Schrödinger PCA: You Only Need Variances for Eigenmodes

Ziming Liu, Sitian Qian, Yixuan Wang, Yuxuan Yan, Tianyi YangarXiv 2020 · OpenReview

Introduction

Principal Component Analysis (PCA) is a fundamental tool for dimensionality reduction in scientific computing, from cosmology to fluid dynamics. However, when applied to random fields—spatial data where values at different points are inherently correlated—PCA faces significant challenges. Standard PCA requires a dense grid of anchor points to capture local correlations; in undersampled regimes, it often fails to produce coherent results. Furthermore, PCA eigenmodes often lack a clear physical interpretation.

This paper introduces Schrödinger PCA, a framework that establishes a formal mathematical duality between PCA and the steady-state Schrödinger equation from quantum mechanics. By treating a random field's variance as a "potential well," this method can recover accurate eigenmodes using up to 100x fewer data points than traditional PCA.

The Schrödinger-PCA Duality

The core of the proposed method is a mapping that translates statistical properties of a random field into physical parameters of a quantum system. In this duality, PCA eigenvectors correspond to quantum eigenstates, and the field's variance acts as a potential that "traps" these modes.

Visualizing the duality: PCA eigenmodes of a random field correspond to the eigenstates of a quantum harmonic oscillator.

The formal mapping is defined as follows:

  • PCA Eigenvector ($\phi_i$) $\longleftrightarrow$ Quantum Eigenstate ($\psi_i$)
  • Field Variance ($A(x)$) $\longleftrightarrow$ Negative Potential ($-V(x)$)
  • Correlation Kernel ($\Sigma$) $\longleftrightarrow$ Inverse Mass Matrix ($\Sigma_m$)
  • PCA Eigenvalue ($\lambda_i$) $\longleftrightarrow$ Negative Eigen-energy ($-E_i$)

Instead of diagonalizing a massive covariance matrix, Schrödinger PCA estimates variance at sparse anchor points, interpolates them to create a continuous potential function, and solves the resulting Schrödinger equation using efficient partial differential equation (PDE) solvers.

Experimental Validation and Results

The researchers evaluated the method using synthetic 2D Gaussian Random Fields and Global Climate Models. The synthetic setup allowed for a direct comparison against "ground truth" fine-grid PCA.

Comparison of eigenmodes across different methods. Schrödinger PCA (SE) accurately recovers the ground truth even on coarse grids where traditional PCA (bottom rows) fails.

Key Findings

  • Sample Efficiency: Schrödinger PCA accurately recovered eigenvalues and eigenvectors using 100x fewer anchor points than standard PCA in undersampled regimes.
  • Robustness: On coarse grids where standard PCA results became decoherent, Schrödinger PCA remained stable and closely matched the fine-grid ground truth.
  • Physical Intuition: The framework explains eigenmode concentration: in quantum terms, the particle (eigenmode) is trapped in potential wells formed by high-variance regions. Conversely, if a field's variance is zero, it acts like an infinitely heavy mass, preventing the eigenmode from existing there—a property the authors note precludes "quantum tunneling" in the PCA dual.

Discussion and Future Work

The primary advantage of Schrödinger PCA is scale separation. While standard PCA needs anchor points close enough to "feel" the correlation length, Schrödinger PCA analytically integrates the correlation. This allows researchers to map the large-scale variance structure with sparse points while the underlying physics "fills in" the small-scale eigenmode structure.

Currently, the duality is based on a second-order Taylor expansion, making it highly accurate for leading eigenmodes but less precise for higher-frequency modes. Future work will explore extending this to higher-dimensional problems and more complex manifolds, such as Riemannian manifolds, where traditional PCA becomes computationally prohibitive.

Summary updated

BibTeX
@misc{liu2020schrodinger,
  author = {Ziming Liu and Sitian Qian and Yixuan Wang and Yuxuan Yan and Tianyi Yang},
  title = {{Schrödinger PCA: You Only Need Variances for Eigenmodes}},
  year = {2020},
  url = {https://openreview.net/forum?id=88dk2KtIHVZ},
}

Loading related content...

Figure