Predictive Structure and Realized Forecast Gain in Nonlinear Fields
We measure forecast accuracy recovered by changing the readout of an unchanged representation. In a controlled known-mode Brusselator ensemble, replacing affine with quadratic readout lowers mean future-field relative L2 error by 26.83% on 720 new initial states generated after method freezing. The prespecified paired bootstrap interval is positive, conditional on three fixed encoders trained with affine decoding heads. A secondary control retains a 26.69% reduction within the same training-derived four-dimensional output space. This measures a realized deficit of the selected affine reader, rather than the code's optimal decodable risk or total retained information. Physical-component diagnostics locate most of the squared-loss reduction in uniform and remaining-field components. Under this same controlled generation law, a richer known-mode quadratic dictionary has lower mean relative L2 but less accurate critical amplitudes; prior and parameter budgets differ. In a complementary driven lattice, frozen sensor codes and predictors are evaluated on new conditional futures. Conservative lower bounds certify additional information at four ages and fixed-predictor score gains at two, although all four score means are positive. This difference describes certificate resolution at the allocated continuation budget, not an information--utility separation. The contribution is a prospective fixed-code readout intervention, common-output-space control and physical localization of the recovered accuracy.
Authors
- Rongzhen Dai (ORCID: https://orcid.org/0009-0005-7165-0080)
Publication Details
- Journal
- Research Square
- Published
- 2026-11-19
- DOI
- https://doi.org/10.21203/rs.3.rs-11566735/v1
- Primary Topic
- Model Reduction and Neural Networks
- Type
- preprint