Clustering Exponents: Theory & Applications

• Clustering exponents are scaling parameters that control the formation, geometry, and error propagation in diverse clustering methods.
• They are applied across fuzzy clustering, percolation theory, and dynamical systems to optimize model performance and interpret critical transitions.
• Examples include tuning the fuzzifier in FCM, defining cluster-size power laws, and deriving exponential complexity bounds in algorithmic clustering.

Across the cited literature, clustering exponents denote several non-equivalent but structurally analogous quantities: exponent parameters that tune fuzziness in soft partitions, critical exponents that govern cluster-size statistics and clustering spectra, Lyapunov-type exponents that quantify spatial concentration in dynamical systems, and statistical or computational exponents that describe the reliability or complexity of clustering procedures. What unifies these usages is that an exponent controls either the formation of clusters, the geometry of clustered states, or the scaling of clustering error, cost, or critical observables (Bora et al., 2014, Cho et al., 2015, Esmaily-Moghadam et al., 2015, Amid et al., 2022).

1. Fuzzifier exponents in soft clustering

In fuzzy $c$-means (FCM), the clustering exponent is the fuzzifier $m$, the exponent applied to the partition matrix $U=[u_{ij}]$. The standard FCM objective is

$$J_m(U,V)=\sum_{i=1}^{c}\sum_{j=1}^{N} u_{ij}^{\,m}\,\|x_j-c_i\|^2,$$

with $m>1$, Euclidean distance, and iterative updates

$$c_i=\frac{\sum_{j=1}^{N}u_{ij}^m x_j}{\sum_{j=1}^{N}u_{ij}^m}, \qquad u_{ij}=\frac{1}{\sum_{k=1}^{c}\left(\frac{\|x_j-c_i\|}{\|x_j-c_k\|}\right)^{\frac{2}{m-1}} }.$$

The factor $2/(m-1)$ is the mechanism by which $m$ controls membership sharpness: as $m\to 1^+$, the partition approaches hard clustering; as $m\to\infty$, memberships become more diffuse. The algorithm converges when $m$0, and the stated time complexity is $m$1 (Bora et al., 2014).

The experimental analysis on the Iris dataset fixes $m$2, maximum iterations $m$3, and minimum improvement factor $m$4, and varies $m$5. Over this range, the final objective decreases monotonically from $m$6 at $m$7 to $m$8 at $m$9, while CPU time decreases from $U=[u_{ij}]$0 s to $U=[u_{ij}]$1 s; the paper also states that the maximum iteration count decreases and is minimal near $U=[u_{ij}]$2. For that dataset and protocol, the reported best performance is therefore at $U=[u_{ij}]$3, while also emphasizing that no universally optimal $U=[u_{ij}]$4 is known and that the choice depends on the “property of dataset” (Bora et al., 2014).

A more localized notion of clustering exponent appears in hedge-algebra FCM. There, the scalar fuzzifier is replaced by a matrix-valued exponent field $U=[u_{ij}]$5, one exponent per data–cluster pair. The construction begins with the relative distance

$U=[u_{ij}]$6

then maps $U=[u_{ij}]$7 linearly to

$U=[u_{ij}]$8

A hedge-algebra reliability score

$U=[u_{ij}]$9

is used to update $$J_m(U,V)=\sum_{i=1}^{c}\sum_{j=1}^{N} u_{ij}^{\,m}\,\|x_j-c_i\|^2,$$0 iteratively, and the FCM objective becomes

$$J_m(U,V)=\sum_{i=1}^{c}\sum_{j=1}^{N} u_{ij}^{\,m}\,\|x_j-c_i\|^2,$$1

This produces sharper memberships near centers and softer memberships near boundaries. On Iris, HAmFCM reports $$J_m(U,V)=\sum_{i=1}^{c}\sum_{j=1}^{N} u_{ij}^{\,m}\,\|x_j-c_i\|^2,$$2 accuracy against $$J_m(U,V)=\sum_{i=1}^{c}\sum_{j=1}^{N} u_{ij}^{\,m}\,\|x_j-c_i\|^2,$$3 for FCM; on Wine, $$J_m(U,V)=\sum_{i=1}^{c}\sum_{j=1}^{N} u_{ij}^{\,m}\,\|x_j-c_i\|^2,$$4 against $$J_m(U,V)=\sum_{i=1}^{c}\sum_{j=1}^{N} u_{ij}^{\,m}\,\|x_j-c_i\|^2,$$5. The same study reports qualitative improvements on color image segmentation, especially for boundary-sensitive regions (Le et al., 2017).

