{"id":"de7b1365-abcd-4834-abcc-0577cac55d3e","arxiv_id":"1908.04741","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"Tensor-train compression makes EDMD and its coherent-set variant CCA tractable for high-dimensional systems with huge product bases.","lead":"The authors show how to analyze high-dimensional dynamical systems with very large basis sets by storing the data as tensor trains and solving a small reduced eigenvalue problem. The approach extends to coherent-set detection and is tested on protein simulations and the ABC flow.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Convergence proof (Prop. 1) uses δ := max_k δ_k where the induction requires min_k δ_k; the written argument does not establish Corollary 1.","rationale":"The reader's conditional verdict is reasonable, and the identified low-rank/HOCUR premise is a genuine practical limitation: the method's efficiency depends on the transformed data tensor having small TT ranks, which is not guaranteed, and HOCUR lacks convergence guarantees. However, the more immediately load-bearing issue is in the central convergence proof. Proposition 1 is the only result giving the infinite-data physical interpretation of AMUSEt, and its induction as printed uses δ = max_k δ_k. Davis–Kahan requires a uniform lower bound on all spectral gaps, so the coefficient must be 1/δ_k; with δ = max, the bound in (16) is not implied whenever δ_k < max_j δ_j. This is not a question of consensus or of missing empirical validation; it is an internal inconsistency in the proof. The defect is plausibly a typo—δ should be min_k δ_k—and the theorem may survive a corrected proof, but as published Corollary 1 is unproven. A secondary mismatch is that the theorem assumes fixed ranks r_k, while the numerical experiments use adaptive rank truncation with threshold ε; the asymptotic statement therefore does not cover the exact algorithm as run. These issues do not invalidate the algorithmic construction or the benchmarks, but they strengthen the case for a conditional verdict pending revision. The proposed test—re-running the induction with δ = min and testing a lopsided-gap example—would settle whether the proof is merely mistyped or genuinely flawed.","tokens_in":21164,"tokens_out":13222,"duration_ms":135922,"concrete_test":"Re-derive the induction in Appendix A with δ := min_k δ_k. Concretely, verify Step 1: with δ = δ_min, confirm that the bound ‖Ĉ00_{V1}−C00_{V1}‖ ≤ εδ / [2^{p−1}(2βδ^{-1}+1)^{p−1}] implies d(G1,Ĝ1) ≤ ε / [2^{p−1}(2βδ^{-1}+1)^{p−1}] via Davis–Kahan, and that the same replacement makes (16)–(17) valid at each step. Then construct a 2×2 counterexample with δ_1 = 1, δ_2 = 10 and run the printed Step 1 with δ = max to show the printed inequality fails. If the corrected proof goes through, the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central convergence theorem, Proposition 1 (Section 3.4), has a concrete defect in the induction constants. After defining the spectral gaps δ_k > 0 of the Galerkin matrices C00_{U_k}, the authors set δ := max_{1≤k≤p} δ_k (Appendix A, after Lemma 3). The Davis–Kahan steps in equations (15)–(17) require a uniform lower bound on the gaps: the coefficient must be 1/δ_k, and the error tolerance must scale with δ only if δ ≤ δ_k for every k. With δ = max, one has δ ≥ δ_k, so 1/δ ≤ 1/δ_k and the stated bound d(G_k, Ĝ_k) ≤ [2δ^{-1}‖Ĉ00_{S_k}‖ + 1] d(H_k, Ĥ_k) + δ^{-1}‖Ĉ00_{U_k}−C00_{U_k}‖ is not implied by the true Davis–Kahan bound with 1/δ_k. For a concrete failure, if δ_1 = 1 and δ = 10, Step 1 yields a claimed bound that is ten times too small; the induction cannot propagate the stated ε-tolerance. Thus Corollary 1 is not established by the proof as written. This is load-bearing because Proposition 1 is the only result providing the infinite-data physical interpretation of AMUSEt. The defect appears to be a correctable typo (δ should be min_k δ_k), but as printed the theorem is unproven. A secondary gap: the theorem assumes fixed ranks r_k, whereas the numerical implementation uses adaptive rank truncation via the relative threshold ε; the asymptotic claim does not directly cover the implemented algorithm.