{"id":"b769bd0e-0c3e-45ef-9636-eecc41c0fb9c","arxiv_id":"2606.26900","paper_version":1,"verdict":"UNVERDICTED","confidence":"LOW","novelty_score":7.0,"correctness_risk":"unknown","formal_verification":"none","parameter_count":0,"one_line_summary":"Presents an incremental TT compression algorithm operating directly on streaming TT-formatted data via core-wise operations, with error bounds and application to ROM construction from low-rank snapshots.","lead":"The paper introduces a deterministic incremental algorithm that compresses streaming tensor-train data directly on TT cores using projection, orthogonalization, and enrichment steps. A smart generalist might read it for potential efficiency gains in large-scale simulations where data arrives already compressed rather than as full tensors.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.3","headline":"POD-analogue claim rests on core-wise operations exactly replicating the dominant subspace of full-order POD without reconstruction","rationale":"The reader's weakest assumption directly identifies the load-bearing step for the strongest claim. The full manuscript supplies the algorithmic details and error bounds absent from the abstract, but does not remove the dependence on incoming TTs already encoding the POD-relevant information; the concrete test above isolates whether that dependence is benign.","tokens_in":1803,"tokens_out":366,"duration_ms":32604,"concrete_test":"Generate a small parametric linear system (e.g., 100×100 matrix with 20 parametric snapshots), compute both (i) standard POD on the dense snapshot matrix and (ii) the incremental TT procedure on the same snapshots first converted to exact TT format at two different tolerances (10^{-4} and 10^{-8}); compare the leading singular values and the ROM projection error on a validation parameter; if the TT-derived subspace misses more than one POD mode or the ROM error exceeds the dense case by >5%, the analogue does not hold.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim requires that the sequence of core-wise projections, residual orthogonalizations, and adaptive enrichments on incoming TT cores produces an accumulated TT whose range is equivalent (up to the tolerance) to the POD subspace obtained from the vectorized full snapshots. Section 4 derives this by showing that the accumulated cores span the same information as the Gram matrix of the snapshots; however, this equivalence implicitly assumes that the TT ranks of the incoming tensors already contain all directions that would be selected by POD on the dense data. If an incoming TT has been compressed below the POD truncation threshold, the orthogonalization step cannot recover the missing components, breaking the claimed correspondence.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.3","summary":"The paper develops a deterministic incremental TT compression algorithm that updates an accumulated TT representation from streaming TT-formatted tensors via core-wise projection, residual orthogonalization, and adaptive enrichment, without reconstructing dense tensors. It derives approximation error bounds, establishes that the accumulated TT is a compressed analogue of POD on full-order snapshots (enabling direct ROM construction from low-rank data), and demonstrates the approach on parametric radiative transfer equations, reporting comparable accuracy with reduced wall time.","tokens_in":1930,"tokens_out":512,"duration_ms":27342,"significance":"If the error bounds and POD equivalence hold, the work enables efficient reduced-order modeling directly from compressed streaming data in high-dimensional applications such as kinetic equations and quantum systems, avoiding prohibitive reconstruction costs. The deterministic core-wise operations and explicit bounds are strengths that support reproducibility and practical deployment in numerical analysis contexts.","major_comments":[{"comment":"§4: The derivation that accumulated cores span the same information as the Gram matrix of snapshots (and thus yield a POD analogue) implicitly assumes incoming TT ranks already contain all directions selected by POD on dense data. If an incoming TT has been pre-compressed below the POD truncation threshold, the orthogonalization step cannot recover missing components, which would break the claimed equivalence; this assumption is load-bearing for the central ROM claim but is not explicitly stated or tested.","section":"§4"},{"comment":"§3.2, Algorithm 1 and Theorem 3.1: The error bound is stated in terms of a prescribed tolerance ε, but the interaction between ε, adaptive rank enrichment, and the underlying physics (e.g., how tolerance choice affects retention of physically relevant modes) is not analyzed; without this, it is unclear whether the bound remains useful when TT ranks grow or when data exhibit slow singular-value decay.","section":"§3.2"}],"minor_comments":[{"comment":"Notation for TT cores and unfolding operators is introduced without a consolidated table; a small reference table would improve readability when following the core-wise projection