Shengyu Zhang CSE Dept. @ CUHK Roadmap • Intro to theoretical computer science • Intro to quantum computing • Export of quantum computing – Formula Evaluation • Solves a classical open question – N-Representability problem • Addresses the failure of many efforts in quantum chemistry • Quantum is natural mathematically – Decision tree complexity – Communication complexity A brief intro to theoretical computer science • Computation: a sequence of elementary instructions. • More than knowing the existence, but a step-by-step way to find it. Efficiency • Efficient Computation: – Algorithm: design fast algorithms – Computational complexity: classify problems according to their computational difficulty • Structural – Measured by resources like time, space, randomness, counting,… • Interactive • Concrete models: Decision Tree, Communication Complexity, Circuit Connections to other sciences • Import: Use of concepts and techniques from – Math: discrete math, analysis, algebra, topology – Physics • Export: – Solve TCS questions appearing naturally in • Statistical Physics, Chemistry, Molecular Biology, Social Science, Economics, Computer & Information Science, – Concepts such as completeness; – Problems such as P vs. NP • One of the seven $1M Millennium Problems*1 *1: http://www.claymath.org/millennium/P_vs_NP/ Roadmap • Intro to theoretical computer science • Intro to quantum computing • Export of quantum computing – Formula Evaluation • Solves a classical open question – N-Representability problem • Addresses the failure of many efforts in quantum chemistry • Quantum is natural mathematically – Decision tree complexity – Communication complexity Areas in quantum computing • • • • • • Quantum algorithms Quantum complexity Quantum cryptography Quantum error correction Quantum information theory Others: Quantum control / game theory / … Area 1: Quantum Algorithms 1994 1996 1998 2000 2002 2004 2006 2008 Shor: Factoring & Discrete Log QFT (Quantum Fourier Transform): exponential speedup; slower than expected. • Factoring: Given an n-bit number, factor it (into product of two numbers). – The reverse problem of Multiplication, which is very easy. • Classical (best known) : ~ O(2n^1/3) • Quantum*1: ~ O(n2) *1: P. Shor. STOC’93, SIAM Journal on Computing, 1997. Area 1: Quantum Algorithms 1994 1996 1998 2000 2002 2004 2006 2008 Shor: Factoring & Discrete Log QFT (Quantum Fourier Transform): exponential speedup; slower than expected. • Implication of fast algorithm for Factoring – Theoretical: Church-Turing thesis – Practical: Breaking RSA-based cryptosystems Area 1: Quantum Algorithms 1994 1996 1998 Shor: Factoring & Discrete Log 2000 2002 2004 2006 2008 Hallgren: Pell’s Equation QFT (Quantum Fourier Transform): exponential speedup; slower than expected. • Pell’s Equation: x2 – dy2 = 1. • Problem: Given d, find solutions (x,y) to the above equation. • Classical (best known): – ~ 2√log d (assuming the generalized Riemann hypothesis) – ~ d1/4 (no assumptions) • Quantum*1: poly(log d). *1: S. Hallgren. STOC’02. Journal of the ACM, 2007. Area 1: Quantum Algorithms 1994 1996 1998 Shor: Factoring & Discrete Log 2000 2002 Hallgren: Pell’s Equation 2004 2006 2008 Kuperberg: HSP-Dihedral QFT (Quantum Fourier Transform): exponential speedup; slower than expected. • Hidden Subgroup Problem (HSP): Given a function f on a group G, which has a hidden subgroup H, s.t. f is – constant on each coset aH, – distinct on different cosets. Task: find the hidden H. • Factoring, Pell’s Equation both reduce to