What kind of thing is a number?

Some numbers you can count your way to. Some you can only measure. Telling which is which turns out to be most of the work — and getting it wrong is most of the mistakes.

This is a plain-language tour of a working guide to mathematics and physics — not the guide itself, which is technical, but the handful of ideas in it that anyone can use. It was first written in the middle of September and has been rewritten at the end of the month, because two more weeks of putting the guide to work changed what looked important. The first seven sections are what survived from the first version. The rest is what the two weeks added, most of it learned by being wrong in a way the guide could catch.

1 · Three kinds of number, and a test that tells them apart

Certain numbers show up again and again in systems that have nothing else in common. It is tempting to treat them all as the same kind of discovery. They are not, and there are three kinds.

Counts. Two hundred and forty spheres can touch a central one in eight dimensions. There are three colours of quark. Nuclei are unusually stable at 2, 8, 20, 28, 50, 82 and 126 particles. These come from counting something, and they are either exactly right or catastrophically wrong. You cannot miscount 240 by three percent.
Exponents from unit-juggling. The 5/3 in the turbulence spectrum, the 3/4 in the metabolic law, the half-powers in ship design. These look like measurements because they are fractions. They are not: they fall out of a table of units with whole numbers in it, and such a table always yields exact fractions.
Critical exponents. The numbers governing a magnet right at the temperature where it loses its magnetism. The same number appears in systems sharing nothing but a symmetry and a dimension — which makes them look like the first two. They are real numbers computed by an optimisation, and they arrive with error bars.

The third kind had been filed with the first two for a long time, and the tell is the error signature. A count has no spread at all. A critical exponent has a spread on both sides — the calculation has an uncertainty and so does the experiment — and occasionally they disagree.

There is a live case. One such exponent was measured aboard the Space Shuttle, because on Earth gravity smears the transition across the sample. It disagrees with the best calculation by about six standard deviations and nobody knows why.

Under the old two-way filing that looked like a paradox: a structural number disagreeing with reality. Under the corrected filing it is an ordinary measurement discrepancy. That removes the mystery about its kind and leaves the six sigmas exactly where they were. Reclassifying something does not explain it.

And a warning that comes with the second kind. An exponent from unit-juggling is exactly rational no matter what — but only once you have decided which quantities belong in the table. That decision is physics, not arithmetic. The famous 5/3 of turbulence is exactly as rational as any other, and rests on an assumption that turned out to be false. Exact does not mean correct.

2 · When you cannot get the number: four ways to fence it in

Some numbers have no route to them at all. The strength of electromagnetism, about 1/137, is the famous one: no argument anywhere derives it. For a long time the guide treated that as the end of the conversation.

It is not. You cannot compute such a number, but you can often fence it in — prove it must lie between two values — and there turn out to be exactly four ways to build a fence.

fence	the idea	an example
positivity	some quantity is not allowed to be negative; push against that	the exact limit on how strongly two distant particles can be correlated
counting	there simply are not enough things to go round	almost every possible computation needs an enormous circuit, because there are more computations than small circuits
conservation	something is conserved, so no process can exceed it	no engine beats the Carnot efficiency; no lens focuses sunlight hotter than the sun
selection	condition on the fact that we are here to ask	the only successful prediction ever made of the cosmic vacuum energy

The positivity fence is the strongest and it is startlingly cheap. The exact ceiling on quantum correlations — a number a great deal of physics rests on — comes out of a computation with two constraints in it and no physics at all: require a certain table of numbers to be non-negative in a technical sense, then maximise. The answer lands on 2.8284271, correct to ten decimal places, with no mention of particles, states or measurements anywhere in the calculation.

A fifth candidate was proposed and rejected, and the reason is worth knowing. There is a classic move that turns an exact equality into an inequality by applying a curved function to it. It works beautifully — but it needs an exact equality to start from, and the whole point of this list is numbers where nothing exact is available. It converts; it does not generate.

That rejection happened because of a one-line habit: before using a list, write down what it is a list of. Without that line the fifth fence would have gone in, and the list would have contained something that cannot address the situation the list exists for.

