pith. sign in

arxiv: 1906.12030 · v1 · pith:UIIW6B4Enew · submitted 2019-06-28 · ⚛️ physics.comp-ph

TensorNetwork on TensorFlow: Entanglement Renormalization for quantum critical lattice models

Martin Ganahl , Ashley Milsted , Stefan Leichenauer , Jack Hidary , Guifre Vidal This is my paper

Pith reviewed 2026-05-25 13:55 UTC · model grok-4.3

classification ⚛️ physics.comp-ph
keywords tensor networksMERAquantum criticalitytransverse-field Ising modelTensorFlowGPU accelerationconformal data
0
0 comments X

The pith

TensorNetwork on TensorFlow optimizes the MERA tensor network to approximate the ground state of the critical transverse-field Ising chain and extracts conformal data, with GPU runtimes up to 200 times faster than CPU.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper shows how the TensorNetwork library, running on a TensorFlow backend, can be used to carry out the full optimization of a Multi-scale Entanglement Renormalization Ansatz for an infinite quantum spin chain. The target is the ground state of the one-dimensional transverse-field Ising model exactly at its quantum critical point. Once optimized, the ansatz supplies scaling dimensions and other conformal data that characterize the critical theory. The same implementation runs the optimization loop on both CPU and GPU hardware, revealing large reductions in wall-clock time on the GPU.

Core claim

TensorNetwork with TensorFlow backend implements a complete MERA optimization algorithm that approximates the ground-state wave function of the infinite critical transverse-field Ising model and extracts conformal data from the optimized tensors, while delivering up to a 200-fold reduction in runtime when the contractions and gradient steps are executed on a GPU rather than a CPU.

What carries the argument

The Multi-scale Entanglement Renormalization Ansatz (MERA), a layered tensor network whose successive coarse-graining layers encode entanglement renormalization and whose variational parameters are updated by gradient descent on the energy expectation value computed via TensorNetwork contractions.

If this is right

  • The same TensorNetwork implementation can be reused to optimize MERA for other one-dimensional quantum critical points whose conformal data are not known analytically.
  • Automatic differentiation through the MERA energy functional removes the need for hand-derived update rules, allowing rapid experimentation with different cost functions or symmetry constraints.
  • The observed GPU speedup scales with system size and bond dimension, making previously prohibitive MERA calculations routine on commodity hardware.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same library interface could be applied without modification to other tensor-network families such as matrix-product states or projected entangled-pair states, inheriting the same GPU acceleration.
  • Because TensorFlow already supports distributed and TPU execution, the reported workflow opens a direct route to multi-device or cloud-scale MERA optimizations that were previously limited by single-GPU memory.
  • The extracted conformal data can be fed back into the same TensorFlow graph to train a supervised model that predicts critical exponents for nearby Hamiltonians, creating a closed loop between tensor-network numerics and machine learning.

Load-bearing premise

The TensorNetwork library and its TensorFlow backend execute all required tensor contractions and automatic differentiation steps accurately enough that the resulting optimized MERA tensors produce conformal data free of library-induced errors.

What would settle it

Run the published MERA optimization code on the critical Ising model and compare the numerically extracted scaling dimensions and central charge against the exact Ising conformal field theory values (central charge exactly 1/2, lowest nontrivial dimension exactly 1/8).

Figures

Figures reproduced from arXiv: 1906.12030 by Ashley Milsted, Guifre Vidal, Jack Hidary, Martin Ganahl, Stefan Leichenauer.

Figure 1
Figure 1. Figure 1: FIG. 1. (a) Binary tree tensor network for a wave function [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: FIG. 2. First four layers of a scale invariant binary MERA. [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: FIG. 3. Computation of the expectation value [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figure 5
Figure 5. Figure 5: FIG. 5. Update to the isometries [PITH_FULL_IMAGE:figures/full_fig_p004_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: FIG. 6. Relative error of the energy of an optimized, scale [PITH_FULL_IMAGE:figures/full_fig_p005_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: FIG. 7. Lowest twelve scaling dimensions (blue dots) obtained [PITH_FULL_IMAGE:figures/full_fig_p005_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: FIG. 8. Comparison of average runtimes per optimization [PITH_FULL_IMAGE:figures/full_fig_p006_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: FIG. 9. Runtimes of individual steps of the MERA optimization using single threaded operation. The top panel shows results [PITH_FULL_IMAGE:figures/full_fig_p007_9.png] view at source ↗
read the original abstract

