The question of correlation between photons in coherent light rays. E. M. Purcell in Nature 178:1449 (1956). What the paper says!?
This is an important paper as the first from a high authority of physics (Edwin Purcell) to side with Hanbury Brown and Twiss in their report of the eponymous HBT effect. It attributes it to the tendency of bosons to "clump", and also predicts the opposite behaviour (antibunching) for fermions. The derivation is based on intensity fluctuations, but are repeatedly cast in bosonic language, invoking Pauli exclusion principle as its counterpart (which has no classical equivalent).
His first paragraph is already a masterpiece of scientific literature:
Brannen and Ferguson¹ have reported experimental results which they believe to be incompatible with the observation by Hanbury Brown and Twiss² of correlation in the fluctuations of two photoelectric currents evoked by coherent beams of light. Brannen and Ferguson suggest that the existence of such a correlation would call for a revision of quantum theory. It is the purpose of this communication to show that the results of the two investigations are not in conflict, the upper limit set by Brannen and Ferguson being in fact vastly greater than the effect to be expected under the conditions of their experiment. Moreover, the Brown–Twiss effect, far from requiring a revision of quantum mechanics, is an instructive illustration of its elementary principles. There is nothing in the argument below that is not implicit in the discussion of Brown and Twiss, but perhaps I may clarify matters by taking a different approach.
then
Now it will be found, even with the steadiest source possible, that the fluctuations in the counting-rate are slightly greater than one would expect in a random sequence of independent events occurring at the same average rate. There is a tendency for the counts to ‘clump’. From the quantum point of view this is not surprising. It is typical of fluctuations in a system of bosons.
Purcell relates the fluctuations of the intensity $P$ («It makes no difference whether we think of $P$ as the square of an electric field-strength or as a photon probability density» to the spectrum through the Siegert relation and the Photoelectric effect. His first main result is then:
we look at the number of counts $n_T$ in a fixed interval $T$, and at the fluctuations in $n_T$ over a sequence of such intervals. From the above relations, the following is readily derived:
I provide here details of this "easy" derivation, in modern notations. Then $I(t)\equiv P(t)=|E(t)|^2$ (Purcell's $P$ is the intensity, written $I$ below), so that $\overline{P(t)P(t+\tau)}$, divided by $\overline{P}^2$, is $g^{(2)}$$(t,\tau)$ while $\rho$ is the first-order correlation function $g^{(1)}(\tau)$ and Eq. (1) reads $g^{(2)}(\tau)=1+|g^{(1)}(\tau)|^{2}$, which is Siegert's relation, which he attributes to Lawson et al.[1]. It holds for chaotic (Gaussian) light, which is what a thermal source gives; the light is also taken as stationary, and the average over the light is identified with Purcell's average over a sequence of intervals, which is the ergodic hypothesis.
The derivation is basically a doubly stochastic Poisson process, admixing a Poisson process (from the photoelectric effect: the instantaneous value of the electric field intensity $I(t)$ results in a probability of photon detection as $\alpha I(t)dt$), and the fluctuation of $I(t)$ itself with $t$. Therefore, the instantaneous probability for two-photon detection at $t$ and $t'$ is, assuming no memory: $$\alpha^{2}I(t)I(t')\,dt\,dt'$$ the average of which, and also using the definition of $g^{(2)}\equiv\langle I(t)I(t+\tau)\rangle/\langle I\rangle^2$, gives \begin{equation} \alpha^{2}\langle I(t)I(t')\rangle\,dt\,dt' =\alpha^{2}\langle I\rangle^{2}\,g^{(2)}(t-t')\,dt\,dt'\,. \label{eq:0URMz} \end{equation}
Now, the quantity of interest is $n_T$ the number of photons detected in a time window $T$ which, still from the photodetection formula and stationarity of the field, reads: $$\langle n_T\rangle=\alpha\langle I\rangle\,T\,,$$ with $\langle I\rangle$ the mean intensity, while the intensity correlations read $$ \int_0^T\!\!\int_0^T\alpha^{2}\langle I(t)I(t')\rangle\,dt\,dt' =\langle n_T(n_T-1)\rangle\,, $$ in term of the photo-detection correlations, since $n_T$ is discrete (integer valued) while $I$ is continuous, and exclude correlations with itself, i.e., we are counting ordered pairs. A more explicit way to show that is to write $n_T=\sum_{k=1}^N b_k$ the sum of $b_k=0$ or $1$ count in each small interval out of $N$ partitioning the interval $T$, in which case $n_T^{2}=\sum_{k}b_k^{2}+\sum_{k\neq l}b_kb_l=n_T+\sum_{k\neq l}b_kb_l$ since $b_k^2=b_k$ for this idempotent (0, 1) variable.