it. • Efficient quantum algorithms are known for Abelian groups. • Main question: HSP for non-Abelian groups? Area 1: Quantum Algorithms 1994 1996 1998 Shor: Factoring & Discrete Log 2000 2002 Hallgren: Pell’s Equation 2004 2006 2008 Kuperberg: HSP-Dihedral QFT (Quantum Fourier Transform): exponential speedup; slower than expected. • Two biggest cases: – HSP for symmetric group Sn: Graph Isomorphism reduce to it. – HSP for dihedral group Dn: Shortest Lattice Vector reduces to it. • HSP(Dn): – Classical (best known): 2log|G| – Quantum*1: 2O(√log|G|) *1: G. Kuperberg. arXiv:quant-ph/0302112, 2003. Area 1: Quantum Algorithms 1994 1996 1998 Shor: Factoring & Discrete Log 2000 2002 Hallgren: Pell’s Equation 2004 2006 2008 Kuperberg: HSP-Dihedral QFT (Quantum Fourier Transform): exponential speedup; slower than expected. Grover: Search QS (Quantum Search): polynomial speedup; most solved. • Given n bits x1,…,xn, find an i with xi = 1. – Given n bits x1,…,xn, decide whether ∃i s.t. xi = 1. • Classical: Θ(n) • Quantum*1: Θ(√n) *1: L. Grover. Physical Review Letters, 1997. Area 1: Quantum Algorithms 1994 1996 1998 Shor: Factoring & Discrete Log 2000 2002 Hallgren: Pell’s Equation 2004 2006 2008 Kuperberg: HSP-Dihedral QFT (Quantum Fourier Transform): exponential speedup; slower than expected. Grover: Many combinatorial Search /graph problems QS (Quantum Search): polynomial speedup; most solved. AAKV*1: Def QW (Quantum Walk): poly and exp speedup; rapidly developed. *1: D. Aharonov, A. Ambainis, J. Kempe, U. Vazirani. STOC'01 Area 1: Quantum Algorithms 1994 1996 1998 2000 2002 2004 2006 2008 • Classical random walk on graphs: starting from some vertex, repeatedly go to a random neighbor – Many algorithmic applications • Quantum walk on graphs: even definition is non-trivial. – For instance: classical random walk converges to a stationary distribution, but quantum walk doesn’t since unitary is reversible. AAKV*1: Def QW (Quantum Walk): poly and exp speedup; rapidly developed. *1: D. Aharonov, A. Ambainis, J. Kempe, U. Vazirani. STOC'01 Area 1: Quantum Algorithms 1994 1996 1998 2000 2002 2004 2006 2008 • Element Distinctness: Given n integers, decide whether they are the all distinct. • Classical: Θ(n) • Quantum: Θ(n2/3) – Apply quantum walk on (n,n2/3)-Johnson graph. AAKV: Def Ambainis*1: Ele. Dist. QW (Quantum Walk): poly and exp speedup; rapidly developed. *1: A. Ambainis, FOCS’04 Area 1: Quantum Algorithms 1994 1996 1998 2000 2002 ∧ ∨ ∨ ¬ ∧ Grover’s search: OR function general formula by {AND-OR-NOT} AAKV: Def 2004 2006 2008 • Classical: Θ(n) • Quantum: ~ Θ(√n) • apply QW on the formula graph with weight carefully designed for inductions to work. Ambainis: Ele. Dist. ACRSZ*1: Formula Evaluation QW (Quantum Walk): poly and exp speedup; rapidly developed. *1: A. Ambainis, A. Childs, B. Reichardt, R. Spalek, S. Zhang. FOCS’07 Area 2: Quantum Complexity • Quantum complexity – Structural: • A sample here: BQP in PSPACE – Interactive: • A sample here: QIP = QIP[3] – Concrete models: DT and CC • More to come in next section BQP in PSPACE • P: problems solvable in polynomial time – One characterization of efficient computation • BPP: problems solvable in probabilistic polynomial time w/ a small error tolerated – Another characterization of efficient computation • BQP: problems solvable in polynomial time by a quantum computer w/ a small error tolerated – Yet another characterization of efficient computation, if you believe large-scale quantum mechanics. Classical upper bound of BQP • Central in