Clusters-of-Centres (CoC) Concepts

• Clusters-of-Centres (CoC) is a term describing methods that coordinate centre-level summaries in both statistical inference and distributed storage systems.
• The statistical CoC method employs multivariate homogeneity tests, sequential merging, and multi-round bootstrap refinements to recover true clusters in federated settings.
• In distributed computing, CoC federates storage across geographically separate centres, optimizing resource utilization and ensuring unified system access.

Searching arXiv for the provided CoC-related papers to ground the article in current literature. Clusters-of-Centres (CoC) denotes at least two technically distinct research constructs on arXiv. In recent statistical methodology, CoC is a test-driven procedure for learning partitions of centres in federated or distributed inference from centre-level summaries, combining multivariate Cochran-type homogeneity tests with sequential merging and a multi-round bootstrap refinement (Debaly et al., 19 Sep 2025). In earlier distributed computing for high-energy physics, “Cluster-of-Centres” described a storage architecture that federates Disk Pool Manager services across geographically separated Tier-2 institutes so that multiple sites expose a single logical namespace and aggregated storage to experiments on the Worldwide LHC Computing Grid (0803.4223). The shared label reflects a common concern with coordinating multiple centres, but the objects being clustered, the mathematical structure, and the operational goals differ substantially.

1. Statistical CoC in federated inference

In the statistical usage, CoC addresses multi-centre studies in which sites share only centre-level summaries and parameter homogeneity across centres may fail. The method is designed to both test equality of centre-specific parameters and learn centre groupings before estimation (Debaly et al., 19 Sep 2025). The basic setup assumes $K$ centres, each providing $\widehat\theta_{n,k}\in\mathbb{R}^p$ together with estimated matrices $\widehat V_{n,k}$ and $\widehat Q_{n,k}$, under regularity conditions ensuring a Bahadur expansion of the local estimator: $$\sqrt n\,(\widehat\theta_{n,k}-\theta_{0,k}) = V_k^{-1}U_{n,k} \;+\;\varepsilon_{n,k}, \quad U_{n,k}\Rightarrow\mathcal N(0,Q_k), \quad \varepsilon_{n,k}=o_p(1).$$ Under the global null $H_0:\theta_{0,1}=\cdots=\theta_{0,K}=\theta_0$, the aggregated estimator is

$$\widehat\theta_n =\Bigl(\sum_{k=1}^K\widehat V_{n,k}\Bigr)^{-1} \sum_{k=1}^K\widehat V_{n,k}\,\widehat\theta_{n,k}.$$

The associated stacked contrast vector is

$$T_n = \bigl(\widehat\theta_n-\widehat\theta_{n,1},\;\dots,\;\widehat\theta_n-\widehat\theta_{n,K}\bigr)^\top\in\mathbb{R}^{Kp}.$$

The resulting test statistic is

$$S_n :=n\;T_n^\top\,\widehat{\overline V}_n^\top\,\widehat{\overline V}_n\,T_n \;\;\xRightarrow{H_0}\;\; \sum_{\ell=1}^{Kp}\widehat\lambda_{\ell,n}\,\chi^2_{1,\ell},$$

where the $\{\widehat\lambda_{\ell,n}\}$ are the eigenvalues of $\widehat\theta_{n,k}\in\mathbb{R}^p$0. Replacing the unknown $\widehat\theta_{n,k}\in\mathbb{R}^p$1 by the plug-in eigenvalues yields an asymptotically level-$\widehat\theta_{n,k}\in\mathbb{R}^p$2 test, while local alternatives of order $\widehat\theta_{n,k}\in\mathbb{R}^p$3 lead to a noncentral chi-square–mixture limit (Debaly et al., 19 Sep 2025).

A two-block analogue is defined for subsets $\widehat\theta_{n,k}\in\mathbb{R}^p$4. After aggregating each block to $\widehat\theta_{n,k}\in\mathbb{R}^p$5 and $\widehat\theta_{n,k}\in\mathbb{R}^p$6, equality is tested using

$\widehat\theta_{n,k}\in\mathbb{R}^p$7

again with an analogous $\widehat\theta_{n,k}\in\mathbb{R}^p$8-mixture limit. This two-block test is the primitive operation used by the clustering algorithm itself (Debaly et al., 19 Sep 2025).

