{"id":"17c07249-9e3f-4bdc-9c62-abd6cb353651","arxiv_id":"2608.06905","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"An iterative thresholding integrator for hierarchical tensors yields quasi-optimal rank bounds for high-dimensional Schrödinger equations, polynomial in dimension.","lead":"This paper presents a time-stepping method for high-dimensional Schrödinger equations that keeps tensor ranks close to the minimal ranks needed for a given accuracy, with rigorous bounds that grow only polynomially in dimension. If the analysis holds, it offers a practical path to simulating many-particle quantum systems without exponential cost.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The quasi-optimality theorem is proved only for an idealized variant (exact F_n, δ_n=0, contraction ρ<1), not for Algorithm 3.1 as implemented; this gap is load-bearing for the central claim.","rationale":"The paper's core analytic work is sound within its stated assumptions: the fixed-point contraction framework, the thresholded iteration, and the soft-thresholding error propagation are internally consistent, and Proposition 5.1 and Theorem 5.7 appear to follow from the stated lemmas. The most serious issue is not a demonstrated algebraic error in those estimates but a scope mismatch between the theorem and the claimed object. Corollary 5.8 bounds ranks of eun for a version with δ_n=0 and exact F_n; Algorithm 3.1 as written and tested uses δ_n>0 and recompressed F_n, and the numerical regime violates the contraction hypothesis. Since the reader's strongest claim is explicitly about Algorithm 3.1, the missing analysis is load-bearing. The reader's verdict already flags this among three issues, but the reader's weakest_assumption emphasizes the contraction regime; I agree with that concern while locating the more decisive mismatch in the exact-F/δ_n=0 restriction. The concrete test above would settle whether the gap is a harmless omission or an actual failure of the theorem to cover the implemented algorithm. I do not see grounds for REJECT: the conditional theorem is plausible and the paper is transparent about both limitations. CONDITIONAL remains appropriate, with the requested revision being either an extension of the analysis to recompressed inexact F and positive δ, or a precise restriction of the central claim to the idealized variant.","tokens_in":28746,"tokens_out":36298,"duration_ms":383972,"concrete_test":"Re-derive Proposition 5.4 with δ_n>0 and inexact F evaluations: add δ_k to the source term in (5.4) and track the accumulated rank bound in the proof of Corollary 5.8. If the resulting bound contains an extra T/h-dependent term in δ_k that is not majorized by r_best(u, c max(ε_k,δ_k)), then the quasi-optimality claim as stated for Algorithm 3.1 fails; if the terms are absorbable, the gap is fixable and the verdict can remain CONDITIONAL rather than REJECT.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"Section 5.4's rank bounds (Theorem 5.7, Corollary 5.8) rest on Proposition 5.4, which explicitly sets δ_n=0 and assumes F_n is evaluated exactly (Section 5.3). Algorithm 3.1, however, includes endpoint recompression R_{δ_n} with δ_n>0 and, as described in Section 3.1.2 and implemented in Section 6, applies recompressions inside evaluations of F_n. Consequently the theorem does not establish the reader's strongest claim as stated for Algorithm 3.1 'at comparable accuracy': the error of the implemented eun contains an additional δ_n, and the iterates are generated by a perturbed fixed-point map, so the relation between the stopping tolerance ε_n, the quantity ∥u−bv(n)∥_Q used in Proposition 5.1, and the soft-thresholding error of the exact solution is no longer governed by the proof. This is not merely a technicality: the numerical tests at h=1/10, Q=10 are admitted to lie outside the contraction regime ρ=C_GΛ_Qh<1 of Section 2.2, so they cannot validate the theorem either. The paper is candid about both limitations, but the abstract and the reader's strongest_claim present Theorem 5.7/Corollary 5.8 as applying to Algorithm 3.1, which is an overstatement of the proven scope.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes and analyzes a rank-adaptive time integrator for high-dimensional linear Schrödinger-type problems. The integrator advances the solution on each time subinterval by a Picard fixed-point iteration for a Gauss collocation formulation, applying hierarchical-tensor soft thresholding to control ranks. The main theoretical contribution is a set of error bounds and rank estimates: under algebraic or exponential decay of the matricization singular values of the exact solution, the ranks of the iterates and of the approximations at interval endpoints are claimed to be quasi-optimal relative to best approximation ranks of the exact solution, up to constants depending polynomially on the dimension and linearly on the number of time steps. Numerical experiments in four and sixty-four dimensions, using an analytic Gaussian reference