A well-known result is that every two-mode covariance matrix

$$ \sigma_{AB} = \pmatrix{\alpha & \gamma \\ \gamma^T & \beta} $$

can be transformed via a symplectic transformation into a standard form with all diagonal 2 × 2 subblocks, $\alpha = diag(a,a)$, $\beta = diag(b,b)$, $\gamma = diag(c,d)$, where $a,b\geq1$, $c\geq|d|\geq0$. Four symplectic invariants can be identified, namely $\det(\alpha )$, $\det(\beta )$, $\det(\gamma )$, and $\det(\sigma_{AB})$.

The transformation can be done in steps by first performing symplectic diagonalization on the blocks $\alpha $ and $\beta$ (Williamson's theorem), leaving the correlations blocks still non-diagonal. The tranformation to standard form is completed using symplectic singular value decomposition of the correlations block; by now, the diagonal blocks are proportional to identity matrix and will not be affected by the orthogonal matrices diagonalizing $\gamma$ (see, e.g., section 2.3 in this reference).

However, aren't the singular values always positive, making $c,d\geq 0$? How can $\det(\gamma )$ be left invariant if we consider initially a $\sigma_{AB}$ with $\det(\gamma )<0$?

I must have misunderstood something about the transformation but I can't figure out what.

  • $\begingroup$ I do not see what kind of transformation would allow you to change the sign of $\det(γ)$? To take a physics example, suppose you have $γ=\text{diag}(c,-c)$: To converting it to $γ=\text{diag}(c,c)$, you would need an operation which convert $P_A↦-P_A$, while keeping $Q_A$, $Q_B$ and $P_B$ constant. Such an operation is neither physical, nor symplectic. $\endgroup$ Jul 24 '17 at 15:41
  • $\begingroup$ Isn't that transformation exactly the singular value decomposition? $\endgroup$
    – Kiro
    Jul 25 '17 at 6:20
  • $\begingroup$ No! Otherwise symplectic diagonalization would be named singular value decomposition. SVD uses orthogonal matrices (which preserve the identity), while symplectic diagonalization uses symplectic matrices (which preserve $Ω=\begin{bmatrix} 0,& 1\\-1,&0\end{bmatrix}^{⊗n}$. A squeezing operation is symplectic but not orthogonal. A rotation exchanging $Q_A$ and $Q_B$ without changing $P_A$ and $P_B$ is orthogonal, but not symplectic $\endgroup$ Jul 31 '17 at 15:03
  • $\begingroup$ Got it. How would I go about finding the diagonal values of $\gamma$ then? I understand that Williamson's theorem applies to positive definite matrices, which $\gamma$ in general will not be, and Bloch-Messiah decomposition is for symplectic matrices which $\gamma$ in general will not be. $\endgroup$
    – Kiro
    Aug 1 '17 at 5:46

$\DeclareMathOperator\diag{diag} \newcommand\Id{\mathbb I}$ Since I’m a physicist — and this question was originally migrated from physics.sx (Why? It’s a standard quantum optic question to me!) — I’ll answer to this question using physics techniques.

I. Symplectic operations in quantum optics

Symplectic algebra offers a way to simply encode the Heisenberg uncertainty principle for multimode quantum optics: for a $n$ modes $2n×2n$ covariance matrix $σ$ where the quadatures are alternating ($Q₁, P₁, Q₂, P₂, …$) the Heisenberg uncertainty principle is \begin{align} σ + iΩ &≥0 & \text{with }& Ω=ω^{⊕n}= \begin{bmatrix}0&1\\-1&0\end{bmatrix}^{⊕n}\\ && &\text{in units where } \frac\hbar2≝1. \end{align} Physical transformations are the one which preserve this relations. Let us restrict ourselves to transformations $M$ linear in quadratures.The covariance matrix transforms as $σ↦M^TσM$. If we want $M$ to preserve Heisenberg uncertainty principle, it should preserve $Ω$: it is a symplectic transformations, defined by $M^TΩM=Ω.$

