It's related to the double-slit experiment.
The basic version of the experiment shows that a single photon can go through both slits, and then "both" photons interfere with each other on the other side.
This was then extended to actually measure which slot it went through -- as soon as you do this, the photon stops going through both. It either goes through one or the other, and the self-interference disappears.
This was again extended to measure which slot it went through, but then erase that information
after the photon had gone through the slots. This discarding of the intermediate measurement causes the photon to retroactively go through both slots again, like in the first experiment, even though it was observed going through a single slot like in the 2nd experiment.
Further, you can rig the intermediate slot-detector to either discard the information or keep it,
after the photon has gone through the slots, and that later decision will change whether it goes through one or the other with 50/50 probability (like in the 2nd), or whether it goes through both and undergoes self-interference on the other side (like in the 1st).
Basically, the path taken by the photon is lazily evaluated. It only actually "decides" which path to take when the universe needs to "know". Until then, it's in a superposition of all possible paths (
including going through one, then the other, then orbiting Jupiter, then going through one again, then hitting the screen).
AFAIK, there are ways to explain this phenomenon that don't violate causality -- which means that the present hasn't actually changed the past -- but I'm no physicist.
[edit]Here's the specific experiment referenced above:
http://news.sciencem...7/02/16-04.html
