{"id":"057c0318-606c-485e-b8e2-337bbaf9ea64","arxiv_id":"2412.13758","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A 3D entanglement-filtered tensor-network RG produces a critical fixed-point tensor for the cubic-lattice Ising model, with scaling dimensions that can match accepted values to within about 0.1-0.4%.","lead":"This paper presents a new way to coarse-grain a 3D lattice model using tensor networks with an added entanglement filtering step, and applies it to the cubic-lattice Ising model. If valid, the method could extract the full universal content of 3D critical points, not just a handful of critical exponents.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"No systematic χ-convergence shown: best-case scaling-dimension errors occur at isolated 'magical' bond dimensions, so 'well-controlled approximations' is not established by the presented data.","rationale":"The reader's verdict is CONDITIONAL, with the primary condition being a proof of the transposition trick. I agree that proof is needed, but I find a more load-bearing concern in the numerical evidence: the scaling-dimension estimates do not converge systematically with bond dimension. The paper's own Table I and Section VII describe the best cases as 'magical', and the errors fluctuate widely across χ. The stated RG errors (2–6%) refer to truncation errors in the block-tensor map, not to the actual observable errors in xσ and xϵ; the relationship between these error measures is never established. Even if the transposition trick is proven exact, the method lacks the key property advertised in the title and abstract: well-controlled approximations. Therefore the central quantitative claims (0.4% and 0.1% errors) should be read as selected best cases, and the method should be considered CONDITIONAL pending a demonstration of systematic improvement with bond dimension. The reader's conditions included this requirement, but their 'weakest_assumption' focused on the transposition trick; my concern is partly overlapping (both are conditions for the central claim) but distinct in that it targets the absence of controlled convergence regardless of the trick's validity.","tokens_in":12293,"tokens_out":3630,"duration_ms":33634,"concrete_test":"Using the published code (Ref. [44]), run the EF-RG for the 3D Ising model at χ=10, 12, and 16 (and rerun χ=6,8,11,14) with hyperparameters chosen by the paper's rule of thumb (χi=χ^1.5, χii=χ^2, and χs, χm as in Table I). For each χ, compute xσ, xϵ and the RG-error curves. If the scaling-dimension errors do not monotonically decrease toward the conformal-bootstrap values and the fixed-point tensor convergence ‖A^(n+1)−A^(n)‖ does not systematically improve, then the claim of well-controlled, systematically improvable approximations is unsupported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that the EF-enhanced block-tensor map yields 'well-controlled approximations' and scaling dimensions with 0.4% and 0.1% relative errors in the best case. Table I and Section VII show the opposite of controlled convergence: for χ=6,8,11,14, the errors for xσ are 5–8%, 4–6%, 3–6%, 0.4–0.5% and for xϵ are 0.1–1%, 4–5%, 1–6%, 2–4%. The best xϵ occurs at the smallest χ=6 and the best xσ at χ=14; errors do not monotonically decrease, and the paper explicitly calls χ=14 'magical'. The RG errors that decrease to ~2% are projective truncation errors in intermediate tensors, not errors in the final scaling dimensions; no argument connects these two error measures. Thus the method is not demonstrated to be systematically improvable, and the headline accuracies are selected best cases, not controlled extrapolations. This issue is independent of the transposition-trick proof, and it directly undermines the 'well-controlled' and 'systematically-improvable' wording of the central claim.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a three-dimensional real-space renormalization group (RG) scheme for classical lattice models, combining a block-tensor (HOTRG-like) transformation with an entanglement filtering (EF) step. The EF step is designed to be graph-independent and to respect lattice reflection symmetry via a transposition trick. The method is benchmarked on the cubic-lattice Ising model. The authors report that the RG errors of the projective truncations are reduced from over 20% (for plain HOTRG) to about 2-6% with the EF enhancement, that a critical fixed-point tensor is obtained with adjacent-step tensor differences of order 4e-3 after gauge fixing, and that scaling dimensions x_sigma and x_epsilon extracted from the linearized