complexity theory: comparisons of different modes of computations • How to compare classical and quantum efficient computation? • An obvious lower bound: BPP ⊆ BQP • An upper bound (of quantum by classical) • [Thm*1] BQP ⊆ PSPACE – PSPACE: problems solvable in polynomial space. *1: Bernstein, Vazirani. STOC’93, SIAM J. on Computing, 1997 Where does BQP sit in? EXP PSPACE PH NPC NP P,BPP • PH: Polynomial Hierarchy • Level 3: – Polynomial time verification V s.t. f(x) = 1 if ∃y1∀y2∃y3 V[x,y1,y2,y3] = 1. • NP is just level 1. Open question: BQP ⊆ PH? Interactive Proof • Interactive Proof: Verifier solves a hard problem with the help of a powerful but untrustworthy Prover. • If YES: P to convince V. P … V P, Pr[P convinces V] > 1-δ (δ: completeness error) • If NO: ∄ P to convince V. Computationally unbounded P, Pr[P convinces V] < ε Probabilistic polynomial time (ε: soundness error) Quantum Interactive Proof • IP: problems solvable by interactive proof system – IP[k]: problems solvable by k-round interactive proof system • QIP: problems solvable by quantum interactive proof system – QIP[k]: problems solvable by k-round quantum interactive proof system • [Thm*1] QIP = QIP[3] • Classically: IP=IP[3] ⇒ PH collapses to AM *1: Kitaev, Watrous. STOC’00. Roadmap • Intro to Theoretical Computer Science • Intro to Quantum Computing • Export of quantum computing – Formula Evaluation • Solves a classical open question – N-Representability problem • Addresses the failure of many efforts in quantum chemistry • Quantum is natural mathematically – Decision tree complexity – Communication complexity Classical implications of quantum algorithms Note that we solved a purely • A classical fact on polynomial threshold degree classical open problem and learnability of a class of functions*1: – thr(f) ≤ by r forgiving all fC ⇒ C can be learned in time nO(r) • Question*2: Any formulaalgorithm. f of size n has a quantum polynomial threshold function thr(f) = O(n1/2)? • Recall that we have O(n1/2)-time quantum algorithm for any AND-OR-NOT formula • Now (roughly): thr(f) ≤ Q(f) ≤ n1/2 • This implies that formulas are learnable in time 2√n. (Matching the known lower bound.) *1: A. Klivans, R. Servedio, STOC’01; A. Klivans, R. Servedio, R. O’Donnell, FOCS’02 *2: O’Donnell, Servedio, STOC’03 Classical implication of quantum arguments • It’s not uncommon. • Quantum computer is not only a potentially more powerful computation machine. • It’s also a different mathematical model. • So studies of quantum computing turn out to provide novel perspectives of old (classical) problems • And some led to complete solutions. Roadmap • Intro to Theoretical Computer Science • Intro to Quantum Computing • Export of quantum computing – Formula Evaluation • Solves a classical open question – N-Representability problem • Addresses the failure of many efforts in quantum chemistry • Quantum is natural mathematically – Decision tree complexity – Communication complexity N-Representability problem • N-Representability problem in quantum • This explains the failure of efforts so chemistry: characterize the allowed set of far. density operators on N-body fermions satisfying 2-body correlations. •given And tells researchers to stop trying to • solve An efficient solution would be a breakthrough. the generic problem. • It had attracted a very large of effort, though not quite successful yet. • [Thm*1] N-Representability is QMA-complete. – QMA: the quantum analog of NP. – Thus QMA-complete is even harder than NP-hard. *1: Liu, Christandl, Verstraete. Physical Review Letters, 2007 Roadmap • Intro to Theoretical Computer Science • Intro to Quantum Computing • Export of quantum computing – Formula Evaluation • Solves a classical open question – N-Representability problem • Addresses the failure of many efforts in quantum chemistry • Quantum is natural mathematically – Decision tree complexity – Communication complexity decision tree computation f(x1,x2,x3) = x1∧(x2∨x3) • Task: compute f(x) • The input x can be x1 = ? accessed by queries 0 1 in the form of “xi = ?”. f(x1,x2,x3)=0 x2 = ? • We only care about the number of 0 1 queries made x3 = ? f(x1,x2,x3)=1 • Query (decision tree) 0 1 complexity: min # queries needed. f(x1,x2,x3)=0 f(x1,x2,x3)=1 Decision tree complexity • DTD(f) = the minimum number of queries needed to compute f (on all inputs x) – Superscript D: “deterministic” • Next we’ll define a natural measure of f and show that it’s a lower bound of DTD(f). degree • ∀ f:{0,1}n→{0,1} can be represented by a multi-variate polynomial of deg ≤ n. – f(001) = f(010) = f(111) = 1, and 0 on other x. – f(x1x2x3) = (x1x2x3=001) OR (x1x2x3=010) OR (x1x2x3=111) = (1-x1)(1-x2)x3 + x1(1-x2)x3 + x1x2x3 – is a deg-3 polynomial. • [Fact] This polynomial representation is unique. Decision tree and degree • [Fact] deg(f) ≤ DT(f) • Collect all 1-leaves. • f = OR of all paths to these 1-leaves. f(x1,x2,x3) = x1∧(x2∨x3) x1 = ? 0 1 f(x1,x2,x3)=0 f(x1,x2,x3) = (x1=1,x2=1) OR (x1=1,x2=0,x3=1) = x1x2 + x1(1-x2)x3 x2 = ? 0 1 x3 = ? 0 f(x1,x2,x3)=1 1 f(x1,x2,x3)=0 f(x1,x2,x3)=1 Randomized decision tree The error prob 0.01 here can be changed to any ε with an extra cost • We can toss coins during about the computation. log(1/ε). • Or equivalently, we have a random string r and a collection of decision tree Tr, s.t. for each input x Er[Tr(x)] ≥ 0.99 if f(x) = 1 Er[Tr(x)] ≤ 0.01 if f(x) = 0 • Thus a randomized d.t. is a collection S of many deterministic d.t. s.t. for any x, most of the d.t. in S give the correct answer f(x). • Randomized DT complexity: the max depth of d.t. in S. --- DTR(f) Quantum query algorithm • Instead of coin-tossing, we ask all variables in superposition. • |i, a, z → |i, axi, z – i: the position we are interested in – a: the register holding the queried variable – z: other part of the work space • i,a,zαi,a,z |i, a, z → i,a,zαi,a,z |i, axi, z • By def: DTQ(f) ≤ DTR(f) ≤ DTD(f) • We’ve shown deg(f) ≤ DTD(f) • Next: We have a similar lower bound for DTR(f). Approximate degree • degε(f) = min {deg(f’): |f(x) – f’(x)| ≤ ε}. • [Fact] degε(f) ≤ DTR(f) • [proof] d.t.’s in DTR gives a polynomial in degε – DTR is a collection of d.t. Tr, each of depth d = DTR(f). – Represent each Tr by a degree≤d polynomial pr. • By the fact in the deterministic case shown just now. – Now let f’ = Er[pr]; it has degree≤d – f’(x) = Er[pr(x)] = Er[Tr on x]: ε-approximating f(x). • By the def of DTR(f) Approximate degree of OR • degε(f) = min {deg(f’): |f(x) – f’(x)| ≤ ε}. • [Fact] degε(f) ≤ DTR(f) • Question: What’s degε(f) for very simple functions, such as AND or OR? – Note that deg(AND) = deg(OR) = n. • AND(x1,…,xn) = x1…xn, • OR(x1,…,xn) = 1-(1-x1)…(1-xn) • Using the above bound? • It gives nothing! Because you are still living in the classical world! Mathematically. – DTR(AND) = DTR(OR) = Ω(n). Welcome to quantum world • So we know DTR(f) ≥ degε(f) • [Theorem*1] DTQ(f) ≥ degε(f)/2 • By this together with Grover’s Search DTQ(OR) = O(√n), we get: degε(OR) = O(√n)! *1: Beals, Buhrman, Cleve, Mosca, de Wolf, STOC’98, J. of the