solution, illustrate the practical behavior of the method, including tests explicitly acknowledged to lie outside the theoretical contraction regime.","tokens_in":29039,"tokens_out":9241,"duration_ms":111529,"significance":"If the advertised rank bounds were established for the implemented algorithm, this would be a valuable contribution to rank-adaptive time integration: it would provide a rigorous balance between accuracy and ranks for hierarchical tensor approximations, with dimension dependence that is only polynomial in the explicit constants. The paper is also commendably candid about its limitations, including the idealized assumptions in the global analysis and the heuristic choices in the numerical section. However, as written, the central quasi-optimality theorem is proved for an idealized variant of the scheme rather than for Algorithm 3.1 as implemented, and the numerical experiments do not operate in the regime of the theorem. These gaps are load-bearing for the paper's main claim, although they appear to be repairable within the manuscript's scope.","major_comments":[{"comment":"The rank-quasi-optimality claim for Algorithm 3.1 is proved only for the idealized variant in which δ_n=0 and F_n is evaluated exactly. Section 5.3 explicitly sets δ_n=0 before Proposition 5.4, and the rank recursion for the endpoint values in Proposition 5.4 and Proposition 5.7 does not include the endpoint recompression R_{δ_n} of Algorithm 3.1, line 19. The implemented algorithm, however, uses δ_n>0 in the experiments, and Section 3.1.2 together with Section 6 recompresses inside every evaluation of F_n with a prescribed relative budget. These two features perturb the contraction map whose fixed point is approximated and introduce an additional error δ_n that is absent from the bound of Proposition 3.4 as used in Proposition 5.4. Consequently, Theorem 5.7 and Corollary 5.8 as stated do not establish quasi-optimality of the ranks produced by the implemented Algorithm 3.1 at the stated accuracy. The paper should either extend the perturbation analysis to cover inexact F_n and positive δ_n, or state the theorem explicitly for the algorithm without internal recompression and with δ_n=0, and describe the numerical experiments as beyond the theorem's scope.","section":"§5.3–5.4, Eq. (5.4), Prop. 5.4, Prop. 5.7, Cor. 5.8"},{"comment":"The text requires α_0 to be chosen so that v_{0,1}=S_{α_0}F(v_{0,0})=0, and the subsequent rank analysis uses v_{0,0}=v_{0,1}=0 and J_0=1. Algorithm 3.1 instead sets α_{0,n}=‖ũ_{n-1}‖/(2d-3). Since the hierarchical soft-thresholding operator S_α zeroes a tensor only if the threshold is at least the largest singular value of every matricization, and since F_n(0) has each stage equal to ũ_{n-1}, the choice α_{0,n}=‖ũ_{n-1}‖/(2d-3) does not in general make v_{0,1}=0. For rank-one initial data, for example, every matricization has singular value ‖ũ_{n-1}‖, so the threshold is too small by a factor of 2d−3. The equality v_{0,1}=0 is therefore not guaranteed by the stated initialization, and the base case of Lemma 4.1 and Proposition 4.3 may be invalid for the implemented algorithm. This discrepancy should be corrected, for instance by setting α_{0,n}=‖ũ_{n-1}‖ as a sufficient threshold, or by modifying the base-case analysis.","section":"§3.1.1, Lemma 4.1, Prop. 4.3, Algorithm 3.1 line 3"},{"comment":"The convergence theory for the outer and inner loops relies on the contraction estimate ρ=C_GΛ_Qh<1 in Eq. (2.9), and the inner stopping criterion (3.2) explicitly contains the factor (1−ρ)/(ρ(1+ρ)). Section 6 states that with h=1/10 and Q=10 contractivity is not ensured, and that in this regime the iteration parameters are chosen heuristically, with the threshold-decrease factor set to 3/5. The numerical experiments therefore do not test the theorem's assumptions and cannot be used to validate Prop. 5.7 or Cor. 5.8. The authors are transparent about this, but the abstract and conclusion should not imply that the experiments confirm the proven quasi-optimality bounds; the scope of the empirical evidence should be stated more precisely.","section":"§2.2, §3.1.1, §6"}],"minor_comments":[{"comment":"The cross-references to 'Theorem 5.7' and 'Theorem 5.8' in Section 5.1 and Section 5.4 do not match the displayed numbering, which is Proposition 5.7 and Corollary 5.8; please harmonize the references.","section":"§5.1, §5.4"},{"comment":"Algorithm 3.1 does not list the internal recompression tolerances described in Section 3.1.2 as input parameters, although these tolerances are needed to reproduce the numerical experiments in Section 6; the input list and the experiment description should be aligned.","section":"Algorithm 3.1, §3.1.2, §6"},{"comment":"Equation (5.4) prescribes ε_n recursively in terms of κ_Q, κ_{2Q}, and η_0; the presentation would be clearer if a closed-form or explicit geometric choice satisfying the recursion were given, since the current formula already suggests exponential growth in n.","section":"§5.3, Eq. (5.4)"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is likely publishable after revision if the authors either tighten the statement of the quasi-optimality theorem to the algorithm actually analyzed or close the gap by analyzing the perturbed fixed-point map and the positive endpoint recompression tolerance. The initialization discrepancy in Algorithm 3.1 is concrete and should be fixed. The numerical section's candor is a strength, but the main advertised claim should not be presented as applying to the implemented algorithm without qualification."