Upload
jovita
View
19
Download
0
Embed Size (px)
DESCRIPTION
A new characterization of ACC 0 and probabilistic CC 0. Kristoffer Arnsfelt Michal Koucký Hansen Aarhus University Institute of Mathematics Denmark Czech Republic. Bounded depth Boolean circuits - PowerPoint PPT Presentation
Citation preview
A new A new characterization of characterization of
ACCACC00 and and probabilistic CCprobabilistic CC00
Kristoffer ArnsfeltKristoffer Arnsfelt MichalMichal KouckýKoucký
HansenHansen Aarhus University Aarhus University Institute of Mathematics Institute of Mathematics
DenmarkDenmark Czech Republic Czech Republic
2
Bounded depth Boolean circuitsBounded depth Boolean circuits
xx11 x x33 xx44 x x77
Constant depth, polynomial size circuitsConstant depth, polynomial size circuits
MOD-q (MOD-q (xx11, , xx22, …, , …, xxn n ) = 0 ) = 0 iff iff i i >0>0 xxii 0 mod q 0 mod q
MAJ (MAJ (xx11, , xx22, …, , …, xxn n ) = 0 ) = 0 iff iff i i >0>0 xxii > > nn/2/2
3
Bounded depth Boolean circuitsBounded depth Boolean circuits
ACAC00: : unbounded fan-in AND, OR and unary NOT gates.unbounded fan-in AND, OR and unary NOT gates.
ACCACC00: : unbounded fan-in AND, OR, MOD-q and unary unbounded fan-in AND, OR, MOD-q and unary NOT.NOT.
CCCC00: : unbounded fan-in MOD-q gates.unbounded fan-in MOD-q gates.
TCTC00: : unbounded fan-in MAJ and unary NOT gates.unbounded fan-in MAJ and unary NOT gates.
NCNC11: : fan-in two AND, OR and unary NOT gates, O(log)-fan-in two AND, OR and unary NOT gates, O(log)-depth.depth.
Constant depth, polynomial size circuitsConstant depth, polynomial size circuits
MOD-q (MOD-q (xx11, , xx22, …, , …, xxn n ) = 0 ) = 0 iff iff i i >0>0 xxii 0 mod q 0 mod q
MAJ (MAJ (xx11, , xx22, …, , …, xxn n ) = 0 ) = 0 iff iff i i >0>0 xxii > > nn/2/2
4
Known relationshipsKnown relationships
ACAC00 ACC ACC00 TC TC00 NC NC11
CCCC00 AC AC0 0 but but CCCC00 ACC ACC00
Open questions: Open questions: NP NP CC CC00 ? ?
CCCC00 ACC ACC0 0 ??
Conjecture (Barrington-Straubing-Conjecture (Barrington-Straubing-
ThThérienérien):):
AND AND CC CC00
5
Our resultsOur results
Thm: Thm: ACCACC00 rand-CC rand-CC00..
Thm: Thm: ACCACC00 AND AND OR OR CC CC00..
Thm: Thm: ACCACC00 = rand-ACC = rand-ACC0 0 = rand-CC= rand-CC00 = rand( log = rand( log n n )-CC)-CC00..
Thm: Thm: ACCACC00 corresponds to planar bounded-with corresponds to planar bounded-with nondeterministic branching programs of nondeterministic branching programs of polynomial size.polynomial size.
6
AND vs CCAND vs CC00
Fact: Fact: 1) For prime 1) For prime pp, CC, CC00[ [ p p ] cannot compute ] cannot compute AND.AND.
2) For prime power 2) For prime power qq, CC, CC00[ [ q q ] cannot ] cannot compute AND.compute AND.
Thm (BST): Thm (BST): MOD-MOD-pp MOD- MOD-qq circuits require circuits require exponential size to compute AND.exponential size to compute AND.
