The name comes from a classic result: if a transmitter knows in advance what interference will corrupt its signal, it can pre-cancel that interference and lose nothing, like writing on dirty paper while knowing exactly where the smudges are.

This work applies that idea to over-the-air computation, where the wireless channel does the arithmetic itself by letting signals sum as they propagate rather than transmitting numbers and adding them afterwards. Each user's neural encoder learns to pre-cancel its own computing symbol as non-causally known interference.

The satisfying detail is what the encoders learn. Given sinusoidal activations, they recover modulo-like periodic structures consistent with lattice-based dirty paper schemes, meaning the networks arrived at the shape theory says is right without being told. That is a good sign for a learned system: it agrees with the analysis where the analysis exists.