= Orthogonal decomposition in a positive-definite metric
A real <positive-definite symmetric matrix> $V$ defines an <inner product> $\langle x,y\rangle_V=x^TVy$ and <norm> $\|x\|_V=\sqrt{x^TVx}$. For a nonzero direction $a$, every vector has the unique decomposition
$$
x=\alpha a+z,\qquad \alpha=\frac{a^TVx}{a^TVa},\qquad a^TVz=0.
$$
Substitution proves the <orthogonality> and gives the corresponding <Pythagorean identity> $\|x\|_V^2=\alpha^2\|a\|_V^2+\|z\|_V^2$. Thus $\alpha a$ is the closest vector to $x$ on the line spanned by $a$, in this metric.
For $V=\operatorname{diag}(1,4)$, $a=(1,1)^T$ and $x=(1,0)^T$, one gets $\alpha=1/5$, unlike the Euclidean projection coefficient $1/2$. The geometry weights coordinate errors differently. Under the invertible change of variables $y=V^{1/2}x$, it becomes ordinary Euclidean orthogonal decomposition. This is useful for separating a constrained direction from free directions in <quadratic optimization>.
Back to article page