Correlation Dimension¶

A fractal-dimension statistic given by the small-radius scaling exponent of the probability or normalized count that two sampled points lie within distance epsilon.

Version: v3 · 2026-09-06 · History
Domain-specific #: 1567
Origin domain: chaos theory
Subdomain: strange-attractor characterization
Aliases: Correlation exponent, Grassberger–Procaccia dimension

Core Idea¶

The correlation dimension is a fractal dimension defined by how the probability that two independently sampled points lie within distance r scales as r becomes small. For a probability measure μ on a metric space, define a correlation integral.

C(r) = ∫ μ(B(x,r)) dμ(x) = P(d(X,Y) < r)

for independent X,Y ~ μ, subject to boundary and convention details. When a stable limit exists,

D_2 = lim_{r→0} log C(r) / log r.

For finite data, the Grassberger–Procaccia estimator counts close pairs over a range of radii and estimates the slope of log C_N(r) against log r in a defensible scaling region.^[1]^[2]

The abstraction is not the plotted slope alone. It comprises a measure or sampled attractor, a metric and embedding, close-pair counting, asymptotic scaling, and diagnostics that separate genuine geometry from finite resolution, noise, and temporal dependence.

Structural Signature¶

Sig role-phrases:

the invariant or sampling measure μ — the distribution of points whose local concentration is studied
the metric space and distance d — the geometry in which closeness is defined
the radius r — the observation scale
the close-pair probability C(r) — mass-weighted probability of two points lying within r
the power-law regime — scales over which C(r) ≈ c r^{D_2}
the logarithmic slope — the estimated exponent
the finite sample and pair counter — empirical approximation to C(r)
the embedding construction — when scalar time-series observations are reconstructed into state vectors
the dependence exclusion — treatment such as a Theiler window preventing temporal neighbors from dominating pair counts
the stability checks — robustness across radii, sample sizes, norms, embeddings, and noise levels

Recognition test. A claimed correlation dimension must arise from close-pair scaling and exhibit a nontrivial slope plateau or convergence argument. A regression through any visually linear portion of a log–log graph, without excluding resolution limits or serial correlation, is not sufficient.

What It Is Not¶

Not correlation coefficient. It does not measure linear association between two variables.
Not correlation length. It is an exponent of pair concentration, not a characteristic spatial scale.
Not automatically Hausdorff dimension. Relationships among dimension notions require hypotheses and may be inequalities rather than equality.
Not box-counting dimension. Box counts weight occupied boxes uniformly; correlation sums weight dense regions more strongly.
Not embedding dimension. The reconstructed coordinate dimension is chosen by the analyst; the estimated attractor dimension may be smaller.
Not proof of deterministic chaos. Noise and stochastic processes can produce apparent scaling.
Not a reliable number from arbitrarily short data. Pair counts are numerous but dependent, and usable small scales shrink under noise and finite precision.

Scope of Application¶

Correlation dimension is used in nonlinear dynamics, fractal geometry, experimental time-series analysis, recurrence analysis, and characterization of invariant measures. Its computational accessibility made it a prominent tool for assessing low-dimensional attractor hypotheses from observations.^[1]

For points sampled independently from a smooth k-dimensional region with regular density, close-pair probability typically scales like r^k away from boundaries. Multifractal measures can yield a correlation dimension reflecting order-two mass concentration and differing from support dimensions.

Time-series use requires reconstruction choices: delay, embedding dimension, sampling interval, stationarity window, and exclusion of temporally adjacent points. Measurement noise often drives the smallest-scale slope toward the ambient dimension, while finite attractor size flattens the largest scales. Only an intermediate regime can carry the intended exponent.

Clarity¶

A common empirical estimator for N distinct sample indices is

C_N(r) = [2/(N(N−1))] Σ_{i<j} 1{d(x_i,x_j)<r}.

Some conventions include ordered pairs or self-pairs with a different normalization. Self-pairs are dangerous because their zero distance dominates as r shrinks. For trajectories, pairs with |i−j| below a chosen exclusion lag should often be omitted.

The local slope d log C_N(r)/d log r is inspected over scale. A report should give the estimator, norm, sample count, radius interval, fit uncertainty, temporal-exclusion rule, and behavior across embedding dimensions. A single slope with no interval is not reproducible.

Manages Complexity¶

A strange attractor may occupy a complicated subset of a high-dimensional phase space. Correlation dimension compresses its mass distribution into one exponent that can distinguish a line-like, surface-like, or fractional concentration pattern. Pair counting avoids explicit construction of an optimal cover.

That efficiency trades away spatial detail. Different measures can share D_2; local dimensions may vary; topology and dynamical equations are not recovered. The exponent is best used as one diagnostic alongside recurrence, Lyapunov, surrogate-data, and model checks.

Abstract Reasoning¶

Start from probability, not the graph. Interpret C(r) as close-pair mass before fitting a line.

Locate scale boundaries. Reject radii dominated by instrument resolution, noise, self-pairs, or finite diameter.

Check slope stability. A candidate exponent should persist over a meaningful interval and across increasing data size.

Vary embedding dimension. Saturation of the estimated exponent can support, but does not prove, a low-dimensional attractor.

Control dependence. Temporally close samples may be spatially close simply because trajectories are continuous.

Preserve the measure. Correlation dimension describes how mass is distributed, not only the geometric support.

Knowledge Transfer¶

The portable skeleton is scale exponent from pairwise neighborhoods: count how rapidly the probability of local coincidence grows with radius and infer an effective dimension. This can organize analyses of point clouds, spatial events, and representation manifolds.

