Recognition Composition Law

• Recognition Composition Law is a set of functional identities that uniquely define how parts combine to yield symmetrical, closed, and optimal invariant aggregates.
• It is applied in representation learning to guide part decomposition, prototype encoding, and reconstruction, thereby enhancing interpretability and transferability.
• In fields such as statistical mechanics and formal algebra, the law underpins canonical cost functions, composition of power series, and invariance properties across diverse formulations.

The Recognition Composition Law encompasses a set of algebraic and functional identities that specify how compositional structure, symmetry, or information can be aggregated or combined across domains as diverse as representation learning, statistical mechanics, formal power series, relaxation processes, and algebraic form composition. In each context, the law formalizes a canonical, often uniquely determined, rule for composing objects—parts, ratios, costs, entropies, functions, or forms—so that resultant invariants (e.g., recognition codes, free energies, or polynomial discriminants) exhibit specified symmetries, closure properties, or optimality. The unifying attribute is that the Recognition Composition Law enforces a closure or compatibility structure that enables interpretable, transferable, and robust aggregation, either at the combinatorial, functional, or geometric level.

1. Functional Definition and Canonical Forms

The core of the Recognition Composition Law is a functional equation governing a scalar penalty, cost, or divergence function $J(x)$ on $(0, \infty)$, which is required to satisfy:

$$J(xy) + J\left(\frac{x}{y}\right) = 2 J(x) + 2 J(y) + 2 J(x) J(y)$$

with the calibration condition $J''(1) = 1$. This condition is referred to as the Recognition Composition Law (RCL) or canonical closure law. Its unique, continuous, nonconstant, positive solution (under normalization and reciprocity) is

$$J(x) = \frac{1}{2} (x + x^{-1}) - 1 = \cosh(\log x) - 1$$

This canonical form arises in diverse fields:

• As the penalty in finite-ratio statistical decision frameworks (Simons et al., 15 May 2026, Washburn et al., 27 Feb 2026)
• As the cost function in constrained Gibbs measures, determining the optimal weight distribution under given recognition costs (Simons et al., 15 May 2026)
• As the geometric divergence in Bregman-type log-coordinates (Washburn et al., 27 Feb 2026)

Reciprocity $J(x) = J(1/x)$ and vanishing at $x=1$ ($J(1)=0$) are immediate consequences of the law. The log-coordinate formulation $J(e^t) = \cosh(t) - 1$ exposes a natural hyperbolic geometry for recognition and divergence.

2. Compositional Recognition in Representation Learning

Within representation learning, the Recognition Composition Law is operationalized at the architectural level by the Recognition as Part Composition (RPC) formulation (Mishra et al., 2022). Here, the law underlies the encoding principle:

1. Decomposition: An image $x$ is divided into $(0, \infty)$0 attended parts via attention maps $(0, \infty)$1, yielding latent part features $(0, \infty)$2.
2. Partwise Mixture: Each part feature $(0, \infty)$3 is encoded as a convex combination over $(0, \infty)$4 learned prototype vectors $(0, \infty)$5, with softmax weights $(0, \infty)$6 such that each $(0, \infty)$7.
3. Full Code Composition: The final code $(0, \infty)$8 is a concatenation of partwise mixtures.
4. Reconstruction: $(0, \infty)$9, i.e., the part feature is reconstructed as a linear mixture of prototypes.

In technical terms, the law is enforced in the learning objective as constraints and autoencoding losses (see equations (4)-(8) in (Mishra et al., 2022)), ensuring a tight linkage between part-based compositionality, cross-class prototype sharing, and transferability. This compositional bottleneck enhances sample efficiency and robustness (few-shot/zero-shot generalization, domain adaptation, adversarial resistance), providing inherent interpretability by tying recognition units to human-interpretable part prototypes.

