Connectivity Dimension in Graphs & Networks

• Connectivity dimension is a family of measures that quantifies how connectivity scales, resolves structure, or constrains global behavior across diverse fields.
• It is applied to determine minimum resolving sets in graphs, infer Euclidean embedding dimensions from connectivity distances, and estimate low-rank parameters in network analyses.
• It also appears as scaling exponents in wireless models and as dimension-shifted bounds in motivic and tropical geometry, linking theory with practical applications.

Connectivity dimension is not a single standardized invariant across contemporary research. In graph theory it denotes the minimum size of a resolving set built from local vertex-connectivity profiles; in network geometry and statistical network analysis it denotes either the smallest Euclidean embedding dimension compatible with connectivity-induced distances or the rank of a latent connectivity matrix; in network physics it appears as an exponent or effective dimension governing connectivity thresholds and scaling laws; and in motivic and tropical geometry it appears as the ambient quantity that shifts or bounds connectivity statements (Gottwald et al., 12 Aug 2025, Grindrod et al., 2023, Jiang et al., 17 Aug 2025, Coon et al., 2016, Druzhinin, 2019, Maclagan et al., 2019).

1. Terminological scope

The common thread is that “dimension” is used to quantify how connectivity information scales, resolves structure, or constrains global behavior. What varies from field to field is the underlying object: a finite graph, a weighted or sampled network, a stochastic spatial model, or a geometric or homotopical category.

Context	Quantity called or serving as connectivity dimension	Role
Graph theory	$\cdim(G)$	Resolving-set invariant from local connectivity profiles
Network embedding and statistics	Euclidean embedding dimension or matrix rank	Parsimonious representation of connectivity structure
Wireless and stochastic models	$\mathcal C=d/\eta$, stochastic dimension $4$, effective $d$, Lifshitz dimension $d_\mathrm{L}$	Threshold and scaling parameter
Motivic and tropical geometry	$(-d)$-connectivity shifts, $(d-l)$-connectivity through codimension one	Dimension-controlled connectivity bounds

A plausible implication is that the phrase functions less as a universal definition than as a family of dimension-sensitive connectivity formalisms. The cited literature does not collapse these into one invariant; instead, it develops domain-specific versions suited to different notions of connectivity.

2. Graph-theoretic connectivity dimension

In graph theory, the most explicit formalization is the invariant introduced for a finite, simple, undirected graph $G$. For vertices $v,w\in V(G)$, the notation $\kappa(v,w)$ denotes the number of independent paths between $\mathcal C=d/\eta$0 and $\mathcal C=d/\eta$1, meaning internally vertex-disjoint $\mathcal C=d/\eta$2-$\mathcal C=d/\eta$3 paths, with the convention $\mathcal C=d/\eta$4. For an ordered vertex set $\mathcal C=d/\eta$5, the connectivity representation of $\mathcal C=d/\eta$6 with respect to $\mathcal C=d/\eta$7 is

$\mathcal C=d/\eta$8

A set $\mathcal C=d/\eta$9 is resolving if distinct vertices have distinct representations, and a minimum-cardinality resolving set is a basis. Its size is the connectivity dimension $4$0 (Gottwald et al., 12 Aug 2025).

This definition is modeled directly on metric dimension, but replaces distance profiles by connectivity profiles. The parameter therefore measures the heterogeneity of a graph’s local connectivity structure rather than its metric geometry. The paper emphasizes that small $4$1 corresponds to strong variation in local connectivity, whereas large $4$2 reflects uniformity (Gottwald et al., 12 Aug 2025).

Several extremal results are known. For a connected graph $4$3, $4$4 holds if and only if $4$5. At the opposite extreme, a connected graph of order $4$6 satisfies $4$7 if and only if it is uniformly $4$8-connected, meaning every pair of distinct vertices has exactly $4$9 independent paths between them. The paper also proves a degree bound

$d$0

where $d$1 is the maximum degree, and a block-structure bound

$d$2

for every connected graph with at least two vertices (Gottwald et al., 12 Aug 2025).

