Global and Relative Phase

A pure quantum state is written as a unit vector $|\psi\rangle$ , but that description carries one coordinate too many. Multiply the whole vector by a complex number of unit modulus, $|\psi\rangle \mapsto e^{i\gamma}|\psi\rangle$ , and nothing observable changes: every probability, every expectation value, every Bloch vector is identical. That redundant factor is the global phase, and its invisibility is not an accident of one experiment but a theorem — the Born rule sees a state only through the density operator $\rho=|\psi\rangle\langle\psi|$ , and $e^{i\gamma}$ cancels there exactly. Physical states are therefore not vectors but rays: points of projective Hilbert space, the Bloch sphere for a single qubit. The relative phase between the amplitudes of a superposition is the opposite story. In $\tfrac{1}{\sqrt2}(|0\rangle + e^{i\phi}|1\rangle)$ the angle $\phi$ is hidden from a computational-basis measurement — both outcomes stay at probability $\tfrac12$ — yet it is fully physical, sitting in the off-diagonal coherence $\rho_{01}=\tfrac12 e^{-i\phi}$ and revealed the instant a Hadamard rotates the measurement basis, where it drives an interference fringe $P(0)=\tfrac12(1+\cos\phi)$ . Relative phase is the quantity every quantum algorithm sculpts. This is the gate-and-circuit reading of that distinction. It is why a $z$ -rotation $R_z(\phi)$ and the diagonal phase gate $\mathrm{diag}(1,e^{i\phi})$ are the same physical operation on one qubit yet different unitaries once controlled; why the Z–Y decomposition quarantines an explicit $e^{i\alpha}$ so the three Euler angles carry all the visible action; and why a controlled operation is the machine that promotes an invisible global phase into an observable relative one — phase kickback, the engine beneath phase estimation, Deutsch–Jozsa, and Shor. Sorting global from relative is the central bookkeeping of circuit analysis and gate synthesis.

Key takeaways

A global phase $e^{i\gamma}$ on the whole state is physically undetectable: $e^{i\gamma}|\psi\rangle$ and $|\psi\rangle$ share one density operator $\rho$ and give identical statistics in every basis.
A relative phase between amplitudes, as in $\alpha|0\rangle+e^{i\phi}\beta|1\rangle$ , is real: it lives in $\rho_{01}$ and becomes visible after a basis change such as a Hadamard.
States are therefore rays in projective Hilbert space; on the Bloch sphere a global phase moves nothing while the relative phase $\phi$ is the azimuth of the Bloch vector.
The rotation $R_z(\phi)=e^{-i\phi/2}\mathrm{diag}(1,e^{i\phi})$ and the phase gate $\mathrm{diag}(1,e^{i\phi})$ differ only by a global phase — identical on one qubit, different once controlled.
A controlled operation converts a global phase into a relative phase on the control (phase kickback): the engine of phase estimation, and why a "controlled global phase" is a genuine $Z$ -type gate.
Key idea: quantum predictions depend only on $\rho=|\psi\rangle\langle\psi|$ ; a global phase leaves $\rho$ unchanged and is unphysical, while a relative phase changes $\rho$ off-diagonally and is observable — until a control makes even a global phase relative.

The core insight

Quantum predictions depend only on the density operator $\rho=|\psi\rangle\langle\psi|$ ; since $e^{i\gamma}|\psi\rangle$ has the same $\rho$ , a global phase is unphysical, whereas a relative phase between amplitudes changes $\rho$ off-diagonally and is therefore observable — and a controlled operation is what turns a hidden global phase into a measurable relative one.

Where this lives today

The global-versus-relative distinction is the daily bookkeeping of running quantum hardware. Superconducting and trapped-ion processors implement single-qubit $z$ -rotations "in software" as virtual- $Z$ gates: instead of driving the qubit, the control system simply advances the phase reference of every subsequent pulse, a zero-duration, zero-error operation introduced by McKay and co-workers in 2017 and now standard on IBM and Google devices. That trick works because $R_z(\phi)$ is only a relative-phase rotation — there is nothing physical to do except relabel the phase — and it is correct only as long as the compiler respects that a controlled $R_z$ still carries the very phase the virtual gate elides. In fault-tolerant compilation the same ledger governs cost: the Gottesman–Knill boundary and the $T$ -count that dominate resource estimates are, at bottom, statements about which relative phases a circuit must synthesize. And phase estimation — Kitaev's promotion of a kicked-back relative phase into a digitizable measurement — remains the algorithmic heart of Shor's factoring, quantum chemistry eigenvalue estimation, and quantum-simulation spectroscopy, all running on hardware that must track relative phases coherently while decoherence attacks exactly those off-diagonal $\rho_{01}$ terms. Dynamical decoupling, spin echoes, and error correction are, in this language, a continuous fight to preserve relative phase against a global-phase-blind environment that would randomize it.

