Augmented Index and Quantum Streaming Algorithms for DYCK(2) ASHWIN NAYAK, AND DAVE TOUCHETTE University of Waterloo, Perimeter Institute CCC 2017, Riga, Latvia 8 July 2017
Communication Complexity β¦ Communication Complexity setting: π π΅ π πΆ π, π π· 1 = π 1 (π, π π΅ ) π· 2 = π 2 (π, π πΆ , π· 1 ) β¦ π· π = π π (π, π πΆ , π· <π ) Output: f(x,y) β¦ How much communication to compute f on x, y β¦ Take information-theoretic view: Information Complexity β¦ How much information to compute f on x, y βΌ π β¦ Information content of interactive protocols? β¦ Classical vs. Quantum?
Communication Complexity β¦ Communication Complexity setting: π π΅ π πΆ π, π π· 1 = π 1 (π, π΅, π π΅ ) π· 2 = π 2 (π, π πΆ , π· 1 ) π π π΅ π π π πΆ π β¦ π· π = π π (π, π πΆ , π· <π ) Output: f(x,y) β¦ How much communication to compute f on x, y β¦ Take information-theoretic view: Information Complexity β¦ How much information to compute f on x, y βΌ π β¦ Information content of interactive protocols? β¦ Classical vs. Quantum?
Communication Complexity β¦ Communication Complexity setting: π π΅ π πΆ π, π π 1 π· 1 π 2 π· 2 π π π΅ π π π πΆ π β¦ π π π· π Output: f(x,y) β¦ How much communication to compute f on x, y β¦ Take information-theoretic view: Information Complexity β¦ How much information to compute f on x, y βΌ π β¦ Information content of interactive protocols? β¦ Classical vs. Quantum?
Communication Complexity β¦ Communication Complexity setting: |πβͺ π, π π 1 π· 1 π 2 π· 2 π π π΅ π π π πΆ π β¦ π π π· π Output: f(x,y) β¦ How much communication to compute f on x, y β¦ Take information-theoretic view: Information Complexity β¦ How much information to compute f on x, y βΌ π β¦ Information content of interactive protocols? β¦ Classical vs. Quantum?
Quantum Communication Complexity π ππ΅ 1 ππ΅ 2 ππ΅ 3 ππ΅ π π Protocol Ξ : π΅ π π 1 π 3 π π π΅ 0 Output: f(X,Y) π· 3 π· 1 π· πβ1 π | πβͺ π· π π· 2 πΆ 0 πΆ π π 2 π π ππΆ πβ1 ππΆ 2 π π
Th.1: Streaming Algorithms for DYCK(2) π¦ π π¦ 1 π¦ 2 |0 π‘(π) βͺ β¦ π π¦ π π π¦ 1 π π¦ 2 Pre Post Repeat T times β¦ Streaming algorithms: Attractive model for early Quantum Computers β¦ Some exponential advantages possible for specially crafted problems [LeG06, GKKRdW07] β¦ DYCK 2 = π + πΈππ·πΏ 2 + πΈππ·πΏ 2 + πΈππ·πΏ 2 β πΈππ·πΏ 2 π β¦ Classical bound: π‘ π β Ξ© [MMN10, JN10, CCKM10] π β¦ Two-way classical algorithm: π‘ π β O(ππππ§πππ(π))
Th.1: Streaming Algorithms for DYCK(2) π β¦ Th. 1: Any T-pass 1-way qu. streaming algo. for DYCK(2) needs space π‘ π β Ξ©( π 3 ) on length N inputs β¦ Even holds for non-unitary streaming operations π β¦ Reduction from multi-party QCC to streaming algorithm to DYCK(2) [MMN10] β¦ Multi-party problem consists of OR of multiple instances of two-party problem β¦ Space s(N) in algorithm corresponds to communication between parties β¦ Consider T-pass, one-way quantum streaming algorithms β¦ Direct sum argument allows to reduce from a two-party problem, Augmented Index β¦ Multi-party QCC lower bounds requires two- party QIC lower bound on βeasy distributionβ β¦ Subtlety for non-unitary streaming operations π
