{"id":"55bf5b79-0920-4e57-9474-01a3a93674ac","arxiv_id":"2608.12837","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"Parametric matrix models learn small matrix representations of expensive parametric physics models from data, preserving the algebraic form of the equations while avoiding explicit projection bases.","lead":"This dissertation develops parametric matrix models, a machine-learning emulator that fits reduced versions of known physics equations to training data instead of building explicit projection bases. It argues these models combine the accuracy of reduced-basis methods with the flexibility of neural networks for expensive nuclear many-body calculations.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Data-efficiency guarantee rests on an unproven and non-unique hidden-snapshot spectrum (§3.4.4); the claimed O(min(m,n)^{-β}) scaling has no demonstrated link to observable PMM training data.","rationale":"The reader's CONDITIONAL verdict is judicious. My stress test identified the same load-bearing weakness: Section 3.4.4's hidden-snapshot singular-value assumption. The claim that PMMs perform \"at least as well as the associated eRBM\" depends on two unsupported links: that a unique hidden snapshot matrix exists and that its singular values decay. For PMMs used as pure regression models, the first link is not even well-defined; for physical eigenproblems, the second is an empirical claim that random-matrix models violate. The thesis is honest that this is an assumption rather than a theorem, but the error scaling is presented as a formal property of the method, and the abstract's trustworthiness claim inherits the gap. The proposed GOE test is a minimal falsifiable check: it uses the same affine Hermitian structure as the thesis examples, makes the hidden eigenvectors explicit, and directly tests whether PMM error follows the claimed scaling when the hidden spectrum does not decay. This concern does not disprove the practical usefulness of PMMs; it shows that the rigorous data-efficiency guarantee is unsupported. That is consistent with the reader's CONDITIONAL verdict, so no verdict change is needed.","tokens_in":53208,"tokens_out":11709,"duration_ms":113591,"concrete_test":"Construct a random-matrix affine family H(μ)=H0+μH1 in dimension N=100, with H0 and H1 independent GOE matrices. Compute the exact ground-state eigenvectors at m=20 training points μ_i; verify that the N×m snapshot matrix has a flat, non-decaying singular-value spectrum. Train a POD-Galerkin-based PMM of dimension n=2,4,...,20 on only the scalar ground-state energies, and measure held-out error at 50 test values of μ. If the error does not decrease as O(n^{-β}), the §3.4.4 data-efficiency claim fails for this physically relevant family. If the error does decrease, repeat the test with a random completion of the same scalar data to show that the hidden spectrum is not identifiable from scalar outputs alone.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that PMMs are data efficient and \"at least as well as the associated eRBM\" is not established. Section 3.4.4 argues that scalar training data hide an unknown matrix of snapshot vectors whose singular values decay exponentially or as a power law, and from this derives an approximate error bound O(min(m,n)^{-β}). This is load-bearing because it is the only quantitative support for PMM data efficiency and trustworthiness. The argument has two gaps. First, for data-driven regression (Sections 3.9 and 3.13) no underlying full-space state vector is specified, so \"the\" hidden snapshot matrix is not unique; different completions of the same scalar data can have completely different singular spectra. Second, even for eigenvalue problems where the hidden vectors are the actual eigenvectors, the decay assumption is an empirical regularity, not a consequence of the PMM construction; random-matrix families relevant to many-body physics have eigenvector snapshot matrices whose singular values do not decay. If the singular values do not decay, Eq. (2.32) gives no small residual, and the O(min(m,n)^{-β}) bound in §3.4.4 has no basis. Moreover, the step from n-dependence to min(m,n)^{-β} is not derived: Eq. (2.31) bounds the projection error by σ_{n+1}, not by a function of m, and for n>m the rank-deficiency argument conflates the number of training examples with the span of hidden vectors. The thesis itself flags the decay as \"overwhelmingly true in practice\" rather than proven, but the abstract's claims of \"nearly unparalleled ... trustworthiness\" inherit the gap.