arXiv:2610.21273
Regret Lower Bound for Online Learning of Quantum States with Logarithmic Loss, with Applications to Quantum State Tomography and Confidence Sequences for Matrix Means
Yen-Huan Li, Wei-Fu Tseng
Abstract
We study the problem of online learning of quantum states with logarithmic loss, which is a quantum generalization of online portfolio selection and has applications to quantum state tomography, the task of estimating an unknown quantum state from measurement outcomes. We first establish a regret lower bound, showing that the regret rates of Exponential Weights and the algorithm of Tseng et al. (2025) have optimal polynomial dependence on the dimension. Our analysis leads to two further results. First, our lower-bound proof leads to a quantum state tomography algorithm that is sample optimal up to polylogarithmic factors in the sample size. Second, we develop an algorithm for constructing confidence sequences for the mean of a matrix-valued random variable. A numerical experiment shows that the latter algorithm is competitive with existing methods.
AI Summary
The paper asks whether the extra factor of $d$ in logarithmic-regret bounds for online learning of $d$-dimensional quantum states is intrinsic. Each round predicts a density matrix $\rho_t$ and incurs loss $-\log\operatorname{tr}(A_t\rho_t)$ for a revealed positive semidefinite matrix $A_t$. For $T\gg d^2\log d$, the main result rules out uniform worst-case regret guarantees of $O(d^\gamma\log T)$ for any $\gamma<2$. Thus, existing $O(d^2\log T)$-type guarantees for Exponential Weights and Q-VB-FTRL have optimal polynomial dependence on dimension. The proof reduces nonadaptive, single-copy tomography to online learning: predictions are shrunk toward the maximally mixed state, then averaged using a high-probability exp-concave online-to-batch conversion. For a projective $4$-design POVM, measurement-distribution error controls trace-norm error. Combining this with the known $\Omega(d^3/\varepsilon^2)$ tomography lower bound yields the regret obstruction. The same construction gives tomography error approximately $\sqrt{d^3\log n/n}$, up to additional polylogarithmic factors, with probability at least $2/3$. A second application constructs anytime-valid confidence sequences for matrix means using nonnegative supermartingales. A general Exponential Weights bound, $\eta^{-1}\log\binom{T+p}{p}$ for $\eta$-exp-concave losses in affine dimension $p$, enables membership tests through convex optimization without running the online algorithm. One covariance-estimation experiment uses $d=8$ and $T=200$. The exact minimax regret remains unresolved: no matching $\Omega(d^2\log T)$ bound is proved. Confidence-set size convergence rates and optimality are also unestablished; the numerical comparison concerns one setting.
Appendix \ref{app_numerical} was initially generated by GPT-6 after multiple rounds of conversations and then modified by the authors.
Written by gpt-6.1-sol from the paper's full text. Automated: it can be wrong.