Pith. sign in

REVIEW 4 major objections 4 minor 28 references

The paper resolves the open problem of identifying the perturbations that finite bond-dimension MPS approximations actually induce in critical systems, and shows they can differ substantially from CFT predictions.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-02 00:03 UTC pith:L3MA3KLG

load-bearing objection A solid implicit-differentiation paper for 2D tensor networks; trust the method, treat the two-Ising conclusion as unproven until the norm choice is justified and the data are shown. the 4 major comments →

arxiv 2607.15124 v2 pith:L3MA3KLG submitted 2026-07-16 quant-ph cond-mat.str-elhep-lat

On the origin of finite entanglement scaling

Luke Hodgkiss , Laurens Lootens , Atsushi Ueda , Bram Vanhecke , Frank Verstraete This is my paper
classification quant-ph cond-mat.str-elhep-lat
keywords finite entanglement scalingmatrix product statesimplicit differentiationtensor networksconformal field theorySchatten 4-normcritical systemsPEPS optimization
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The paper identifies, for the first time, the actual perturbation that a finite bond-dimension matrix product state approximation imposes on a critical system: it is the direction in the local tensor that best improves the fidelity of the MPS approximation, measured by a Schatten 4-norm. For a single Ising chain this matches the CFT expectation—the most relevant spin operator. But for two decoupled critical Ising chains, the induced perturbation has only about two-thirds overlap with the CFT-predicted operator and contains symmetry-breaking inter-chain couplings. This means finite entanglement scaling is not a simple application of CFT scaling, and it opens the way to a purely MPS-based construction of RG flows and to more stable PEPS optimization.

Core claim

The central discovery is a method and a result. The method: implicit differentiation of the fixed-point equations of a 2D tensor network contraction (pulling-through or CTMRG), preconditioned by embedding tensors in their natural environment, yields forward and backward derivatives without automatic differentiation. The result: the effective perturbation induced by a finite bond dimension χ is the direction δT that maximally decreases the Schatten 4-norm fidelity (Eq. 13). In the case of two decoupled critical Ising chains, this optimal perturbation does not converge to the pure spin operator predicted by CFT; instead its overlap with the spin operator saturates around 2/3 and a significant

What carries the argument

The paper's argument rests on the fixed-point equations that characterise the optimal MPS approximation of the boundary vector (left-canonical MPS tensor AL, corner tensor X, eigenvalue λ). Differentiating these equations gives a large sparse linear system for the implicit derivatives; preconditioning by the natural environment (using X to make the local metric Euclidean) makes the system well-conditioned. The relevant perturbation is selected by minimising a Schatten 4-norm that the paper posits as the natural analogue of the MPS variance in the corner-transfer-matrix setting.

Load-bearing premise

The paper identifies the perturbation induced by the MPS with the direction that optimally improves a Schatten 4-norm fidelity (Eq. 13), but this norm is posited, not derived from the objective that MPS algorithms actually minimize.

What would settle it

Run a variational MPS optimisation directly on the transfer matrix of two decoupled critical Ising chains at a fixed bond dimension, extract the actual perturbation to the local tensor by comparing fixed-point tensors at consecutive bond dimensions, and check whether the overlap of that perturbation with the spin operator is indeed about 2/3 as predicted; if the overlap saturates to 1, the Schatten-4-norm premise fails.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Finite entanglement scaling can be interpreted quantitatively: the finite-χ perturbation is a specific direction in the local tensor space, computable by the new solver.
  • The method gives a purely MPS-based way to construct RG flows, by following the direction of maximal correlation-length increase or decrease.
  • Implicit differentiation of 2D tensor networks provides a primitive for stable variational PEPS optimization that avoids the instabilities of automatic differentiation.
  • The result overturns the assumption that finite-entanglement scaling is governed only by the most relevant CFT operator; multi-chain simulations must account for entangling perturbations.

