{"id":"3fcfe98b-b7bc-426d-bf42-ea148cac7f72","arxiv_id":"2504.14721","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"The authors extend balanced truncation, balanced POD, and eigensystem realization to T-product tensor systems by replacing SVD with T-SVD, claiming comparable accuracy at lower cost.","lead":"This paper builds model order reduction methods for tensor-based dynamical systems that multiply states with the T-product, including T-balanced truncation, T-balanced proper orthogonal decomposition, and a T-eigensystem realization algorithm. The goal is to reduce memory and computation while keeping similar accuracy, with image and video processing in mind.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Proposition 4's error-bound proof is invalid: the inverse balancing identity is false, and tube-Frobenius truncation does not imply the per-block tail sums used in (13).","rationale":"The central claim is the H-infinity error bound for T-BT, and Proposition 4 is its only proof. The inverse identity is not a minor typo: it changes the state transformation used in the proof, and the balanced-Gramian calculation is not shown. The second gap is conceptually deeper: T-SVD tubal-rank truncation is optimal for Frobenius norm, but H-infinity balanced truncation needs per-block information, and the paper moves from tube-level truncation to per-block tail sums without an argument. These are internal correctness issues, not disagreements with consensus. The numerical tables showing essentially identical T-BT and BT errors across all k strengthen the concern that the comparison is not exercising the T-SVD truncation or that the generated examples have degenerate block structure; the authors should report retained index sets and per-block singular values. I do not see bad faith; the framework may be repairable by using Q^⊤ as T^{-1} and either bounding by Σ_{j∈K} σ_j^{(i)} or changing the truncation rule to per-block tails. Since the reader's verdict already conditioned acceptance on fixing the proof and numerical comparison, I leave the verdict unchanged.","tokens_in":19548,"tokens_out":24545,"duration_ms":236640,"concrete_test":"Run a single diagnostic script on the released code with a small random stable TPDS (e.g., n=6, s=3): (i) form T = Z_c ⊛ V ⊛ S^{-1/2} and test whether (Z_o ⊛ U ⊛ S^{-1/2}) ⊛ T equals I or whether (S^{-1/2} ⊛ U^⊤ ⊛ Z_o^⊤) ⊛ T equals I; (ii) record the common index set K chosen by tube-Frobenius truncation and compare Σ_{j∈K} σ_j^{(i)} with the tail sum in (13) for each Fourier block i. If the first identity fails and the second comparison exceeds the right-hand side, Proposition 4 as stated is incorrect and the numerical tables should be regenerated with a non-comonotonic example.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The proof of Proposition 4 does not establish the stated H-infinity bound. With H = Z_o^⊤ ⊛ Z_c = U ⊛ S ⊛ V^⊤ and T = Z_c ⊛ V ⊛ S^{-1/2}, the paper asserts T^{-1} = Z_o ⊛ U ⊛ S^{-1/2}. The correct inverse is Q^⊤ with Q = Z_o ⊛ U ⊛ S^{-1/2}, because Q^⊤ ⊛ T = I; the displayed expression satisfies no such identity for generic Z_c, Z_o, U, V. The proof then builds a 'balanced' realization using T^{-1} = Q rather than Q^⊤, so it is not the realization used in Algorithm 1, and the equality of balanced Gramians is unsupported. Even after correcting this, bound (13) requires that truncating k singular tuples by Frobenius norm removes the k trailing singular values in every Fourier block. Tubal-rank truncation deletes one common index set K in all blocks; if the block Hankel singular values are not comonotonic, K is not the per-block tail, and the correct discarded sum is Σ_{j∈K} σ_j^{(i)}, which can exceed the tail sum in (13). Thus the central error bound is not proved as written.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper proposes three tensor-native model order reduction methods for discrete-time input-output T-product-based dynamical systems (TPDSs): T-balanced truncation (T-BT), T-balanced proper orthogonal decomposition (T-BPOD), and the T-eigensystem realization algorithm (T-ERA). The methods construct a generalized Hankel tensor, apply T-SVD, and truncate singular tuples to obtain reduced TPDSs, thereby preserving the block-circulant structure that is lost when classical MOR is applied to the unfolded representation. The authors provide algorithmic descriptions, memory and complexity counts, an H-infinity error bound for T-BT (Proposition 