Download presentation

Presentation is loading. Please wait.

Published byRomeo Dempster Modified over 3 years ago

1
Lower bounds for small depth arithmetic circuits Chandan Saha Joint work with Neeraj Kayal (MSRI) Nutan Limaye (IITB) Srikanth Srinivasan (IITB)

2
Arithmetic Circuit: A model of computation + xxxx ++++ xxxx …. ….. x1x1 x2x2 x n-1 xnxn f(x 1, x 2, …, x n )--> multivariate polynomial in x 1, …, x n x g h gh + g h g+h Product gate Sum gate There are `field constants’ on the wires

3
Arithmetic Circuit: A model of computation + xxxx ++++ xxxx …. ….. x1x1 x2x2 x n-1 xnxn f(x 1, x 2, …, x n ) Depth = 4

4
Arithmetic Circuit: A model of computation + xxxx ++++ xxxx …. ….. x1x1 x2x2 x n-1 xnxn f(x 1, x 2, …, x n ) Size = no. of gates and wires

5
The lower bound question Is there an explicit family of n-variate, poly(n) degree polynomials {f n } that requires… …super-polynomial in n circuit size ?

6
The lower bound question Is there an explicit family of n-variate, poly(n) degree polynomials {f n } that requires… …super-polynomial in n circuit size ? Note : A random polynomial has super-poly(n) circuit size

7
The Permanent – an explicit family Perm n = ∑ ∏ x i σ(i) σ є S n i є [n]

8
The Permanent – an explicit family Degree of Perm n is low. i.e. bounded by poly(n) Perm n = ∑ ∏ x i σ(i) σ є S n i є [n]

9
The Permanent – an explicit family Degree of Perm n is low. Coefficient of any given monomial can be found efficiently. …given a monomial, there’s a poly-time algorithm to determine the coefficient of the monomial. Perm n = ∑ ∏ x i σ(i) σ є S n i є [n]

10
The Permanent – an explicit family Degree of Perm n is low. Coefficient of any given monomial can be found efficiently. These two properties characterize explicitness Perm n = ∑ ∏ x i σ(i) σ є S n i є [n]

11
The Permanent – an explicit family Degree of Perm n is low. Coefficient of any given monomial can be found efficiently. Define class VNP Perm n = ∑ ∏ x i σ(i) σ є S n i є [n]

12
The Permanent – an explicit family Degree of Perm n is low. Coefficient of any given monomial can be found efficiently. Define class VNP Perm n = ∑ ∏ x i σ(i) σ є S n i є [n] Class VP: Contains families of low degree polynomials {f n } that can be computed by poly(n)-size circuits.

13
The Permanent – an explicit family Degree of Perm n is low. Coefficient of any given monomial can be found efficiently. Perm n = ∑ ∏ x i σ(i) σ є S n i є [n] VP vs VNP: Does Perm n family require super-poly(n) size circuits?

14
A strategy for proving arithmetic circuit lower bound Step 1: Depth reduction Step 2: Lower bound for small depth circuits

15
A strategy for proving arithmetic circuit lower bound Step 1: Depth reduction Step 2: Lower bound for small depth circuits

16
Notations and Terminologies Notations: n = no. of variables in f n d = degree bound on f n = n O(1) Homogeneous polynomial: A polynomial is homogeneous if all its monomials have the same degree (say, d). Homogeneous circuits: A circuit is homogeneous if every gate outputs/computes a homogeneous polynomial. Multilinear polynomial: In every monomial, degree of every variable is at most 1.

17
Reduction to depth ≈ log d Valiant, Skyum, Berkowitz, Rackoff (1983). Homogeneous, degree d, f n computed by poly(n) circuit f n computed by homogeneous poly(n) circuit of depth O(log d) arbitrary depth ≈ log d poly(n)

18
Reduction to depth 4 Agrawal, Vinay (2008); Koiran (2010); Tavenas (2013). Homogeneous, degree d, f n computed by poly(n) circuit f n computed by homogeneous depth 4 circuit of size n O(√d) ≈ log d 4 n O(√d) poly(n)

19
Reduction to depth 4 Agrawal, Vinay (2008); Koiran (2010); Tavenas (2013). Homogeneous, degree d, f n computed by poly(n) circuit f n computed by homogeneous depth 4 circuit of size n O(√d) ≈ log d 4 n O(√d) poly(n) … f n can have n O(d) monomials !

