Multiple Consensus Technique

• Multiple consensus technique is a set of approaches that construct agreement from multiple partitions, views, or agents rather than a single aggregation step.
• It addresses structural heterogeneity by integrating methods like ensemble clustering, iterative averaging, and rank-one nonnegative matrix factorization.
• Applications span network community detection, phylogenetics, decentralized optimization, and multi-agent decision-making, offering refined summaries for complex data.

Multiple consensus technique denotes a family of agreement procedures in which consensus is constructed from multiple partitions, multiple views, multiple agents, multiple communication rounds, or multiple ethical criteria rather than from a single aggregation step. Across the literature, the term covers multiresolution consensus clustering in networks, iterative consensus over ensembles of clustering algorithms, clustering-based inference of multiple consensus trees and supertrees, repeated averaging in decentralized optimization, ordinal agreement across several metrics, multi-agent deliberation, and layered blockchain validation (Jeub et al., 2017, Race et al., 2014, Makarenkov et al., 2023, Liu et al., 7 Nov 2025).

1. Conceptual scope and recurring motivations

The principal motivation for multiple-consensus methods is that a single “best” summary is often too coarse. In network community detection, stochastic optimization and multiscale structure make one partition insufficient, so consensus is built from an ensemble of partitions sampled across resolutions (Jeub et al., 2017). In phylogenetics, one consensus tree can collapse into a star tree when gene trees are genuinely heterogeneous, so trees are first clustered and then summarized cluster by cluster (Makarenkov et al., 2023). In decentralized optimization, one communication step per iteration may not diffuse information sufficiently, especially when agents perform several local updates between communication rounds, so multiple consensus steps are introduced to control drift and disagreement (Zhang et al., 13 Aug 2025).

Taken together, these formulations suggest a common pattern: multiple consensus is used when disagreement is structural rather than incidental. The relevant heterogeneity may come from scale parameters, algorithmic stochasticity, incompatible tree topologies, nonconvex data heterogeneity, asynchronous communication, unequal ethical principles, or distinct reasoning styles among agents. The resulting consensus object is likewise domain-dependent: a hierarchical community tree, a final clustering, a supertree, a stationary point with agentwise agreement, a maximal ordinally consistent subset, or a unanimous or graded decision (Jeub et al., 2017, Liu et al., 7 Nov 2025, Salas-Molina et al., 2024, Pokharel et al., 2 Apr 2025).

2. Ensemble-based clustering and community detection

A major line of work treats multiple consensus as aggregation over an ensemble of clusterings. In multiresolution community detection, the input ensemble is generated by optimizing the Reichardt–Bornholdt modularity

$$Q(\vec{g}, \gamma)=\sum_{i,j=1}^{n} \left(A_{ij}-\gamma P_{ij}\right)\delta(g_i,g_j)$$

over a carefully sampled range of resolution parameters $\gamma$. The ensemble is summarized by the co-classification matrix

$$C_{ij}(\mathbf{g}) = \frac{1}{|\mathbf{g}|}\sum_{t=1}^{|\mathbf{g}|}\delta(g_i^{(t)},g_j^{(t)}),$$

and consensus is obtained not by thresholding $C$, but by optimizing a modified modularity with a null model derived from the ensemble itself. The procedure is then applied recursively within detected clusters to produce a hierarchical consensus structure (Jeub et al., 2017).

A related but broader framework is Iterative Consensus Clustering. There the ensemble may combine multiple algorithms, multiple dimensionality reductions, and multiple values of $k$. The consensus matrix is defined by pairwise co-occurrence counts,

$$M(C)_{ij} = \# \text{ of times object } x_i \text{ was placed in the same cluster as } x_j \text{ in the ensemble } C,$$

with an optional intolerance parameter $\tau$ that removes weakly supported edges. The matrix is viewed as a graph, a random walk on that graph yields a transition matrix $P=D^{-1}M$, and the number of clusters is estimated from the Perron cluster near eigenvalue $1$. The voting process can then be iterated until a majority solution emerges (Race et al., 2014).

