ABSTRACT
Multigrid relaxation algorithms for discretized partial differential equations require learning steps when disorder is present. They have to determine the interpolation operators from coarse to fine grids (disordered ‘wavelets’). The matrix elements of these operators are considered as connection strengths of a neural net. Learning by backward propagation is too slow. An efficient alternative algorithm is presented. It is based on the multiscale philosophy where objects on larger scales are built from objects of smaller scales. Applications include gauge-covariant propagators in lattice gauge theory, fissures in materials, and so on.
