Correlation Dimension Rate

• Correlation Dimension Rate is the scaling exponent from the log–log slope of the correlation integral, defining how pairwise proximities grow with resolution.
• It is applied in diverse fields such as dynamical systems, network structures, and statistical manifolds, revealing intrinsic geometric or effective dimensionality.
• Empirical studies demonstrate its utility in diagnosing dynamical regimes and quantifying effective dimensions across systems from regular lattices to complex language models.

Searching arXiv for recent and foundational papers on correlation dimension and related “rate” interpretations. arxiv_search(query="correlation dimension Grassberger Procaccia correlation dimension rate graphs random walk statistical manifold causal networks", max_results=10) Correlation dimension rate is the scaling exponent that governs how a correlation integral or correlation sum grows with resolution scale in a metric space. In the Grassberger–Procaccia framework, for a point set or trajectory, the basic object is the fraction of pairs closer than a threshold $r$, and the rate is the log–log slope in the small-$r$ scaling regime,

$$C(r) \propto r^{D_2}, \qquad \beta \equiv D_2 = \frac{\mathrm{d}\,\log C(r)}{\mathrm{d}\,\log r}.$$

In this sense, “rate” usually denotes the scaling rate itself rather than a temporal derivative. Across later extensions to graphs, statistical manifolds, causal-network inference, binary data, and LLMs, the same core object persists: a slope of $\log C$ versus $\log r$, estimated on an appropriate scaling window and interpreted as an intrinsic geometric or effective dimensional quantity (Lacasa et al., 2014).

1. Formal definition and classical framework

In the standard Grassberger–Procaccia setting, for a point set $\{x_1,\dots,x_N\}$ in a metric space with norm $\|\cdot\|$, the correlation integral counts the fraction of pairs within radius $r$: $$C(r) = \lim_{N\to\infty} \frac{2}{N(N-1)} \sum_{i<j} \Theta\big(r - \|\mathbf{x}_i - \mathbf{x}_j\|\big) \propto r^{D_2}.$$ The correlation dimension is the corresponding scaling exponent,

$$D_2 = \lim_{r\to 0}\frac{\mathrm{d}\log C(r)}{\mathrm{d}\log r},$$

or, equivalently, the slope of the linear part of the $r$0 versus $r$1 curve in a scaling regime (Lacasa et al., 2014).

This usage of “rate” is explicit in several later formulations. In complex networks, the correlation dimension rate at scale $r$2 is written as

$r$3

with a constant plateau indicating a well-defined dimension (Lacasa et al., 2012). In causal-network inference, the correlation dimension itself is described as the scaling rate of the correlation sum,

$r$4

estimated by linear regression over a scale range where $r$5 versus $r$6 is approximately linear (Usta et al., 2024). In measure-theoretic treatments, the same quantity appears as

$r$7

with $r$8, or equivalently through $r$9-energy thresholds and cylinder-sum formulas on suitable filtrations (Yang et al., 2015).

A recurring theme is that the “rate” is not restricted to Euclidean point clouds. What changes across applications is the representation of states, the metric, and the finite-sample estimator; what remains fixed is the asymptotic slope interpretation.

2. Delay embeddings, trajectories, and graph-based generalizations

A major extension replaces a static point cloud by a trajectory. For a scalar or vector time series, delay-embedding constructs

$$C(r) \propto r^{D_2}, \qquad \beta \equiv D_2 = \frac{\mathrm{d}\,\log C(r)}{\mathrm{d}\,\log r}.$$0

and the embedded correlation sum is

$$C(r) \propto r^{D_2}, \qquad \beta \equiv D_2 = \frac{\mathrm{d}\,\log C(r)}{\mathrm{d}\,\log r}.$$1

The embedded exponent

$$C(r) \propto r^{D_2}, \qquad \beta \equiv D_2 = \frac{\mathrm{d}\,\log C(r)}{\mathrm{d}\,\log r}.$$2

is then estimated from the log–log slope in a scaling window (Lacasa et al., 2014).

For graphs, Lacasa and Gómez-Gardeñes adapt Grassberger–Procaccia by generating a node-coordinate time series from an ergodic random walk on a spatially embedded graph. Each node $$C(r) \propto r^{D_2}, \qquad \beta \equiv D_2 = \frac{\mathrm{d}\,\log C(r)}{\mathrm{d}\,\log r}.$$3 carries a coordinate vector $$C(r) \propto r^{D_2}, \qquad \beta \equiv D_2 = \frac{\mathrm{d}\,\log C(r)}{\mathrm{d}\,\log r}.$$4, and an unbiased walker generates $$C(r) \propto r^{D_2}, \qquad \beta \equiv D_2 = \frac{\mathrm{d}\,\log C(r)}{\mathrm{d}\,\log r}.$$5. Delay vectors are built from this sequence, and the same pair-counting logic yields a graph-based correlation sum (Lacasa et al., 2012). The method relies only on local information provided by the walker and does not require global topology or box coverings.