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know. First, this is a real extension: the d=2 iterative rank-adaptive integrator from your earlier matrix work is carried to hierarchical tensors with bounds that depend only polynomially on dimension, and the new threshold-decrease rule is what makes that work. The analysis of soft thresholding for hierarchical tensors is nontrivial, and the proofs are detailed. Second, the headline quasi-optimality result (Prop. 5.7, Cor. 5.8) is proved for an idealized version of the algorithm: exact F_n evaluations and no endpoint recompression (Section 5.3 sets δ_n = 0, and Section 3.1.3 assumes exact F_n). The implemented Algorithm 3.1 recompresses inside F_n and uses δ_n > 0. The authors are candid about both choices, but the abstract and the way the results are stated in Section 5.4 make it easy to overread the theorem as covering Algorithm 3.1 as written. The stress-test note has it right: this is a real gap, and it is load-bearing for the claim that the implemented method is quasi-optimal at comparable accuracy.\n\nWhat the paper does well: the threshold-steering strategy genuinely differs from the d=2 predecessor, avoiding the dimension-exponential factor that a direct extension would produce; the local rank bounds (Prop. 4.3, Cor. 4.4) are clean; and the numerical experiments in 4D and 64D show the method works in practice, with modest ranks and honest reporting. The paper also says plainly that the tests at h = 1/10, Q = 10 lie outside the contraction regime ρ < 1, and that in that regime the iteration parameters are chosen heuristically. That is the right way to report.\n\nSoft spots, in proportion. The theory/implementation gap is real but not hidden; I would want the authors to state the scope of the main theorem explicitly in the abstract and to add a remark on what would be needed to cover δ_n > 0 and inexact F_n. The α_0 initialization concern the reader raised is weaker than it looks: with α_0 = ||ũ_{n-1}||/(2d−3), the successive soft thresholdings do zero out a rank-one tensor, so Lemma 4.1 is fine. The global error bound in Prop. 5.4 grows exponentially in t_n, which the paper acknowledges as pessimistic; that is not a flaw but it limits the practical relevance of the error guarantee.\n\nWho is this for? Numerical analysts working on low-rank tensor time integration, especially people comparing step-truncation and dynamical low-rank methods. It deserves a serious referee: the analysis is original, the dependence on dimension is a genuine improvement, and the remaining gap is fixable by repositioning the claims, not by redoing the mathematics. I would recommend acceptance after revision, with the scope of the quasi-optimality theorem made precise and the implemented-versus-analyzed discrepancy addressed head-on.","headline":"Genuine extension of the matrix-level iterative thresholding analysis to hierarchical tensors with polynomial dimension dependence, but the quasi-optimality theorem is proved for an idealized variant of the algorithm and the paper is only partially clear about that.","tokens_in":29607,"tokens_out":4013,"would_cite":true,"duration_ms":44638,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["65D40","65F55","65M12","65Y20","65L70"],"pacs":[],"model":"deepseek-v4-flash","headline":"Under algebraic or exponential singular-value decay, the iterative soft-thresholding time integrator for hierarchical tensors produces ranks bounded by the best approximation ranks of the exact solution, up to polynomial-in-dimension and…","keywords":["low-rank approximation","hierarchical tensor format","soft thresholding","time-dependent Schrödinger equation","quasi-optimal ranks","fixed-point iteration","collocation methods","curse of dimensionality"],"falsifier":"Run Algorithm 3.1 on a high-dimensional Schrödinger-type problem with a known exact solution whose matricization singular values decay algebraically with a known exponent, and compare the maximum hierarchical rank of the computed endpoint approximations at error level $\\varepsilon$ to the best approximation rank $r_{\\mathrm{best}}(u,\\varepsilon)$. If the observed rank grows faster than $\\mathrm{poly}(d)\\cdot (T/h)\\cdot r_{\\mathrm{best}}(u,\\varepsilon)$ as $\\varepsilon\\to0$, or if the iteration diverges whenever $\\rho\\ge 1$, the quasi-optimality claim is wrong.","tokens_in":28510,"feed_emoji":"🧮","tokens_out":10632,"duration_ms":104877,"temperature":0.7,"pith_summary":"This paper