4), and numerical comparisons with classical BT, BPOD, and ERA on synthetic examples and an image-dynamics case study. The main claims are that the T-product-based methods reduce memory and computation while achieving errors comparable to classical methods and that T-BT admits a tensor-analog of the standard balanced truncation error bound.","tokens_in":19806,"tokens_out":4118,"duration_ms":40810,"significance":"If the central claims hold, the paper offers a useful tensor-native extension of three standard data-driven MOR tools, with clear potential for image/video-type state data. The algorithms are concretely specified, complexity estimates are given, and the authors provide open-source code and reproducible numerical experiments, which is a strength. The main scientific stake is Proposition 4: the proposed H-infinity bound would be a genuine tensor analogue of the classical balanced truncation bound. The numerical validation also supports the practical claims, although the reported error coincidences need explanation. Because the two central load-bearing points, the inverse identity in the balancing construction and the passage from tensor truncation to per-block tail sums, are not properly established, the paper is not yet ready in its current form; the issues are local and repairable rather than fundamental.","major_comments":[{"comment":"The proof asserts that for T = Z_c ⊛ V_tilde ⊛ S_tilde^{-1/2}, the inverse is T^{-1} = Z_o ⊛ U_tilde ⊛ S_tilde^{-1/2}. This identity is not correct in general. With H = Z_o^⊤ ⊛ Z_c = U_tilde ⊛ S_tilde ⊛ V_tilde^⊤, the quantity Q = Z_o ⊛ U_tilde ⊛ S_tilde^{-1/2} satisfies Q^⊤ ⊛ T = I, so the inverse of T is Q^⊤, not Q. The displayed expression does not reduce to the T-identity for generic Z_c, Z_o, U_tilde, V_tilde. Since the balanced realization in the proof is built from T^{-1} = Q rather than Q^⊤, it is not the realization used in Algorithm 1, and the claimed equality of the balanced Gramians is unsupported. This is a load-bearing step for Proposition 4 and must be corrected or the proof must be rewritten with the correct inverse.","section":"Proposition 4, proof of the balancing inverse identity"},{"comment":"The derivation of bound (13) applies the classical balanced truncation error bound to each Fourier block and then uses a tail sum over j = n-k+1, ..., n in every block. However, T-BT truncates k singular tuples by their Frobenius norm, which selects one common index set K across all Fourier blocks. If the block Hankel singular values are not comonotonic, K is not the per-block tail set, and the discarded sum in block i is Σ_{j∈K} σ_j^{(i)}, not the tail sum appearing in (13). The displayed inequality therefore does not follow from the classical bound. The authors need either to prove a common-index version of the bound or to replace (13) by the appropriate sum over the chosen index set and justify why that set is the per-block tail.","section":"Proposition 4, bound (13) and the truncation set"},{"comment":"The numerical tables report relative errors that are identical, or identical to two or three significant digits, between the tensor methods and their unfolded counterparts at nearly every truncation level (e.g., 1.2×10^{-14} vs 1.2×10^{-14} in Table I at k=55, 2.19×10^{-3} vs 2.19×10^{-3} at k=90, and similarly repeated in Tables II and III). Since T-BT, T-BPOD, and T-ERA reduce in a different space than standard BT, BPOD, and ERA, exact agreement at this precision is not expected and is unexplained. This undermines the validation of the central claim of comparable accuracy. Please report the actual reduced models or their differences, state whether the experiments were repeated across random seeds, and explain the mechanism that produces identical error values.","section":"Section IV, Tables I–III"}],"minor_comments":[{"comment":"The T-SVD definition as printed writes A = U ⊛ S ⊛ U^⊤ although V is introduced in the same sentence; it should be A = U ⊛ S ⊛ V^⊤.","section":"Definition 7, Eq. (2)"},{"comment":"The dimensions of the middle tensor S are not stated explicitly; since A ∈ R^{n×m×s}, S should be F-rectangle diagonal in R^{n×m×s}, and V should be in R^{m×m×s}. Please clarify.","section":"Section II.A, Definition 7"},{"comment":"There are several typographical errors that should be corrected in revision, including 'nonlineaar' and 'other other' in Section II.A and 'Gramain' in the paragraph preceding Proposition 2.","section":"Throughout"},{"comment":"The