{"id":"10e4bdfb-eff5-4c55-902d-09b4b286a57b","arxiv_id":"2411.14242","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"Approximate constrained lumping relaxes exact lumping by a tolerance epsilon, yielding polynomial-time reductions whose output error provably scales with epsilon.","lead":"This paper presents an approximate version of constrained lumping, a model reduction technique for chemical reaction networks, allowing reductions that preserve user-selected outputs only up to a tunable error tolerance epsilon. It gives polynomial-time algorithms, proves that the output error is of order epsilon, and shows coarser reductions than exact lumping on published biochemical models.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 3's error bound omits the condition that the reduced trajectory \\bar{L}y(t) stays in the compact set Ω; for rational models it can reach a denominator zero, invalidating the order-ε claim.","rationale":"Reading in good faith: the paper's main contribution is an algorithmic relaxation of exact lumping with a provable O(ε) error bound and polynomial-time computation. The strongest claim is the error bound. The most load-bearing assumption is the compact-domain condition in Theorem 3. The proof of Proposition 1 relies on a Lipschitz estimate for Lf(\\bar{L}y) − Lf(\\bar{L}Lx), which requires both \\bar{L}y and \\bar{L}Lx to be in the domain where f has the Lipschitz constant C. The theorem statement only assumes x(t) and \\bar{L}Lx(t) are in Ω; it omits \\bar{L}y(t) ∈ Ω. For rational drifts, the reduced trajectory may approach or reach a denominator zero even when the original trajectory is well-separated from it, as shown by the concrete counterexample. This does not mean the method is wrong; it means the theorem is mis-stated and the abstract's claim is not justified for rational systems without an additional condition. The paper's Section 4.1 checks the condition post hoc for the experiments, which mitigates the practical impact but does not fix the theorem. The reader's weakest_assumption identifies exactly this issue; I agree. Because the gap is amendable and the experimental evidence supports the method's utility, I do not change the verdict: it remains CONDITIONAL.","tokens_in":24071,"tokens_out":14443,"duration_ms":127660,"concrete_test":"Simulate the counterexample above: \\dot{x}_1 = 1/(1−x_1−x_2), \\dot{x}_2 = −10, L = [1 0], x(0) = (0.5, 0), T = 0.2. Verify that the original solution is well-defined and bounded in a compact set excluding the singularity (e.g., x_1(t) < 0.9), while the reduced solution \\dot{y}_1 = 1/(1−y_1) reaches y_1 = 1 before T. If it does, Theorem 3's statement is incomplete: it must add the condition \\bar{L}y(t) ∈ Ω (or at least require the reduced solution to stay in the analyticity domain) for the bound to hold.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central O(ε) error claim rests on Theorem 3, whose proof (Proposition 1) bounds ||Lf(\\bar{L}y) − Lf(\\bar{L}Lx)|| by C||L|| ||\\bar{L}|| ||e(t)|| using a Lipschitz constant for f on Ω. This requires both arguments \\bar{L}y(t) and \\bar{L}Lx(t) to lie in Ω (or in a common domain where f has that Lipschitz constant). Theorem 3 assumes only x(t) and \\bar{L}Lx(t) ∈ Ω; it never assumes \\bar{L}y(t) ∈ Ω. For rational f, \\bar{L}y can reach a denominator zero even when x and \\bar{L}Lx stay in a safe compact set, so the reduced system is undefined (or the Lipschitz inequality fails) and the error bound is false. Section 4.1 checks this only post hoc ('the denominators of rational f(x) vanished only for negative values of x'), so the theorem does not justify the abstract's claim in general. A concrete counterexample is \\dot{x}_1 = 1/(1−x_1−x_2), \\dot{x}_2 = −10, L = [1 0], x(0) = (0.5, 0): the original solution stays in a compact set where f is analytic, but the reduced equation \\dot{y}_1 = 1/(1−y_1) blows up at t ≈ 0.125, so the error is undefined on [0, 0.2].","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes approximate constrained lumping for systems of ODEs with analytic right-hand sides, relaxing exact linear lumping through a tolerance parameter ε. For a user-specified observable subspace, Algorithm 4 computes a lumping matrix in polynomial time by adding rows only when a Jacobian-related vector is more than ε away from the current row space. The central theoretical claim is that the difference between the original observable and the observable of the reduced model is O(ε) (Theorems 3 and 5). A heuristic (Algorithm 5) selects ε by binary search to meet a target reduced size. The approach is evaluated on eight literature models (polynomial and rational) and on a scalable multisite phosphorylation model, reporting significantly smaller reductions than exact lumping at acceptable