IPOP-CMA-ES

⚠️ Curated Limitations & Negative Results

• The 2005 author-hosted copy includes an explicit no-effect-coordinate erratum, used above. Its separate author-hosted ERRATA text also reports that the F4 initial mean was mistakenly sampled with Xup rather than Xlow in the offset, and that boundary settings were omitted on some functions. The authors suggest limited or favorable effects of correction, but those effects were not reproduced. No numerical superiority or repaired benchmark result is curated.
• The original inner routine is the 2004 positive-weight rank-mu and evolution-path CMA-ES, not merely the maintained 2001 rank-one CMA description and not later active/negative-weight variants. The 2004 conclusion suggests a restart multiplier of 3; the concrete 2005 procedure chooses 2. The 2005 discussion of factors between 1.5 and 5 is heuristic context, not a proven optimal range or an identification of every such schedule with the original procedure.
• The literal small-path stopping prose says sigma*p_c is smaller than TolX in all components without explicitly printing absolute-value syntax. The draft retains the stopping category and does not claim a verified executable interpretation. The prose also does not fully specify partial-generation evaluation or budget overshoot behavior. The referenced cmaes.m version 2.35 was not acquired or executed.
• The fixed objective-value tolerance, known-optimum error target, coordinate no-effect checks, condition-number threshold, initialization box and boundary policy limit claims about the whole protocol. An increasing transformation that preserves rankings can change a numeric function-range stopping test. Corresponding transformations of inner distributions alone do not certify invariance of every configured restart run.
• Population growth changes the allocation of evaluations across attempts, but numerical stopping and a finite budget do not establish a global optimum, monotonic distribution quality or universal improvement over constant-population restarts. The source empirical comparison also changes initial scale. No optimizer, benchmark or author code was run for this draft.
• The success-runtime expression in the 2005 Section 3 concerns independent repetition of a complete algorithm with positive success probability. It must not be applied without qualification to heterogeneous internal IPOP attempts with changing populations. Table 2 on PDF p. 5 displays SP2=0 for some zero-success rows, although Equation 4 on PDF p. 3 divides by that estimated success probability; those entries are not evidence of zero expected runtime. No success-rate or expected-runtime result is claimed.
• The 2004 base-paper errata correct a benchmark penalty and identify the initialized point as the recombination mean. Source build dates and later author errata are distinct from the 2004 and 2005 publication years. The source version and acquisition records are retained; no complete certification of all baseline proofs, numerical safeguards or benchmark tables is asserted.
• The local-restart companion is cited only for the defining method and initialization comparison. BIPOP is retained solely as a separate unresolved boundary: its full primary methods were not read and no substantive BIPOP construction, equivalence or performance claim is made. Existing rank-mu, active-CMA and other source-family leads are not resolved by this one named algorithm.
• Supporting research read all nine pages of the IPOP author copy including its erratum, selected defining and discussion pages of the 2004 base, and the first four pages of the local-restart companion. Exact text/visual scopes and the narrower root reading are retained separately. The historical wrapper/no-match proposal is preserved; the current mapping follows the accepted whole-method scope. Identity and frontier checks remain bounded to their recorded snapshot and must be refreshed before curation.