Statistical physics approaches to the complex Earth system

1.1. The Earth as a Complex System

• Earth System Science treats geosphere, atmosphere, hydrosphere, cryosphere, biosphere, and human processes as an integrated system with nonlinear interactions and feedbacks.Its development combined observations, proxy records, mathematical models, simulations, and assessments to study Earth dynamics across space and time scales.Source: §1.1 · pp. 3–4

1.2. Why Statistical Physics?

• Statistical physics addresses unresolved extreme climate events, earthquakes, and early-warning questions by improving understanding and prediction of high-impact disruptions.The review applies critical phenomena, fractals, self-organized criticality, entropy, and complex networks to Earth-system problems.Source: §1.2 · pp. 4, 6

1.3. Outline of the Report

• The report introduces formal methods before reviewing applications to climate systems, Earth geometric relief, and earthquakes, then concludes with perspectives.Its theoretical framework includes climate networks, percolation, tipping-point analysis, entropy, and complexity.Source: §1.3 · p. 6

2.1. Climate Networks

• Climate networks represent geographical grid points as nodes and similarity or causality between climate time series as weighted links, enabling analysis and prediction of global climate behavior.They have been applied to temperature, pressure, wind, and precipitation data to study interactions and transfers of energy and matter across locations.Source: §2.1 · pp. 6–7, §1.2 · p. 6

2.1.1. Network Characteristics

• A network is a graph of vertices and edges, with N denoting the number of vertices and M the number of edges.The review focuses on graph-theoretic characteristics relevant to complex Earth systems and distinguishes network terminology across disciplines.Source: §2.1.1 · p. 8

The Adjacency Matrix

• For simple networks, the adjacency matrix A encodes whether an edge connects vertices i and j, with diagonal entries zero and weighted or directed variants extending this representation.Weighted matrices store connection strengths, whereas directed-network matrices are generally nonsymmetric.Source: The Adjacency Matrix · p. 8

Degree

• The degree k_i of vertex i counts its incident edges, while the total edge count satisfies 2M = sum of all undirected vertex degrees.Directed networks distinguish in-degree k_in and out-degree k_out for incoming and outgoing edges.Source: Degree · p. 9

Degree Distributions

• Degree distributions measure the fraction P(k) of nodes with k connections and are a fundamental directly measurable network property.Source: Degree Distributions · p. 9
• Erdős-Rényi networks have Poissonian degree distributions, and random edge placement produces relatively homogeneous graphs.Source: Degree Distributions · p. 9
• Real networks, including Internet, WWW, email, metabolic, airline, and neuronal networks, generally do not follow the Poisson law.Source: Degree Distributions · p. 9
• Approximate power-law degree distributions define scale-free networks, typically with exponent 2 ≤ λ ≤ 3, although values slightly outside this range also occur.A log-log degree-distribution histogram is a simple way to assess scale-free behavior, while the Barabási–Albert model and its variants model scale-free network evolution.Source: Degree Distributions · p. 9

Clustering Coefficient

• The clustering coefficient measures how strongly nodes cluster at very short scales and is related to the number of triangles in the network.Clustering is high when two nodes sharing a neighbor have a high probability of being connected to each other.Source: Clustering Coefficient · pp. 9–10
• The global clustering coefficient is defined as C = (number of triangles) × 3 / (number of connected triples).The factor of three accounts for each triangle being counted three times among connected triples.Source: Clustering Coefficient · p. 10
• The local coefficient C_i is the fraction of connected neighbor pairs around vertex i, representing the probability that two of i’s friends are connected.For vertices with degree 0 or 1, C_i = 0 because both numerator and denominator are zero.Source: Clustering Coefficient · p. 10
• The whole-network clustering coefficient is obtained by averaging C_i, and clustering values satisfy 0 ≤ C ≤ 1.Source: Clustering Coefficient · p. 10

Subgraph types

• Subgraphs comprise subsets of nodes and edges, including cycles, trees with exactly one path between node pairs, and completely connected k-cliques.A tree of order k has k − 1 edges, whereas a k-clique contains k(k − 1)/2 edges.Source: Subgraphs · p. 10

Subgraph emergence

• In the Erdős–Rényi model, threshold probabilities scale as cN −k/(k−1) for trees, cN −1 for cycles, and cN −2/(k−1) for k-cliques.These thresholds mark the appearance of subgraphs with specified node and edge structures.Source: Subgraphs · p. 10

