pith. sign in

arxiv: 2604.16645 · v2 · submitted 2026-04-17 · 📊 stat.ME · math.ST· stat.TH

Strang splitting estimator for nonlinear multivariate stochastic differential equations with Pearson-type multiplicative noise

Predrag Pilipovi\'c , Adeline Samson , Susanne Ditlevsen This is my paper

Pith reviewed 2026-05-10 07:19 UTC · model grok-4.3

classification 📊 stat.ME math.STstat.TH
keywords Strang splittingparameter estimationPearson diffusionstochastic differential equationsmultiplicative noiseconsistencyasymptotic efficiencyStudent Kramers oscillator
0
0 comments X

The pith

Strang splitting estimator achieves consistency and asymptotic efficiency for nonlinear Pearson-type diffusions

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

The paper develops a parameter estimation method for nonlinear multivariate stochastic differential equations whose multiplicative noise follows a Pearson-type structure. Strang splitting separates each trajectory into a deterministic nonlinear ordinary differential equation and a linear-drift Pearson diffusion, whose exact mean and covariance are obtained via matrix exponential integrals. The estimator composes the two flows and approximates the transition density by a Gaussian whose moments match those of the Pearson component exactly. The authors prove that the resulting estimator is consistent and asymptotically efficient. A reader would care because these models arise in population genetics, climate records, and other fields where standard discretizations lose accuracy.

Core claim

We derive closed-form expressions for the mean and covariance matrix of multivariate Pearson diffusions using matrix exponential integrals and extend the framework to nonlinear diffusions with Pearson-type multiplicative noise. The main contribution is a parameter estimator based on Strang splitting that decomposes the system into a deterministic nonlinear ODE and a multivariate Pearson diffusion, composes their flows, and applies a Gaussian transition approximation with exact moments from the Pearson component. We prove that the estimator is consistent and asymptotically efficient. We also introduce the Student Kramers oscillator, prove existence and uniqueness of the strong solution and of

What carries the argument

Strang splitting decomposition of the SDE into a deterministic nonlinear ODE flow and a multivariate Pearson diffusion flow, followed by composition and a Gaussian transition whose moments are taken exactly from the Pearson component

If this is right

  • The estimator outperforms Euler-Maruyama, Gaussian approximation, and local linearization methods in finite-sample simulations for the Student Kramers oscillator and the multivariate Wright-Fisher diffusion.
  • Closed-form mean and covariance formulas are available for any multivariate Pearson diffusion.
  • The Student Kramers oscillator possesses a unique strong solution and a unique invariant measure.
  • The estimator can be applied directly to irregularly sampled real data such as Greenland ice-core records.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same splitting-plus-exact-moment idea may extend to other multiplicative-noise classes where only part of the dynamics admits closed moments.
  • High-dimensional population-genetics models beyond Wright-Fisher could become tractable for likelihood-based inference.
  • Finite-sample robustness could be checked by applying the estimator to additional benchmark diffusions with known ground-truth parameters.

Load-bearing premise

The Strang splitting produces a Gaussian transition approximation whose moments stay accurate enough for the consistency and asymptotic efficiency proofs to carry through when the drift is nonlinear.

What would settle it

In a Monte Carlo study of the Student Kramers oscillator, the estimator's bias fails to vanish and its variance fails to attain the Cramér-Rao bound as the time step tends to zero.

Figures

Figures reproduced from arXiv: 2604.16645 by Adeline Samson, Predrag Pilipovi\'c, Susanne Ditlevsen.

