Complexity Theory of Polynomial-Time Problems

Transcription

1 Complexity Theory of Polynomial-Time Problems Lecture 11: Nondeterministic SETH Karl Bringmann

2 Complexity Inside P SAT 2 n 3SUM n 2 OV n 2 Colinearity n 2 3SUM-hard APSP n 3 APSP equivalent Radius n 3 LCS n 2 SETH-hard BMM n " Negative Triangle n 3 EDIT n 2 Fréchet n 2 SlidingWindowHD n " diameter n 2 BMM-hard can we relate SAT / 3SUM / APSP / BMM?

3 Relating Hypotheses only very weak relations are known e.g. ETH implies that k-sum has no n #(%) algorithm OPEN: SETH implies that 3SUM has no O(n ()* ) algorithm? today we will see a barrier for tighter connections

4 I. Nondeterministic SETH

5 Nondeterministic Algorithms Turing machine: can choose any applicable transition at any point in time RAM: operation guess() fills a cell with an integer YES-instance: at least one accepting path NO-instance: all paths reject guesses on an accepting path = proof that we have a YES-instance NP: all problems solvable in polytime by a nondet. Turing machine k-sat algorithm: guess a satisfying assignment, check correctness O(n + m) guess a short proof of satisfiability and check it 3SUM algorithm: guess a, b, c A and check a + b + c = 0 O(1)/O(n)

6 Co-Nondeterministic Algorithms Turing machine: can choose any applicable transition at any point in time RAM: operation guess() fills a cell with an integer YES-instance: all paths accept NO-instance: at least one path rejects guesses on a rejecting path = proof that we have a NO-instance co-np: all problems solvable in polytime by a co-nondet. Turing machine k-sat: guess a short proof of unsatisfiability and check it Is this possible? classic computational complexity: if NP co-np, then k-sat has no O(poly(n)) co-nondet. algorithm we believe that NP co-np, since otherwise the polynomial hierarchy collapses

7 (Co-)Nondeterministic SETH if NP co-np, then k-sat has no O(poly(n)) co-nondet. algorithm not even a ) co-nondet. algorithm is known! Nondeterministic SETH: [CGIMPS 16] k-sat has no no ) co-nondet. algorithm do not allow randomization! NSETH implies SETH (without randomization) barely anyone believes that NSETH is true but it formalizes a current barrier NSETH can be used to conditionally rule out reductions

8 Potential Reduction from SAT to 3SUM an deterministic algorithm A for k-sat with oracle access to 3SUM s.t.: Properties: k-sat fomula φ n variables, m n % clauses reduction total time r(n) 3SUM instance I 1 size n? instance I k size n % for any fomula φ, algorithm A(φ) correctly solves k-sat on φ A runs in time r(n) = ) for some γ > 0 for any ε > 0 there is a δ (0, γ) s.t. % ()* KL? n K e.g. k = 1 and n? = m O, then n? ()* n O%

9 Potential Reduction from SAT to 3SUM an deterministic algorithm A for k-sat with oracle access to 3SUM s.t.: Properties: k-sat fomula φ n variables, m n % clauses reduction total time r(n) 3SUM instance I 1 size n? instance I k size n % for any fomula φ, algorithm A(φ) correctly solves k-sat on φ A runs in time r(n) = ) for some γ > 0 for any ε > 0 there is a δ (0, γ) s.t. % ()* KL? n K ) algorithm O(n ()* ) algorithm

10 Potential Reduction from SAT to 3SUM an deterministic algorithm A for k-sat with oracle access to 3SUM s.t.: Properties: k-sat fomula φ n variables, m n % clauses reduction total time r(n) 3SUM instance I 1 size n? instance I k size n % for any fomula φ, algorithm A(φ) correctly solves k-sat on φ A runs in time r(n) = ) for some γ > 0 for any ε > 0 there is a δ (0, γ) s.t. % ()* KL? n K nondet. ) algorithm and co-nondet. ) algorithm nondet. O(n ()* ) algorithm and co-nondet. O(n ()* ) algorithm

