Cosine Similarity Definition and Applications

• Cosine similarity is a measure that quantifies the angular alignment of two non-zero vectors, ranging from −1 (opposite direction) to 1 (same direction), with 0 indicating orthogonality.
• Key applications of cosine similarity include document retrieval, clustering, classification, and embedding evaluation, enhancing tasks like bibliometrics and search engines where direction similarity is critical. It enables techniques such as PF normalization or XOR-Friendly Binary Quantization for efficient sorting, including in 76.8% accurate text classification tasks.
• Despite its drawbacks— such as insensitivity to centering and sensitivity to vector magnitude— cosine similarity is computationally efficient, appropriate for large-scale datasets, and interacts well with sparse linear algebra. It also partially supports pragmatic uses, aiding orientations in data analysis functions.

Cosine similarity is a normalized inner product that measures the angular alignment of two nonzero vectors. For $\mathbf{x},\mathbf{y}\in\mathbb{R}^n$, it is defined as

$$\operatorname{cosine}(\mathbf{x},\mathbf{y}) = \frac{\mathbf{x}^{\mathsf T}\mathbf{y}} {\|\mathbf{x}\|_2\|\mathbf{y}\|_2} = \frac{\sum_i x_i y_i} {\sqrt{\sum_i x_i^2}\sqrt{\sum_i y_i^2}}.$$

Equivalently, it is the cosine of the angle $\theta$ between the vectors, $\cos(\theta)$. For nonzero real-valued vectors, its range is $[-1,1]$: $1$ denotes identical direction, $0$ orthogonality, and $-1$ opposite direction. For nonnegative vectors, including many term-frequency and probability representations, the range is generally $[0,1]$. Cosine similarity is invariant to multiplication of either vector by a positive scalar, but this invariance does not imply invariance to arbitrary transformations of the underlying data, embedding geometry, covariance structure, or learned parameterization.

1. Mathematical foundation and geometric properties

Cosine similarity compares direction after Euclidean normalization. Defining

$$\widehat{\mathbf{x}}=\frac{\mathbf{x}}{\|\mathbf{x}\|_2}, \qquad \widehat{\mathbf{y}}=\frac{\mathbf{y}}{\|\mathbf{y}\|_2},$$

gives

$$\operatorname{cosine}(\mathbf{x},\mathbf{y}) = \frac{\mathbf{x}^{\mathsf T}\mathbf{y}} {\|\mathbf{x}\|_2\|\mathbf{y}\|_2} = \frac{\sum_i x_i y_i} {\sqrt{\sum_i x_i^2}\sqrt{\sum_i y_i^2}}.$$0

Consequently,

$$\operatorname{cosine}(\mathbf{x},\mathbf{y}) = \frac{\mathbf{x}^{\mathsf T}\mathbf{y}} {\|\mathbf{x}\|_2\|\mathbf{y}\|_2} = \frac{\sum_i x_i y_i} {\sqrt{\sum_i x_i^2}\sqrt{\sum_i y_i^2}}.$$1

The measure is therefore primarily sensitive to orientation rather than absolute magnitude. Two vectors with different lengths can have similarity $$\operatorname{cosine}(\mathbf{x},\mathbf{y}) = \frac{\mathbf{x}^{\mathsf T}\mathbf{y}} {\|\mathbf{x}\|_2\|\mathbf{y}\|_2} = \frac{\sum_i x_i y_i} {\sqrt{\sum_i x_i^2}\sqrt{\sum_i y_i^2}}.$$2 when they point in the same direction. This property makes cosine useful for document vectors, word embeddings, image features, classifier representations, and nearest-neighbor search.

The zero vector is excluded because its norm appears in the denominator. For signed data, cosine can be negative; for nonnegative data, negative values cannot occur. The interpretation also depends on whether the vectors are centered. Pearson correlation is equivalent to cosine similarity after subtracting the sample mean from each vector:

$$\operatorname{cosine}(\mathbf{x},\mathbf{y}) = \frac{\mathbf{x}^{\mathsf T}\mathbf{y}} {\|\mathbf{x}\|_2\|\mathbf{y}\|_2} = \frac{\sum_i x_i y_i} {\sqrt{\sum_i x_i^2}\sqrt{\sum_i y_i^2}}.$$3

Thus, non-centered cosine is affected by shared offsets and the direction of the mean, whereas Pearson correlation removes the mean direction. This distinction is important in bibliometrics, biological perturbational data, and embedding analysis (Smith et al., 2023).

