Phase Kickback

Phase kickback arises from the structure of controlled unitary operations. When a controlled-$U$ gate acts on a control qubit and a target qubit, and the target is in an eigenstate $|v\rangle$ of $U$ with eigenvalue $e^{i\theta}$, the gate applies $e^{i\theta}$ as a global phase on the $|v\rangle$ branch. If the control qubit is in a superposition $a|0\rangle + b|1\rangle$, the result is $a|0\rangle|v\rangle + b\, e^{i\theta}|1\rangle|v\rangle$: the target state is unchanged, but the eigenvalue phase now multiplies the $|1\rangle$ amplitude of the control. From the control qubit’s perspective, it has picked up a relative phase of $e^{i\theta}$ without the target qubit changing at all; the phase has “kicked back.”

A concrete example illustrates the mechanism. Prepare the control qubit in $|+\rangle = (|0\rangle + |1\rangle) / \sqrt{2}$ and the target qubit in $|1\rangle$. Apply a CNOT (controlled-X). Since $|1\rangle$ is an eigenstate of $X$ with eigenvalue $-1$, the kickback leaves the target in $|1\rangle$ while the control becomes $(|0\rangle - |1\rangle) / \sqrt{2} = |-\rangle$. Measuring the control in the Hadamard basis now yields a deterministic outcome encoding the eigenvalue. This two-qubit circuit is the seed from which much of quantum algorithm design grows.

Grover’s search algorithm uses phase kickback through its oracle. The oracle marks a solution state $|x^*\rangle$ by applying a phase flip of $-1$ to it, implemented by preparing an ancilla qubit in $|-\rangle = (|0\rangle - |1\rangle) / \sqrt{2}$ and applying a controlled-X that flips the ancilla when the input is $|x^*\rangle$. Because $|-\rangle$ is an eigenstate of $X$ with eigenvalue $-1$, the kickback writes $-1$ onto the $|x^*\rangle$ amplitude in the search register while leaving the ancilla unchanged in $|-\rangle$. The ancilla can be reused across every Grover iteration, so the cost of marking is just the oracle query itself.

Quantum phase estimation (QPE) extends this idea across multiple control qubits. Each control qubit $k$ is prepared in $|+\rangle$ and controls $U^{2^k}$. The kickback writes $e^{i \cdot 2^k \theta}$ onto control qubit $k$, so after all controlled applications the control register encodes the binary expansion of $\theta / 2\pi$ in its phases. A final inverse quantum Fourier transform converts those phases into a measurement outcome that reads out $\theta$ to $n$ bits of precision using $n$ ancilla qubits. The requirement that the target register start in an eigenstate of $U$ is what makes exact kickback possible; approximate eigenstates introduce estimation error proportional to the deviation from the true eigenstate.