A third formulation replaces pairwise co-occurrence by an object–candidate preference matrix

$A_{ij}= \begin{cases} w_j & \text{if the %%%%6%%%%-th element belongs to the %%%%7%%%%-th group},\ 0 & \text{otherwise}, \end{cases}$

and seeks consensus as a bi-cluster in that matrix. The resulting framework unifies clustering, community detection, and multiple parametric model estimation, and uses iterative rank-one nonnegative $\gamma$0-NMF rather than binary approximation of an $\gamma$1 co-occurrence matrix (Tepper et al., 2014).

A fourth variant, specialized to hierarchical clustering ensembles, generates multiple dendrograms by bagging, converts each to a Cophenetic Difference matrix, and searches for a weight vector $\gamma$2 with $\gamma$3. The consensus matrix is

$\gamma$4

and the search is guided by a genetic algorithm whose fitness is the cophenetic correlation coefficient between the candidate consensus and the original Euclidean distance matrix. The reported winning frequency is $\gamma$5 across the experiments summarized in the paper (Rashedi et al., 2018).

Setting	Aggregated object	Consensus mechanism
Multiresolution network clustering	Co-classification matrix $\gamma$6	Modified modularity and recursive hierarchical consensus
Iterative consensus clustering	Consensus matrix $\gamma$7	Perron-cluster analysis and repeated reclustering
Bi-clustering framework	Preference matrix $\gamma$8	Rank-one nonnegative $\gamma$9-NMF bi-clustering
Hierarchical ensemble fusion	CD matrices $$C_{ij}(\mathbf{g}) = \frac{1}{|\mathbf{g}|}\sum_{t=1}^{|\mathbf{g}|}\delta(g_i^{(t)},g_j^{(t)}),$$0	GA-weighted averaging into $$C_{ij}(\mathbf{g}) = \frac{1}{|\mathbf{g}|}\sum_{t=1}^{|\mathbf{g}|}\delta(g_i^{(t)},g_j^{(t)}),$$1

These approaches differ in representation, but they share a shift from single-run output to ensemble-derived evidence. They also differ in how arbitrariness is handled. Multiresolution consensus replaces ad hoc thresholding by a significance-based null model (Jeub et al., 2017); ICC replaces a fixed $$C_{ij}(\mathbf{g}) = \frac{1}{|\mathbf{g}|}\sum_{t=1}^{|\mathbf{g}|}\delta(g_i^{(t)},g_j^{(t)}),$$2 by Perron-cluster estimation (Race et al., 2014); bi-clustering retains higher-order object–group structure rather than collapsing it to pairwise overlap (Tepper et al., 2014); and GA-based fusion treats the fusion function itself as an optimization variable (Rashedi et al., 2018).

3. Multiple consensus in decentralized optimization and networked control

In decentralized optimization, multiple consensus usually means repeated communication or averaging between local computation phases. The clearest recent example is MILE, which studies fully decentralized nonconvex optimization

$$C_{ij}(\mathbf{g}) = \frac{1}{|\mathbf{g}|}\sum_{t=1}^{|\mathbf{g}|}\delta(g_i^{(t)},g_j^{(t)}),$$3

Each agent performs $$C_{ij}(\mathbf{g}) = \frac{1}{|\mathbf{g}|}\sum_{t=1}^{|\mathbf{g}|}\delta(g_i^{(t)},g_j^{(t)}),$$4 local updates between communication rounds, and at communication times transmits only

$$C_{ij}(\mathbf{g}) = \frac{1}{|\mathbf{g}|}\sum_{t=1}^{|\mathbf{g}|}\delta(g_i^{(t)},g_j^{(t)}),$$5

The paper’s main technical device is a periodic-system formulation with lifting, yielding a closed-form evolution across one period and the stability condition $$C_{ij}(\mathbf{g}) = \frac{1}{|\mathbf{g}|}\sum_{t=1}^{|\mathbf{g}|}\delta(g_i^{(t)},g_j^{(t)}),$$6. Under exact gradients, MILE proves both optimality and consensus with an $$C_{ij}(\mathbf{g}) = \frac{1}{|\mathbf{g}|}\sum_{t=1}^{|\mathbf{g}|}\delta(g_i^{(t)},g_j^{(t)}),$$7 convergence rate; under stochastic gradients it proves bounded-neighborhood guarantees (Liu et al., 7 Nov 2025).