Some example of symplectic and non symplectic operations:

  • Special orthogonal operations on a single mode, defined by rotation matrices $R⊕\mathbb I_{2(n-1)}$, corresponds to a dephasing in this mode. This operation is both orthogonal and symplectic.
  • Squeezing operations, $M=\diag\left(s_1,\frac1{s_1},s_2,\frac1{s_2}, \dots\right)$, are symplectic, but not orthogonal.
  • A beamsplitter between modes $1$ and $1$ is corresponds to the following symplectic and orthogonal operation $M=\begin{bmatrix}\Id_2\cosθ & -\Id_2\sinθ \\ \Id_2\sinθ & \Id_2\cosθ\end{bmatrix}⊕\Id_{2(n-2)}$.

The above operation generates the whole symplectic group. Note that it contains a non-orthogoal operations (squeezing operations) and that some orthogonal operations are excluded (like rotations involving $P_1$ and $Q_2$: e.g. $\begin{bmatrix} 1& 0& 0& 0\\ 0&0&1&0\\0&1&0&0\\0&0&0&1 \end{bmatrix}⊕\Id_{2(n-2)}$). Therefore, symplectic diagonalizations is not SVD, since the set of allowed transformations is different.

Another set of orthogonal but not symplectic operations are orthogonal operations of determinant $-1$ (reflections), which is essentially the answer to your question about why $c$ and $d$ cannot always be made positive.

II. Local transformation of a 2-mode covariance matrix $σ_{AB}$ in its standard form

The task here is to transform a a two mode covariance matrix $σ_{AB}$ $$ σ_{AB}=\begin{bmatrix} α & γ \\ γ^T & β\end{bmatrix}$$ in a standard form where the $2×2$ submatrices $α$, $β$ and $γ$ are diagonal and, if possible, $∝\Id_2$, using local operations, that is dephasing and squeezing.

Since $α$ is symmetric, it can be diagonalized into $\diag(a_1, a_2)$ using the local special orthogonal transformation $R_a⊕\Id_2$. This operation changes $γ$, but not $β$. Applying the squeezing operation $S_a=\diag\left(\sqrt{\frac{a_2}{a_1}}, \sqrt{\frac{a_1}{a_2}}, 1, 1\right)$ finally transforms $α$ into $a\Id_2$, with $a=\sqrt{a_1a_2}$, further changing $γ$, but not transforming $β$. The same method, using $\Id_2⊕S_bR_b$, transforms $β$ into $b\Id_2$ without changing $a\Id_2$.

Our covariance matrix is now under the form $\begin{bmatrix} a\Id_2 & γ' \\ γ'^T & b\Id_2\end{bmatrix}$.The SVD of $γ'$ gives us $c$,$d$ and the orthogonal operations $O_1$ and $O_2$ s.t. $\diag(|c|,|d|)=O_1^T γ' O_2$. Let us define the rotations $R_i≝\diag(1,\det(O_i))O_i$. We have then $\diag(c,d)=O_1^T γ' O_2$. Let us now apply $R_1⊕R_2$ to our covariance matrix:

  • It diagonalizes $γ'$ and $γ'^T$, as expected.
  • On the mode $A$, it has the effect of the rotation $R_1$, but since the covaraince matrix of this mode is $a\Id_2$, it is invariant by rotation, and it does not change.
  • Similarly, the mode $B$ does not change.

The matrix is now under the standard form $\begin{bmatrix} a& &c\\ &a& &d\\c& &b\\ &d& &b\end{bmatrix}$. One could be tempted to change the balance beween $c$ and $d$ by squeezing operations, but it would destroy the balance between $a_1$ and $a_2$. All we can do is exhanging $c$ and $d$, and change both their signs simultaneously.

  • $\begingroup$ Thanks a lot for taking the time to write this answer, this clarified the process for me. Cheers! $\endgroup$
    – Kiro
    Aug 4 '17 at 9:03

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

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