We use TensorNetwork [C. Roberts et al., arXiv: 1905.01330], a recently developed API for performing tensor network contractions using accelerated backends such as TensorFlow, to implement an optimization algorithm for the Multi-scale Entanglement Renormalization Ansatz (MERA). We use the MERA to approximate the ground state wave function of the infinite, one-dimensional transverse field Ising model at criticality, and extract conformal data from the optimized ansatz. Comparing run times of the optimization on CPUs vs. GPU, we report a very significant speed-up, up to a factor of 200, of the optimization algorithm when run on a GPU.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

0 major / 3 minor

Summary. The paper implements the optimization of the Multi-scale Entanglement Renormalization Ansatz (MERA) using the TensorNetwork library with a TensorFlow backend. The implementation is applied to approximate the ground state of the infinite one-dimensional transverse-field Ising model at criticality; conformal data (scaling dimensions and central charge) are extracted from the optimized tensors, and runtime benchmarks report GPU speedups of up to a factor of 200 relative to CPU execution.

Significance. If the reported implementation and benchmarks hold, the work supplies a concrete, reusable demonstration of the TensorNetwork API for a non-trivial tensor-network algorithm on a standard quantum-critical model. The extraction of conformal data provides an internal consistency check, while the timing results quantify the practical benefit of the TensorFlow backend for MERA contractions and gradient-based optimization.

minor comments (3)
  1. §4 (timing benchmarks): the hardware specifications (CPU model, GPU model, TensorFlow version, and batch sizes) should be stated explicitly so that the reported 200x factor can be reproduced.
  2. Figure 3 (conformal data table): the quoted error bars on the scaling dimensions are not accompanied by a description of how they were obtained from the MERA tensors; a brief statement on the fitting procedure would improve clarity.
  3. Reference list: the citation to the authors' prior TensorNetwork paper (arXiv:1905.01330) is appropriate, but the manuscript should also cite the original MERA literature (Vidal 2007, 2008) when describing the ansatz and the conformal-data extraction method.

Simulated Author's Rebuttal

0 responses · 0 unresolved

We thank the referee for the positive summary, significance assessment, and recommendation of minor revision. No specific major comments were provided in the report.

Circularity Check

0 steps flagged

No significant circularity: implementation benchmark on standard model

full rationale

The paper implements a known MERA optimization algorithm via the cited TensorNetwork library (arXiv:1905.01330) and applies it to the infinite critical TFIM, a well-studied model with independently known conformal data. Extracted scaling dimensions and central charge are compared to external benchmarks, and GPU/CPU timings are reported as direct measurements. No derivation reduces by construction to fitted inputs, self-definitions, or load-bearing self-citations; the library citation supports the computational backend but the benchmark results and speedup claims remain externally falsifiable and independent of the present paper's fitted values.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The paper is an implementation and benchmarking study that relies on established tensor-network contraction rules and the pre-existing TensorNetwork library; no new free parameters, axioms, or invented entities are introduced.

axioms (1)
  • standard math Standard MERA tensor network structure and contraction rules are assumed to be correctly implemented by the TensorNetwork library.
    The work invokes the library without re-deriving the underlying tensor operations.

pith-pipeline@v0.9.0 · 5648 in / 1165 out tokens · 33813 ms · 2026-05-25T13:55:08.717022+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

38 extracted references · 38 canonical work pages · 18 internal anchors

  1. [1]

    At the largest bond dimension χ = 16 the cost for calculating ρ∗ is still about 40 times larger than the SVD of the environment of u

    Note that the SVDs in both cases were performed on CPU. At the largest bond dimension χ = 16 the cost for calculating ρ∗ is still about 40 times larger than the SVD of the environment of u. This is in contrast to the case of the TTN [30], where the factor between the most expensive tensor contractions and the SVD is on the order of 3. For this reason, the...

    work page

  2. [2]

    TensorNetwork: A Library for Physics and Machine Learning

    C. Roberts, A. Milsted, M. Ganahl, A. Zalcman, B. Fontaine, Y. Zou, J. Hidary, G. Vidal, and S. Le- ichenauer, arXiv:1905.01330 [cond-mat, physics:hep-th, physics:physics, stat] (2019), arXiv: 1905.01330

    work page internal anchor Pith review Pith/arXiv arXiv 1905

  3. [3]

    TensorNetwork: A library for physics and machine learning,

    C. Roberts et al. , “TensorNetwork: A library for physics and machine learning,” (2015), The TensorNet- work library and MERA code can be downloaded from https://github.com/google/TensorNetwork

    work page 2015

  4. [4]

    TensorFlow: Large- scale machine learning on heterogeneous systems,

    M. Abadi, A. Agarwal, P. Barham, E. Brevdo, Z. Chen, C. Citro, G. S. Corrado, A. Davis, J. Dean, M. Devin, S. Ghemawat, I. Goodfellow, A. Harp, G. Irving, M. Is- ard, Y. Jia, R. Jozefowicz, L. Kaiser, M. Kudlur, J. Lev- enberg, D. Man´ e, R. Monga, S. Moore, D. Murray, C. Olah, M. Schuster, J. Shlens, B. Steiner, I. Sutskever, K. Talwar, P. Tucker, V. Van...

