{"id":"96afa1cd-6133-4165-8e9a-8838e2223710","arxiv_id":"2512.06939","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"Energy minimization over tensor-train states has a well-defined Rayleigh-Ritz degree; for small systems homotopy continuation enumerates all critical points, showing ALS often stops at suboptimal local minima and rank-truncated DMRG leaves the manifold.","lead":"This paper counts and computes the stationary points of energy minimization when a quantum state is constrained to tensor-train low-rank form, and uses them to test two standard solvers. It matters because these solvers are widely used in quantum chemistry but have no guarantee of finding the ground state; the paper measures how often they get stuck.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Algorithm 1's XR parametrization covers only a Zariski-open subset U/W; the paper's own Example 7.4 shows four of 48 critical points (including the global minimum) lie outside W, so 'all critical points' and the ALS/DMRG benchmarks are not exhaustive as stated.","rationale":"The reader's weakest assumption is exactly the load-bearing concern here. The paper's central claim is that all complex critical points can be computed exhaustively via the new birational parametrization and homotopy continuation, and that this yields the first exhaustive ALS/DMRG benchmark. That claim fails for critical points outside the open set U/W, as the paper itself documents in Examples 7.3 and 7.4. The concern is not internal inconsistency or a disagreement with consensus; it is a correctness risk in the numerical completeness of the main pipeline. The paper deserves credit for disclosing these limitations in the examples rather than hiding them, and for providing reproducible code and data. The appropriate verdict is therefore the reader's CONDITIONAL: accept with claims restricted to the open set, and with the ALS/DMRG conclusions labeled as valid only for critical points reachable by the parametrization. No verdict change is needed beyond what the reader already recommended.","tokens_in":30510,"tokens_out":5155,"duration_ms":55401,"concrete_test":"For the Hamiltonian of Example 7.4, solve the critical equations using the globally surjective Segre parametrization P^1 x P^1 x P^1 (Proposition 2.5) instead of Algorithm 1. If the four missing points, including the global minimum, are recovered, then the open-set computation is demonstrably incomplete. Then repeat the same comparison for the ten Hamiltonians of Example 7.3 with k=(2)^4, r=(1,2,1), using the Segre parametrization P^3 x (P^1)^2; if any previously unreported local minimum (e.g., for H5) appears, Table 3 and the ALS/DMRG conclusions in Table 5 must be revised.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The birational parametrization in Theorem 3.3 is only a bijection between a Zariski-open U of the tensor-train manifold and an open set W of the Grassmannian product, defined by requiring that the first r_i rows of each flattening are linearly independent (Assumption 3.1, eq. (5)). The numerical pipeline solves the parametric critical equations on these affine coordinates, so it can only find critical points whose image lies in U; any critical point in V\\U is invisible to the homotopy. This is not a merely hypothetical defect: Example 7.4 states that for k=(2)^3, r=(1)^2 only 44 of 48 critical points lie in W, and the global minimum is among the four missing points. Example 7.3 likewise reports that H5 produces no extrema because the expected solution lies in the complement of W. Since the Section 7.3 ALS/DMRG comparison is presented as a benchmark against 'all critical points' of (6), the central exhaustive-computation claim is missing exactly the low-energy points that matter for the conclusions. This is structural to Algorithm 1, not a tuning or accuracy issue: no affine parameter homotopy on W can reach V\\U.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies Rayleigh-quotient minimization over tensor train (TT) varieties, a model problem in quantum-chemistry DMRG/ALS computations. It introduces the Rayleigh–Ritz degree as the number of complex critical points for generic symmetric H, defines the Rayleigh–Ritz discriminant, and proves a birational parametrization of the TT manifold from a product of Grassmannians (Algorithm 1/2, Theorem 3.3). The authors then use homotopy continuation to compute all critical points for small TT and determinantal varieties, report RR degrees and discriminant degrees, and benchmark ALS and DMRG against these exhaustive critical-point lists. The theoretical sections include several proved results (Lemma 2.2, Theorem 3.3, Theorem 5.1, Proposition 5.6) as well as conjectures, and the numerical pipeline is implemented in TensorTrainOptimization.jl with data on Zenodo.","tokens_in":30872,"tokens_out":3358,"duration_ms":33848,"significance":"If the claims hold, the paper provides a useful algebro-geometric framework for a problem of genuine computational interest: it formalizes the number of critical points of the constrained Rayleigh quotient, gives a parametrization that makes numerical homotopy methods applicable, and offers the first systematic comparison of ALS/DMRG with the complete critical set. The availability of reproducible code and data is a clear strength. The main theoretical contributions (RR correspondence, RR discriminant, birational parametrization) are mathematically substantial and appear sound in their local statements. However, the central numerical claim of exhaustive computation is undermined by the authors' own examples, as detailed below.","major_comments":[{"comment":"The birational parametrization of Theorem 3.3 is only a bijection between a Zariski-open U ⊆ V⁼_{k,r} and an open W of the Grassmannian product, defined by Assumption 3.1 (first r_i rows of each flattening independent). The homotopy pipeline of §7 solves the parametric critical equations on W, so it can only find critical points whose image lies in U. This is not a hypothetical restriction: Example 7.4 states that for k=(2)^3, r=(1)^2 only 44 of 48 critical points lie in W, with the global minimum among the four missing points; Example 7.3 reports that for H5 no extrema are found because the solution is expected to lie in the complement of W. Consequently, the statement in the abstract and §7 that the method computes 'all critical points' is not supported. This is structural to Algorithm 1, not a tuning issue: no parameter homotopy on W can reach V⁼_{k,r} \\ U. The authors should either r","section":"§3, Theorem 3.3 and §7, Examples 7.3, 7.4"},{"comment":"The benchmark conclusion that 'ALS frequently gets stuck in local minima with suboptimal energy' is based on comparing ALS outcomes with the computed critical-point set. Since the global minimum can lie outside W (as in Example 7.4), the comparison may systematically miss the true optimal low-energy points. Thus the strength of the empirical conclusions in Table 5 and the associated discussion is not justified by the data. Unless the missing critical points are computed by an independent method, the claims should be restricted to critical points lying in W, or explicitly qualified as applying only to the computed subset.","section":"§7.3, Table 5 and Example 7.4"},{"comment":"The proof of Proposition 5.6 asserts that the image of the ramification locus under the birational map ψ×Id is dense in R_ram^V, but the argument only establishes this for a dense open subset of the domain. A rigorous proof should show that the birational map induces a surjection (or at least a dominant map) on the ramification loci, not merely on a dense open piece. This is a gap in a theoretical statement that is used to justify the discriminant-degree computations in Example 7.5. I recommend either supplying a complete proof or stating the result as a heuristic/proposition with a weaker conclusion.","section":"§5.2, Proposition 5.6 and §7.5"}],"minor_comments":[{"comment":"The notation V⁼_{k,r} appears in §3 but is not formally defined in a displayed equation; please define it explicitly. Also, the overline notation in V_{k,r} = im(Ψ) is used inconsistently with the open set U.","section":"§2, after (1)"},{"comment":"Typo: 'Lagragian' should be 'Lagrangian'.","section":"§4.1, Proposition 4.6"},{"comment":"Typo: 'distnace' should be 'distance'. Also, the phrase 'essentially parametrizes' is vague; please specify the sense in which the BW correspondence is parametrized.","section":"§6, line before Definition 6.1"},{"comment":"The statement that the RR correspondence of P¹×P¹ has degree 17 is given without a reference to the supplementary material; please include the computation or point to the relevant file in [4].","section":"§5.1, Example 5.5"},{"comment":"The table would be easier to read if the units of 'time' were stated (seconds, minutes, hours are mixed). Please unify or add a note.","section":"§7.2, Table 2"},{"comment":"Reference [5] is listed as 'J. Software for Algebra and Geometry' without page numbers; please complete the bibliographic data.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The paper's core theoretical framework is valuable and the numerical infrastructure is commendable, but the exhaustive-computation claim is not tenable as stated. The authors' own Examples 7.3 and 7.4 demonstrate that low-energy critical points can lie in the complement of the parametrized open set, which is exactly the region of interest for the ALS/DMRG benchmark. This is not a minor caveat but a structural limitation of the proposed pipeline. I recommend major revision: the authors should either (i) extend the numerical method to cover all critical points via a chart atlas or implicit methods, or (ii) substantially rewrite the abstract, §7, and the ALS/DMRG conclusions to state clearly that the comparison covers only critical points in the open set U. The theoretical sections (RR degree, discriminant, birational parametrization) are likely publishable even with the numerical limitations, but the current framing overstates the results."