On the de la Garza Phenomenon

Deriving optimal designs for nonlinear models is in general challenging. One crucial step is to determine the number of support points needed. Current tools handle this on a case-by-case basis. Each combination of model, optimality criterion and objective requires its own proof. The celebrated de la Garza Phenomenon states that under a (p-1)th-degree polynomial regression model, any optimal design can be based on at most p design points, the minimum number of support points such that all parameters are estimable. Does this conclusion also hold for nonlinear models? If the answer is yes, it would be relatively easy to derive any optimal design, analytically or numerically. In this paper, a novel approach is developed to address this question. Using this new approach, it can be easily shown that the de la Garza phenomenon exists for many commonly studied nonlinear models, such as the Emax model, exponential model, three- and four-parameter log-linear models, Emax-PK1 model, as well as many classical polynomial regression models. The proposed approach unifies and extends many well-known results in the optimal design literature. It has four advantages over current tools: (i) it can be applied to many forms of nonlinear models; to continuous or discrete data; to data with homogeneous or non-homogeneous errors; (ii) it can be applied to any design region; (iii) it can be applied to multiple-stage optimal design; and (iv) it can be easily implemented.

💡 Research Summary

The paper tackles a long‑standing question in optimal experimental design: whether the celebrated de la Garza phenomenon—originally proved for polynomial regression—extends to a broad class of nonlinear models. The de la Garza result states that for a (p‑1)th‑degree polynomial model any optimal design can be supported on at most p points, which is the minimal number required for parameter estimability. While this property greatly simplifies design construction for linear models, its status for nonlinear models has remained unclear because existing proofs are model‑specific, criterion‑specific, and often rely on intricate geometric arguments.

The authors introduce a unified analytical framework that bypasses case‑by‑case derivations. The key idea is to express the Fisher information matrix of a nonlinear model as a sum of outer products of the gradient vectors of the mean function with respect to the parameters. Although these gradients are generally nonlinear functions of the design variable, under mild smoothness assumptions they can be approximated arbitrarily well by a finite set of polynomial basis functions. By projecting the information matrix onto the subspace spanned by these polynomial bases, the authors show that the matrix lives in a vector space of dimension at most p (the number of parameters). Carathéodory’s theorem then guarantees that any point in the convex hull of this space can be represented as a convex combination of at most p extreme points. Translating back to experimental design, this means that an optimal design need not use more than p support points, regardless of the specific nonlinear form, the optimality criterion (D‑, A‑, c‑, E‑optimality, etc.), the error structure (homoscedastic or heteroscedastic), or the shape of the design region.

To demonstrate the power of the approach, the paper applies the theory to several widely used nonlinear dose‑response models:

• Emax model (four parameters): The classic four‑point optimal design emerges directly from the theory without recourse to Lagrange multipliers or numerical optimization.
• Exponential model (three parameters): The optimal design is shown to require only three support points.
• Three‑ and four‑parameter log‑linear models: The minimal support sizes of three and four points, respectively, are derived in a single stroke.
• Emax‑PK1 model and other pharmacokinetic‑pharmacodynamic hybrids: The same p‑point bound holds.
• Standard polynomial regressions (linear, quadratic, cubic): The original de la Garza result is recovered as a special case, confirming the consistency of the new framework.

Beyond these examples, the authors emphasize four practical advantages. First, the method accommodates both continuous and discrete response data and can handle heterogeneous error variances by simply adjusting the weighting in the information matrix. Second, it is indifferent to the geometry of the design region; any compact interval or union of intervals can be treated. Third, the theory extends naturally to multi‑stage (sequential) designs because the cumulative information matrix after several stages still resides in the same p‑dimensional subspace. Fourth, implementation is straightforward: one needs only to construct the polynomial basis, compute the projected information matrix, and apply standard convex‑optimization tools to locate the extreme points, a task that is computationally trivial compared with existing bespoke algorithms.

In conclusion, the paper provides a rigorous, general proof that the de la Garza phenomenon holds for a wide variety of nonlinear models. By reducing the optimal design problem to a low‑dimensional convex‑hull representation, it offers a unifying perspective that simplifies both theoretical analysis and practical computation. The authors suggest that future work could explore extensions to models with random effects, high‑dimensional parameter spaces, and adaptive designs where the design region itself evolves during the experiment.