Linear computation coding (LCC) has been developed in as a new framework
for the computation of linear functions. LCC significantly reduces the complexity
of matrix-vector multiplication. In basic LCC, storage is not restricted i.e. the
wiring exponents are arbitrary integer exponents of 2.
In this work, we constrain the set of wiring exponents to be finite. From an
information-theoretic viewpoint, this problem is function compression with a finite-alphabet
representation by atoms. We show that by an efficient choice of this set,
the impact of finite alphabet size is negligible. Our numerical experiments
reveal that these algorithms can closely track the LCC performance in
the ideal case and demonstrate the trade-off between accuracy and storage
requirements empirically. The gains offered by LCC are closely achievable in practice with
finite-precision arithmetic operations.