2. Critical exponents of cluster populations and clustering spectra

In percolation and aggregation, clustering exponents are critical exponents attached to cluster-size distributions. In the restricted Erdős–Rényi cluster-merging process, the central observable is

$$J_m(U,V)=\sum_{i=1}^{c}\sum_{j=1}^{N} u_{ij}^{\,m}\,\|x_j-c_i\|^2,$$6

At the hybrid percolation transition $$J_m(U,V)=\sum_{i=1}^{c}\sum_{j=1}^{N} u_{ij}^{\,m}\,\|x_j-c_i\|^2,$$7, finite clusters obey

$$J_m(U,V)=\sum_{i=1}^{c}\sum_{j=1}^{N} u_{ij}^{\,m}\,\|x_j-c_i\|^2,$$8

where $$J_m(U,V)=\sum_{i=1}^{c}\sum_{j=1}^{N} u_{ij}^{\,m}\,\|x_j-c_i\|^2,$$9 is the cluster-size exponent. The paper shows that $m>1$0 varies continuously with the control parameter $m>1$1 and satisfies

$m>1$2

with numerical estimates $m>1$3 at $m>1$4, $m>1$5 at $m>1$6, $m>1$7 at $m>1$8, and $m>1$9 as $$c_i=\frac{\sum_{j=1}^{N}u_{ij}^m x_j}{\sum_{j=1}^{N}u_{ij}^m}, \qquad u_{ij}=\frac{1}{\sum_{k=1}^{c}\left(\frac{\|x_j-c_i\|}{\|x_j-c_k\|}\right)^{\frac{2}{m-1}} }.$$0. Post-transition critical exponents are tied directly to $$c_i=\frac{\sum_{j=1}^{N}u_{ij}^m x_j}{\sum_{j=1}^{N}u_{ij}^m}, \qquad u_{ij}=\frac{1}{\sum_{k=1}^{c}\left(\frac{\|x_j-c_i\|}{\|x_j-c_k\|}\right)^{\frac{2}{m-1}} }.$$1 through

$$c_i=\frac{\sum_{j=1}^{N}u_{ij}^m x_j}{\sum_{j=1}^{N}u_{ij}^m}, \qquad u_{ij}=\frac{1}{\sum_{k=1}^{c}\left(\frac{\|x_j-c_i\|}{\|x_j-c_k\|}\right)^{\frac{2}{m-1}} }.$$2

The same work proposes that a necessary condition for hybrid transitions in cluster-merging models is a power-law finite-cluster distribution with exponent in the restricted range $$c_i=\frac{\sum_{j=1}^{N}u_{ij}^m x_j}{\sum_{j=1}^{N}u_{ij}^m}, \qquad u_{ij}=\frac{1}{\sum_{k=1}^{c}\left(\frac{\|x_j-c_i\|}{\|x_j-c_k\|}\right)^{\frac{2}{m-1}} }.$$3 (Cho et al., 2015).

A distinct but related exponent appears in multilayer random graphs through the clustering spectrum

$$c_i=\frac{\sum_{j=1}^{N}u_{ij}^m x_j}{\sum_{j=1}^{N}u_{ij}^m}, \qquad u_{ij}=\frac{1}{\sum_{k=1}^{c}\left(\frac{\|x_j-c_i\|}{\|x_j-c_k\|}\right)^{\frac{2}{m-1}} }.$$4

the degree-dependent local clustering coefficient. For overlays of Bernoulli random graph layers with power-law layer-size distribution

$$c_i=\frac{\sum_{j=1}^{N}u_{ij}^m x_j}{\sum_{j=1}^{N}u_{ij}^m}, \qquad u_{ij}=\frac{1}{\sum_{k=1}^{c}\left(\frac{\|x_j-c_i\|}{\|x_j-c_k\|}\right)^{\frac{2}{m-1}} }.$$5

the limiting degree distribution has exponent

$$c_i=\frac{\sum_{j=1}^{N}u_{ij}^m x_j}{\sum_{j=1}^{N}u_{ij}^m}, \qquad u_{ij}=\frac{1}{\sum_{k=1}^{c}\left(\frac{\|x_j-c_i\|}{\|x_j-c_k\|}\right)^{\frac{2}{m-1}} }.$$6

whereas the clustering-spectrum exponent is

$$c_i=\frac{\sum_{j=1}^{N}u_{ij}^m x_j}{\sum_{j=1}^{N}u_{ij}^m}, \qquad u_{ij}=\frac{1}{\sum_{k=1}^{c}\left(\frac{\|x_j-c_i\|}{\|x_j-c_k\|}\right)^{\frac{2}{m-1}} }.$$7

For $$c_i=\frac{\sum_{j=1}^{N}u_{ij}^m x_j}{\sum_{j=1}^{N}u_{ij}^m}, \qquad u_{ij}=\frac{1}{\sum_{k=1}^{c}\left(\frac{\|x_j-c_i\|}{\|x_j-c_k\|}\right)^{\frac{2}{m-1}} }.$$8, the spectrum is asymptotically degree-independent, $$c_i=\frac{\sum_{j=1}^{N}u_{ij}^m x_j}{\sum_{j=1}^{N}u_{ij}^m}, \qquad u_{ij}=\frac{1}{\sum_{k=1}^{c}\left(\frac{\|x_j-c_i\|}{\|x_j-c_k\|}\right)^{\frac{2}{m-1}} }.$$9. This explicitly separates a degree exponent controlled by both $2/(m-1)$0 and $2/(m-1)$1 from a clustering exponent controlled only by $2/(m-1)$2, showing partial decoupling between heavy-tailed degree statistics and the decay of degree-dependent clustering (Bloznelis et al., 2019).