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript develops parametric matrix models (PMMs) as a class of emulators that combine reduced-basis structure with trainable reduced-space operators. The thesis positions PMMs as implicit reduced basis methods: instead of constructing an explicit projection from full-space snapshots, one specifies only the reduced dimension and structural constraints (e.g., Hermiticity) and optimizes the reduced operators against scalar or vector training data. The paper claims that PMMs inherit the interpretability, physical consistency, and systematic improvability of explicit reduced basis methods while being numerically parametric and data efficient, with an approximate error scaling O(min(m,n)^(-beta)) derived from the singular-value decay of an assumed hidden snapshot matrix. The manuscript also surveys existing emulation methods, details training via gradient-based optimization and effective parameterizations, and describes an open-source implementation (pyPMM).","tokens_in":53560,"tokens_out":4724,"duration_ms":47831,"significance":"If the central claims were established, PMMs would be a valuable new tool for nuclear and many-body emulation: they promise a numerically parametric, adaptive framework that can exploit partial knowledge of the governing equations while avoiding explicit full-space snapshots. The conceptual unification of reduced-basis methods and parametric machine learning, together with the open-source package, could lower the barrier to adoption. The toy example in Eqs. (3.4)-(3.5) is a clear and instructive demonstration that explicit reduced basis methods can fail while an implicit reduced model exists. The review of existing methods and the property tables are useful contributions. However, the theoretical support for the headline data-efficiency and trustworthiness claims is not yet rigorous: the error bound rests on an unproven and, in data-driven settings, non-unique hidden-snapshot construction, and the training procedure carries no convergence or generalization guarantee. The work is promising but requires substantial clarification and, in places, a downgrade of the strength of the claims.","major_comments":[{"comment":"The claimed error scaling O(min(m,n)^(-beta)) is the sole quantitative support for the data-efficiency and 'at least as well as the associated eRBM' claims, but it is not derived from the training data. The argument postulates hidden snapshot vectors that are never constructed; for data-driven regression (Sections 3.9 and 3.13) no underlying full-space state is specified, so the hidden snapshot matrix is not unique and different completions can have completely different singular spectra. Even where eigenvectors exist, the decay assumption in Eq. (2.33) is acknowledged in the manuscript itself as 'overwhelmingly true in practice' rather than proven, and random-matrix counterexamples are admitted. Furthermore, Eq. (2.31) bounds the projection error by sigma_{n+1}; the conversion of this to a rate in the number of training examples m is not shown, and the rank-deficiency step for n > m conflates the number of training examples with the rank of the hidden snapshot matrix. The bound should either be proven under stated assumptions or replaced by a clearly empirical claim, with the data-efficiency property reworded accordingly.","section":"Section 3.4.4 and Eqs. (2.31)-(2.33)"},{"comment":"The statement 'By construction, the resulting PMM—if trained properly—will perform at least as well as the associated eRBM' conditions the central inheritance claim on training that the manuscript does not guarantee. Section 3.5 describes gradient descent on a generally nonconvex loss, with only heuristic arguments (e.g., prevalence of near-global local minima for high-dimensional parameter spaces) and no convergence or generalization theorem. The toy example in Eqs. (3.4)-(3.5) exhibits a family of implicit models that reproduce the exact energies, but it does not show that the training procedure in Section 3.5 can locate such a family from scalar data. The property tables and the abstract present data efficiency and trustworthiness as intrinsic PMM properties; without an operational criterion for 'trained properly' or empirical certification, these claims are stronger than what is established.","section":"Section 3.4.4, first paragraph; Section 3.5"},{"comment":"The inheritance of eRBM error bounds and physical-consistency properties is claimed for PMMs, but the effective parameterizations used in practice introduce approximations whose errors are not analyzed. In particular, Eq. (3.49) replaces the POD projector P in the tensor hypernetwork expression by a trainable matrix Theta, and the resulting truncation error is not bounded. Consequently, the 'at least as well as the associated eRBM' statement does not automatically extend to the implemented PMMs for nonlinear problems; the relationship between the trained PMM and the corresponding eRBM solution is not quantified. The manuscript