In addition to the optimality criteria just listed, we also implement several ordering criteria for block designs (optimality criteria are ordering criteria that meet conditions described fully in a later subsection).
The number of distinct . For balanced incomplete block designs this value is 1. A balance criterion; the fewer variances a design produces, the easier are the results to understand.
The ratio of largest to smallest canonical variance ( ), called the canonical variance ratio. Again, the value for a balanced incomplete block design is 1. Values close to one correspond to variances that are quite similar.
The number of distinct . Analogous to no_distinct_canonical_variances, but for pairwise variances rather than canonical variances.
The ratio of largest to smallest pairwise variance ( ), called the pairwise variance ratio. Analogous to max_min_ratio_canonical_variances, but for pairwise variances rather than canonical variances.
The trace of the square of . This is called the S-criterion. Typically invoked as part of an (M,S)-optimality argument (minimize S subject to maximizing the trace of ). No direct statistical interpretation, though usually leads to reasonably ``good'' designs.
It was mentioned above that a complete block design (each block size is and each treatment is assigned to one unit in each block) is a ``good'' design. Now we state why. Over all possible assignments of treatments to blocks of size , a complete block design minimizes all of the criteria defined above (save for tr(), which it minimizes subject to the mean of the unsquared components). The same statement holds for a balanced incomplete block design for constant block size less than (whenever a BIBD exists). Otherwise, the optimal block design problem can be quite tricky, with such uniform optimality hard to come by.
An optimality_value for any of the optimality criteria above has three elements: its numerical value and two associated numbers absolute_efficiency and calculated_efficiency (for other_ordering_criteria, the same concepts are implemented under the names absolute_comparison and calculated_comparison so are not separately discussed here - see the later subsection on design orderings). Given any two designs, and say, they can be compared on any of the listed optimality criteria. The relative efficiency of design with respect to criterion , compared to design , is . If is in fact an optimal design as measured by ( minimizes over all ), then the relative efficiency of any compared to is the absolute_efficiency of . Both of these efficiencies are between 0 and 1, with smaller criterion values corresponding to larger efficiencies; the absolute efficiency of an optimal design is 1.
The concept of absolute efficiency depends on what is meant by the phrase ``all ''. It has already been explained that comparisons are for designs with the same , , and block sizes. In the external representation, an absolute_efficiency is for the class of all binary designs with the same , , and block size distribution, called the reference universe. When the minimum criterion value over the reference universe is not known, absolute_efficiency takes the value ``unknown.'' For a disconnected design absolute_efficiency takes the value ``0'' regardless of whether the optimal value is known or not. It happens, only rarely, that a smaller value of a criterion can be found for a nonbinary design with the same , , and block sizes, in which case the absolute_efficiency of the nonbinary design will be greater than 1. Nonbinary designs are not at present considered in the external representation. Relative efficiencies when the best value over the reference universe is not known, or within a subclass of the reference universe, can be calculated on a case-by-case basis; in external representation terminology, this is a calculated_efficiency. For instance, one may wish to compare only resolvable designs. calculated_efficiency takes the value ``0'' for all disconnected designs.