Skip to content

Theory in brief

For a reader who wants the model stated in one place: definitions only, no proofs; the paper has them.

The invariant form

Let \(m\in\mathbb{R}^d_{>0}\) be the metrics of a design, positive and larger-is-better, and \(u=\log m\). The model scores a design by a weighted geometric mean

\[ \begin{aligned} V(m)&=\prod_{j=1}^{d} m_j^{\,w_j},\\ w_j&\ge 0,\quad \sum_j w_j=1 . \end{aligned} \]

Because the weights sum to one, \(V(\lambda m)=\lambda V(m)\): doubling every metric doubles V. So V reads as a factor of improvement, which is what "invariant" refers to. A sum of such products, a posynomial, extends one term to K terms and keeps the programs inside geometric programming. The model is a root of that sum, so the surface V = 1 is where the posynomial equals one. Fitting one term is a geometric program (GP), solved as a convex program after the logarithmic change of variables. Fitting several terms alternates assigning the designs to terms with a separate convex refit of each term. By default (term_temperature=np.inf) the package takes the sharp limit of the weights, one label per design, and the labels drive the refit. Reading off which term scores a design highest is analysis of the fitted model, not part of the fit.

V = 1 is the fitted reference surface

The fit chooses \(w\) and an offset \(\beta\) so that the data lie on or below the surface \(\log V=w\cdot u-\beta=0\), using the log-mean-exp (entropic) soft maximum of \(w\cdot u\) over the designs. With one term this is a convex program. Three objects answer to the word frontier:

  • the fitted reference surface, V = 1 at the fit temperature tau; it is smooth, so a few designs lie beyond it;
  • the supporting envelope, the same fit at tau = infinity: the exact maximum, a linear program, with no design above it;
  • the Pareto boundary, the nondominated set of the attainable designs, a property of the system and not of the fit. The fit does not identify the other two with it.

Inside the surface V < 1; beyond it V > 1, and V - 1 says by how much. The logarithm, \(\log V\), is the energy differential (level offset) to the surface.

The chart

Let \(B\) be a basis of the sum-zero subspace of \(\mathbb{R}^d\) (a Helmert log-ratio basis by default). Each design has the coordinates

\[ \big(Z,\ \log V\big),\qquad Z=Bu\in\mathbb{R}^{d-1}, \]

where \(Z\) is free of overall scale and says where along the surface the design sits, and \(\log V\) says how far from it. The map is reversible for every row with strictly positive metrics. Regions of the \(Z\) plane that no evaluated design occupies are white space.

Elasticities and the n-lemma

The surface \(w\cdot u=\beta\) is flat in log space, so moving along it trades the contending metrics at constant elasticities:

\[ \frac{\mathrm{d}\log m_j}{\mathrm{d}\log m_k}=-\frac{w_k}{w_j}. \]

The metrics with positive weight at tau = infinity are the contending set, and its size n is the n-lemma, drawn in Figure 1 of the paper. A dilemma is n = 2, a trilemma n = 3. The set is found from the data and not assumed. The fit at tau = infinity is a linear program whose solution is generically sparse. At a finite fit temperature a small weight means weak participation, set by a participation threshold. With K terms there is one contending set per term, and elasticities vary between terms.

What the model assumes

Metrics are positive and larger-is-better (conditioning is yours to choose); the surface is log-convex; at one term the elasticities are constant; designs are chosen by an evaluator and not drawn from the attainable set. The limitations page says what each means when you act on a result.