Generalizing Analogical Inference from Boolean to Continuous Domains

Conference: AAAI 2026 arXiv: 2511.10416
Code: None
Area: AI Foundations / Analogical Reasoning Keywords: analogical reasoning, Boolean domain generalization, continuous domain regression, generalized mean, error bounds

TL;DR

This paper revisits the theoretical foundations of analogical inference: it first constructs a counterexample demonstrating the failure of classical generalization bounds in the Boolean domain, then proposes a unified analogical inference framework based on parameterized generalized means, extending discrete classification to continuous regression domains.

Background & Motivation

Background: Analogical reasoning (four-term relations of the form \(a:b::c:d\)) is a fundamental mechanism of human cognition, with broad applications in few-shot learning, transfer learning, morphological analysis, and word vector evaluation. Theoretical foundations in the Boolean domain were established by Couceiro et al. (2017, 2018): analogical inference is exact for affine functions and admits a \(4\epsilon\) probabilistic error bound for near-affine functions.

Limitations of Prior Work: - Existing theory is restricted to discrete attribute spaces and binary classification, and cannot handle regression tasks or continuous domains. - More critically, even within the Boolean domain, the classical generalization bound itself is flawed.

Key Challenge: Analogical reasoning is widely applied in practice over continuous domains (e.g., word vector analogies), yet theoretical guarantees are entirely absent; moreover, the core theorem (\(4\epsilon\) bound, Theorem 3) underpinning discrete domain theory is incorrect.

Goal: To correct the erroneous theoretical bound in the Boolean domain and extend analogical inference to continuous domains, providing theoretical guarantees for regression-based analogical reasoning.

Key Insight: A parameterized analogical proportion is defined via Hölder generalized means, constructing a unified framework that encompasses both Boolean classification and continuous regression.

Core Idea: The generalized mean parameter \(p\) defines the analogical proportion \(a:b::^p c:d\) over continuous domains, characterizes the class of functions preserving analogical structure, and derives worst-case and average-case error bounds under smoothness assumptions.

Method

Overall Architecture

The theoretical framework proceeds in three progressive layers:

1. Counterexample Construction: Refuting the classical Boolean domain generalization bound.
2. Continuous Domain Analogy Framework: Parameterized analogy definition based on generalized means.
3. Error Analysis: Characterizing analogy-preserving functions and deriving error bounds.

Key Designs

1. Counterexample to the Classical Bound (Section 3):

   • A function \(f: \mathbb{B}^4 \to \mathbb{B}\) is constructed that outputs 1 only at \(\mathbf{x} = \mathbf{1}\) and 0 elsewhere.
   • The distance from \(f\) to the affine class \(\mathcal{L}\) is \(d(f, \mathcal{L}) = 1/16\).
   • An exhaustive algorithm (enumerating \(2^{15}\) subsets) yields \(P(\text{err}_{S,f} > 0) \geq 0.42\).
   • However, the classical Theorem 3 predicts an upper bound of \(4 \times 1/16 = 0.25\), yielding a contradiction.
   • This constitutes an important theoretical correction, refuting a core theorem that has been widely cited in the field for a decade.
   • Intuition behind the counterexample: when a training set of all-zero instances meets an all-one test point, the analogical inference system systematically predicts the unique label 1 as 0.

2. Parameterized Analogy via Generalized Means (Section 4):

   • Generalized mean definition: \(m_p(x_1, ..., x_n) = \lim_{r \to p} (\frac{1}{n}\sum x_i^r)^{1/r}\)
   • \(p = 1\) corresponds to the arithmetic mean, \(p = 0\) to the geometric mean, and \(p = -1\) to the harmonic mean.
   • Analogy definition: \((a,b,c,d) \in \mathbb{R}_+^4\) satisfies \(a:b::^p c:d\) if and only if \(m_p(a,d) = m_p(b,c)\).
   • Key properties: for any four strictly increasing positive reals, there exists a unique analogical exponent \(p\); any such analogy can be reduced to an equivalent arithmetic analogy; solutions always exist for increasing sequences.
   • The notions of analogical root and analogical extension are generalized from the Boolean domain to the continuous domain, governed by the parameter pair \((\mathbf{p}; q)\) controlling the analogical exponents for attribute and label domains respectively.

3. Characterization of Analogy-Preserving Functions (Sections 4.3 and 5):

   • Core theorem (Proposition 9): A continuous function \(f\) belongs to \(AP_{(\mathbf{p};q)}\) if and only if \(f\) maps analogies of exponent \(\mathbf{p}\) to analogies of exponent \(q\).
   • When \(p = q = 1\) (arithmetic analogy), the analogy-preserving functions are exactly affine functions, recovering the classical Boolean domain result.
   • In the general case, analogy-preserving functions form a family of generalized power functions with rich structural properties.
   • This provides a function-theoretic foundation for analogical inference in regression settings.

