rbtheory.py

"""
RB-related functions of gates and models
"""
#***************************************************************************************************
# Copyright 2015, 2019 National Technology & Engineering Solutions of Sandia, LLC (NTESS).
# Under the terms of Contract DE-NA0003525 with NTESS, the U.S. Government retains certain rights
# in this software.
# Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except
# in compliance with the License.  You may obtain a copy of the License at
# http://www.apache.org/licenses/LICENSE-2.0 or in the LICENSE file in the root pyGSTi directory.
#***************************************************************************************************

from . import rbtools as _rbtls
from . import optools as _optls
from . import matrixtools as _mtls
from .. import objects as _objs
#from .. import construction as _cnst
#from .. import algorithms as _algs

import numpy as _np
import warnings as _warnings


def predicted_rb_number(model, target_model, weights=None, d=None, rtype='EI'):
    """
    Predicts the RB error rate from a model.

    Uses the "L-matrix" theory from Proctor et al Phys. Rev. Lett. 119, 130502
    (2017). Note that this gives the same predictions as the theory in Wallman
    Quantum 2, 47 (2018).

    This theory is valid for various types of RB, including standard
    Clifford RB -- i.e., it will accurately predict the per-Clifford
    error rate reported by standard Clifford RB. It is also valid for
    "direct RB" under broad circumstances.

    For this function to be valid the model should be trace preserving
    and completely positive in some representation, but the particular
    representation of the model used is irrelevant, as the predicted RB
    error rate is a gauge-invariant quantity. The function is likely reliable
    when complete positivity is slightly violated, although the theory on
    which it is based assumes complete positivity.

    Parameters
    ----------
    model : Model
        The model to calculate the RB number of. This model is the
        model randomly sampled over, so this is not necessarily the
        set of physical primitives. In Clifford RB this is a set of
        Clifford gates; in "direct RB" this normally would be the
        physical primitives.

    target_model : Model
        The target model, corresponding to `model`. This function is not invariant
        under swapping `model` and `target_model`: this Model must be the target model,
        and should consistent of perfect gates.

    weights : dict, optional
        If not None, a dictionary of floats, whereby the keys are the gates
        in `model` and the values are the unnormalized probabilities to apply
        each gate at each stage of the RB protocol. If not None, the values
        in weights must all be non-negative, and they must not all be zero.
        Because, when divided by their sum, they must be a valid probability
        distribution. If None, the weighting defaults to an equal weighting
        on all gates, as this is used in many RB protocols (e.g., Clifford RB).
        But, this weighting is flexible in the "direct RB" protocol.

    d : int, optional
        The Hilbert space dimension.  If None, then sqrt(model.dim) is used.

    rtype : str, optional
        The type of RB error rate, either "EI" or "AGI", corresponding to
        different dimension-dependent rescalings of the RB decay constant
        p obtained from fitting to Pm = A + Bp^m. "EI" corresponds to
        an RB error rate that is associated with entanglement infidelity, which
        is the probability of error for a gate with stochastic errors. This is
        the RB error rate defined in the "direct RB" protocol, and is given by:

            r =  (d^2 - 1)(1 - p)/d^2,

        The AGI-type r is given by

            r =  (d - 1)(1 - p)/d,

        which is the conventional r definition in Clifford RB. This r is
        associated with (gate-averaged) average gate infidelity.

    Returns
    -------
    r : float.
        The predicted RB number.
    """
    if d is None: d = int(round(_np.sqrt(model.dim)))
    p = predicted_rb_decay_parameter(model, target_model, weights=weights)
    r = _rbtls.p_to_r(p, d=d, rtype=rtype)
    return r


def predicted_rb_decay_parameter(model, target_model, weights=None):
    """
    Computes the second largest eigenvalue of the 'L matrix' (see the `L_matrix` function).

    For standard Clifford RB and direct RB, this corresponds to the RB decay
    parameter p in Pm = A + Bp^m for "reasonably low error" trace preserving and
    completely positive gates. See also the `predicted_rb_number` function.

