Quantum Measurement, How Reading a Qubit Works

Quantum measurement is the act of pulling a definite, classical answer out of a quantum state, which sounds simple enough until you look at what it costs. It is where quantum mechanics is at its strangest, because the act of looking changes the thing you are looking at and forces a random outcome. Every quantum computation ends in a measurement, and on real hardware that read is a specification in its own right, with its own speed and its own error rate.Measurement is an active event that destroys information, not a passive reading. One rule, named after Max Born, turns the quantum state’s amplitudes into the odds of each outcome, and everything else about reading a qubit follows from it. The act is irreversible. It takes many repetitions to learn anything about an unknown state, and it needs no conscious observer, only a detector that ends up correlated with the qubit.Key TakeawaysIt extracts a classical answer. A measurement maps a quantum state to a single classical outcome plus a changed state, turning quantum information into ordinary bits.The Born rule sets the odds. The probability of each outcome is the squared magnitude of its amplitude, the bridge from quantum states to observed results.You choose what to measure. The same state gives different results depending on which observable you decide to measure, so there is no single privileged measurement.It cannot be undone. Unlike a reversible gate, a measurement discards the part of the state that disagrees with the outcome, making it the one irreversible step in a circuit.One shot gives one number. Learning an unknown state takes many identically prepared copies, and the statistical error falls only as the square root of the number of runs.No observer is required. Any sufficiently strong interaction with a detector or the environment counts, so consciousness plays no role.On This PageA definite value only in one set of special statesThe Born rule squares an amplitude to get a probabilityThe number you read is one of the observable’s allowed valuesMeasurement is the one step in a circuit that cannot be undoneYou choose what to measureYou cannot dodge the disturbance by copying the qubitNot clumsy instruments, whatever the story saysOne more decimal place of precision costs a hundred times more shotsThe general measurement allows more outcomes than the system has statesThe measurement problem nobody has closedReal machines read a qubit with a microwave tone or a scattered photonError correction measures whether two qubits agree, not what they hold Slow readout caps the error-correction cycle whatever the gates do Measurement-based computing applies no gates once the calculation begins Frequently asked questionsA definite value only in one set of special statesA qubit in a superposition has no definite value for most properties. It has one only when it sits in one of the observable’s special states, and for every other state the theory supplies probabilities, not a value waiting to be found. Quantum measurement forces the issue by producing a single classical outcome you can record, along with a post-measurement state that is usually different from what you started with.This is why measurement sits at the end of every quantum algorithm. The delicate quantum manipulation is useless until you read out an answer. Reading it out is exactly this act. The answer is generally random, drawn from a distribution the quantum state defines, which is why quantum algorithms are built to make the useful answers overwhelmingly probable rather than merely possible.The Born rule squares an amplitude to get a probabilityMax Born published the rule in 1926, in the collision-process paper that gave quantum measurement its probabilities (Zeitschrift für Physik 37, 863). The recipe is arithmetic. Square the magnitude of an amplitude and you have the probability of the matching outcome. That single step is the bridge from the smooth deterministic mathematics of quantum states to the random results we actually see.An amplitude is not a probability but a complex number with a magnitude and a phase, and only the squared magnitude comes out as a probability. The relative phase between two amplitudes fixes how they interfere, and that changes the odds you see when you measure a different way, so the phase is not decorative. Born took the 1954 Nobel Prize in Physics, shared with Walther Bothe, for that statistical interpretation of the wavefunction.The Born rule is a postulate, not a theorem. Several derivations from more basic assumptions have been proposed, and none has yet been generally accepted. That leaves the most useful rule in the subject standing on an assumption, though it is also among the most precisely confirmed statements in physics.The number you read is one of the observable’s allowed valuesThree ideas get blurred together constantly. The first is the observable, the property you choose to measure. Written out mathematically, it is a particular kind of matrix called a Hermitian operator. Each observable comes with its own set of special states, called eigenstates, which together form a complete list of mutually exclusive alternatives. The number you actually read off the instrument is one of those states’ values, an eigenvalue.The mathematics