ContactPerson: aduri@cse.buffalo.edu ### Begin Citation ### Do not delete this line ### %R 2001-10 %U /home/csgrad/aduri/comp/thesis/thesis.ps %A Pavan, A. %T Average-case complexity theory and polynomial-time reductions %D August 26, 2001 %I Department of Computer Science and Engineering, SUNY Buffalo %K average-case complexity, reasonable distributions, distributionally-hard languages, polynomial-time reducibilities, separating NP-completeness notions %X This thesis studies average-case complexity theory and polynomial-time reducibilities. The issues in average-case complexity arise primarily from Cai and Selman's extension of Levin's definition of average polynomial time. We study polynomial-time reductions between distributional problems. Under strong but reasonable hypotheses we separate ordinary $\NP$-completeness notions. The average-case complexity of a problem is, in many cases, a more significant measure than the worst-case complexity. It is known that some $\NP$-complete problems, such as $k$-coloring and Hamiltonian cycle, can be solved quickly on average. This has motivated the development of the study of average-case analysis of algorithms. Levin~\cite{Levin86} initiated a general study of average-case complexity by defining a robust notion of average polynomial time and the notion of distributional $\NP$-completeness. Cai and Selman ~\cite{CaiSelman99} gave a general definition of {\em $T$ on aver age} for arbitrary time bounds $T$. They observed difficulties with Levin's definition of average polynomial time when applied to unreasonable input distributions. Their definition of average polynomial time slightly modifies Levin's definition to avoid these difficulties. In this thesis we study reasonable distributions, distributionally-hard languages, and reductions among distributional problems. We prove several results demonstrating that it suffices to restrict one's attention to reasonable distributions. This is important because both Levin's definition and Cai and Selman's definition yield unsatisfactory consequences when we permit unreasonable distributions. We show that if $\NP$ has a $\DTIME(2^n)$-bi-immune language, then every $\DistNP$-complete problem must have a reasonable distribution. We prove that the class $\ppcomp$, the set of languages that can be decided in average polynomial time with respect to every polynomial-time computable distribution, remains unchanged when restricted to reasonable distributions. We strengthen and present a simpler proof of a result of Belanger and Wang~\cite{BelangerWang96}, which shows that Cai and Selm an's definition of average-polynomial time is not closed under many-one reductions. Cai and Selman showed that every is $\P$-bi-immune language is dis\-tri\-bu\-tionally-hard. We study the question of whether there exist distributionally-hard languages that are not $\P$-bi-immune. First we show that such languages exist if and only if $\P$ contains $\P$-printable-immune sets. Then we extend this characterization significantly to include assertions about several traditional questions about immunity, about finding witnesses for $\NP$-machines, and about the existence of one-way functions. Next we study polynomial-time reducibilities. Ladner, Lynch, and Selman~\cite{LadnerLynchSelman75} showed, in the context of worst-case complexity, that various poly\-no\-mi\-al-t ime reductions differ in $\E$. We show similar results for reductions between distributional problems and we show that most of the completeness notions for $\DistEXP$ are different. Finally, we turn our attention to the question of whether various notions of $\NP$-completeness are different. Lutz and Mayordomo, under the hypothesis that the $p$-measure of $\NP$ is not zero, and Ambos-Spies and Bentzien, under the hypothesis that $\NP$ has a $p$-generic language, showed that various completeness notions in $\NP$ are different. We introduce a structural hypothesis not involving stochastic properties and prove that the existence of a Turing complete language for $\NP$ that is not truth-table complete follows from our hypothesis. This result is the first to provide evidence that truth-table completeness for $\NP$ is weaker than Turing completeness for $\NP$.