Quantum Academy is now open. Enroll in live online training – see the schedule.
Skip to content
Browse courses

What Happens When You Measure a Qubit

Quantum Computing

Quantum Academy6 min read

Measuring a qubit gives you one classical bit, 0 or 1. Which bit you get is random, with odds that depend on the qubit’s amplitudes. Afterwards the qubit is in the state that matches the result. Whatever superposition it was in beforehand is gone.

Three things practitioners meet early follow from that. Quantum programs run the same circuit thousands of times and report counts. Algorithms require interference before the readout, because it returns one bit string. And an eavesdropper who measures the photons on a quantum key distribution (QKD) link introduces errors that the two ends can detect.

The Born rule

A qubit’s state is written a|0⟩ + b|1⟩, where |0⟩ and |1⟩ are the states a standard measurement reports as 0 and 1. The numbers a and b are called amplitudes. If both are nonzero, the qubit is in a superposition. Measure it and you get 0 with probability |a|² and 1 with probability |b|², where |a| and |b| are the magnitudes of the amplitudes. That’s the Born rule, named after the physicist Max Born.

Amplitudes can be complex numbers, which is why the rule uses magnitudes. Take a = 1/2 and b = i√3/2, where i is the square root of −1. The magnitude of b is √3/2, about 0.866, and its square is 3/4, so this qubit reads 0 with probability 25% and 1 with the remaining 75%.

After the readout, the qubit is in |0⟩ if the result was 0 and in |1⟩ if it was 1. Qiskit’s documentation calls the operation “non-reversible,” and the change of state is often called collapse.

Why amplitudes can’t be read directly

Each readout gives you a 0 or a 1, never the numbers 1/2 and i√3/2. To estimate |b|², you prepare and measure the same state many times and count the ones. You can’t keep a copy of the qubit to measure later, because no device can copy an unknown quantum state. William Wootters and Wojciech Zurek published that result, now called the no-cloning theorem, in Nature in 1982.

From these counts you can’t learn an amplitude’s phase either, meaning its sign or complex angle. With b = √3/2 or −√3/2 in place of i√3/2, the qubit still reads 1 with probability 75%, so the counts are the same for all three states. You can tell them apart by measuring in a different basis, covered below.

Measuring many qubits returns one bit string, chosen with Born-rule probabilities, and every amplitude in the superposition is lost. For a quantum algorithm to be useful, its gates have to cancel the amplitudes of wrong answers and reinforce the right one before the readout. That’s called interference, and in our explainer on quantum parallelism we work through the simplest example, Deutsch’s algorithm.

Shots and histograms

A shot is one run of a circuit, from preparing the qubits to measuring them. Quantum software runs a circuit for the number of shots you set and returns counts, the number of times each result occurred, which you can plot as a histogram.

You estimate probabilities from counts, and the typical error is proportional to one over the square root of the number of shots. For the 25/75 qubit above, 100 shots give about 75 ones with a standard deviation of about 4, so a count of 71 or 79 is unremarkable. With 10,000 shots you’d expect about 7,500 ones with a standard deviation of about 43, which puts the estimate typically within half a percentage point of 75%. Going from 100 shots to 10,000 gives ten times the precision, which is why quantum programs run thousands of times.

Measurement bases

For one qubit, a measurement basis is a pair of states that you can always tell apart with a single measurement. The pair |0⟩ and |1⟩ is the Z or computational basis, and IBM’s documentation states that a measurement projects the qubit onto it. The X basis is the pair

  • |+⟩ = (|0⟩ + |1⟩)/√2
  • |−⟩ = (|0⟩ − |1⟩)/√2

To measure in the X basis, you apply a Hadamard gate, written H, which turns |+⟩ into |0⟩ and |−⟩ into |1⟩, and then measure as usual. You then read a 0 as + and a 1 as −.

The plus state in two bases

Apply H to a qubit that starts in |0⟩ and you get |+⟩. Run each circuit below for 10,000 shots on a simulator.

Measurement Gates after preparing the plus state Expected counts
Z basis Measure About 5,000 zeros and 5,000 ones
X basis H, then measure 10,000 zeros
Z, then X Measure, H, measure About 5,000 zeros and 5,000 ones in the second result

In the Z basis, |+⟩ gives a coin flip, because amplitudes of 1/√2 give probabilities of 1/2. In the X basis the same state gives 0 every time, because H turns |+⟩ into |0⟩ before the measurement.

In the third circuit, the first measurement returns 0 or 1 at random and leaves the qubit in |0⟩ or |1⟩. Each of those is an equal mix of |+⟩ and |−⟩, since |0⟩ = (|+⟩ + |−⟩)/√2 and |1⟩ = (|+⟩ − |−⟩)/√2, so the X-basis result is now random too. The Z-basis measurement destroyed |+⟩, and an answer that was certain became a coin flip.

The state |0⟩ is the reverse, certain in the Z basis and random in the X basis. In either basis, one of |0⟩ and |+⟩ gives a random result. QKD depends on that.

Eavesdropping on QKD

BB84, the QKD protocol that Charles Bennett and Gilles Brassard presented in 1984, encodes key bits on single photons in two polarization bases, horizontal and vertical for one and the two diagonals for the other. They correspond to the Z and X bases. Alice, the sender, picks the basis at random for each photon, so an eavesdropper has to guess it. One who measures a diagonal photon in the horizontal-vertical basis is running the third circuit. Her result is random, the photon she forwards is polarized to match it, and the receiver, Bob, reads the wrong value half the time when he measures in Alice’s basis.

In their paper, Bennett and Brassard calculated that measuring and resending every photon in one basis causes disagreements in 1/4 of the bits Bob later measures in Alice’s basis. Alice and Bob find them by comparing a sample of bits in public, which is how BB84 detects an eavesdropper.

Readout on superconducting hardware

On Google’s Sycamore processor, every qubit has its own microwave resonator for readout. The control electronics probe the resonator with a microwave pulse, and the signal they record depends on whether the qubit is in 0 or 1. In a Google Quantum AI error-correction experiment published in Nature in February 2023, single-qubit gates took 25 ns and two-qubit gates 34 ns, while measurement took 500 ns and reset another 160 ns. Those two steps took up more than two-thirds of each 921 ns error-correction cycle, in which the processor measured its helper qubits.

Among the readout error mechanisms the team modeled, listed in the paper’s supplementary material, are a finite signal-to-noise ratio and qubit relaxation, a 1 decaying to 0 before the measurement ends. Google’s spec sheet for its Willow processor, published in December 2024, gives a mean measurement error of 0.77% for its error-correction chip, against 0.035% for single-qubit gates and 0.33% for two-qubit gates. That’s about one wrong bit in every 130 readouts, a higher error rate than either kind of gate.

Our Quantum Computing Fundamentals course covers hardware modalities and their trade-offs, quantum error correction and cloud platforms, alongside qubits, gates and algorithms at a conceptual and practical level.

Measuring your own readout error

Before you read meaning into a small difference in a hardware histogram, check the readout error of the qubits you used. IBM’s calibration data gives the error and length of each qubit’s readout, with the error calculated as the average of P(0|1), the probability of reading 0 after preparing 1, and P(1|0).

To measure it yourself, run two circuits on the same physical qubit with the same number of shots. One only measures and should read 0 every time. The other applies an X gate first, flipping |0⟩ to |1⟩, and should read 1 every time. The share of wrong results in each is that qubit’s error in each direction, mostly from readout, at the time you ran it. Any effect smaller than those shares could be readout error alone.