And the fences are not equal. For the three hardest numbers in physics — the strength of electromagnetism, the energy of empty space, and the puzzle of why the Higgs particle is as light as it is — the only fence currently working is selection, which is the weakest of the four and the one most often disputed. That is a sharper thing to say than "we don't know", because it says what progress would look like: a positivity or counting argument, not a better anthropic one.

3 · The one theorem: count it twice

Strip away the vocabulary and an enormous amount of deep mathematics is one move. Take a single quantity, evaluate it two completely different ways, and set the answers equal. One side is usually easy and uninformative; the other is hard and full of content; and the equation between them does the work.

The version that matters most in mathematics equates a list of frequencies with a geometric measurement — the sound a shape makes, against the shape itself. It has been rediscovered four separate times under four different names.

But why you are allowed to count twice is not always the same reason, and there turn out to be at least three licences:

licence	what it needs
swapping the order of a double count	things must not add up to infinity — and this is free if nothing is negative
an edge matching an inside	a smooth space with a sensible boundary — a structural condition, not an arithmetic one
two sides of an optimisation meeting	the problem must be convex

That first row explains something that had looked like luck. There is a law in queueing theory — the average number of people waiting equals how fast they arrive times how long they stay — which is famous for needing no assumptions at all. No assumption about arrivals, service, ordering, or number of servers. The reason is that it counts a region two ways and the thing being counted is never negative, so the licence is free. And it is the same theorem as the frequencies-and-shapes one, with the queue's customer index playing the role of the frequency index.

4 · The hypotheses nobody states

Almost every mistake found in building this guide had one shape: a tool was used whose fine print nobody had read. Not an arithmetic slip — a theorem applied where it does not apply.

A method for converting between two views of a system needs the system's disorder to curve a certain way. Gravity breaks that, which is why a star heats up as it loses energy and why there is no ordinary thermodynamics of gravitating matter.
A formula for queue delays assumes the spread in service times is finite. Real service times — web requests, phone calls, file sizes — often are not, and then the formula says nothing.
The black hole information paradox assumes the inside and the outside can be treated as separate systems. Gravity forbids exactly that split, and the field had known it for other reasons for decades.

None of those was found by a search. All three turned up in the ordinary course of using the tool, which is the point: check what a theorem requires, not just whether your object is well-behaved. Two weeks later this rule earned its keep again, in a place nobody expected: a count of "the three open corners" of a problem turned out to be a count of what happened to be known. Finding the classification the corners came from turned three into six — and showed that one of the "open" three had been closed all along by a theorem everyone had already cited and nobody had applied to that axis.

One audit was run deliberately, on a tool used twice with its fine print never stated, and it found nothing — the same inequality both times, no confusion. That matters. Four earlier audits each found a collision, and it was becoming easy to believe audits usually do.

They do not. Those four were run on words that already looked suspicious, so four-for-four measures the suspicion, not the rate. A hit rate on investigations you chose to open tells you about your nose, not about the world.

It did find one real thing, though. The tool's constant is not known exactly, which is harmless when the conclusion is a trend and not harmless when the conclusion is a number. So one widely-quoted figure in fluid dynamics should be read as a bound with unknown slack rather than as an estimate.

5 · What this apparatus cannot do, proved from outside

Everything in this guide is built from three kinds of move: diagonalise (a self-referential trick, the ancestor of most famous impossibility results), compute an invariant (find a number that does not change and use it to tell things apart), and linearise (make a hard problem straight, solve it, translate back).

Computer science has proved, separately and for its own reasons, that each of the three is provably too weak to settle its central question. Three theorems, three barriers, one per move. The middle one rules out invariants by name — which is this guide's entire method.

There is a fourth move, and it has no barrier against it. Instead of computing something, you assert that some computation is complete — that nothing else could have produced this answer. That is a different kind of claim, it evades all three barriers, and it is where the serious attempts now live. It is also the only move with nothing proved against it, which is worth knowing even though it has not yet worked.

