White-box, path-based, testing is largely used for the validation of programs. Given the control-flow graph (CFG) of the program under test, a test suit is generated by selecting a collection of paths of interest, then trying to provide, for each path, some concrete input values that will make the program follow that path during a run.For the first step, there are various ways to define paths of interest: structural testing methods select some set of paths that fulfills coverage criteria related to elements of the graph; in random-based techniques, paths are selected according to a given distribution of probability over these elements (for instance, uniform probability over all paths of length less than a given bound). Both approaches can be combined as in structural statistical testing. The random-based methods above have the advantage of providing a way to assess the quality of a test set as the minimal probability of covering an element of a criterion.The second step requires to produce for each path its path predicate, i.e. the conjunction of the constraints over the input parameters that must hold for the system to run along that path. This is done using symbolic execution. Then, constraint-solving is used to compute test data. If there is no input values such that the path predicate evaluates to true, the path is infeasible. It is very common for a program to have infeasible paths and such paths can largely outnumber feasible paths. Infeasible paths selected during the first step will not contribute to the final test suite, and there is no better choice than to select another path, hoping for its feasibility. Handling infeasible paths is the serious limitation of structural methods since most of the time is spent selecting useless paths. It is also a major challenge for all techniques in static analysis of programs, since the quality of the approximations they provide is lowered by data computed for paths that do not correspond to actual program runs.To overcome this problem, different methods have been proposed, like concolic testing or random testing based on the input domain. In path-biased random testing, paths are drawn according to a given distribution and their feasibility is checked in a second step. We present an algorithm that builds better approximations of the behavior of a program than its CFG, providing a transformed CFG, which still over-approximates the set of feasible paths but with fewer infeasible paths. This transformed graph is used for drawing paths at random.We modeled our graph transformations and formally proved, using the interactive theorem proving environment Isabelle/HOL, the key properties that establish the correctness of our approach.Our algorithm uses symbolic execution and constraint solving, which allows to detect whether some paths are infeasible. Since programs can contain loops, their graphs can contain cycles. In order to avoid to follow infinitely a cyclic path, we enrich symbolic execution with the detection of subsumptions. A subsumption can be interpreted as the fact that some node met during the analysis is a particular case of another node met previously: there is no need to explore the successors of the subsumed node: they are subsumed by the successors of the subsumer. Our algorithm has been implemented by a prototype, whose design closely follows said formalization, giving a good level of confidence in its correctness.In this thesis, we introduce the theoretical concepts on which our approach relies, its formalization in Isabelle/HOL, the algorithms our prototype implements and the various experiments done and results obtained using it.