{"id":"ba0c0e2a-c4a5-44bd-8d27-31cfb38d4a1d","arxiv_id":"2603.21141","paper_version":2,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":7.0,"correctness_risk":"low","formal_verification":"none","parameter_count":5,"one_line_summary":"Tucker tensor train Taylor series (T4S) makes high-order local Taylor surrogates tractable for high-dimensional implicit maps using symmetric derivative probes and Riemannian rank-adaptive fitting.","lead":"High-order Taylor surrogates for PDE-based maps become practical by representing each derivative tensor as a Tucker tensor train trained only from cheap symmetric probes at one point. This gives local, derivative-accurate surrogates for outer-loop UQ and optimization without many nonlinear solves.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.5","headline":"No significant objection identified beyond the paper's own stated limitations on spectral decay and locality.","rationale":"The paper is a methods contribution whose strongest claim is carefully scoped. The theory supplies explicit rank/error bounds under spectral decay; the random-tensor experiments show that the Riemannian fitting procedures recover quasi-optimal accuracy from directionally symmetric probes alone up to data-limited ranks; the PDE examples confirm that the full pipeline recovers the expected high-order Taylor structure once dimension reduction and ranks are adequate. The free algorithmic parameters (n_chunk, τ, batch size, rank-continuation thresholds) and absence of released code justify the reader's CONDITIONAL rather than unconditional ACCEPT, but they do not undermine the mathematical or experimental support for the claim inside its stated regime. The single most load-bearing condition is precisely the one the reader already named; no stronger, unacknowledged vulnerability was identified. Hence the verdict remains CONDITIONAL with high confidence.","tokens_in":56270,"tokens_out":537,"duration_ms":5660,"concrete_test":"Independently re-derive the peeling step of Lemma 3 (the bound ∥F−F′∥≤2ϵ_{k−1}∥B∥ after truncating the Kronecker-product eigenvalues of C_1⊗⋯⊗C_{k−1}) without using the eigenvector-basis assumption; if the induced-norm bound fails for a non-diagonal C, the finite-dimensional foundation of Theorem 8 is weaker than claimed.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The reader's weakest assumption correctly identifies the two conditions the paper itself flags in §1.1 and uses as hypotheses for Theorem 8 / Corollary 9: (i) sufficiently rapid spectral decay of C (or additional low-rank structure in D^j q) so that moderate Tucker/TT ranks suffice, and (ii) local validity of the Taylor expansion about a single point. These are not hidden soft spots; they are explicit scope conditions. The central claim—that high-order Taylor surrogates become tractable via T4S under those conditions—is supported by the peeling argument (Proposition 2 + Lemmas 3–5), the infinite-dimensional reduction (Lemma 7 + Theorem 8), the probe-cost analysis (Table 1, §3.5), the matching of T3-SVD accuracy from probes alone (Figs. 11–14), and the recovery of high-order Taylor structure on the Poisson examples (Figs. 16–22). No internal inconsistency or unstated assumption that would invalidate the claim under the stated hypotheses was found.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.5","summary":"The paper constructs local high-order Taylor surrogates for covariance-preconditioned maps that depend implicitly on the solution of a nonlinear state equation (e.g., a PDE). Each derivative tensor D^j f(0) is represented as a Tucker tensor train (T4S), fit from random directionally symmetric forward/reverse probes at a single expansion point after a derivative-informed dimension reduction. The authors supply Riemannian Gauss–Newton (TR-RMGN) and Cauchy-step SGD (MC-SGD) algorithms with edge-condition rank continuation, fast sweeping methods for the Riemannian Jacobian and its adjoint, and representational error bounds (Theorem 8, Corollary 9) that depend on the spectral decay of C and the induced norm of D^k q. Numerical experiments show that the fitting procedures match quasi-optimal T3-SVD accuracy from probes alone up to data-limited ranks on random preconditioned tensors, and that T4S recovers high-order Taylor structure on two Poisson PDE examples.","tokens_in":56649,"tokens_out":988,"duration_ms":8948,"significance":"High-order Taylor surrogates for high-dimensional implicit maps have long been regarded as intractable because the derivative tensors are enormous and accessible only through probes. The combination of directionally symmetric probing (O(mk) shared-operator linearized solves), Tucker-tensor-train compression, and derivative-informed sketching makes such surrogates practical under spectral decay of C. The representational theory (peeling argument, symmetry-to-Tucker reduction, hyperbolic-cross eigenvalue sums) is carefully developed, the algorithms are specified at the level of gauged tangent vectors and sweeping contractions, and the random-tensor experiments provide an independent T3-SVD baseline. If the claims hold under the stated hypotheses, the work supplies a concrete, derivative-accurate alternative to global operator learning for outer-loop problems that only need local accuracy near a design or prior mean.","major_comments":[{"comment":"The central claim is supported under the paper’s own hypotheses (spectral decay of C or additional low-rank structure in D^j q; local Taylor validity). No load-bearing internal inconsistency was found in the peeling argument (Proposition 