{"id":"74681805-4406-4526-b3fa-28b82521f6e7","arxiv_id":"2412.14460","paper_version":2,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"A new reduced-order modeling framework uses tensor-train decomposition of finite element snapshots to cut offline costs and reduce subspace dimensions for parameterized PDEs on Cartesian grids.","lead":"The paper introduces a tensor-train reduced basis method that builds reduced models of parameterized PDEs on Cartesian grids cheaply by compressing snapshots in a tensor format. It reports faster offline construction and smaller reduced subspaces than standard space-time reduced basis methods, at the cost of somewhat larger online errors.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The central accuracy guarantee rests on an unproved rank-1 crossnorm heuristic for rank-K stiffness norms; if that heuristic fails, TT-RB error no longer tracks the prescribed tolerance.","rationale":"The reader's weakest-assumption analysis correctly identifies the unproved rank-1 crossnorm heuristic in Section 3.3 as the most load-bearing gap. The paper's central accuracy statements rely on the rank-K stiffness norm for all numerical benchmarks, yet Theorem 2's bound (21) is explicitly not guaranteed for that case. My independent reading confirms that the subsequent X_s,s-orthogonalization does not repair a rank deficiency introduced by using the wrong norm during truncation, so the issue is not merely cosmetic. The numerical results do show error decay with tolerance, which is genuine supporting evidence but is limited to the tested configurations; no theorem or parameter-uniform analysis covers anisotropic or otherwise unfavorable cases. The other possible concern, the empirically asserted unfolding equivalence in Section 3.2, is less central because it concerns only the rank-1 case and is likely provable by row-space invariance. I therefore agree with the reader's conditional verdict and propose a targeted computational test that would either substantiate or refute the heuristic in a regime where it could plausibly fail.","tokens_in":33537,"tokens_out":7386,"duration_ms":69203,"concrete_test":"Run Algorithm 4 for a 2D Poisson problem on a highly anisotropic rectangle (e.g. aspect ratio 100:1) with X_s,s = A_1⊗M_2 + M_1⊗A_2 and |X_s,s chosen as in Remark 4. For ε = 1e-2, 1e-3, 1e-4, compare the resulting X_s,s-orthogonal basis's energy-norm projection error against the bound (21), and also against a reference basis obtained by explicit Cholesky factorization of X_s,s followed by standard TT-SVD and inverse rescaling. If the observed error exceeds roughly 10×ε√(d+1)‖U‖_X, or fails to decrease linearly with ε, the central accuracy claim is unsupported; if it tracks the reference, the heuristic is acceptable for this regime.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing assumption is in Section 3.3. For the actual stiffness norm X_s,s of rank K (Eq. 11), Algorithm 4 first builds TT cores using a rank-1 norm |X_s,s that is 'similar' to X_s,s (Eq. 32 and Remark 4), then X_s,s-orthogonalizes the final core. The paper states explicitly that 'the error bound (21) is no longer guaranteed in this case' and offers only numerical evidence that the error decays with tolerance. All three benchmarks (Poisson, heat, elasticity) measure error in the rank-K norm X_s,s of Eq. (11), so the advertised accuracy guarantee in Theorem 4 inherits this gap. The heuristic choice q = arg max over the K terms does not imply that |X_s,s is equivalent to X_s,s with constants independent of h or of domain aspect ratio; when the norm is dominated by a different term, the selected rank can be too small, and the subsequent orthogonalization cannot recover the lost component. Separately, the claimed equivalence of the two diadic decompositions in Section 3.2 is only 'empirically verified,' though the rank-K case is the one used in practice. Both gaps weaken the central claim that TT-RB achieves controlled accuracy for a prescribed tolerance.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces TT-RB, a reduced-basis solver for parameterized PDEs on Cartesian grids, in which the snapshot tensor is compressed with a tensor-train SVD (TT-SVD) instead of the space-time TPOD used by ST-RB. The reduced subspace is built in a split-axes format, the residual and Jacobian are hyper-reduced with a tensor-train variant of MDEIM (TT-MDEIM), and an a posteriori error estimate is adapted from the authors' earlier ST-RB analysis. Numerical experiments on Poisson, heat, and linear elasticity problems in 2D and 3D report offline and online speedups relative to ST-RB and compare accuracy at matching tolerances. The central claims are that TT-RB lowers the offline