steps.","section":null},{"comment":"Numerical experiments section reports wall-time reductions but does not include a direct comparison of the accumulated TT ranks versus ranks obtained from batch POD on reconstructed snapshots; adding this would strengthen the POD-analogue validation.","section":null}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":{"model":"grok-4.3","summary":"We thank the referee for the careful and constructive review. The two major comments identify implicit aspects of the analysis that warrant explicit clarification. We address each point below and indicate the planned revisions.","responses":[{"response":"We agree that the assumption is implicit. The algorithm and the claimed POD equivalence are formulated for streaming TT-formatted data; therefore the accumulated representation is exactly the POD analogue of the given (already compressed) snapshots. If an incoming TT has been truncated below the POD threshold of the underlying dense data, lost directions cannot be recovered, which is an inherent limitation of operating on compressed inputs. We will revise Section 4 to state this assumption explicitly, qualify the ROM claim accordingly, and note that the method still yields the optimal low-rank representation of the information actually present in the stream.","revision_made":"yes","referee_comment":"[§4] §4: The derivation that accumulated cores span the same information as the Gram matrix of snapshots (and thus yield a POD analogue) implicitly assumes incoming TT ranks already contain all directions selected by POD on dense data. If an incoming TT has been pre-compressed below the POD truncation threshold, the orthogonalization step cannot recover missing components, which would break the claimed equivalence; this assumption is load-bearing for the central ROM claim but is not explicitly stated or tested."},{"response":"Theorem 3.1 provides a deterministic a-priori bound that holds independently of the singular-value decay rate; the adaptive enrichment step guarantees that the tolerance ε is met by increasing the relevant TT ranks. We acknowledge that a detailed examination of how ε interacts with physically important modes or with slow decay is absent. In the revision we will add a short discussion in §3.2 clarifying that (i) the bound remains valid for any decay rate, (ii) slower decay simply produces higher ranks while still satisfying the tolerance, and (iii) the numerical experiments on the radiative-transfer problem already illustrate practical behavior under the chosen ε. A fuller parametric study of ε versus modal retention lies beyond the present scope.","revision_made":"partial","referee_comment":"[§3.2] §3.2, Algorithm 1 and Theorem 3.1: The error bound is stated in terms of a prescribed tolerance ε, but the interaction between ε, adaptive rank enrichment, and the underlying physics (e.g., how tolerance choice affects retention of physically relevant modes) is not analyzed; without this, it is unclear whether the bound remains useful when TT ranks grow or when data exhibit slow singular-value decay."}],"tokens_in":1427,"tokens_out":547,"duration_ms":50460,"standing_objections":[]},"desk_editor":{"model":"grok-4.3","letter":"The paper presents an incremental algorithm for compressing a stream of TT tensors by working exclusively at the core level. It updates an accumulated TT through core-wise projection, residual orthogonalization, and adaptive enrichment, all without reconstructing dense data. The authors also link the result to a compressed form of proper orthogonal decomposition.\n\nThis approach is new in handling TT-to-TT streaming directly. Prior work often starts from dense or sketched inputs, so avoiding that step is the practical advance. The error bounds and the connection to ROM construction from low-rank streams are laid out clearly. The experiments on parametric radiative transfer show matching accuracy with reduced computation time, which supports the claims for that setting.\n\nThe soft spot is the POD analogue. The derivation shows the accumulated cores relate to the snapshot Gram matrix, but it implicitly requires that the incoming TT ranks already include all directions POD would retain. If an input TT is compressed more aggressively, missing components cannot be recovered, weakening the equivalence. The abstract does not call out this condition, though the math in section 4 appears to rest on it.\n\nThe citation pattern covers the relevant TT and streaming literature without obvious omissions. The method is deterministic and the numerical verification is present.\n\nThis paper is for numerical analysts and engineers doing reduced-order modeling on high-dimensional streaming data that arrives pre-compressed. A reader who needs to build ROMs without ever forming full snapshots will get usable steps and evidence that it works in practice.\n\nIt deserves a serious referee. The algorithmic idea is solid and the application area is important, even if the POD claim could use tighter discussion of its assumptions.