is chosen to fit the physics. A Hermitian operator is guaranteed to have real eigenvalues. That matters because a measured quantity has to be a real number rather than a complex one. Its eigenstates are also guaranteed to be mutually orthogonal, meaning no two of them overlap, which is what makes the outcomes genuinely distinct alternatives.One trap catches people constantly. It is the average of many measurements, a quantity known as the expectation value. That average is not the result of any single measurement and need not even be one of the allowed outcomes. Measure a qubit in the standard basis and you always get zero or one, yet the average can be any number in between. An algorithm quoting an expectation value is reporting a statistic gathered over many repetitions.Measurement is the one step in a circuit that cannot be undoneEvery quantum gate is reversible. Gates are unitary operations that rotate the state without losing any information. Running the inverse rotation puts you exactly back where you started. Measurement is the exception, a non-unitary operation that hands you a definite classical outcome instead of another quantum state, and there is no inverse to run.The reason it cannot be undone is mechanical, since the projection keeps only the part of the state that agrees with the outcome and discards everything at right angles to it. The amplitudes and phases carried by the discarded part are simply gone. Many different starting states land on the same post-measurement state, so the outcome and the state afterwards tell you nothing about which one you began with. That collapse, the many-to-one squashing of possibilities into a single recorded answer, is where the coherence powering the machine is finally lost.You choose what to measureThe same qubit gives different odds depending on the axis you measure along, and no axis is privileged over the others. Diagram by Quantum Zeitgeist.Quantum measurement is not one fixed operation. You pick which observable to measure. Measuring a qubit’s spin along one axis is a genuinely different measurement from measuring it along a perpendicular one. Because the same state gives different probabilities for each, there is no single privileged way to measure a quantum system.The Born rule makes this concrete and geometric. Write the state in the basis belonging to whichever observable you chose, and the probability of each outcome is the squared length of the state’s projection onto that basis direction. A state sitting exactly on the plus-X axis is certain to give the plus result along X, and a perfect coin flip along Z. Rotate the state instead of the axis and the roles swap.That freedom is used constantly. Hardware can usually only measure in one fixed basis, so revealing the hidden phase of a superposition means applying a rotation first and then measuring. That is exactly equivalent to having measured along a different axis. Observables that share a set of eigenstates are said to commute, and they can be measured together, while non-commuting ones such as spin along perpendicular axes cannot both have definite values at once.You cannot dodge the disturbance by copying the qubitQuantum measurement generally disturbs the state it acts on. Physicists summarise this by saying there is no measurement without back-action. Back-action is the kick the apparatus gives the system in return for the information, and it has exactly one exception. If the state is already an eigenstate of the observable you measure, the measurement leaves it unchanged and confirms the value, and in every other case measuring changes the state.The obvious dodge is to copy the qubit first and measure the copies in different bases. The no-cloning theorem forbids it. An unknown quantum state cannot be duplicated by any physical process at all, because the linearity of quantum mechanics rules out a machine that copies arbitrary inputs. That single restriction is why eavesdropping on a quantum key distribution link is detectable, and it underpins much of quantum cryptography.Not clumsy instruments, whatever the story saysHeisenberg’s uncertainty principle is regularly explained as clumsiness, the idea that looking at a particle inevitably knocks it about. That story describes something real. It is not what the textbook inequality says. The standard relation between position and momentum is a statement about the quantum state itself, and it says no state exists whose position spread and momentum spread are both arbitrarily small.The difference matters because the inequality is about preparation, not clumsiness. Prepare an ensemble of identical systems, measure position on some of them and momentum on the others, each as precisely as you like, and the two spreads you obtain still cannot both be small. No improvement in instruments removes that limit. It was already present in the state before any instrument touched it.A separate family of relations, the measurement-disturbance relations, does deal with disturbance, trading off how much information a measurement extracts against how much it disturbs a complementary property. They are real physics. They formalise the reasoning in Heisenberg’s microscope argument, where looking at a particle disturbs it, but they remain a different statement from the inequality