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper combines tensor-train (TT) decompositions with the AMUSE formulation of extended dynamic mode decomposition (EDMD) and canonical correlation analysis (CCA). It introduces a greedy higher-order CUR (HOCUR) scheme for representing transformed data tensors, derives a reduced matrix eigenvalue problem (Eq. (13)) whose spectrum coincides with the EDMD/CCA problem on the selected subspace, proves a subspace interpretation (Lemma 2) and an infinite-data convergence result (Proposition 1, Corollary 1) for the exact TT decomposition, and reports numerical experiments on deca-alanine, NTL9, and ABC flow.","tokens_in":21566,"tokens_out":9290,"duration_ms":97672,"significance":"The derivation is elegant and the algorithmic idea is valuable: for product bases the reduced matrix can be assembled by contracting only a few TT cores, avoiding dense N x N Galerkin matrices. The numerical results are encouraging, showing close agreement with Markov state models and substantial scalability (up to 3^20 basis functions). If the convergence proof is repaired, the paper provides a useful and principled tensor framework for transfer operator approximation. The main current weaknesses are that the central theorem is not proven as printed and the practical HOCUR/adaptive-rank algorithm is not covered by the theory.","major_comments":[{"comment":"In the proof of Proposition 1, the constant is defined as delta := max_k delta_k, where delta_k are the spectral gaps of the matrices C00_U_k. The Davis-Kahan applications in Eqs. (15)-(16) require the inverse of the gap of the specific matrix C00_U_k, i.e. a factor 1/delta_k; with delta = max_k delta_k the coefficient 1/delta is smaller than 1/delta_k, so the stated bounds are not implied. Consequently Eq. (17) and Corollary 1 are not established as written. Replacing delta by min_k delta_k would repair the argument, but the printed proof needs this correction.","section":"Section 3.4 / Appendix A (Proposition 1)"},{"comment":"Proposition 1 assumes fixed ranks r_k and the exact TT decomposition (9). The numerical method in Section 3.2 and the experiments in Section 4 use adaptive rank truncation with a relative tolerance epsilon (e.g., epsilon = 10^-3 in Figure 6c) and the HOCUR approximation from Algorithm 3, for which no convergence or error bound is provided. Therefore the infinite-data subspace interpretation is not shown for the actually implemented algorithm. The authors should either extend the analysis to these cases or explicitly state that the convergence result applies only to the exact fixed-rank decomposition.","section":"Section 3.4 vs. Sections 3.1-3.2 and 4"},{"comment":"The greedy HOCUR iteration in Algorithm 3 is described as running until convergence, but no convergence criterion, fixed-point property, or error bound is given. Since all large-scale examples in Section 4 rely on HOCUR, the efficiency and accuracy claims rest on an unverified assumption that the iteration produces a small fixed-rank TT approximation. A formal analysis would be preferable; at minimum the authors should state clearly that this is a heuristic and report the observed rank growth or stopping behavior.","section":"Section 3.1 (Algorithm 3)"}],"minor_comments":[{"comment":"In Eq. (13), the symbol U_x should read U_X for consistency with the surrounding notation.","section":"Section 3.2, Eq. (13)"},{"comment":"The text refers to the 'HOSVD method' although earlier sections call the procedure 'global SVD'; please align the terminology.","section":"Section 4.1"},{"comment":"Algorithm 3 does not specify a stopping criterion for the outer loop; please state that the iteration is run for N sweeps and describe which convergence monitor is used.","section":"Section 3.1, Algorithm 3"},{"comment":"The color scale in Figure 6(c) is not explained in the caption; please add the mapping from color to t2 values.","section":"Figure 6"}],"recommendation":"major_revision","confidential_remarks":"The delta = max versus delta = min issue in Proposition 1 appears to be a correctable typo, and I would not reject on that basis alone. The larger open point is the status of HOCUR: either provide some guarantee or clearly delimit the theoretical claims to the exact decomposition. The paper is within scope and the numerical results are compelling; after the proof correction and appropriate caveats, it could be acceptable."