Figure 1
Figure 1. Figure 1: A trajectory of the Wright-Fisher diffusion. [PITH_FULL_IMAGE:figures/full_fig_p006_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: A trajectory of Student Kramers oscillator. [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Wright-Fisher diffusion. Violin plots of parameter estimation errors θˆN − θ0 based on 1000 simulated datasets (T = 20, N = 100, h = 0.2). Colors indicate estimator. Inside each violin plot, black dashed lines show medians and the 25th and 75th percentiles. Black violin plots represent the asymptotic distributions obtained with the central limit theorem (CLT). The LL estimator returned NaN for each dataset… view at source ↗
Figure 4
Figure 4. Figure 4: Student Kramers oscillator. Normalized distributions of parameter estimation errors (θˆN − θ0) ⊘ θ0 based on 1000 simulated datasets with a fixed time interval of length T = 50. Different colors indicate the type of estimator. Each column corresponds to a different parameter, and each row corresponds to a different value of h, and consequently N. Black density lines represent the asymptotic distributions. … view at source ↗
Figure 5
Figure 5. Figure 5: Median wall-clock estimation time (in seconds) to estimate parameters in the Student Kramers oscillator for step sizes h ∈ {0.01, 0.02}, based on 1000 simulated datasets (T = 50, N ∈ {5000, 2500}). Outliers and failed runs are excluded. Colors indicate estimator. 4.2.1 Strang splitting estimator To define the SS estimator, we rewrite (16) as d  Xt Vt  =  Vt −ηVt + aX3 t + bX2 t + cXt + d  | {z } F(Xt,V… view at source ↗
Figure 6
Figure 6. Figure 6: Ice core data from Greenland. Left: Trajectories (black) over time (in 1000 years) of the centered negative logarithm of the Ca2+ measurements (top) and forward difference approximations of its rate of change (bottom). Simulated trajectory from the estimated model (dark red). Right: Histograms of marginal distributions (gray) and estimated approximated invariant densities (black line) with prediction inter… view at source ↗
Figure 7
Figure 7. Figure 7: Violin plots of parameter estimation errors θˆN − θ0 based on 1000 simulated datasets (T = 20, N = 1000, h = 0.02). Colors indicate estimator. Inside each violin plot, black dashed lines show medians and the 25th and 75th percentiles. Black violin plots represent the asymptotic distributions. 29 [PITH_FULL_IMAGE:figures/full_fig_p029_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Median wall-clock estimation time (in seconds) across step sizes h ∈ {0.02, 0.2}, based on 1000 simulated datasets (T = 20, N ∈ {1000, 100}). Outliers and failed runs are excluded. Colors indicate estimator. S3 Supplementary Material: Results from the Multivariate Wright-Fisher diffusion For the finer time step h = 0.02 ( [PITH_FULL_IMAGE:figures/full_fig_p030_8.png] view at source ↗
read the original abstract

Multivariate Pearson diffusions are characterized by a linear drift and a diffusion matrix that is quadratic in the state variables. We derive closed-form expressions for the mean and covariance matrix of this class using matrix exponential integrals, and extend this framework to a broader class of nonlinear diffusions with Pearson-type multiplicative noise. The main contribution is a new parameter estimator for these nonlinear multiplicative models based on Strang splitting, which decomposes the stochastic system into a deterministic nonlinear ordinary differential equation and a multivariate Pearson diffusion. The estimator is constructed by composing their respective flows and applying a Gaussian transition approximation with exact moments from the Pearson component. We prove that the estimator is consistent and asymptotically efficient. We also introduce a new model within this class, the Student Kramers oscillator, and prove existence and uniqueness of the strong solution and of an invariant measure. We evaluate the estimator through simulation studies on this oscillator and on the multivariate Wright-Fisher diffusion from population genetics, where it outperforms the Euler-Maruyama, Gaussian approximation, and local linearization estimators. We conclude with an application to Greenland ice core data using the Student Kramers oscillator.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

1 major / 3 minor

Summary. The paper derives closed-form mean and covariance expressions for multivariate Pearson diffusions via matrix exponential integrals, extends the framework to nonlinear SDEs with Pearson-type multiplicative noise, and introduces a Strang splitting estimator that composes the flow of a deterministic nonlinear ODE with a Gaussian transition approximation using exact Pearson moments. It proves consistency and asymptotic efficiency of the estimator, introduces the Student Kramers oscillator with proofs of strong solution existence/uniqueness and invariant measure existence, demonstrates via simulations that the estimator outperforms Euler-Maruyama, Gaussian approximation, and local linearization methods on the oscillator and multivariate Wright-Fisher diffusion, and applies the model to Greenland ice core data.