11 Potential Reduction from SAT to 3SUM an deterministic algorithm A for k-sat with oracle access to 3SUM s.t.: k-sat fomula φ n variables, m n % clauses reduction total time r(n) 3SUM instance I 1 size n? instance I k size n % for each instance I R : guess whether it is YES- or NO-instance if we guessed YES: guess a proof π R that I R is a YES-instance if we guessed NO: guess a proof π R that I R is a NO-instance if we guessed correctly: π = (π?,, π % ) forms a proof that φ is satisfiable or unsatisfiable algorithm A is the proof checker

12 Potential Reduction from SAT to 3SUM an deterministic algorithm A for k-sat with oracle access to 3SUM s.t.: k-sat fomula φ n variables, m n % clauses reduction total time r(n) 3SUM instance I 1 size n? instance I k size n % nondet. ) algorithm and co-nondet. ) algorithm nondet. O(n ()* ) algorithm and co-nondet. O(n ()* ) algorithm no nondet. ) algorithm or no co-nondet. ) algorithm =NSETH no nondet. O(n ()* ) algorithm or no co-nondet. O(n ()* ) algorithm

13 Ruling Out Reductions If NSETH holds and 3SUM has a O(n ()* ) co-nondeterministic algorithm we will show this then there is no deterministic reduction from k-sat to 3SUM either 3SUM has strongly subquadratic algorithms or 3SUM is hard for a different reason than k-sat or NSETH fails has drawbacks, but this is the only tool for negative results in this area

14 Co-Nondeterministic Algorithm for 3SUM 3SUM: given set A of integers in { n Y,, n Y }, are there a, b, c A s.t. a + b + c = 0? Thm: 3SUM has a co-nondeterministic algorithm in time OV(n P/( ) [CGIMPS 16] OV hides polylogarithmic factors in n OV f n = O(f n polylog n) OV f n = \ O f n log O n O^_

15 Co-Nondeterministic Algorithm for 3SUM 3SUM: given set A of integers in { n Y,, n Y }, are there a, b, c A s.t. a + b + c = 0? Thm: 3SUM has a co-nondeterministic algorithm in time OV(n P/( ) 1) guess prime p n P/( logn [CGIMPS 16] OV(p) 2) compute t = a, b, c A P a + b + c = 0 mod p OV(p)

16 Recall: 3SUM for Small Numbers 3SUM is in time O n + OV(U) for numbers in {0,, U} define polynomial P X has degree at most U k l X k compute Q X P X P X P X = ( k l X k )( k l X k )( X k k l ) what is the coefficient of X n in Q(X)? it is the number of (a, b, c) summing to t (X k X p X O = X kqpqo ) use efficient polynomial multiplication (via Fast Fourier Transform): polynomials of degree d can be multiplied in time OV(d)

17 Co-Nondeterministic Algorithm for 3SUM 3SUM: given set A of integers in { n Y,, n Y }, are there a, b, c A s.t. a + b + c = 0? Thm: 3SUM has a co-nondeterministic algorithm in time OV(n P/( ) 1) guess prime p n P/( logn [CGIMPS 16] 2) compute t = a, b, c A P a + b + c = 0 mod p OV(p) let B a mod p a A} (in general B is a multi-set!) let r _ a, b, c B P a + b + c = 0 let r? a, b, c B P a + b + c = p universe size U = p let r ( a, b, c B P a + b + c = 2p then t = r _ + r? + r (

18 Co-Nondeterministic Algorithm for 3SUM 3SUM: given set A of integers in { n Y,, n Y }, are there a, b, c A s.t. a + b + c = 0? Thm: 3SUM has a co-nondeterministic algorithm in time OV(n P/( ) 1) guess prime p n P/( logn [CGIMPS 16] 2) compute t = a, b, c A P a + b + c = 0 mod p 3) if t > α n P/( logn: accept (constant α to be fixed later) OV(p) 4) guess distinct a?, b?, c?,, a n, b n, c n A P such that a K + b K + c K = 0 mod p i 5) check that for all a K, b K, c K we have a K + b K + c K 0 6) if everything works out: reject (otherwise accept) time OV(n P/( ) if we reject then we have a NO-instance YES-instance: all paths accept NO-instance: at least one path rejects