caption says results for k=0 and k=10 are similar to those for k=20 and are omitted; since Figure 1(b) is already sparse, it would be clearer to show all levels or at least state the visible levels explicitly in the figure itself.","section":"Section IV.D, Figure 1(b)"},{"comment":"The paper notes that T-BPOD and T-ERA may not be equivalent and that this 'warrants further theoretical exploration.' This is an important structural difference from the matrix case; a brief explanation or a reference would strengthen the paper, even if a full proof is deferred.","section":"Section IV.C, final paragraph"},{"comment":"The proof overloads the symbol Wc for both the tensor Gramian and the matrix Gramian of the unfolded system and writes Wc = ξ(Wc). Please distinguish the two objects notationally to avoid circularity in the definition.","section":"Proposition 2, proof"}],"recommendation":"major_revision","confidential_remarks":"The paper is within scope for a systems/control journal, and the self-citations are relevant to T-product dynamical systems rather than to the MOR results themselves. The main issue is the Proposition 4 proof: the inverse identity and the truncation-tail step are both load-bearing, and the numerical validation has unexplained error coincidences. These are repairable within the scope of a revision, but the current version should not be accepted as is."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's my take on 2504.14721. The paper does something genuinely new: the first MOR framework for input-output T-product dynamical systems, with three algorithms (T-BT, T-BPOD, T-ERA) built on T-SVD. The construction is a natural extension—replace SVD with T-SVD and work per Fourier block—and the complexity and memory claims are carefully derived. The image-data case study is a nice demonstration.\n\nThe soft spots are real. Proposition 4's proof is invalid as written. They assert T^{-1} = Zo ⊛ U ⊛ S^{-1/2}, but the correct inverse is Q^T with Q = Zo ⊛ U ⊛ S^{-1/2}, since Q^T ⊛ T = I. That is a concrete algebraic error, not a gap in exposition. Even after fixing that, the bound (13) assumes truncating k singular tuples by Frobenius norm removes the trailing k singular values in every Fourier block. That only holds if the block Hankel singular values are comonotonic; otherwise the discarded sum is Σ_{j∈K} σ_j^{(i)} over the common index set K, which can exceed the tail sum. So the advertised H-infinity bound is unsupported.\n\nThe numerics also bother me. In Tables I–III the relative errors for T-BT and BT are identical to two significant digits across nearly every k. That is not what I'd expect from two genuinely different reduction procedures, and the paper gives no explanation. The comparison is also tilted: T-BT truncates k tuples per frequency, effectively k·s singular values, while BT truncates k singular values total. The memory advantage is real, but the error comparison is not apples-to-apples.\n\nThat said, the core algorithms are plausible and likely work; the T-SVD machinery and the per-frequency interpretation are sound. The problem is the paper claims a theoretical guarantee it has not actually established, and the validation does not convincingly support the narrative. This is fixable: correct the inverse identity, state precisely which truncation set gives the bound (or weaken it), and rerun experiments with a fair baseline and an explanation of the matching errors.\n\nWho is this for? People working on tensor dynamical systems, MOR, and multidimensional signal processing will find the algorithms useful. It deserves peer review because the question is relevant and the approach is legitimate. My recommendation: send it to referees, but expect major revision—the error-bound proof and experimental comparison both need real work.","headline":"Tensor-native MOR for T-product systems: a real extension with a broken error-bound proof and too-clean numerics.","tokens_in":20293,"tokens_out":3071,"would_cite":false,"duration_ms":28648,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["93B11","15A69","93C05"],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper claims that balanced truncation, balanced POD, and the eigensystem realization algorithm can be lifted into T-product tensor algebra, so image and video dynamical systems are reduced without flattening and with less memory and…","keywords":["model order reduction","T-product","third-order tensors","balanced