The methodological significance of this formulation is that it operates entirely on summary statistics. This suggests a direct fit to federated settings in which raw observations remain local but centre-level estimators, sensitivities, and variability estimates can be exchanged.

2. Test-driven clustering and bootstrap refinement

The one-shot CoC algorithm is explicitly sequential and test-driven. At level $\widehat\theta_{n,k}\in\mathbb{R}^p$9, it first applies the global homogeneity test to all $\widehat V_{n,k}$0 centres. If homogeneity is not rejected, the output is a single cluster $\widehat V_{n,k}$1. Otherwise, the algorithm initializes $\widehat V_{n,k}$2, then processes centres $\widehat V_{n,k}$3 one at a time. For each existing cluster $\widehat V_{n,k}$4, it performs the two-block test of $\widehat V_{n,k}$5 versus $\widehat V_{n,k}$6, collects those clusters with p-value $\widehat V_{n,k}$7, and either creates a new singleton if none qualify or merges $\widehat V_{n,k}$8 into the cluster with the largest p-value, breaking ties by smallest index (Debaly et al., 19 Sep 2025).

The paper states two complementary finite-sample asymptotic features of this one-shot rule. First, it never merges heterogeneous centres, with power tending to one. Second, each pair of truly homogeneous centres fails to merge with probability up to $\widehat V_{n,k}$9 (Debaly et al., 19 Sep 2025). The latter motivates the bootstrap extension.

The multi-round bootstrap CoC fixes the original $\widehat Q_{n,k}$0 and generates $\widehat Q_{n,k}$1 i.i.d. bootstrap collections $\widehat Q_{n,k}$2 using any scheme satisfying conditional CLT, including nonparametric bootstrap, multiplier or weighted resampling, and a universal Gaussian resampling scheme: $\widehat Q_{n,k}$3 Round 1 applies one-shot CoC to the first bootstrap set to obtain $\widehat Q_{n,k}$4. For rounds $\widehat Q_{n,k}$5, the procedure starts from $\widehat Q_{n,k}$6 and re-evaluates every potential merge by two-block tests using the $\widehat Q_{n,k}$7-th resampled summaries, again merging when p-value $\widehat Q_{n,k}$8 with the same largest-p tie-break rule (Debaly et al., 19 Sep 2025).

Under Assumptions A1–A6 and the separation condition A7, the method satisfies the Golden-Partition Recovery property

$\widehat Q_{n,k}$9

where $$\sqrt n\,(\widehat\theta_{n,k}-\theta_{0,k}) = V_k^{-1}U_{n,k} \;+\;\varepsilon_{n,k}, \quad U_{n,k}\Rightarrow\mathcal N(0,Q_k), \quad \varepsilon_{n,k}=o_p(1).$$0 is the true centre partition (Debaly et al., 19 Sep 2025). In the terminology of the paper, the number of rounds grows with $$\sqrt n\,(\widehat\theta_{n,k}-\theta_{0,k}) = V_k^{-1}U_{n,k} \;+\;\varepsilon_{n,k}, \quad U_{n,k}\Rightarrow\mathcal N(0,Q_k), \quad \varepsilon_{n,k}=o_p(1).$$1, and the true partition is recovered with probability tending to one. A plausible implication is that the bootstrap rounds are not only a variance-reduction device but also a mechanism for correcting false splits introduced by the one-shot level-$$\sqrt n\,(\widehat\theta_{n,k}-\theta_{0,k}) = V_k^{-1}U_{n,k} \;+\;\varepsilon_{n,k}, \quad U_{n,k}\Rightarrow\mathcal N(0,Q_k), \quad \varepsilon_{n,k}=o_p(1).$$2 decision rule.