    work page 2015

  5. [5]

    Fannes, B

    M. Fannes, B. Nachtergaele, and R. F. Werner, Commu- nications in Mathematical Physics 144, 443 (1992)

    work page 1992

  6. [6]

    S. R. White, Physical Review Letters 69, 2863 (1992). 7 FIG. 9. Runtimes of individual steps of the MERA optimization using single threaded operation. The top panel shows results for optimization on GPU, the bottom panel for optimization on CPU

    work page 1992

  7. [7]

    J. C. Bridgeman and C. T. Chubb, Journal of Physics A: Mathematical and Theoretical 50, 223001 (2017)

    work page 2017

  8. [8]

    Vidal, Physical Review Letters 99 (2007), 10.1103/PhysRevLett.99.220405

    G. Vidal, Physical Review Letters 99 (2007), 10.1103/PhysRevLett.99.220405

    work page doi:10.1103/physrevlett.99.220405 2007

  9. [9]

    A class of quantum many-body states that can be efficiently simulated

    G. Vidal, quant-ph/0610099 (2006), doi:10.1103/PhysRevLett.101.110501, phys. Rev. Lett. 101, 110501 (2008)

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.1103/physrevlett.101.110501 2006

  10. [10]

    Verstraete, M

    F. Verstraete, M. M. Wolf, D. Perez-Garcia, and J. I. Cirac, Physical Review Letters 96, 220601 (2006)

    work page 2006

  11. [11]

    Renormalization algorithms for Quantum-Many Body Systems in two and higher dimensions

    F. Verstraete and J. I. Cirac, arXiv:cond-mat/0407066 (2004), arXiv: cond-mat/0407066

    work page internal anchor Pith review Pith/arXiv arXiv 2004

  12. [12]

    S. R. White and R. L. Martin, The Journal of Chemical Physics 110, 4127 (1999), arXiv: cond-mat/9808118

    work page internal anchor Pith review Pith/arXiv arXiv 1999

  13. [13]

    G. K.-L. Chan, J. J. Dorando, D. Ghosh, J. Hachmann, E. Neuscamman, H. Wang, and T. Yanai, arXiv:0711.1398 [cond-mat] (2007), arXiv: 0711.1398

    work page internal anchor Pith review Pith/arXiv arXiv 2007

  14. [14]

    Tensor product methods and entanglement optimization for ab initio quantum chemistry

    S. Szalay, M. Pfeffer, V. Murg, G. Barcza, F. Verstraete, R. Schneider, and . Legeza, International Journal of Quantum Chemistry 115, 1342 (2015), arXiv: 1412.5829

    work page internal anchor Pith review Pith/arXiv arXiv 2015

  15. [15]

    Fermionic orbital optimisation in tensor network states

    C. Krumnow, L. Veis, . Legeza, and J. Eisert, Physical Review Letters 117, 210402 (2016), arXiv: 1504.00042

    work page internal anchor Pith review Pith/arXiv arXiv 2016

  16. [16]

    S. R. White and E. M. Stoudenmire, Physical Review B 99, 081110 (2019), arXiv: 1809.10258

    work page internal anchor Pith review Pith/arXiv arXiv 2019

  17. [17]

    Bauernfeind, M

    D. Bauernfeind, M. Zingl, R. Triebl, M. Aichhorn, and H. G. Evertz, Physical Review X 7, 031013 (2017)

    work page 2017

  18. [18]

    Ganahl, M

    M. Ganahl, M. Aichhorn, H. G. Evertz, P. Thunstrm, K. Held, and F. Verstraete, Physical Review B 92, 155132 (2015)

    work page 2015

  19. [19]

    Ganahl, P

    M. Ganahl, P. Thunstrm, F. Verstraete, K. Held, and H. G. Evertz, Physical Review B 90, 045144 (2014)

    work page 2014

  20. [20]

    Verstraete and J

    F. Verstraete and J. I. Cirac, Physical Review Letters 104, 190405 (2010)

    work page 2010

  21. [21]

    Haegeman, T

    J. Haegeman, T. J. Osborne, H. Verschelde, and F. Ver- straete, Physical Review Letters 110, 100402 (2013)

    work page 2013

  22. [22]

    Ganahl, J

    M. Ganahl, J. Rinc´ on, and G. Vidal, Physical Review Letters 118, 220402 (2017)

    work page 2017

  23. [23]