A closely related online saddle-point setting uses a distributed online stochastic mirror descent convex-concave optimization algorithm with predictive mappings. The baseline performs one local mirror-descent step and one communication step, whereas the multiple-consensus variant repeats the averaging $$C_{ij}(\mathbf{g}) = \frac{1}{|\mathbf{g}|}\sum_{t=1}^{|\mathbf{g}|}\delta(g_i^{(t)},g_j^{(t)}),$$8 times: $$C_{ij}(\mathbf{g}) = \frac{1}{|\mathbf{g}|}\sum_{t=1}^{|\mathbf{g}|}\delta(g_i^{(t)},g_j^{(t)}),$$9 Equivalently, it replaces one-step mixing by powers of the mixing matrix $C$0. The dynamic saddle point regret keeps the same asymptotic structure but improves the network-dependent factor when $C$1, at the cost of higher communication overhead (Zhang et al., 13 Aug 2025).

The NEAR-DGD family studies the same design question from the standpoint of nested consensus and gradient operators. With $C$2 consensus steps and $C$3 gradient steps per outer iteration,

$C$4

The analysis shows a sharp asymptotic distinction: fixed finite $C$5 yields convergence only to a neighborhood of the optimizer, whereas increasing consensus steps over time while reducing gradient repetitions to one yields $C$6-linear convergence to the exact solution (Berahas et al., 2020).

In control-theoretic consensus over clustered networks, multiple consensus appears as intermittent and asynchronous output-feedback updates. Agents are identical LTI systems, cluster and inter-cluster subgraphs have independent timers, disagreement information is stored in auxiliary variables $C$7 and $C$8, and the closed-loop ensemble is modeled as a hybrid system. A Lyapunov-based analysis gives sufficient conditions under which the agreement subspace is globally exponentially stable, and an input-to-state stability argument shows robustness to a class of perturbations (Zegers et al., 2024).

An earlier network-dynamics formulation addresses unconstrained gossip when nodes receive multiple simultaneous transmissions. Conflict resolution converts a transmission matrix $C$9 into a row-stochastic binary adoption matrix $k$0, preventing semantically invalid linear combinations of states. The resulting process can be analyzed as an absorbing Markov chain, which yields exact probabilities of converging to specific consensus states and expected times to consensus (Hollander et al., 2013).

4. Trees, metrics, and multi-model structural synthesis

In phylogenetics, multiple consensus is motivated by the failure of a single consensus tree when input trees are strongly discordant. The review literature distinguishes strict consensus, majority-rule consensus, extended majority-rule consensus, and Nelson consensus, and argues that these can become highly unresolved or degenerate into a star tree when gene trees reflect different histories induced by horizontal gene transfer, hybridization, recombination, or gene duplication and loss. The alternative is to cluster topologically similar trees—typically using the Robinson–Foulds distance $k$1—and infer one consensus tree or supertree per cluster. The same idea extends from trees on identical taxa to supertrees on overlapping taxon sets (Makarenkov et al., 2023).

The mathematical justification for this approach is that the square root of the Robinson–Foulds distance has Euclidean properties, which makes k-means-style formulations workable. The review discusses consensus-based and pairwise-distance-based objective functions, as well as tree-adapted validity indices such as Caliński–Harabasz, Ball–Hall, Silhouette, and Gap. It also emphasizes important limitations: results depend on the chosen distance, the number of clusters, and the validity criterion, and the centroid in k-means tree clustering is not necessarily itself a phylogenetic tree or a consensus tree (Makarenkov et al., 2023).

A more abstract version of multiple consensus appears in the study of metric systems $k$2, where $k$3 are metrics on the same point set. The objective is to find the largest subset $k$4 on which all metrics are ordinally consistent. Strong consistency requires exact agreement on every comparison between two pairwise distances; weak consistency uses plurality over the relations $k$5, $k$6, and $k$7, and deems a subset consistent if the induced relation set is valid. The paper characterizes weak validity via an auxiliary mixed graph and proves that the relation set is valid iff the auxiliary graph contains no directed cycle. Both strong and weak versions are NP-complete in restricted metric families, and the paper gives a $k$8-approximation for the minimum strongly inconsistent subset and a $k$9-approximation for the minimum weakly inconsistent subset (Wang et al., 2020).