20
A depth 4 circuit + xxxx ++++ xxxx …. ….. x1x1 x2x2 x n-1 xnxn ∑ ∏ ∑ ∏

21
A depth 4 circuit + xxxx ++++ xxxx …. ….. x1x1 x2x2 x n-1 xnxn ∑ ∏ Q ij ij sum of monomials Q ij

22
Reduction to depth 3 Gupta, Kamath, Kayal, Saptharishi (2013); Tavenas (2013). Homogeneous, degree d, f n computed by poly(n) circuit f n computed by depth 3 circuit of size n O(√d) 3 n O(√d) 4

23
Reduction to depth 3 Gupta, Kamath, Kayal, Saptharishi (2013); Tavenas (2013). Homogeneous, degree d, f n computed by poly(n) circuit f n computed by depth 3 circuit of size n O(√d) 3 n O(√d) 4 not homogeneous!

24
A depth 3 circuit + xxxx ++++ …. x1x1 x2x2 x n-1 xnxn ∑ ∏ l ij ij linear polynomial l ij bottom fanin

25
Implication of the depth reductions Let {f n } be an explicit family of polynomials. if f n takes n ω(√d) size homogeneous if f n takes n ω(√d) size VP ≠ VNP or 4 3

26
A strategy for proving arithmetic circuit lower bound Step 1: Depth reduction Step 2: Lower bound for small depth circuits

27
Lower bound for homogeneous depth 4 Theorem: There is a family of homogeneous polynomials {f n } in VNP (with deg f n = d) such that… …any homogeneous depth-4 circuit computing f n has size n Ω(√d) size = n Ω(√d) 4 fnfn

28
Lower bound for homogeneous depth 4 Theorem: There is a family of homogeneous polynomials {f n } in VNP (with deg f n = d) such that… …any homogeneous depth-4 circuit computing f n has size n Ω(√d) size = n Ω(√d) 4 fnfn f n = i ∑ ∏ Q ij … has size n Ω(√d) j sum of monomials

29
Lower bound for homogeneous depth 4 Theorem: There is a family of homogeneous polynomials {f n } in VNP (with deg f n = d) such that… …any homogeneous depth-4 circuit computing f n has size n Ω(√d) size = n Ω(√d) 4 fnfn …joint work with Kayal, Limaye, Srinivasan

30
Lower bound for homogeneous depth 4 Theorem: There is a family of homogeneous polynomials {f n } in VNP (with deg f n = d) such that… …any homogeneous depth-4 circuit computing f n has size n Ω(√d) size = n Ω(√d) 4 fnfn …the technique appears to be using homogeneity crucially

31
Lower bound for depth 3 Theorem: There is a family of homogeneous polynomials {f n } in VNP (with deg f n = d) such that… any depth-3 circuit (bottom fanin ≤ √d) computing f n has size n Ω(√d) size = n Ω(√d) 3 fnfn

32
Lower bound for depth 3 Theorem: There is a family of homogeneous polynomials {f n } in VNP (with deg f n = d) such that… any depth-3 circuit (bottom fanin ≤ √d) computing f n has size n Ω(√d) size = n Ω(√d) 3 fnfn needn’t be homogeneous

33
Lower bound for depth 3 Theorem: There is a family of homogeneous polynomials {f n } in VNP (with deg f n = d) such that… any depth-3 circuit (bottom fanin ≤ √d) computing f n has size n Ω(√d) size = n Ω(√d) 3 fnfn Note: Even for bottom fanin ≤ √d, depth-3 circuits n ω(√d) VP ≠ VNP

34
Lower bound for depth 3 Theorem: There is a family of homogeneous polynomials {f n } in VNP (with deg f n = d) such that… any depth-3 circuit (bottom fanin ≤ t) computing f n has size n Ω(d/t) size = n Ω(d/t) 3 fnfn …joint work with Kayal

35
Lower bound for depth 3 Theorem: There is a family of homogeneous polynomials {f n } in VNP (with deg f n = d) such that… any depth-3 circuit (bottom fanin ≤ t) computing f n has size n Ω(d/t) size = n Ω(d/t) 3 fnfn … answers a question by Shpilka & Wigderson (1999)

36
Proof ideas

37
Homogeneous depth-4 lower bound

38
Complexity measure A measure is a function μ: F[x 1, …, x n ] -> R. We wish to find a measure μ such that 1.If C is a circuit (say, a depth 4 circuit) then μ(C) ≤ s. “small quantity”, where s = size(C) 2.For an “explicit” polynomial f n, μ(f n ) ≥ “large quantity” Implication: If C = f n then s ≥ “large quantity” “small quantity” Upper bound Lower bound

39
Some complexity measures Measure Model Partial derivatives (Nisan & Wigderson) homogeneous depth-3 circuits Evaluation dimension (Raz) multilinear formulas Hessian (Mignon & Ressayre) determinantal complexity permanent Jacobian (Agrawal et. al.) occur-k, depth-4 circuits Incomplete list ?

