A topic in the Open Knowledge Graph — a free, open map of 15,290 topics and the order to learn them in.
Linear regression fits a model y = Xβ + ε to data by minimizing ||y − Xβ||². The optimal coefficients are β* = (XᵀX)⁻¹Xᵀy (normal equations), found via orthogonal projection of y onto the column space of X. Residuals r = y − Xβ* are orthogonal to the fitted subspace. QR decomposition is preferred numerically over normal equations for stability.
Linear regression, viewed through linear algebra, is a problem about projection. You have a vector of observations y in ℝⁿ and a matrix X whose columns span a subspace — the column space of X. No vector β will generally make Xβ equal y exactly (the system is overdetermined: more equations than unknowns), so instead you find the β that makes Xβ as close to y as possible. "Closest" means minimizing the Euclidean distance ||y − Xβ||², which is the sum of squared residuals.
The key geometric insight is that the minimizer is the orthogonal projection of y onto col(X). From your work on orthogonal projections, you know that the closest point in a subspace to a given vector is the foot of the perpendicular from that vector to the subspace. The fitted values ŷ = Xβ* are exactly that foot, and the residual vector r = y − ŷ is perpendicular to col(X). Algebraically, this perpendicularity condition is Xᵀr = 0, which expands to Xᵀ(y − Xβ*) = 0 and rearranges to the normal equations: (XᵀX)β* = Xᵀy. If XᵀX is invertible, β* = (XᵀX)⁻¹Xᵀy.
The matrix P = X(XᵀX)⁻¹Xᵀ is called the hat matrix (because ŷ = Py "puts a hat on y"). It is an orthogonal projection matrix: P² = P and Pᵀ = P. The complementary matrix I − P projects onto the orthogonal complement of col(X), and (I − P)y = r is the residual. This decomposition y = ŷ + r = Py + (I − P)y partitions the response into explained variation and unexplained noise — exactly what R² measures.
In practice, computing β* by inverting XᵀX explicitly is avoided. Forming XᵀX squares the condition number of X, making the calculation fragile when predictors are nearly collinear. The preferred approach is QR decomposition: write X = QR where Q has orthonormal columns and R is upper triangular. Then β* = R⁻¹Qᵀy, which is numerically stable and requires only back-substitution rather than a matrix inverse. This is what virtually all statistical software computes under the hood when you call a regression function.
One thing to notice: the normal equations β* = (XᵀX)⁻¹Xᵀy look like ordinary division scaled up to matrices — you are "dividing y by X" in a matrix sense. The factor (XᵀX)⁻¹Xᵀ is the Moore-Penrose pseudoinverse of X when X has full column rank, and generalizes to all X via SVD. Linear regression is thus not a statistical trick but a straightforward application of projection geometry — the statistics (error distributions, inference, R²) are layered on top of this geometric foundation.