Q-SINDy: Quantum-Kernel Sparse Identification of Nonlinear Dynamics with Provable Coefficient Debiasing

1. Problem Statement

Sparse Identification of Nonlinear Dynamics (SINDy) is a popular framework for discovering governing equations of dynamical systems from time-series data by selecting active terms from a library of candidate functions (typically polynomials).

The authors investigate a natural extension: augmenting the standard polynomial library ($P$) with a quantum feature library ($Q$) derived from quantum feature maps (e.g., ZZ-angle encoding, IQP, data re-uploading). The goal is to leverage the expressive power of quantum embeddings to capture residual dynamics that polynomials miss.

The Core Issue: The authors identify a systematic failure mode in naive implementations called "Coefficient Cannibalization." When quantum features are simply concatenated with polynomial features ($\Theta = [P, Q]$), the quantum features absorb explanatory power that rightfully belongs to the polynomial basis. This biases the estimated polynomial coefficients ($\hat{\xi}_P$) away from their true values, corrupting the recovery of the underlying physical equations, even if the overall fit error is low.

2. Methodology

2.1 Q-SINDy Framework

The method solves the sparse regression problem:
$$\min_{\Xi} \| \dot{X} - [P, Q] \Xi \|_F^2 + \lambda \| \Xi \|_0$$
where:

• $P$: Polynomial library (e.g., $1, x, y, x^2, xy, \dots$).
• $Q$: Quantum library containing expectation values of observables from a parametrized quantum circuit $U_\phi(x)$.
• $\dot{X}$: Time derivatives of the state.

2.2 Theoretical Analysis: Coefficient Cannibalization

Using least-squares theory (relaxing the sparsity constraint), the authors derive the exact bias formula for the polynomial coefficients in a naive augmented model:
$$\hat{\xi}_P = \hat{\xi}_P^{\text{vanilla}} - \underbrace{(P^\top P)^{-1} P^\top Q \hat{\xi}_Q}_{\Delta \xi_P}$$

• $\Delta \xi_P$ represents the "cannibalization" bias.
• It occurs when two conditions are met: (1) $P$ and $Q$ are not orthogonal ($P^\top Q \neq 0$), and (2) quantum features have explanatory power for the dynamics ($\hat{\xi}_Q \neq 0$).
• Since quantum feature maps often have high overlap with polynomial bases on trajectory data, this bias is severe and systematic.

2.3 The Solution: Orthogonalization

The authors propose a geometric correction: Orthogonalizing the quantum features against the polynomial column space before regression.

1. Compute the projection of $Q$ onto the orthogonal complement of $P$:
   $$Q_\perp = Q - P(P^\top P)^{-1} P^\top Q$$
2. Construct the library as $\Theta = [P, Q_\perp]$.
3. Theorem 3 proves that with this modification, the estimated polynomial coefficients become identical to the vanilla SINDy estimates ($\hat{\xi}_P^{\text{orth}} = \hat{\xi}_P^{\text{vanilla}}$), regardless of the quantum coefficients. The quantum features are restricted to explaining only the variance that $P$ cannot.

2.4 Diagnostic Tool ($R^2_Q$)

To predict whether cannibalization will occur for a given system and feature map, the authors introduce a dynamics-aware diagnostic:
$$R^2_Q = 1 - \frac{\| \dot{X} - Q(Q^\dagger \dot{X}) \|_F^2}{\| \dot{X} - \bar{\dot{X}} \|_F^2}$$
This measures how well the quantum features alone can predict the time derivative. A high $R^2_Q$ indicates high risk of cannibalization.

3. Key Contributions

1. Discovery of Coefficient Cannibalization: Identification of a specific failure mode where quantum features corrupt the interpretability of classical polynomial terms in SINDy.
2. Provable Correction: Derivation of the exact bias formula and a proof that simple orthogonalization eliminates this bias exactly, preserving the structural recovery of the governing equations.
3. Predictive Diagnostic: Development of the $R^2_Q$ metric, which predicts cannibalization severity with statistical significance ($r=0.70, p=0.023$), outperforming simple column-space overlap metrics.
4. Extensive Validation: Comprehensive testing across six dynamical systems, three quantum architectures, and various noise levels.

4. Experimental Results

The authors evaluated the method on six canonical systems (Duffing, Van der Pol, Lorenz, Lotka-Volterra, cubic oscillator, Rössler) and Burgers' equation (PDE).

• Structural Recovery:
  • Naive Q-SINDy: Suffered severe degradation. For example, on the Duffing oscillator with noise, the True Positive Rate (TPR) for recovering correct terms dropped from 1.0 (vanilla) to 0.40. In some cases (Lotka-Volterra), TPR dropped to 0 even without noise.
  • Orthogonalized Q-SINDy: Consistently matched the performance of vanilla SINDy (TPR $\approx$ 1.0) across all noise levels and systems, confirming the theoretical elimination of bias.
• Ruling Out Feature Count: An RBF kernel baseline with 20 different hyperparameter configurations failed more severely than the quantum variants, proving that the issue is not merely "too many features" but a specific geometric interaction between the bases.
• Robustness: The orthogonalization method remained robust under simulated depolarizing hardware noise (up to 2% per gate).
• PDE Extension: The framework successfully extended to Burgers' equation. In this case, the $R^2_Q$ diagnostic correctly predicted low cannibalization risk (due to low quantum explanatory power), and both naive and orthogonalized methods recovered the truth.

5. Significance and Conclusion

• Methodological Rigor: The paper provides a rigorous mathematical foundation for integrating quantum features into scientific machine learning, moving beyond "black box" augmentation to geometrically principled integration.
• Interpretability: By eliminating coefficient bias, the method ensures that the recovered symbolic equations remain physically interpretable and accurate, which is critical for safety-critical applications (e.g., digital twins).
• Generalizability: While focused on quantum features, the authors note that the "cannibalization" mechanism applies to any kernel-augmented sparse regression (e.g., RBF, neural networks). The orthogonalization fix is substrate-independent.
• Future Outlook: The work suggests that while current quantum hardware is not yet faster than classical simulation for these small-scale problems, the methodological contribution (orthogonalization) is immediately applicable. It paves the way for reliable hybrid quantum-classical pipelines in scientific discovery.

In summary, Q-SINDy demonstrates that while naive quantum augmentation fails to improve equation discovery due to coefficient cannibalization, a simple orthogonalization step restores the benefits of quantum expressivity without sacrificing the accuracy of the underlying physical laws.