40
Some complexity measures Measure Model Partial derivatives (Nisan & Wigderson) homogeneous depth-3 circuits Evaluation dimension (Raz) multilinear formulas Hessian (Mignon & Ressayre) determinantal complexity permanent Jacobian (Agrawal et. al.) occur-k, depth-4 circuits Shifted partials (Kayal; Gupta et. al.) homog. depth-4 with low bottom fanin Projected shifted partials homogeneous depth-4 circuits; depth-3 circuits (with low bottom fanin)

41
Space of Partial Derivatives Notations: ∂ =k f : Set of all k th order derivatives of f(x 1, …, x n ) : The vector space spanned by F-linear combinations of polynomials in S Definition: PD k (f) = dim( ) Sub-additive property: PD k (f 1 + f 2 ) ≤ PD k (f 1 ) + PD k (f 2 )

42
Space of Shifted Partials Notation: x =ℓ = Set of all monomials of degree ℓ Definition: SP k,ℓ (f) := dim ( ) Sub-additivity: SP k,ℓ (f 1 + f 2 ) ≤ SP k,ℓ (f 1 ) + SP k,ℓ (f 2 )

43
Space of Shifted Partials Notation: x =ℓ = Set of all monomials of degree ℓ Definition: SP k,ℓ (f) := dim ( ) Sub-additivity: SP k,ℓ (f 1 + f 2 ) ≤ SP k,ℓ (f 1 ) + SP k,ℓ (f 2 ) Why do we expect SP(C) to be small ?

44
Shifted partials – the intuition C = Q 11 Q 12 …Q 1m + … + Q s1 Q s2 …Q sm (homog. depth 4) Q ij = Sum of monomials

45
Shifted partials – the intuition C = Q 11 Q 12 …Q 1m + … + Q s1 Q s2 …Q sm (homog. depth 4) Observation: ∂ =k Q i1 …Q im has “many roots” if k << m << n … any common root of Q i1 …Q im is also a common root of ∂ =k Q i1 …Q im

46
Shifted partials – the intuition C = Q 11 Q 12 …Q 1m + … + Q s1 Q s2 …Q sm (homog. depth 4) Observation: Dimension of the variety of ∂ =k Q i1 …Q im is large if k << m << n

47
Shifted partials – the intuition C = Q 11 Q 12 …Q 1m + … + Q s1 Q s2 …Q sm (homog. depth 4) Observation: Dimension of the variety of ∂ =k Q i1 …Q im is large if k << m << n [Hilbert’s] Theorem (informal): If dimension of the variety of {g} is large then dim ( ) is small.

48
Shifted partials – the intuition C = Q 11 Q 12 …Q 1m + … + Q s1 Q s2 …Q sm (homog. depth 4) Observation: Dimension of the variety of ∂ =k Q i1 …Q im is large if k << m << n [Hilbert’s] Theorem (informal): If dimension of the variety of {g} is large then dim ( ) is small. … so we expect SP k,ℓ (Q i1 …Q im ) to be a `small quantity’

49
Shifted partials – the intuition C = Q 11 Q 12 …Q 1m + … + Q s1 Q s2 …Q sm (homog. depth 4) Observation: Dimension of the variety of ∂ =k Q i1 …Q im is large if k << m << n [Hilbert’s] Theorem (informal): If dimension of the variety of {g} is large then dim ( ) is small. … by subadditivity, SP k,ℓ (C) ≤ s. `small quantity’

50
Depth-4 with low bottom degree C = Q 11 Q 12 …Q 1m + … + Q s1 Q s2 …Q sm (homog. depth 4) Q ij = Sum of monomials of degree ≤ t (w.l.o.g m ≤ 2d/t )

51
Depth-4 with low bottom degree C = Q 11 Q 12 …Q 1m + … + Q s1 Q s2 …Q sm ∂ =k Q i1 …Q im = Q i1 Q i2 …Q ik …Q im + Q i1 Q i2 …Q ik Q i k+1 …Q im + … X...... = Q i k+1 … Q im + Q i1 Q i k+2 … Q im + … degree ≤ k.t