The physical picture

A global phase is the setting of a clock sealed inside a box: every internal process runs identically whatever the offset, and no measurement reaching outside can read it. A relative phase is the difference between two such clocks brought together — it shows up the moment their signals interfere. Probabilities come from $|\text{amplitude}|^2$ , so a factor common to all amplitudes squares away as $|e^{i\gamma}|^2=1$ , while a phase that weights amplitudes differently survives in the cross-terms $\propto\cos\phi$ when paths recombine. That surviving cosine is interference, and it is the entire observable content of phase. The Bloch sphere is the screen of physical reality, and the global phase is a hidden dimension the screen never displays. The Hopf fibration makes this exact: the three-sphere $S^3$ of normalized qubit vectors projects onto the Bloch sphere $S^2$ , collapsing each circle of global phases to a single point. What survives the projection is two real numbers — the polar angle $\theta$ and the azimuthal angle $\phi$ — and that azimuth is the relative phase. The first figure looks straight down the $z$ -axis so the whole action of phase is a rotation in the equatorial plane.

Figure

Looking down the Bloch z-axis: the relative phase φ is the azimuth of the Bloch vector. The state (1/√2)(zero + e^{iφ} one) with φ = π/3 = 60° points to ⟨X⟩ = cos φ = 0.50, ⟨Y⟩ = sin φ = 0.87. A global phase e^{iγ} maps every vector to the same point, the Hopf fiber the sphere never shows

To make a global phase real, make it conditional. If a black box multiplies a target by a phase only when a control qubit is $|1\rangle$ , the control's two branches now differ by that phase — a relative phase a Hadamard turns into a visible probability. The second figure is the quantitative punchline: sweep the relative phase $\phi$ and read the interference fringe a Hadamard produces, the cosine that certifies $\phi$ is physical.

Figure

The interference fringe P_H(0) = ½(1 + cos φ) after a Hadamard on the state (1/√2)(zero + e^{iφ} one). At φ = 0 the paths add (P = 1, constructive); at φ = π they cancel (P = 0, destructive); at φ = π/3 the fringe reads P = 0.75. The relative phase is fully observable — the Z-basis, by contrast, would give a flat 0.5 line

The core equations

The first statement is the equivalence itself: two vectors differing by an overall unit-modulus factor are the same physical state, a ray rather than a vector.

|\psi'\rangle = e^{i\gamma}|\psi\rangle \quad\Longleftrightarrow\quad |\psi'\rangle \sim |\psi\rangle \qquad (1)

where: $|\psi\rangle$ — normalized pure state vector; $e^{i\gamma}$ — global phase factor, $\gamma\in[0,2\pi)$ ; $\sim$ — physical equivalence (same ray). Reading: the map $|\psi\rangle\mapsto e^{i\gamma}|\psi\rangle$ is a symmetry that changes the label but not the state. The reason is that a state enters every prediction only through its density operator, and the global phase cancels there identically.

\rho = |\psi\rangle\langle\psi|, \qquad \big(e^{i\gamma}|\psi\rangle\big)\big(e^{i\gamma}|\psi\rangle\big)^{\dagger} = e^{i\gamma}e^{-i\gamma}\,|\psi\rangle\langle\psi| = \rho \qquad (2)

where: $\rho$ — density operator, the complete physical description of a pure state; $\dagger$ — Hermitian conjugate. Reading: the ket picks up $e^{i\gamma}$ and the bra its conjugate $e^{-i\gamma}$ ; their product is $1$ , so $\rho$ — and every quantity built from it — is untouched. Fixing the global phase by convention leaves exactly two real parameters, and they are the Bloch angles: a polar $\theta$ and the relative phase $\phi$ .

|\psi\rangle = \cos\tfrac{\theta}{2}\,|0\rangle + e^{i\phi}\sin\tfrac{\theta}{2}\,|1\rangle \qquad (3)

where: $\theta$ — polar angle, $\theta\in[0,\pi]$ ; $\phi$ — relative (azimuthal) phase, $\phi\in[0,2\pi)$ . Reading: the amplitude of $|0\rangle$ is taken real and non-negative to spend the global-phase freedom; everything physical then sits in $\theta$ (populations) and $\phi$ (coherence). Relative phase is exactly what the diagonal gates write, in tick sizes $\pi$ , $\pi/2$ , $\pi/4$ .

Z=\begin{pmatrix}1&0\\0&-1\end{pmatrix},\quad S=\begin{pmatrix}1&0\\0&i\end{pmatrix},\quad T=\begin{pmatrix}1&0\\0&e^{i\pi/4}\end{pmatrix} \qquad (4)

where: $Z$ — Pauli- $Z$ , relative phase $\pi$ ; $S$ — phase gate, relative phase $\pi/2$ , $S=\sqrt{Z}$ ; $T$ — the $\pi/8$ gate, relative phase $\pi/4$ , $T=\sqrt{S}$ . Reading: each fixes $|0\rangle$ and phases $|1\rangle$ , so each is a pure $z$ -axis rotation of the Bloch vector by its tick angle. The continuous $z$ -rotation is the same physical operation as the diagonal phase gate, but written with a symmetric generator so it carries an explicit global phase.

