Eigenvector Centrality¶

A network node score recursively weighted by the scores of connected nodes and selected from a leading adjacency eigenvector.

Version: v1 · 2026-10-03 · History
Domain-specific #: 13180
Domain group: Formal Sciences
Origin domain: Mathematics
Subdomain: Network Analysis → Mathematics

Core Idea¶

Eigenvector centrality ranks nodes of a network by a recursive rule: a connection contributes more when the connected node also has a high score. For a nonnegative adjacency or similarity matrix \(A\), a score vector \(x\) satisfies \(Ax=\lambda x\) (with an explicitly chosen transpose convention for directed networks). The conventional centrality vector is a nonnegative eigenvector for the spectral radius, normalized because multiplying \(x\) by a positive constant changes no ranking. Bonacich's original account gives the summed-neighbor equation and distinguishes it from his more flexible parameterized centrality family.^[1]^[2]

This is a graph-relative measure, not an assertion about intrinsic merit, social power or neural causation. A connected undirected nonnegative network gives a unique strictly positive Perron vector up to scale; a strongly connected directed network has the analogous guarantee under the selected left/right convention. A disconnected or reducible graph can yield zero entries or nonunique leading rankings. Even with a unique Perron vector, a simple undamped power iteration may need extra conditions to converge; existence of the measure is not a guarantee for every computational routine.^[2]

Structural Signature¶

Sig role-phrases: specified network — nonnegative adjacency representation — recursively weighted node scores — leading eigenvector and normalization — connectivity qualification.

Specified network. The nodes and edges, including direction and weight meaning, determine what counts as a neighbor. A social tie and a similarity between fMRI voxel time series are unlike relations but each produces a network.^[1]^[3]
Nonnegative adjacency representation. Matrix entry \(A_{ij}\) records the relevant tie/weight under a stated orientation. The standard Perron-ranking reading does not automatically extend to arbitrary signed similarities.^[3]^[2]
Recursively weighted scores. Each \(x_i\) is proportional to a weighted sum of neighbors' \(x_j\), rather than merely to the number of adjacent nodes. This is why one strong connection can outrank many weak ones.^[1]^[3]
Leading eigenvector and normalization. The principal nonnegative spectral solution is selected, then a normalization fixes arbitrary scale. An arbitrary eigenvector is not the conventional centrality score.^[1]^[3]
Connectivity qualification. Irreducibility supports positivity and uniqueness up to scale. Without it, multiple components can compete for the dominant eigenvalue and leave some nodes with zero or ambiguous rank.^[2]

What It Is Not¶

It is not degree centrality: degree sums immediate ties, while eigenvector centrality weights them by the other nodes' scores. Lohmann and colleagues compute both and report different patterns in their fMRI example. It is also not a direct measure of causal influence: score depends on how nodes and edges were constructed, and Bonacich explicitly cautions that network-derived position can omit information quality, rights or resource values.^[3]^[1]

It is not identical to Katz centrality, PageRank or every member of Bonacich's parameterized family. Baselines, attenuation, damping and degree normalization can address different data conditions but change the equation and interpretation. Calling these related descendants does not make their values interchangeable on a given graph.^[1]^[3]

Scope of Application¶

The conventional positive-score reading uses a nonnegative graph matrix and a specified edge interpretation. For undirected connected graphs, the adjacency matrix is symmetric and irreducible. Directed networks require explicit score orientation (incoming versus outgoing importance); NetworkX's official implementation uses a left eigenvector and states strong-connectivity conditions for a unique positive result. Neither software convention is a universal statement that all fields orient their adjacency matrices the same way.^[2]

The measure is used in social-network analysis and in weighted similarity networks such as Lohmann et al.'s voxel-wise fMRI maps. These uses share the eigenvector relation but not the substantive meaning of an edge: a social connection is not a neural functional connection. A study-specific centrality map cannot by itself prove a node's real-world influence or physiological role.^[1]^[3]

Clarity¶

The measure makes recursive prominence exact. Instead of vaguely saying that a node is important because it connects to important nodes, it asks whether the entire score vector is self-consistent under the adjacency operator. This distinguishes quality-weighted contact from raw degree and exposes why the whole network, not an isolated ego neighborhood, matters.^[1]

It also forces two modeling questions into view: what does an edge mean, and when is the score uniquely interpretable? A voxel correlation graph and a social relation graph can use the same algebra, but neither receives the same substantive interpretation merely from matching mathematics. A disconnected component can be assigned a zero centrality in a global Perron vector even if it has active internal ties; that zero is a consequence of the global scoring convention, not proof of local inactivity.^[3]^[2]

Manages Complexity¶

A network with many pairwise ties can be summarized by a single node-score vector, permitting comparison or mapping without narrating every edge. In Lohmann et al., voxel similarities become a matrix and then a centrality map, making a many-thousand-node pattern inspectable at the voxel level. That compression preserves the selected spectral relation but discards other properties such as community structure, direction-specific mechanisms and information not encoded in edges.^[3]

