{"id":"f591a494-946a-422f-9906-3a3cdb19dbc5","arxiv_id":"2607.15030","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"PEPS energy gradients can be computed by implicit differentiation of characteristic equations for the contraction environment, avoiding unstable subroutine backpropagation and reducing asymptotic cost.","lead":"This paper shows how to compute gradients in projected entangled-pair state (PEPS) optimizations by solving a single linear equation per gradient, instead of differentiating through the iterative contraction algorithm. The approach is faster, numerically stable, and much simpler to implement for three common contraction methods.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Boundary MPS characteristic equations have a scaling zero mode, making ∂yF singular and the gradient ambiguous.","rationale":"The reader identified the invertibility of ∂yF as the weakest assumption, with the scalar proof in Appendix B not covering the full coupled systems. My analysis confirms and sharpens this: for the boundary MPS characteristic equations of Section IV B, there is an explicit scaling symmetry—simultaneously rescaling G_L and G_R by inverse factors leaves every equation invariant because each term is bilinear in these variables and λ_t, λ are defined as bilinear inner products. Consequently the Jacobian has a null vector, the root is not isolated, and Eq. (17) is singular. This directly threatens the central claim that the gradient is obtained from a single well-posed linear solve. The issue is fixable by adding explicit normalization conditions (e.g., fixing the trace or norm of each environment) to the characteristic equations, but this must be done and the invertibility of the resulting Jacobian verified. The reader’s CONDITIONAL verdict remains appropriate; the concern is more specific than the reader stated but does not move the verdict. I see no reason to reject the paper outright, since the framework is sound in principle and the flaw appears localized to the boundary MPS equations as presented.","tokens_in":35230,"tokens_out":13215,"duration_ms":146975,"concrete_test":"Construct the Jacobian ∂F/∂y at a VUMPS fixed point for a small boundary MPS example (e.g., D=2, χ=8) using the equations of Section IV B, and estimate its smallest singular value via a Krylov method on the normal equations. If the smallest singular value is at machine precision (≈1e-12), the zero mode is confirmed. As a cross-check, solve Eq. (17) with two different initial guesses for ˘F; if the resulting gradient differs beyond solver tolerance, the linear system is singular and the gradient is ambiguous.","verdict_should_be":"UNCHANGED","load_bearing_attack":"For the boundary MPS case, Section IV B defines characteristic equations for G_tL and G_tR (Eqs. 54-55) and for the mixed environments G_L and G_R (Eqs. 56-57). These equations are each invariant under the simultaneous rescaling (G_L, G_R) → (α G_L, α^{-1} G_R), because λ_t in Eq. (50) and λ in Eq. (58) are bilinear in G_L and G_R and hence unchanged, and every other term is also bilinear. The same holds for the top and bottom environment pairs with α and α^{-1}. Thus the full 12-variable system has at least two independent null directions at the root. The root y* is not isolated; the Jacobian ∂yF has a zero eigenvalue. Equation (17), ˘F ∂yF = ˘y, is then singular, and the gradient from Eq. (16) is not unique unless the right-hand side is orthogonal to the left null space, which is not established. The paper’s remarks about normalizing G_tL, G_tR, G_L, G_R (Sec. IV A, Eq. (59)) describe the forward fixed point, but these normalization conditions are not included in F itself. This is not merely a missing proof; the system as written is singular, so the central claim for boundary MPS is not well-defined as stated.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes an implicit differentiation framework for computing energy gradients in variational PEPS optimization. Instead of backpropagating through the iterative CTMRG or boundary-MPS contraction routines, the authors introduce an algebraic characteristic equation F(y,p)=0 whose root characterizes the converged contraction environment, and evaluate the gradient by solving a single linear system \\bar F ∂_yF = \\bar y. Explicit characteristic equations are given for C4v-symmetric CTMRG, boundary MPS, and general asymmetric CTMRG, and benchmarks compare the approach with fixed-point differentiation. The central claim is that this un-nests the gradient computation