Polynomial bounds for decoupling, with applications Ryan OβDonnell, Yu Zhao Carnegie Mellon University
Block-multilinearity A homogeneous polynomial function π with degree π is Block-multilinear
Block-multilinearity A homogeneous polynomial function π with degree π is Block-multilinear if we can partition the input variables into π blocks π 1 , β¦ , π π
Block-multilinearity A homogeneous polynomial function π with degree π is Block-multilinear if we can partition the input variables into π blocks π 1 , β¦ , π ' such that each monomial in π contains exactly 1 variable in each block. π π¦ * , π¦ + ,π¦ , , π¦ - = 1 2 π¦ * π¦ + + 1 2 π¦ + π¦ , + 1 2 π¦ , π¦ - β 1 2 π¦ * π¦ - Anti-concentrations of π * = π¦ * , π¦ , , π + = π¦ + , π¦ - Gaussian polynomial Max-E3-Lin-2 [Khot Naor 08, Lovett 10, Kane Meka13, Aaronson Ambainis15] PRG for Liptschitz Classical simulation for functions of polynomials quantum query algorithm
AA Conjecture Let π: {β1,1} 5 β [β1,1] be a bounded Boolean polynomial with degree at most π . Then MaxInf[π] β₯ poly(Var[π]/π) . π π¦ * ,π¦ + ,π¦ , ,π¦ - = 1 2 π¦ * π¦ + + 1 2 π¦ + π¦ , + 1 2 π¦ , π¦ - β 1 Def: J(π) Kπ¦ L π π¦ = I π 2 π¦ * π¦ - Nβ[5] LβN J(π) + Var[π] = 1 Var[π] = I π NPβ Inf * [π] = 1 J(π) + Inf L [π] = I π 2 NβL MaxInf π = 1 MaxInf π = max Lβ[5] {Inf L π } 2
AA Conjecture Let π: {β1,1} 5 β [β1,1] be a bounded Boolean polynomial with degree at most π . Then MaxInf[π] β₯ poly(Var[π]/π) . Suppose AA Conjecture holds: 1. There exists some deterministic simulation of a quantum algorithm; 2. P = P#P implies BQP π΅ β AvgP π΅ with probability 1 for a random oracle π΅ .
AA Conjecture, weak version Let π: {β1,1} 5 β [β1,1] be a Boolean polynomial with degree at most π . Then MaxInf[π] β₯ Var[π] + /exp (π) .
AA Conjecture, weak version Let π: {β1,1} 5 β [β1,1] be a Boolean polynomial with degree at most π . Then MaxInf[π] β₯ Var[π] + /exp (π) . There exists an easy proof for block-multilinear function!! π π§,π¨ = Iπ§ L π L (π¨) L First block Then use hypercontractivity and Cauchy-Schwartz Rest variables
AA Conjecture, weak version Let π: {β1,1} 5 β [β1,1] be a Boolean polynomial with degree at most π . Then MaxInf[π] β₯ Var[π] + /exp (π) . There exists an easy proof for block-multilinear function!! Can we extend this proof to arbitrary Boolean polynomials? Yes, via decoupling!
Decoupling decoupling c π π general polynomial block-multilinear degree π degree π π variables ππ variables ( π blocks of π variables) π copies of π¦ c(π¦, π¦, β¦ ,π¦) 1. π π¦ = π c and π has similar properties 2. π
Examples of decoupling π π¦ * , π¦ + ,π¦ , = π¦ * π¦ + π¦ , c π§ * , π§ + , π§ , , π¨ * , π¨ + , π¨ , , π₯ * , π₯ + ,π₯ , π = * e π§ * π¨ + π₯ ,
Examples of decoupling π π¦ * , π¦ + ,π¦ , = π¦ * π¦ + π¦ , c π§ * , π§ + , π§ , , π¨ * , π¨ + , π¨ , , π₯ * , π₯ + ,π₯ , π = * e π§ * π¨ + π₯ , + * e π§ * π₯ + π¨ ,
Examples of decoupling π π¦ * , π¦ + ,π¦ , = π¦ * π¦ + π¦ , c π§ * , π§ + , π§ , , π¨ * , π¨ + , π¨ , , π₯ * , π₯ + ,π₯ , π = * e π§ * π¨ + π₯ , + * e π§ * π₯ + π¨ , + * e π¨ * π§ + π₯ , + * e π¨ * π₯ + π§ , + * e π₯ * π§ + π¨ , + * e π₯ * π¨ + π§ , c = 1 1 c = Var π π! Var π Inf g h π π! i π Inf j h π
AA Conjecture, weak version Let π: {β1,1} 5 β [β1,1] be a Boolean function with degree at most π . Then MaxInf[π] β₯ Var[π] 2 /exp (π). decoupling c π π c = 1 Var π π! Var π 1 c = MaxInf π π! i π MaxInf π
Decoupling inequality ( π is the degree of π ) Theorem 1. Let Ξ¦: β mn β β mn be convex and non-decreasing. c π¦ * , β¦, π¦ ' E Ξ¦ π β€ E[Ξ¦(π· ' |π(π¦)|)] [de la PeΓ±a 92] Theorem 2. For all π’ > 0 , c π¦ * , β¦ , π¦ ' Pr π > π· ' π’ β€ πΈ ' Pr π(π¦) > π’ [PeΓ±a Montgomery-Smith 95, GinΓ© 98] Comments: 1. π· ' , πΈ ' = exp (π logπ) 2. The inputs can be any independent random variables with all moments finite. 3. The reverse inequality also holds with some worse constants. 4. f does not need to be multilinear neccesarily
