Design and analysis of fractional factorial experiments from the viewpoint of computational algebraic statistics

We give an expository review of applications of computational algebraic statistics to design and analysis of fractional factorial experiments based on our recent works. For the purpose of design, the techniques of Gr"obner bases and indicator functions allow us to treat fractional factorial designs without distinction between regular designs and non-regular designs. For the purpose of analysis of data from fractional factorial designs, the techniques of Markov bases allow us to handle discrete observations. Thus the approach of computational algebraic statistics greatly enlarges the scope of fractional factorial designs.

💡 Research Summary

This paper provides a comprehensive review of how computational algebraic statistics can be employed for both the construction and the analysis of fractional factorial experiments. The authors focus on three central algebraic tools: Gröbner bases, indicator functions, and Markov bases, illustrating how each contributes to a unified treatment of regular and non‑regular designs and to the handling of discrete response data.

In the design phase, the set of treatment combinations is encoded as a polynomial ideal. Computing a Gröbner basis of this ideal yields a canonical generating set of algebraic constraints that the design must satisfy. This approach eliminates the traditional distinction between regular designs—characterized by simple defining relations—and non‑regular designs, whose constraints are often more intricate. The indicator function, a 0‑1 valued polynomial that signals whether a given point belongs to the design, is derived directly from the Gröbner basis. It enables straightforward verification of projection properties (e.g., marginal balance) and facilitates the assessment of design optimality criteria such as minimum aberration or uniformity.

For data analysis, the paper turns to Markov bases, which arise from the theory of toric ideals associated with the sufficient statistics of a log‑linear model. A Markov basis provides a set of moves that connect all contingency tables sharing the same sufficient statistics, guaranteeing that a Markov chain Monte Carlo (MCMC) sampler can explore the exact conditional distribution of the data. This is particularly valuable for fractional factorial experiments where the response is discrete (counts, binary outcomes, etc.) and where traditional chi‑square or asymptotic likelihood‑ratio tests may be unreliable, especially under non‑regular designs. By using the Markov basis, exact p‑values can be obtained without resorting to large‑sample approximations.

The authors demonstrate the practical implementation of these concepts using open‑source software such as 4ti2 and Macaulay2. They present case studies involving both a 2⁵‑1 non‑regular design and a 3⁴‑2 regular design. For each case, they construct the design ideal, compute its Gröbner basis, and derive the indicator function to confirm design properties. In the analysis stage, they generate the corresponding Markov basis, run MCMC simulations, and compare the resulting exact tests with conventional approximations. The results show that the algebraic approach yields more conservative and reliable inference, while also offering a systematic way to handle designs of arbitrary complexity.

Beyond the examples, the paper discusses broader implications. By framing design construction and analysis within a single algebraic language, researchers can seamlessly transition from design generation to data interpretation without switching methodological frameworks. This unification is especially advantageous when dealing with mixed‑level factors, unbalanced designs, or when extending to higher‑order interactions that would be cumbersome to treat with classical combinatorial methods. Moreover, the algebraic perspective opens avenues for automated design optimization, such as searching for designs that minimize aberration or maximize D‑efficiency, through Gröbner basis computations.

Finally, the authors outline future research directions. Key challenges include scaling Gröbner and Markov basis computations to experiments with many factors (e.g., ten or more) and to designs that combine continuous and discrete factors. Development of more efficient algorithms, parallel implementations, and integration with modern statistical software (R, Python) are identified as critical steps toward making computational algebraic statistics a routine tool for experimental scientists. In summary, the paper convincingly argues that computational algebraic statistics dramatically expands the scope and robustness of fractional factorial experimentation, offering exact, unified, and computationally tractable methods for both design and analysis.