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 criterion for turning an exact sum-of-squares decomposition of centered Gaussian variation into a joint distributional statement. Let \(U\sim N_n(0,I_n)\), let \(B_1,\ldots,B_k\) be real symmetric matrices, and suppose \(\sum_i B_i=I_n\). Define \(Q_i=U^\mathsf{T}B_iU\) and \(r_i=\operatorname{rank}(B_i)\). In the classical central version, the rank equality \(\sum_i r_i=n\) holds if and only if the \(Q_i\) are jointly independent central chi-square variables with \(r_i\) degrees of freedom each. James's original note explicitly restates this necessary-and-sufficient form of Cochran's 1934 Theorem II.[1][2]
The recurring abstraction is a rank-to-distribution certificate. In the stated setting, adding the component ranks to the ambient dimension checks whether the squared Gaussian energy has been partitioned into disjoint orthogonal subspaces. If it has, each subspace contributes as many independent standard-normal squares as its dimension, and different components are independent. Thus one algebraic condition supplies both degrees of freedom and joint independence, which are essential for normal-theory variance and \(F\) calculations.[2]
The assumptions carry the result. An arbitrary quadratic form is generally a weighted sum of normal squares, not a chi-square. A numerical decomposition without symmetric matrices summing to the identity, or a Gaussian vector with a nonzero projected mean or nonspherical covariance, cannot simply inherit the central conclusion. The theorem's name should therefore denote its conditional formal structure rather than “any sum of squares in statistics.”[2][3]
Structural Signature¶
Sig role-phrases: centered spherical Gaussian carrier → exact symmetric quadratic-form partition → rank-additivity criterion → orthogonal component structure → joint independent chi-square certificate.
- Centered spherical Gaussian carrier. After any known-scale normalization, the coordinates have independent standard-normal distributions. Orthogonal changes of coordinates preserve that joint law; this symmetry is why subspace energies have tractable chi-square distributions. It is constitutive of the central version.[1][2]
- Exact quadratic-form partition. The component matrices are real symmetric and add to \(I_n\), making \(\sum_iQ_i=U^\mathsf{T}U\) as an identity of forms. This is stronger than having a few sums of squares that happen to resemble one another numerically.[2]
- Rank-additivity criterion. With \(r_i=\operatorname{rank}(B_i)\), the equality \(\sum_i r_i=n\) says the component dimensions exactly exhaust the full space. The rank check is meaningful only after the Gaussian carrier and exact partition are established.[2]
- Orthogonal component structure. Under those hypotheses, the rank condition forces the components to behave as projections onto pairwise orthogonal subspaces. In the projection formulation, \(B_i^2=B_i\), \(B_iB_j=0\) for \(i\ne j\), and \(\sum_iB_i=I_n\). One can equivalently rotate the Gaussian vector and partition its independent coordinates.[2]
- Joint distributional certificate. Each \(Q_i\) has a central \(\chi^2_{r_i}\) law, and the forms are mutually independent. The conclusion is joint, not merely a list of marginal distributions. Its practical value is that ratios of suitably scaled nonzero components have the relevant \(F\) law.[2]
The rank equality is not a substitute for checking the preceding roles. It is a compact certificate within their scope.
What It Is Not¶
It is not a distribution law for every symmetric quadratic form. If a matrix has eigenvalues other than zero and one, its centered Gaussian quadratic form generally weighs independent normal squares unequally; it need not be chi-square. Even a positive-semidefinite matrix is not automatically a projection. In the classical theorem, projection structure follows from the complete-partition and rank conditions, or may be made an explicit sufficient starting assumption.[2]
It is not a conclusion from rank counts alone. Two unrelated matrices can have ranks summing to \(n\) while failing to add to the identity. Nor is a decomposition of nonnormal or correlated errors automatically covered. Whitening may sometimes produce a new spherical-Gaussian formulation, but that is an additional justified model change, not part of the bare rank statement.
It is not equivalent to saying any two sums of squares are uncorrelated. Statistical independence is stronger, and the theorem establishes it from disjoint Gaussian projection spaces under its hypotheses. A nonzero projected mean leads instead toward noncentral quadratic-form results, treated in distinct original work by Madow.[3]
Scope of Application¶
The theorem lives in mathematical statistics, especially normal linear-model inference. It is a reusable schema because sample size, number of components, subspaces and design matrix can change while the same Gaussian–partition–rank argument recurs. Cochran's original paper motivated the joint distribution of quadratic forms used in analysis of variance, covariance and regression inference.[1]
For a normal sample, it identifies the one-dimensional mean-axis square and the \((n-1)\)-dimensional residual square. For one-way ANOVA under a common-mean null, the grand mean, between-group contrasts and within-group residuals occupy orthogonal spaces with dimensions $1$, \(g-1\) and \(n-g\). The between and within components then yield a central \(F\) ratio when their positive degrees of freedom exist. These are applications of one theorem, not different definitions of it.[2]
Outside a centered spherical Gaussian model, the exact central chi-square and \(F\) statements need not hold. ANOVA may use approximations or other robustness arguments in practice, but those are separate justifications. Under unequal group means, the between-group projection generally carries a nonzero mean and may be noncentral.[3]
Clarity¶
The theorem disentangles an accounting identity from a probabilistic independence claim. The equality “total sum of squares equals components” only conserves a scalar total. It does not by itself say that observing one component gives no information about another. The rank condition, together with the full Gaussian and matrix assumptions, reveals whether the components are disjoint subspace energies and supplies that stronger claim.[2]
It also clarifies degrees of freedom. A component's rank is the dimension of its contributing subspace, not an arbitrary denominator selected for an \(F\) statistic. In a normal sample, the mean uses one direction and residuals use the remaining \(n-1\); in one-way ANOVA, \(g-1\) independent between-group contrasts remain after removing the grand mean.