Components

• Network components are internally connected groups, and a component whose size is proportional to N is called a giant component.Nodes in different components have no connecting path.Source: Subgraphs · p. 10

Communities

• Communities are tightly connected groups with fewer links between groups, often overlapping, making community detection important for understanding network function.Overlapping k-clique communities are illustrated at k = 4.Source: Subgraphs · p. 10, Figure 4 · p. 11

Network Models

• Erdős-Rényi, Barabási-Albert, and Watts-Strogatz models capture random, preferential-growth, and small-world network structures, respectively.The ER model has Poissonian degree distributions; Barabási-Albert growth with preferential attachment produces a power-law degree distribution; Watts-Strogatz interpolates between regular lattices and random graphs with high clustering and L ∝log N [37].Source: Network Models · p. 11

2.1.2. Pearson Correlation Climate Network

• Climate networks represent geographical regions as nodes and statistical dependencies between climatological time series as links, interpreted through underlying climate dynamics.Construction involves selecting nodes and variables, preprocessing time series, measuring interdependencies, thresholding significant edges, and interpreting network structure climatologically.Source: §2.1.2 · pp. 11–12
• Pearson correlation climate networks can use time-delayed cross-correlations to define weighted directed links, with positive lag indicating direction from j to i.In the example, θi,j = 2 and W + i,j = 5.71, representing 5.71 standard deviations above noise and therefore a highly significant link.Source: §2.1.2 · p. 14, Figure 6 · p. 15

2.1.3. Event Synchronization Climate Network

• Event synchronization constructs climate networks from relative event timings, quantifying synchronization and precedence-based delays, making it especially suitable for extreme events such as rainfall above the 99th percentile.The normalized measures satisfy 0 ⩽Qxy ⩽1 and −1 ⩽qxy ⩽1; Qxy = 1 denotes full synchronization, while qxy = 1 means X always precedes Y.Source: §2.1.3 · pp. 15–16

2.1.4. Mutual Information Climate Network

• Mutual information measures dependence and nonlinearity between variables, with MI = 0 exactly for independence, and can construct networks capturing westward propagation of Atlantic multidecadal-oscillation SST anomalies.It is non-negative, symmetric, and computable with a time lag.Source: §2.1.4 · p. 16

2.2. Percolation Theory

• Percolation studies connectivity as occupied sites or bonds form clusters, with critical behavior near pc marking formation or collapse of a giant component containing O(N) nodes.For networks without geometric sides, a giant component replaces the spanning-cluster criterion; site and bond percolation occupy sites or bonds with probability p.Source: §2.2 · pp. 16–17

2.2.1. Phase Transition

• Percolation is a geometrical phase transition in which pc separates finite-cluster and infinite-cluster phases, with occupied probability p serving as the control parameter.This parallels thermal phase transitions, where temperature controls changes between ordered and disordered phases.Source: §2.2.1 · p. 18

Percolation Order Parameter

• P∞ is the fraction of sites belonging to the infinite cluster and serves as the percolation system’s order parameter.For p < pc, only finite clusters exist and P∞ = 0; for p > pc, it behaves near pc as a power law.Source: Percolation Order Parameter · p. 18
• In Fig. 8, P∞ increases monotonically with p for bond and site percolation on the 2D square lattice and site percolation on an ER network.Decreasing p from 1 represents removing or attacking nodes or links, causing collapse after a fraction 1 − pc is removed.Source: Percolation Order Parameter · p. 18
• The quantity 1 − pc measures system resilience by indicating how close the system is to collapse under node or link removal.This interpretation follows from the system collapsing after removal of a fraction 1 − pc when p is gradually decreased from 1.Source: Percolation Order Parameter · p. 18

• Near criticality, the finite-component distribution follows a scaling form.Source: Cluster Size Distribution · p. 18
• At the percolation threshold, sξ diverges as |p − pc|−σ, while the distribution tail follows a power law with critical exponent σ.Here, s denotes component size and ns the number of components of size s.Source: Cluster Size Distribution · p. 18

Average Cluster Size

• The probability that a site belongs to a cluster of size s defines the finite-cluster probability ρ(p).Source: Average Cluster Size · p. 19
• The cluster-size probability distribution w_s(p) is normalized over finite clusters, with its first moment giving the average cluster size.Source: Average Cluster Size · p. 19
• The average cluster size ⟨s(p)⟩, excluding the infinite cluster, diverges near the critical point.Source: Average Cluster Size · p. 19

