N. Misra on Polynomial Kernelization

Neeldhara Misra, PhD student from IMSc (Chennai, India), recently visited our department at Chalmers. I had the opportunity to attend her well-given talk about the occasional infeasibility of polynomial kernelization. I liked the ideas she presented so I want to share them; you can read her own summary and download her complete work on her website.

Idea of Kernelization

Kernelization is about formalising preprocessing for hard problems and figuring out what can and cannot be achieved. In particular, NP-complete problems are studied; for these problems no fast (i.e. with polynomial bound on runtime) algorithms have been found yet. The idea is to crunch down a given instance of size to a size that is manageable by — more or less — naive algorithms while preserving equivalence, that is the reduced instance should have a solution if and only if the original instance has one. The notion of manageability is captured by a problem and instance dependent parameter ; we say that we have a polynomial kernel if there is an equivalent problem whose size is bounded by a polynomial in (i.e. independent of ). If is small — which is often the case in practical scenarios — the kernel can be solved reasonably quickly. Of course, the reduction process should then also be fast. Remains to mention that has to be chosen wisely; only knowing at least an upper bound for enables useful kernelizations.


Consider the well-known VERTEX COVER problem in its decision version, that is:

Given a graph and , check wether there exists a set such that while .

This problem is NP-complete. Now we do some preprocessing, that is thinking. What about a vertex with more than edges? Well, it needs to be included in our set because if it was not, all adjacent vertices had to be chosen in order to cover all those edges — but there are more than , so that cannot yield a feasible solution. So we remember as taken and remove it from the graph. We now do the same thing with as target size; you can see the recursion.

This process either terminates when we have chosen more than nodes or we have arrived at a reduced graph with some parameter for which we can not eliminate any vertices. In the former case, we can immediately output — after polynomial time in . In the other case, we are left with an instance that has only vertices with less than edges. Note that at most of such vertices can only cover at most edges. Therefore, by counting edges, we can immediately output or we know that we have only vertices left, dropping isolated edges that are obviously no help at all. The (connected) graph that remains is our kernel with polynomial size in .

Note that can really be small compared to the number of vertices. Intuitively, if has some relation to the sparsity of the given graph, effects can be tremendous, since many graphs out there are very sparse. Think of Facebook with hundreds of millions of vertices but very small cliques of maybe dozens.

No Free Lunch

In the above example, choosing the classical decision problem parameter as kernelization parameter was both obvious and rewarding. This need not always be the case. Neeldhara assured me that there are problems that have polynomial kernels but only for far more obscure choices of .

Also, kernelization does not always work. Even if we consider CONNECTED VERTEX COVER, only a simple variation of VC that demands the resulting graph must be connected, no polynomial kernel exists in general. In particular, the above algorithm fails since we can no longer simply drop nodes with degree greater but must take care of connectivity.


Looking at the above example, we immediately recognize a strong relation of polynomial kernelizability and fixed parameter tractability (FPT); performing any efficient preprocessing, we get a total runtime bounded by with a polynomial and any function. This is exactly the kind of runtime that makes a problem treatable, that is runtime is bounded polynomially, for a fixed .

The reverse inclusion does not hold under , though, that is there are FPT problems that have no polynomial kernel; see for example Bodlaender et al. Consequently, a whole theory has been built in order to investigate the precise capabilities of kernelization. There is apparently much to do. Good hunting!

Comments are closed.