Automated scientific discovery aims to extract governing physical laws from experimental data. A fundamental challenge in this process is identifiability—determining whether a discovered symbolic model is the unique representation of the underlying system or if multiple models could explain the data equally well. While previous research often assumed noise-free data, real-world applications must contend with measurement errors that can obscure the true physics.
Theoretical Framework and Algorithms
The authors introduce a mathematical framework to assess identifiability under noise. They leverage Weyl’s Theorem to bound noise-induced perturbations in singular values and derive new error bounds for finite difference (FD) approximations of derivatives. Two primary algorithms are proposed:
- NR-FRanCo (Noise Robust Stable Feature Rank Computation): This algorithm targets linear PDEs by analyzing the spectral properties of a feature matrix. It identifies non-uniqueness when the ratio of the smallest to largest singular value falls below a noise-dependent threshold.
- NR-JRC (Noise Robust Jacobian Rank Computation): Designed for more complex algebraic and analytic PDEs, this method utilizes the Jacobian of feature functions. It is inherently more sensitive to noise due to the requirement of higher-order derivatives.
Experimental Results and Robustness
The algorithms were tested on simulated datasets with varying levels of Gaussian noise. The study found that the ability to distinguish between unique and non-unique systems is highly dependent on both the noise level and the order of the numerical differentiation scheme used.
As shown in Figure 2, the regions where a PDE can be confidently classified as unique or non-unique shift as noise increases. The "no definitive conclusion" zone expands, reflecting the mathematical ill-posedness of the problem at high noise levels.
Figure 4 illustrates that NR-FRanCo is remarkably robust for identifying unique linear PDEs even at noise levels up to $10^{-1}$. However, identifying non-uniqueness is significantly more fragile, with performance degrading rapidly beyond $10^{-5}$ noise. NR-JRC exhibits even greater sensitivity, struggling with non-uniqueness at noise levels as low as $10^{-8}$.
Key Insights
A critical finding is that proving uniqueness is fundamentally more robust than proving non-uniqueness. The smallest singular values, which indicate non-uniqueness, are highly susceptible to even tiny perturbations. Furthermore, the "cost of derivatives" is evident: algorithms requiring higher-order derivatives (like NR-JRC) amplify noise exponentially, limiting their practical range. The paper argues that when an algorithm fails to reach a conclusion, it is often a reflection of the fact that the problem has become mathematically ill-posed due to excessive noise, rather than a failure of the algorithm itself.
Conclusion and Future Directions
This work provides a rigorous foundation for verifiable symbolic discovery. Future research may explore probabilistic error bounds—specifically tailored for Gaussian noise—to achieve tighter, more practical thresholds. Additionally, scaling these robust checks to high-dimensional systems remains a challenge due to the "curse of dimensionality" in derivative calculation.
Summary updated