Thm (ThThm (Théérien): rien): CCCC00 circuits for AND require circuits for AND require ΩΩ( ( n n ) ) gates in their first layer.gates in their first layer.
p,qp,q co-prime integers co-prime integers
7
AND vs CCAND vs CC00
Thm (BST): Thm (BST): CCCC00[ [ pq pq ] circuits of exponential size ] circuits of exponential size can compute any Boolean function, in particular can compute any Boolean function, in particular AND.AND.
Cor: Cor: CCCC00[ [ pq pq ] circuits of size 2] circuits of size 2nn and depth O(1/ and depth O(1/) )
can compute AND.can compute AND.
Thm(BBR): Thm(BBR): CCCC00[ [ q q ] circuits of size 2] circuits of size 2n n 1/1/rr and depth 4 and depth 4 can compute AND if can compute AND if qq has has rr distinct prime distinct prime factors.factors.
p,qp,q co-prime integers co-prime integers
8
Thm:Thm: AND is computable by rand-CC AND is computable by rand-CC00[ [ pq pq ] circuits with ] circuits with error <1/error <1/n n log log nn if if pp and and qq are co-prime integers. are co-prime integers.
Pf: Pf: Razborov-Smolensky methodRazborov-Smolensky method
Fixed input Fixed input xx11, , xx22, …, , …, xxnn
Take a random set S Take a random set S {1, …, {1, …, n n }}
with probability at least 1/2 over random choice of Swith probability at least 1/2 over random choice of S
OR(OR(xx11, , xx22, …, , …, xxnn ) = MOD-q { ) = MOD-q { xxi i , , ii S } S }
take take kk=log=log22 nn independent random sets S independent random sets S11, S, S22, …, S, …, Skk
with probability at least 1/with probability at least 1/n n log log nn over random over random choices of S’schoices of S’s
OR(OR(xx11, , xx22, …, , …, xxnn ) = OR ) = ORjj MOD-q { MOD-q { xxi i , , ii S Sjj } }
Cor:Cor: ACC ACC00 rand-CC rand-CC00..
9
Previous construction requires Previous construction requires nn log log22 nn random bits. random bits.
One can reduce the number of random bits to One can reduce the number of random bits to O(log O(log nn) while keeping the error below 1/) while keeping the error below 1/n n kk by by use of:use of:
1.1. Valiant-Vazirani isolation technique, andValiant-Vazirani isolation technique, and
2.2. Randomness efficient sampling using random Randomness efficient sampling using random walks on expanders.walks on expanders.
Similar to [AJMV]Similar to [AJMV]
logspace uniformitylogspace uniformity
Cor:Cor: ACC ACC00 rand(log rand(log nn)-CC)-CC00..
10
DerandomizationDerandomization
Thm (Ajtai, Ben-Or):Thm (Ajtai, Ben-Or): 1) rand-AC1) rand-AC00 AC AC00..2) rand-ACC2) rand-ACC00 ACC ACC00..
Open: Open: rand-CC rand-CC00 CC CC00 ? ?
Claim: Claim: rand-CCrand-CC00 = CC = CC0 0 iff AND iff AND CC CC00..
Thm: Thm: ACCACC00 AND AND OR OR CC CC00..
11
Thm: Thm: ACCACC00 AND AND OR OR CC CC00. . (non-(non-uniformly)uniformly)
Pf:Pf: Technique of Ajtai and Ben-Or Technique of Ajtai and Ben-Or
CCnn a rand-CC a rand-CC00 circuit computing f circuit computing fnn with error <1/3 with error <1/3nn..
Take OR of Take OR of nn independent copies of C independent copies of Cnn
if fif fn n ( ( x x ) = 1 then OR ) = 1 then OR C Cnn( ( xx ) = 0 with probability < ) = 0 with probability < ( 1/3( 1/3n n ))nn
if fif fn n ( ( x x ) = 0 then OR ) = 0 then OR C Cnn( ( xx ) = 1 with probability < ) = 1 with probability < 1/31/3
12
CCnn a rand-CC a rand-CC00 circuit computing f circuit computing fnn with error <1/3 with error <1/3nn..
