Theory Seminar<br><br>Title: Computing Requires Larger Formulas than Approximating<br>Speaker: Avishay Tal<br>Date: November 2, 2017<br>Time: 4:15pm<br>Location: Gates Building, 463A<br><br>Abstract: <br>A de-Morgan formula over Boolean variables x_1, ..., x_n is a binary tree whose internal nodes are marked with AND or OR gates and whose leaves are marked with variables or their negation. We define the size of the formula as the number of leaves in it. Proving that some explicit function (in P or NP) requires large formulas is a central open question in computational complexity.<p>In this work, we introduce a size-amplification hardness reduction for de-Morgan formulas. We show that average-case hardness implies worst-case hardness for a larger size. More precisely, if a function f cannot be computed correctly on more than 1/2 + eps of the inputs by any formula of size s, then computing f correctly on all inputs requires size ~s*log(1/eps). The tradeoff is essentially tight. Quite surprisingly, the proof relies on a result from quantum query complexity by Reichardt [SODA, 2011].</p><p>As an application, we improve the best known formula size lower bounds for explicit functions by logarithmic factors to ~n^3/log(n). In addition, we propose candidates for explicit functions that we believe have formula size ~n^4, and prove non trivial super-quadratic formula size lower bounds for them using our reduction. Learning discrete Markov Random Fields with nearly optimal runtime and sample complexity.</p>