From numerical proportions to analogical proportions between probabilities

Direct Answer

The paper introduces a formal framework for defining analogical proportions between probabilities—that is, a way to say “the probability of A is to the probability of B as the probability of C is to the probability of D.” It matters because it extends analogical reasoning, a cornerstone of human cognition, into the probabilistic domain, opening new pathways for classification, recommendation, and decision‑making systems that operate on uncertainty.

Background: Why This Problem Is Hard

Analogical reasoning has long been used to model “a is to b as c is to d” relationships in domains such as geometry, language, and vector‑based profiles. Traditional analogical proportions work well when the items are concrete numbers or fixed‑length vectors, because component‑wise operations preserve the underlying structure.

When the items become probability distributions, two major challenges arise:

• Normalization constraint: Probabilities must sum to one, so any transformation that breaks this property yields an invalid distribution.
• Non‑linear interaction: Simple arithmetic ratios that work for raw numbers often distort the shape of a distribution, especially when dealing with sparse or multimodal data.

Existing analogical classification methods assume that if four feature vectors satisfy a component‑wise proportion, their class labels will also respect the same proportion. Extending this intuition to probabilistic profiles has never been rigorously explored, leaving a gap in AI systems that need to reason about uncertainty in an analogical way.

What the Researchers Propose

Henri Prade and Gilles Richard propose a two‑pronged definition of analogical proportions for probabilities:

• Arithmetic proportion: Directly applying the classic “a : b = c : d” formula to probability values, while checking that the resulting fourth probability remains within the [0, 1] interval.
• Hybrid arithmetic‑geometric proportion: Combining the arithmetic ratio with a geometric scaling factor to better respect the multiplicative nature of probability spaces.

Both definitions are evaluated against a set of algebraic properties (symmetry, reflexivity, and preservation of normalization) to determine which formulation behaves most like a true analogical relation in the probabilistic realm.

How It Works in Practice

The practical workflow can be broken down into three stages:

1. Profile construction: Each entity (e.g., a user, a product, or a sensor reading) is represented by a discrete probability distribution over a set of possible outcomes.
2. Proportion testing: Given four profiles a, b, c, d, the system computes both the arithmetic and hybrid ratios. If the computed d matches the observed distribution of the fourth profile within a tolerance threshold, the four profiles are said to form an analogical proportion.
3. Application layer: The proportion relationship can be fed into downstream tasks such as analogical classification (predicting the class of d from the classes of a, b, c) or recommendation (inferring missing probabilities for a new user based on analogical peers).

What distinguishes this approach from earlier vector‑based analogies is the explicit handling of the probability simplex: every operation is constrained to keep the sum of probabilities equal to one, and the hybrid formulation introduces a multiplicative correction that aligns with the way likelihoods combine in Bayesian reasoning.

Evaluation & Results

The authors conducted experiments on three benchmark datasets that contain discrete attributes with known class labels:

• UCI Mushroom – binary edible/poisonous classification based on categorical features.
• Adult Income – predicting income brackets from demographic distributions.
• Custom synthetic distributions – designed to stress‑test normalization preservation.

For each dataset, they compared four setups:

1. Baseline k‑Nearest‑Neighbour (k‑NN) on raw probability vectors.
2. Analogical classification using pure arithmetic proportion.
3. Analogical classification using the hybrid arithmetic‑geometric proportion.
4. State‑of‑the‑art probabilistic analogical method from prior literature (used as a reference point).

Key findings include:

• The hybrid proportion consistently outperformed the pure arithmetic version, achieving up to a 7 % boost in classification accuracy on the Mushroom dataset.
• Both analogical methods preserved the probability simplex, whereas the baseline sometimes produced invalid probability vectors after interpolation.
• When the underlying distributions were highly skewed, the hybrid approach maintained stability, confirming its robustness to real‑world data imbalance.

These results demonstrate that analogical reasoning can be meaningfully transferred to probabilistic representations without sacrificing mathematical soundness.

Why This Matters for AI Systems and Agents

Analogical proportions between probabilities give AI engineers a new primitive for reasoning under uncertainty. Practical implications include:

• Enhanced classification pipelines: By leveraging analogical relations, models can infer class labels for sparse or partially observed profiles, reducing the need for large labeled datasets.
• Agent‑to‑agent knowledge transfer: In multi‑agent environments, an agent can communicate its belief state to peers using analogical proportions, enabling rapid alignment of expectations without transmitting full distributions.
• Decision‑support systems: Probabilistic analogies can suggest “what‑if” scenarios (e.g., “If the probability of event A rises as it did for user X, what will happen to event B for user Y?”) in a mathematically consistent way.

Enterprises looking to embed such reasoning into their workflows can start with the AI marketing agents offered on the UBOS platform, which already support probabilistic profiling and could be extended with analogical proportion modules.

What Comes Next

While the study establishes a solid theoretical foundation, several avenues remain open:

• Scalability: Extending the proportion calculations to high‑dimensional distributions (e.g., language models with thousands of tokens) will require efficient approximation techniques.
• Continuous domains: The current work focuses on discrete probability vectors; adapting the framework to continuous densities (Gaussian mixtures, Dirichlet processes) is an exciting challenge.
• Integration with deep learning: Embedding analogical proportion loss functions into neural networks could enable end‑to‑end training of models that respect analogical constraints.
• Real‑time orchestration: Deploying these calculations in streaming environments (e.g., IoT sensor networks) will test the latency and robustness of the hybrid formulation.

Developers interested in prototyping these ideas can explore the Enterprise AI platform by UBOS, which provides the necessary compute, data pipelines, and integration points (such as Chroma DB or OpenAI ChatGPT) to experiment with probabilistic analogies at scale.

References

Prade, H., & Richard, G. (2026). From numerical proportions to analogical proportions between probabilities. arXiv preprint arXiv:2606.23029.