a near optimal depth hierarchy theorem for small depth
play

A Near-Optimal Depth-Hierarchy Theorem for Small-Depth Multilinear - PowerPoint PPT Presentation

A Near-Optimal Depth-Hierarchy Theorem for Small-Depth Multilinear Circuits Nutan Limaye Compuer Science and Engineering Department, Indian Institute of Technology, Bombay, (IITB) India. Joint work with Suryajith Chillara, Christian Engels,


  1. Multilinear polynomials Let X = { x 1 , . . . , x N } . Let p ( X ) ∈ F [ X ] be a degree d multilinear polynomial. � � p ( x ) = c S · x i , i ∈ S S ∈ [ n ]: | S |≤ d Many interesting polynomials are multilinear. σ ∈ S n sgn ( σ ) � n Determinant: Det( X ) = � i =1 x i σ ( i )

  2. Multilinear polynomials Let X = { x 1 , . . . , x N } . Let p ( X ) ∈ F [ X ] be a degree d multilinear polynomial. � � p ( x ) = c S · x i , i ∈ S S ∈ [ n ]: | S |≤ d Many interesting polynomials are multilinear. σ ∈ S n sgn ( σ ) � n Determinant: Det( X ) = � i =1 x i σ ( i ) � n Permanent: Perm( X ) = � i =1 x i σ ( i ) σ ∈ S n

  3. Multilinear polynomials Let X = { x 1 , . . . , x N } . Let p ( X ) ∈ F [ X ] be a degree d multilinear polynomial. � � p ( x ) = c S · x i , i ∈ S S ∈ [ n ]: | S |≤ d Many interesting polynomials are multilinear. σ ∈ S n sgn ( σ ) � n Determinant: Det( X ) = � i =1 x i σ ( i ) � n Permanent: Perm( X ) = � i =1 x i σ ( i ) σ ∈ S n Matrix Multiplication: ( X × Y ) i , j = � n k =1 x ik × y kj

  4. Multilinear polynomials Let X = { x 1 , . . . , x N } . Let p ( X ) ∈ F [ X ] be a degree d multilinear polynomial. � � p ( x ) = c S · x i , i ∈ S S ∈ [ n ]: | S |≤ d Many interesting polynomials are multilinear. σ ∈ S n sgn ( σ ) � n Determinant: Det( X ) = � i =1 x i σ ( i ) � n Permanent: Perm( X ) = � i =1 x i σ ( i ) σ ∈ S n Matrix Multiplication: ( X × Y ) i , j = � n k =1 x ik × y kj

  5. Multilinear formulas A formula is multilinear if every gate in it computes a multilinear polynomial.

  6. Multilinear formulas A formula is multilinear if every gate in it computes a multilinear polynomial. Many tools and techniques A breakthrough result of Raz [Raz04] gave a strong lower bound.

  7. Multilinear formulas A formula is multilinear if every gate in it computes a multilinear polynomial. Many tools and techniques A breakthrough result of Raz [Raz04] gave a strong lower bound. Multilinear formulas for Det/Perm must have superpolynomial size.

  8. Multilinear formulas A formula is multilinear if every gate in it computes a multilinear polynomial. Many tools and techniques A breakthrough result of Raz [Raz04] gave a strong lower bound. Multilinear formulas for Det/Perm must have superpolynomial size. A set of tools introduced in [Raz04].

  9. Multilinear formulas A formula is multilinear if every gate in it computes a multilinear polynomial. Many tools and techniques A breakthrough result of Raz [Raz04] gave a strong lower bound. Multilinear formulas for Det/Perm must have superpolynomial size. A set of tools introduced in [Raz04]. Extended and appended by a line of work. [Raz06,RSY07,RY09,DMPY12,KV17].

  10. Small depth formulas We will focus on small product-depth multilinear formulas.

  11. Small depth formulas We will focus on small product-depth multilinear formulas. Product-depth ∆ = 1: � � or � � � formulas

  12. Small depth formulas We will focus on small product-depth multilinear formulas. Product-depth ∆ = 1: � � or � � � formulas � � formulas are not succinct.

  13. Small depth formulas We will focus on small product-depth multilinear formulas. Product-depth ∆ = 1: � � or � � � formulas � � formulas are not succinct. What about � � � formulas?

  14. Small depth formulas We will focus on small product-depth multilinear formulas. Product-depth ∆ = 1: � � or � � � formulas � � formulas are not succinct. What about � � � formulas? � � p ( x ) = L i , j , where, L i , j are linear polynomials in X . i ∈ [ s ] j ∈ [ s ′ ] The model is surprisingly powerful!

