250 likes | 264 Views
Explore the relationship between communication or query complexity and partitioning methods in computational protocols, showcasing how optimal partitions can optimize communication. The study analyzes the minimization of communication while computing functions, highlighting key findings and breakthroughs in the field.
E N D
Nearly optimal separations between Communication (or query) complexity and partitions Andris Ambainis (Latvia), Martins Kokainis (Latvia), Robin Kothari (MIT)
communication complexity • Alice holds xX, Bob holds yY. • Dcc(f) - minimum amount of communication to compute F(x, y).
communication complexity • k bit protocol 2k possible transcripts. • Transcript = rectangle XiYi, Xi, Yi– sets of inputs for Alice and Bob. k bit protocol partition of XY into 2k rectangles, F constant on each rectangle
Not all partitions into 2k rectangles correspond to k bit communication protocols. Partition communication protocol?
Communication vs. Partition number • (F) – min number of rectangles XiYi in a partition with F constant on every rectangle. • Dcc(F) log (F) [Yao, 1979]. • Dcc(F) = O(log2(F)) [Aho, Ullman, Yannakakis, 1983]. F: Dcc(F) 2 log (F) [Kushilevitz, Linial, Ostrovsky, 1999] F: Dcc(F) = (log1.5(F)) [Goos, Pitassi, Watson, 2015]
x2 1 0 1 x3 1 0 0 1 QUERY COMPLEXITY x1 • Task: compute F(x1, ..., xN), xi{0, 1}. • Operation: queries to variables xi. 1 0 1
x2 1 0 1 x3 1 0 0 1 QUERY COMPLEXITY Decision tree with k queries x1 1 0 1 Partition of {0, 1}n into subcubes Si of dim n-k with F constant on each Si
Queries vs. partitions • D(F) – deterministic query complexity; • UC(F) – smallest k with a partition of {0, 1}n into subcubes defined by fixing k bits. • k D(F) k2; F: D(F) = (UC1.5(F)) [Goos, Pitassi, Watson, 2015]
Our separations • Deterministic query: D(F)=(UC2-o(1)(F)); • Deterministic communication: Dcc(F) = (log2-o(1)(F)); • Query result + lifting [GPW15]. • Randomized query: R(F)=(UC2-o(1)(F)); • Quantum query: Q(F)=(UC1.5-o(1)(F));
certificates • Certificate = partial assignment C such that: • x1 ... xnsatisfies C F(x1... xn)=a. • OR(x1, x2, ... xN): • 1-certificates: 1***, *1**, ..., ***1. • 0-certificates: 0000.
Subcubes vs. certificates • Certificate subcube consisting of all x1 ... xn that satisfy C. • Partition of {0, 1}ninto subcubes collection of certificates with each x1 ... xnsatisfying exactly one C. Unambiguous certificates.
Notation • UC(F) – smallest k such that there is an unambiguous collection of C covering {0, 1}n. • UCa(F) – smallest k such that there is an unambiguous collection of C covering x:F(x)=a. • C(F) – smallest k such that there is a collection of C covering {0, 1}n.
Cheat-sheets • [Aaronson, Ben-David, Kothari, STOC’16]. • Show R(F) = (Q2.5-o(1)(F)). • Number of other separations.
Cheat-sheets (Part 1) 0 1 0 ... 0 1 0 F(x(1)) F(x(2)) 0 0 0 ... 0 1 1 k a1a2... ak ... 1 1 1 ... 0 0 0 F(x(k))
Cheat-sheets (part 2) C1, C2, ..., Ck 0...00 0 1 0 ... 0 1 0 C1, C2, ..., Ck 0...01 0 0 0 ... 0 1 1 2k ... Ci – certificate for F(x(i))=ai 1...11 1 1 1 ... 0 0 0
Cheat-sheets • Fcs=1 if: • F(x(1))=a1, ..., F(x(k))=ak; • Cheat-sheet a1...akcontains descriptions of correct certificates for F(x(1))=a1, ..., F(x(k))=ak.
Certificates for fcs=1 0 1 0 ... 0 1 0 0 0 0 ... 0 1 1 C1, C2, ..., Ck UC1 kC ... cheat sheet 1 1 1 ... 0 0 0 Cheat-sheets decrease UC1!
Other complexity measures • UC1 (FCS) k C(F); • UC0 (FCS) k UC(F); • C(FCS) k C(F); • D(FCS) k D(F); • Typically, k log N, can be omitted. decreases does not decrease
STAGE 1 Cheat-sheet ORn ANDn ... ... ANDn ANDn
STAGE k, K>1 Cheat-sheet ORn ... ANDn ... ANDn ANDn F F F F F F F F F
D C0 C1 UC0 UC1 AND n 1 n n n OR n2 n n n2 n2 cheat-sheet n2 n n n2n
D C0 C1 UC0 UC1 AND n2i+1 ni ni+1 ni+1 ni+1 cheat-sheet n2i ni ni ni+1ni OR n2i+2 ni+1 ni+1 ni+2 ni+2 cheat-sheet n2i+2 ni+1 ni+1 ni+2ni+1
Result • D n2i, UC ni+1. • TheoremD(F) = (UC2-o(1)(F)). R(F) = (UC2-o(1)(F)) follows similarly.
Quantum results • TheoremQ(F)=(UC1.5-o(1)(F)). • TheoremQ(F)=(UC12-o(1)(F)). • Cheat-sheet+OR+BKK (instead of AND).
Conclusion • Almost quadratic gap between partitions and deterministic communication/query complexity. • Simpler construction? • Gap for randomized communication vs. partitions? • Quantum?