Let $X$ separable Hilbert space and $B \in \mathcal{L}(X)$ (i.e. a bounded linear operator) be strictly positive and self-adjoint. Then I can define the fractional power $B^{1/2}$ and if I set $$|x|_{-1}=| B^{1/2}x|$$ then $|\cdot|_{-1}$ is a norm on $X$.
Now $(X,|\cdot|_{-1})$ is not a Hilbert space because it is not complete but I can define $X_{-1}$ to be the completion of $X$ under $|\cdot|_{-1}$, i.e. see [Fabbri G, Gozzi F, Swiech A. Stochastic optimal control in infinite dimension. Probability and Stochastic Modelling. Springer. 2017.] p.172.
Now my question is: does $X_{-1}$ admit an orthonormal basis made of elements of $X$ as it is claimed in [Fabbri G, Gozzi F, Swiech A. Stochastic optimal control in infinite dimension. Probability and Stochastic Modelling. Springer. 2017.] p.189?
It could be useful to note that $B^{1/2}$ can be extended to $B^{1/2} \in \mathcal{L}(X_{-1},X)$ and it is an isometry, i.e. for $x \in X_{-1}$ we have: $$|B^{1/2}x|=|x|_{-1}$$ Morally I would like do something like this: let $\{ e_k\}$ be orthoormal basis for $X$. Then for $x \in X_{-1}$ there exist $x_n=\sum_{k=1}^{\infty}x_{nk}e_k$ in $X$ such that $|x_n-x|_{-1}\to 0$, i.e. \begin{align} x=\lim_{n}\sum_{k=1}^{\infty}x_{nk}e_k =\sum_{k=1}^{\infty} \lim_{n} x_{nk}e_k=\sum_{k=1}^{\infty} x_{\infty k}e_k \end{align} where the limit is in the $X_{-1}$-topology. This of course can't be done.
Any suggestion?