These results use the same word, clustering, in two adjacent but distinct senses. In the r-ER model, the exponent describes the abundance of finite clusters at criticality; in the multilayer overlay model, it describes how local triangle closure decays with node degree. The papers jointly suggest that clustering exponents are often most informative when treated as scaling laws on observables, not as a single universal number (Cho et al., 2015, Bloznelis et al., 2019).

3. Geometric, transport, and non-linear clustering exponents

Percolation theory introduces another family of clustering exponents through transport on the incipient infinite cluster. In that setting, the paper distinguishes standard exponents $2/(m-1)$3 from walk-based exponents $2/(m-1)$4 and $2/(m-1)$5, and the spectral dimension

$2/(m-1)$6

The key relations are

$2/(m-1)$7

Numerically, $2/(m-1)$8 in 2D and $2/(m-1)$9 in 3D, leading to $m$0 and $m$1, with spectral dimensions $m$2 and $m$3. Here the relevant clustering exponents are the exponents that encode how transport probes the geometry of critical clusters (Siclen, 2016).

Planar random geometry uses generalized disconnection exponents

$m$4

which extend Brownian disconnection exponents from $m$5 to all $m$6. They are defined by

$m$7

and admit the explicit formula

$m$8

For $m$9, the exponents have a loop-soup interpretation and yield predictions for the Hausdorff dimension of multiple points on cluster boundaries. In particular, the predicted dimension of double points on loop-soup cluster boundaries is strictly positive for $m\to 1^+$0 and equals zero at the critical intensity $m\to 1^+$1 (Qian, 2019).

A cosmological usage appears in one-dimensional scale-free gravitational clustering, where the strongly non-linear two-point function obeys

$m\to 1^+$2

The paper derives a stable-clustering prediction

$m\to 1^+$3

with $m\to 1^+$4 the initial power-spectrum exponent and $m\to 1^+$5 the expansion parameter. Simulations then divide $m\to 1^+$6 space into two regions: one in which $m\to 1^+$7 agrees well with $m\to 1^+$8, and another in which $m\to 1^+$9 is approximately universal, $m\to\infty$0, with weak dependence on $m\to\infty$1 and $m\to\infty$2. The paper identifies the boundary empirically with a critical value $m\to\infty$3 (Benhaiem et al., 2012).

4. Dynamical, temporal, and Lyapunov clustering exponents

In inertial-particle turbulence, clustering exponents are finite-time Lyapunov exponents. For a small particle cloud, the principal-axis growth rates are $m\to\infty$4, and the cloud volume satisfies

$m\to\infty$5

The paper defines the clustering index

$m\to\infty$6

so that $m\to\infty$7 means net volume contraction. The central result is

$m\to\infty$8

where $m\to\infty$9 is the difference between Lagrangian strain-rate and rotation-rate spectra sampled along particle trajectories. For homogeneous isotropic turbulence, the analysis predicts maximum clustering at intermediate Stokes number, with $m$00 for $m$01, $m$02 for $m$03, and a maximum near $m$04 in the reported DNS (Esmaily-Moghadam et al., 2015).

A closely related but distinct Lyapunov-based notion appears for passive tracers in compressible random velocity fields. In one dimension, the Lyapunov exponent

$m$05

is always negative. The small-$m$06 expansion begins

$m$07

and Padé–Borel resummation remains accurate up to $m$08. At large Kubo number, the asymptotic estimate is

$m$09

In two dimensions, the sign of $m$10 depends on compressibility and Kubo number, and the small-$m$11 transition line obeys

$m$12

Here the clustering exponents determine whether particles form fractal clusters or enter a path-coalescing phase (Gustavsson et al., 2013).

In dielectric relaxation, the clustering exponent is neither a Lyapunov exponent nor a critical exponent but the low-frequency fractional exponent

$m$13

appearing in

$m$14

The frequency-domain response

$m$15

decouples the low-frequency clustering exponent $m$16 from the high-frequency stop–move exponent $m$17. The paper attributes $m$18 to the undershooting process $m$19, which generates a random partition, or clustering, of temporal changes (Stanislavsky et al., 2011).

