Generalized least squares¶

A linear-model estimator that minimizes residuals in the inverse-covariance metric when errors have known nonconstant variance or correlation.

Version: v2 · 2026-09-19 · History
Domain-specific #: 4697
Domain group: Formal Sciences
Origin domain: Experimental Design & Statistics
Subdomain: Regression and Econometrics → Experimental Design & Statistics

Core Idea¶

Whitening converts GLS to ordinary least squares, giving the best linear unbiased estimator under the generalized Gauss–Markov assumptions; feasible GLS estimates the covariance and adds model-risk and finite-sample uncertainty. A factor of the error covariance transforms responses and design so residuals become spherical, ordinary least squares is applied in transformed coordinates, and the result maps back to the original coefficients. The abstraction is therefore identified by a declared carrier, a transformation or constraint over that carrier, and an invariant that tells an analyst whether the named structure is genuinely present.

How would you explain it like I'm…

Fix the Dots, Then Draw

Imagine drawing a straight line through dots, but some dots were measured with shaky hands and some dots copy each other. Generalized least squares first fixes the dots so they're all equally trustworthy and not copying. Then it draws the best line, and turns the answer back into the original picture.

Fair-Error Best Fit

Ordinary least squares fits a line by treating every data point's mistake as equally sized and unconnected to the others. Real data often isn't like that: some points are noisier, and some mistakes are linked together. Generalized least squares uses a description of how the mistakes behave to reshape the data so the mistakes become even and unlinked. Then the usual line-fitting works, and the answer is changed back to fit the original data. If you don't know how the mistakes behave, you have to estimate it first, which adds extra uncertainty.

Whitened Least Squares Regression

Generalized least squares (GLS) is a regression method for when errors have unequal variances or are correlated, described by an error covariance matrix. Ordinary least squares assumes errors are 'spherical': equal variance and uncorrelated. GLS uses a factor of the covariance matrix to transform both the responses and the design matrix so the transformed errors are spherical. It then runs ordinary least squares in these 'whitened' coordinates, and the coefficients apply to the original model. Under the right assumptions this gives the best linear unbiased estimator. When the covariance must be estimated from data, the method is called feasible GLS, and it carries extra risk if the covariance model is wrong.

Generalized least squares estimates the coefficients of a linear mean model y = Xβ + ε when Var(ε) = Σ is a positive-definite but non-spherical covariance. Choose a whitening factor W with WΣWᵀ = I; transforming y and X by W makes the errors spherical, and ordinary least squares on the transformed data minimizes the GLS objective (y − Xβ)ᵀΣ⁻¹(y − Xβ). Under the generalized Gauss–Markov assumptions (full rank, exogeneity, known Σ) this gives the best linear unbiased estimator. In practice Σ is usually unknown, so feasible GLS estimates it in a first stage and then applies the GLS formula; this adds model-misspecification risk and finite-sample uncertainty beyond the textbook sampling distribution. A common alternative is OLS with robust standard errors. Good practice states whether Σ is known or estimated, the whitening factor, the assumptions, and diagnostic checks of the fitted covariance.

Scope of Application¶

Generalized least squares belongs to regression and econometrics and is useful where the analyst can specify the typed regression and econometrics carrier, defining objects and relations, parameters, conventions, evidence, boundary cases, and comparison targets, then evaluate the response and design matrix, coefficient vector, linear mean model, error covariance and positive-definiteness, known versus estimated covariance, whitening factor, GLS objective and estimator, rank and exogeneity assumptions, sampling distribution, robust alternative, feasible-GLS stages and diagnostic validation are explicit.

Clarity¶

The abstraction clarifies a crowded vocabulary by making the response and design matrix, coefficient vector, linear mean model, error covariance and positive-definiteness, known versus estimated covariance, whitening factor, GLS objective and estimator, rank and exogeneity assumptions, sampling distribution, robust alternative, feasible-GLS stages and diagnostic validation are explicit the center of the account. A claim should name the carrier, the governing operation or relation, the applicable assumptions, and the recognition test.

Manages Complexity¶

Without the abstraction, an analyst must reason directly over many local details: the carrier roles, admissibility assumptions, competing conventions, derived invariants, boundary cases, and proof or validation obligations specific to Generalized least squares. Generalized least squares compresses them into the roles in the structural signature. That compression permits comparison across instances without erasing the variables that determine validity. It also exposes which details may be varied safely and which are constitutive.

Abstract Reasoning¶

Identify the carrier. State what the elements, states, objects, or observations are: the typed regression and econometrics carrier, defining objects and relations, parameters, conventions, evidence, boundary cases, and comparison targets. Reject examples whose alleged carrier belongs to a different problem. 2.

Knowledge Transfer¶

Knowledge transfers strongly among subfields of regression and econometrics because they reuse the typed regression and econometrics carrier, defining objects and relations, parameters, conventions, evidence, boundary cases, and comparison targets, A factor of the error covariance transforms responses and design so residuals become spherical, ordinary least squares is applied in transformed coordinates, and the result maps back to the original coefficients., and type the carrier, state every parameter and convention in the definition, test that the response and design matrix, coefficient vector, linear mean model, error covariance and positive-definiteness, known versus estimated covariance, whitening factor, GLS objective and estimator, rank and exogeneity assumptions, sampling distribution, robust alternative, feasible-GLS stages and diagnostic validation are explicit, compare the nearest accepted identity, and report counterexamples, uncertainty, and limiting cases.

Relationships to Other Abstractions¶

Current abstraction Generalized least squares Domain-specific

Parents (1) — outgoing typed relations

Generalized least squares is a kind of Estimation Prime

It remains its own entry because its identity is fixed by the response and design matrix, coefficient vector, linear mean model, error covariance and positive-definiteness, known versus estimated covariance, whitening factor, GLS objective and estimator, rank and exogeneity assumptions, sampling distribution, robust alternative, feasible-GLS stages and diagnostic validation.

Hierarchy path (1) — routes to 1 parentless root

Generalized least squares → is a kind of Estimation → is a kind of Approximation → is a decomposition of Representation → presupposes Abstraction

Neighborhood in Abstraction Space¶

Generalized least squares sits in a crowded region of the domain-specific corpus (23^rd percentile for distinctiveness): several abstractions share nearly its structure, so a description that fits it tends to fit its neighbors too.

Family — Regression Diagnostics & Linear Models (11 abstractions)

Nearest neighbors

Computed from structural-signature embeddings · 2026-10-25