RG map have relative errors as low as 0.4% and 0.1% in the best cases. The paper emphasizes that a fixed-point tensor contains much more information than few observables and argues that the method is a promising systematically improvable real-space RG in 3D.","tokens_in":12597,"tokens_out":2998,"duration_ms":28935,"significance":"If the central claims hold, the paper would be an important step toward quantitative real-space RG in three dimensions: it demonstrates a concrete entanglement-filtering construction that is graph-independent and symmetry-aware, provides a numerical fixed-point tensor for the 3D Ising model, reports substantial improvement over plain HOTRG, and makes the code publicly available. The scaling-dimension extraction from the linearized RG map is methodologically interesting and, if controlled, would complement conformal bootstrap and Monte Carlo results. However, the significance as currently stated depends crucially on the 'well-controlled' and 'systematically improvable' claims, which the presented data do not yet establish. The paper is transparent about the lack of clear improvement with bond dimension, but this transparency also exposes the gap between the headline best-case errors and a controlled convergence statement.","major_comments":[{"comment":"The central claim that the proposed RG has 'well-controlled approximations' and is 'systematically improvable' is not supported by the data in Table I and Section VII. The relative errors for x_sigma are 5-8% at chi=6, 4-6% at chi=8, 3-6% at chi=11, and 0.4-0.5% at chi=14, while the errors for x_epsilon are 0.1-1%, 4-5%, 1-6%, and 2-4% for the same sequence. The best x_epsilon error occurs at the smallest bond dimension (chi=6) and the best x_sigma error at chi=14, with no monotonic trend; Section VII itself states there is 'no clear improvement when χ increases from 6 to 14' and refers to chi=14 as a 'magical' bond dimension. The RG errors that do decrease to about 2% are the projective truncation errors of the intermediate tensors (Section VI, Figure 3), not the errors in the scaling dimensions, and no argument is given that connects these two error measures. To support the headline claim, the authors should either show a systematic chi-convergence of the scaling dimensions (e.g., a sequence of chi values with decreasing error bars) or substantially soften the wording to present the accuracies as selected best cases rather than controlled extrapolations.","section":"§VII and Table I"},{"comment":"The transposition trick that imposes lattice reflection symmetry and is used to define the cube-filtering approximation in Eq. (9a) is asserted to preserve the partition function exactly for reflection-symmetric models, but no proof is given. Appendix A says 'we will expound how to exploit lattice reflection symmetry in a TNRG setting in a coming paper' and only states that 'one can show' the partition function is invariant. This is load-bearing because the filtering matrices sx, sy, sz are optimized to maximize the overlap between the filtered state and the transposed target state; if the transposition identity is not exact, the optimized filtering matrices solve the wrong approximation problem and the fixed-point tensor and scaling dimensions inherit an uncontrolled bias. The authors should either include a complete proof of the partition-function invariance in this paper, or explicitly state the approximation status and provide a numerical test of the transposition step (for example, by comparing the partition function before and after the transposition on small finite lattices).","section":"§V, Eq. (9a); Appendix A"},{"comment":"The demonstration of a fixed-point tensor relies on the Frobenius norm difference between adjacent RG steps reaching about 4e-3 at a single finite bond dimension (chi=8). While this is an improvement over HOTRG, it does not by itself constitute a controlled demonstration of a critical fixed point: the norm difference is not extrapolated to chi to infinity, and the dependence of the fixed-point tensor on chi is not shown. Since the subsequent extraction of scaling dimensions requires a genuine fixed point, the authors should report the chi-dependence of the fixed-point convergence and of the extracted scaling dimensions, or explicitly discuss the finite-bond-dimension uncertainty in the reported values.","section":"§VI, Figure 6"}],"minor_comments":[{"comment":"There are a few typographical issues: 'tensor-network representation' is followed by 'tenors' in the sentence about the partition function, and the phrase 'a nature