5. Exponent parameters in probabilistic and information-geometric clustering

Model-based clustering uses exponent parameters to modify cluster shape, tail weight, and robustness. In mixtures of multivariate power exponential distributions, each component has density

$m$20

with $m$21 the shape parameter. The interpretation is explicit: $m$22 gives leptokurtic heavy-tailed components, $m$23 gives the Gaussian, $m$24 gives platykurtic light-tailed components, $m$25 yields the multivariate Laplace, and $m$26 tends to a multivariate uniform distribution on an ellipsoid. In the mixture

$m$27

the $m$28 therefore function as cluster-level exponents that separate tail behavior from orientation and scale. The paper combines these with eigen-decomposed scale matrices, derives a GEM algorithm, and reports that allowing $m$29 to vary improves clustering accuracy or parsimony in several simulations and benchmark datasets (Dang et al., 2015).

A different generalization appears in tempered exponential measures (TEMs), where the exponent parameter is $m$30. TEMs are defined by

$m$31

with co-density normalization

$m$32

The corresponding information-theoretic distortion satisfies

$m$33

where

$m$34

is a conformal Bregman divergence. The right population minimizer becomes a weighted average,

$m$35

and the paper proves bounded-influence robustness for $m$36 under growth conditions on $m$37. In this literature, the exponent $m$38 is a tempering parameter that directly deforms clustering geometry and centroid robustness (Amid et al., 2022).

6. Exponential constructions and collectiveness measures

Some clustering methods use exponentials not as fitted parameters but as the organizing device of the algorithm itself. In agglomerative clustering via path integrals, data are represented by a weighted directed graph $m$39, and the contribution of all walks is collected by the matrix exponential

$m$40

The $m$41 entry,

$m$42

is the path-integral descriptor of an edge. From it the set descriptor

$m$43

is defined as a normalized collectiveness measure. The paper proves $m$44, derives asymptotic growth

$m$45

and uses conditional versions of $m$46 to define an agglomerative affinity

$m$47

Because the coefficients $m$48 define an exponential generating function, the method emphasizes paths of length roughly comparable to the maximum out-degree $m$49, while retaining contributions from all lengths (Ren et al., 2015).

This usage broadens the notion of clustering exponents. The exponent is not estimated from data; rather, an exponential generating function regularizes multi-step connectivity and turns collectiveness into a bounded graph functional. A plausible implication is that “clustering exponent” in algorithmic graph settings may refer as much to the exponential weighting scheme as to a fitted critical exponent, provided the exponential controls how multi-scale connectivity enters the clustering criterion (Ren et al., 2015).

7. Estimation, exact complexity exponents, and error exponents

Finite-size methodology introduces yet another layer. For Monte Carlo estimates of cluster observables, the scaling ansatz

$m$50

assumes multiple observables $m$51 share the same set of exponents $m$52. The paper then constructs

$m$53

and shows that minima of $m$54 identify the exponents $m$55. Error bars are defined by

$m$56

Applied to uncorrelated percolation hulls, the method gives

$m$57

using sizes $m$58 to $m$59, in agreement with the exact $m$60, and also extracts correction exponents from the same data (Mandre et al., 2013).

Algorithmic complexity yields exact exponential clustering exponents in the runtime sense. For standard $m$61-Median and $m$62-Means on $m$63 points, the trivial exact algorithm is $m$64, while the paper gives the first non-trivial exact algorithm with runtime

$m$65

uniformly over all $m$66. For supplier variants, the paper gives $m$67 algorithms via subset convolution, and proves that under ETH there is no $m$68 exact algorithm for standard $m$69-Median/$m$70-Means, while under the Set Cover Conjecture there is no $m$71 exact algorithm for supplier versions (Fomin et al., 2022).

Finally, short-read clustering in DNA-storage models introduces a statistical error exponent

$m$72

under the scaling $m$73. Reads are produced by an unknown set of $m$74 source sequences through a memoryless channel $m$75, and the optimal rule is the MAP estimator of the canonical source-index pattern. The finite-length upper bound is expressed through Bhattacharyya quantities

$m$76

and the maximal pairwise Bhattacharyya coefficient among sources. Here the “clustering exponent” is the asymptotic rate at which the exact-recovery error of the optimal clusterer decays with read length (Chachamovitz et al., 19 Jun 2026).

Taken together, these literatures show that clustering exponents are not a single invariant but a family of scaling descriptors. Depending on context, they control fuzziness in a partition matrix, the tail of a critical cluster-size law, the decay of clustering spectrum with degree, the contraction of particle clouds, the shape of probabilistic clusters, the weighting of graph paths, the exact exponential complexity of a clustering problem, or the large-deviation rate of clustering error. The common role is to compress a multi-scale clustering phenomenon into a parameter or scaling law that remains stable under asymptotic analysis.