PreprintarXiv licencearXiv
Physical policy gradient theorem for in situ stochastic-adjoint training
Abstract
In situ adjoint training extracts parameter gradients directly from measurement, but has so far been limited to reciprocal or restricted systems. Here, we introduce the physical counterpart of the policy gradient theorem: a stochastic-adjoint gradient estimator that lifts these constraints by trading reciprocity for nondegenerate diffusion. As validation, we train a nonlinear resonator network, whose own dynamics supply the policy, against antagonistic temporal modulations with gradients from measured stochastic trajectories alone, without finite differences or a separate adjoint experiment.
Related
- Symmetry-driven in-situ physical adjoint computing beyond reciprocity
- Physical Adjoint Learning in Wave Systems with Asymmetric Couplings
- Wasserstein-2 gradient flows and the geometry of entropy production in classical and quantum stochastic thermodynamics
- Stochastic learning flow: moment equations for stochastic gradient descent at finite step size
- Dilation theorem for continuum quantum stochastic processes
BibTeX
@misc{tuxbury2026physical,
title = {{Physical policy gradient theorem for in situ stochastic-adjoint training}},
author = {William Tuxbury and Zin Lin},
howpublished = {arXiv},
year = {2026},
month = sep,
doi = {10.48550/arxiv.2609.05808},
eprint = {2609.05808},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/2609.05808},
}
RIS
TY - UNPB
TI - Physical policy gradient theorem for in situ stochastic-adjoint training
AU - William Tuxbury
AU - Zin Lin
PY - 2026
DA - 2026/09/05
PB - arXiv
DO - 10.48550/arxiv.2609.05808
UR - https://arxiv.org/abs/2609.05808
AB - In situ adjoint training extracts parameter gradients directly from measurement, but has so far been limited to reciprocal or restricted systems. Here, we introduce the physical counterpart of the policy gradient theorem: a stochastic-adjoint gradient estimator that lifts these constraints by trading reciprocity for nondegenerate diffusion. As validation, we train a nonlinear resonator network, whose own dynamics supply the policy, against antagonistic temporal modulations with gradients from measured stochastic trajectories alone, without finite differences or a separate adjoint experiment.
ER -