How Spectral Bias in Machine-Learning Forecast Emulators Propagates Through the Data-Assimilation Cycle: A Linear–Gaussian Spectral Theory

Machine-learning forecast emulators systematically under-represent small-scale variance, a spectral bias documented across architectures. Its effect inside a data-assimilation cycle has been conjectured rather than derived. Here, a closed-form theory is developed in the simplest setting that retains the following mechanism: a linear–Gaussian cycle on a periodic domain with statistically homogeneous errors, in which the covariance recursion separates into independent scalar recursions per wavenumber k. The emulator is modelled as the true propagator composed with a spectral damping d(k) ∈ [0,1]. Three results are proved. First, the filter is overconfident at damped scales by a closed-form gap that grows monotonically with the damping. Second, below an exact damping threshold, the Kalman gain falls beneath a prescribed tolerance, provided that the tolerance exceeds the gain floor set by the process noise, so that observations are assimilated with a weight below that tolerance; at weakly forced, locally growing modes, the actual error diverges. Third, without injected model-error variance, ensemble spread collapses geometrically at d(k)2m(k)2 per cycle wherever d(k) m(k) < 1, against m(k)2 for a perfect-model ensemble at decaying modes. Phase and scale-transfer errors and inflation are quantified as departures from this baseline, and localization is examined numerically. The closed forms of Theorems 1–4 are confirmed by computer algebra and reproduced in 50-digit arithmetic to better than 10−42, and the full matrix recursions agree with the closed forms to better than 10−10. Within the linear theory, the gain-miscalibration formulae bound the error of a fully emulator-driven cycle from below at decaying modes. A Lorenz-96 experiment shows that nonlinearity carries the degradation to undamped scales, where no such bound holds.

Authors

Institutions

Publication Details

Journal
Entropy
Published
2026-11-19
DOI
https://doi.org/10.3390/e28111276
Primary Topic
Meteorological Phenomena and Simulations
Type
article
Field-Weighted Citation Impact
0.00
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

How Spectral Bias in Machine-Learning Forecast Emulators Propagates Through the Data-Assimilation Cycle: A Linear–Gaussian Spectral Theory

Dimitris G. Kaskaoutis, Mohamed Zhran, Nitin Lohan, Fahdah Falah Ben Hasher et al.
Entropy
Meteorological Phenomena and Simulations
article

How Spectral Bias in Machine-Learning Forecast Emulators Propagates Through the Data-Assimilation Cycle: A Linear–Gaussian Spectral Theory

Dimitris G. Kaskaoutis, Mohamed Zhran, Nitin Lohan, Fahdah Falah Ben Hasher, R. K. Giri, Pankaj Goel, Sushil Kumar
article en

Abstract

Machine-learning forecast emulators systematically under-represent small-scale variance, a spectral bias documented across architectures. Its effect inside a data-assimilation cycle has been conjectured rather than derived. Here, a closed-form theory is developed in the simplest setting that retains the following mechanism: a linear–Gaussian cycle on a periodic domain with statistically homogeneous errors, in which the covariance recursion separates into independent scalar recursions per wavenumber k. The emulator is modelled as the true propagator composed with a spectral damping d(k) ∈ [0,1]. Three results are proved. First, the filter is overconfident at damped scales by a closed-form gap that grows monotonically with the damping. Second, below an exact damping threshold, the Kalman gain falls beneath a prescribed tolerance, provided that the tolerance exceeds the gain floor set by the process noise, so that observations are assimilated with a weight below that tolerance; at weakly forced, locally growing modes, the actual error diverges. Third, without injected model-error variance, ensemble spread collapses geometrically at d(k)2m(k)2 per cycle wherever d(k) m(k) < 1, against m(k)2 for a perfect-model ensemble at decaying modes. Phase and scale-transfer errors and inflation are quantified as departures from this baseline, and localization is examined numerically. The closed forms of Theorems 1–4 are confirmed by computer algebra and reproduced in 50-digit arithmetic to better than 10−42, and the full matrix recursions agree with the closed forms to better than 10−10. Within the linear theory, the gain-miscalibration formulae bound the error of a fully emulator-driven cycle from below at decaying modes. A Lorenz-96 experiment shows that nonlinearity carries the degradation to undamped scales, where no such bound holds.

EntropyVol. 28(11)
Princess Nourah bint Abdulrahman University (SA), Cyprus Institute (CY), Mansoura University (EG), India Meteorological Department (IN), Ministry of Earth Sciences (IN), National Observatory of Athens (GR), G.L. Bajaj Institute of Technology and Management Greater Noida (IN), Gautam Buddha University (IN)
Meteorological Phenomena and Simulations
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon