Louis Schatzki

We make progress in understanding the relationship between learning models with access to entangled, separable and statistical measurements in the quantum statistical query (QSQ) model. We show the following results. Entangled versus separable measurements: The goal is to learn an unknown f from the concept class C containing functions from 0,1^n to [k] given copies of a uniform superposition over |x,f(x)>. We show that, if T copies suffice to learn f using entangled measurements, O(nT^2) copies suffice to learn f using only separable measurements. Entangled versus statistical measurements: The goal is to learn a function f in C given access to separable measurements or statistical measurements. We exhibit a concept class C based of degree-2 functions with exponential separation between QSQ learning and quantum learning with entangled measurements (even in the presence of noise). This proves the ""quantum analogue"" of the seminal result of Blum et al. that separates classical SQ learning from classical PAC learning with classification noise. QSQ lower bounds for learning states: We introduce a quantum statistical query dimension (QSD), and use it to give lower bounds on the QSQ complexity of learning. We prove superpolynomial QSQ lower bounds for testing purity of quantum states, shadow tomography, learning coset states for the Abelian hidden subgroup problem, degree-2 functions, planted biclique states, and learning output states of Clifford circuits of depth polylog(n). We also show that an extension of QSD characterizes the complexity of general search problems. Further applications: We give an unconditional separation between weak and strong error mitigation and prove lower bounds for learning distributions in the QSQ model. Prior works by Quek et al., Hinsche et al., and Nietner et al. proved analogous results assuming diagonal measurements and our work removes this assumption.

In this work, we study the sample complexity of two variants of product testing when restricted to single-copy measurements. In particular, we consider both bipartite product testing (i.e., does there exist at least one non-trivial cut across which the state is product) and multipartite product testing (i.e., is the state fully product across every cut). For the first variant, we prove an exponential lower bound on the sample complexity of any algorithm for this task which utilizes only single-copy measurements. When comparing this with known efficient algorithms that utilize multi-copy measurements, this establishes an exponential separation for this and several related entanglement learning tasks. For the second variant, we prove another sample lower bound that establishes a separation between single- and multi-copy strategies. To obtain our results, we prove a crucial technical lemma that gives a lower bound on the overlap between tensor products of permutation operators acting on subsystems of states that themselves carry a tensor structure. Finally, we provide an algorithm for multipartite product testing using only single-copy, local measurements, and we highlight several interesting open questions arising from this work.

Quantum computers are believed to provide expo- nential speedups in solving certain classes of opti- mization problems, with the recently introduced Decoded Quantum Interferometry (DQI) frame- work (Jordan et al., 2025) providing a template for discovering novel applications of this form. Since many quantum speedups apply only to discrete problems with algebraic structure, it is of great in- terest to understand when speedups are possible in less structured settings that more closely resemble natural problems. We consider constraint satisfac- tion problems with continuous, unstructured con- straints that are inspired by problems in machine learning. We analyze the performance of a quan- tum algorithm based on DQI that leverages the sparse recovery guarantees of compressed sens- ing. We prove that this approach outperforms certain classical algorithms with provable guar- antees, and suggest regimes where it might also outperform classical heuristic algorithms like sim- ulated annealing. Our work presents a novel way in which the powerful toolkit of continuous sparse recovery algorithms can be used to design novel quantum algorithms.