SWAP gate

The SWAP gate maps $|x⟩|y⟩$ to $|y⟩|x⟩$.

The SWAP gate and local unitary operators preserve product states. The displayed identity moves every exchange past a pair of local gates, reducing any such two-qubit quantum circuit to a local operation, optionally followed by SWAP gate. On more wires the same argument leaves a permutation of the tensor factors. By induction, a fully factored input remains factored. Since universal quantum computation includes operations that create quantum entanglement from a product state, these gates cannot form a universal quantum gate set.

For two spin-one-half qubits, the SWAP gate acts as one on the triplet sector and minus one on the singlet sector. The displayed unitary squares to SWAP. An isotropic exchange interaction with integrated angle $\pi /2$ implements it up to a common phase. There are other square-root branches; a pulse decomposition must use the declared branch consistently.

Let $P_{-}$ project onto $(|01⟩-|10⟩)/\sqrt{2}$ and let $P_{+}=I-P_{-}$ project onto the symmetric sector. An isotropic exchange interaction, with a chosen common energy offset, gives the continuous family displayed above. A full exchange has $\theta =\pi$ and equals SWAP gate; the conventional half exchange has $\theta =\pi /2$ and equals $P_{+}+i{P}_{-}$. Reversing the sign of the exchange gives the same full exchange at $\theta =-\pi$ but the inverse half exchange at $\theta =-\pi /2$. Thus even continuous time-independent Hamiltonian evolution and unchanged $|00⟩,|11⟩$ do not let the full-time truth table determine a unique half-time gate. A chosen generator or spectral branch is necessary. The half-exchange eigenvalues $1,i$ are the conventional principal scalar square-root choices for $1,-1$; no analytic matrix-function branch across the negative real axis is being assumed.

Two square roots of the SWAP gate, separated and followed by local $z$ rotations, implement a controlled phase gate. Local $y$ rotations of its target convert it into a CNOT gate. The construction assumes isotropic exchange and independent one-qubit control; pulse errors and extra Hamiltonian terms change the implemented operation.

With $S=P_{+}+i{P}_{-}$ the conventional square root of the SWAP gate, $P=\mathrm{diag}(1,i)$ the S gate, and operators acting rightmost first, $S{Z}_{1}S=\mathrm{diag}(1,i,-i,-1)$. Therefore $H_{2}S{Z}_{1}S{P}_2{H}_2=|0⟩⟨0|\otimes X-i|1⟩⟨1|\otimes I$. Appending $X_2$ changes the zero-controlled flip into a one-controlled flip, and $P_1$ removes the control-dependent phase. This proves the displayed exact CNOT gate identity. Both the exchange branch and the phase gate square-root convention matter; changing either requires recalculating the correction.

Applying controlled-NOT gates with directions $1\to 2$, $2\to 1$, and $1\to 2$ mapsand therefore implements the SWAP gate.