Where Pith is reading between the lines

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

  • If the 4-norm identification is accepted, the same machinery could be applied to gapped systems, characterising finite-χ effects away from criticality where CFT does not apply.
  • The presence of symmetry-breaking inter-chain couplings suggests that simulations of coupled critical chains at finite bond dimension may mis-estimate universal data unless these couplings are explicitly controlled.
  • A testable extension is to measure the overlap between the predicted T' and the actual change in the fixed-point MPS when the bond dimension is increased by one, for various critical models.
  • The condition-number improvement from natural-environment preconditioning might carry over to other tensor-network optimisation problems, making implicit gradients a general replacement for automatic differentiation.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

4 major / 4 minor

Summary. The paper develops an implicit-differentiation framework for infinite two-dimensional tensor networks with matrix-product-state (MPS) boundary approximations. The authors construct a preconditioned, full-rank linear system (Eqs. 5–7) whose solution gives forward and backward derivatives of the MPS tensors with respect to the local tensor, and they propose this as a primitive for PEPS optimization that avoids the instabilities of automatic differentiation. They then define a Schatten 4-norm fidelity measure (Eq. 13) and identify the perturbation of the local tensor that most decreases this measure with the perturbation actually induced by finite bond dimension. For the critical Ising model, the resulting perturbation overlaps strongly with the CFT spin operator (Table I). For two decoupled critical Ising chains, the same calculation reportedly yields an overlap of only about 2/3 with the CFT-predicted spin-sum operator, with significant symmetry-breaking inter-chain couplings, leading the authors to conclude that finite entanglement scaling is not a simple application of CFT scaling.

Significance. If the framework is correct, the implicit-differentiation solver is a valuable technical contribution: it provides a way to compute stable derivatives of tensor-network contractions without automatic differentiation, potentially enabling more robust PEPS optimization. The single-Ising benchmark in Table I is a concrete, falsifiable check that supports the preconditioned derivative construction and the identification of the most relevant perturbation with the spin operator. The two-Ising result, if fully substantiated, would challenge a standard assumption in finite-entanglement scaling. The paper also correctly emphasizes the distinction between correlation-length gradients and fidelity gradients, which is conceptually useful. However, the central claim about the 'actual' perturbation induced by MPS algorithms rests on a posited norm choice and is not yet backed by numerical data for the headline two-chain case, so the significance of the main physical claim remains conditional.

