Copying Mechanism¶

Grow a network by sampling an existing prototype node and giving a new node some of the prototype's neighbors, optionally mixed with random links, deletions, or other mutations, so local imitation can generate global degree heterogeneity and clustering.

Version: v2 · 2026-08-30 · History
Domain-specific #: 1562
Origin domain: network science
Subdomain: generative models of growing networks
Aliases: Copying model, Copy model, Node-copying model

Core Idea¶

A copying mechanism, more commonly a copying model in network-science literature, is a stochastic network-growth rule in which a new vertex samples an existing vertex as a prototype and inherits some of that prototype's adjacency pattern. Copying may be mixed with direct attachment to the prototype, uniformly random links, loss of copied links, or new links. The invariant is not a single formula but new node + sampled prototype + copied neighborhood structure + probabilistic retention or mutation + repeated graph growth.^[1]^[2]

The mechanism was developed to explain how a creator can build a new web page by finding a page on a relevant topic and reusing part of its link list. This local action creates statistically dependent edges and can yield highly unequal in-degree distributions without requiring each arriving node to know the degree of every existing node.^[1]

Node duplication and divergence in biological-network models is a closely related realization. A duplicate protein begins with the ancestor's interaction neighborhood, after which some inherited interactions are removed and others may be added. The biological interpretation differs, but the graph-generating move remains copy and modify.^[3]

Structural Signature¶

The recognition roles are:

Growing graph: a directed or undirected graph evolves through discrete arrivals or epochs.
Seed condition: growth starts from a stated finite graph or initial network distribution.
New vertex: at least one vertex is introduced at a growth step.
Prototype selection: an existing vertex is sampled by a declared rule, commonly uniformly.
Prototype neighborhood: outgoing, incoming, undirected, or multi-hop adjacency supplies the structure eligible for copying.
Copy rule: some candidate edges of the new vertex are derived from the prototype's neighbors rather than chosen independently.
Retention or mutation rule: copied edges may be retained, deleted, rewired, supplemented, or selected with stated probabilities.
Innovation channel: a variant may add random targets, direct prototype attachment, or de novo edges so the process is not closed to inherited neighborhoods.
Iteration: the copy-and-modify step repeats, making earlier outcomes possible prototypes for later arrivals.
Emergent statistics: degree distribution, edge count, clustering, clique counts, densification, component structure, and realization variance are consequences to analyze rather than inputs to assume.

The decisive test is causal: did a new vertex obtain adjacency because another vertex's neighborhood served as its template? Merely producing a power law or adding a similar-looking vertex is insufficient.

What It Is Not¶

It is not copying arbitrary data, cloning a whole graph, duplicating a file, or replicating a process. The copied object is neighborhood structure within a growing-network model.

It is not ordinary independent-edge random graph generation. Copying makes new edges dependent through a shared prototype, precisely the feature used to reproduce correlations and local clustering absent from simple independent choices.^[1]

It is not identical to preferential attachment. Preferential attachment directly weights an existing vertex by degree or another attractiveness measure. Copying samples locally; high-degree vertices can receive more links indirectly because they occur in more prototype neighborhoods. Similar degree effects do not erase the different microscopic rule.

It is not every duplication–divergence model. Biological models may duplicate nodes, interactions, genes, or whole networks under specialized evolutionary assumptions. They instantiate copying only when the new node's adjacency is inherited from a selected prototype and then altered.

It is not a claim that an observed network truly grew by copying. A copying model can reproduce a statistic without uniquely identifying the historical mechanism.

Scope of Application¶

The abstraction belongs to random graph theory, complex networks, web graph modeling, citation and bibliographic networks, social-network formation, and duplication-based biological-network models. It is used both as an explanatory hypothesis about local formation behavior and as a generative null or comparison model.

The family includes directed fixed-outdegree models, whole-neighborhood node-copying models, partial-duplication models, exponential-arrival variants, and models with deletion or second-neighbor copying. Bhat and colleagues show that even a simple rule—attach to a random target and independently to each target neighbor with probability (p)—has sparse and dense regimes and nontrivial clique transitions.^[2]

Scope excludes algorithms that copy vertex attributes but generate edges independently, graph-rewrite systems without prototype-neighborhood inheritance, and static similarity measures that never generate a network.

Clarity¶

For a candidate model, specify: Is the graph directed? What is copied—out-neighbors, in-neighbors, all neighbors, or a sampled subset? How is the prototype chosen? Does the new node link to the prototype? Which probabilities retain, delete, or add edges? Can vertices or edges die? Does one node arrive per step or a whole epoch? Are multiple edges and self-loops allowed?