52
Depth-4 with low bottom degree C = Q 11 Q 12 …Q 1m + … + Q s1 Q s2 …Q sm ∂ =k Q i1 …Q im = Q i1 Q i2 …Q ik …Q im + Q i1 Q i2 …Q ik Q i k+1 …Q im + … X...... = Q i k+1 … Q im + Q i1 Q i k+2 … Q im + … at most ( ) terms m k

53
Depth-4 with low bottom degree C = Q 11 Q 12 …Q 1m + … + Q s1 Q s2 …Q sm ∂ =k Q i1 …Q im = Q i1 Q i2 …Q ik …Q im + Q i1 Q i2 …Q ik Q i k+1 …Q im + … X...... = Q i k+1 … Q im + Q i1 Q i k+2 … Q im + … u. ∂ =k Q i1 …Q im = Q i k+1 … Q im + Q i1 Q i k+2 … Q im + … X degree = ℓ degree ≤ ℓ + k.t

54
Depth-4 with low bottom degree C = Q 11 Q 12 …Q 1m + … + Q s1 Q s2 …Q sm ∂ =k Q i1 …Q im = Q i1 Q i2 …Q ik …Q im + Q i1 Q i2 …Q ik Q i k+1 …Q im + … X...... = Q i k+1 … Q im + Q i1 Q i k+2 … Q im + … u. ∂ =k Q i1 …Q im = Q i k+1 … Q im + Q i1 Q i k+2 … Q im + … X n + ℓ + kt n m k SP k,ℓ (Q i1 …Q im ) ≤ ( ). ( )

55
Depth-4 with low bottom degree C = Q 11 Q 12 …Q 1m + … + Q s1 Q s2 …Q sm ∂ =k Q i1 …Q im = Q i1 Q i2 …Q ik …Q im + Q i1 Q i2 …Q ik Q i k+1 …Q im + … X...... = Q i k+1 … Q im + Q i1 Q i k+2 … Q im + … u. ∂ =k Q i1 …Q im = Q i k+1 … Q im + Q i1 Q i k+2 … Q im + … X n + ℓ + kt n m k SP k,ℓ (C) ≤ s. ( ). ( ) Upper bound

56
Reduction to low bottom degree C = Q 11 Q 12 …Q 1m + … + Q s1 Q s2 …Q sm (homog. depth 4) Q ij = Sum of monomials (NO degree restriction)

57
Reduction to low bottom degree C = Q 11 Q 12 …Q 1m + … + Q s1 Q s2 …Q sm Idea: Reduce to the case of low bottom degree using Random restriction Multilinear projection

58
Reduction to low bottom degree C = Q 11 Q 12 …Q 1m + … + Q s1 Q s2 …Q sm Random restriction: Set every variable to zero independently at random with a certain probability. …denoted naturally by a map σ

59
Reduction to low bottom degree C = Q 11 Q 12 …Q 1m + … + Q s1 Q s2 …Q sm Random restriction: Set every variable to zero independently at random with a certain probability. …denoted naturally by a map σ σ(C) = σ(Q 11 ) σ(Q 12 )…σ(Q 1m ) + … + σ(Q s1 ) σ(Q s2 )…σ(Q sm ) Obs: If a monomial u has many variables (high support) then σ(u) = 0 w.h.p

60
Reduction to low bottom degree C = Q 11 Q 12 …Q 1m + … + Q s1 Q s2 …Q sm Random restriction: Set every variable to zero independently at random with a certain probability. …denoted naturally by a map σ σ(C) = σ(Q 11 ) σ(Q 12 )…σ(Q 1m ) + … + σ(Q s1 ) σ(Q s2 )…σ(Q sm ) w.l.o.g σ(Q ij ) = sum of ‘low support’ monomials

61
Reduction to low bottom degree C = Q 11 Q 12 …Q 1m + … + Q s1 Q s2 …Q sm Random restriction: Set every variable to zero independently at random with a certain probability. Homogeneous depth 4 homogenous depth 4 with low bottom support … w.l.o.g assume that C has low bottom support

62
Reduction to low bottom degree C = Q 11 Q 12 …Q 1m + … + Q s1 Q s2 …Q sm Projection map: π (g) = sum of the multilinear monomials in g

63
Reduction to low bottom degree C = Q 11 Q 12 …Q 1m + … + Q s1 Q s2 …Q sm Projection map: π (g) = sum of the multilinear monomials in g Observation: π (sum of ‘low support’ monomials) = sum of ‘low degree’ monomials

64
Reduction to low bottom degree C = Q 11 Q 12 …Q 1m + … + Q s1 Q s2 …Q sm Projection map: π (g) = sum of the multilinear monomials in g Observation: π (Q ij ) = sum of ‘low degree’ monomials

