Weak Form Estimation of Nonlinear Dynamics

Updated 24 March 2026

WENDy is a computational framework that transforms ODE systems into a weak form using smooth, compactly supported test functions to mitigate noisy differentiation.
It employs Iteratively Reweighted Least Squares and likelihood maximization to address errors-in-variables in both linear and nonlinear-in-parameter scenarios.
WENDy extends to high-dimensional, latent-space, and complex noise applications, offering enhanced accuracy and computational speed-ups compared to traditional methods.

Weak Form Estimation of Nonlinear Dynamics (WENDy) is a family of computational methods for parameter estimation and model selection in nonlinear dynamical systems, particularly ordinary differential equations (ODEs), via transformation of the system into a weak (variational) form. The principal innovation is the use of integration against smooth, compactly supported test functions, which eliminates the need for pointwise differentiation of noisy time-series data, and replaces it with linear operations that yield noise-robust, computationally efficient algorithms. WENDy encompasses linear-in-parameter as well as nonlinear-in-parameter models and has been extended to a range of application domains including high-dimensional systems, systems with unobserved variables, and models with complex noise structures (Bortz et al., 2023, Rummel et al., 13 Feb 2025, Chawla et al., 3 Oct 2025, Heitzman-Breen et al., 20 Jun 2025).

1. Weak-Formulation of Nonlinear Dynamics

The starting point for WENDy is the recasting of an ODE system into its weak form. For a system

$\dot{x}(t) = f(x(t), \theta),$

with unknown parameters $\theta \in \mathbb{R}^J$ , WENDy multiplies both sides by a compactly supported, smooth test function $\varphi(t)$ and integrates over the observation interval $[0,T]$ . Integrating by parts (using $\varphi(0)=\varphi(T)=0$ ) yields:

$- \int_0^T \varphi'(t) x(t)\,dt = \int_0^T \varphi(t) f(x(t), \theta)\,dt.$

If $f(x, \theta)$ is linear in $\theta$ , the right-hand side becomes a linear combination over features; for nonlinear-in-parameter models, a nonlinear system in $\theta$ emerges (Bortz et al., 2023, Rummel et al., 13 Feb 2025).

Discretization replaces integrals with weighted sums (typically via numerical quadrature, e.g., the composite trapezoidal rule), forming a set of weak-form equations for each test function. For $K$ test functions $\varphi_k$ , the system assembles into a linear or nonlinear system in $\theta$ :

$G \theta \approx b,$

where $G \in \mathbb{R}^{K \times J}$ encodes the integrals of features against test functions, and $b \in \mathbb{R}^K$ encodes the data side.

2. Algorithmic Structure and Statistical Foundations

WENDy for linear-in-parameter ODEs reduces to a regression problem, where parameter estimation is performed via least squares or its weighted variant, accounting for the noise-induced covariance of residuals—arising from the fact that both sides of the regression equation depend on the noisy data. This is the Errors-in-Variables (EiV) problem, for which ordinary least squares is biased.

To address EiV, WENDy employs Iteratively Reweighted Least Squares (IRLS) (Bortz et al., 2023, Chawla et al., 3 Oct 2025). At each iteration, the covariance $C^{(n)}$ of the residuals is estimated by linearizing the regression with respect to noise and propagating the estimated noise through the quadrature structure. The weighted least squares update is then:

$\theta^{(n+1)} = (G^\top (C^{(n)})^{-1} G)^{-1} G^\top (C^{(n)})^{-1} b.$

For nonlinear-in-parameter ODEs, WENDy maximizes the weak-form likelihood, making use of closed-form expressions for the likelihood, gradient, and Hessian, which enables robust and efficient second-order trust-region or regularized Newton optimization (Rummel et al., 13 Feb 2025).

The design of test functions is critical: C-infinity bump functions, piecewise polynomials, and their orthonormalizations provide bases with rapid spectral decay, ensuring high numerical stability and denoising via integration (Bortz et al., 2023).

3. Practical Identifiability and (e, q)-Mapping

Parameter identifiability in the presence of observational noise is assessed via the $(e, q)$ -identifiability criterion, where $e$ denotes the ratio of noise standard deviation to the root-mean-square of the observed variable, and $q$ fixes the maximal mean-square error (relative to the true parameter value) deemed tolerable. For each parameter $\theta_k$ , identifiability is declared if

$\mathrm{MSE}(\hat{\theta}_k; e) < (q |\theta_k^\star|)^2$

across simulated datasets with additive noise ratio $e$ (Heitzman-Breen et al., 20 Jun 2025). This mapping directly relates noise magnitude to estimation accuracy and can be computed efficiently via WENDy without repeated forward simulations, in contrast to output-error methods.

WENDy enables rapid, noise-robust identifiability assessment even for systems with unobserved compartments through the use of differential algebra elimination, transforming the system to observable-only weak-form equations prior to regression (Heitzman-Breen et al., 20 Jun 2025).

4. Extensions and Applications

Nonlinear-in-Parameter Extensions

WENDy extends to nonlinear-in-parameter ODEs by formulating a weak-form negative log-likelihood under additive Gaussian or multiplicative log-normal noise. The likelihood structure incorporates the covariance of the residuals, analytically differentiates the likelihood for gradient and Hessian computation, and is minimized by robust global solvers (Rummel et al., 13 Feb 2025). This accommodates accurately estimating parameters in rational, Hill-type, or other nonlinear parametric structures, and for both additive and multiplicative (e.g., log-normal) noise models.