Cosine similarity is not itself a metric. The commonly used quantity

$$\operatorname{cosine}(\mathbf{x},\mathbf{y}) = \frac{\mathbf{x}^{\mathsf T}\mathbf{y}} {\|\mathbf{x}\|_2\|\mathbf{y}\|_2} = \frac{\sum_i x_i y_i} {\sqrt{\sum_i x_i^2}\sqrt{\sum_i y_i^2}}.$$4

does not generally satisfy the metric triangle inequality. Two related metric quantities are angular distance,

$$\operatorname{cosine}(\mathbf{x},\mathbf{y}) = \frac{\mathbf{x}^{\mathsf T}\mathbf{y}} {\|\mathbf{x}\|_2\|\mathbf{y}\|_2} = \frac{\sum_i x_i y_i} {\sqrt{\sum_i x_i^2}\sqrt{\sum_i y_i^2}}.$$5

and the square-root cosine distance,

$$\operatorname{cosine}(\mathbf{x},\mathbf{y}) = \frac{\mathbf{x}^{\mathsf T}\mathbf{y}} {\|\mathbf{x}\|_2\|\mathbf{y}\|_2} = \frac{\sum_i x_i y_i} {\sqrt{\sum_i x_i^2}\sqrt{\sum_i y_i^2}}.$$6

For normalized vectors, the latter is exactly their Euclidean distance. Angular distance is a metric on the unit sphere, and both angular distance and normalized Euclidean distance induce the same ranking as cosine similarity.

2. Representations, preprocessing, and semantic interpretation

In information retrieval, documents are represented as vectors over a vocabulary. After tokenization, stop-word removal, and stemming or lemmatization, a document can be written as

$$\operatorname{cosine}(\mathbf{x},\mathbf{y}) = \frac{\mathbf{x}^{\mathsf T}\mathbf{y}} {\|\mathbf{x}\|_2\|\mathbf{y}\|_2} = \frac{\sum_i x_i y_i} {\sqrt{\sum_i x_i^2}\sqrt{\sum_i y_i^2}}.$$7

where $$\operatorname{cosine}(\mathbf{x},\mathbf{y}) = \frac{\mathbf{x}^{\mathsf T}\mathbf{y}} {\|\mathbf{x}\|_2\|\mathbf{y}\|_2} = \frac{\sum_i x_i y_i} {\sqrt{\sum_i x_i^2}\sqrt{\sum_i y_i^2}}.$$8 is a term weight. These weights may be raw term frequency, binary occurrence, or TF–IDF. A standard TF–IDF weight is

$$\operatorname{cosine}(\mathbf{x},\mathbf{y}) = \frac{\mathbf{x}^{\mathsf T}\mathbf{y}} {\|\mathbf{x}\|_2\|\mathbf{y}\|_2} = \frac{\sum_i x_i y_i} {\sqrt{\sum_i x_i^2}\sqrt{\sum_i y_i^2}}.$$9

Cosine similarity then measures whether two documents assign high weights to the same vocabulary terms. It is insensitive to positive document-length scaling, but it is not mean-centered and remains affected by term distribution and concentration. In an AG’s News experiment, TF–IDF cosine achieved a reported top-1 accuracy of $\theta$0, compared with $\theta$1 for cosine over averaged Word2Vec vectors and $\theta$2 for soft cosine over Word2Vec features (Sitikhu et al., 2019).

Averaged word embeddings represent a document by

$\theta$3

This can capture relationships between different words, but averaging discards word order and document structure. Out-of-vocabulary words assigned zero vectors further weaken the representation. Ordinary cosine therefore treats “John loves Mary” and “Mary loves John” as identical when their term vectors are identical, despite the difference in semantic roles (Crocetti, 2015).

Textual Spatial Cosine Similarity combines ordinary cosine with a positional measure. If $\theta$4 and $\theta$5 are the positions of corresponding occurrences, the normalized positional discrepancy is

$\theta$6

The resulting Textual Space Similarity is

$\theta$7

and the combined measure is

$\theta$8

The cases $\theta$9 and $\cos(\theta)$0 recover ordinary cosine and pure positional similarity, respectively. For “John loves Mary” versus “Mary loves John,” the reported positional similarity is approximately $\cos(\theta)$1, while with $\cos(\theta)$2 the combined score is approximately $\cos(\theta)$3.

