The hypotheses Theorem 1 of Top Program Construction needs: small witnesses and corrected statements (Comment on arXiv:2101.05050)
Year: 2026
Type: preprint
Abstract: Patsantzis and Muggleton (Machine Learning 110:755-778, 2021) define the Top program of a meta-interpretive learning problem as the set of clauses that each entail a positive example and no negative example against the background knowledge B, and state (Theorem 1) that the Top program is a correct hypothesis whenever a correct hypothesis exists. The printed proof assumes that every clause of a correct hypothesis entails a positive example on its own and that individually consistent clauses are jointly consistent. Neither holds when clauses interact. The first fails through recursion and invented predicates. The second fails through a background that mentions the target (for Definition 1), through recursion (for the test of the paper's Section 4.1), and through recursion and invented predicates (for the description of the Top program as the union of all correct hypotheses). We give a small witness for each failure, including a printed example. We then state and prove two regimes in which a corrected form holds. Definition 1's Top program is correct for arbitrary examples when B mentions no target symbol, no invented predicate is used, and some correct hypothesis has no target symbol in a clause body. The test of Section 4.1, Louise's single-clause mode, is correct when every target atom is labelled and no invented predicate is needed, even with recursion and with B reading the target. Witnesses show that these conditions cannot simply be dropped. The condition on labels is made explicit in earlier work on extensional coverage; to our knowledge, what is new here is the refutation of this particular theorem and its repair. The named witnesses are machine-checked. Scripts accompany the note, with a brute-force suite of 489,072 instances that found no counterexample to the corrected statements (466,056 test the theorems directly; the rest are cross-checks and boundary sweeps). Files. top_program_theorem1_note.pdf is the note and top_program_theorem1_note.tex its LaTeX source. theorem1_checks.zip holds the checking scripts and the full brute-force suite with its recorded outputs; start with README.txt inside it. Tooling. Large-language-model tooling was used throughout: to search the literature, to find and minimise the witnesses, to write and run the checking code and the brute-force suite, to attack the proofs, and to draft the text. The author takes responsibility for the mathematics, the witnesses and the text, and the scripts accompany the note so that they can be checked independently. A copy of an earlier version of this note was sent to the authors of the paper on 29 October 2026.
Author: Siddiqui, Raed
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7225313655