Significance. If the consistency and efficiency results hold, the work would provide a computationally attractive estimation method for a practically relevant class of nonlinear diffusions with multiplicative noise, supported by closed-form moment calculations and the splitting construction. The introduction of the Student Kramers oscillator together with its well-posedness results is a useful addition, and the simulation comparisons plus real-data example strengthen the practical case. The explicit use of exact Pearson moments within the approximation is a clear technical strength.

major comments (1)
  1. [Proof of asymptotic efficiency] Proof of asymptotic efficiency (central claim in the Strang splitting estimator section): The argument that the composed flow (nonlinear ODE + Pearson diffusion) approximated by a Gaussian with exact Pearson moments yields an asymptotically efficient estimator extends the linear Pearson case. However, the nonlinear deterministic flow can amplify local approximation errors in a state-dependent manner, potentially affecting the score or information matrix at a rate that prevents efficiency. Explicit bounds on the transition approximation error (e.g., total variation or KL divergence between true and approximate transitions, or on the difference in estimating functions) are required to close the proof; their absence makes the efficiency claim rest on an unverified extension to the nonlinear setting.
minor comments (3)
  1. [Abstract] The abstract states that closed-form moments are derived using 'matrix exponential integrals' but does not indicate the precise integral form or the matrix exponential expression used; adding one sentence or an equation reference would improve readability.
  2. [Simulation studies] In the simulation studies, the number of Monte Carlo replications and the precise definition of the reported standard errors (e.g., whether they are Monte Carlo standard deviations or asymptotic) should be stated explicitly in the table captions or text for reproducibility.
  3. [Model definition] Notation for the diffusion matrix in the Pearson class (quadratic in state variables) could be clarified with an explicit component-wise definition early in the model section to avoid ambiguity when extending to the nonlinear case.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for the positive overall assessment of our manuscript and for the constructive major comment. We address it point by point below.

read point-by-point responses

  1. Referee: Proof of asymptotic efficiency (central claim in the Strang splitting estimator section): The argument that the composed flow (nonlinear ODE + Pearson diffusion) approximated by a Gaussian with exact Pearson moments yields an asymptotically efficient estimator extends the linear Pearson case. However, the nonlinear deterministic flow can amplify local approximation errors in a state-dependent manner, potentially affecting the score or information matrix at a rate that prevents efficiency. Explicit bounds on the transition approximation error (e.g., total variation or KL divergence between true and approximate transitions, or on the difference in estimating functions) are required to close the proof; their absence makes the efficiency claim rest on an unverified extension to the nonlinear setting.

    Authors: We agree that the current proof sketch for asymptotic efficiency in the nonlinear setting would be strengthened by explicit quantitative bounds on the approximation error. The manuscript currently relies on the exact moment matching of the Pearson component together with the Strang splitting structure to argue that the composed estimating function converges to the true score; however, we did not supply uniform bounds on the total-variation or Wasserstein distance between the true transition and the Gaussian approximation after the nonlinear flow is applied. In the revised manuscript we will insert a new lemma that derives such bounds under the standing Lipschitz and linear-growth assumptions on the drift, using the contractivity properties of the deterministic flow over small time steps and the fact that the Pearson Gaussian approximation matches the first two moments exactly. These bounds will be used to show that the difference between the approximate and true estimating functions is o_p(n^{-1/2}), thereby closing the argument for asymptotic efficiency. We view this as a clarification rather than a change in the main claims. revision: yes

Circularity Check

0 steps flagged

No circularity: estimator consistency and efficiency derived from independent moment calculations and standard splitting analysis

full rationale

The derivation begins with closed-form mean and covariance expressions for multivariate Pearson diffusions obtained via matrix exponential integrals, which are then used exactly in the Gaussian transition approximation for the Strang-split nonlinear system. The consistency and asymptotic efficiency proofs are stated to follow from this construction applied to the decomposed flows (nonlinear ODE plus Pearson diffusion), without any reduction of the target claims to fitted parameters, self-defined quantities, or load-bearing self-citations. The new Student Kramers oscillator is introduced with separate existence/uniqueness proofs, and simulation comparisons are external to the analytic claims. No step equates the efficiency result to its own inputs by construction.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 1 invented entities

Only the abstract is available; no explicit free parameters, axioms, or invented entities beyond the new model name can be extracted.

