Source-linked AI summary
Complex networks in climate dynamics - Comparing linear and nonlinear network construction methods
Jonathan F. Donges, Yong Zou, Norbert Marwan, Jürgen Kurths
TL;DR
The paper addresses whether climate-network topology depends on using linear Pearson correlation or nonlinear mutual information to measure interdependence. It systematically compares the resulting networks across local, mesoscopic, and global scales, finding strong local and mesoscopic agreement but larger global differences, especially in betweenness centrality.
Problem
Linear correlation may not fully capture nonlinear climate relationships, motivating evidence on how Pearson- and mutual-information-based climate networks compare across topological scales.
Method
The study constructs networks from surface air temperature data using Pearson correlation and mutual information, tests statistically significant connections, and compares network measures at equal edge densities.
Results
The networks agree well on local and mesoscopic scales for AOGCM and reanalysis data, while larger differences occur globally, particularly in betweenness centrality.
Takeaways & Limitations
Mutual-information climate networks offer new perspectives for detecting network structures based on nonlinear physical processes.
Takeaways & Limitations
Further work is needed to establish whether the observed global-scale deviations are due to nonlinear physical processes in the climate system.
Abstract
from arXiv · showhide
Complex network theory provides a powerful framework to statistically investigate the topology of local and non-local statistical interrelationships, i.e. teleconnections, in the climate system. Climate networks constructed from the same global climatological data set using the linear Pearson correlation coefficient or the nonlinear mutual information as a measure of dynamical similarity between regions, are compared systematically on local, mesoscopic and global topological scales. A high degree of similarity is observed on the local and mesoscopic topological scales for surface air temperature fields taken from AOGCM and reanalysis data sets. We find larger differences on the global scale, particularly in the betweenness centrality field. The global scale view on climate networks obtained using mutual information offers promising new perspectives for detecting network structures based on nonlinear physical processes in the climate system.
1 Introduction
Climate networks represent spatial grid points as vertices and statistical interdependence between anomaly time series as edges, enabling analysis of local, mesoscopic, and global climate-system structure. This study motivates comparing Pearson correlation with mutual information because nonlinear climate processes may not be fully captured by linear dependence.
- Climate networks map global-data grid points to vertices and connect pairs according to statistical interdependence between their anomaly time series.
- Teleconnections are richly structured long-range correlations that extend beyond the locality assumed in climate-system dynamics.
- Local degree, mesoscopic clustering, and global path-based measures reveal climate-network topology across multiple spatial scales.High-degree regions can be associated with atmospheric teleconnection patterns such as the North Atlantic Oscillation.
- Pearson correlation has been widely used, but mutual information is introduced to capture nonlinear relationships between climate time series.
- The paper describes data processing, network theory, construction methods, systematic comparison, climatological interpretation, and conclusions.
2 Data
The study constructs climate networks from monthly global surface air temperature fields using reanalysis and model data. The time series are converted to anomalies, normalized, and evaluated over the full data set after seasonal-bias checks.
- Monthly global surface air temperature fields provide a common basis for studying atmospheric and oceanic dynamics.The data include NCEP/NCAR reanalysis and a CMIP3 HadCM3 twentieth-century reference run.
- The data sets associate a time series x_i(t) with every spatial grid point i on a regular spatiotemporal grid.
- 2.2 Filtering and normalization: Anomaly values are obtained by removing the mean annual cycle through phase averaging, then each anomaly time series is normalized to unit variance.
- 2.2 Filtering and normalization: Using one particular season instead of the whole data set does not substantially alter the results, so the full data set is retained.
3 Elements of complex network theory
The paper models climate networks as undirected, unweighted simple graphs and organizes network measures by the topological information they use. Geographic grid inhomogeneity motivates area-weighted measures.
- A climate network is an undirected, unweighted simple graph G := (V, E) with vertices and edges but no multiple edges or self-loops.
- Local measures use direct-neighborhood information, mesoscopic measures use neighbors and next neighbors, and global measures rely on shortest paths.
- Fields assign a real-valued measure to each vertex, whereas scalar measures produce one value for the whole graph.
- Because grid vertices represent unequal surface areas, area-weighted generalizations are used to reduce geographic bias in network measures.Studies across interpolated grids and resolutions found that the reported results were not altered.
3.1 Local measures
The paper’s local-scale measures characterize direct connectivity and graph similarity, including degree-based quantities, area-weighted connectivity, and Hamming distance. Area weighting accounts for unequal geographic cell areas.
