300 likes | 314 Views
Dual VP Classes. Joint work with Anna Gál (U. Texas) and Ian Mertz (Rutgers). MFCS, Milan, August 27, 2015. Our Contributions. New characterizations of ACC 1 and TC 1 . New examples of fan-in reduction. Highlight connections between ACC 1 and VP.
E N D
Dual VP Classes Joint work with Anna Gál (U. Texas) and Ian Mertz (Rutgers) MFCS, Milan, August 27, 2015
Our Contributions • New characterizations of ACC1 and TC1. • New examples of fan-in reduction. • Highlight connections between ACC1 and VP. • Revisit the Immerman-Landau Conjecture, and offer some new conjectures about circuit complexity classes. • But first …let’s review the relevant complexity classes.
NP P AC1 NL L NC1 AC0 Fan-in is Important! Log-Depth Poly-size Unbounded Fan-in Fan-in 2
NP P AC1 SAC1=LogCFL NL L NC1 AC0 Fan-in is Important! Log-Depth Poly-size Semi-unbounded fan-in Λ fan-in 2 V fan-in nk
NP P TC1 AC1 SAC1=LogCFL NL L NC1 TC0 AC0 Components are Important! Log depth Majority gates O(1) depth
P#P NP P TC1 AC1 SAC1=LogCFL NL L NC1 TC0 AC0 L#L = LDet(Q)
P#P=PVNP(Q) NP P TC1 AC1 SAC1=LogCFL NL L NC1 TC0 AC0 L#LogCFL = LVP(Q) L#L = LDet(Q)
Valiant’s Class VP • VP(R) is the class of families (fn) of multivariate polynomials over R such that • fn has degree nO(1). • There is a family of arithmetic circuits (Cn) of size poly(n) such that Cn computes fn. • Furthermore, Cn can be assumed to have depth O(log n) with fan-in 2 x and unbounded fan-in +. (Semiunbounded fan-in arithmetic circuits.) • #SAC1 = the functions in VP(N).
P#P=PVNP(Q) NP P TC1 AC1 SAC1=LogCFL NL L NC1 TC0 AC0 L#LogCFL = LVP(Q) L#L = LDet(Q)
P#P =PVNP(Q) NP P TC1 AC1 SAC1=LogCFL NL L #NC1(Q) NC1 TC0 AC0 # AC1(Q) Not contained in P for a trivial reason: The output has more than poly-many bits. L#LogCFL = LVP(Q) =L#SAC1 L#L = LDet(Q)
P#P =PVNP(Q) NP P TC1 AC1 SAC1=LogCFL NL L #NC1(Q) NC1 TC0 AC0 = # AC1(Fpn) The meaning of Fpn is: Circuit Cn is interpreted modulo the nth prime. LVP(Fpn) = L#LogCFL = LVP(Q) =L#SAC1 L#L = LDet(Q) ≈ # NC1(Fpn) ≈ # AC0(Q) = # AC0(Fpn)
P#P =PVNP(Q) NP P TC1 ACC1 AC1 SAC1=LogCFL NL L #NC1(Q) NC1 TC0 ACC0 AC0 = # AC1(Fpn) = Uq # AC1(Fq) LVP(Fpn) = L#LogCFL = LVP(Q) =L#SAC1 ACCi = Um ACi[m] L#L = LDet(Q) = Uq # AC0(Fq)
P#P =PVNP(Q) NP P TC1 ACC1 AC1 SAC1=LogCFL NL L #NC1(Q) NC1 TC0 ACC0 AC0 = # AC1(Fpn) LVP(Fpn) = L#LogCFL = LVP(Q) =L#SAC1 Our focus lies here. L#L = LDet(Q)
Dual VP Classes SAC1=LogCFL = VP(B2): Unbounded V Bounded Λ VP(R): Unbounded + Bounded x But LogCFL is closed under complement! [BCDRT]
Dual VP Classes SAC1=LogCFL = VP(B2): Unbounded V Bounded Λ VP(R): Unbounded + Bounded x = Unbounded Λ Bounded V
Dual VP Classes SAC1=LogCFL = VP(B2): Unbounded V Bounded Λ VP(R): Unbounded + Bounded x = ΛP(B2): Unbounded Λ Bounded V ΛP(R): Unbounded x Bounded + Is this interesting??