  15. Small depth formulas We will focus on small product-depth multilinear formulas. Product-depth ∆ = 1: � � or � � � formulas � � formulas are not succinct. What about � � � formulas? � � p ( x ) = L i , j , where, L i , j are linear polynomials in X . i ∈ [ s ] j ∈ [ s ′ ] The model is surprisingly powerful! [AV08,Koi09,Tav10,GKKS12] Any polynomial on n variables of degree d computable by a size s circuit can be computed by

  16. Small depth formulas We will focus on small product-depth multilinear formulas. Product-depth ∆ = 1: � � or � � � formulas � � formulas are not succinct. What about � � � formulas? � � p ( x ) = L i , j , where, L i , j are linear polynomials in X . i ∈ [ s ] j ∈ [ s ′ ] The model is surprisingly powerful! [AV08,Koi09,Tav10,GKKS12] Any polynomial on n variables of degree d computable by a size s √ circuit can be computed by � � � formula of size s O ( d ) .

  17. Small depth formulas We will focus on small product-depth multilinear formulas. Product-depth ∆ = 1: � � or � � � formulas � � formulas are not succinct. What about � � � formulas? � � p ( x ) = L i , j , where, L i , j are linear polynomials in X . i ∈ [ s ] j ∈ [ s ′ ] The model is surprisingly powerful! [AV08,Koi09,Tav10,GKKS12] Any polynomial on n variables of degree d computable by a size s √ circuit can be computed by � � � formula of size s O ( d ) . (Assume characteristic 0.)

  18. Small depth formulas We will focus on small product-depth multilinear formulas. Product-depth ∆ = 1: � � or � � � formulas � � formulas are not succinct. What about � � � formulas? � � p ( x ) = L i , j , where, L i , j are linear polynomials in X . i ∈ [ s ] j ∈ [ s ′ ] The model is surprisingly powerful! [AV08,Koi09,Tav10,GKKS12] Any polynomial on n variables of degree d computable by a size s √ circuit can be computed by � � � formula of size s O ( d ) . (Assume characteristic 0.) This � � � realization is non-multilinear!

  19. Product-depth ∆ = 1 Non-multilinear to multilinear formula conversion.

  20. Product-depth ∆ = 1 Non-multilinear to multilinear formula conversion. Let p ( X ) be a multilinear polynomial computable by a � � � formula of size s .

  21. Product-depth ∆ = 1 Non-multilinear to multilinear formula conversion. Let p ( X ) be a multilinear polynomial computable by a � � � formula of size s . Does p ( X ) have a � � � multilinear formula of size s O (1) ?

  22. Product-depth ∆ = 1 Non-multilinear to multilinear formula conversion. Let p ( X ) be a multilinear polynomial computable by a � � � formula of size s . Does p ( X ) have a � � � multilinear formula of size s O (1) ? [Chillara, L , Srinivasan, 18] prove that the answer is no.

  23. Product-depth ∆ = 1 Non-multilinear to multilinear formula conversion. Let p ( X ) be a multilinear polynomial computable by a � � � formula of size s . Does p ( X ) have a � � � multilinear formula of size s O (1) ? [Chillara, L , Srinivasan, 18] prove that the answer is no. Product-depth ∆ = 2 to ∆ = 1 conversion

  24. Product-depth ∆ = 1 Non-multilinear to multilinear formula conversion. Let p ( X ) be a multilinear polynomial computable by a � � � formula of size s . Does p ( X ) have a � � � multilinear formula of size s O (1) ? [Chillara, L , Srinivasan, 18] prove that the answer is no. Product-depth ∆ = 2 to ∆ = 1 conversion Let p ( X ) be a multilinear polynomial computable by a � � � � multilinear formula of size s .

  25. Product-depth ∆ = 1 Non-multilinear to multilinear formula conversion. Let p ( X ) be a multilinear polynomial computable by a � � � formula of size s . Does p ( X ) have a � � � multilinear formula of size s O (1) ? [Chillara, L , Srinivasan, 18] prove that the answer is no. Product-depth ∆ = 2 to ∆ = 1 conversion Let p ( X ) be a multilinear polynomial computable by a � � � � multilinear formula of size s . Does p ( X ) have a � � � multilinear formula of size s O (1) ?

  26. Product-depth ∆ = 1 Non-multilinear to multilinear formula conversion. Let p ( X ) be a multilinear polynomial computable by a � � � formula of size s . Does p ( X ) have a � � � multilinear formula of size s O (1) ? [Chillara, L , Srinivasan, 18] prove that the answer is no. Product-depth ∆ = 2 to ∆ = 1 conversion Let p ( X ) be a multilinear polynomial computable by a � � � � multilinear formula of size s . Does p ( X ) have a � � � multilinear formula of size s O (1) ? [Kayal, Nair, Saha, 15] show that this is not possible.

