Effective Decentralisation Is the Minimum, Not the Average
A protocol cannot be more decentralised than its most concentrated effective control layer. Reporting one layer and ignoring the others is not measurement; it is selection bias.
Keywords: decentralisation, decentralization, Nakamoto coefficient, Gini coefficient, Herfindahl index, validator concentration, mining concentration, governance concentration, client diversity, reference implementation, sponsor concentration, exchange concentration, staking concentration, correlated failure, common control, effective decentralisation, protocol governance, blockchain measurement, layer concentration, institutional economics, infrastructure concentration
Decentralisation is reported as a number. It is not a number. It is a vector, and the vector has at least four components, and the economically meaningful summary of the vector is its minimum, not its average. The dominant practice in the field is to compute one component — usually a measure of consensus operator distribution — and to report it as if it summarised the whole. This is not measurement. It is selection of the most flattering dimension and presentation of it as if it answered a question it does not answer.
The argument of this essay is straightforward: a system cannot be more decentralised than its most concentrated effective control layer. Effective decentralisation is the minimum across layers, not the average. A protocol with thousands of independent validators and a single dominant client maintained by a small team is governed by the team, not by the validators, and reporting the validator count as the decentralisation metric is misleading. I will set out the layers, explain why the minimum is the binding constraint, address the standard metrics on their own terms, and end with what a complete decentralisation report would actually contain.
1. What “decentralisation” is supposed to mean
The economic content of decentralisation, in the context of a protocol economy, is that no small set of actors can unilaterally determine outcomes for the system. The reason this matters is not aesthetic. It matters because participants make investments under the system’s rules, and they need confidence that those rules cannot be revised, suspended, or selectively enforced by parties whose interests diverge from theirs. Decentralisation is, in this analytical sense, the property of having no actor or coalition with the unilateral capacity to act against the participants’ interests.
This is not the same as “many participants.” A system with a million participants can be highly centralised if a small number of those participants control the operations that matter. Counting participants is a proxy, sometimes a useful one, but the underlying question is about effective control, not headcount. The proxy fails whenever counting and control diverge, which is more often than the field acknowledges.
Several distinct dimensions of effective control exist in any operational protocol. They include who can submit transactions and have them processed, who can validate the consensus, who can revise the rule set, who can disrupt operations, and who can coordinate the actors who do any of the above. These are different dimensions and they are populated by different actors. A measurement programme that captures one dimension and ignores the others is not measuring decentralisation in the economic sense. It is measuring one of its components.
The temptation to collapse the dimensions into a single number is understandable. Single numbers are easier to communicate, easier to compare across systems, and easier to track over time. The cost is that the single number always represents one of the dimensions and silently treats the others as if they were either equal to it or irrelevant. They are typically neither.
2. The four layers
For analytical purposes, distinguish four layers at which control can be concentrated or dispersed.
Admission. The admission layer concerns who may participate in the system at all — who may submit transactions, who may run validating software, who may write applications on top of the protocol. A permissionless system has open admission; anyone meeting basic technical requirements may participate. A permissioned system has gated admission; participation requires authorisation from a controlling entity. Most public blockchains have nominally open admission at the protocol level, although effective admission to specific functions (operating an exchange node, running validator infrastructure at scale, accessing development resources) is often more restricted than the nominal openness suggests.
Consensus operation. The consensus operation layer concerns who actually performs the work of validating the chain — running miners, operating validators, producing blocks. This is the layer that the standard decentralisation metrics typically measure. The relevant dimensions include the number of independent operators, the distribution of capacity among them, and the correlations among their operations (shared clients, shared infrastructure, shared sponsors).
Governance. The governance layer concerns who has effective authority to revise the protocol rule set. This includes maintainers of reference implementations, sponsors who fund development, foundations that publish roadmaps, validators or miners whose adoption decisions determine activation, exchanges and infrastructure providers whose coordination is required for major changes, and token holders where on-chain governance applies. Governance authority is typically distributed across these categories, but the distribution is not uniform and is rarely captured by any single metric.
Mutability. The mutability layer concerns whether the base-layer rule set can be changed at all, and through what mechanism. A system with formal procedures for rule change has high mutability; a system whose base-layer rules are treated as fixed and whose changes are restricted to compatible extensions has low mutability. Mutability is not the same as governance — a system can have concentrated governance over a low-mutability rule set, in which case the governance authority is constrained — but the two interact.
