---
title: Intensified Morris-Mitchell Criterion
url: https://www.emergentmind.com/topics/intensified-morris-mitchell-criterion
type: topic
---

# Intensified Morris-Mitchell Criterion

The intensified Morris–Mitchell criterion is a size-invariant generalization of the original Morris–Mitchell space-filling metric, designed to provide a standardized and scalable means of quantifying design coverage in multi-dimensional input spaces. By introducing a normalization over the number of pairwise distances, this criterion enables direct and fair comparison of experimental designs irrespective of the number of samples, a feature indispensable in adaptive or incremental optimization workflows. It preserves the lexicographic max–min emphasis of the original metric while ensuring its scale remains stable as the design size changes, thus serving as a robust tool for monitoring and guiding space-filling improvements in sequential and batch design construction [2512.21989].

## 1. Mathematical Formulation and Properties

The classical Morris–Mitchell criterion for a design $X = \{x_1,…,x_n\} \subset \mathbb{R}^k$ is defined as
\[
\Phi_q(X) = \left( \sum_{i=1}^{m} J_i\, d_i^{-q} \right)^{1/q}
\]
where $d_i$ are the distinct pairwise Minkowski distances $\|x_j - x_k\|_p$ sorted in ascending order, $J_i$ is the multiplicity of each distance $d_i$, and $q > 0$ is an exponent that intensifies the max–min emphasis as $q \to \infty$.

The intensified (size-invariant) Morris–Mitchell criterion is:
\[
\Phi^*_q(X) = \left( \frac{1}{M} \sum_{i=1}^{m} J_i\, d_i^{-q} \right)^{1/q}
\]
with $M = n(n-1)/2$ being the number of unordered pairs. Equivalently,
\[
\Phi^*_q(X) = \frac{\Phi_q(X)}{M^{1/q}}
\]
This intensive normalization yields a criterion that remains on a consistent scale, regardless of the number of design points.

| Formula           | Description                                   | Scale Dependence      |
|-------------------|-----------------------------------------------|-----------------------|
| $\Phi_q(X)$       | Standard Morris–Mitchell criterion            | Grows with $n$        |
| $\Phi^*_q(X)$     | Intensive, size-invariant criterion           | Stable across $n$     |

## 2. Algorithmic Implementation

Batch and sequential calculation of $\Phi^*_q(X)$ are straightforward. The batch pseudocode involves:

1. Compute all pairwise distances and sort them.
2. Group unique distances into $d[1…m]$ each with multiplicity $J[i]$.
3. Calculate $\Phi_q(X)$ by summing $J_i d_i^{-q}$.
4. Normalize by $M$, yielding $\Phi^*_q(X)$.

Efficient sequential updates allow rapid computation when adding points:
- Compute new distances from $x_\mathrm{new}$ to all existing points.
- Update multiplicities $J$ and possibly create new distinct distances.
- Increment $M$ by $n$.
- Update the summed value and recompute $\Phi_q(X)$ and $\Phi^*_q(X)$.

The Python package **spotoptim** implements both batch and efficient update routines:
```python
import numpy as np
from spotoptim import mmphi_intensive, mmphi_intensive_update
X = np.array([[0.0, 0.0], [1.0, 0.0], [0.0, 1.0]])
phi_int, J, d = mmphi_intensive(X, q=2, p=2)
new_point = np.array([0.5,0.5])
phi_new, J_new, d_new = mmphi_intensive_update(X, new_point, J, d, q=2, p=2)
```
As demonstrated, $\Phi^*_2(X) \simeq 0.9129$ for three points, and remains stable in scale upon addition [2512.21989].

## 3. Space-Filling Properties and Theoretical Context

$\Phi^*_q(X)$ preserves all max–min and anti-clustering characteristics of the original Morris–Mitchell criterion, with the critical distinction that its scale is insensitive to the size of $X$. Lower values of $\Phi^*_q(X)$ correspond to more uniform coverage and reduced clustering. As $q\rightarrow\infty$, the criterion is governed by the minimum pairwise distance, then the number of instances at that distance, and so forth. The monotonicity between $\Phi_q(X)$ and $\Phi^*_q(X)$ ensures that the ranking of designs is unchanged for fixed size, but comparability and interpretability across varying sizes are significantly improved.

Empirical studies confirm that for Latin Hypercube Sampling (LHS) in two dimensions, standard $\Phi_q(X)$ increases linearly with $n$, while $\Phi^*_q(X)$ remains tightly bounded—facilitating longitudinal monitoring and cross-comparison of evolving designs.

## 4. Applications in Multi-Objective Optimization

The intensified Morris–Mitchell criterion is directly employed in multi-objective optimization scenarios, where industry experimental designs often lack representative coverage of input space. By integrating $\Phi^*_q(X)$ into multi-objective desirability functions, practitioners can jointly optimize experimental outcome predictions (from surrogate models such as Random Forest) and the space-filling quality of the design. The **spotdesirability** package leverages this approach, combining model-based desirability with the intensified Morris–Mitchell improvement metric into a unified score for candidate infill points.

Case study application in compressor development demonstrates optimization of performance objectives alongside $\Phi^*_q(X)$, producing designs with improved input coverage and balancing exploration–exploitation trade-offs [2512.21989].

## 5. Practical Implementation and Numerical Behavior

A numerical example illuminates the stability and comparability of $\Phi^*_q(X)$:
For $X = \{(0,0), (1,0), (0,1)\}$ in $\mathbb{R}^2$, $q=2$:
- Pairwise distances: $1, 1, \sqrt{2}$
- $\Phi_2(X) = \sqrt{2 + 0.5} \simeq 1.5811$
- $M=3 \implies \Phi^*_2(X) = 1.5811 / \sqrt{3} \simeq 0.9129$

Adding $(0.5, 0.5)$, $\Phi^*_2$ shifts in a manner that reflects the spatial configuration but does not change scale drastically, allowing comparability across sizes.

## 6. Interpretation and Significance

The key innovation of the intensified Morris–Mitchell criterion is its statistical intensiveness, achieved by normalizing the $L_q$-norm over unordered pairs, thus enabling operational use in adaptive sampling designs and real-time monitoring of space-filling improvements. This property is particularly relevant in scenarios where the number of design points is not fixed a priori or is expected to evolve, as in sequential optimization, surrogate-based experimental design, and active learning frameworks.

The robust, monotonic linkage between standard and intensive metrics ensures backward compatibility in ranking for fixed design sizes. A plausible implication is that size-invariance supports meta-optimization routines where batch and incremental improvements must be assessed on a unified scale, which is a recurring challenge in practical multi-objective design optimization.

## 7. Related Criteria and Generalization

While the intensified Morris–Mitchell criterion is specialized for space-filling assessment, its statistical normalization principle is generically applicable to other coverage metrics that depend on exhaustive pairwise or set-wise enumeration. The methodology outlined in Bartz-Beielstein (2025b) extends readily to other Minkowski norms, various $q$ exponents, and arbitrary input dimensionality $k$. Its implementation within spotoptim facilitates efficient usage even for large design sets, with computational complexity of $O(n^2)$ in batch mode and amortized $O(n)$ for sequential updates.

The widespread utility and straightforward implementation of the intensified criterion suggest it will remain central to modern design of experiments and surrogate-based optimization methodologies [2512.21989].

Source: https://www.emergentmind.com/topics/intensified-morris-mitchell-criterion