Kiefer's Measures in Optimal Design

• Kiefer's measures are a family of φ_p criteria that generalize D-, A-, and E-optimality by evaluating the nonzero eigenvalues of symmetric positive semidefinite matrices.
• They possess properties like concavity, orthogonal invariance, and positive homogeneity that facilitate efficient optimization in statistical design and network applications.
• The framework enables practical algorithms with rank-one updates and support-delimitation bounds, accelerating convergence in both experimental design and graph optimization.

Kiefer's measures, also known as the $\phi_p$-criteria or Kiefer–$\Phi_p$ optimality criteria, are a parametric family of spectral information functions central to optimal experimental design and network optimization. These criteria generalize and interpolate between classical $A$-, $D$-, and $E$-optimal objectives, providing a unified analytic framework for problems ranging from statistical design to spectral graph theory. The measures are formulated in terms of the nonzero spectrum (eigenvalues) of a symmetric positive semidefinite matrix—typically an information matrix or Laplacian—with properties such as concavity, orthogonal invariance, and positive homogeneity that are essential for tractable optimization and robust algorithmic deployment (Rosa et al., 22 Dec 2025, Pronzato, 2013).

1. Formal Definition of Kiefer's Measures

Let $M$ denote a symmetric positive semidefinite matrix of dimension $m$ (such as an information matrix in experimental design or a reduced Laplacian in graph optimization) with strictly positive eigenvalues $\{\lambda_j\}_{j=1}^r$. Kiefer's $\phi_p$-criterion is defined for $p \in (-1,\infty]$ as: $$\Phi_p(M) = \left( \frac{1}{r} \sum_{j=1}^r \lambda_j(M)^{-p} \right)^{-1/p}$$ The continuous extension covers singular $M$ (for $p\ge0$) by assigning $\Phi_p(M)=0$ if $M$ is singular. Key special cases:

• $p=0$ ($D$-optimality): $$\Phi_0(M) = \left( \prod_{j=1}^r \lambda_j(M) \right)^{1/r}$$ (geometric mean or determinant-based)
• $p=1$ ($A$-optimality): $\Phi_1(M) = r / \sum_j \lambda_j(M)^{-1}$ (harmonic mean or trace-inverse-based)
• $p \to \infty$ ($E$-optimality): $\Phi_\infty(M) = \min_j \lambda_j(M)$ (smallest eigenvalue)

In network design, the same analytic form is used with $M$ as the grounded Laplacian or its pseudoinverse (Rosa et al., 22 Dec 2025). The mapping $p \to \Phi_p(M)$ is nonincreasing for fixed $M$.

2. Core Properties: Concavity, Invariance, and Algorithmic Suitability

Kiefer's measures possess the following essential properties:

• Orthogonal Invariance: $\Phi_p(M)$ depends only on the spectrum of $M$.
• Isotonicity: For $M \preceq M'$, $\Phi_p(M) \le \Phi_p(M')$.
• Positive Homogeneity: $\Phi_p(\alpha M) = \alpha \Phi_p(M)$ for $\alpha>0$.
• Concavity on the Cone: $\Phi_p$ is concave on the positive semidefinite cone, facilitating both greedy and convex optimization.
• Upper Semicontinuity: Guarantees existence of maxima under compactness.

These features establish $\Phi_p$ as an information function in both statistical design and combinatorial network applications (Rosa et al., 22 Dec 2025).

3. Directional Derivatives, Node Dissimilarities, and Algorithmic Updates

The directional derivative of $\Phi_p$ at $M$ in the direction $\Delta$ is given by: $$\partial\Phi_p(M;\Delta) = \Phi_p(M)\left[ \frac{\operatorname{tr}\left( M^{-(p+1)} \Delta \right)}{\operatorname{tr}(M^{-p})}-1 \right]$$ For rank-one updates $\Delta = u u^\top$, this becomes: $$\partial\Phi_p(M; uu^\top) = \Phi_p(M) \frac{u^\top M^{-(p+1)} u}{\operatorname{tr}(M^{-p})} = (\Phi_p(M))^{p+1} v_p(u)$$ where $v_p(u) = u^\top M^{-(p+1)} u$ is called the node dissimilarity in the network optimization context. This enables computationally efficient (often $O(n^2)$ per iteration) greedy or exchange-based optimization methods via rank-one update formulas (Rosa et al., 22 Dec 2025).

4. Support Delimitation and Accelerated Algorithms

In approximate design, not every point $x$ in the design space $X$ can serve as a support point in an optimum design for $\phi_p$. Harman-Pronzato bounds and their generalizations (Pronzato, 2013) show that, for any current design $\xi$, any support point $x_*$ of a $\phi_p$-optimal design must satisfy: $F_{\phi_p}(\xi, x_*) \ge h_p[M(\xi), \delta]$ where $F_{\phi_p}(\xi,x)$ is the directional derivative at $\xi$ in the direction of $x$, $\delta=\max_{x\in X}F_{\phi_p}(\xi,x)$, and $h_p$ is a computable threshold depending on the current information matrix and polynomial root-finding in $p$ (simplified for integer $p$). Exclusion of $x$ with $F_{\phi_p}(\xi, x) < h_p$ causes significant computational speed-up by reducing the candidate support set as the algorithm converges (Pronzato, 2013).

5. Interpolation Between Classical Optimalities

Kiefer's measures interpolate between well-known design criteria: | Criterion | $p$ | $\Phi_p$ Expression | |------------------|----------|----------------------------------------------| | D-optimality | $p=0$ | Geometric mean of eigenvalues | | A-optimality | $p=1$ | Harmonic mean of eigenvalues | | E-optimality | $p\to\infty$ | Minimum eigenvalue |

Intermediate $p$ yields compromise criteria, suitable for applications requiring tradeoffs between robustness (as in E-optimality), information gain (D-optimality), and variance minimization (A-optimality). Theoretical monotonicity in $p$ follows from power-mean inequalities (Rosa et al., 22 Dec 2025, Pronzato, 2013).

6. Applications: Experimental Design and Laplacian Network Optimization

In statistical experimental design, $\phi_p$-criteria govern the selection of designs maximizing information content under linear models. The information matrix $M(\xi)$ of a design $\xi$ determines the expected estimation accuracy, and optimization under $\phi_p$ yields designs best suited for parameter inference with respect to the chosen criterion. Efficient support-delimitation bounds and directional derivative formulas accelerate practical algorithms for large discrete or continuous design spaces (Pronzato, 2013).

In network design, Kiefer’s measures are used to optimize spectral properties of the Laplacian matrix—boosting network robustness, information flow, and mixing times. Applications include resilient communication topologies, Markov chain mixing, and controlling effective resistance in electrical networks. The rank-one updates and node dissimilarities derived from $\Phi_p$ supply efficient primitives for greedy addition/exchange algorithms, with performance guarantees tied to the concavity of the criterion (Rosa et al., 22 Dec 2025).

7. Algorithmic Frameworks and Tradeoffs

Algorithms leveraging Kiefer's measures operate either via greedy addition (selecting the edge or design point maximizing the incremental derivative) or via edge/exchange heuristics that swap elements to improve the objective, guided by the node dissimilarity or support-delimitation lower bounds. Rank-one formulas allow per-iteration complexity that scales quadratically with problem size, and support-exclusion accelerates convergence as the optimal set is approached. The selection of $p$ determines the tradeoff between exploration (D-optimality), averaging (A-optimality), and minimax robustness (E-optimality), with intermediate values offering tailored blends for specific application constraints (Rosa et al., 22 Dec 2025, Pronzato, 2013).