{"id":"a101b02a-5b19-4d4d-a862-8e1cc0af35e3","arxiv_id":"2411.13344","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"Reducing a subsystem against a low-order abstraction of its environment gives accuracy close to full closed-loop reduction at a fraction of the cost, with robust error and stability certificates when the full model is available.","lead":"This paper introduces abstracted model reduction, a way to shrink large interconnected engineering models by reducing each component while it is attached to a small, approximate stand-in for the rest of the system. The goal is to make structure-preserving model reduction computationally feasible for high-order systems such as lithography machines.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The robust error guarantee in Algorithm 2/Theorems 2-4 requires the true environment E to build the nominal model N in (16); for the modular unknown-environment scenario, Remark 1 concedes Section IV cannot be applied, so the strongest advertised guarantee excludes the motivating case.","rationale":"The reader's weakest_assumption identifies the same load-bearing concern: the robust error certificate in Algorithm 2 and Theorem 2 needs the true environment E to build N(s) and to verify budgets, and Remark 1 explicitly says the formal analysis does not apply when only a preliminary Ê is available. I agree this is the most consequential limitation for the paper's advertised guarantee. The paper is otherwise internally coherent: Lemma 2 and Theorem 1 provide a plausible exact error decomposition, Theorem 2 is a standard robust-performance certificate, and the numerical study in Section VI-A supports the empirical efficiency/accuracy claim for Algorithm 1 when E is known. The remaining issues mentioned by the reader—conservatism of the automatic order selection, pointwise frequency-grid certification, and absence of released code/data—are real but secondary; they do not shift the verdict beyond CONDITIONAL. Because the concern is already acknowledged by the authors and does not break the stated problem formulation (Section II-C assumes E is given), it conditions the scope of the guarantee rather than falsifying the framework. Therefore the reader's CONDITIONAL verdict stands unchanged.","tokens_in":25504,"tokens_out":33184,"duration_ms":362963,"concrete_test":"On the lithography benchmark, run Algorithm 2 twice: once constructing N(s) from the true E22 as specified in (16)-(17), and once replacing E22 in those equations by the abstraction Ê22 from a previous design iteration, while validating the final reduced interconnected model against the true E on a substantially finer frequency grid than the 250-point grid used in Section VI-B. Compare the certified (rE,rΣ), the feasibility of the budget optimization, and the measured ||ΛC|| against the prescribed εC(ω). If the surrogate-based certificate is infeasible or fails to bound the true error, the full-E requirement is confirmed as load-bearing for the modular scenario; if it succeeds, a modified theorem might support the motivating case.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's strongest advertised deliverable—automatic order selection with a stability and accuracy guarantee—rests on the nominal model N(s) in (16), which contains M(s)=Σ(I−E22Σ)^{-1} from (17), i.e., the exact coupled dynamics of the true environment E. Constructing N and solving the budget optimization in Theorem 3 or Theorem 4 therefore requires evaluating full-order transfer functions of the interconnection before any reduction is performed. This sits awkwardly with the modular-design motivation in Section II-B, which states that in a parallel design process the environment model E is typically unavailable or only a rough estimate Ê is available. Remark 1 concedes the point explicitly: if only a preliminary model Ê is known, the formal error analysis of Section IV cannot be applied. Consequently, Algorithm 2 provides a guarantee only in the case where E is already known and evaluable—the case where the framework's computational benefit is limited to the reduction step but not to the modeling/design workflow. This does not invalidate the empirical Table II comparison, but it means the central claim's guarantee component is conditional on access to the very full-order environment model that the abstraction idea is intended to avoid. The limitation is acknowledged rather than hidden, so it is a scope restriction on the central claim rather than an internal inconsistency.