6 · How it caught its own mistakes

The corrections in this guide did not come from staring harder at its contents. Four habits produced nearly all of them, and they transfer.

Make a list say what it is a list of. Both of the largest corrections came from this, and neither came from examining entries. A classifier never asked what it classifies will happily classify the wrong thing forever. One list turned out to cover only half of a two-sided thing; another turned out to answer a kind of question the guide almost never actually asked.
Audit the words you invented, not just the ones you borrowed. Borrowed words get checked, because someone else's definition exists to compare against. Invented words never do, because inventing a word feels like defining it. The guide's most-used invented word was carrying five meanings.
Be specific enough to be wrong. One computation here hit four separate bugs, and every one produced a plausible-looking answer. They were caught only because the prediction said exactly which numbers should appear. A vaguer claim would have been "confirmed" by all four. Specificity is not a style preference; it is the error-detection mechanism.
Run a control. Always include the case that should fail. Without it you have a definition wearing a measurement's clothes.

7 · Two things actually measured

Two claims in this guide were argued for a long time before anyone computed them. Both now hold.

The crystal that should not diffract. The theorem underneath most of this needs a regular grid. Quasicrystals have no grid and produce sharp diffraction spots anyway — discovered on photographic film, disbelieved for years, eventually a Nobel Prize. The question is what survives when the precondition fails.

What survives, measured: the spots are indexed by two whole numbers instead of one, which is why they are densely packed rather than separated; their brightnesses follow exactly the pattern predicted by treating the quasicrystal as a slice through a higher-dimensional grid, matching to one part in six thousand across twenty-three spots; and the spots stay sharp as the crystal grows, which is what makes them real.

And the control, which is what makes this a measurement: a different non-repeating pattern was tested and produces no sharp spots at all, with its brightness fading at precisely the predicted rate. Being non-repeating is not enough. There is a specific extra condition, and now we know it is the one that matters.

A fundamental limit, from two constraints. The ceiling on how correlated two separated quantum systems can be was computed here for the first time rather than quoted: require a table of numbers to be non-negative, maximise, get 2.8284271 to ten decimal places. Classical physics allows 2. Pure cause-and-effect would allow 4. Reality sits in between, and where it sits follows from positivity alone.

8 · Write the answer down before you look

The single change that made the biggest difference in the last two weeks was mechanical. Before any computation is run, two numbers go into a file with the date on them: the answer we expect, as a range, and the answer that would prove us wrong. The file is committed before the computer starts. Afterwards the result is scored against those numbers as written, never against what we would have liked to have written.

Seven such tests were run in one weekend on a single object — a strange graph described in the physics companion. The scorecard: two predictions confirmed, two flatly wrong, one that landed on the edge of its band, one withdrawn because the instrument failed its own control, and one still running. Nobody would have predicted that spread in advance, and that is the point. A method that is right most of the time is not being asked hard enough questions.

The two wrong ones were the most useful. One predicted that a certain system would always settle into agreement, on the reasoning that its loops were too short to hold anything. Every one of two hundred trials disagreed — and the analysis showed the reasoning had the sign backwards: short loops in a slow system are exactly where things get stuck. The other predicted a particular anatomy for a pattern we had found; the measurement showed the pattern was a different shape entirely, and the different shape was the real explanation. In both cases the written prediction is what turned a vague "hm" into a specific "wrong about this, in this direction".

And one seal was set on a wrong reference. The band for one test was placed "within a factor of two of the textbook value" — but the textbook value had been computed wrongly, by mixing up two ways of measuring the same wall. The control — the ordinary flat grid whose answer is known exactly — landed outside the band, which is how the mistake showed. The seal stands as written and the result is scored against it; the record says the band described our arithmetic, not the object.

The rule that came out: before sealing a prediction, compute what the control should give under the same definition and check it sits where your words say it should. A band the control fails is a statement about the instrument.

9 · When "three" means "the three I know about"

A question came up of the form "why exactly three?" — three known ways of doing a certain thing. The honest answer was that three is how many happened to be known. That is not a count; it is an inventory.

