BPResult

class pyqit.utils.diagnostic.BPResult(n_qubits: int, n_samples: int, layer_variances: dict[str, float], layer_ratios: dict[str, float], overall_variance: float, expected_variance: float, is_barren: bool, quantum_variance: float, classical_variance: float | None = None, n_executions: int | None = None, n_parameters: int | None = None, n_dead_parameters: int | None = None)

Bases: object

Result of check_barren_plateau.

repr(result) or print(result) renders a table (rich if installed, ASCII otherwise).

n_qubits

Width of the circuit that sets the baseline. For a pipeline, that is its one trainable quantum stage.

Type:

int

n_samples

Random weight draws the gradients were sampled at.

Type:

int

layer_variances

Gradient variance per weight tensor, keyed by its name in model.weights, such as “main_circuit.weights”. Circuit tensors count live parameters only. Classical tensors are included.

Type:

dict of {str: float}

layer_ratios

Each entry of layer_variances divided by expected_variance. The table marks a ratio below 1 as a plateau.

Type:

dict of {str: float}

overall_variance

The same value as quantum_variance.

Type:

float

expected_variance

Theoretical floor from McClean et al., scaled by bp_scale_factor. For a local cost it is about twice the gradient variance of a random circuit, so a flagged model is within a factor of two of random.

Type:

float

is_barren

True when quantum_variance is below expected_variance.

Type:

bool

quantum_variance

Mean of the circuit tensors’ variances. This is the number the verdict is made on. NaN when the model has no circuit weights.

Type:

float

classical_variance

Mean of the classical tensors’ variances. None when the model has no classical weights.

Type:

float, optional

n_parameters

Circuit parameters sampled, dead ones included.

Type:

int, optional

n_dead_parameters

Circuit parameters whose gradient was zero in every sample because they cannot reach the measured wires. They are left out of the variance.

Type:

int, optional

n_executions

Circuit executions the sampling cost, as counted by the device: one per sample under backprop, one plus two per parameter under parameter-shift. None when the model runs no QNode.

Type:

int, optional