Recos: Similarity Metric and Neural Transform

• Recos is a dual-concept term referring both to a rearrangement-inequality-based similarity metric for measuring ordinal concordance and to a neural transform used in CNNs.
• The recos metric leverages sorted vector orderings to capture monotonic relationships, leading to improved semantic retrieval performance over traditional cosine similarity.
• RECOS transforms employ spherical clustering with rectified activations to highlight the pivotal role of nonlinearity in deep neural network architectures and signal processing.

The term Recos has two main usages in contemporary machine learning and information retrieval research. Most importantly, "recos" refers to a mathematically principled similarity metric, the Rearrangement-inequality-based Cosine Similarity, introduced as a generalization of classical cosine similarity to address the limitations of linear dependence in measuring semantic similarity in vector spaces. Separately, but less commonly, the acronym "RECOS" (REctified-COrrelations on a Sphere) also describes a family of geometric transforms underpinning convolutional neural networks and signal processing algorithms, emphasizing spherical data clustering with rectified activations. Both concepts are foundational for advanced recommender systems and the analysis of neural architectures.

1. Rearrangement-Inequality-Based Cosine Similarity: recos

Formal Definition and Derivation

The recos similarity metric is designed to supersede standard cosine similarity for vector comparison, moving beyond its reliance on the Cauchy–Schwarz inequality and expanding the class of relationships identified as "perfect similarity" (Ai, 5 Feb 2026). It operationalizes this transition via the rearrangement inequality, which offers a tighter upper bound on the dot product $u \cdot v$ than the classic Cauchy–Schwarz bound.

Given vectors $u, v \in \mathbb{R}^d$ with $u \cdot v \neq 0$, define:

• $u^{\uparrow}$ as $u$ sorted in non-decreasing order;
• $v^{\uparrow}$, $v^{\downarrow}$ as $v$ sorted ascending/descending;
• $v^{\updownarrow} = v^{\uparrow}$ if $u \cdot v > 0$, and $u, v \in \mathbb{R}^d$0 if $u, v \in \mathbb{R}^d$1.

The metric is: $u, v \in \mathbb{R}^d$2 The denominator $u, v \in \mathbb{R}^d$3 is the tightest possible upper bound on $u, v \in \mathbb{R}^d$4 for given component values under any coordinate ordering. This normalization ensures that $u, v \in \mathbb{R}^d$5, with saturation (score $u, v \in \mathbb{R}^d$6) under mere ordinal concordance rather than strict proportionality.

Distinction from Cosine Similarity

Whereas cosine similarity saturates only for vectors differing by a positive or negative scalar multiple ($u, v \in \mathbb{R}^d$7), recos saturates for any pair of vectors that have strictly concordant (respectively, strictly discordant) per-coordinate orderings. In other words, recos is sensitive to monotonic (ordinal) relationships, even if these are not strictly linear.

This makes recos particularly adept in complex, real-world embedding spaces where high-dimensional semantic vectors often encode non-proportional yet concordant information.

Theoretical Properties

The key theoretical guarantees, proved via the rearrangement inequality, include:

• For all $u, v \in \mathbb{R}^d$8, $u, v \in \mathbb{R}^d$9.
• $u \cdot v \neq 0$0 if and only if all coordinate pairs are strictly concordant (or in perfect reverse order for $u \cdot v \neq 0$1).
• Metric hierarchy: $u \cdot v \neq 0$2, with each equality’s condition strictly more general.

Empirical Benchmarking

recos was evaluated across 77 settings (11 embedding models $u \cdot v \neq 0$3 7 STS datasets), including static (Word2Vec, FastText, GloVe), contextualized (BERT, SGPT, DPR), and universal (BGE, E5, GTE, SPECTER, CLIP-ViT) embeddings. Over 92% of the time recos outperformed standard cosine similarity, with significant micro-average gains in Spearman’s correlation with human similarity judgments (66.12 for recos vs. 65.83 for cos). The improvement was especially marked for contextual and universal embedding models (e.g., CLIP-ViT, $u \cdot v \neq 0$4 points) (Ai, 5 Feb 2026).

Statistical significance was established via the Wilcoxon signed-rank test ($u \cdot v \neq 0$5), sign test (win-rate $u \cdot v \neq 0$6), and mixed-effects $u \cdot v \neq 0$7-test ($u \cdot v \neq 0$8).

Practical Considerations

• recos should be preferred when embeddings are likely to encode meaningful ordinal (monotonic) structure, such as contextualized or multi-modal representations, or when standard cosine underestimates similarity in non-proportional relationships.
• The extra computational overhead ($u \cdot v \neq 0$9 for sorting) is negligible except at extremely large scale.
• recos does not necessarily replace cosine in all pipelines, but augments conventional metrics where sensitivity to ordinal concordance yields higher alignment with semantic ground truth.

