Noisy Majority Vote Process

• The noisy majority vote process is a nonequilibrium stochastic dynamics model that updates binary or multistate variables using a local majority rule perturbed by a controlled noise parameter.
• Variants incorporating heterogeneous noise, ancillary flipping, and memory effects allow researchers to extract critical exponents and examine universality across lattices, networks, and adaptive systems.
• This framework underpins applications in consensus dynamics, distributed optimization, and sensing, highlighting phenomena such as metastability, non-ergodicity, and phase coexistence.

The noisy majority vote process is a class of nonequilibrium stochastic dynamics in which binary or multistate variables are updated by a local majority rule perturbed by explicit randomness. In its canonical form, each site or agent carries an opinion or spin, the neighborhood majority is computed, and the update is applied with a noise parameter that permits misalignment, random flipping, communication corruption, or other departures from deterministic majority adoption. Across square lattices, complex and adaptive networks, multiplex systems, complete graphs, trees, and application-specific aggregation settings, these processes have been used to study continuous and discontinuous phase transitions, universality, metastability, robustness, and non-ergodicity (Lima, 2013, Lima, 2013, d'Amore et al., 2021, Ding et al., 27 Jan 2025).

1. Canonical stochastic formulation

A standard binary majority-vote dynamics places a spin variable $\sigma_i\in\{+1,-1\}$ on each site or node and updates it according to the sign of the local field. On square lattices and networks, the flip probability is commonly written as

$$w_i=\frac12\Bigl[1-(1-2q)\,\sigma_i\,S\!\Bigl(\sum_j \sigma_j\Bigr)\Bigr],$$

or with the local field $H_i=\sum_{j\in \mathrm{neigh}(i)}\sigma_j$ as

$w_i=\frac12[1-(1-2q)\sigma_i S(H_i)].$

Here $q$ is the noise or misalignment probability, $S(0)=0$, and ties are handled by random or null-majority conventions depending on the formulation (Lima, 2013, Lima, 2013, Lima, 2011).

Several extensions replace the single homogeneous noise parameter by other noise mechanisms. In the heterogeneous-agent square-lattice model, each site carries its own noise $q_i$, drawn independently from the uniform distribution on $[0,q]$, so that the control parameter remains the upper bound $q$ rather than a single global flip probability (Lima, 2013). In the noisy majority-vote model with ancillary noise, a node flips independently of its neighborhood with probability $p$, while with probability $$w_i=\frac12\Bigl[1-(1-2q)\,\sigma_i\,S\!\Bigl(\sum_j \sigma_j\Bigr)\Bigr],$$0 it follows the majority-vote rule with misalignment $$w_i=\frac12\Bigl[1-(1-2q)\,\sigma_i\,S\!\Bigl(\sum_j \sigma_j\Bigr)\Bigr],$$1, yielding

$$w_i=\frac12\Bigl[1-(1-2q)\,\sigma_i\,S\!\Bigl(\sum_j \sigma_j\Bigr)\Bigr],$$2

on arbitrary uncorrelated networks (Encinas et al., 2018). In the non-Markovian model, the standard majority-vote kernel is further modulated by an age-dependent activation $$w_i=\frac12\Bigl[1-(1-2q)\,\sigma_i\,S\!\Bigl(\sum_j \sigma_j\Bigr)\Bigr],$$3, so that the probability of changing state depends on how long the agent has held its current opinion (Chen et al., 2020).

