EFiGP: Eigen-Fourier Physics-Informed Gaussian Process for Inference of Dynamic Systems

Jianhong Chen; Shihao Yang

REVIEW 4 major objections 4 minor 22 references

EFiGP: Eigen-Fourier Physics-Informed Gaussian Process for Inference of Dynamic Systems

T0 review · 4 major / 4 minor · reviewed 2026-08-10 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read EFiGP moves the ODE constraint into the Fourier domain and truncates high frequencies and eigenmodes, making dense-grid Bayesian inference of parameters and trajectories faster and more accurate than MAGI.

desk verdict A credible, useful extension of MAGI with a genuinely new Fourier-domain conditioning trick, but the flat-runtime and accuracy claims rest on untested band-limited assumptions and unreported tuning details. read the letter →

arxiv 2501.14107 v1 pith:OOV4F5LN submitted 2025-01-23 stat.ML cs.LG

Jianhong Chen , Shihao Yang This is my paper · ORCID

classification stat.MLcs.LG MSC 62F1565L0965T5062M20

keywords physics-informedGaussianprocessODEinverseproblemFouriertransformeigen-decompositionBayesianinferenceMAGIcomparisonoscillatorysystemstrajectoryrecovery

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper proposes EFiGP, a Bayesian method for inferring parameters and trajectories of nonlinear ODE systems from sparse, noisy observations without ever running a numerical integrator during fitting. Its central claim is that enforcing the ODE constraint in the Fourier domain, while truncating both Fourier frequencies and GP eigencomponents, makes the posterior cheaper to evaluate and more accurate than MAGI as the discretization grid grows denser. On three benchmark oscillatory systems, EFiGP's runtime stays flat while MAGI's rises linearly, giving roughly a sixfold speedup at 1,281 grid points and better trajectory RMSE. The significance is that ODE inverse problems, which usually require repeated numerical integration, become practical at much higher resolution.

What carries the argument

The load-bearing object is the truncated Fourier-domain ODE residual $W^F_I$, together with the low-rank eigen-reparametrization of the GP. Lemma 2.2 guarantees the Fourier-transformed GP is Gaussian, turning the physics constraint into a tractable Gaussian quadratic form; Lemma 2.1 gives the spectral reparametrization that lets the algorithm discard small-eigenvalue directions. Together they replace an $n \times n$ covariance inversion with operations on $l$ frequencies and $j$ eigencomponents, which is why runtime stops scaling with the discretization size.

What would settle it

Run EFiGP on a stiff or sharply spiking system, such as the van der Pol oscillator with large stiffness or a relaxation oscillator whose true residual carries significant energy above the $l=41$ cutoff, on grids from 41 to 1,281 points; if parameter or trajectory RMSE degrades markedly once the residual is no longer band-limited, the central truncation assumption fails.

Watch

Extended reading notes

Core claim

EFiGP replaces the pointwise manifold constraint of MAGI with a Fourier-domain constraint: instead of forcing $\dot{X}(t) = f(X(t), \theta, t)$ at each discretization point, it forces agreement of the discrete Fourier coefficients of the GP derivative and of $f$, keeping only the first $l$ frequencies. Because the Fourier transform of a Gaussian process is again Gaussian, the physics condition remains a quadratic form in the posterior. The GP trajectory is then reparametrized through its eigen-decomposition, $X(I) = \mu(I) + V_{(j)}\Lambda_{(j)}^{1/2}z$, truncated to $j$ eigencomponents, so that dense grids no longer force dense covariance matrices. The authors claim that with $l=41$ and $j=81$ the posterior stabilizes, runtime stays flat as grids grow from 41 to 1,281 points, and trajectory error stays below MAGI's while MAGI's runtime grows linearly and, on the Hes1 system, MAGI fails to converge beyond 321 points.

Load-bearing premise

The ODE residual on the discretized grid is effectively band-limited, so truncating the Fourier constraint at $l=41$ removes only high-frequency noise and not information needed to identify $\theta$ or reconstruct the trajectory.