Correlation Length

• The correlation function g(r) gives the probability that a site at distance r from an occupied site belongs to the same finite cluster.Source: Correlation Length · p. 19
• For large r, g(r) typically has an exponential cutoff g(r) ∼ e^-r/ξ, where ξ is the correlation length measuring mean separation between sites on the same finite cluster.Source: Correlation Length · p. 19
• P∞ and ⟨s⟩ are analogous to magnetization m and susceptibility χ, while β, γ, and ν describe critical behavior at the percolation phase transition.Source: Correlation Length · p. 19

2.2.2. Structural Properties

• Fundamental measurements are introduced to characterize the structural properties of percolation clusters, with further details provided in Ref..Source: §2.2.2 · p. 19

Definition and percolation

• Fractal dimension df measures how the average cluster mass M within radius r changes with r, describing self-similar clusters at the percolation threshold pc.The infinite percolation cluster is self-similar across length scales and can therefore be regarded as a fractal.Source: Fractal Dimension · p. 19
• For scales below the correlation length ξ, clusters exhibit fractal structure, whereas above ξ they become homogeneous.Source: Fractal Dimension · p. 19

Percolation scaling

• Percolation-cluster fractal dimension df is related to the critical exponents β and ν through the scaling relation derived from the preceding equations.The supplied passage states the relation as df = d − β/ν.Source: Fractal Dimension · p. 20

Network fractal analysis

• Complex-network fractal dimensions are commonly estimated using either box counting or cluster growing.Box counting fits the scaling of the number of covering boxes NB with box size lB, while cluster growing fits the mass within a characteristic size.Source: Fractal Dimension · p. 20
• Because networks are generally not spatially embedded, fractal measurement combines box covering with renormalization: covered boxes are replaced by nodes, links are preserved between connected boxes, and the procedure is repeated.The number of boxes NB scales with chemical box size ℓB according to a power law involving df.Source: Fractal Dimension · p. 20

Graph Dimensions dmin and dℓ

• The shortest-path chemical distance scales with spatial distance through the fractal dimension dmin, while dℓ is the graph or chemical dimension related by dℓ = df dmin.Backbone and red-bond exponents additionally describe substructures composing percolation clusters.Source: Graph Dimensions dmin and dℓ · p. 20

2.2.3. Scaling Theory

• Percolation critical exponents satisfy scaling and hyperscaling relations, while universality makes them depend on system dimensionality rather than microscopic interaction details.Finite-size scaling functions characterize behavior near phase transitions, and systems sharing critical-exponent values belong to the same universality class.Source: §2.2.3 · p. 21, §2.2.4 · p. 22

2.2.4. Universal Gap Scaling

• Universal gap scaling uses finite-size scaling of the largest order-parameter gap and introduces six critical exponents applicable to continuous and discontinuous percolation.The framework applies across various lattice and network models through finite-size scaling functions.Source: §2.2.4 · p. 22

Gap Exponents

• β1 = β2 ≈0.104, 1/ν1 = 1/ν2 ≈1/ν, rc(∞) ≈0.5, and df1 = df2 ≈1.895 agree with theoretical values for 2D bond percolation.The six gap exponents characterize the percolation universality class and are related to standard percolation exponents; β1 = 0 indicates a first-order transition, whereas 0 < β1 < 1 indicates a continuous transition.Source: Gap Exponents · p. 23, Figure 10 · p. 24

2.3. Tipping Points Analysis

• Tipping points are critical thresholds where small changes can produce large, potentially abrupt and irreversible shifts across climate, ecological, social, financial, and medical systems.Early-warning analysis commonly uses time-series memory, lag-1 autocorrelation, detrended-fluctuation scaling, and power-spectrum methods, although prediction remains difficult.Source: §2.3 · pp. 24–25

2.3.1. Basic Concepts

• Tipping mechanisms comprise bifurcation-induced, noise-induced, and rate-dependent transitions, with reversible nonlinear tipping also possible.Noise-induced transitions can be fundamentally unpredictable and show no early-warning signals, whereas rate-dependent tipping reflects failure to track a changing quasi-static attractor and has been observed in a global energy-balance model.Source: §2.3.1 · pp. 25–26

