Isometrization of Tensor Network States via Gauge Propagation

Abstract:We introduce a gauge-propagation approach for approximately converting generic tensor-network states into an isometric tensor-network form with a prescribed orthogonality center. In one dimension, this propagation is exact because the non-isometric factor produced by a QR or singular-value decomposition is supported on a single virtual bond. In higher-dimensional networks, however, a local step can have several outgoing directions, and the residual factor is generally not separable into independent single-bond contributions. We address this local obstruction by approximating a local tensor, or a contracted local cluster, by structured terms consisting of an isometric factor multiplied by a tensor product of output-leg factors. The isometric factor is retained at the current site or cluster, while the output-leg factors are absorbed into neighboring tensors along the propagation directions. This construction applies to general local input-output partitions for which the input-side dimension is no smaller than the output-side dimension and provides a local truncation criterion for gauge propagation. Benchmarks on random tensors show that the proposed decomposition is effective in the low-term regime and that the residual decreases for larger local clusters. For the loop-gas tensor representation of the Kitaev spin liquid, two structured terms reduce the local residual to numerical precision, and the same cluster refinement further lowers the leading-term truncation error and reduces error accumulation during gauge propagation on a finite honeycomb network. These results identify a propagation-compatible local decomposition as a useful building block for approximate isometrization and as a potential initializer or preconditioner for variational isoTNS algorithms.

Strongly Correlated Electrons (cond-mat.str-el); Computational Physics (physics.comp-ph); Quantum Physics (quant-ph)

arXiv:2606.22816 [cond-mat.str-el]

(or arXiv:2606.22816v2 [cond-mat.str-el] for this version)

https://doi.org/10.48550/arXiv.2606.22816 Focus to learn more arXiv-issued DOI via DataCite

arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.

Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.

Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.