Linked Nets: Node Sharing & Entanglement

• Linked Nets are diverse network structures that encompass interconnecting bilayer networks (coupled by shared nodes) and ensnarled spatial network pairs (defined by topological interlocking of cycles).
• The node-sharing formulation quantifies structural differences by comparing degree and betweenness metrics across shared nodes in multilayer systems spanning biology, transportation, and more.
• The topological framework employs linking invariants and greedy edge-cutting algorithms to analyze and measure spatial entanglement, offering actionable metrics for network decoupling.

Linked nets denote distinct forms of network coupling in contemporary research. In one formulation, a linked system is an interconnecting bilayer network in which two network layers are coupled because they share some of their nodes, called interconnecting nodes. In another formulation, a linked system is a pair of finite, 2-component spatial nets embedded in $\mathbb{R}^3$ whose cycles are topologically interlocked so that the two networks cannot be separated without removal of a subset of edges; this state is termed ensnarled. The two constructions address different problems—identity overlap across layers in the first case, topological entanglement of spatial graphs in the second—but both treat coupling as a structural property of a network pair rather than as a property of a single isolated graph (Xu et al., 2011, Kramer et al., 2022).

1. Terminological scope and principal constructions

Within the literature considered here, “linked nets” does not refer to a single canonical object. The phrase covers at least two non-equivalent constructions. In interconnecting bilayer networks, linkage is created by shared nodes. In ensnarled spatial network pairs, linkage is created by topological interlocking of cycles in space. The distinction is substantive: the former is a multilayer network problem, whereas the latter is a spatial graph and topological graph theory problem.

Aspect	Interconnecting bilayer networks	Ensnarled spatial network pairs
Coupling principle	Layers share some common nodes	Embedded cycles are topologically linked
Basic objects	Two network layers with $V_1 \cap V_2 \neq \emptyset$	Two finite graphs embedded in $\mathbb{R}^3$
Central quantities	$n$, $U_k$, $U_b$	$L$, $\Lambda_0$, $\rho_i$, $\rho_{\max}$, $V_1 \cap V_2 \neq \emptyset$0, $V_1 \cap V_2 \neq \emptyset$1

These constructions are also distinct from adjacent categories explicitly discussed in the sources. Interconnecting bilayer networks differ from interdependent networks, where layers are coupled by dependency links, and from multiplex networks, where layers typically share the same node set but differ in edge type. Ensnarled nets, by contrast, generalize classical knot and link theory from disjoint closed curves to mesh-like spatial graphs with many cycles (Xu et al., 2011, Kramer et al., 2022).

2. Interconnecting bilayer networks as node-sharing systems

An interconnecting bilayer network consists of an upper layer with node set $V_1 \cap V_2 \neq \emptyset$2 and a lower layer with node set $V_1 \cap V_2 \neq \emptyset$3, subject to $V_1 \cap V_2 \neq \emptyset$4. A real entity that appears in both layers is represented as node $V_1 \cap V_2 \neq \emptyset$5 in the upper layer and node $V_1 \cap V_2 \neq \emptyset$6 in the lower layer; these two layer-specific representations are treated as a single interconnecting node in the bilayer system. If there are $V_1 \cap V_2 \neq \emptyset$7 such shared nodes, the normalized number of interconnecting nodes is

$V_1 \cap V_2 \neq \emptyset$8

This formulation treats coupling as node identity overlap rather than as inter-layer edges. The paper analyzes 8 bilayer systems spanning food, biology, medicine, movies, and transportation. Representative examples include the HP–CD bilayer, in which the upper layer is a network of Chinese herb prescriptions and the lower layer is a network of Chinese cooked dishes, with interconnecting nodes given by items used both as herbs and as food ingredients; the coach–train bilayer, where interconnecting nodes are cities that have both coach and train stations; and the YPI–YM bilayer, where interconnecting nodes are proteins that are both present in the yeast protein interaction network and act as enzymes in the yeast metabolic network.

The empirical analysis focuses on two standard node properties in each layer: degree $V_1 \cap V_2 \neq \emptyset$9 and betweenness $\mathbb{R}^3$0. For an interconnecting node, with values $\mathbb{R}^3$1 in the upper layer and $\mathbb{R}^3$2 in the lower layer, where $\mathbb{R}^3$3 stands for either $\mathbb{R}^3$4 or $\mathbb{R}^3$5, the topological difference is defined as

$\mathbb{R}^3$6

and the averaged topological difference over all $\mathbb{R}^3$7 interconnecting nodes is

$\mathbb{R}^3$8

The paper studies $\mathbb{R}^3$9 and $n$0 as bilayer-level measures of how differently shared entities function structurally in the two layers. Across all 16 single layers, the degree and betweenness distributions are fitted by a shifted power-law,

$n$1

which interpolates between a power law at $n$2 and exponential-like decay for large $n$3 (Xu et al., 2011).