As a result, $$\langle n_T(n_T-1)\rangle=\alpha^{2}\langle I\rangle^{2}\int_0^T\!\!\int_0^T g^{(2)}(t-t')\,dt\,dt'\,.$$
Now, from $\langle n_T^{2}\rangle=\langle n_T(n_T-1)\rangle+\langle n_T\rangle$, and using the Siegert's relation (Purcell's Eq. (1)), we arrive to: \begin{equation} \langle n_T^{2}\rangle-\langle n_T\rangle^{2} =\underbrace{\langle n_T\rangle}_{\text{normal}} +\underbrace{\alpha^{2}\langle I\rangle^{2} \int_0^T\!\!\int_0^T\bigl|g^{(1)}(t-t')\bigr|^{2}\,dt\,dt'}_{\text{extra}}\,. \label{eq:0UuCB} \end{equation} where the "normal" term is the shot noise (Poisson fluctuations), from «a random sequence of independent events occurring at the same average rate», and the "extra" one originates from $|g^{(1)}|^2$, i.e., from the noise of the light with a correlation time. The extra term is, in modern language, the intensity noise of the light, or the variance of the number of counts $\alpha\int_0^TI\,dt$.
The double integral, upon substituting $\tau\equiv t-t'$ in the windows $0\le t\le T$ and $0\le t'\le T$, i.e., $0\le t'+\tau\le T$, and since for any function $f$, \begin{equation} \int_0^T\!\!\int_0^Tf(t-t')\,dt\,dt' =\int_{-T}^{T}\bigl(T-|\tau|\bigr)\,f(\tau)\,d\tau\,, \label{eq:0UFX1} \end{equation} then for $f=|g^{(1)}|^2$, \begin{equation}\label{eq:0UyR0}\langle n_T^{2}\rangle-\langle n_T\rangle^{2}=\langle n_T\rangle\left(1+\alpha\langle I\rangle\int_{-T}^{T}\Bigl(1-{|\tau|\over T}\Bigr)\,|g^{(1)}(\tau)|^2\,d\tau\right)\end{equation} since $\alpha\langle I\rangle T=\langle n_T\rangle$. Assuming $|g^{(1)}(\tau)|^2$ is nonzero on a small window as compared to $T$, so that $|\tau|/T\approx 0$ and for large values of $T$, the integrand is zero because $|g^{(1)}|^2\approx 0$, so that it can be approximated to $\int_{-\infty}^{+\infty}|g^{(1)}(\tau)|^2\,d\tau=\tau_0$, which is Eq. (2) of Purcell, reproduced above in Fig. (0U8IU).
Not entirely the «readily derived» result, especially as some details have been left out (e.g., validity of the integral approximation, which overestimates the excess at any finite $T$).
Then:
τ₀, a correlation time determined by the light spectrum, is approximately the reciprocal of the spectral bandwidth Δν
This makes the fractional excess of the variance over the normal amount, $\alpha\langle I\rangle\tau_0$, «independent of T, and is about equal to the number of counts expected in an interval 1/Δν.»
Although Purcell doesn't discuss the opposite limit $T\ll\tau_0$, it is instructive to do so: then $|g^{(1)}|^{2}\simeq1$ over the whole double integral, which becomes $T^{2}$, and the extra term thus reads $\alpha^{2}\langle I\rangle^{2}T^{2}=\langle n_T\rangle^{2}$. The intensity does not change during the interval, so the interval carries the whole fluctuation of the light, $\langle I^{2}\rangle-\langle I\rangle^{2}=\langle I\rangle^{2}$ (chaotic light, $g^{(2)}(0)=2$), which gives the extra term $\langle n_T\rangle^{2}$. The normal term $\langle n_T\rangle$ is separate: it comes from the counts being discrete and is there for any light. Expressed as an intensity, by dividing by $(\alpha T)^{2}$, the variance above reads $\operatorname{Var}(n_T/\alpha T)=\langle I\rangle/(\alpha T)+\langle I\rangle^{2}$: the first term is the counting noise, which decreases as $1/(\alpha T)$ and is not part of the variance of $I$, the second is $\langle I^{2}\rangle-\langle I\rangle^{2}$. For $T\gg\tau_0$ it is $\langle n_T\rangle\,\alpha\langle I\rangle\tau_0=\langle n_T\rangle^{2}\,\tau_0/T$, a factor $\tau_0/T$ smaller: \begin{equation} \text{extra}=\langle n_T\rangle^{2}\times \begin{cases}1 & T\ll\tau_0\,,\\[0.5ex] \tau_0/T & T\gg\tau_0\,.\end{cases} \label{eq:0UWRy} \end{equation} The interval contains about $T/\tau_0$ stretches of light independent of one another, and averaging over them divides the variance by that number. $\tau_0$ is therefore the duration over which the intensity stays correlated, and the crossover between the two behaviours is at $T\sim\tau_0$.
