What does the E refer to?

It is associated with an extreme-eigenvalue criterion: the smallest information eigenvalue or corresponding largest covariance eigenvalue.

Does E-optimality reduce total uncertainty most?

Not necessarily. Total variance and uncertainty volume are different objectives, represented by A- and D-type criteria.

Can E-optimality fix an unidentifiable model?

No. If every feasible design lacks information in a required direction, the experiment or model must change.

AI Glossary / Causal inference and experimentation / Advanced

E-optimal design

Definition

E-optimal design chooses an experimental allocation that maximizes the smallest eigenvalue of its information matrix. In standard settings with covariance proportional to inverse information, it minimizes the largest covariance eigenvalue, protecting the least precisely estimated parameter direction.

What it means

An experiment can estimate most combinations of parameters well while leaving one combination poorly determined. E-optimality concentrates on that weakest direction.

A direction can combine several parameters; it is not necessarily the parameter with the largest individual variance. Correlations affect the geometry of joint uncertainty, so checking only the covariance diagonal can miss the most uncertain combination.

This criterion is useful when a large uncertainty in any allowed unit-length parameter direction is undesirable. Its meaning depends on the parameter scales and the model. A direction measured in an arbitrary mixture of incompatible units is not automatically a sensible business priority.

How it works

Construct an information matrix M from the candidate measurements and their proposed allocation. E-optimality seeks the largest value t such that M − tI is positive semidefinite, where I is the identity matrix. This requires every information eigenvalue to be at least t.

When M is positive definite and covariance equals a common positive factor times M inverse, the largest covariance eigenvalue corresponds to the reciprocal of the smallest information eigenvalue. Maximizing one therefore minimizes the other in the required direction.

The optimization still needs a feasible experiment set, cost constraints and an appropriate noise model. If the candidate measurements cannot identify every included parameter direction, rearranging their allocation cannot manufacture missing information.

Three practical examples

A calibration team protects against one poorly measured combination of sensor parameters.
A researcher compares trial allocations when the widest direction of a model-based uncertainty region is the main concern.
A preference-study designer tries to avoid an experiment that learns some trade-offs accurately but leaves another nearly unidentified.

Think of it this way

Think of an oval uncertainty region and ask how to shorten its longest radius. A smaller total area alone does not answer that question: a thin, very long oval can still have modest area.

The analogy describes covariance geometry under the chosen model. It does not assert that all real uncertainty is elliptical or that the longest statistical axis necessarily matches the most consequential business error.

Worked example

Balancing the worst variance

Take two independent quantities measured in compatible units. Their single-observation noise variances are 1 and 4. With independent sample means based on n1 and n2 measurements, the covariance matrix is diagonal with entries 1/n1 and 4/n2.

There are fifteen measurements available, with positive counts. Because the covariance is diagonal, its largest eigenvalue is max(1/n1, 4/n2). Balancing these terms gives n2 = 4n1. Together with n1 + n2 = 15, this gives n1 = 3 and n2 = 12.

Both variances are then 1/3. An allocation of five and ten gives variances 0.2 and 0.4, whose maximum is larger. The three-and-twelve allocation minimizes the maximum here: achieving a maximum below 1/3 would require both n1 > 3 and n2 > 12, exceeding the budget.

This argument uses diagonal covariance. With correlated estimates, the largest diagonal entry need not equal the largest eigenvalue, so the same shortcut would be invalid. The example also assumes known noise variances and equal measurement costs.

Illustrative case study

Fictional case: reviewing a sensor experiment

A fictional robotics laboratory compares candidate calibration schedules. One schedule produces a small determinant of covariance but leaves an elongated uncertainty region. The engineers worry that the corresponding combination of parameters could affect an important motion estimate.

They compare an E-optimal schedule and examine the direction of its largest remaining uncertainty. They also check whether that statistical direction actually corresponds to the operational concern; if not, a targeted criterion may be more useful.

The laboratory documents the trade-off between joint volume and worst-direction precision. It then plans independent checks of model fit. The case describes a decision process, not a claim that a particular robot became safer or more accurate.

Common misconceptions and mistakes

E-optimality is not the same as minimizing the largest individual variance when off-diagonal covariance terms are present. For example, a covariance matrix with diagonal entries 1 and off-diagonal entries 0.9 has eigenvalues 1.9 and 0.1. Its weakest precision direction has variance 1.9, exceeding either diagonal entry.

The criterion is sensitive to parameter scaling. Changing units can change which direction counts as worst unless the scaling is handled deliberately.

An optimal value also does not establish model correctness. Systematic errors and omitted effects can remain even when the smallest information eigenvalue looks favorable.

Business use

Use E-optimality when the weakest estimated direction is a defensible target for an experiment budget. State why that target matches the intended decision, and compare the result with simpler allocations and other optimality criteria.

Have the analyst explain the critical direction in terms the project team can interpret. If it is a combination of parameters, provide the combination rather than only a matrix score. Check practical feasibility after translating weights into actual observations, and revisit the plan if preliminary data challenge the assumed noise structure.

Check your understanding

Two diagonal covariance matrices have entries (1, 9) and (4, 4). Which is preferred by minimizing the largest covariance eigenvalue?

Answer: The second. Its largest eigenvalue is 4, compared with 9 for the first. The first nevertheless has the smaller determinant, illustrating that E and D criteria can disagree.

Related terms

A-optimal design
D-optimal design
Eigenvalue
Covariance matrix
information matrix
identifiability
experimental design

Learn next

Plot a two-parameter covariance ellipse and relate its axes to eigenvectors and eigenvalues. Then compare how rescaling one parameter changes the A and E objectives before choosing a meaningful scale.

Sources

Primary technical source: Boyd and Vandenberghe, experiment-design example accompanying Convex Optimization. The executable E-design formulation maximises the information lower bound: https://cvxr.com/cvx/examples/cvxbook/Ch07_statistical_estim/html/expdesign.html

The allocations, covariance illustration and laboratory case are original teaching examples.

Previous: Dynamical system Next: E-values