invented entities (1)
  • Student Kramers oscillator no independent evidence
    purpose: New example model inside the class of nonlinear diffusions with Pearson-type multiplicative noise
    Introduced and analyzed for existence, uniqueness, and invariant measure in the paper.

pith-pipeline@v0.9.0 · 5501 in / 1181 out tokens · 50438 ms · 2026-05-10T07:19:19.429711+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. Splitting schemes and estimators for stochastic differential equations with H\"older multiplicative noise

    stat.ME 2026-05 unverdicted novelty 6.0

    New splitting-scheme-based pseudo-likelihood estimators for SDEs with Hölder multiplicative noise that achieve strong convergence, state-space preservation, consistency, and asymptotic normality.

Reference graph

Works this paper leans on

17 extracted references · 9 canonical work pages · cited by 1 Pith paper · 1 internal anchor

  1. [1]

    URLhttp://www.jstor.org/stable/3318679

    ISSN 13507265. URLhttp://www.jstor.org/stable/3318679. M. Blondel, Q. Berthet, M. Cuturi, R. Frostig, S. Hoyer, F. Llinares-López, F. Pedregosa, and J.-P. Vert. Efficient and modular implicit differentiation.arXiv preprint arXiv:2105.15183,

    work page arXiv

  2. [2]

    URL http://github.com/jax-ml/jax. F. Carbonell, J. C. Jimenez, and L. M. Pedroso. Computing multiple integrals involving matrix exponentials.Journal of Computational and Applied Mathematics, 213(1):300–305, 2008a. doi:10.1016/j.cam.2007.01.007. F. Carbonell, J. Jímenez, and L. Pedroso. Computing multiple integrals involving matrix exponentials.Journal of ...

    work page doi:10.1016/j.cam.2007.01.007 2007

  3. [3]

    doi:10.1038/s41467-023-39810-w. P. D. Ditlevsen, S. Ditlevsen, and K. K. Andersen. The fast climate fluctuations during the stadial and interstadial climate states.Annals of Glaciology, 35:457–462,

    work page doi:10.1038/s41467-023-39810-w

  4. [4]

    doi:10.1016/j.aml.2015.04.009. J. C. Jimenez and T. Ozaki. Linear estimation of continuous–discrete linear state space models with multiplicative noise.Systems & Control Letters, 47(2):91–101,

    work page doi:10.1016/j.aml.2015.04.009 2015

  5. [5]

    doi:10.1016/S0167-6911(02)00150-0. J. C. Jimenez and T. Ozaki. Local linearization filters for non-linear continuous–discrete state space models with multiplicative noise.International Journal of Control, 76(12):1159–1170,

    work page doi:10.1016/s0167-6911(02)00150-0

  6. [6]

    doi:10.1080/0020717031000138214. M. Kessler. Estimation of an Ergodic Diffusion from Discrete Observations.Scand. J. Stat., 24(2):211–229,

    work page doi:10.1080/0020717031000138214

  7. [7]

    URLhttps://arxiv.org/abs/1412.6980. P. Kloeden and E. Platen.Numerical Solution of Stochastic Differential Equations. Stochastic Modelling and Applied Probability. Springer Berlin Heidelberg,

    work page internal anchor Pith review Pith/arXiv arXiv

  8. [8]

    doi:10.1007/978-0-387- 40065-5. G. A. Pavliotis.Stochastic Processes and Applications: Diffusion Processes, the Fokker-Planck and Langevin Equations, volume 60 ofTexts in Applied Mathematics. Springer,

    work page doi:10.1007/978-0-387-

  9. [9]

    doi:10.1007/978-1-4939-1323-7. G. A. Pavliotis and A. M. Stuart.Multiscale Methods: Averaging and Homogenization, volume 53 ofTexts in Applied Mathematics. Springer Science & Business Media,

    work page doi:10.1007/978-1-4939-1323-7

  10. [10]

    Sørensen

    M. Sørensen. Efficient estimation for ergodic diffusions sampled at high frequency. CREATES Research Papers 2007-46, Department of Economics and Business Economics, Aarhus University, Jan

    2007

  11. [11]

    Each case presents specific conditions for the existence and uniqueness of ergodic solutions and corresponding invariant distributions

    27 Strang estimator for nonlinear multivariate SDEs with Pearson-type noiseA PREPRINT S1 Supplementary Material: One-dimensional Pearson diffusions The ergodic one-dimensional Pearson diffusions (1) can be classified into six cases based on the form of the squared diffusion coefficient σ2(x) =αx 2 +βx+γ . Each case presents specific conditions for the exi...

