Skip to content
Crypto.world

Research

Defining Decentralization: An Ontological Perspective

Decentralization has been a computer science concept for over half a century, yet there is no universally accepted definition that works across communication systems, and the term is routinely…

Defining Decentralization: An Ontological Perspective
arXiv
2608.09748
Published
2026-08-10
Authors
Jakub Kacper Szeląg, Aydin Abadi, Mohammad Naseri

AI summary

Overview

Research area: Distributed computing and computer communication systems, with a formal-ontology and formal-semantics treatment of "decentralization" that spans blockchain, federated learning, decentralized AI, cloud and IoT architectures.

Technical level: Advanced. The paper builds a four-layer ontology using description-logic notation, formal semantics of gradable adjectives, and graph-theoretic definitions; the prose is readable, but the framework itself assumes comfort with formal logic and graph abstractions.

Scope (one sentence): The paper diagnoses why decentralization has never been given a transferable formal definition and proposes a graph-based, anchor-relative ontology, two analytical metrics (Void Tolerance and Imperviousness), and a browser-based sandbox for classifying and comparing arbitrary graph-representable systems.

What This Paper Is About

Decentralization has been a computer science concept for over half a century, yet there is no universally accepted definition that works across communication systems, and the term is routinely conflated with mere distribution of data or components. The authors contrast Baran's classic topology taxonomy, in which federated learning's star-shaped server-client architecture counts as centralized, with machine learning literature that calls the very same architecture decentralized — and then further splits federated learning into "Centralized FL" and "Decentralized FL". The goal is to replace intuition and domain-specific proxies with a formal, transferable rule that still makes the context of every claim explicit.

Key Contributions

  1. Diagnosis of the "Decentralization Problem." The authors break the gap into four concrete research problems: definitional ambiguity, conceptual confusion with distribution, lack of structural formalism (node-count metrics that ignore relationships), and absence of transferable quantification beyond blockchains. A comparison table scores representative prior work (Baran; Lamport et al.; Brewer; Nakamoto; Srinivasan and Lee / Kwon et al.; Lin et al.; Rossi et al.; Troncoso et al.; Shahsavari et al.) against four requirements, and only the authors' own framework is marked as addressing all of them.

  2. A graph-based ontology separating subject, anchor, and center. A system is a non-empty graph G = (V, E). A subject is whatever is being projected over the graph (data locality, model training, and so on); an anchor is the interpretation against which centralization is judged (ownership, authority, trust). Anchors are represented extensionally as graph-realistic center regions, so a subject can be spread across many vertices while still being centralized with respect to a single center. The number of subject–anchor pairs judged decentralized defines the system's dimensionality.

  3. Formal separation of decentralization from distribution. Three functions — δ (realization particulars), λ (supporting vertices), and μ (center regions reached under an anchor) — give a rigorous semantic distinction between the two terms. Centralization then falls out as the degenerate zero-dimensional case of decentralization rather than as an opposite or a separate concept. The authors ground the argument in formal semantics, treating "decentralized" as a multidimensional gradable (mixed) adjective in Sassoon's sense.

  4. Two analytical metrics plus a working tool. Void Tolerance measures the impact of nodes and Imperviousness measures the impact of connections; together they form a decentralization vector intended to compare systems where a single scalar cannot. The framework is implemented as a browser-based application supporting interactive modeling, automated classification, metric computation, and large-scale deterministic simulation.

Main Findings

  • No prior framework satisfies all four requirements. In the comparison table, Baran is credited with relational structure only partially; Rossi et al. and Troncoso et al. each receive two full marks (the decentralization/distribution distinction and relational structure); Lin et al., Srinivasan and Lee / Kwon et al., and Shahsavari et al. receive only partial credit for multi-dimensionality or transferable quantification. Only the authors' work is marked as meeting formal definition, distribution distinction, relational structure, transferable quantification, and multi-dimensionality.

  • Identical architectures are labeled both ways in the same field. Vanilla federated learning relies on a central server that aggregates model updates, forming a star topology that Baran's taxonomy classes as centralized, yet it is widely described as decentralized because data stays local. The paper treats this not as a terminological slip but as evidence that decentralization claims silently depend on an undeclared dimension.

  • Decentralization is a mixed multidimensional adjective. Following Sassoon, multidimensional adjectives can be read conjunctively (all dimensions must hold), disjunctively (one suffices), or in a pragmatically determined mixture. The paper argues "decentralized" behaves as the mixed case, which is why disputes about whether a system is decentralized are often disputes about which dimensions were selected and how they were bound together.

  • Distribution is used disjunctively; decentralization is not. A system may be called distributed if at least one relevant class of components is placed across multiple locations or entities, whereas decentralization does not tolerate the same looseness without specifying its dimensions and the standard for satisfying them.

  • A subject can be physically dispersed and still centralized under an anchor. Because centers are modeled as graph-realistic regions rather than single nodes, the framework can express cases the paper highlights, such as cryptocurrency nodes physically clustered in one data center or cloud region, while still being centralized with respect to ownership or authority.