3. Assumptions, estimators, and implementation

The regularity structure is stated through Assumptions A1–A5 for the test statistics and A6–A7 for the bootstrap recovery theorem. Assumption A5 requires that each $$\sqrt n\,(\widehat\theta_{n,k}-\theta_{0,k}) = V_k^{-1}U_{n,k} \;+\;\varepsilon_{n,k}, \quad U_{n,k}\Rightarrow\mathcal N(0,Q_k), \quad \varepsilon_{n,k}=o_p(1).$$3 admit estimators $$\sqrt n\,(\widehat\theta_{n,k}-\theta_{0,k}) = V_k^{-1}U_{n,k} \;+\;\varepsilon_{n,k}, \quad U_{n,k}\Rightarrow\mathcal N(0,Q_k), \quad \varepsilon_{n,k}=o_p(1).$$4, ensuring that the statistics depending on $$\sqrt n\,(\widehat\theta_{n,k}-\theta_{0,k}) = V_k^{-1}U_{n,k} \;+\;\varepsilon_{n,k}, \quad U_{n,k}\Rightarrow\mathcal N(0,Q_k), \quad \varepsilon_{n,k}=o_p(1).$$5 remain asymptotically valid (Debaly et al., 19 Sep 2025). Assumptions A1–A4 are described as sufficient for the Bahadur representation at each centre.

The scope of these assumptions is broad within the paper. The authors state that they hold for smooth M-estimators, including GLMs and robust M-estimators, under standard second-order differentiability and moment conditions; for U- and V-statistics via Hoeffding decomposition; and for quantile regression and debiased high-dimensional Lasso via available empirical process results (Debaly et al., 19 Sep 2025). Each example in Section 3 is reported to verify A1–A4 and A5.

The practical guidance is intentionally tuning-light. The procedure needs no additional tuning parameters beyond $$\sqrt n\,(\widehat\theta_{n,k}-\theta_{0,k}) = V_k^{-1}U_{n,k} \;+\;\varepsilon_{n,k}, \quad U_{n,k}\Rightarrow\mathcal N(0,Q_k), \quad \varepsilon_{n,k}=o_p(1).$$6 and the choice of $$\sqrt n\,(\widehat\theta_{n,k}-\theta_{0,k}) = V_k^{-1}U_{n,k} \;+\;\varepsilon_{n,k}, \quad U_{n,k}\Rightarrow\mathcal N(0,Q_k), \quad \varepsilon_{n,k}=o_p(1).$$7, with $$\sqrt n\,(\widehat\theta_{n,k}-\theta_{0,k}) = V_k^{-1}U_{n,k} \;+\;\varepsilon_{n,k}, \quad U_{n,k}\Rightarrow\mathcal N(0,Q_k), \quad \varepsilon_{n,k}=o_p(1).$$8 given as an example to satisfy A7 (Debaly et al., 19 Sep 2025). For stopping the multi-round algorithm, the paper defines $$\sqrt n\,(\widehat\theta_{n,k}-\theta_{0,k}) = V_k^{-1}U_{n,k} \;+\;\varepsilon_{n,k}, \quad U_{n,k}\Rightarrow\mathcal N(0,Q_k), \quad \varepsilon_{n,k}=o_p(1).$$9, the number of clusters after round $H_0:\theta_{0,1}=\cdots=\theta_{0,K}=\theta_0$0, and notes that $H_0:\theta_{0,1}=\cdots=\theta_{0,K}=\theta_0$1 is nonincreasing and takes at most

$H_0:\theta_{0,1}=\cdots=\theta_{0,K}=\theta_0$2

distinct values. Accordingly, the sequence must stabilize. A plateau-based rule stops once $H_0:\theta_{0,1}=\cdots=\theta_{0,K}=\theta_0$3 remains constant for $H_0:\theta_{0,1}=\cdots=\theta_{0,K}=\theta_0$4 consecutive rounds, while practical implementations use a smaller window such as $H_0:\theta_{0,1}=\cdots=\theta_{0,K}=\theta_0$5 (Debaly et al., 19 Sep 2025). If computational governance or software libraries limit new resamples, stored replicates can be cycled while retaining the plateau check.