Other variants shift the noise source from local spin-flip uncertainty to communication or aggregation uncertainty. In the noisy 3-majority process on the complete graph, each agent samples three neighbors with repetition, each received opinion is independently corrupted with probability $$w_i=\frac12\Bigl[1-(1-2q)\,\sigma_i\,S\!\Bigl(\sum_j \sigma_j\Bigr)\Bigr],$$4, and the node updates from the noisy triple by majority or random tie-breaking (d'Amore et al., 2021). In label aggregation, agents independently report correct labels with competence $$w_i=\frac12\Bigl[1-(1-2q)\,\sigma_i\,S\!\Bigl(\sum_j \sigma_j\Bigr)\Bigr],$$5, while an additional group-level flip probability $$w_i=\frac12\Bigl[1-(1-2q)\,\sigma_i\,S\!\Bigl(\sum_j \sigma_j\Bigr)\Bigr],$$6 corrupts the final majority output, giving

$$w_i=\frac12\Bigl[1-(1-2q)\,\sigma_i\,S\!\Bigl(\sum_j \sigma_j\Bigr)\Bigr],$$7

for the probability that the final majority decision is correct (Vilone, 2022).

Variant	Noise specification	Reported consequence
Heterogeneous square lattice	$$w_i=\frac12\Bigl[1-(1-2q)\,\sigma_i\,S\!\Bigl(\sum_j \sigma_j\Bigr)\Bigr],$$8 uniform on $$w_i=\frac12\Bigl[1-(1-2q)\,\sigma_i\,S\!\Bigl(\sum_j \sigma_j\Bigr)\Bigr],$$9	different universality class (Lima, 2013)
SHP adaptive network	homogeneous $H_i=\sum_{j\in \mathrm{neigh}(i)}\sigma_j$0 plus rewiring of disagreeing links	weak universality discussion (Lima, 2013)
Three-state square lattice	$H_i=\sum_{j\in \mathrm{neigh}(i)}\sigma_j$1 with noise $H_i=\sum_{j\in \mathrm{neigh}(i)}\sigma_j$2	Ising-like critical ratios (Lima, 2011)
Non-Markovian majority vote	age-dependent activation $H_i=\sum_{j\in \mathrm{neigh}(i)}\sigma_j$3	non-monotonic $H_i=\sum_{j\in \mathrm{neigh}(i)}\sigma_j$4 (Chen et al., 2020)

2. Macroscopic observables and critical diagnostics

The standard macroscopic observables are the magnetization, susceptibility, and Binder cumulant. On the square lattice with $H_i=\sum_{j\in \mathrm{neigh}(i)}\sigma_j$5 and

$H_i=\sum_{j\in \mathrm{neigh}(i)}\sigma_j$6

the measured quantities are

$H_i=\sum_{j\in \mathrm{neigh}(i)}\sigma_j$7

Near the critical noise $H_i=\sum_{j\in \mathrm{neigh}(i)}\sigma_j$8, they obey the finite-size scaling forms

$H_i=\sum_{j\in \mathrm{neigh}(i)}\sigma_j$9

on lattices, with analogous $w_i=\frac12[1-(1-2q)\sigma_i S(H_i)].$0-based scaling relations on networks (Lima, 2013, Lima, 2013).

These observables support the extraction of exponent ratios $w_i=\frac12[1-(1-2q)\sigma_i S(H_i)].$1, $w_i=\frac12[1-(1-2q)\sigma_i S(H_i)].$2, and $w_i=\frac12[1-(1-2q)\sigma_i S(H_i)].$3, the critical noise, and a Binder fixed point $w_i=\frac12[1-(1-2q)\sigma_i S(H_i)].$4. In the three-state model on the square lattice, the shift of the susceptibility peak is written as $w_i=\frac12[1-(1-2q)\sigma_i S(H_i)].$5, while on the SHP network the analogous expression is $w_i=\frac12[1-(1-2q)\sigma_i S(H_i)].$6 (Lima, 2011, Lima, 2013). Binder-cumulant crossings are also used in the non-Markovian model and in the feedback-coupled two-layer model to locate continuous transitions (Chen et al., 2020, Liu et al., 2023).