Soft cosine relaxes the assumption that distinct coordinates are unrelated. Given a feature-similarity matrix $\cos(\theta)$4,

$\cos(\theta)$5

When $\cos(\theta)$6, soft cosine reduces to ordinary cosine. Off-diagonal entries can encode similarity between terms such as “car” and “automobile,” using Levenshtein distance, WordNet similarity, or word-embedding similarity (Sitikhu et al., 2019).

A different interpretive approach applies centering, whitening, and Independent Component Analysis to embeddings. For normalized ICA vectors $\cos(\theta)$7 and $\cos(\theta)$8,

$\cos(\theta)$9

Each product

$[-1,1]$0

is interpreted as the contribution of an independent-component axis. ICA preserves the total cosine after an orthogonal transformation but can concentrate similarity into sparse, manually interpretable axes such as [spectrum], [space], [chemistry], or animals.

3. Generalized cosine geometries

The ordinary cosine assumes the identity Euclidean metric. A learned metric tensor replaces the standard inner product with

$[-1,1]$1

and defines

$[-1,1]$2

With $[-1,1]$3, this becomes ordinary cosine. A parameterization

$[-1,1]$4

makes the metric symmetric and positive semidefinite and yields the equivalent transformation view

$[-1,1]$5

A full metric can reweight dimensions, introduce cross-dimensional interactions, and alter effective angles without modifying the original embeddings. Context-specific matrices can be learned for semantic contexts such as clothing, vehicles, or birds. In contextualized word-similarity experiments, learned metric tensors improved correlations with human judgments, including a combined contextualized-BERT result increasing from Pearson correlation $[-1,1]$6 to $[-1,1]$7 and Spearman correlation $[-1,1]$8 to $[-1,1]$9 (Vos et al., 2022).

Covariance-adjusted cosine applies a whitening transformation before calculating cosine. If

$1$0

then

$1$1

removes variance and correlation effects in the transformed coordinates. Cosine is then computed between the whitened vectors. The adjustment reduces to ordinary cosine when the covariance is spherical, for example $1$2 with $1$3. On the Wisconsin Breast Cancer Diagnostic Dataset, the reported class-specific covariance adjustment achieved $1$4 test accuracy and $1$5 accuracy in the stated cross-validation procedures, while the expected inverse sample-Cholesky transformation achieved approximately $1$6 held-out accuracy (Sahoo et al., 4 Feb 2025).

A convex-cost generalization defines similarity through surface normals of the graph of a convex function $1$7. For differentiable $1$8,

$1$9

and the normalized normal is

$0$0

The proposed similarity is

$0$1

This compares the local orientation of the cost surface rather than the input coordinates directly. For nondifferentiable convex functions, subgradients replace gradients, and the normals producing the smallest angle are selected. Negative entropy, total variation, and filtered variation are proposed cost functions. For a quadratic cost, the tangent-based alternative becomes ordinary cosine similarity (Gunay et al., 2014).

4. Statistical behavior and limitations in embedding spaces

Cosine similarity is often treated as a semantic measure, but its numerical behavior depends on the distribution and geometry of the representation space. For independent vectors drawn from a centered isotropic Gaussian distribution, the cosine distribution satisfies

$0$2

with

$0$3

For centered data with covariance eigenvalues $0$4, the high-dimensional approximation is

$0$5

The corresponding effective-rank quantity is

$0$6

Equal covariance eigenvalues minimize the variance for a fixed overall scale. Dominant eigenvalues reduce the effective dimension and broaden the null distribution (Smith et al., 2023).

For a nonzero mean, the asymptotic expected cosine is

$0$7

Thus, a shared mean direction can produce positive cosine similarity even for independent observations. The null distribution depends on both the mean and covariance structure, not dimension alone.

Contextual embeddings introduce another source of distortion. BERT contextual representations for frequent words occupy broader spatial regions, and cosine similarity tends to underestimate human-perceived similarity for frequent words. In WiC, for different-meaning examples, a reported regression is

$0$8

with $0$9 and $-1$0. For same-meaning examples, the frequency coefficient is approximately $-1$1. In SCWS, the frequency coefficient remains negative after controlling for human similarity ratings and polysemy. The authors associate this effect with broader contextual sibling cohorts, anisotropy, and frequency-dependent representational geometry rather than vector norm alone (Zhou et al., 2022).