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces \"abstracted model reduction,\" a framework for reducing a subsystem of an interconnected system while it is coupled to a low-order abstraction of its environment rather than to the full environment. Algorithm 1 performs four steps: abstract E to \\hat{E}, augment the abstraction to form \\hat{F}, apply a structure-preserving reduction method to Fl(\\hat{F},\\Sigma), and finally substitute the original E back to obtain Fl(E,\\hat{\\Sigma}). Section IV derives an LFT expression for the coupled error \\Lambda_C = Fl(E,\\hat{\\Sigma}) - Fl(E,\\Sigma) in terms of the abstraction error \\Lambda_{E,22} and the reduction error \\tilde{\\Lambda}_F, namely \\Lambda_C = Fu(N,\\mathrm{diag}(\\Lambda_{E,22},\\tilde{\\Lambda}_F,\\Lambda_{E,22})) with nominal model N in (16). Using \\mu-analysis and the authors' prior robust-performance results, Theorem 2 and Corollary 1 convert a prescribed bound on \\Lambda_C into bounds on the two error sources, and Theorem 3 / Theorem 4 formulate optimization problems for allocating these error budgets. Algorithm 2 is the resulting \"robust abstracted reduction\" routine with automatic order selection. The paper evaluates both frameworks on a 2136-th-order lithography structural-dynamics model with five components. In the fixed-order comparison, abstracted closed-loop balanced reduction (aCLBR) achieves \\|\\Lambda_C\\|_2 = 7.70e-6 in 39 s, close to full ISBR (6.24e-6 in 231 s) and much better than subsystem balanced reduction (2.56e-5 in 19 s).","tokens_in":25796,"tokens_out":6225,"duration_ms":74292,"significance":"If the central claims hold, the framework is a genuinely useful bridge between cheap open-loop reduction and expensive structure-preserving reduction: it gives a principled way to choose a low-order environment abstraction, relates abstraction and reduction errors to the final interconnected-system error through a clean LFT formula, and demonstrates on a realistic industrial model that most of the accuracy benefit of structure-preserving reduction can be retained at a small fraction of its cost. The use of \\mu-analysis is appropriate, and the paper is notably honest: Remark 1 acknowledges that the formal error analysis requires the true environment model, and Section VI-B transparently reports the conservatism of the robust routine. The empirical comparison in Table II is informative, and the 80% environment-reduction result is a useful, falsifiable statement about the motivating application. The main limitation, discussed below, is that the strongest advertised guarantee does not cover the modular/unknown-environment scenario that motivates the framework in Section II-B; this is a scope restriction rather than an internal inconsistency.","major_comments":[{"comment":"The robust guarantee advertised in the abstract and Section V is conditional on having the true environment E, not on having only a preliminary model \\hat{E}. Algorithm 2 and Theorem 3 construct the nominal model N(s) in (16), which contains M(s)=\\Sigma(I-E_{22}\\Sigma)^{-1} from (17); this is the exact closed-loop map of the true E. Similarly, verifying the budgets \\Lambda_{E,22}\\in \\overline{\\Lambda}_{E,22} and \\tilde{\\Lambda}_F\\in \\overline{\\Lambda}_F requires evaluating the full-order interconnection before reduction. Remark 1 explicitly concedes that if only a preliminary \\hat{E} is known, \"the formal error analysis of Section IV can not be applied.\" Consequently, Algorithm 2 does not deliver a guarantee for the modular design scenario described in Section II-B, where E is typically unavailable or only a rough estimate. The abstract and conclusions should be qualified to state that the automatic stability/accuracy guarantee applies when E is already known and evaluable, and the paper should either extend the analysis to environment-model uncertainty or clearly mark this as a key open problem. The empirical Table II comparison is not affected, but the strongest advertised deliverable is.","section":"Sec. II-B, II-C, Remark 1, Algorithm 2"},{"comment":"The formal error analysis requires square invertible weighting matrices Gu and Gy, but the main empirical comparison of the general framework uses Gu=Gy=O. Lemma 2 states the inversion result for \"square, invertible matrices Gu and Gy,\" yet Theorem 1, which follows immediately, does not restate this assumption and defines \\tilde{\\Lambda}_F := G_y^{-1}\\Lambda_{F,22}G_u^{-1}. In Section VI-A the authors write \"For aCLBR, we use Gy = Gu = Om\\times m, i.e., no augmentation.