2, Lemmas 3–5), the infinite-dimensional reduction (Lemma 7, Theorem 8), or the probe-cost analysis (Table 1, §3.5). The experiments match T3-SVD from probes alone (Figs. 11–14) and recover high-order Taylor structure on the Poisson examples (Figs. 16–22). I therefore raise no major technical objections that would require a rewrite of the core contribution.","section":null}],"minor_comments":[{"comment":"§1.1 and the abstract correctly flag locality and spectral-decay requirements; a short forward pointer in the abstract to the precise hypotheses of Theorem 8 / Corollary 9 would help readers who stop at the abstract.","section":null},{"comment":"Figure 2 and the surrounding discussion of graphical tensor notation are clear, but a one-sentence reminder that the output mode is the last index would reduce momentary confusion when reading the T3 definition (Definition 4).","section":null},{"comment":"In §4.4.1 the “useless rank removal” three-phase sweep is described only in prose; a short algorithmic box or pseudocode would make the procedure easier to re-implement.","section":null},{"comment":"The MC-SGD stopping criterion (§4.3.2) uses fixed constants C_τ=1, C_t=3 and |B|=⌊n_s/10⌋. A brief sensitivity remark (or a single additional panel) would strengthen the claim of “little hyperparameter tuning.”","section":null},{"comment":"Typographical: “dimen-sion” hyphenation artifact appears in a figure caption in §7.2.1; “co-vector” in Fig. 21 is fine but could be “covector” for consistency with the rest of the text.","section":null},{"comment":"Related-work placement: the connection to [65] and [15,16] is noted, but a sentence contrasting T4S training data (symmetric probes at one point) with those works’ moment/correlation constructions would clarify novelty for readers coming from the stochastic-PDE literature.","section":null}],"recommendation":"accept","confidential_remarks":"The manuscript is long and dense but technically solid; it is a natural fit for a computational mathematics / scientific computing journal. The authors are transparent about scope (locality, spectral decay). I see no citation or novelty-disclosure issues that would require editorial intervention."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"This is a clean methods paper that reopens high-order local Taylor surrogates for maps of the form f(x)=q(θ0+Cx) when C has decaying spectrum. The actual novelty is the combination: each D^j f(0) is a Tucker tensor train, trained only from random directionally symmetric probes (O(mk) shared-operator linear solves instead of O(m 2^k) asymmetric ones), plus derivative-informed shared bases, TR-RMGN / Cauchy-SGD with edge-condition rank continuation, and sweeping Riemannian Jacobians. That package is new relative to their earlier TT-Taylor work and standard TT/Tucker geometry.\n\nWhat works well is the theory and the random-tensor experiments. The peeling argument (Prop. 2 + Lemmas 3–5), the symmetry-to-Tucker reduction, the infinite-dimensional reduction, and the hyperbolic-cross bound (Thm. 8 / Cor. 9) are carefully done under the stated Hilbert–Schmidt and bounded-derivative hypotheses. On dense random preconditioned tensors the fitting methods track quasi-optimal T3-SVD accuracy from probes alone up to the data-limited rank; that is the right baseline and it holds. The two 2-D Poisson examples recover the true Taylor structure once enough probes are used, and the logistic case shows the local-Gaussian trick for non-Gaussian parameters works at least illustratively.\n\nSoft spots are the ones the paper itself flags: locality of a single expansion point and the requirement that spectrum of C (or extra low-rank structure in the derivatives) keep ranks moderate. Those are scope conditions, not hidden flaws. Free parameters exist (k, ε, n_s, continuation knobs), but the rank-continuation and Cauchy step are designed to keep tuning light. No code is shipped, and the PDE examples stay 2-D Poisson; that is fine for a methods paper but means the practical outer-loop payoff is still prospective.\n\nThis is for people who already care about derivative-informed surrogates, UQ outer loops, or tensor-network geometry for Fréchet derivatives. The math and the probe-vs-T3-SVD evidence are solid enough that a serious referee should see it. I would engage with it and expect to cite the probe-cost analysis and the representational bounds.","headline":"Solid methods paper that makes high-order local Taylor surrogates for covariance-preconditioned implicit maps tractable via Tucker-TT + symmetric probes; theory and probe experiments are the real strength.","tokens_in":57238,"tokens_out":583,"would_cite":true,"duration_ms":8320,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["65N21","15A69","65K10","65F30"],"pacs":[],"model":"grok-4.5","headline":"High-order Taylor surrogates for implicit maps become tractable by representing each derivative tensor as a Tucker tensor train fit from cheap random probes.","keywords":["Tucker tensor train","Taylor series surrogate","derivative probes","Riemannian optimization","implicit maps","PDE-constrained maps","rank continuation","covariance preconditioning"],"falsifier":"On a family of random preconditioned tensors or Poisson problems whose covariance eigenvalues decay only as i^{-1} or slower, measure whether the relative forward error of the fitted T4S continues to drop with increasing rank, or whether the ranks needed already exceed the data budget before the error reaches the T3-SVD baseline.","tokens_in":57201,"feed_emoji":"📐","tokens_out":983,"duration_ms":11718,"temperature":0.7,"pith_summary":"High-order