Quantum information · Theory
Quantum robustness via sandwiched Rényi divergences
arXiv:2611.12738v1 ↗PDF ↗Submitted 9 November 2026
- Question
- Tomography, certification and classical shadows return a state known only up to a confidence region, usually a ball in fidelity or relative entropy. To turn that into a guarantee on an observable one needs the worst case: the largest expectation \(\mathrm{tr}(L\rho)\) over states \(\rho\) within divergence \(q\) of the reference \(\sigma\). Classically this distributionally robust value has closed forms; quantum mechanically it was known only at the endpoints, and how noncommutativity of \(L\) with \(\sigma\) affects it was unclear.
- Result
- For the sandwiched Rényi divergences of every order \(\alpha\in[\tfrac12,\infty]\), the worst case is a one-dimensional optimisation over a sandwiched Schatten norm, ranging from a fidelity risk at \(\alpha=\tfrac12\) through the Gibbs variational formula at \(\alpha=1\) to a quantum conditional value-at-risk at \(\alpha=\infty\). The main theorem is classicalization: for each \(\alpha\) there is a probability measure \(\omega_\alpha\) on the spectrum of \(L\) such that the quantum problem equals the classical one for \(\omega_\alpha\). For small budgets \(V_\alpha(q)=\mathrm{tr}(L\sigma)+s_\alpha\sqrt{2q}+o(\sqrt q)\), and the variance of \(\omega_\alpha\) decreases strictly in \(\alpha\) when \([\sigma,L]\neq0\).
The paper in detail
Fix a reference state \(\sigma\), a Hermitian observable \(L\) and a budget \(q\). The robust value is \(V_\alpha(q)=\sup\{\mathrm{tr}(L\rho):H_\alpha(\rho\|\sigma)\le q\}\), where \(H_\alpha=(Q_\alpha-1)/(\alpha-1)\) is the sandwiched Rényi divergence in Hellinger normalisation and \(Q_\alpha(\rho\|\sigma)=\mathrm{tr}\bigl[(\sigma^{(1-\alpha)/2\alpha}\rho\,\sigma^{(1-\alpha)/2\alpha})^{\alpha}\bigr]\). At \(\alpha=\tfrac12\) the ball is a fidelity ball, at \(\alpha=1\) a relative-entropy ball, at \(\alpha=\infty\) a max-divergence ball. Balls are nested in \(\alpha\), so \(V_\alpha\) decreases with the order.
The questions are whether \(V_\alpha\) has a usable formula, and how it compares with the classical answer for the Born distribution of \(L\) in \(\sigma\), which is what one gets by ignoring noncommutativity.
Convex duality gives, for \(\alpha>1\), \[V_\alpha(q)=\inf_{c}\Bigl\{c+\bigl(1+(\alpha-1)q\bigr)^{1/\alpha}\bigl\|\bigl(\sigma^{\beta}(L-c)\sigma^{\beta}\bigr)_{+}\bigr\|_{\alpha^{\prime}}\Bigr\},\qquad\beta=\tfrac{\alpha-1}{2\alpha},\] the quantum optimized certainty equivalent, with a resolvent version for \(\alpha<1\). Classicalization says each such formula is the classical one for a measure \(\omega_\alpha\) with the same mean and support as the Born measure: Born itself at \(\alpha=\tfrac12\), the Laplace-transform measure of Stahl's theorem at \(\alpha=1\), and a measure built from Heinävaara's joint spectral measure for \(\alpha>1\). Its variance is the inverse curvature of the divergence, so the small-budget slope \(s_\alpha\) is the inverse of the sandwiched Fisher metric applied to \(L\).
Open directions are the same programme for Petz and measured Rényi divergences, whose duals are of a different nature, and explicit computation of \(\omega_\alpha\) in larger systems, which would make the deflated bounds practical for certification.
Transferable methods
Closed-form worst case at \(\alpha=2\)
The dual formula is a one-dimensional infimum, and at \(\alpha=2\) it can be done by hand because the Schatten norm is Hilbert–Schmidt. This gives an exact square-root law and an explicit worst-case state, and shows how the variance that controls robustness arises.