\n\nI recommend sending it to peer review rather than desk rejecting it.","headline":"Incremental TT-to-TT compression works without dense reconstruction but the POD equivalence depends on input ranks being sufficient.","tokens_in":2454,"tokens_out":407,"would_cite":false,"duration_ms":45605,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.3","headline":"Streaming tensor-train data can be incrementally compressed into a compressed analogue of proper orthogonal decomposition without reconstructing full tensors.","keywords":["tensor-train format","streaming data","incremental compression","reduced-order modeling","proper orthogonal decomposition","radiative transfer equations","low-rank tensors"],"falsifier":"A numerical test in which the reduced-order model built from the accumulated TT cores produces errors larger than the prescribed tolerance when compared with a standard POD model built from the same data after full reconstruction.","tokens_in":2672,"feed_emoji":"","tokens_out":630,"duration_ms":43767,"temperature":0.7,"pith_summary":"The paper develops a deterministic algorithm that updates an accumulated tensor-train representation from new TT tensors arriving in sequence. It does so by projecting new cores onto the current basis, orthogonalizing the residual, and adaptively enriching to keep only information outside a prescribed tolerance. All steps act directly on the TT cores, so neither the incoming tensor nor the accumulated object is ever expanded to full size. The resulting accumulated representation functions as a compressed version of proper orthogonal decomposition on full-order snapshots, which in turn permits reduced-order models to be assembled from the TT cores alone.","feed_headline":"Streaming TT data yields compressed POD analogue for ROMs","feed_subtitle":"The algorithm updates accumulated TT cores directly, enabling reduced-order models without full snapshot reconstruction.","key_machinery":"Core-wise projection onto the current TT basis followed by residual orthogonalization and adaptive enrichment that updates the accumulated TT representation while operating entirely at the level of the cores.","core_discovery":"The authors present an incremental TT compression procedure that, given a new TT tensor, performs core-wise projection, residual orthogonalization, and adaptive enrichment to retain only the complementary information beyond a given tolerance. They prove approximation error bounds for the accumulated representation and establish that it serves as a compressed analogue of standard proper orthogonal decomposition for full-order snapshot data, thereby allowing reduced-order models to be constructed directly from streaming low-rank solution data through operations on TT cores without first reconstructing full snapshots.","pith_inferences":["The same core-wise update pattern could be applied to other low-rank tensor formats that admit similar projection and enrichment steps.","Memory savings would be especially pronounced in real-time or embedded simulations where solution data is generated already in TT form.","Choice of tolerance may need to be linked to quantities of physical interest rather than left as a fixed input."],"forward_implications":["Approximation error bounds hold for the incremental updates.","Reduced-order models can be assembled solely through operations on TT cores.","Numerical tests on parametric radiative transfer equations achieve reconstruction accuracy comparable to full methods while using less wall time.","The method produces efficient and accurate reduced-order models directly from compressed low-rank data."],"fun_headline_variants":["Core-wise projection compresses TT streams for ROMs","Incremental TT enrichment retains data for compressed ROMs","TT cores enable ROMs from streaming data without snapshots","Compressed POD analogue built incrementally from TT streams"],"cache_read_input_tokens":2112,"weakest_assumption_plain":"Incoming data arrives already in exact TT format and the projection plus enrichment steps capture all essential information beyond the chosen tolerance.","fun_headline_variants_meta":{"raw":{"variants":["Core-wise projection compresses TT streams for ROMs","Incremental TT enrichment retains data for compressed ROMs","TT cores enable ROMs from streaming data without snapshots","Compressed POD analogue built incrementally from TT streams"]},"model":"grok-4.3","cost_usd":0.00508,"raw_usage":{"total_tokens":2496,"prompt_tokens":713,"num_sources_used":0,"completion_tokens":58,"cost_in_usd_ticks":50799500,"prompt_tokens_details":{"text_tokens":713,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":1725,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":713,"tokens_out":58,"duration_ms":25429,"temperature":1.0,"reasoning_tokens":1725,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-06-26T04:07:17.057922+00:00","model_set":{"reader":"grok-4.3"},"falsifier":"A numerical test in which the reduced-order model built from the accumulated TT cores produces errors larger than the prescribed tolerance when compared with a standard POD model built from the same data after full reconstruction.","supporting_citations":[],"review_version":1}