Literal transfer requires a metric, a sampling measure, and a defensible asymptotic or scaling regime. A log–log slope in network statistics or market data is only analogous until those elements and dependence controls are supplied.

The broader lesson is that scale laws require an interval of validity and adversarial checks against artifacts, not only a high regression fit.

Examples¶

Canonical: uniform points on a planar region¶

For many points sampled nearly uniformly from the interior of a two-dimensional region, the mass of a small radius-r ball is proportional to its area, about r². Averaging over centers gives C(r) ∝ r² away from boundaries, so the correlation dimension is 2.

Mapped back: the planar distribution is the measure; Euclidean distance supplies the metric; close pairs supply C(r); the small-scale area law gives the power law; and slope 2 gives the dimension.

Applied / In Practice: reconstructed laboratory signal¶

A scalar sensor record is delay-embedded at several coordinate dimensions. The analyst excludes temporally adjacent vectors, computes pair sums, and finds a slope near 2.4 over the same intermediate radii for embeddings 5 through 8. The plateau is reported as evidence consistent with a low-dimensional invariant measure, while surrogate stochastic series and longer recordings are used to test alternatives.

Mapped back: delay vectors form the point cloud; the exclusion window controls serial dependence; the shared interval is the scaling regime; saturation is a stability diagnostic; and cautious wording respects non-identification of chaos.

Structural Tensions¶

T1: Small-scale limit vs finite resolution. The definition sends r to zero, while data fail at the smallest scales. Diagnostic: Is there an expanding intermediate regime as data improve?

T2: More pairs vs dependent pairs. N² comparisons look abundant, but share points and temporal structure. Diagnostic: Are uncertainty and exclusion rules dependency-aware?

T3: Geometric support vs mass concentration. Dense regions dominate correlation sums. Diagnostic: Is D_2 being mislabeled as a support dimension?

T4: Scaling fit vs mechanistic inference. A straight segment can arise from noise or crossover. Diagnostic: Do surrogates and alternate embeddings reproduce or refute it?

T5: Dimension saturation vs data insufficiency. A plateau across embeddings is suggestive, not conclusive. Diagnostic: Does it also persist across sample sizes and radius windows?

T6: Scalar summary vs heterogeneous structure. One exponent hides multifractality. Diagnostic: Do local slopes or other Rényi dimensions vary materially?

T7: Domain autonomy vs prime reduction. Dimension, Metric, and Scaling Law explain the skeleton. Diagnostic: Do close-pair probability, order-two mass weighting, and time-series safeguards still prescribe distinctive computation? If so, the domain node is nonredundant.

Structural–Framed Character¶

The five-criterion aggregate is 0.05 (structural). The abstraction is formal, nonevaluative, and portable across sampled metric measures. Its identity is recognized from the pair-scaling invariant rather than imported from a social or institutional frame.

Structural Core vs. Domain Accent¶

Structural core: estimate an exponent from how pairwise-neighborhood probability changes with scale.

Domain accent: invariant measures and strange attractors, the Grassberger–Procaccia correlation sum, delay embeddings, Theiler exclusions, and interpretation as the order-two fractal dimension.

Generalization yields a pair-scaling statistic. The accent yields correlation dimension with its established inferential discipline.

Dimension is instantiated as the sought scaling exponent. Allometry and Scaling Law is presupposed through C(r) ∝ r^{D_2}. Metric determines pairwise neighborhoods. Fractal Geometry supplies the broader family of noninteger dimension concepts. None alone contains the measure-weighted close-pair estimator.

Relationships to Other Abstractions¶

Current abstraction Correlation Dimension Domain-specific

Parents (3) — more general patterns this builds on

Correlation Dimension is a kind of Dimension Prime

Dimension is instantiated as the sought scaling exponent.
Correlation Dimension presupposes Allometry and Scaling Law Prime

Dimension is instantiated as the sought scaling exponent.
Correlation Dimension presupposes Metric Prime

Dimension is instantiated as the sought scaling exponent.

Hierarchy paths (3) — routes to 3 parentless roots

Correlation Dimension → Dimension

Show alternative paths (2)

Neighborhood in Abstraction Space¶

Correlation Dimension sits in a sparse region of the domain-specific corpus (70^th percentile for distinctiveness): few abstractions share its structure, so a faithful description tends to retrieve it precisely.

Family — Unclustered & Miscellaneous (1565 abstractions)

Nearest neighbors

Empirical Measure — 0.88
Benford's Law — 0.85
Variogram — 0.84
Particle Filter — 0.83
Least-Squares Adjustment — 0.83

Computed from structural-signature embeddings · 2026-09-08

Not to Be Confused With¶

Pearson, rank, or autocorrelation coefficients
correlation length or correlation time
Hausdorff, box-counting, or information dimension
embedding dimension
recurrence rate at one fixed threshold
a generic slope on log–log axes
proof that a process is deterministic or chaotic

References¶

[1] Peter Grassberger and Itamar Procaccia, “Characterization of Strange Attractors”, Physical Review Letters 50 (1983), 346–349. Introduces the correlation-integral approach to attractor dimension. registry ↩a ↩b

[2] Peter Grassberger and Itamar Procaccia, “Measuring the Strangeness of Strange Attractors”, Physica D 9(1–2), 1983, 189–208; accessible manuscript. Develops the estimator, scaling interpretation, and comparisons with other dimensions. registry ↩