truncation","balanced proper orthogonal decomposition","eigensystem realization algorithm","T-singular value decomposition","tensor dynamical systems"],"falsifier":"Compute the truncated T-SVD of the Hankel tensor $H\\approx U\\circledast S\\circledast V^{\\top}$ for a small random stable TPDS, form $T=Z_c\\circledast V\\circledast S^{-1/2}$, and test numerically whether $(Z_o\\circledast U\\circledast S^{-1/2})\\circledast T$ and $T\\circledast (Z_o\\circledast U\\circledast S^{-1/2})$ both equal the T-identity tensor. A single generic counterexample would invalidate the balancing construction and Proposition 4.","tokens_in":19385,"feed_emoji":"🎞️","tokens_out":10443,"duration_ms":86117,"temperature":0.7,"pith_summary":"Model order reduction normally treats a dynamical system as vectors and matrices, so tensor-valued data such as image sequences must be flattened before reduction. This paper argues that the T-product, a way to multiply third-order tensors through block-circulant convolution, lets balanced truncation, balanced POD, and the eigensystem realization algorithm operate directly on the tensor structure. The central claim is that truncating the singular tuples of a generalized Hankel tensor yields reduced T-product systems whose input-output behavior stays close to the original, while using less memory and computation than reducing the unfolded system. For T-balanced truncation the paper derives a worst-case error bound of at most twice the sum of the discarded singular-tuple norms. If correct, the result gives image and video dynamics a reduction method that preserves their multilinear structure.","feed_headline":"T-product reduction cuts memory for image dynamical systems","feed_subtitle":"Tensor-native algorithms match classical reduction errors while keeping data structure and using fewer parameters.","key_machinery":"The machinery is the T-product and its singular value decomposition. Under the T-product, a third-order tensor acts on a matrix by block-circulant convolution, and the T-SVD factors a tensor into two T-orthogonal factor tensors and a block tensor whose frontal slices are diagonal, with the diagonal tubes called singular tuples. A discrete Fourier transform diagonalizes the T-product, so each frontal slice becomes an independent matrix problem; truncating the singular tuples by their Frobenius norm gives the low-tensor-rank approximation that drives all three reduction algorithms. The balancing transform $P=Z_c\\circledast V\\circledast S^{-1/2}$, $Q=Z_o\\circledast U\\circledast S^{-1/2}$ then maps the system to a reduced TPDS.","core_discovery":"The paper's core discovery is that the entire balanced-truncation pipeline can be transplanted into T-product algebra. For an input-output T-product dynamical system $X(t+1)=A\\circledast X(t)+B\\circledast U(t)$, $Y(t)=C\\circledast X(t)$, the controllability and observability Gramians solve T-Lyapunov equations, and the discrete Fourier transform decouples each into $s$ independent matrix Lyapunov equations. The generalized Hankel tensor $H=Z_o^{\\top}\\circledast Z_c$ is reduced by truncating its singular tuples, ranked by Frobenius norm, rather than by truncating singular values of a flattened matrix. The paper claims that the reduced system is again a TPDS, that it preserves controllability and observability, and that T-balanced truncation obeys the error bound $\\|G-G_{\\mathrm{red}}\\|_\\infty \\le 2\\max_i \\sum_{j=n-k+1}^{n}\\sigma_j^{(i)}$ over the discarded singular tuples in the Fourier domain. T-BPOD and T-ERA obtain the same kind of reduction from snapshot or Markov-parameter data, and the numerical examples report relative errors nearly identical to their matrix counterparts with substantially fewer parameters.","pith_inferences":["Because one singular tuple contains $s$ scalars, truncating $k$ tuples in the tensor method discards about $s$ times as much spectral content as truncating $k$ singular values of the unfolded system; a matched-parameter comparison would test whether the reported savings persist when the reduced models have equal total parameter counts.","The observed difference between the T-ERA and T-BPOD reduced models, where classical ERA and BPOD coincide, suggests the two tensor variants realize genuinely different systems and would merit a separate equivalence analysis.","The same Hankel-tensor truncation idea could in principle be combined with other tensor factorizations, such as tensor-train or hierarchical