The way to get a real count turned out to be simple, and it transfers. Find the classification the instances are drawn from. There is usually a theorem somewhere that sorts all possible objects into a few types. Cross it with whatever distinction had already split the known instances, and read the no-go theorems' fine print as the boundaries of the cells that are closed. The result is a grid. In this case: six cells, not three corners. One cell was structurally empty — the classification produces it and a theorem then rules it out — which is the best evidence the grid is the right grid. One cell had never been examined by anyone, as far as searching could tell. And one of the original "open" corners had been closed all along by a theorem cited on the same page.

The standing rule behind this: never base a count on what happens to have been found when no theorem says that is all there is. A coincidence of inventory is an assumption, and assumptions are for turning over whenever you can.

10 · A coincidence, chased and broken

The same quantity, N divided by the logarithm of N, showed up in three places that have no business sharing anything: the number of primes below N, a weight used in a famous open problem about primes, and a property of the strange graph mentioned above. The pull to declare them "the same thing" was strong. The guide's rule is the opposite: chase the coincidence, then try to break it.

It broke twice. First, the logarithm was counting different things in the three cases — in two of them the depth of a tree, in the third its width, which are an exponential apart, so the shapes agreed by an accident of how big each object is. Second, and this one was measured: the next term after the leading one is where the prime-counting side keeps all its content, and on the graph's side that term was a plain constant — 11/16, no structure like the prime side's at all. Two things that agree at the first term and disagree at the second are different things. What survives is a record of the chase and the one sentence it left standing.

11 · Instruments fail, and the control is how you know

Two of the weekend's instruments broke, in ways that would have produced confident wrong answers if the controls had not been there.

The first estimator of a wall's cost walked a system slowly across a threshold and read the cost off the histogram. On the plainest possible test object — a flat grid, whose answer is known exactly from a 1944 calculation — it came out thirty percent high, because the system got stuck on one side of a transition halfway through and never crossed. The fix was a different route to the same number, checked to one percent of the exact answer before any real object was touched.
The second was a standard test for a "glassy" landscape, run at a temperature that turned out to be too warm to see one. It reported "no glass" on the object that is the textbook example of a glass. By the rule written into the test beforehand, that voided it. The cruder measurement recorded alongside — the plain energy cost of walking from one valley to the next — was the one that worked, and it was not the one that had been sealed.

Neither failure was visible from the object under study. Both were visible from the controls. A control is not a courtesy; it is the only way an instrument can tell you it is lying.

The error this whole apparatus exists to catch was committed by its author. A few days before the weekend, one run of a chaotic simulation was reported as a measurement, to four decimal places. A second machine got a number three to four times different. Chaotic systems do not reproduce run by run; only averages over many runs do. The number was withdrawn the same day and replaced by an average over thirty runs with an error bar. The withdrawal is in the file, above the correction, because deleting it would erase the lesson.

The whole guide in fourteen lines

Numbers that recur across unrelated systems come in three kinds, and you tell them apart by whether they have error bars.
Exact does not mean correct: an exact fraction can rest on a false assumption about what belongs in the calculation.
A number you cannot compute can often still be fenced in, and there are four ways to build a fence.
Most deep theorems are one move: count something two ways and set the answers equal.
Why you are allowed to count twice is not always the same reason. There are at least three licences, with different fine print.
Almost every mistake is a tool used outside its fine print — not an arithmetic slip.
A hit rate on investigations you chose to open measures your nose, not the world.
Three theorems prove that the three moves this guide is built from are each too weak for the hardest question in computing. A fourth move has no barrier against it.
Make every list say what it is a list of. That one habit produced the two largest corrections here.
Be specific enough to be wrong, and always run the case that should fail.
Write the expected answer and the disproving answer down, dated, before you compute. Score against them as written.
Check the control against your band before you seal it; a band the control fails is about your arithmetic.
"Three" usually means "the three I know about". Find the classification and count the cells.
Chase coincidences, then try to break them. Two things that agree at the first term and disagree at the second are different things.