major comments (4)
  1. [Section 'Relevant perturbations induced by χ', Eq. (13)] The central identification of the MPS-induced perturbation with the direction that maximally decreases the Schatten 4-norm is introduced as a posit ('We posit that this 4-norm ... is the natural corner transfer matrix analogue of the variance'), not derived from the pulling-through, CTMRG, or variational MPS algorithm that actually produces the finite-χ fixed point. Since the entire two-Ising conclusion depends on this identification, this is load-bearing. The authors should either derive the objective from the algorithm, or reframe the claim as characterizing the optimal perturbation for a chosen fidelity measure rather than the perturbation actually induced by an MPS algorithm. A concrete sensitivity test, e.g. comparing the predicted perturbation direction with the difference between finite-χ and exact fixed-point tensors, would also strengthen the claim.
  2. [Section 'The curious case of two Ising models'] The headline two-Ising result is stated only qualitatively: the overlap with the spin operator 'hovers around 2/3, even when chi gets large,' and overlaps with inter-chain terms 'prevail when increasing the bond dimension.' No numerical table, no list of χ values, no operator-basis decomposition, and no error estimates are provided. This makes the central claim impossible to check. The paper should include a table analogous to Table I, with the overlap of the gradient with the spin-sum operator, the leading inter-chain terms, and the residual overlap, as a function of χ, together with the condition numbers of M′ for those runs.
  3. [Equations (7), (12), and 'The curious case of two Ising models'] The reliability of the solver rests on the assertion that the preconditioning makes M′ well-conditioned: the text states 'We have tested extensively that this preconditioning leads to an M′ with a small condition number,' but no condition-number data are reported. Since the backward derivative is computed by solving a linear system with M′, conditioning is directly load-bearing for both the Ising check and the two-Ising prediction. Please report condition numbers (or spectra) for the systems used, at least for representative χ and d.
  4. [Section 'The incurable case of the Ising model' and Eq. (13)] The paper notes that MPS algorithms at criticality tend to break the Z2 symmetry, while the derivative framework assumes a D4-symmetric real fixed point. For the two-Ising case, the symmetric fixed point may not be the one selected by actual MPS algorithms; symmetry-broken fixed points could lead to a different induced perturbation. This is not just a technical caveat: it could change the numerical values of the overlaps and the operator content. The authors should either compute the perturbation at the symmetry-broken fixed point or give a convincing argument that the symmetric fixed point controls the local perturbation relevant for finite-entanglement scaling.
minor comments (4)
  1. [Eq. (3)] The parametrization dAL = A⊥L dQ X^{-2} is introduced with a reference to [26] but the notation is not fully explained in the main text; please clarify the dimensions and the role of X^{-2} as a natural-environment preconditioner.
  2. [Table I] Table I would benefit from a caption stating the norm used for the overlap and the precision/error bars of the reported values. The text also says 'exactly the same scaling is observed' for the fidelity optimization, but no data are shown for that case.
  3. [Eq. (16)] The displayed equation for the correlation-length gradient appears incomplete: it contains only '(vL) dAL AL (vR)' without the full contraction terms and the normalization condition. Please provide the complete expression.
  4. [Text around Eq. (5)] The dimension formula for dT says 'd(d+1)/2 + (d−1)d(d+1)(d+2)/8 − (d(d−1)/2 +1) components' but the preceding text lists d(d+1)/2 + (d−1)d(d+1)(d+2)/8 linearly independent components; the subtraction is explained only obliquely. Please state explicitly which gauge and 'gradient proportional to T' degrees of freedom are removed.

Circularity Check

1 steps flagged

Partial definitional circularity: the two-Ising 'MPS-induced perturbation' is the gradient of a posited Schatten-4 norm, not a quantity derived from the actual algorithm's objective; the single-Ising benchmark supplies independent support.

specific steps
  1. self definitional [Section 'Relevant perturbations induced by χ', Eqs. (13)–(15) and the following paragraph]
    "We posit that this 4-norm, with the conditioning chosen in this particular way, is the natural corner transfer matrix analogue of the variance in usual MPS-based algorithms. ... If we set E(T, A_l, X, λ) from Eq. (8) equal to this nor, we can easily determine the direction dT in which it decreases maximally. This answers the question of determining the perturbation that is induced by working with a finite bond dimension: MPS algorithms effectively induce this perturbation as it leads to the best possible variational MPS approximation."

    The paper defines the 'perturbation induced by MPS' as the steepest-descent direction of the Schatten 4-norm (Eq. 13), a norm it posits rather than derives from the CTMRG/pulling-through optimization that actually produces the MPS. The two-Ising result—overlap ~2/3 with the CFT spin sum and significant inter-chain couplings—is exactly the gradient of that posited norm in the posited natural metric (Eq. 12). Thus the headline prediction is forced by the choice of cost function and metric; a different norm or metric would likely change the overlap. The single-Ising benchmark (Table I) gives independent external support, so this is partial circularity rather than a fully tautological derivation.

full rationale

The paper's implicit-differentiation algorithm is self-contained and the single-Ising table provides an external consistency check against the CFT spin operator, lowering the circularity burden. However, the paper's central two-Ising 'prediction' is not an independent first-principles result: it is computed as the derivative of a Schatten 4-norm that the authors explicitly posit ('We posit...'), and the natural metric used for overlaps is likewise chosen. The claimed 'actual perturbation induced by MPS' is therefore, by construction, the steepest-descent direction of that chosen cost function. A norm-sensitivity test or a derivation from the actual fixed-point optimization would be needed to establish that the reported 2/3 overlap and inter-chain couplings are not artifacts of the norm choice. The self-citations ([22], [23]) are not load-bearing for the central derivation; the companion paper [22] provides benchmarks for the derivative solver, not for the norm identification. Score 4 reflects partial definitional circularity offset by the single-Ising benchmark.