2.3.2. Tipping Elements in the Earth’s Climate System

• Policy-relevant climate tipping elements are evaluated by their changing feature, control parameter, critical value, transition timescale, and impacts, including sea ice, ice sheets, AMOC, ENSO, monsoons, and forests.Examples include Greenland’s approximately 3°C threshold and 300-year timescale, West Antarctica’s approximately 5–8°C threshold and 300-year timescale, and AMOC’s approximately 0.1–0.5 Sv threshold and 100-year timescale.Source: §2.3.2 · pp. 27–28

2.3.3. Early Warning of Tipping Points

• Critical slowing down provides an early-warning signal because recovery slows near a threshold, increasing lag-1 autocorrelation and fluctuation variance.The lag-1 autocorrelation increased near modeled thermohaline-circulation bifurcations and before eight documented climate transitions, while DFA exponents over 10 ≤ n ≤ 100 are sensitive to dynamical changes.Source: §2.3.3 · pp. 29–31
• The power-spectrum indicator behaves similarly to ACF(1) and DFA and provides an early warning for tropical cyclones, where ACF(1) fails.Near critical transitions, spectral reddening reflects greater low-frequency variation, and the scaling exponents satisfy α = 1+β 2 = 1 −γ.Source: §2.3.3 · p. 31
• Early-warning precursors include increased variance, flickering between stable states, and changes in skewness and kurtosis near critical transitions.Increased variance was demonstrated near a lake-model bifurcation, while flickering occurred in lake eutrophication and trophic-cascade models.Source: §2.3.3 · p. 31

2.3.4. Precursors of Transitions in Real Systems

• Climate transitions can exhibit increased autocorrelation and flickering, including eight documented transitions and the abrupt end of the Younger Dryas cold period.Abrupt climate changes are framed as critical transitions across events including the Younger Dryas and Dansgaard-Oeschger events.Source: §2.3.4 · p. 32
• Early-warning signals also appear in exploited fish stocks, lake and marine systems, epileptic seizures, asthma attacks, and financial market shocks.Examples include increased fluctuations in harvested fish populations, flickering before epileptic seizures, and increased trade volatility before major shocks.Source: §2.3.4 · p. 32

2.4.1. Introduction

• Entropy is presented as a general complexity measure with interpretations as irreversible change, disorder, and uncertainty, using thermodynamic, statistical-mechanical, information, and quantum definitions.Different entropy definitions can perform better or worse across applications including DNA, cosmology, economics, and climate systems [233–241].Source: §2.4.1 · p. 32

2.4.2. Entropy as Irreversibility

• For irreversible processes, entropy does not decrease in isolated systems; reversible Carnot cycles create no entropy, whereas real irreversible engines increase it.The Carnot-cycle treatment relates heat flow and entropy change, distinguishing reversible entropy transfer from entropy production in irreversible engines.Source: §2.4.2 · pp. 33–34, Figure 14 · p. 33

2.4.3. Entropy as Disorder

• Mixing two equal ideal gases increases entropy by ∆S = NkB log 2, linking spontaneous diffusion and randomized molecular distributions to disorder and complexity.The increase is kB log 2 per molecule because equal temperatures and pressures make mixing occur without heat transfer.Source: §2.4.3 · p. 34

2.4.4. Entropy as Uncertainty

• Information entropy measures uncertainty about system states and, in bits, is maximal for equal probabilities while supporting communication and data-compression applications.For non-equilibrium and equilibrium systems, entropy is defined from probability distributions; quantum non-equilibrium entropy uses a density matrix.Source: §2.4.4 · pp. 34–35

ApEn

• ApEn, introduced by Pincus in 1991, quantifies regularity in time-series data, particularly relatively short and noisy datasets.It was motivated by exact regularity statistics, including Kolmogorov–Sinai entropy and K2 entropy.Source: ApEn · p. 35
• ApEn constructs length-m vectors from the time series, compares them using maximum componentwise distance, and counts matches within a positive tolerance r.The embedding length is m, while r specifies the filtering level.Source: ApEn · p. 35
• Compared with K-S, E-R, and K2 entropy, ApEn is less noise-sensitive, robust to outliers, computationally lighter, and finite for stochastic, noisy deterministic, and composite processes.Increasing ApEn also corresponds intuitively to increasing process complexity.Source: ApEn · p. 36
• ApEn has been applied to classify EEG in psychiatric diseases, including schizophrenia, epilepsy, and addiction.Other reported applications include heart-rate data, finance, psychology, and human factors engineering.Source: ApEn · pp. 35–36