    Parameters
    ----------
    model : Model
        The model to calculate the RB decay parameter of. This model is the
        model randomly sampled over, so this is not necessarily the
        set of physical primitives. In Clifford RB this is a set of
        Clifford gates; in "direct RB" this normally would be the
        physical primitives.

    target_model : Model
        The target model corresponding to model. This function is not invariant under
        swapping `model` and `target_model`: this Model must be the target model, and
        should consistent of perfect gates.

    weights : dict, optional
        If not None, a dictionary of floats, whereby the keys are the gates
        in `model` and the values are the unnormalized probabilities to apply
        each gate at each stage of the RB protocol. If not None, the values
        in weights must all be non-negative, and they must not all be zero.
        Because, when divided by their sum, they must be a valid probability
        distribution. If None, the weighting defaults to an equal weighting
        on all gates, as this is used in many RB protocols (e.g., Clifford RB).
        But, this weighting is flexible in the "direct RB" protocol.

    Returns
    -------
    p : float.
        The second largest eigenvalue of L. This is the RB decay parameter
        for various types of RB.
    """
    L = L_matrix(model, target_model, weights=weights)
    E = _np.absolute(_np.linalg.eigvals(L))
    E = _np.flipud(_np.sort(E))
    if abs(E[0] - 1) > 10**(-12):
        _warnings.warn("Output may be unreliable because the model is not approximately trace-preserving.")

    if E[1].imag > 10**(-10):
        _warnings.warn("Output may be unreliable because the RB decay constant has a significant imaginary component.")
    p = abs(E[1])
    return p


def rb_gauge(model, target_model, weights=None, mx_basis=None, eigenvector_weighting=1.0):
    """
    Computes the gauge transformation required so that the RB number matches the average model infidelity.

    This function computes the gauge transformation required so that, when the
    model is transformed via this gauge-transformation, the RB number -- as
    predicted by the function `predicted_rb_number` -- is the average model
    infidelity between the transformed `model` model and the target model
    `target_model`. This transformation is defined Proctor et al
    Phys. Rev. Lett. 119, 130502 (2017), and see also Wallman Quantum 2, 47
    (2018).

    Parameters
    ----------
    model : Model
        The RB model. This is not necessarily the set of physical primitives -- it
        is the model randomly sampled over in the RB protocol (e.g., the Cliffords).

    target_model : Model
        The target model corresponding to model. This function is not invariant under
        swapping `model` and `target_model`: this Model must be the target model, and
        should consistent of perfect gates.

    weights : dict, optional
        If not None, a dictionary of floats, whereby the keys are the gates
        in `model` and the values are the unnormalized probabilities to apply
        each gate at each stage of the RB protocol. If not None, the values
        in weights must all be non-negative, and they must not all be zero.
        Because, when divided by their sum, they must be a valid probability
        distribution. If None, the weighting defaults to an equal weighting
        on all gates, as this is used in many RB protocols (e.g., Clifford RB).
        But, this weighting is flexible in the "direct RB" protocol.

    mx_basis : {"std","gm","pp"}, optional
        The basis of the models. If None, the basis is obtained from the model.

    eigenvector_weighting : float, optional
        Must be non-zero. A weighting on the eigenvector with eigenvalue that
        is the RB decay parameter, in the sum of this eigenvector and the
        eigenvector with eigenvalue of 1 that defines the returned matrix `l_operator`.
        The value of this factor does not change whether this `l_operator` transforms into
        a gauge in which r = AGsI, but it may impact on other properties of the
        gates in that gauge. It is irrelevant if the gates are unital.

    Returns
    -------
    l_operator : array
        The matrix defining the gauge-transformation.
    """
    gam, vecs = _np.linalg.eig(L_matrix(model, target_model, weights=weights))
    absgam = abs(gam)
    index_max = _np.argmax(absgam)
    gam_max = gam[index_max]

    if abs(gam_max - 1) > 10**(-12):
        _warnings.warn("Output may be unreliable because the model is not approximately trace-preserving.")

    absgam[index_max] = 0.0
    index_2ndmax = _np.argmax(absgam)
    decay_constant = gam[index_2ndmax]
    if decay_constant.imag > 10**(-12):
        _warnings.warn("Output may be unreliable because the RB decay constant has a significant imaginary component.")

    vec_l_operator = vecs[:, index_max] + eigenvector_weighting * vecs[:, index_2ndmax]

    if mx_basis is None:
        mx_basis = model.basis.name
    assert(mx_basis == 'pp' or mx_basis == 'gm' or mx_basis == 'std'), "mx_basis must be 'gm', 'pp' or 'std'."

    if mx_basis in ('pp', 'gm'):
        assert(_np.amax(vec_l_operator.imag) < 10**(-15)), "If 'gm' or 'pp' basis, RB gauge matrix should be real."
        vec_l_operator = vec_l_operator.real

    vec_l_operator[abs(vec_l_operator) < 10**(-15)] = 0.
    l_operator = _mtls.unvec(vec_l_operator)

    return l_operator


def transform_to_rb_gauge(model, target_model, weights=None, mx_basis=None, eigenvector_weighting=1.0):
    """
    Transforms a Model into the "RB gauge" (see the `RB_gauge` function).