Taylor expansions of maps that depend on the solution of a nonlinear system (for example a PDE) have long been regarded as unusable in high dimensions: the derivative tensors are enormous and can be touched only by probing. This paper shows that those tensors can be replaced by compact Tucker tensor trains, yielding a local surrogate called the Tucker Tensor Train Taylor Series (T4S). The surrogate is trained not from many input-output pairs, but from random directionally symmetric probes of the derivatives at a single expansion point; each such probe needs only a handful of linear solves that share the same operator. Under spectral decay of the input covariance the paper proves that moderate ranks already give controllable approximation error, and the numerical algorithms (derivative-informed dimension reduction, Riemannian Gauss-Newton or Cauchy SGD with rank continuation, and fast sweeping Jacobians) recover nearly optimal accuracy from probes alone. The result matters because outer-loop tasks such as inverse problems, optimal design, and uncertainty quantification repeatedly need both the map and its derivatives; a derivative-accurate local surrogate can replace the expensive implicit solve inside those loops.","feed_headline":"High-order Taylor surrogates made tractable with tensor trains","feed_subtitle":"Each derivative is a Tucker tensor train fit from cheap random probes at one point","key_machinery":"The Tucker Tensor Train Taylor Series (T4S): each Fréchet derivative tensor is written as a Tucker decomposition whose central core is itself a tensor train, fitted on the Riemannian manifold of fixed-rank trains by trust-region Gauss-Newton or Cauchy-step SGD with rank continuation, using fast sweeping routines for the Riemannian Jacobian.","core_discovery":"A truncated Taylor series whose derivative tensors are each replaced by a Tucker tensor train (T4S) is a computationally tractable local surrogate for a covariance-preconditioned, implicitly defined map. The trains can be fitted from random directionally symmetric probes that cost far less than function evaluations or asymmetric probes, and spectral decay of the covariance supplies explicit rank-error bounds that guarantee the representation exists with moderate ranks.","pith_inferences":["A mixture of several T4S expansions centered at different points could extend the local surrogate into a piecewise-global model without changing the core fitting machinery.","The same probe-and-fit pipeline could be used to compress high-order derivatives that appear inside Newton or Gauss-Newton outer loops themselves, turning each outer iteration into a low-rank linear algebra step.","If the covariance spectrum is only moderately decaying, hybrid bases that combine the leading eigenmodes of C with a few active-subspace directions may keep ranks practical."],"forward_implications":["Outer-loop algorithms that need many evaluations of an implicit map and its derivatives can replace each nonlinear solve by a cheap T4S evaluation once the trains are built at a single point.","Training cost scales linearly with derivative order rather than exponentially, because only directionally symmetric probes are required.","Rank-continuation with edge-condition balancing and Cauchy step sizes removes most hyper-parameter tuning from the fitting stage.","The same representational guarantees apply to any map whose derivatives are preconditioned by a Hilbert-Schmidt operator with decaying spectrum, not only PDE maps."],"fun_headline_variants":["Tucker tensor trains make high-order Taylor surrogates tractable","T4S: derivative tensors as Tucker trains from cheap probes","High-order local surrogates via Tucker tensor train Taylor series","Fit high-dim Taylor expansions from symmetric derivative probes","Tensor-train Taylor series recovers high-order structure from probes"],"cache_read_input_tokens":49280,"weakest_assumption_plain":"The method works only when the input covariance spectrum decays fast enough (or the derivative tensors themselves are low-rank) so that the required Tucker and tensor-train ranks stay moderate; otherwise storage and fitting cost explode, and the local Taylor expansion is valid only near the chosen expansion point.","fun_headline_variants_meta":{"raw":{"variants":["Tucker tensor trains make high-order Taylor surrogates tractable","T4S: derivative tensors as Tucker trains from cheap probes","High-order local surrogates via Tucker tensor train Taylor series","Fit high-dim Taylor expansions from symmetric derivative probes","Tensor-train Taylor series recovers high-order structure from probes"]},"model":"grok-4.5","effort":"low","cost_usd":0.005064,"raw_usage":{"total_tokens":1434,"prompt_tokens":789,"num_sources_used":0,"completion_tokens":89,"cost_in_usd_ticks":50640000,"prompt_tokens_details":{"text_tokens":789,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":556,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":789,"tokens_out":89,"duration_ms":5123,"temperature":1.0,"reasoning_tokens":556,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-13T21:00:31.198933+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"On a family of random preconditioned tensors or Poisson problems whose covariance eigenvalues decay only as i^{-1} or slower, measure whether the relative forward error of the fitted T4S continues to drop with increasing rank, or whether the ranks needed already exceed the data budget before the error reaches the T3-SVD baseline.","supporting_citations":[],"review_version":1}