    E. M. Stoudenmire and D. J. Schwab, arXiv:1605.05775 [cond-mat, stat] (2016), arXiv: 1605.05775

    work page internal anchor Pith review Pith/arXiv arXiv 2016

  24. [24]

    E. M. Stoudenmire, Quantum Science and Technology 3, 034003 (2018), arXiv: 1801.00315

    work page internal anchor Pith review Pith/arXiv arXiv 2018

  25. [25]

    Number-State Preserving Tensor Networks as Classifiers for Supervised Learning

    G. Evenbly, arXiv:1905.06352 [quant-ph, stat] (2019), arXiv: 1905.06352

    work page internal anchor Pith review Pith/arXiv arXiv 1905

  26. [26]

    Swingle, Physical Review D 86, 065007 (2012)

    B. Swingle, Physical Review D 86, 065007 (2012)

    work page 2012

  27. [27]

    Causal structure of the entanglement renormalization ansatz

    C. Bny, New Journal of Physics15, 023020 (2013), arXiv: 1110.4872

    work page internal anchor Pith review Pith/arXiv arXiv 2013

  28. [28]

    Tensor Networks from Kinematic Space

    B. Czech, L. Lamprou, S. McCandlish, and J. Sully, Journal of High Energy Physics 2016 (2016), 10.1007/JHEP07(2016)100, arXiv: 1512.01548

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.1007/jhep07(2016)100 2016

  29. [29]

    N. Bao, C. Cao, S. M. Carroll, and A. Chatwin-Davies, Physical Review D 96, 123536 (2017), arXiv: 1709.03513

    work page internal anchor Pith review Pith/arXiv arXiv 2017

  30. [30]

    Geometric interpretation of the multi-scale entanglement renormalization ansatz

    A. Milsted and G. Vidal, arXiv:1812.00529 [cond- mat, physics:hep-th, physics:quant-ph] (2018), arXiv: 1812.00529

    work page internal anchor Pith review Pith/arXiv arXiv 2018

  31. [31]

    TensorNetwork on TensorFlow: A Spin Chain Application Using Tree Tensor Networks

    A. Milsted, M. Ganahl, S. Leichenauer, J. Hidary, and G. Vidal, arXiv:1905.01331 [cond-mat, physics:hep-th, physics:physics, stat] (2019), arXiv: 1905.01331

    work page internal anchor Pith review Pith/arXiv arXiv 1905

  32. [32]

    TensorNetwork for Machine Learning

    S. Efthymiou, J. Hidary, and S. Leichenauer, arXiv:1906.06329 [cond-mat, physics:physics, stat] (2019), arXiv: 1906.06329

    work page internal anchor Pith review Pith/arXiv arXiv 1906

  33. [33]

    Evenbly and G

    G. Evenbly and G. Vidal, Physical Review B 79, 144108 (2009)

    work page 2009

  34. [34]

    R. N. C. Pfeifer, G. Evenbly, and G. Vidal, Physical Review A 79, 040301 (2009)

    work page 2009

  35. [35]

    Quantum Criticality with the Multi-scale Entanglement Renormalization Ansatz

    G. Evenbly and G. Vidal, arXiv:1109.5334 [cond-mat, physics:quant-ph] (2011), arXiv: 1109.5334

    work page internal anchor Pith review Pith/arXiv arXiv 2011

  36. [36]

    www.tensors.net,

    G. Evenbly, “www.tensors.net,” (2015), Everything you need to begin your exciting journey into the world of tensor networks! 8 = ρτ hτ−1 wτ−1 wτ−1wτ−1wτ−1 uτ−1uτ−1 uτ−1 uτ−1 wτ−1 ρτ hτ−1 wτ−1 wτ−1wτ−1wτ−1 uτ−1uτ−1 uτ−1 uτ−1 wτ−1 ρτ hτ−1 wτ−1wτ−1wτ−1 uτ−1uτ−1 uτ−1 uτ−1 wτ−1 wτ−1 + + (a) ρτ hτ−1 wτ−1 uτ−1 uτ−1 uτ−1 wτ−1 wτ−1 wτ−1 wτ−1 wτ−1 = ρτ hτ−1 wτ−1 uτ...

    work page 2015

  37. [37]

    Di Francesco, P

    P. Di Francesco, P. Mathieu, and D. Senechal,Conformal Field Theory (Springer, 1997)

    work page 1997

  38. [38]

    Since the SVDs contribute only a small fraction to the total runtime, the overall runtimes were essentially unaffected by this choice

    SVDs were computed on the host CPU, which was seen to be faster than performing the SVD on GPU. Since the SVDs contribute only a small fraction to the total runtime, the overall runtimes were essentially unaffected by this choice

    work page