The analytically clean case is the integer lattice $$C(r) \propto r^{D_2}, \qquad \beta \equiv D_2 = \frac{\mathrm{d}\,\log C(r)}{\mathrm{d}\,\log r}.$$6. For $$C(r) \propto r^{D_2}, \qquad \beta \equiv D_2 = \frac{\mathrm{d}\,\log C(r)}{\mathrm{d}\,\log r}.$$7, the paper proves that, up to first order in $$C(r) \propto r^{D_2}, \qquad \beta \equiv D_2 = \frac{\mathrm{d}\,\log C(r)}{\mathrm{d}\,\log r}.$$8, the correlation dimension equals the Euclidean or Hausdorff/topological dimension,

$$C(r) \propto r^{D_2}, \qquad \beta \equiv D_2 = \frac{\mathrm{d}\,\log C(r)}{\mathrm{d}\,\log r}.$$9

The underlying argument uses lattice homogeneity and isotropy, independence of Cartesian displacements, and the small-$\log C$0 behavior of folded-normal coordinate differences under the $\log C$1 norm. For $\log C$2, $\log C$3, so $\log C$4; for $\log C$5, $\log C$6, so $\log C$7; in general $\log C$8, $\log C$9 for all $\log r$0 (Lacasa et al., 2014).

The contrast case is a fully connected graph. There, $\log r$1 increases without saturation, which is interpreted as an “infinite-dimensional” object (Lacasa et al., 2014). This establishes a central distinction in the literature: on regular spatial networks the rate recovers geometric dimension, whereas on highly nonlocal graphs it can reflect effective infinite-dimensionality.

3. Metrics, manifolds, and generalized state spaces

The correlation dimension rate is metric-dependent in representation, but not arbitrary in meaning. In LLM-state analyses based on next-token log-probabilities, the state vector is

$\log r$2

and the correlation integral is computed with Euclidean distances in FP32 over these vectors (Du et al., 24 Oct 2025). In “Correlation Dimension of Natural Language in a Statistical Manifold,” the same logic is reformulated on the statistical manifold of multinoulli distributions, with Fisher–Rao distance

$\log r$3

The empirical estimator is again

$\log r$4

with $\log r$5 in the reported analyses (Du et al., 2024).

The same structural generalization appears in abstract multivariable spaces. “Multivariable-based correlation dimension analysis for generalized space” defines a generalized space by choosing a distance $\log r$6 over multivariable observations, including Euclidean, Minkowski, Chebychev, precisely weighted, and Mahalanobis distances. The global correlation function is

$\log r$7

or equivalently $\log r$8, with $\log r$9 estimated as the slope of $\{x_1,\dots,x_N\}$0 versus $\{x_1,\dots,x_N\}$1 in a scaling range (Chen, 2022).

Binary data require a different metric geometry. For $\{x_1,\dots,x_N\}$2 datasets, the relevant distance is Hamming $\{x_1,\dots,x_N\}$3, and the correlation function is

$\{x_1,\dots,x_N\}$4

where $\{x_1,\dots,x_N\}$5 is the Hamming distance between two randomly chosen points. The corresponding dimension is the least-squares slope of $\{x_1,\dots,x_N\}$6 over a chosen interval. Because the unnormalized values are difficult to interpret, the paper introduces normalized fractal dimension: the number of independent Bernoulli columns required to match the observed unnormalized fractal dimension (Tatti et al., 2019).

These constructions suggest a broad but coherent notion of correlation dimension rate: it is the scaling exponent of pairwise proximity in a chosen metric representation, whether the underlying states are points in $\{x_1,\dots,x_N\}$7, random-walk embeddings of graphs, categorical distributions on a Fisher–Rao manifold, or binary vectors under Hamming distance.

4. Relations to dynamical rates, fluctuations, and conditional information flow

In dynamical-systems theory, correlation dimension rate interacts with genuinely temporal rates. “Fluctuations of separation of trajectories in chaos and correlation dimension” introduces the scaled cumulant generating function

$\{x_1,\dots,x_N\}$8

and identifies the only non-trivial zero of $\{x_1,\dots,x_N\}$9 at

$\|\cdot\|$0

For incompressible flows, $\|\cdot\|$1, so $\|\cdot\|$2. This links the pair-correlation scaling exponent to the exponential growth rates of separation moments and to large deviations of finite-time stretching (Fouxon et al., 2019).

A different link appears in inertial-particle dynamics. There, the logarithmic separation $\|\cdot\|$3 acquires a stationary tail $\|\cdot\|$4, and matching this to $\|\cdot\|$5 gives

$\|\cdot\|$6

In the tilted-generator formulation, $\|\cdot\|$7 is the spectral parameter in a solvability condition $\|\cdot\|$8, so the dimension plays the role of an eigenvalue-like rate (Gustavsson et al., 2015).

