Handwritten Text Recognition for manuscripts and early printed texts
Survey on Polynomial Identity Testing Reveals Important Connections Between Circuit Lower Bounds and PIT Algorithms
1. Survey on Polynomial Identity Testing Amir Shpilka Technion and MSR NE June 18, 2011 PIT Survey – CSR 1
2. Exam n=1. Is the following polynomial identically 0? June 18, 2011 2 Prove it! Will do so later.
3. Goal of talk Survey known results Explain some proof techniques Give an interesting set of `accessible’ open questions June 18, 2011 3
4. Talk outline Definition of the problem Connection to lower bounds (hardness) Survey of positive results Some proofs Connection to polynomial factorization June 18, 2011 4
5. Playground: Arithmetic Circuits Field: Variables: X1,...,Xn Gates: , Every gate in the circuit computes a polynomial in [X1,...,Xn] Example:(X1 X2) (X2 + 1) Size = number of wires Depth = length of longest input-output path Degree = max degree of internal gates June 18, 2011 5
7. Bounded depth circuits circuits: depth-2 circuits with + at the top and at the bottom. Size s circuits compute s-sparse polynomials. circuits: depth-3 circuits with + at the top, at the middle and + at the bottom. Compute sums of products of linear functions. I.e. a sparse polynomial composed with a linear transformation. circuits: depth-4 circuits. Compute sums of products of sparse polynomials. June 18, 2011 7
8. Why Arithmetic Circuits? Most natural model for computing polynomials For many problems (e.g. Matrix Multiplication, Det) best algorithm is an arithmetic circuit Great algorithmic achievements: Fourier Transform Matrix Multiplication Polynomial Factorization Structured model (compared to Boolean circuits) P vs. NP may be easier June 18, 2011 8
9. Polynomial Identity Testing Input:Arithmetic circuit computing f Problem: Is f0? f(x1,...,xn) C x1 x2 xn Randomized algorithm [Schwartz, Zippel, DeMillo-Lipton]: evaluate f at a random point Goal: A proof. I.e., a deterministic algorithm June 18, 2011 9
10. Black Box PIT Explicit Hitting Set Input: A Black-Box circuit computing f. Problem:Is f=0 ? f(x1,...,xn) (x1,...,xn) C Goal: deterministic algorithm (a.k.a. Hitting Set) S,Z,DM-L: small Hitting Set (not explicit) June 18, 2011 10
11. Motivation Natural and fundamental problem Strong connection to circuit lower bounds Algorithmic importance: Primality testing [Agrawal-Kayal-Saxena] Parallel algorithms for finding matching [Karp-Upfal-Wigderson, Mulmuley-Vazirani-Vazirani] May help you solve exams! June 18, 2011 11
12.
13. Hardness: PIT lower bounds [Kabanets-Impagliazzo]: 2(n) lower bound for Permanent PIT in npolylog(n)time PIT P super-polynomial lower bounds: Boolean for NEXP or arithmetic for Permanent [Dvir-S-Yehudayoff]:(almost) same as K-I for bounded depth circuits [Heintz-Schnorr,Agrawal]: Polynomial time Black-Box PIT Exponential lower bounds for arithmetic circuits Lesson: Derandomizing PIT essentially equivalent to proving lower bounds for arithmetic circuits June 18, 2011 13
14. Black-Box PIT Lower Bounds [Heintz-Schnorr,Agrawal]: BB PIT for size s circuits in time poly(s) (i.e. poly(s) size hitting set) exp. lower bounds for arithmetic circuits. Proof: Given ={pi}, find non-zero polynomial f inlog(||)variablse, such that f(pi)=0 for all i. f does not have size s circuits Gives lower bounds for f in EXP (PSPACE) Conjecture[Agrawal]: ={(y1,…, yn) : yi=yki mod r, k,r < s20} is a hitting set for size s circuits June 18, 2011 14