• SampEn improves on ApEn by better matching random-number theory across broad operating conditions and maintaining relative consistency, though it retains residual bias for very short records.Source: SampEn · p. 36
• SampEn compares length-m and length-(m+1) vector matches within tolerance r×SD, using N, m, and r as fixed calculation parameters.The distance between vectors is defined by the maximum absolute difference between corresponding components.Source: SampEn · p. 36
• For each calculation with N, m, and r fixed, a smaller SampEn value indicates greater self-similarity in the data set or less noise.Source: SampEn · p. 36

Method

• SysSampEn generalizes entropy analysis from univariate or bivariate series to multivariate time series with spatial-temporal structure.It constructs template vectors from interdependent series and compares similarity at lengths m and m + p to quantify system disorder.Source: SysSampEn · pp. 36–37

Climate-system application

• A strong positive correlation between previous-year SysSampEn values and El Niño magnitude enabled accurate one-year-ahead forecasting in the Niño 3.4 region.Spatial asynchrony and temporal disorder tests identified effective parameter combinations, with Fig. 15C showing the accuracy-level calculation.Source: SysSampEn · pp. 36, 38, Figure 15 · p. 37

Method

• When N = 1, p = 1, and l_eff = l, SysSampEn is equivalent to classical SampEn.Parameter combinations must be selected carefully because only some provide accurate estimates of system complexity.Source: SysSampEn · p. 38

Complexity validation

• Higher SysSampEn values track higher Lyapunov exponents in logistic-map systems, showing that the measure captures increasing system complexity [241].The logistic-map example used 10 time series generated with identical µ values but different initial conditions, supporting accurate complexity estimation [241].Source: SysSampEn · p. 38, Figure 16 · p. 38

3.1. Climate System

• The climate system comprises atmosphere, hydrosphere, cryosphere, land surface, and biosphere components that exchange mass, heat, and momentum across broad spatial and temporal scales.Greenhouse gases, ice-albedo effects, land-surface processes, and biospheric carbon exchange materially influence these interactions, and the climate system is part of the broader Earth System.Source: §3.1 · pp. 38–39
• Climate models range from simplified energy-balance models to AOGCMs, state-of-the-art ESMs with biogeochemical feedbacks, and regional models for downscaling.These models support climate prediction and projection but can be computationally costly and approximate, while extreme-event forecasting remains challenging because of coupled feedbacks and rare conditions.Source: §3.1 · p. 40

3.1.1. El NiñoâĂŞSouthern Oscillation

• ENSO is operationally classified using the Niño 3.4-region ONI, with El Niño requiring ONI > +0.5°C and La Niña ONI < −0.5°C for at least five consecutive months.ENSO arises from ocean–atmosphere feedbacks: positive Bjerknes feedback supports development, while delayed, western-Pacific, recharge-discharge, and advective-reflective mechanisms contribute to oscillation [273–285].Source: §3.1.1 · pp. 42–44
• ENSO forecasts generally have high skill for only about six months because the spring predictability barrier amplifies errors in coupled ocean–atmosphere feedbacks.Climate-network methods found a cooperative mode one year before warming events and achieved better accuracy than 12-month climate-model forecasts, successfully predicting the 2014–2016 El Niño.Source: §3.1.1 · p. 44
• 70% accuracy with 4% false alarms was achieved by a climate-network and percolation method forecasting El Niño one year ahead.The method detected violent network-structure changes approximately one year before events using the largest size change of the largest percolation cluster, with discontinuous percolation identified by finite-size scaling.Source: §3.1.1 · p. 46
• The forecasting index successfully predicted the onset of the 2018–2019 El Niño event.Source: §3.1.1 · p. 46, Figure 21 · p. 47
• ENSO network analysis identified strongly localized impacts: high in-weight activity occurred in affected regions while relatively few nodes were influenced globally.Figures 23 and 24 compare El Niño and La Niña in-weight maps, temperature anomalies, and the temporal evolution of influenced nodes and average in-weights.Source: §3.1.1 · p. 49, Figure 23 · p. 50, Figure 24 · p. 51
• Network indices objectively discriminated eastern-Pacific and central-Pacific El Niño events, with sharp peaks for eastern-Pacific events but near-normal values for central-Pacific events.Climate-network analysis also supported advance estimation of El Niño impacts and prediction of event types.Source: §3.1.1 · p. 52

