Aristotelian Representation Hypothesis (ARH)

• ARH is a conceptual framework that integrates neural representation learning with formal Aristotelian logic to capture local relational structures.
• In neural representation learning, ARH demonstrates that calibrated local metrics, unlike global ones, reliably reveal shared neighborhood relationships across models.
• In mathematical domains, ARH rigorously represents Aristotelian diagrams, continuum, and topos via Boolean algebras and intuitionistic fuzzy systems.

The Aristotelian Representation Hypothesis (ARH) is a label used in contemporary research for the claim that Aristotelian structures can be represented faithfully within precise formal systems, although the term is not used uniformly across fields. In neural representation learning, ARH is explicitly defined as the hypothesis that “Neural networks, trained with different objectives on different data and modalities, are converging to shared local neighborhood relationships” (Gröger et al., 16 Feb 2026). In adjacent work on logic and philosophy of mathematics, a closely related representational program concerns the rigorous mathematical realization of Aristotelian diagrams, the square of opposition, and Aristotelian concepts such as continuum and topos within Boolean algebras, intuitionistic fuzzy Boolean algebras, locales, and topoi (Syropoulos, 2024, Protin, 2021). These usages are not identical, but they share an emphasis on structural representation, relational organization, and the formalization of distinctions originally articulated in Aristotelian terms.

1. Principal senses of the term

Current usage separates into a machine-learning sense and a broader philosophical-mathematical sense. The former is explicitly named and contrasted with the Platonic Representation Hypothesis. The latter is sometimes presented without the acronym itself, but it advances the feasibility and fruitfulness of interpreting Aristotelian concepts within modern mathematics (Gröger et al., 16 Feb 2026, Protin, 2021).

Context	Core claim	Formal setting
Neural representation learning	Convergence to shared local neighborhood relationships	Representational similarity metrics and null calibration
Aristotelian logic and diagrams	Classical Aristotelian logic can be faithfully represented in precise mathematical structures	Boolean algebra, intuitionistic fuzzy Boolean algebra, fuzzy category
Aristotelian concepts in mathematics	Continuum, topos, and related notions admit rigorous modern interpretation	Topology, geometry, category theory, locale theory, topos theory

In the machine-learning literature, ARH is a revision of a stronger representational thesis. In the mathematical literature, the relevant question is whether Aristotelian forms can be rendered as exact objects of contemporary formal theory. This suggests a shared concern with representational adequacy, but the objects represented differ substantially: neural embeddings in one case, logical and conceptual structures in the other.

2. ARH in neural representation learning

The immediate background to the machine-learning formulation is the Platonic Representation Hypothesis (PRH), which suggests that representations from neural networks are converging to a common statistical model of reality. The Aristotelian reformulation is proposed after a critical reanalysis of representational similarity metrics showing that existing measures are confounded by network scale (Gröger et al., 16 Feb 2026).

Two confounders are central. The width confounder arises because, for two random, independent representations $X \in \mathbb{R}^{n \times d_x}$ and $Y \in \mathbb{R}^{n \times d_y}$, the expected squared Frobenius norm of the sample cross-covariance under independence is

$$\mathbb{E}_{H_0}[\|\widetilde{C}\|_F^2] = \frac{d_x d_y}{n-1},$$

so the null expectation scales with representation dimension rather than vanishing (Gröger et al., 16 Feb 2026). The depth confounder comes from comparing many layer pairs and then summarizing by a maximum or top-$k$ statistic; under the null,

$$\mathbb{E}_{H_0}[\max S_{\ell,\ell'}] \leq \mu + C \sigma \sqrt{\log M},$$

so increasing the number of layer pairs inflates apparent similarity through a look-elsewhere effect (Gröger et al., 16 Feb 2026).