This implementation profile differentiates CoC from clustering procedures that require metric choices, penalties, or user-specified linkage criteria. Here the operative decisions are statistical tests on centre-level summaries.

4. Empirical behaviour in simulations and real data

The simulation study in Section 5.1 uses logistic regression on $H_0:\theta_{0,1}=\cdots=\theta_{0,K}=\theta_0$6 centres with a planted 3-cluster structure $H_0:\theta_{0,1}=\cdots=\theta_{0,K}=\theta_0$7, covariates i.i.d. $H_0:\theta_{0,1}=\cdots=\theta_{0,K}=\theta_0$8, mixed intercept shifts of order $H_0:\theta_{0,1}=\cdots=\theta_{0,K}=\theta_0$9, sample sizes $$\widehat\theta_n =\Bigl(\sum_{k=1}^K\widehat V_{n,k}\Bigr)^{-1} \sum_{k=1}^K\widehat V_{n,k}\,\widehat\theta_{n,k}.$$0, and 5000 Monte Carlo runs (Debaly et al., 19 Sep 2025). Three bootstrap schemes are compared for generating $$\widehat\theta_n =\Bigl(\sum_{k=1}^K\widehat V_{n,k}\Bigr)^{-1} \sum_{k=1}^K\widehat V_{n,k}\,\widehat\theta_{n,k}.$$1: nonparametric bootstrap, weighted exponential multipliers, and universal Gaussian resampling. The reported metrics are Adjusted Rand Index and rounds to plateau.

The simulation summary states that the weighted bootstrap gave the highest ARI, near 1, and fastest convergence; universal Gaussian was nearly identical in accuracy; and nonparametric resampling was slightly slower and less stable (Debaly et al., 19 Sep 2025). This suggests that, within the reported design, the bootstrap choice affects computational behaviour more than the recovered partition.

The real-data illustration uses the 2007 U.S. airline on-time performance data, with $$\widehat\theta_n =\Bigl(\sum_{k=1}^K\widehat V_{n,k}\Bigr)^{-1} \sum_{k=1}^K\widehat V_{n,k}\,\widehat\theta_{n,k}.$$2 flights at each of $$\widehat\theta_n =\Bigl(\sum_{k=1}^K\widehat V_{n,k}\Bigr)^{-1} \sum_{k=1}^K\widehat V_{n,k}\,\widehat\theta_{n,k}.$$3 destination airports (Debaly et al., 19 Sep 2025). The response is $$\widehat\theta_n =\Bigl(\sum_{k=1}^K\widehat V_{n,k}\Bigr)^{-1} \sum_{k=1}^K\widehat V_{n,k}\,\widehat\theta_{n,k}.$$4, and predictors are distance, day-of-week, month, and arrival-hour bins. Each centre shares its MLE $$\widehat\theta_n =\Bigl(\sum_{k=1}^K\widehat V_{n,k}\Bigr)^{-1} \sum_{k=1}^K\widehat V_{n,k}\,\widehat\theta_{n,k}.$$5, sensitivity $$\widehat\theta_n =\Bigl(\sum_{k=1}^K\widehat V_{n,k}\Bigr)^{-1} \sum_{k=1}^K\widehat V_{n,k}\,\widehat\theta_{n,k}.$$6, and variability $$\widehat\theta_n =\Bigl(\sum_{k=1}^K\widehat V_{n,k}\Bigr)^{-1} \sum_{k=1}^K\widehat V_{n,k}\,\widehat\theta_{n,k}.$$7. Across nonparametric, weighted, and universal resampling schemes, CoC is reported to find stable clusters such as $$\widehat\theta_n =\Bigl(\sum_{k=1}^K\widehat V_{n,k}\Bigr)^{-1} \sum_{k=1}^K\widehat V_{n,k}\,\widehat\theta_{n,k}.$$8 and $$\widehat\theta_n =\Bigl(\sum_{k=1}^K\widehat V_{n,k}\Bigr)^{-1} \sum_{k=1}^K\widehat V_{n,k}\,\widehat\theta_{n,k}.$$9, with only minor local differences such as how $$T_n = \bigl(\widehat\theta_n-\widehat\theta_{n,1},\;\dots,\;\widehat\theta_n-\widehat\theta_{n,K}\bigr)^\top\in\mathbb{R}^{Kp}.$$0 are arranged (Debaly et al., 19 Sep 2025).