printed in the textbook.One more decimal place of precision costs a hundred times more shotsOne shot returns one outcome. No single measurement can reveal an unknown state’s amplitudes. To learn the distribution you prepare the same state and measure it over and over, building up frequencies that approximate the underlying probabilities. Reconstructing the state completely is called quantum state tomography, and it needs several different bases, because no single basis reveals the phases.The statistics are unforgiving. The uncertainty in an estimated probability falls only as one over the square root of the number of runs. One more decimal place of precision costs a hundred times more shots, so estimating an expectation value to a part in a thousand takes on the order of a million repetitions. That is why the sampling cost, rather than the gate count, is often the real bottleneck for variational algorithms in the NISQ era.Scaling makes it worse. Full tomography of an n-qubit state needs a number of distinct measurement settings that grows exponentially with n, which makes it a practical tool for two or three qubits and hopeless for twenty. Useful quantum algorithms are therefore designed to concentrate probability onto a few correct answers rather than to have you read out the whole state.The general measurement allows more outcomes than the system has statesThe clean picture of a measurement that projects onto orthogonal alternatives is only the simplest case. The most general measurement quantum mechanics allows can have more outcomes than the system has states, and its alternatives need not be sharply distinct at all. The bookkeeping is simple. A general measurement is described by a set of positive operators that add up to the identity. That sum is what forces the probabilities to total one, and the object itself is called a POVM. In practice you build one by measuring the system together with an extra qubit brought along for the purpose, called an ancilla.Two variants matter here. The first is the weak measurement, which couples the apparatus to the system only slightly. It extracts a little information per shot and disturbs the state correspondingly little, with the statistics built up over many runs (Aharonov, Albert and Vaidman, Physical Review Letters 60, 1351, 1988). A quantum non-demolition measurement, or QND, goes the other way, targeting an observable the system’s own dynamics leave unchanged. Repeated measurements then agree, and the back-action is pushed into a variable you do not care about (Braginsky, Vorontsov and Thorne, Science 209, 547, 1980).Mid-circuit measurement matters most for computing. Some qubits are measured partway through the circuit, and the outcome steers the operations that follow. That is the mechanism behind quantum error correction, which measures combinations of qubits chosen so that the answer reveals whether an error occurred and nothing about the protected information. The same feed-forward trick makes quantum teleportation work, with a measurement outcome sent over an ordinary classical channel and used to pick the right correction.The measurement problem nobody has closedBeneath the practical rules lies a genuine puzzle about how the smooth and the abrupt fit together. Between measurements a quantum system evolves smoothly and reversibly, following an equation that never picks out one outcome over another, and then a measurement arrives and looks abrupt, random and final. Nobody has reconciled the two pictures. The gap between them is called the measurement problem, and it is still open.Decoherence, the leaking of quantum information into the environment, explains a great deal of it. It shows why interference between alternatives becomes unobservable in practice, and why only certain stable states ever get recorded. One thing it does not explain is the outcome. Suppressing interference between possibilities is not the same as eliminating all but one of them, so the question of why this result and not that one survives untouched.The competing interpretations divide on exactly that residue, and textbook collapse treats the projection as a brute fact while many-worlds denies that anything is eliminated at all. Pilot-wave theories add definite particle positions, while objective-collapse models change the equations so that collapse becomes a physical process rather than a rule applied by hand. All of them reproduce today’s predictions, with a partial exception in the objective-collapse models, which predict tiny deviations that experiments keep bounding more tightly without ever finding one.Real machines read a qubit with a microwave tone or a scattered photonMeasurement in real hardware is precision engineering. The two leading platforms do it in quite different ways. In a superconducting quantum computer each qubit is coupled to a small microwave resonator, a scrap of circuitry that rings at its own frequency. The two are deliberately kept far apart in frequency, a setting called the dispersive regime, and there the qubit’s state shifts the resonator’s ringing frequency without either one handing energy to the other. You therefore bounce a weak microwave tone off the resonator and read the phase it comes back with, and that phase says whether the qubit is a zero or a one. The probe