"Efficient" least- and greatest fixpoint computations?

Question

I am trying to compute two finite sets of some enumerable type (let's say char) using a least- and greatest- fixpoint computation, respectively. I want my definitions to be extractable to SML, and to be "semi-efficient" when executed. What are my options?

From exploring the HOL library and playing around with code generation, I have the following observations:

1. I could use the complete_lattice.lfp and complete_lattice.gfp constants with a pair of additional monotone functions to compute my sets, which in fact I currently am doing. Code generation does work with these constants, but the code produced is horribly inefficient, and if I understand the generated SML code correctly is performing an exhaustive search over every possible set in the powerset of characters. Any use, no matter how simple, of these two constants at type char therefore causes a divergence when executed.
2. I could try to make use of the iterative fixpoint described by the Kleene fixpoint theorem in directed complete partial orders. From exploring, there's a ccpo_class.fixp constant in the theory Complete_Partial_Order, but the underlying iterates constant that this is defined in terms of has no associated code equations, and so code cannot be extracted.

Are there any existing fixpoint combinators hiding somewhere, suitable for use with finite sets, that produce semi-efficient code with code generation that I have missed?

Answer 1

None of the general fixpoint combinators in Isabelle's standard library is meant to used directly for code extraction because their construction is too general to be usable in practice. (There is another one in the theory ~~/src/HOL/Library/Bourbaki_Witt_Fixpoint.) But the theory ~~/src/HOL/Library/While_Combinator connects the lfp and gfp fixpoints to the iterative implementation you are looking for, see theorems lfp_while_lattice and gfp_while_lattice. These characterisations have the precondition that the function is monotone, so they cannot be used as code equations directly. So you have two options:

1. Use the while combinator instead of lfp/gfp in your code equations and/or definitions.

2. Tell the code preprocessor to use lfp_while_lattice as a [code_unfold] equation. This works if you also add all the rules that the preprocessor needs to prove the assumptions of these equations for the instances at which it should apply. Hence, I recommend that you also add as [code_unfold] the monotonicity statement of your function and the theorem to prove the finiteness of char set, i.e., finite_class.finite.