    This notion was introduced in Proctor et al Phys. Rev. Lett. 119, 130502
    (2017). This gauge is a function of both the model and its target. These may
    be input in any gauge, for the purposes of obtaining "r = average model
    infidelity" between the output :class:`Model` and `target_model`.

    Parameters
    ----------
    model : Model
        The RB model. This is not necessarily the set of physical primitives -- it
        is the model randomly sampled over in the RB protocol (e.g., the Cliffords).

    target_model : Model
        The target model corresponding to model. This function is not invariant under
        swapping `model` and `target_model`: this Model must be the target model, and
        should consistent of perfect gates.

    weights : dict, optional
        If not None, a dictionary of floats, whereby the keys are the gates
        in `model` and the values are the unnormalized probabilities to apply
        each gate at each stage of the RB protocol. If not None, the values
        in weights must all be non-negative, and they must not all be zero.
        Because, when divided by their sum, they must be a valid probability
        distribution. If None, the weighting defaults to an equal weighting
        on all gates, as this is used in many RB protocols (e.g., Clifford RB).
        But, this weighting is flexible in the "direct RB" protocol.

    mx_basis : {"std","gm","pp"}, optional
        The basis of the models. If None, the basis is obtained from the model.

    eigenvector_weighting : float, optional
        Must be non-zero. A weighting on the eigenvector with eigenvalue that
        is the RB decay parameter, in the sum of this eigenvector and the
        eigenvector with eigenvalue of 1 that defines the returned matrix `l_operator`.
        The value of this factor does not change whether this `l_operator` transforms into
        a gauge in which r = AGsI, but it may impact on other properties of the
        gates in that gauge. It is irrelevant if the gates are unital.

    Returns
    -------
    model_in_RB_gauge : Model
        The model `model` transformed into the "RB gauge".
    """
    l = rb_gauge(model, target_model, weights=weights, mx_basis=mx_basis,
                 eigenvector_weighting=eigenvector_weighting)
    model_in_RB_gauge = model.copy()
    S = _objs.FullGaugeGroupElement(_np.linalg.inv(l))
    model_in_RB_gauge.transform_inplace(S)
    return model_in_RB_gauge


def L_matrix(model, target_model, weights=None):  # noqa N802
    """
    Constructs a generalization of the 'L-matrix' linear operator on superoperators.

    From Proctor et al Phys. Rev. Lett. 119, 130502 (2017), the 'L-matrix' is
    represented as a matrix via the "stack" operation. This eigenvalues of this
    matrix describe the decay constant (or constants) in an RB decay curve for
    an RB protocol whereby random elements of the provided model are sampled
    according to the `weights` probability distribution over the model. So, this
    facilitates predictions of Clifford RB and direct RB decay curves.

    Parameters
    ----------
    model : Model
        The RB model. This is not necessarily the set of physical primitives -- it
        is the model randomly sampled over in the RB protocol (e.g., the Cliffords).

    target_model : Model
        The target model corresponding to model. This function is not invariant under
        swapping `model` and `target_model`: this Model must be the target model, and
        should consistent of perfect gates.

    weights : dict, optional
        If not None, a dictionary of floats, whereby the keys are the gates
        in `model` and the values are the unnormalized probabilities to apply
        each gate at each stage of the RB protocol. If not None, the values
        in weights must all be non-negative, and they must not all be zero.
        Because, when divided by their sum, they must be a valid probability
        distribution. If None, the weighting defaults to an equal weighting
        on all gates, as this is used in many RB protocols (e.g., Clifford RB).
        But, this weighting is flexible in the "direct RB" protocol.