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Bottom line: the paper is worth reading and mostly earns its claims. It scales EDMD and coherent-set CCA to tensor product bases far beyond dense matrix limits, and the NTL9 demonstration with a 3^20 basis and about 1.4 million frames is the right kind of stress test. The main thing to know before citing the convergence result: Proposition 1 has a max-vs-min spectral gap slip, so as printed the infinite-data corollary is not proven.\n\nWhat is new: AMUSEt for EDMD, the tensorized CCA variant, and the HOCUR construction for transformed data tensors are not restatements of [31,32,34]. The reduction in Eq. (13) is clean, and Lemma 2 gives the right subspace interpretation. The benchmarks compare against external MSM references; the timescales and contact maps line up. That is real evidence the method works at scale. The ABC flow example is more of a sanity check, but fine.\n\nThe soft spot is the convergence analysis. The stress-test note is correct: they define δ := max_k δ_k, while every Davis-Kahan step needs a lower bound on the gap, i.e. δ := min_k δ_k. With max, the error tolerances in Steps 1 and k are too large and the induction does not close. This is almost certainly a one-line typo, not a structural flaw, but Corollary 1 is not established by the printed proof. The abstract's claim that the infinite-data limit is established should be qualified until this is fixed.\n\nTwo smaller issues. The theorem covers only the exact decomposition (9) at fixed ranks, while the implementation uses adaptive relative truncation and HOCUR, so the formal result is narrower than the numerical section. HOCUR itself has no convergence or error bound; the method's efficiency depends on unproven low TT-rank of the transformed data tensor. That is a genuine limitation, though honestly framed. Also, the benchmarks have no error bars, so the agreement with references is suggestive rather than quantitative. Minor.\n\nCitation pattern is fine. There is heavy self-citation, but to the building blocks (AMUSE, TT-SVD, CUR, prior tensor EDMD), and the new algorithmic content is distinguishable. The paper is for anyone who wants to use high-dimensional product bases with Koopman or transfer-operator methods, and for tensor numerical linear algebra readers. It deserves a serious referee. Recommendation: send to peer review; ask for the δ fix, a qualified infinite-data claim, and error bars or at least a statement of variability in the experiments.","headline":"A practical and mostly sound tensor-train EDMD/CCA paper with impressive molecular benchmarks; the convergence proof has a fixable max-vs-min spectral gap slip that should be corrected.","tokens_in":22081,"tokens_out":4650,"would_cite":true,"duration_ms":48904,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["15A69","37L65","37M10","62H20","92C40"],"pacs":[],"model":"deepseek-v4-flash","headline":"Tensor-train decompositions let EDMD and CCA run on high-dimensional data without forming the full basis matrix.","keywords":["Koopman operator","extended dynamic mode decomposition","tensor train format","canonical correlation analysis","coherent sets","metastable dynamics","high-dimensional systems","forward-backward operator"],"falsifier":"Take a system for which the exact TT ranks of $\\Psi(X)$ at a fixed accuracy are known to be large, such as a tensor-product basis on strongly interacting coordinates with no low-rank structure, compute AMUSEt with a deliberately small rank cap, and compare the slowest eigenvalues against a dense EDMD solution on the same data; if the reduced spectrum diverges while dense EDMD is still feasible, the rank assumption underlying the method's efficiency is violated and the advantage over dense EDMD disappears as the ranks grow.","tokens_in":20982,"feed_emoji":"🧮","tokens_out":7452,"duration_ms":75697,"temperature":0.7,"pith_summary":"Koopman-based analysis of dynamical systems is usually limited by the cost of building and diagonalizing large matrices of basis-function evaluations. This paper removes that bottleneck for