Manages Complexity¶
Without the theorem, a proposed decomposition invites separate calculations of each quadratic form's distribution and their joint dependence. Cochran's criterion compresses that joint problem into carrier law, exact partition, matrix ranks. The accompanying projection interpretation explains why this economy works: a rotation turns the decomposition into sums of squares of non-overlapping standard-normal coordinates.[2]
This compression does not erase modeling work. One still has to establish a common error variance, centering under the relevant null, and the actual ranks of the design subspaces. An incorrect model can pass a matrix-rank calculation yet invalidate the probability law assumed behind it.
Abstract Reasoning¶
Given a proposed normal-theory sum-of-squares analysis, the logical order is: first identify a centered spherical Gaussian vector; then express every component as a symmetric quadratic form and verify that the matrices sum to the identity; then determine ranks. If their sum equals the ambient dimension, the theorem licenses the joint independent central chi-square conclusion. If one premise fails, the inference stops; a familiar ANOVA-looking formula does not repair it.[2]
The conclusion then explains why an \(F\) ratio can be formed from two nonzero components under a suitable null: each contributes a chi-square divided by its degrees of freedom, and their independence is part of the theorem. The theorem does not itself decide whether a scientific null hypothesis is true; it gives the sampling law conditional on its stated model.
Knowledge Transfer¶
The same rank-and-projection reasoning transfers literally from normal sample variance to ANOVA and other Gaussian linear-model partitions. What changes is the choice of subspaces—sample mean and residuals in one case, grand mean, group contrasts and within-group residuals in the other. The Gaussian carrier, exact partition and dimension accounting remain the same.[1][2]
A looser analogy can be made to decomposing information or energy into orthogonal channels, but the chi-square certificate does not transfer to arbitrary non-Gaussian data or arbitrary notions of orthogonality. Live Projection captures a broader geometric operation, and live Statistical Independence captures the probabilistic relation; neither alone supplies Cochran's theorem.
Examples¶
Normal sample: mean axis and residual space¶
Let \(X_1,\ldots,X_n\) be independent \(N(\mu,\sigma^2)\) values with \(n\ge2\), and set \(U_i=(X_i-\mu)/\sigma\). The mean-axis projector is \(P_m=\mathbf1\mathbf1^\mathsf{T}/n\) with rank $1$; its complement \(P_r=I_n-P_m\) has rank \(n-1\). The familiar identity is
The two right-hand forms are independent central \(\chi^2_1\) and \(\chi^2_{n-1}\). In particular, \((n-1)S^2/\sigma^2\sim\chi^2_{n-1}\) for the usual \(S^2=\sum_i(X_i-\bar X)^2/(n-1)\). Strictly, the quadratic-form conclusion gives independence of the squared mean displacement and residual square; independence of the signed \(\bar X\) and \(S^2\) follows from the stronger fact that orthogonal Gaussian projected vectors are independent.[2]
Mapped back: carrier = standardized independent normal sample; partition = \(P_m+P_r=I_n\); rank criterion = \(1+(n-1)=n\); orthogonal structure = mean axis perpendicular to zero-sum residual space; certificate = independent chi-square squares with the stated degrees of freedom.
One-way ANOVA: group contrasts and residuals¶
Consider \(n\) independent observations divided among \(g\) nonempty groups, with \(n>g\ge2\), common variance \(\sigma^2\) and a common mean \(\mu\) under the null. Let \(P_G\) project onto vectors constant within each group and \(P_1\) onto the all-ones grand-mean axis. Then \(P_B=P_G-P_1\) and \(P_W=I_n-P_G\) are the between-group and within-group projectors. Their ranks are \(g-1\) and \(n-g\), while \(P_1+P_B+P_W=I_n\).[1][2]
With \(U=(Y-\mu\mathbf1)/\sigma\) under the null, \(U^\mathsf{T}P_BU\sim\chi^2_{g-1}\) and \(U^\mathsf{T}P_WU\sim\chi^2_{n-g}\) independently. Hence their mean-square ratio has \(F_{g-1,n-g}\) distribution. If the group means are unequal, the between-group component is generally noncentral and that central-null statement no longer applies.[3]
Mapped back: carrier = centered, common-variance Gaussian group observations; partition = \(P_1+P_B+P_W=I_n\); rank criterion = \(1+(g-1)+(n-g)=n\); orthogonal structure = grand mean, between contrasts and within residuals; certificate = independent central between/within chi-squares under the common-mean null.
