Aurora Mugnai

In a previous work [Senno et al., Phys. Rev. Lett. 131, 130202 (2023)], we provided a framework to quantify the amount of intrinsic randomness produced by characterized but untrusted prepare-and-measure (P\&M) setups. While the cases of pure state preparations or extremal measurements were shown to be defined by semidefinite programs (SDPs), up until now we were missing computational methods for the general scenario of mixed states and nonextremal measurements. In this work, we present a hierarchy of SDP relaxations to lower bound the device-dependent conditional min-entropy. Benchmarking against known special cases, we find that the first level of the hierarchy already attains the optimal value. We then provide two applications. First, for setups affected by global depolarizing noise, we compute a matching lower bound to the analytical attack derived in [Curran et al., arXiv:2506.22294 (2025)], thus showing its optimality. Finally, we show that restricting the correlations between the P\&M boxes to be classical strictly decreases an adversary's predictive power, already in the most elementary setup of a qubit binary measurement.