Axiom & Free-Parameter Ledger

1 free parameters · 5 axioms · 0 invented entities

The framework relies on the standard pulling-through fixed-point equations and on two posited modeling choices: the Schatten 4-norm as the fidelity measure and the identification of the maximal-fidelity-improvement direction with the perturbation actually induced by MPS. No data are fitted, but the norm exponent is ad hoc, and the empirical claim that the preconditioned linear system is well-conditioned is asserted without supporting data.

free parameters (1)
  • Schatten 4-norm exponent = 4
    The perturbation induced by MPS is defined as the gradient of the Schatten 4-norm (Eq. 13); p=4 is posited as 'the natural corner transfer matrix analogue of the variance' rather than derived from the MPS algorithm, so the central result depends on this hand-chosen value.
axioms (5)
  • domain assumption The leading eigenvector of the transfer matrix exists, is real, and does not break the D4 lattice symmetry (possibly after blocking).
    The entire derivative computation is restricted to this setting; the paper itself notes MPS algorithms at criticality tend to break Z2 symmetry, so the symmetric fixed point may not be the one actual algorithms find (Section 'Implicit Derivatives', first paragraph).
  • domain assumption The optimal MPS fixed point satisfies the pulling-through fixed-point equations (Eqs. 1-2) with the stated gauge and normalization Tr(X^4)=1.
    This turns the MPS approximation problem into a fixed-point equation, enabling implicit differentiation; it is a standard but nontrivial assumption about the algorithm's output (Eqs. 1-2).
  • ad hoc to paper The Schatten 4-norm in Eq. (13) is the natural measure of MPS approximation quality: an MPS is an exact eigenstate iff the norm is zero, and the perturbation decreasing it maximally is the perturbation induced by finite bond dimension.
    Posited in the text ('we posit that this 4-norm ... is the natural corner transfer matrix analogue of the variance'), not derived from the actual optimization objective of pulling-through or CTMRG (Eq. 13 and following paragraph).
  • ad hoc to paper The perturbation T' that provides the optimal improvement to the fidelity of the MPS approximation at fixed bond dimension is the perturbation actually induced by MPS algorithms.
    This identification connects the gradient computation to the physical claim; it is an assumption about what MPS algorithms effectively do and is the load-bearing step for the two-Ising conclusion (Section 'Relevant perturbations induced by χ', first paragraph).
  • ad hoc to paper The preconditioned linear system M' is full rank and has small condition number, so LSQR/LSMR solutions are reliable.
    Stated as 'tested extensively' but no proof or numerical condition numbers are supplied; the backward-derivative results depend on this (after Eqs. 7 and 11).

pith-pipeline@v1.3.0-alltime-deepseek · 7851 in / 13826 out tokens · 185355 ms · 2026-08-02T00:03:48.524445+00:00 · methodology

0 comments
read the original abstract

The concept of finite entanglement scaling forms one of the pillars on which the tensor network ecosystem is built. In this paper, we resolve the open problem of determining the actual perturbations induced by matrix product state approximations of critical systems, and we demonstrate that these can be quite different than the ones predicted by conformal field theory. To that aim, we develop a sparse linear solver to calculate the forward and backward derivatives of 2-dimensional tensor networks with respect to their defining parameters in an implicit way. This algorithm is of independent interest as it provides a primitive for the variational optimization of projected entangled pair states that circumvents the instabilities plaguing traditional automatic differentiation methods.

Figures

Figures reproduced from arXiv: 2607.15124 by Atsushi Ueda, Bram Vanhecke, Frank Verstraete, Laurens Lootens, Luke Hodgkiss.