analyzes a time-stepping method for high-dimensional linear Schrödinger-type equations in which the solution is kept in hierarchical tensor form and the tensor ranks are controlled by iterative soft thresholding. The central claim is that the ranks produced by the algorithm stay close to the minimal ranks needed to approximate the exact solution at the same accuracy: under algebraic or exponential decay of the matricization singular values, the computed ranks are bounded by the best approximation ranks up to factors polynomial in the dimension and linear in the number of time steps. This matters because it turns an otherwise heuristic rank-adaptation strategy into a provable balance between accuracy and complexity, extending to higher-order tensors what was previously known only for matrices. The practical payoff is a way to simulate high-dimensional quantum dynamics without the curse of dimensionality while retaining a priori rank control.","feed_headline":"Soft-thresholded time stepping keeps tensor ranks near-optimal","feed_subtitle":"A new proof shows the integrator's ranks track best approximation ranks up to mild dimension and step factors.","key_machinery":"The central object is the hierarchical tensor soft-thresholding operator $S_\\alpha = S_{E,\\alpha}\\circ\\cdots\\circ S_{1,\\alpha}$, which applies soft thresholding to each matricization associated with the nodes of the binary dimension tree in succession. Because this operator is non-expansive, the composed map $S_{\\alpha_i}F$ is still a contraction whenever the fixed-point map $F$ (built from the Duhamel/twisted-variable integral formulation) has contraction ratio $\\rho=C_G\\Lambda_Q h<1$; each such map has a unique fixed point $v_{\\alpha_i}$. The algorithm decreases the threshold geometrically ($\\alpha_{i+1}=\\theta\\alpha_i$) only after the inner Picard iteration has converged to $v_{\\alpha_i}$, and the analysis transfers the resulting bounds on the residual $\\|Fv_{i,j}-v_{i,j}\\|_Q$ into bounds on $\\|u-S_{\\alpha_i}u\\|_Q$, the soft-thresholding error of the exact solution. Standard estimates then convert singular-value decay of $u$ into the quasi-optimal rank bounds.","core_discovery":"The authors prove that Algorithm 3.1—a Gauss-collocation integrator whose fixed-point equations are solved by Picard iteration with hierarchical-tensor soft thresholding—produces approximations whose hierarchical ranks are quasi-optimal. Theorem 5.7 and Corollary 5.8 state that if the exact solution's matricization singular values decay algebraically or exponentially, then for every subinterval the maximal rank of the computed iterates and endpoint approximations is bounded by a constant (depending only on the decay, the contraction ratio, and the threshold parameters) times $E^{C} r_{\\mathrm{best}}(u,\\varepsilon)$ with $C$ of order 2 to 3 and logarithmic corrections in the exponential case, up to a linear factor in the number of time steps. The proof works by relating the residual of the thresholded iteration to the soft-thresholding error of the exact solution, and by choosing the recompression tolerance at interval endpoints so that the endpoint ranks are controlled by best approximation ranks of the local fixed-point solution.","pith_inferences":["The reported stability of the iteration for parameters where $\\rho\\ge 1$ suggests that the contraction hypothesis may be stronger than necessary; a natural test is whether thresholding, rather than step-size restriction, is what keeps the iteration convergent for unbounded potentials.","Because the linear-in-$N$ factor appears unavoidable for comparisons with the exact solution, the method is best used with a small number of large time steps; combining it with step-truncation schemes on finer grids could give a practical two-level strategy.","A direct consequence not pursued in the paper is that the same soft-thresholded fixed-point iteration can be applied to nonlinear Schrödinger and parabolic problems, provided the generator admits a hierarchical low-rank application and a Lipschitz constant.","The two-tier rank comparison—against local fixed points and against the exact solution—provides a decomposition of rank growth into intrinsic solution complexity and error-propagation effects, which could be used as a diagnostic in numerical experiments."],"forward_implications":["With Gauss–Legendre nodes the step map is isometric, so the global error bound grows linearly in the final time $T$ instead of exponentially.","The rank bounds carry a linear factor in the number of time steps $N=T/h$; this is the price of the fixed-point formulation and it matches the general limitation identified in the cited step-truncation counterexamples.","The local error of a $Q$-stage Gauss–Legendre method is of order $h^{2Q+1}$, so taking a few large steps with moderate $Q$ keeps both error and rank growth small.","The constants in the rank bounds involve only low-degree polynomials in the dimension via $E=2d-3$, so the quasi-optimality does not deteriorate