Two polynomial representations of experimental design

In the context of algebraic statistics an experimental design is described by a set of polynomials called the design ideal. This, in turn, is generated by finite sets of polynomials. Two types of generating sets are mostly used in the literature: Groebner bases and indicator functions. We briefly describe them both, how they are used in the analysis and planning of a design and how to switch between them. Examples include fractions of full factorial designs and designs for mixture experiments.

💡 Research Summary

The paper provides a comprehensive comparison and integration of two algebraic representations of experimental designs that are widely used in algebraic statistics: Gröbner bases and indicator functions. It begins by framing experimental design as a design ideal— the set of all polynomials that vanish on the design points— and explains that, for computational purposes, this ideal can be generated by a finite set of polynomials. The two dominant generating sets are then introduced.

First, Gröbner bases are defined with respect to a chosen monomial order (lexicographic, graded reverse lexicographic, etc.). Because a Gröbner basis is unique for a given order, it furnishes a canonical normal form of the design ideal. This normal form makes the algebraic structure of the design transparent: the dimension of the design, the number of free variables, and the presence of hidden symmetries can all be read directly from the basis. Gröbner bases also enable rigorous tests of design equivalence (isomorphism), minimal generating sets, and algebraic diagnostics such as the Hilbert series.

Second, indicator functions are introduced as a single polynomial that takes the value 1 on points belonging to the design and 0 elsewhere. This compact representation is particularly useful in the planning stage, where one wishes to verify quickly whether a candidate design satisfies certain constraints, or to embed the design problem into an optimization framework that can be expressed as a polynomial program. Indicator functions also expose the combinatorial structure of fractional factorial designs and mixture designs in a way that is often more intuitive than the Gröbner basis.

The core contribution of the paper lies in the explicit algorithms for converting between the two representations. To obtain an indicator function from a Gröbner basis, the authors compute the standard monomials of the design ideal and construct a linear combination that reproduces the 0‑1 pattern on the design points. Conversely, to derive a Gröbner basis from an indicator function, the indicator polynomial is added to the defining equations of the design space (e.g., sum‑to‑one constraints for mixtures) and a Gröbner basis computation is performed on the resulting ideal. These conversion procedures are implemented in standard computer algebra systems such as CoCoA, Singular, and Macaulay2, demonstrating that the workflow can be fully automated.

Two illustrative examples are presented. The first deals with fractional factorial designs derived from a full factorial lattice. In this case, the Gröbner basis is relatively simple, reflecting the regular grid structure, while the indicator function clearly displays the defining aliasing relations and the selection of treatment combinations. The second example concerns mixture experiments, where the variables are constrained to sum to one. Here the indicator function delineates the feasible region of the simplex and flags which points belong to the chosen design, whereas the Gröbner basis captures the algebraic constraints (including the sum‑to‑one equation) and reveals the effective dimensionality of the design space.

The authors conclude that Gröbner bases and indicator functions are complementary tools. Gröbner bases excel at uncovering the underlying algebraic geometry of a design, making them ideal for analysis, dimension counting, and equivalence testing. Indicator functions, on the other hand, are more suited to design construction, feasibility checking, and integration with polynomial optimization. By providing concrete conversion algorithms, the paper equips practitioners with the flexibility to switch between the two viewpoints as the problem demands, thereby enhancing the efficiency of both experimental design analysis and planning. The work also suggests that these algebraic techniques can be combined with traditional statistical design methods to generate novel designs and to improve model selection in complex experimental settings.