High-Dimensional and Latent-Space Reduction

The weak form has been successfully adapted to parameter estimation and dynamics learning in latent spaces of autoencoder-based reduced-order models. In this setting, WENDy is integrated within frameworks such as WgLaSDI, where an autoencoder and weak-form latent ODE identification are trained jointly. The weak-form approach provides enhanced noise-robustness and enables training on high-dimensional datasets where standard strong-form approaches degrade catastrophically in noise (He et al., 2024).

Applications to Biological, Physical, and Compartmental Models

WENDy has demonstrated performance on biological models (logistic growth, Lotka–Volterra, SIR epidemic, blood–tissue diffusion), neuroscience (FitzHugh–Nagumo, Hindmarsh–Rose), and biochemical networks (Protein Transduction Benchmark), exhibiting up to 100× computational speed-up and higher estimation accuracy compared to direct ODE solving, particularly with sparse or noisy data (Bortz et al., 2023, Heitzman-Breen et al., 20 Jun 2025, Chawla et al., 3 Oct 2025).

5. Noise Robustness, Consistency, and Statistical Properties

WENDy's robustness stems from two aspects: (1) integration against smooth test functions low-pass filters high-frequency measurement noise and (2) IRLS weighting corrects bias from noisy regressors. The method achieves asymptotic unbiasedness and nominal confidence interval coverage under mild regularity for additive Gaussian noise (Chawla et al., 3 Oct 2025).

Key empirical findings, summarized in the table below, illustrate coverage and bias as a function of noise structure, problem dimension, and data resolution:

Model	Max $\gamma$ (AN) with $\geq$ 50% CI Coverage	Min Points Needed
Logistic (2D)	0.70	~$120$
Lotka-Volterra (4D)	0.40	~$120$
FitzHugh-Nagumo	0.07	~$200$
Hindmarsh–Rose	0.0025 (MLN)	~$300$
Protein Transd.	0.80 (AN, ACN)	~$130$

Bias typically remains below 10% for moderate noise; coverage is close to nominal at appropriate resolutions until noise becomes extreme (Chawla et al., 3 Oct 2025).

For large noise, the weak-form methods retain identifiability up to a critical threshold on the noise standard deviation, above which the algorithm may select spurious terms unless simple denoising (e.g., moving average) is applied (Messenger et al., 2022). The explicit threshold is given by

$\sigma_c \simeq \sqrt{2\log K}\,\frac{\max_{k} \|\varphi f_k\|_2}{\min_{k \in S} \|\varphi f_k\|_2^2}$

where $K$ is the number of candidate terms and $\varphi$ is the test function (Messenger et al., 2022).

6. Comparison with Other Weak-Form and Sparse Identification Methods

WENDy builds upon and is closely related to the Weak-SINDy (WSINDy) framework (Messenger et al., 2020, Messenger et al., 2020), which emphasizes model structure discovery (sparse identification) in arbitrary libraries of candidate functions. WENDy is typically employed when model structure is known (or prescribed by the application domain) and parameterization is the primary interest, whereas WSINDy targets model selection from large candidate bases with $\ell_1$ or sequential-thresholding regularization. Both methods leverage the weak form for noise robustness and avoid direct data differentiation (Wang et al., 2024).

A key distinction: WENDy uses weighted least squares (via IRLS or analytic likelihood maximization) to neutralize errors-in-variables and directly incorporates statistical uncertainty quantification, while WSINDy employs sparse regression (thresholded least squares or LASSO) for support selection (Bortz et al., 2023, Rummel et al., 13 Feb 2025, Messenger et al., 2020).

In the context of partial differential equations, weak-form SINDy (WSINDy-PDE) has demonstrated optimal scaling and robustness by discretizing convolutional weak forms and leveraging fast Fourier transform separability, but the essential noise-mitigation principle remains shared across these frameworks (Messenger et al., 2020).

7. Implementation, Efficiency, and Practical Guidelines

Efficient implementation of WENDy involves the following components:

Precompute test function matrices and quadrature weights (bump, polynomial, or Hartley bases).
For observed data, assemble the weak-form linear (or nonlinear) regression system using vectorized weighted sums.
Employ IRLS for linear-in-parameter estimation or trust-region/Newton methods for nonlinear-in-parameter likelihood maximization.
Propagate residual covariance estimates for confidence intervals and coverage assessment (Bortz et al., 2023, Rummel et al., 13 Feb 2025, Chawla et al., 3 Oct 2025).

Best-practices recommendations: select $K=200–500$ test functions (for moderate ODE dimension), choose test function radii adaptively to balance integration error against noise, use the trapezoidal rule for quadrature when $M\geq 100$ , and apply mild low-pass filtering for stiff or oscillatory systems at high noise (Chawla et al., 3 Oct 2025, Rummel et al., 13 Feb 2025). Open-source software implementations are available in MATLAB and Julia (Bortz et al., 2023, Rummel et al., 13 Feb 2025).

WENDy exhibits linear scaling with data size and remains computationally efficient for stiff, high-dimensional, or partially observed systems, with empirical speedup factors of 5–100× compared to traditional output-error-based inference using forward ODE solves (Heitzman-Breen et al., 20 Jun 2025, Bortz et al., 2023).

References