The Frame Kernel Method for Multiscale Operator Learning

What the math gives to ML

The paper introduces a nested hierarchy of compactly supported kernel centers for representing functions at multiple spatial resolutions. The transferable asset is an explicit coarse-to-fine coefficient representation whose local support yields sparse evaluation and whose hierarchy exposes a multiscale decomposition of the predicted field. A neural operator can learn mappings between these coefficient sets instead of operating directly on dense grid values, potentially improving irregular-point-cloud handling and reducing computation. The most direct test is a sparse kernel-frame encoder and decoder wrapped around a small coefficient-space operator, evaluated against an FNO or U-Net on multiscale PDE data.