\" For zero weights, G_y^{-1} and G_u^{-1} do not exist, Lemma 2's expression (9) does not follow, and the error relation (15) is not meaningful as stated. If the VI-A experiment is intended only as an empirical demonstration of Algorithm 1 without invoking the Section IV guarantees, this should be stated explicitly; if the error analysis is claimed to cover the configuration used in the comparison, the assumptions of Lemma 2 and Theorem 1 must be aligned with the actual choice of Gu and Gy.","section":"Sec. IV, Lemma 2 / Theorem 1 and Sec. VI-A"},{"comment":"The abstract promises a single systematic approach that \"preserve[s] stability and guarantee[s] a given frequency-dependent error specification,\" but no single algorithm in the paper delivers both simultaneously in the frequency-dependent sense. Algorithm 2/Theorem 3 provides stability and a weighted H\\infty guarantee, with a fixed (bistable, biproper) weighting function; an arbitrary user-defined frequency-dependent specification is handled by Theorem 4/Corollary 1, which, as Remark 6 states, comes \"at the loss of any guarantees on well-posedness, stability or on an error bound for other frequency points.\" The claims should be reworded to distinguish the H\\infty-based guarantee (stability, one weighted norm) from the frequency-grid-based method (no stability guarantee), so that the reader is not led to believe that both properties are guaranteed by the frequency-dependent procedure alone.","section":"Abstract and Section V"}],"minor_comments":[{"comment":"The phrase \"while having a computational cost almost competitive wth ssBR\" contains a typo: \"wth\" should be \"with.\"","section":"Sec. VI-A, text before Table II"},{"comment":"The quantity called the \"bounded L2-norm\" is an integral over the finite frequency band [10,10^6] Hz; this is not the standard L2/H2 norm on the imaginary axis, so the terminology should be clarified or the integral should be identified as a finite-band approximation.","section":"Eq. (46)"},{"comment":"In Figure 12, the use of \\phi in the axis labels (e.g., \"n_{phi} - r_{phi}\") is not defined in the caption; the reader must infer that \\phi denotes the thin-plate system \\Sigma. Please add an explicit definition.","section":"Sec. VI-B, Figs. 12-13"},{"comment":"The sentence \"The full execution of the robust abstracted reduction framework of Algorithm 2 using these settings takes approximately 450 s\" is duplicated in the main text and in Remark 9; keep the information once.","section":"Sec. VI-B, Remark 9"},{"comment":"The statement that the specification \\Lambda_C(\\omega) with \\epsilon_C=10^{-7} is achievable by aCLBR with r_E=20 and r_\\Sigma=8 and by ssBR with r_\\Sigma=22 is valuable and should be placed more prominently, because it quantifies the conservatism of the robust routine and supports the general abstracted-reduction conclusion.","section":"Sec. VI-B, final paragraph"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is honest and technically careful; the reader's circularity concern does not land, since the error budgets are derived from the nominal closed-loop model and verified a posteriori. The main issue is scope: the robust guarantee requires the true environment model, which is exactly what the motivating modular-design scenario lacks. This is acknowledged in Remark 1, but the abstract and conclusions overstate the deliverable. The second issue (invertibility assumptions versus Gu=Gy=O in the main comparison) is also fixable by clarifying the assumptions and scope. I would recommend major revision rather than rejection, because the core LFT error relation and the empirical demonstration are sound and the required changes are local: qualify the claims, align assumptions with the numerical configuration, and clearly separate the two guarantee regimes. No concerns about citation practices or novelty disclosure."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"I read the Poort et al. paper on abstracted model reduction. Here is my take: the core idea is sound and genuinely useful, but the strongest advertised guarantee is narrower than the framing suggests.