The multifractal formalism for functions relates some functional norms
of a signal to its ''Holder spectrum'' (which is the dimension of the
set of points where the signal has a given Holder regularity). This f
ormalism was initially introduced by Frisch and Parisi in order to num
erically determine the spectrum of fully turbulent fluids: it was late
r extended by Arneodo, Bacry, and Muzy using wavelet techniques and ha
s since been used by many physicists. Until now, it has only been supp
orted by heuristic arguments and verified for a few specific examples.
Our purpose is to investigate the mathematical validity of these form
ulas; in particular, we obtain the following results: The multifractal
formalism yields for any function an upper bound of its spectrum. We
introduce a ''case study,'' the self-similar functions; we prove that
these functions have a concave spectrum (increasing and then decreasin
g) and that the different formulas allow us to determine either the wh
ole increasing part of their spectrum or a part of it. One of these me
thods (the wavelet-maxima method) yields the complete spectrum of the
self-similar functions. We also discuss the implications of these resu
lts for fully developed turbulence.