3. Statistical Mechanics and Ratio-Based Gibbs Constructions

In statistical mechanics, the Recognition Composition Law supplies the structural axiom that links ratio-based costs to the canonical Gibbs construction (Simons et al., 15 May 2026):

• Each outcome $$J(xy) + J\left(\frac{x}{y}\right) = 2 J(x) + 2 J(y) + 2 J(x) J(y)$$0 is assigned a positive ratio $$J(xy) + J\left(\frac{x}{y}\right) = 2 J(x) + 2 J(y) + 2 J(x) J(y)$$1.
• The recognition cost $$J(xy) + J\left(\frac{x}{y}\right) = 2 J(x) + 2 J(y) + 2 J(x) J(y)$$2 defines the effective "energy" for the Gibbs measure.
• Maximizing entropy at fixed mean cost yields the canonical weights

$$J(xy) + J\left(\frac{x}{y}\right) = 2 J(x) + 2 J(y) + 2 J(x) J(y)$$3

where $$J(xy) + J\left(\frac{x}{y}\right) = 2 J(x) + 2 J(y) + 2 J(x) J(y)$$4 is the partition function, and $$J(xy) + J\left(\frac{x}{y}\right) = 2 J(x) + 2 J(y) + 2 J(x) J(y)$$5 is set by the mean cost constraint.

The RCL imposes that the free energy function $$J(xy) + J\left(\frac{x}{y}\right) = 2 J(x) + 2 J(y) + 2 J(x) J(y)$$6 satisfies a Helmholtz–Kullback–Leibler identity, and that cost-induced distributions exhibit a canonical exponential family structure. Comparisons with alternative surrogates (squared-log, affinity-as-energy, Tsallis) demonstrate that the RCL uniquely determines the fluctuation properties and sample-size detectability thresholds for empirical distinctions (Simons et al., 15 May 2026).

4. Formal Power Series, Algebraic Composition, and Right Distributive Laws

The Recognition Composition Law extends to the algebraic composition of formal power series, dictating when compositions and distributive identities hold in one or multiple variables (Bugajewski et al., 2022):

• For $$J(xy) + J\left(\frac{x}{y}\right) = 2 J(x) + 2 J(y) + 2 J(x) J(y)$$7 and $$J(xy) + J\left(\frac{x}{y}\right) = 2 J(x) + 2 J(y) + 2 J(x) J(y)$$8, the composition $$J(xy) + J\left(\frac{x}{y}\right) = 2 J(x) + 2 J(y) + 2 J(x) J(y)$$9 is

$J''(1) = 1$0

• Existence criteria reduce to block convergence properties:

$J''(1) = 1$1

where $J''(1) = 1$2 and $J''(1) = 1$3 is the degree-$J''(1) = 1$4 homogeneous block from $J''(1) = 1$5.

• The Right Distributive Law

$J''(1) = 1$6

holds iff all three series exist, with automatic validity for nonnegative real coefficients. This provides structural closure properties for algebraic manipulations in combinatorial and generating function contexts.

5. Composition Laws in Relaxation, Entropy, and Higher Algebra

a) Fractional Relaxation (Cole–Cole Law)

In time-fractional relaxation, the Cole-Cole (CC) composition law replaces the multiplicative factorization of classical Debye relaxation (Górska et al., 2018). The integro-differential composition

$J''(1) = 1$7

prescribes a bilinear law matching the long-memory, non-Markovian character of CC relaxation. The law's semigroup property leads directly to Caputo or Riemann–Liouville fractional differential equations and generalizes to subordination in anomalous diffusion frameworks.

b) Entropic Nonlinearities ($J''(1) = 1$8-Entropy)