65
Projected Shifted Partials PSP k,ℓ (f) := dim (π (x =ℓ. ∂ =k f) ) (obeys subadditivity)

66
Projected Shifted Partials PSP k,ℓ (f) := dim (π (x =ℓ. ∂ =k f) ) (obeys subadditivity) multilinear shifts only!

67
Projected Shifted Partials PSP k,ℓ (f) := dim (π (x =ℓ. ∂ =k f) ) (obeys subadditivity) multilinear derivatives!

68
Depth-4 with low bottom support C = Q 11 Q 12 …Q 1m + … + Q s1 Q s2 …Q sm support of every monomial bounded by t

69
Depth-4 with low bottom support C = Q 11 Q 12 …Q 1m + … + Q s1 Q s2 …Q sm Q ij = Q’ ij + Every variable in every monomial has degree 2 or less

70
Depth-4 with low bottom support C = Q 11 Q 12 …Q 1m + … + Q s1 Q s2 …Q sm Q ij = Q’ ij + Every monomial has a variable with degree 3 or more

71
Depth-4 with low bottom support C = Q 11 Q 12 …Q 1m + … + Q s1 Q s2 …Q sm Q ij = Q’ ij + Q i1 Q i2 …Q im = Q’ i1 Q’ i2 …Q’ im + Every monomial has a variable with degree 3 or more

72
Depth-4 with low bottom support C = Q 11 Q 12 …Q 1m + … + Q s1 Q s2 …Q sm Q ij = Q’ ij + Q i1 Q i2 …Q im = Q’ i1 Q’ i2 …Q’ im + PSP k,ℓ (Q i1 Q i2 …Q im ) ≤ PSP k,ℓ (Q’ i1 Q’ i2 …Q’ im ) + PSP k,ℓ ( )

73
Depth-4 with low bottom support C = Q 11 Q 12 …Q 1m + … + Q s1 Q s2 …Q sm Q ij = Q’ ij + Q i1 Q i2 …Q im = Q’ i1 Q’ i2 …Q’ im + PSP k,ℓ (Q i1 Q i2 …Q im ) ≤ PSP k,ℓ (Q’ i1 Q’ i2 …Q’ im ) + PSP k,ℓ ( ) 0

74
Depth-4 with low bottom support C = Q 11 Q 12 …Q 1m + … + Q s1 Q s2 …Q sm Q ij = Q’ ij + Q i1 Q i2 …Q im = Q’ i1 Q’ i2 …Q’ im + PSP k,ℓ (Q i1 Q i2 …Q im ) ≤ PSP k,ℓ (Q’ i1 Q’ i2 …Q’ im ) + PSP k,ℓ ( ) 0 degree ≤ 2t

75
Depth-4 with low bottom support C = Q 11 Q 12 …Q 1m + … + Q s1 Q s2 …Q sm Q ij = Q’ ij + Q i1 Q i2 …Q im = Q’ i1 Q’ i2 …Q’ im + PSP k,ℓ (Q i1 Q i2 …Q im ) ≤ PSP k,ℓ (Q’ i1 Q’ i2 …Q’ im ) Abusing notation: Call Q’ ij as Q ij

76
Depth-4 with low bottom support ∂ =k Q i1 …Q im = Q i1 Q i2 …Q ik …Q im + Q i1 Q i2 …Q ik Q i k+1 …Q im + … X...... = Q i k+1 … Q im + Q i1 Q i k+2 … Q im + … degree ≤ 2kt

77
Depth-4 with low bottom support ∂ =k Q i1 …Q im = Q i1 Q i2 …Q ik …Q im + Q i1 Q i2 …Q ik Q i k+1 …Q im + … X...... = Q i k+1 … Q im + Q i1 Q i k+2 … Q im + … u. ∂ =k Q i1 …Q im = u. Q i k+1 … Q im + u. Q i1 Q i k+2 … Q im + X degree = ℓ degree ≤ 2kt

78
Depth-4 with low bottom support ∂ =k Q i1 …Q im = Q i1 Q i2 …Q ik …Q im + Q i1 Q i2 …Q ik Q i k+1 …Q im + … X...... = Q i k+1 … Q im + Q i1 Q i k+2 … Q im + … π(u.∂ =k Q i1 …Q im ) = π( Q i k+1 … Q im ) + π( Q i1 Q i k+2 … Q im ) + X multilinear, degree ≤ ℓ + 2k.t