Hyperscaling relations are a recurring diagnostic. For the heterogeneous square-lattice model, the reported exponents satisfy, within error bars,

$w_i=\frac12[1-(1-2q)\sigma_i S(H_i)].$7

consistent with an effective dimension $w_i=\frac12[1-(1-2q)\sigma_i S(H_i)].$8 (Lima, 2013). On the SHP network,

$w_i=\frac12[1-(1-2q)\sigma_i S(H_i)].$9

and the effective dimension is reported as $q$0 (Lima, 2013). In the noisy majority-vote model with ancillary noise, the reported finite-size exponents for RR and ER networks satisfy $q$1, while BA networks satisfy $q$2 with different individual ratios (Encinas et al., 2018). This suggests that the interplay between update rule and substrate geometry is central to the scaling structure.

3. Regular lattices: universality, multistate extensions, and heterogeneity

On the two-dimensional square lattice, the majority-vote model is historically tied to the conjecture that up-down-symmetric nonequilibrium models on regular lattices fall into the Ising universality class. The three-state generalization studied by Lima and Stauffer retains the majority-vote form but introduces a neutral state $q$3. Its reported critical quantities are

$q$4

with exponent ratios

$q$5

These agree within statistical error with the spin-$q$6 and spin-1 Ising values, and the zero state is reported not to change the universality class (Lima, 2011).

Introducing quenched heterogeneity at the level of individual noise changes that conclusion. In the heterogeneous-agent square-lattice model, each $q$7 is drawn independently from the uniform distribution on $q$8, and Monte Carlo finite-size scaling yields

$q$9

together with

$S(0)=0$0

The reported comparison is explicit: $S(0)=0$1 and $S(0)=0$2 differ significantly from the homogeneous square-lattice majority-vote model and from the two-dimensional Ising model, while $S(0)=0$3 and $S(0)=0$4 remain close to the homogeneous-agent values. The paper therefore concludes that agent-to-agent heterogeneity in the noise drives the system into a different nonequilibrium universality class (Lima, 2013).

A further modification couples the vote layer to a dynamic noise layer. With externally imposed homogeneous noise on the square-lattice vote layer, the reported critical point is $S(0)=0$5 and the exponents are consistent with the two-dimensional Ising values $S(0)=0$6, $S(0)=0$7, and $S(0)=0$8 (Liu et al., 2023). Under feedback, however, the behavior depends strongly on topology: for SL–RRN and SL–SL combinations there is no crossing of $S(0)=0$9 and no divergent $q_i$0, and the critical phenomena vanish. This provides a precise example in which local structure combined with dynamical noise feedback suppresses the standard order-disorder criticality (Liu et al., 2023).

4. Complex, adaptive, and multiplex networks

On adaptive opinion-dependent networks of Stauffer–Hohnisch–Pittnauer type, each node has $q_i$1 directed outgoing links, and disagreeing links are rewired with probability $q_i$2 while agreeing links remain unchanged. The majority-vote process on this substrate exhibits a continuous transition located by Binder crossings at

$q_i$3

with exponent ratios

$q_i$4

The paper reports that the model belongs to a different universality class than the equilibrium Ising model on the same SHP network, while sharing $q_i$5 and $q_i$6 with some other majority-vote models on complex graphs but differing in $q_i$7, which is described there as an example of weak universality (Lima, 2013).

The noisy majority-vote model with ancillary noise generalizes the dynamics to arbitrary uncorrelated networks, including RR, ER, and BA topologies. Mean-field theory and simulations are reported to agree well, and the central qualitative result is that the additional independent flip probability $q_i$8 does not alter the continuous character of the phase transition. The critical line depends on degree moments through formulas such as

$q_i$9

and the ordered phase disappears above a threshold

$[0,q]$0

The reported numerical exponent ratios are $[0,q]$1, $[0,q]$2, and $[0,q]$3 for RR and ER, while BA networks show $[0,q]$4 and $[0,q]$5 (Encinas et al., 2018).