and removes gauge-induced instabilities from the differentiation of eigensolvers and SVDs.","tokens_in":35480,"tokens_out":7198,"duration_ms":85848,"significance":"If the framework is correct, it is a valuable methodological contribution: it replaces nested differentiation of iterative subroutines with one linear solve, simplifies the AD implementation, and ships with open-source code and benchmarks. The C4v and general CTMRG derivations are carefully constructed, and the numerical speedups for those cases are plausible. However, the boundary-MPS formulation, which is one of the three main examples, has a load-bearing singularity issue that is not addressed in the manuscript. The claim that the characteristic equations are well-posed for implicit differentiation is therefore not established for that section.","major_comments":[{"comment":"The characteristic system is scale-invariant. For any α, the transformation G_tL → α G_tL, G_tR → α^{-1}G_tR leaves λ_t in Eq. (50) unchanged and leaves each of Eqs. (51)–(55) satisfied at the root; the same holds for the bottom pair (G_bL, G_bR) and for the mixed pair (G_L, G_R) with λ in Eq. (58). Thus the 12-variable system has at least three independent null directions, ∂_yF is singular at the root, and Eq. (17) is not well-posed as written. The normalization statements around Eq. (59) are forward fixed-point conventions, not constraints included in F. The authors must either add explicit gauge-fixing/normalization equations to F, or prove that \\bar y lies in the range of ∂_yF and that any solution of Eq. (17) yields the same physical gradient. As it stands, the central boundary-MPS claim is not defined.","section":"§IV B, Eqs. (50)–(57)"},{"comment":"Appendix B analyzes only a scalar eigenvalue problem and explicitly shows that without a normalization/phase-fixing condition the tangent is not unique. The manuscript asserts that the analogous zero-eigenvalue elimination works for the full 7-, 12-, and 20-variable systems, but no corresponding argument is given for the boundary-MPS scale modes. This is not merely a missing proof: the equations as displayed are singular, so the application of the implicit function theorem and the linear solve in Eq. (17) requires additional hypotheses that are not stated.","section":"Appendix B and §II C, Eq. (17)"},{"comment":"The gradient accuracy benchmark in Fig. 1 compares implicit differentiation with finite differences only for C4v CTMRG and asymmetric CTMRG. The boundary-MPS section is validated only by timing comparisons in Figs. 4–5. Given the singular Jacobian identified above, a numerical correctness test for the boundary-MPS gradient is necessary to support the claim that the approach computes the correct gradient in that case.","section":"§VI, Fig. 1 and Figs. 4–5"}],"minor_comments":[{"comment":"The use of C^{*-1} as a preconditioner should mention that C* is assumed invertible; if the corner tensor can have zero or small eigenvalues, the conditioning of Eq. (17) needs discussion.","section":"§III B, Eq. (28)"},{"comment":"The notation λ_t ≡ λ_t^L = λ_t^R = λ_t^AC is introduced informally; it would help to define the normalization convention exactly once, since normalization plays a role in the boundary-MPS zero modes.","section":"§IV A, text after Eq. (42)"},{"comment":"The notation \"˘y←(˘x ∂_y x,0)\" is slightly confusing because x is a function of y; rewriting as the adjoint of the parametrization map y→x would improve clarity.","section":"Algorithm 2, line 6"}],"recommendation":"major_revision","confidential_remarks":"The boundary-MPS scale invariance is the key technical issue. It is fixable in principle by including normalization constraints in F or by proving consistency of the singular system, but without that fix the main claim for Section IV is not well-defined. I would not reject the paper outright, because the general idea and the other examples are promising, but the revision must address this point and add a boundary-MPS gradient accuracy test."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper is worth taking seriously, but the boundary MPS section has a gap that needs fixing before the method is usable as advertised. What's genuinely new: the explicit characteristic equations for boundary MPS (12 variables) and asymmetric CTMRG (20 variables), plus the null-space parameterization that removes the need to differentiate unstable subroutines. The benchmarks are honest, the code is public, and the accuracy check against finite differences is the right kind of validation. For C4v CTMRG the derivative is essentially a rephrasing of Ref. 29, so the novelty is concentrated in the other two cases—that's fine.