2. RECOS Transform in Neural Architecture and Signal Processing

A second, less commonly referenced, meaning of "RECOS" appears in the analysis of neural networks, particularly convolutional architectures (Kuo, 2017, Kuo, 2016, Su, 2017). Here, RECOS stands for REctified-COrrelations on a Sphere.

Mathematical Formulation

A RECOS layer is defined by a set of anchor vectors $u^{\uparrow}$0 on the unit sphere $u^{\uparrow}$1. The transform is as follows:

1. Spherical correlations: Output $u^{\uparrow}$2 measures geodesic proximity between $u^{\uparrow}$3 and each $u^{\uparrow}$4.
2. Rectification: Apply a monotonic, non-negative "rectifier" (e.g., ReLU, power-law, or sigmoid) elementwise, yielding $u^{\uparrow}$5. This enforces non-negativity and enables meaningful stacking of layers.

Stacked RECOS transforms (linear correlation + rectification) form the architectural primitive of CNNs and certain deep signal-processing systems. The rectification step is essential; absent it, all layers collapse linearly and "depth" provides no representational advantage (Kuo, 2016).

Interpretation and Applications

• Clustering View: Each layer clusters inputs on the sphere into soft cluster memberships over the anchors (filters).
• Deep Learning: Multi-layer cascades enable increasingly peaked, position-aware clustering, with higher layers aggregating spatial motifs and supporting the emergence of global class or object representations (Kuo, 2017).
• Signal Processing: RECOS cascades explain the equivalence of multi-layer DNNs and deep cepstral or autocorrelation transforms in music transcription and pitch estimation; nonlinearity (e.g., power-law rectification) restores missing fundamentals and enhances noise robustness (Su, 2017).

Key Theoretical Insights

• Without rectification, information about negative correlations is misused in deeper layers; rectification ensures monotonic, interpretable flow of similarity, preventing re-flipping of "anti-patterns" (Kuo, 2016).
• The RECOS formalism reveals the paramount importance of activation functions (especially ReLU and its variants) for the expressive power of deep networks.
• Empirical analyses on classic datasets (e.g., MNIST) and architectures (LeNet-5, AlexNet) illustrate these principles and clarify the role of anchor vectors as learned prototypes.

3. Impact and Applications Across Research Areas

Similarity-Based Retrieval and Embedding Analysis

recos finds immediate application in:

• Semantic textual similarity: Outperforms cosine on STS benchmarks due to greater alignment with human-judged ordinal concordances.
• Universal and contextual embeddings: Especially effective when vector distributions violate assumptions of isotropy or uniform norm.
• Billion-scale retrieval, reranking, and search: As a drop-in alternative or complement to cosine similarity, particularly in modalities beyond text where true similarity often correlates with ordinal agreement.

Neural Architecture, Signal Processing, and Deep Learning

RECOS transforms:

• Clarify the mathematical role of ReLU and its generalizations, providing a geometric rationale for their necessity in modern architectures.
• Unify perspectives on deep neural networks and homomorphic signal processing, showing that classical cepstral cascades and CNNs share a foundational principle.
• Enable efficient quantization and binarization: Since only sign order matters after rectification, filters and activations can be aggressively quantized for high-throughput, energy-efficient inference.

4. Limitations and Statistical Guarantees

recos Metric

• The computational overhead is minimal at moderate scale, but sorting may be replaced by partial or approximate schemes for massive retrieval tasks (Ai, 5 Feb 2026).
• recos, while consistently outperforming cosine, yields improvements in correlation that are typically modest (ca. $u^{\uparrow}$6 points micro-averaged). Its main value is where cosine dramatically underestimates similarity due to non-proportional but concordant structure.

RECOS Perspective

• The anchor-vector and rectification formalism most accurately describes models with explicit filter banks and clear geometric interpretations; its utility for highly entangled or transformer architectures may require further extension (Kuo, 2017, Su, 2017).
• RECOS-based explanations do not obviate the need for empirical evaluation and hyperparameter tuning in practice.

5. Conclusion

The term recos principally refers to a similarity metric grounded in the rearrangement inequality, which extends cosine similarity by accounting for monotonic (ordinal) concordance rather than strict linear dependence in vector space similarity comparison (Ai, 5 Feb 2026). This metric yields tangible empirical improvements in tasks where embedding spaces are non-uniform and semantic similarity is governed as much by relative ranking as by proportional scaling. In parallel, "RECOS" transforms provide a geometric, rectified, spherical-clustering model of learning in modern neural networks, clarifying the role of activation nonlinearity and cascaded architecture in deep learning and homomorphic signal processing (Kuo, 2017, Kuo, 2016). Both conceptions underpin key advances in recommender system theory, neural network analysis, and high-dimensional information retrieval.