  27. How expensive ∆ = 2 − → ∆ = 1? Consider � � � � formula of size s .

  28. How expensive ∆ = 2 − → ∆ = 1? Consider � � � � formula of size s . Consider the � � layer � i ∈ [ t ] Q i .

  29. How expensive ∆ = 2 − → ∆ = 1? Consider � � � � formula of size s . Consider the � � layer � i ∈ [ t ] Q i . That is, � �� [ t ] �� � .

  30. How expensive ∆ = 2 − → ∆ = 1? Consider � � � � formula of size s . Consider the � � layer � i ∈ [ t ] Q i . That is, � �� [ t ] �� � . Open up the multiplication of summands as a sum of multiplications.

  31. How expensive ∆ = 2 − → ∆ = 1? Consider � � � � formula of size s . Consider the � � layer � i ∈ [ t ] Q i . That is, � �� [ t ] �� � . Open up the multiplication of summands as a sum of multiplications. � �� [ t ] �� � − → � �� [exp( t )] �� �

  32. How expensive ∆ = 2 − → ∆ = 1? Consider � � � � formula of size s . Consider the � � layer � i ∈ [ t ] Q i . That is, � �� [ t ] �� � . Open up the multiplication of summands as a sum of multiplications. � �� [ t ] �� � − → � �� [exp( t )] �� � − → � [exp( t )] �

  33. How expensive ∆ = 2 − → ∆ = 1? Consider � � � � formula of size s . Consider the � � layer � i ∈ [ t ] Q i . That is, � �� [ t ] �� � . Open up the multiplication of summands as a sum of multiplications. � �� [ t ] �� � − → � �� [exp( t )] �� � − → � [exp( t )] � The conversion incurs an exponential blow-up.

  34. How expensive ∆ = 2 − → ∆ = 1? Consider � � � � formula of size s . Consider the � � layer � i ∈ [ t ] Q i . That is, � �� [ t ] �� � . Open up the multiplication of summands as a sum of multiplications. � �� [ t ] �� � − → � �� [exp( t )] �� � − → � [exp( t )] � The conversion incurs an exponential blow-up. [Kayal, Nair, Saha, 15] show that this exponential blow-up is essential while going from ∆ = 2 to ∆ = 1.

  35. Larger product depth ∆ + 1 − → ∆ Consider � � � � . . . � � � formula of size s and product depth ∆ + 1.

  36. Larger product depth ∆ + 1 − → ∆ Consider � � � � . . . � � � formula of size s and product depth ∆ + 1. Consider the � � layer � i ∈ [ t ] Q i , such that t ≤ s O (1 / ∆) .

  37. Larger product depth ∆ + 1 − → ∆ Consider � � � � . . . � � � formula of size s and product depth ∆ + 1. Consider the � � layer � i ∈ [ t ] Q i , such that t ≤ s O (1 / ∆) . That is, � � . . . � �� [( s O (1 / ∆) )] �� � . . . � � � .

  38. Larger product depth ∆ + 1 − → ∆ Consider � � � � . . . � � � formula of size s and product depth ∆ + 1. Consider the � � layer � i ∈ [ t ] Q i , such that t ≤ s O (1 / ∆) . That is, � � . . . � �� [( s O (1 / ∆) )] �� � . . . � � � . Open up the multiplication of summands as a sum of multiplications.

  39. Larger product depth ∆ + 1 − → ∆ Consider � � � � . . . � � � formula of size s and product depth ∆ + 1. Consider the � � layer � i ∈ [ t ] Q i , such that t ≤ s O (1 / ∆) . That is, � � . . . � �� [( s O (1 / ∆) )] �� � . . . � � � . Open up the multiplication of summands as a sum of multiplications. � � . . . � �� [( s O (1 / ∆) )] �� � . . . � � � − → � � . . . � �� [exp(( s O (1 / ∆) ))] �� � . . . � � �

  40. Larger product depth ∆ + 1 − → ∆ Consider � � � � . . . � � � formula of size s and product depth ∆ + 1. Consider the � � layer � i ∈ [ t ] Q i , such that t ≤ s O (1 / ∆) . That is, � � . . . � �� [( s O (1 / ∆) )] �� � . . . � � � . Open up the multiplication of summands as a sum of multiplications. � � . . . � �� [( s O (1 / ∆) )] �� � . . . � � � − → � � . . . � �� [exp(( s O (1 / ∆) ))] �� � . . . � � � − → � � . . . �� [exp(( s O (1 / ∆) ))] �� . . . � � �

  41. Larger product depth ∆ + 1 − → ∆ Consider � � � � . . . � � � formula of size s and product depth ∆ + 1. Consider the � � layer � i ∈ [ t ] Q i , such that t ≤ s O (1 / ∆) . That is, � � . . . � �� [( s O (1 / ∆) )] �� � . . . � � � . Open up the multiplication of summands as a sum of multiplications. � � . . . � �� [( s O (1 / ∆) )] �� � . . . � � � − → � � . . . � �� [exp(( s O (1 / ∆) ))] �� � . . . � � � − → � � . . . �� [exp(( s O (1 / ∆) ))] �� . . . � � � Careful analysis shows a blow-up of exp( s 1 / ∆+ o (1) ).

