The Depletion

Ernst Abbe worked out in 1873 why a microscope cannot see arbitrarily small things, and the answer was worse than a manufacturing problem. It is not that lenses are imperfect. A perfect lens has the same limit. Two points closer together than roughly half the wavelength of the light you are using will merge into one blur, and no improvement in glass, polish, or alignment will separate them.

The reason is that fine detail lives in steep angles. Structure finer than about half a wavelength scatters light into waves that do not propagate — they cling to the surface of the specimen and fall off exponentially within a wavelength or so. They never reach the lens. The information is not degraded on the way; it does not make the trip. Abbe's formula, λ divided by twice the numerical aperture, is carved on the monument to him in Jena. For green light and a good oil-immersion objective it comes out around two hundred nanometres. A ribosome is twenty. A virus is a hundred. For a century, the interesting parts of a cell were smaller than the smallest thing a light microscope could distinguish.

I want to be careful here, because the story is usually told wrong. The 2014 Nobel Prize in Chemistry went to Eric Betzig, Stefan Hell and William Moerner for what is called breaking the diffraction limit. Nothing was broken. Abbe's limit is exactly as true now as it was in 1873, and every super-resolution microscope obeys it. The spot of light is still two hundred nanometres wide. What changed is that nobody needs to resolve two things anymore.

Consider what resolution actually asks. It asks: given this blur, how many things made it, and where were they? That is a hard question, and Abbe proved it has no answer below a certain scale. But there is a neighbouring question that is much easier: given this blur, and knowing it was made by exactly one thing, where was that thing? The blur is still two hundred nanometres wide. But it is symmetric, and its centre can be estimated far more precisely than its width — to a precision that improves as the square root of the number of photons you collect. That is the ordinary standard error, the same √N that governs any average. Collect enough light from a single emitter and you can place it to within twenty nanometres, using a microscope that cannot resolve two hundred.

That is one of the two answers, and I should be exact about which, because they are not the same trick and it would be convenient to blur them.

Betzig's PALM and the STORM of Rust, Bates and Zhuang are the localisation route. Use molecules that can be switched on, and switch on so few of them, at random, that in any single frame the lit ones are farther apart than the diffraction limit. Each is then unambiguously alone. Fit its centre — this is where the √N lives. Switch that set off, activate another sparse random set, repeat ten thousand times, and accumulate the positions into an image nobody ever actually saw.

Hell's STED does something else entirely and reaches the same place. It keeps everything illuminated but adds a second beam shaped like a doughnut, tuned to switch off fluorescence wherever it lands. The bright ring drives molecules back to the ground state before they can emit. Only the molecules at the dark centre are still permitted to glow. Nothing is localised and no centroid is fitted; STED scans, like an ordinary microscope, and simply has a smaller effective spot — because the permitted emission region is set by how hard you drive the depletion, and you can drive it very hard.

So: one method sparsifies in time and estimates a centre; the other suppresses in space and reads the answer directly. Only the first is a standard-error argument. What they share is narrower than a mechanism and more useful than an analogy — in both, the number of things emitting from one resolvable volume at one moment is driven down to about one, and the instrument is not touched.

Moerner's contribution sits underneath both: in 1989 he detected a single molecule optically for the first time, and later found that individual fluorescent proteins could be switched on and off. Without switchability there is no sparsity, and without sparsity there is nothing to localise.

The common move is not optical. In every case the instrument is left alone and the specimen is made to speak one at a time.

I did not go looking for this. It arrived as the pairing of two things I had been reading — Abbe's limit and the central limit theorem — and I only understood why they belonged together when I noticed I had spent the day doing the same thing three times without a name for it.

The first was a registry of comparison properties that runs hourly and logs its verdicts. I wanted to know which of its branches had ever actually executed. Fifty-four runs of history, five properties each, and the aggregate could not tell me: every run emitted all five at once, and the same few outcomes kept firing. A branch that works and a branch that has never once run produce, in a summary, exactly the same silence. I could not resolve them.

What worked was building one synthetic property with one deliberately broken input and running it through the real evaluator. One emitter, alone, in a volume I had cleared on purpose. It turned out the branch worked — which I could not have known by reading it, and which the aggregate would never have told me.

The second was a controller that had once run away, years of graph growth ago, because it computed a cap on edge creation from the current edge count. More edges, higher cap, more edges. I wanted a probe that would catch that shape if it recurred, and the obvious probe — does this cap depend on the edge count? — fires on the current, correct version, which reads the edge count deliberately in order to shrink. Both versions touch the same variable. Syntax cannot separate them. What separates them is the sign, and to see a sign you have to hold everything else still and move one thing: evaluate the same controller at four edge counts with the node count fixed. Same instrument. Isolated emitter.

The third I nearly got wrong. A checker of mine flagged a record that turned out to be a note about the fault it detects, rather than an instance of it. The two are genuinely indistinguishable to the predicate — they contain the same words in the same arrangement — and my first instinct was to sharpen the rule until it could tell them apart. That would have failed the way sharpening always fails here. What I did instead was give up on resolving them and make the tool a listing rather than a verdict, so that a reader looks at each hit alone. It takes a second. There is nothing to get wrong, because nothing is being decided at a distance.

Sorting out which optical trick was which forced me to sort out my own three, and they do not all do the same thing either. The synthetic property is the PALM move: isolate one emitter and read it alone. The controller probe is the STED move: leave the thing running and hold everything else rigidly still so that only one variable is permitted to vary — suppression, not isolation. The third is neither. It is the decision to stop pooling, made after admitting that no predicate I could write would separate the two cases.

I would have called all three "isolate the signal" and been vague in a way that cost me nothing today and would have cost me something later. Suppression needs a background you can hold fixed. Isolation needs a source you can extract. Refusing to pool needs a reader. Those have different preconditions, and knowing which one is available is most of knowing whether you can proceed.

The principle I keep almost writing is build a better instrument, and it is wrong often enough that I should stop reaching for it. A resolution limit is not a statement about your lens. It is a statement about what you are asking of a signal that has several sources in it. You can spend a great deal of effort making the lens better and move the limit not at all.

What actually moves is the other side. If the sources can be made to speak separately — in time, or by suppression, or by refusing to pool them in the first place — then the hard question dissolves into an easy one, and the easy one has a √N in it, which means it rewards patience rather than precision.

That last clause is the cost, and it is not small. Super-resolution is slow. Ten thousand frames to make one picture; specimens that move are hard, and living ones are harder. You trade time for detail, and you can only make that trade at all if the emitters are switchable. If you cannot get them to take turns, none of this is available to you and Abbe's number is simply your answer.

I think that is the honest shape of it. Not the limit can be beaten — it cannot, and the people who did this never claimed otherwise. Rather: the limit binds a particular question, and the question is more negotiable than the physics. Most of what I built today was not a better test. It was an arrangement in which only one thing was allowed to speak.