These choices are not interchangeable. A fixed-outdegree web model can preserve sparsity by construction. Whole-neighborhood copying can densify when the copying probability is large. Symmetric biological divergence may remove edges from both prototype and duplicate, whereas an asymmetric model changes only the new vertex.

The phrase “mutation” is also model-relative. It can mean failure to copy an eligible edge, addition of a random edge, rewiring, or a biological divergence event. A draft must state the operator rather than relying on the metaphor.

Manages Complexity¶

Copying mechanisms explain how global structure can arise from local information. An arriving node need not calculate a graph-wide degree ranking; it consults one prototype and its neighborhood. This compresses a difficult network-formation problem into a repeated local rule.

The abstraction also unifies families that otherwise appear unrelated. Web authors reusing link lists, duplicated genes retaining interactions, and papers copying references can be represented by the same prototype-neighborhood-mutation skeleton, while domain-specific parameters remain explicit.

Finally, the model separates mechanism from outcome. Power-law degree, clustering, or densification becomes a prediction to derive and compare, preventing the observed statistic from being mistaken for the generative rule itself.

Abstract Reasoning¶

Let (G_t=(V_t,E_t)). At step (t+1), add \(v\notin V_t\), sample prototype \(u\in V_t\), and form candidate edges from a neighborhood (N(u)). In a simple undirected version,

\[ P[(v,w)\in E_{t+1}\mid w\in N(u)] = p. \]

A direct edge ((v,u)) may be mandatory, optional, or absent; innovation may add edges to random \(w\in V_t\). In a directed fixed-outdegree version, each new out-edge may independently choose between a random target and a target copied from the prototype's outgoing links.^[1]

The local rule induces an effective rich-get-richer channel. A vertex of degree (k) appears in (k) neighborhoods, so under uniform prototype sampling its opportunity to receive copied edges often scales with (k). That resembles linear preferential attachment at the aggregate level while retaining a different information requirement and correlation structure.^[3]

Parameter reasoning is regime-sensitive. In Bhat and colleagues' whole-neighborhood model, (p<½) yields a sparse regime while \(p\geq1/2\) yields densification and stronger realization-to-realization fluctuations.^[2] This threshold belongs to that model, not to the entire copying family.

Counterfactual diagnosis asks whether shuffling prototype identities while preserving degree would change adjacency correlations. If not, the claimed copying channel may be unnecessary. Matching only a degree distribution cannot distinguish copying from other growth rules.

Knowledge Transfer¶

Literal transfer is valid when vertices arrive, one or more existing vertices serve as neighborhood templates, and inherited edges undergo controlled retention or innovation. Web, citation, social, and protein-interaction models can meet those roles despite different semantics.

The mechanism's portable residue is copy-with-modification along a constructed lineage. Live prime:inheritance captures a parent, derivative, default structural carry-over, and selective override. Network science adds stochastic prototype sampling, adjacency, graph growth, and emergent topological statistics.

Transfer becomes metaphor when no graph grows, no prototype neighborhood is copied, or “copying” merely means behavioral imitation. Similarity between two mature networks is not evidence of this mechanism.

Examples¶

Web graph. A new page samples an existing topical page and copies some of its outgoing links, while other links are chosen independently. Repetition creates dependent hyperlink neighborhoods and potentially heavy-tailed in-degree.^[1]

Whole-neighborhood copying. A new node attaches to a randomly selected target and copies each target-neighbor connection independently with probability (p). Changing (p) changes sparsity, densification, cliques, and fluctuation behavior.^[2]

Protein interaction network. A protein node is duplicated with the ancestor's interactions; divergence probabilistically removes common links and may add a duplicate–ancestor edge. The graph model instantiates copy-and-modify, while the biological hypothesis supplies its interpretation.^[3]

Negative—Barabási–Albert attachment. A new node selects targets directly in proportion to degree. It may produce a similar tail but has no sampled prototype neighborhood.

Negative—attribute clone. A system duplicates a user's profile fields but chooses all social edges independently. Data was copied, not network adjacency.

Negative—graph bootstrap. Resampling whole observed graphs for uncertainty analysis does not grow one graph by prototype-node copying.

Structural Tensions¶

T1: Local information versus global inequality. A node consults one neighborhood, yet the repeated rule can create hubs and heavy tails.

T2: Fidelity versus innovation. High copying preserves clusters and inherited structure; innovation prevents closure and introduces new regions.

T3: Sparsity versus densification. Copying more of a growing neighborhood can cause edges to grow superlinearly with vertices.