4. Error Bounds in Continuous Domains (Section 5):

   • An appropriate function distance metric suited to generalized analogies is introduced.
   • Under smoothness assumptions, the following are derived:
   • Worst-case bound (uniform bound): For functions that are \(\epsilon\)-close to the analogy-preserving function class, the maximum error of analogical inference is bounded.
   • Average-case bound (probabilistic bound): Under random selection of training sets, the expected inference error is bounded.
   • These bounds provide PAC-learning-style theoretical guarantees for analogical inference in continuous domains.

Loss & Training

This paper is a purely theoretical work and does not involve training. The core contributions are theorems and their proofs.

Key Experimental Results

Main Results

The paper is a theoretical contribution; the central "experiment" is counterexample verification:

Theoretical Quantity	Value
\(d(f, \mathcal{L})\)	1/16 = 0.0625
Classical theorem predicted upper bound \(4\epsilon\)	0.25
Exhaustively computed actual lower bound	≥ 0.42
Violation gap	0.42 > 0.25 (contradiction)

The exhaustive algorithm completes in approximately 30 seconds for \(n=4\) (\(2^{15}\) subsets).

Ablation Study

Not applicable.

Key Findings

• Theorem 3 of Couceiro et al. (2018) (\(P(\text{err}_{S,f} > \delta) \leq 4\epsilon(1-\delta)\)) fails even at \(\delta = 0\).
• The generalized mean parameter \(p\) provides a natural "knob" that unifies different types of analogies (arithmetic, geometric, harmonic, etc.).
• Analogy-preserving functions in the continuous domain are generalized power functions, structurally richer than the affine functions of the Boolean domain.

Highlights & Insights

• Refuting a theoretical result that has been widely cited for a decade demands both intellectual courage and rigor. The counterexample, while elementary (a 4-dimensional Boolean function), points to a fundamental flaw in the theoretical framework.
• The parameterization scheme based on generalized means is particularly elegant: a single parameter \(p\) subsumes the classical analogical concepts of arithmetic, geometric, and harmonic means.
• Connecting analogical reasoning to regression tasks opens new territory for a theoretical field long confined to classification.

Limitations & Future Work

• The framework considers only the positive real domain \(\mathbb{R}_+\) and does not directly extend to the general real line including negative values (though many applications such as image processing are naturally non-negative).
• The new error bounds have not yet been connected to practical analogical classification or regression algorithms, and empirical validation is lacking.
• While the counterexample identifies the failure of the old bound, it does not provide a corrected and improved bound for the Boolean domain.
• The selection of the analogical exponent \(p\) may pose challenges in practice — how should the optimal \(p\) be determined for a given dataset?

Related Work & Insights

• The Boolean domain analogical inference theory of Couceiro et al. (2017, 2018) serves as the direct point of departure for this work.
• The parameterized analogy based on generalized means proposed by Lepage (2024) is the inspirational source for the continuous domain framework.
• The word2vec analogy task of Mikolov et al. (2013) (king:queen::man:woman) represents a canonical application scenario for continuous domain analogy.
• Insight: The theoretical foundations of analogical reasoning may require comprehensive reconstruction, particularly in high-dimensional continuous spaces.

Rating

• Novelty: ⭐⭐⭐⭐⭐ Refutation of a classical theorem combined with an entirely new continuous domain framework constitutes a strong theoretical breakthrough.
• Experimental Thoroughness: ⭐⭐⭐ Purely theoretical work; the counterexample verification is rigorous, but empirical data are absent.
• Writing Quality: ⭐⭐⭐⭐⭐ Mathematical derivations are rigorous, the logical flow is clear, and the narrative arc from counterexample to generalization is well-paced.
• Value: ⭐⭐⭐⭐ Provides important corrections and advances to the theory of analogical reasoning.

Related Papers

• [ICML 2026] Correcting Split Selection in Online Decision Trees via Anytime-Valid Inference
• [NeurIPS 2025] How Many Domains Suffice for Domain Generalization? A Tight Characterization via the Domain Shattering Dimension
• [AAAI 2026] Streaming Generated Gaussian Process Experts for Online Learning and Control: Extended Version
• [AAAI 2026] A Switching Framework for Online Interval Scheduling with Predictions
• [ICML 2026] MMD-Balls as Credal Sets: A PAC-Bayesian Framework for Epistemic Uncertainty in Test-Time Adaptation