At \(\alpha=2\), \(\beta=\tfrac14\). For \(c\) below the spectrum, with \(u=\mathrm{tr}(L\sigma)-c\) and \(\hat L=L-\mathrm{tr}(L\sigma)\), \[\bigl\|\sigma^{1/4}(L-c)\sigma^{1/4}\bigr\|_2^{2}=2s_2^{2}+u^{2},\qquad 2s_2^{2}=\mathrm{tr}\bigl(\sigma^{1/2}\hat L\sigma^{1/2}\hat L\bigr),\] since \(\mathrm{tr}(\sigma\hat L)=0\). Minimising \(\mathrm{tr}(L\sigma)-u+\sqrt{1+q}\sqrt{2s_2^{2}+u^{2}}\) gives \(u_*=s_2\sqrt{2/q}\) and \(V_2(q)=\mathrm{tr}(L\sigma)+s_2\sqrt{2q}\) for \(q\le q_0=2s_2^{2}/\lambda_{\min}(\hat L)^{2}\).
The worst case is \(\rho^{*}=\sigma^{1/2}\bigl(I+\sqrt{q/2}\,\hat L/s_2\bigr)\sigma^{1/2}\). Editorial check: \(\mathrm{tr}\rho^{*}=1\), \(Q_2(\rho^{*}\|\sigma)=1+q\) and \(\mathrm{tr}(L\rho^{*})=\mathrm{tr}(L\sigma)+s_2\sqrt{2q}\) follow by expanding the square; for a random qubit, direct constrained maximisation agrees with the formula to six digits at \(q=0.2q_0\) and \(0.8q_0\). The same check for the \(\alpha=1\) Gibbs dual agrees at \(q=10^{-4},10^{-2},5\times10^{-2}\).
How to reuse it
Whenever a robust or worst-case value has a dual involving a Hilbert–Schmidt norm of a shifted operator, the shift cancels the cross term and the problem reduces to a scalar quadratic, giving an exact square-root law with an explicit maximiser.
Variance kernels order the orders
The small-budget slope is set by a variance, and comparing orders requires comparing these variances. In the eigenbasis of \(\sigma\) each is a weighted sum of \(|\hat L_{ij}|^{2}\) with a kernel in the eigenvalues, and the kernels are classical means that are ordered.
With \(\sigma=\sum_ip_i\ketbra{i}{i}\), Theorem 5.3 gives \(\mathrm{Var}_{\omega_\alpha}=\sum_{ij}\alpha\,m_\alpha(p_i,p_j)|\hat L_{ij}|^{2}\), decreasing in \(\alpha\), from the Born variance \(\mathrm{tr}(\sigma\hat L^{2})\) at \(\alpha=\tfrac12\) (kernel \((p_i+p_j)/2\)) to \[\mathrm{Var}_{\omega_\infty}=\sum_{i,j}\frac{p_ip_j\log(p_i/p_j)}{p_i-p_j}|\hat L_{ij}|^{2}\] at \(\alpha=\infty\). Editorial remark: at \(\alpha=1\) the kernel is the logarithmic mean \((p_i-p_j)/\log(p_i/p_j)\), giving the Kubo–Mori variance, and at \(\alpha=2\) it is the geometric mean \(\sqrt{p_ip_j}\), matching \(2s_2^{2}\) above.
All kernels equal \(p_i\) on the diagonal, so they agree when \([\sigma,L]=0\) and separate otherwise. Editorial check: for a random qubit the four variances are \(0.360\), \(0.194\), \(0.127\) and \(0.094\) (Born, \(\alpha=1,2,\infty\)), in the stated order, and the \(\alpha=1\) dual approaches \(\mathrm{tr}(L\sigma)+\sqrt{2q\,\mathrm{Var}_{\mathrm{KM}}}\) as \(q\to0\).
How to reuse it
Quantum variances built from monotone metrics differ only in a two-variable kernel; to compare them, identify each kernel as a mean (arithmetic, logarithmic, geometric, and so on) and use the classical inequalities between means.
Context & limitations
Why it matters. Classicalization lets classical concentration bounds (Hoeffding, Bernstein, Markov–Krein) be used verbatim on quantum robustness problems, and reading them on the Helstrom observable recovers Pinsker and Fuchs–van de Graaf inequalities with Bernstein-type refinements. The message the author draws is a quantum deflation: for \(\alpha>\tfrac12\), noncommutativity strictly lowers the worst case below what the Born distribution of \(L\) alone would allow.
Limitations. The measures \(\omega_\alpha\) are defined through Heinävaara's tracial joint spectral measure and, at \(\alpha=1\), Stahl's theorem; they exist but are not easy to compute beyond small cases. Results are for finite-dimensional, faithful reference states. Other quantum Rényi families (Petz, measured) and hockey-stick \(f\)-divergences are left open; since the sandwiched divergence is the smallest of the family, its bounds are the most conservative. The author states that Claude was used extensively for drafting, cross-checking proofs and references, numerical verification and figures, and that the author verified all mathematical claims.