Sparse Equation Matching (SEM)

• Sparse Equation Matching (SEM) is a unified framework that reformulates differential equations using Green’s functions to bypass direct derivative estimation.
• It employs ℓ1-regularized sparse regression on integral features to robustly identify linear operator coefficients and sparse driving terms from noisy data.
• SEM exhibits superior recovery and prediction performance in both simulated systems and real-world applications such as EEG analysis.

Sparse Equation Matching (SEM) is a unified framework for equation discovery in general-order dynamical systems that eliminates the need for direct derivative estimation. SEM leverages Green’s functions to reformulate differential equations as integral equations, enabling robust, derivative-free learning of both the underlying linear differential operator and the sparse driving terms. The approach accommodates arbitrary system order, supports high-dimensional data, and demonstrates improved empirical performance compared to derivative-based methods—particularly in the inference of dynamic biological and physical systems from noisy time series (Li et al., 26 Jul 2025).

1. Mathematical Foundations

Let $X(t) = (X_1(t), \dotsc, X_p(t))^\top$ denote a $p$-dimensional trajectory observed on $t \in [0, C]$. SEM assumes each component satisfies a linear $K$th-order ODE of the form

$$P_i^K X_i(t) = f_i(X(t), t), \quad i = 1, \dotsc, p,$$

where the operator $P_i^K$ is

$$P_i^K = \frac{\mathrm{d}^K}{\mathrm{d} t^K} + \sum_{l=1}^{K-1} \omega_{i\,l} \frac{\mathrm{d}^l}{\mathrm{d} t^l},$$

and $f_i(X(t), t)$ is an unknown driving function. By introducing Green’s functions $G^k(t, s)$ for $\frac{\mathrm{d}^k}{\mathrm{d} t^k}$, these ODEs are reformulated as integral equations. Notably, for $p$0, the formulation is fully integral and eliminates all derivatives of $p$1: $p$2 where $p$3 and $p$4 are feature maps constructed via integrals of Green’s functions applied to the state $p$5 and predetermined basis functions $p$6; $p$7 concatenates the negative linear coefficients and the sparse coefficients $p$8 for the dictionary expansion of $p$9.

2. Integral-Based Sparse Regression

Equation discovery proceeds by representing $t \in [0, C]$0 as a sparse linear expansion in a chosen dictionary of $t \in [0, C]$1 candidate basis functions: $t \in [0, C]$2 with basis maps $t \in [0, C]$3. The estimation problem is formulated as an $t \in [0, C]$4-regularized (LASSO) least squares minimization: $t \in [0, C]$5 where $t \in [0, C]$6 discretize $t \in [0, C]$7, $t \in [0, C]$8 is a pre-smoothed trajectory, and $t \in [0, C]$9 is selected by cross-validation. The design matrix encodes the time-integrated effect of the candidate dynamical terms, permitting the extraction of a parsimonious set of active basis functions even in the presence of measurement noise and without direct computation of derivatives.

3. Algorithmic Workflow

The SEM procedure involves the following sequence:

1. Pre-smoothing: Each signal coordinate is denoised by RKHS regression in a Sobolev space $K$0 (typically with $K$1), balancing fidelity and smoothness penalties ($K$2 via generalized cross-validation).
2. Feature construction: For each time point, compute integrals involving Green’s functions applied to the smoothed trajectories and basis functions.
3. Sparse regression: Assemble the design matrix and solve the LASSO problem for each variable coordinate, yielding estimates of linear operator coefficients and dictionary sparsity pattern.
4. Outputs: The procedure recovers system parameters $K$3 and driving term coefficients $K$4.

Computational complexity is dominated by the pre-smoothing stage and the formation of integral features, with naive implementation at $K$5 for integrals (improved to $K$6 via FFT) and $K$7 for LASSO. Hyperparameters include the smoothing order $K$8, regularization parameters $K$9 and $$P_i^K X_i(t) = f_i(X(t), t), \quad i = 1, \dotsc, p,$$0, and the integration window as set by data span.