Take OR of Take OR of nn independent copies of C independent copies of Cnn
if fif fn n ( ( x x ) = 1 then OR ) = 1 then OR C Cnn( ( xx ) = 0 with probability < ) = 0 with probability < ( 1/3( 1/3n n ))nn
if fif fn n ( ( x x ) = 0 then OR ) = 0 then OR C Cnn( ( xx ) = 1 with probability < 1/3 ) = 1 with probability < 1/3
Take AND of Take AND of nn independent copies of OR independent copies of OR C Cnn
if fif fn n ( ( x x ) = 1 then AND ) = 1 then AND OR OR C Cnn( ( xx ) = 0 with p. < ) = 0 with p. < n n ( 1/3( 1/3n n ))nn
if fif fn n ( ( x x ) = 0 then AND ) = 0 then AND OR OR C Cnn( ( xx ) = 1 with p. < ( 1/3 ) = 1 with p. < ( 1/3 ))nn
In both cases the probability of error is less than 2In both cases the probability of error is less than 2n n so we so we can fix a particular random bits that will give the correct can fix a particular random bits that will give the correct answer for all answer for all xx..
13
Previous construction requires to fix >Previous construction requires to fix >nn22 random random bits so it is non-uniform.bits so it is non-uniform.
One can get uniform construction using:One can get uniform construction using:
1.1. Lautemann’s technique, andLautemann’s technique, and
2.2. Randomness efficient sampling using random Randomness efficient sampling using random walks on expanders.walks on expanders.
Similar to [AH, V]Similar to [AH, V]
Thm: Thm: ACCACC00 AND AND OR OR CC CC00. . (uniformly) (uniformly)
14
Geometric restrictions of circuits and Geometric restrictions of circuits and branching programsbranching programs
15
Geometric restrictions of circuits and Geometric restrictions of circuits and branching programsbranching programs
16
Constant width circuitsConstant width circuits
Thm (Barrington): Thm (Barrington): NC NC11 corresponds to constant corresponds to constant width circuits.width circuits.
Thm (Hansen’06): Thm (Hansen’06): ACC ACC00 corresponds to constant corresponds to constant width planar circuits.width planar circuits.
Thm (BLMS’99): Thm (BLMS’99): ACAC00 corresponds to constant width corresponds to constant width upward planar circuits.upward planar circuits.
17
Constant width nondeterministic branching Constant width nondeterministic branching programsprograms
Thm (Barrington): Thm (Barrington): NC NC11 corresponds to constant corresponds to constant width nondeterministic branching programs.width nondeterministic branching programs.
Thm: Thm: ACC ACC00 corresponds to constant width planar corresponds to constant width planar nondeterministic branching programs.nondeterministic branching programs.
Thm (BLMS’98): Thm (BLMS’98): ACAC00 corresponds to constant width corresponds to constant width upward planar nondeterministic branching upward planar nondeterministic branching programs.programs.
18
Constant width nondeterministic branching Constant width nondeterministic branching programsprograms
Thm (Hansen’08): Thm (Hansen’08): Quasi-polynomial size ACC Quasi-polynomial size ACC00 corresponds to quasi-polynomial size constant width corresponds to quasi-polynomial size constant width planar nondeterministic branching programs.planar nondeterministic branching programs.
Thm (HMV): Thm (HMV): Functions computable by constant width Functions computable by constant width planar nondeterministic branching programs are in planar nondeterministic branching programs are in ACCACC00..
Thm (Hansen’08): Thm (Hansen’08): Functions from AND Functions from AND OR OR CC CC00 are are computable by constant width planar computable by constant width planar nondeterministic branching programs.nondeterministic branching programs.
Thm: Thm: ACC ACC00 corresponds to constant width planar corresponds to constant width planar nondeterministic branching programs.nondeterministic branching programs.