formats, to push memory savings further, but the paper does not test that combination."],"forward_implications":["Reduced systems stay in TPDS form, so the multilinear structure of image and video state data is preserved instead of being flattened into a long vector.","The Fourier-domain decoupling lowers the cost from roughly $O(n^3s^3)$ for balanced truncation on the unfolded system to $O(n^3s+n^2s\\log s)$ for T-BT.","T-balanced truncation carries an explicit worst-case error bound, so the truncation level can be chosen with a guarantee on the reduced model's accuracy.","T-BPOD and T-ERA work from data; T-ERA in particular needs no model equations or adjoint simulations, so it can identify reduced tensor systems from experimental impulse responses.","In the reported examples the tensor methods match the relative errors of BT, BPOD, and ERA while using roughly a third of the reduced-model parameters."],"supporting_citations":[{"why":"Supplies the classical balanced truncation construction that T-BT generalizes.","marker":"[7]"},{"why":"Supplies the ERA procedure that T-ERA lifts to tensors.","marker":"[13]"},{"why":"Introduces T-product-based dynamical systems as the model class under study.","marker":"[31]"},{"why":"Provides the T-product and T-SVD theory that all three algorithms build on.","marker":"[44]"},{"why":"Supplies the snapshot balancing method behind T-BPOD.","marker":"[50]"},{"why":"Establishes the ERA-BPOD equivalence that motivates comparing T-ERA with T-BPOD.","marker":"[52]"},{"why":"Supplies the discrete-time balanced truncation error bound applied blockwise in Proposition 4.","marker":"[54]"},{"why":"Establishes data-driven analysis of TPDSs, including the empirical Gramian notions T-BPOD uses.","marker":"[35]"}],"fun_headline_variants":["T-product MOR slashes memory for tensor dynamical systems","Tensor-balanced truncation cuts memory, matches errors","T-product balanced truncation for third-order tensor dynamics","Data-driven tensor ROM preserves structure, saves memory","T-product MOR: less storage, same accuracy for image dynamics"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the balancing tensor can be inverted by a specific formula that the paper asserts without proof; if that formula is wrong, the balanced realization and the error guarantee do not follow.","fun_headline_variants_meta":{"raw":{"variants":["T-product MOR slashes memory for tensor dynamical systems","Tensor-balanced truncation cuts memory, matches errors","T-product balanced truncation for third-order tensor dynamics","Data-driven tensor ROM preserves structure, saves memory","T-product MOR: less storage, same accuracy for image dynamics"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000516,"raw_usage":{"total_tokens":2524,"prompt_tokens":985,"completion_tokens":1539,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":601,"completion_tokens_details":{"reasoning_tokens":1463}},"tokens_in":601,"tokens_out":1539,"duration_ms":9742,"temperature":1.0,"reasoning_tokens":1463,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T11:42:58.308994+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute the truncated T-SVD of the Hankel tensor $H\\approx U\\circledast S\\circledast V^{\\top}$ for a small random stable TPDS, form $T=Z_c\\circledast V\\circledast S^{-1/2}$, and test numerically whether $(Z_o\\circledast U\\circledast S^{-1/2})\\circledast T$ and $T\\circledast (Z_o\\circledast U\\circledast S^{-1/2})$ both equal the T-identity tensor. A single generic counterexample would invalidate the balancing construction and Proposition 4.","supporting_citations":[{"cited_title":"An eigensystem realization algorithm for modal parameter identification and model reduction,","cited_arxiv_id":null,"evidence_quote":"Supplies the ERA procedure that T-ERA lifts to tensors."},{"cited_title":"Model reduction for flow analysis and control,","cited_arxiv_id":null,"evidence_quote":"Supplies the snapshot balancing method behind T-BPOD."},{"cited_title":"Reduced-order models for control of fluids using the eigensystem realization algorithm,","cited_arxiv_id":null,"evidence_quote":"Establishes the ERA-BPOD equivalence that motivates comparing T-ERA with T-BPOD."},{"cited_title":"An error bound for a discrete reduced order model of a linear multivariable system,","cited_arxiv_id":null,"evidence_quote":"Supplies the discrete-time balanced truncation error bound applied blockwise in Proposition 4."}],"review_version":1}