This work sits at the intersection of several fields: data assimilation for forecasting chaotic systems, Koopman operator theory, and quantum information theory. Nearby in the archive are works on extracting signals from weak data, tracking heat with a single sensor, applying physical analogies to social processes, self-repairing quantum computers, digital processing of gravitational lenses, and taming chaos with algorithms. These studies share a common goal: building computational schemes that preserve key structures of the original problem, such as positivity, causality, and invariance. This paper offers an algebraic framework in which positivity preservation is built into the construction itself, rather than imposed as a constraint.
The authors construct three levels of description. At the first, abelian level, observables are functions on the state space, and states are probability densities. At the second, nonabelian level, observables become multiplication operators on a Hilbert space, and states become density operators, as in quantum mechanics. The forecast is given by the Koopman operator acting on states by conjugation, and the analysis by a quantum effect that generalizes Bayes' formula. For computations, the system is projected onto a finite-dimensional matrix algebra via a positivity-preserving projection. All operators are approximated from data using kernel methods: a basis of eigenfunctions of a kernel integral operator constructed from the time series of observations allows the Koopman operator, the multiplication operator, and the observation effect to be represented as matrices.
Numerical experiments were conducted for two systems. For the multiscale Lorenz-96 model with 9 slow and 8 fast variables per slow variable in the chaotic regime, data assimilation on the slow variables yields forecasts with a normalized root mean square error of about 0.35 and an anomaly correlation of about 0.95 at zero lead time. As lead time increases, the error grows and the correlation drops, while the uncertainty estimate from the forecast variance tracks the actual error well. For El Niño in the CCSM4 climate model with 44,414 sea surface temperature observation points, useful forecasts of the Niño-3.4 index persist up to 12 months at an anomaly correlation threshold of 0.6. The quality is comparable to the kernel analog forecast method, but additionally provides full probability distributions rather than just point estimates.
The main difference from neighboring works is the systematic use of nonabelian operator algebras as a setting where positivity is preserved automatically under any finite-dimensional projection. Classical projection methods operating in Hilbert spaces of functions lose positivity: for example, Fourier filtering of a positive function yields oscillations with negative values. Here, embedding into an operator algebra and subsequent projection onto a matrix subalgebra guarantees that positive observables remain positive. This property is critical for physical quantities such as temperature, mass, or probability density, where a sign violation renders the forecast meaningless.
A natural direction is porting the scheme to quantum computers. Koopman operators of ergodic systems can be approximated by shallow quantum circuits, promising exponential speedup over classical algorithms. The iterative nature of the forecast-analysis cycle will require repeated interaction between the quantum computer and the classical system, and developing such hybrid protocols is an open problem. Another direction is extending to systems with model error and incomplete observations, where the algebraic structure may suggest new regularizations.
The method is applicable in weather forecasting, climate science, oceanography, and other fields where large volumes of observations need to be assimilated into nonlinear chaotic models. Positivity preservation is important for any sign-definite quantity, and the ability to handle high-dimensional data opens the way to assimilating satellite imagery and temperature fields.
Immediate next steps include testing on more complex climate models with real observational data, as well as implementing a quantum version of the algorithm on available quantum simulators. It would also be useful to investigate the behavior of the scheme in the presence of model error and observational noise.
In the field of data assimilation, the problem of preserving physical constraints (positivity, conservation laws) when passing to finite-dimensional approximations remains unsolved. This work shows that moving to nonabelian algebras provides a natural mechanism for such preservation, but the question of extending to non-ergodic or non-stationary systems remains open.
The Lorenz-96 model was proposed by Edward Lorenz as a simplified model of atmospheric circulation on a latitude circle, and its chaotic regime at F=10 has become a standard test bed for data assimilation algorithms.
The idea of representing probabilistic states as quantum operators resonates with Greg Egan's novel 'Permutation City', where consciousness is modeled as a quantum state evolving according to the laws of quantum mechanics.