    Returns
    -------
    L : float
        A weighted version of the L operator from Proctor et al Phys. Rev. Lett.
        119, 130502 (2017), represented as a matrix using the 'stacking' convention.
    """
    if weights is None:
        weights = {}
        for key in list(target_model.operations.keys()):
            weights[key] = 1.

    normalizer = _np.sum(_np.array([weights[key] for key in list(target_model.operations.keys())]))
    L_matrix = (1 / normalizer) * _np.sum(
        weights[key] * _np.kron(
            model.operations[key].to_dense().T, _np.linalg.inv(target_model.operations[key].to_dense())
        ) for key in target_model.operations.keys())

    return L_matrix


def R_matrix_predicted_rb_decay_parameter(model, group, group_to_model=None, weights=None):  # noqa N802
    """
    Returns the second largest eigenvalue of a generalization of the 'R-matrix' [see the `R_matrix` function].

    Introduced in Proctor et al Phys. Rev. Lett. 119, 130502 (2017).  This
    number is a prediction of the RB decay parameter for trace-preserving gates
    and a variety of forms of RB, including Clifford and direct RB. This
    function creates a matrix which scales super-exponentially in the number of
    qubits.

    Parameters
    ----------
    model : Model
        The model to predict the RB decay paramter for. If `group_to_model` is
        None, the labels of the gates in `model` should be the  same as the labels of the
        group elements in `group`. For Clifford RB this would be the clifford model,
        for direct RB it would be the primitive gates.

    group : MatrixGroup
        The group that the `model` model contains gates from (`model` does not
        need to be the full group, and could be a subset of `group`). For
        Clifford RB and direct RB, this would be the Clifford group.

    group_to_model : dict, optional
        If not None, a dictionary that maps labels of group elements to labels
        of `model`. If `model` and `group` elements have the same labels, this dictionary
        is not required. Otherwise it is necessary.

    weights : dict, optional
        If not None, a dictionary of floats, whereby the keys are the gates in `model`
        and the values are the unnormalized probabilities to apply each gate at
        each stage of the RB protocol. If not None, the values in weights must all
        be positive or zero, and they must not all be zero (because, when divided by
        their sum, they must be a valid probability distribution). If None, the
        weighting defaults to an equal weighting on all gates, as used in most RB
        protocols.

    Returns
    -------
    p : float
        The predicted RB decay parameter. Valid for standard Clifford RB or direct RB
        with trace-preserving gates, and in a range of other circumstances.
    """
    R = R_matrix(model, group, group_to_model=group_to_model, weights=weights)
    E = _np.absolute(_np.linalg.eigvals(R))
    E = _np.flipud(_np.sort(E))
    p = E[1]
    return p


def R_matrix(model, group, group_to_model=None, weights=None):  # noqa N802
    """
    Constructs a generalization of the 'R-matrix' of Proctor et al Phys. Rev. Lett. 119, 130502 (2017).

    This matrix described the exact behaviour of the average success
    probablities of RB sequences. This matrix is super-exponentially large in
    the number of qubits, but can be constructed for 1-qubit models.

    Parameters
    ----------
    model : Model
        The noisy model (e.g., the Cliffords) to calculate the R matrix of.
        The correpsonding `target` model (not required in this function)
        must be equal to or a subset of (a faithful rep of) the group `group`.
        If `group_to_model `is None, the labels of the gates in model should be
        the same as the labels of the corresponding group elements in `group`.
        For Clifford RB `model` should be the clifford model; for direct RB
        this should be the native model.

    group : MatrixGroup
        The group that the `model` model contains gates from. For Clifford RB
        or direct RB, this would be the Clifford group.

    group_to_model : dict, optional
        If not None, a dictionary that maps labels of group elements to labels
        of model. This is required if the labels of the gates in `model` are different
        from the labels of the corresponding group elements in `group`.

    weights : dict, optional
        If not None, a dictionary of floats, whereby the keys are the gates in model
        and the values are the unnormalized probabilities to apply each gate at
        for each layer of the RB protocol. If None, the weighting defaults to an
        equal weighting on all gates, as used in most RB protocols (e.g., Clifford
        RB).