R_z(\phi)=e^{-i\phi Z/2}=\begin{pmatrix}e^{-i\phi/2}&0\\0&e^{i\phi/2}\end{pmatrix}=e^{-i\phi/2}\begin{pmatrix}1&0\\0&e^{i\phi}\end{pmatrix} \qquad (5)

where: $R_z(\phi)$ — rotation by $\phi$ about the Bloch $z$ -axis; $e^{-i\phi/2}$ — the global phase distinguishing $R_z(\phi)$ from $\mathrm{diag}(1,e^{i\phi})$ . Reading: on one qubit the scalar $e^{-i\phi/2}$ is invisible, so $R_z(\phi)$ and the phase gate induce the identical azimuthal rotation by $\phi$ ; the difference matters only when the gate is controlled or when total circuit phase is tracked. The Z–Y decomposition makes the split structural: pull the global phase out front and the three Euler angles carry all the visible action.

U=e^{i\alpha}\,R_z(\beta)\,R_y(\gamma)\,R_z(\delta) \qquad (6)

where: $U$ — arbitrary single-qubit unitary in $U(2)$ ; $e^{i\alpha}$ — the explicit global phase; $\beta,\gamma,\delta$ — Euler angles giving the $SU(2)$ (Bloch) action. Reading: any $2\times2$ unitary is a $SU(2)$ rotation times one overall phase; the rotation is what a Bloch vector feels, the $e^{i\alpha}$ is bookkeeping until it is controlled. A Hadamard converts the hidden relative phase into a measured probability — the fringe.

H\,\tfrac{1}{\sqrt2}\big(|0\rangle+e^{i\phi}|1\rangle\big)=\cos\tfrac{\phi}{2}\,|0\rangle-i\sin\tfrac{\phi}{2}\,|1\rangle\ \ (\text{up to global phase}),\qquad P(0)=\tfrac12(1+\cos\phi) \qquad (7)

where: $H$ — Hadamard gate, mapping the $Z$ -basis to the $X$ -basis; $P(0)$ — probability of outcome $0$ after $H$ . Reading: the relative phase, flat in the $Z$ -basis, drives a full-contrast cosine fringe in the $X$ -basis. Finally, a controlled operation writes a global phase onto the control as a relative phase — phase kickback.

\mathrm{c}\text{-}U\,|c\rangle\!\otimes\!|u\rangle=e^{ic\theta}\,|c\rangle\!\otimes\!|u\rangle,\qquad U|u\rangle=e^{i\theta}|u\rangle \qquad (8)

where: $|c\rangle$ — control qubit, $c\in\{0,1\}$ ; $|u\rangle$ — eigenstate of $U$ with eigenvalue $e^{i\theta}$ ; $e^{ic\theta}$ — the phase kicked onto the control, now a relative phase between its $|0\rangle$ and $|1\rangle$ branches. Reading: a phase that is global on $|u\rangle$ becomes relative on the control, and a Hadamard reads it out.

Derivation ANo measurement can see a global phase

The goal is to prove — for the most general measurement quantum mechanics allows, not merely projective ones — that $|\psi\rangle$ and $e^{i\gamma}|\psi\rangle$ are experimentally indistinguishable, which is what makes "physical state = ray" a theorem rather than a convention. We take an arbitrary POVM, compute the outcome probability for the phase-shifted state, and watch the phase cancel; then we check that no prior unitary and no adaptive strategy can rescue it, so the invisibility is complete and permanent. Let $\{E_m\}$ be any POVM: $E_m\succeq 0$ with $\sum_m E_m=\mathbb{1}$ .

Step 1. Write the general Born rule for the original state, $P_\psi(m)=\langle\psi|E_m|\psi\rangle$ , which contains projective measurements as the special case $E_m=|m\rangle\langle m|$ . Why: starting from POVMs rather than projectors means the conclusion cannot be dodged by some cleverer generalized measurement.

Step 2. Insert the phase-shifted state $e^{i\gamma}|\psi\rangle$ : $P_{e^{i\gamma}\psi}(m)=\langle\psi|\,e^{-i\gamma}E_m\,e^{i\gamma}|\psi\rangle$ , where the bra contributes $e^{-i\gamma}$ and the ket $e^{i\gamma}$ . Why: the phase and its conjugate sit on opposite sides of the operator, exactly the arrangement that will cancel.

Step 3. The factors are c-numbers, so they commute through $E_m$ and multiply to $e^{-i\gamma}e^{i\gamma}=1$ , giving $P_{e^{i\gamma}\psi}(m)=\langle\psi|E_m|\psi\rangle=P_\psi(m)$ for every $m$ and every $\gamma$ . Why: identical probabilities for every outcome of every POVM is precisely the statement that no single measurement distinguishes the two vectors.

