Mon Mar 30 13:16:47 CEST 2009

normal numbers / autodiff -> memoized expressions

Combining data structures reminds me of normal numbers and automatic
differentiation: instead of producing a giant expression, autodiff
produces a memoized version that has lower complexity than a fully
subsituted single expression.  Differentiation doesn't really change
complexity, but it is a "convolution" operation that increases only
memory use.

Maybe this can be generalized to composition of any
signal/image-processing operations: instead of working with
intermediate representations, use deforestation to construct memoized
expression iterators.