1

I am trying to calculate the distance between sets of vectors. To do this I can create a SwapTest circuit (described here on page 34) and encode a feature vector with a U3 gate, then apply H and CSWAP gates to calculate the overlap. When your feature vectors are length 3 this seems fairly easy, as this can be encoded entirely by one U3. Each U3 parameter will represent a feature. I've done this with vectors of length 2, and have used the overlap as a measure to decide which 2 vectors are "closest".

However, what if we add a 4th or 5th or 10th feature? I imagine we could then also use a second qubit, encode the 5 features over two U3 gates. However, I have trouble seeing how to construct the rest of the circuit.

If we take two states, A and B, each with 6 features we could split this over 2 U3 gates each. Do we perform the "normal" SwapTest on each half of each state. Ie, find the overlap of the upper half of A and the upper half of B, then do the same for the lower half? This seems reasonable, but how would I be able to combine this into 1 value I can use to compare to another set of vectors?

R.W
  • 2,468
  • 7
  • 25
Andrew
  • 333
  • 1
  • 6

1 Answers1

2

You can employ a method described in paper Transformation of quantum states using uniformly controlled rotations. This allows you to prepare arbitrary quantum state. If you have $x$ features, you can encode them into amplitudes of $\log_2 x$ qubits state because $n$ qubits states are described by vectors of length $2^n$.

Just note that in Qiskit the algorithm in the article above is implemented in function initialize, for example:

quantumState = [
    1 / math.sqrt(2) * complex(1, 0),
    1 / 2 * complex(1, 1),]

q = QuantumRegister(1, name = 'q')
c = ClassicalRegister(1, name = 'c')


circuit = QuantumCircuit(q,c)


circuit.initialize(quantumState, [q[0]])

If you have two such states (vector), you can use a swap test for multiqubit states are described by KAJ226 here.

Martin Vesely
  • 15,244
  • 4
  • 32
  • 75