4. Advantages and Benchmarks versus Derivative-Based Methods

SEM generalizes and unifies existing methods—gradient matching ($$P_i^K X_i(t) = f_i(X(t), t), \quad i = 1, \dotsc, p,$$1) and integral matching ($$P_i^K X_i(t) = f_i(X(t), t), \quad i = 1, \dotsc, p,$$2)—within a single Green’s function–based integral formulation. Its main theoretical advantages are:

• Derivative-free formulation avoids the instability and bias that afflict high-order numerical differentiation (e.g., in SINDy).
• Simultaneous learning of differential operator coefficients and sparse driving terms from noisy, finite-sampled data.
• Flexibility with respect to system order, permitting modeling and inference in arbitrary-order dynamical systems.

Benchmarks demonstrate substantially improved recovery and prediction fidelity:

• For a nonlinear pendulum ($$P_i^K X_i(t) = f_i(X(t), t), \quad i = 1, \dotsc, p,$$3), relative error rate (RER) over 500 runs with $$P_i^K X_i(t) = f_i(X(t), t), \quad i = 1, \dotsc, p,$$4, noise $$P_i^K X_i(t) = f_i(X(t), t), \quad i = 1, \dotsc, p,$$5: SINDy RER ≈ 0.15 vs. SEM RER ≈ 0.05. Vector field reconstructions show SEM more faithfully recovers system dynamics.
• For synthetic 40-node directed networks, mean adjacency matrix recovery accuracy: SINDy ≈ 75%, SEM ≈ 92%.

A notable limitation is the requirement of reliable pre-smoothing, as the two-step approach presumes sufficiently dense sampling for accurate functional reconstruction. In sparsely sampled regimes, a joint estimation strategy may be needed (Li et al., 26 Jul 2025).

5. Application to Multichannel EEG Data

SEM was applied to EEG recordings from 52 participants engaged in oculomotor tasks (eye-blinking, horizontal/vertical movement, resting; 64 channels, 512 Hz). Each signal was pre-smoothed by RKHS regression ($$P_i^K X_i(t) = f_i(X(t), t), \quad i = 1, \dotsc, p,$$6), and the system was modeled as a second-order ODE: $$P_i^K X_i(t) = f_i(X(t), t), \quad i = 1, \dotsc, p,$$7 where $$P_i^K X_i(t) = f_i(X(t), t), \quad i = 1, \dotsc, p,$$8 includes polynomial basis terms up to degree 2 in $$P_i^K X_i(t) = f_i(X(t), t), \quad i = 1, \dotsc, p,$$9.

On one-step-ahead prediction tasks (using 4 s training and 1 s validation per trial), SEM attained significantly lower relative prediction errors than SINDy across all channels and participants. Example time series from channels C3, CP5, FT7 demonstrated that SEM tracks held-out signal remarkably closely, whereas SINDy drifts. These predictive gains confirm the advantage of an integral matching approach in high-noise, high-dimensional biological systems (Li et al., 26 Jul 2025).

6. Network Inference and Neuroscientific Interpretation

SEM yields a directed adjacency matrix for each subject and task, revealing brain connectivity patterns at the population and individual levels. Population-level edges were assessed via one-sided binomial tests with false discovery rate (FDR) control at 5% (Benjamini–Hochberg), and task-specific connectivity identified by Fisher exact test (FDR 5%).

Empirical results:

• Resting state showed no consistent directed connectivity.
• Eye-blinking exhibited dense prefrontal connectivity, with FP2, F7, F5, FPz as hubs.
• Horizontal saccades were characterized by bilateral prefrontal hub nodes (AF4, AF3, AF7, AF8).
• Vertical saccades exhibited a posterior-to-anterior propagation (notably PO3 → F5), with weaker bilateral coupling; key nodes included FT7 and AF4.

Network metrics (in/out-degree, betweenness, closeness) identified FP2 as a stable oculomotor hub, consistent with established neurophysiological literature implicating the right prefrontal region in voluntary saccade control. Observed differences between horizontal (interhemispheric) and vertical (anterior–posterior) oculomotor circuitry corroborate documented neuroanatomical pathways (Li et al., 26 Jul 2025).

7. Summary and Context

Sparse Equation Matching introduces a derivative-free regime for sparse equation discovery in dynamical systems of arbitrary order. Through the use of Green’s function–based integral operators, SEM unifies and generalizes earlier matching paradigms, supports robust sparse regression, and demonstrably improves both recovery and prediction in simulated and physiological data. Its application to EEG data enables the extraction of directed functional networks with high anatomical plausibility, highlighting SEM’s value in neuroscientific and complex system inference settings (Li et al., 26 Jul 2025).