tensor-product basis sets by storing the data tensor, the evaluations of all product basis functions at all snapshots, in the tensor-train (TT) format, and deriving tensor-based versions of the EDMD and canonical correlation analysis algorithms. The key claim is that these tensor algorithms, AMUSEt for EDMD and tensor CCA for coherent sets, yield a reduced matrix eigenvalue problem with the same spectrum as the full problem on a fixed-rank subspace, and that this subspace converges to a deterministic Galerkin subspace as the number of snapshots grows. If correct, this makes slow metastable dynamics and coherent sets computable for high-dimensional systems with very large basis sets, as the paper demonstrates on peptide simulations and the ABC flow.","feed_headline":"Tensor trains speed up dynamical-system analysis in high dimensions","feed_subtitle":"AMUSEt recovers the same spectra as EDMD and CCA on a provably convergent low-rank subspace.","key_machinery":"The central object is the transformed data tensor $\\Psi(X)$, an order-$(p+1)$ tensor formed by evaluating every product basis function $\\psi_1(x_k)\\otimes\\cdots\\otimes\\psi_p(x_k)$ at every snapshot and appending the snapshot index. Its TT decomposition represents this tensor as a chain of small cores with linking ranks $r_k$, a low-parameter representation of a high-dimensional array. AMUSEt uses a global SVD of the TT tensor, namely a segment $U$, a diagonal matrix $\\Sigma$, and an orthonormal $V$, to build the reduced matrix $M$; orthonormality of the cores makes almost all contractions cancel, so only the final cores and $\\Sigma^{-1}$ enter. The paper also introduces HOCUR, a greedy higher-order CUR decomposition of $\\Psi(X)$ based on maximum-volume index selection, which avoids building the full tensor. The convergence argument is carried by the recursive subspaces $F_k$, obtained by compressing $U_k=F_{k-1}\\otimes V_k$ to its leading $r_k$ eigenvectors, and by Davis-Kahan type perturbation estimates showing that the distance $d(F_k,\\hat{F}_k)$ shrinks as more data are used.","core_discovery":"The paper's central claim is that for a tensor-product trial space, EDMD and forward-backward/CCA analysis can be carried out on a low-rank tensor-train representation of the transformed data tensor $\\Psi(X)=\\sum_{k}\\psi_1(x_k)\\otimes\\cdots\\otimes\\psi_p(x_k)\\otimes e_k$, without ever forming the full Galerkin matrices $C_{00}, C_{01}, C_{10}, C_{11}$. The tensor algorithm AMUSEt first computes a global SVD of $\\Psi(X)$ in TT form; because the TT cores are orthonormal, the reduced matrix $M=V_X^{\\top}M_Y^{\\top}U_X\\Sigma_X^{-1}$ contains the same spectrum as the standard EDMD eigenproblem on the subspace spanned by the left singular vectors. Analogously, tensor CCA constructs a reduced matrix from separate global SVDs of $\\Psi(X)$ and $\\Psi(Y)$, yielding eigenfunctions of the forward-backward operator and hence coherent sets. The paper proves that, under ergodicity and a singular-value gap assumption, the data-driven subspace converges to a deterministic fixed subspace $F_p$ in the infinite-data limit, so that AMUSEt converges to a Galerkin approximation of the Koopman or forward-backward operator on that subspace.","pith_inferences":["An implicit consequence is that the fixed-rank subspace $F_p$ could serve as a built-in feature-selection or coarse-graining layer: the algorithm decides which products of univariate functions matter, making it a data-driven alternative to hand-crafted collective variables.","A natural testable extension is to use the AMUSEt spectrum as a cheap surrogate during basis construction, adaptively adding univariate functions while monitoring whether the slowest timescales change and avoiding a full refit.","Because the convergence proof is stated for the exact TT decomposition, extending it to bound the error introduced by the greedy HOCUR approximation would be a direct next step; until then, practical users may need residual-based rank-adaptivity or stopping criteria.","If TT ranks turn out to be large for strongly coupled variables, combining HOCUR with nonlinear feature maps or kernel evaluations is a plausible route, and the subspace interpretation would likely