3. Analytical node-sharing mechanism and empirical regularity

The core empirical finding is that networks with smaller averaged topological differences of the interconnecting nodes tend to share more nodes. The paper formulates a “very simple node sharing mechanism” to explain this relation analytically. The model adopts three simplifying assumptions: $n$4, identical distributions $n$5 and equal averages $n$6; $n$7 for interconnecting nodes; and a proportional selection rule

$n$8

where $n$9 depends on the bilayer system index $U_k$0. The parameter $U_k$1 is interpreted as a measure of role difference: larger $U_k$2 means that, for shared entities, the upper-layer centrality is much larger than the lower-layer centrality.

With $U_k$3 distributed on $U_k$4 according to the shifted power-law, the number of interconnecting nodes can be expressed as an integral over the admissible range determined by $U_k$5. After substitution, the normalized number of interconnecting nodes becomes

$U_k$6

with $U_k$7 and $U_k$8 defined in terms of the shifted power-law parameters. The paper also derives a closed-form expression $U_k$9. Eliminating $U_b$0 yields a functional relation $U_b$1. Its qualitative content is explicit: $U_b$2 decreases as $U_b$3 increases, $U_b$4 increases as $U_b$5 increases, and therefore $U_b$6 is a decreasing function of $U_b$7.

The empirical data from eight bilayers are reported to follow this monotonic tendency for both degree and betweenness. Bilayers such as HP–CD exhibit small $U_b$8 and large $U_b$9, meaning that few entities are shared and those shared entities play quite different structural roles. Bilayers such as coach–train and YPI–YM exhibit larger $L$0 and smaller $L$1, meaning that many entities are common to both layers and have more similar centrality profiles. To support the proportional rule indirectly, the paper also examines an averaged scaling relation between $L$2 and $L$3, reported in logarithmic form as

$L$4

with small $L$5, which suggests a statistical tendency consistent with $L$6 rather than an exact nodewise identity (Xu et al., 2011).

4. Ensnarled network pairs as topologically linked spatial graphs

In the second usage, linked nets are pairs of spatially embedded graphs whose edges and cycles are topologically interlocked in three dimensions. The paper defines an ensnarled pair as two finite graphs $L$7 and $L$8, each embedded in $L$9 as tubal structures, such that one cannot move $\Lambda_0$0 and $\Lambda_0$1 apart to infinity by continuous deformations in $\Lambda_0$2 while keeping all edges intact; unlinking requires removal of a subset of edges. This extends classical link theory from closed curves to mesh-like networks.

Each network is represented as a simple graph $\Lambda_0$3, where $\Lambda_0$4 is the vertex set, $\Lambda_0$5 the edge set, and $\Lambda_0$6 give the start and end vertices of each edge. A connected graph with $\Lambda_0$7 vertices and $\Lambda_0$8 edges has cycle-space dimension, or cyclomatic number,

$\Lambda_0$9

A spanning tree uses $\rho_i$0 edges, and the remaining $\rho_i$1 edges generate $\rho_i$2 fundamental cycles forming a cycle basis $\rho_i$3. Because the graphs are spatially embedded, each cycle is represented as a polygonal curve in $\rho_i$4.

For two closed oriented curves $\rho_i$5, the linking number is defined by the Gauss integral

$\rho_i$6

The paper uses this invariant to evaluate pairwise linking of cycles from two graph cycle bases. If $\rho_i$7 and $\rho_i$8, the linking matrix is

$\rho_i$9

For arbitrary cycles $\rho_{\max}$0 and $\rho_{\max}$1, represented through basis coefficients $\rho_{\max}$2 and $\rho_{\max}$3, the linking number can be written as

$\rho_{\max}$4

Within this framework, Hopf-link identification means finding pairs of basis cycles with linking number $\rho_{\max}$5, interpreted as primitive topological locks between the two spatial nets (Kramer et al., 2022).

5. Edge-space factorization, greedy unlinking, and ensnarledness metrics

To transfer topological information from cycle space to edge space, the paper introduces the mesh matrix

$\rho_{\max}$6

With mesh matrices $\rho_{\max}$7 and $\rho_{\max}$8 for the two graphs, the linking matrix is factorized through an edge-space operator $\rho_{\max}$9. Because the mesh matrices are not full rank, $V_1 \cap V_2 \neq \emptyset$00 is not unique; the paper selects the canonical least-squares solution

$V_1 \cap V_2 \neq \emptyset$01

where $V_1 \cap V_2 \neq \emptyset$02 denotes a generalized inverse. This yields an edge-space expression for arbitrary cycle linking numbers,

$V_1 \cap V_2 \neq \emptyset$03

with $V_1 \cap V_2 \neq \emptyset$04 and $V_1 \cap V_2 \neq \emptyset$05 the edge-incidence vectors of the cycles. Two positive-definite edge priority operators are then defined:

$V_1 \cap V_2 \neq \emptyset$06

The diagonals of $V_1 \cap V_2 \neq \emptyset$07 and $V_1 \cap V_2 \neq \emptyset$08 assign scalar priorities to edges, indicating how critical each edge is for maintaining nonzero linking. An edge that lies on many Hopf-linked cycles, or on cycles with large linking numbers, acquires high priority.