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know two things before reading: the birational parametrization and the numerical pipeline are real and useful, but the phrase \"all critical points\" is doing more work than it should. The paper mostly knows this, which is why I'd still referee it.\n\nThe genuinely new pieces are Algorithm 1/2 (the XR-decomposition parametrization of TT manifolds from products of Grassmannians), the TT-specific RR correspondence and discriminant computations, and the small-scale ALS/DMRG benchmark. The proofs I checked are fine: Lemma 2.2, Theorem 3.3, and the RR-correspondence results are coherent. The code and data on Zenodo are a real asset; that is the right way to do this kind of computation. Attribution to [38] for the general RR framework is explicit and correct.\n\nThe soft spot is the one flagged by the examples themselves. Assumption 3.1 means the parametrization only sees tensors whose flattenings have their first r rows independent. The homotopy runs on that open set W. Example 7.4 says for k=(2)^3, r=(1)^2, only 44 of 48 critical points lie in W, and the global minimum is among the four missing. Example 7.3's H5 yields no extrema because the solution is in the complement. So the \"all critical points\" claim in the abstract and in the ALS/DMRG comparison is too strong. The comparison shows what ALS/DMRG do among the points the homotopy can reach — which may systematically exclude low-energy points. That is a structural mismatch, not an accuracy problem in the solver.\n\nA few smaller issues: the RR degree of the non-Segre TT variety in Proposition 7.2 is only a lower bound from a monodromy run that did not terminate, and some discriminant degrees in Table 2 come from nonterminating or heuristic computations. The paper usually labels these, but the tables themselves can read as exact values.\n\nThe limitation section is honest about this, which is not nothing. The theoretical core holds up; the \"exhaustive\" framing needs to be pulled back to \"over the open set W,\" with the possibility that ground states sit outside it addressed or at least discussed.\n\nThis is a paper for people in algebraic geometry and numerical tensor methods. I'd let it through peer review, with the authors asked to reword the exhaustive claims and to add an explicit statement that the computed list is a subset of the true critical set for examples like 7.4.","headline":"Solid birational-parametrization paper whose 'all critical points' claim overreaches; the authors disclose their own counterexamples, and the core theory holds.","tokens_in":31314,"tokens_out":2235,"would_cite":true,"duration_ms":23275,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["14Q20","14M12","65H10","15A69"],"pacs":[],"model":"deepseek-v4-flash","headline":"The constrained Rayleigh quotient over a tensor-train variety has a well-defined number of complex critical points, the Rayleigh-Ritz degree, and for small systems those points can be computed exhaustively, exposing when standard algorithms","keywords":["tensor train varieties","Rayleigh-Ritz degree","Rayleigh quotient","critical points","homotopy continuation","alternating linear scheme","DMRG","quantum chemistry"],"falsifier":"Construct a small Hamiltonian whose ground state lies on the tensor-train variety with the first r rows of some flattening linearly dependent, then run the parametrization-based homotopy solver and compare to direct eigenvector computation: if the solver's reported global minimum energy exceeds the true constrained minimum, the exhaustiveness claim fails.","tokens_in":30412,"feed_emoji":"🧮","tokens_out":5796,"duration_ms":51564,"temperature":0.7,"pith_summary":"The paper tries to settle a practical question: when a quantum-chemistry Hamiltonian is minimized over the low-rank tensor-train format, what does the optimization landscape actually look like? It defines the Rayleigh-Ritz degree — the number of complex critical points of the constrained Rayleigh quotient for a generic Hamiltonian — and proves it is well defined for tensor-train varieties. The authors give a birational parametrization of the tensor-train variety from a product of Grassmannians, turning the critical-point equations into a polynomial system that homotopy continuation can solve exhaustively for small systems. With the full critical-point list in hand, they benchmark the standard algorithms ALS and DMRG: ALS frequently lands in local minima with energies far from the global minimum, and DMRG's rank truncation means it does not converge to critical points of the fixed-rank problem at all. The payoff is the first exhaustive picture of where these heuristics can go wrong.","feed_headline":"Small tensor-train energy landscapes fully mapped","feed_subtitle":"Full critical-point census shows ALS and DMRG can miss the ground state.","key_machinery":"The load-bearing object is the birational parametrization of