Alternatives to the swap test using a smaller number of qubits

Question

I need to compute the inner product between two generic quantum states. To this aim, one can use the swap test as explained in https://en.wikipedia.org/wiki/Swap_test .

In my case, both of the quantum states are given by two parametrized quantum circuits having the same structure but different parameters.

Since the swap test requires a relatively big register (one ancilla, a set of $n$ qubits for $\psi$ and a set of $n$ qubits for $\phi$), I am wondering if there are alternatives where I do not have to double the quantum register size for $\psi$ and $\phi$, so use $n+1$ qubits instead of $2n+1$. Or, maybe there is no way to do that.

Answer 1

Let us say that you want to evaluate $|\langle\psi|\varphi\rangle|^2$. Let us denote $U$ and $V$ the unitaires associated to your parametrized quantum circuits, producing respectively $|\psi\rangle$ and $|\varphi\rangle$. Using these notations, you actually want to evaluate $|\langle0|U^\dagger V|0\rangle|^2$. Note that this is equal to the probability of getting $|0\rangle$ when measuring the $U^\dagger V|0\rangle$ state in the computational basis.

I'm assuming here that you are able to apply $U^\dagger$, which is true if you have access to $U$'s circuit description.

Thus, what you can do is:

1. Start from the $|0\rangle$ state.
2. Apply $V$.
3. Apply $U^\dagger$.
4. Measure in the computational basis.

You're interested in the frequency at which you've measured $|0\rangle$ state, which will provide a good estimate for the quantity you're interested in. You can furthermore upper-bound the absolute error of this estimation using a simple confidence interval.

I'm not sure this is the most efficient way to do it though, but it uses the least number of qubits you could ask for.

Answer 2

The SWAP test is great when you don't have access to $U$ or $V$ and have no other way to prepare $|\psi\rangle$ or $|\phi\rangle$, and uses $2n+1$ qubits - while Tristan's test uses only $n$ qubits, in addition to any ancillary registers used for $U$ or $V$. Tristan's approach assumes we can deduce $U^\dagger$ from $U$, which, as indicated, is almost as simple as running $U$ backwards.

Alternatively and morally equivalently you could use a Hadamard test on $n+1$ qubits (+other ancillary qubits) as follows:

1. Prepare a single-qubit control register as $|0\rangle$;
2. Apply a Hadamard gate to $|0\rangle$;
3. Apply $U$ to $|00\cdots 0\rangle$, controlled off of the control qubit being $|0\rangle$;
4. Apply $V$ to the same register, controlled off of the control qubit being $|1\rangle$;
5. Hadamard the control qubit;
6. Measure the control qubit in the computational basis.

This test runs both $U$ and $V$ forward, but uses at least one more qubit than Tristan's ($n+1$) while also having to assume that we have a way to control $U$ and $V$.

(We almost always assume that if we know how to execute $U$ then we can execute $U^\dagger$, and/or that we can execute controlled versions of $U$ - these assumptions may be risky though, especially the second one and especially on real-world devices!)