Step 4. Close the last loophole: any preparation-and-measurement protocol applies some unitary $V$ before measuring, which only replaces $E_m$ by $V^{\dagger}E_m V$ — still a POVM element — so Step 3 applies verbatim, and adaptive or repeated strategies are just sequences of such POVMs. Why: this upgrades "one measurement fails" to "no experiment whatsoever succeeds," establishing the ray picture. *Note: the cancellation $e^{-i\gamma}e^{i\gamma}=1$ is exact, and $V^{\dagger}E_mV\succeq0$ with $\sum_m V^{\dagger}E_mV=\mathbb{1}$ , confirming that conjugation by any unitary preserves the POVM structure.*

Go deeper · Derivation B · 1 minWhy a relative phase is observable: the interference fringe

Derivation BWhy a relative phase is observable: the interference fringe

The aim is to show that the very phase invisible to a computational-basis measurement completely controls the statistics after a single Hadamard, turning an abstract off-diagonal number into a measured cosine and proving the relative phase is physically real. We first confirm the $Z$ -basis blindness, then apply $H$ and use the Born rule, and finally read off the constructive and destructive extremes that a lab actually sees. Start from $|\psi\rangle=\tfrac{1}{\sqrt2}(|0\rangle+e^{i\phi}|1\rangle)$ with Hadamard $H=\tfrac{1}{\sqrt2}\left(\begin{smallmatrix}1&1\\1&-1\end{smallmatrix}\right)$ .

Step 1. In the computational basis $P_Z(0)=|\tfrac{1}{\sqrt2}|^2=\tfrac12$ and $P_Z(1)=|\tfrac{e^{i\phi}}{\sqrt2}|^2=\tfrac12$ , independent of $\phi$ . Why: the modulus-squared discards $e^{i\phi}$ , so in this basis the relative phase is hidden — it lives only in the coherence $\rho_{01}=\tfrac12 e^{-i\phi}$ , off the diagonal the $Z$ -measurement reads.

Step 2. Apply the Hadamard: $H|\psi\rangle=\tfrac12\big[(1+e^{i\phi})|0\rangle+(1-e^{i\phi})|1\rangle\big]$ . Why: $H$ rotates the measurement axis from $z$ to $x$ , mixing the two amplitudes so that their phase difference $\phi$ now appears inside the new amplitudes rather than hiding between them.

Step 3. Take the modulus squared using $|1+e^{i\phi}|^2=(1+\cos\phi)^2+\sin^2\phi=2+2\cos\phi$ : $P_H(0)=\tfrac14|1+e^{i\phi}|^2=\tfrac12(1+\cos\phi)$ , and likewise $P_H(1)=\tfrac12(1-\cos\phi)$ . Why: the interference between the two paths produces the cross-term $\cos\phi$ — the fringe — which is the observable signature of the relative phase.

Step 4. Read the extremes: $\phi=0$ gives $P_H(0)=1$ (fully constructive) and $\phi=\pi$ gives $P_H(0)=0$ (fully destructive), a swing from certainty to certainty as $\phi$ moves. Why: full-contrast fringes are direct laboratory evidence that $\phi$ is not a bookkeeping artifact but a physical property of the state. *Note: the simplification $\tfrac14(2+2\cos\phi)=\tfrac12(1+\cos\phi)$ is elementary, and the extremes $\phi=0,\pi$ reproduce $P_H(0)=1,0$ as expected.*

Go deeper · Derivation C · 1 minPhase kickback: a controlled operation makes a global phase relative

Derivation CPhase kickback: a controlled operation makes a global phase relative

The purpose is to derive the mechanism that promotes an unobservable phase into an observable one — a controlled- $U$ acting on an eigenstate of $U$ transfers the eigenphase onto the control register — because this single step is the working core of phase estimation and of every controlled-phase construction in the QFT. We prepare the control in a superposition and the target in an eigenstate, apply the controlled gate, and factor the target back out to expose the phase now riding on the control. Let $U|u\rangle=e^{i\theta}|u\rangle$ with the control in $|+\rangle$ .

Step 1. Write the joint input $|+\rangle|u\rangle=\tfrac{1}{\sqrt2}(|0\rangle+|1\rangle)|u\rangle$ , control in the equal superposition, target in the eigenstate. Why: the superposed control provides two branches whose eventual phase difference is exactly what will become measurable.

Step 2. Apply $\mathrm{c}\text{-}U=|0\rangle\langle0|\otimes\mathbb{1}+|1\rangle\langle1|\otimes U$ : the $|0\rangle$ branch is untouched and the $|1\rangle$ branch applies $U$ to the target, giving $\tfrac{1}{\sqrt2}\big(|0\rangle|u\rangle+|1\rangle\,U|u\rangle\big)$ . Why: only the control- $|1\rangle$ branch samples $U$ , so any phase $U$ carries is imprinted on that branch alone — the seed of a relative phase.