cost, reduces the projection-subspace dimension, and achieves accuracy comparable to ST-RB.","tokens_in":33854,"tokens_out":5058,"duration_ms":42178,"significance":"If the accuracy claims are fully substantiated, TT-RB would be a useful addition to the reduced-basis literature: the split-axes construction has the potential to lower offline memory and time substantially, the hyper-reduction is naturally adapted to tensor formats, and the reported reduction factors are large. The paper has clear strengths: it gives explicit complexity estimates for the offline phase, the numerical comparisons are made against independent full-order solutions, and the main heuristic is disclosed rather than hidden. The code is available through the authors' GridapROMs.jl library, which supports reproducibility. However, the significance is currently limited by a load-bearing gap: the algorithm used in all benchmarks (rank-K norm) is not covered by the proved a priori estimate, and several accuracy claims go beyond what the numerical evidence establishes.","major_comments":[{"comment":"The rank-K basis construction used in all numerical benchmarks is based on a rank-1 crossnorm |X_s,s whose equivalence to the actual rank-K stiffness norm X_s,s is only asserted, not proved. The text explicitly states that 'the error bound (21) is no longer guaranteed in this case', so Theorem 2 does not cover the algorithm as implemented. Since every benchmark measures error in the rank-K norm (11), the central accuracy guarantee advertised in the abstract and in Theorem 4 is not established for the method actually tested. The numerical evidence is consistent with this gap: TT-RB errors are typically several times larger than ST-RB errors at matched tolerance, in some cases more than an order of magnitude larger (e.g., Table 6, M=40, ε=10^-4: E/ε = 50.30 for TT-RB vs 2.38 for ST-RB). Please either prove a norm-equivalence bound for |X_s,s versus X_s,s with constants independent of h and the domain aspect ratio, provide a rigorous error bound for Algorithm 4, or restrict the a priori and a posteriori claims to the rank-1 case and present the rank-K results explicitly as heuristic.","section":"Section 3.3, Eq. (32), Algorithm 4"},{"comment":"The paper claims in Section 1 that 'TT-RB decompositions offer significantly higher accuracy than ST-RB for a given reduced subspace dimension' and in Section 5 that accuracy is 'comparable'. No experiment compares the two methods at equal reduced-subspace dimension; the reported comparisons are at equal tolerance ε, where TT-RB uses a much smaller subspace (rt vs rst) but has larger normalized errors. For example, Table 6 shows E/ε = 50.30 for TT-RB versus 2.38 for ST-RB at M=40, ε=10^-4, and Table 8 shows E/ε = 89.90 versus 5.57 at M=40, ε=10^-2. The abstract's 'lower dimensionality of the projection subspaces for a given accuracy' is therefore only qualitatively supported. Please provide equal-dimension accuracy comparisons, or revise the accuracy claims to match the presented evidence.","section":"Section 1 and Section 5; Tables 3-8"},{"comment":"The equivalence of the two diadic decompositions (23) and (24), on which Algorithm 3 and its accuracy guarantee rest, is only 'empirically verified' up to machine precision. This is not a proof that the sequentially rescaled TT-SVD produces the same projection subspace as the full Cholesky rescaling, and the rank-K extension inherits this gap. A rigorous statement, or an explicit assumption in Theorem 2 that this equivalence holds, is needed before Algorithm 3 can be presented as a variant of TT-SVD with the accuracy of Theorem 2.","section":"Section 3.2, Eqs. (23)-(24)"},{"comment":"The a posteriori bound (42) contains an explicit residual term that is not bounded a priori, and no effectivity indices are reported in the numerical section. Since the paper's stated contribution includes 'a posteriori error estimates to validate the accuracy of the method', the numerical validation should report the ratio of the estimated bound to the true error, or at least the size of the residual term, for the three benchmarks. Without this, the claim that the estimates validate the method is not quantitatively supported.","section":"Theorem 4, Section 4"}],"minor_comments":[{"comment":"The word 'diadic' should be 'dyadic' in Sections 3.1-3.2 and in the captions of Eqs. (23)-(24).","section":"Throughout"},{"comment":"Several references have truncated DOIs or broken line breaks, including [7], [22], and [37]; please correct these in the bibliography.","section":"References"},{"comment":"The quantity E/ε is used throughout the tables but is not defined in a table caption; please state explicitly that E is the averaged relative