Single-qubit and Controlled Z rotations

[mpharrigan]

Update single-qubit Z rotations ZPowGate and Rz. These are atomic operations.

Introduce actual constructions for CZPowGate and CRz for better resource counting. The previous hack would just count them as "one rotation". Now the former is a rotation and two Toffolis (TODO: can these be And?); the latter is two rotations and two cliffords.

Connect the single-qubit gates to their controlled versions via the controlled protocol.

There's still hacks for Controlled-S and Controlled-T, which I will port in a follow-up.

How should we handle X and Y rotations? One possibility is to make them compile to ZPow/Rz. Otherwise: duplicate the controlled construction for each flavor.

@tanujkhattar there are two tests failing that I'd like your help with

• the TextbookQFT uses cirq controlled_by which is causing breaking simulation tests in a weird way. This triggers BloqAsCirqGate.__pow__ probably broken #1446 but even after hack-fixing that it runs into a different issue. Switching to qualtran tensor contraction fixes the test but there may be outstanding issues
• I don't understand the flamegraph format. test_get_flame_graph_data_qft_textbook has different numbers now but I don't know how to verify that these are correct.
• fixes BloqAsCirqGate.__pow__ probably broken #1446

[anurudhp]

TODO: can these be And?

Yep you can use CZPow = And^dagger . ZPow . And

[mpharrigan]

I don't know why this is failing specifically for n=123 (and by a lot). There's (n-1) phase gradients, each of size i and we want to skip the first two (S, T). cc @tanujkhattar

[tanujkhattar]

The formula looks correct to me. Did you try debugging by checking which of the phase gradient bloqs results in lower rotation counts?

[mpharrigan]

I didn't look at the individual ones. Do you mind taking a look?

[anurudhp]

It fails from n=41 onwards. Any rotation with angle 0.5 / 2**b with b >= 40 is getting ignored as the _ANGLE_ATOL = 1e-12 which is roughly 2**40.

[mpharrigan]

oooh interesting

[mpharrigan]

@tanujkhattar do you have any thoughts on the cirq simulation failure

[anurudhp]

I think this can be replaced with the new get_ctrl_system_1bit_cv

[mpharrigan]

yep this is some vestigial code where I wanted to try my hand at factoring it out :) thanks for actually doing it

[anurudhp]

please override the ctrl system for this (and other controlled bloqs), and add a test to verify gate counts for CRz().controlled() (should be CRz() + And() + And.adjoint())

[mpharrigan]

why would I need to override the ctrl system? there isn't a native CCRz bloq

[anurudhp]

Because right now, CRz.controlled would return Controlled(CRz), i.e. does not merge the two controls using And.

The second case of get_ctrl_system_1bit_cv essentially handles this

[anurudhp]

(sync discussion) I did not notice the decomposition, in which case the override is not needed.

But I still think it's good to override to get a much more cost effective circuit. The current CRz(2*pi*t).controlled() gives

toffoli: 2, and_bloq: 2, clifford: 2, rotation: 2, measurement: 2

But with the override it becomes

and_bloq: 1, clifford: 3, rotation: 2, measurement: 1

---

the subbloq CRz(2*pi*theta) for reference:

full code

import attrs
import numpy as np
import sympy

from qualtran import Bloq, DidNotFlattenAnythingError
from qualtran.bloqs.basic_gates import CRz
from qualtran.bloqs.mcmt import And
from qualtran.drawing import show_bloq
from qualtran.resource_counting import get_cost_value, QECGatesCost

def display_bloq_and_cost(bloq, *, flat=True):
    def _pred(binst):
        return not binst.bloq_is(And)
        
    cbloq = bloq.as_composite_bloq()
    if flat: cbloq = cbloq.flatten(pred=_pred)

    show_bloq(cbloq, 'latex')
    cost = get_cost_value(bloq, QECGatesCost())
    print(cost)

theta = 2 * np.pi * sympy.Symbol("t")
crz = CRz(theta)
display_bloq_and_cost(crz.controlled())

@attrs.frozen
class CRzWithOverride(CRz):
    def get_ctrl_system(self, ctrl_spec):
        from qualtran.bloqs.mcmt.specialized_ctrl import get_ctrl_system_1bit_cv_from_bloqs

        return get_ctrl_system_1bit_cv_from_bloqs(
            self,
            ctrl_spec,
            current_ctrl_bit=1,
            bloq_with_ctrl=self,
            ctrl_reg_name='ctrl'
        )

display_bloq_and_cost(CRzWithOverride(theta).controlled())
print(cost.total_t_count())

[mpharrigan]

that can be done in a follow-up. I don't think it's worth blocking this PR any more than it already has been

[anurudhp]

should this be += 2? (as there's already an XGate() in L186)

[mpharrigan]

that's guarded by if control_val is None and this line test if control_val == 0