Each photon has a twin that records which slit it used, so the screen shows no stripes. Measure the twin in a way that scrambles that record and the stripes come back, but only if you know where to look.
Look at the top panel: no stripes, whatever you do to the twin. Now turn on the delay, let grey dots pile up, pick an angle, and press Measure stored twins. The dots that are already on the screen get sorted into stripes and gaps.
A crystal splits each laser photon into two twins. One goes through a double slit to the screen. In this version each slit marks its photon so that the twin ends up horizontally polarized (H) for the top slit and vertically polarized (V) for the bottom one. The twin now holds a which-path record, so the screen shows a plain blob.
The twin then meets a polarizer that sends it to detector D1 or D2. At 0° it reads the record: D1 means top slit, D2 means bottom. At 45° each outcome is equally likely from either slit, so the record is erased. A sorter matches every screen hit with its twin's detector.
It looks as if a choice made later reaches back and decides whether earlier photons acted as waves. It doesn't. The top panel never changes: no setting, early or late, puts stripes on the bare screen. Stripes only appear after you sort the hits using the twin's result, and that list has to be carried over by ordinary means, no faster than light.
What is truly strange is the bookkeeping. The stripes and the anti-stripes were hiding inside the blob all along, and which way you are able to split it depends on how you question the twin.
Proposed by Scully and Drühl in 1982. Done with a delayed choice by Kim, Yu, Kulik, Shih and Scully (published 2000), and with polarization tags by Walborn and colleagues in 2002.
The eraser is a clean lesson in one rule: interference needs the paths to be indistinguishable, and correlations only show up when you compare records. Engineers use both halves of that rule.
In the BB84 protocol (1984) each photon is measured in one of two polarization settings 45° apart. Measured the matching way, it gives a definite bit; measured the other way, it gives a random one. An eavesdropper can't know which setting to use, so tapping the line leaves detectable errors. Commercial systems already guard some bank and government fibre links.
In the demo: 0° and 45° are those two settings; one reveals the path, the other makes it random.Aircraft, ships and spacecraft sense rotation by sending light both ways round a fibre coil and watching the interference shift. If the two beams' polarizations drift apart they become distinguishable and the fringes fade, so these gyros use polarization-maintaining fibre and polarizers to keep them identical.
In the demo: H and V tags make the slits distinguishable and the stripes vanish.In 2014 a Vienna team imaged a cat-shaped stencil using twin photons where the photons that passed the stencil were never detected. The image appeared only in their partners, through interference that depended on the twins being indistinguishable. It lets cheap cameras see objects lit at wavelengths those cameras can't detect.
In the demo: the twin's detector decides which pattern the screen hits belong to.The same crystal trick, spontaneous parametric down-conversion, is the workhorse photon-pair source in quantum optics labs. Detecting one twin announces that the other exists, giving single photons on demand for experiments, quantum networks and photonic quantum computers.
In the demo: every screen hit is paired with a twin; the sorter relies on it.Quantum algorithms work by making computational paths interfere. Any scratch qubit left holding a record of which path was taken kills that interference, exactly like the tag here. So algorithms routinely run steps backwards to erase scratch records before the final measurement.
In the demo: a leftover record means no stripes until it is erased.In 2017 an Italian team ran Wheeler's delayed-choice test with photons bounced off satellites thousands of kilometres up, choosing how to measure each photon only while it was already on its way. Quantum predictions held, with no sign of any influence running backwards in time.
In the demo: turn on the delay; the totals are the same either way.