{"id":"3a177bce-1924-488e-9b4b-00a77a72da8f","arxiv_id":"2508.10126","paper_version":1,"verdict":"UNVERDICTED","confidence":"LOW","novelty_score":6.0,"correctness_risk":"unknown","formal_verification":"none","parameter_count":1,"one_line_summary":"A tensor-based DMD using the star-M-product is proposed, achieving equal or better accuracy at lower storage and computational cost.","lead":"This paper proposes a tensor-based version of dynamic mode decomposition (DMD) using the star-M-product, claiming better compression and lower cost than matrix-based DMD. It also introduces a randomized streaming variant and draws connections to existing DMD frameworks.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Storage comparison may undercount tensor-method parameters, threatening the claimed Pareto improvement.","rationale":"The reader's verdict was UNVERDICTED because the full text was unavailable. My stress-test does not overturn that; rather, it identifies a specific, load-bearing condition that must hold for the abstract's central claim to be true. The reader's weakest assumption concerned the choice of M and faithful tensor representation; my concern is complementary and more concrete: even if the tensor representation is faithful, the claimed Pareto improvement over standard DMD depends on an apples-to-apples storage accounting. Standard DMD's storage is straightforward, but the tensor method's storage includes M and the factors; if the comparison omits M or uses different truncation ranks, the central claim is unsupported. This warrants a CONDITIONAL recommendation: the paper should be accepted only if the numerical section provides a transparent parameter count and M is not selected using test data. I agree with the reader's cautious stance, hence partial agreement.","tokens_in":616,"tokens_out":2975,"duration_ms":36815,"concrete_test":"Retrieve the numerical experiments from the full text and reconstruct the storage counts: for each tested example, list the total number of stored real scalars for (i) standard DMD (mode matrix entries + eigenvalues) and (ii) the star-M-DMD (all tensor factor entries + the entries of M, unless M is a fixed orthogonal/structured matrix with negligible storage or is not needed at deployment). Then, for each example, compare accuracy at equal storage values. If the tensor method's accuracy is not at least as good as DMD's at every matched storage point, or if including M changes the ranking, the abstract's claim of 'same storage' advantage fails. Additionally, check the experimental protocol for how M is chosen: if M is fitted to the training data and not re-derived in the streaming experiments, rerun with M fixed prior to any data to rule out selection bias.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The abstract's central claim is that the proposed star-M-product tensor DMD achieves 'equal or better accuracy for the same storage compared to the standard DMD' and is 'more efficient to compute.' For this claim to be meaningful, 'storage' must count all parameters the tensor method actually uses. Standard DMD stores an n×r mode matrix plus r eigenvalues (or their real-scalar equivalents), totaling roughly nr + r real scalars. The tensor method, as described, stores the tensor factors of the star-M decomposition and likely also the transformation M. If M is a dense n×n matrix and is counted as part of the model, the reported 'same storage' comparison could be off by O(n²), making the Pareto improvement illusory on problems where n is not tiny. Alternatively, if M is fixed ad hoc as a known transform (e.g., identity or Fourier), the comparison may be fair only when M is truly cost-free in storage and computation, which is not stated. The abstract also does not specify how M is selected. If M is learned from the same data used to evaluate accuracy, the claim of generalizability to the streaming setting is unsupported; the experiments would be vulnerable to overfitting. Thus, the central empirical assertion depends on a precise and complete accounting of stored parameters, and on M being chosen in a principled, non-test-set-dependent way. The abstract omits these details, so the headline claim is not yet established.