📊 statistics

A Bayesian-Geometric Reading of Classical Model-Based Optimal Design

This paper unifies classical model-based optimal design criteria (D-, T-, L-, C-, and A-optimality) under a single Bayesian-geometric framework by demonstrating that they asymptotically emerge as leading-order terms of expected dissimilarities between prior and posterior distributions, with E-optimality arising as a robustification of C-optimality.

Original authors: Mohamed Tarek

Published 2026-11-04

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Mohamed Tarek

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Scientists who build statistical models to understand the world face a constant, practical dilemma: how to gather the most useful information with the fewest resources. Whether they are tracking the spread of a disease, testing a new drug, or mapping the interior of a planet, they must decide where to look, when to measure, and how many times to repeat an experiment. If they choose poorly, they might collect a mountain of data that tells them nothing new. If they choose well, a handful of observations can reveal the hidden rules governing a system. This challenge is known as optimal experimental design. For decades, researchers have relied on a set of standard rules to make these choices, treating the problem as a mathematical puzzle involving the precision of their measurements. These rules, often named with letters like D, T, and A, have been justified by looking at how much a model's uncertainty shrinks as data accumulates. However, a new perspective suggests these rules are not just about shrinking error bars, but about the fundamental shift in what we believe before and after an experiment.

A researcher has re-examined these classic rules through the lens of Bayesian geometry, a way of thinking that treats probability distributions as shapes in space. In this view, an experiment is a journey that moves our understanding from a starting point, called the prior, to a destination, called the posterior. The prior represents what we believe before seeing any data, while the posterior represents our updated beliefs after the data arrives. The goal of a good experiment is to create the largest possible gap between these two states of mind. The researcher asked a simple question: what if we measure the "distance" between our starting beliefs and our final beliefs, and then choose the experiment that maximizes this distance on average? By applying this idea, they discovered that the five most famous rules for designing experiments are not arbitrary formulas, but rather the leading terms of five different ways to measure this distance.

The study shows that these rules emerge naturally when we assume the experiment involves a large amount of data and that our initial beliefs can be described by a smooth, bell-shaped curve. The researcher found that if you measure the distance between the starting and ending beliefs using a specific method called the Kullback-Leibler divergence, where you look at how much the new data changes the old beliefs, you arrive at the rule known as D-optimality. This rule, which has been the gold standard for decades, essentially seeks to maximize the volume of information gained. However, if you reverse the direction of this measurement—asking how much the old beliefs would have to change to fit the new data—you arrive at a different rule called T-optimality, which focuses on the total amount of information available.

The researcher also explored a different way to measure distance, one that treats the probability distributions as physical objects that can be moved and reshaped. When they used a measure known as the Wasserstein distance, which calculates the effort required to transform one distribution into another, they uncovered the remaining three rules: L-optimality, C-optimality, and A-optimality. These rules, which have traditionally been justified by how they shrink the uncertainty of specific parameters, were shown to be the natural result of minimizing the effort needed to shift beliefs. The paper further revealed that a sixth rule, E-optimality, does not come from a single distance measure but arises when a designer wants to be safe against the worst-case scenario of not knowing exactly which direction their beliefs should move.

This work unifies a fragmented field by showing that these distinct rules are actually different faces of the same geometric principle. The choice of which rule to use depends entirely on how one chooses to measure the change in belief and how one defines the starting point of that change. For instance, the weight given to different parameters in a model is shown to be equivalent to the shape of the initial belief curve. The researcher emphasizes that these findings are asymptotic, meaning they describe what happens when the amount of data becomes very large. In situations with limited data, the shape of the initial beliefs matters more, and the simple geometric rules might need adjustment. Nevertheless, the study provides a powerful new way to understand why these rules work, grounding them in the intuitive idea that a good experiment is one that moves our understanding as far as possible from where we started.

Technical Summary: A Bayesian-Geometric Reading of Classical Model-Based Optimal Design

Problem Statement
Classical model-based optimal experimental design (OED) criteria—specifically D-, T-, L-, C-, and A-optimality—are traditionally motivated through the Cramér–Rao bound. In this classical framework, these criteria are defined as scalar functions of the Fisher Information Matrix (FIM), $F(\theta_0, \delta)$ , aimed at shrinking the asymptotic covariance of the maximum likelihood estimator. While effective, this perspective treats the criteria as distinct algebraic manipulations of the FIM. This paper seeks to provide a complementary, unified "Bayesian-geometric" reading. The central problem is to demonstrate that these five classical criteria (and a sixth, E-optimality) emerge as leading-order asymptotic terms ( $N \to \infty$ ) of expected dissimilarities between a prior distribution $p(\theta)$ and a posterior distribution $p(\theta | D, \delta)$ , where the FIM enters via the Bernstein–von Mises (BvM) limit.