These priorities drive a greedy edge-cutting algorithm for unlinking. One graph is designated as reference and the other as target. The procedure computes cycle bases, the linking matrix $V_1 \cap V_2 \neq \emptyset$09, $V_1 \cap V_2 \neq \emptyset$10, and the target priority operator, then iteratively removes the target edge with maximum priority, recomputing all quantities after each cut. The process terminates when the updated linking matrix becomes the zero matrix. The resulting number of removed edges is $V_1 \cap V_2 \neq \emptyset$11. The paper notes that this heuristic “reliably identifies efficient cuts,” but it may fail to find the global optimum, particularly when linking is near the periphery of the meshes or when degree-2 vertices distort priorities; coarse-graining such vertices can improve performance.

From these cut sets, the paper defines several topological metrics. For each network $V_1 \cap V_2 \neq \emptyset$12, with cyclomatic number $V_1 \cap V_2 \neq \emptyset$13 and required cut count $V_1 \cap V_2 \neq \emptyset$14,

$V_1 \cap V_2 \neq \emptyset$15

This is interpreted as a normalized linkage cost: if $V_1 \cap V_2 \neq \emptyset$16, only a small fraction of the cycle space is essential for linkage, whereas $V_1 \cap V_2 \neq \emptyset$17 means that essentially all cycles are involved in ensnarement. When $V_1 \cap V_2 \neq \emptyset$18, the graph is a tree and cannot be topologically linked via linking numbers. To quantify worst-case burden and symmetry, the paper defines

$V_1 \cap V_2 \neq \emptyset$19

The minimal absolute cut cost is

$V_1 \cap V_2 \neq \emptyset$20

These metrics are illustrated on dual ladder meshes, a ring linked to a hexagonal mesh, finite pieces of Laves nets, and random spatial networks. In the dual ladder example, translation changes the edge-priority pattern and eventually drives all linking numbers to zero. In the ring–hexagonal mesh example, the greedy algorithm removes four edges from the mesh to achieve unlinking. The methodology is implemented in the open-source Python package SnarlPy (Kramer et al., 2022).

6. Conceptual comparison, limitations, and open directions

The two linked-net constructions encode different structural questions. Interconnecting bilayer networks ask how many entities can be shared between two layers, and how the degree or betweenness of those entities differs across layers. Ensnarled network pairs ask which cycles and edges sustain topological interlocking in $V_1 \cap V_2 \neq \emptyset$21, and how many cuts are required to eliminate it. In the first case, coupling is by node identity and is quantified by $V_1 \cap V_2 \neq \emptyset$22 and $V_1 \cap V_2 \neq \emptyset$23. In the second, coupling is by spatial linkage and is quantified by $V_1 \cap V_2 \neq \emptyset$24, $V_1 \cap V_2 \neq \emptyset$25, edge priorities, and cut-based measures such as $V_1 \cap V_2 \neq \emptyset$26 and $V_1 \cap V_2 \neq \emptyset$27. The two uses are therefore complementary rather than interchangeable.

Each framework is explicitly limited. The interconnecting bilayer study is restricted to two layers, static undirected unweighted networks, a symmetric analytical model with $V_1 \cap V_2 \neq \emptyset$28 and $V_1 \cap V_2 \neq \emptyset$29, and a stylized proportional rule $V_1 \cap V_2 \neq \emptyset$30. It focuses primarily on degree and betweenness, though the authors report trying average nearest neighbor degree and node strength with similar qualitative results. Suggested extensions include more than two layers, combining node-sharing with inter-layer edges or dependency links, studying dynamical processes on interconnecting networks, and treating directed or weighted networks.

The ensnarled-network framework is inherently two-component and based on pairwise linking numbers between cycles. The paper states that it does not capture higher-order entanglements such as Brunnian links and suggests that multi-component nets would require higher-order invariants such as Milnor invariants. Additional open directions include construction algorithms that synthesize spatial graphs from a desired $V_1 \cap V_2 \neq \emptyset$31, treatment of dynamic embeddings, and systematic coarse-graining strategies for large meshes and degree-2 vertices. A plausible implication of considering the two literatures together is that “linked nets” should be read contextually: in one domain it denotes multilayer overlap of entities, while in another it denotes topological entanglement of spatial network components (Xu et al., 2011, Kramer et al., 2022).