Figure 1
Figure 1. Figure 1: FIG. 1. Renormalisation group flows for the 2D classical Ising [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗

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

28 extracted references · 9 linked inside Pith

  1. [1]

    Schollw¨ ock, The density-matrix renormalization group in the age of matrix product states, Annals of physics 326, 96 (2011)

    U. Schollw¨ ock, The density-matrix renormalization group in the age of matrix product states, Annals of physics 326, 96 (2011)

    2011

  2. [2]

    Haegeman and F

    J. Haegeman and F. Verstraete, Diagonalizing transfer matrices and matrix product operators: A medley of ex- act and computational methods, Annual Review of Con- densed Matter Physics8, 355–406 (2017)

    2017

  3. [3]

    Verstraete and J

    F. Verstraete and J. I. Cirac, Renormalization al- gorithms for quantum-many body systems in two and higher dimensions (2004), arXiv:cond-mat/0407066 [cond-mat.str-el]

    Pith/arXiv arXiv 2004

  4. [4]

    Vanderstraeten, J

    L. Vanderstraeten, J. Haegeman, P. Corboz, and F. Ver- straete, Gradient methods for variational optimization of projected entangled-pair states, Phys. Rev. B94, 155123 (2016)

    2016

  5. [5]

    Corboz, Variational optimization with infinite pro- jected entangled-pair states, Phys

    P. Corboz, Variational optimization with infinite pro- jected entangled-pair states, Phys. Rev. B94, 035133 (2016)

    2016

  6. [7]

    Naumann, E

    J. Naumann, E. L. Weerda, M. Rizzi, J. Eisert, and P. Schmoll, An introduction to infinite projected entangled-pair state methods for variational ground state simulations using automatic differentiation, SciPost Phys. Lect. Notes , 86 (2024)

    2024

  7. [8]

    Evenbly and G

    G. Evenbly and G. Vidal, Local scale transformations on the lattice with tensor network renormalization, Phys. Rev. Lett.116, 040401 (2016)

    2016

  8. [9]

    Yang, Z.-C

    S. Yang, Z.-C. Gu, and X.-G. Wen, Loop optimization for tensor network renormalization, Phys. Rev. Lett.118, 110504 (2017)

    2017

  9. [10]

    M. Bal, M. Mari¨ en, J. Haegeman, and F. Verstraete, Renormalization group flows of hamiltonians using tensor networks, Phys. Rev. Lett.118, 250602 (2017)

    2017

  10. [11]

    Vanthilt, A

    V. Vanthilt, A. Naravane, C. Meng, and A. Ueda, A prac- tical introduction to tensor network renormalization with tnrkit.jl (2026), arXiv:2604.06922 [cond-mat.str-el]

    Pith/arXiv arXiv 2026

  11. [12]

    A. A. Belavin, A. M. Polyakov, and A. B. Zamolodchikov, Infinite Conformal Symmetry in Two-Dimensional Quan- tum Field Theory, Nucl. Phys. B241, 333 (1984)

    1984

  12. [13]

    Cardy,Scaling and renormalization in statistical physics, Vol

    J. Cardy,Scaling and renormalization in statistical physics, Vol. 5 (Cambridge university press, 1996)

    1996

  13. [14]

    Nishino, K

    T. Nishino, K. Okunishi, and M. Kikuchi, Numerical renormalization group at criticality, Physics Letters A 213, 69 (1996)

    1996

  14. [15]

    Tagliacozzo, T

    L. Tagliacozzo, T. R. de Oliveira, S. Iblisdir, and J. I. Latorre, Scaling of entanglement support for Ma- trix Product States, Phys. Rev. B78, 024410 (2008), arXiv:0712.1976 [cond-mat.stat-mech]

    Pith/arXiv arXiv 2008

  15. [16]

    Pollmann, S

    F. Pollmann, S. Mukerjee, A. M. Turner, and J. E. Moore, Theory of finite-entanglement scaling at one- dimensional quantum critical points, Phys. Rev. Lett. 102, 255701 (2009)

    2009

  16. [17]

    Pirvu, G

    B. Pirvu, G. Vidal, F. Verstraete, and L. Tagliacozzo, Matrix product states for critical spin chains: Finite- size versus finite-entanglement scaling, Phys. Rev. B86, 075117 (2012), arXiv:1204.3934 [cond-mat.stat-mech]

