Why two networks with the same size can have different dimensions
A graph can have the same number of vertices and the same overall growth rate as another—and still behave as if it has a completely different dimension. In one example, a long path joined to a complete graph has a mass dimension of 1 but a uniform energy dimension of 0. The reason is not a change in how many vertices the network contains. It is where those vertices sit: a dense cluster of mutually close points overwhelms the distance-based measurement.
That distinction matters because “network dimension” is often treated as a compact description of how a complex system scales. But a single number can quietly combine different questions. How quickly does the graph grow? How much of its mass collects in small neighborhoods? How far apart are typical pairs? Jiantao Liu, Cheng Zeng, Yumei Xue, and Qingcheng Zeng develop a framework that separates those questions and shows precisely when their answers agree—and when they do not (Liu et al., 2026).
This is a mathematical paper, not a measurement study of transport, social, ecological, or technological networks. Its contribution is a set of definitions, bounds, and carefully chosen graph examples. The payoff is conceptual: it clarifies what a discrete fractal dimension can and cannot tell us about a network, and gives researchers tools to avoid mistaking overall size for spatial organization.
The Science
A fractal dimension describes how complexity changes as we look at finer scales. In an ordinary finite graph, shortest-path distances give a perfectly valid geometry—but classical Hausdorff dimension assigns any finite space dimension zero. To get a meaningful scaling quantity, network scientists instead study a family of graphs whose diameter grows without bound. The dimension then captures how the family’s structure changes as the scale expands.
The authors focus on connected, unweighted graphs. They rescale each graph’s shortest-path distances by its diameter, so every graph fits into a normalized distance range from near zero to one. A one-step separation becomes smaller and smaller in this normalized view as the graph grows; the reciprocal of the diameter sets the smallest meaningful scale.
Their first tool is an energy: a weighted sum over pairs of vertices, where close pairs count more heavily. For the usual power-law version, a pair’s contribution grows like the inverse of its normalized distance raised to a chosen exponent. The exponent asks how sensitive the measurement is to closeness. If a graph contains many pairs packed tightly together, the energy rises rapidly as that exponent increases.
There is a technical snag: a vertex is distance zero from itself, which would make the inverse-distance kernel blow up. The researchers regularize the kernel by replacing distances below the graph’s smallest positive normalized distance with that minimum. This makes the energy finite and allows them to optimize over probability distributions on the vertices, not just the uniform distribution that gives each vertex equal weight.
The second tool is ball mass: how much probability lies within a given radius of any vertex. A network measure has strong concentration if some small neighborhoods can hold a large share of its mass. The paper connects this local picture to energy, and also compares both with a mass dimension based on the total number of vertices relative to the graph’s smallest scale.
The authors generalize the familiar power-law scale as well. Instead of measuring scale only with the logarithm of inverse radius, they allow a broader class of decreasing functions. That makes it possible, in principle, to study graph families whose growth is not well summarized by a single power law. Their main results require a modest regularity condition: doubling a radius cannot change the scale function by an arbitrarily large amount.
What They Found
The first result is a discrete variational principle. Under that regularity condition, the best energy dimension obtainable by choosing a probability distribution equals the best ball-mass concentration dimension obtainable over sequences of distributions. In plain terms: optimizing the pairwise interaction measure and optimizing the control of local crowding lead to the same critical scaling exponent.
The result also says that the uniform energy dimension cannot exceed the optimized energy dimension, which in turn cannot exceed the upper mass dimension. This ordering is useful because it identifies what each measure is allowed to claim. A uniform measure may miss a more favorable distribution over the graph, while a total vertex count can place an upper ceiling on the optimized dimension without determining it exactly.
The paper’s sharpest demonstration is a graph made by attaching a complete graph of size to a path of size with one edge. The total number of vertices is , and the diameter grows in proportion to , giving the family a mass dimension of 1. Yet all pairs inside the complete-graph portion are separated by just one edge. Their normalized distances shrink as the path gets longer, making their contribution to uniform energy explode for every positive exponent. The resulting uniform energy dimension is 0.
That is a striking mismatch: vertex growth says “one-dimensional,” while the energy calculation says the dense cluster dominates so thoroughly that no positive exponent remains bounded. The paper’s Figure 1 depicts this construction.
The example is deliberately simple, but it reveals a general problem: a network’s diameter and vertex count do not tell us how its vertices are distributed across distances.
One graph family, two dimensions
Reported dimensions for the family formed by attaching a complete graph of size n to a path of size n.
| Label | Value |
|---|---|
| Mass dimension | 1 dimension |
| Uniform energy dimension | 0 dimension |
A second example shows that even the lower and upper versions of a growth dimension can matter. The authors alternate graph families: odd-indexed graphs are paths, while even-indexed graphs are square grids. Along the path subsequence, the ratio of logarithmic vertex count to logarithmic diameter tends to 1. Along the grid subsequence, it tends to 2. The classical energy-based Hausdorff dimension of the full alternating family is 2, so the lower mass dimension is strictly smaller than that energy dimension.
