The ITransverse.jl library for transverse tensor network contractions

Transverse contraction methods are extremely promising tools for the efficient contraction of tensor networks associated with the time evolution of quantum many-body systems, allowing in some cases to circumvent the entanglement barrier that would normally prevent the study of quantum dynamics with classical resources. We present here the ITransverse.jl package, written in Julia and based on ITensors.jl, containing several of these high-level algorithms, including novel prescriptions for efficient truncations of temporal matrix product states.

💡 Research Summary

**
The paper introduces ITransverse.jl, a Julia‑based open‑source library that implements state‑of‑the‑art transverse tensor‑network contraction algorithms for simulating the real‑time dynamics of quantum many‑body systems. Traditional time‑evolution methods such as TEBD, tDMRG, or TDVP suffer from the “entanglement barrier”: the bipartite entanglement of the wavefunction grows linearly with time, forcing the MPS bond dimension to increase exponentially and rendering long‑time simulations infeasible.

Transverse contraction circumvents this problem by treating the D‑dimensional quantum system together with the time direction as a (D+1)‑dimensional tensor network. Instead of contracting rows (the usual Schrödinger picture), the network is contracted column‑wise, i.e., along the temporal direction. The left and right boundaries become Temporal Matrix Product States (tMPS) while the bulk columns act as Temporal Matrix Product Operators (tMPO). The central computational task is to repeatedly apply tMPO layers to the boundary tMPS while keeping the bond dimension under control.

Two truncation schemes are provided. The first is the conventional Reduced Density Matrix (RDM) truncation: after each tMPO application the tMPS is brought into canonical form, a Schmidt decomposition is performed, and only the largest χ singular values are retained. The second, novel scheme is based on Reduced Transition Matrices (RTM). Because the final observable is the overlap ⟨L|R⟩ of the left and right tMPS, the authors propose to truncate directly the transition operator |R⟩⟨L| rather than the separate density matrices. By tracing out parts of this operator one obtains reduced transition matrices whose spectra guide an optimal compression. In practice RTM truncation yields a substantially lower bond dimension when temporal entanglement is strong, especially in the “folded” representation where forward and backward evolution tensors are combined.

At the algorithmic level ITransverse.jl implements two high‑level methods. The Power Method iteratively applies the bulk tMPO to the left and right tMPS until convergence, effectively finding the dominant eigenvectors of the infinite‑system transfer matrix. Convergence is monitored via changes in temporal entanglement entropy and the overlap error. The Light‑Cone method exploits the finite Lieb‑Robinson velocity: for a local operator expectation value, tensors outside the causal cone reduce to identities after folding, allowing the contraction to be restricted to a narrow spatio‑temporal region. This reduces the computational cost from O(Nₜ²) to O(v·Nₜ), where v is the cone width.

The paper also defines generalized temporal entropies, obtained from the Schmidt spectrum of the tMPS, and uses them to quantify the difficulty of a given simulation. A linear growth of temporal entropy signals a volume‑law regime where even transverse methods become costly; however, many physically relevant quenches exhibit sub‑linear growth, making the presented techniques highly advantageous. Complexity analysis shows that the overall cost scales as O(χ³·Nₜ) for a tMPO with bond dimension χ, and the authors provide symmetric SVD and symmetric eigen‑decomposition routines that halve memory usage.

ITransverse.jl is built on top of ITensors.jl and ITensorMPS.jl, inheriting their high performance, automatic differentiation, and GPU support. Installation is a single Julia package manager command, and the library offers a clean API for defining custom Hamiltonians, constructing MPOs, and selecting truncation strategies. Its modular design (Core, Truncation, Algorithms, Utilities) enables users to plug in new truncation schemes or extend the light‑cone algorithm with minimal effort.

Benchmark results on 1D spin chains (transverse‑field Ising, XXZ) demonstrate that, for comparable accuracy, the transverse methods achieve a 2–3× reduction in bond dimension relative to TEBD, and that the folded‑RTM combination remains stable even when temporal entanglement approaches saturation.

In conclusion, ITransverse.jl provides the first publicly available, production‑ready implementation of transverse tensor‑network contraction techniques. By focusing on temporal entanglement rather than spatial entanglement, it opens the door to longer‑time simulations of quantum dynamics, with potential extensions to higher‑dimensional lattices, open quantum systems, and connections to quantum field‑theoretic concepts such as influence functionals. The library is poised to become a valuable tool for both methodological research and practical many‑body simulations.