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a tensor-based extension of dynamic mode decomposition (DMD) built on the star-M-product tensor algebra. The abstract claims that the method offers excellent compression, computational efficiency, connections to standard and physics-informed DMD, a randomized streaming algorithm, and numerical results showing equal or better accuracy for the same storage compared to standard DMD on the tested examples. No derivations, equations, or numerical data appear in the abstract, so the technical content and evidence for these claims are not available in the reviewed text.","tokens_in":953,"tokens_out":1592,"duration_ms":19567,"significance":"If the claims are substantiated in the full manuscript, the contribution would be relevant to tensor-based data-driven modeling, particularly for problems where state snapshots naturally form tensors. The proposed Pareto improvement over standard DMD—better or equal accuracy for the same storage with higher computational efficiency—would be a meaningful advance, and a randomized streaming variant would extend its practical utility. However, the abstract alone provides no verifiable derivations, complexity analyses, or experimental data. The significance assessment is therefore conditional on the full text providing rigorous proofs, complete storage/performance accounting, and reproducible numerical experiments.","major_comments":[{"comment":"The claim of \"equal or better accuracy for the same storage\" is not supported by any storage accounting. Standard DMD stores an n-by-r mode matrix plus r eigenvalues (or equivalent real scalars), O(nr). The tensor method's storage must include all tensor factors and the transformation matrix M. If M is dense n-by-n, the comparison is off by O(n^2) unless M is excluded or is cost-free (e.g., identity or fixed transform). The abstract does not state how M is stored or selected, so the Pareto claim is unverified.","section":"Abstract, central claim"},{"comment":"The assertion that the method is \"more efficient to compute\" lacks any complexity comparison. The abstract mentions detailing computational costs, but no such details are given. A specific, complete complexity analysis relative to standard DMD—including the cost of applying or constructing M—must be presented before the efficiency claim can be assessed.","section":"Abstract, computational efficiency"},{"comment":"The randomized streaming algorithm is mentioned without specification of the streaming model, memory limits, update rules, or error guarantees. In particular, it is unclear how M is obtained or updated in streaming mode and whether the same data are used for both learning M and evaluating accuracy, which could lead to overfitting and invalidate generalizability claims.","section":"Abstract, streaming setting"},{"comment":"The claimed connections to standard DMD and physics-informed DMD are stated but not demonstrated. Without explicit equations showing how the star-M-product DMD reduces to or generalizes these methods, the relationship remains an assertion. The full manuscript must provide the derivations and identify any assumptions needed for the connections to hold.","section":"Abstract, connections to existing frameworks"}],"minor_comments":[{"comment":"The notation \\(\\star_{\\boldsymbol{M}}\\) is undefined in the abstract. A brief definition or reference is needed for readers unfamiliar with the product.","section":"Abstract, notation"},{"comment":"No references are cited in the abstract. Key prior work on DMD, physics-informed DMD, and the star-M product should be cited to position the contribution.","section":"Abstract, references"},{"comment":"The phrase \"on these examples\" appropriately limits the numerical claim, but it should be made explicit that the reported improvements are empirical and not guaranteed in general. Also, the number and nature of the examples should be stated.","section":"Abstract, empirical scope"}],"recommendation":"uncertain","confidential_remarks":"The submission as provided is abstract-only, which is unusual for a full review. The central claims—accuracy, storage, and efficiency—are plausible but entirely unverified from the abstract. If the full manuscript supplies rigorous derivations, a complete parameter/storage count including M, and reproducible experiments, the paper may be suitable. At present, insufficient evidence is available to reach a positive verdict; I recommend requesting the full text before a substantive decision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"This is an abstract-only review, so my verdict is provisional. The core idea is a genuine extension: using the star-M-product tensor framework inside DMD, plus a randomized streaming variant, and drawing explicit links to physics-informed DMD. That is enough novelty to take seriously. The abstract claims equal-or-better accuracy at the same storage and lower compute cost, which is a concrete, falsifiable Pareto claim. If it holds up, it is a useful incremental contribution to numerical linear algebra, not a field redefinition.\n\nWhat I like: the method is framed as a natural tensor generalization, not a toy. The randomized streaming algorithm is a practical addition that could matter for large-scale data. The acknowledgments of connections to standard DMD and physics-informed DMD suggest the authors are positioning the work honestly rather than overselling.