Ex. 18.7

Ex. 18.7

Consider a linear regression problem where pN, and assume the rank of X is N. Let the SVD of X=UDVT=RVT,where R is N×N nonsingular, and V is p×N with orthonormal columns.

(a) Show that there are infinitely many least-squares solutions all with zero residuals.

(b) Show that the ridge-regression estimate for β can be written

β^λ=V(RTR+λI)1RTy

(c) Show that when λ=0, the solution β^0=VD1UTy has residuals all equal to zero, and is unique in that it has the smallest Euclidean norm amongst all zero-residual solutions.

Soln. 18.7

(a) Since XRp×N has rank Np, we know there exists α0 such that Xα=0. If β^0 has zero residuals, so does β^0+kα for any kR. Therefore there are infinitely many least-squares solutions all with zero residuals.

(b) This is the same as Ex. 18.4.

(c) Note that

Xβ^0=UDVTVD1UTy=y,

so β^0 has zero residual.

Suppose that β^0+β has zero residual for some β0, that is, X(β^0+β)=y. Since β^0 has zero residual, we know

Xβ=RVTβ=0.

Note that R is N×N nonsingular, so we have VTβ=0. Now consider the Euclidean norm of β^0+β, we have

(β^0+β)T(β^0+β)=β^0Tβ^0+βTβ+2β^0Tβ=β^0Tβ^0+βTβ+2yTUD1VTβ=β^0Tβ^0+βTβ+0.

Since βTβ>0, we know that β^0 has the smallest Euclidean norm.