This research starts with the building of a firm theoretical model. This model rethinks the working of distributed consensus algorithms and institutes them as complex dynamic systems. Based on well-known principles from non-linear dynamics (Strogatz, 2024), ²⁴ we formally define a consensus cluster with N nodes as a discrete time dynamical system. The state of each node is represented by a state vector X i ( k) that consists of important variables about each node i at communication round k e.g., its current role (leader, follower, candidate), its committed log index, its current term etc. The collective state of the whole system therefore is described by a vector X ( k) = [ X ₁( k), X ₂( k).... X _N ( k)] of dimension N . The function that is described by the deterministic consensus logic function Λ (e.g., the rules of the Raft algorithm) and is subject to stochastic perturbations in the network functioning η( k). This relation given in terms of difference equation: X ( k + 1 ) = Λ ( X ( k ) , η ( k ) ) (1)

The essential chaos theoretic hypothesis, which was inspired by a seminal work by Lorenz (1963) ¹³ : The function Λ, although deterministic, is defined so as to be sensitive to the initial conditions, and thus is appreciably affected by the perturbations η( k). These perturbations represent inherent network stochasticity, the effects of latency fluctuations (modeled by heavy tailed distributions Sheluhin and Rybakov 2024 ²² ) and random failures of nodes. Small variations in η( k) get multiplied in the iterative consensus process and may produce complex (fractal) patterns for measurable output time series such as consensus latency and message count.

In order to empirically confirm this hypothesis, a sophisticated discrete event network simulator was implemented in python. The simulation environment is based on a distributed network programmed with the Raft consensus algorithm, chosen because of its architectural clarity and well-defined states (Mazzoni et al. 2022). ¹⁵ The model includes several input parameters that are systematically varied to help replicate a variety of operational conditions as specified in Table 1.

The chosen parameter values span a representative range of scales for deployment and load. Network sizes of 5, 10, 21, and 50 nodes were selected that represent small size clusters, common number of nodes in mid-size deployments. The larger configurations commonly studied in consensus algorithm evaluations. Request rates of 1, 10, and 100 requests per second correspond to low, moderate, and high traffic scenarios, respectively. The selection enables the assessment of system behavior under varying stress levels.

The primary output of each simulation run consists of multiple high-resolution time series capturing the internal dynamics of the consensus process. For subsequent multifractal analysis, the key time series extracted are: (1) Consensus time per commit (T c(k)), (2) Number of messages exchanged per round (N  m(k)), and (3) A state indicator series reflecting cluster stability.

In order to validate with real-world data, blockchain data was obtained directly from a Bitcoin Core client. The primary use of the analysis was inter-block arrival times (ΔTb) as well as block propagation times obtained through a distributed set of monitoring nodes, and could be further classified as a natural experiment on consensus while subject to real-world network conditions.

The basic analytical methodology is based on the Multifractal Detrended Fluctuation Analysis (MF-DFA) technique, using the established analysis procedure by Zhang et al. (2021). ²⁷ The analysis comprises five computation steps: 1.

Integration: The time series x ( i) of length L is integrated to give Y ( i) is profile.

The scaling behaviour of Fq ( s) is analyzed for various values of q, normally between −5 and +5. If the series is multifractal in nature, then the generalized Hurst exponent h ( q) will display dependence on q. Similar scaling-based methods have been used for self-similar network traffic, for which the estimation of the Hurst exponent plays a central role (Millan, 2021). ¹⁶ The singularity strength α and multifractal spectrum f( α ) are then obtained by the Legendre transform: α = h ( q ) + q · dh ( q ) dq ( 4 ) f ( α ) = q ( α − h ( q ) ) + 1 (4)

The width of the singularity spectrum Δ α =  αmax− αmin is the main quantitative estimate of the multifractality strength with wider spectra reflecting, more so, rich multifractal structure and higher complexity of the considered system.

A second order polynomial ( m = 2) was chosen to remove linear and quadratics trends, which is sufficient to manage the non-stationary behaviour of consensus latency series. The range of values for s was set from s = 16 up to s = 4096, where the lower limit guarantees that the local trends can be well estimated and the upper limit follows the L/4 rule with respect to the total series length ( L = 16,000).

The last analytical phase is the correlation of the multifractal metrics with the traditional performance metrics. Some of the most important metrics that are assessed are: •

Traditional Performance Metrics: Average throughput (requests/second), mean latency ( ms), and failure recovery time.

The analysis specifically examines how changes in input parameters ( Table 1) affect both traditional and multifractal metrics, testing the hypothesis that network conditions characterized by heavy-tailed distributions (Pareto latency) will produce significantly wider multifractal spectra (Δ α) in the consensus time series, thereby establishing a direct relationship between network microstructure and consensus dynamics.

The time series were segmented into non overlapping groups of 500 rounds each, for each group of windows the multifractal spectrum width Da and the traditional performance indicators were calculated and the correlations (Pearson and Spearman) were assessed.

