The threshold logic filter is an important nonlinear filter class, whi
ch is defined by a threshold logic function of binary input values. It
is proved in this paper that not all positive self-dual logical funct
ions are threshold functions if the number of input variables is 5 or
more. The positive self-dual logical filter is a limited class but inc
ludes almost all filters for noise removal. Our result means that not
all positive self-dual filters can be expressed by one operation of th
e threshold logic filters. In the sense of the filter expression by ne
ural networks, the two-layer network cannot always optimize even this
limited class.