This distinction yields a methodological split between global metrics and local metrics. Global metrics such as CKA and CCA are reported as severely affected by width and depth confounding, whereas local neighborhood metrics such as mutual $k$-Nearest Neighbors (mKNN) are much less sensitive. For mKNN, the null baseline under independence is

$$\mathbb{E}_{H_0}[\mathrm{mKNN}(X, Y)] = \frac{k}{n-1},$$

with $k \ll d$ typically (Gröger et al., 16 Feb 2026).

On this basis, ARH rejects the claim that global geometry or global statistical structure demonstrably converges across large models after proper calibration. It instead maintains that the empirically robust convergence lies in the relational question “who is near whom?” in representation space (Gröger et al., 16 Feb 2026).

3. Calibration, evidence, and the scope of the neural ARH

The technical mechanism used to support the neural ARH is a permutation-based null-calibration framework. Its stated purpose is to transform any representational similarity metric into a calibrated score with statistical guarantees. The procedure consists of four steps: compute the observed similarity score; generate null scores by permuting sample correspondence in one representation; calculate an empirical right-tail $p$-value and determine a null critical value $\tau_\alpha$; and form a calibrated score for bounded metrics. For aggregate scores such as max-over-layers summaries, calibration is performed on the aggregate itself, which the paper terms aggregation-aware calibration (Gröger et al., 16 Feb 2026).

The paper attributes several formal properties to this framework: finite-sample guarantees, type I error controlled at chosen $Y \in \mathbb{R}^{n \times d_y}$0, and applicability to arbitrary bounded similarity statistics without closed-form assumptions (Gröger et al., 16 Feb 2026). Empirically, the reported result is a “nuanced picture.” When the framework is applied to synthetic experiments and to image/text and video/text models, the increase in similarity previously reported by global spectral measures largely disappears after calibration, while local neighborhood similarity remains significant across different modalities (Gröger et al., 16 Feb 2026).

The paper further narrows the content of convergence. The retained agreement is not reported for fine-grained local distances; rather, it persists at the level of ordinal neighborhood relationships. In this formulation, ARH is a relational thesis rather than a metric one: it concerns agreement about neighborhood structure, not exact global geometry and not exact local distances (Gröger et al., 16 Feb 2026).

A common misreading is therefore to treat ARH as a weaker rhetorical restatement of PRH. The paper’s actual claim is sharper: after calibration, the global-convergence claim is not supported, while the local-neighborhood claim is (Gröger et al., 16 Feb 2026).

4. Aristotelian diagrams as formal objects

A different but related use of the representational idea appears in work on Aristotelian diagrams. The point of departure is the traditional square of opposition, which visually represents relations among four categorical propositions: A (universal affirmative), E (universal negative), I (particular affirmative), and O (particular negative). The classical relations are contradictory, contrary, subcontrary, and subalternation (Syropoulos, 2024).

The paper “Fuzzy Aristotelian Diagrams” gives an explicit algebraic formalization. A Boolean algebra is specified by a set $Y \in \mathbb{R}^{n \times d_y}$1, binary operations $Y \in \mathbb{R}^{n \times d_y}$2 and $Y \in \mathbb{R}^{n \times d_y}$3, a unary operation $Y \in \mathbb{R}^{n \times d_y}$4, constants $Y \in \mathbb{R}^{n \times d_y}$5 and $Y \in \mathbb{R}^{n \times d_y}$6, and the usual idempotent, commutative, associative, distributive, absorption, and complementation laws. Within that setting, an Aristotelian diagram $Y \in \mathbb{R}^{n \times d_y}$7 is defined as a pair $Y \in \mathbb{R}^{n \times d_y}$8 where $Y \in \mathbb{R}^{n \times d_y}$9 is a Boolean algebra and $$\mathbb{E}_{H_0}[\|\widetilde{C}\|_F^2] = \frac{d_x d_y}{n-1},$$0 is a set of designated elements standing for logical forms (Syropoulos, 2024).

