iqm.benchmarks.entanglement.ghz_optimal.TreeSearchOptions

class iqm.benchmarks.entanglement.ghz_optimal.TreeSearchOptions(objective: str = 'estimated_fidelity', idle_penalty: float = 1.0, depth_penalty: float = 0.05, cz_duration: float = 4e-08)

Bases: object

How the tree search trades gate fidelity against circuit depth.

Attributes

cz_duration	Duration of one CZ layer in seconds, used by the idle term of "estimated_fidelity".
depth_penalty	Cost of one CZ layer for the "depth_penalty" objective, in units of -log fidelity.
idle_penalty	Weight of the idle-decoherence term of the "estimated_fidelity" objective.
objective	Which cost to minimize, one of TREE_OBJECTIVES.

Methods

__delattr__(name)	Implement delattr(self, name).
__eq__(other)	Return self==value.
__hash__()	Return hash(self).
__repr__()	Return repr(self).
__setattr__(name, value)	Implement setattr(self, name, value).

Parameters:

• objective (str)

• idle_penalty (float)

• depth_penalty (float)

• cz_duration (float)

objective: str = 'estimated_fidelity'

Which cost to minimize, one of TREE_OBJECTIVES.

"estimated_fidelity" scores a candidate by its estimated GHZ infidelity, gate_cost + idle_penalty * sum(idle time * (1/T1 + 1/T2)), where gate_cost is _gate_cost(): the -log fidelity of the tree’s CZs plus that of the single-qubit gates around them. "depth_penalty" uses the simpler gate_cost + depth_penalty * number of CZ layers, which needs no coherence times. "min_error" scores by the gate fidelity alone, which is what this routine did before the idle term existed: it selects the same edges as a minimum spanning tree, so it is the setting to use to reproduce older results.

idle_penalty: float = 1.0

Weight of the idle-decoherence term of the "estimated_fidelity" objective.

0.0 disables it (leaving the gate fidelity product), values above 1.0 buy depth more aggressively than the T1 and T2 estimate alone would justify.

depth_penalty: float = 0.05

Cost of one CZ layer for the "depth_penalty" objective, in units of -log fidelity.

The scale that matters is what a layer costs the whole register, roughly num_qubits * layer duration / T2; on a 45-qubit chip with 100 ns layers and T2 = 40 us that is about 0.05, which is the default. Values well below it never buy depth, since one layer of a 45-qubit tree is worth several CZ gates’ worth of error.

cz_duration: float = 4e-08

Duration of one CZ layer in seconds, used by the idle term of "estimated_fidelity".