Tailoring tensor network techniques to the quantics representation for highly inhomogeneous problems and few body problems

Jheng-Wei Li, Nicolas Jolly, Xavier Waintal · 2026

Read Tailoring tensor network techniques to the quantics representation for highly inhomogeneous problems and few body problems by Jheng-Wei Li, Nicolas Jolly, Xavier Waintal (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2604-09337.txt.

Introduction

Li, Jolly, and Waintal tailor tensor network techniques to the quantics representation for inhomogeneous and few-body problems. Tensor networks solve partial differential equations in quantics form, with resolution depending only logarithmically on grid point count. First-generation techniques used out-of-the-box algorithms designed for settings like quantum magnetism, where degrees of freedom play equal roles. In quantics form, however, degrees of freedom correspond to physics at different scales. By tailoring algorithms to this case, the authors obtain faster and more robust convergence.

Catalog record

CREATORJheng-Wei Li, Nicolas Jolly, Xavier Waintal
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE59942
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2604.09337

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Tailoring tensor network techniques to the quantics representation for highly inhomogeneous problems and few body problems Jheng-Wei Li, Nicolas Jolly, Xavier Waintal Abstract Tensor network techniques are becoming increasingly popular tools to solve partial differential equations within the so-called quantics representation. Their popularity stems from the fact that their spatial resolution depends only logarithmically on the number of grid points, making them very tempting approaches in situations where two or more characteristic length scales are vastly different. A first generation of technique used “out-of-the-box” algorithms of the tensor network toolkit (e.g. the celebrated Density Matrix Product State (DMRG) algorithm) to solve these problems. These techniques were designed for situations (e.g. quantum magnetism) where the different degrees of freedom (e.g. spins) play equivalent roles. In the quantics representation, however, the different degrees of freedom correspond to the physics at different scales and therefore play inequivalent role. Here we show that by tailoring the tensor network algorithms to this particular case, in the spirit of the multigrid approach, we obtain faster and more robust convergence of the algorithms. We showcase the approach on linear (Poisson equation) and eigenvalue (Schrödinger equation) problems in two, three and four dimensions. Our simulations involve up to 2 80 2^{80} grid points and would represent, we argue, a very strong challenge for conventional approaches. I Introduction As a general rule, large differences in magnitude between different characteristic length scales are hard to address numerically because one needs to use a discretization dictated by the smallest of these scales. This is somewhat paradoxical because the same property can be harvested in analytical works to solve the problem. For instance, if one studies the dynamics of electrons with wavelength \lambda subject to scattering with a mean free path \ell\gg\lambda , one can integrate out the ballistic motion taking place at scales smaller than \ell and derive an effective diffusion equation valid on scales much larger than \ell . In that case, the existence of a small parameter \lambda/\ell\ll 1 provides an efficient route to solve the problem; the hard case is the crossover regime \ell\sim\lambda (where Anderson localization occurs) Akkermans and Montambaux 2020 . Another example is geometrical optics which emerges when one considers mirrors and lenses of characteristic sizes L\gg\lambda . A last example is homogenization problems in e.g. diffusion equation in rapidly varying media Sanchez 2009 . There, one can show that an effective diffusion equation emerges at large scale (albeit with a diffusion constant that depends non-trivially on the microscopic media). Surely, if we can take advantage of the separation of scales analytically, there should be a way to do it numerically too, if only through an algorithmic path that follows closely what the analytics do. I.1 The quantics representation An implementation of such a program is the quantics tensor train (QTT) representation which naturally expresses functions in terms of their behavior at different scales. QTT encodes discretized functions on 2^{N} grid points using only N binary degrees of freedom and an O(N) memory footprint Oseledets 2009 ; Oseledets and Tyrtyshnikov 2011 ; Oseledets 2012 ; Dolgov et al. 2012 ; Chen et al. 2023 ; Lindsey 2023 . In recent years, QTT methods have rapidly gained attraction across multiple disciplines including classical turbulence Gourianov et al. 2022 ; Peddinti et al. 2024 ; Kornev et al. 2023 ; Hölscher et al. 2025 ; Gourianov et al. 2025 ; van Hülst et al. 2025 , plasma physics Ye and Loureiro 2022 , multidimensional Fokker–Planck equations Chertkov and Oseledets 2021 , quantum chemistry Khoromskaia and Khoromskij 2015 ; Jolly et al. 2025 , Gross-Pitaevskii equation Bou-Comas et al. 2025 ; Chen et al. 2025 ; Connor et al. 2025 ; Niedermeier et al. 2025 and quantum (many-body) physics Benner et al. 2017 ; Shinaoka et al. 2023 ; Erpenbeck et al. 2023 ; Ritter et al. 2024 ; Murray et al. 2024 ; Otero Fumega et al. 2024 ; Kiese et al. 2025 ; Rohshap et al. 2025 ; Kim and Werner 2025 ; Gidi et al. 2025 . By representing data with only logarithmically many parameters, QTT provides a compact and powerful computational framework with practical tools to perform important operations such as differentiation, integration, Fourier transform…with an O(N) complexity (see Waintal et al. 2026 for an introduction). The key point is that the problem must possess a certain separation of scales for the associated functions to be easily compressible in the QTT representation. Figure 1: Top : The V-cycle of the multi-grid method. When descending, the problem ( L ) is coarse-grained down to a sparser grid. At the coarsest level, the problem is initially solved. Its solution then ascends by successive interpolation and is refined at every level until reaching the finest grid. Bottom : QTT calculation of the lowest eigenvalue of the hydrogen molecule ion H_{2}^{+} using plain DMRG (static resolution, N=60 ) or the multigrid DMRG approach (dynamical resolution N=12\cdots 60 ). The ultimate resolution of this 3D problem is 2^{N/3}\times 2^{N/3}\times 2^{N/3}=2^{N}\approx 10^{18} grid points for our highest resolution N=60 . The plot show the error on the energy as a function of the number DMRG optimization steps. In the dynamical resolution strategy, we solve at a given resolution N , increase N to N+3 by interpolation, solve again…and repeat until the final resolution N=60 is reached. Clearly in this case, the convergence is much more robust than in the static resolution approach. This article is concerned with the problem of scaling up QTT algorithms to use-cases beyond the proof of principle stage. We will consider two such problems. The first is solving the Poisson equation with a highly oscillatory charge density. Here, the linear problem of the form \Delta u=\rho where \Delta (the discretized Laplacian) and \rho (the discretized charge density) can be put, respectively, in the form of a Matrix Product Operator (MPO) and Matrix Product State (MPS) while we seek the solution u as an MPS. The second problem is a three body quantum system in four dimensions. After discretization, it will take the form of an eigenvalue problem of the form H\Psi=E\Psi where H is a MPO while we seek the eigenvector \Psi in the form of a MPS. Traditionally, one uses solvers developed in the context of many-body physics: an Alternating Least Square (ALS) solver Holtz et al. 2012 for the linear problem and the Density Matrix Renormalization Group (DMRG) solver White 1992 for the eigenvalue problem. However, these techniques have been designed for problems where the degrees of freedom are equivalent and may not be optimal for QTT where they are not. Here, we seek to leverage on the specific properties of QTT to speedup the convergence. I.2 Multi-resolution approaches for iterative solvers Specifically, we shall make use of the fact that QTT naturally implement the concepts needed in multigrid calculations. Let us mention that upon completion of this work, we learned of Ref. Lubasch et al. 2018 where this idea has already been pursued, although for different problems and smaller dimensions. The basic idea is straightforward: from a QTT at a given resolution, it is easy to average out the smaller scales to obtain a coarser QTT or to increase the resolution by interpolation

Read the complete text