errors.","tokens_in":24364,"tokens_out":20558,"duration_ms":175491,"significance":"If the theoretical claims are correct, this is a useful contribution: it gives a polynomial-time algorithm for approximate model reduction with a provable, user-tunable error bound, extending the exact-lumping tool CLUE to rational and other analytic systems. The experimental evaluation is a genuine strength: it covers both polynomial and rational models from the literature, reports reduced sizes and errors, and the implementation is provided in a public repository. The idea of selecting ε by a size-based binary search is practical and likely to interest the systems-biology modelling community. However, the central error-bound theorem has a domain-assumption gap, and the pseudocode of the heuristic algorithm is internally inconsistent, so the manuscript needs substantial revision before the claims are fully supported.","major_comments":[{"comment":"The error bound in Theorem 3 is not justified because the proof of Proposition 1 bounds ∥Lf(\\bar{L}y) − Lf(\\bar{L}Lx)∥ by C∥L∥∥\\bar{L}∥∥e(t)∥ using a Lipschitz constant of f on Ω, but the theorem assumes only x(t) and \\bar{L}Lx(t) lie in Ω, not the reduced trajectory \\bar{L}y(t). For rational f, \\bar{L}y(t) can leave Ω or reach a denominator zero even when x(t) and \\bar{L}Lx(t) stay in a safe compact set, in which case the reduced system is undefined and the error bound is false. A concrete example is \\dot{x}_1 = 1/(1−x_1−x_2), \\dot{x}_2 = −10, L = [1 0], x(0) = (0.5, 0), where the original solution is well-defined on [0,0.2] but the reduced equation \\dot{y}_1 = 1/(1−y_1) blows up at t ≈ 0.125. Section 4.1's post hoc check ('the denominators of rational f(x) vanished only for negative values of x') only verifies x and \\bar{L}Lx, not \\bar{L}y. The same domain-tracking issue appears in Proposition 2, where the mean-value argument evaluates f at points v+x_R that are not guaranteed to lie in Ω. The theorem and its proof must either assume \\bar{L}y(t) ∈ Ω (and explain how this can be verified or guaranteed), or state the error bound as conditional on the reduced solution remaining in the domain.","section":"Algorithm 5, Section 3.3"},{"comment":"The pseudocode of Algorithm 5 is inconsistent with the surrounding text and with Example 9. With mmin = nrows(L_{εmin}) (the size at ε=0, i.e., the exact-lumping size) and mmax = nrows(L_{εmax}) (the minimal size at maximal tolerance), the branch 'if m∗ < mmin then return εmin' returns ε=0 whenever the target size is smaller than the exact size, which is precisely the case where an approximate reduction is needed. For Model 2, mmin=17, m∗=11, this branch fires and returns the exact lumping of size 17, contradicting Example 9 and the reported reduction to size 7. The branches appear to have the roles of mmin and mmax swapped, and the comparison directions also seem to be reversed relative to the goal of finding the smallest ε with size ≤ m∗. This needs to be corrected and the example rechecked.","section":"Section 3.3, before Algorithm 5"},{"comment":"The assertion that 'the size of a reduced model decreases monotonically with ε' is load-bearing for Algorithm 5's binary search, but no proof or reference is provided. Monotonicity is not immediate for an iterative greedy algorithm like Algorithm 4, because adding a row early can enlarge the row space and reduce later projection distances, potentially causing a smaller threshold to block rows that a larger threshold would add. Without a proof, the binary search may return a value of ε that does not correspond to the intended reduced size. Please provide a formal proof or a counterexample; if the property does not hold in general, the heuristic should be presented as a heuristic with no correctness guarantee and the experiments should be re-evaluated under that caveat.","section":"Theorem 3 statement"}],"minor_comments":[{"comment":"The sentence 'Here, C is the Lipschitz constant of f over the set of initial conditions S' is not well-formed, since f's argument is a state, not an initial condition. It should say 'the Lipschitz constant of f on Ω' (or on a suitable set containing the trajectories).","section":"Theorem 3 statement"},{"comment":"The notation J_i is used both for the coefficient matrices in the decomposition J(x)=Σ J_i μ_i(x) and for the spanning matrices returned by Algorithm 1 (which are sampled Jacobian evaluations). This makes statements like 'all rows s of L J_i' ambiguous. Please use distinct notation, e.g., B_i for the coefficient matrices and Â_i for the sampled basis.","section":"Proposition 3 and Theorem 5"},{"comment":"The text says 'we want to obtain the smallest ε whose induced size is below the cutoff size' while the earlier description says 'find the largest ε such