19 Co-Nondeterministic Algorithm for 3SUM 3SUM: given set A of integers in { n Y,, n Y }, are there a, b, c A s.t. a + b + c = 0? NO-instance: at least one path rejects: show that there exists a prime p n P/( logn such that t = a, b, c A P a + b + c = 0 mod p < α n P/( logn M := # tuples (a, b, c, p) with a, b, c A and prime p s.t. a + b + c = 0 mod p each a + b + c is in 3n Y,, 3n Y {0}, so it has at most log(3n Y ) prime factors thus M n P log 3n Y 3s n P log n by prime number theorem: there are at least n P/( /β primes p n P/( log n thus there is a prime p contained in at most M/(n P/( /β) tuples (a, b, c, p) thus there is a prime p with t M/(n P/( /β) 3s β n P/( log(n) set α 3s β

20 Co-Nondeterministic Algorithm for 3SUM 3SUM: given set A of integers in { n Y,, n Y }, are there a, b, c A s.t. a + b + c = 0? Thm: 3SUM has a co-nondeterministic algorithm in time OV(n P/( ) 1) guess prime p n P/( logn [CGIMPS 16] 2) compute t = a, b, c A P a + b + c = 0 mod p 3) if t > α n P/( logn: accept (constant α to be fixed later) OV(p) 4) guess distinct a?, b?, c?,, a n, b n, c n A P such that a K + b K + c K = 0 mod p i 5) check that for all a K, b K, c K we have a K + b K + c K 0 6) if everything works out: reject (otherwise accept) time OV(n P/( ) if we reject then we have a NO-instance YES-instance: all paths accept NO-instance: at least one path rejects

21 I. Randomized Nondeterministic SETH

22 Randomized Nondeterministic SETH Nondeterministic SETH: k-sat has no no ) co-nondet. algorithm what if we allow randomization? then the hypothesis is wrong! Thm: k-sat has a randomized co-nondeterministic poly(n)) algorithm with error probability 2 ) (@) [Williams 16] Thm: OV has a randomized co-nondeterministic O(n poly(d, log n)) algorithm with error probability n ) (?)

23 Tools: Basics on Polynomials fix field Z and assume that field operations can be performed in O(1) time univariate polynomials P X KL_ a K X K, Q X = b K X K, m n ƒ KL_ multiplication P X Q(X): OV n (by FFT, without proof) division with remainder: OV n (without proof) P X = S X Q X + R(X), where R(X) has degree < m we write R X = P X mod Q X evaluate P(X) at a given point x: O(n) Horner s method: P x = a _ + x a? + x a ( + x

24 Tools: Multipoint Evaluation on Polynomials fix field Z and assume that field operations can be performed in O(1) time univariate polynomial P X KL_ a K X K multipoint evaluation: OV n evaluate P(X) at given points X = x?,, 1) let L X (X x? ) (X ) and R X (X ) (X ) 2) let P X P X mod L(X) and P Š X P X mod R(X) 3) recursively compute P (x? ),, P ) and P Š ),, P Š ) polynomial division: T n = 2T n/2 + OV n = OV n P X = S X L X + P (X) P x K = S x K L x K + P x K = P x K = 0 for i n/2

25 Tools: Multipoint Evaluation on Polynomials fix field Z and assume that field operations can be performed in O(1) time univariate polynomial P X KL_ a K X K computing L X (X x? ) (X ): logn 0 straight-forward binary tree (X x K ) computes canonical polynomials defined by P Y ( Ž q?, Yq? ( Ž P K,R X : = X x % R %LK in layer i: n/2 K multiplications of polynomials of degree 2 K total time OV n X