These four layers are conceptually distinct, but in any operational system they interact. A system with open admission, dispersed consensus operation, concentrated governance, and high mutability has a particular profile. A system with permissioned admission, dispersed consensus operation, dispersed governance, and low mutability has a different profile. The two systems are not equivalent on a single decentralisation scale, and the differences matter for how participants should evaluate them.
3. The min-rule
The central claim of this essay is that effective decentralisation is the minimum across the layers, not the average. Formally:
Deff = min(Dadmission, Dconsensus, Dgovernance, Dmutability)
where each D is a measure of dispersion at the relevant layer, normalised so that higher values represent greater dispersion.
The reason for the minimum is that an actor or coalition seeking to act against the participants’ interests does not need to dominate every layer. It needs to dominate the layer at which its action operates. A coalition that controls the consensus layer can subvert transaction processing. A coalition that controls the governance layer can revise the rules. A coalition that controls the admission layer can deny participation. A coalition that controls the mutability constraint can enable revisions that would otherwise be foreclosed. Each layer is its own attack surface, and the system’s resistance to action by such a coalition is the minimum resistance across all the relevant surfaces.
This is not novel reasoning. The min-rule is the standard rule for analysing systems with sequential constraints in many economic contexts. The capacity of a supply chain is the capacity of its bottleneck. The strength of a chain is the strength of its weakest link. The security of a multi-step authentication is the security of its weakest step. In each case, the binding constraint is the minimum, not the average, because the system fails when any one constraint fails.
The same logic applies to decentralisation when the layers are sequential constraints on coordinated action against participants. To act against participants through transaction censorship, an actor needs control of the consensus layer. To act through rule revision, the actor needs control of the governance layer. To act through admission denial, the actor needs control of the admission layer. To act through retroactive revision, the actor needs the mutability constraint to permit it. Different attacks operate through different layers, but each layer is independently sufficient for some class of attacks. A system that is dispersed at one layer and concentrated at another is exposed to whichever attacks operate through the concentrated layer.
The averaging approach implicitly assumes that an attacker must overcome all layers simultaneously, so that strength at one layer compensates for weakness at another. This is wrong. The layers are alternatives, not complements. Strength at the consensus layer does not protect against governance attacks. Strength at the governance layer does not protect against consensus attacks. Strength at neither protects against admission attacks if the admission layer is concentrated.
The min-rule has a sharper consequence. A system’s effective decentralisation does not improve by improving its strongest layer; it improves only by improving its weakest. A protocol that has dispersed validators, dispersed token holdings, and dispersed network nodes but a single dominant client is bottlenecked at the client. Adding more validators does not help. Adding more nodes does not help. The bottleneck is the client, and the only intervention that increases Deff is one that disperses the client layer.
This is uncomfortable because it implies that many of the standard activities reported as decentralisation improvements have no effect on the binding constraint. Increasing validator count from 5,000 to 10,000 does not change Deff if governance was the binding layer at 5,000. The metric the system is reporting as improving is not the metric that bounds its actual decentralisation.
4. The standard metrics, layer by layer
The decentralisation metrics in common use can be classified by which layer they measure. Once classified, their incompleteness becomes visible.
Hash rate distribution and the Nakamoto coefficient at the mining or validation level. These measure consensus operation. They count the number of operators required to control more than half (or some other threshold) of consensus capacity. They capture Dconsensus, partially. They do not capture Dadmission, Dgovernance, or Dmutability. They are also incomplete measures of Dconsensus itself, because they treat operators as independent when they may share clients, infrastructure providers, sponsors, or jurisdictional exposure. Two operators using the same client are subject to a common point of failure that the operator count does not capture.
Validator count and stake distribution in proof-of-stake systems. Similar to the above. These measure consensus operation. They are subject to similar critiques: validator count does not adjust for client diversity, and stake distribution does not adjust for delegated stake or for stake held by entities under common control. A system with 100,000 validators all running the same client and operated by a small number of staking services has a much smaller effective Dconsensus than the validator count suggests.
