Broadband multilevel fast multipole algorithm for locally refined meshes

Multilevel fast multipole algorithm (MLFMA) based on the standard Rokhlin translation function cannot handle effectively cases where the element size is very small compared to the wavelength. Broadband MLFMA allows the creation of additional levels with smaller cubes, and so also these cases can be computed. However, if the element size varies a lot across the model and the childless cubes are all on the same level, either the sparse system matrix or the cube structure can require large amounts of memory. This paper presents an algorithm that modifies broadband MFLMA so that the memory requirements stay reasonable in all cases.