    2008

  12. [12]

    A unique ergodic solution exists for all α >0 and m≥α/(2λ)

    , the process is defined on the positive half-line (0,∞) . A unique ergodic solution exists for all α >0 and m≥α/(2λ) . The invariant distribution is a scaled F-distribution with −4am/α and 2(1−2λ/α) degrees of freedom. If 0< m < α/(2λ) , the boundary at zero can be reached, but with an instantaneous reflecting boundary, the process remains stationary wit...

    2008

  13. [13]

    In the following we propose diagonal ˇαwhich leads to faster and more numerically stable algorithms

    We see that the choice of α(l)ij as in (8) leads to non-diagonal ˇα(21) . In the following we propose diagonal ˇαwhich leads to faster and more numerically stable algorithms. First, we write the full squared diffusion matrix as a function ofXas follows ΣΣ⊤(X) = diag(X)− LX l=1 ΠlXX⊤Πl,(S3) 30 Strang estimator for nonlinear multivariate SDEs with Pearson-t...

    1978

  14. [14]

    We have Ut =ψ(V t) = Z Vt dvp αv2 +βv+γ = 1√α arcsinh 2αVt +βp 4αγ−β 2 !

    for a one-dimensional nonlinear student-type Pearson diffusion. We have Ut =ψ(V t) = Z Vt dvp αv2 +βv+γ = 1√α arcsinh 2αVt +βp 4αγ−β 2 ! . Since we assume thatα >0and4αγ−β 2 ≤0, then Vt =ψ −1(Ut) = p 4αγ−β 2 sinh(√αUt)−β 2α . We transform (16) by applying Itô’s lemma toeψ(Xt, Vt) = (Xt, ψ(Vt)) dXt =F (1)(eψ−1(Xt, Ut)) dt, dUt = ∂ψ ∂v (eψ−1(Xt, Ut))F (2)(e...

    1967

  15. [15]

    We choose V(x, v) = 1 2 v2 +U(x)as a Lyapunov function

    to prove that SDE(16) has an invariant measure. We choose V(x, v) = 1 2 v2 +U(x)as a Lyapunov function. Sincea <0,lim ∥x∥→∞ V(x) = +∞. For SDE (16), the infinitesimal generator is Lϕ(x, v) =v ∂ϕ ∂x + (−ηv−U ′(x)) ∂ϕ ∂v + 1 2(αv2 +βv+γ) ∂2ϕ ∂v 2 . Then, LV(x, v) = α 2 −η v2 + β 2 v+ γ 2 .(S21) Sinceα <2η, we can find a compact setK⊂R 2 and constantsc 1 >0,...

    1993

  16. [16]

    tk−1 =x] =ΣΣ ⊤(x) + (s−t k−1)L[1]ΣΣ⊤(x) +R((s−t k−1)2,x). Then, (35) becomes Ωh(x) =hΣΣ ⊤(x) + h2 2 AΣΣ⊤(x) +ΣΣ ⊤(x)A⊤ +L [1]ΣΣ⊤(x) +R(h 3,x).(S32) 36 Strang estimator for nonlinear multivariate SDEs with Pearson-type noiseA PREPRINT As Strang splitting approximation (39) uses Ωh(fh/2(x)), we approximate it using Lemma S5.2 and the fact that fh/2(x) =x+h/...

    2025

  17. [17]

    More precisely, the consistency of the diffusion parameter is proved by studying the limit of 1 N L[SS] N (θ)

    and study the limit ofL[SS] N (θ) from (S37) rescaled by the correct rate of convergence. More precisely, the consistency of the diffusion parameter is proved by studying the limit of 1 N L[SS] N (θ). In contrast, the consistency of the drift parameter is proved by studying the limit of 1 N h(L[SS] N (θ(1),θ (2))− L [SS] N (θ(1) 0 ,θ (2)). To prove the co...

    1997