe s om e oth e r f?

• • • • • •

A > B ?

B > C?

C > A?

“C >

A >

B”

“A >

B >

C”

“B >

C >

A”

f( )

f( )

f( )

= +

= +

= −

S oc ie ty: “A > B > C”

+

+

+

+

+

+

+

+

+

+

+

+

+

+

Page 56: Some topics in analysis of boolean functions

R a nking 3 c a ndida te s

Condorc e t [1775] E le c tion:

=> (x_i, y_i, z _i) a re Not All E qua l (no 111 -1-1-1)

Condorc e t: Try f = Ma j. Outc om e c a n be “irra tiona l” A

> B > C > A. [e a s y e g ]

Ma ybe s om e oth e r f?

• • • • • •

A > B ?

B > C?

C > A?

“C >

A >

B”

“A >

B >

C”

“B >

C >

A”

f( )

f( )

f( )

= +

= +

= −

S oc ie ty: “A > B > C”

+

+

+

+

+

+

+

+

+

+

+

+

+

+

+

Page 57: Some topics in analysis of boolean functions

R a nking 3 c a ndida te s

Condorc e t [1775] E le c tion:

=> (x_i, y_i, z _i) a re Not All E qua l (no 111 -1-1-1)

Condorc e t: Try f = Ma j. Outc om e c a n be “irra tiona l” A

> B > C > A. [e a s y e g ]

Ma ybe s om e oth e r f?

• • • • • •

“C >

A >

B”

“A >

B >

C”

“B >

C >

A”

S oc ie ty: “A > B > C”

A > B ?

B > C?

C > A?

f( )

f( )

f( )

= +

= +

= +

+

+

+

+

+

+

+

+

+

+

+

+

+

+

+

Page 58: Some topics in analysis of boolean functions

R a nking 3 c a ndida te s

Condorc e t [1775] E le c tion:

=> (x_i, y_i, z _i) a re Not All E qua l (no 111 -1-1-1)

Condorc e t: Try f = Ma j. Outc om e c a n be “irra tiona l” A

> B > C > A. [e a s y e g ]

Ma ybe s om e oth e r f?

• • • • • •

“C >

A >

B”

“A >

B >

C”

“B >

C >

A”

S oc ie ty: “A > B > C > A”?A > B ?

B > C?

C > A?

f( )

f( )

f( )

= +

= +

= +

+

+

+

+

+

+

+

+

+

+

+

+

+

+

+

Page 59: Some topics in analysis of boolean functions

Arrow’s Impos s ib ility Th e ore m [1950]:

If

f : {−1,+1}n {−1,+1} never g ive s

irra tiona l outc ome in Condorc e t e le c tions ,

th e n

f is a Dictator or a negated-Dic tator.

Page 60: Some topics in analysis of boolean functions

G il Ka la i’s P roof [2002]:

Page 61: Some topics in analysis of boolean functions

• • • • • •

“C >

A >

B”

“A >

B >

C”

“B >

C >

A”

A > B ?

B > C?

C > A?

f( )

f( )

f( )

= +

= +

= −

+

+

+

+

+

+

+

+

+

+

+

+

+

+

Page 62: Some topics in analysis of boolean functions

• • • • • •

“C >

A >

B”

“A >

B >

C”

“B >

C >

A”

A > B ?

B > C?

C > A?

f( )

f( )

f( )

= +

= +

= −

+

+

+

+

+

+

+

+

+

+

+

+

+

+

Page 63: Some topics in analysis of boolean functions

G il Ka la i’s P roof:

Page 64: Some topics in analysis of boolean functions

G il Ka la i’s P roof:

Page 65: Some topics in analysis of boolean functions

G il Ka la i’s P roof, c onc lude d:

f never g ive s irra tiona l outc ome s ⇒ e qua lity

⇒ a ll F ourie r we ig h t “a t le ve l 1”

⇒ f(x) = ±xj for s om e j (e xe rc is e ).

