Wilks' theorem

Wilks' theorem
Name	Samuel S. Wilks
Birth date	1906
Death date	1964
Field	Statistics
Known for	Likelihood ratio tests

Wilks' theorem Wilks' theorem gives the asymptotic distribution of the likelihood ratio statistic in parametric hypothesis testing, forming a cornerstone of modern statistical inference. The result connects parametric models, such as those used by Ronald Fisher, Jerzy Neyman, Egon Pearson, and Karl Pearson, with asymptotic theory developed in the traditions of Andrey Kolmogorov, Harald Cramér, and Norbert Wiener. It underlies practical procedures employed in institutions like National Institutes of Health and Food and Drug Administration and informs work in fields from Astronomy to Econometrics.

Statement

Wilks' theorem states that for nested parametric models with parameter vectors in Euclidean space, under a simple null hypothesis the twice log-likelihood ratio statistic converges in distribution to a chi-squared law. The formulation is standard in texts associated with Jerzy Neyman, Egon Pearson, R.A. Fisher, and later expositions by George Box, David Cox, and Bradley Efron. Formally, if the null restricts r parameters within a p-dimensional model estimated by maximum likelihood, then -2 times the log-likelihood ratio converges to a chi-squared distribution with r degrees of freedom; this result is used in practice by researchers at Harvard University, Princeton University, and Stanford University.

Assumptions and Regularity Conditions

The asymptotic chi-squared conclusion requires regularity conditions that mirror those in classical works by Andrey Kolmogorov, Harald Cramér, Wassily Hoeffding, and Lehmann and Romano. Conditions include differentiability of the log-likelihood, identifiability of parameters, interiority of true parameter values relative to parameter space, and nonsingularity of the Fisher information matrix. Additional technical hypotheses echo the frameworks used by Louis Bachelier and Norbert Wiener: existence of third derivatives locally, dominated convergence conditions, and uniform law of large numbers properties as in results associated with Andrei Kolmogorov and S.R. Srinivasa Varadhan.

Proof Sketch and Intuition

The proof proceeds by Taylor expansion of the log-likelihood around the true parameter value, invoking the asymptotic normality of maximum likelihood estimators established by methods of C.R. Rao, Harris Teeter? (note: remove spurious), and Harald Cramér. The score vector converges to a multivariate normal distribution with covariance equal to the Fisher information matrix, a fact linked historically to works by R.A. Fisher and C.R. Rao. Substituting the quadratic approximation yields that the likelihood ratio is asymptotically equivalent to a quadratic form in the score, which through an orthogonal transformation reduces to a sum of squares of independent standard normals and hence a chi-squared law. This line of reasoning parallels methods used in asymptotic analysis by Andrey Kolmogorov, Norbert Wiener, and Sergei Bernstein.

Applications and Examples

Wilks' theorem is applied across scientific domains that rely on parametric models developed at institutions like Bell Labs, Bell Laboratories, and Los Alamos National Laboratory. Examples include likelihood-ratio tests in generalized linear models used in analyses by Bradley Efron and Peter McCullagh, model comparison in nested time-series models influenced by Clive Granger and Robert Engle, and mixture-model inference in genetics informed by work at Cold Spring Harbor Laboratory. In phylogenetics and systematics, practitioners at Smithsonian Institution and Natural History Museum, London use likelihood-ratio approaches; in particle physics experiments at CERN, likelihood-based methods assess signal significance. Econometric model selection employing likelihood ratios reflects traditions from John Maynard Keynes-inspired statistical economics and later developments at Cowles Commission.

Limitations and Extensions

Wilks' theorem can fail when regularity conditions are violated: boundary parameters, nonidentifiability, singular Fisher information, or models with increasing number of parameters as in high-dimensional statistics. Such failures motivated extensions by researchers like Bradley Efron, Persi Diaconis, Terry Speed, and contemporary work by groups at Princeton University and MIT. Alternatives and refinements include Chernoff-type results, bootstrap methods advocated by Bradley Efron and R.J. Diakonikolas? (note: remove spurious), higher-order asymptotics due to H. Daniels and John W. Tukey, and nonstandard limits studied by researchers associated with Institute for Advanced Study and Courant Institute. In modern machine-learning contexts at Google, OpenAI, and DeepMind, practitioners adapt likelihood-based testing with penalization, resampling, and Bayesian model comparison inspired by Thomas Bayes and Harold Jeffreys.

Historical Context and Attribution

The theorem is attributed to Samuel S. Wilks in the mid-1930s, situated in the lineage of statistical theory that includes R.A. Fisher, Jerzy Neyman, Egon Pearson, and C.R. Rao. Wilks's contributions were contemporaneous with foundational work emerging from centers at Columbia University, Princeton University, and University of London, and his results were influential for subsequent developments in hypothesis testing, asymptotics, and the formalization of likelihood methods used in institutions such as Royal Society and American Statistical Association.

Category:Statistical theorems