79
Depth-4 with low bottom support ∂ =k Q i1 …Q im = Q i1 Q i2 …Q ik …Q im + Q i1 Q i2 …Q ik Q i k+1 …Q im + … X...... = Q i k+1 … Q im + Q i1 Q i k+2 … Q im + … π(u.∂ =k Q i1 …Q im ) = π( Q i k+1 … Q im ) + π( Q i1 Q i k+2 … Q im ) + X Upper bound ℓ + 2kt n m k SP k,ℓ (C) ≤ s. ( ). ( )

80
How large can PSP(f) be? Trivially, PSP k,ℓ (f) ≤ min { ( ).( ), ( ) } n k n ℓ n ℓ + d - k

81
How large can PSP(f) be? Trivially, PSP k,ℓ (f) ≤ min { ( ).( ), ( ) } n k n ℓ n ℓ + d - k Size of the set { x =ℓ. ∂ =k f } ≤ ( ).( ) Number of monomials in any polynomial in π (x =ℓ. ∂ =k f) ≤ ( ) n k n ℓ n ℓ + d - k Let f be a multilinear polynomial

82
How large can PSP(f) be? Trivially, PSP k,ℓ (f) ≤ min { ( ).( ), ( ) } Best lower bound for s s ≥ n k n ℓ n ℓ + d - k min {( ).( ), ( )} ( ).( ) m k n ℓ + 2kt n k n ℓ n ℓ + d - k = n Ω(d/t) After setting k and ℓ appropriately

83
How large can PSP(f) be? Trivially, PSP k,ℓ (f) ≤ min { ( ).( ), ( ) } Best lower bound for s s ≥ There’s an explicit f such that PSP k,ℓ (f) is close to the trivial upper bound. (lower bound) n k n ℓ n ℓ + d - k min {( ).( ), ( )} ( ).( ) m k n ℓ + 2kt n k n ℓ n ℓ + d - k = n Ω(d/t)

84
Depth-3 lower bound

85
Trading depth for homogeneity Idea: Depth-3 with low bottom fanin Homogeneous depth-4 with low bottom support Size = s Bottom fanin = t 3 fnfn 4 (homogeneous) fnfn Size = s. 2 O(√d) Bottom support = t

86
Depth-3 to Depth-4 Implicit in Shpilka & Wigderson ; Hrubes & Yehudayoff (2011) C = α 1.(1 + l 11 )(1 + l 12 )…(1 + l 1m ) + …. + α s.(1 + l s1 )(1 + l s2 )…(1 + l sm ) linear forms field constants

87
Depth-3 to Depth-4 Implicit in Shpilka & Wigderson ; Hrubes & Yehudayoff (2011) C = (1 + l 11 )(1 + l 12 )…(1 + l 1m ) + …. + (1 + l s1 )(1 + l s2 )…(1 + l sm ) Notation: [g] d = d-th homogeneous part of g Easy observation: If C = f, which is homogeneous deg d polynomial, then [C] d = f.

88
Depth-3 to Depth-4 Implicit in Shpilka & Wigderson ; Hrubes & Yehudayoff (2011) C = (1 + l 11 )(1 + l 12 )…(1 + l 1m ) + …. + (1 + l s1 )(1 + l s2 )…(1 + l sm ) [C] d = [(1 + l 11 )(1 + l 12 )…(1 + l 1m )] d +….+ [(1 + l s1 )(1 + l s2 )…(1 + l sm )] d idea: transform these to homogeneous depth-4

89
Newton’s identities E d (y 1, y 2, …, y m ) := ∑ ∏ y j P r (y 1, y 2, …, y m ) := ∑ y j r S in 2 [m] |S| = d j in S (elementary symmetric polynomial of degree d) j in [m] (power symmetric polynomial of degree r)

90
Newton’s identities E d (y 1, y 2, …, y m ) := ∑ ∏ y j P r (y 1, y 2, …, y m ) := ∑ y j r S in 2 [m] |S| = d j in S j in [m] Lemma: E d (y) = ∑ β a ∏ P r (y) a = (a 1, …, a d ) ∑ r. a r = d r in [d] arar e.g. 2y 1 y 2 = (y 1 + y 2 ) 2 – y 1 2 – y 2 2 = P 1 2 – P 2 field constant

91
Newton’s identities E d (y 1, y 2, …, y m ) := ∑ ∏ y j P r (y 1, y 2, …, y m ) := ∑ y j r S in 2 [m] |S| = d j in S j in [m] Lemma: E d (y) = ∑ β a ∏ P r (y) a = (a 1, …, a d ) ∑ r. a r = d r in [d] arar Hardy-Ramanujan estimate: The number of a = (a 1, …, a d ) such that ∑ r.a r = d is 2 O(√d)

