Fitting Power-laws in empirical data with estimators that work for all exponents

Abstract

It has been repeatedly stated that maximum likelihood (ML) estimates of exponents of power-law distributions can only be reliably obtained for exponents smaller than minus one. The main argument that power laws are otherwise not normalizable, depends on the underlying sample space the data is drawn from, and is true only for sample spaces that are unbounded from above. Here we show that power-laws obtained from bounded sample spaces (as is the case for practically all data related problems) are always free of such limitations and maximum likelihood estimates can be obtained for arbitrary powers without restrictions. Here we first derive the appropriate ML estimator for arbitrary exponents of power-law distributions on bounded discrete sample spaces. We then show that an almost identical estimator also works perfectly for continuous data. We implemented this ML estimator and discuss its performance with previous attempts. We present a general recipe of how to use these estimators and present the associated computer codes.

arXiv comment11 pages, 3 figures

References in corpus (11)

• Emergence of scaling in random networks
• Statistical mechanics of complex networks
• Power-law distributions in empirical data
• Power laws, Pareto distributions and Zipf's law
• How Popular is Your Paper? An Empirical Study of the Citation Distribution
• Problems with Fitting to the Power-Law Distribution
• On the Frequency of Severe Terrorist Events
• Power-law distributions in binned empirical data
• Parameter estimation for power-law distributions by maximum likelihood methods
• Understanding scaling through history-dependent processes with collapsing sample space
• How multiplicity determines entropy and the derivation of the maximum entropy principle for complex systems

Cited by in corpus (9)

• Testing statistical laws in complex systems
• Traditional and Heavy-Tailed Self Regularization in Neural Network Models
• The distinct flavors of Zipf's law in the rank-size and in the size-distribution representations, and its maximum-likelihood fitting
• Sample space reducing cascading processes produce the full spectrum of scaling exponents
• Tail of the distribution of fatalities in epidemics
• On the survival of the quantum depletion of a condensate after release from a magnetic trap
• Non-parametric power-law surrogates
• Independent Approximates enable closed-form estimation of heavy-tailed distributions
• Power Law in Sparsified Deep Neural Networks