couples to the qubit’s energy rather than flipping it, so it does not drive transitions between the two states. That leaves the readout approximately non-demolition, meaning the qubit survives in the state it just reported (Blais et al., Reviews of Modern Physics 93, 025005, 2021).Trapped-ion machines take a different route. A laser is tuned to a transition that only one of the two qubit states can absorb. That state scatters photons repeatedly and shows up as a bright spot on a camera, while the other stays dark. Counting those photons for a short window separates bright from dark with high confidence. Vendors quote high readout fidelities on both platforms, and those are their own figures, measured on their own devices, so any headline number of that kind dates quickly.Both methods are, to a good approximation, the projective measurement of the textbook, delivering one classical bit per qubit per run. Neither needs an observer in any mystical sense. An ordinary physical interaction with a detector is enough. What matters is that the detector ends up correlated with the qubit and that the correlation is recorded somewhere classical. Our guide to quantum computing sets out the rest of the machine that surrounds this one irreversible step. Error correction measures whether two qubits agree, not what they hold There is a problem here that ought to make error correction impossible. Finding an error means asking the machine a question, asking a question means measuring, and measuring destroys the superposition you were trying to protect. On that reading, fault-tolerant quantum computing could not exist in the first place. The way out is to change the question. You measure something that carries information about the error and none about the encoded state. The machine does not ask what value a qubit holds, but whether two neighbouring qubits agree with each other, which reveals that something flipped without revealing which state the pair was in. The superposition survives the question. Those agree-or-disagree questions are called parity checks. Stabiliser codes run entirely on these checks, so the whole scheme reduces to running and reading them. The string of answers they produce is called the syndrome, and a decoder reads it to work out which correction to apply. No one ever learns the logical value the qubits hold, which is precisely what makes the trick work. The measurement is real and its result is recorded, and it has been chosen so that what it extracts is exactly what you are allowed to know. The result is stranger than it first sounds, because the obstacle running through the whole subject is that you cannot look at a quantum state without changing it. That obstacle turns out to be navigable if you are careful about what you look at. Error correction does not defeat the measurement rule. It works inside it. Our guide to quantum error correction follows what happens once the syndrome has been read. Slow readout caps the error-correction cycle whatever the gates do Measurement on a real machine is not the instantaneous, perfect operation of the textbook. It takes time and it can be wrong. Both of those facts propagate into everything built on top of it, which is why readout fidelity is quoted separately from gate fidelity. That figure says how often the machine reports the state the qubit was actually in. The two directions of error are usually not symmetric, because a qubit in the excited state can decay during the readout window and be reported as a zero. The reverse mistake is rarer. The error rate depends on which answer was correct. Hardware teams report the two separately because a decoder that assumes they are equal will make worse corrections than one that knows the asymmetry. Speed matters as much as accuracy, and this is where readout stops being a detail. Error correction has to measure the syndrome, decide on a correction and apply it, over and over, while the computation is still running. A slow read caps that cycle. A machine with excellent gates and slow readout is limited by the readout, not by the gates. When a specification sheet arrives, measurement deserves the same scrutiny as the gates. That comes down to three questions. Ask how long a readout takes, ask what the fidelity is in each direction, and ask whether the qubit is still usable afterwards. Those answers say more about what a machine can actually run than the qubit count printed above them. Measurement-based computing applies no gates once the calculation begins Measurement has been the thing you do at the end here, after the gates have done the work. One whole family of machines reverses that ordering. Knowing it exists changes what you think measurement is for. In those machines the measuring is the computing rather than the full stop after it. The approach begins by preparing a large entangled state, often called a cluster state, before any calculation has been specified. The computation then consists of measuring the qubits one at a time in a chosen order, with the basis for each measurement decided by the results of the earlier ones. No gates are applied during the calculation at all. The entanglement was created up front and is simply consumed as the measurements proceed. This is called measurement-based or one-way quantum computing, and