Cosine can also be non-unique when applied to learned embeddings. In a matrix-factorization model trained on dot products, the transformation

$-1$2

for invertible diagonal $-1$3 preserves the product $-1$4 and therefore all predictions. It does not, however, preserve item-item or user-user cosine similarities. Consequently, different factorizations of the same predictive model can yield different cosine structures, including a full-rank factorization in which all distinct item embeddings are orthogonal. Separate-factor regularization removes this diagonal rescaling freedom, but the resulting cosine geometry still reflects the regularizer’s spectral weighting rather than an objective semantic structure (Steck et al., 2024).

These results do not imply that cosine is universally invalid. They show that a cosine score is conditional on centering, covariance, embedding training, regularization, frequency, normalization, and representation parameterization.

5. Applications in learning and retrieval

Cosine similarity is used in document retrieval, clustering, classification, nearest-neighbor search, embedding evaluation, knowledge distillation, OOD detection, topological data analysis, and model analysis.

In bibliometrics, an occurrence matrix $-1$5 contains case-by-variable values, while a co-occurrence matrix is often

$-1$6

Its entries satisfy

$-1$7

Ochiai normalization on $-1$8 is exactly cosine similarity on the underlying occurrence vectors:

$-1$9

Applying cosine or Pearson correlation directly to $[0,1]$0 performs a second normalization over already aggregated inner products and generally inflates apparent similarity. The recommended rule is cosine on occurrence data and Ochiai on corresponding co-occurrence data (Zhou et al., 2015).

For large-scale search, normalized database and query vectors reduce cosine similarity to an inner product:

$[0,1]$1

XOR-Friendly Binary Quantization approximates this inner product using sign-valued binary expansions, XOR, bit shifts, integer additions, and POPCNT operations. The method retains an exhaustive scan but uses quantized scores to construct a candidate set, then recomputes exact floating-point similarities for final top-$[0,1]$2 selection. Its approximate search complexity is

$[0,1]$3

where $[0,1]$4 and $[0,1]$5 are document and query bit depths. Reported experiments include approximately $[0,1]$6 faster full search than an optimized GPU brute-force implementation and recall above $[0,1]$7 under stated extra-distance settings (Jian et al., 2020).

A transitive inequality enables metric-style pruning without turning cosine into a metric. For an intermediate vector $[0,1]$8,

$[0,1]$9

The corresponding upper bound uses a plus sign. These bounds are tight under spherical geometry and can support pruning in VP-trees, cover trees, M-trees, ball trees, LAESA, GNAT, and related structures, provided suitable similarity intervals or angular summaries are stored (Schubert, 2021).

Class Typical Matching uses cosine between a test feature and class-typical feature means. For class $$\widehat{\mathbf{x}}=\frac{\mathbf{x}}{\|\mathbf{x}\|_2}, \qquad \widehat{\mathbf{y}}=\frac{\mathbf{y}}{\|\mathbf{y}\|_2},$$0,

$$\widehat{\mathbf{x}}=\frac{\mathbf{x}}{\|\mathbf{x}\|_2}, \qquad \widehat{\mathbf{y}}=\frac{\mathbf{y}}{\|\mathbf{y}\|_2},$$1

and the OOD score is

$$\widehat{\mathbf{x}}=\frac{\mathbf{x}}{\|\mathbf{x}\|_2}, \qquad \widehat{\mathbf{y}}=\frac{\mathbf{y}}{\|\mathbf{y}\|_2},$$2

An input is classified as in-distribution when this score exceeds a threshold calibrated on ID validation data. CTM is post hoc, stores one mean per class, and avoids dependence on feature magnitude. On ImageNet with ResNet-50, it reported average FPR95 $$\widehat{\mathbf{x}}=\frac{\mathbf{x}}{\|\mathbf{x}\|_2}, \qquad \widehat{\mathbf{y}}=\frac{\mathbf{y}}{\|\mathbf{y}\|_2},$$3 and AUROC $$\widehat{\mathbf{x}}=\frac{\mathbf{x}}{\|\mathbf{x}\|_2}, \qquad \widehat{\mathbf{y}}=\frac{\mathbf{y}}{\|\mathbf{y}\|_2},$$4 (Ngoc-Hieu et al., 2023).

