When we say that two things are similar, what exactly are we comparing? Color, shape, size, function, purpose, origin, and countless other characteristics are possible criteria. The pair that seems most similar depends on which characteristics matter. Behind this familiar observation lies a fundamental problem in logic and classification.
Satosi Watanabe’s Challenge
Physicist and information scientist Satosi Watanabe developed the Ugly Duckling Theorem, named after Hans Christian Andersen’s story, to illuminate a basic limitation of similarity-based classification. We may confidently recognize that the supposedly ugly duckling resembles a swan, but that recognition already relies on a decision about which properties are significant.
The theorem shows, under a formal setup that treats every possible predicate equally, that any two distinct objects share the same number of predicates. If every conceivable property receives equal weight, merely counting shared properties cannot tell us which objects are more similar.
A Simple Example with Four Objects
Imagine four distinct objects A, B, C, and D. A predicate can be represented by the set of objects that satisfy it: for example, a set containing A and B, one containing only B, or one containing none of the objects. There are 16 possible subsets of four objects.
Across all 16 subsets, A and B are both included in four subsets and both excluded in another four. So they receive the same yes/no result for eight predicates. The same count holds for A and C, B and D, and every other pair of distinct objects. If only shared positive membership is counted, each pair still has the same count of four.
Real classifications do not treat all 16 possible subsets as equally meaningful. Biological taxonomy gives weight to characteristics and evidence relevant to biological relationships. Retail merchandising may instead prioritize purpose, price, or purchasing behavior. Such choices make meaningful classifications possible.
Classification Requires a Point of View
A yellow ball and a yellow chick resemble each other in color. A chick and an adult chicken are more closely related biologically. Which pair is more similar depends on the question being asked.
This does not establish that all classifications are arbitrary or equally good. Feature choices can be assessed against a research question, predictive performance, independent evidence, and domain knowledge. The theorem identifies a formal limitation: similarity cannot be extracted from an indiscriminate inventory of all possible properties without selecting what matters.
What This Means for Data Analysis and AI
In multivariate analysis, similarity and clustering depend on variables, scaling, weighting, and the selected distance measure. In a two-axis chart such as the BCG growth-share matrix, the choice of axes strongly influences how cases appear. Using two axes is not inherently wrong, but it is worth asking what important dimensions might be missing.
Machine learning makes similar choices through feature representations, training objectives, datasets, and similarity metrics. An automatically generated clustering or embedding map is not a view of the world’s one uniquely natural classification. Its distances are meaningful in relation to the model and the task for which they were constructed.
When examining a classification, ask: Which properties made these items close? Are those properties appropriate for our purpose? Would another choice of features change the result?
Removing a Hidden Assumption
The value of the Ugly Duckling Theorem extends beyond its mathematical result. It challenges the often-unnoticed assumption that similarity is a single fixed property waiting to be discovered. Categories and clusters can be very useful, but making them always involves decisions about what counts as the same or different.
Recognizing those decisions provides a common perspective for understanding concept formation, qualitative analysis, crosscharts, multivariate methods, and machine learning.
Related Pages
- Concept Formation and Categorization
- Foundations of Logic
- Analogy and Concept Exploration
- Information Organization and Multivariate Analysis