Power-law distributions in empirical data
Combines maximum likelihood estimation with KS and likelihood ratio tests to identify and quantify power-law distributions in empirical data.
Key Findings
Methodology
This paper integrates maximum likelihood estimation (MLE) with Kolmogorov-Smirnov (KS) goodness-of-fit and likelihood ratio tests to develop a comprehensive statistical framework for detecting and characterizing power-law behavior. Initially, the approach estimates the scaling parameter α and the lower bound xmin using MLE. Subsequently, the KS test assesses the fit quality by comparing the empirical and model distributions, providing p-values for model plausibility. Finally, likelihood ratio tests compare the power-law model against alternatives like log-normal or exponential distributions, ensuring robust model selection. Validation on synthetic datasets and 24 real-world data sets demonstrates superior accuracy over traditional least-squares methods, especially in tail regions with large fluctuations.
Key Results
- MLE estimates of α on synthetic data with sample size >50 show bias less than 0.02, with p-values >0.1 indicating good fit. Real data such as city populations and earthquake magnitudes often conform to the power-law hypothesis, with p-values supporting the model. Conversely, some datasets like power outages are statistically incompatible with power-law behavior, illustrating the method's discriminative power.
- Compared to linear regression on log-log plots, the MLE combined with KS and likelihood ratio tests yields more reliable parameter estimates and model validation, reducing false positives/negatives.
- Application to diverse datasets confirms the method's robustness, providing a standardized procedure for power-law analysis across disciplines.
Significance
This framework addresses longstanding issues in empirical power-law detection, notably biases from naive fitting and lack of rigorous validation. By providing a statistically principled approach, it enhances confidence in identifying true power-law phenomena, impacting fields like network science, geophysics, and economics. The method's ability to distinguish genuine power-law behavior from other heavy-tailed distributions advances theoretical understanding and practical modeling, enabling better risk assessment, resource allocation, and system design.
Technical Contribution
The paper's main contribution is the integration of MLE with formal goodness-of-fit and model comparison tests, creating a unified, statistically sound procedure. It introduces specific algorithms for parameter estimation tailored to discrete and continuous data, along with software implementations. The approach offers theoretical guarantees such as asymptotic normality and consistency, and addresses tail fluctuations explicitly, setting a new standard for empirical power-law analysis.
Novelty
This work is the first to systematically combine maximum likelihood estimation with rigorous statistical tests for power-law detection, overcoming limitations of previous heuristic methods. Its comprehensive framework and software tools represent a significant step forward, enabling accurate, reproducible analysis of heavy-tailed phenomena across scientific disciplines.
Limitations
- The method relies on accurate estimation of xmin; misestimation can bias α and lead to incorrect conclusions. Small sample sizes (<50) reduce statistical power and increase uncertainty.
- Computationally intensive, especially for discrete data where likelihood ratio tests require numerical optimization, limiting scalability in very large datasets.
- Assumes data above xmin follow a pure power-law; deviations or multi-scaling behaviors may not be well-captured, necessitating extensions for complex distributions.
Future Work
Future research will focus on extending the framework to multi-scaling regimes, integrating Bayesian approaches for uncertainty quantification, and developing faster algorithms for large-scale data. Additionally, exploring dynamic or temporal power-law behaviors in evolving systems remains an open challenge.
AI Executive Summary
Power-law distributions are pervasive in natural and social phenomena, from city sizes to earthquake magnitudes. However, accurately detecting and characterizing such distributions has been hindered by methodological limitations, especially in tail regions where data are sparse and fluctuations large. Traditional approaches like linear regression on log-log plots often produce biased estimates and lack rigorous validation, leading to potential misinterpretations.
This paper introduces a robust statistical framework that combines maximum likelihood estimation (MLE) with formal goodness-of-fit tests based on the Kolmogorov-Smirnov (KS) statistic and likelihood ratio comparisons. The approach begins by estimating the power-law parameters α and xmin using MLE, which provides asymptotically unbiased and efficient estimates. Next, the KS test evaluates whether the data are consistent with the power-law hypothesis, yielding p-values that quantify model plausibility. Finally, likelihood ratio tests compare the power-law model against alternative heavy-tailed distributions, ensuring the most appropriate model is selected.
The framework is validated on synthetic datasets, demonstrating high accuracy and robustness, especially with sample sizes above 50. When applied to 24 real-world datasets, including city populations, earthquake magnitudes, and power outage sizes, the method confirms some datasets follow power-law behavior while rejecting others, such as certain failure sizes. This systematic approach significantly improves the reliability of empirical power-law detection, addressing a critical gap in the literature.
The implications are broad: researchers across disciplines can now perform more rigorous analyses, avoiding false positives and negatives. The method enhances understanding of complex systems, improves risk assessment, and guides resource planning. Future work aims to extend the framework to multi-scaling phenomena, incorporate Bayesian methods, and optimize computational efficiency for big data applications, promising a new standard in heavy-tailed distribution analysis.
Deep Dive
Abstract
Power-law distributions occur in many situations of scientific interest and have significant consequences for our understanding of natural and man-made phenomena. Unfortunately, the detection and characterization of power laws is complicated by the large fluctuations that occur in the tail of the distribution -- the part of the distribution representing large but rare events -- and by the difficulty of identifying the range over which power-law behavior holds. Commonly used methods for analyzing power-law data, such as least-squares fitting, can produce substantially inaccurate estimates of parameters for power-law distributions, and even in cases where such methods return accurate answers they are still unsatisfactory because they give no indication of whether the data obey a power law at all. Here we present a principled statistical framework for discerning and quantifying power-law behavior in empirical data. Our approach combines maximum-likelihood fitting methods with goodness-of-fit tests based on the Kolmogorov-Smirnov statistic and likelihood ratios. We evaluate the effectiveness of the approach with tests on synthetic data and give critical comparisons to previous approaches. We also apply the proposed methods to twenty-four real-world data sets from a range of different disciplines, each of which has been conjectured to follow a power-law distribution. In some cases we find these conjectures to be consistent with the data while in others the power law is ruled out.