The simplification is conditional. Global scores depend on the entire represented graph, the weighting choices and the treatment of disconnected parts. A more regularized related measure can be easier to calculate or compare on a reducible graph, yet then one is no longer reporting the same unmodified eigenvector centrality.^[1]^[2]

Abstract Reasoning¶

Suppose a nonnegative connected undirected network has adjacency \(A\) and Perron vector \(x\). The equation \(x_i=\lambda^{-1}\sum_j A_{ij}x_j\) means that merely adding a tie is not the whole story: the tied node's score affects the change in the receiving node's recursive equation. The score vector must be solved jointly because each neighbor's value itself depends on others. Perron–Frobenius supplies a positive, scale-unique solution under irreducibility; it does not assign a causal direction to the substantive network.^[1]^[2]

Now split a graph into disconnected components. Its adjacency becomes block diagonal. If one block has a strictly larger spectral radius, a global dominant eigenvector can have zero entries in other blocks; if blocks tie at the top, the eigenspace can have multiple eligible vectors. Therefore a statement such as “all connected nodes have positive, uniquely comparable scores” needs a graph-wide connectivity condition, not merely nonzero local degree.^[2]

Knowledge Transfer¶

The structural roles transfer literally from social ties to voxel similarities: define nodes, weights, adjacency, a nonnegative leading eigenvector and its normalization. What changes is the interpretation of edges and scores. Bonacich's structural status does not entail bargaining power; Lohmann's voxel-wise prominence does not entail a voxel causally drives another. The mathematical transfer is real, but domain claims require independent evidence.^[1]^[3]

At a higher level, live Network captures nodes and relations without imposing a Perron score. Live Eigenvalue and Eigenvector captures an operator's invariant direction and associated scalar; eigenvector centrality requires that spectral pair but specializes it to a leading nonnegative ranking of graph nodes. These are two different strict prerequisites. Fixed Point remains related because normalized spectral iterations can have fixed vectors, but no iteration is constitutive of the score.^[4]^[5]^[6]

Examples¶

Bonacich writes a relation matrix \(R\) and an eigenvector status score \(e\) satisfying \(Re=\lambda e\). A position receives score through the scores of its contacts. The article notes use in interlocking-directorate research yet warns that structural centrality can diverge from bargaining power and omit substantive differences in resources or ties. This is a model of network-derived status, not an empirical guarantee that an actor wields more influence.^[1]

Mapped back: The specified network is positions and their relation ties; the nonnegative adjacency representation is \(R\) for the positive-status reading; recursive node scores weight each contact by its score; the leading eigenvector and normalization select \(e\) up to scale; the connectivity qualification governs whether a positive unique ranking follows. The status interpretation remains bounded by the represented ties.

Voxel-wise fMRI similarity network¶

Lohmann et al. treat voxels in a brain region of interest as nodes, construct nonnegative symmetric similarities between time series, and compute a normalized leading-eigenvector score for each voxel. They compare maps with degree centrality and report state-related differences. The example shows transfer to a weighted measurement network, not a claim that central voxels cause the observed states.^[3]

Mapped back: The specified network is voxels connected by time-series similarity; the nonnegative adjacency representation is their symmetric similarity matrix; recursively weighted scores favor voxels similar to other high-score voxels; the leading eigenvector and normalization produce the centrality map; the connectivity qualification is the study's irreducibility assumption. The domain meaning of a tie is wholly different from social contact.

Structural Tensions¶

T1 — Global recursive context versus local interpretability. The leading eigenvector captures the network-wide pattern that degree misses, but a node's score then depends on distant edges and is harder to explain from its own immediate ties. Restricting attention to local degree is easier to interpret but loses recursively weighted status. Diagnostic: Is the research question about immediate connection count or prominence through high-score neighbors across the network?^[1]^[3]

T2 — Pure spectral score versus reducible-network coverage. Keeping the unmodified eigenvector equation preserves a transparent identity, but a reducible graph may have zero or nonunique global scores. Adding a baseline, damping or other regularization can make a ranking usable across components, at the cost of changing the measure. Diagnostic: Do the graph's connectivity and comparison needs justify the pure Perron score, or is a separately named regularized measure required?^[1]^[2]

Structural–Framed Character¶

Eigenvector centrality lies toward the structural end within network analysis: its eigen-equation is exact, though the graph, edge meaning and normalization are constitutive. Its evaluative weight is low as mathematics; “central” need not mean admirable, powerful or causally decisive. Its human-practice dependence is moderate in empirical use because people choose nodes, similarity rules and direction, while the algebra is not a social convention. Its institutional origin is mathematical social-network analysis, later reused in neuroscience; no institution's endorsement is required for the definition. Its vocabulary travels literally between those domains—node, edge, adjacency, Perron score—while their substantive interpretations change. For import versus recognition, applying the equation to a new network is a real import of a graph measure, not proof that the new substrate itself obeys social-status theory. Its character: a reusable but still domain-specific network measure, with the broader Network prime as a constitutive prerequisite.^[1]^[3]^[4]

