1. Chris Beck, Russell Impagliazzo, Shachar Lovett

2. History of Sampling Problems Earliest uses of randomness in algorithms were for sampling, not decision problems – “Equation of State Calculations by Fast Computing Machines” [Metropolis, Rosenbluth, Rosenbluth, Teller, Teller ‘53] – Complexity Theory of Randomized Sampling [Jerrum Valiant Vazirani ‘86] – Markov Chain Monte Carlo method [Jerrum Sinclair ‘89]

3. History of Sampling Problems

4. The task of explicitly exhibiting a distribution which is hard to sample for concrete models was raised more recently: – [Ambainis, Schulman, Ta-shma, Vazirani, Wigderson ‘03] “The Quantum Communication Complexity of Sampling” – [Goldreich, Goldwasser, Nussboim ‘10] “On the implementation of huge random objects” – [Viola’10] “On the Complexity of Distributions”

5. Prior Work

6. Applications to Data Structures

7. Second Main Result

8. Sketch of Proof

11. Detour: Regularity in Ckt. Complexity Given a circuit, get a small decision tree. Each circuit at a leaf is the restriction of original circuit, according to path to that leaf. All or “most” leaves are “balanced” / “regular” / “pseudorandom”.

12. Actual Sketch of Proof

13. Balanced Decision Forests

16. Contrast with Other Results

17. Handy Corollary of Main Result

18. Conclusion: “Tail-Bound LMN”

19. Open Questions

20. Thanks!

22. Second Main Result In proving the strengthened sampling lower bound, we were led to discover a new concentration bound for sums of random variables computed by decision trees. This concentration bound extends a long line of work aimed at obtaining Chernoff-like concentration despite limited independence. We believe it will have other applications.

23. Decision Forests

26. Balanced Decision Trees

27. Applications of [LV’11] Lovett and Viola also gave several interesting applications for their work: – Lower Bounds for Succinct Data Structures

29. Prior Work