3.1.2. Indian Summer Monsoon

• 74% of ISM onset predictions and 84% of withdrawal predictions were successful under the stated timing thresholds, with forecasts made on April 10 and July 25, respectively.The tipping-elements scheme combines extreme-rainfall climate networks with nonlinear-dynamics fluctuations between Eastern Ghats and North Pakistan reference points.Source: §3.1.2 · pp. 52, 54
• ∼0.50 forecast skill at a 5-month lead matched state-of-the-art ISM rainfall models while substantially extending their lead time [324].The network predictor used the previous calendar year’s data, and the 2019 AIRI forecast was 949.87 mm versus an observed 968.3 mm.Source: §3.1.2 · pp. 55–56
• The ISM rainfall network predictor used a South Atlantic key node with r = 0.49 and p < 0.01, while its mechanism involved delayed ENSO and South Atlantic teleconnections [324].Warming of 2.5 ◦C in the key-node region coincided with changing network structure and substantially improving rainfall-prediction skill [324].Source: §3.1.2 · pp. 55–56

3.1.3. Extreme Rainfall

• The event-synchronization network revealed extreme-rainfall organization over distances up to 250 km, predicting spatial extent with accuracy above 70% and up to 100% in some locations.The analysis used 94th- and 90th-percentile thresholds to represent very heavy and heavy rainfall events across South Asia.Source: §3.1.3 · p. 59
• More than 60% of Central Andes rainfall events above the 99th percentile were predicted, reaching 90% during El Niño conditions.The proposed mechanism involves northward-migrating frontal systems and a low-level wind channel from the western Amazon to the subtropics.Source: §3.1.3 · p. 59
• Extreme rainfall in South Eastern South America precedes events along the eastern Andean slopes to western Bolivia, enabling prediction in the eastern Central Andes within 2 days.The passage states that these events could not be forecasted using other methods.Source: §3.1.3 · p. 60
• Global extreme-rainfall networks show a power-law distance distribution below 2,500 km and distinct regional-weather and global-teleconnection links, with Rossby waves underlying the teleconnections.South-central Asia connects prominently with eastern Asia, the African tropics, Europe, North America, and Southern Hemisphere extratropics.Source: §3.1.3 · p. 60, Figure 31 · p. 62

3.1.4. Atmospheric Circulation and Global Warming

• Climate change is associated with a 2° poleward Hadley-cell expansion from 1979 to 2005, weakening circulation, drier tropical/subtropical regions, and poleward tropical-cyclone migration.These effects are linked to the Hadley cell’s role in transporting energy and heat poleward.Source: §3.1.4 · p. 63
• The network-percolation approach detected abrupt climate-network transitions, robust poleward expansion and weakening of the tropical cluster in reanalysis data and 31 CMIP5 models.The largest-cluster relative size jumped at rc ≈0.53, and projected impacts strengthen in northern India, southern Africa, and western Australia but weaken in the Northeast Pacific.Source: §3.1.4 · pp. 65, 67

3.1.5. Atlantic Meridional Overturning Circulation

• The AMOC weakened by about 3±1 Sv, around 15%, since the mid-twentieth century, with the decline since the 1950s very likely largely anthropogenic.The slowdown appears in CM2.6 simulations and HadISST observations and is mainly attributed to gradually varying freshwater forcing in the northern North Atlantic.Source: §3.1.5 · p. 67, Figure 34 · p. 67
• AMOC network indicators increase toward collapse, while the kurtosis Kd in a FAMOUS-model hosing experiment provides a signal at least 100 years before transition when variance and lag-1 autocorrelation do not.In the idealized model, Ed, Ec, Kd, and Kc increased near the tipping point, with Kd and Kc showing especially strong increases.Source: §3.1.5 · pp. 68, 71, Figure 36 · p. 70

3.2. Earth Geometric Surface Relief

• Percolation analysis identifies present mean sea level as the critical threshold of global topography, with abrupt relief-network transitions indicating continental or cluster aggregation.This framework may identify critical locations more exposed to global climate change.Source: §3.2 · p. 71