Token Gini coefficient and concentration measures. These measure the distribution of token holdings, which is relevant for systems with on-chain governance via token voting. They capture an aspect of Dgovernance for those systems. They do not capture other aspects of governance — implementation authority, sponsor influence, foundation control — and they are subject to severe identification problems because token holdings are often held by entities under common control through multiple addresses, custody arrangements, or wrapping mechanisms.
Node count. This counts the number of full nodes running the protocol. It is sometimes presented as a decentralisation metric. It is, in fact, a metric of network observation rather than control. A node that observes the chain but does not produce blocks, vote on rules, or coordinate with other nodes is not exercising any form of authority. Counting nodes therefore measures something — perhaps the breadth of independent observation — but it does not measure decentralisation in the sense of dispersed control. The conflation of node count with decentralisation is one of the more durable category errors in the field.
Client diversity. The number of distinct client implementations and their relative usage shares. This is a measure of Dgovernance at one specific layer (implementation authority). It is informative because a system dominated by a single client is governed by the maintainers of that client at the implementation level, regardless of how dispersed other layers are. Few systems report this metric prominently, and where they do, the figures often show concentrations that other metrics obscure.
Geographic distribution. The distribution of operators or nodes by geographic location. This is sometimes presented as a decentralisation metric. It captures something — exposure to single-jurisdiction risks, primarily — but it does not capture control concentration. A system with operators evenly distributed across forty jurisdictions but operated by a single corporate group is not decentralised by virtue of its geographic spread. The geographic metric measures jurisdictional risk diversification, which is a different property.
Developer count and commit distribution. The number of contributors to the reference implementation and the distribution of commits among them. This measures one aspect of Dgovernance: the breadth of code contribution. It is informative but incomplete. Commit count does not measure who has merge authority, who has release authority, who has the authority to designate official versions, or who has the authority to direct contributors. A repository with hundreds of contributors and three maintainers with merge authority has its governance concentrated in the three, not the hundreds.
Each of these metrics is informative about something. None of them is sufficient. A complete decentralisation report would require all of them, and would require additional metrics that are rarely reported at all.
5. The metrics that are missing
Several dimensions of control concentration are routinely absent from public decentralisation reporting, and their absence is consequential.
Sponsor and funding concentration. Who pays the maintainers, sponsors the foundation, funds the development teams, and underwrites the infrastructure? Funding concentration is one of the most reliable predictors of governance influence, because the entities that fund development have leverage over development priorities, hiring, release timing, and roadmap content. The metric is rarely reported because the data are not always public, but in many cases the funding sources can be identified through public filings, foundation disclosures, and ecosystem reports. The exercise has been done occasionally and the results are typically surprising in the direction of greater concentration than the consensus metrics would suggest.
Adoption-critical infrastructure concentration. Which entities, if they failed to adopt a rule change, would prevent its activation? In most operational protocols, the answer is a small set: the largest exchanges, the largest staking services, the largest mining pools, and the dominant infrastructure providers. The concentration of adoption-critical infrastructure is a measure of governance bottleneck that is rarely captured by validator counts or token distributions.
Custody concentration. What proportion of tokens, particularly tokens used for staking or governance voting, is held in custodial arrangements that aggregate the holdings of many beneficial owners under the operational control of a few custodians? The custodian, not the beneficial owner, exercises operational control over staking, voting, and signalling. A system with widely distributed beneficial ownership but concentrated custodial control has its effective governance closer to the custodial concentration than to the beneficial concentration.
Cloud and hosting concentration. What fraction of consensus operators run their infrastructure on a small number of cloud providers or hosting services? A consensus layer with thousands of operators all running on three cloud providers has correlated failure modes — and correlated coercion possibilities — that the operator count does not capture. This metric is occasionally reported and the figures are typically concerning.
Legal jurisdiction concentration of identified actors. Where the entities exercising authority — maintainers, foundations, sponsors, custodians — are identifiable, where are they domiciled? A system whose governance actors are all subject to the law of a single jurisdiction is, for any matter that jurisdiction can reach, governed by that jurisdiction’s regulators and courts as well as by the system’s nominal governance processes. This is not necessarily a vulnerability; it depends on the jurisdiction and the matter. But it is an aspect of effective decentralisation that the standard metrics do not capture.