New Characterizations of ACC1 • ACC1= Uq#AC1(Fq) • = UqΛP(Fq) • Fan-in Reduction (from unbounded to semiunbounded) • #AC1(Fq) = AC1[q(q-1)] • ΛP(Fq) = AC1[q-1]
…and TC1 • ACC1= Uq#AC1(Fq) • = UqΛP(Fq) • Fan-in Reduction (from unbounded to semiunbounded) • #AC1(Fq) = AC1[q(q-1)] • ΛP(Fq) = AC1[q-1] • TC1= # AC1(Fpn) = LΛP(Fpn)
Boolean Fan-In Reduction • By definition, AC1[m] has poly size, log depth, with unbounded fan-in MODm, V and Λ gates. • Theorem: The fan-in of the V and Λ gates can be reduced to log n, with no loss of computational power. • In symbols: AC1[m] = log-AC1[m]. • Theorem: If m is not a prime power, then the fan-in can be reduced to 2, with no loss of power. AC1[m] = 2-AC1[m]. • …and to ZERO! AC1[m] = 0-AC1[m].
ACC1 and VP • That is: ACC1 corresponds to uniform families of MODm gates (with no other hardware). • Compare the circuit characterization of ACC1 with the circuit characterization of VP(Fq): • For any odd prime q, VP(Fq) is the class of languages accepted by uniform families of MODq gates (with no other hardware).
ACC1 and VP • That is: ACC1 corresponds to uniform families of MODm gates (with no other hardware). • Compare the circuit characterization of ACC1 with the circuit characterization of VP(Fq): • For any odd prime q, VP(Fq) is the class of languages accepted by uniform families of MODq gates (with no other hardware). • Thus, over finite fields, the difference between VP and ΛP (=ACC1) boils down to the difference between primes and composites.
Degree Reduction • We have seen examples of fan-in reduction for Boolean circuits (such as AC1[5] = log-AC1[5]). • And we have seen examples of fan-in reduction for arithmetic circuits (such as Uq #AC1(Fq) = Uq ΛP(Fq))… • …which only reduced the fan-in of + gates – and hence did not result in a reduction of the degree of the polynomial represented. • Should we expect any reduction of the fan-in of x gates to be possible?
Degree Reduction • Should we expect any reduction of the fan-in of x gates to be possible? • Consider the Immerman-Landau conjecture: • TC1=LDet(Q) • Equivalently: # AC1(Fpn) = LDet(Q) = LVP(Q) = LVP(Fpn) • [Buhrman et al] argued that it would be unlikely for a high-degree arithmetic class to coincide with a polynomial-degree arithmetic class.
Degree Reduction • We present examples where degree reduction is possible. • Define #WSAC1 to be circuits with a “weak” form of the semiunbounded fan-in restriction: poly-size, log depth circuits with unbounded fan-in + gates, and logarithmic-fan-in x gates. • Theorem: For any prime q, AC1[q] = #WSAC1(Fq). • Corollary: #AC1(F2) = #WSAC1(F2).
Degree Reduction • Consider #AC1(F2) = #WSAC1(F2). • Polynomials in #AC1(F2) have degree nO(log n). • Polynomials in #WSAC1(F2) have degree nO(log log n). • This is proved using off-the-shelf techniques (isolation lemma, derandomization using walks on expanders). We see no reason why degree nO(log log n) should be optimal. • If it can be reduced to nO(1), then #AC1(F2) = VP(F2).
Degree Reduction • Consider #AC1(F2) = #WSAC1(F2). • Polynomials in #AC1(F2) have degree nO(log n). • Polynomials in #WSAC1(F2) have degree nO(log log n). • This is proved using off-the-shelf techniques (isolation lemma, derandomization using walks on expanders). We see no reason why degree nO(log log n) should be optimal. • If it can be reduced to nO(1), then #AC1(F2) = VP(F2) = ΛP(F3).
Open Questions • We believe that the arguments presented against the Immerman-Landau conjecture – which are based on degree-reduction being unlikely – are weakened by examples of degree-reduction. Can one improve the degree reduction? • Can the connection between ACC1 and VP be strengthened? • Is Um LVP(Zm) equal to Um AC1[m] (= ACC1)? • This would imply AC1 is contained in LVP[Zm] for some m.
Open Questions • We believe that the arguments presented against the Immerman-Landau conjecture – which are based on degree-reduction being unlikely – are weakened by examples of degree-reduction. Can one improve the degree reduction? • Can the connection between ACC1 and VP be strengthened? • Is Um LVP(Zm) equal to Um AC1[m] (= ACC1)? • This would imply AC1 is contained in LVP[Zm] for some m. (SAC1is there, nonuniformly.)
#P NP P TC1 ACC1 AC1 SAC1=LogCFL NL L #NC1 NC1 TC0 ACC0 AC0