26 Tools: Polynomial Interpolation fix field Z and assume that field operations can be performed in O(1) time univariate polynomial P X = polynomial interpolation: OV KL_ a K X K given pairs x?, y?,, ) find a polynomial P(X) with P x K = y K for all i Lagrange s formula: P X = y K K R K X x R x K x R Caveat: division by x in Z means multiplication with the inverse x )? extended Euclidean algorithm: computes s, t with s x + t p = gcd x, p = 1 modulo p: s x = 1 so s = x )? is the inverse of x

27 Tools: Polynomial Interpolation 1st goal: compute P X = y K X x R K R K (for y K := y K? R K ) ) let L X (X x? ) (X ) and R X (X ) (X ) recursion: y K X x R = y K X x R R X K R K? R K + y K X x R L K R K

28 Tools: Polynomial Interpolation 2nd goal: compute factors s K = R K 1 x K x R compute the derivative Q X X x K = a K X K KL_ Q X = K R K X x = i a K X K)? KL_ can compute Q X in time OV n then we have Q x % = x % x R K R K = x % x R R % = s % )? compute all Q x % for k = 1,, n in time OV n by multipoint evaluation

29 Tools: Polynomial Interpolation fix field Z and assume that field operations can be performed in O(1) time univariate polynomial P X = polynomial interpolation: OV KL_ a K X K given pairs x?, y?,, ) find a polynomial P(X) with P x K = y K for all i Lagrange s formula: P X = y K can be computed in time OV n! K R K X x R x K x R

30 Tools: Arithmetic Circuits fix field Z and assume that field operations can be performed in O(1) time arithmetic circuits are a (succinct) representation of (multivariate) polynomials input gates labeled with variables X?,, X % - univariate if k = 1 each other gate is a + or (unbounded fanin) or a (fanin 1) or a constant (fanin 0) fanout is unbounded one output gate (no cyclic dependencies) given an input x?,, x % Z the output C(x?,, x % ) is the number computed by the output gate (in Z ) 42 + = 42 = 40 = 2 = 2 X? X ( = 640 = 1 = 1 + = 8 = 1 = 7 X P X? X ( X? X ( X ( X P + X P 7 = 7

31 Tools: Arithmetic Circuits fix field Z and assume that field operations can be performed in O(1) time representation as circuit is not unique X 1 + 2X 2X 3 X = circuit = unstructured, succinct X 4X P + 10X ( + 6X polynomial = structured, verbose

32 Tools: Arithmetic Circuits fix field Z and assume that field operations can be performed in O(1) time working modulo p is necessary over Z numbers can get very large: 2 = 2 (

33 Tools: Arithmetic Circuits fix field Z and assume that field operations can be performed in O(1) time k inputs size s = number if wires degree d = largest degree of any monomial assuming no cancelations degree(input gate) = 1 6 degree(constant gate) = 0 degree( gate) = degree of child degree( + gate) = maximum of degrees of children degree( gate) = sum of degrees of children degree of circuit = degree(output gate) X? X ( X P X? X ( X? X ( X ( X P + X P 7

34 Tools: Evaluation of Circuits fix field Z and assume that field operations can be performed in O(1) time k inputs size s = number if wires degree d evaluating a circuit at given input: O(s) X 1 + 2X 2X 3 X

35 Tools: Identity Testing fix field Z and assume that field operations can be performed in O(1) time k inputs size s = number if wires degree d given two circuits C?,C (, do they represent the same polynomial over Z? this problem is called polynomial identity testing assume that C?,C ( are univariate and assume p 2d Schwarz-Zippel-Lemma: this has a randomized OV(s)-time algorithm with error probability 1 s ) (?) 1) for logs rounds: 2) pick random x Z 3) if C? (x) C ( (x): return not identical 4) return identical C? X C ( (X) is a polynomial Q(X) of degree d if C? (X) C ( (X) then Q(X) is not the 0-polynomial and thus has at most d roots with probability 1/2 we pick a non-root x

36 Tools: Evaluation of Circuits fix field Z and assume that field operations can be performed in O(1) time k inputs size s = number if wires degree d evaluating a circuit at given input: O(s) multipoint evaluation (at n points): no near-linear time algorithm known trivial algorithm: O(n s) converting a univariate circuit 2 X to a polynomial: OV(s d) (write each gate as a degree polynomial) 1 + 2X 2X 3 X conversion + multipoint evaluation for polynomials = multipoint evaluation for univariate circuits in OV(s d + n + d) (for multivariate: degree d polynomial has up to d + k + 1 k monomials )