\n\nThe soft spots are exactly where the abstract is silent. First, storage accounting: the headline claim \"same storage\" must include every stored parameter. The star-M decomposition stores tensor factors, and unless M is a fixed, storage-free transform (e.g., identity or a named fast transform), a dense n×n M would add O(n²) and wreck the comparison. The abstract does not say whether M is stored or how it is selected. Second, if M is learned from the same training data used to evaluate accuracy, the streaming claim could be overfit territory. The stress-test note flagging this is on target. That said, the concern is a question, not a demonstrated flaw; many papers in this area fix M ahead of time or use a diagonalizable M with negligible storage. The lack of equations and experimental detail in the abstract means I cannot tell.\n\nThe paper is short on verifiable detail in this form, but that is normal for an abstract. The underlying tensor algebra is established, and the authors are known in this community. This deserves a serious referee who can check the parameter count, the computational cost tables, and the numerical comparisons. If those hold, it is citable. If the storage comparison is sloppy, that is fixable in revision.\n\nWho is this for? Researchers in model reduction and data-driven dynamics who work with tensor-valued states. A reading group would get a good discussion out of it, especially on the trade-off between representation flexibility and parameter overhead.\n\nRecommendation: send it to peer review. The novelty is real, the claim is concrete, and the potential flaw is precisely the kind of thing referees should catch.","headline":"The star-M-tensor DMD idea is new and plausible enough to referee; the main open question is whether the storage comparison honestly counts the transformation matrix M.","tokens_in":1351,"tokens_out":912,"would_cite":true,"duration_ms":12023,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["15A69","65F30"],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims a tensor-product version of dynamic mode decomposition that matches or beats standard DMD in accuracy at the same storage, while being cheaper to compute, on the examples tested.","keywords":["dynamic mode decomposition","star-M-product","tensor-tensor product","data-driven dynamics","randomized algorithm","streaming computation","tensor compression","physics-informed DMD"],"falsifier":"Pick a dynamical system whose states are naturally vectors, force the snapshots into third-order tensors by padding, and compare $\\star_{\\boldsymbol{M}}$-DMD against standard DMD at equal storage using the paper's fixed, un-tuned choice of $\\boldsymbol{M}$: if the tensor method is strictly less accurate, or if finding a competitive $\\boldsymbol{M}$ costs more computation than the matrix SVD it replaces, the claimed Pareto improvement fails for that setting. The most direct check on the paper's own examples is reproducing the reported accuracy-at-equal-storage numbers from the public data.","tokens_in":596,"feed_emoji":"🧮","tokens_out":12660,"duration_ms":127237,"temperature":0.7,"pith_summary":"Dynamic mode decomposition (DMD) extracts the dominant linear dynamics of a system from snapshots of its state. This paper argues that when states are naturally tensors — fields on a grid, for example — the usual matrix formulation throws away structure that a tensor product can exploit. Its central claim is that DMD built on the $\\star_{\\boldsymbol{M}}$-product matches or beats standard DMD in accuracy at the same storage and computes the result more cheaply on the examples tested. The paper also derives a streaming, randomized variant for settings where snapshots arrive one at a time, and shows how the construction sits inside both classical DMD and physics-informed DMD. If right, practitioners get a drop-in alternative that captures the same dynamics from fewer stored numbers.","feed_headline":"Match DMD accuracy while cutting storage and compute","feed_subtitle":"Keeping state snapshots in tensor form lets the same dynamics be captured from fewer stored numbers.","key_machinery":"The $\\star_{\\boldsymbol{M}}$-product: a tensor-tensor product for third-order tensors defined by transforming both tensors along their tube (mode-3) dimension with a fixed invertible matrix $\\boldsymbol{M}$, multiplying the resulting faces as ordinary matrices, and transforming back. It carries the entire argument: DMD's snapshot matrix, its low-rank SVD, and the projected dynamics are all redefined with this product, so the data never