Before this lesson: qubits and measurement.
Equal probabilities do not specify a quantum state
Consider the two column vectors
Measuring either in the computational basis gives outcomes 0 and 1 with equal probabilities. Yet the two states are different: a suitable operation before measurement distinguishes them perfectly. The missing information is the relative phase of their amplitudes.
This note derives a complete experiment that makes that phase observable. We will also distinguish a global phase, which multiplies the whole state and changes no measurement probability, from a relative phase, which can affect interference. All computations use one ideal qubit; there is no noise, sampling simulation or hidden physical assumption about the hardware needed for the linear algebra below.
Kets, amplitudes and the Born rule
The computational basis consists of the orthonormal vectors
Definition 8.1.1
(Normalized pure qubit state)
A pure qubit can be represented by
where are complex amplitudes. For , . Computational-basis measurement returns 0 with probability and 1 with probability . The amplitudes themselves are not probabilities.
The dagger means conjugate transpose. For a unit vector , write . More generally, measurement in an orthonormal basis gives outcome with probability . This is the Born rule in a different basis.
A one-qubit gate is represented by a unitary matrix , meaning . Hence : the output remains normalized. Matrix products act from right to left.
Why a shared global phase cannot be measured
Let be real. The states and represent the same physical pure state. This equivalence uses a factor of modulus one, not an arbitrary change in either amplitude separately.
Definition 8.1.2
(Global phase and relative phase)
A real number gives the global phase factor in . For a state with , the relative phase is , taken modulo .
Proposition 8.1.3
(A global phase changes no measurement probability)
For every real , every unitary and every orthonormal basis vector ,
Proof
(Pull the phase out by linearity)
For any unitary and any measurement basis vector , linearity gives
Thus changing the basis or adding further unitary evolution cannot turn a shared phase into a different measurement distribution. For instance, is the same state, using .
Proof X-ray
What the cancellation depends on
Only two facts enter. Linearity lets the scalar pass through and the inner product, and makes it disappear under the squared modulus. Unitarity of is not used here; it only keeps the state normalized. The hypothesis that the factor multiplies the entire state is exactly what allows it to be pulled out. In the scalar multiplies one term only and cannot be factored, so the two contributions to an output amplitude no longer rotate together. That failure of factoring is how a relative phase becomes observable.
In an amplitude diagram, changing rotates both arrows together. That changes the chosen vector representative, not an observable outcome.
By contrast, the family
changes the phase between its two nonzero amplitudes. If two members of this family differ only by a global factor, their equal nonzero first amplitudes force that factor to be 1, so their second amplitudes must also agree. Thus they represent the same state exactly when their phases agree modulo . Equal computational-basis probabilities do not erase this distinction.
Counterexample mode
A sign on one branch is not a global phase
Claim: and are the same state, because they differ only by a sign. This is false. The vector is a global-phase copy of , but changes the sign of one amplitude only. Applying gives and , so a final computational-basis measurement returns 0 with certainty in one case and 1 with certainty in the other. A sign is a global phase only when it multiplies every amplitude; on one branch it is the relative phase .
Prepare phase, then choose where to read out
Use the Hadamard gate and a phase gate :
Direct multiplication gives and . Also , since . Starting from , first apply , then , and optionally multiply the whole state by the global factor :
Immediately before a final Hadamard, both outcome probabilities are , for every and . The phase information is in the amplitudes, while this particular measurement sees only their squared magnitudes.
The two possible readout points describe alternative freshly prepared experiments. Reading out before the final means stopping and measuring there. It does not mean performing that measurement, keeping its result, and then continuing the same coherent state through . Measurement is a different operation from inspecting a mathematical state vector on a page.
Hadamard turns relative phase into different probabilities
Apply the final Hadamard without an intervening measurement:
The amplitude for outcome 0 is the sum of two contributions, and . For outcome 1, the second contribution changes sign. Add these amplitudes first, then take the squared magnitude. Adding their squared magnitudes separately would discard the interference term.
Theorem 8.1.4
(Hadamard interference law)
For the state , a computational-basis measurement after one Hadamard gate returns 0 with probability and 1 with probability , for every real .
Proof
(Derive both probabilities and check normalization)
Using gives
The common factor has squared magnitude 1. Therefore
Both are nonnegative and . Equivalently, the final state is
This form separates the global factor from the two observable squared magnitudes. Its signs and complex phases still matter for later gates; probabilities alone are not a replacement state vector.
Common mistake
(Adding probabilities instead of amplitudes)
Treating the two contributions to outcome 0 as separate alternatives gives for every , so no phase would ever matter. At the true probability is : the amplitudes and cancel before squaring.
Exact checkpoints for the interference experiment
Mathematical angles below are in radians; the controls may display degrees. A displayed angle degrees means . Take for the amplitude columns in the table. A different multiplies both columns by the same .
| Relative phase | Final amplitude 0 | Final amplitude 1 | ||
|---|---|---|---|---|
| () | ||||
| () | ||||
| () | ||||
| () | ||||
| () |
Worked example 8.1.5
(A phase flip changes a certain 0 into a certain 1)
At , the phase gate is the identity and . The final outcome is therefore 0 with probability 1.