that the reduced model's size m_ε satisfies m_ε ≤ m∗'. These are contradictory; please align the wording with the intended meaning (smallest ε that achieves the size target).","section":"Section 3.3, text before Algorithm 5"},{"comment":"Several references are incomplete or malformed in the extracted text, including '45? ?', '[16? ]', and '[ ? ]' in the related-work paragraph. Please ensure all citations are properly resolved.","section":"Section 1"},{"comment":"For Model 5, the rows with reduced sizes 9, 8, and 1 report eRel(T)=1.00E+00, i.e., 100% relative error. The text explains that aggressive reductions collapse the dynamics, but the table could explicitly mark these as failed reductions to avoid confusion with the low-error rows.","section":"Section 4.1, Table 2"}],"recommendation":"major_revision","confidential_remarks":"The paper presents a promising and practically relevant extension of constrained lumping, with a public implementation and a broad experimental study. The main blocker is the domain-assumption gap in the central error theorem and the clearly incorrect pseudocode of Algorithm 5. Both are fixable within the scope of the manuscript, but they affect the core claims, so I cannot recommend acceptance in the current form. The monotonicity assumption also needs to be addressed. If the authors can close these gaps, the paper would likely be suitable for publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"This is a useful, honest extension of exact constrained lumping to approximate lumping, with a proof that observable error is O(epsilon), support for rational drifts, a polynomial-time algorithm, and a heuristic for choosing epsilon. The core idea is incremental over the authors' CMSB 2023 paper, but the generalization is real. The central theorem, however, is missing a domain assumption: the proof of the error bound needs the reduced trajectory to stay in the compact set Omega, and the theorem only assumes the original trajectory and its projection do. That gap is fixable but should not be ignored.\n\nWhat is new and good: the paper extends the exact CLUE algorithm to allow a tolerance epsilon, proves a bound on the deviation, shows how to compute the approximate lumping in polynomial time, and provides a binary-search heuristic that finds an epsilon to meet a target reduced size. The experiments on eight published models (five polynomial, three rational) show coarser reductions than exact lumping with often small errors, and the scalability test on multisite phosphorylation is nice. The code and data are public. The paper is also honest that the worst-case bound is conservative and the tolerance search is heuristic.\n\nSoft spots: the stress-test concern is valid. In Theorem 3, the proof applies a Lipschitz estimate on f between \\bar{L}y(t) and \\bar{L}Lx(t), which requires both arguments to lie in Omega. The theorem only states x(t), \\bar{L}Lx(t) in Omega. A simple rational example can make the reduced solution blow up while the original stays safe; I checked the counterexample and it works. Section 4.1 only checks the condition post hoc. This is the main issue. Second, Algorithm 5 relies on the reduced size being monotone in epsilon; this is asserted but not proved, and it is not obvious because skipping a row early changes the rowspace for later checks. Third, the evaluation compares only against exact lumping, not other approximate reduction methods, which is a minor weakness.\n\nWho this is for: systems biology modelers and people working on model reduction algorithms. It deserves a serious referee. Recommendation: engage with it, but ask the authors to fix the theorem statement, or at least make the reduced-trajectory condition explicit and show how to verify it; and discuss the monotonicity assumption.","headline":"Solid incremental extension of exact CLUE to approximate lumping with an O(epsilon) error bound and a tolerance heuristic; the bound is credible, but Theorem 3 understates the domain condition for the reduced trajectory.","tokens_in":24915,"tokens_out":5189,"would_cite":true,"duration_ms":48392,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["34C41","92C42"],"pacs":[],"model":"deepseek-v4-flash","headline":"Relaxing exact lumping gives provably bounded approximations of ODE models.","keywords":["approximate lumping","model reduction","chemical reaction networks","ordinary differential equations","constrained lumping","error bounds","polynomial-time algorithm","biochemical models"],"falsifier":"Take a rational ODE model where the reduced trajectory crosses a pole of f within the simulation horizon, run Algorithm 4, simulate the reduced system, and compute max_t ||e(t)||; if the deviation dev_L(f, \\bar{L}y(t)) grows without bound and the error exceeds the Theorem 3 