This paper studies a sequential decision problem in quantum information: a learner repeatedly predicts a quantum state and incurs logarithmic loss based on the actual state. The setup is a quantum generalization of online portfolio selection, where the goal is to minimize cumulative regret over time.

The main contribution is a regret lower bound, which establishes a fundamental performance limit for any online learning algorithm in this setting. By quantifying the unavoidable worst-case error, the result clarifies how hard it is to learn quantum states adaptively.

The authors connect this bound to practical tasks. In quantum state tomography, the bound informs the difficulty of estimating an unknown state from sequential measurements. The same framework also yields confidence sequences for matrix means, offering sequential statistical guarantees for matrix-valued data.