Quantum Phase Estimation turns an eigenvalue hidden as a complex phase into an ordinary binary estimate. It works by creating a phase gradient with controlled powers of a unitary, then reading that gradient with the inverse Quantum Fourier Transform.
Guide 9 showed that the inverse QFT converts a structured phase ramp into a basis label. Quantum phase estimation manufactures exactly such a ramp from repeated controlled applications of a unitary operator. The result is one of the most important mathematical subroutines in quantum algorithms because eigenvalues appear throughout physics, linear algebra, number theory and dynamical systems.
Eigenvalue phase → controlled powers → binary phase ramp → inverse QFT → measured estimate.
1. The eigenphase problem
Let U be unitary and suppose |u⟩ is one of its eigenvectors:
U|u⟩=e^{2πiφ}|u⟩,
where 0≤φ<1. Because unitary eigenvalues have magnitude one, each eigenvalue lies on the complex unit circle and can be written as a phase e^{2πiφ}.
The goal of phase estimation is to estimate φ.
2. Why unitary eigenvalues are phases
If U|u⟩=λ|u⟩ and U is unitary, norm preservation gives
||U|u⟩||=||λ|u⟩||=|λ| ||u||=||u||.
For non-zero |u⟩, this implies |λ|=1. Therefore λ=e^{iθ} for some real θ. Writing θ=2πφ places the phase on the unit interval.
3. The controlled-U operation
A controlled-U acts on a control qubit and a target register. If the control is |0⟩, nothing happens. If the control is |1⟩, U is applied to the target.
If the target is already the eigenstate |u⟩, then
|0⟩|u⟩ → |0⟩|u⟩
and
|1⟩|u⟩ → e^{2πiφ}|1⟩|u⟩.
The target state remains |u⟩; the eigenvalue appears as a phase on the control branch.
4. Phase kickback
Prepare the control in |+⟩=(|0⟩+|1⟩)/√2. Then controlled-U gives
(1/√2)(|0⟩+|1⟩)|u⟩ → (1/√2)(|0⟩+e^{2πiφ}|1⟩)|u⟩.
The eigenphase has been kicked back onto the control qubit as a relative phase. This is one of the most reusable patterns in quantum algorithms.
If the target is an eigenstate, controlled application can transfer eigenvalue information into control-register phase without changing the target direction.
5. One control qubit gives only coarse phase information
After phase kickback, applying a Hadamard to the control produces probabilities that depend on φ. The probability of measuring 0 is
cos²(πφ),
and the probability of 1 is sin²(πφ). This contains phase information but not a direct multi-bit estimate. To obtain t bits of phase resolution, use t control qubits and controlled powers of U.
6. Why controlled powers appear
If
U|u⟩=e^{2πiφ}|u⟩,
then
U^{2^j}|u⟩=e^{2πi2^jφ}|u⟩.
Each doubling of the exponent doubles the phase angle modulo one. This is the quantum analogue of reading successive binary digits through powers of two.
7. Build the t-qubit control register
Start with t control qubits in |0⟩^{⊗t} and the target eigenstate |u⟩. Apply Hadamards to the control register:
(1/√(2^t))Σ_{k=0}^{2^t-1}|k⟩|u⟩.
Then apply controlled U^{2^j} operations according to the binary bits of k. The control register becomes
(1/√(2^t))Σ_{k=0}^{2^t-1}e^{2πikφ}|k⟩.
The target remains |u⟩. All the unknown phase information has been encoded into a Fourier-like phase gradient across the control basis.
8. The inverse QFT reads the phase register
Compare the control state
(1/√N)Σ_k e^{2πikφ}|k⟩
with Guide 9’s Fourier state
QFT_N|m⟩=(1/√N)Σ_k e^{2πikm/N}|k⟩.
If φ=m/N exactly, the two expressions are identical. Therefore the inverse QFT maps the control register exactly to |m⟩.