For reference, the paper’s reported empirical settings are summarized below.

Setting	Reported specification	Reported outcome
Simulation	Logistic regression; $$T_n = \bigl(\widehat\theta_n-\widehat\theta_{n,1},\;\dots,\;\widehat\theta_n-\widehat\theta_{n,K}\bigr)^\top\in\mathbb{R}^{Kp}.$$1; planted 3-cluster structure $$T_n = \bigl(\widehat\theta_n-\widehat\theta_{n,1},\;\dots,\;\widehat\theta_n-\widehat\theta_{n,K}\bigr)^\top\in\mathbb{R}^{Kp}.$$2; $$T_n = \bigl(\widehat\theta_n-\widehat\theta_{n,1},\;\dots,\;\widehat\theta_n-\widehat\theta_{n,K}\bigr)^\top\in\mathbb{R}^{Kp}.$$3; 5000 MC runs	Weighted bootstrap highest ARI near 1; universal Gaussian nearly identical; nonparametric slightly slower/less stable
Real data	2007 U.S. airline on-time performance; $$T_n = \bigl(\widehat\theta_n-\widehat\theta_{n,1},\;\dots,\;\widehat\theta_n-\widehat\theta_{n,K}\bigr)^\top\in\mathbb{R}^{Kp}.$$4 flights at each of $$T_n = \bigl(\widehat\theta_n-\widehat\theta_{n,1},\;\dots,\;\widehat\theta_n-\widehat\theta_{n,K}\bigr)^\top\in\mathbb{R}^{Kp}.$$5 destination airports	Stable clusters across resampling schemes; only minor local differences

The empirical record in the paper therefore emphasizes partition stability under different resampling mechanisms, rather than sensitivity to hyperparameters, because the method is formulated to avoid most such hyperparameters in the first place.

5. Cluster-of-Centres in distributed Tier-2 storage

An earlier and independent use of the term appears in distributed high-energy computing. In that setting, a Cluster-of-Centres is a storage federation across distributed Tier-2 sites in which the storage of multiple institutes is unified and presented as a single system to experiments using the Worldwide LHC Computing Grid (0803.4223). The motivation is operational rather than inferential: geographically spread Tier-2s may expose CPU and storage in smaller units than experiments would ideally use, and mismatches between storage and CPU at individual centres can impede efficient exploitation (0803.4223).

The architecture described for the ScotGrid distributed Tier-2 uses Disk Pool Manager components at each site: the dpm server, dpnsd, rfiod, dpm-gsiftp, and SRM front-ends (0803.4223). By federating dpnsd across multiple institutes, all participating disk servers export a single “/scotgrid” namespace, so clients do not need to know the physical site location of data. Clients link against the RFIO client library and perform rfio_open and rfio_read calls into dpnsd to resolve logical paths. No kernel-level NFS mount is required; access is via GSI-authenticated RFIO or GridFTP. The summary also notes future exposure of the same namespace via NFS v4.1 or xrootd once implemented or GSI-enabled (0803.4223).

The network path between sites required explicit WAN traversal adjustments: opening TCP ports 5001, 5010, and 5015; tuning Linux TCP buffers to the bandwidth–delay product; and using the throughput estimate

$$T_n = \bigl(\widehat\theta_n-\widehat\theta_{n,1},\;\dots,\;\widehat\theta_n-\widehat\theta_{n,K}\bigr)^\top\in\mathbb{R}^{Kp}.$$6

