F-Test for Linear Constraints
Theorem
Let \(R \in \mathbb R^{q \times p}\) be a constraint matrix, with \(q \leq p\). If \(\text{rank}(X) = p\) (with \(p < n\)), \(\text{rank}(R) = q\), and \(\varepsilon \sim N(0, \sigma^2 I_n)\), then under \(H_0: R\beta = 0\):
\[F = \frac{n-p}{q} \cdot \frac{SSR_c - SSR}{SSR} \sim F(q, n-p)\]
where:
- \(SSR = \|Y - X\hat\beta\|^2\): sum of squares of residuals in the unconstrained model
- \(SSR_c\): sum of squares of residuals in the constrained model, i.e., in the sub-model satisfying \(R\beta = 0\)
Proof
Define the constrained subspace
- \(V_0 = \{X\beta : \beta \in \mathbb R^p, \ R\beta = 0\} = X(\text{Ker}(R))\)
Dimension of \(V_0\). By the rank–nullity theorem applied to \(R\), \(\dim(\text{Ker}(R)) = p - \text{rank}(R) = p - q\).
Since \(\text{rank}(X) = p\), the map \(\beta \mapsto X\beta\) is injective. Therefore: \(\dim(V_0) = \dim(X(\text{Ker}(R))) = \dim(\text{Ker}(R)) = p - q\).
The two \(SSR\)s are squared norms of projections. Fitting a linear model by least squares amounts to projecting \(Y\) orthogonally onto the corresponding subspace, so
- \(SSR = \|Y - P_{[X]}Y\|^2 = \|P_{[X]^\perp} Y\|^2\)
- \(SSR_c = \|Y - P_{V_0}Y\|^2 = \|P_{V_0^\perp}Y\|^2\)
Let us now decompose the constrained residual vector as
\[ P_{V_0^\perp}Y = P_{[X]^\perp} Y + (P_{V_0^\perp} - P_{[X]^\perp})Y \]
The first term is the vector of residuals of the unconstrained model; let us analyse the second term. For this we use the following useful Lemma.
Here, \(V_0 \subset [X]\) so that \([X]^\perp \subset V_0^{\perp}\), and the Lemma gives \(P_{[X]^\perp}P_{V_0^\perp} = P_{V_0^\perp}P_{[X]^\perp} = P_{[X]^\perp}\). Hence
\[(P_{V_0^\perp} - P_{[X]^\perp})^2 = P_{V_0^\perp} - P_{[X]^\perp} - P_{[X]^\perp} + P_{[X]^\perp} = P_{V_0^\perp} - P_{[X]^\perp} = (P_{V_0^\perp} - P_{[X]^\perp})^T\]
so that \(P_{V_0^\perp} - P_{[X]^\perp}\) is an orthogonal projector (it projects onto \([X] \cap V_0^\perp\)) on a subspace of dimension \(\text{tr}(P_{V_0^\perp}) - \text{tr}(P_{[X]^\perp}) = n-(p-q)-(n-p) = q\).
The same identity gives \(P_{[X]^\perp}(P_{V_0^\perp} - P_{[X]^\perp}) = P_{[X]^\perp} - P_{[X]^\perp} = 0\), so the two projectors in the decomposition project onto orthogonal subspaces. By Pythagoras, \(\|P_{V_0^\perp}Y\|^2 = \|P_{[X]^\perp}Y\|^2 + \|(P_{V_0^\perp} - P_{[X]^\perp})Y\|^2\), that is
\[ SSR_c - SSR = \|(P_{V_0^\perp} - P_{[X]^\perp})Y\|^2 \geq 0 \enspace . \]
Distributions. Since \(P_{[X]^\perp}X = 0\), we always have \(P_{[X]^\perp}Y = P_{[X]^\perp}\varepsilon\). Under \(H_0\), \(X\beta \in V_0\), so \(P_{V_0^\perp}X\beta = 0\) and also \((P_{V_0^\perp} - P_{[X]^\perp})Y = (P_{V_0^\perp} - P_{[X]^\perp})\varepsilon\).
By the Cochran Theorem applied to \(\varepsilon \sim \mathcal N(0, \sigma^2I_n)\) and to the two orthogonal projectors \(P_{[X]^\perp}\) and \(P_{V_0^\perp} - P_{[X]^\perp}\), whose product is \(0\):
- \(P_{[X]^\perp} \varepsilon\) and \((P_{V_0^\perp} - P_{[X]^\perp})\varepsilon\) are independent — not merely uncorrelated: this step uses the Gaussianity of \(\varepsilon\);
- \(SSR = \|P_{[X]^\perp} \varepsilon\|^2\) satisfies \(SSR/\sigma^2 \sim \chi^2(n-p)\) (projector of rank \(n-p\));
- Under \(H_0\), \(SSR_c - SSR = \|(P_{V_0^\perp} - P_{[X]^\perp}) \varepsilon\|^2\) satisfies \((SSR_c - SSR)/\sigma^2 \sim \chi^2(q)\) (projector of rank \(q\));
- \(SSR\) and \(SSR_c - SSR\) are independent, as functions of the two independent vectors above.
Conclusion. Under \(H_0\),
\[F = \frac{n-p}{q} \cdot \frac{SSR_c - SSR}{SSR} = \frac{\big[(SSR_c - SSR)/\sigma^2\big]/q}{\big[SSR/\sigma^2\big]/(n-p)}\]
is the ratio of two independent \(\chi^2\) random variables, each divided by its number of degrees of freedom: by definition, \(F \sim F(q, n-p)\). \(\square\)