it is provably as powerful as the gate model rather than a cut-down version of it. The one-way name comes from the resource being used up, because a measured qubit cannot be put back into the cluster. What looks like destruction in the gate picture is the mechanism itself here. Some hardware finds this ordering easier. Photonic machines are the clearest case. They are good at producing entangled states and poor at holding photons still long enough to run a sequence of gates on them. A model that front-loads the entanglement and then only measures suits that physics, and it corrects the common assumption that measurement is where a quantum computation goes to end. One consequence lands on the electronics, because each measurement basis depends on earlier outcomes. A measurement-based machine therefore needs classical control fast enough to pick the next question while the state is still alive. Error correction imposes the same requirement, so fast classical hardware sitting beside the quantum hardware is now an engineering discipline of its own rather than a supporting detail. None of this changes the underlying rules. The Born rule still governs the probabilities, and the choice of basis still decides what you can learn. The state is still disturbed by being read, and what changes is the job those facts do. A constraint in one model turns out to be the engine of another. Frequently asked questionsWhat is quantum measurement in simple termsIt is the act of extracting a definite classical answer from a quantum state. Because a quantum system has no fixed value for most properties, the measurement forces a single outcome and usually changes the state. Every quantum computation ends in a measurement.What is the Born ruleThe Born rule, introduced by Max Born in 1926, gives the probability of each measurement outcome as the squared magnitude of its amplitude in the quantum state. It is the bridge from the deterministic mathematics of quantum states to the random results we observe, and it earned Born the 1954 Nobel Prize in Physics.Does quantum measurement need a conscious observerNo. Any sufficiently strong interaction with a detector or the environment counts as a measurement. Consciousness plays no role, and the outcome is recorded whether or not anyone is watching. This is one of the most persistent myths in the subject.Why can a measurement not be reversedBecause the projection keeps only the part of the state that agrees with the outcome and discards everything at right angles to it. Many different starting states end up in the same post-measurement state, so there is no way to work backwards. Unlike a reversible gate, a measurement destroys the coherence it acts on.Does measuring a qubit always destroy its stateNot always, though it disturbs it in general. If the qubit is already an eigenstate of the observable being measured, the measurement leaves it unchanged and just confirms the value. In every other case, measuring changes the state.Is the uncertainty principle about the instrument disturbing the systemNot the textbook version. The standard position-momentum inequality is a property of the quantum state itself, saying no state has both a sharp position and a sharp momentum, and it holds even if every individual measurement is perfect. There is a separate family of measurement-disturbance relations, but the two statements are not the same thing.How many measurements do you need to learn a quantum stateMany. One shot gives one outcome, so you must prepare and measure the same state repeatedly, and the statistical error falls only as the square root of the number of runs. Full reconstruction also needs several measurement bases, and the number of settings grows exponentially with the qubit count, which makes complete tomography impractical beyond a few qubits.Can you copy a qubit to measure it many timesNo. The no-cloning theorem forbids copying an unknown quantum state, so one copy gives one outcome. To learn an unknown state you need many identically prepared copies, measuring each. This restriction is the basis of quantum cryptography.Does measuring an entangled particle send a signalNo. Measuring one half of an entangled pair changes the correlations but does not change the local statistics on the other side, so no information travels faster than light. A separate classical message is always needed, as in teleportation.What is the measurement problemIt is the open puzzle of why a quantum system evolves smoothly between measurements yet appears to collapse abruptly and randomly when measured. Different interpretations of quantum mechanics answer it differently while agreeing on every prediction. Decoherence explains why the world looks classical but does not say why one outcome occurs.Read NextSuperpositionQuantum DecoherenceQuantum Error CorrectionQuantum TeleportationWhat Is Quantum ComputingMore like thisQuantum FeaturesMeasuring a Photon’s Past. It Didn’t Exist Until We LookedQuantum PhysicsQuantum Mechanics: Purpose Over RealityQuantum ComputingQuantum Paradoxes Unlock Faster ComputationQuantum FeaturesThe Quantum Zeno Effect: How Watching Stops Quantum SystemsStay currentSee today’s quantum computing news on Quantum Zeitgeist for the latest breakthroughs in qubits, hardware, algorithms, and industry deals. Tags: