Parisi and Frisch proposed some time ago an explanation for ''multisca
ling'' of turbulent velocity structure functions in terms of a ''multi
fractal hypothesis,'' i.e., they conjectured that the velocity field h
as local Holder exponents in a range [h(min), k(max)], with exponents
<h occurring on a set S(h) with a Fractal dimension D(h). Heuristic re
asoning led them to an expression for the scaling exponent z(p) of pth
order as the Legendre transform of the codimension d-D(h). We show he
re that a part of the multifractal hypothesis is correct under even we
aker assumptions: namely, if the velocity field has L(p)-mean Holder i
ndex s, i.e., if it les in the Besov space B-p(s,infinity), then local
Holder regularity is satisfied. If s<d/p, then the hypothesis is true
in a generalized sense of Holder space with negative exponents and we
discuss the proper definition of local Holder classes of negative ind
ex. Finally, if a certain ''box-counting dimension'' exists, then the
Legendre transform of its codimension gives the scaling exponent z(p),
and, more generally, the maximal Besov index of order p, as s(p)=z(p)
/p. Our method of proof is derived from a recent paper of S. Jaffard u
sing compactly-supported, orthonormal wavelet bases and gives an exten
sion of his results. We discuss implications of the theorems for ensem
ble-average scaling and fluid turbulence.