bound, the compactness assumption is violated and the O(ε) guarantee does not hold.","tokens_in":23862,"feed_emoji":"🧪","tokens_out":5101,"duration_ms":47574,"temperature":0.7,"pith_summary":"This paper introduces approximate constrained lumping, a model-reduction method for systems of ordinary differential equations with analytic right-hand sides, including polynomial and rational models common in biochemistry. Instead of requiring a lumping matrix to preserve user-chosen linear observables exactly, it allows a pointwise deviation bounded by a tolerance and proves that the resulting error between original and reduced outputs stays proportional to that tolerance. The reduction is computed in polynomial time by a row-space iteration that adds new rows only when the Jacobian action leaves the current lumping subspace by more than a numerical threshold. On eight published biochemical models, the method yields coarser aggregations than exact lumping while keeping simulated observables close, and a heuristic selects the largest feasible tolerance for a target reduced size. If the error bound holds in practice, modelers can trade reduction size against accuracy in a principled way instead of hoping an exact reduction exists.","feed_headline":"Approximate lumping cuts ODE models with error in O(ε)","feed_subtitle":"Relax exact lumping to a tolerance and provably keep observable error proportional to it — coarser, still faithful.","key_machinery":"The operative device is the deviation functional dev_L(f,x) = ||Lf(\\bar{L}Lx) - Lf(x)||_2, which compares the original dynamics to the dynamics evaluated after projecting onto the row space of L; exact lumping is the zero set of this functional. Algorithm 4 operationalizes it: with an orthonormal L (so \\bar{L} = L^T), it forms π_i = rJ_i L^T L and only grows L when the residual rJ_i - π_i exceeds ε, appending the normalized residual as a new row. This keeps L orthonormal, keeps the observable rows inside rowsp(L), and, via the spanning representation J(x) = Σ_i J_i μ_i(x), converts a symbolic invariance check into finitely many numerical checks. The proof engine is a Grönwall-type inequality that turns the pointwise deviation bound into the exponential-in-time error bound of Theorem 3.","core_discovery":"The central claim is that near-lumpability can be made quantitative. Given a full-rank matrix L, the deviation dev_L(f,x) = ||Lf(\\bar{L}Lx) - Lf(x)||_2 measures how far L is from being an exact lumping at state x; exact lumping is exactly the case where this deviation is identically zero. An approximate (S,T,η)-lumping requires the deviation along all trajectories of interest to stay below η, and Theorem 3 bounds the observable error e(t) = y(t) - Lx(t) by η times the factor (1/(C||L||_2||\\bar{L}||_2))($e^{{C||L||_2||\\bar{L}}$||_2 T} - 1), with C the Lipschitz constant of f on a compact set containing the trajectories. Algorithm 4 constructs such an L by orthonormal row addition: starting from the observable rows, it appends the normalized residual rJ_i - (rJ_i)L^T L whenever its norm exceeds a numerical tolerance ε, and Theorem 5 shows the resulting L has deviation η = O(√m C' C K ε). The paper argues this is a polynomial-time approximate constrained lumping with an error guarantee proportional to the tolerance, covering rational and polynomial models alike.","pith_inferences":["The continuous tolerance knob could be repurposed for parameter-sensitivity analysis: one could estimate how much uncertainty in kinetic parameters a reduced model can absorb before the O(ε) error guarantee degrades, a direction the paper mentions as future work.","For rational models, the compactness assumption is the real bottleneck; the authors only verify post hoc that denominators vanish only for negative states. A checkable sufficient condition, such as positivity preservation of the projected trajectory, would make the method usable without prior simulation.","The exponential-in-time factor in the bound means the O(ε) guarantee is most meaningful over finite horizons; extending to steady-state or long-time statements would likely require additional dissipativity or contraction assumptions.","The ε_max normalization from Lemma 1 gives a dimensionless difficulty index for reduction across models, though the paper does not develop that comparison explicitly."],"forward_implications":["Models with no exact lumping can still be reduced: on several benchmark models, approximate reductions were obtained where exact lumping was impossible or much coarser.","The reduced size can be chosen by the user: the binary-search heuristic finds the largest tolerance whose reduced model stays below a target size, and the error remains O(ε).","Rational models arising from Hill and Michaelis-Menten kinetics are