Reproduce “Operator learning with PCA-Net: upper and lower complexity bounds” on OpenResearch

About this paper

Samuel Lanthaler

This work establishes a comprehensive approximation theory for PCA-Net, providing both upper and lower complexity bounds to quantify its efficiency in learning operators between function spaces. It rigorously proves a "curse of parametric complexity" for general operators while demonstrating algebraic convergence rates for specific partial differential equations like Darcy flow and Navier-Stokes.

Key insights

• The "curse of parametric complexity" rigorously proves that algebraic convergence rates are unattainable for broad classes of general operators when learning mappings between infinite-dimensional spaces.
• The efficiency of PCA-Net is directly limited by the inherent complexity of the output distribution, specifically the decay rate of the PCA eigenvalues of the push-forward measure.
• Operator properties like analytic regularity (for Darcy flow) or the ability to efficiently emulate numerical methods (for Navier-Stokes) are crucial for PCA-Net to achieve efficient, algebraic approximation rates, effectively overcoming the general curse.

Summary by alphaXiv.