Threshold graphs provide explicit realizations of prescribed non-extremal values. If a connected threshold graph is generated by

$d$3

then

$d$4

This yields infinitely many graphs for every prescribed $d$5 (Gottwald et al., 12 Aug 2025).

The invariant is also computationally difficult. The decision problem “given $d$6 and $d$7, decide whether $d$8” is NP-complete via a reduction from 3-SAT. Membership in NP follows because the values $d$9 can be verified by standard maximum-flow methods (Gottwald et al., 12 Aug 2025).

A recurrent misconception is that connectivity dimension should behave similarly to metric dimension. The paper shows that this is false in both directions: for every $d_\mathrm{L}$0, there exist graphs with $d_\mathrm{L}$1 and graphs with $d_\mathrm{L}$2. In particular, the path $d_\mathrm{L}$3 satisfies $d_\mathrm{L}$4 but $d_\mathrm{L}$5 (Gottwald et al., 12 Aug 2025).

3. Connectivity-induced embedding dimension and statistical estimation

A second usage appears in network geometry, where connectivity dimension is the smallest Euclidean embedding dimension in which nodes can be placed so that pairwise distances reflect the connectivity structure. For weighted networks, this is treated as an intrinsic embedding dimension suggested by the edge-weight geometry; for unweighted networks, it becomes an effective dimension recovered after constructing an embedding proxy (Grindrod et al., 2023).

The principal estimator in this setting adapts the twoNN method. For each node $d_\mathrm{L}$6, one defines the nearest-neighbor and second-nearest-neighbor distances $d_\mathrm{L}$7 and $d_\mathrm{L}$8, together with the ratio

$d_\mathrm{L}$9

Under locally uniform sampling in a $(-d)$0-dimensional setting, the cumulative distribution satisfies

$(-d)$1

After sorting the ratios, the pointwise estimators are

$(-d)$2

and the final estimate $(-d)$3 is the mean of $(-d)$4 over the stable middle range $(-d)$5. For weighted graphs this can be applied directly to the dissimilarity matrix. For unweighted graphs the proposed procedure is: spectrally embed the graph into trial dimension $(-d)$6, apply twoNN to the embedded coordinates, repeat for increasing $(-d)$7, and select the dimension where the estimate stabilizes or forms a plateau (Grindrod et al., 2023).

This estimator is explicitly contrasted with the common practice of visually searching for a gap in the Laplacian spectrum. The paper argues that spectral gaps may be unclear, absent, or noise-sensitive, whereas the twoNN-based estimate is more direct and robust in such cases (Grindrod et al., 2023).

A related but distinct line of work estimates intrinsic dimension from local connectivity occupancy rather than distance ratios. In the eDCF framework, adjacency is defined on a gridded Moore neighborhood in $(-d)$8, where each point has $(-d)$9 possible immediate neighbors. For a gridded set $(d-l)$0, the Connectivity Factor is

$(d-l)$1

The paper derives the embedding relation

$(d-l)$2

introduces Information Percentage to choose the grid scale, and then replaces hard dimension assignment by empirical triangular weighting. On synthetic benchmarks with noisy samples, the method achieves comparable mean absolute error to leading estimators and higher exact intrinsic-dimension match rates, reaching $(d-l)$3 versus $(d-l)$4 for MLE and $(d-l)$5 for TWO-NN under the reported high-noise, large-sample setting (Gupta et al., 18 Oct 2025).

In statistical network analysis, dimensionality is also identified with the rank of a community-level connectivity matrix. For a sample of network layers with shared memberships, the proposed estimator solves

$(d-l)$6

so that the nuclear norm acts as a convex surrogate for rank. The optimization is implemented by ADMM with a singular-value soft-thresholding step, and the number of surviving singular values provides the estimated dimension $(d-l)$7. The paper proves Frobenius- and nuclear-norm error bounds, shows improvement over averaging when perfect membership recovery holds and the true rank is much smaller than the number of communities, and uses a primate brain dataset to show that posited connectivity need not be full rank in practice (Jiang et al., 17 Aug 2025).

