pdft.coherence.certify_flat_modulus

pdft.coherence.certify_flat_modulus(basis, frozen_indices=None, *, atol=1e-08)

Certify that training cannot raise mu above 1.

Two conditions, both necessary. The basis must be flat-modulus now, and every tensor left trainable must be diagonal — then by the proposition in the module docstring, mu == 1 for every value those parameters can take.

frozen_indices is the same argument train_basis_batched accepts, so:

cert = certify_flat_modulus(basis, frozen_indices=frozen)
if not cert:
    raise ValueError(cert.reason)
result = train_basis_batched(basis, frozen_indices=frozen, ...)

Frozen non-diagonal gates are fine: a fixed Hadamard is what the proposition assumes. It is training them that voids it.

Parameters:

• frozen_indices (list[int] | None)

• atol (float)

Return type:

FlatModulusCertificate