The interpreter in the meeting room hears the shift in tone before it is spoken aloud. They catch subtle changes in breathing, pauses, and intonations, and already know that in a few minutes the conversation will turn heated. The X-ray binary 4U 1705-44 has been having such a conversation for thirty years: its radiation flux sometimes quiets down, sometimes flares up, and until now no one could catch the moment when a calm phase gives way to a turbulent one.
Traditional analysis methods, such as the Fourier transform, are good at capturing periodic oscillations but smear sharp switches into broadband noise. The authors took a different approach: they applied the Koopman operator, a mathematical tool that turns chaotic dynamics into linear dynamics. To do this, the system's state is lifted into a space of observable functions, where evolution becomes simple multiplication by a number. Each function in this space changes at its own rate, and the slowest functions reveal how the system prepares for a transition.
The slow eigenfunction of the Koopman operator divided the phase space of 4U 1705-44 into two basins: low and high flux. The sign of this function changed days or weeks before the transition became visible in the light curve itself. This is like an interpreter who notices that the speaker has started pausing more often and lowering their voice, even though the words themselves have not changed yet. Flux forecasts on a horizon of two to eight days were most accurate near the middle of the transition, where the curvature of the curve is small, and less accurate on steep segments.
The authors also derived analytically that each eigenfunction contributes a Lorentzian peak to the power spectrum, and process noise broadens this peak and shortens the coherence time. This provides a dynamical interpretation of quasi-periodic oscillations: nearly periodic flux variations often associated with the corona, jet, or disk instabilities. Now their width is explained by mode damping due to noise, and their frequency by the imaginary part of the eigenvalue. In simulations of a Duffing oscillator with added noise, the forecast error grew proportionally to the noise variance and the trajectory curvature, confirming the theoretical formula.
The interpreter does not just hear the change in tone; they can predict when it will happen and prepare the listeners for it. Likewise, the new method allows telescopes to be pointed at the X-ray binary in advance, catching the transition across all wavelengths, from radio to gamma rays. This is the first test of Koopman theory on real X-ray data, and it opens the way to interpretable machine learning in high-energy astrophysics, where prediction and understanding arise from a single decomposition.
4U 1705-44 has been observed almost continuously for thirty years, and during that time more than fifty transitions between high and low states have been recorded, a rare long record for a chaotic system. One of the transitions predicted by the method occurred four days after the eigenfunction changed sign.