State preparation operations for common ansatz

[cgranade]

Is your feature request related to a problem? Please describe.
It would be helpful in targeting some variational applications to have state preparation operations for common variational ansatz such as RyRz.

Additional context
At the moment, these ansatz can be implemented by users using existing features (see example below), but it would be nice to have "pre-canned" versions with a common signature (e.g.: (Double[], Qubit[]) => Unit is Adj + Ctl) so that it's easier to test out and compare various ansatz.

Example using current features

function Pairs<'T>(elements : 'T[]) : ('T, 'T)[] {
    mutable allPairs = [];
    for idxFirst in 0..Length(elements) - 2 {
        for idxSecond in idxFirst + 1..Length(elements) - 1 {
            set allPairs += [(elements[idxFirst], elements[idxSecond])];
        }
    }
    return allPairs;
}

operation ApplyRotationLayer(parameters : Double[], targets : Qubit[]) : Unit is Adj + Ctl {
    let yParams = parameters[...Length(parameters) / 2 - 1];
    let zParams = parameters[Length(parameters) / 2...];
    ApplyToEachCA(Ry, Zipped(yParams, targets));
    ApplyToEachCA(Rz, Zipped(zParams, targets));
}

operation ApplyEntanglingLayer(targets : Qubit[]) : Unit is Adj + Ctl {
    ApplyToEachCA(CNOT, Pairs(targets));
}

operation ApplyRyRz(parameters : Double[], targets : Qubit[]) : Unit is Adj + Ctl {
    let layers = Chunks(2 * Length(targets), parameters);
    ApplyRotationLayer(layers[0], targets);
    for layer in layers {
        ApplyEntanglingLayer(targets);
        ApplyRotationLayer(layer, targets);
    }
}

[msoeken]

Thanks for the proposal. I have a few questions:

1. Should there be (2 + 1) * Length(targets) parameters? Currently, the first set of parameters layers[0] is passed to the first and second rotation layer.
2. Is it the only or a common choice for the entangling layer? I have also seen ansatz implementations that use a CNOT staircase as the entangling layer.
3. Is there a reference that explains these constructions to point to in the documentation?

[cgranade]

Thanks for taking a look, @msoeken!

I think you're right about the duplicated first layer, sorry, that's what I get for writing up a quick proof of concept.

For the choice of entangling layer and rotation gates, I agree, those are often replaced by cyclic, nearest-neighbor, or other kinds of patterns, or CNOT may be replaced by CZ in some cases. It may be good to see what we can consolidate with the QML library, QAOA ansatz, or other places where parameterized families of state preparations are used to identify what the right kinds of configuration options might be.