Cochran's Theorem¶
A rank criterion that identifies when a complete quadratic-form partition of centered spherical Gaussian variation yields mutually independent chi-square components.
Core Idea¶
Cochran's theorem is a reusable criterion for a complete sum-of-squares decomposition of centered Gaussian variation. For \(U\sim N_n(0,I_n)\) and real symmetric matrices \(B_i\) satisfying \(\sum_iB_i=I_n\), set \(Q_i=U^\mathsf{T}B_iU\) and \(r_i=\operatorname{rank}(B_i)\). The classical central result states that \(\sum_i r_i=n\) if and only if the \(Q_i\) are jointly independent central chi-square variables with \(r_i\) degrees of freedom. The rank condition forces an orthogonal partition of Gaussian coordinate directions under these full hypotheses.[ref-98f6d3248833][ref-dd04a6ed4b0e]
It is a theorem schema, not merely one named ANOVA calculation. The dimension, number of pieces and projection spaces vary between applications while the Gaussian–partition–rank relation stays fixed.
Scope of Application¶
In an independent normal sample, the one-dimensional mean axis and \((n-1)\)-dimensional residual space split standardized squared deviations from the true mean. Their squared lengths are independent \(\chi^2_1\) and \(\chi^2_{n-1}\) forms. Independence of the signed sample mean from sample variance follows additionally from independence of the orthogonal Gaussian projected vectors, not solely from the theorem's quadratic-form wording.[^ref-dd04a6ed4b0e]
In one-way ANOVA under the common-mean, common-variance Gaussian null, between-group contrasts have rank \(g-1\) and within-group residuals rank \(n-g\). Along with the grand-mean rank $1$, they form a full orthogonal partition and support a central \(F_{g-1,n-g}\) ratio when \(n>g\ge2\). Outside the null, the between-group form can be noncentral.[ref-98f6d3248833][ref-dd04a6ed4b0e][^ref-e1af04ef3057]
Clarity¶
The theorem distinguishes an accounting identity from a distributional certificate. Writing “total squares = component squares” does not prove independence. One must identify the Gaussian carrier, symmetric matrices that sum exactly to the identity, and ranks exhausting the dimension. A generic positive-semidefinite quadratic form need not be a projection or chi-square.[^ref-dd04a6ed4b0e]
Manages Complexity¶
Instead of separately deriving every marginal law and joint dependence, the theorem reduces the problem to a small set of matrix and distribution checks. Orthogonal coordinates explain the economy: disjoint subspaces collect non-overlapping standard-normal squares, with dimension becoming degrees of freedom. A wrong normal or variance assumption invalidates the shortcut even if rank arithmetic is correct.[^ref-dd04a6ed4b0e]
Abstract Reasoning¶
The inference runs from centered spherical Gaussian input and an exact identity decomposition to a rank test, then to mutually independent central chi-square components. That joint conclusion licenses ratios of scaled components as \(F\) variables under the appropriate null. The theorem does not grant the same conclusion to arbitrary, correlated, heteroscedastic or nonnormal data, nor to a nonzero projected mean.[ref-dd04a6ed4b0e][ref-e1af04ef3057]
Knowledge Transfer¶
The same structural criterion transfers literally from normal-sample variance to ANOVA and other Gaussian linear-model partitions, but the actual subspaces and degrees of freedom must be reconstructed each time. Live Projection and Statistical Independence are related ingredients or consequences, not established taxonomic parents of this theorem; the draft remains an explicitly unparented workspace node pending independent DAG review.
[^ref-98f6d3248833]: W. G. Cochran, “The distribution of quadratic forms in a normal system, with applications to the analysis of covariance”, Proceedings of the Cambridge Philosophical Society 30:178–191, 1934, original theorem paper. [^ref-dd04a6ed4b0e]: G. S. James, “Notes on a theorem of Cochran”, Proceedings of the Cambridge Philosophical Society 48:443–446, 1952, §1 theorem statement and orthogonal-transform rationale. [^ref-e1af04ef3057]: W. G. Madow, “The Distribution of Quadratic Forms in Non-Central Normal Random Variables”, Annals of Mathematical Statistics 11:100–103, 1940, original noncentral-normal extension.
Neighborhood in Abstraction Space¶
Cochran's Theorem sits in a sparse region of the domain-specific corpus (71st percentile for distinctiveness): few abstractions share its structure, so a faithful description tends to retrieve it precisely.
Family — Multivariate & Spectral Signal Analysis (10 abstractions)
Nearest neighbors
- Covariance Matrix — 0.84
- Schur decomposition — 0.84
- Exponentially Modified Gaussian Distribution — 0.84
- Fredholm Kernel — 0.83
- Complex normal distribution — 0.83
Computed from structural-signature embeddings · 2026-10-08