Why use the logarithm of the determinant?

On positive definite matrices, logarithm is increasing, so maximizing log determinant gives the same optimum while providing a useful optimization formulation.

Is D-optimal design always better than A-optimal design?

No. D targets joint volume; A targets a variance sum. Their preferred allocations can differ.

Does D-optimality guarantee good prediction everywhere?

No. Parameter-information quality under one model is not a universal guarantee of prediction quality across every input region.

AI Glossary / Causal inference and experimentation / Advanced

D-optimal design

Definition

D-optimal design selects an experimental allocation that maximizes the determinant of its information matrix. In standard estimation settings, this is equivalent to minimizing the determinant of the parameter covariance matrix and, under appropriate confidence-region assumptions, its uncertainty volume.

What it means

Experiments often estimate several quantities together. Looking at one variance at a time can miss how their uncertainties interact. D-optimality offers a joint criterion rather than choosing a separate target for every parameter.

The determinant summarizes a multidimensional scale. It is not a probability that the experiment succeeds, and its numerical value depends on the parameterization. When designs use the same parameterization, model and budget, it can support a meaningful comparison.

The word optimal remains conditional. A design can optimize the specified information model while performing poorly if the response relationship or noise assumptions are wrong. It can also be unsuitable operationally if important constraints were omitted from the search.

How it works

For a linear model, each candidate measurement contributes to an information matrix M. The planner chooses allowed observation counts or approximate sampling weights. The D criterion maximizes det(M), often implemented by maximizing log det(M) when M is positive definite.

The inverse relationship between information and covariance explains the direction of optimization: more determinant of information corresponds to less determinant of covariance, with common noise factors treated consistently. Singularity indicates that the selected experiment cannot identify all the included parameter directions.

A continuous-weight solution must be translated into feasible observations if experiments come in indivisible units. That translation can change performance, so the final allocation should be scored again rather than inheriting the ideal solution's label.

Three practical examples

A laboratory chooses combinations of experimental settings to estimate several response coefficients jointly.
A preference-study team selects comparisons intended to identify a model's trade-off parameters.
A sensor planner allocates measurement opportunities to reduce the overall uncertainty volume in a calibration model.

Think of it this way

Think of shrinking a flexible oval drawn around plausible combinations of two unknown quantities. D-optimality focuses on the oval's area in this two-dimensional picture.

An oval can become smaller while remaining long in one direction. That is why shrinking total volume does not necessarily protect the least precisely estimated combination. The analogy also assumes an ellipsoidal uncertainty description; real uncertainty regions need not have that shape.

Worked example

Why equal allocation can be D-optimal

Consider two independent quantities with single-observation noise variances 1 and 4. Estimate them using independent sample means with positive counts n1 and n2. Their covariance matrix is diagonal, with entries 1/n1 and 4/n2.

With twelve measurements, n1 + n2 = 12. The covariance determinant is 4/(n1n2). Minimizing it is equivalent to maximizing n1n2. For a fixed sum of twelve, the product is largest at n1 = n2 = 6.

The determinant is then 4/36 = 1/9, approximately 0.1111. Allocating four and eight measurements instead gives 4/32 = 0.125, a larger determinant. The six-and-six allocation is therefore D-optimal within this simple setup.

Interestingly, the four-and-eight allocation has a smaller variance sum: 0.75 instead of 5/6. It wins under the A criterion while losing under the D criterion. This is a difference in objectives, not an arithmetic contradiction. The example assumes known noise variances, equal observation costs and no shared measurement errors.

Illustrative case study

Fictional case: choosing a trial schedule

A fictional manufacturing research team plans a controlled study of two adjustable settings. The analyst proposes a D-optimal schedule based on a specified response model. The operations lead notices that one suggested setting combination is impractical during normal production.

They add feasibility restrictions and regenerate the schedule. They also retain a separate check for lack of fit because an efficient coefficient estimate is unhelpful if the model omits an important interaction.

The resulting report includes the final integer schedule, assumptions and comparison with the original feasible baseline. It does not claim improved product quality merely because the information determinant increased. That claim would require actual outcome evidence.

Common misconceptions and mistakes

D-optimality does not minimize every parameter variance. A design can reduce joint uncertainty volume while leaving a particular direction relatively uncertain. If that direction drives the decision, another criterion or an additional constraint may be appropriate.

A positive determinant is not proof that the measurements are unbiased or representative. It only describes the specified information structure.

Changing parameter units changes determinant values, so raw scores from unrelated models should not be compared casually. However, a fixed invertible linear reparameterization applied to every candidate multiplies their determinants by the same positive factor. Their D-optimal ranking is preserved in that comparison.

Business use

D-optimal planning can make a limited experiment budget more purposeful when the organization needs to estimate a joint parameter model. Define the candidate conditions, costs and operational restrictions before interpreting the optimizer's output.

Inspect individual uncertainties as well as the determinant, especially when one coefficient has a direct business consequence. Use sensitivity analysis to examine whether different plausible model assumptions produce substantially different schedules. An unstable recommendation signals that gathering preliminary information may be worthwhile before committing the full study budget.

Check your understanding

Two designs have diagonal covariance matrices with entries (1, 9) and (4, 4). Which has the smaller covariance determinant?

Answer: The first: 1 × 9 = 9, versus 4 × 4 = 16. It is preferred by this determinant comparison despite having a larger variance sum and a worse largest variance.

Related terms

A-optimal design
E-optimal design
Determinant
information matrix
confidence region
experimental design
identifiability

Learn next

Compare determinant, trace and largest-eigenvalue criteria on two-dimensional covariance examples. Then examine how practical constraints turn an ideal fractional design into an executable schedule of whole experiments.

Determinant

Sources

Primary technical source: Boyd and Vandenberghe, experimental-design example accompanying Convex Optimization, with D-optimal information-matrix formulations: https://cvxr.com/cvx/examples/cvxbook/Ch07_statistical_estim/html/expdesign.html

The twelve-measurement comparison and manufacturing case are original teaching examples.

Previous: D* Next: D-Separation