3.2.1. Self-similarity and Long-range Correlations

• Earth’s topography exhibits self-similarity, long-range correlations, scaling spectra, and a bimodal continental–oceanic distribution, with H ≃0.5 globally but H = 0.46, 0.66, and 0.77 for bathymetry, continents, and margins.The scaling relation is S(k) ∼k^-βc, with βc = 2H + 1.Source: §3.2.1 · pp. 71–72

3.2.2. Landmass and Oceanic Clusters

• The landmass order parameter jumps and χ and ξ diverge at the present mean sea level h = 0, identifying a geometrical percolation threshold.For oceanic clusters, the order parameter jumps near −3640 m; lunar topography instead shows equal land and ocean areas at its threshold.Source: §3.2.2 · pp. 72, 74

3.2.3. Origin of the Discontinuity

• Earth’s relief network exhibits discontinuous landmass and oceanic transitions, with critical thresholds pc ≈0.321 at h = −43 m and pc ≈0.379 at h = −3621 m.At the landmass threshold, four major communities appear, and a critical node connects Afro-Eurasia with the Americas; the oceanic critical node connects the Atlantic+Indian Ocean Plate to the Pacific Plate.Source: §3.2.3 · pp. 74–75, Figure 40 · p. 75

3.3. Earthquake Systems

• Earthquake research emphasizes probabilistic forecasting because deterministic prediction of event sizes, locations, and times remains challenging for complex spatio-temporal seismic processes.The review covers phenomenological occurrence laws, scale invariance, intertime distributions, and memory analysis for improving forecasts from real earthquake catalogs.Source: §3.3 · p. 77

3.3.1. Scale Invariance and the inter-event Distribution

• The Gutenberg–Richter b-value for the Israeli catalog is 0.97±0.02, while inter-event distributions from multiple catalogs collapse onto a universal scaling function.Typical tectonic earthquakes have b ≃1, although later analyses report that universality may be approximate or absent, with bimodality from correlated aftershocks and independent events.Source: §3.3.1 · pp. 78, 80, Figure 42 · p. 79, Figure 44 · p. 80

3.3.2. Modeling Seismic Time Series by the Point Process Approach

• The ETAS point-process model represents earthquake clustering through conditional occurrence rates and has supported operational forecasts for the Amatrice-Norcia sequence, Japan, and California.Its parameters characterize regional seismic activity and can vary spatially and temporally.Source: §3.3.2 · pp. 80–81

3.3.3. Memory Analysis

• DFA finds robust long-term seismic memory, with α ≈0.75 across Californian catalogs, Israeli and Italian records, magnitude thresholds, time windows, event counts, released energy, and an ETAS model.For the ETAS example, parameters A = 6.26, µ = 0.2, p = 1.1, αM = 1.5 and c = 0.007 yield α ≈0.75.Source: §3.3.3 · pp. 82–83, Figure 45 · p. 82
• Conditional probability finds that short recurrence times tend to follow short ones and long recurrence times follow long ones, indicating dependence on the preceding interval [389].The method compares recurrence-time distributions conditioned on extreme quantiles Q1 and Q4 with the unconditional distribution.Source: §3.3.3 · p. 83
• Earthquake recurrence memory is robust across catalogs and magnitudes, with Israeli values ρ1 = 0.248 and ρ4 = −0.193 and Italian values ρ1 = 0.260 and ρ4 = −0.153.Higher |ρ1| or |ρ4| indicates stronger memory; the reported correlations do not depend on Mz.Source: §3.3.3 · p. 83
• ETAS reproduces empirically observed earthquake memory only over a narrow parameter range; smaller background rate μ strengthens memory through overlapping aftershock subsequences.The branching-ratio dependence also strengthens memory for smaller Omori exponent p and larger aftershock-productivity exponent αM.Source: §3.3.3 · p. 83
• Lagged conditional probability measures long-term memory in interevent times and distances by comparing the distributions following Q1 and Q3 conditions.The memory measure S lies in [0, 1] and is defined from the nonoverlapping area of the conditional probability densities.Source: §3.3.3 · p. 83
• For the Italian catalog, lagged-memory scaling parameters are du = 0.14 and a = 0.09 for interevent times, versus du = −0.08 and a = 0.24 for distances.These parameters describe the reported scaling functions for S(τk|τ0) and S(rk|r0).Source: §3.3.3 · p. 83

