Dynamic State Preparation #306
Conversation
| # n is static (known at compile time), so we can use normal numpy here | ||
| n = int(np.log2(target_array.shape[0])) | ||
|
|
||
| thetas, u_params, phases = _preprocess(target_array) |
There was a problem hiding this comment.
This classical preprocessing function for n=20 takes 0.3 seconds on my laptop (each layer is vectorized with vmap). Therefore, most of the time is spent in the quantum part, which basically just calls qswitch (with a bit reverse operation, as we cannot slice a DynamicQubitArray backward)
|
|
||
| int_encoder(self, self.encoder(value)) | ||
|
|
||
| def init_state_qswitch(self, state_array, method="auto"): |
There was a problem hiding this comment.
@positr0nium Instead of having a separate init_state_qswitch method, what about modifying the existing init_state in such a way that the workflow is dispatched with different implementations depending on the type of the provided argument?
That is, if a dict is provided, the qiskit implementation is used, which is currently based on a different (and currently more efficient) algorithm, while if an array is provided we dispatch to the implementation introduced in this PR. Please let me know what you think : )
1 participant
Context: This PR implements the state preparation algorithm based on the technique using multiplexers. However, instead of the latter, we use the
qswitchfunction.Description of the Change: As above. For a description of the algorithm, we refer to the documentation.
Benefits: Qrisp can now initialize a quantum circuit without resorting to
qiskit.Possible Drawbacks: None that I can think of, although we can probably implement a more efficient algorithm in the future. The reason is that the
qswitchfunction introduces several auxiliary qubits, which can easily cause overhead.Related GitHub Issues: None