Methodology
The paper adopts an asymptotic Bayesian framework where the model structure is assumed true, but parameters $\theta$ are unknown. The methodology proceeds through the following steps:

Asymptotic Setup: The paper assumes standard regularity conditions where the maximum likelihood estimator (MLE) is asymptotically Gaussian. By the BvM theorem, the posterior distribution $p(\theta | D, \delta)$ converges to a Gaussian distribution centered at the MLE with covariance $N^{-1}F(\theta_0, \delta)^{-1}$ .
Expected Dissimilarity Objective: The design objective is defined as the expected dissimilarity $U(\delta)$ between the prior and the posterior, where the expectation is taken over the true data-generating distribution $p(D | \theta_0, \delta)$ :
$U(\delta) = \mathbb{E}_{D|\theta_0, \delta} [d(p(\theta), p(\theta | D, \delta))]$
Here, $d$ represents a measure of dissimilarity (distance or divergence).
Derivation via Specific Dissimilarities: The paper derives the classical criteria by selecting specific dissimilarity functions and prior parametrizations, then expanding the expected dissimilarity for large $N$ $N$ :
- KL Divergence (Forward): $d = \text{KL}(p(\theta|D) \parallel p(\theta))$ .
- KL Divergence (Reverse): $d = \text{KL}(p(\theta) \parallel p(\theta|D))$ .
- Wasserstein Distance: $d = W_2^2(p(\theta), p(\theta|D))$ , specifically using a first-order approximation of the squared 2-Wasserstein distance involving the matrix square root.
Prior Parametrization: For weighted criteria, the Gaussian prior covariance $\Sigma_0$ is explicitly chosen to encode the weight matrices (e.g., $L$ or $c$ ) associated with the specific optimality criterion.
Robustification: E-optimality is derived not as a direct limit of a single dissimilarity, but as a min-max robustification of C-optimality against uncertainty in the direction of the parameter of interest.
Designer vs. Modeler Priors: The framework is extended to distinguish between a "designer's prior" (used to average over plausible true parameters $\theta_0$ at design time) and a "modeler's prior" (used for inference).

Key Contributions and Results

Unified Template: The paper establishes that D-, T-, L-, C-, and A-optimality are special cases of a single Bayesian-geometric template, distinguished by three choices: the dissimilarity measure, the prior parametrization, and the approximation scheme.
D-Optimality: Recovered as the leading-order term of the expected forward KL divergence, $\mathbb{E}[\text{KL}(p(\theta|D) \parallel p(\theta))]$ . This reproduces the classical result (Stone, 1959; Chaloner and Verdinelli, 1995) where maximizing information gain corresponds to maximizing $\log \det F(\theta_0, \delta)$ .
T-Optimality (Trace): Novelly derived as the leading-order term of the expected reverse KL divergence, $\mathbb{E}[\text{KL}(p(\theta) \parallel p(\theta|D))]$ . The paper shows that maximizing the trace of the FIM, $\text{tr}(F(\theta_0, \delta))$ , corresponds to minimizing the reverse KL divergence, with the prior covariance acting as the weight matrix.
L-, C-, and A-Optimality: Novelly derived as the leading-order terms of the expected squared 2-Wasserstein distance.
- L-optimality arises from a general positive semi-definite weight matrix $B$ (where $B=LL^T$ ) in the prior covariance.
- A-optimality is recovered when the weight matrix is a scalar multiple of the identity.
- C-optimality is recovered when the weight matrix is rank-1 ( $B = \alpha^2 cc^T$ ), corresponding to a specific linear functional of interest.
E-Optimality: Recovered not as an independent dissimilarity limit, but as the min-max robustification of C-optimality. By maximizing the minimum eigenvalue of the FIM, the design guards against the worst-case direction of the parameter vector $c$ when the designer is uncertain about the modeler's specific linear functional of interest.
Distinction of Priors: The paper clarifies the distinction between the designer's prior (over $\theta_0$ ) and the modeler's prior (over $\theta$ ). It shows that when these coincide and the dissimilarity is the forward KL divergence, the framework recovers the classical Expected Information Gain (EIG) objective.

Significance and Claims
The paper claims that the primary contribution is the presentation of all five classical criteria under this unified Bayesian-geometric template. Specifically, the author states that the derivations for T-, L-, C-, and A-optimality from their respective expected dissimilarities appear novel. While the recovery of D-optimality via KL divergence is a known result, the paper offers a new perspective by showing how different dissimilarity measures (forward vs. reverse KL, Wasserstein) naturally lead to different classical criteria.

The framework positions these criteria not as arbitrary algebraic choices, but as geometric separations between prior and posterior distributions. The paper modestly notes that while the asymptotic derivations are rigorous under the BvM limit, the approach relies on large sample sizes ( $N \to \infty$ ) and correct model specification. It acknowledges that for moderate $N$ , non-Gaussian posteriors (e.g., in nonlinear models) may reduce the accuracy of these approximations. The paper does not propose new computational algorithms or specific experimental applications but rather provides a theoretical unification of existing design criteria.

Technical Summary: A Bayesian-Geometric Reading of Classical Model-Based Optimal Design

More like this