Measurement: one answer, once

Ninety-six per cent, and then it is gone

You have run the search. The state is exactly where you wanted it: the right answer sits at 96%, the fifteen wrong ones at a fraction of a per cent each. You can picture the whole arrangement of amplitudes.

You measure.

Out comes a single number. One box. And the arrangement you spent three moves building is destroyed — every amplitude collapsed to a single definite outcome. There is no way to look again, because there is nothing left to look at.

This is the tax on the whole enterprise, and it is easy to miss because the popular story stops at “the answer is in there”.

What you get is a sample, not a readout

The distinction matters. A quantum machine does not hand you its amplitudes. It hands you one draw from them.

You built a distribution. You get one sample of it. And you cannot look at that sample and tell whether it came from the 96% or from the 0.3%. It is just a number.

So a quantum program is not “run it and read the answer”. It is:

prepare the state → run the circuit → measure → write down what came out → throw everything away → do it all again.

Each repetition is called a shot. Real programs take thousands.

The arithmetic of being sure

How many shots do you need? It depends entirely on how good the state was.

If the right answer has probability p, then the chance of missing it n times running is (1−p)ⁿ. Set that below one in a thousand and you get the shot count:

• p = 96% → 3 shots
• p = 50% → 10 shots
• p = 30% → 20 shots
• p = 10% → 66 shots
• p = 5% → 135 shots
• p = 1% → 688 shots

Notice how badly this falls apart at the bottom. Halving your success rate does not double the work; near the bottom it multiplies it. A great deal of quantum engineering is really about keeping p high enough that the shot count stays sane.

And you must be able to check

There is a second condition, and it disqualifies more proposals than the first.

You get a number out. You do not know whether it is the good one. So you need a cheap, ordinary way to test a candidate answer.

For factoring, that test is trivial: multiply the two factors and see if you get the original number back. For the box search, open the box. For a chemistry simulation, you are sampling an energy, and repeated samples build the distribution you actually wanted — the samples are the answer.

But for a problem where checking a proposed answer is itself hard, sampling buys you nothing at all. You are handed a guess you cannot evaluate, thousands of times. A quantum speed-up needs a classical checker, and when someone describes a quantum application without saying what the checker is, that is usually the missing piece.

The cost nobody quotes

Put the shot count together with the run length and something slightly awkward appears.

Take the sixteen-box search. Three moves gets the answer to 96.1%, which needs 3 shots — 3 × 4 units of machine time, call it 12. But two moves gets to 90.8%, which also needs only 3 shots, at 3 × 3 = 9 units.

Two moves is cheaper than three, even though three is more accurate.

The “optimal” number of repetitions depends on what you are counting. Optimising the success probability and optimising the total machine time are different problems with different answers, and the headline figure is almost always the first one.

The general habit

The step that gets forgotten is nearly always the one where the work leaves the system and has to become usable outside it.

A model that predicts well but cannot say why. A study with a real effect that no clinic can administer. A machine that builds a beautiful state and hands you one sample. In each case the impressive part is genuine, and the question that decides whether it matters is the dull one: how does the result get out, and what does that cost?