- Degree centrality: Degree centrality k_v counts the first neighbors of vertex v, and exceptionally high-degree vertices are called hubs or super-nodes.
- Area weighted connectivity: Area weighted connectivity generalizes degree centrality for grid cells representing different surface areas.For angularly equidistant grids, cell area is proportional to cos(λ_v), and AWC_v is the connected fraction of Earth’s surface area.
- Hamming distance: Hamming distance measures the fraction of edges that must change to transform one labeled simple graph into another.In this application, the compared graphs have approximately equal edge density ρ.
- Hamming distance: The Erdős-Rényi reference distance is H_R(ρ) = 2ρ(1 − ρ) for two independent random graphs with edge density ρ.
3.2 Mesoscopic measures
Mesoscopic measures characterize neighborhood structure through clustering coefficients, including local and graph-level variants.
- Local clustering coefficient: The local clustering coefficient C_v measures the probability that two randomly chosen first neighbors of v are also neighbors.It is also called the Watts-Strogatz clustering coefficient.
- Global clustering coefficient: The global clustering coefficient C is defined as the mean Watts-Strogatz clustering coefficient.
3.3 Global measures
Global measures summarize how vertices relate to the entire network, using distances and shortest-path mediation.
- Closeness centrality: Closeness centrality CC_v measures the inverse average topological distance from vertex v to all other vertices.Under this definition, larger CC_v means that v is topologically closer to the rest of the network.
- Closeness centrality: Topological distance d_ij is the minimum number of edges crossed on a path from vertex i to vertex j.For disconnected pairs, d_ij = N − 1 is used, and closeness centrality is normalized to 0 ≤ CC_v ≤ 1.
- Betweenness centrality: Betweenness centrality identifies vertices that mediate information transport by counting how many shortest paths traverse them.Shortest-path contributions are weighted by their multiplicity σ_ij.
- Average path length: The average or characteristic path length L is the average topological distance between all pairs of vertices.Disconnected pairs are excluded from this average.
4 Constructing climate networks
The climate-network construction is motivated by synchronization in nonlinear dynamical systems and is designed to represent dynamical correlations between climate regions.
- Physical rationale: Dynamical correlations in a discretized climate model can be viewed as partial synchronization among nonlinear oscillators on a locally connected grid.Even this simple topology can produce nontrivial spatial synchronization patterns.
- Physical rationale: The synchronization framework guides understanding of nonlinear teleconnections and motivates more advanced measures for detecting them in measured climate data.
- Physical rationale: The authors propose embedding network construction from multivariate climatological data within the framework of synchronization in complex networks.
4.1 Correlation measures
The study uses Pearson correlation as a standard linear measure and mutual information as a nonlinear cross-check for climate-network construction.
- Pearson correlation: Pearson correlation is used first for simplicity and consistency with the literature, then compared with mutual information.
- Mutual information: Mutual information is introduced to investigate nonlinear dynamical relationships that Pearson correlation may not fully detect.A strongly nonlinear relationship can yield large M_ij but small P_ij.
- Pearson correlation: The empirical Pearson coefficient estimates the strength of a linear relationship between two normalized time series under normality.It can produce spurious results for non-normal observables and nonlinear relationships, so it requires care in network construction.
- Pearson correlation: The absolute Pearson correlation P_ij = |R_ij| treats both strong negative and strong positive correlations as strong linear interdependence.
- Mutual information: Mutual information M_ij is estimated from the marginal and joint probability density functions of paired time series and is symmetric.Its standard unit is the bit when logarithms to base 2 are used.
- Mutual information: A histogram estimator with 64 equally sized bins is applied consistently across all pairs to improve comparability of M_ij.For typical time series of length O(10^3), this meets the Cochran criterion of at least 5 samples per bin.
4.2 Obtaining the network adjacency matrix
The climate network is constructed by thresholding the correlation-measure matrix, linking vertex pairs whose measure exceeds τ. The resulting adjacency matrix is symmetric and represents an undirected, unweighted simple graph.
- Pairs of vertices {i, j} are linked when their correlation measure Cij exceeds the threshold τ.
- The adjacency matrix Aij is defined from the thresholded correlation-measure matrix using the Heaviside function Θ(x).
- Because Aij inherits symmetry from Cij, the climate network is an undirected and unweighted simple graph.