\n\nWhat is actually new: they replace the full environment E(s) in a structure-preserving reduction by a low-order abstraction Ê(s), reducing the cost of reducing a subsystem while keeping most of the closed-loop relevance. The error analysis in Theorem 1 and Lemma 2 is the right machinery: they express the coupled error ΛC as an LFT of the abstraction and reduction errors, then use robust performance (their prior work [17,27]) to allocate error budgets. That is a real extension, not a routine application. The lithography case study is convincing: aCLBR gets ||ΛC||2 = 7.7e-6 at 39 s versus 6.24e-6 at 231 s for ISBR, and clearly beats ssBR. For the fixed-order experiment, the efficiency-accuracy trade-off is demonstrated.\n\nThe soft spots, in order of severity. First, the robust automatic framework (Algorithm 2, Theorems 3-4) requires the true environment E to build N(s), which contains M(s)=Σ(I−E22Σ)^{-1}. So the stability and accuracy guarantee applies only when E is known and evaluable—not in the modular-design scenario where only a preliminary Ê is available. Remark 1 concedes this explicitly, so it is a scope limitation rather than a hidden flaw, but it undercuts the \"automatic\" selling point for the motivating case. Second, the frequency-grid variant (Theorem 4) is pointwise and gives no stability guarantee; the paper says this. Third, the robust procedure is very conservative: actual error is two orders of magnitude below the allowed budget, and rE=184/rΣ=63 versus manually achievable rE=20/rΣ=8. Fourth, the tuning of β and Gu/Gy is fiddly; their own Figure 13 shows the weighting behaves non-intuitively with CLBR. Minor: no code or data, and the certification is done on a finite grid.\n\nNone of these are load-bearing flaws. The paper's central claim—that low-order environment abstractions can substitute for the full environment in structure-preserving reduction—is supported by the case study. The analysis is honest about its limitations.\n\nWho it is for: researchers working on model reduction of interconnected and structural systems, especially with industrial-scale models. It deserves a serious referee. I would send it to review and ask the authors to clarify the scope of the guarantee in the abstract and introduction, since the current framing overstates coverage of the unknown-environment case.","headline":"Solid framework paper: abstracted reduction is a real step forward for structure-preserving reduction, but the advertised automatic guarantee needs the full environment model, so it does not cover the unknown-environment case.","tokens_in":26316,"tokens_out":2475,"would_cite":true,"duration_ms":27552,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["93B11","93A15","93C05"],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper proposes abstracted model reduction: substitute a low-order stand-in for the environment before applying structure-preserving reduction, and keep most of the closed-loop accuracy at a fraction of the cost.","keywords":["abstracted model reduction","structure-preserving model reduction","interconnected systems","robust performance","linear fractional transformation","balanced reduction","error budget allocation","structural dynamics"],"falsifier":"Construct an example where the environment abstraction $\\hat{E}$ satisfies the $\\Lambda_{E,22}$ bound but is deliberately poor in $\\hat{E}_{11}, \\hat{E}_{12}, \\hat{E}_{21}$; the paper's Theorem 1 says only $\\Lambda_{E,22}$ enters $\\Lambda_C$ directly, but $\\Lambda_F$ depends on all of $\\hat{E}$, so a large measured $\\Lambda_C$ would show that the off-diagonal abstraction error cannot be ignored.","tokens_in":25320,"feed_emoji":"⚙️","tokens_out":7250,"duration_ms":71579,"temperature":0.7,"pith_summary":"Abstracted model reduction is a framework for reducing one subsystem of a high-order interconnected model by first replacing the rest of the model—its environment—with a much smaller approximation, then running a structure-preserving reduction method on the subsystem coupled to that stand-in. The paper's central claim is that this swap preserves most of the accuracy benefit of full closed-loop reduction while cutting the computational cost from cubic scaling in the full interconnected order to cubic scaling in the abstracted order. In the lithography-structure case study, the abstracted implementation of closed-loop balanced reduction costs 39 seconds