Editorial extensions

If this is right

ODE inverse problems can be solved without any numerical integration during fitting, removing the repeated-solver bottleneck that dominates classical least-squares calibration.
Runtime becomes essentially independent of discretization size, so practitioners can use dense grids to resolve fast transients without paying MAGI's linear cost.
Fourier truncation acts as a denoiser: high-frequency noise is discarded before it enters the physics constraint, which the benchmarks show as improved trajectory RMSE.
On the Hes1 system, MAGI fails to converge beyond 321 discretization points while EFiGP continues to produce stable results at 1,281, indicating that Fourier conditioning also stabilizes the optimization.
Parameter estimation remains limited by weak identifiability of the ODE parameters even when trajectory recovery is accurate, exactly as the authors observe in the Lotka-Volterra results.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

Because the speed-up depends on the residual being band-limited, EFiGP's flat runtime should transfer to other smooth or oscillatory inverse problems but not automatically to stiff systems; an adaptive cutoff $l$ would be a natural extension.
The Fourier-domain conditioning could be applied to other probabilistic differential-equation solvers, since differentiation becomes multiplication in frequency space there too.
The truncation parameters $j$ and $l$ set the effective cost, so a user could in principle tune grid resolution independently of computational budget, a decoupling that MAGI does not offer.
Reporting the Fourier residual spectrum after fitting would give a direct diagnostic of whether the band-limited assumption holds in a new application.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Quick take: this is a genuine algorithmic contribution. Enforcing the ODE constraint on truncated Fourier components within the physics-informed GP posterior is new relative to MAGI/PIGP, and the eigen-decomposition reparameterization is a sensible way to cut the cost of dense grids. The three benchmark experiments show EFiGP often beats MAGI in trajectory RMSE and runtime, with the runtime roughly flat from n=41 to n=1281. That's a practically useful result for oscillatory biological systems. The authors also ship code, which helps reproducibility.

The soft spots are real but not disqualifying. The biggest is the band-limited residual assumption: the posterior conditions on only l=41 Fourier modes, so increasing discretization beyond the point where those modes are resolved adds no new physics information. For smooth oscillators this is fine, but the paper never checks the spectral decay of the ODE residual and makes no claim about stiff or nonsmooth dynamics. That's the load-bearing gap between the benchmarks and the title's 'dynamic systems' promise; the stress-test note lands. The second issue is experimental reporting: the truncation orders l and j are chosen by increasing them on the same benchmark data until results stabilize, the GP kernel length-scale and variance are never reported, and the flat-runtime claim appears without a complexity analysis even though evaluating f at n points is O(n). These are exactly the details a referee would need to judge whether the method is robust or tuned to these examples. The parameter degradation at high discretization for Lotka–Volterra is acknowledged and plausibly attributed to weak identifiability, so I don't count that as a flaw.

The math itself looks coherent to me. The posterior in Eq. (12) is a standard conditioning argument, the change of variables is legitimate, and the paper correctly notes the Jacobian cancels. There's no circularity: parameters and trajectories are inferred from data and the prediction period is genuinely held out.

Bottom line: this deserves a serious referee, not a desk reject. I'd suggest the authors be asked to report kernel hyperparameters, describe how l and j are selected without peeking at test performance, and either add a spectral analysis of the residual or explicitly scope the method to smooth oscillatory systems. With those revisions the paper would be a solid contribution to the GP-for-ODE inference literature.

Referee Report

4 major / 4 minor

Summary. The paper proposes EFiGP, an extension of the MAGI framework for Bayesian inference of ODE parameters and trajectories. EFiGP enforces the ODE constraint in Fourier space, truncating high-frequency terms, and uses eigen-decomposition with truncation to reparameterize the Gaussian process. The method is evaluated on FitzHugh-Nagumo, Lotka-Volterra, and Hes1 systems, comparing accuracy and runtime against MAGI. The central claim is that EFiGP yields stable accuracy and flat runtime as discretization becomes dense, making it a scalable alternative to MAGI.