Q Q Q

• Real earthquake memory measures collapse onto a single rescaled function with a crossover near k · 10^bMz ≃ 10^5, whereas ETAS memory is weaker at short timescales, stronger at long timescales, and lacks this crossover.The crossover appears for different grid sizes and magnitude thresholds in the Italian catalog, while the ETAS discrepancy distinguishes model from observed catalogs.Source: Q Q Q · p. 86, Figure 48 · p. 86

3.3.4. Earthquake Forecasting

• Earthquake forecasting estimates the conditional occurrence probability PΣ(ω|ℓ) by integrating the local conditional rate λ over a hypercell, with accuracy affected by cell volume and temporal extent.Forecasting horizons range from long-term decades to centuries, intermediate-term months, and short-term seconds to weeks; ETAS-based models support long-term and post-seismic forecasting more reliably than foreshock forecasting.Source: §3.3.4 · pp. 86–87
• ROC analysis varies an alarm threshold to measure hits, false alarms, misses, and correct rejections, with perfect prediction at TPR = 1 and FPR = 0.The Molchan diagram complements ROC by plotting miss rate against alarmed space-time volume, where (F = 0, M = 0) is optimal and small ΓM indicates high skill.Source: §3.3.4 · pp. 87, 89
• The N-test evaluates whether forecast and observed event counts agree, while the S-test evaluates spatial likelihood consistency, with values near 1 indicating optimal spatial forecasts.Additional evaluation methods include the L-test, R-test, and M-test.Source: §3.3.4 · p. 89
• More than 20%: incorporating memory improved 14-day cumulative aftershock forecasting after the Capitignano M5.7 earthquake, while a revised ETAS model improved forecasts for Italian and Southern California catalogs [402].The memory-consistent ETAS parameterization was μ = 0.2, A = 6.26, αM = 1.5, p = 1.1, and c = 0.007.Source: §3.3.4 · p. 90

Methodological perspectives

• Statistical physics, complex networks, percolation, tipping points, entropy, and complexity methods remain promising but require a comprehensive framework grounded in physical mechanisms, multilayer networks, nonequilibrium critical phenomena, and multivariate tipping-point principles.The authors identify these four issues as central unresolved methodological challenges for complex Earth-system analysis.Source: §4 · pp. 90–91

Climate systems

• Climate-network interactions and patterns can significantly improve predictions across ENSO, monsoons, extreme rainfall, atmospheric circulation, and Atlantic meridional overturning circulation, while future work should address human forcing, models, abrupt paleoclimate transitions, and severe convection.Other applications include Atlantic multidecadal oscillation propagation, teleconnections, causal gateways, polar-vortex prediction, and CO2 impacts on global surface air temperature.Source: §4 · p. 91

Earth surface topography

• Earth-relief networks show abrupt transitions consistent with continental or cluster aggregation, while fractional Brownian-motion analysis suggests long-range correlations and motivates studying tectonic-plate movement and critical-node evolution.These findings may clarify geometrical phase transitions and help identify critical nodes relevant to future global change.Source: §4 · p. 91

Earthquakes

• DFA and CP memory analyses can significantly improve short-term forecasting rates for real earthquake catalogs, but earthquake prediction remains controversial and requires models of foreshocks, precursors, and occurrence-time or seismic-signal networks.The review also covers Gutenberg–Richter, Omori, productivity, inter-event scaling, ETAS, and forecast-evaluation methods.Source: §4 · p. 91

Artificial intelligence and networks

• Combining network-based methods with artificial intelligence and deep learning could strengthen forecasting of climate phenomena, aftershocks, and earthquakes while revealing computational-network structure.The authors view these approaches as complementary skills that can boost one another across Earth-system science and statistical physics.Source: §4 · pp. 91–92

Resilience and climate–health research

• Future research should use complex networks to model climate resilience, cascading failures in interdependent infrastructure, and climate–health relationships, including climate-sensitive diseases and epidemiological models.Such integration is presented as important for understanding climate-change health consequences and achieving climate and infrastructure resilience.Source: §4 · p. 92

Appendix A. Acronyms

• The appendix organizes acronyms into scientific and institutional categories, presented in Tables A.2 and A.3, respectively.Table A.2 contains scientific acronyms, while Table A.3 contains institutional acronyms.Source: Appendix A · p. 92, Table A.2 · p. 93, Table A.3 · p. 93