Skip to main content
To KTH's start page To KTH's start page

On Long Proofs of Simple Truths

Time: Thu 2022-10-27 14.00

Location: F3, Lindstedtsvägen 26 & 28, Stockholm

Video link:

Language: English

Subject area: Computer Science

Doctoral student: Kilian Risse , Teoretisk datalogi, TCS

Opponent: Associate Professor Benjamin Rossman, Duke University, Levine Science Research Center, Durham, NC, USA

Supervisor: Per Austrin, Teoretisk datalogi, TCS; Johan Håstad, Matematik (Avd.); Jakob Nordström, Teoretisk datalogi, TCS

Export to calendar

QC 20221005


Propositional proof complexity is the study of certificates of infeasibility. In this thesis we consider several proof systems with limited deductive ability and unconditionally show that they require long refutations of the feasibility of certain Boolean formulas. We show that the depth $d$ Frege proof system, restricted to linesize $M$, requires proofs of length at least $\exp\bigl(n/(\log M)^{O(d)}\bigr)$ to refute the Tseitin contradiction defined over the $n \times n$ grid graph, improving upon the recent result of Pitassi et al. [PRT21]. Along the way we also sharpen the lower bound of Håstad [Hås20] on the depth $d$ Frege refutation size for the same formula from exponential in $\tilde{\Omega}(n^{1/59d})$ to exponential in$\tilde{\Omega}(n^{1/(2d-1)})$. We also consider the perfect matching formula defined over a sparse random graph on an odd number of vertices $n$. We show that polynomial calculus over fields of characteristic $\neq 2$ and sum of squares require size exponential in $\Omega(n/\log^2 n)$ to refute said formula. For depth $d$ Frege we show that there is a constant $\delta > 0$ such that refutations of these formulas require size $\exp\bigl(\Omega(n^{\delta/d})\bigl)$. The perfect matching formula has a close sibling over bipartite graphs: the graph pigeonhole principle. There are two methods to prove resolution refutation size lower bounds for the pigeonhole principle. On the one hand there is the general width-size tradeoff by Ben-Sasson and Wigderson [BW01] which can be used to show resolution refutation size lower bounds in the setting where we have a sparse bipartite graph with $n$ holes and $m \ll n^2$pigeons. On the other hand there is the pseudo-width technique developed by Razborov [Raz04] that applies for any number of pigeons, but requires the graph to be somewhat dense. We extend the latter technique to also cover the previous setting and more: for example, it has been open whether the functional pigeonhole principle defined over a random bipartite graph of bounded degree and $\poly(n) \ge n^2$ pigeons requires super-polynomial size resolution refutations. We answer this and related questions. Finally we also study the circuit tautology which claims that a Boolean function has a circuit of size $s$ computing it. For $s = \poly(n)$ we prove an essentially optimal Sum of Squares degree lower bound of $\Omega(s^{1-\eps})$ to refute this claim for any Boolean function. Further, we show that for any $0 < \alpha < 1$ there are Boolean functions on $n$ bits with circuit complexity larger than$2^{n^\alpha}$ but the Sum of Squares proof system requires size $2^{\bigl(2^{\Omega(n^\alpha)}\bigr)}$ to prove this. Lastly we show that these lower bounds can also be extended to the monotone setting.