4. Dimension as a threshold or scaling parameter in network and stochastic models

In wireless network theory, a dimension-like exponent appears explicitly as the connectivity exponent

$(d-l)$8

the ratio of the dimension of the network domain to the path-loss exponent. The local connectivity mass $(d-l)$9 then scales as $G$0 in parameters such as transmit power or antenna-related gains; in particular,

$G$1

The interpretation given in the paper is that $G$2 quantifies the balance between geometric opportunity and propagation attenuation, so it acts as a connectivity dimension in the sense of a universal scaling exponent (Coon et al., 2016).

In random distance graphs on the lattice

$G$3

with edge probability

$G$4

the ambient lattice dimension $G$5 determines the connectivity threshold. The critical exponent is

$G$6

If $G$7, the graph is connected asymptotically almost surely; if $G$8, it is disconnected asymptotically almost surely and has $G$9 isolated vertices (Flynn et al., 2015).

In branching interlacements on $v,w\in V(G)$0, the relevant notion is stochastic dimension. The paper proves that the basic relation induced by a conditioned critical branching random walk has stochastic dimension $v,w\in V(G)$1. Consequently, every two vertices visited by the branching interlacement are connected by at most

$v,w\in V(G)$2

underlying branching trajectories, and this bound is sharp. The same dimension $v,w\in V(G)$3 yields the threshold for pairwise intersection of branching random walks: two such traces intersect if and only if $v,w\in V(G)$4 (Procaccia et al., 2016).

A spectral version of connectivity dimension appears in linear network dynamics. If the density of nearly critical eigenvalues satisfies

$v,w\in V(G)$5

the exponent $v,w\in V(G)$6 is interpreted as an effective spatial dimension of the connectivity eigenmodes. It governs long-time autoresponse and autocorrelation scaling,

$v,w\in V(G)$7

and produces a principal-component spectrum with exponent $v,w\in V(G)$8 in the equal-time covariance case. The paper further identifies a transition between high- and low-dimensional activity at $v,w\in V(G)$9 (Tiberi et al., 2023).

Hierarchical-modular brain networks provide a contrasting case in which the usual spectral dimension is not defined. Instead of a low-end density of states with a power law, the paper finds Lifshitz tails and introduces the anomalous Lifshitz dimension $\kappa(v,w)$0. The resulting return probability obeys

$\kappa(v,w)$1

so connectivity-related slowing is controlled by $\kappa(v,w)$2 rather than a classical spectral dimension (Esfandiary et al., 2020).

5. Dimension-shifted connectivity in motivic and tropical geometry

In motivic homotopy theory, “connectivity dimension” arises through dimension-dependent bounds on $\kappa(v,w)$3-localization. Over a field, Morel’s stable connectivity theorem states that if $\kappa(v,w)$4 is a connective $\kappa(v,w)$5-spectrum, then $\kappa(v,w)$6 remains connective. Over a general base scheme $\kappa(v,w)$7, this fails in that form. The corrected statement is that if $\kappa(v,w)$8 and $\kappa(v,w)$9 is connective, then $\mathcal C=d/\eta$00 should be $\mathcal C=d/\eta$01-connective. For arbitrary base schemes of finite Krull dimension, the paper proves precisely this base-scheme version without residue-field hypotheses, and concludes that for a smooth scheme $\mathcal C=d/\eta$02 over $\mathcal C=d/\eta$03,

$\mathcal C=d/\eta$04

(Druzhinin, 2019).

An unstable analogue replaces Krull dimension by valuative dimension. If $\mathcal C=d/\eta$05 is qcqs of valuative dimension $\mathcal C=d/\eta$06 and $\mathcal C=d/\eta$07 is an $\mathcal C=d/\eta$08-connective pointed Nisnevich space, then its motivic localization is $\mathcal C=d/\eta$09-connective. The paper introduces perverse $\mathcal C=d/\eta$10-connectivity, defined by the requirement that for every local essentially smooth $\mathcal C=d/\eta$11-scheme $\mathcal C=d/\eta$12 of valuative dimension $\mathcal C=d/\eta$13, the henselization $\mathcal C=d/\eta$14 satisfies $\mathcal C=d/\eta$15 being $\mathcal C=d/\eta$16-connective. This unstable theorem yields a stable version, recovers earlier results over fields and noetherian bases, extends them to possibly non-noetherian qcqs schemes, and implies convergence of the slice filtration on homotopy $\mathcal C=d/\eta$17-theory for qcqs schemes of finite valuative dimension (Bouis et al., 11 Dec 2025).