the tensor-train manifold V=_{k,r} from a product of Grassmannians (Algorithm 1 and Theorem 3.3), built from successive XR decompositions — a factorization of a matrix into a row-selection factor and a coefficient matrix, valid under the assumption that the first r rows of each flattening are linearly independent. This converts the constrained Rayleigh quotient into an unconstrained rational function on the parameter space, whose critical equations form a polynomial system. The paper also introduces the Rayleigh-Ritz correspondence (the incidence variety of critical-point–Hamiltonian pairs) and its discriminant, the set of Hamilton","core_discovery":"The central claim is that constrained Rayleigh-quotient critical points on a tensor-train variety can be computed completely for small systems, and that the count is governed by a constant, the Rayleigh-Ritz degree, independent of the generic Hamiltonian. This is achieved by identifying the tensor-train variety (when ranks satisfy certain inequalities) as a Segre product, by giving a birational map from a product of Grassmannians to the tensor-train manifold (Theorem 3.3), and by defining the Rayleigh-Ritz correspondence and discriminant, which describe Hamiltonians with a deficient number of critical points. Numerical homotopy continuation then computes all isolated complex critical points","pith_inferences":["A natural next step is to compute the average number of real critical points using the discriminant's regions; the examples here show that counts vary widely, so a random Hamiltonian's landscape is far from typical.","Because the birational map misses critical points outside its open set, a robust implementation should union left-to-right and right-to-left XR/CX charts to capture low-energy states the current method cannot see.","If energy-adaptive truncation (rather than Frobenius-norm SVD truncation) is used inside DMRG, the algorithm might converge to fixed-rank critical points; the paper does not test this, leaving a concrete open modification.","The same Rayleigh-Ritz degree and discriminant framework could be applied to other tensor-network formats, such as matrix product operators or hierarchical Tucker varieties, to benchmark their optimizers."],"forward_implications":["For small tensor-train and determinantal varieties, the Rayleigh-Ritz degree and the actual critical points are now known, giving a ground truth for testing any approximate eigensolver.","The Rayleigh-Ritz discriminant splits the space of Hamiltonians into regions with different numbers of real critical points; for example, P1×P1 has 4, 6, or 8 real critical points depending on the Hamiltonian.","ALS can converge to any local minimum, and the global minimum is not always the most frequent output; for random symmetric matrices the energy gap can be large.","DMRG with rank truncation is not a solver for the fixed-rank Rayleigh quotient, so its output should not be compared directly to manifold critical points.","When critical points lie outside the parametrized open set, the computed census is incomplete; Example 7.4 shows that the global minimum can be one of the missing points."],"fun_headline_variants":["All tensor-train energy critical points counted","Small tensor-train: all energy states counted","Complete critical-point census for tensor-train energies","Rayleigh-Ritz degree counts tensor-train energy states","Homotopy finds all tensor-train critical points"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The exhaustive count is only guaranteed on the open set of tensors where, at every step, the first r rows of each flattened tensor are linearly independent; any critical point outside this open set is invisible to the computation, and the paper itself shows a case where the global minimum is one of the missing points.","fun_headline_variants_meta":{"raw":{"variants":["All tensor-train energy critical points counted","Small tensor-train: all energy states counted","Complete critical-point census for tensor-train energies","Rayleigh-Ritz degree counts tensor-train energy states","Homotopy finds all tensor-train critical points"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.002208,"raw_usage":{"total_tokens":8351,"prompt_tokens":679,"completion_tokens":7672,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":423,"completion_tokens_details":{"reasoning_tokens":7599}},"tokens_in":423,"tokens_out":7672,"duration_ms":47164,"temperature":1.0,"reasoning_tokens":7599,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-03T18:01:18.551687+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Construct a small Hamiltonian whose ground state lies on the tensor-train variety with the first r rows of some flattening linearly dependent, then run the parametrization-based homotopy solver and compare to direct eigenvector computation: if the solver's reported global minimum energy exceeds the true constrained minimum, the exhaustiveness claim fails.","supporting_citations":[],"review_version":1}