metric' should read 'a natural metric'.","section":"§II"},{"comment":"The sentence 'near the critical fixed point, the RG errors are reduced from more than 20% to about 6%' should clarify which bond dimension is used, since Figure 3 shows this for chi=6 and chi_s=chi_m=4, while nearby text discusses chi up to 22.","section":"§VI"},{"comment":"The word 'magical' to describe chi=14 is informal and, more importantly, signals the absence of a systematic trend. The authors should either explain the origin of the particularly good result at chi=14 or remove the term and replace it with a quantitative description.","section":"§VII"},{"comment":"The hyperparameter choice chi_i = chi^1.5 and chi_ii = chi^2 is described as a 'rule of thumb' from numerical experiments; it would strengthen the paper to show at least one example demonstrating that these choices indeed make the inner-leg projective truncation errors smaller than the outer-leg errors, as claimed.","section":"Appendix C"}],"recommendation":"major_revision","confidential_remarks":"The paper is likely to be of interest to the tensor-network RG community, and the release of the code is a positive feature. My main concern for the editor is that the abstract and Section VII advertise 'well-controlled approximations' and 'systematically improvable', but Table I shows errors that fluctuate with bond dimension and the best values are selected from different chi. The authors are honest about the lack of monotonic improvement, yet the framing still overstates what is demonstrated. A revision that either provides systematic convergence data or clearly re-scopes the claims to 'best-case accuracy at finite bond dimension' would make the contribution sound. The missing proof of the transposition trick is a second issue that should be addressed, at minimum by a numerical check."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's the short version. The genuinely new thing is a graph-independent cube entanglement filtering scheme in 3D, using lattice reflection symmetry to reduce the filtering matrices from 24 to 3, and it actually produces a critical fixed-point tensor for the cubic-lattice Ising model where plain HOTRG fails. That is real progress: the gauge-fixed fixed-point norm difference reaches about 4e-3 and the RG errors drop from more than 30% to 2-6%. But the headline phrase 'well-controlled approximations' is not yet supported by the scaling-dimension data. Table I shows errors for x_sigma and x_epsilon fluctuate with bond dimension: the best x_epsilon (0.1%) is at chi=6, the best x_sigma (0.4%) is at chi=14, and in between errors are 4-8%. The paper itself says there is no clear improvement when chi goes from 6 to 14, and calls chi=14 'magical'. So the reported accuracies are selected best cases, not a controlled extrapolation. The decreasing RG errors are projective truncation errors in intermediate tensors; no argument connects them to the final scaling-dimension errors. This is the main soft spot.\n\nThere is also the transposition trick. The paper says the proof that it preserves the partition function is deferred to a coming paper. If the trick is wrong, the cube filtering solves the wrong approximation problem. That is a real gap, but it is a well-defined one: either provide the proof or show numerically that the filtered tensors match an independent calculation.\n\nCredit where due: the method construction is careful, the paper is unusually honest about its limitations (Section VII essentially concedes the systematic-improvement point), the code is on GitHub, and the bootstrap comparison is genuinely independent. The fixed-point tensor, if it holds up, is a much richer object than a few exponents, and that makes the approach worth pursuing.\n\nWho is this for? People working on tensor-network RG and real-space RG in 3D. They should read it. It deserves a serious referee, but I would send it back for major revision: the authors need to either demonstrate systematic improvement with chi (or explain why it is not expected), remove or heavily qualify the 'well-controlled' language, and supply the transposition proof or an equivalent numerical check. If those conditions are met, the paper could be an important one.","headline":"Genuinely new 3D cube filtering scheme that produces a critical fixed-point tensor where HOTRG fails, but the 'well-controlled' claim is not yet earned by the non-monotonic scaling-dimension data.","tokens_in":13096,"tokens_out":2493,"would_cite":true,"duration_ms":20501,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":["05.10.Cc","64.60.F-"],"model":"deepseek-v4-flash","headline":"Cube