The measurement that never touched it

Suppose a large number of bombs, each with a trigger so sensitive that a single photon falling on it sets it off. Suppose that some fraction of them are duds, whose triggers are jammed and absorb nothing. The problem is to find a bomb that is definitely live, without setting it off.

Classically the problem has no solution, and the reason is worth stating in a form that makes it look inevitable. Light is delivered in indivisible lumps, so there is no such thing as illuminating the trigger a little; either a photon arrives or it does not. Testing a trigger means putting a photon on it, and a live trigger that receives a photon explodes. Any test that can distinguish live from dud must interact with the trigger, and any interaction with a live trigger destroys the bomb.

A quarter of the attempts succeed. Half destroy the bomb. The remaining quarter are inconclusive and can be repeated. And every success is a photon that established the presence of an opaque object without depositing a single quantum of energy in it.

The apparatus, and the port that receives nothing #

The instrument is two beam splitters and two mirrors, and everything depends on one property of it.

That exact zero is the instrument. A detector placed where nothing ever arrives is enormously sensitive, because anything at all that arrives there is news, and there is no background to subtract.

That is also why the instrument has to be balanced rather than merely stable: two beams interfere only if they are coherent, and an arm-length difference beyond the coherence length turns the dark port grey.

The zero comes from cancellation. There are two ways for a photon to reach the dark port — reflect then transmit, or transmit then reflect — and the amplitudes for them are equal and opposite. A beam splitter imposes a quarter-turn of phase on the reflected part relative to the transmitted part, which is not an arbitrary convention but a requirement of energy conservation: a splitter whose two outputs did not differ in phase by ninety degrees would send more light out than came in for some inputs.

When two waves meet they add, reinforcing along some directions and cancelling along others. That is all an interferometer does — and the port that receives nothing is the direction along which the two paths cancel. What makes the device useful here is that the cancellation is complete: a detector at that port sees nothing at all, so anything arriving there is evidence that something has changed.

What cancels, when there is only one photon #

The whole argument would be unremarkable for a beam. It is remarkable because it holds for one photon at a time.

Each particle arrives whole and at one place, and the pattern is a property of the accumulation rather than of any arrival. That is worth having in mind throughout, because the argument below is about single particles and the interference is a statistical statement — the dark port is dark because arrivals there are absent over many runs, not because a particle was seen to be cancelled.

So the dark port is dark for a single photon, and the reason it is dark is that the photon’s amplitude reached it by two routes that cancelled. That statement requires both routes to exist. A photon that had gone down one arm and not the other would have one amplitude, not two, and would arrive at each port half the time.

Blocking one arm #

Now put something opaque in one arm. The two-route cancellation needs both routes and one has been removed, so the dark port is no longer dark. Working the amplitudes through:

• The photon is absorbed with probability one half, because it reaches the blocked arm half the time.
• Of the remaining half, the surviving amplitude meets the second beam splitter alone and is divided evenly. A quarter of all photons reach the bright port and a quarter reach the dark one.

A click at the dark port therefore means an object is in the arm. It also means the photon was not absorbed, so it did not go into the blocked arm — a photon that goes into a blocked arm is absorbed, that being what blocked means.

The object has done nothing except be there. It has not scattered a photon, absorbed one, or been touched by a field. It has removed a possibility, and the removal is detectable because the possibility was interfering with another one.

A wavefunction crossing a barrier it has not the energy for leaves amplitude on the far side, and no particle is ever seen inside. The lesson transfers directly: amplitude is present where the particle is not, and it is the amplitude that interferes. So blocking one arm removes an amplitude rather than a particle, and the interference fails even in runs where nothing was absorbed.

The exact trade #

The dark port going bright is a loss of interference, and interference is lost in exact proportion to the availability of path information.

Two features of that relation matter here. It is an equality rather than an inequality, because nothing has been discarded: the photon and its marker are in one pure state. And nothing has to be read from the marker for the interference to disappear. It is enough that the information exists somewhere — a point that is often stated as a slogan and is here a computed curve.

The opaque object is the extreme case. It marks the path perfectly, because a photon that survives has certainly gone the other way, so $D = 1$ and $V = 0$: the fringes are gone entirely, which is precisely why the dark port is receiving a quarter of the light.

This is complementarity with a number attached, and it is worth setting beside the uncertainty relation, which is often presented as the same statement and is not. The uncertainty relation is about the widths of two distributions for one system. This is about how much a second system knows, and the trade is exact rather than bounded.

What a measurement does instead of disturbing #

The reason the phrase measurement disturbs the system misleads here is that it names the wrong thing as the mechanism.