Significance. If substantiated, the central contribution is a practically useful reformulation that decouples computational cost from discretization density for oscillatory systems. Strengths of the paper include a transparent derivation of the posterior, available code, and an honest discussion of parameter identifiability limitations. However, the empirical claims rest on data-dependent choice of truncation orders and an untested band-limited residual assumption, and the runtime claim lacks a complexity analysis. These issues make the current evidence suggestive rather than conclusive.

major comments (4)

[Section 3, Eq. (10); Section 4] The constraint W^F_I = 0 is enforced only on the first l retained Fourier modes, and for all dense grids l is fixed at 41. The paper provides no estimate of the spectrum of the ODE residual r(t) = ẋ(t) - f(x(t), θ, t) for the benchmark systems, so there is no evidence that the residual energy above mode 41 is negligible. For systems with fast transients or stiffness, truncating the constraint at l = 41 will bias both θ and trajectory inference, and the claim of reliable inference for general dynamic systems is therefore not established. The authors should report the residual spectra for the benchmark systems or provide a consistency or error analysis in l.
[Section 4, EFiGP Setting; Tables 6-8] The truncation orders l and j are chosen per system and per discretization level by increasing them until results stabilize on the same datasets that are then used to report performance. This data-dependent selection makes the reported accuracy and runtime results optimistic and prevents a straightforward interpretation of the benchmark comparisons. The paper should either use a fixed rule for l and j, report sensitivity to these choices, or evaluate on separate tuning and validation data.
[Table 1; Section 4] The claim that EFiGP runtime no longer increases with discretization size is presented without a complexity analysis. Since the cost after eigen-decomposition and Fourier truncation should depend on l, j, and the optimization dimension rather than n, the authors should state the per-iteration and total complexity in these parameters and verify it empirically. Without this, the observed flat runtime on three systems does not support a general scalability claim.
[Section 4, EFiGP Setting] The Matérn kernel is specified with degree of freedom 2.01, but the length-scale and variance hyperparameters are not reported, nor is the procedure for setting or estimating them described. This is an essential reproducibility gap and raises the possibility that the accuracy gains over MAGI reflect favorable hyperparameter choices rather than the method itself.

minor comments (4)

[Section 2.4] The text states that the augmented Fourier matrix maps R^n to R^{2n}, but later says the resulting covariance matrix is of dimension (2n-1) x (2n-1); please clarify the correct dimensions and reconcile the notation.
[Eq. (12)] The quadratic norm subscript is typeset as (C^F_(l))^{-1} inside the exponent, which is difficult to read; please use brackets or a clear norm notation for readability.
[Abstract; Section 4] The abstract claims EFiGP 'eliminates the need for numerical integration,' but Section 4 states that numerical integration is used for evaluation and forecasting after fitting; please qualify the abstract to refer to the fitting procedure.
[Section 6.5, Tables 6-8] The captions for Tables 6-8 should explicitly state that the reported E and F values are tuned truncation numbers, not fixed a priori choices, and define E and F in the caption or nearby text.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: the EFiGP posterior conditions on data and a projected ODE constraint; truncation-order tuning is model selection, not a fitted prediction.

full rationale

The central Bayesian posterior (Eq. 12) is built from the observed-data likelihood, the GP prior z^T z, and a truncated Fourier ODE residual; none of these terms is defined in terms of the quantities being predicted (theta and the trajectory). The physics constraint in Eq. (10), F[dot X] = F[f(x, theta)], is a genuine projected ODE condition, not an identity that encodes the answer. Truncation orders l and j are selected by gradually increasing them until results stabilize on the benchmark systems (Section 4 and Tables 6-8), which is standard model selection rather than fitting the reported RMSE or runtime. Out-of-sample trajectory RMSE is evaluated on a held-out second half-period with no observations, so the predictions are not forced by the training fit. The only notable self-citation is the MAGI baseline [23], co-authored by the corresponding author, but MAGI is used as an empirical comparison method and as guidance for choosing discretization numbers, not as a premise that entails EFiGP's claims. No uniqueness theorem or ansatz is imported from the authors' prior work to rule out alternatives. The band-limited residual assumption is a stated approximation that could be wrong for stiff systems, but that is a correctness/robustness concern, not circularity. Thus no step in the derivation chain reduces to its own inputs by construction.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The central claim rests on four tuned or unstated numerical choices (l, j, kernel hyperparameters, noise level) and on a band-limitedness assumption for the ODE residual that is never tested. No new physical or mathematical entities are introduced.

free parameters (4)