4.3 Choosing the threshold
Threshold selection balances statistically significant connections against the richness of network structures, with teleconnections providing important long-range shortcuts. The comparison fixes edge density across Pearson-correlation and mutual-information networks, while network measures vary smoothly over the studied density range.
- Threshold selection balances statistical significance against the richness of network structures revealed by the climate network.Different thresholds expose different features of the correlation-measure matrix.
- Long-distance edges beyond approximately 15000 km enter at τ ≲0.65 for Pearson correlation and τ ≲0.3 for mutual information.These edges are identified as teleconnections and provide spatial shortcuts associated with non-trivial network features.
- Teleconnections are treated as a necessary criterion when choosing thresholds for interesting climate-network analysis.
- At edge density ρ = 0.005, the corresponding thresholds are τ = 0.682 for Pearson correlation and τ = 0.398 for mutual information.The thresholds are obtained from the correlation-measure probability distributions.
- Edge density ρ(τ) decreases monotonically with τ and, for connected empirical supports, establishes a one-to-one correspondence between threshold and edge density.Its approximately exponential decay follows from the shape of the correlation-measure distribution.
- The clustering coefficient remains approximately constant at intermediate ρ but decays to zero at small ρ, where the network splits into smaller components.Average path length decreases approximately as a power law with increasing ρ and shows discontinuities associated with high-edge-betweenness links.
- The studied edge-density range keeps the giant component size at O(1), making non-giant components negligible for average path length and closeness centrality.Very small edge densities require efficiency measures for more robust analysis of disconnected components.
- Comparisons between Pearson and mutual-information networks fix edge density ρ, yielding different thresholds because their empirical measure distributions differ.The selected density balances structural richness and statistical significance.
5 Results
The study compares Pearson-correlation and mutual-information climate networks across local, mesoscopic, and global topological scales. Similarity is high locally and mesoscopically, while global differences are more pronounced, especially for betweenness centrality.
- Comparison design: The comparison varies edge density from ρmin = 0 to ρmax = 0.1 and evaluates network properties across local, mesoscopic, and global scales.Fields are also compared visually at ρ = 0.005 using area weighted connectivity, clustering, closeness, and betweenness measures.
- Local comparison: At low edge densities, Pearson and mutual-information networks are highly similar locally, with high area weighted-connectivity rank correlations at ρ = 0.005 and ρ = 0.01.Their Hamming distance remains below the random-network expectation and the difference is statistically significant across the considered edge densities.
- Mesoscopic comparison: Pearson and mutual-information networks also show high mesoscopic similarity, with nearly indistinguishable local clustering fields and global clustering deviations of O(10^-2).The largest local deviations cluster along coastlines, where differing oceanic and continental SAT dynamics increase disagreement over adjacent edges.
- Global comparison: At the global scale, closeness fields differ little, whereas betweenness fields show more pronounced regional differences, particularly over oceanic regions.Betweenness rank correlations are notably smaller than those of the other fields, while closeness rank correlations are close to unity.
- Global comparison: Betweenness centrality may reveal local differences between Pearson and mutual-information networks that trace nonlinear physical processes.Its sensitivity to a small number of bridging edges produces large changes and a dynamic range spanning 20 orders of magnitude.
- Climatological interpretation: The climatological interpretation links network structures to teleconnections, ENSO-related tropical-Pacific similarity, Coriolis effects, land–sea contrasts, and current-like betweenness features.The equatorial East Pacific combines low connectivity, closeness, and betweenness with high local clustering, indicating a dense but nearly detached network cluster.
6 Conclusions
Pearson-correlation and mutual-information climate networks agree well locally and mesoscopically across AOGCM and reanalysis surface air temperature data, but show localized, structured deviations globally, especially in betweenness centrality. These deviations may reflect nonlinear physical processes, although further work is needed to establish that interpretation.
- A systematic comparison across local, mesoscopic, and global scales found strong agreement between Pearson-correlation and mutual-information climate networks.
- The local and mesoscopic agreement held for AOGCM and reanalysis surface air temperature data, while surface pressure networks yielded the same qualitative conclusions.
- Global differences in surface air temperature networks were qualitatively and quantitatively most evident in the betweenness centrality field.
- The remaining global deviations were highly localized and structured, pointing to possible involvement of nonlinear physical processes.
- Further work is needed to determine whether the global-scale deviations arise from nonlinear physical processes detectable only with mutual information.