Cosine-based adversarial training uses the squared cosine between encoder features and subsidiary-classifier weight vectors:

$$\widehat{\mathbf{x}}=\frac{\mathbf{x}}{\|\mathbf{x}\|_2}, \qquad \widehat{\mathbf{y}}=\frac{\mathbf{y}}{\|\mathbf{y}\|_2},$$5

Minimizing this loss encourages the encoder representation to become orthogonal to nuisance directions associated with channel or domain labels. In contrast to inverted cross-entropy, which can encode nuisance information negatively, the cosine objective seeks to eliminate the corresponding directional information. The method reported lower error rates in speaker identification and improved unsupervised domain adaptation results (Heo et al., 2019).

Cosine similarity is also used in knowledge distillation. Cosine Similarity Knowledge Distillation compares teacher and student predictions across samples for each class rather than only comparing class distributions within each sample. For class $$\widehat{\mathbf{x}}=\frac{\mathbf{x}}{\|\mathbf{x}\|_2}, \qquad \widehat{\mathbf{y}}=\frac{\mathbf{y}}{\|\mathbf{y}\|_2},$$6,

$$\widehat{\mathbf{x}}=\frac{\mathbf{x}}{\|\mathbf{x}\|_2}, \qquad \widehat{\mathbf{y}}=\frac{\mathbf{y}}{\|\mathbf{y}\|_2},$$7

Cosine Similarity Weighted Temperature assigns lower temperatures when teacher and student predictions are already similar and higher temperatures when they differ. On the reported CIFAR-100 and ImageNet experiments, the method generally improved over the standalone student and listed distillation baselines (Ham et al., 2023).

6. Specialized and emerging uses

Cosine similarity can be applied after vectorizing non-Euclidean objects. Persistence diagrams, for example, can be mapped to persistence landscapes. If $$\widehat{\mathbf{x}}=\frac{\mathbf{x}}{\|\mathbf{x}\|_2}, \qquad \widehat{\mathbf{y}}=\frac{\mathbf{y}}{\|\mathbf{y}\|_2},$$8 denotes the landscape of a diagram $$\widehat{\mathbf{x}}=\frac{\mathbf{x}}{\|\mathbf{x}\|_2}, \qquad \widehat{\mathbf{y}}=\frac{\mathbf{y}}{\|\mathbf{y}\|_2},$$9, the landscape cosine is

$$\operatorname{cosine}(\mathbf{x},\mathbf{y}) = \frac{\mathbf{x}^{\mathsf T}\mathbf{y}} {\|\mathbf{x}\|_2\|\mathbf{y}\|_2} = \frac{\sum_i x_i y_i} {\sqrt{\sum_i x_i^2}\sqrt{\sum_i y_i^2}}.$$00

Because persistence landscapes are nonnegative,

$$\operatorname{cosine}(\mathbf{x},\mathbf{y}) = \frac{\mathbf{x}^{\mathsf T}\mathbf{y}} {\|\mathbf{x}\|_2\|\mathbf{y}\|_2} = \frac{\sum_i x_i y_i} {\sqrt{\sum_i x_i^2}\sqrt{\sum_i y_i^2}}.$$01

The value $$\operatorname{cosine}(\mathbf{x},\mathbf{y}) = \frac{\mathbf{x}^{\mathsf T}\mathbf{y}} {\|\mathbf{x}\|_2\|\mathbf{y}\|_2} = \frac{\sum_i x_i y_i} {\sqrt{\sum_i x_i^2}\sqrt{\sum_i y_i^2}}.$$02 occurs if and only if the persistence diagrams are equal, while $$\operatorname{cosine}(\mathbf{x},\mathbf{y}) = \frac{\mathbf{x}^{\mathsf T}\mathbf{y}} {\|\mathbf{x}\|_2\|\mathbf{y}\|_2} = \frac{\sum_i x_i y_i} {\sqrt{\sum_i x_i^2}\sqrt{\sum_i y_i^2}}.$$03 corresponds to orthogonality and disjoint filtration support. The associated quantity $$\operatorname{cosine}(\mathbf{x},\mathbf{y}) = \frac{\mathbf{x}^{\mathsf T}\mathbf{y}} {\|\mathbf{x}\|_2\|\mathbf{y}\|_2} = \frac{\sum_i x_i y_i} {\sqrt{\sum_i x_i^2}\sqrt{\sum_i y_i^2}}.$$04 is called a cosine distance but is not asserted to be a metric. In experiments on planar shapes, landscape cosine distance more sharply separated a disk from an annulus or circle than bottleneck, Wasserstein, and landscape norm distances (Nordin et al., 6 Apr 2025).

