REVIEW 3 major objections 4 minor 8 references
Parameter Estimation in ODE Models with Certified Polynomial System Solving
T0 review · 3 major / 4 minor · reviewed 2026-08-16 · deepseek-v4-flash
Pith's one-line read A certified algebraic solver cracks parameter-estimation systems that homotopy continuation cannot finish.
desk verdict A genuinely useful benchmark with reproducible code, but the 'vice versa' claim in the abstract is not supported by the reported data; the paper deserves referee time but needs a tempered conclusion. read the letter →
The pith
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
The reading
What carries the argument
The load-bearing object is the rational univariate representation (RUR): a certified description of all solutions of a zero-dimensional polynomial system as rational functions of a single variable, produced from a lexicographic Groebner basis and combined with real root isolation. Because RUR certifies the solution set rather than tracking numerical paths, it is not disturbed by ill-conditioned or non-identifiable coordinate blow-ups; this is what lets it succeed where homotopy continuation times out or returns no solutions. The comparison baseline is homotopy continuation, which tracks solution paths from a generic start system to the target system; its speed depends on the number of paths, which scales with the Bezout bound, so systems with 80 actual solutions but astronomically large Bezout bounds remain slow or impractical. RUR instead pays for the actual solution count, which the paper observes is typically small for parameter-estimation systems.
What would settle it
Take a second biologically plausible parameter set for the Akt-2 model and run both solvers under the same one-day limit: if the continuation solver finishes before the RUR solver, the claim that Akt-2 is out of reach for homotopy continuation is not universal.
Extended reading notes
Core claim
The central claim is that the new RUR solver can tackle examples that are out of reach for homotopy continuation methods, and vice versa. Concretely, the Akt-2 parameter estimation system has 69 unknowns, a Bezout bound of $2^{26}\cdot 33^7$, but only 80 actual solutions; the RUR-based solver found them in about an hour, while the continuation solver did not finish within a day. The paper also reports cases where continuation returns no solutions on structurally non-identifiable models, while RUR finds them. The opposite direction is demonstrated by Crauste-2 and Crauste-3, artificially constructed by squaring parameters to introduce symmetries and increase the solution count from 32 to 128 and 512; there, homotopy continuation is faster and more accurate than RUR. The paper's conclusion is that a certified algebraic backend can be competitive and sometimes more robust on parameter estimation problems with models up to roughly ten states and ten parameters.
Load-bearing premise
The benchmark treats one tested parameter set per model as representative of solver behavior, so the reported winners and losers may not generalize to other parameter values or data sets.
Editorial extensions
If this is right
- For parameter-estimation systems with few actual solutions, certified RUR solving can replace long or unsuccessful numerical continuation runs with a one-hour computation and a guaranteed answer.
- Structurally non-identifiable states or parameters, which often destabilize numerical path tracking, can be handled by the certified solver where continuation returns no solutions.
- On systems with many solutions created by symmetries, homotopy continuation remains the faster and more accurate backend, so the practical solver choice depends on structural features of the model.
- The observed gap between Bezout bounds and actual solution counts makes algebraic certification feasible for a wider range of biological models than previously assumed.
Reading between the lines
- Editorial inference: if the small-actual-solution-count pattern holds broadly, a cheap structural pre-analysis could predict which backend will win, allowing a hybrid pipeline to be scheduled automatically.
- Editorial inference: the RUR solver's slower, less accurate runs on Crauste-2 and Crauste-3 suggest that a hybrid strategy—certify with RUR, then refine with homotopy—could combine reliability with speed.
- Editorial inference: because only one parameter set per model was tested, backend rankings may shift on other data; a natural extension is a parameter sweep measuring how often each solver wins.
Signed reviews
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. The manuscript compares a new certified polynomial system solver based on rational univariate representation (RUR) and real root isolation with HomotopyContinuation.jl (HC) on parameter estimation problems in rational ODE models. The authors use a differential-algebra construction that turns the estimation task into a polynomial system, and benchmark both solvers on a set of models from the literature. Table 1 reports running times and parameter estimation errors. The main claim is that the RUR solver can solve systems that homotopy continuation cannot, and vice versa, with the RUR solver also being more reliable on some structurally non-identifiable models.
