Technical Summary: Knowledge-Informed Kernel State Reconstruction from Heterogeneous Partial Observations

Problem Formulation

Real-world scientific systems, particularly in clinical and biological domains, are rarely observed through complete, regularly sampled state trajectories. Instead, measurements are often partial, noisy, and heterogeneous, providing fragmented views of latent dynamical states. For instance, in oncology, tumor burden may be observed via sparse imaging while dense but indirect biomarkers are derived from blood panels. These observations differ in temporal resolution, noise characteristics, and their relationship to the latent state.

The core challenge is a data fusion problem: recovering physically meaningful latent trajectories and their time derivatives from fragmented observations with heterogeneous structure. Existing approaches fail to address this holistically:

• Classical smoothing (e.g., splines) ignores domain constraints and heterogeneous observation operators.
• State-space models (e.g., Kalman filters) require specifying transition dynamics a priori.
• Deep latent models (e.g., Neural ODEs) rely on black-box representations that obscure mechanistic structure.

Consequently, state reconstruction is often treated as a mere preprocessing step, creating a bottleneck for downstream mechanistic discovery methods like symbolic regression, which require accurate, analytic derivatives.

Methodology: MAAT

The authors introduce MAAT (Model Aware Approximation of Trajectories), a framework for knowledge-informed kernel state reconstruction. MAAT formulates the reconstruction as a risk minimization problem within a Reproducing Kernel Hilbert Space (RKHS).

1. Optimization Framework

MAAT seeks a trajectory function $\hat{x} \in \mathcal{H}_K$ that minimizes a regularized empirical risk incorporating three components:
$$\hat{x} = \arg\min_{x \in \mathcal{H}_K} \underbrace{\sum_{i=1}^N \|y_i - H_i(x(t_i))\|^2}_{\text{Data Fidelity}} + \underbrace{\lambda_1 \|x\|_{\mathcal{H}_K}^2}_{\text{Smoothness}} + \underbrace{\lambda_2 R_{\text{phys}}(x, C)}_{\text{Physical Priors}}$$

• Data Fidelity: Accounts for heterogeneous observation operators $H_i$, allowing the model to handle diverse measurement types (e.g., sparse direct measurements vs. dense aggregated signals) and noise levels simultaneously.
• Smoothness: Enforced via the RKHS norm, ensuring continuous trajectories.
• Physical Priors ($R_{\text{phys}}$): Incorporates semantic and structural constraints such as non-negativity, conservation of mass, and domain-specific measurement models.

2. Kernel State Reconstruction (KSR)

The trajectory for each state variable is modeled as a linear combination of kernel functions: $\hat{x}_j(t) = \sum_{\ell=1}^N u_{\ell j} \kappa(t, t_\ell)$. The coefficients $U$ are learned by minimizing a composite loss function that balances:

• Fidelity to full-state snapshots.
• Fidelity to linear signals (via observation operators $H_i$).
• Deviation from prior hypotheses on system dynamics (e.g., penalizing deviation from a known vector field $F$).
• Regularization of the coefficients.

3. Analytic Derivatives and Robustness

A key feature of MAAT is its ability to provide analytic time derivatives. Since the time derivative of the reconstructed trajectory acts only on the kernel function (which is $C^\infty$ smooth, e.g., Gaussian), derivatives admit closed-form expressions.

• Theoretical Justification: The authors prove that the composite loss is a calibrated surrogate for the true $L_2$ reconstruction error.
• Noise Robustness: Unlike finite difference methods, which suffer from an irreducible error floor due to noise amplification ($O(\sigma^2/\Delta t^2)$), MAAT's analytic derivative estimator scales with the regularization parameter and sample size ($O(\lambda) + O(\sigma^2/n)$), avoiding high-frequency noise amplification.

Key Contributions

1. Conceptual: The paper identifies knowledge-informed state reconstruction from heterogeneous partial observations as a central bottleneck for mechanistic modeling.
2. Technical: MAAT embeds latent trajectories in an RKHS and directly incorporates heterogeneous observation operators and structural priors into the reconstruction objective, rather than treating them as separate preprocessing steps.
3. Empirical: The framework yields smooth, physically consistent state estimates with analytic derivatives, providing a principled interface for downstream symbolic regression.

Experimental Results

The authors evaluated MAAT on nine scientific benchmarks (spanning epidemiology, oncology, and pharmacology) and a real-world COVID-19 dataset.

• Performance: MAAT substantially reduced trajectory and derivative reconstruction errors compared to strong baselines, including RBF smoothing, Cubic Splines, Gaussian Processes, Kalman Filters, and Neural ODEs.
• Noise Regimes: The method demonstrated robustness across Gaussian, correlated Gaussian, and heavy-tailed Student-t noise regimes.
• Structural Priors: Incorporating domain knowledge (e.g., mass conservation in SEIR models, non-negativity constraints) significantly reduced the Mean Squared Error (MSE) of recovered trajectories.
• Downstream Impact: When coupled with symbolic regression algorithms (SINDy and PySR), the trajectories reconstructed by MAAT led to more accurate equation discovery compared to those derived from baseline methods.
• Real-World Application: On the COVID-19 dataset, MAAT achieved significantly lower test MSE than all baselines when used to recover disease dynamics.

Significance and Claims

The paper positions MAAT as a solution to the "representational bottleneck" in scientific machine learning. By moving state reconstruction from a purely numerical preprocessing step to a knowledge-informed inference problem in function space, MAAT enables:

• Physically Consistent Estimates: Trajectories that respect domain constraints (e.g., non-negativity, conservation laws).
• Mechanistic Discovery: High-quality analytic derivatives that are critical inputs for successful symbolic regression, allowing researchers to discover governing equations from fragmented data.
• Clinical Relevance: The framework aligns with workflows in Quantitative Systems Pharmacology (QSP), where models are assembled from interpretable modules. MAAT supports this by fusing sparse, subsystem-specific measurements with denser indirect signals while enforcing clinically meaningful constraints, making it suitable for model-informed drug development and regulatory decision-making.

The authors note that the effectiveness of MAAT depends on the informativeness of available observations and the validity of injected priors. If measurements are too sparse or priors are misspecified, reconstruction may be biased. However, within the context of structured scientific domains where such priors are available, MAAT offers a principled path from fragmented data to mechanistic understanding.