Page 66: Some topics in analysis of boolean functions

G uilba ud’s Th e ore m [1952]

Guilba ud’s Numbe r ≈ .912

Page 67: Some topics in analysis of boolean functions

Corolla ry o f “Ma jority Is S ta b le s t” [MOO05]:

If In fi(f) ≤ o (1) for a ll i,

th e n

Pr[ra tiona l outc ome with f]

Page 68: Some topics in analysis of boolean functions
Page 69: Some topics in analysis of boolean functions

P a rt 2:

A. Th e Hype rc ontra c tive Ine qua lity

B . Alg orith m ic G a ps

Page 70: Some topics in analysis of boolean functions

2A. Th e Hype rc ontra c tive Ine qua lity

AKA B ona m i-B e c kne r Ine qua lity

Page 71: Some topics in analysis of boolean functions

a ll us e “Hype rc ontra c tive Ine qua lity”

Page 72: Some topics in analysis of boolean functions

Hoe ffd ing Ine qua lity:

Le t

F = c 0 + c 1 x1 + c 2 x2 + ··· + c n xn,

wh e re x i’s a re inde p., un if. ra ndom ±1.

Page 73: Some topics in analysis of boolean functions

Me a n: μ = c 0 Va ria nc e :

Hoe ffd ing Ine qua lity:

Le t

F = c 0 + c 1 x1 + c 2 x2 + ··· + c n xn,

Page 74: Some topics in analysis of boolean functions

Me a n: μ = Va ria nc e :

Hype rc ontra c tive Ine qua lity*:

Le t

Page 75: Some topics in analysis of boolean functions

Th e n for a ll q ≥ 2,

Hype rc ontra c tive Ine qua lity:

Le t

Page 76: Some topics in analysis of boolean functions

Th e n F is a “re a s ona b le d” ra ndom va ria b le .

Hype rc ontra c tive Ine qua lity:

Le t

Page 77: Some topics in analysis of boolean functions

Th e n for a ll q ≥ 2,

Hype rc ontra c tive Ine qua lity:

Le t

Page 78: Some topics in analysis of boolean functions

Th e n

“q = 4” Hype rc ontra c tive Ine qua lity:

Le t

Page 79: Some topics in analysis of boolean functions

Th e n

“q = 4” Hype rc ontra c tive Ine qua lity:

Le t

Page 80: Some topics in analysis of boolean functions

a ll us e Hype rc ontra c tive Ine qua lity

Page 81: Some topics in analysis of boolean functions

ju s t us e “q = 4” Hype rc ontra c tive Ine qua lity

Page 82: Some topics in analysis of boolean functions

“q = 4” Hype rc ontra c tive Ine qua lity:

Le t F be de g re e d ove r n i.i.d. ±1 r.v.’s .

Th e n

P roof [MOO’05]: Induc tion on n.

Obvious s te p.

Us e induc tion h ypoth e s is .

Us e Ca uc h y-S c h wa rz on th e obvious th ing .

Us e induc tion h ypoth e s is .

Obvious s te p.

Page 83: Some topics in analysis of boolean functions

2B . Alg orith m ic G a ps

Page 84: Some topics in analysis of boolean functions

Opt

be s t po ly-tim eg ua ra nte e

ln(N)

“S e t-Cove r is NP -h a rd to

a pproxim a te to fa c tor ln(N)”

Page 85: Some topics in analysis of boolean functions

Opt

LP -R a nd-R oundingg ua ra nte e

ln(N)

“F a c tor ln(N) Algorithmic Gap

for LP -R a nd-R ounding ”

Page 86: Some topics in analysis of boolean functions

Opt(S )

LP -R a nd-R ounding (S )

ln(N)

“Algorithmic Gap Ins tance S

for LP -R a nd-R ounding ”

Page 87: Some topics in analysis of boolean functions

Alg orith mic G a p ins ta nc e s

a re o fte n “ba s e d on” {−1,+1}n.

Page 88: Some topics in analysis of boolean functions

S pa rs e s t-Cut:

Alg orith m : Arora -R a o-Va z ira n i S DP .