A multi-panel visualization of the major time series which emerge out of the simulation, a contiguous excerpt of 10,000 rounds of the simulation for all four scenarios outlined in the canonical analysis, is presented Figure 1. The upper panel shows the consensus-time accumulated per round (consensus_time_ms) showing a logarithmic axis - we use the logarithm to highlight heavy tailed excursions; the middle panel shows the numbers of messages exchanged per round (num_messages); and the bottom panel shows the per round mean inter-node latency (latency_mean_ms). The temporal segmentation in the four scenarios is marked by a weak vertical banding on the time axis to ease the visual comparison task.

This figure illustrates that a normal situation is low mean consensus latency and low variance coupled with occasional micro-bursts due to the stochastic component of the latency. The portion of the workload with the high_load shows high baseline message counts and correspondingly moderate increases in consensus times, consistent with the queuing effects. The dosattack segment exhibits high frequency and large amplitude positive excursions over consensus_time_ms as well as a notable propagation latency variance for these excursions that show clustering in time. The segment partial_failures show intermittent persistent invalid increases of the consensus time with discrete steps in the node_failure_count (documented in the simulation CSV), which constitutes injected node outages. This type of composite visualization is the motivation to each scenario as a different dynamical regime for multifractal characterization.

Table 2 summarizes the results for each of the scenarios calculated on the full length of simulated series: median and mean consensus time (ms), standard deviation, interquartile range, median number of messages per round, and empirical failure incidence (fraction of rounds with node_failure_count >0). These metrics put the multifractal analysis into context by relating traditional measures of performance solutions to the properties of the distribution of the time series.

The values observed in Table 2 are a systematic increase in central tendency and dispersion in the attacks and failure scenarios compared to the normal operation. Specifically, the mean in dos_attack is pulled upwards because of heavy tail of high latency events, the median is still lower than the mean which indicates that the distributions are skewed. This situation is distributional asymmetry which is the situation under which the multifractal analysis is most valuable, because the multifractality is an estimation about the heterogeneity of the scaling exponents that come as a result of intermittent bursts and clustered extremes.

For each, we embarked on MF DFA on the ‘consensus_time_ms’ series. This process consisted of a second order polynomial detrending ( m = 2), a series of window sizes s, varying logarithmically from 16 to 4096, and generalized moments q, varying from −5 to +5 in increments of 1. From this the generalized Hurst exponents h ( q) and singularity spectra f( α) were obtained and used to extract two different descriptors: spectrum width Δ\ α = 〖 α〗 (max) – a_(min) and spectral asymmetry (the skew of f( α)). The singularity spectra f( α) curves vs. as for the normal condition are shown in Figure 2. The spectra are a multifold unimodal bell shape with finite width which implies a continuous distribution of the singularity strength over a range, not a delta function. Such multifractal is a feature of the consensus latency time series and is evidence of an inherent multifractal structure even in baseline conditions. It represents the multifractal structure of the time series as a mix of scaling behaviors associated with both the smaller fluctuations and less frequent and larger excursions.

A plot of the f( α) curve for this normal situation is shown to be centred around α ≈ 0.52 and of reasonable width (Δ α ≈ 0.21). Interpretation: a finite Δ α under ‘normal’ operation suggests that even in the absence of attacks, even under load, latencies, message scheduling etc., the consensus process is not monofractal, there is a little more to ‘microscopical’ variability that induce a spectrum of local regularities. The presence of the spectral peak around α ≈ 0.52 indicates that near-diffusive first-order temporal scaling is suggested for the dominant fluctuations and more persistent but rarer events are encoded in the tails.

Figure 3 illustrates an overlay of f( α) curves for all four scenarios (normal and high_load, dos_attack, partial_failures). The shapes and supports are compared in a direct manner to see the effect of operational stressors on multifractal structure.

The overlay shows the progressive broadening and shifting to the left of the spectra as the operational stress increases. In particular, the spectra of both dos_attack and partial_failures are significantly broader than normal and high_load (Δα _dos ≈ 0.98; Δα _partial ≈ 0.) compared to normal and high_load (Δα _normal ≈ 0.21; Δα _highload ≈ 0.34). The dos_attack spectrum shows significant left-skewed asymmetry that suggests the extreme slow events (big consensus times) have a disproportionate contribution to the multifractal signature. Heavy-tailed latencies create a rich crop of very slow consensus rounds, which aligns to small Hölder exponents a and so stretches the singularity spectrum towards smaller a value, which results in left skew. This behaviour is similar to that predicted by a hypothesis that heavy tailed perturbations (Pareto-type latencies) and injected failures enhance heterogeneity of local scaling exponents (by enabling richer multifractal behaviour).

The multifractal descriptors obtained from MF-DFA for different scenarios: Δ α, a peak (location information of maximum f( α)), spectral skewness and the q-dependence range of h ( q) ( h (−5) −  h(+5)) are summarized in Table 3.

These in turn agree with the graphical observations made. The dos_attack scenario is the one with the highest multifractality in terms of Δ α and shows the largest variability of h ( q) as well; the spectral skewness is negative: confirming the predominance of extreme slow events. The high_load scenario results in moderate multifractal enhancement from normal and indicates the increase in heterogeneity of queuing and resource contention, not due to the extreme heavy tail shocks bridged by the DOS phase.