Significance. If the reported performance differences persist across a broader set of instances, the paper would provide practitioners with a useful empirical comparison of a certified algebraic solver against a widely used numerical continuation solver. The strengths include the use of an independently developed external solver (HC) as a baseline, a reproducible repository, and the certified nature of the RUR solver, which avoids numerical non-convergence. However, the evidence base is narrow: one parameter set per model, no error bars or repeated runs, and the only examples favoring HC are artificially constructed. The significance is therefore conditional on the representativeness of the benchmark.
major comments (3)
- [Section 2 (Main Result), Table 1 and bullet list] The abstract's claim 'and vice versa' — that homotopy methods can tackle examples out of reach for the RUR solver — is not supported by Table 1. In the rows where HC Julia is faster and more accurate (Crauste-2 and Crauste-3), the RUR solver still solves the systems; the paper itself labels these as 'More challenging for RUR,' not 'out of reach.' No row shows HC succeeding within the stated limits while RUR fails. The authors should either provide a genuine example where RUR does not terminate or fails, or revise the abstract and the concluding sentence of the introduction to state the actual finding: RUR extends the class of solvable instances on some problems, while on other problems HC is more efficient but RUR still succeeds.
- [Section 2 (Main Result), paragraph 'We tested one set of parameters per model'] The generalization of the benchmark conclusions rests on an untested assumption that 'the solvers behave similarly for different numerical values.' Since the polynomial systems are constructed from specific data, the number of solutions and the conditioning can vary substantially with the parameter values and observations. With a single parameter set per model, the 'out of reach' classification for Akt-2 and the relative performance ordering may be specific to the tested instances. The authors should test multiple parameter sets per model (or at least provide a sensitivity analysis) or explicitly restrict the claims to the tested data.
- [Section 2 (Main Result), bullet 'More challenging for RUR'] The only evidence for the HC-favorable direction is based on Crauste-2 and Crauste-3, which the authors acknowledge are artificially constructed by squaring parameters in Crauste-1 because they 'had difficulties to illustrate that with our small set of examples.' This is a significant weakness: the comparison does not demonstrate that the HC advantage occurs in natural parameter estimation problems. If the 'vice versa' claim is retained, the authors should find real ODE models where RUR is out of reach, or clearly state that the observed advantage is limited to a constructed class of systems.
minor comments (4)
- [Section 2, Table 1] The table caption defines n/a, OOM, and timeout, but the body would be clearer if each n/a entry were accompanied by a reason (e.g., convergence failure, numerical blow-up) rather than leaving the reader to infer it from the text.
- [Section 2, Table 1] The 'relative error' column is not defined: it is unclear whether it is the maximum relative error over all estimated parameters, or the relative residual of the polynomial system. Please define the error metric.
- [Section 1, Introduction] The sentence 'We have discovered that such polynomial systems typically have very few solutions' is supported only by the toy example and the observed solution counts in Table 2. As stated, it is a heuristic claim; the paper should either provide more systematic evidence or temper the wording.
- [Section 2, Experimental setup] The comparison would be more reproducible if the exact version of HomotopyContinuation.jl and the chosen tolerance/precision settings were reported, since continuation solvers are sensitive to these parameters.
Circularity Check
No significant circularity: the performance comparison is an empirical benchmark against an external solver, not a derivation that reduces to its own inputs.
full rationale
The paper's central comparison is between the authors' RUR-based solver and HomotopyContinuation.jl, an independently developed external package. The reported running times and errors in Table 1 are measured outcomes, not quantities fitted from the models and then relabeled as predictions. The polynomial systems solved in the benchmarks are generated by the prior differential-algebra method of Bassik et al. (2023); they are not defined in terms of the solver outputs, so there is no self-definitional reduction. The only self-citation of consequence is Demin et al. (2024), which supplies the RUR implementation; this is a code dependency, and the benchmark's evidence base includes an external solver and a public repository for reproduction, so the self-citation is not load-bearing in a circular sense. The stated limitation that only one parameter set per model was tested affects generalizability, not circularity. Likewise, the abstract's 'vice versa' claim is not demonstrated by Table 1, since RUR solves every example that HC Julia solves, but this is an evidentiary weakness rather than a circular step: no prediction is identical by construction to an input or to a fitted parameter. Consequently the circularity score is 0.