  42. Larger product depth ∆ + 1 − → ∆ Consider � � � � . . . � � � formula of size s and product depth ∆ + 1. Consider the � � layer � i ∈ [ t ] Q i , such that t ≤ s O (1 / ∆) . That is, � � . . . � �� [( s O (1 / ∆) )] �� � . . . � � � . Open up the multiplication of summands as a sum of multiplications. � � . . . � �� [( s O (1 / ∆) )] �� � . . . � � � − → � � . . . � �� [exp(( s O (1 / ∆) ))] �� � . . . � � � − → � � . . . �� [exp(( s O (1 / ∆) ))] �� . . . � � � Careful analysis shows a blow-up of exp( s 1 / ∆+ o (1) ). Is the blow-up essential?

  43. Depth hierarchy theorem More Resources

  44. Depth hierarchy theorem More Resources

  45. Depth hierarchy theorem More Resources More power?

  46. Depth hierarchy theorem More Resources More power? More Product-depth

  47. Depth hierarchy theorem More Resources More power? More Product-depth

  48. Depth hierarchy theorem More Resources More power? More Product-depth More power?

  49. Depth hierarchy theorems Arithmetic circuit complexity world

  50. Depth hierarchy theorems Arithmetic circuit complexity world [Raz and Yehudayoff, 2009] prove quasipolynomial depth-hierarchy theorem

  51. Depth hierarchy theorems Arithmetic circuit complexity world [Raz and Yehudayoff, 2009] prove quasipolynomial depth-hierarchy theorem, i.e. they show quasipolynomial size lower bound for converting product-depth ∆ + 1 multilinear formula into product-depth ∆ formula, as long as ∆ is small .

  52. Depth hierarchy theorems Arithmetic circuit complexity world [Raz and Yehudayoff, 2009] prove quasipolynomial depth-hierarchy theorem, i.e. they show quasipolynomial size lower bound for converting product-depth ∆ + 1 multilinear formula into product-depth ∆ formula, as long as ∆ is small . [Kayal, Nair, Saha, 2015] show an exponential size lower bound for converting product-depth 2 multilinear formulas into product-depth 1 formulas.

  53. Depth hierarchy theorems Arithmetic circuit complexity world [Raz and Yehudayoff, 2009] prove quasipolynomial depth-hierarchy theorem, i.e. they show quasipolynomial size lower bound for converting product-depth ∆ + 1 multilinear formula into product-depth ∆ formula, as long as ∆ is small . [Kayal, Nair, Saha, 2015] show an exponential size lower bound for converting product-depth 2 multilinear formulas into product-depth 1 formulas. Boolean circuit complexity world

  54. Depth hierarchy theorems Arithmetic circuit complexity world [Raz and Yehudayoff, 2009] prove quasipolynomial depth-hierarchy theorem, i.e. they show quasipolynomial size lower bound for converting product-depth ∆ + 1 multilinear formula into product-depth ∆ formula, as long as ∆ is small . [Kayal, Nair, Saha, 2015] show an exponential size lower bound for converting product-depth 2 multilinear formulas into product-depth 1 formulas. Boolean circuit complexity world [Ajtai,Frust et al.,Yao, H˚ astad, 1980s] proved quasipolynomial depth-hierarchy theorem.

  55. Depth hierarchy theorems Arithmetic circuit complexity world [Raz and Yehudayoff, 2009] prove quasipolynomial depth-hierarchy theorem, i.e. they show quasipolynomial size lower bound for converting product-depth ∆ + 1 multilinear formula into product-depth ∆ formula, as long as ∆ is small . [Kayal, Nair, Saha, 2015] show an exponential size lower bound for converting product-depth 2 multilinear formulas into product-depth 1 formulas. Boolean circuit complexity world [Ajtai,Frust et al.,Yao, H˚ astad, 1980s] proved quasipolynomial depth-hierarchy theorem. [H˚ astad, 1986] proved exponential depth-hierarchy theorem.

  56. Depth hierarchy theorems Arithmetic circuit complexity world

  57. Depth hierarchy theorems Arithmetic circuit complexity world [Raz and Yehudayoff, 2009] prove quasipolynomial depth-hierarchy theorem

  58. Depth hierarchy theorems Arithmetic circuit complexity world [Raz and Yehudayoff, 2009] prove quasipolynomial depth-hierarchy theorem, i.e. they show quasipolynomial size lower bound for converting product-depth ∆ + 1 multilinear formula into product-depth ∆ formula, as long as ∆ is small .

Recommend


More recommend