A related structural view treats consensus over groupings as pattern discovery in an object-by-candidate relation space rather than in an averaged co-occurrence matrix. This perspective is especially important for multiple parametric model estimation, where each candidate group is the consensus set of a hypothesized model and bi-clustering directly extracts coherent subsets of observations and model hypotheses, including overlapping structures (Tepper et al., 2014).

5. Decision making, ethics, and deliberative multi-agent agreement

In decision-theoretic multi-agent systems, a classical multiple-consensus technique is centralized majority voting. One coordinating agent polls $$M(C)_{ij} = \# \text{ of times object } x_i \text{ was placed in the same cluster as } x_j \text{ in the ensemble } C,$$0 agents solving the same dichotomous problem, and the probability that the consensus judgment is correct is

$$M(C)_{ij} = \# \text{ of times object } x_i \text{ was placed in the same cluster as } x_j \text{ in the ensemble } C,$$1

Under equal competence and equal prior odds, the paper states that if $$M(C)_{ij} = \# \text{ of times object } x_i \text{ was placed in the same cluster as } x_j \text{ in the ensemble } C,$$2, then $$M(C)_{ij} = \# \text{ of times object } x_i \text{ was placed in the same cluster as } x_j \text{ in the ensemble } C,$$3 and increases monotonically with $$M(C)_{ij} = \# \text{ of times object } x_i \text{ was placed in the same cluster as } x_j \text{ in the ensemble } C,$$4, while if $$M(C)_{ij} = \# \text{ of times object } x_i \text{ was placed in the same cluster as } x_j \text{ in the ensemble } C,$$5, then $$M(C)_{ij} = \# \text{ of times object } x_i \text{ was placed in the same cluster as } x_j \text{ in the ensemble } C,$$6 and decreases with $$M(C)_{ij} = \# \text{ of times object } x_i \text{ was placed in the same cluster as } x_j \text{ in the ensemble } C,$$7. Extensions cover heterogeneous competence classes and unequal prior odds, leading to the criterion $$M(C)_{ij} = \# \text{ of times object } x_i \text{ was placed in the same cluster as } x_j \text{ in the ensemble } C,$$8 for prior-adjusted correctness $$M(C)_{ij} = \# \text{ of times object } x_i \text{ was placed in the same cluster as } x_j \text{ in the ensemble } C,$$9 (O'Leary, 2013).

A distinct normative line of work computes consensus in group decision making under multiple ethical principles. The consensus judgment $\tau$0 is obtained by minimizing a weighted sum of several $\tau$1-metric disagreement terms, where $\tau$2 corresponds to maximum freedom or Benthamite utilitarianism and $\tau$3 corresponds to maximum fairness or Rawlsian maximin. The key technical contribution is a multi-$\tau$4-norm approximation problem together with a re-weighting scheme that divides each norm by the value it attains in the corresponding single-norm optimization. This is introduced to prevent the unweighted formulation from being heavily biased toward the $\tau$5 ethical principle (Salas-Molina et al., 2024).

In LLM-based multi-agent systems, deliberation can itself be the consensus mechanism. One blockchain-oriented proposal models a deliberation game in which multiple LLM agents generate initial responses, revise them over several rounds, and converge either to unanimous consensus for definitive problems or graded consensus for policy problems. For policy sets, agreement on a policy $\tau$6 is quantified by

$\tau$7

and accepted policies are aggregated into a confidence score $\tau$8. The paper maps the procedure to blockchain properties such as consistency, agreement, liveness, and determinism, and reports a $\tau$9 success rate on arithmetic problems in the tested subset (Pokharel et al., 2 Apr 2025).

Another facilitation-oriented use of multiple consensus appears in automated online consensus building through parallel thinking. The Parallel Thinking-based Facilitation Agent performs the six roles of the Six Thinking Hats technique—White, Red, Black, Yellow, Green, and Blue—in a text-based online discussion. The intended logic is a two-phase process of divergent thinking and convergent thinking, but the reported pilot study finds that the system often stays too long in idea-generation mode, sometimes interrupting an in-progress consensus or introducing new suggestions after agreement had emerged (Gu et al., 16 Mar 2025).