should either provide error estimates for these parameterizations or restrict the inheritance claims to the exactly parameterized cases.","section":"Section 3.4.4 and Section 3.6.6"}],"minor_comments":[{"comment":"The expected relative error formula is presented without stating the sample distribution or normalization of the test vectors; specify the assumptions under which this expectation is taken.","section":"Section 2.11.1.1, Eq. (2.32)"},{"comment":"The notation O(min(m,n)^(-beta)) mixes asymptotic notation in m with a problem-dependent exponent beta; clarify whether beta is a constant or a quantity that may depend on the data distribution.","section":"Section 3.4.4, last paragraph"},{"comment":"The term 'tensor hypernetwork' is nonstandard in machine learning, where 'hypernetwork' commonly refers to a network generating weights; consider renaming to 'high-order tensor contraction' or another descriptive term to avoid confusion.","section":"Section 3.6.6"},{"comment":"The legend line 'IIntrusive' appears to be a formatting artifact; the column header is missing and should be corrected.","section":"Table 2.10"},{"comment":"Phrases such as 'nearly unparalleled adaptability' are more evaluative than a formal property statement; consider rewriting these as concrete, testable claims.","section":"Abstract and Section 3.4"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is a PhD dissertation posted to arXiv, and its structure (long review chapters, personal footnotes, self-referential remarks) is not that of a typical journal article. If the journal is open to thesis-like manuscripts, the scope is acceptable; otherwise the authors should be asked to condense. The most important issue for the editor is whether the data-efficiency claim can be made rigorous or must be downgraded; the empirical sections were not fully available in the portion I reviewed, so the practical record may partially compensate. The novelty relative to the author's prior paper [62] should be clearly delineated."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague, the useful thing here is the formal packaging. The thesis takes the PMM idea from the author's earlier paper and builds a genuine framework around it: implicit reduced basis methods, a Galerkin-style derivation, a canonical matrix-valued treatment of elementwise nonlinearities, tensor hypernetwork parameterizations, and worked examples that look internally consistent. The toy model in Section 3.1 clearly shows a case where an eRBM fails and a PMM succeeds. That is real content, not just a survey. The pyPMM package is a plus, though I would want a commit hash before treating results as reproducible.\n\nThe soft spot is exactly where the stress-test note lands. Section 3.4.4 claims an O(min(m,n)^-beta) error scaling from the singular values of a hidden matrix of snapshot vectors. Those vectors are never constructed. For data-driven regression they are not even unique: different completions of the same scalar data can have completely different singular spectra. For eigenvalue problems, the decay is an empirical regularity, not a consequence of the PMM construction. The step from n-dependence to min(m,n)^-beta is also not derived; rank deficiency for n>m does not automatically produce that exponent. The thesis itself flags the decay as \"overwhelmingly true in practice\" rather than proven, so this is not a fraud, but the abstract's \"nearly unparalleled trustworthiness\" claim inherits the gap. If the author recasts this as a heuristic with caveats, the paper's core value barely changes.\n\nA few smaller items: training has no convergence guarantees, but that is normal for gradient-based ML emulators and not a disqualifier. The empirics on held-out data are the right kind of evidence, and the examples are evaluated against high-fidelity data rather than fitted values. The citation pattern is fine; citing your own prior method is not a flaw when the extension is real.\n\nWho is this for? Nuclear and many-body practitioners who want a self-contained bridge between reduced-basis methods and trainable emulators. It deserves a serious referee. I would send it out, but with a clear request to rewrite Section 3.4.4 so that the data-efficiency claim is stated as an observed regularity with identified failure modes, not a proven bound. That revision is within reach.","headline":"A well-built thesis that formalizes PMMs as implicit reduced-basis emulators, but its headline data-efficiency guarantee rests on an unproven hidden-snapshot decay assumption; soften that claim and the paper is still solid and worth refereeing.","tokens_in":702,"tokens_out":769,"would_cite":true,"duration_ms":22832,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Parametric matrix models train a reduced-space, structure-preserving