    Returns
    -------
    R : float
        A weighted, a subset-sampling generalization of the 'R-matrix' from Proctor
        et al Phys. Rev. Lett. 119, 130502 (2017).
    """
    if group_to_model is None:
        for key in list(model.operations.keys()):
            assert(key in group.labels), "Gates labels are not in `group`!"
    else:
        for key in list(model.operations.keys()):
            assert(key in group_to_model.values()), "Gates labels are not in `group_to_model`!"

    d = int(round(_np.sqrt(model.dim)))
    group_dim = len(group)
    R_dim = group_dim * d**2
    R = _np.zeros([R_dim, R_dim], float)

    if weights is None:
        weights = {}
        for key in list(model.operations.keys()):
            weights[key] = 1.

    normalizer = _np.sum(_np.array([weights[key] for key in list(model.operations.keys())]))

    for i in range(0, group_dim):
        for j in range(0, group_dim):
            label_itoj = group.labels[group.product([group.inverse_index(i), j])]
            if group_to_model is not None:
                if label_itoj in group_to_model:
                    gslabel = group_to_model[label_itoj]
                    R[j * d**2:(j + 1) * d**2, i * d**2:(i + 1) * d**2] = weights[gslabel] * model.operations[gslabel]
            else:
                if label_itoj in list(model.operations.keys()):
                    gslabel = label_itoj
                    R[j * d**2:(j + 1) * d**2, i * d**2:(i + 1) * d**2] = weights[gslabel] * model.operations[gslabel]

    R = R / normalizer

    return R

### COMMENTED OUT SO THAT THIS FILE DOESN'T NEED "from .. import construction as _cnst".
### THIS SHOULD BE ADDED BACK IN AT SOME POINT.
# def exact_rb_asps(model, group, m_max, m_min=0, m_step=1, success_outcomelabel=('0',),
#                   group_to_model=None, weights=None, compilation=None, group_twirled=False):
#     """
#     Calculates the exact RB average success probablilites (ASP).

#     Uses some generalizations of the formula given Proctor et al
#     Phys. Rev. Lett. 119, 130502 (2017). This formula does not scale well with
#     group size and qubit number, and for the Clifford group it is likely only
#     practical for a single qubit.

#     Parameters
#     ----------
#     model : Model
#         The noisy model (e.g., the Cliffords) to calculate the R matrix of.
#         The correpsonding `target` model (not required in this function)
#         must be equal to or a subset of (a faithful rep of) the group `group`.
#         If group_to_model is None, the labels of the gates in model should be
#         the same as the labels of the corresponding group elements in `group`.
#         For Clifford RB `model` should be the clifford model; for direct RB
#         this should be the native model.

#     group : MatrixGroup
#         The group that the `model` model contains gates from. For Clifford RB
#         or direct RB, this would be the Clifford group.

#     m_max : int
#         The maximal sequence length of the random gates, not including the
#         inversion gate.

#     m_min : int, optional
#         The minimal sequence length. Defaults to the smallest valid value of 0.

#     m_step : int, optional
#         The step size between sequence lengths. Defaults to the smallest valid
#         value of 1.

#     success_outcomelabel : str or tuple, optional
#         The outcome label associated with success.

#     group_to_model : dict, optional
#         If not None, a dictionary that maps labels of group elements to labels
#         of model. This is required if the labels of the gates in `model` are different
#         from the labels of the corresponding group elements in `group`.

#     weights : dict, optional
#         If not None, a dictionary of floats, whereby the keys are the gates in model
#         and the values are the unnormalized probabilities to apply each gate at
#         for each layer of the RB protocol. If None, the weighting defaults to an
#         equal weighting on all gates, as used in most RB protocols (e.g., Clifford
#         RB).

#     compilation : dict, optional
#         If `model` is not the full group `group` (with the same labels), then a
#         compilation for the group elements, used to implement the inversion gate
#         (and the initial randomgroup element, if `group_twirled` is True). This
#         is a dictionary with the group labels as keys and a gate sequence of the
#         elements of `model` as values.

#     group_twirled : bool, optional
#         If True, the random sequence starts with a single uniformly random group
#         element before the m random elements of `model`.

#     Returns
#     -------
#     m : float
#         Array of sequence length values that the ASPs have been calculated for.

#     P_m : float
#         Array containing ASP values for the specified sequence length values.
#     """
#     if compilation is None:
#         for key in list(model.operations.keys()):
#             assert(key in group.labels), "Gates labels are not in `group`, so `compilation must be specified."
#         for label in group.labels:
#             assert(label in list(model.operations.keys())
#                    ), "Some group elements not in `model`, so `compilation must be specified."