For sentence-embedding corpora, similarity distributions can be modeled rather than treated only as rankings. Shifted gamma distributions and gamma mixtures have been proposed for query-specific cosine scores. A shifted component has density proportional to

$$\operatorname{cosine}(\mathbf{x},\mathbf{y}) = \frac{\mathbf{x}^{\mathsf T}\mathbf{y}} {\|\mathbf{x}\|_2\|\mathbf{y}\|_2} = \frac{\sum_i x_i y_i} {\sqrt{\sum_i x_i^2}\sqrt{\sum_i y_i^2}}.$$05

with shape $$\operatorname{cosine}(\mathbf{x},\mathbf{y}) = \frac{\mathbf{x}^{\mathsf T}\mathbf{y}} {\|\mathbf{x}\|_2\|\mathbf{y}\|_2} = \frac{\sum_i x_i y_i} {\sqrt{\sum_i x_i^2}\sqrt{\sum_i y_i^2}}.$$06, shift $$\operatorname{cosine}(\mathbf{x},\mathbf{y}) = \frac{\mathbf{x}^{\mathsf T}\mathbf{y}} {\|\mathbf{x}\|_2\|\mathbf{y}\|_2} = \frac{\sum_i x_i y_i} {\sqrt{\sum_i x_i^2}\sqrt{\sum_i y_i^2}}.$$07, and rate $$\operatorname{cosine}(\mathbf{x},\mathbf{y}) = \frac{\mathbf{x}^{\mathsf T}\mathbf{y}} {\|\mathbf{x}\|_2\|\mathbf{y}\|_2} = \frac{\sum_i x_i y_i} {\sqrt{\sum_i x_i^2}\sqrt{\sum_i y_i^2}}.$$08. Mixture parameters can be estimated through EM or ECM procedures. The motivation is to calibrate unusually high similarities, estimate tail probabilities, and support anomaly or significance detection. The method is empirical and query-dependent; exact truncation to the cosine range $$\operatorname{cosine}(\mathbf{x},\mathbf{y}) = \frac{\mathbf{x}^{\mathsf T}\mathbf{y}} {\|\mathbf{x}\|_2\|\mathbf{y}\|_2} = \frac{\sum_i x_i y_i} {\sqrt{\sum_i x_i^2}\sqrt{\sum_i y_i^2}}.$$09 requires an explicit normalization term that is not included in the displayed fitting equations (Player, 6 Oct 2025).

Several alternatives modify the normalization itself. Recos replaces the Euclidean norm product with a rearrangement-based denominator:

$$\operatorname{cosine}(\mathbf{x},\mathbf{y}) = \frac{\mathbf{x}^{\mathsf T}\mathbf{y}} {\|\mathbf{x}\|_2\|\mathbf{y}\|_2} = \frac{\sum_i x_i y_i} {\sqrt{\sum_i x_i^2}\sqrt{\sum_i y_i^2}}.$$10

Its positive perfect-similarity condition is ordinal concordance rather than strict proportionality. The method requires sorting and therefore approximately $$\operatorname{cosine}(\mathbf{x},\mathbf{y}) = \frac{\mathbf{x}^{\mathsf T}\mathbf{y}} {\|\mathbf{x}\|_2\|\mathbf{y}\|_2} = \frac{\sum_i x_i y_i} {\sqrt{\sum_i x_i^2}\sqrt{\sum_i y_i^2}}.$$11 time rather than the $$\operatorname{cosine}(\mathbf{x},\mathbf{y}) = \frac{\mathbf{x}^{\mathsf T}\mathbf{y}} {\|\mathbf{x}\|_2\|\mathbf{y}\|_2} = \frac{\sum_i x_i y_i} {\sqrt{\sum_i x_i^2}\sqrt{\sum_i y_i^2}}.$$12 cost of ordinary cosine. Across seven STS datasets and eleven embedding models, the reported mean score was $$\operatorname{cosine}(\mathbf{x},\mathbf{y}) = \frac{\mathbf{x}^{\mathsf T}\mathbf{y}} {\|\mathbf{x}\|_2\|\mathbf{y}\|_2} = \frac{\sum_i x_i y_i} {\sqrt{\sum_i x_i^2}\sqrt{\sum_i y_i^2}}.$$13 for recos and $$\operatorname{cosine}(\mathbf{x},\mathbf{y}) = \frac{\mathbf{x}^{\mathsf T}\mathbf{y}} {\|\mathbf{x}\|_2\|\mathbf{y}\|_2} = \frac{\sum_i x_i y_i} {\sqrt{\sum_i x_i^2}\sqrt{\sum_i y_i^2}}.$$14 for cosine (Ai, 5 Feb 2026).