Structural Core vs. Domain Accent¶

The portable core is a score determined recursively by weighted relations, with a nonnegative leading-eigenvector solution under appropriate graph assumptions. It transfers among kinds of network, but cannot lose network adjacency, nodes and spectral selection without becoming a different abstraction. Live Network carries the broader node–edge relation and live Eigenvalue and Eigenvector carries the invariant-direction/scalar pair. Their combination is necessary but not sufficient: eigenvector centrality adds a nonnegative leading graph-node ranking.^[4]^[5]^[1]

The domain accents are the meaning of each tie and the claim made from the resulting score: social standing in Bonacich's motivating setting, similarity-derived voxel prominence in Lohmann's. Neither accent changes the equation, but each governs what the score can responsibly mean. Fixed Point is a neighboring prime, not an asserted parent; defining a centrality score does not require a particular power-iteration dynamics.^[1]^[3]^[6]

This entry presupposes Network and presupposes Eigenvalue And Eigenvector. Adjacency-based centrality presupposes a defined network relation. The score is necessarily a leading eigenvector of the adjacency operator with an associated eigenvalue.

Relationships to Other Abstractions¶

Current abstraction Eigenvector Centrality Domain-specific

Parents (2) — more general patterns this builds on

Eigenvector Centrality presupposes Eigenvalue And Eigenvector Prime

The score is necessarily a leading eigenvector of the adjacency operator with an associated eigenvalue.
Eigenvector Centrality presupposes Network Prime

Adjacency-based centrality presupposes a defined network relation.

Hierarchy paths (3) — routes to 3 parentless roots

Eigenvector Centrality → Eigenvalue And Eigenvector → Linearity

Show alternative paths (2)

Neighborhood in Abstraction Space¶

Eigenvector Centrality sits in a sparse region of the domain-specific corpus (74^th percentile for distinctiveness): few abstractions share its structure, so a faithful description tends to retrieve it precisely.

Family — Graph Structures & Combinatorial Objects (44 abstractions)

Nearest neighbors

Join Count Statistic — 0.84
Metric dimension (graph theory) — 0.83
Hadwiger number — 0.83
Narrative network — 0.83
Generalized blockmodeling of binary networks — 0.83

Computed from structural-signature embeddings · 2026-10-08

Not to Be Confused With¶

Degree centrality: immediate tie total, not recursively weighted neighbor status.^[3]
Katz centrality and PageRank: related but altered equations involving attenuated paths, baselines or damping/normalization; not unconditional aliases.^[1]^[3]
Bonacich's whole parameterized power-centrality family: includes negative dependence on partners' power in exchange networks, unlike the nonnegative leading-eigenvector score.^[1]
Actual influence or causation: a graph score can omit resources, information quality, physiology and edge-construction uncertainty.^[1]^[3]
Guaranteed strictly positive ranking on any graph: the Perron guarantee needs irreducibility/strong connectivity conditions; reducible graphs need qualification.^[2]

References¶

[1] Phillip Bonacich, “Power and Centrality: A Family of Measures”, American Journal of Sociology 92(5), 1987, pp.1170–1182; especially printed pp.1170–1174, Eq. (1)–(4). Original full article. registry ↩a ↩b ↩c ↩d ↩e ↩f ↩g ↩h ↩i ↩j ↩k ↩l ↩m ↩n ↩o ↩p ↩q ↩r ↩s ↩t

[2] NetworkX project, eigenvector centrality official documentation, Notes and NumPy implementation sections, inspected 2026-10-01. registry ↩a ↩b ↩c ↩d ↩e ↩f ↩g ↩h ↩i ↩j ↩k

[3] Gabriele Lohmann et al., “Eigenvector Centrality Mapping for Analyzing Connectivity Patterns in fMRI Data of the Human Brain”, PLOS ONE 5(4):e10232, 2010; Abstract, Materials and Methods: Eigenvector Centrality, Results and Discussion. Original full article. registry ↩a ↩b ↩c ↩d ↩e ↩f ↩g ↩h ↩i ↩j ↩k ↩l ↩m ↩n ↩o ↩p ↩q

[4] Encyclopedia of Abstractions, live prime_abstractions/v2/network.md, Core Idea and Structural Signature, inspected 2026-10-01. registry ↩a ↩b ↩c

[5] Encyclopedia of Abstractions, live prime_abstractions/v2/eigenvalue_and_eigenvector.md, Core Idea and Structural Signature, inspected 2026-10-01. registry ↩a ↩b

[6] Encyclopedia of Abstractions, live prime_abstractions/v2/fixed_point.md, Core Idea and Structural Signature, inspected 2026-10-01. registry ↩a ↩b