Xst,st error defined in (43) and that the table reports E divided by the tolerance ε.","section":"Tables 3-8"},{"comment":"The identity (35) is justified only by reference to a figure; either provide a short proof or label it as an empirically observed identity.","section":"Section 3.4, Remark 5"},{"comment":"The notation `r:, js` and `pUt, psµ` in Eq. (14) is not defined in the text; a brief explanation would improve readability.","section":"Section 2.2, Eq. (14)"}],"recommendation":"major_revision","confidential_remarks":"For the editor: the manuscript is transparent about the rank-K heuristic, but the abstract and conclusions overstate the theoretical support. If the authors can supply a norm-equivalence argument for the rank-K construction or explicitly reframe the rank-K algorithm as heuristic, major revision is appropriate; otherwise the central accuracy claim should be narrowed. The error spikes at ε=10^-3 and ε=10^-4 in Tables 3-8 also deserve a concrete explanation rather than a qualitative remark."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Worth a look. The paper does something new: it builds the reduced basis from a tensor-train decomposition of the snapshot tensor, with spatial axes split per Cartesian direction, and it customizes TT-SVD for energy norms. The split-axes trick is the real idea; it avoids forming and factoring the global stiffness matrix, and makes the offline cost scale like one spatial axis rather than the whole grid. The reported offline speedups, especially in 3D, are large and consistent. TT-MDEIM is also a reasonable tensor extension of MDEIM, and the numerical results show it tracks the tolerance in most cases. The authors are honest about weak spots: they flag that the rank-K norm construction in Section 3.3 has no guaranteed error bound, and they report the larger constants of TT-RB relative to ST-RB without hiding them. The self-citation to [7] is legitimate because the a posteriori theorem is genuinely an adaptation.\n\nThe central accuracy guarantee for the actual stiffness norm is exactly the Section 3.3 heuristic. The algorithm replaces the rank-K norm X_s,s by a rank-1 crossnorm, builds the TT cores in that norm, then orthogonalizes the last core in X_s,s. The paper states that the error bound (21) is no longer guaranteed in this case and relies on numerical decay. That is load-bearing: the measured errors are in the rank-K norm, so the advertised tolerance-controlled accuracy is not actually proved. The empirically verified equivalence of the two diadic decompositions in Section 3.2 is another unproved step, though less worrying because the rank-1 case is not the one used in practice. The numerical evidence also shows TT-RB errors routinely 10 to 50 times larger than ST-RB at matched tolerance, with occasional rank-stagnation spikes in the hyper-reduction. That does not kill the method—the offline savings and the smaller subspace dimension are real—but it undercuts the \"comparable accuracy\" phrasing in the conclusions.\n\nThis paper is for people building ROMs on structured grids, especially time-dependent 3D problems, and for anyone working on tensor-structured reduced bases. It deserves a serious referee. The referee should push for either a proof of norm equivalence for the rank-K case or a clear statement that the tolerance is heuristic and should be set conservatively, plus a sensitivity study for the q = arg max choice in (32). With that revision, this is a solid contribution.","headline":"A genuinely useful tensor-train reduced basis method with solid offline gains, but the accuracy certificate for the rank-K energy norm is a heuristic, so the paper needs a careful revision before the claims are taken at face value.","tokens_in":34316,"tokens_out":3273,"would_cite":true,"duration_ms":29517,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["65N30","65M60","15A69"],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper introduces a tensor-train reduced basis method that builds projection subspaces for parameterized PDEs by decomposing snapshots in the split-axes format, claiming cheaper offline construction, a smaller reduced subspace, and…","keywords":["tensor-train decomposition","reduced basis methods","model order reduction","space-time reduced basis","hyper-reduction","empirical interpolation","parameterized PDEs","Cartesian grids"],"falsifier":"Run the rank-$K$ algorithm on 2D or 3D Poisson problems where the terms of the stiffness norm $X_{s,s}$ differ strongly in magnitude, and decrease $\\epsilon$ at fixed mesh size $M$; if the measured energy-norm error stops decaying with $\\epsilon$, or violates the bound (21) beyond the algorithm's constants, then the rank-one crossnorm