A translator records a multi-hour conversation. If they try to hold the entire dialogue in their head, they're bound to garble details by the end. That's roughly how gravitational observatories work: they listen to the 'conversation' of massive bodies spiraling together. The future space-based detector LISA (a laser antenna in space) will catch gravitational waves, ripples in spacetime, from such events for months or even years. The signal changes constantly, and the longer it lasts, the harder it is to decode accurately.

Standard methods require a single mathematical model for the entire observation window. But if the model deviates even slightly from reality, errors snowball. The result can be wrong estimates of object masses or distances. The authors propose a different approach: cut the long recording into short, independent chunks, like breaking a multi-hour conversation into individual lines. Each chunk is analyzed separately, then results are combined without trying to 'stitch' the signal phase between chunks.

The method is implemented in a program called SPLIT (Semi-coherent Posteriors for Long-Inspiral Templates), running on graphics cards for speed. Tests on a model signal mimicking environmental effects showed that standard analysis deviated from the truth by nearly five standard errors, while the new method stayed within half a standard error. It's like comparing shooting with your eyes closed to aiming with a clear sight.

A translator who breaks the conversation into short phrases no longer gets tangled in details, and the conversation sounds clear from start to finish.