covered, not just polynomial systems, extending the earlier polynomial-only version.","The method runs in polynomial time and scales to a 1032-variable multisite phosphorylation model, reducing it to a handful of species with small error.","Exact lumping is recovered in the limit ε → 0, so the method is a strict generalization that never does worse than the exact approach."],"supporting_citations":[{"why":"Characterizes exact lumping via rowspace invariance, which the deviation functional relaxes in Definition 4.","marker":"[48]"},{"why":"Provides the finite basis {J_i} of the Jacobian via sampling and automatic differentiation, used by Algorithm 4.","marker":"[24]"},{"why":"Presents the CLUE algorithm for exact constrained lumping in polynomial time, the baseline extended here.","marker":"[35]"},{"why":"Conference version of this work for polynomial ODEs, which the present paper extends to analytic and rational drifts.","marker":"[27]"},{"why":"Classic analysis of approximate lumping in chemical kinetics, motivating the relaxation of exactness.","marker":"[30]"},{"why":"Introduces constrained lumping with user-specified preserved observables, formalized in Definition 3.","marker":"[31]"},{"why":"Grönwall-type inequality used to convert the differential error inequality into the exponential error bound of Theorem 3.","marker":"[23, Lemma 2]"},{"why":"Mean value theorem used in Proposition 2 to bound differences of f through Jacobian norms.","marker":"[40, Theorem 9.19]"}],"fun_headline_variants":["Approx lumping: coarser ODEs, error O(ε) guaranteed","Near-lumping ODEs: error scales with ε, polynomial time","Shrink ODEs with approximate lumping, error O(ε)","Approx constrained lumping: coarser, error ~ ε","Lump ODEs approximately: error O(ε), polynomial time"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The error bound requires both the original trajectory and its projection onto the lumping subspace to stay inside a compact region where the dynamics are analytic and Lipschitz; for rational models this can fail when a denominator vanishes along the reduced trajectory, and the paper only checks this after the fact.","fun_headline_variants_meta":{"raw":{"variants":["Approx lumping: coarser ODEs, error O(ε) guaranteed","Near-lumping ODEs: error scales with ε, polynomial time","Shrink ODEs with approximate lumping, error O(ε)","Approx constrained lumping: coarser, error ~ ε","Lump ODEs approximately: error O(ε), polynomial time"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000853,"raw_usage":{"total_tokens":3740,"prompt_tokens":1011,"completion_tokens":2729,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":627,"completion_tokens_details":{"reasoning_tokens":2632}},"tokens_in":627,"tokens_out":2729,"duration_ms":19030,"temperature":1.0,"reasoning_tokens":2632,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T15:23:01.573020+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a rational ODE model where the reduced trajectory crosses a pole of f within the simulation horizon, run Algorithm 4, simulate the reduced system, and compute max_t ||e(t)||; if the deviation dev_L(f, \\bar{L}y(t)) grows without bound and the error exceeds the Theorem 3 bound, the compactness assumption is violated and the O(ε) guarantee does not hold.","supporting_citations":[{"cited_title":"Tomlin, Genyuan Li, Herschel Rabitz, and J´ anos T´ oth","cited_arxiv_id":null,"evidence_quote":"Characterizes exact lumping via rowspace invariance, which the deviation functional relaxes in Definition 4."},{"cited_title":"Exact linear reduction for rational dynamical systems","cited_arxiv_id":null,"evidence_quote":"Provides the finite basis {J_i} of the Jacobian via sampling and automatic differentiation, used by Algorithm 4."},{"cited_title":"CLUE: exact maximal reduction of kinetic models by constrained lumping of differential equations","cited_arxiv_id":null,"evidence_quote":"Presents the CLUE algorithm for exact constrained lumping in polynomial time, the baseline extended here."},{"cited_title":"Approximate Constrained Lumping of Polynomial Differential Equations","cited_arxiv_id":null,"evidence_quote":"Conference version of this work for polynomial ODEs, which the present paper extends to analytic and rational drifts."},{"cited_title":"A general analysis of approximate lumping in chemical kinetics","cited_arxiv_id":null,"evidence_quote":"Classic analysis of approximate lumping in chemical kinetics, motivating the relaxation of exactness."},{"cited_title":"New approaches to determination of constrained lumping schemes for a reaction system in the whole composition space","cited_arxiv_id":null,"evidence_quote":"Introduces constrained lumping with user-specified preserved observables, formalized in Definition 3."}],"review_version":1}