tv_apply_transpose_from_sweep#

t3toolbox.backend.apply.tv_apply_transpose_from_sweep(c, ww, frame, frame_sweep, sum_over_probes=False)#

def tv_apply_transpose_from_sweep(
        c:          NDArray,                # residual, shape = W + K + C (K optional)
        ww:         typ.Sequence[NDArray],  # apply vectors (one-hot e_index for entries), len=d, elm_shape=W+(Ni,)
        frame:       typ.Tuple[
            typ.Sequence[NDArray],          # up_tucker_cores  U. len=d  (unused; uniform call signature)
            typ.Sequence[NDArray],          # down_tt_cores    O. len=d
            typ.Sequence[NDArray],          # left_tt_cores    P. len=d  (unused)
            typ.Sequence[NDArray],          # right_tt_cores   Q. len=d
        ],                                  # frame order = T3Frame.data = (up, down, left, right)
        frame_sweep: typ.Tuple[
            typ.Sequence[NDArray],          # xis
            typ.Sequence[NDArray],          # mus
        ],                                  # = tv_precompute_apply_frame_sweep(frame, ww)  (lean: no nu/eta)
        sum_over_probes: bool = False,
) -> typ.Tuple[typ.Sequence[NDArray], typ.Sequence[NDArray]]:  # (dU_tildes, dG_tildes) = T3Variations.data

Transpose of the all-modes apply reusing a precomputed frame sweep – the bare 𝒥ᵀ, by the adjoint-state method (the scalar residual c seeds one reverse sigma_hat sweep; see _apply_transpose_adjoint()). Takes the lean (xis, mus) sweep + the frame cores O, Q (it recomputes the right context rather than storing nu/eta – half the memory). Reuse hook for fitting.py (one frame sweep feeds the forward and this transpose). Full W + K + C (the residual c may carry the tangent stack K). No gauge projector Π.