Mean-zero L2 space
= Mean-zero L2 space
{title2=$L^2_0(P)=\{h\in L^2(P):Ph=0\}$}
For a <probability measure> $P$, the mean-zero L2 space is
$$
L^2_0(P)=\{h\in L^2(P):\int h\,dP=0\}.
$$
The expectation functional is bounded: <Cauchy-Schwarz inequality> gives $|\int h\,dP|\leq\|h\|_2$ because $P(\Omega)=1$. Its kernel is consequently a <closed subspace of a Hilbert space>. The inherited <inner product> is
$$
\langle h,k\rangle=\int h\overline k\,dP.
$$
For real functions the conjugate can be omitted; for complex functions the product $\int hk\,dP$ alone is not an inner product.
Every square-integrable function has the orthogonal decomposition $h=(h-Ph)+Ph$, with the first term in $L^2_0(P)$ and the second constant. The squared norm of the centered term is the <variance> in the real-valued case. This is the natural space for centered <score functions> and <canonical gradients>.