Fourier truncation order l = 11, 21, or 41 depending on system and discretization; 41 for dense grids
Chosen by increasing l until results stabilize on each benchmark; controls how much of the ODE constraint is enforced and is central to denoising and speed.
Eigen truncation order j = 21, 41, or 81 depending on system and discretization; 81 for dense grids
Chosen by the same stabilization heuristic; controls the low-rank approximation of the GP prior and directly affects cost and accuracy.
GP kernel length-scale and variance = not reported
The Matérn kernel is fixed only by degrees of freedom 2.01; length-scale and variance enter the covariance K and conditional covariance C but are not given or described as estimated.
Observation noise level sigma = 0.2 (FN), 0.1 (Hes1 and LV, on log scale)
Treated as known in the experiments; the paper does not describe estimating sigma in real applications.

assumptions (5)

standard math The DFT plus real/imag augmentation of a Gaussian random vector is Gaussian (Lemma 2.2).
Used to derive the Fourier-domain physics constraint in Section 3.
domain assumption The discretized maximum event W_I=0 is a faithful surrogate for the continuous ODE constraint W=0.
MAGI already relies on this; EFiGP inherits it and applies it after truncation.
domain assumption The ODE residual is band-limited over [0,T], so truncating Fourier terms above l discards only high-frequency noise.
Introduced to justify denoising and the constant-cost claim; no spectral analysis of the residual is provided.
domain assumption The GP kernel is twice differentiable, so cross-derivative covariances are well-defined.
Matérn with df=2.01 is chosen to guarantee this; also needed for the conditional covariance C.
domain assumption Numerically integrated trajectories are accurate enough to serve as ground truth and evaluation.
All benchmark data and all forecasting evaluations use numerical integration; no error control is discussed.

how reviews work

0 comments

Cite this review

Pith. "Pith review of EFiGP: Eigen-Fourier Physics-Informed Gaussian Process for Inference of Dynamic Systems." pith.science (2026). https://pith.science/paper/OOV4F5LN

@misc{pith2026250114107,
  author       = {Pith},
  title        = {Pith review of: EFiGP: Eigen-Fourier Physics-Informed Gaussian Process for Inference of Dynamic Systems},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/OOV4F5LN}},
  note         = {Machine review of arXiv:2501.14107}
}

read the original abstract

Parameter estimation and trajectory reconstruction for data-driven dynamical systems governed by ordinary differential equations (ODEs) are essential tasks in fields such as biology, engineering, and physics. These inverse problems -- estimating ODE parameters from observational data -- are particularly challenging when the data are noisy, sparse, and the dynamics are nonlinear. We propose the Eigen-Fourier Physics-Informed Gaussian Process (EFiGP), an algorithm that integrates Fourier transformation and eigen-decomposition into a physics-informed Gaussian Process framework. This approach eliminates the need for numerical integration, significantly enhancing computational efficiency and accuracy. Built on a principled Bayesian framework, EFiGP incorporates the ODE system through probabilistic conditioning, enforcing governing equations in the Fourier domain while truncating high-frequency terms to achieve denoising and computational savings. The use of eigen-decomposition further simplifies Gaussian Process covariance operations, enabling efficient recovery of trajectories and parameters even in dense-grid settings. We validate the practical effectiveness of EFiGP on three benchmark examples, demonstrating its potential for reliable and interpretable modeling of complex dynamical systems while addressing key challenges in trajectory recovery and computational cost.

Figures

Figures reproduced from arXiv: 2501.14107 by the authors.

**Figure 2.** Predicted trajectory from EFiGP (red solid, dashed and dotted line) and from MAGI [PITH_FULL_IMAGE:figures/full_fig_p010_2.png] view at source ↗

**Figure 3.** Predicted trajectory from EFiGP (red solid and dashed line) and from MAGI (blue solid [PITH_FULL_IMAGE:figures/full_fig_p012_3.png] view at source ↗

Figures from the paper (1 more)

**Figure 4.** Figure 4: Single-dataset comparison between the predicted trajectory (red solid and dashed lines) [PITH_FULL_IMAGE:figures/full_fig_p018_4.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

22 extracted references · 16 canonical work pages

[1]

Y. Bard. Nonlinear parameter estimation. SIAM Review , 17(4):703–704, Oct 1975. ISSN 0036-1445. doi: 10.1137/1017088. URL https://doi.org/10.1137/1017088

work page doi:10.1137/1017088 1975
[2]