Multiplexity introduces rule-level heterogeneity even without changing the local state space. In the two-layer model, AND-rule voters update only when both layers have the same majority sign, whereas OR-rule voters choose one layer uniformly at random and follow its majority. The OR transition is reported as continuous, with smaller consensus but faster convergence. The AND model reaches the largest consensus below $[0,q]$6, requires much longer time to reach consensus, and for even-degree regular duplexes exhibits a genuinely discontinuous jump in the stationary magnetization. Approximate master-equation calculations qualitatively support these numerical findings (Choi et al., 2018).

5. Communication noise, memory, and dynamical feedback

The noisy 3-majority process on the complete graph is a synchronous consensus dynamics rather than the usual asynchronous spin-flip model, but it preserves the same majority-with-noise logic. In the binary case, the signed bias

$[0,q]$7

obeys the mean-drift relation

$[0,q]$8

The process exhibits a sharp phase transition at $[0,q]$9. For $q$0, the dynamics reaches in logarithmic time an almost-consensus metastable phase that lasts for a polynomial number of rounds with high probability, with attractive equilibrium bias

$q$1

For $q$2, no form of consensus is possible and information about the initial majority is lost in logarithmic time with high probability (d'Amore et al., 2021).

Memory modifies the effective update rate even when the local majority rule is unchanged. In the non-Markovian majority-vote model, the flip probability is

$q$3

where the activation $q$4 is decreasing in the aging regime and increasing in the anti-aging regime. A heterogeneous mean-field analysis yields a self-consistency equation for the edge-state probability and a critical noise $q$5 whose non-Markovian dependence is encoded by a function $q$6. The reported result is non-monotonicity: in the aging regime $q$7 has a maximum, while in the anti-aging regime it has a minimum, and the transition remains continuous for all $q$8 on RR, ER, SF, and lattice topologies (Chen et al., 2020).

Dynamic feedback can also be imposed through an explicit noise layer. In the two-layer feedback model, the local vote-layer flip rate is

$q$9

while the mean-field noise evolution is

$p$0

This system exhibits two kinds of third-order transitions. The isolated-spin density $p$1 has a size-independent local maximum corresponding to an independent third-order transition, while a local minimum of $p$2 or $p$3 marks a dependent third-order transition. The dependent third-order transition appears only when a second-order critical transition exists, and the paper proposes it as a precursor of that critical transition (Liu et al., 2023).

6. Algorithmic, crowdsourcing, and sensing formulations

In group evaluation and label aggregation, the noisy majority-vote process is formulated in terms of competence and aggregation error rather than spins on a graph. If $p$4 is the number of correct votes in a group of size $p$5, the noise-free majority success probability is

$p$6

With aggregation noise $p$7, the final success probability becomes

$p$8

An equivalent “effective competence” is

$p$9

The reported conclusion is that majority rule is still, in most cases, the best way to get a reliable group evaluation, but noise and mesoscopic fluctuations can undermine its efficiency in non-trivial situations (Vilone, 2022).

In distributed optimization, majority voting appears as a coordinate-wise aggregation operator in signSGD. Each of $$w_i=\frac12\Bigl[1-(1-2q)\,\sigma_i\,S\!\Bigl(\sum_j \sigma_j\Bigr)\Bigr],$$00 workers sends only the sign vector $$w_i=\frac12\Bigl[1-(1-2q)\,\sigma_i\,S\!\Bigl(\sum_j \sigma_j\Bigr)\Bigr],$$01, and the server computes