AA Conjecture, weak version Let π: {β1,1} 5 β [β1,1] be a Boolean function with degree at most π . Then MaxInf[π] β₯ Var[π] + /exp (π). decoupling c c/π· ' π π π [β1,1] [βπ· ' , π· ' ] [β1,1] c/π· ' = 1 c = 1 c Var π + Var π π· ' = exp π logπ Var π π! Var π π· ' from decoupling inequality 1 c/π· ' = 1 c = c MaxInf π π! i π MaxInf π MaxInf π + MaxInf π π· '
Summary of classical decoupling Advantage: Transfer a general function π to a block- multilinear function. Disadvantage: Introduce an exponential factor on π in decoupling inequality. L
Summary of classical decoupling Sometimes we donβt need the function to be all- blocks-multilinear. We only need π to be a linear map on π§ . π π§,π¨ = Iπ§ L π L (π¨) L First block Then use hypercontractivity and Cauchy-Schwartz Rest variables
One-block-multilinear A polynomial function π with degree π is one- block-multilinear if there exists a subset of the input variables π such that each monomial (except the constant term) in π contains exactly 1 variable in π . π π§,π¨ = Iπ§ L π L (π¨) L
Partial decoupling, with polynomial bounds Our result: Partial decoupling w π π general function One-block-multilinear function degree π degree π π variables 2π variables ( 2 blocks of π variables)
Examples of partial decoupling π π¦ * , π¦ + ,π¦ , = π¦ * π¦ + π¦ , w π§ * ,π§ + , π§ , , π¨ * , π¨ + , π¨ , π = * , π§ * π¨ + π¨ , + * , π¨ * π§ + π¨ , + * , π¨ * π¨ + π§ , w = 1 w = 1 Var π π Var π Inf g h π π + Inf j h π w = π β 1 Inf x h π Inf j h π π +
Partial decoupling, with polynomial bounds Our result: Theorem 1. Let Ξ¦:β mn β β mn be convex and non-decreasing. y π§, π¨ E Ξ¦ π β€ E[Ξ¦(π· ' |π(π¦)|)] Theorem 2. For all π’ > 0 , w π§, π¨ Pr π > π· ' π’ β€ πΈ ' Pr π(π¦) > π’ poly(π) With constants: π(π + ) Boolean π(π ,/+ ) Boolean, homogeneous π· ' = { πΈ ' = exp (π logπ) . π(π) standard Gaussian
AA Conjecture, weak version Let π: {β1,1} 5 β [β1,1] be a Boolean function with degree at most π . Then MaxInf[π] β₯ Var[π] + /exp (π). w w/π· ' π π π [β1,1] [βπ· ' , π· ' ] [β1,1] w/π· ' = 1 w = 1 w Var π + Var π π· ' = poly π Var π π Var π π· ' from new decoupling inequality w β₯ 1 w/π· ' = 1 w MaxInf π π + MaxInf π MaxInf π + MaxInf π π· '
AA Conjecture Let π: {β1,1} 5 β [β1,1] be a Boolean function with degree at most π . Then MaxInf[π] β₯ Var[π] + /poly (π). w w/π· ' π π π [β1,1] [βπ· ' , π· ' ] [β1,1] w/π· ' = 1 w = 1 w Var π + Var π π· ' = poly π Var π π Var π π· ' from new decoupling inequality w β₯ 1 w/π· ' = 1 w MaxInf π π + MaxInf π MaxInf π + MaxInf π π· '
AA Conjecture Let π: {β1,1} 5 β [β1,1] be a Boolean function with degree at most π . Then MaxInf[π] β₯ Var[π] + /poly (π). The conjecture holds for one-block-multilinear functions. π π§,π¨ = Iπ§ L π L (π¨) L
Comparisons Full decoupling Partial decoupling One-block-multilinear Block-multilinear π· ' = exp (π logπ) π· ' = poly (π) General inputs Boolean or Gaussian with all finite moments
Decoupling with polynomial bounds Main result: Prove the decoupling inequalities for one-block decoupling with polynomial bounds. Applications: 1. Give an easy proof for the weak version of AA Conjecture. Show that AA Conjecture holds iff it holds for all one-block-multilinear functions; 2. Generalize a randomized algorithm to arbitrary Boolean functions with the same query complexity; 3. Prove the tight bounds for DFKO Theorems.
Application 2 f :{ β 1,1} n β [ β 1,1] Theorem in [AA15] Let be any bounded block-multilinear Boolean function with degree k . Then there exists a randomized algorithm that, on input , non-adaptively queries 2 O ( k ) ( n / Ξ΅ 2 ) 1-1/ k x β { β 1,1} n bits of x , and then estimate the output of f within error Ξ΅ with high probability. ! / C k ! ! ( x ,..., x ) Ξ΅ ' = Ξ΅ / C k f ( x ) = f f f f C k = 2 O ( k ) [ β 1,1] [ β C k , C k ] [ β 1,1]
Recommend
More recommend