has to be flattened, which is exactly what produces the storage savings and the efficiency gain.","core_discovery":"The core discovery is that dynamic mode decomposition can be carried out entirely inside the $\\star_{\\boldsymbol{M}}$-product tensor framework without giving up accuracy. Instead of flattening tensor-valued snapshots into vectors and working with one large data matrix, the method keeps snapshots as third-order tensors and replaces DMD's matrix multiplications with $\\star_{\\boldsymbol{M}}$-products: the tensors are transformed along their third dimension by a fixed invertible matrix $\\boldsymbol{M}$, multiplied face-by-face as ordinary matrices, and transformed back. The tensor structure is preserved and exploited, and since the transformation is applied once up front, the tensor SVD costs le","pith_inferences":["The paper fixes the transformation matrix $\\boldsymbol{M}$ rather than learning it from the data; a natural extension the authors do not pursue is a data-dependent $\\boldsymbol{M}$, which could push the storage savings further on systems where a single fixed choice is suboptimal.","The framework only requires a tensor product and an SVD, so the same construction should transfer to other snapshot-based methods — Koopman-style approximations or reduced-order models of grid-valued PDE states — whenever the state is naturally a tensor; the paper demonstrates only the DMD case.","If the accuracy parity holds beyond the paper's examples, the working rule that emerges is: leave snapshot data in tensor form instead of flattening it, a principle that could guide other data-driven linear-algebra pipelines.","The streaming random variant implies the method is aimed at long-horizon sensing, and a natural test would be measuring how the sketch size must scale with tensor dimensions to preserve the accuracy parity — a scaling law the paper does not state."],"forward_implications":["On problems with tensor-valued states, $\\star_{\\boldsymbol{M}}$-DMD can be used as a direct replacement for standard DMD: at the same storage it matches or improves accuracy, and the computation is cheaper.","The formal link to physics-informed DMD means the tensor formulation can carry known physical constraints into the fit without losing the storage advantage.","The randomized streaming variant extends the savings to settings where snapshots cannot all be held in memory, the regime where standard DMD's storage costs are most severe.","The paper's cost analysis gives users an exact way to compute the flop and storage ratio for their own tensor dimensions before they implement the method."],"supporting_citations":[],"fun_headline_variants":["Tensor DMD: same accuracy, lower storage and compute","Star-M tensor DMD matches accuracy, uses less","Tensor DMD: compress more, compute faster","Same DMD accuracy, tensor-sized storage","Tensor DMD: efficient, accurate, and compact"],"cache_read_input_tokens":2816,"weakest_assumption_plain":"The approach relies on choosing, in advance, a single transformation matrix (the $\\boldsymbol{M}$ in the $\\star_{\\boldsymbol{M}}$-product) such that the tensor multiplication it defines faithfully preserves the system's dynamics; if no such fixed $\\boldsymbol{M}$ exists for a given problem, the demonstrated accuracy and storage advantages should not be expected to carry over.","fun_headline_variants_meta":{"raw":{"variants":["Tensor DMD: same accuracy, lower storage and compute","Star-M tensor DMD matches accuracy, uses less","Tensor DMD: compress more, compute faster","Same DMD accuracy, tensor-sized storage","Tensor DMD: efficient, accurate, and compact"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000212,"raw_usage":{"total_tokens":1215,"prompt_tokens":665,"completion_tokens":550,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":409,"completion_tokens_details":{"reasoning_tokens":476}},"tokens_in":409,"tokens_out":550,"duration_ms":6925,"temperature":1.0,"reasoning_tokens":476,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T20:36:37.205748+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Pick a dynamical system whose states are naturally vectors, force the snapshots into third-order tensors by padding, and compare $\\star_{\\boldsymbol{M}}$-DMD against standard DMD at equal storage using the paper's fixed, un-tuned choice of $\\boldsymbol{M}$: if the tensor method is strictly less accurate, or if finding a competitive $\\boldsymbol{M}$ costs more computation than the matrix SVD it replaces, the claimed Pareto improvement fails for that setting. The most direct check on the paper's own examples is reproducing the reported accuracy-at-equal-storage numbers from the public data.","supporting_citations":[],"review_version":1}