$$w_i=\frac12\Bigl[1-(1-2q)\,\sigma_i\,S\!\Bigl(\sum_j \sigma_j\Bigr)\Bigr],$$02

Under an i.i.d. sign-flip model with per-coordinate error probability $$w_i=\frac12\Bigl[1-(1-2q)\,\sigma_i\,S\!\Bigl(\sum_j \sigma_j\Bigr)\Bigr],$$03, the probability that the aggregate sign is correct is the upper tail of a $$w_i=\frac12\Bigl[1-(1-2q)\,\sigma_i\,S\!\Bigl(\sum_j \sigma_j\Bigr)\Bigr],$$04 distribution. The paper proves $$w_i=\frac12\Bigl[1-(1-2q)\,\sigma_i\,S\!\Bigl(\sum_j \sigma_j\Bigr)\Bigr],$$05 nonconvex convergence rates, states that communication is reduced by a factor of $$w_i=\frac12\Bigl[1-(1-2q)\,\sigma_i\,S\!\Bigl(\sum_j \sigma_j\Bigr)\Bigr],$$06, and shows robustness when up to $$w_i=\frac12\Bigl[1-(1-2q)\,\sigma_i\,S\!\Bigl(\sum_j \sigma_j\Bigr)\Bigr],$$07 of workers behave adversarially (Bernstein et al., 2018).

A spatial sensing formulation appears in local event boundary detection with unreliable sensors. Sensors are distributed in a region by a Poisson process of intensity $$w_i=\frac12\Bigl[1-(1-2q)\,\sigma_i\,S\!\Bigl(\sum_j \sigma_j\Bigr)\Bigr],$$08, each sensor makes an initial binary decision with error probability $$w_i=\frac12\Bigl[1-(1-2q)\,\sigma_i\,S\!\Bigl(\sum_j \sigma_j\Bigr)\Bigr],$$09, and then revises its decision by the majority vote of all neighbors within radius $$w_i=\frac12\Bigl[1-(1-2q)\,\sigma_i\,S\!\Bigl(\sum_j \sigma_j\Bigr)\Bigr],$$10, breaking ties in favor of its previous bit. The expected number of post-vote misclassifications decomposes as $$w_i=\frac12\Bigl[1-(1-2q)\,\sigma_i\,S\!\Bigl(\sum_j \sigma_j\Bigr)\Bigr],$$11, where outside the boundary band $$w_i=\frac12\Bigl[1-(1-2q)\,\sigma_i\,S\!\Bigl(\sum_j \sigma_j\Bigr)\Bigr],$$12 the error bounds decay exponentially in $$w_i=\frac12\Bigl[1-(1-2q)\,\sigma_i\,S\!\Bigl(\sum_j \sigma_j\Bigr)\Bigr],$$13, while inside $$w_i=\frac12\Bigl[1-(1-2q)\,\sigma_i\,S\!\Bigl(\sum_j \sigma_j\Bigr)\Bigr],$$14 the bound depends on $$w_i=\frac12\Bigl[1-(1-2q)\,\sigma_i\,S\!\Bigl(\sum_j \sigma_j\Bigr)\Bigr],$$15, the perimeter $$w_i=\frac12\Bigl[1-(1-2q)\,\sigma_i\,S\!\Bigl(\sum_j \sigma_j\Bigr)\Bigr],$$16, and the number of components $$w_i=\frac12\Bigl[1-(1-2q)\,\sigma_i\,S\!\Bigl(\sum_j \sigma_j\Bigr)\Bigr],$$17. Under convexity and radius-of-curvature assumptions, the boundary-layer estimate improves substantially (Brass et al., 2013).