filtering turns Kadanoff's block-spin idea into a reliable 3D renormalization group.","keywords":["real-space renormalization group","tensor network renormalization","entanglement filtering","cube filtering","3D Ising model","scaling dimensions","critical fixed-point tensor","HOTRG"],"falsifier":"Compare, for a small cubic lattice, the partition function obtained by contracting the transposed-cube tensor network against exact enumeration; any discrepancy at finite lattice size would falsify the transposition assumption. Alternatively, run the RG on a lattice model that lacks reflection symmetry with and without the transposition trick: if the estimated scaling dimensions shift outside the reported error bars, the trick is not symmetry-neutral.","tokens_in":1792,"feed_emoji":"🧊","tokens_out":3337,"duration_ms":63675,"temperature":0.7,"pith_summary":"This paper claims that Kadanoff's block-spin renormalization idea, once reformulated as a tensor-network map and augmented with an entanglement-filtering step called cube filtering, becomes a quantitatively reliable real-space RG in three dimensions. The authors apply it to the cubic-lattice Ising model and report RG errors near the critical fixed point reduced from more than 20% for plain higher-order tensor renormalization to about 2% when more couplings are retained, and scaling dimensions of the spin and energy-density fields with relative errors as low as 0.4% and 0.1%. If correct, the method provides a systematically controllable real-space RG in 3D and, for the first time, a numerical critical fixed-point tensor in a high-dimensional tensor space, which carries a complete description of the universality class. The paper is explicit that the best exponent estimates occur at particular bond dimensions and that improvement with increasing bond dimension is not yet monotonic.","feed_headline":"Cube filtering makes real-space RG precise in 3D","feed_subtitle":"Entanglement filtering tames the error that kept Kadanoff's block idea qualitative in three dimensions.","key_machinery":"The machinery is a tensor-network RG map $A \\to A'$ composed of two stages. First, a cube filtering squeezes the bond dimension of each outer leg of the anchor tensor through filtering matrices $s_x, s_y, s_z$, optimized by maximizing the overlap between the $2\\times2\\times2$ cube built from the tensor and its transpositions and the filtered cube; the transposition trick is what imports lattice reflection symmetry and cuts the number of independent filtering matrices from 24 to 3. Second, an HOTRG-like block-tensor transformation coarse-grains the filtered network direction by direction (z, then y, then x) using isometric tensors with separate bond dimensions for inner and outer legs. The EF step targets the corner entanglement that a plain block-tensor map cannot eliminate, and the total computational cost is $O(\\chi^{12.5})$.","core_discovery":"On the paper's own terms, the central discovery is that a block-tensor renormalization map in 3D can be made to flow to a genuine critical fixed point if each RG step is preceded by a cube filtering that removes short-range entanglement located at block corners. The cube filtering optimizes three filtering matrices, one per spatial direction, to maximize the overlap between the original 2x2x2 tensor patch and a filtered patch, using a transposition trick that enforces lattice reflection symmetry and reduces the number of independent filtering matrices from 24 to 3. With this enhancement, applied to the cubic-lattice Ising model, the RG errors stay stable near the fixed point and decrease from about 6% at bond dimension 6 to about 2% at bond dimension 14, and the linearized RG map yields estimates of $x_\\sigma$ and $x_\\epsilon$ whose best relative errors are 0.4% and 0.1% relative to conformal bootstrap values. The paper also reports higher scaling dimensions exhibiting conformal tower structure in 3D, which it notes has previously been seen numerically only in a fuzzy-sphere construction.","pith_inferences":["If the deferred proof of the transposition trick goes through, the same construction should apply to other reflection-symmetric lattice models, such as $n$-state Potts models, and plausibly to (2+1)D quantum systems via the quantum-classical map.","The non-monotonic improvement with bond dimension, including the 'magical' parameter choices that give the best exponents, suggests that truncation errors in the filtering and block-tensor stages partially cancel; a diagnostic tracking each truncation error separately