Extreme-value theory produces another rate interpretation. For the observable $\|\cdot\|$9 on product dynamics, block maxima obey a Gumbel law whose scale parameter yields $r$0, while the Dynamical Extremal Index $r$1 captures clustering of extremes and is related to contraction or instability rates such as positive Lyapunov exponents or $r$2 under suitable assumptions (Faranda et al., 2017). Here the geometric scaling exponent and temporal instability rates are extracted from the same extreme-value construction but remain conceptually distinct.

Conditional variants shift the emphasis from intrinsic geometry to information flow. In the oGeoC framework, one defines

$r$3

and then

$r$4

In this setting, conditional differences of correlation-dimension rates quantify geometric information flow and distinguish direct from indirect couplings (Usta et al., 2024).

5. Empirical interpretations across domains

In spatial networks, the rate recovers effective geometry. For synthetic and real-world networks, the original network paper reported $r$5 for a 2D lattice at $r$6, $r$7 for the world air-transportation network at $r$8, $r$9 for San Joaquin at $$C(r) = \lim_{N\to\infty} \frac{2}{N(N-1)} \sum_{i<j} \Theta\big(r - \|\mathbf{x}_i - \mathbf{x}_j\|\big) \propto r^{D_2}.$$0, and no clear scaling regime for Oldenburg (Lacasa et al., 2012). The lattice results align with the analytic proof that $$C(r) = \lim_{N\to\infty} \frac{2}{N(N-1)} \sum_{i<j} \Theta\big(r - \|\mathbf{x}_i - \mathbf{x}_j\|\big) \propto r^{D_2}.$$1 has $$C(r) = \lim_{N\to\infty} \frac{2}{N(N-1)} \sum_{i<j} \Theta\big(r - \|\mathbf{x}_i - \mathbf{x}_j\|\big) \propto r^{D_2}.$$2 to first order (Lacasa et al., 2014).

In natural-language modeling on the Fisher–Rao statistical manifold, the reported global correlation dimension is around $$C(r) = \lim_{N\to\infty} \frac{2}{N(N-1)} \sum_{i<j} \Theta\big(r - \|\mathbf{x}_i - \mathbf{x}_j\|\big) \propto r^{D_2}.$$3. The paper gives language-wise values of English $$C(r) = \lim_{N\to\infty} \frac{2}{N(N-1)} \sum_{i<j} \Theta\big(r - \|\mathbf{x}_i - \mathbf{x}_j\|\big) \propto r^{D_2}.$$4, Chinese $$C(r) = \lim_{N\to\infty} \frac{2}{N(N-1)} \sum_{i<j} \Theta\big(r - \|\mathbf{x}_i - \mathbf{x}_j\|\big) \propto r^{D_2}.$$5, Japanese $$C(r) = \lim_{N\to\infty} \frac{2}{N(N-1)} \sum_{i<j} \Theta\big(r - \|\mathbf{x}_i - \mathbf{x}_j\|\big) \propto r^{D_2}.$$6, and German $$C(r) = \lim_{N\to\infty} \frac{2}{N(N-1)} \sum_{i<j} \Theta\big(r - \|\mathbf{x}_i - \mathbf{x}_j\|\big) \propto r^{D_2}.$$7, with SEP texts concentrated around $$C(r) = \lim_{N\to\infty} \frac{2}{N(N-1)} \sum_{i<j} \Theta\big(r - \|\mathbf{x}_i - \mathbf{x}_j\|\big) \propto r^{D_2}.$$8 and $$C(r) = \lim_{N\to\infty} \frac{2}{N(N-1)} \sum_{i<j} \Theta\big(r - \|\mathbf{x}_i - \mathbf{x}_j\|\big) \propto r^{D_2}.$$9 for over $$D_2 = \lim_{r\to 0}\frac{\mathrm{d}\log C(r)}{\mathrm{d}\log r},$$0 of samples (Du et al., 2024). Long memory is identified as key: as context length decreases from $$D_2 = \lim_{r\to 0}\frac{\mathrm{d}\log C(r)}{\mathrm{d}\log r},$$1 to $$D_2 = \lim_{r\to 0}\frac{\mathrm{d}\log C(r)}{\mathrm{d}\log r},$$2, the scaling region shrinks and the dimension decreases to about $$D_2 = \lim_{r\to 0}\frac{\mathrm{d}\log C(r)}{\mathrm{d}\log r},$$3.