Three analysers in a row, the third aligned with the first, produce an output the first had removed entirely — and the middle one did not knock anything about. It asked a question, and asking it destroyed the answer to the earlier one. That is what a measurement does instead of disturbing: the interaction-free label means no momentum was exchanged, and it does not mean nothing happened.

What a measurement does is make one description definite at the cost of another. Sometimes that requires an interaction with the system and sometimes, as here, it does not — the interferometer’s dark port is definite about whether an obstacle is present and it obtains that definiteness from the geometry of the apparatus rather than from anything the photon did to the obstacle.

Objects that are only partly there #

Real objects are not perfectly opaque, and the arithmetic degrades smoothly.

The ratio of information to damage improves as the object becomes more transparent, and the amount of information falls faster. That is the same bargain in a different currency, and it is why the technique is used in practice for imaging fragile samples rather than for detecting bombs: a biological specimen that would be bleached by ordinary illumination can be imaged with a fraction of the dose, and the fraction is set by exactly this curve.

Asking the question in small pieces #

The quarter is not a limit of the idea. It is the price of asking the whole question in one shot.

Replace the single interferometer with a cycle that rotates the photon’s polarisation by a small angle $\pi/2N$ and then asks, gently, whether the object is present. If it is absent, the rotations accumulate and after $N$ cycles the polarisation has turned right over. If it is present, each interrogation projects the state back to where it started, and the chance of an absorption in any one cycle is $\sin^2(\pi/2N)$ — small, and getting smaller as the questions get gentler.

The mechanism has a name — the quantum Zeno effect — and its content is that frequent gentle questioning freezes a system in place. The quadratic short-time behaviour it depends on is the same one that makes a decaying state’s survival probability start flat rather than exponential, which is a general property of amplitudes and not a special feature of this apparatus. Each near-measurement projects the state back, and the probability of a transition in a short interval goes as the square of the interval, so N of them cost only 1/N of what one of them would.

What was actually built #

The thought experiment is Elitzur and Vaidman’s, from 1993, and the reason it is worth more than a paradox is that it was performed.

Kwiat and colleagues built the interferometer in 1994, with a mirror in one arm standing in for the bomb, and measured the outcome fractions. The single-shot version gave the quarter. The Zeno version, built a year later with polarising optics and a cavity, reached about 70 per cent efficiency in six cycles — short of the ideal for the ordinary reasons of loss, and unambiguously above the quarter.

The technique has since been used for what it is actually good for, which is not bombs. Interaction-free imaging of a sample that is damaged by light, quantum-secured detection of an object that must not know it has been looked at, and — the largest current use — ghost imaging and induced-coherence imaging, in which an object is illuminated at one wavelength and imaged at another because the two are entangled and the interference carries the information.

None of those is a loophole. Each is an application of the one fact this essay is about: an amplitude that is never realised as a detection can still decide what is detected.

The classical version, and what makes it different #

There is an objection to all of this that deserves a full answer rather than a clause, because the answer locates precisely where the quantum content sits — and it is not where most retellings put it.

Send an ordinary classical light beam into the balanced interferometer. All the power leaves by the bright port. Now block one arm. Half the power is absorbed by the obstacle; the surviving half meets the second splitter alone and divides evenly, so a quarter of the original power emerges from each port, including the one that was dark.

Those are the same three numbers. The whole outcome ledger of this experiment — a half, a quarter, a quarter — is reproduced exactly by a classical wave, with no quantisation anywhere in the argument. So nothing about the statistics is quantum mechanical, and any account that presents the numbers themselves as the surprise has missed the point.

What is not classical is what the numbers are about. In the classical case, the field genuinely did reach the obstacle and genuinely was absorbed by it — half of it, continuously, at the same time as the other half was going round the other way. There is no sense in which the obstacle was untouched. The energy that came out of the dark port is the remainder of a field that was partly eaten.

In the quantum case the energy arrives in indivisible units. A photon that is detected at the dark port was not partly absorbed, because there is no such thing as partly: it either deposited its whole energy in the obstacle or it deposited none. The detection at the dark port certifies that it deposited none. And the same detection certifies that the obstacle is present, because with the arm clear that detector receives nothing at all.

So the interaction-free claim rests on exactly two things, and both need saying. It rests on the indivisibility of the photon, which is the fact a single-arrival experiment establishes and which no wave theory supplies. And it rests on the exactness of the dark port’s null, which is a statement about interference and is entirely classical. Neither half alone is interesting; the combination is.

That also settles what the experiment actually has to demonstrate to be worth anything. Measuring the three fractions with a bright beam proves nothing. What has to be shown is that the apparatus is operating one photon at a time — which is why every real version of it runs with heralded single photons or with attenuated pulses and a measured second-order correlation, and why the paper’s persuasiveness is in that part rather than in the interferometry.