15. Lower Bounds Black-Box PIT [Kabanets-Impagliazzo]: arithmetic version of N-W generator: Combinatorial design: S1,…, Sn [m]s.t. |Si| = k and |Si Sj| < log n Assumef is a k-variate polynomial of arithmetic circuit complexity 2(k) For every ymgeneratewnas wi = f(y|Si) I.e. = {(f(y|S1),…, f(y|Sn) : ym} Proof Idea: C0 iff (roughly) for some i, xi – f(y|Si) is a factor of C. Gives circuit for f. June 18, 2011 15
16. Importance of circuits [Agrawal-Vinay,Raz]: Exponential lower bounds for circuitsimply exponential lower bounds for general circuits. Proof: 1. Depth reduction a-la P=NC2[Valiant-Skyum-Berkowitz-Rackoff] 2. Break the circuit in the middle and interpolate each part using circuits. Cor[Agrawal-Vinay]: Polynomial time PIT of circuits gives quasi-polynomial time PIT for general circuits. Proof: By [Heintz-Schnorr,Agrawal] polynomial time PIT exponential lower bounds for circuits. By [Agrawal-Vinay, Raz] exponential lower bounds for general circuits. Now use [K-I]. June 18, 2011 16
17.
18. Connection to lower bounds (hardness)Survey of positive results Some proofs Connection to polynomial factorization June 18, 2011 17
19. Deterministic algorithms for PIT circuits (a.k.a.,sparse polys) [BenOr-Tiwari, Grigoriev-Karpinski, Klivans-Spielman,…] Black-Box in polynomial time Non-commutative formulas [Raz-S] White-Box in polynomial time (k) circuits [Dvir-S,Kayal-Saxena,Karnin-S,Kayal-Saraf,Saxena-Seshadri] Black-Box in time nO(k) Read-k multilinear formulas [S-Volkovich, Anderson-van Melkebeek-Volkovich] White-Box in time nkO(k) Black-Box in nO(log(n)+kO(k)) Multilinear (k)[Karnin-Mukhopadhyay-S-Volkovich, Saraf-Volkovich] Black-Box in time npoly(k) June 18, 2011 18
20. Why study restricted models [Agrawal-Vinay] PIT for circuits implies PIT for general depth. Gaining insight into more general questions: Intuitively:lower bounds imply PIT Multilinear formulas: super polynomial bounds [Raz,Raz-Yehudayoff] but no PIT algorithms Not even for Depth-3 multilinear formulas! Read-k, depth-3,4 multilinear formulas relaxations of the more general problem Interesting results: Structural theorems for (k) and (k) circuits. June 18, 2011 19
23. Survey of positive resultsSome proofs: Depth-3 circuits Depth-4 circuits Connection to polynomial factorization June 18, 2011 20
24. Proofs – tailored for the model Proofs usually use `weakness’ inherent in model Depth 2: few monomials. Substituting yai to xiwe can control `collapses’ of different monomials. Non Commutative formulas: Polynomial has few linearly independent partial derivatives [Nisan]. Keep track of a basis for derivatives to do PIT. (k): setting a linear function to zero reduces top fan-in. If k=2 then multiplication gates must be the same. Calls for induction. Multilinear(k): in some sense `combination’ of sparse polynomials and multilnear(k). Read-k-Formulas: subformulas of root contain ½ of variables and are (roughly) read-(k-1). June 18, 2011 21
25. Solution to Exam n=1. Is the following polynomial identically 0? June 18, 2011 22 Prove it! Will do so later now
26.
