The Schmidt decomposition theorem says:

If $H_1,H_2$ are Hilbert-spaces (for simplicity: of same dimension) and $x\in H_1\otimes H_2$, then there exist orthonormal bases $\alpha_i,\beta_i$ of $H_1,H_2$, and reals $\lambda_i\geq0$ such that $x=\sum_i \lambda_i\alpha_i\otimes\beta_i$.

I know that this holds for separable Hilbert spaces (e.g., Breuer, Petruccione, The theory of open quantum systems).

Does this also hold for nonseparable Hilbert spaces? (If possible, please provide a reference to a paper or textbook.)

  • 7
    $\begingroup$ If $x=\sum_n c_n v_n\otimes w_n$, then you can just consider $x$ as a vector from $\overline{L(v_n)}\otimes\overline{L(w_n)}$, and you're back in the separable case. $\endgroup$ Commented May 29, 2017 at 18:51

1 Answer 1


Yes. Christian Remling's comment in essence already gave an affirmative answer, and this question is a bit old, but I think it's worthwhile to write a proof that does not require reducing to the separable situation.

I encountered the same question myself and have not found any reference---but with the polar decomposition and spectral theorems, the result follows with little fuss; this leads me to be believe it is almost certainly known, if not written down somewhere on some ancient scroll of functional analysis. (But the fact that I haven't seen it written down in general form compels me to put it out into the internet ether.)

I take my Hermitian inner products to be antilinear in the first entry and linear in the second. If $\mathcal{H}$ is a Hilbert space, the continuous dual of a Hilbert space $\mathcal{H}$ will be denoted $\mathcal{H}^{\vee}$, the complex conjugate space will be denoted $\overline{\mathcal{H}}$. Every linear map $f: \mathcal{H} \rightarrow \mathcal{K}$ is functorially mapped to a linear map $\overline{f}: \overline{\mathcal{H}} \rightarrow \overline{\mathcal{K}}$. If you don't like conjugate spaces, just decorate the correct terms with ``anti".

First note that there is an isometric isomorphism of Hilbert spaces $I: \mathcal{H}_{1} \otimes \mathcal{H}_{2} \rightarrow \mathcal{H}_{1} \otimes \overline{\mathcal{H}_{2}}^{\vee}$ given by linearization and continuous extension of the map $\alpha \otimes \beta \mapsto \alpha \otimes \langle -, \beta \rangle_{2}$. The Hilbert space $\mathcal{H}_{1} \otimes \overline{\mathcal{H}_{2}}^{\vee}$ is canonically identifiable with the space of Hilbert-Schmidt maps from $\overline{\mathcal{H}}_{2}$ to $\mathcal{H}_{1}$; in particular, it is canonically identifiable with a subspace of the space of bounded operators from $\overline{\mathcal{H}}_{2}$ to $\mathcal{H}_{1}$.

Define $X := I(x)$. By the polar decomposition theorem

\begin{align*} X = U \sqrt{X^{*} X} \end{align*}

for a partial isometry $U: \overline{\mathcal{H}}_{2} \rightarrow \mathcal{H}_{1}$. The operator $X^{*} X: \overline{\mathcal{H}}_{2} \rightarrow \overline{\mathcal{H}}_{2}$ is positive semidefinite and trace class (with trace $\|x\|^2$), so, it is in particular self-adjoint and compact. By self-adjoint compact operator spectral theory we have that the non-vanishing spectrum $\sigma_{\neq 0}$ of $X^{*} X$ is countable, consists only of eigenvalues, and moreover the eigenspaces $E_{\lambda}$ associated to each $\lambda \in \sigma_{\neq 0}$ are finite dimensional and orthogonal. Positive semidefiniteness ensures that each eigenvalue is positive. By the spectral theorem

\begin{align*} X^{*} X = \sum_{\lambda \in \sigma_{\neq 0}} \lambda \mathrm{Proj}_{\lambda}, \end{align*}

where $\mathrm{Proj}_{\lambda}$ is the orthogonal projection onto $E_{\lambda}$. Choosing an orthonormal basis $(\beta^{\lambda}_{i})_{i = 1}^{\dim E_{\lambda}}$ for each finite dimensional $E_{\lambda}$

\begin{align*} X^{*} X &= \sum_{\lambda \in \sigma_{\neq 0}} \lambda \left[\sum_{i_{\lambda} = 1}^{\dim E_{\lambda}} \beta^{\lambda}_{i_{\lambda}} \otimes (\beta^{\lambda}_{i_{\lambda}})^{\vee} \right], \end{align*}

