Lossy Kernelization for (Implicit) Hitting Set Problems

We re-visit the complexity of kernelization for the $d$-Hitting Set problem. This is a classic problem in Parameterized Complexity, which encompasses several other of the most well-studied problems in this field, such as Vertex Cover, Feedback Vertex Set in Tournaments (FVST) and Cluster Vertex Deletion (CVD). In fact, $d$-Hitting Set encompasses any deletion problem to a hereditary property that can be characterized by a finite set of forbidden induced subgraphs. With respect to bit size, the kernelization complexity of $d$-Hitting Set is essentially settled: there exists a kernel with $O(k^d)$ bits ($O(k^d)$ sets and $O(k^{d-1})$ elements) and this it tight by the result of Dell and van Melkebeek [STOC 2010, JACM 2014]. Still, the question of whether there exists a kernel for $d$-Hitting Set with fewer elements has remained one of the most major open problems~in~Kernelization. In this paper, we first show that if we allow the kernelization to be lossy with a qualitatively better loss than the best possible approximation ratio of polynomial time approximation algorithms, then one can obtain kernels where the number of elements is linear for every fixed $d$. Further, based on this, we present our main result: we show that there exist approximate Turing kernelizations for $d$-Hitting Set that even beat the established bit-size lower bounds for exact kernelizations -- in fact, we use a constant number of oracle calls, each with ``near linear'' ($O(k^{1+\epsilon})$) bit size, that is, almost the best one could hope for. Lastly, for two special cases of implicit 3-Hitting set, namely, FVST and CVD, we obtain the ``best of both worlds'' type of results -- $(1+\epsilon)$-approximate kernelizations with a linear number of vertices. In terms of size, this substantially improves the exact kernels of Fomin et al. [SODA 2018, TALG 2019], with simpler arguments.