version of a system's own equations from scalar data, aiming to emulate eigenvalue problems, nonlinear dynamics, and regression without constructing snapshots.","keywords":["parametric matrix models","implicit reduced basis methods","emulation","eigenvector continuation","nuclear many-body physics","nonlinear dynamics","machine learning","uncertainty quantification"],"falsifier":"Construct a parametric family whose scalar outputs come from a high-dimensional response with a flat, non-decaying singular-value spectrum (for example, random i.i.d. Gaussian output vectors at each parameter value), train PMMs of increasing size and data count, and check whether the relative error follows $O(\\min(m,n)^{-\\beta})$; if the error stalls or converges only logarithmically, the assumed hidden-vector decay is violated.","tokens_in":52962,"feed_emoji":"⚛️","tokens_out":7885,"duration_ms":74310,"temperature":0.7,"pith_summary":"The thesis contends that parametric matrix models (PMMs) offer a single emulation framework that can replace explicit snapshot-based reduced basis emulators for parametric eigenvalue problems, nonlinear dynamics, and data-driven regression. Instead of choosing a reduced subspace and projecting the full equations onto it, a PMM starts from the known or assumed form of the governing equations, derives the form its Galerkin projection would take, and replaces every unknown projected operator with a trainable matrix or tensor that respects known structure such as Hermiticity. The trained reduced model runs entirely in a low-dimensional space, so the cost of evaluation and fitting is independent of the full model size. If correct, PMMs would make uncertainty quantification and parameter fitting feasible for expensive many-body models while retaining the interpretability of reduced basis methods. The thesis also positions PMMs as 'implicit reduced basis methods' and develops the accompanying open-source pyPMM package.","feed_headline":"PMMs replace snapshot-based emulators with trained reduced equations","feed_subtitle":"Training the reduced operators directly from scalar data keeps the physics and cuts cost to reduced-space size.","key_machinery":"The load-bearing object is the parametric matrix model itself: a reduced-space matrix or tensor model whose form is derived from the Galerkin projection of the known or supposed full model, with all unknown projected objects replaced by trainable parameters. The training is carried by effective parameterizations that enforce constraints in closed form, such as Hermiticity via $H=(\\Theta+\\Theta^\\dagger)/2$, unitarity via matrix exponentials, and rank via sums of rank-one terms. For elementwise nonlinearities, a canonical matrix-valued map $Y(A)=V_A\\,\\mathrm{diag}(y(\\lambda_A))\\,V_A^{-1}$ reduces the nonlinearity to the same third-order tensor $G^{(2)}$ that handles quadratic products, preserving the exact nonlinear form without hyper-reduction. The optimization itself uses complex-valued gradient descent (a Wirtinger-modified Adam) with automatic differentiation, and uncertainty can be added through bootstrap aggregation or conformal prediction.","core_discovery":"The central discovery is that an emulator need not construct the reduced basis at all. For any system whose equations have a known parametric form, one can write down the reduced-space problem exactly as Galerkin projection would produce it, then treat the projected operators as free parameters constrained only by properties the projection would preserve (Hermiticity, vector normalization, unitarity, and the like). Training these parameters on scalar or vector data yields a numerically parametric emulator that inherits the interpretability and physical consistency of the explicit reduced basis method but does not require full-space snapshots. The thesis demonstrates the framework on affine and nonlinear eigenproblems, nonlinear dynamics, data-driven regression, and emulation of in-medium similarity renormalization group nuclear matter calculations, and it argues that for typical data the error scales as $O(\\min(m,n)^{-\\beta})$ with the number $m$ of training examples and reduced dimension $n$. This makes PMMs, in the thesis's framing, the first adaptive emulation method: it can be fully intrusive, fully data-driven, or any hybrid in between.","pith_inferences":["Beyond the paper, the hidden-snapshot argument implies that PMM-style training should be tested as a general regression method on non-physics datasets with known singular-value decay, such as image or sensor datasets, to see whether the $O(\\min(m,n)^{-\\beta})$ scaling holds there too.","The canonical matrix-valued nonlinearity suggests PMMs could serve as structure-preserving surrogates for any elementwise analytic term in large PDE solvers, for