G ua ra nte e : F a c tor

Page 89: Some topics in analysis of boolean functions
Page 90: Some topics in analysis of boolean functions

Opt = 1/n

Page 91: Some topics in analysis of boolean functions

Opt = 1/n

Page 92: Some topics in analysis of boolean functions

Opt = 1/n

Page 93: Some topics in analysis of boolean functions

Opt = 1/n

f(x) = s g n( )

Page 94: Some topics in analysis of boolean functions

Opt = 1/n

f(x) = s g n(r1x1 + ••• + rnxn)

AR V g e ts

Page 95: Some topics in analysis of boolean functions

Opt = 1/n

AR V g e ts

g a p:

Page 96: Some topics in analysis of boolean functions

Alg orith mic G a ps Ha rdne s s -o f-Approx→

LP / S DP -rounding Alg . G a p ins ta nc e

• n optim a l “Dic ta tor” s o lutions

• “g e ne ric m ixture o f Dic ta tors ” muc h wors e

+ P CP te c h nolog y

= s a me -g a p h a rdne s s -o f-a pproxim a tion

Page 97: Some topics in analysis of boolean functions

Alg orith mic G a ps Ha rdne s s -o f-Approx→

LP / S DP -rounding Alg . G a p ins ta nc e

• n optim a l “Dic ta tor” s o lutions

• “g e ne ric m ixture o f Dic ta tors ” muc h wors e

+ P CP te c h nolog y

= s a me -g a p h a rdne s s -o f-a pproxim a tion

Page 98: Some topics in analysis of boolean functions

KKL / Ta la g ra nd Th e ore m :

If f is b a la nc e d,

In fi(f) ≤ 1/n .01 for a ll i,

th e n

a vg In fi(f) ≥

G a p: Θ(log n) = Θ(log log N).

Page 99: Some topics in analysis of boolean functions

[CKKR S 05]: KKL + Unique G a me s Conje c ture

⇒ Ω(log log log N) h a rdne s s -o f-a pprox.

Page 100: Some topics in analysis of boolean functions

2-Colora b le 3-Uniform h ype rg ra ph s :

Input: 2-c o lora b le , 3-un if. h ype rg ra ph

Output: 2-c o loring

Ob j: Ma x. fra c tion o f le g a lly

c o lore d h ype re dg e s

Page 101: Some topics in analysis of boolean functions

2-Colora b le 3-Uniform h ype rg ra ph s :

Alg orith m : S DP [KLP 96].

G ua ra nte e :

[Zwic k99]

Page 102: Some topics in analysis of boolean functions

Alg orith m ic G a p Ins ta nc e

Ve rtic e s : {−1,+1}n

6 n h ype re dg e s :{ (x,y,z ) : pos s . pre fs in

a Condorc e t e le c tion}

(i.e., triples s.t. (xi,y i,z i) NAE for all i)

Page 103: Some topics in analysis of boolean functions

E lts : {−1,+1}n E dg e s : Condorc e t vote s (x,y,z )

2-c o loring = f : {−1,+1}n {−1,+1}→

fra c . le g a lly c o lore d h ype re dg e s

= Pr[“ra tiona l” outc ome with f]

Ins ta nc e 2-c o lora b le ? ✔

(2n optima l s o lutions : ±Dic ta tors )

Page 104: Some topics in analysis of boolean functions

E lts : {−1,+1}n E dg e s : Condorc e t vote s (x,y,z )

S DP rounding a lg . m a y output

R a ndom we ig h te d m a jority a ls o

ra tiona l-with -prob .-.912! [s a m e CLT a rg .]

f(x) = s g n(r1x1 + ••• + rnxn)

Page 105: Some topics in analysis of boolean functions

Alg orith mic G a ps Ha rdne s s -o f-Approx→

LP / S DP -rounding Alg . G a p ins ta nc e

• n optim a l “Dic ta tor” s o lutions

• “g e ne ric m ixture o f Dic ta tors ” muc h wors e

+ P CP te c h nolog y

= s a me -g a p h a rdne s s -o f-a pproxim a tion

Page 106: Some topics in analysis of boolean functions

Corolla ry o f Ma jority Is S ta b le s t:

If In fi(f) ≤ o (1) for a ll i,

th e n

Pr[ra tiona l outc ome with f]

Cor: th is + Unique Ga me s Conje c ture

⇒ .912 ha rdne s s -of-a pprox*

Page 107: Some topics in analysis of boolean functions

2C. F uture Dire c tions

Page 108: Some topics in analysis of boolean functions

De ve lop th e “s truc ture vs . ps e udora ndom ne s s ”

th e ory for B oole a n func tions .

Page 109: Some topics in analysis of boolean functions