Th.2: Augmented Index β¦ Index π¦ 1 β¦ π¦ π β¦ π¦ π , π = π¦ π β¦ Augmented Index: π΅π½ π π¦ 1 β¦ π¦ π , π, π¦ 1 β¦ π¦ <π , π = π¦ π β π π¦ π§ π¦ 1 π¦ 1 π¦ 2 π¦ 2 β¦ β¦ π¦ πβ1 π¦ πβ1 π¦ π π β [π] β¦ π¦ πβ1 b β {0, 1} π¦ π β π π¦ π
Th.2: Augmented Index β¦ Th. 2.2: For any r-round protocol Ξ for π΅π½ π , either π β¦ π π½π· π΅βπΆ Ξ , π 0 β Ξ© or π 2 1 β¦ π π½π· πΆβπ΅ Ξ , π 0 β Ξ© with π 2 β¦ π 0 the uniform distribution on zeros of π΅π½ π (βeasy distributionβ) β¦ Classical bounds: π β¦ π½π· π΅βπΆ Ξ , π 0 β Ξ© 2 π or β¦ π½π· πΆβπ΅ Ξ , π 0 β Ξ© π β¦ [MMN10, JN10, CCKM10, CK11] β¦ We Build on direct sum approach of [JN10] β¦ General approach uses two main Tools (Sup.-Average Encoding Th., Qu. Cut-and-Paste) β¦ More specialized approach uses one more Tool (Information Flow Lemma)
Warm-up: Disjointness πΈππ‘π π (π¦, π§) = Β¬Ϊ πβ[π] (π¦ π β§ π§ π ) π·π· πΈππ‘π π β Ξ©(π) π¦ π§ π¦ 1 π§ 1 π¦ 2 π§ 2 β¦ β¦ π¦ πβ1 π§ πβ1 π¦ π AND π§ π β¦ β¦ π¦ πβ1 π§ πβ1 π¦ π π§ π
Warm-up: Disjointness CC πΈππ‘π π β₯ π½π· 0 πΈππ‘π π = π π½π· 0 π΅ππΈ [BJKS02] π½π· β€ π·π·, π½π· satisfies direct sum property (needs private and public randomness) 2 2 π½π· 0 (π΅ππΈ) = 3 π½ π: π π = 0 + 3 π½(π: π|π = 0) Comparing message transcript π on 01, 00, 10 inputs: ?
Tool: Average Encoding Theorem β¦ Average encoding theorem [KNTZ01]: E π β 2 π π , π β β€ π½(π: π) β¦ π β = E π [π π ] , average message β¦ h 2 π 1 π 2 , Heilinger distance β¦ Follows from Pinskerβs inequality β¦ Low information messages are close to average message β¦ For AND, π = 0 : 1 2 β 2 π 00 , π β0 + 1 2 β 2 π 10 , π β0 β€ π½(π: π|π = 0) β¦ Using Jensen and triangle inequality: 1 4 β 2 π 00 , π 10 β€ π½ π: π π = 0 β¦ Similarly, for X=0: 1 4 β 2 π 00 , π 01 β€ π½ π: π π = 0 1 8 β 2 π 10 , π 01 β€ π½ π: π π = 0 + π½ π: π π = 0 = 3 β¦ Comparing 01, 10 inputs: 2 π½π· 0 (π΅ππΈ)
Warm-up: Disjointness CC πΈππ‘π π β₯ π½π· 0 πΈππ‘π π = π π½π· 0 π΅ππΈ [BJKS02] π½π· β€ π·π·, π½π· satisfies direct sum property, needs private and public randomness 2 2 π½π· 0 (π΅ππΈ) = 3 π½ π: π π = 0 + 3 π½(π: π|π = 0) 12 β 2 π 10 , π 01 β€ π½π· 0 (π΅ππΈ) 1 Comparing message transcript π on 01, 00, 10 inputs:
Warm-up: Disjointness CC πΈππ‘π π β₯ π½π· 0 πΈππ‘π π = π π½π· 0 π΅ππΈ [BJKS02] π½π· β€ π·π·, π½π· satisfies direct sum property, needs private and public randomness 2 2 π½π· 0 (π΅ππΈ) = 3 π½ π: π π = 0 + 3 π½(π: π|π = 0) 12 β 2 π 10 , π 01 β€ π½π· 0 (π΅ππΈ) 1 Comparing message transcript π on 01, 00, 10 inputs: Comparing π on 00, 11 inputs: ?