Social coordination concentration. Which forums, communication channels, and informal networks are used to coordinate governance discussions? A governance layer that nominally involves thousands of contributors but in which the substantive coordination occurs in a small set of private channels or among a small set of personal relationships has a different effective concentration than its public face suggests. This is hard to measure and rarely attempted.
None of these metrics is exotic. Each is grounded in identifiable economic relationships and could be reported with reasonable rigour by entities motivated to do so. The fact that they are typically not reported is partly a function of the difficulty of measurement and partly a function of the fact that what they measure is generally less flattering than the consensus metrics.
6. Correlation across layers and what it means
The min-rule treats the layers as independent, but the layers are correlated, and the correlations matter for how the overall system actually behaves.
A common correlation is between sponsor concentration and implementation concentration. Where a small set of sponsors funds development, those sponsors typically fund a small set of implementations, and the implementation concentration is downstream of the funding concentration. Reporting client diversity without reporting sponsor concentration captures only half of the structure.
Another correlation is between validator concentration and custody concentration. Where validation requires substantial operational expertise and capital, validation tends to be performed by professional services, and the professional services market itself is concentrated. The result is that nominal validator counts overstate effective validator independence, because many validators are operationally identical to a small number of underlying services.
A third correlation is between governance concentration and adoption concentration. The actors who adopt rule changes early (large exchanges, infrastructure providers, dominant clients) are often the same actors whose involvement in governance discussions has the most weight. The pattern is not a coincidence: actors with adoption authority have leverage in governance, and actors with governance leverage have incentives to maintain adoption authority. The two reinforce each other.
These correlations have a critical implication for the min-rule. The minimum across layers is not just the lowest of four independent figures; it is the lowest after accounting for correlations that make the layers less independent than they appear. A system that is dispersed at every layer measured separately may be more concentrated than the layer-by-layer figures suggest, because the same small set of actors appears across multiple layers under different labels.
This is one of the harder analytical problems in the field, and it is not obvious how to address it without identifying actors and tracing their roles across layers. The identification work has been done in some cases, typically by independent researchers rather than by the protocols themselves, and the results have generally been more concentrated than the standard metrics indicated. Where the work has not been done, the standard metrics should be treated as upper bounds on dispersion, not point estimates.
7. The practical consequences of the min-rule
If effective decentralisation is the minimum across layers, several practical consequences follow.
Improvement is layer-specific. A system that wishes to improve its effective decentralisation must improve its weakest layer. Resources spent on improving layers that are not the binding constraint produce no change in Deff. A protocol with a strong consensus layer and a weak governance layer cannot improve its decentralisation by adding more validators; it has to disperse its governance. This is not what most decentralisation improvement programmes actually do.
Comparison across systems is multi-dimensional. Two systems with the same nominal validator count or same Nakamoto coefficient may have very different effective decentralisations because their other layers differ. A complete comparison requires assessing all four layers, ideally with the additional metrics noted above. Single-metric rankings are systematically misleading because they collapse a vector to a scalar in a way that loses the binding information.
Reporting requirements should be mandatory. Where decentralisation is treated as a property the protocol claims, the claim should be supported by reports on all relevant layers, not on a self-selected subset. The current practice — selecting one favourable metric and reporting it as the decentralisation figure — is the kind of selective disclosure that in other contexts would be regarded as misleading.
External evaluation matters. Layer-by-layer assessment is hard to do credibly when conducted by parties with an interest in the result. Independent measurement of governance concentration, sponsor concentration, custody concentration, and adoption concentration is more reliable than self-reporting, particularly because the unfavourable metrics are typically the ones the protocol does not voluntarily disclose.
Architectural choices have decentralisation consequences. Some architectural decisions affect specific layers. A protocol that depends on a single reference client is exposed to client-layer concentration regardless of how its other layers are managed. A protocol that requires high operational expertise for validation is exposed to professional-service concentration at the consensus layer. A protocol that designates a foundation with substantial development control is exposed to foundation-layer concentration regardless of token distribution. The architectural decisions are choices, and they have measurable consequences for Deff.
Scaling does not automatically increase decentralisation. Adding participants increases D at the layer where they participate, but does not change D at other layers. A system that scales its user base, its application ecosystem, or its transaction throughput without dispersing its governance, its implementation authority, or its sponsor concentration becomes more centralised in the relative sense — more participants depending on a control structure of unchanged size. The growth in the dispersed layers does not compensate for stagnation in the concentrated layers when the binding constraint is the latter.