to guide the choice of window size $$T_n = \bigl(\widehat\theta_n-\widehat\theta_{n,1},\;\dots,\;\widehat\theta_n-\widehat\theta_{n,K}\bigr)^\top\in\mathbb{R}^{Kp}.$$7 MiB for best average performance (0803.4223). The measured production-link characteristics were RTT $$T_n = \bigl(\widehat\theta_n-\widehat\theta_{n,1},\;\dots,\;\widehat\theta_n-\widehat\theta_{n,K}\bigr)^\top\in\mathbb{R}^{Kp}.$$8 ms and iperf peak $$T_n = \bigl(\widehat\theta_n-\widehat\theta_{n,1},\;\dots,\;\widehat\theta_n-\widehat\theta_{n,K}\bigr)^\top\in\mathbb{R}^{Kp}.$$9 Mb/s, corresponding to $$S_n :=n\;T_n^\top\,\widehat{\overline V}_n^\top\,\widehat{\overline V}_n\,T_n \;\;\xRightarrow{H_0}\;\; \sum_{\ell=1}^{Kp}\widehat\lambda_{\ell,n}\,\chi^2_{1,\ell},$$0 MiB/s (0803.4223).

Performance measurements showed the practical costs of “distant” access. For sequential reads, a single client achieved $$S_n :=n\;T_n^\top\,\widehat{\overline V}_n^\top\,\widehat{\overline V}_n\,T_n \;\;\xRightarrow{H_0}\;\; \sum_{\ell=1}^{Kp}\widehat\lambda_{\ell,n}\,\chi^2_{1,\ell},$$1–$$S_n :=n\;T_n^\top\,\widehat{\overline V}_n^\top\,\widehat{\overline V}_n\,T_n \;\;\xRightarrow{H_0}\;\; \sum_{\ell=1}^{Kp}\widehat\lambda_{\ell,n}\,\chi^2_{1,\ell},$$2 MiB/s over the WAN versus approximately $$S_n :=n\;T_n^\top\,\widehat{\overline V}_n^\top\,\widehat{\overline V}_n\,T_n \;\;\xRightarrow{H_0}\;\; \sum_{\ell=1}^{Kp}\widehat\lambda_{\ell,n}\,\chi^2_{1,\ell},$$3 MiB/s over the LAN, while aggregate throughput at 32 clients reached about $$S_n :=n\;T_n^\top\,\widehat{\overline V}_n^\top\,\widehat{\overline V}_n\,T_n \;\;\xRightarrow{H_0}\;\; \sum_{\ell=1}^{Kp}\widehat\lambda_{\ell,n}\,\chi^2_{1,\ell},$$4 MiB/s in RFIO_STREAM or READAHEAD modes before tailing off beyond 40 clients as disk seeks saturated (0803.4223). For partial reads, NORMAL mode outperformed buffered modes when skipping through data, reaching up to about $$S_n :=n\;T_n^\top\,\widehat{\overline V}_n^\top\,\widehat{\overline V}_n\,T_n \;\;\xRightarrow{H_0}\;\; \sum_{\ell=1}^{Kp}\widehat\lambda_{\ell,n}\,\chi^2_{1,\ell},$$5 MiB/s aggregate at approximately 18 clients, because buffered modes pre-fetch unwanted data (0803.4223). File open times increased linearly from about $$S_n :=n\;T_n^\top\,\widehat{\overline V}_n^\top\,\widehat{\overline V}_n\,T_n \;\;\xRightarrow{H_0}\;\; \sum_{\ell=1}^{Kp}\widehat\lambda_{\ell,n}\,\chi^2_{1,\ell},$$6 s at 5 clients to about $$S_n :=n\;T_n^\top\,\widehat{\overline V}_n^\top\,\widehat{\overline V}_n\,T_n \;\;\xRightarrow{H_0}\;\; \sum_{\ell=1}^{Kp}\widehat\lambda_{\ell,n}\,\chi^2_{1,\ell},$$7 s at 20 clients and exceeded $$S_n :=n\;T_n^\top\,\widehat{\overline V}_n^\top\,\widehat{\overline V}_n\,T_n \;\;\xRightarrow{H_0}\;\; \sum_{\ell=1}^{Kp}\widehat\lambda_{\ell,n}\,\chi^2_{1,\ell},$$8 s for 64 simultaneous opens, with error rate remaining below $$S_n :=n\;T_n^\top\,\widehat{\overline V}_n^\top\,\widehat{\overline V}_n\,T_n \;\;\xRightarrow{H_0}\;\; \sum_{\ell=1}^{Kp}\widehat\lambda_{\ell,n}\,\chi^2_{1,\ell},$$9 even under stress (0803.4223).

