Nonparametric model reconstruction for stochastic differential equation from discretely observed time-series data

A scheme is developed for estimating state-dependent drift and diffusion coefficients in a stochastic differential equation from time-series data. The scheme does not require to specify parametric forms for the drift and diffusion coefficients in advance. In order to perform the nonparametric estimation, a maximum likelihood method is combined with a concept based on a kernel density estimation. In order to deal with discrete observation or sparsity of the time-series data, a local linearization method is employed, which enables a fast estimation.

💡 Research Summary

**
The paper presents a novel, fully non‑parametric framework for reconstructing stochastic differential equations (SDEs) from discretely observed time‑series data, without any prior specification of the functional forms of the drift f(x) and diffusion g(x). The authors begin by motivating the need for such methods in small‑scale biological systems where stochasticity is intrinsic and experimental techniques now provide single‑molecule trajectories. Traditional approaches—parametric maximum‑likelihood, Markov‑chain Monte‑Carlo, or variational methods—require pre‑chosen functional forms or incur prohibitive computational costs, making them unsuitable for many real‑world applications.

The core of the proposed method combines three ideas: (1) kernel density estimation (KDE) to obtain a smooth estimate of the marginal distribution of the observed states; (2) a reformulation of KDE within a maximum‑likelihood (ML) framework, treating the density itself as the parameter to be estimated; and (3) a local linearization of the SDE to handle the fact that observations are spaced at finite, possibly large, time intervals.

In the KDE‑ML step, the authors derive the log‑likelihood of a candidate density p̂(x) as L(p̂|X)=∑₁ᴺ∫log p̂(x) K_h(x−X_i)dx, where K_h is a Gaussian kernel with bandwidth h. Introducing a Lagrange multiplier to enforce the normalization constraint yields the familiar KDE solution p̂(x)= (1/N)∑₁ᴺ K_h(x−X_i). This derivation shows that the “distributed” pseudo‑observations induced by the kernel are mathematically equivalent to a weighted likelihood, providing a principled justification for KDE beyond heuristic smoothing.

To estimate the conditional density p(ΔX|X) needed for SDE inference, the authors extend the ML‑KDE argument to pairs (X_i,ΔX_i). However, unlike standard conditional KDE, the pairs are not identically distributed because ΔX_i depends on the varying time step Δt_i. To overcome this, they approximate the dynamics locally: within each observation interval the nonlinear SDE dx = f(x)dt + g(x)dW is linearized around the current state X_i, yielding a locally linear SDE dx ≈ a_i (x−X_i)dt + b_i dW. The coefficients a_i and b_i are directly related to the unknown f and g evaluated at X_i (via first‑order Taylor expansion). For a linear SDE the transition density is Gaussian with mean a_i X_i Δt_i and variance b_i² Δt_i. Substituting this Gaussian form into the log‑likelihood and weighting each term with the kernel K_h(x−X_i) leads to an overall objective that is a sum of analytically tractable terms.

Maximizing this objective with respect to the unknown functions f and g (represented implicitly through a_i and b_i) yields closed‑form expressions that are essentially kernel‑weighted averages of the observed increments ΔX_i. The result is a non‑parametric estimator for the drift and diffusion that automatically adapts to the local density of data points, preserving fine‑scale features while remaining robust to noise.

The authors validate the method on a synthetic one‑dimensional SDE with a double‑well potential (drift f_true(x)=−4x³+4x) and a state‑dependent diffusion g_true(x)=0.2 sin(πx). They generate a high‑resolution trajectory (Δt=0.001) and then down‑sample to Δt=0.05, obtaining 2000 observations. Applying the proposed algorithm, they recover f̂(x) and ĝ(x) that closely match the true functions, even though the data set is relatively small and the observation interval is not infinitesimal. Comparisons with naïve parametric fits demonstrate that the non‑parametric approach avoids misspecification bias and captures the bimodal stationary distribution accurately.

Key advantages highlighted include: (i) complete avoidance of parametric assumptions, (ii) computational complexity linear in the number of observations (O(N)), (iii) automatic bandwidth selection via cross‑validation of a risk functional, and (iv) flexibility to handle irregular sampling intervals through the local linearization step. The paper concludes by suggesting extensions to multivariate SDEs, time‑varying coefficients, and real experimental data from single‑molecule fluorescence or intracellular signaling studies. Overall, the work provides a practical, theoretically grounded tool for scientists needing to infer stochastic dynamics directly from noisy, sparsely sampled time‑series.