92
Depth-3 to Depth-4 Implicit in Shpilka & Wigderson ; Hrubes & Yehudayoff (2011) [(1 + l i1 )(1 + l i2 )…(1 + l im )] d = E d ( l i1, …, l im ) = ∑ β a ∏ P r ( l i1, …, l im ) a = (a 1, …, a d ) ∑ r. a r = d r in [d] arar 2 O(√d) summands

93
Depth-3 to Depth-4 Implicit in Shpilka & Wigderson ; Hrubes & Yehudayoff (2011) [(1 + l i1 )(1 + l i2 )…(1 + l im )] d = E d ( l i1, …, l im ) = ∑ β a ∏ P r ( l i1, …, l im ) a = (a 1, …, a d ) ∑ r. a r = d r in [d] arar 2 O(√d) summands Suppose every l ij has at most t variables, then…

94
Depth-3 to Depth-4 Implicit in Shpilka & Wigderson ; Hrubes & Yehudayoff (2011) [(1 + l i1 )(1 + l i2 )…(1 + l im )] d = E d ( l i1, …, l im ) = ∑ β a ∏ P r ( l i1, …, l im ) a = (a 1, …, a d ) ∑ r. a r = d r in [d] arar = ∑ β a ∏ Q i,a,r a = (a 1, …, a d ) ∑ r. a r = d r in [d] every monomial has support ≤ t

95
Depth-3 to Depth-4 Implicit in Shpilka & Wigderson ; Hrubes & Yehudayoff (2011) [(1 + l i1 )(1 + l i2 )…(1 + l im )] d = E d ( l i1, …, l im ) = ∑ β a ∏ P r ( l i1, …, l im ) a = (a 1, …, a d ) ∑ r. a r = d r in [d] arar = ∑ β a ∏ Q i,a,r a = (a 1, …, a d ) ∑ r. a r = d r in [d] [C] d = ∑ ∑ β a ∏ Q i,a,r a = (a 1, …, a d ) ∑ r. a r = d r in [d] i in [s]

96
Depth-3 to Depth-4 Implicit in Shpilka & Wigderson ; Hrubes & Yehudayoff (2011) [(1 + l i1 )(1 + l i2 )…(1 + l im )] d = E d ( l i1, …, l im ) = ∑ β a ∏ P r ( l i1, …, l im ) a = (a 1, …, a d ) ∑ r. a r = d r in [d] arar = ∑ β a ∏ Q i,a,r a = (a 1, …, a d ) ∑ r. a r = d r in [d] [C] d = ∑ ∑ β a ∏ Q i,a,r a = (a 1, …, a d ) ∑ r. a r = d r in [d] i in [s] Homogeneous depth-4 with low bottom support and size s.2 Ω(√d)

97
An explicit family with high PSP k,ℓ

98
An explicit family of polynomials Nisan-Wigderson family of polynomials: NW r := ∑ ∏ x i, h(i) d2d2 h(z) in F [z], deg(h) ≤ r i in [d] identifying the elements of F with {1,2, …, d 2 } d2d2

99
An explicit family of polynomials Nisan-Wigderson family of polynomials: NW r := ∑ ∏ x i, h(i) d2d2 h(z) in F [z], deg(h) ≤ r i in [d] `Disjointness’ property: Two monomials can share at most r ≈ d/3 variables. = + + … d r r d 2(r+1) monomials

100
Projected Shifted Partials of NW r The set π (x =ℓ. ∂ =k NW r ) has ( ).( ) elements. Every polynomial in π (x =ℓ. ∂ =k NW r ) is multilinear & homogeneous of degree (ℓ + d – k). n k n ℓ

101
Projected Shifted Partials of NW r The set π (x =ℓ. ∂ =k NW r ) has ( ).( ) elements. Every polynomial in π (x =ℓ. ∂ =k NW r ) is multilinear & homogeneous of degree (ℓ + d – k). PSP k,ℓ (NW r ) = rank (M) n k n ℓ M := ( ).( ) rows π (x =ℓ. ∂ =k NW r ) (0/1)-matrix of coefficients n ℓ + d - k ( ) columns n k n ℓ

102
Projected Shifted Partials of NW r Because of the `disjointness property’ of NW r, the columns of M are almost orthogonal. Hence, B := M T M is diagonally dominant. Observe, rank (M) ≥ rank (B).