\n\nThe soft spots, in order of seriousness. First, the abstract says \"reduce the cost\" without qualification, but Fig. 6 shows the opposite for non-symmetric CTMRG at small bond dimensions. The body states the caveat, so it's a presentation flaw, not a scientific one. Second, and more serious: I checked the boundary MPS equations (51)–(57), and I think the stress-test note is right. The equations for G_tL, G_tR and G_L, G_R are invariant under simultaneous rescaling (G_L,α; G_R,α^{-1}) because λ is bilinear in those variables. The same holds for the top and bottom environment pairs. The normalization conditions mentioned in the text—e.g., Eq. (59)—are not built into F itself. So at the root, ∂yF has at least two null directions. Equation (17) is then singular, and the gradient from Eq. (16) is not defined unless the right-hand side is orthogonal to the left null space, which the paper does not show. This is not a missing proof: the system as written is singular. The C4v and asymmetric CTMRG cases may avoid this because the corner/edge tensors are normalized inside the equations, but the boundary MPS case needs either a proof of condition (invertibility after projection), or additional normalization equations in F.\n\nAll that said, the authors know what they are doing. The derivations are careful, the limitations that are disclosed are disclosed, and the framework is modular and appealing. If the boundary MPS singularity is fixed—either by adding gauge-fixing constraints or by showing the null space is harmless—this becomes a citeable contribution. As written, the central claim for one of the three flagship examples is not well-defined.\n\nSend it to peer review with a clear ask: the authors must address the scaling degeneracy. A referee who reads the equations will catch it, and it's fixable.","headline":"Honest, useful framework with a real flaw: the boundary-MPS characteristic equations look scale-singular, so the advertised linear solve is not well-posed as written.","tokens_in":36038,"tokens_out":4944,"would_cite":false,"duration_ms":56177,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"PEPS energy gradients become one linear solve instead of nested backpropagation.","keywords":["implicit differentiation","tensor networks","PEPS","gradient optimization","CTMRG","boundary MPS","automatic differentiation","characteristic equation"],"falsifier":"Take a PEPS parameter p where the CTMRG truncation step has exactly degenerate leading eigenvalues or singular values and check whether the GMRES solution of Eq. (17) converges and matches a finite-difference gradient. A second check: scan the optimization trajectory and monitor the smallest singular value of ∂yF; if it touches zero away from phase-transition-like gaps, the core well-posedness assumption is violated.","tokens_in":35092,"feed_emoji":"🧮","tokens_out":3302,"duration_ms":34460,"temperature":0.7,"pith_summary":"The paper aims to make gradient-based optimization of projected entangled-pair states (PEPS) cheaper and more stable. It argues that the energy gradient does not need to be computed by differentiating the iterative contraction algorithm (CTMRG or boundary MPS) subroutine by subroutine. Instead, the converged contraction environment can be characterized as the root of a purely algebraic equation F(y,p)=0, and the gradient follows from solving a single linear system. This un-nests the gradient problem, removes instabilities caused by differentiating eigensolvers and SVDs, and makes implementation easier because only tensor contractions need to be differentiated.","feed_headline":"PEPS energy gradient from a single linear solve","feed_subtitle":"Implicit differentiation of the contraction environment removes nested backpropagation and gauge instabilities.","key_machinery":"The key object is the characteristic equation F(y,p)=0 — a set of tensor-contraction equations that implicitly define the converged contraction environment, augmented with the defining relations of the truncated eigenspace or SVD and parametrized by variables y that live only in the null spaces of the fixed-point isometries. This parametrization removes the gauge-induced zero modes and divergent terms: variations along the kept subspace (skew-Hermitian ω terms) are dropped, while the corner/edge tensors are treated as generic complex Hermitian matrices in the reverse pass. The adjoint linear