(where $\omega^{\vee}$ denotes the linear form $\langle \omega, - \rangle$). Of course I can just reindex things here to get a sum

\begin{align*} X^{*} X=\sum_{j = 1}^{r} \lambda_{j} \beta_{j} \otimes \beta_{j}^{\vee}, \end{align*}

where $r = \mathrm{rank}(X^{*} X) = \sum_{\lambda \in \sigma_{\neq 0}} \dim E_{\lambda} \leq \aleph_{0}$, the $\lambda_{j}$ are eigenvalues, and the $\beta_{j} \in \overline{\mathcal{H}}_{2}$ are associated eigenvectors with $\langle \beta_{i}, \beta_{j} \rangle = \delta_{ij}$. Using our polar decomposition:

\begin{align*} X = \sum_{j = 1}^{r} \sqrt{\lambda_{j}} (U \beta_{j}) \otimes \beta_{j}^{\vee}. \end{align*}

Because $U$ is a partial isometry, and one can verify $\ker(U) \leq \ker(X^{*} X)^{\perp}$, the $U \beta_{j} \in \mathcal{H}_{1}$ remain orthonormal. Applying $I^{-1}$, then we simply have \begin{align*} x = \sum_{j = 1}^{r} \sqrt{\lambda_{j}} (U \beta_{j}) \otimes \tilde{\beta}_{j}. \end{align*} where $\tilde{\beta}_{j} \in \mathcal{H}_{2}$ is the image of $\beta^{\vee}_{j}$ under the canonical isomorphism $\overline{\mathcal{H}}^{\vee}_{2} \overset{\sim}{\rightarrow} \mathcal{H}_{2}$. It is an eigenvector of $\overline{X^{*}X}: \mathcal{H}_{2} \rightarrow \mathcal{H}_{2}$ with eigenvalue $\lambda_{j}$.

A few side remarks:

  • $U \beta_{i}$ is an eigenvector of $X X^{*}: \mathcal{H}_{1} \rightarrow \mathcal{H}_{1}$ associated to eigenvalue $\lambda_{i}$. This follows by looking at the explicit construction of $U$ as in, e.g. Douglas' Lemma (a generalization of the polar decomposition theorem). I mention this because $\overline{X^{*} X}$ and $X X^{*}$ are the reduced density states associated to the pure state $x \otimes x^{\vee}$. The proof above is just a generalization of the usual finite-dimensional proofs that the Schmidt decomposition is given by tensor products of eigenstates of reduced density states.
  • Because Douglas' lemma constructs a unique $U$, the only choices involved in the procedure above are the choices of bases on each finite dimensional eigenspace $E_{\lambda}$. Any two choices are then related by the action of $\prod_{\lambda \in \sigma_{\neq 0}} U(E_{\lambda})$ (the infinite product of underlying groups). One can check that this group can be realized as a subgroup of $U(\bigoplus_{\lambda \in \sigma_{\neq 0}} E_{\lambda})$, and hence $U(\mathcal{H}_{2})$, or even $U(\mathcal{H}_{1} \otimes \mathcal{H}_{2})$ in a natural way.
  • If one wants to be overly pedantic about the question you asked: you mentioned something about bases. Indeed, every Hilbert space has an orthornormal basis, so there exists a (possibly uncountable) basis $(e_{k})_{k \in K}$ for $\mathrm{span}(\beta_{i})^{\perp} \leq \mathcal{H}_{2}$ (I'm taking all spans to be closed) and a basis $(f_{l})_{l \in L}$ for $\mathrm{span}(U\beta_{i})^{\perp} \leq \mathcal{H}_{1}$; we can combine these with the bases we constructed on the spans to get orthonormal bases $(\tilde{e}_{k})_{k \in K'}$ and $(\tilde{f}_{l})_{l \in L'}$. Then, in a rather stupid way, we have \begin{align*} x = \sum_{m \in M} \omega_{m} \tilde{e}_{m} \otimes \tilde{f}_{m} \end{align*} where $M$ is some (possibly uncountable) indexing set with $|M| = \mathrm{min}(|K'|,|L'|)$. Only countably many of the $\omega_{m}$ are non-zero.

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.