Step 3. Use the eigenvalue $U|u\rangle=e^{i\theta}|u\rangle$ and factor the target out: $\tfrac{1}{\sqrt2}\big(|0\rangle+e^{i\theta}|1\rangle\big)|u\rangle$ . The target returns unchanged; the eigenphase $e^{i\theta}$ — global on $|u\rangle$ — is now a relative phase on the control. Why: this is kickback: a controlled operation launders an invisible target phase into a visible control phase, readable by a Hadamard as in the fringe derivation.

Step 4. Specialize to a pure global phase $U=e^{i\gamma}\mathbb{1}$ , whose every state is an eigenstate with eigenvalue $e^{i\gamma}$ : then $\mathrm{c}\text{-}(e^{i\gamma}\mathbb{1})=\mathrm{diag}(1,1,e^{i\gamma},e^{i\gamma})$ acts as $\mathrm{diag}(1,e^{i\gamma})$ on the control. Why: the "unobservable" global phase of a gate becomes a genuine $Z$ -type phase gate on the control — which is why controlled- $R_z(\phi)$ and controlled- $\mathrm{diag}(1,e^{i\phi})$ are different operations. *Note: the algebra $\mathrm{c}\text{-}U|+\rangle|u\rangle=\tfrac{1}{\sqrt2}(|0\rangle+e^{i\theta}|1\rangle)|u\rangle$ and the controlled-global-phase reduction are exact.*

Limiting cases

How the two kinds of phase manifest across regimes:

Case	Result	Meaning
Pure global phase	$\vert\psi'\rangle=e^{i\gamma}\vert\psi\rangle \Rightarrow \rho'=\rho$	Completely unobservable: same density operator, same statistics in every basis, same Bloch vector.
Relative phase in its own basis	$\tfrac{1}{\sqrt2}(\vert0\rangle+e^{i\phi}\vert1\rangle),\ P_Z(0)=P_Z(1)=\tfrac12$	Hidden in the $Z$ -basis; the phase lives entirely in the coherence $\rho_{01}=\tfrac12 e^{-i\phi}$ .
Relative phase revealed by interference	$P_H(0)=\tfrac12(1+\cos\phi)$	After a Hadamard, $\phi$ sets the probability; $\phi=0$ and $\phi=\pi$ give opposite certainties.
$R_z(\phi)$ versus $\mathrm{diag}(1,e^{i\phi})$ on one qubit	differ by $e^{-i\phi/2}$ , act identically	The distinguishing factor is a global phase, invisible to any single-qubit measurement.
The same two, controlled	$\mathrm{c}\text{-}R_z(\phi)\ne \mathrm{c}\text{-}\mathrm{diag}(1,e^{i\phi})$	The formerly-global $e^{-i\phi/2}$ becomes a relative phase on the control — a real, measurable difference.
Controlled pure global phase	$\mathrm{c}\text{-}(e^{i\gamma}\mathbb{1})\to\mathrm{diag}(1,e^{i\gamma})_{\text{ctrl}}$	Phase kickback: the target's global phase becomes a $Z$ -type gate on the control.

Common trapA phase you can factor out is always global and always harmless

Not so. A phase is global only relative to the whole state at one instant; the moment that branch interferes with another, or is conditioned on a control, the "same" phase becomes relative and fully physical. Two traps follow. First, $R_z(\phi)=e^{-i\phi/2}\mathrm{diag}(1,e^{i\phi})$ and $\mathrm{diag}(1,e^{i\phi})$ are interchangeable on one qubit but their controlled versions are different unitaries — the discarded $e^{-i\phi/2}$ reappears as a relative phase on the control, so silently swapping them corrupts any phase-estimation or QFT circuit. Second, never carry a global phase into a probability expecting it to matter: within a single amplitude-squared it always cancels. The skill is knowing which situation you are in — isolated state (global, drop it) or interfering/controlled branch (relative, keep it).

Worked examples

1 · Reading a relative phase off the Bloch sphere and out of an interferometer (fully worked)

A qubit is prepared in $|\psi\rangle=\tfrac{1}{\sqrt2}(|0\rangle+e^{i\phi}|1\rangle)$ with $\phi=\pi/3$ . Find its density matrix, its Bloch vector, the computational-basis statistics, and the statistics after a Hadamard. Then say what a global phase $e^{i\pi/5}$ on the whole state would change.

Step 1. Amplitudes are $c_0=\tfrac{1}{\sqrt2}$ , $c_1=\tfrac{1}{\sqrt2}e^{i\phi}$ . The density matrix is $\rho=|\psi\rangle\langle\psi|=\tfrac12\left(\begin{smallmatrix}1&e^{-i\phi}\\e^{i\phi}&1\end{smallmatrix}\right)$ , so the populations are $\rho_{00}=\rho_{11}=\tfrac12$ and the coherence is $\rho_{01}=\tfrac12 e^{-i\phi}=\tfrac12 e^{-i\pi/3}$ — this off-diagonal number is where the relative phase lives.

