In the laboratory, a hundred or so microwave pulses leave an electronics rack sitting at room temperature, travel down coaxial cables through the fridge’s stages, strike the chip and come back. The whole thing lasts a few microseconds. For all that time, nobody looks at anything.

A suspended dilution refrigerator, its three golden plates bristling with coaxial cables running downwards.
The coaxial cables of a dilution refrigerator, at IQM in Espoo. Microwave pulses travel down these cables to the chip, and the reflected signal comes back up the same ones. (Photographie Ragsxl, 2020, Wikimedia Commons, licence CC BY-SA 4.0, source page)

The qubit France bought is a current turning in a superconducting loop, at ten millikelvin, carrying its information in the way it oscillates. What remains is to know what it is made to do.

You have heard the answer already, and it circulates everywhere: a qubit would be worth zero and one at the same time, so a quantum machine would explore every solution in one go, where an ordinary computer reviews them one by one.

It is false. And it goes on to produce every faulty piece of reasoning in the field, including among serious people.

An amplitude is not a probability

Physicists describe it differently. A quantum system does not occupy several states at once, the way a drawer might hold several socks piled on top of each other: they attach to it a list of numbers, one per possible outcome, which they call amplitudes. The word matters, and the nuance separates two worlds. A probability lives between zero and one, so two paths leading to the same outcome can never do anything but swell the total. An amplitude carries a sign.

A comparison between two probabilities that always add up and two amplitudes that can subtract.
Two probabilities can only add. Two amplitudes of opposite sign subtract, and the result can vanish. (Schéma Shy Robotics)

Two paths can therefore subtract instead of adding, and a perfectly possible outcome becomes impossible because another has erased it. Physicists call that interference, the same word they use for two waves crossing.

The rule comes in two steps, in this order. First the amplitudes of all paths are added. Then the result of that addition is squared to obtain the probability.

Two waves in phase reinforcing each other, then two waves in opposition cancelling out, with the measured result underneath.
In phase, the two paths reinforce. In opposition, they cancel. That possibility of cancellation exists nowhere in ordinary computing. (Schéma Shy Robotics)

Reversing those two steps changes everything. An ordinary computer squares first, then adds. Nothing ever cancels in its world.

A quantum algorithm orchestrates cancellations

The whole art lies in an orchestration: making the amplitudes of the wrong answers cancel, making those of the right one add, so that at the end the measurement almost always falls on what is left standing.

Four drawn steps: prepare the possibilities, cancel the wrong ones, reinforce the right one, measure.
The four movements of a quantum algorithm. Cancelling the wrong answers does all the work; the measurement merely records. (Schéma Shy Robotics)

The machine tries nothing, then. It sculpts.

This distinction has a formidable consequence, and it explains why so few problems qualify: a problem becomes accelerable only if someone has found how to orchestrate the cancellation for that precise problem, and nobody knows how to do that on demand.

Three movements, and no looking allowed

With an ordinary computer everyone roughly sees what is going on: a sequence of elementary operations on bits, executed one after another, each reading values somewhere and writing others elsewhere. With a quantum machine, nothing of the sort.

On 9 December 2024, when Google published its logical memory, the sequence had already been settled for years. Physicists first prepare the qubits in a known starting state. They then apply a series of elementary operations, the quantum gates, which the CEA presents as the equivalents of the logic gates of an ordinary circuit. Finally they measure, once, at the end.

Above, five steps in a row each reading and writing; below, three lines of qubits crossed by a series of gates, with a single measurement at the end.
A classical computation reads and writes at every step. A quantum computation chains gates without ever looking, and measures only once, at the end. (Schéma Shy Robotics)

They write no instruction into a memory. To apply a gate, engineers send a microwave pulse of precise duration and frequency, onto one qubit to flip its state, or onto two qubits to bind them together. A hundred gates, a hundred pulses, a few microseconds.

They never look between the preparation and the final measurement. Not out of elegance: looking would destroy the superposition, exactly as an outside disturbance would. The CEA ranks measurement among the interactions that collapse the state.

A quantum computation is therefore not a sequence of steps whose intermediate result is read, but a piece of choreography in one continuous movement, danced blind from start to finish, whose result is known only when the eyes open. That choreography has a name: a circuit. So does its length: the circuit depth.

Factoring fifteen, from start to finish

Take an example, the smallest one that means anything: breaking 15 into 3 times 5. Fifteen is the smallest composite number meeting the conditions of the test, the product of two primes and not even, and a team in Santa Barbara factored it as early as 2012 on a superconducting processor. The input is the number 15, written in binary on the qubits at preparation time, and the circuit then chains a few hundred gates that cancel the wrong tracks. So far, nothing unfamiliar.

Four numbered steps from input to post-processing, and on the right a histogram of the draws obtained by repeating the circuit.
The full journey of a quantum computation, from the number to be factored to its factors. The raw output is not the answer: it is a draw, which an ordinary computer then interprets. (Schéma Shy Robotics)

It is the output that throws you. The measurement does not return 3 and 5: it returns a string of zeros and ones, drawn at random, which is not the answer but a clue, from which an ordinary computer then deduces the factors by a very simple classical calculation.

And it has to be done again. One execution gives one draw, and it is the distribution of thousands of draws that carries the information: in Google’s laboratories, in December 2024, a typical experiment counted a hundred thousand repetitions of the same circuit.

There lies the real difference in kind from an ordinary function. You do not ask it for a result: you ask it to bias a die, then you roll that die thousands of times to see which way it leans. The machine does not compute the answer. It makes it far more probable than the others.

One constraint has run through all of the above without being named. The whole choreography must reach the end in one continuous movement, without a single dancer stumbling, because if one falters along the way no intermediate step has been kept: the entire computation is lost.

How long does a qubit stay on its feet?

The next episode gives the figure, and explains why France bought eighteen rather than a thousand.

Sources

Partager

ARTICLES CONNEXESPLUS DE L'AUTEUR

Leave a Reply