Floating-point computations are difficult to handle soundly due to pervasive rounding. This is particularly the case when inferring relations between variables, as relational analyses rely heavily on algebraic identities of real numbers that no longer hold in float. In this talk, we will discuss some solutions and address two problems: how to soundly analyze programs that feature floating-point computations, and how to implement a sound relational static analyzer that itself uses floating-point computations. We will discuss the application of the methods to the detection of run-time errors.