choice (32) does not preserve enough of the true norm.","tokens_in":33340,"feed_emoji":"🧮","tokens_out":9794,"duration_ms":71743,"temperature":0.7,"pith_summary":"This paper introduces TT-RB, a reduced-order modeling method that constructs low-dimensional subspaces for parameterized PDEs by applying a tensor-train (TT) decomposition to solution snapshots whose spatial degrees of freedom are split into Cartesian directions. The central claim is that this construction is substantially cheaper offline than the standard space-time reduced basis (ST-RB) approach, because the expensive operations act on one spatial direction at a time rather than on the full spatial mesh. It further claims that the resulting reduced subspace has dimension equal to a single tensor rank instead of a product of spatial and temporal ranks, so that for a given accuracy the online system is smaller, and for a given subspace dimension the projection is more accurate. The paper supports these claims with a priori and a posteriori error estimates and benchmarks on the Poisson, heat, and transient linear elasticity equations in two and three dimensions.","feed_headline":"Tensor-train compression speeds up reduced-order PDE models","feed_subtitle":"Offline basis building scales with one grid direction at a time, and subspace size drops to a single tensor rank.","key_machinery":"The central object is the tensor-train (TT) decomposition, a low-rank format that writes a tensor as a product of three-dimensional cores, combined with the 'split-axes' representation of finite-element unknowns on a Cartesian grid. The TT format carries the argument because every operation in the offline phase—Cholesky rescaling, SVD compression, empirical interpolation, and Galerkin projection—is applied to one core at a time, so cost scales with a single axis length rather than the full spatial dimension $N_s$, and the projection subspace dimension is the temporal rank $r_t$ rather than the product $r_s r_t$. For the rank-$K$ stiffness norm, the load-bearing device is a rank-one crossnorm $|X_{s,s}|$ chosen 'similar' to $X_{s,s}$, used to compute the initial TT cores before an orthogonalization sweep on the cores.","core_discovery":"The central discovery is that the TT format can carry the whole reduced-basis pipeline—subspace construction, hyper-reduction, and Galerkin projection—with all operations performed on the tensor cores. Given snapshots in the split-axes format, a modified TT-SVD imposes orthogonality in the finite-element energy norm: for rank-one norm matrices this is done by sequential Cholesky rescaling of one-dimensional factors, and for the rank-$K$ stiffness norm by recursively building a reduced norm matrix on the cores. Hyper-reduction is handled by TT-MDEIM, an empirical interpolation of residual and Jacobian tensors directly in TT form. The paper's numerical claim is that TT-RB matches ST-RB accuracy with a much smaller projection subspace and a cheaper offline phase, with gains that grow with the spatial dimension $d$, and that for a fixed subspace dimension the TT-RB projection is more accurate. Theorem 4 gives the a posteriori error bound tying total error to the user tolerance through both the subspace and the interpolation steps.","pith_inferences":["One extension the paper leaves implicit: if the rank-one crossnorm heuristic holds beyond these benchmarks, the same pipeline should transfer to non-Cartesian geometries through unfitted finite element methods, since the split-axes structure is the only geometry-dependent ingredient.","A direct test of the weakest assumption would be to run the rank-$K$ algorithm with several different choices of the rank-one norm $|X_{s,s}|$; if the energy-norm error is insensitive to that choice, the heuristic is robust, and if not, the divergence marks its limits.","The method's advantage is largest when the temporal rank is small; for problems with high temporal complexity, permuting the TT axes so that time is compressed later or earlier could trade rank for accuracy, an option the paper does not explore.","Because the residual and Jacobian assembly in TT-MDEIM is algebraic and parameter-agnostic, the hyper-reduction should carry over to nonlinear and saddle-point problems, potentially extending TT-RB beyond the linear cases tested here."],"forward_implications":["For spatial dimension $d>2$, the asymptotic offline cost of the TT basis construction is lower than TPOD: roughly $O((d+1)M^{d+2}\\log M)$ versus $O(M^{3d-2}+M^{d+2}\\log M)$ under the paper's equal-size assumptions.","The dimension of the TT projection subspace is $r_t$ instead of $r_s r_t$, which in the heat-equation benchmark yields reduction factors tens to hundreds of times larger than