Speaker
Description
Quantum autoencoders (QAEs) compress families of quantum states onto a few latent qubits, but their latent space is a black box: one can neither certify what it contains nor read it in physical terms. We show that one physically motivated constraint, restricting the encoder to conserve particle number, resolves both problems at once. The latent becomes operationally certifiable (measuring the trash register and accepting the all-zeros outcome projects it exactly into the correct $N$-electron sector) and interpretable (the trained one-body encoder reproduces the state-averaged natural-orbital active space, an identification proved from the optimality condition and verified numerically, so the latent is a learned, state-averaged active space in the literal SA-CASSCF sense). We note that this interpretability comes at a cost: the constrained encoder compresses to at least as many qubits as a hardware-efficient ansatz (HEA). We map where the latent's physical meaning holds and report actual numbers, including the negative results, on an exactly-solvable Hubbard control and on H$_2$/H$_4$/H$_6$ chains.
| Topic | Quantum variational circuits and their applications |
|---|---|
| Additional Topic | Encoding and processing of data in quantum systems |