Skip to content

Limitations

Conditioning defines the question. Moving the reference or the scale of one metric changes which term is tangent. Refit under a second reasonable conditioning and report what moves (Conditioning).

The surface is log-convex. Dents and disconnected regions are invisible to it, and no number of terms represents them. White space is an opportunity only if the trade space is log-convex around it.

One term has one trade-off. A single term has constant elasticities and finds one trade-off, at its balanced point. Away from that point a design can score below 1 while locally on the frontier. If a weight is exactly zero at the balanced point, exponents are not unique and some elasticities are undefined. For regionally varying elasticities fit several terms.

The surface is not the Pareto boundary. The fitted reference surface and the supporting envelope are determined by the sample you give; the Pareto boundary is a property of the system.

Noise and temperature. At a finite temperature some designs lie beyond V = 1 by construction. Judge a score against the band 1/tau. A high tau rests the fit on a few extreme designs. Below tau of about 3 the soft maximum approaches the mean.

Small samples. With few designs, or designs that barely differ near the edge, exponents are weakly determined. Read n_eff_ and conditioning_ (Reading termination). Designed samples, such as grids, are not independent observations: the exponents describe the sample and carry no sampling error.

Contending sets are a reading. The set is exact at infinite temperature with min_exponent at 0. At a fit temperature participation is graded. A trade-off is a binding constraint, not a correlation, and negative association near the surface is partly a closure effect.

Local fits. A multi-term fit is a local optimum chosen among seeded restarts. A trade-off with few frontier rows can be missed. More restarts do not cure this. A larger sample can help.

Novelty is not preference. V - 1 says how far a design lies beyond the fitted reference surface, not that it is wanted, feasible or safe.

Prediction is conditional. A prediction inherits the regressor's error and the fitted trade space, and inverting a level into a bounded outcome is left to you (Predicting).

Machine checks confirm that formulas follow from stated conditions and that the code agrees with them. Whether the conditions suit your problem is for you to judge. The lemma scripts replace pencil-and-paper derivation, not formal verification. Appendix E of the paper states what they do and do not establish. A proof in a theorem prover would fix each lemma's statement over the real numbers. It would not bound the double-precision error, which the certificates check.