DrChinese said:
Doesn't this imply there is no such things as a truly "free" photon? And if so, how are photons detected as CMB? Only those photons that are ultimately detected (absorbed) later could exist.
There are oscillations in a free electromagnetic field (to a good approximation a beam in the lab is such a free field).
This is what exists (i.e., is describable by states in quantum field theory) at the fundamental level. The intensity of the field determines the average frequency of detector events. That's the QED picture. Individual photons that one could talk of don't appear at all, the state is typically a thermal mixture of complicated superpositions of multiphoton states with arbitrarily many photons without any direct interpretation, and not even the mean photon number is well-defined since there is an unbounded number of very soft (low energy) photons. Photons are just building blocks to set up the theory - nothing more.
However, to obtain some kind of intuitive picture, these oscillations are somehow portioned by means of a somewhat vague concept of a photon. However, the process of portioning is not really well-defined, especially when the beam is pulsed, as in photon on demand experiment. One has a time-dependent signal in the unobserved part of the beam, which must be short time Fourier transformed within each pulse to obtain its frequency content and hence the photon content. The length of the chosen window affects the result. If a beam contains many photons, the window of the transform is not really determined by the signal; so the result is ambiguous. If a beam contains very few photons (dim laser) they would appear at random times (in the usual way of speaking) and again, it is not clear how to window it. Only if the preparation guarantees that each pulse contains exactly one photon, the beam content is predictable. But in this case losses mean that the pulse will contain less energy than prepared, which again interferes with the notion of a photon. Moreover, this is a description of the unobserved beam. If we observe it we interfere and change everything. Measuring the full state is impossible (unless one knows it in advance and can do a non-demolition measurement). What is measured is only an event count that tells very little about the state except after sufficient time averaging, assuming the beam or the pulse sequence is stationary.
None of these problems appear when working statistically with a large number of photons. This is why
sufficiently long statistics of photon events is the only thing about photons that has scientific content, i.e., is repeatable and hence testable. Therefore
all talk about single photons or photon pairs must be understood as inadequate but partially intuitive language for dealing with a complex, indivisible process happening in the electromagnetic field.
Note that the above problematic is exactly the same as what one has classically when interpreting 1-dimensional acoustic signals (phonons in place of photons) in traditional signal analysis. We can resolve music in time and frequency only because music is not arbitrary sound but structured with a fixed rhythm that allows one to recognize where one should place the windows. The musical notes correspond to prepared single phonons, a chord to prepared phonon pairs, triples, etc. But speech or acoustic noise is already very different - no sensible division into frequencies, hence phonons is possible. Instead one divides it into ''formants'' - complex entities adapted to human sound recognition. This analogy between light and sound should be kept in mind when thinking about photons.