Step 2. The Bloch components are $\langle X\rangle=2\,\mathrm{Re}(c_0^{*}c_1)=\cos\phi$ , $\langle Y\rangle=2\,\mathrm{Im}(c_0^{*}c_1)=\sin\phi$ , $\langle Z\rangle=|c_0|^2-|c_1|^2=0$ . At $\phi=\pi/3$ : $(\langle X\rangle,\langle Y\rangle,\langle Z\rangle)=(0.50,\,0.87,\,0)$ , an equatorial vector at azimuth $60°$ — exactly the arrow in the first figure. The azimuth is $\phi$ .

Step 3. Computational-basis measurement gives $P_Z(0)=|c_0|^2=\tfrac12$ and $P_Z(1)=|c_1|^2=\tfrac12$ : flat, $\phi$ invisible. The relative phase never entered because $|e^{i\phi}|^2=1$ .

Step 4. Apply a Hadamard and measure: $P_H(0)=\tfrac12(1+\cos\phi)=\tfrac12(1+\cos 60°)=\tfrac12(1+0.5)=\mathbf{0.75}$ , and $P_H(1)=0.25$ . The same $\phi$ that was invisible now biases the outcome three-to-one — this is the interferometer reading the phase.

Step 5. Now multiply the whole state by $e^{i\pi/5}$ : $c_0$ and $c_1$ both acquire the factor, so $\rho$ is unchanged (Step 1), the Bloch vector is unchanged (Step 2), and both $P_Z$ and $P_H$ are unchanged (Steps 3–4). A global phase changes nothing measurable; only the relative phase $\phi$ did any work.

2 · When the "invisible" global phase of $R_z$ becomes visible (fully worked)

The gates $S=\mathrm{diag}(1,i)$ and $R_z(\pi/2)=\mathrm{diag}(e^{-i\pi/4},e^{i\pi/4})$ differ only by a global phase. Show they act identically on one qubit, then build their controlled versions and show these are genuinely different — and measurably so.

Step 1. Factor the rotation: $R_z(\pi/2)=e^{-i\pi/4}\,\mathrm{diag}(1,e^{i\pi/2})=e^{-i\pi/4}\,\mathrm{diag}(1,i)=e^{-i\pi/4}S$ . So $R_z(\pi/2)$ and $S$ differ by the global phase $e^{-i\pi/4}$ , and on any single qubit they induce the identical $90°$ azimuthal rotation — no measurement separates them.

Step 2. Control on the first qubit, ordering $\{|00\rangle,|01\rangle,|10\rangle,|11\rangle\}$ . The controlled- $S$ is $\mathrm{c}\text{-}S=\mathrm{diag}(1,1,1,i)$ , while $\mathrm{c}\text{-}R_z(\pi/2)=\mathrm{diag}\big(1,1,e^{-i\pi/4},e^{i\pi/4}\big)$ . These are different unitaries: the global $e^{-i\pi/4}$ of $R_z$ has landed on both control- $|1\rangle$ entries, i.e. as a relative phase between the control's $|0\rangle$ and $|1\rangle$ branches.

Step 3. Make the difference measurable. Take control $|+\rangle$ , target $|0\rangle$ . Under $\mathrm{c}\text{-}S$ the target's amplitude is unchanged in both branches ( $S|0\rangle=|0\rangle$ ), so the state stays $|+\rangle|0\rangle$ : a Hadamard on the control gives $P(0)=1$ . Under $\mathrm{c}\text{-}R_z(\pi/2)$ the control- $|1\rangle$ branch multiplies the target by $R_z(\pi/2)|0\rangle=e^{-i\pi/4}|0\rangle$ , giving $\tfrac{1}{\sqrt2}(|0\rangle+e^{-i\pi/4}|1\rangle)|0\rangle$ — a relative phase $-\pi/4$ kicked onto the control.

Step 4. Read it: with control phase $-\pi/4$ , a Hadamard gives $P(0)=\tfrac12(1+\cos\tfrac{\pi}{4})=\tfrac12(1+0.7071)=\mathbf{0.8536}=\cos^2(\pi/8)$ , versus $P(0)=1$ for $\mathrm{c}\text{-}S$ . The two gates that were indistinguishable on one qubit now give different measurement statistics — the exact reason a compiler must never substitute $R_z(\phi)$ for $\mathrm{diag}(1,e^{i\phi})$ under a control.

Physics of the portrait

The distinction between a phase you can gauge away and a phase that drives an interference fringe is not a computing nicety — it is the physical lesson of one of the twentieth century's most surprising effects.