Boundary: a noncentral group contrast¶
Keep the same projection geometry but let true group means differ. The between-group projected mean is then typically nonzero. Orthogonal Gaussian components can still be independent, but the between-group square no longer has the central chi-square law promised by the theorem's centered form. The missing role is the centered Gaussian carrier on that projected subspace.[3]
Structural Tensions¶
T1 — Algebraic economy versus probabilistic assumptions. Rank addition offers a strikingly short route to a joint law, but only after centering, common spherical variance and exact partition have been justified. Leaning on rank alone risks false central inference; declining the theorem despite satisfied assumptions forfeits an exact calculation. Diagnostic: Are the errors truly modeled as centered spherical Gaussian in the coordinates used for the forms?[2]
T2 — Conservation of total squares versus independence of pieces. A total can always be split algebraically in many ways, but only a disjoint orthogonal component structure gives the named independence result. Treating a shared direction twice overstates degrees of freedom; insisting on orthogonal rank accounting exposes it. Diagnostic: Do the component matrices sum to \(I_n\) and do their ranks add to \(n\)?[2]
Structural–Framed Character¶
Evaluative weight: The theorem itself is an exact conditional mathematical statement, though judgments about whether a model is appropriate are evaluative. Human-practice dependence: Humans choose a design and formulate null hypotheses, but the Gaussian and matrix implication is not constituted by those practices. Institutional origin: Cochran's name is historical provenance, not a limit on where the relation can hold.[1][2]
Vocabulary travel: “Rank,” “projection” and “independence” travel across mathematics, while central chi-square quadratic-form laws remain statistical. Import versus recognition: Recognizing a new instance requires reconstructing its Gaussian carrier, sum-to-identity partition and ranks, not importing the conclusion from a familiar ANOVA table. Its character: a structural formal rule within mathematical statistics, with broad reuse across Gaussian designs but not an unqualified prime outside its probability domain.
Structural Core vs. Domain Accent¶
The structural core is certifying a complex decomposition from independent dimensions: a complete partition of a carrier into non-overlapping components allows the components to be analyzed separately. Live Projection captures the geometric act of splitting by subspaces; live Statistical Independence captures the factorization consequence. Whether a further general prime of rank-certified independent decomposition exists is an unadmitted future-prime question, not established merely by this Gaussian result.
The domain accent is decisive: a centered spherical normal vector, symmetric quadratic forms, rank equality, central chi-square distributions and their \(F\)-test use. Remove these and one has an analogy about decomposition, not Cochran's theorem. This is why the named entry remains domain-specific despite a portable-sounding rank-counting slogan.[2]
Instantiates / Related Primes¶
Projection is a derived geometric description of the admissible quadratic-form pieces, not a general theorem genus; statistical independence is a conclusion, not a superclass; generalized chi-square distribution concerns a wider distribution family. An unparented theorem node is more honest than a lexical attachment and can be reconsidered during graph densification.
One-way ANOVA and normal-sample variance are child applications of the theorem's schema, not aliases. The theorem applies to their squared forms only after their specific normal-model and rank conditions are verified.
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
Not to Be Confused With¶
- Any quadratic form of Gaussian variables: arbitrary eigenvalue weights generally do not give a central chi-square.[2]
- An ANOVA sum-of-squares identity alone: an accounting equality is not a proof of independent components.
- Uncorrelatedness: weaker than mutual independence.
- A central law under unequal means: a nonzero projected mean can produce a noncentral form.[3]
- The signed sample mean's independence from sample variance as a direct quadratic-form conclusion: the theorem certifies independence of squares; the full signed-mean result additionally uses independence of orthogonal Gaussian projected vectors.
- A distribution-free or heteroscedastic theorem: changing the carrier requires a new justification.
References¶
[1] 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; publisher extract states its quadratic-form and variance/covariance-analysis scope. registry ↩a ↩b ↩c ↩d ↩e ↩f
[2] G. S. James, “Notes on a theorem of Cochran”, Proceedings of the Cambridge Philosophical Society 48:443–446, 1952. See §1, directly restating Cochran's Theorem II and explaining the orthogonal-transform proof. registry ↩a ↩b ↩c ↩d ↩e ↩f ↩g ↩h ↩i ↩j ↩k ↩l ↩m ↩n ↩o ↩p ↩q ↩r ↩s ↩t ↩u
[3] 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; cited only to delimit the central theorem. registry ↩a ↩b ↩c ↩d ↩e ↩f