He also already observes that the effect does not depend on the detection efficiency.
After this purely stochastic explanation, he then proceeds:
If one insists on representing photons by wave packets and demands an explanation in those terms of the extra fluctuation, such an explanation can be given. But I shall have to use language which ought, as a rule, to be used warily.
And his explanation is then the closest I know to understand it as a manifestation of Simpson's paradox, although he maintains (somehow reluctantly) the quantum interpretation:[2]
Think, then, of a stream of wave packets, each about c/Δν long, in a random sequence. There is a certain probability that two such trains accidentally overlap. When this occurs they interfere and one may find (to speak rather loosely) four photons, or none, or something in between as a result. It is proper to speak of interference in this situation because the conditions of the experiment are just such as will ensure that these photons are in the same quantum state. To such interference one may ascribe the ‘abnormal’ density fluctuations in any assemblage of bosons.
At this point, he "predicts" the opposite for fermions. It is interesting that for Purcell, a bosonic wavefunction can represent a classical field:
Nor would we be entitled in that case to treat the wave function as a classical field.
The case in question being fermions.
He then turns to the HBT setup proper, also adding polarization. Each branch, $n_1$ and $n_2$ behaves similarly as the previous case, and he then considers the sum $n=n_1+n_2$ and arrives to (he silently assumes $\bar n_1=\bar n_2$):
which, he comments, «This is the positive cross-correlation effect of Brown and Twiss, although they express it in a slightly different way.» He relates it again to "clumping of photons":
It is merely another consequence of the ‘clumping’ of the photons.
And even more forcefully:
As usual in fluctuation phenomena, the behaviour of fermions and the behaviour of bosons deviate in opposite directions from that of classical particles. The Brown–Twiss effect is thus, from a particle point of view, a characteristic quantum effect.
Note that he takes the precaution to write «from a particle point of view» (cf. the above comment on being entitled to treat the wavefunction as classical).
The effect as described is very small, since this excess is the photon degeneracy parameter, or the mean number of photons per mode. The failure of Brannen and Ferguson is then explained by Purcell from those numbers:
| Hanbury Brown and Twiss | Brannen and Ferguson | |
|---|---|---|
| Measured quantity | correlation between the fluctuations of the two photocurrents, amplified over 3–27 MHz and multiplied in a linear mixer (analog) | excess of the coincidence rate between two channels over the accidental rate, counting single photoelectrons |
| Source | mercury arc, 4358 Å line, source aperture 0.13 mm × 0.15 mm, cathodes 2.65 m away (apertures 9.0 mm × 8.5 mm) | not specified; the 4358 Å line is assumed by Purcell |
| Optical bandwidth, | not stated | not specified by the authors; Purcell takes Hz for illustration, so s |
| Counting rate or current | photomultiplier gain , output current 140 µA, quantum efficiency about 15% | single-channel rate s−1, accidental coincidences about 20 s−1 |
| Primary photocurrent | about times that of B&F (Purcell) | 1 (reference) |
| Photoelectrons per second per Hz of optical bandwidth (derived) | about ( quanta s−1 Hz−1) | (with Purcell's assumed bandwidth) |
| Resolving or integration time | 4 runs of 90 min | resolving time about s |
| Expected effect | correlation over r.m.s. deviation to (theory) | fractional change in coincidence rate , with s |
| Observed | to in the four runs; about 0 with the cathodes 1.8 cm apart | statistical uncertainty of the coincidence rate about 0.01, twenty times the expected effect |
The HB values are taken from Hanbury Brown and Twiss[3]; the B&F values are those quoted by Purcell, whose own assumptions are marked. Rows marked derived are arithmetic on those numbers.
And Purcell to conclude:
counting periods of the order of years would be needed to demonstrate the effect with the apparatus of Brannen and Ferguson. This only adds lustre to the notable achievement of Brown and Twiss.