carry over."],"forward_implications":["Users can apply AMUSEt to a tensor-product basis with very many basis functions without storing the full Galerkin matrices, since the reduced matrix never exceeds the number of snapshots.","For reversible systems the resulting slowest implied timescales respect the variational principle, so smaller truncation thresholds select larger subspaces and give closer approximations.","For non-reversible or time-dependent dynamics, tensor CCA extracts finite-time coherent sets directly from trajectory data, as demonstrated on the ABC flow.","In the infinite-data limit the method is a Galerkin approximation of the Koopman or forward-backward operator on the deterministic subspace $F_p$, giving the procedure a physical interpretation.","The HOCUR variant can process data sets and basis sizes that are intractable for direct construction, as shown on a high-dimensional protein data set with millions of frames."],"supporting_citations":[{"why":"Introduces extended dynamic mode decomposition, the EDMD formulation this paper tensorizes.","marker":"[6]"},{"why":"Provides the Galerkin approximation of transfer operators and the EDMD eigenvalue problem that AMUSEt solves in reduced form.","marker":"[7]"},{"why":"Gives the variational principle connecting EDMD eigenvalues to the leading eigenvalues of the transfer operator, used to interpret the slow timescales.","marker":"[8]"},{"why":"Establishes canonical correlation analysis as a data-based approximation of the forward-backward operator, the basis of tensor CCA.","marker":"[13]"},{"why":"Defines the tensor-train format used to represent the transformed data tensors.","marker":"[20]"},{"why":"Shows how transformed data tensors for product bases can be written in TT form, the starting point for AMUSEt.","marker":"[26]"},{"why":"Supplies the AMUSE algorithm that AMUSEt adapts to tensor trains.","marker":"[33]"},{"why":"Provides the global SVD for data matrices in TT format that AMUSEt relies on.","marker":"[34]"},{"why":"Supplies the greedy higher-order CUR decomposition used by the HOCUR algorithm.","marker":"[35]"}],"fun_headline_variants":["Tensor trains tame high-dimensional Koopman analysis","Low-rank tensors for metastable and coherent sets","TT format accelerates dynamical system analysis","Tensor method computes spectra of high-dimensional systems","Koopman meets tensor trains for large-scale dynamics"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method's speed and memory advantage rests on the transformed data tensor admitting a small tensor-train rank, and on the greedy HOCUR iteration actually finding that small-rank representation; when that fails, the reduced eigenvalue problem no longer captures the dynamics of interest.","fun_headline_variants_meta":{"raw":{"variants":["Tensor trains tame high-dimensional Koopman analysis","Low-rank tensors for metastable and coherent sets","TT format accelerates dynamical system analysis","Tensor method computes spectra of high-dimensional systems","Koopman meets tensor trains for large-scale dynamics"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000289,"raw_usage":{"total_tokens":1691,"prompt_tokens":940,"completion_tokens":751,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":556,"completion_tokens_details":{"reasoning_tokens":682}},"tokens_in":556,"tokens_out":751,"duration_ms":7272,"temperature":1.0,"reasoning_tokens":682,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T13:48:15.206653+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a system for which the exact TT ranks of $\\Psi(X)$ at a fixed accuracy are known to be large, such as a tensor-product basis on strongly interacting coordinates with no low-rank structure, compute AMUSEt with a deliberately small rank cap, and compare the slowest eigenvalues against a dense EDMD solution on the same data; if the reduced spectrum diverges while dense EDMD is still feasible, the rank assumption underlying the method's efficiency is violated and the advantage over dense EDMD disappears as the ranks grow.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Introduces extended dynamic mode decomposition, the EDMD formulation this paper tensorizes."}],"review_version":1}