27. Depth 3 identities C = M1+ … + MkMi = j=1...diLi,j Rank: dimension of space spanned by {Li,j} In the exam: Rank=3 Turns out: this is (almost) the general case! Theorem [Dvir S]: If C 0 is a basic identity then dim(C) ≤ Rank(k,d) = (log(d))k White-Box Algorithm: find partition to sub-circuits of low dimension (after removal of g.c.d.) and brute force verify that they vanish. Improved nO(k) algorithm by [Kayal-Saxena]. June 18, 2011 25
28. Black-Box PIT Black-Box Algorithm[KarninS]:Intuitively, if we project the inputs to a `low’ dimensional space in a way that does not collapse the dimension below Rank(k,d) then the circuit should not become zero. Theorem [GabizonRaz]: "small" explicit set of D-dimensional subspaces V1,...,Vm such that space of linear functions :dim(|Vi) = min(dim(),D)for most i June 18, 2011 26 In other words: the linear functions in remain as independent as possible on Vi
29. Black-Box PIT If C’ has the same rank as C’|V, then C’ and C’|V are isomorphic. Hence, C’|V 0 C 0 Corollary: i, C|Vihas low "rank“ Chas low "rank" Corollary: if i, C|Vi 0 then Chas structure (i.e. Cis sum of circuits of low "rank") Theorem: if i, C|Vi 0 then C 0. Algorithm: For every i, brute force compute C|Vi Time: poly(n)ddim(Vi) = dO(Rank(k,d)) If Chas high rank then by Gabizon-Raz, for some i,C|Vi has high rank. If Cis not a sum of low rank circuits then for some i, C|Viis not a sum of low rank circuits. This contradicts the structural theorem. Cis sum of low rank subcircuitsVis.t. rank of subcircuits remain the same. C|Viis zero each subcircuit vanishes on Vi. subcircuitscompute the zero polynomial. 27
30. Depth 3 identities Lesson 1: depth 3 identities arevery structured! Lesson 2: Rank is an important invariant to study. Improvements [Kayal-Saraf,Saxena-Seshadri]: finite, klog(d) < Rank(k,d) < k3log(d) over , k < Rank(k,d) < k2log(k) Improves [Dvir-S] + [Karnin-S] (plug and play) NEW: [Saxena-Seshadri] BB-PIT in time nO(k) June 18, 2011 28
31. Bounding the rank Basic observation: Consider C = M1 + M2 M1 = M2= Fact: linear functions are irreducible polynomial. Corollary: C 0 then M1, M2 have same factors. Corollary: matching i (i) s.t. Li ~ L'(i)
34. Black-Box PIT for multilinear (k) C = T1+ … + Tk Ti = j=1...diPi,j where Ti are multilinear and Pi,j are sparse Rank: sparsity of Ti Theorem [Saraf-Volkovich]: If C 0 is a basic identity then #monom(C) ≤ spoly(k) White-Box Algorithm: find partition to sub-circuits of low dimension (after removal of g.c.d.) and brute force verify that they vanish. Black-Box Algorithm: If circuit not sparse can fix many variables and get sparse but not too sparse circuit. Use interpolation for circuits. June 18, 2011 31
41. PIT and Factoring f is composed if f(X) = g(X|S)h(X|T) where S and T are disjoint [S-Volkovich]: PIT is equivalent to factoring to decomposable factors. : f 0 ifff+yz has two decomposable factors. : Claim: If we have a PIT for all circuits of the form C1 + C2C3, whereCi Mthen given C Mwe can deterministically output all decomposable factors of C. June 18, 2011 33
42. PIT and Factoring Deterministic decomposable factoring is equivalent to lower bounds: Deterministic factoring implies NEXP does not have small arithmetic circuits Lower bounds imply Deterministic decomposable factoring PIT factoring formultilinear polynomials Deterministic decomposable factoring for depth-2, (k), sum of read-once… Open problem: is PIT equivalent to general factorization? June 18, 2011 34
50. Some `accessible’ open problems Give a Black-Box PIT algorithm for non-commutative formulas Solve PIT for depth-3 circuits Solve PIT for multilinear depth-3 circuits Black-Box PIT for set-multilinear depth-3 circuits (degree d tensors) Polynomial time BB-PIT for read-k formulas PIT for (non-multilinear) depth-4 with restricted fan-in Is PIT equivalent to general factorization? June 18, 2011 36