Origins · 1 minYakir Aharonov and how this was discovered

Yakir Aharonov

born 1932, Haifa · the Aharonov–Bohm effect and the physical reality of quantum phase

Yakir Aharonov gave quantum phase its sharpest physical demonstration. In 1959, working with his doctoral supervisor David Bohm in Bristol, he predicted what is now the Aharonov–Bohm effect: a charged particle travelling through a region of zero magnetic field can still be shifted by the electromagnetic potential, provided its two interfering paths enclose a confined flux. Classically the potential is a mathematical convenience fixed only up to a gauge — an overall, "global" freedom with no consequence. Quantum mechanically each path accumulates a phase $\tfrac{q}{\hbar}\!\int \mathbf{A}\cdot d\mathbf{l}$ , and while that phase is gauge-dependent on any single path, the relative phase between two paths that encircle the flux is gauge-invariant, physical, and directly visible as a shift of the interference fringes. It is precisely the theme of this article — a phase that looks removable becomes observable the moment it is made relative around a loop — realized in the laboratory of electrons and magnetic flux rather than qubits and controlled gates. The prediction, initially controversial, was confirmed by Robert Chambers in 1960 and definitively by Akira Tonomura's electron-holography experiments in the 1980s. Aharonov went on to reshape how physicists think about quantum measurement, introducing (with Bergmann and Lebowitz) time-symmetric quantum mechanics and, with Albert and Vaidman in 1988, the theory of weak values. The through-line of his work is a single conviction that this article shares: in quantum mechanics the phase is not scaffolding to be discarded but a physical quantity to be measured — you only have to make it relative.

History

1926

— Born's rule ties probabilities to squared amplitudes, making an overall phase unobservable

1927

— Weyl and Wigner frame quantum states as rays in projective Hilbert space

1931

— Hopf describes the fibration S³ → S² that later models the qubit's global-phase freedom

1959

— Aharonov and Bohm predict an observable phase shift from potentials in a field-free region

1985

— Berry isolates the geometric (relative) phase acquired under adiabatic cyclic evolution

1995

— Kitaev's phase-estimation algorithm turns kickback of a relative phase into a measurement primitive

Going further

Texts. Nielsen & Chuang, Quantum Computation and Quantum Information, §1.3.1, §2.2.7, §4.2 (single-qubit gates and the Z–Y decomposition), §5.2 (phase estimation and kickback) · Preskill, Lecture Notes on Quantum Computation, Ch. 2–3 · Watrous, The Theory of Quantum Information, Ch. 2 · Wilde, Quantum Information Theory, 2nd ed., Ch. 3 · Kitaev, Shen & Vyalyi, Classical and Quantum Computation, §8 · Mermin, Quantum Computer Science, Ch. 1–2 · Aharonov & Bohm, "Significance of electromagnetic potentials in the quantum theory," Phys. Rev. 115, 485 (1959). Key numerical checks: $\rho$ is invariant under $e^{i\gamma}$ ; $P_H(0)=\tfrac12(1+\cos\phi)$ from $H|\psi\rangle$ , giving $0.75$ at $\phi=\pi/3$ ; $R_z(\pi/2)=e^{-i\pi/4}S$ and hence $\mathrm{c}\text{-}S=\mathrm{diag}(1,1,1,i)\ne\mathrm{c}\text{-}R_z(\pi/2)=\mathrm{diag}(1,1,e^{-i\pi/4},e^{i\pi/4})$ , with control readout $\cos^2(\pi/8)\approx0.8536$ ; and the kickback identity $\mathrm{c}\text{-}U|+\rangle|u\rangle=\tfrac{1}{\sqrt2}(|0\rangle+e^{i\theta}|1\rangle)|u\rangle$ together with its $U=e^{i\gamma}\mathbb{1}$ specialization.

Prerequisites. The Quantum Circuit Model · Bloch Sphere Representation of Gates · Pauli Gates (X, Y, Z). Leads to. Phase Gates (S and T) · Single-Qubit Rotation Gates · Z-Y Decomposition of Single-Qubit Gates · Controlled-U Gates.

Ask the tutor: ask me to track which phases in your circuit are global versus relative, work an interference or phase-kickback readout with real numbers, or show why a controlled $R_z$ differs from a controlled phase gate.

If a global phase is truly unobservable, why not just drop it from every calculation?

On an isolated state you should — it never affects a probability, an expectation value, or a Bloch vector, so carrying it is pure clutter. The danger is that "global" is a statement about a whole state at one moment, and quantum circuits routinely change the context. The instant a state becomes one branch of a superposition — the control- $|1\rangle$ branch of a controlled gate, or one arm of an interferometer — a phase that was global on that branch becomes relative to the other branch, and now it is fully physical. So the correct rule is contextual: drop a phase when it multiplies the entire state you will ever measure, but keep it the moment that state might be conditioned or interfered. Compilers encode exactly this rule, which is why they track an overall phase symbolically rather than discarding it outright.

What is the difference between R_z(φ) and the phase gate diag(1, e^{iφ}), and does it ever matter?