and gives an error norm $7.70\\times10^{-6}$, compared with 231 seconds and $6.24\\times10^{-6}$ for the full structure-preserving method and 19 seconds and $2.56\\times10^{-5}$ for plain subsystem reduction. The paper also develops a robust-performance-based procedure that allocates a user-specified error budget between environment abstraction and subsystem reduction, and proves stability and accuracy guarantees when the true environment model is available.","feed_headline":"Cheap environment stand-in cuts model-reduction cost sixfold","feed_subtitle":"Replacing a subsystem's environment with a low-order abstraction preserves closed-loop accuracy at a fraction of the computation.","key_machinery":"The central object is the abstracted interconnection $F_l(\\hat{F}, \\Sigma)$, obtained by replacing the environment $E(s)$ with a low-order $\\hat{E}(s)$ and augmenting it with weighting matrices $G_u$ and $G_y$. The paper shows that $\\hat{\\Sigma}$ can be recovered from the augmented closed-loop error $\\Lambda_F$ through an upper LFT inversion, and that the final error satisfies $\\Lambda_C = F_u(N, \\operatorname{diag}(\\Lambda_{E,22}, \\tilde{\\Lambda}_F, \\Lambda_{E,22}))$, where $M = \\Sigma(I - E_{22}\\Sigma)^{-1}$ is the closed-loop map of the true system. This identity carries the argument: it separates the error budget into an abstraction part and a reduction part, allows the robust-performance optimization of Theorems 3 and 4, and justifies substituting the original environment after reduction.","core_discovery":"The core discovery is that the environment of a subsystem acts mainly as a weighting that determines which subsystem dynamics matter for the interconnected response, and that this weighting can be replaced by a low-order abstraction without losing the essential information. Concretely, the paper expresses the total interconnected reduction error as $\\Lambda_C = F_u(N, \\operatorname{diag}(\\Lambda_{E,22}, \\tilde{\\Lambda}_F, \\Lambda_{E,22}))$, with $M = \\Sigma(I - E_{22}\\Sigma)^{-1}$, showing that only the $22$-block of the abstraction error enters this relation directly, while the reduction error is measured through the augmented closed loop $F_l(\\hat{F}, \\Sigma)$. This error relation is turned into a certificate: if the abstraction and reduction errors stay within frequency-weighted bounds that satisfy a scaled structured-singular-value condition, then the reduced interconnected model is stable and meets the prescribed accuracy specification. The numerical demonstration shows that reducing the environment by over 80% does not degrade the accuracy of the reduced interconnected model, which is the concrete sense in which the abstraction is sufficient.","pith_inferences":["The error-relation theorem suggests that the off-diagonal blocks $\\hat{E}_{11}, \\hat{E}_{12}, \\hat{E}_{21}$ of the abstraction only influence the final error through the reduction step; a natural testable extension is to check whether cheaply approximating only $E_{22}$ is enough when the reduction method is insensitive to the other blocks.","Because the certificate requires the true environment, the practical modular-design setting discussed in Remark 1 would need a separate validation procedure, for example bracketing the unknown environment with a family of models and checking the error bound over that family.","The $\\beta$ trade-off suggests an automated hyperparameter selection based on estimated reduction cost, rather than the manual tuning currently needed to balance $r_E$ and $r_\\Sigma$.","If abstraction is viewed as a learned or data-driven surrogate, the same framework could allow reduction when only input-output data of the environment is available, though the formal error analysis would need replacing."],"forward_implications":["Structure-preserving reduction becomes computationally feasible for interconnected systems whose environment order dominates, because the cost drops from $(n_E+n_\\Sigma)^c$ to $(r_E+n_\\Sigma)^c$.","A low-order abstraction of the environment is sufficient to indicate which subsystem dynamics to retain: in the case study, reducing the environment by over 80% did not significantly degrade the reduced interconnected model.","User-specified frequency-dependent accuracy specifications can be met automatically by allocating error budgets between abstraction and reduction, with stability preserved when a true