  • Existing quantitative proxies measure concentration, not decentralization. The Nakamoto coefficient takes blockchain architecture for granted and derives control from node counts; Gini coefficient and Shannon entropy capture dispersion of resources or participation. All three see attributes more readily than relations, which the paper identifies as decisive since decentralization is relational.

  • Consensus and trade-off results formalize neighboring questions. Byzantine fault tolerance establishes that reliable consensus cannot be guaranteed once adversarial participants reach or exceed one third of the total, and CAP states that consistency, availability, and partition tolerance cannot all be guaranteed simultaneously — but both take the decentralized setting as given rather than defining it.

  • Cross-domain demonstration. The ontology is instantiated on blockchain models and federated learning, and the authors report that it produces consistent classifications and comparable assessments when the graph abstraction, evaluation profile, and center interpretations are disclosed.

Methodology in Plain English

The authors began with a structured literature review spanning topology and resilience, Byzantine consensus, CAP, blockchain trust-minimization, and concentration metrics, and used it to derive four requirements any transferable definition would need. They then analyzed how the word "decentralized" actually functions in language, borrowing from formal semantics the idea that adjectives like "healthy" need a dimension parameter and a rule for combining dimensions — and showed decentralization has the same structure but rarely gets its dimensions spelled out.

On that basis they built the definition in four explicitly separated layers: an ontological layer with commitments, competency questions, signature, and taxonomy; a logical layer with formal semantics and axioms for producing truth values; a modeling layer that binds the abstraction to graphs; and an analytical layer with the metrics. Two scope commitments shape everything: systems must be graph-representable, and evaluation happens at a fixed point in time, so a protocol with distinct phases is treated as a family of static topologies rather than one evolving one. Finally they encoded the framework in a browser-based sandbox that classifies graphs and computes the metrics, and ran it against blockchain and federated learning models to check that the outputs were consistent across domains.

The paper is conceptual and definitional. It reports no benchmark accuracies, dataset sizes, or performance numbers; the validation is a demonstration of classification consistency and metric computation, not an empirical competition with existing metrics.

Why This Matters

Impact on research. The work targets a methodological failure that undermines comparison: systems are routinely described as more or less decentralized without a shared rule, so cross-paper conclusions are often incompatible. By forcing every claim into a subject–anchor pair with a disclosed graph abstraction and evaluation profile, the framework makes decentralization claims falsifiable and comparable, and it corrects what the authors call an inversion in the literature where the definition follows from the application rather than the reverse.

Real-world applications:

  • Federated learning and collaborative training: deciding whether a deployment is genuinely decentralized, and distinguishing coordination bottlenecks (the aggregation server) from the privacy benefit of local data.
  • Blockchain and distributed ledger assessment: replacing or supplementing concentration metrics such as the Nakamoto coefficient with an evaluation that accounts for relationships and can be used beyond one architecture.
  • Decentralized AI and agentic systems: evaluating distributed inference pipelines and agent coordination where decentralization is a stated design objective but is usually asserted rather than shown.
  • Cloud and IoT infrastructure: surfacing cases where components are spread across locations but remain under a single owner or authority — the physical-distribution-versus-control gap the paper emphasizes.

Industry relevance. Teams building or auditing decentralized systems need defensible answers when they claim decentralization to regulators, users, or investors. A tool that classifies a system and reports a two-component vector, while disclosing the assumptions behind it, gives engineering and compliance groups something reviewable rather than rhetorical. The sandbox lowers the barrier to running such an assessment without building custom analysis.

Future Directions

  • Extending beyond graph-representable systems. The ontology explicitly restricts itself to systems that can be drawn as graphs of computationally capable entities; architectures that resist that encoding remain outside its reach.
  • Handling dynamics rather than snapshots. Because evaluation is fixed in time and phase changes are treated as separate topologies, a principled account of structural evolution — and of how to compare a family of phase topologies as one system — is left open.
  • Settling the non-uniqueness of the metrics. Void Tolerance and Imperviousness are presented as explicitly non-unique analytical degree functions, so how sensitive comparisons are to the chosen functions, and whether canonical choices can be justified, remains unresolved.
  • Standardizing comparison across evaluation profiles. The framework promises comparability only when the graph abstraction, evaluation profile, and center interpretations are disclosed, which raises the practical question of whether community conventions or reference profiles can be established for federated learning, blockchains, and decentralized AI.

Target Audience

Researchers and graduate students in distributed systems, blockchain, and decentralized machine learning who need to define or defend decentralization claims; protocol designers and architects choosing between centralized, hierarchical, and peer-to-peer coordination; and technical analysts, auditors, or standards groups who must evaluate whether a deployed system's decentralization claims hold up. Readers wanting an empirical benchmarking study should look elsewhere — this is a foundational, definitional contribution with a supporting tool.

Authors’ abstract

Decentralization as a concept in computer science has existed for over half a century. Despite its fundamental role across domains such as security, distributed computing, artificial intelligence, cloud infrastructures, and Internet of Things (IoT) architectures, there remains no universally accepted definition of decentralization applicable across computer communication systems. This has become increasingly problematic with the emergence of decentralized AI and machine…

The full abstract on arXiv
Read the original paper