For the $J''(1) = 1$9-entropy $$J(x) = \frac{1}{2} (x + x^{-1}) - 1 = \cosh(\log x) - 1$$0 (Kaniadakis et al., 2017), the composition law for independent systems $$J(x) = \frac{1}{2} (x + x^{-1}) - 1 = \cosh(\log x) - 1$$1 and $$J(x) = \frac{1}{2} (x + x^{-1}) - 1 = \cosh(\log x) - 1$$2 is

$$J(x) = \frac{1}{2} (x + x^{-1}) - 1 = \cosh(\log x) - 1$$3

encoding superadditivity and deformation from Shannon additivity, with explicit parentropy factors dependent on the shifted entropic functionals. This generalizes compositional entropy closure to the relativistic $$J(x) = \frac{1}{2} (x + x^{-1}) - 1 = \cosh(\log x) - 1$$4-framework.

c) Higher Algebraic Composition

Bhargava's higher composition laws (Chinta et al., 6 Feb 2026) provide explicit multilinear composition identities on spaces such as binary cubic forms, pairs of quadratics, and alternated forms, preserving polynomial discriminants and group structures on projectivity classes. The guiding principle is the discriminant-preserving, ideal-class-based group action, supplying a concrete compositional law generalizing Gauss composition.

6. Uniqueness, Calibration, and Structural Consequences

The Recognition Composition Law and its calibration have several intrinsic mathematical consequences:

• Uniqueness: The functional equation and curvature calibration uniquely select $$J(x) = \frac{1}{2} (x + x^{-1}) - 1 = \cosh(\log x) - 1$$5, ruling out all nontrivial alternatives up to trivial rescalings or sign flips, as proven via reduction to a d'Alembert-type functional equation (Washburn et al., 27 Feb 2026, Simons et al., 15 May 2026).
• Normalization and Reciprocity: RCL imposes $$J(x) = \frac{1}{2} (x + x^{-1}) - 1 = \cosh(\log x) - 1$$6 and $$J(x) = \frac{1}{2} (x + x^{-1}) - 1 = \cosh(\log x) - 1$$7 without extra assumptions (Washburn et al., 27 Feb 2026).
• Integrability and Bregman Geometry: The canonical solution aligns the divergence geometry to hyperbolic cosines, enabling explicit Bregman-type coordinate embeddings for error and projection analysis (Washburn et al., 27 Feb 2026).
• Coercivity and Identifiability: Quantitative lower bounds, such as $$J(x) = \frac{1}{2} (x + x^{-1}) - 1 = \cosh(\log x) - 1$$8, control distortion and enforce stability in finite-data identifiability and inference (Washburn et al., 27 Feb 2026).

7. Interpretability, Robustness, and Practical Applications

Recognition Composition Laws confer multiple operational advantages:

• Interpretability: Structured decomposition (e.g., in RPC) yields codes aligned with human-intuitive parts or prototypical components (Mishra et al., 2022).
• Robustness: Convexity and closure at the core of the law provide resistance to adversarial perturbations and noise, with empirical validations in low-shot and domain-shifted settings (Mishra et al., 2022).
• Generalizability: The recirculation and alignment of parts/prototypes enables zero- or few-shot knowledge transfer and synthetic attribute inference, with comparable or superior performance to annotation-based methods (Mishra et al., 2022).
• Canonical Decision: In finite-data certification and projection, procedures based on the canonical cost induced by RCL are maximally informative under identifiability constraints, with no other rule exceeding their local resolving power (Washburn et al., 27 Feb 2026).

The law unifies domains as disparate as image recognition, statistical mechanics, algebra, dynamical systems, and combinatorics by formalizing the only consistent closure structure compatible with natural symmetry, calibration, and compositionality requirements.

---

Key References:

• Recognition as Part Composition and compositional representation learning (Mishra et al., 2022)
• Canonical cost construction and Gibbs measures (Simons et al., 15 May 2026, Washburn et al., 27 Feb 2026)
• Power series and distributive composition (Bugajewski et al., 2022)
• Cole–Cole relaxation and fractional evolution (Górska et al., 2018)
• $$J(x) = \frac{1}{2} (x + x^{-1}) - 1 = \cosh(\log x) - 1$$9-entropy composition (Kaniadakis et al., 2017)
• Higher algebraic composition laws (Chinta et al., 6 Feb 2026)