Iterative expansion

⚠️ Curated Limitations & Negative Results

• The monotonic correspondence concerns the static evaluation supplied to both procedures, not merely stored values after IE has raised them through pathmax. Full RBFS uses conditional inheritance and SRBFS uses static reinitialization; local overlap does not make IE a global alias or establish equality on every nonmonotonic instance.
• The outline returns infinity when a node has no children, but the caller shows a bare recursive call followed by a minimum over mutable child values. How that infinity is stored back in the child and how successful return propagates are not fully specified. Preserve the no-children guard before child selection; no corrected executable assignment or actual implementation failure is asserted.
• The printed do-until/or-while wording is not a precise programming-language loop. Under its bounded-search reading, IE tests the parent’s stored value before backing it up, whereas Korf tests the best child before recursion; an immediate cutoff call can therefore occur in IE when the returned bound agrees. Exact call-trace equivalence is not claimed.
• Equal-value child ordering is not fully specified in the IE outline. An exact expansion trace or choice among tied goals requires an explicit shared tie and successor order; SMA*’s deepest-leaf rule must not be transferred to IE merely because it appears in the same paper.
• For a single child, the minimum over other children needs an empty-alternative convention so the recursive bound remains the caller’s bound. The outline also does not fully specify failure after an exhausted root with an infinite bound; infinity <= infinity does not itself signal failure. No unrestricted unsolvable-instance termination guarantee follows.
• The source treatment excludes graph-transposition complications. The IE outline supplies no general closed-list, duplicate-state or reopening policy. Korf’s current-path cycle check for nonincreasing evaluation is a separate scoped condition, not a global duplicate-elimination guarantee for IE.
• Ordering nodes by an evaluation function does not alone establish minimum path cost, completeness or termination. Stronger results require appropriate cost/admissibility, goal and finiteness assumptions. Korf’s summary omits full proofs, and no general theorem is certified here for infinite spaces, zero-cost cycles, arbitrary negative costs or unqualified unsolvable instances.
• The storage bound counts retained node information in a path-and-sibling representation. It does not bound all bytes, heuristic tables, call-stack overhead or temporary allocations, and it is not independent of branching or attained depth. Reduced storage can require regeneration of earlier paths.
• The papers’ puzzle comparisons do not establish universal speed or node-count equality between IE and RBFS. Korf’s related-work expectation about IE is not a reported direct IE-versus-RBFS benchmark. No runtime ratio, performance dominance, implementation equivalence or benchmark reproduction is claimed.
• Evidence is limited to Russell’s inspected five-page 1992 manuscript and Korf’s six-page 1992 proceedings summary. The Korf 1991 predecessors, 1993 journal article, mismatched Bratko citations, later RBFS paper and 1992 ITS report remain separate source/version questions. All 30 bibliography occurrences and seven pending questions are retained without asserting that every cited original was read or any source family is exhausted.