ST-RB.","TT-MDEIM gives the same form of error bound as matrix MDEIM, with an extra $\\sqrt{d+1}$ factor entering through the TT-SVD compression.","Theorem 4's a posteriori estimate connects the total error to the user tolerance $\\epsilon$ through both the TT subspace error and the TT-MDEIM interpolation errors.","On the Poisson, heat, and transient elasticity benchmarks, TT-RB reaches comparable or better accuracy than ST-RB at the same tolerance, with larger online speedups in wall time and memory."],"supporting_citations":[{"why":"Supplies the TT-SVD algorithm and its error analysis, the foundation the paper modifies for energy-norm orthogonality.","marker":"[14]"},{"why":"Supplies the space-time MDEIM framework and the a posteriori error estimate that Theorem 4 adapts to the TT setting.","marker":"[7]"},{"why":"Provides the TT-cross-DEIM collocation procedure that TT-MDEIM simplifies, including the full-rank lemmas used in Theorem 3.","marker":"[25]"},{"why":"Supplies the DEIM interpolation error lemma used to prove the TT-MDEIM accuracy bound.","marker":"[36]"},{"why":"Supplies the matrix DEIM method and its error analysis, the baseline that TT-MDEIM extends to tensor format.","marker":"[37]"},{"why":"Supplies the reduced basis methodology and the reducibility criterion that justifies low-rank snapshot compression.","marker":"[1]"},{"why":"Supplies the randomized SVD used as the compression primitive in both TPOD and TT-SVD.","marker":"[29]"}],"fun_headline_variants":["Tensor trains shrink reduced-basis cost for PDEs","TT-RB: Faster offline stage for parametric PDEs","Reduced-basis PDE solver gets tensor-train boost","Compressed subspaces cut PDE solver offline time","Tensor-train MDEIM trims reduced-order PDE build"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method's central accuracy guarantee for energy norms rests on the heuristic that a rank-one crossnorm 'similar' to the true rank-$K$ stiffness norm preserves enough of the norm for the TT error bound to hold; the paper states this bound is no longer guaranteed and verifies the behavior numerically instead of proving it.","fun_headline_variants_meta":{"raw":{"variants":["Tensor trains shrink reduced-basis cost for PDEs","TT-RB: Faster offline stage for parametric PDEs","Reduced-basis PDE solver gets tensor-train boost","Compressed subspaces cut PDE solver offline time","Tensor-train MDEIM trims reduced-order PDE build"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000498,"raw_usage":{"total_tokens":2436,"prompt_tokens":936,"completion_tokens":1500,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":552,"completion_tokens_details":{"reasoning_tokens":1421}},"tokens_in":552,"tokens_out":1500,"duration_ms":8842,"temperature":1.0,"reasoning_tokens":1421,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T12:12:43.306514+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the rank-$K$ algorithm on 2D or 3D Poisson problems where the terms of the stiffness norm $X_{s,s}$ differ strongly in magnitude, and decrease $\\epsilon$ at fixed mesh size $M$; if the measured energy-norm error stops decaying with $\\epsilon$, or violates the bound (21) beyond the algorithm's constants, then the rank-one crossnorm choice (32) does not preserve enough of the true norm.","supporting_citations":[{"cited_title":"Tensor-train decomposition","cited_arxiv_id":null,"evidence_quote":"Supplies the TT-SVD algorithm and its error analysis, the foundation the paper modifies for energy-norm orthogonality."},{"cited_title":"Model order reduction with nove l discrete empirical interpolation methods in space–time","cited_arxiv_id":null,"evidence_quote":"Supplies the space-time MDEIM framework and the a posteriori error estimate that Theorem 4 adapts to the TT setting."},{"cited_title":"Collocation methods for nonlinear differential equations on low-rank manifolds","cited_arxiv_id":"2402.18721","evidence_quote":"Provides the TT-cross-DEIM collocation procedure that TT-MDEIM simplifies, including the full-rank lemmas used in Theorem 3."},{"cited_title":"Nonlinear model r eduction via discrete empirical interpolation","cited_arxiv_id":null,"evidence_quote":"Supplies the DEIM interpolation error lemma used to prove the TT-MDEIM accuracy bound."},{"cited_title":"Efficient model r eduction of parametrized systems by matrix discrete em- pirical interpolation","cited_arxiv_id":null,"evidence_quote":"Supplies the matrix DEIM method and its error analysis, the baseline that TT-MDEIM extends to tensor format."},{"cited_title":"Quarteroni, A","cited_arxiv_id":null,"evidence_quote":"Supplies the reduced basis methodology and the reducibility criterion that justifies low-rank snapshot compression."}],"review_version":1}