In order to relate multifractal descriptors with common performance indicators, the Pearson correlation coefficients and the robust (Spearman) rank correlations between Δ α and the mean of the consensus time in each simulation window, as well as the Δ α and the empirical round failure rate, were computed. A scatter plot of Δ α versus mean consensus time with each point representing a non-overlapping block of 500 rounds sampled over the entire simulated timeline is given in Figure 4. A least-squares line and a nonparametric LOWESS smoothing curve are superimposed to visualize a linear and local monotonic relationship, respectively.

The scatter suggests a linear monotonic relationship, with large Δ α blocks being linked to systematically higher mean consensus times. Correlation coefficients and regression summary statistics measuring this relationship can be found in Table 4. The Pearson correlation r = 0.72 ( p < 1 e−6) and Spearman ρ = 0.68 show a strong and significant association between A simple linear regression of mean_consensus_tim on Δ α gives an R 2 ≈ 0.52, indicating that Δ α accounts for a large proportion of the cross-block variability in mean consensus latency.

The equivalent exercise using Δ α versus the round failure rate gives Figure 5 and the statistics of Table 5. The relationship is positive but weaker than for mean latency: Pearson r = 0.51( p < 1 e−4), Spearman ρ = 0.49. The moderate strength of this association means that whereas multifractal widening is sensitive for interpreting intermittent failures, Δ α is more closely related to ongoing delay dynamics (consensus timing) than it is for the binary occurrence of round failures.

Taken together, these results suggest that the multifractal spectrum width Δ α is a valuable and quantitatively interpretable descriptor of system performance: increases in Δ α are a reliable predictor of the system’s degradation as measured by consensus latency, and of significantly increasing order, of the increased incidence of the failed rounds. This result agrees with the main hypothesis introduced in this paper about the relationship between microstructural network stochasticity and the consensus dynamics at the macroscopic level through multifractal indices.

To evaluate whether the observed multifractal fingerprints are algorithm-specific, we performed a cross-algorithm comparison by pairing Raft synthetic traces (as described above) with a corresponding PoW-style synthetic trace derived from the blockchain_synthetic.csv file (block propagation mean times and inter-block arrival series). For the PoW analogue we treated inter-block intervals Δ T _b and propagation_mean_ms as primary series and applied identical MF-DFA settings (polynomial detrend m = 2, scales s ∈ [16,4096], q ∈ [−5,+5]). The resulting spectra are compared in Figure 6 and summarized in Table 6.

A comparison can show commonalities as well as distinct signatures. Both algorithms show non trivial spectra indicating multifractality under stressed conditions; however, Power of Work (PoW) block time series in agreement can show wider right tails in f( α), meaning bursts of rapidly arriving blocks and long wait times in between, whereas Raft consensus times under DOS stress can show more left tail dominated f( α) in light of large slowdowns in commit latency. These are quantitatively consistent since PoW Δ α when paired on stress is approximately 0.86, whereas Raft Δ α is approximately 0.98 in dos_attack. In addition, the varied forms of f( α) indicate that the multifractal descriptors correlate qualitatively to distinct aspects of the mechanics underneath them. This is to say that consensus dynamics for PoW reduce to the stochastic timing of events at the source (that source being mining/leader acquisition); whereas consensus dynamics for Raft improve the communication-and coordination-related effects, and commit latency delays.

The results suggest multifractal analysis can be used for two things, good: showing that there is complex scaling in consensus setups and spotting the kind of intermittency methodical - disorganized. So, multifractal descriptors could be considered in case of diagnosis and comparison of different distributed ledgers and cluster consensus systems.

Table 7 presents some practical suggestions and data-based cutoffs from this study for linking Δ α values to various levels of performance where action may be required. These thresholds are given as a tentative guideline to those practitioners who are interested in the hint of multifractal monitoring for an early warning indicator. The thresholds should be understood in the context of the particular deployment and may need empirical recalibration in the application to real-life networks.

Practical suggestions with data-based cutoffs from this study relating Δ α values to different levels of performance where action may be needed are presented in Table 7. These thresholds are provided as a rough guide for the practitioners who want to implement multifractal monitoring as early warning indicator. Its thresholds have to be taken in the context of goes on specific deployment and can need empirical recalibration when apply experiments on real world networks.

The analysis presented here gives an example as to how the consensus protocols, analyzed as nonlinear dynamical systems and multifractals reveal a complex structure, to above the simple analysis of performance measured by first moment and second moment. The synthetic experiments revealed that heavy tailed network perturbations and node outages had a significant effect in increasing the multifractal spectrum of consensus latency as Δ α spectral width showed a high correlation with mean latency and a moderate correlation with round failures. The investigation done on a comparative basis using the various algorithms showed that although the overall shapes of the spectral width are similar between algorithms under duress the spectral signatures are algorithm specific and can be used for diagnostic and comparative purposes.

Results in summary suggest sealing the central thesis to this work within the multifractal descriptors being sensitive and scale aware and transitory proxies of macroscale undefined measures from microscale is networks irregularities consensus performance. Future investigations should evaluate these findings on empirical traces from production clusters, public blockchains and continue to look into multifractal indicator to adaptive consensus controls and anomaly detection pipelines.