instance Gross-Pitaevskii or Hartree-Fock mean fields, with the same $G^{(2)}$ tensor overhead, a connection the thesis does not develop.","Interpreting every explicit reduced-basis method as a particular PMM parameterization opens the possibility of unifying POD, eigenvector continuation, and DMD under one training objective, which the thesis only gestures at.","A direct empirical test of the central claim would compare PMM error curves with explicit emulators on the same problems while monitoring the actual singular-value decay of the underlying snapshots, where available; the thesis reports individual examples but no systematic benchmark."],"forward_implications":["PMMs can emulate parametric eigenvalue problems from scalar energies alone, matching or beating explicit eigenvector-continuation emulators without storing or computing full eigenvectors.","Nonlinear dynamics can be emulated at reduced-space cost without hyper-reduction, because the canonical matrix-valued nonlinearity preserves the exact elementwise analytic form.","A single PMM framework covers intrusive, hybrid, and fully data-driven settings, so it remains applicable when the full model is only partially known or when snapshots are too large to store.","Because PMMs are numerically parametric, new training data can be incorporated by further gradient steps without retraining from scratch.","Attaching conformal prediction to PMMs yields calibrated prediction intervals, making the emulator usable for uncertainty-quantified parameter fitting."],"supporting_citations":[{"why":"Introduces the PMM method and the non-interacting spin example showing explicit reduced basis failure; the central technique this thesis formalizes.","marker":"[62]"},{"why":"Supplies eigenvector continuation's exponential convergence bounds and the Galerkin framing that PMMs reinterpret as explicit reduced basis methods.","marker":"[22]"},{"why":"DEIM hyper-reduction for nonlinear POD-Galerkin; the baseline nonlinearity handling that PMM's canonical matrix-valued construction improves on.","marker":"[41]"},{"why":"Petrov-Galerkin emulator with gradient-learned projectors; the closest prior method PMMs contrast with.","marker":"[123]"},{"why":"Data-driven emulator via quadratic matrix inequality; the related non-intrusive approach PMMs distinguish from.","marker":"[124]"},{"why":"Tractable a priori dimensionality reduction for quantum dynamics; motivates emulation in settings where explicit snapshots are unavailable.","marker":"[77]"},{"why":"The pyPMM package used to construct, train, and deploy all PMM examples in the thesis.","marker":"[1]"}],"fun_headline_variants":["PMMs skip snapshots, train reduced operators directly","Emulate without snapshots: PMMs train reduced operators","Parametric matrix models: snapshots optional, physics kept","Reduced operators trained from data: PMM emulation","Adaptive emulation: PMMs blend physics and data"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The data-efficiency guarantee rests on the unproven assumption that scalar training data always hide an underlying set of snapshot vectors whose singular values decay exponentially or as a power law, so that a small reduced space can capture almost all of the information.","fun_headline_variants_meta":{"raw":{"variants":["PMMs skip snapshots, train reduced operators directly","Emulate without snapshots: PMMs train reduced operators","Parametric matrix models: snapshots optional, physics kept","Reduced operators trained from data: PMM emulation","Adaptive emulation: PMMs blend physics and data"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000755,"raw_usage":{"total_tokens":3408,"prompt_tokens":1046,"completion_tokens":2362,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":662,"completion_tokens_details":{"reasoning_tokens":2281}},"tokens_in":662,"tokens_out":2362,"duration_ms":15872,"temperature":1.0,"reasoning_tokens":2281,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T22:16:16.732830+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Construct a parametric family whose scalar outputs come from a high-dimensional response with a flat, non-decaying singular-value spectrum (for example, random i.i.d. Gaussian output vectors at each parameter value), train PMMs of increasing size and data count, and check whether the relative error follows $O(\\min(m,n)^{-\\beta})$; if the error stalls or converges only logarithmically, the assumed hidden-vector decay is violated.","supporting_citations":[{"cited_title":"Colloquium: Eigenvector continuation and projection-based emulators","cited_arxiv_id":null,"evidence_quote":"Supplies eigenvector continuation's exponential convergence bounds and the Galerkin framing that PMMs reinterpret as explicit reduced basis methods."}],"review_version":1}