9. Exact binary phases
Suppose t=3 and
φ=3/8=0.011₂.
Because N=2^3=8, the phase is exactly on the Fourier grid. After controlled powers, the control register is QFT_8|3⟩. Applying the inverse QFT returns |3⟩=|011⟩ with probability one.
Worked example: diagonal unitary
Let
U=diag(1,e^{2πi(3/8)}).
The state |1⟩ is an eigenvector with eigenphase 3/8. Three-bit phase estimation using |1⟩ as the target returns 011 exactly in the ideal model.
10. What happens when φ is between grid points
If φ cannot be represented exactly with t binary digits, the inverse QFT does not produce one basis state with certainty. Instead, the probability is concentrated near integers m satisfying
m/2^t≈φ.
The output amplitudes are finite geometric sums. If δ=φ-m/N, then the amplitude for m is proportional to
Σ_{k=0}^{N-1}e^{2πikδ}.
When δ is small, the phase increments are nearly aligned and the probability is large. Farther bins cancel more strongly.
11. Precision versus number of control qubits
With t control qubits, the Fourier grid has spacing 1/2^t. Increasing t therefore increases phase resolution exponentially in the number of control bits, but the circuit also requires higher powers of U such as U^{2^{t-1}}.
Precision is therefore not free. The cost depends critically on whether controlled powers of U can be implemented efficiently.
12. The eigenstate requirement
The clean derivation assumes the target register starts in an eigenstate of U. But suppose instead
|ψ⟩=Σ_j c_j|u_j⟩
where |u_j⟩ are eigenvectors with phases φ_j. By linearity, phase estimation creates a superposition of phase estimates correlated with those eigenstates.
Measuring the control register yields an estimate near one of the φ_j, with weight related to |c_j|², and correspondingly projects the target toward the associated eigenspace in the idealised case.
Phase estimation can act simultaneously as eigenvalue estimation and probabilistic eigenstate filtering.
13. Phase estimation as spectral analysis
In ordinary linear algebra, spectral decomposition asks for eigenvalues and eigenvectors. Phase estimation performs a quantum version of spectral interrogation when the relevant operator is unitary or has been encoded into a unitary.
If a Hermitian operator H generates
U=e^{-iHt},
then an energy eigenvalue E becomes a unitary phase e^{-iEt}. Estimating that phase can therefore reveal spectral information about H, subject to scaling and phase-wrapping constraints.
14. Phase wrapping and aliasing
Phases are measured modulo one. The eigenvalues e^{2πiφ} and e^{2πi(φ+n)} are identical for every integer n. Therefore any physical quantity encoded into phase must be scaled so that the relevant range can be reconstructed without ambiguity.
This is analogous to aliasing in classical sampling: periodic encoding loses absolute information beyond the chosen phase interval.
15. Controlled powers are the real resource question
The compact circuit diagram of QPE can hide its main cost. To obtain t bits, the algorithm requires controlled U,U²,U⁴,...,U^{2^{t-1}}. If these powers can be produced by repeated squaring, modular arithmetic structure, Hamiltonian simulation or another efficient construction, QPE can be powerful. If implementing high powers is expensive, the apparent precision advantage may disappear.
16. Relation to order finding
In order-finding problems, a modular multiplication operator has eigenvalues whose phases encode rational numbers with denominators related to the desired period. Phase estimation samples approximations to those rational phases. Classical continued-fraction processing can then reconstruct the period under suitable conditions.
This is the mathematical bridge from the QFT to Shor-type factoring algorithms: the quantum part estimates spectral phase, and the classical part interprets the result as number-theoretic period information.
17. Relation to Hamiltonian eigenvalues
Suppose H|E_j⟩=E_j|E_j⟩ and define
U=e^{-iHτ}.
Then
U|E_j⟩=e^{-iE_jτ}|E_j⟩.
Estimating the eigenphase can estimate E_jτ modulo 2π. Choosing τ and the energy scaling correctly is therefore part of the mathematical design.