They differ by the global phase $e^{-i\phi/2}$ : $R_z(\phi)=e^{-i\phi/2}\,\mathrm{diag}(1,e^{i\phi})$ . On a single qubit that factor is invisible, so the two gates produce the identical $90°$ -per- $\pi/2$ azimuthal rotation of the Bloch vector and no measurement can tell them apart. The difference becomes real under control. The controlled- $R_z(\phi)$ places $e^{-i\phi/2}$ on both control- $|1\rangle$ diagonal entries, which is a relative phase between the control's branches; the controlled- $\mathrm{diag}(1,e^{i\phi})$ does not. They are different unitaries with different measurement statistics, so substituting one for the other inside a controlled construction — a QFT, a phase-estimation subroutine — silently changes the answer. This is one of the most common subtle bugs in hand-compiled circuits.

Why does the density matrix make the global-versus-relative distinction so cleanly?

Because the density matrix keeps exactly the physical information and throws away exactly the redundant phase. Forming $\rho=|\psi\rangle\langle\psi|$ pairs the ket's $e^{i\gamma}$ with the bra's $e^{-i\gamma}$ , so any global phase annihilates itself and never appears in $\rho$ . Meanwhile a relative phase between amplitudes shows up as a complex off-diagonal entry: for $\tfrac{1}{\sqrt2}(|0\rangle+e^{i\phi}|1\rangle)$ you get $\rho_{01}=\tfrac12 e^{-i\phi}$ , the coherence. The diagonal entries are the populations (measurement probabilities in the $Z$ -basis); the off-diagonal entries are the coherences (interference capacity). So $\rho$ literally separates the two: everything on the diagonal is what a $Z$ -measurement sees, everything off it is relative phase waiting to be revealed by a basis change — and nothing in $\rho$ records a global phase at all.

How does phase estimation turn an invisible phase into a number I can read?

By repeatedly making it relative and then interfering it. A unitary $U$ imprints an eigenphase $e^{i\theta}$ on its eigenstate $|u\rangle$ — global there, invisible. Put a control in $|+\rangle$ and apply a controlled- $U$ : the phase kicks back, leaving the control in $\tfrac{1}{\sqrt2}(|0\rangle+e^{i\theta}|1\rangle)$ , a relative phase. One control qubit and a Hadamard already extract one bit of $\theta$ through the fringe $P(0)=\tfrac12(1+\cos\theta)$ . Phase estimation scales this up: a register of control qubits applies controlled powers $U^{2^k}$ , so each control acquires a relative phase $2^k\theta$ , and the inverse quantum Fourier transform interferes the whole register to read $\theta$ to many bits at once. The entire algorithm is a machine for promoting one hidden global eigenphase into a digitized relative phase — the reason it sits under Shor's algorithm and quantum eigenvalue estimation.

Is the relative phase the same thing as Berry's geometric phase?

They are related but not identical. The relative phase discussed here is the total phase difference between two amplitudes, however it arose. A geometric (Berry) phase is a specific contribution to a relative phase that depends only on the path a state traces through projective Hilbert space, not on how fast or by what dynamics — it is what survives after the dynamical part is stripped away. The Aharonov–Bohm phase is a geometric phase of exactly this kind: it depends on the enclosed flux (a topological feature of the loop) rather than on the particle's speed. In a qubit, when you steer the Bloch vector slowly around a closed loop, it picks up a geometric phase equal to half the solid angle enclosed, on top of any dynamical phase. So geometric phase is a particularly robust source of relative phase; all the observability arguments of this article apply to it, and its path-dependence is what makes holonomic and topological quantum computing attractive.

Why is decoherence described as the loss of relative phase rather than global phase?

Because the environment attacks exactly the off-diagonal coherences where relative phase lives. Model dephasing as the environment randomly sampling a $z$ -phase: it multiplies $\rho_{01}$ by $\langle e^{i\delta}\rangle$ , an average over random $\delta$ that shrinks toward zero, while leaving the populations $\rho_{00},\rho_{11}$ untouched. The state slides from a definite superposition toward a classical mixture — the Bloch vector collapses onto the $z$ -axis — as the relative phase is scrambled. A global phase, having no physical presence in $\rho$ at all, is not something the environment can "destroy"; there is nothing there to lose. This is why $T_2$ (dephasing time) is quoted separately from $T_1$ (relaxation), why coherence is the resource error correction protects, and why dynamical decoupling refocuses relative phase the way a spin echo refocuses precession.

Global and Relative Phase

Where this lives today

The physical picture

The core equations

Limiting cases

Worked examples

1 · Reading a relative phase off the Bloch sphere and out of an interferometer (fully worked)

2 · When the "invisible" global phase of $R_z$ becomes visible (fully worked)

Physics of the portrait

Going further

Before this

Where this leads

Global and Relative Phase

Where this lives today

The physical picture

The core equations

Limiting cases

Worked examples

1 · Reading a relative phase off the Bloch sphere and out of an interferometer (fully worked)

2 · When the "invisible" global phase of RzR_zRz​ becomes visible (fully worked)

Physics of the portrait

Going further

Before this

Where this leads

2 · When the "invisible" global phase of $R_z$ becomes visible (fully worked)