103
Projected Shifted Partials of NW r Because of the `disjointness property’ of NW r, the columns of M are almost orthogonal. Hence, B := M T M is diagonally dominant. Observe, rank (M) ≥ rank (B). Alon’s rank bound (for diagonally dominant matrix): If B is a real symmetric matrix then rank (B) ≥ Tr (B) 2 Tr (B 2 )

104
Projected Shifted Partials of NW r [Main lemma]: Using Alon’s bound and settings r, k and ℓ appropriately, PSP k,ℓ (NW r ) ≥ η. min {( ).( ), ( )} n k n ℓ n ℓ + d - k small factor

105
An explicit family in VP [Kumar-Saraf (2014)] : Showed the same lower bound using the Iterated Matrix multiplication polynomial, which is in VP

106
An explicit family in VP [Kumar-Saraf (2014)] : Showed the same lower bound using the Iterated Matrix multiplication polynomial, which is in VP VNP Circuits (VP) ABPs Formulas Depth-4 exponential separation

107
An explicit family in VP [Kumar-Saraf (2014)] : Showed the same lower bound using the Iterated Matrix multiplication polynomial, which is in VP VNP Circuits (VP) ABPs Formulas Open: separation ? …known in the multilinear setting [Dvir, Malod, Perifel, Yehudayoff (2012)]

108
An explicit family in VP [Kumar-Saraf (2014)] : Showed the same lower bound using the Iterated Matrix multiplication polynomial, which is in VP VNP Circuits (VP) ABPs Formulas Open: separation ? …improve n Ω(√d) to n ω(√d)

109
Some other open questions 1.Prove a n Ω(√d) lower bound for general depth-3 circuits (i.e. without the low bottom fanin restriction).

110
Some other open questions 1.Prove a n Ω(√d) lower bound for general depth-3 circuits. 2.Prove a n Ω(√d) lower bound for homogeneous depth-5 circuits. [open problem in Nisan & Wigderson (1996)] (2) (1)

111
Some other open questions 1.Prove a n Ω(√d) lower bound for general depth-3 circuits. 2.Prove a n Ω(√d) lower bound for homogeneous depth-5 circuits. 3.Prove a n Ω(d) lower bound for multilinear depth-3 circuits. (current best is 2 Ω(d) ) …interestingly, one can get this using PSP measure

112
Some other open questions 1.Prove a n Ω(√d) lower bound for general depth-3 circuits. 2.Prove a n Ω(√d) lower bound for homogeneous depth-5 circuits. 3.Prove a n Ω(d) lower bound for multilinear depth-3 circuits. 4.A separation between homogeneous formulas and homogeneous depth-4 formulas.

113
Some other open questions 1.Prove a n Ω(√d) lower bound for general depth-3 circuits. 2.Prove a n Ω(√d) lower bound for homogeneous depth-5 circuits. 3.Prove a n Ω(d) lower bound for multilinear depth-3 circuits. 4.A separation between homogeneous formulas and homogeneous depth-4 formulas. 5.A separation between homogeneous formulas and multilinear homogeneous formulas. …exhibiting the power of non-multilinearity

114
Some other open questions 1.Prove a n Ω(√d) lower bound for general depth-3 circuits. 2.Prove a n Ω(√d) lower bound for homogeneous depth-5 circuits. 3.Prove a n Ω(d) lower bound for multilinear depth-3 circuits. 4.A separation between homogeneous formulas and homogeneous depth-4 formulas. 5.A separation between homogeneous formulas and multilinear homogeneous formulas. Thanks!

Similar presentations

OK

THE FUNDAMENTAL THEOREM OF ALGEBRA. Descartes’ Rule of Signs If f(x) is a polynomial function with real coefficients, then *The number of positive real.

THE FUNDAMENTAL THEOREM OF ALGEBRA. Descartes’ Rule of Signs If f(x) is a polynomial function with real coefficients, then *The number of positive real.

© 2018 SlidePlayer.com Inc.

All rights reserved.

To ensure the functioning of the site, we use **cookies**. We share information about your activities on the site with our partners and Google partners: social networks and companies engaged in advertising and web analytics. For more information, see the Privacy Policy and Google Privacy & Terms.
Your consent to our cookies if you continue to use this website.

Ads by Google

A ppt on android Ppt on object-oriented programming concepts in c++ Best ppt on forest society and colonialism in india Ppt on some new technology Ppt on machine translation online Ppt on pin diode application Ppt on suspension type insulators Ppt on polynomials for class 9 free download Ppt on quadrilaterals and its properties for class 9 Ppt on principles of object-oriented programming encapsulation