7. Infinite-volume behavior, invariant measures, and non-ergodicity

On infinite regular trees, the noisy majority-vote process is defined in discrete time by alternating vote events and noise events. With probability $$w_i=\frac12\Bigl[1-(1-2q)\,\sigma_i\,S\!\Bigl(\sum_j \sigma_j\Bigr)\Bigr],$$18, a vertex adopts the majority of its neighbors and keeps its own state in case of tie; with probability $$w_i=\frac12\Bigl[1-(1-2q)\,\sigma_i\,S\!\Bigl(\sum_j \sigma_j\Bigr)\Bigr],$$19, it is reset to $$w_i=\frac12\Bigl[1-(1-2q)\,\sigma_i\,S\!\Bigl(\sum_j \sigma_j\Bigr)\Bigr],$$20 with probability $$w_i=\frac12\Bigl[1-(1-2q)\,\sigma_i\,S\!\Bigl(\sum_j \sigma_j\Bigr)\Bigr],$$21 each. For every integer $$w_i=\frac12\Bigl[1-(1-2q)\,\sigma_i\,S\!\Bigl(\sum_j \sigma_j\Bigr)\Bigr],$$22, there exists $$w_i=\frac12\Bigl[1-(1-2q)\,\sigma_i\,S\!\Bigl(\sum_j \sigma_j\Bigr)\Bigr],$$23 such that for all $$w_i=\frac12\Bigl[1-(1-2q)\,\sigma_i\,S\!\Bigl(\sum_j \sigma_j\Bigr)\Bigr],$$24 the process on the infinite $$w_i=\frac12\Bigl[1-(1-2q)\,\sigma_i\,S\!\Bigl(\sum_j \sigma_j\Bigr)\Bigr],$$25-regular tree is non-ergodic, meaning that there are at least two distinct invariant measures $$w_i=\frac12\Bigl[1-(1-2q)\,\sigma_i\,S\!\Bigl(\sum_j \sigma_j\Bigr)\Bigr],$$26. Starting from all $$w_i=\frac12\Bigl[1-(1-2q)\,\sigma_i\,S\!\Bigl(\sum_j \sigma_j\Bigr)\Bigr],$$27, the one-site marginal converges to a limit $$w_i=\frac12\Bigl[1-(1-2q)\,\sigma_i\,S\!\Bigl(\sum_j \sigma_j\Bigr)\Bigr],$$28; starting from all $$w_i=\frac12\Bigl[1-(1-2q)\,\sigma_i\,S\!\Bigl(\sum_j \sigma_j\Bigr)\Bigr],$$29, it converges to $$w_i=\frac12\Bigl[1-(1-2q)\,\sigma_i\,S\!\Bigl(\sum_j \sigma_j\Bigr)\Bigr],$$30 (Ding et al., 27 Jan 2025).

The proof strategy reported there introduces a comparison minus-biased process and derives cluster-separation bounds for the probability that finite sets of vertices are simultaneously in the minority state. These bounds are organized through odd clusters, trifurcation counts, and combinatorial enumeration of ancestor sets, and they show that sufficiently small noise does not destroy the persistent bias inherited from the initial condition (Ding et al., 27 Jan 2025).

A broader nonamenable-graph version establishes low-noise non-uniqueness of invariant measures under suitable edge- or vertex-expansion assumptions. For connected infinite graphs of bounded degree satisfying the required expansion conditions, there exists $$w_i=\frac12\Bigl[1-(1-2q)\,\sigma_i\,S\!\Bigl(\sum_j \sigma_j\Bigr)\Bigr],$$31 such that the noisy majority-vote process admits at least two distinct invariant measures $$w_i=\frac12\Bigl[1-(1-2q)\,\sigma_i\,S\!\Bigl(\sum_j \sigma_j\Bigr)\Bigr],$$32. The analysis proceeds via decorated set systems, space-time zero clusters, and exponentially decaying tails for the cluster diameter. The paper further states that on hyperbolic lattices the result is essentially sharp and that regular trees are recovered as a concrete example (Gaspari, 9 Aug 2025).

Taken together, these results show that the noisy majority vote process is not a single fixed model but a broad framework for studying how local majoritarian alignment competes with stochastic perturbations. On regular lattices it is a testbed for nonequilibrium universality; on adaptive, multiplex, and feedback-coupled networks it probes the effect of topology and endogenous noise; in algorithmic and sensing contexts it formalizes aggregation under uncertainty; and on infinite graphs it exhibits genuine phase coexistence and symmetry breaking at low noise (Lima, 2013, Choi et al., 2018, Gaspari, 9 Aug 2025).