Abstract
In the fields of Algorithmic and Complexity, a large area of research is based on the assumption that P ≠ NP (the classesPolynomial time and Non deterministic Polynomial time are different), which means that there are problems for which a solution can be verified but not constructed in polynomial time. Many natural problems are not in P, which means that they have no efficient algorithm. In order to tackle such problems, many different branches of Algorithmics have been developed. One of them is called Parametric Complexity. It consists of developing exact algorithms whose complexity is measured as a function of the size of the instance and of a parameter. Such a parameter allows a more precise analysis of the complexity. In this context, an algorithm will be considered to be efficient if it is fixed parameter tractable (fpt), that is, if it has a complexity which exponentially depends on the parameter and polynomially depends on the size of the instance. Problems that can be solved by such an algorithm form the FPT class.Kernelisation is one technique that produces fpt algorithms, among others. It can be viewed as a preprocessing of the instance, with a guarantee on the compression of the data. More formally, a kernelisation is a polynomial reduction from a problem to itself, with the additional constraint that the size of the kernel, the reduced instance, is bounded by a function of the parameter. In order to obtain an fpt algorithm, it is sufficient to solve the problem in the reduced instance, by brute-force for example (which has exponential complexity, in the parameter). Hence, the existence of a kernelisiation implies the existence of an fpt algorithm. It holds that the converse is true also. Nevertheless, the existence of an efficient fpt algorithm does not imply a small kernel, meaning a kernel with a linear or polynomial size. Under certain hypotheses, it can be proved that some problems can not have a kernel (that is, are not in FPT) and that some problems in FPT do not have a polynomial kernel.One of the main results in the field of Kernelisation is the construction of a linear kernel for the Dominating Set problem on planar graphs, by Alber, Fellows and Niedermeier.To begin with, the region decomposition method proposed by Alber, Fellows and Niedermeier has been reused many times to develop kernels for variants of Dominating Set on planar graphs. Nevertheless, this method had quite a few inaccuracies, which has invalidated the proofs. In the first part of our thesis, we present a more thorough version of this method and we illustrate it with two examples : Red Blue Dominating Set and Total Dominating Set.Next, the method has been generalised to larger classes of graphs (bounded genus, minor-free, topological-minor-free), and to larger families of problems. These meta-results prove the existence of a linear or polynomial kernel for all problems verifying some generic conditions, on a class of sparse graphs. As a price of generality, the proofs do not provide constructive algorithms and the bound on the size of the kernel is not explicit. In the second part of our thesis, we make a first step to constructive meta-results. We propose a framework to build linear kernels based on principles of dynamic programming and a meta-result of Bodlaender, Fomin, Lokshtanov, Penninkx, Saurabh and Thilikos.