Cos-Mix applies cosine similarity and cosine-distance retrieval in a hybrid-plus-fallback RAG architecture. BM25 and dense retrieval are used initially; an LLM validates the response; if validation fails, a sparse-information subset is searched through a distance-based path and reranked. The supplied formulation does not specify a numerical weighted fusion equation, so the method is best characterized as a conditional retrieval architecture rather than a fully specified score-level mixture (Juvekar et al., 2024).

7. Interpretation, validation, and limitations

Cosine similarity measures angular alignment, not semantic truth, causal importance, or statistical equivalence by itself. In Transformer analysis, a layer can produce a small angular change and remain functionally indispensable because later nonlinearities, residual pathways, or downstream amplification can magnify a small signal. A theoretical construction shows that a layer can have cosine-based relevance arbitrarily close to zero while its removal reduces perfect accuracy to zero. Empirical correlations between cosine-based layer relevance and performance degradation were reported as $$\operatorname{cosine}(\mathbf{x},\mathbf{y}) = \frac{\mathbf{x}^{\mathsf T}\mathbf{y}} {\|\mathbf{x}\|_2\|\mathbf{y}\|_2} = \frac{\sum_i x_i y_i} {\sqrt{\sum_i x_i^2}\sqrt{\sum_i y_i^2}}.$$15 for Pythia, $$\operatorname{cosine}(\mathbf{x},\mathbf{y}) = \frac{\mathbf{x}^{\mathsf T}\mathbf{y}} {\|\mathbf{x}\|_2\|\mathbf{y}\|_2} = \frac{\sum_i x_i y_i} {\sqrt{\sum_i x_i^2}\sqrt{\sum_i y_i^2}}.$$16 for Mistral, and $$\operatorname{cosine}(\mathbf{x},\mathbf{y}) = \frac{\mathbf{x}^{\mathsf T}\mathbf{y}} {\|\mathbf{x}\|_2\|\mathbf{y}\|_2} = \frac{\sum_i x_i y_i} {\sqrt{\sum_i x_i^2}\sqrt{\sum_i y_i^2}}.$$17 for OLMo; cosine misestimated layer relevance in $$\operatorname{cosine}(\mathbf{x},\mathbf{y}) = \frac{\mathbf{x}^{\mathsf T}\mathbf{y}} {\|\mathbf{x}\|_2\|\mathbf{y}\|_2} = \frac{\sum_i x_i y_i} {\sqrt{\sum_i x_i^2}\sqrt{\sum_i y_i^2}}.$$18 of layer–task cases (Hinostroza et al., 13 May 2026).

For model pruning, direct accuracy degradation after layer removal is more task-specific but substantially more expensive. Cosine is useful as a fast descriptive statistic or screening heuristic, whereas causal ablation is preferable when pruning safety or functional relevance is the target.

The principal validation requirements for cosine-based systems are therefore:

• Representation validation: determine whether direction, magnitude, order, or covariance structure carries the relevant signal.
• Normalization validation: check whether centering, whitening, TF–IDF weighting, or metric learning is appropriate.
• Distributional validation: inspect score distributions, effective rank, frequency dependence, and query or corpus dependence.
• Task validation: compare cosine with dot product, Euclidean distance, Pearson correlation, learned metrics, and application-specific baselines.
• Numerical validation: handle zero vectors, near-zero norms, finite precision, quantization error, and truncated or approximate score models.
• Causal validation: when cosine is used to infer importance, perform interventions such as ablation rather than relying exclusively on local angular change.

Cosine similarity remains attractive because it is simple, symmetric, scale-invariant under positive rescaling, computationally efficient, and compatible with sparse linear algebra and highly parallel hardware. Its limitations arise when magnitude contains useful information, when covariance makes Euclidean geometry inappropriate, when embedding parameterizations are non-unique, when contextual dispersion varies across words, or when the desired relationship is nonlinear, ordinal, distributional, or causal. Its reliable use therefore depends less on the formula alone than on the representation, geometry, statistical assumptions, and downstream objective to which the formula is applied.