In 1978, John Archibald Wheeler proposes a thought experiment. A photon is sent toward a Mach-Zehnder interferometer. At the entrance, a beam splitter divides the beam into two possible paths. At the exit, a second beam splitter recombines them. If the second beam splitter is in place, the two paths interfere. The photon is described as having traversed both paths. If the second beam splitter is removed, one detector clicks, not the other. The photon is described as having traversed one path.
Wheeler proposes to decide whether to remove or insert the second beam splitter after the photon has already passed through the first.
Jacques et al. (2007) carry out the experiment at the École Normale Supérieure. A quantum random number generator decides, after the photon has entered the interferometer, whether the measurement will be wave-like or particle-like. The results confirm the prediction of quantum mechanics. When the second beam splitter is inserted late, interference appears. When it is removed late, interference disappears. The statistics conform to the measurement that is eventually performed.
Kim et al. (2000) push further with the delayed-choice quantum eraser. Two entangled photons are produced. One is detected immediately. The other is sent to a device that can, after the detection of the first, erase or preserve the which-path information. When the information is erased, the correlations between the two photons show an interference pattern. When it is preserved, the pattern disappears. The pattern is recovered only by sorting the already-recorded detections into subsets defined by the later measurement.
This sorting is the key point. The interference pattern never appears in the total raw data of the early detector. It appears only in the post-selected subsets, once the outcome of the later measurement is known and used to partition the earlier record. No signal travels backward. No information can be sent. The early detector registers what it registers, independently of what happens later. What changes, after the fact, is which subensemble one is looking at.
Standard quantum mechanics accounts for the full set of results without invoking any retrocausal mechanism. The wave function does not collapse in the past. The correlations were already present in the entangled pair at preparation. The later choice determines which statistical structure within those correlations becomes visible.
Doctrine
The experiment does not show that the past changes. It shows that the description of what happened at $t_1$ is not complete until one specifies what is measured at $t_2$. Two different late measurements partition the same early record into two different sets of subensembles. Each subensemble has its own statistics. None of the statistics existed in the raw data alone.
Retrocausal language describes the phenomenon economically. It does not describe the mechanism.
Open vector
Quantum mechanics forbids sending signals to the past, and it also forbids assigning the photon a single definite trajectory before the measurement basis is fixed. These are two different constraints. The first is a no-go theorem about information transfer. The second is a statement about what the formalism will and will not let one say about the intermediate history.
The question raised by delayed-choice experiments is not whether the past is fixed. It is which questions about the past have definite answers before the later measurement is specified, and which do not. The line between the two is not a frontier of physics. It is a property of the formalism. Whether the formalism is complete is a separate question, and an older one.