Consensus can also be made easier by weakening disagreement rather than resolving it directly. In a three-valued framework based on Kleene logic, agents assign $P=D^{-1}M$0, $P=D^{-1}M$1, or $P=D^{-1}M$2 to propositions, and a consensus operator replaces direct truth-value conflicts by the borderline value $P=D^{-1}M$3. Simulations show convergence to a smaller set of shared beliefs, and when the choice of interacting agents is payoff-dependent, the system converges to beliefs with higher average payoff (Crosscombe et al., 2016).

A layered blockchain formulation pushes the idea further by separating explicit agreement from semantic validity. In implicit consensus, nodes run conventional BFT only on checkpoint blocks, whereas transactions are validated locally by collecting proofs from the relevant personal blockchains. Consensus is thus explicit for checkpoints and implicit for validated transactions, with the design relying on self-interest and spontaneous sharding rather than on global agreement over every transaction (Ren et al., 2017).

6. Limitations, trade-offs, and consensus-aware divergence

Multiple-consensus methods are not uniformly beneficial, and the literature is unusually explicit about their failure modes. In network community detection, hierarchical consensus can produce some spurious intermediate levels, even though the finest and/or best cuts are often strong, and the local permutation null model behaves better than the simple permutation model for recursive splitting (Jeub et al., 2017). In phylogenetic tree clustering, outcomes depend on the chosen distance, number of clusters, and validity criterion, and the field remains comparatively underdeveloped for direct supertree clustering (Makarenkov et al., 2023).

In distributed optimization and control, additional consensus rounds tighten disagreement bounds but increase communication load. The online convex-concave setting states this trade-off directly: repeated averaging can better approximate global agreement and reduce consensus error, but the improvement comes at the cost of higher communication overhead (Zhang et al., 13 Aug 2025). The NEAR-DGD analysis likewise shows that more local gradient steps can accelerate early progress while enlarging the asymptotic neighborhood, whereas exact convergence requires increasing consensus steps over time (Berahas et al., 2020). Control-theoretic clustered consensus adds further requirements in the form of timer bounds, graph assumptions, and matrix inequalities that ensure global exponential stability and ISS robustness (Zegers et al., 2024).

Decision and deliberation settings expose a different set of constraints. Majority voting is actively harmful when individual competence is below chance, and even heterogeneous extensions require competence-sensitive agent selection (O'Leary, 2013). LLM-deliberation systems face hallucinations, degeneration of thought, malicious models and nodes, resource consumption, and scalability (Pokharel et al., 2 Apr 2025). Facilitation agents based on parallel thinking remain sensitive to phase awareness, timing, intervention rate, and sensitivity to the discussion state (Gu et al., 16 Mar 2025).

A notable recent inversion of the theme is consensus-aware divergence. The Semantic Repulsion Technique does not seek a shared answer; instead, it estimates a consensus centroid from $P=D^{-1}M$4 sampled outputs, visualizes the consensus region as a “Yellow Zone,” and pushes generation away from that region using contrastive decoding, phrase-level penalties, and fluency controls. In the reported computational assessment, SRT increases semantic diversity by $P=D^{-1}M$5 while reducing consensus phrases by $P=D^{-1}M$6 across task modes; in the user study, originality and coherence ratings were positively correlated across all systems with $P=D^{-1}M$7 to $P=D^{-1}M$8 (Khan et al., 8 Jun 2026). This does not redefine multiple consensus as an anti-consensus method, but it demonstrates that, in generative systems, consensus can itself become the object to be measured, visualized, and deliberately avoided.

The broad implication is that multiple-consensus techniques are best understood not as a single algorithmic paradigm but as a recurring response to structured heterogeneity. Where the underlying object is a clustering ensemble, a metric system, a set of gene trees, a decentralized optimization trajectory, or a group deliberation, the central question is the same: whether agreement should be forced into one summary, decomposed into several internally coherent summaries, or regulated through repeated interaction until the relevant form of consensus becomes statistically, topologically, or operationally justified.