ACM, 2001 Roadmap • Intro to Theoretical Computer Science • Intro to Quantum Computing • Export of quantum computing – Formula Evaluation • Solves a classical open question – N-Representability problem • Addresses the failure of many efforts in quantum chemistry • Quantum is natural mathematically – Decision tree complexity – Communication complexity Communication complexity*1 • Two parties, Alice and Bob, jointly compute a function F(x,y) with x known only to Alice and y only to Bob. x y Alice Bob F(x,y) F(x,y) • Communication complexity: how many bits are needed to be exchanged? --- CCD(F) *1. A. Yao. STOC’79. Why CC is interesting? • Reason 1: Mathematically interesting and challenging. • Reason 2: Rich connections to other areas in TCS • Though defined in an information theoretical setting, it turned out to provide lower bounds to many computational models. – Data structures, circuit complexity, streaming algorithms, decision tree complexity, VLSI, algorithmic game theory, optimization, pseudo-randomness… Rank lower bound • Two-variable function f(x,y) ↔ matrix Af = [f(x,y)] – Two-variable Boolean function ↔ Boolean matrix • Rank lower bound*1: CCD(f) ≥ log2 rank(Mf), where Mf = [f(x,y)]x,y • [proof] – Decompose A into monochromatic combinatorial rectangles. • CCD(f) ≥ log2 # monochromatic combinatorial rectangles – Each rectangle has rank 1. – Rank is subadditive. *1. K. Melhorn and E. Schmidt. STOC’82. Log Rank Conjecture • Big open problem: • Log Rank Conjecture*1: ∀ total Boolean f, CCD(f) = poly(log2 rank(Mf)) – Largest known gap*2: CCD(f) = (log2 rank(Mf))1.63… *1. L. Lovász and M. Saks. FOCS’88 *2. N. Nisan and A. Wigderson. Combinatorica, 1995. Variant of rank • Next: we’ll introduce a natural variant of rank and show that it’s a lower bound of CCR(f) • One cute question2as a bait: 3 IN 1 0 ¢¢¢ 0 6 0 1 ¢¢¢ 07 6 7 = 6. . . .. 7 4 .. .. 5 0 0 ¢¢¢ 1 the N-dim identity matrix has rank N. • Question: If you can perturb each entry by 0.01, how much can you decrease the rank? Approximate rank • Approximate rank: For M = [mij] rankε(M) = min{rank(M’): |Mij – M’ij| ≤ ε}. • [Thm*1] CCR(A) ≥ log2 rankε(A) • Back to our question of rankε(IN): It’s2nothing but 3 1 0 ¢¢¢ 0 the Equality problem where 6 7 – f(x,y) = 1 iff x=y. • [Fact*2] CCR(Eq) = O(1). • So, quite counterintuitively, rankε(IN) = O(1) *1. M. Krause. Theoretical Computer Science, 1996. *2. M. Rabin, A. Yao. Unpublished. IN 6 0 1 ¢¢¢ = 6. . 4 .. .. 0 0 ¢¢¢ 07 .. 7 .5 1 Not always work • Another matrix M of dimension 2n2n M[x,y] = 1 iff ∃i s.t. xi = yi = 1. – An important matrix in TCS. • CCR(M) ≥ log2 rankε(M) doesn’t work – [Thm*1] CCR(A) = Ω(n). – So this only gives rankε(A) = 2O(n). • [Thm*2] CCQ(A) ≥ log2 rankε(A) / 2 • Thus rankε(A) = 2O(√n). *1. Kalyanasundaram and Schintger, SIAM Journal on Discrete Mathematics, 1992. Razborov. Theoretical Computer Science, 1992. *2. H. Buhrman and R. de Wolf. CCC’01. Natural mathematically Algebraic Parameter (degree, rank) Allow perturbation ≤ Complexity Measure (DTD, CCD) • The quantum complexity is closer to the natural math lower bound. • The tightening gives nontrivial results randomized complexity can’t yield. Approximate Parameter ≤ (degε/rankε) Quantum ≤ Complexity (DTQ, CCQ) ≤ Allow error Randomized Complexity (DTR, CCR) A brief intro to quantum computing • Feymann’82: Idea • Deutsch’85,’89: quantum Turing machine and quantum circuit • Bernstein-Vazirani’93, Yao’93: ground of quantum complexity theory • Shor’94: fast quantum algorithm for Factoring and Discrete Log