T4: Mechanistic plausibility versus observational non-identifiability. The rule is behaviorally plausible, but multiple generative models can fit the same degree distribution.

T5: Family unity versus parameter sensitivity. The prototype-copy skeleton is stable while direction, arrival rate, mutation operator, and threshold results differ sharply.

T6: Inherited dependence versus analytical tractability. Shared prototypes generate realistic correlations while making edges harder to analyze than independent choices.

Structural–Framed Character¶

Copying Mechanism is highly structural within network science. Graph state, prototype selection, copied neighborhood, probabilities, iteration, and resulting statistics admit formal definitions and simulation.

The explanatory interpretation is framed. A fitted copying probability does not prove that authors, proteins, or people literally executed the modeled step. Empirical use must distinguish a generative representation from a uniquely identified history.

Structural Core vs. Domain Accent¶

The structural core is transmission of a parent's structure to a derivative with selective override. Live prime:inheritance supplies that copy-with-modification relation.

The domain accent includes vertices, directed or undirected adjacency, prototype-selection distributions, edge-retention probabilities, random attachment, deletion, seed graphs, growth epochs, degree distributions, clustering, cliques, scale-free regimes, and densification. Without these, one has generic inheritance rather than a copying network model.

The minimal prospective placement is a strict composition/instantiates edge to live prime:inheritance. At each step, the new vertex is a derivative of a sampled prototype for purposes of adjacency: neighborhood structure carries over by default or probability, with selective deletion, addition, or rewiring. Copying Mechanism is not a subtype of all inheritance relations; it instantiates the relation inside stochastic graph growth.

prime:variation_strategies is a frozen semantic neighbor but false coverage. It deliberately generates alternatives and selects among results. A copying model can be an unguided stochastic process with no evaluator or selection step. Variation is a consequence of mutation, not the whole identity.

prime:template_instantiation captures reuse of a stable slot frame, but a prototype's neighborhood changes over time and is copied as graph structure rather than bound as typed slots. Preferential attachment is a related network mechanism, not a catalog parent.

Relationships to Other Abstractions¶

Current abstraction Copying Mechanism Domain-specific

Parents (1) — more general patterns this builds on

Copying Mechanism is a kind of Inheritance Prime

The minimal prospective placement is a strict composition/instantiates edge to live prime:inheritance.

Hierarchy path (1) — routes to 1 parentless root

Copying Mechanism → Inheritance → Dependency

Neighborhood in Abstraction Space¶

Copying Mechanism sits in a sparse region of the domain-specific corpus (82^nd percentile for distinctiveness): few abstractions share its structure, so a faithful description tends to retrieve it precisely.

Family — Unclustered & Miscellaneous (1565 abstractions)

Nearest neighbors

Computed from structural-signature embeddings · 2026-09-08

Not to Be Confused With¶

Preferential attachment: direct degree-weighted selection rather than prototype-neighborhood copying.

Duplication–divergence: biologically motivated family that can instantiate copying but adds gene or protein evolution assumptions.

Random attachment: independent target selection without neighborhood inheritance.

Graph cloning: reproduction of an entire graph rather than incremental node arrival.

Data copying: reproduction of values, files, or objects without graph-generation roles.

Inheritance: substrate-independent parent-to-derivative carry-over; the proposed parent, not the complete network model.

Variation Strategies: controlled exploration-and-selection pattern, not stochastic graph growth by itself.

Empirical power law: an outcome compatible with many mechanisms, not evidence uniquely identifying copying.

References¶

[1] Kumar, Ravi, Prabhakar Raghavan, Sridhar Rajagopalan, D. Sivakumar, Andrew Tomkins, and Eli Upfal. “Stochastic Models for the Web Graph.” 41^st IEEE Symposium on Foundations of Computer Science (2000): 57–65. Primary introduction and analysis of evolving dependent-edge Web graph models. https://research.ibm.com/publications/stochastic-models-for-the-web-graph. registry ↩a ↩b ↩c ↩d ↩e

[2] Bhat, U. N., P. L. Krapivsky, R. Lambiotte, and S. Redner. “Densification and Structural Transitions in Networks That Grow by Node Copying.” Physical Review E 94 (2016): 062302. https://doi.org/10.1103/PhysRevE.94.062302. registry ↩a ↩b ↩c ↩d

[3] Vázquez, Alexei, Alessandro Flammini, Amos Maritan, and Alessandro Vespignani. “Modeling of Protein Interaction Networks.” Complexus 1 (2003): 38–44. Primary duplication-and-divergence graph model. https://doi.org/10.1159/000067642. registry ↩a ↩b ↩c