This article was accepted into the corpus but its outbound wikilinks were never NER-processed — typical at the deepest BFS hop or when the run's entity cap was reached. No expansion funnel to show.
| Cochran's theorem | |
|---|---|
| Name | Cochran's theorem |
| Field | Statistics |
| Introduced | 1934 |
| Named after | William Gemmell Cochran |
Cochran's theorem is a result in mathematical statistics concerning quadratic forms in normally distributed variables and the decomposition of sums of squares. It gives necessary and sufficient conditions under which the sum of independent quadratic forms is distributed as a sum of independent chi-squared variables and when associated projection matrices have specified ranks. The theorem is fundamental in the development of classical inference for linear models, analysis of variance, and multivariate techniques.
Let X be an n-dimensional random vector with multivariate normal distribution having mean vector μ and covariance matrix σ^2 I_n, where σ^2>0. Let Q_1, Q_2, ..., Q_k be symmetric n×n matrices such that the quadratic forms X^T Q_i X are involved in the decomposition of the total sum of squares. Cochran's theorem states that if the matrices Q_i are idempotent and pairwise orthogonal in the sense that Q_i Q_j = 0 for i ≠ j, and if Σ_i Q_i = I_n, then the quadratic forms X^T Q_i X are independent and each has a chi-squared distribution with degrees of freedom equal to rank(Q_i), scaled by σ^2. Conversely, if the X^T Q_i X are independent chi-squared variables with appropriate degrees of freedom and Σ_i Q_i = I_n, then each Q_i must be idempotent and have the stated ranks. The theorem links matrix algebraic conditions on the Q_i (ranks, idempotency, orthogonality) to probabilistic conclusions about independence and chi-squared distributions.
The proof proceeds by diagonalizing the sum of the Q_i via an orthogonal transformation. One constructs an orthogonal matrix P that simultaneously transforms the idempotent, pairwise-orthogonal family into diagonal projection matrices onto orthogonal coordinate subspaces. Under the multivariate normal assumption with covariance σ^2 I_n, orthogonal transformations preserve independence and distributional form; the transformed vector P^T X is again normal with covariance σ^2 I_n. Each quadratic form X^T Q_i X becomes a sum of squared, independent normal coordinates, yielding a scaled chi-squared distribution with degrees of freedom equal to the trace (or rank) of Q_i. Independence follows from the orthogonality of the ranges of the Q_i, which ensures the corresponding coordinate blocks of P^T X are independent. The converse uses linear-algebraic necessity: independence and chi-squared structure imply that the generating matrices act as orthogonal projections onto independent coordinate subspaces, hence are idempotent and orthogonal, with ranks matching the observed degrees of freedom.
Cochran's theorem underpins the classical theory of the Analysis of variance and the decomposition of the total sum of squares into components associated with factors in models such as the one-way analysis of variance and two-way analysis of variance. It justifies the distributional claims for residual sums of squares and treatment sums of squares that lead to the F-test used in Ronald Fisher's frameworks. In linear regression, Cochran's theorem explains why the residual sum of squares is independent of the least-squares estimator and follows a scaled chi-squared law, a fact exploited in constructing Student's t-test and confidence intervals introduced by William Sealy Gosset. Multivariate versions inform inference in Hotelling's T-squared distribution, Wilks' lambda, and tests developed by Harold Hotelling and Samuel S. Wilks. The theorem also appears in the derivation of likelihood-ratio tests in Jerzy Neyman-Egon Pearson style hypothesis testing and in derivations used by the National Institute of Standards and Technology for uncertainty quantification.
Generalizations relax the normality or orthogonality assumptions. Extensions consider quadratic forms in elliptically contoured distributions and in settings with structured covariance matrices, connecting to results attributed to Andrey Kolmogorov-type decompositions and to matrix factorization techniques used by Carl Friedrich Gauss in least-squares history. Related results include the spectral decomposition theorems of John von Neumann and Marshall Stone, and rank-additivity facts in linear algebra linked to work by Évariste Galois and Arthur Cayley. In multivariate analysis, Cochran-type decompositions relate to canonical correlations studied by Harold Hotelling and to decomposition approaches in the Karhunen–Loève theorem as developed by Norbert Wiener and Harald Cramér. Robust and nonparametric analogues have been proposed in literature influenced by Frank Wilcoxon and Jerzy Neyman's colleagues.
Examples: In the classical balanced one-way ANOVA with a total of n observations across g groups, projection matrices for group means and residuals satisfy the hypotheses of Cochran's theorem; the treatment sum of squares and residual sum of squares are independent chi-squared variables with degrees of freedom g−1 and n−g respectively. In ordinary least squares with full column rank p, the hat matrix and residual projection observe the decomposition Σ = H + (I−H), producing independent quadratic forms with degrees p and n−p.
Counterexamples: If the covariance matrix of X is not a scalar multiple of the identity (for example, heteroscedastic errors associated with Gauss–Markov theorem violations) or if the matrices Q_i fail idempotency or orthogonality (such as overlapping projection matrices), the conclusions fail: quadratic forms need not be independent nor chi-squared. Another counterexample arises when error distributions are heavy-tailed (e.g., those studied by Paul Lévy), where sums of squares do not follow chi-squared laws and independence properties break down.
Category:Theorems in statistics