exponentially with the spatial dimension.","The arguments carry over to other evolution problems that admit a contractive fixed-point formulation, including parabolic equations and Lipschitz nonlinearities."],"supporting_citations":[{"why":"Supplies the matrix-case soft-thresholding integrator and the contraction-based fixed-point analysis that this paper generalizes to hierarchical tensors.","marker":"[6]"},{"why":"Defines hierarchical tensor soft thresholding, proves its non-expansiveness, and provides the fixed-point and rank lemmas used in the core proofs.","marker":"[8]"},{"why":"Provides the hierarchical tensor format background and the HSVD truncation error estimates assumed throughout.","marker":"[4]"},{"why":"Gives the recompression lemma (restated as Lemma 4.5) that links endpoint truncation to best approximation ranks.","marker":"[5]"},{"why":"Shows that a $h^{-1}$ factor in rank bounds relative to the exact solution is generally unavoidable, which motivates the linear-in-$N$ form of the global bounds.","marker":"[17]"},{"why":"Foundational reference for the hierarchical tensor (HT) format and its rank structure.","marker":"[22]"},{"why":"Introduces tensor trains, the notable special case of the HT format that the method also covers.","marker":"[38]"},{"why":"Establishes the isometry (norm-preservation) properties of Gauss–Legendre collocation used in Lemma 3.1.","marker":"[24]"}],"fun_headline_variants":["Iterative thresholding in time integration achieves near-optimal tensor ranks","Low-rank time integrator with soft thresholding keeps ranks near-optimal","Soft-thresholded time stepping provably attains near-optimal tensor ranks","Thresholded time integration balances error and rank near-optimally","Provably quasi-optimal ranks for hierarchical tensor time integration"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole proof stands on the assumption that the time step is small enough for the fixed-point iteration to be a contraction; in the reported experiments, the chosen step is large enough that this condition is not guaranteed to hold.","fun_headline_variants_meta":{"raw":{"variants":["Iterative thresholding in time integration achieves near-optimal tensor ranks","Low-rank time integrator with soft thresholding keeps ranks near-optimal","Soft-thresholded time stepping provably attains near-optimal tensor ranks","Thresholded time integration balances error and rank near-optimally","Provably quasi-optimal ranks for hierarchical tensor time integration"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000246,"raw_usage":{"total_tokens":1462,"prompt_tokens":791,"completion_tokens":671,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":407,"completion_tokens_details":{"reasoning_tokens":580}},"tokens_in":407,"tokens_out":671,"duration_ms":7125,"temperature":1.0,"reasoning_tokens":580,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T18:41:29.020632+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run Algorithm 3.1 on a high-dimensional Schrödinger-type problem with a known exact solution whose matricization singular values decay algebraically with a known exponent, and compare the maximum hierarchical rank of the computed endpoint approximations at error level $\\varepsilon$ to the best approximation rank $r_{\\mathrm{best}}(u,\\varepsilon)$. If the observed rank grows faster than $\\mathrm{poly}(d)\\cdot (T/h)\\cdot r_{\\mathrm{best}}(u,\\varepsilon)$ as $\\varepsilon\\to0$, or if the iteration diverges whenever $\\rho\\ge 1$, the quasi-optimality claim is wrong.","supporting_citations":[{"cited_title":"Iterative methods based on soft thresholding of hierarchical tensors.Found","cited_arxiv_id":null,"evidence_quote":"Defines hierarchical tensor soft thresholding, proves its non-expansiveness, and provides the fixed-point and rank lemmas used in the core proofs."},{"cited_title":"Adaptive near-optimal rank tensor approximation for high- dimensional operator equations.Foundations of Computational Mathematics, 15(4):839–898, 2015","cited_arxiv_id":null,"evidence_quote":"Gives the recompression lemma (restated as Lemma 4.5) that links endpoint truncation to best approximation ranks."},{"cited_title":"Adaptive low-rank integration for ordi- nary differential equations","cited_arxiv_id":null,"evidence_quote":"Shows that a $h^{-1}$ factor in rank bounds relative to the exact solution is generally unavoidable, which motivates the linear-in-$N$ form of the global bounds."},{"cited_title":"Springer, Cham, 2019","cited_arxiv_id":null,"evidence_quote":"Foundational reference for the hierarchical tensor (HT) format and its rank structure."},{"cited_title":"Oseledets","cited_arxiv_id":null,"evidence_quote":"Introduces tensor trains, the notable special case of the HT format that the method also covers."},{"cited_title":"Springer, Heidelberg, second edition, 2006","cited_arxiv_id":null,"evidence_quote":"Establishes the isometry (norm-preservation) properties of Gauss–Legendre collocation used in Lemma 3.1."}],"review_version":1}