8. Standard objections
“The min-rule is too strict; it makes decentralisation impossibly hard to achieve.” The min-rule is not normative; it is descriptive. It does not say that achieving high Deff is impossible. It says that achieving high Deff requires achieving high D at every relevant layer. If that is hard, then high Deff is hard, and reports that suggest a system has achieved it through high D at one layer alone are wrong about the achievement, not about the standard. The standard is what it is because the layers are alternative attack surfaces, not because the analyst chose to make life difficult.
“Some layers matter more than others, and the relevant metric should weight them.” Weighting requires a rationale, and the rationale must come from the analytical purpose. If the purpose is to assess vulnerability to attacks operating through any layer, weighting is hard to justify because each layer is an attack surface and the system fails when any layer fails. If the purpose is something else — for example, expected harm averaged across attack scenarios — weighting can be defended, but the weights must be derived from the probabilities and consequences of attacks at each layer, not chosen for analytical convenience. The standard practice of weighting by ease of measurement is not defensible on either ground.
“Governance and consensus are different kinds of decentralisation; they should be reported separately rather than combined into a single figure.” This is partly correct and partly wrong. Reporting layer-specific figures separately is preferable to averaging or to reporting a single figure. But for participants who must make decisions on the basis of the reports, the layer-specific figures need to be combined into a summary, and the relevant summary is the minimum, because the minimum is the binding constraint on the attacks the participant is exposed to. Reporting the layers separately and refusing to summarise is a partial improvement; reporting the layers separately and summarising by the minimum is the full improvement.
“Many of the additional metrics are unmeasurable in practice.” Some are hard to measure precisely; few are unmeasurable in principle. Sponsor concentration, custody concentration, cloud concentration, and adoption-critical infrastructure concentration can all be measured with available data, sometimes with substantial effort. Where the measurement is approximate, ranges are still informative. The fact that a metric is hard to measure is not a reason to omit it; it is a reason to report it with appropriate uncertainty bounds. Replacing a hard metric with an easier one that captures something different is a worse alternative than reporting the hard metric imprecisely.
“The min-rule treats decentralisation as a binary property, which is too crude.” The min-rule does not treat decentralisation as binary. It treats decentralisation as a continuous variable whose value is bounded above by the lowest of several layer-specific continuous variables. The bound is sharp, but the underlying scale is continuous. A system can have Deff = 0.3 because every layer is 0.3, or because three layers are 0.9 and one is 0.3. The implications are different — the second system has more room to improve through layer-specific intervention — but the bound is the same.
“Centralisation at one layer can be benign or even desirable; the min-rule conflates concentration with vulnerability.” Whether concentration at a particular layer is concerning depends on the layer, on the actors who are concentrated there, and on the constraints under which they operate. A concentrated reference implementation maintained by a team subject to strong fiduciary duties, transparent decision-making, and credible accountability is less concerning than a concentrated reference implementation maintained without those constraints. The min-rule does not prejudge which concentration is concerning; it identifies where concentration is, and leaves to further analysis whether the constraints on the concentrated actors are sufficient to make the concentration tolerable. The analysis becomes possible only after the concentration is identified.
9. What complete reporting would look like
A complete decentralisation report for an operational protocol would have several components.
First, a layer-by-layer dispersion measurement, with metrics chosen for the relevant layer. For consensus, validator or miner count and stake or hash distribution, adjusted for client diversity and shared infrastructure. For governance, implementation maintainer count and merge authority distribution, sponsor count and funding distribution, foundation control structure, and adoption-critical infrastructure concentration. For admission, the formal openness of participation and any practical barriers to specific functions. For mutability, the formal procedures for rule change, the historical frequency of changes, and the typical coordination structure of significant revisions.
Second, an identification of cross-layer correlations. Which actors appear in multiple layers? What proportion of consensus capacity is operated by entities that also exercise governance authority? What proportion of governance authority is held by entities that fund the implementation? The correlations should be reported as adjustments to the layer-specific figures, not separately.
Third, an explicit identification of the binding constraint. Which layer has the lowest D, and what does its concentration look like? This is the layer at which the system is bottlenecked, and the layer at which improvement would actually increase Deff. Reporting this clearly invites the question of why the bottleneck is where it is, which is a productive question even when the answer is uncomfortable.