environment model is available.","Abstracted reduction is compatible with any structure-preserving reduction method, so improved reduction methods can be plugged into the framework without changing the abstraction step.","In the benchmark, abstracted closed-loop balanced reduction achieved $\\|\\Lambda_C\\|_2 = 7.70\\times10^{-6}$ in 39 seconds, close to the full structure-preserving result $6.24\\times10^{-6}$ in 231 seconds."],"supporting_citations":[{"why":"Supplies the ISBR method that serves as the expensive structure-preserving baseline in the comparison.","marker":"[9]"},{"why":"Supplies the closed-loop balanced reduction method used inside the abstracted reduction algorithm.","marker":"[18, 19]"},{"why":"Provides the robust performance perspective and the scaled structured-singular-value condition used in Theorem 2.","marker":"[27]"},{"why":"Supplies the prior top-down modular reduction approach and the iterative solution used for the error-budget optimization.","marker":"[17]"},{"why":"Supplies the component-mode-synthesis method used to abstract the environment model in the case study.","marker":"[34]"},{"why":"Provides the residualization and balanced-reduction background used for the subsystem reduction step.","marker":"[6]"},{"why":"Provides the LFT definitions and internal-stability conditions on which the interconnected-system formulation rests.","marker":"[20]"}],"fun_headline_variants":["Environment abstraction slashes model-reduction cost","80% smaller environment, same interconnected accuracy","Low-order stand-in for environment speeds up reduction","Structure-preserving reduction gets a tractable shortcut"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The formal error and stability guarantees require that the full, true environment model $E(s)$ is available before reduction, so the certificate does not cover the modular-design situation in which only a rough environment model is known.","fun_headline_variants_meta":{"raw":{"variants":["Environment abstraction slashes model-reduction cost","80% smaller environment, same interconnected accuracy","Low-order stand-in for environment speeds up reduction","Structure-preserving reduction gets a tractable shortcut"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000191,"raw_usage":{"total_tokens":1361,"prompt_tokens":982,"completion_tokens":379,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":598,"completion_tokens_details":{"reasoning_tokens":322}},"tokens_in":598,"tokens_out":379,"duration_ms":4391,"temperature":1.0,"reasoning_tokens":322,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T16:33:26.995015+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Construct an example where the environment abstraction $\\hat{E}$ satisfies the $\\Lambda_{E,22}$ bound but is deliberately poor in $\\hat{E}_{11}, \\hat{E}_{12}, \\hat{E}_{21}$; the paper's Theorem 1 says only $\\Lambda_{E,22}$ enters $\\Lambda_C$ directly, but $\\Lambda_F$ depends on all of $\\hat{E}$, so a large measured $\\Lambda_C$ would show that the off-diagonal abstraction error cannot be ignored.","supporting_citations":[{"cited_title":"Model Reduction of Interconnected Systems,","cited_arxiv_id":null,"evidence_quote":"Supplies the ISBR method that serves as the expensive structure-preserving baseline in the comparison."},{"cited_title":"Modular model reduction of interconnected systems: A robust performance analysis perspective,","cited_arxiv_id":null,"evidence_quote":"Provides the robust performance perspective and the scaled structured-singular-value condition used in Theorem 2."},{"cited_title":"Modular Model Reduction of Interconnected Systems: A Top-Down Approach,","cited_arxiv_id":null,"evidence_quote":"Supplies the prior top-down modular reduction approach and the iterative solution used for the error-budget optimization."},{"cited_title":"A general purpose, multi-stage, component modal synthesis method,","cited_arxiv_id":null,"evidence_quote":"Supplies the component-mode-synthesis method used to abstract the environment model in the case study."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the residualization and balanced-reduction background used for the subsystem reduction step."},{"cited_title":"Zhou and J","cited_arxiv_id":null,"evidence_quote":"Provides the LFT definitions and internal-stability conditions on which the interconnected-system formulation rests."}],"review_version":1}