= Gram-Schmidt orthogonalization for a symmetric bilinear form
{c}
For a <symmetric bilinear form> $B$, the <Gram-Schmidt process> projection formula becomes
$$
u_k=v_k-\sum_{j<k}\frac{B(v_k,u_j)}{B(u_j,u_j)}u_j.
$$
It produces <orthogonal vectors> whenever every denominator used is nonzero. A <symmetric bilinear form> can have nonzero <vectors> with zero self-pairing, so <linear independence> alone does not justify those divisions. For an <inner product>, <positive definiteness> excludes this obstruction.
For example, on $\mathbb R^2$ let $B((x_1,x_2),(y_1,y_2))=x_1y_2+x_2y_1$. Both <standard basis vectors> have zero self-pairing, so reordering them cannot provide the first pivot. Replacing them by $u=(1,1)$ and $v=(1,-1)$ works: $B(u,u)=2$, $B(v,v)=-2$ and $B(u,v)=0$. Changing <linear combinations> can therefore be necessary, rather than merely reordering the input. This example also separates the existence of an <orthogonal basis> from the success of an unmodified pivot order.
Back to article page