In autoregressive LLMs, the same exponent is computed from sequences of next-token log-probability vectors. The paper reports that natural-language documents across multiple models show $$D_2 = \lim_{r\to 0}\frac{\mathrm{d}\log C(r)}{\mathrm{d}\log r},$$4–$$D_2 = \lim_{r\to 0}\frac{\mathrm{d}\log C(r)}{\mathrm{d}\log r},$$5, that context length produces a two-stage curve in which $$D_2 = \lim_{r\to 0}\frac{\mathrm{d}\log C(r)}{\mathrm{d}\log r},$$6 rises from about $$D_2 = \lim_{r\to 0}\frac{\mathrm{d}\log C(r)}{\mathrm{d}\log r},$$7 to about $$D_2 = \lim_{r\to 0}\frac{\mathrm{d}\log C(r)}{\mathrm{d}\log r},$$8 up to $$D_2 = \lim_{r\to 0}\frac{\mathrm{d}\log C(r)}{\mathrm{d}\log r},$$9 and then declines to about $r$00, and that training produces a three-stage trajectory: decrease, increase, decline (Du et al., 24 Oct 2025). That paper also introduces a distinct but related use of “Correlation Dimension Rate”: $r$01 These are derivatives of the correlation dimension with respect to training step, context length, or generation index rather than the original log–log scaling slope.

The same paper reports several diagnostic associations. Across SEP articles for Qwen2.5-32B, Pearson’s correlation between $r$02 and perplexity is about $r$03. In degeneration detection for Falcon3-10B, $r$04 decreases across repetitive $r$05, incoherent $r$06, and bland $r$07 outputs versus normal $r$08, all $r$09. On a knowledge-intensive list, models with $r$10 hallucinated, while Falcon3-7B $r$11 and Falcon3-10B $r$12 aligned with accurate recall. In a long-text stress test, Spearman’s $r$13 between $r$14 and HelloEval long-text scores (Du et al., 24 Oct 2025).

In causal-network inference, the empirical interpretation is structural rather than semantic. On coupled logistic networks, the oGeoC algorithms recover direct links with TPR approaching $r$15 and FPR approaching $r$16 as $r$17 grows in the directed $r$18 case; for $r$19 Erdős–Rényi networks, mean TPR increases toward $r$20 and FPR decreases with $r$21, while larger significance threshold $r$22 degrades performance (Usta et al., 2024). Here the correlation dimension rate functions as a geometric signature whose conditional changes reveal direct and indirect influence.

6. Estimation, finite-size effects, and common pitfalls

All formulations depend on scaling-window selection. In classical GP estimation, one fits the slope where $r$23 versus $r$24 is approximately linear, excluding very small $r$25 where counts are poor and very large $r$26 where saturation dominates (Lacasa et al., 2014). For graph trajectories and lattice walks, finite-size effects arise from discreteness, saturation near system diameter, and bias when trajectories are too short for central-limit approximations to hold (Lacasa et al., 2014). In complex-network practice, the recommended procedure is to compute $r$27 over log-spaced radii, identify an intermediate scaling regime, fit $r$28, and verify saturation in $r$29 (Lacasa et al., 2012).

The literature repeatedly emphasizes that the rate is asymptotic. In causal-network inference, increasing $r$30 stabilizes slopes, while noise and high dimensionality can obscure the linear region (Usta et al., 2024). In language-model applications, finite $r$31 constrains the usable range of $r$32, and the estimator is standardized by fitting only scales with $r$33, with $r$34 by default (Du et al., 24 Oct 2025). In Fisher–Rao language analysis, the scaling region is selected by maximizing linear-fit $r$35 and ensuring several decades of scale (Du et al., 2024).

A standard misconception is to treat any apparent slope as intrinsic dimension. The binary-data literature shows that unnormalized fractal dimension is hard to interpret and is strongly shaped by sparsity; normalized fractal dimension was introduced precisely because raw slopes can be misleading in sparse $r$36 settings (Tatti et al., 2019). Another misconception is that a single global exponent captures heterogeneous data. “On the Estimation of Pointwise Dimension” calls this “dimension blindness”: for mixtures, correlation dimension collapses to the minimum local dimension, so the global rate can miss heterogeneity (Hidaka et al., 2013).

Data corruption can also manufacture or destroy scaling. In gappy time series, the estimated $r$37 is strongly biased when the mean gap size $r$38 is near the delay $r$39, and no clean saturation may appear; cubic spline interpolation can spuriously produce finite saturated $r$40 even for Gaussian noise or a non-chaotic light curve (George et al., 2014). The reported example is SS Cygni, a known non-chaotic light curve whose 5-day binned series yields a spurious $r$41 after cubic-spline interpolation (George et al., 2014).

A broader synthesis is therefore warranted. The correlation dimension rate is best regarded as a scale-local geometric exponent that requires a demonstrable scaling regime, an appropriate metric representation, and careful separation from other notions of “rate.” In most of the literature, it is the slope $r$42. In some recent work, especially on LLMs, derivatives such as $r$43 or $r$44 are introduced as secondary rates. The two notions are related but not identical (Du et al., 24 Oct 2025).