​Excited to share our latest quantum chemistry preprint led by Clemens Giuliani. We employ a "simple" variational wavefunction composed of a few hundred optimized non-orthogonal Slater determinants, achieving energy accuracies comparable to state-of-the-art methods. arxiv.org/abs/2503.14502 1/5​

An interesting open question is whether NQS can saturate the magic and recover the value of Haar random states (I speculate, they can, but no proof yet!) (3/4)

The general idea is to minimize the ensemble energy over a suitable distribution of Hamiltonian parameters P(gamma). The wave function (a transformer) depends on the Hamiltonian, and generalizes across many instances of different Hamiltonians. We give a version of SR specialized for this task (3/5)

These models allows for applications that are hard to do with the standard framework, e.g.:
1. Automatic ensemble averages for hamiltonians with disorder
2. Computing fidelity susceptibility for automatic, and from first-principles, detection of phase transitions
3. Fine tuning of models (2/5)

(3/n) We use a Galerkin-inspired ansatz combining Neural Quantum States with Fourier decomposition to capture the relevant dynamical frequencies. The time-depedent state is the linear combination fo time-independent basis states, fully optimized by the loss function.

(2/n) Unlike conventional step-by-step approaches, our method minimizes a physically-motivated loss function across the whole time window. The loss is general and works for arbitrary, unnormalized time-dep states.

Hi there Bluesky! for those of you who don't know me, an image is worth 300 characters: (main keywords in my publications, courtesy of Scholar Goggler)