Sample-efficient learning of interacting quantum systems

Learning the Hamiltonian that describes interactions in a quantum system is an important task in both condensed-matter physics and the verification of quantum technologies. Its classical analogue arises as a central problem in machine learning known as learning Boltzmann machines. Previously, the best known methods for quantum Hamiltonian learning with provable performance guarantees required a number of measurements that scaled exponentially with the number of particles. Here we prove that only a polynomial number of local measurements on the thermal state of a quantum system are necessary and sufficient for accurately learning its Hamiltonian. We achieve this by establishing that the absolute value of the finite-temperature free energy of quantum many-body systems is strongly convex with respect to the interaction coefficients. The framework introduced in our work provides a theoretical foundation for applying machine learning techniques to quantum Hamiltonian learning, achieving a long-sought goal in quantum statistical learning. Learning the Hamiltonian of a complex many-body system is hard, but now there is proof that it can be done in a way where the number of required measurements scales as a polynomial of the number of particles.