The sandwiched Renyi divergence is defined as in
$$ \tilde{D}_\alpha(\rho\|\sigma):=\frac{1}{\alpha−1}\log tr[(\sigma^{\frac{1−\alpha}{2\alpha}}\rho \sigma^{\frac{1−\alpha}{2 \alpha }})^\alpha] $$

The divergence measure takes on finite values when $\rho, \sigma$ are not orthogonal to each other and is $\infty$ otherwise.

If we take $\alpha > 1$, the above expression involves fractional powers of $\sigma^{-1}$. Since non-orthogonality (or overlap of support of $\rho, \sigma$) is important, how should we consider the inverse of $\sigma$? Will the generalized inverse of a matrix suffice?


1 Answer 1


Firstly, the sandwiched divergence can be infinite even when $\rho$ and $\sigma$ are not orthogonal. For example, consider $\rho = \frac{|0\rangle \langle 0| + |1\rangle\langle 1|}{2}$ and $\sigma = |0\rangle \langle 0|$. We can think of this as the limiting case of $\sigma_\epsilon = (1-\epsilon) |0\rangle \langle 0| + \epsilon |1 \rangle \langle 1|$. Computing the divergence with $\sigma_\epsilon$ we don't have any issues because everything is full rank and we find $$ D_\alpha(\rho\|\sigma_\epsilon) = \frac{1}{\alpha-1}\log \left( 2^{-\alpha}\left((1-\epsilon)^{1-\alpha} + \epsilon^{1-\alpha}\right) \right) $$ which tends to $+\infty$ as $\epsilon \to 0$ when $\alpha > 1$.

The correct condition for finiteness is that for every vector $|v\rangle$ such that $\sigma |v\rangle = 0$ we must also have $\rho |v \rangle =0$. In other words the kernel of $\sigma$ is contained in the kernel of $\rho$. So for example if we were to swap the roles of $\rho$ and $\sigma$ in the previous example, so $\sigma = \frac{|0\rangle \langle 0| + |1\rangle\langle 1|}{2}$ and $\rho = |0\rangle \langle 0|$, then this satisfies our condition for finiteness. Moreover we can see it as a limiting case of $\rho_\epsilon = (1-\epsilon) |0\rangle \langle 0| + \epsilon |1 \rangle \langle 1|$ which gives $$ D_\alpha(\rho_\epsilon\|\sigma) = \frac{1}{\alpha-1}\log \left( 2^{\alpha-1}\left((1-\epsilon)^{\alpha} + \epsilon^{\alpha}\right) \right) $$ which tends to $1$ as $\epsilon \to 0$ whenever $\alpha > 1$ which also agrees with a direct computation of $D_\alpha(\rho\|\sigma)$.

To answer your question, yes you can take the pseudoinverse which for Hermitian operators can be computed readily via the spectral decomposition. If $\sigma = \sum_i \lambda_i |v_i\rangle \langle v_i|$ is the spectral decomposition of $\sigma$ then $\sigma^{-1} = \sum_i \lambda_i^{-1} |v_i\rangle \langle v_i|$.


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.