Assumptions & free parameters
assumptions (3)
- domain assumption The differential-algebra pipeline of Bassik et al. (2023) correctly converts parameter estimation into a polynomial system whose solutions are the parameters.
- ad hoc to paper Solver behavior is expected to be similar across different numerical parameter values.
- domain assumption The RUR and real root isolation algorithms in RationalUnivariateRepresentation.jl and RS.jl are implemented correctly.
Cite this review
Pith. "Pith review of Parameter Estimation in ODE Models with Certified Polynomial System Solving." pith.science (2026). https://pith.science/paper/26AM4PWS
@misc{pith2026250417268,
author = {Pith},
title = {Pith review of: Parameter Estimation in ODE Models with Certified Polynomial System Solving},
year = {2026},
howpublished = {\url{https://pith.science/paper/26AM4PWS}},
note = {Machine review of arXiv:2504.17268}
}
read the original abstract
We consider dynamical models given by rational ODE systems. Parameter estimation is an important and challenging task of recovering parameter values from observed data. Recently, a method based on differential algebra and rational interpolation was proposed to express parameter estimation in terms of polynomial system solving. Typically, polynomial system solving is a bottleneck, hence the choice of the polynomial solver is crucial. In this contribution, we compare two polynomial system solvers applied to parameter estimation: homotopy continuation solver from HomotopyContinuation.jl and our new implementation of a certified solver based on rational univariate representation (RUR) and real root isolation. We show how the new RUR solver can tackle examples that are out of reach for the homotopy methods and vice versa.
Reference graph
Works this paper leans on
-
[1]
, " * write output.state after.block = add.period write newline
ENTRY address author booktitle chapter doi edition editor eid howpublished institution journal key month note number organization pages publisher school series title type url volume year label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block FUNCTION init.state.consts #0 'before.all := #1 'mid.sent...
-
[2]
write newline
" write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION word.in bbl.in capitalize " " * FUNCT...
-
[3]
Balsa-Canto, E., Henriques, D., Gábor, A., and Banga, J.R. (2016). AMIGO2, a toolbox for dynamic modeling, optimization and control in systems biology . Bioinformatics, 32(21), 3357--3359
work page 2016
-
[4]
Barreiro, X.R. and Villaverde, A.F. (2023). On the origins and rarity of locally but not globally identifiable parameters in biological modeling. IEEE Access, 11, 65457--65467
work page 2023
-
[5]
Bassik, O., Berman, Y., Go, S., Hong, H., Ilmer, I., Ovchinnikov, A., Rackauckas, C., Soto, P., and Yap, C. (2023). Robust P arameter E stimation for R ational O rdinary D ifferential E quations
work page 2023
-
[6]
Breiding, P. and Timme, S. (2018). H omotopy C ontinuation.jl: A P ackage for H omotopy C ontinuation in J ulia. In International Congress on Mathematical Software, 458--465. Springer
work page 2018
-
[7]
Crauste, F., Mafille, J., Boucinha, L., Djebali, S., Gandrillon, O., Marvel, J., and Arpin, C. (2017). Identification of N ascent M emory CD8 T C ells and M odeling of T heir O ntogeny. Cell Systems, 4(3), 306--317.e4
work page 2017
-
[8]
Demin, A., Rouillier, F., and Ruiz, J. (2024). Reading R ational U nivariate R epresentations on lexicographic G roebner bases. ://arxiv.org/abs/2402.07141
arXiv 2024
Reviewed August 16, 2026 · model on record in the stance chip above.
Discussion (0). Continue with ORCID to comment.