Periods, prequantization, and rigidity in relative multisymplectic geometry

What the math gives to ML

The paper's most transferable construction is the mapping-cone treatment of a target object together with a source-side trivialization or correction. It gives an exact algebraic residual, F*omega - d eta, rather than an ad hoc penalty: the target form must be closed, while its pullback must be exact on the source. This suggests a bulk-boundary or cross-domain neural regularizer for models with a known map between two representation spaces, especially neural operators, geometric encoders, and models with interface data. The available extraction does not expose the paper's explicit period and rigidity formulas, so the most defensible experiment uses the concrete mapping-cone differential and its induced conservation identity.