At , the state before the final Hadamard is . The two contributions to output 0 cancel, while those to output 1 add, giving . The final outcome is 1 with probability 1. Both preparations had computational-basis probabilities before that final gate.
Now change only , for example to . Both final amplitudes change sign, but the certain outcome remains 1. Relative phase changed the distribution after interference; global phase did not.
Worked example 8.1.6
(Amplitudes and probabilities at )
Here . With the final amplitudes are
Their squared magnitudes are and . These agree with and and sum to 1. The phase gives the same two probabilities.
At and , the final distributions are identical. This readout depends only on , so it cannot distinguish from modulo . It does not recover an arbitrary phase completely. A different measurement basis, or a known phase gate before readout, can supply different information. For example, a separate experiment inserting immediately before the last gives instead. The two settings probe different parts of the relative phase.
Concept lensGeometric
Phase as an angle on the Bloch sphere
On the Bloch sphere, is the equatorial point , so the relative phase is its azimuth, while a global phase moves nothing. A computational readout gives . The Hadamard gate exchanges the and axes, which turns the final readout into . Inserting rotates the azimuth by a quarter turn, so becomes and .
Worked example 8.1.7
(Separating from )
The ordinary readout gives for both phases, because cosine is even. Now insert before the final , so that . This equals for and for . The two preparations are different states that the first readout cannot separate. Together, the values of and fix modulo .
A coherent superposition is not a classical mixture
Compare a qubit prepared as with a classical procedure that prepares half the time and half the time. Both give a fair computational-basis measurement. But after , the pure state gives outcome 0 certainly, while each member of the mixture gives a fair outcome, so the mixture stays fair overall.
Proposition 8.1.8
(Superposition and mixture differ after interference)
Let and . Both give computational-basis probabilities . After one , the first gives and the second gives .
Proof
(Density matrices and the off-diagonal entry)
A pure state has density matrix . For our phase family and for the classical mixture respectively,
The diagonal entries give the same computational-basis probabilities, but the off-diagonal entries differ. Gate evolution is , because . For a Hermitian of trace one, the entry of is . This is for . Since , the mixture remains fair after . A global phase cancels already in .
Proof X-ray
Why only the off-diagonal entry matters
After the Hadamard, the probability of outcome 0 is . The diagonal entries enter only through their sum, which is one, so all dependence on the phase comes from the coherence . This is the key move. The classical coin has by construction: it has erased the relative phase, and a computational-basis measurement whose outcome is discarded applies the same erasure. For the pure family, positivity bounds the entry by , and the pure state attains this largest value. Purity is therefore the hypothesis that makes the interference as visible as it can be.
If an actual computational-basis measurement is made before the final Hadamard and its outcome is discarded, the phase-family preparation becomes this mixture. Continuing with then gives , not the unmeasured interference distribution. If the outcome is retained, each conditional state is or , and each also gives a fair result after . This explains why switching a readout location must not be interpreted as inserting a measurement midway through one run. That measured-and-continued circuit is a third experiment, distinct from either choice of a single readout location.
The probabilities describe repeated identical preparations. They do not promise exactly half zeros in every finite collection of trials, nor do they turn a single observed bit into a complete description of the state.
Solved exercises
Quick check
(Which of and represents the same pure state as ? Can their equal computational-basis probabilities decide the question?)
Compare the ratio of the second amplitude to the first in each state.
Solution · Answer
, so it differs only by global phase. For , the equal nonzero first amplitudes would force a global factor of 1, but the second amplitudes then disagree. It is a different state, although all three have probabilities . For example, applying and then maps to and to , distinguishing them. The same operations map to , with the same certain outcome as .
Quick check
(With relative phase () and , compute both final amplitudes and their squared magnitudes. What changes if only the global phase becomes ()?)
Add the amplitudes before squaring.
Solution · Answer
Since , the amplitudes are and . Their squared magnitudes are and . Changing the global phase multiplies both amplitudes by but leaves these two probabilities unchanged. The pre-Hadamard probabilities remain in both cases.
Quick check
(Someone argues that the circuit always returns because . Find a phase giving a counterexample, and compare an experiment with no intermediate measurement to one that measures in the computational basis before the last .)
The matrix sits between the two Hadamards.
Solution · Answer
Take . Without an intervening measurement, the final state is , so the certain outcome is 1. The equality does not allow us to remove the different matrix between the factors. If we actually measure beforehand, outcomes 0 and 1 each occur with probability , leaving the corresponding basis state. Either state followed by gives a fair final result, so the final distribution is instead. These are different experiments.
Quick check
(Compare a known preparation with a preparation that chooses or by a fair classical coin. Give the outcome distributions after and computational-basis measurement. If a single trial returns 0, does that certify that the source was ?)
Compute each distribution first, then ask which outcomes are possible under both.
Solution · Answer
The coherent preparation gives , while the classical mixture gives . A single outcome 0 is possible under both preparations and therefore does not certify the coherent one. In the ideal model an outcome 1 rules out , but repeated trials are needed to estimate an unknown source's distribution. The different distributions establish an observable distinction without claiming perfect identification from one arbitrary observation.