The classical logical relations are then expressed algebraically. For $$\mathbb{E}_{H_0}[\|\widetilde{C}\|_F^2] = \frac{d_x d_y}{n-1},$$1:

• Contradictory: $$\mathbb{E}_{H_0}[\|\widetilde{C}\|_F^2] = \frac{d_x d_y}{n-1},$$2 and $$\mathbb{E}_{H_0}[\|\widetilde{C}\|_F^2] = \frac{d_x d_y}{n-1},$$3, equivalently $$\mathbb{E}_{H_0}[\|\widetilde{C}\|_F^2] = \frac{d_x d_y}{n-1},$$4.
• Contrary: $$\mathbb{E}_{H_0}[\|\widetilde{C}\|_F^2] = \frac{d_x d_y}{n-1},$$5 and $$\mathbb{E}_{H_0}[\|\widetilde{C}\|_F^2] = \frac{d_x d_y}{n-1},$$6, equivalently $$\mathbb{E}_{H_0}[\|\widetilde{C}\|_F^2] = \frac{d_x d_y}{n-1},$$7.
• Subcontrary: $$\mathbb{E}_{H_0}[\|\widetilde{C}\|_F^2] = \frac{d_x d_y}{n-1},$$8 and $$\mathbb{E}_{H_0}[\|\widetilde{C}\|_F^2] = \frac{d_x d_y}{n-1},$$9, equivalently $k$0.
• Bi-implication: $k$1.
• Subalternation: described as left or right implication under the order of the algebra (Syropoulos, 2024).

The significance of this formalization for ARH is explicit in the paper’s discussion: classical Aristotelian logic, and especially the square of opposition, can be treated as a precise mathematical object rather than only as a historical or pedagogical diagram. That is the form of representational faithfulness at stake in this line of work.

5. Fuzzy generalization and categorical universes

The same paper extends the representational program from bivalent logic to vagueness. Its stated motivation is that classical Aristotelian diagrams are crisp and bivalent, whereas real-world and natural language reasoning is often vague. The proposed mathematical vehicle is intuitionistic fuzzy logic (Syropoulos, 2024).

A fuzzy set on a set $k$2 is given by a membership function $k$3. An intuitionistic fuzzy set is a pair $k$4 satisfying

$k$5

where $k$6 is the degree of membership and $k$7 the degree of non-membership (Syropoulos, 2024). To host fuzzy Aristotelian diagrams, the supporting Boolean algebra is replaced by an intuitionistic fuzzy Boolean algebra whose underlying set $k$8 is equipped with an intuitionistic fuzzy partial order $k$9, described as reflexive, perfectly antisymmetric, and transitive. The relation is written as

$$\mathbb{E}_{H_0}[\max S_{\ell,\ell'}] \leq \mu + C \sigma \sqrt{\log M},$$0

with the same sum condition on membership and non-membership degrees (Syropoulos, 2024).

Within this setting, a fuzzy Aristotelian diagram is again a pair $$\mathbb{E}_{H_0}[\max S_{\ell,\ell'}] \leq \mu + C \sigma \sqrt{\log M},$$1, now with $$\mathbb{E}_{H_0}[\max S_{\ell,\ell'}] \leq \mu + C \sigma \sqrt{\log M},$$2 an intuitionistic fuzzy Boolean algebra and $$\mathbb{E}_{H_0}[\max S_{\ell,\ell'}] \leq \mu + C \sigma \sqrt{\log M},$$3. The paper characterizes this as a graded generalization in which logical forms and their relations can be assigned degrees rather than strict truth values. It also introduces fuzzy categories and intuitionistic fuzzy categories, where morphisms carry plausibility and non-plausibility degrees; a morphism is denoted

$$\mathbb{E}_{H_0}[\max S_{\ell,\ell'}] \leq \mu + C \sigma \sqrt{\log M},$$4

with $$\mathbb{E}_{H_0}[\max S_{\ell,\ell'}] \leq \mu + C \sigma \sqrt{\log M},$$5 and $$\mathbb{E}_{H_0}[\max S_{\ell,\ell'}] \leq \mu + C \sigma \sqrt{\log M},$$6 (Syropoulos, 2024).

In this framework, the “universe” of fuzzy Aristotelian diagrams is categorical: objects are fuzzy Aristotelian diagrams $$\mathbb{E}_{H_0}[\max S_{\ell,\ell'}] \leq \mu + C \sigma \sqrt{\log M},$$7 and morphisms are functions between diagrams equipped with plausibility and non-plausibility degrees and constrained to respect fuzzy logical relations (Syropoulos, 2024). The paper presents this as a rigorous, graded generalization of ARH, extending representational faithfulness from crisp logical forms to contexts involving vagueness.

6. Continuum, topos, and adjacent Aristotelian representation programs

A broader philosophical-mathematical version of the representational idea appears in work on continuum (sunekhês), place (topos), infinity (apeiron), and related notions. In this literature, the central claim is not that a named ARH has already been fully codified, but that Aristotelian concepts can be given rigorous interpretations in topology, geometry, and category theory, and that such interpretation is both feasible and fruitful (Protin, 2021).

For the continuum, Aristotle’s thesis that it “does not consist of points” is mapped against modern formalisms. Point-set topology can model connectivity, but the paper argues that it does not fully satisfy Aristotle’s demand that the continuum is not constructed from points. Point-free topology and topos theory represent a more faithful modern realization (Protin, 2021). Locale theory is used to formulate connectedness and decomposability in lattice-theoretic terms, including expressions such as

$$\mathbb{E}_{H_0}[\max S_{\ell,\ell'}] \leq \mu + C \sigma \sqrt{\log M},$$8

for global connectivity, and

$$\mathbb{E}_{H_0}[\max S_{\ell,\ell'}] \leq \mu + C \sigma \sqrt{\log M},$$9

for local connectivity (Protin, 2021). In topos theory, the relevant parallel is the locally connected or “molecular” topos, written as

$k$0

where every object is a sum of connected or molecular objects (Protin, 2021).

For topos as place, the paper uses sheaf theory and algebraic geometry. Given a closed set $k$1, the “qualities at” $k$2 are represented by the inverse limit

$k$3

taken over open neighborhoods $k$4 containing $k$5 (Protin, 2021). For $k$6 a point, this is the stalk of the sheaf at that point. The paper presents this as an exact match to the Aristotelian notion of place.

The paper is explicit about interpretive limits. The internal logic of a topos shares significant, though not full, alignment with Aristotle’s local and qualitative logic; the claimed relationship is a structural parallel—not identity (Protin, 2021). That caveat is important for the encyclopedia treatment of ARH: the hypothesis, in this mathematical sense, is about rigorous representation of structure, not about reducing Aristotle’s concepts without remainder to present-day formal systems.

A nearby but distinct research program appears in computational argumentation. The Trichotomic Argument Interchange Format (T-AIF) realizes Aristotle’s trichotomy of Logos, Ethos, and Pathos within a weighted, multi-relational graph. Logos is represented through the AIF/AIF+ tradition of I-nodes, S-nodes, and attack/support structure; Ethos through weighted trust edges between E-nodes; and Pathos through weighted commitment edges from E-nodes to I-nodes. The underlying abstraction, the T-AF, uses fuzzy predicates for scheme interpretation and actor-specific belief functions $k$7, and the semantics are stated in $k$8 fuzzy logic (Göttlinger et al., 2018). This work is not presented as ARH, but it belongs to the same general effort to formalize Aristotelian distinctions as operational representational structures.

Across these literatures, the term ARH therefore denotes either a specific empirical hypothesis about neural representation or a broader formal aspiration to render Aristotelian logical and conceptual structures mathematically exact. What unifies them is not a single theorem but a common representational strategy: local relations, structural organization, and formal comparability are treated as the primary bearers of Aristotelian content.