The choice can be made after the photon is inside #

One further modification removes the last comfortable way of describing what happens, and it has been carried out.

The natural story about the apparatus is that the photon meets the first beam splitter and, in some sense, commits: it either becomes a two-path superposition, because the apparatus is set up to interfere, or it takes one path, because the apparatus is set up to determine which. That story is not available, because the apparatus does not have to be set up until afterwards.

Remove the second beam splitter and the two paths land on two separate detectors, and each click says unambiguously which arm the photon took. Insert it and the paths recombine, the ports interfere, and no which-path statement is possible. The two configurations differ by one optical element sitting at the far end of the apparatus — well after the point where the photon would have had to decide anything.

Wheeler’s proposal was to make that insertion while the photon is in flight, using a switch fast enough that the choice is spacelike separated from the photon’s passage through the first splitter. It has been done, with an interferometer tens of metres long, an electro-optic modulator as the switch, and a random number generated after the photon entered. The results are the ordinary ones: interference when the splitter is in, which-path information when it is out, with no trace of the photon having anticipated either.

The consequence for this essay is direct. A photon that reaches the dark port is a photon whose being a superposition of two paths was, in the delayed version, not settled until after it had passed the obstacle. So the account in which the photon “went both ways because it had to interfere later” is not a description of a decision made early; nothing was decided early, and the amplitudes are not a record of anything the photon did.

The correct summary is the deflationary one, and it is worth having plainly. The apparatus, as a whole, determines the probabilities of its outcomes. Questions about what the photon did between the splitters have no answers, not because the answers are hidden but because nothing in the theory computes them and nothing in any experiment measures them. Every paradox in this area is generated by insisting on such an answer, and the delayed-choice arrangement is the cleanest demonstration that the insistence is what fails.

Where the model stops #

The photon is idealised. Everything above assumes a single mode, perfect splitters, lossless mirrors and exactly balanced arms. Every real component contributes a leak at the dark port, and the achievable extinction — thirty or forty decibels — is what limits the technique rather than any of the arithmetic here.

The object is assumed to be a perfect absorber that does not scatter. An object that scatters a photon out of the mode without absorbing it behaves like an absorber for this purpose, which is convenient; one that reflects it back into the apparatus does not, and has to be treated as part of the interferometer.

Nothing here is free. The photon that succeeded had an energy and it went somewhere; what it did not do is deposit that energy in the object. Calling the measurement interaction-free is a statement about the object’s energy budget rather than about the universe’s, and the Zeno version costs more photons, not fewer, for its higher success rate.

The which-path relation assumes a pure state. V² + D² = 1 is an equality only when nothing has been thrown away; with a mixed state it becomes an inequality, and the missing amount measures how much has leaked into an environment nobody is tracking. In practice that leak is what limits every interference experiment, and it is the same decoherence that keeps a superposition from being visible in anything large.

And the two-state idealisation hides the imaging problem. Detecting whether one object is present at one place is a single bit. Building an image means asking the question at every pixel, and the dose saving then depends on how much of the scene is opaque — which is why the technique’s practical gains are modest and real rather than the unlimited saving the single-object case suggests.

What the pictures cannot show #

The outcome figure draws percentages as bars, and a bar implies a population. Every statement in it is about one photon, and the percentages are its probabilities. Nothing in the drawing distinguishes a probability from a fraction of a beam, and the whole argument fails if it is read as the second.

Nor can any figure here show what the photon did. There is no fact of the matter about which arm it took, and every drawing of an interferometer implies two paths with something on them. The lines in the diagram are the geometry of the apparatus rather than trajectories, and the essay’s central claim — that a click at the dark port was caused by an object the photon did not visit — is only intelligible once that distinction is accepted.

Where this ladder goes next #

The rung below this one established that a measurement can produce an answer that did not exist beforehand. This one establishes something stranger and more specific: that the absence of a possibility is physical. Blocking a path a particle did not take changes what happens, because what happens was determined by a sum over paths, and removing a term from a sum changes it.

The habit worth carrying away is a reading rule. When a quantum argument turns on something not happening, look for what it was interfering with. A dark port, a forbidden transition, a suppressed decay and a stopped chemical reaction are all the same structure: two contributions that cancel, with the cancellation destroyed by anything that distinguishes them. The interesting physics is in the destruction rather than in the cancellation, because the cancellation is the quiet state and the destruction is what is observed.

It is also the structure behind a correlation that no list of instructions can produce: there too, what settles the outcome is a set of alternatives none of which is realised on its own.

What is left on this ladder is what happens when the object in the arm is itself quantum — a single atom that may or may not be in the way, which can then be put in a superposition of the two, and the interferometer’s two ports become entangled with it rather than measuring it.