Phase estimation converts an eigenvalue into a measurement distribution. Its simplest illustration uses a phase that fits exactly into the control register; real phases need not do so. The calculation below shows what the circuit returns in that case and what resources are hidden inside a controlled power.
From an eigenphase to a Fourier sum
Definition 2.4.1
(Phase-estimation setting)
Let be a unitary on a finite-dimensional target register, and suppose a normalized state satisfies , where . Choose control qubits and write . Use the Fourier convention
Starting from , Hadamards on the control register and controlled powers produce
Here , and control bit controls . The inverse Fourier transform is the exact map in the integer-labeled computational basis defined above; a circuit convention that reverses output-bit order must include the corresponding swaps or reinterpret the measured bits before reading . The product of the phases is . Applying to the first register gives the amplitude of outcome :
If modulo one, every summand is one and . Otherwise the finite geometric-series identity gives
When the denominator vanishes, use the original sum, which gives probability one at the matching grid point. Probabilities sum to one because the circuit is unitary. The fractional phase is defined modulo one: an answer near may be close to a phase near .
A nearest-grid guarantee
Definition 2.4.2
(Circular distance)
For real the circular distance is , the distance on the circle on which phases live.
Theorem 2.4.3
(Nearest-grid guarantee)
Let be any grid point for which is smallest; if there is a tie, choose either. Then , and the probability of this particular outcome obeys
Proof
(Shift to a small offset, then bound the sine quotient)
Choose a representative with . The probability is unchanged by this integer shift. If , the probability is one. Otherwise, and the elementary bounds for and imply
Squaring proves the bound. It applies even at the wraparound between and .
Proof X-ray
Where the bound comes from
The proof has one reduction and one estimate. The reduction is the integer shift: for integer , both and have period one, so the probability is unchanged when moves by an integer, and this is why the wraparound between and is harmless. The estimate bounds the numerator below by the chord and the denominator above by . The hypothesis is exactly what the chord bound needs, and it is where “nearest” is used: for a farther grid point it fails. Both inequalities are nearly sharp when and is large, so the constant cannot be improved by this argument.
Repetition on fresh preparations can raise the chance of seeing a near-grid result, but extracting a final estimate from repeated outcomes requires a specified classical rule; the bound alone does not claim a particular rule succeeds.
Worked example 2.4.4
(A phase halfway between two grid points)
Take and , exactly halfway between and . For or ,
The two remaining outcomes each have probability . Thus either nearest outcome occurs with total probability about , while neither is guaranteed. The lower bound above concerns a particular nearest point; at a tie it does not double automatically without analyzing both events.
Worked example 2.4.5
(Wraparound near one)
Take and . The nearest grid point on the circle is , because is within of . With , . The most likely outcome reads as a phase near , which is close to on the circle although not on the interval .
Counterexample mode
More bits do not make every phase exact
A tempting claim says that with enough control qubits the circuit returns the phase exactly. The phase is not for any integers : a dyadic fraction has denominator a power of two, while is in lowest terms with denominator . So no binary control register has a grid point equal to it. For the nearest grid point has , not . The nearest-grid bound still gives probability at least for a particular nearest grid point, at circular distance at most from ; an exact binary output is impossible.
When the target is not an eigenstate
Let be an orthonormal eigenbasis of , with phases , and let the prepared target be . Linearity gives the joint state after inverse Fourier transform:
Theorem 2.4.6
(Outcome distribution for a general target)
With the notation above, the probability of observing is
Proof
(Orthogonality removes the cross terms)
The squared norm of the component of the joint state with first register is . Because the are orthonormal, this equals ; the cross terms with vanish.
It is a mixture of the eigenphase distributions, not an estimate of an average phase. Conditional on an outcome of nonzero probability, the target becomes the normalized vector . Distinct phases can both contribute to one outcome; finite precision does not generally project onto a single eigenspace. If several eigenvectors share an eigenvalue, phase estimation cannot distinguish vectors within that eigenspace.
Worked example 2.4.7
(Two eigenphases in one target)
For example, if , and the target is , a two-qubit control register reports and with probability each. The target collapses to or respectively. There is no output at the average of the phases and .
Common mistake
(Looking for the average phase)
A target that mixes the eigenphases and does not produce an outcome near their average . In the example above outcome has probability , because . Each run returns one eigenphase, never a blend, and the target collapses accordingly.
What is counted
The circuit assumes access to controlled as well as preparation of the target and an inverse Fourier transform. If a controlled is made by repeating controlled , then the control stage uses calls to controlled . In some applications those powers have a more efficient problem-specific implementation, but this must be established rather than assumed. The exact Fourier transform uses controlled rotations; approximating those over a fixed gate set introduces an additional implementation error. The distribution and guarantee here describe ideal gates and a perfectly prepared eigenstate.
Practice with solutions
Quick check
(With and eigenphase , give the output and its probability.)
Solution · Answer
The output is (binary ), with probability one.
Quick check
(With and phase , find the probability that the outcome is neither of the two nearest grid points.)
Solution · Answer
The two remaining probabilities sum to .
Quick check
(For the two-eigenvalue example above, replace by . Give the outcome probabilities.)
Solution · Answer
The outcomes are and with probabilities and . Orthogonality eliminates cross terms.
Quick check
(Why does using controlled-power blocks not automatically mean only calls to controlled ?)
Solution · Answer
If implemented by repeated controlled , the blocks cost calls each, totaling ; faster special implementations require separate justification.