This is why the upper mass dimension, rather than the lower one, appears as the general upper bound for optimized energy dimension. A liminf energy definition can stay bounded along a favorable subsequence even if another subsequence behaves very differently. In this alternating construction, the grid subsequence sets the energy dimension, while the lower mass dimension tracks the path subsequence.
Alternating paths and grids separate growth dimensions
Reported lower and upper mass dimensions and energy-based Hausdorff dimension for the alternating path-grid family.
| Label | Value |
|---|---|
| Lower mass dimension | 1 dimension |
| Hausdorff dimension | 2 dimension |
| Upper mass dimension | 2 dimension |
The authors also introduce q-energy dimensions, which use a moment of order of the potential generated by a measure. The parameter changes how strongly the calculation responds to uneven mass distribution. For a fixed measure, the resulting dimension is nonincreasing as rises. But after optimizing over all vertex measures, the dimension becomes independent of : for every , the optimized q-energy dimension equals the optimized concentration dimension.
This distinction between fixed and optimized measures is important. In a path-and-tree example described by the paper, the uniform q-energy dimension depends on both the tree’s relative size and q, while the optimized value is always 1. The uniform measure on the path itself achieves that optimized value for every q greater than one. Optimization therefore identifies a structural capacity of the graph family that a particular choice of weights may obscure.
The paper further compares energy with Rényi entropy, a measure of how unevenly probability is distributed across outcomes. For the uniform measure, Rényi entropy at every order is simply the logarithm of the number of vertices. That means entropy records global growth, but it cannot by itself see whether vertices are tightly clustered or spread across the graph. Energy adds the missing distance information. The authors also show that under uniform polynomial bounds on the sizes of metric balls, the power-scale energy and mass dimensions agree.
Why This Changes Things
Network dimensions are often used to compare systems or summarize scaling in a single figure. This paper makes clear that the figure is meaningful only after specifying what is being measured: total growth, local concentration, average pairwise separation, or the best possible distribution of mass. Two families can have the same number of vertices at the same scale and still have different energy dimensions because one hides a dense cluster inside a long-diameter structure.
That insight is relevant well beyond abstract graph theory, even though the paper does not test real-world networks. In a transport network, a large hub or tightly connected downtown core may dominate measures that reward short paths. In a social graph, a dense community can coexist with long bridges between communities. In an ecological network, a highly connected local cluster may sit inside a broader, sparse system. In each case, a global growth exponent and a distance-sensitive measure answer different questions.
The framework also provides a way to reason about the effect of weighting. The uniform measure treats every vertex equally. An optimized measure can place more probability on regions that are less concentrated, revealing the largest scaling exponent the graph family can support. That is mathematically informative, but it should not automatically be interpreted as the “true” dimension of a real system: the optimized distribution may not correspond to how people, organisms, traffic, or information actually move through it.
The use of generalized scale functions widens the potential reach of the framework. Power laws are convenient, but complex systems do not always scale as clean powers. A flexible scale function can represent other forms of growth, provided it changes regularly enough when the radius is doubled. That helps distinguish a failure of power-law fit from a deeper absence of measurable structure.
The paper’s examples are not empirical proof that real networks exhibit these exact mismatches. They are counterexamples to overly broad assumptions: vertex count and diameter alone do not determine energy dimension, lower mass growth need not bound it, and conclusions based on a uniform distribution need not survive optimization. Those are useful guardrails for future empirical work.
What’s Next
The supplied paper text establishes the main variational results and begins a section on energy bounds for concentrated vertex sets, but it cuts off during a proposition. The abstract and introduction also summarize further analyses of graph families and measure invariance, while the full proofs and numerical details for those later sections are not included in the supplied material. So the conclusions here should be read as a digest of the material available, not as a reconstruction of missing results.
A natural next step is to apply these definitions to observed networks and ask which measure is appropriate for a particular scientific question. Researchers would need to test how sensitive the dimensions are to sampling, missing edges, changing network boundaries, and the choice of probability measure. In many empirical systems, the network itself evolves, adding another scale of variation that these graph-family definitions do not by themselves resolve.
There is also a practical interpretive challenge: an optimized dimension is an upper structural possibility, not necessarily a description of typical activity. Comparing it with the uniform dimension—or with a measure weighted by actual traffic, interaction frequency, or abundance—could reveal whether a network’s geometry is broadly accessible or only present in a carefully selected subset.
The larger contribution is a clearer vocabulary for talking about network complexity. Dimension is not a single property waiting to be read off a graph. It depends on what counts as mass, how distances are weighted, and which scales are compared. By making those choices explicit, Liu and colleagues give network scientists a better chance of turning a compact number into a meaningful explanation.