18. Iterative phase estimation
The textbook algorithm uses a t-qubit control register and an inverse QFT. Iterative variants recycle one control qubit, extracting phase bits sequentially with classical feed-forward and adaptive rotations.
The mathematical information is the same, but the resource trade-off changes: fewer ancilla qubits, more sequential depth and classical control.
19. Statistical estimation still matters
Real devices are noisy and finite-shot. Even when the ideal algorithm predicts a sharply concentrated binary output, hardware errors and approximate gates broaden the observed distribution. Reliable phase estimation therefore requires repeated runs, calibration and statistical inference.
Quantum mathematics does not eliminate estimation theory. It changes the experiment whose statistics are being analysed.
20. Common misconception: QPE reveals every eigenvalue of U in one run
No. If the target is one eigenstate, one run produces one phase estimate associated with that eigenstate. If the target is a superposition, the run samples one spectral component according to the state’s overlap weights. Mapping an entire spectrum generally requires repeated preparation and sampling.
21. Common misconception: more control qubits automatically mean accurate answers
More qubits refine the ideal Fourier grid, but practical accuracy also depends on controlled-U fidelity, inverse-QFT accuracy, target-state preparation, decoherence and sampling. Formal resolution is only one part of total error.
22. Common misconception: phase estimation is only about quantum physics
The mathematics is spectral. Whenever a useful problem can be encoded into eigenphases of an efficiently controlled unitary, phase estimation may become relevant. Physics supplies many natural operators, but number theory and algorithm design supply others.
23. Worked synthesis problem
Let U have eigenstate |u⟩ with eigenvalue
e^{2πi(5/8)}.
Use three control qubits.
Step 1: Binary phase. 5/8=0.101₂.
Step 2: Control-register state after controlled powers.
(1/√8)Σ_{k=0}^{7}e^{2πik(5/8)}|k⟩.
This equals QFT_8|5⟩.
Step 3: Apply inverse QFT. The control register becomes |5⟩.
Step 4: Measure. The ideal output is 101 with probability one.
Step 5: Verify using powers. The highest control sees phase 4φ=20/8=2+1/2, equivalent modulo one to 1/2. Powers expose successive binary structure modulo one.
24. Practice set
- Why can a unitary eigenvalue be written
e^{2πiφ}? - Explain phase kickback in one sentence.
- If
U|u⟩=e^{2πiφ}|u⟩, what isU^8|u⟩? - What state does the control register have immediately before the inverse QFT?
- If t=4, what is the phase grid spacing?
- Convert
0.1011₂to a fraction. - What happens if the input target is a superposition of eigenstates?
- Why must phase wrapping be considered when encoding energy?
- What is the main implementation cost hidden inside QPE?
- Why is QPE naturally connected to spectral problems?
Answers
- Unitary eigenvalues have magnitude one.
- An eigenvalue phase appears as relative phase on a control branch when a controlled unitary acts on an eigenstate target.
e^{2πi8φ}|u⟩.(1/√(2^t))Σ_ke^{2πikφ}|k⟩.1/16.11/16.- The phase register becomes correlated with eigencomponents and measurement samples one according to overlap weights.
- Eigenphases are defined modulo one, so different raw values can encode the same phase.
- Efficient controlled powers of U.
- It estimates phases of unitary eigenvalues.
25. Series navigation
- Guide 9: Quantum Fourier Transform, Roots of Unity and Interference
- Guide 10: Quantum Phase Estimation, Eigenphases and Controlled Powers
- Guide 11: Amplitude Amplification, Grover Rotations and Search Geometry
- Guide 12: Stabilizer Mathematics, Pauli Groups, Syndromes and Quantum Error Correction
- Return to the BTT Mathematics Learning Hub
Educational note: practical phase-estimation performance depends on the implementation of controlled powers, state preparation, gate precision, coherence time and statistical post-processing.