Chapter 4 established that fringes die when a which-path record exists. Two famous follow-ups push on that. What if you decide whether to make the record after the particle has passed the slits? And what if you make the record, then destroy it? These experiments have grown a folklore of time-travel talk. The experiments are real and wonderful; the folklore is wrong, and this page shows both.
John Wheeler, 1978: let the photon pass the slits first, then choose the apparatus. Leave the screen in place (fringes: "it went both ways") or snatch it away, revealing two telescopes aimed one at each slit (a click in one: "it went one way"). Whatever the photon "did" at the slits, your later choice seems to decide it retroactively. Done for real in 2007, with the choice made by a quantum random-number generator while the photon was in flight, the statistics came out exactly as quantum mechanics predicts: fringes when you look late for waves, path counts when you look late for particles.
Wheeler's own conclusion is the sober one: no photon is a wave or a particle while in flight. The mistake is assuming it was ever carrying one of the two labels for your choice to overwrite. Nothing travels backwards except your bookkeeping.
Tag the slits with orthogonal polarisations. That is a perfect which-path record, so no fringes. Now put an analyser in front of the detector at angle θ. Behind it, an H-tagged and a V-tagged path can end up in the same state. The record is erased, and fringes return. The catch that kills the magic: the analyser has two output ports, and what one port gains as fringes, the other gains as anti-fringes. Add them back together and the screen total is as fringeless as ever.
Kim et al. (2000) made the erasure decision after the screen photon had already landed: each screen photon has an entangled idler partner that reaches its own apparatus later, where it either reveals the path or erases it. The pop-science telling ("the choice reaches back and repaints the screen") is exactly what the data forbid. The screen, taken alone, never shows fringes. Fringes appear only when you use the idler outcomes to sort the screen dots into subsets, and the fringed subsets cancel to the same smooth band you already had.
The simulation below is that experiment. Dots land drawn from the true joint distribution (θ = 45°): the screen marginal is smooth. Every dot secretly carries its idler outcome. Press sort to reveal what was there all along.
Slide θ: every angle that buys fringes in one port buys equal anti-fringes in the other. The grey sum curve never moves. Erasure redistributes; it cannot create.
Why it must cancelErase before the screen, after the screen, next Tuesday: the joint statistics are the same. Records don't care when you read them, only whether they exist.
Kim's actual numbersThe unsorted screen is identical whatever is done to the idlers. A friend at the idler station can signal you nothing. You both need the coincidence records, which travel forwards, by post.
The no-signalling rule