system Fbar ∂yF = ybar is then solved once with a Krylov solver; evaluating ∂yF and ∂pF requires only","core_discovery":"On the paper's own terms, the central claim is that the derivative de/dp = xbar ∂p x + pbar can be evaluated without ever differentiating the contraction algorithm's inner subroutines. One writes the converged environment as the unique root y*(p) of a characteristic equation F(y,p)=0 chosen so that it is purely algebraic and so that its variables parametrize only the physical variations of the environment — isometry variations along the fixed-point subspace are removed by construction. Differentiating F at the root gives the adjoint equation Fbar ∂yF = ybar, and the desired vector–Jacobian product is −Fbar ∂pF. The paper works out explicit characteristic equations for the three most common c","pith_inferences":["The same 'characteristic equation plus one linear solve' recipe likely transfers to other tensor network algorithms beyond PEPS, such as time evolution or excitation ansätze, whenever the fixed point can be characterized algebraically.","The alternative boundary-MPS decomposition into three smaller linear problems suggests a general trade-off: one large linear solve can be exchanged for smaller parallel ones, which may matter at very large environment bond dimensions.","Because the conditioning of the linear system is set by the spectral gap between kept and truncated eigenvalues, near a phase transition where that gap closes the gradient may become ill-conditioned — but this shows up visibly as solver non-convergence rather than hidden in AD pullbacks.","A direct test of the zero-mode concern would be to intentionally construct a degenerate truncation subspace and check whether GMRES still converges for Eq. (17)."],"forward_implications":["Gradient evaluation for PEPS becomes a single un-nested linear solve, asymptotically subleading to the contraction itself in the tested settings.","No subroutine of the contraction algorithm (eigensolvers, SVDs, QR) needs to be made differentiable; only tensor contractions are differentiated.","Numerical instabilities from degenerate eigenvalues or singular values in the pullback are removed by the gauge-aware parametrization, since the divergent in-subspace rotations are not part of the variables.","The same characteristic-equation approach applies on top of any existing CTMRG or boundary MPS implementation, including QR-based and pulling-through variants, by mapping the forward environment to the root y*.","For poorly converged environments, gradient accuracy matches black-box reverse-mode AD when the linear solver tolerance is adjusted accordingly."],"fun_headline_variants":["PEPS energy gradient via a single linear solve","Implicit differentiation stabilizes PEPS gradients","Avoid nested backprop for PEPS gradients","PEPS gradient from one characteristic equation","Cost and instability cut in PEPS gradient"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The load-bearing premise is that the algebraic equations F(y,p)=0 have a unique, isolated root at the contraction fixed point and that the Jacobian ∂yF is invertible there for every parameter p encountered during optimization — if a residual gauge or normalization freedom leaves a zero mode, the single linear solve fails and the gradient is undefined.","fun_headline_variants_meta":{"raw":{"variants":["PEPS energy gradient via a single linear solve","Implicit differentiation stabilizes PEPS gradients","Avoid nested backprop for PEPS gradients","PEPS gradient from one characteristic equation","Cost and instability cut in PEPS gradient"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000221,"raw_usage":{"total_tokens":1261,"prompt_tokens":693,"completion_tokens":568,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":437,"completion_tokens_details":{"reasoning_tokens":502}},"tokens_in":437,"tokens_out":568,"duration_ms":6071,"temperature":1.0,"reasoning_tokens":502,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-02T00:21:25.598530+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a PEPS parameter p where the CTMRG truncation step has exactly degenerate leading eigenvalues or singular values and check whether the GMRES solution of Eq. (17) converges and matches a finite-difference gradient. A second check: scan the optimization trajectory and monitor the smallest singular value of ∂yF; if it touches zero away from phase-transition-like gaps, the core well-posedness assumption is violated.","supporting_citations":[],"review_version":1}