The reported best practices are correspondingly workload-specific: use RFIO_STREAM or READAHEAD for large sequential scans; switch to RFIO_NORMAL for sparse or indexed reads; leave IOBUFSIZE at 128 KiB unless the workload is pure streaming; and prefer aggregate parallelism over pushing a single stream’s TCP window above 1 MiB (0803.4223). Namespace federation is recommended via a single dpnsd serving a unified “/$\{\widehat\lambda_{\ell,n}\}$0VO$\{\widehat\lambda_{\ell,n}\}$1” namespace, with replicated metadata service for resilience and dynamic failover to avoid a single site outage taking down the CoC (0803.4223).

In this usage, “Cluster-of-Centres” does not denote a statistical clustering algorithm. It denotes a federated storage and namespace design for distributed computing centres.

6. Terminology, distinctions, and common conflations

The acronym “CoC” is not unique on arXiv. Besides Clusters-of-Centres and Cluster-of-Centres, it is also used for “Context Clustering” in “CoC-GAN: Employing Context Cluster for Unveiling a New Pathway in Image Generation” (Wang et al., 2023). That work treats an image as an unordered set of points, defines cosine-similarity-based clustering of point features around centres $\{\widehat\lambda_{\ell,n}\}$2, and integrates the resulting module with a Point Increaser inside a GAN (Wang et al., 2023). Its central objects are points, clusters, and generated images, not centres in a multi-site study or data centre federation.

The three usages can be separated by what is being grouped. In statistical CoC, the units are study centres and the output is a partition of centres inferred from summary-level estimators (Debaly et al., 19 Sep 2025). In Tier-2 Cluster-of-Centres, the units are computing sites whose storage is federated into a unified namespace and access fabric (0803.4223). In CoC-GAN, the units are image points clustered for feature aggregation and dispatch (Wang et al., 2023).

A common misconception is therefore to treat CoC as a single method family. The available arXiv evidence does not support that interpretation. The shared acronym masks unrelated technical lineages: distributed systems engineering in one case, summary-statistic inference in another, and point-set image generation in a third. A plausible implication is that citations to “CoC” require domain-specific disambiguation, especially in bibliographic databases or automated literature surveys.

7. Significance across domains

Within federated inference, CoC provides a fully test-driven and tuning-light mechanism for detecting heterogeneity and learning centre partitions using only centre-level summaries, with a bootstrap enhancement that yields golden-partition recovery under mild separation and increasing rounds of resampling (Debaly et al., 19 Sep 2025). Its significance lies in replacing ad hoc grouping with a sequence of formal homogeneity tests and in doing so without requiring raw-data pooling.

Within distributed high-energy computing, the Cluster-of-Centres model showed that federating DPM services across multiple Tier-2 centres and exposing a single logical namespace over GSI-authenticated RFIO could simplify VO data management and increase overall resource utilization, even over a WAN with RTT around 12 ms, provided that mode selection, buffering, security hardening, and monitoring were handled carefully (0803.4223). The paper reports aggregate read rates of $\{\widehat\lambda_{\ell,n}\}$3–$\{\widehat\lambda_{\ell,n}\}$4 MiB/s under such conditions (0803.4223).

Taken together, these works show that “Clusters-of-Centres” has become a domain-dependent term for coordination across distributed centres, but with distinct meanings determined by the surrounding technical problem. In statistics, it is a partition-learning procedure over centre-specific parameters; in grid computing, it is a unified storage architecture across geographically dispersed centres; and it should not be conflated with “Context Clustering” in image generation (Debaly et al., 19 Sep 2025, 0803.4223, Wang et al., 2023).