might allow a more controlled extrapolation to the $\\chi \\to \\infty$ limit.","A direct test of systematic improvability would be to push the bond dimension beyond 14 using the cost-reduction ideas the paper cites; if the roughly 2% error keeps decreasing and the exponent estimates converge, the method becomes a practical alternative to Monte Carlo for 3D critical exponents."],"forward_implications":["A numerical critical fixed-point tensor for the 3D Ising universality class is obtained, so the full RG spectrum, not just a handful of exponents, becomes accessible.","Linearization around the fixed point yields scaling dimensions whose best estimates (relative errors 0.4% for $x_\\sigma$ and 0.1% for $x_\\epsilon$) lie close to accepted high-precision values.","The RG error is stable with RG step rather than growing, which is the qualitative failure of the plain HOTRG approach in 3D.","The method is graph-independent, so the cube filtering can be inserted into other block-tensor schemes beyond the particular HOTRG-like map used here.","The fixed-point tensor can be used to extract operator product expansion coefficients, and the flow provides a link between real-space RG and conformal field theory data."],"supporting_citations":[{"why":"Supplies the graph-independent local truncation idea that the cube filtering is built on.","marker":"[13]"},{"why":"Provides the optimization method for the filtering matrices by maximizing overlap in closed-loop tensor networks.","marker":"[14]"},{"why":"Defines the HOTRG block-tensor map that the proposed method enhances and uses as its baseline.","marker":"[21]"},{"why":"Establishes the qualitative difference between 2D and 3D entanglement growth that motivates the need for entanglement filtering.","marker":"[20]"},{"why":"Supplies the conformal bootstrap estimates of $x_\\sigma$ and $x_\\epsilon$ used as the benchmark for the reported errors.","marker":"[15]"},{"why":"Provides updated conformal bootstrap results for the 3D Ising stress tensor, used in the scaling-dimension comparison.","marker":"[16]"},{"why":"Introduces the tensor renormalization group reformulation that gives a natural measure of RG approximation errors.","marker":"[11]"},{"why":"Supplies the projective truncation technique with separate inner and outer isometric tensors used in the block-tensor map.","marker":"[12]"}],"fun_headline_variants":["Cube filtering sharpens 3D renormalization group to 0.4% error","Entanglement filtering makes block RG quantitative in 3D","Cube filter yields precise 3D critical fixed point","Filtering tames 3D real-space RG errors"],"cache_read_input_tokens":15232,"weakest_assumption_plain":"The method's accuracy rests on the claim that transposing the tensor in the cube-filtering step leaves the partition function unchanged for the cubic-lattice Ising model; the proof of this is deferred to a later paper, and if it fails, the filtering matrices solve the wrong optimization problem.","fun_headline_variants_meta":{"raw":{"variants":["Cube filtering sharpens 3D renormalization group to 0.4% error","Entanglement filtering makes block RG quantitative in 3D","Cube filter yields precise 3D critical fixed point","Filtering tames 3D real-space RG errors"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000652,"raw_usage":{"total_tokens":3007,"prompt_tokens":983,"completion_tokens":2024,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":599,"completion_tokens_details":{"reasoning_tokens":1951}},"tokens_in":599,"tokens_out":2024,"duration_ms":14033,"temperature":1.0,"reasoning_tokens":1951,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T12:49:06.861987+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compare, for a small cubic lattice, the partition function obtained by contracting the transposed-cube tensor network against exact enumeration; any discrepancy at finite lattice size would falsify the transposition assumption. Alternatively, run the RG on a lattice model that lacks reflection symmetry with and without the transposition trick: if the estimated scaling dimensions shift outside the reported error bars, the trick is not symmetry-neutral.","supporting_citations":[{"cited_title":"Hauru, C","cited_arxiv_id":null,"evidence_quote":"Supplies the graph-independent local truncation idea that the cube filtering is built on."},{"cited_title":"Evenbly, Algorithms for tensor network renormaliza- tion, Phys","cited_arxiv_id":null,"evidence_quote":"Supplies the projective truncation technique with separate inner and outer isometric tensors used in the block-tensor map."}],"review_version":1}