Tool: Cut-and-Paste Lemma β¦ Consider input subset {π¦ 1 , π¦ 2 } Γ {π§ 1 , π§ 2 } π¦ 1 π§ 1 π¦ 2 π§ 2 β¦ Triangle inequality implies for π on π¦ 1 , π§ 2 and (π¦ 2 , π§ 1 ) : β¦ β π π¦ 1 π§ 2 , π π¦ 2 π§ 1 β€ β π π¦ 1 π§ 1 , π π¦ 1 π§ 2 + β(π π¦ 1 π§ 1 , π π¦ 2 π§ 1 ) π¦ 1 π§ 1 π¦ 1 π§ 1 π¦ 1 π§ 1 + β π¦ 2 π§ 2 π¦ 2 π§ 2 π¦ 2 π§ 2 β¦ What about π on π¦ 1 , π§ 1 and (π¦ 2 , π§ 2 ) : ? β¦ Cut-and-paste Lemma [BJKS02]: β π π¦ 1 π§ 1 , π π¦ 2 π§ 2 = β π π¦ 1 π§ 2 , π π¦ 2 π§ 1 π¦ 1 π§ 1 π¦ 1 π§ 1 β π¦ 2 π§ 2 π¦ 2 π§ 2
Warm-up: Disjointness CC πΈππ‘π π β₯ π½π· 0 πΈππ‘π π = π π½π· 0 π΅ππΈ [BJKS02] π½π· β€ π·π·, π½π· satisfies direct sum property, needs private and public randomness 2 2 π½π· 0 (π΅ππΈ) = 3 π½ π: π π = 0 + 3 π½(π: π|π = 0) 12 β 2 π 10 , π 01 β€ π½π· 0 (π΅ππΈ) 1 Comparing message transcript π on 01, 00, 10 inputs: 12 β 2 π 00 , π 11 = 1 12 β 2 π 10 , π 01 β€ π½π· 0 (π΅ππΈ) 1 Comparing π on 00, 11 inputs:
Warm-up: Disjointness CC πΈππ‘π π β₯ π½π· 0 πΈππ‘π π = π π½π· 0 π΅ππΈ [BJKS02] π½π· β€ π·π·, π½π· satisfies direct sum property, needs private and public randomness 2 2 π½π· 0 (π΅ππΈ) = 3 π½ π: π π = 0 + 3 π½(π: π|π = 0) 12 β 2 π 10 , π 01 β€ π½π· 0 (π΅ππΈ) 1 Comparing message transcript π on 01, 00, 10 inputs: 12 β 2 π 00 , π 11 = 1 12 β 2 π 10 , π 01 β€ π½π· 0 (π΅ππΈ) 1 Comparing π on 00, 11 inputs: 1 Statistical interpretation: β π 1 , π 2 β₯ 4 | π 1 β π 2 | ππ ||π 00 β π 11 || ππ β Ξ© 1 since for AND, π 00 β 0, π 11 β 1 π·π· πΈππ‘π π β Ξ©(π) Quantum?
Warm-up: Disjointness π π·π· πΈππ‘π π β Ξ( π) π π·π· π πΈππ‘π π β O( π π ) , r round protocols 1 π πΈππ‘π π = 1 π π΅ππΈ [JRS03] QCC r πΈππ‘π π β₯ π ΰ·ͺ π π ΰ·ͺ π π½π· 0 π π½π· 0 ΰ·ͺ ΰ·ͺ π π½π· β€ π π π·π·, π π½π· satisfies direct sum property, requires private and public βrandomnessβ Comparing π on 01, 00, 10 inputs: ?
Recommend
More recommend