M. Benson. Parameter fitting in dynamic models. Ecological Modelling, 6(2):97–115, 1979

work page 1979
[3]

Accelerating bayesian inference over nonlinear differential equations with gaussian processes

Ben Calderhead, Mark Girolami, and Neil D Lawrence. Accelerating bayesian inference over nonlinear differential equations with gaussian processes. In Advances in neural information processing systems, pages 217–224. Citeseer, 2009

work page 2009
[4]

Ode parameter inference using adaptive gradient matching with gaussian processes

Frank Dondelinger, Dirk Husmeier, Simon Rogers, and Maurizio Filippone. Ode parameter inference using adaptive gradient matching with gaussian processes. In Artificial Intelligence and Statistics , pages 216–228. PMLR, 2013

work page 2013
[5]

FitzHugh

R. FitzHugh. Impulses and physiological states in theoretical models of nerve membrane. Biophysical journal, 1(6):445–466, 1961

work page 1961
[7]

Osborne, and Mark Girolami

Philipp Hennig, Michael A. Osborne, and Mark Girolami. Probabilistic numerics and un- certainty in computations. Proceedings of the Royal Society A: Mathematical, Physical and Engineering Sciences, 471(2179):20150142, 2015. 13

work page 2015
[8]

Hirata, S

H. Hirata, S. Yoshiura, T. Ohtsuka, Y. Bessho, T. Harada, K. Yoshikawa, and R. Kageyama. Oscillatory expression of the bhlh factor hes1 regulated by a negative feedback loop. Science, 298(5594):840–843, 2002

work page 2002
[9]

Bayesian calibration of computer models

Marc C Kennedy and Anthony O’Hagan. Bayesian calibration of computer models. Journal of the Royal Statistical Society: Series B (Statistical Methodology) , 63(3):425–464, 2001

2001

Show all 22 references

[10]

Lapidus and J

L. Lapidus and J. H. Seinfeld. Numerical solution of ordinary differential equations . Academic Press, 1971

1971
[11]

Parameter inference based on gaussian processes informed by nonlinear partial differential equations.SIAM/ASA Journal on Uncertainty Quan- tification, 12(3):964–1004, 2024

Zhaohui Li, Shihao Yang, and CF Jeff Wu. Parameter inference based on gaussian processes informed by nonlinear partial differential equations.SIAM/ASA Journal on Uncertainty Quan- tification, 12(3):964–1004, 2024

2024
[12]

Fourier neural operator for parametric partial differential equations

Zongyi Li, Nikola Kovachki, Kamyar Azizzadenesheli, Burigede Liu, Kaushik Bhattacharya, Andrew Stuart, and Anima Anandkumar. Fourier neural operator for parametric partial differential equations. In International Conference on Learning Representations , 2021. URL https://openr...

2021
[13]

A. J. Lotka. The growth of mixed populations: Two species competing for a common food supply. Journal of the Washington Academy of Sciences , 22(16/17):461–469, 1932. ISSN 0043-0439. URL http://www.jstor.org/stable/24530449

1932
[14]

Nagumo, S

J. Nagumo, S. Arimoto, and S. Yoshizawa. An active pulse transmission line simulating nerve axon. Proceedings of the IRE, 50(10):2061–2070, 1962

1962
[15]

Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations

Maziar Raissi, Paris Perdikaris, and George E Karniadakis. Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations. Journal of Computational Physics , 378:686–707, 2019

2019
[16]

Parameter estimation for differential equations: a generalized smoothing approach

James O Ramsay, Giles Hooker, David Campbell, and Jiguo Cao. Parameter estimation for differential equations: a generalized smoothing approach. Journal of the Royal Statistical Society: Series B (Statistical Methodology) , 69(5):741–796, 2007

2007
[17]

Carl Edward Rasmussen and Christopher K. I. Williams. Gaussian Processes for Machine Learning. MIT Press, 2006. URL https://gaussianprocess.org/gpml/chapters/RW.pdf

2006
[18]

Patrick Seifner, Kostadin Cvejoski, and Ramses J. Sanchez. Foundational inference models for dynamical systems. arXiv preprint arXiv:2402.07594 , 2024. URL https://arxiv.org/abs/ 2402.07594