37 OV as Multipoint Evaluation on Circuits OV: Given sets A, B 0,1 of size n Decide whether there are a A, b B such that a b circuit C(a, b) for testing orthogonality of a, b: C(a, b) = (1 a K b K ) + KL? 1 1 a? b? a b

38 OV as Multipoint Evaluation on Circuits OV: Given sets A, B 0,1 of size n Decide whether there are a A, b B such that a b circuit C(a) for testing orthogonality of a with any b B: C(a) = C(a, b) = (1 a K b K ) + p p KL? - d inputs (for the coordinates of a) - size O(nd) - degree 2d C(a, b) for each b B - for any a {0,1} : 0 C(a) n - pick prime p n and work modulo p, i.e., over field Z

39 (Co-)Nondet. Multipoint Evaluation on Circuits Given circuit C on inputs X?,, X % with size s and degree d over Z % Given inputs z?,, Z want to evaluate C on each z R = (z R 1,, z R k ) can assume p 2nd, p n (?) 1) compute polynomials R? X,, R % (X) such that R K j = z R [i] OV(kn) by polynomial interpolation new goal: evaluate univariate circuit C X = C(R? X,, R % (X)) on X = 1,, n C has size O(s + kn) and depth dn 2) guess a polynomial Q X of degree at most dn 3) check that Q X = C(R? X,, R % (X)) by polynomial identity testing O(dn) OV(s + kn + dn) 4) multipoint evaluate Q X on X = 1,, n and return these values OV(dn)

40 Co-Nondet. Algorithm for OV use circuit C(a) for testing orthogonality of a with any b B {0,1} evaluate C(a) at each a A d inputs, size O(dn), degree 2d YES-instance: all paths accept NO-instance: at least one path rejects...with high probability 1) compute polynomials R? X,, R % (X) such that R K j = z R [i] OV(kn) by polynomial interpolation new goal: evaluate univariate circuit C X = C(R? X,, R % (X)) on X = 1,, n C has size O(s + kn) and depth dn 2) guess a polynomial Q X of degree at most dn 3) check that Q X = C(R? X,, R % (X)) by polynomial identity testing, if not: ACCEPT O(dn) OV(s + kn + dn) 4) multipoint evaluate Q X on X = 1,, n and return these values OV(dn) 5) ACCEPT RL? Q(j) 1 OV(dn)

41 Co-Nondet. Algorithm for OV NO-instance: If we correctly guess Q X then the identity test works with prob. 1 and we correctly report non-existence of an orthogonal pair = REJECT YES-instance: If we correctly guess Q X then we ACCEPT If we wrongly guess Q X then identity test fails with prob. 1 n ) (?) for any guess: we ACCEPT with probability 1 n ) (?) 1) compute polynomials R? X,, R % (X) such that R K j = z R [i] OV(kn) by polynomial interpolation new goal: evaluate univariate circuit C X = C(R? X,, R % (X)) on X = 1,, n C has size O(s + kn) and depth dn 2) guess a polynomial Q X of degree at most dn 3) check that Q X = C(R? X,, R % (X)) by polynomial identity testing, if not: ACCEPT O(dn) OV(s + kn + dn) 4) multipoint evaluate Q X on X = 1,, n and return these values OV(dn) 5) ACCEPT RL? Q(j) 1 OV(dn)

42 Conclusion Nondeterministic SETH: k-sat has no no ) co-nondet. algorithm No randomization allowed If it holds, then there is no deterministic reduction from SETH to 3SUM, since 3SUM has a OV(n P/( ) co-nondeterministic algorithm This is the only tool for ruling out reductions! Randomized Nondeterministic SETH : is wrong! We have seen a OV(dn) co-nondeterministic algorithm for OV uses many tools for computing with polynomials and arithmetic circuits