Fourth, a comparison to the actor population that would be required to act adversely against participants at each layer. Effective decentralisation is meaningful insofar as it makes adverse coordination expensive. The relevant question is not just “how many independent actors exist?” but “how few of them would be required to act jointly to cause harm of various kinds?” The answer is layer-specific and is the operational meaning of D at each layer.
Fifth, an acknowledgement of what is not measured. There are aspects of effective control that are difficult to quantify — informal social coordination, personal relationships among key actors, unstated agreements about decision-making — and any honest report should acknowledge that the quantifiable metrics underestimate concentration in the presence of these informal mechanisms. The acknowledgement is not a failure of the report; it is a feature of honest measurement.
Reports of this form are rare. The reasons they are rare include the difficulty of measurement, the unfavourable picture they typically paint, and the absence of any external entity with both the authority and the incentive to require them. None of these is a good reason; they are explanations for the gap between what could be reported and what typically is. The gap is one of the more serious data problems in the field.
10. Closing
Decentralisation is not a number. It is a vector. The vector has at least four components: admission, consensus operation, governance, and mutability. The economically meaningful summary of the vector is its minimum, not its average, because the layers are alternative attack surfaces and the system’s resistance to adverse coordination is the resistance at its weakest layer. Standard reporting collapses the vector to a single component — usually consensus operation — and treats the component as if it were the whole. This practice systematically misranks systems and gives participants information that is precise about the wrong question.
The min-rule is not a refinement of standard practice. It is a different framework, derived from the recognition that the layers are sequential constraints rather than alternative substitutes, and that the binding constraint at any moment is the most concentrated layer. A system with thousands of independent validators and a single dominant client is governed at the client level, not at the validator level. A system with diffuse token holdings and concentrated custody is governed at the custodian level, not at the holder level. A system with open admission and concentrated governance is open at the door and closed at the rule book. In each case, the metric the system is reporting as decentralisation captures the dispersed layer rather than the concentrated one, and the participant who relies on the metric is making decisions on the basis of an incomplete picture.
What follows for measurement is that complete decentralisation reporting requires layer-by-layer assessment, identification of cross-layer correlations, and explicit identification of the binding constraint. The relevant metrics include not only the consensus-layer measures that dominate current reporting but also sponsor concentration, custody concentration, adoption-critical infrastructure concentration, client diversity, and several others that the standard reporting omits. The combined report is harder to produce than the standard report, but the standard report is partly so easy because it omits what is hard.
What follows for analysis is that decentralisation cannot be substituted for the underlying property it is supposed to proxy. The underlying property is the cost of adverse coordination by a small set of actors against participant interests, and that cost is determined by the most concentrated layer of effective control. A system can score well on the easiest-to-measure metric and score poorly on the binding metric, in which case its reported decentralisation overstates its actual resistance to adverse coordination. The analysis required to detect this is more work than the analysis required to compute the standard metric, but it is the analysis that answers the question the standard metric is supposed to answer.
What follows for design is that improving effective decentralisation requires identifying the binding constraint and acting on it. Improvements at non-binding layers do not change Deff. A system whose governance layer is the binding constraint cannot improve its decentralisation by adding consensus operators; it has to disperse its governance. This is harder than scaling the consensus layer, which is part of why the latter is the more frequently attempted improvement. The fact that something is hard does not make doing something easier instead of it equivalent. The binding constraint is the binding constraint.
The basic claim of this essay can be stated as a single equation. Effective decentralisation is the minimum across the layers of effective control: Deff = min(Dadmission, Dconsensus, Dgovernance, Dmutability). Every other formulation either implicitly averages across layers — which is incorrect when the layers are alternative attack surfaces — or measures one layer and treats it as the whole, which is also incorrect. The min-rule is the rule that follows from the analytical purpose decentralisation is supposed to serve. Anything else is measurement of something other than what the term claims to measure.
References mentioned in passing: O. E. Williamson, The Economic Institutions of Capitalism (1985); D. C. North, Institutions, Institutional Change and Economic Performance (1990); A. O. Hirschman, Exit, Voice, and Loyalty (1970); R. H. Coase, “The Nature of the Firm,” Economica (1937).