2024 arXiv
[19]

Understanding and mitigating gradient flow pathologies in physics-informed neural networks

Sifan Wang, Yujun Teng, and Paris Perdikaris. Understanding and mitigating gradient flow pathologies in physics-informed neural networks. SIAM Journal on Scientific Comput- ing, 43(5):3055–3081, 2021. doi: 10.1137/20M1318043. URL https://doi.org/10.1137/ 20M1318043

2021 doi
[20]

Gaussian processes for bayesian estimation in ordinary differential equations

Zhenwen Wang and David Barber. Gaussian processes for bayesian estimation in ordinary differential equations. In International Conference on Machine Learning , pages 1485–1493. PMLR, 2014. 14

2014
[21]

Gorbach, Andreas Krause, and Joachim M

Philippe Wenk, Athanasios Gotovos, Stefan Bauer, Nils S. Gorbach, Andreas Krause, and Joachim M. Buhmann. Fast gaussian process based gradient matching for parameter identi- fication in systems of nonlinear odes. In Kamalika Chaudhuri and Masashi Sugiyama, edi- tors, Proceedin...

2019
[22]

Osborne, Bernhard Sch¨ olkopf, Andreas Krause, and Stefan Bauer

Philippe Wenk, Giacomo Abbati, Michael A. Osborne, Bernhard Sch¨ olkopf, Andreas Krause, and Stefan Bauer. Odin: Ode-informed regression for parameter and state inference in time- continuous dynamical systems. Proceedings of the AAAI Conference on Artificial Intelligence , 34(...

2020 doi
[23]

Inference of dynamic systems from noisy and sparse data via manifold-constrained gaussian processes

Shihao Yang, Sophia W Wong, and SC Kou. Inference of dynamic systems from noisy and sparse data via manifold-constrained gaussian processes. Proceedings of the National Academy of Sciences, 118(15), 2021. 6 supplementary materials 6.1 GP Smoothing Gaussian Process (GP) smoothi...

2021

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.

[1] [1]

Y. Bard. Nonlinear parameter estimation. SIAM Review , 17(4):703–704, Oct 1975. ISSN 0036-1445. doi: 10.1137/1017088. URL https://doi.org/10.1137/1017088

work page doi:10.1137/1017088 1975

[2] [2]

M. Benson. Parameter fitting in dynamic models. Ecological Modelling, 6(2):97–115, 1979

work page 1979

[3] [3]

Accelerating bayesian inference over nonlinear differential equations with gaussian processes

Ben Calderhead, Mark Girolami, and Neil D Lawrence. Accelerating bayesian inference over nonlinear differential equations with gaussian processes. In Advances in neural information processing systems, pages 217–224. Citeseer, 2009

work page 2009

[4] [4]

Ode parameter inference using adaptive gradient matching with gaussian processes

Frank Dondelinger, Dirk Husmeier, Simon Rogers, and Maurizio Filippone. Ode parameter inference using adaptive gradient matching with gaussian processes. In Artificial Intelligence and Statistics , pages 216–228. PMLR, 2013

work page 2013

[5] [5]

FitzHugh

R. FitzHugh. Impulses and physiological states in theoretical models of nerve membrane. Biophysical journal, 1(6):445–466, 1961

work page 1961

[6] [7]

Osborne, and Mark Girolami

Philipp Hennig, Michael A. Osborne, and Mark Girolami. Probabilistic numerics and un- certainty in computations. Proceedings of the Royal Society A: Mathematical, Physical and Engineering Sciences, 471(2179):20150142, 2015. 13

work page 2015

[7] [8]

Hirata, S

H. Hirata, S. Yoshiura, T. Ohtsuka, Y. Bessho, T. Harada, K. Yoshikawa, and R. Kageyama. Oscillatory expression of the bhlh factor hes1 regulated by a negative feedback loop. Science, 298(5594):840–843, 2002

work page 2002

[8] [9]

Bayesian calibration of computer models

Marc C Kennedy and Anthony O’Hagan. Bayesian calibration of computer models. Journal of the Royal Statistical Society: Series B (Statistical Methodology) , 63(3):425–464, 2001

2001