1. Convergence Speed of Binary Interval Consensus Moez Draief Imperial College London Milan Vojnović Microsoft Research IEEE Infocom 2010, San Diego, CA, March 2010

2. Binary Consensus Problem 0 1 0 1 0 0 1 1 0 1 0 Each node wants answer to: was 0 or 1 initial majority? 0 2 Requirements:local interactions limited communication limited memory per node

3. Related Work Hypothesis testing with finite memory (ex. Hellman & Cover 1970’s...) – But typically not for dependent observations in network settings Ternary protocol ( Perron, Vasudevan & V. 2009) – Diminishing probability of error for some graphs – Ex. complete graphs – exponentially diminishing probability of error with the network size n; logarithmic convergence time in n Interval consensus (Bénézit, Thiran & Vetterli, 2009) – Convergence with probability 1 for arbitrary connected graphs – Limited results on convergence time 3

4. Our Problem Q: What is the expected convergence time for binary interval consensus over arbitrary connected graphs? 4

5. Binary Interval Consensus Four states 01e0e0 e1e1 e0e0 0 e0e0 0 e1e1 0 e0e0 0 01 e0e0 e1e1 e0e0 e1e1 e0e0 e1e1 e0e0 e1e1 1 1e1e1 1 e1e1 1 5 Update rules – Swaps – Annihilation

6. Outlook Upper bound on expected convergence time for arbitrary connected graphs Application to particular graphs – Complete – Star-shaped – Erdös-Rényi Conclusion 6

7. General Bound on Expected Convergence Time Let for every nonempty set of nodes S, : 7 Each edge (i, j) activated at instances a Poisson process (q i,j )

8. General Bound on Expected Convergence Time (cont’d) 8 Without loss of generality we assume that initial majority are state 0 nodes  n = initial fraction of nodes in state 0, other nodes in state 1,  > 1/2

9. General Bound on Expected Convergence Time (cont’d) Key observation: two phases – In phase 1 nodes in state 1 are depleted – In phase 2 nodes in state e 1 are depleted Phase 1 1 if node i in state 1 1 if node i in state 0 9

10. Phase 1 Dynamics: S k = set of nodes in state 0 The result follows by using a “spectral bound” on the expected number of nodes in state 1 10

11. Outlook Upper bound on expected convergence time for arbitrary connected graphs Application to particular graphs – Complete – Star-shaped – Erdös-Rényi Conclusion 11

12. Complete graph Each edge activated with rate 1/(n-1) Inversely proportional to the voting margin Can be made arbitrary large! 12

13. Complete graph (cont’d) The general bound is tight 0 and 1 state nodes annihilate after a random time that has exponential distribution with parameter cut(S 0 (t), S 1 (t)) / (n-1) 13

14. Star-shaped graph Each edge activated with rate 1/(n-1) 14

15. Star-shaped graph (cont’) By first step analysis: Same scaling, different constant 15

16. Erdös-Rényi graph Each edge age e activated with rate X e /np n where X e ~ Ber(p n ) 16

17. Erdös-Rényi graph (cont’d) For sufficiently large expected degree, the bound is approximately as for the complete graph – In conformance with intuition 17

18. Conclusion Established a bound on the expected convergence time of binary interval consensus for arbitrary connected graphs The bound is inversely proportional to the smallest absolute eigenvalue of some matrices derived from the contact rate matrix The bound is tight – Achieved for complete graphs – Exact scaling order for star-shaped and Erdös-Rényi graphs Future work – Expected convergence time for m-ary interval consensus? – Lower bounds on the expected convergence time? 18