The same general theme appears in the theory of motives. Using the homotopy $\mathcal C=d/\eta$18-structure on $\mathcal C=d/\eta$19, motivic connectivity is detected by Chow-weight homology. The basic criterion is

$\mathcal C=d/\eta$20

More generally, vanishing ranges of Chow-weight homology detect both effectivity and weight bounds, so connectivity is expressed through precise homological disappearance conditions (Bondarko et al., 2014).

In tropical geometry, higher connectivity is controlled by geometric dimension minus lineality. If $\mathcal C=d/\eta$21 is an irreducible $\mathcal C=d/\eta$22-dimensional variety over a characteristic-$\mathcal C=d/\eta$23 field and $\mathcal C=d/\eta$24 is a pure $\mathcal C=d/\eta$25-dimensional rational polyhedral complex with support $\mathcal C=d/\eta$26, then $\mathcal C=d/\eta$27 is $\mathcal C=d/\eta$28-connected through codimension one, where $\mathcal C=d/\eta$29 is the dimension of the lineality space. The paper also proves that the intersection of $\mathcal C=d/\eta$30 with a generic rational affine hyperplane is again the tropicalization of an irreducible variety, and deduces that the $\mathcal C=d/\eta$31-dimensional skeleton of the normal fan of a rational full-dimensional polytope is $\mathcal C=d/\eta$32-connected through codimension one (Maclagan et al., 2019).

6. Comparative perspective and recurrent misconceptions

The surveyed literature shows that connectivity dimension is a genuinely plural concept. The graph-theoretic invariant $\mathcal C=d/\eta$33 is a discrete resolving-set parameter; the network-geometric version is an embedding dimension inferred from connectivity-preserving distances; the statistical-network version is a low-rank parameter of a connectivity matrix; the wireless and stochastic versions are scaling exponents or stochastic dimensions; and the motivic and tropical versions are dimension-controlled connectivity bounds rather than independent invariants (Gottwald et al., 12 Aug 2025, Grindrod et al., 2023, Jiang et al., 17 Aug 2025, Coon et al., 2016, Druzhinin, 2019, Maclagan et al., 2019).

Several misconceptions recur across these settings. First, connectivity dimension is not generally interchangeable with metric dimension: the graph-theoretic literature gives explicit families where the two differ by arbitrarily large factors, and even the path $\mathcal C=d/\eta$34 separates them maximally in one direction (Gottwald et al., 12 Aug 2025). Second, spectral heuristics are not always reliable proxies for connectivity-induced dimension: in network embedding, the twoNN paper argues that visually identifying a Laplacian spectral gap may be ambiguous or misleading, especially under noise (Grindrod et al., 2023). Third, full-rank connectivity is not a universal default: the low-rank estimation paper shows that community-level connectivity matrices in multilayer data may be strictly lower rank than the number of communities (Jiang et al., 17 Aug 2025). Fourth, perfect connectivity preservation under localization is not universal in motivic homotopy theory: over general bases, the correct statement is dimension-shifted $\mathcal C=d/\eta$35-connectivity rather than $\mathcal C=d/\eta$36-connectivity (Druzhinin, 2019, Bouis et al., 11 Dec 2025).

A plausible synthesis is that the phrase “connectivity dimension” is most useful when read relationally: it identifies the numerical parameter that determines how connectivity distinguishes vertices, supports embeddings, scales with resources, survives localization, or propagates through a category. The parameter itself may be a cardinality, an embedding dimension, a matrix rank, a ratio such as $\mathcal C=d/\eta$37, a stochastic dimension, or a geometric dimension corrected by lineality or base dimension.