    Pith/arXiv arXiv 2012

  17. [18]

    Zauner, D

    V. Zauner, D. Draxler, L. Vanderstraeten, M. Degroote, J. Haegeman, M. M. Rams, V. Stojevic, N. Schuch, and F. Verstraete, Transfer Matrices and Excitations with Matrix Product States, New J. Phys.17, 053002 (2015), arXiv:1408.5140 [quant-ph]

    Pith/arXiv arXiv 2015

  18. [19]

    M. M. Rams, P. Czarnik, and L. Cincio, Precise ex- trapolation of the correlation function asymptotics in uniform tensor network states with application to the Bose-Hubbard and XXZ models, Phys. Rev. X8, 041033 (2018), arXiv:1801.08554 [cond-mat.str-el]

    Pith/arXiv arXiv 2018

  19. [20]

    Vanhecke, J

    B. Vanhecke, J. Haegeman, K. Van Acoleyen, L. Vander- straeten, and F. Verstraete, Scaling hypothesis for matrix product states, Phys. Rev. Lett.123, 250604 (2019). 7

    2019

  20. [21]

    J. T. Schneider, A. Ueda, Y. Liu, A. M. L¨ auchli, M. Os- hikawa, and L. Tagliacozzo, Self-congruent point in crit- ical matrix product states: An effective field theory for finite-entanglement scaling, SciPost Phys.18, 142 (2025), arXiv:2411.03954 [cond-mat.stat-mech]

    Pith/arXiv arXiv 2025

  21. [22]

    and includes benchmarks for PEPS optimizations. Indeed, combined with a variational pulling through al- gorithm, the new solver of the backwards derivative can serve as an efficient primitive for state-of-the-art PEPS optimization algorithms without the need for memory- intensive automatic differentiation subroutines. arXiv:2607.15124v2 [quant-ph] 25 Jul ...

    Pith/arXiv arXiv 2026

  22. [23]

    Burgelman, A

    L. Burgelman, A. Francuz, P. Brehmer, L. Devos, J. Haegeman, F. Verstraete, and B. Vanhecke, Im- plicit differentiation of tensor network algorithms (2026), arXiv:2607.15030 [quant-ph]

    Pith/arXiv arXiv 2026

  23. [24]

    Lootens, C

    L. Lootens, C. Delcamp, and F. Verstraete, Entangle- ment and the density matrix renormalization group in the generalized Landau paradigm, Nature Physics21, 1657 (2025), arXiv:2408.06334 [quant-ph]

    arXiv 2025

  24. [25]

    M. T. Fishman, L. Vanderstraeten, V. Zauner-Stauber, J. Haegeman, and F. Verstraete, Faster methods for con- tracting infinite two-dimensional tensor networks, Phys- ical Review B98, 10.1103/physrevb.98.235148 (2018)

    doi:10.1103/physrevb.98.235148 2018

  25. [26]

    Nishino and K

    T. Nishino and K. Okunishi, Corner transfer matrix renormalization group method, Journal of the Physical Society of Japan65, 891 (1996)

    1996

  26. [27]

    Vanderstraeten, J

    L. Vanderstraeten, J. Haegeman, and F. Verstraete, Tangent-space methods for uniform matrix product states, SciPost Physics Lecture Notes 10.21468/scipost- physlectnotes.7 (2019)

    doi:10.21468/scipost- 2019

  27. [28]

    Vanderstraeten, J

    L. Vanderstraeten, J. Haegeman, P. Corboz, and F. Ver- straete, Gradient methods for variational optimization of projected entangled-pair states, Physical Review B94, 10.1103/physrevb.94.155123 (2016)

    doi:10.1103/physrevb.94.155123 2016

  28. [29]

    Liao, J.-G

    H.-J. Liao, J.-G. Liu, L. Wang, and T. Xiang, Differ- entiable programming tensor networks, Phys. Rev. X9, 031041 (2019)

    2019