Given a decomposition for a unitary $U$, how do you decompose the corresponding controlled unitary gate $C(U)$?

Suppose we have a circuit decomposition of a unitary $U$ using some universal gate set (for example CNOT-gates and single qubit unitaries). Is there a direct way to write down the circuit of the corresponding controlled unitary $C_U$ using the same universal gate set?

For example take $U=i Y = H X H X$, as a circuit: We can replace the $X$ gates by $C_X$ (CNOT) gates to obtain $C_U$: This works because if the control qubit is in state $|0\rangle$ the action on the target is $H^2=\mathbb{I}$, while for $|1\rangle$ it applies the circuit for $U$. For different $U$, in particular if it acts on several qubits, coming up with such a circuit might be cumbersome. Is there a recipe to obtain the circuit of $C_U$ given that you know how to build $U$?

• are you asking how to build a C-U out of an arbitrary one-qubit U? A method to do that can be found in chapter 4 of N&C (see e.g. figure 4.6 in the last edition), which is basically the generalisation of the decomposition you showed
– glS
Mar 26 '18 at 17:05
• @glS oh wow, I was not aware of that. Looks exactly like my example. Good to see how it implements the phase $\alpha$. But they don't seem to discuss the generalization to more target qubits? Mar 26 '18 at 18:47

The question may not be entirely well-defined, in the sense that to ask for a way to compute $C(U)$ from a decomposition of $U$ you need to specify the set of gates that you are willing to use. Indeed, it is a known result that any $n$-qubit gate can be exactly decomposed using $\text{CNOT}$ and single-qubit operations, so that a naive answer to the question would be: just decompose $C(U)$ using single-qubit and $\text{CNOT}$s.

A different interpretation of the question is the following: given $U$, can I compute $C(U)$ using a set of single-qubit operations and $\text{CNOT}$s not on the control qubit, and $\text{CNOT}$s with the control being the first qubit? This can be done generalising a result found in chapter four of Nielsen & Chuang.

Let $U$ be a single-qubit gate. It can then be proved that $U$ can always be written as $U = e^{i\alpha} AXBXC$, where $X$ is the Pauli X gate, and $A, B$ and $C$ are single-qubit operations such that $ABC=I$ (see N&C for a proof). It follows that $$C(U)=\Phi_1(\alpha)A_2C(X)B_2C(X) C_2,$$ where $\Phi_1(\alpha)\equiv\begin{pmatrix}1&0\\0&e^{i\alpha}\end{pmatrix}\otimes I$ is a phase gate applied to the first qubit, and $A_2, B_2, C_2$ are $A, B, C$ applied to the second qubit. This is immediate once you realise that, if that first qubit is $|0\rangle$, then $C(X)$ becomes an identity, and on the second qubit you have the operations $ABC$, which give the identity. On the other hand, if the first qubit is $|1\rangle$, then on the second rail you have $AXBXC$, which (together with the phase) equals $U$ by definition.

The above decomposition can be used to find a naive way to compute $C(U)$ for a general $n$-qubit unitary gate. The main observation is that if $U=A_1 A_2\cdots A_m$ for any set of gates $\{A_1,..,A_m\}$, then $$C(U)=C(A_1)C(A_2)\cdots C(A_m).$$ But we also know that any $n$-qubit $U$ can be decomposed in terms of CNOTs and single-qubit operations. It follows that $C(U)$ is a sequence of CCNOT and $C(V)$ operations, where CCNOT is here an $X$ gate applied to some qubit conditioned to two other qubits being $|1\rangle$, and $V$ is a single-qubit operation on some qubit. But again, any CCNOT operation (also called Toffoli), can be decomposed as shown in Figure 4.9 in N&C, and the $C(V)$ are decomposed as shown in the first part of the answer.

This method allows decomposing a general $n$-qubit unitary gate $U$ using only $\text{CNOT}$ and single-qubit gates. You may then go further and generalise this to find a decomposition for the case of multiple control qubits. For this you only now need a way to decompose the Toffoli gates, which is again found in Figure 4.9 of N&C.

• I think that's what I was looking for. Just to make sure: Lets say the known decomposition $U=A_1 A_2 \dots A_m$ contains $C(X)$ and single qubit gates. Then for single-qubit gates we replace $A_i$ by $C(A_i)$, which is constructed following the description in N&C. And the $C(X)$ are replaced by Toffoli gates (which might also be decomposed). Right? Mar 26 '18 at 22:12
• @M.Stern well almost. If $U$ contains a $C(X)$ (which more precisely would be a $C(X)_{ij}$, acting between $i$-th and $j$-th qubit, with $i, j>1$), then the equivalent gate in $C(U)$ is already a Toffoli gate, with first and $i$-th qubits as control and $j$-th qubit as target. You can therefore go and replace the Toffolis using the known decompositions
– glS
Mar 26 '18 at 23:50

Although this might not answer your question completely, I think it might provide some direction of thinking. Here are two important facts:

• Any unitary $2^{n}\times 2^{n}$ matrix $M$, can be realized on a quantum computer with $n$-quantum bits by a finite sequence of controlled-not and single qubit gates1.

• Suppose $U$ is a unitary $2\times 2$ matrix satisfying $\text{tr } U \neq 0$, $\text{tr} (UX) \neq 0$, and $\text{det } U \neq 1$. Then six elementary gates are necessary and sufficient to implement a controlled $U$-gate2.

It should be possible to extend the second case to the general $n\times n$ case, given the first point, although I haven't found any paper which does that explicitly.