#     i_max = _np.floor((m_max - m_min) / m_step).astype('int')
#     m = _np.zeros(1 + i_max, int)
#     P_m = _np.zeros(1 + i_max, float)
#     group_dim = len(group)
#     R = R_matrix(model, group, group_to_model=group_to_model, weights=weights)
#     success_prepLabel = list(model.preps.keys())[0]  # just take first prep
#     success_effectLabel = success_outcomelabel[-1] if isinstance(success_outcomelabel, tuple) \
#         else success_outcomelabel
#     extended_E = _np.kron(_mtls.column_basis_vector(0, group_dim).T, model.povms['Mdefault'][success_effectLabel].T)
#     extended_rho = _np.kron(_mtls.column_basis_vector(0, group_dim), model.preps[success_prepLabel])

#     if compilation is None:
#         extended_E = group_dim * _np.dot(extended_E, R)
#         if group_twirled is True:
#             extended_rho = _np.dot(R, extended_rho)
#     else:
#         full_model = _cnst.create_explicit_alias_model(model, compilation)
#         R_fullgroup = R_matrix(full_model, group)
#         extended_E = group_dim * _np.dot(extended_E, R_fullgroup)
#         if group_twirled is True:
#             extended_rho = _np.dot(R_fullgroup, extended_rho)

#     Rstep = _np.linalg.matrix_power(R, m_step)
#     Riterate = _np.linalg.matrix_power(R, m_min)
#     for i in range(0, 1 + i_max):
#         m[i] = m_min + i * m_step
#         P_m[i] = _np.dot(extended_E, _np.dot(Riterate, extended_rho))
#         Riterate = _np.dot(Rstep, Riterate)

#     return m, P_m

### COMMENTED OUT SO THAT THIS FILE DOESN'T NEED "from .. import construction as _cnst"
### THIS SHOULD BE ADDED BACK IN AT SOME POINT.
# def L_matrix_asps(model, target_model, m_max, m_min=0, m_step=1, success_outcomelabel=('0',),  # noqa N802
#                   compilation=None, group_twirled=False, weights=None, gauge_optimize=True,
#                   return_error_bounds=False, norm='diamond'):
#     """
#     Computes RB average survival probablities, as predicted by the 'L-matrix' theory.

#     This theory was introduced in Proctor et al Phys. Rev. Lett. 119, 130502
#     (2017). Within the function, the model is gauge-optimized to target_model. This is
#     *not* optimized to the gauge specified by Proctor et al, but instead performs the
#     standard pyGSTi gauge-optimization (using the frobenius distance). In most cases,
#     this is likely to be a reasonable proxy for the gauge optimization perscribed by
#     Proctor et al.

#     Parameters
#     ----------
#     model : Model
#         The noisy model.

#     target_model : Model
#         The target model.

#     m_max : int
#         The maximal sequence length of the random gates, not including the inversion gate.

#     m_min : int, optional
#         The minimal sequence length. Defaults to the smallest valid value of 0.

#     m_step : int, optional
#         The step size between sequence lengths.

#     success_outcomelabel : str or tuple, optional
#         The outcome label associated with success.

#     compilation : dict, optional
#         If `model` is not the full group, then a compilation for the group elements,
#         used to implement the inversion gate (and the initial random group element,
#         if `group_twirled` is True). This is a dictionary with the group labels as
#         keys and a gate sequence of the elements of `model` as values.

#     group_twirled : bool, optional
#         If True, the random sequence starts with a single uniformly random group
#         element before the m random elements of `model`.

#     weights : dict, optional
#         If not None, a dictionary of floats, whereby the keys are the gates in model
#         and the values are the unnormalized probabilities to apply each gate at
#         for each layer of the RB protocol. If None, the weighting defaults to an
#         equal weighting on all gates, as used in most RB protocols (e.g., Clifford
#         RB).

#     gauge_optimize : bool, optional
#         If True a gauge-optimization to the target model is implemented before
#         calculating all quantities. If False, no gauge optimization is performed.
#         Whether or not a gauge optimization is performed does not affect the rate of
#         decay but it will generally affect the exact form of the decay. E.g., if a
#         perfect model is given to the function -- but in the "wrong" gauge -- no
#         decay will be observed in the output P_m, but the P_m can be far from 1 (even
#         for perfect SPAM) for all m. The gauge optimization is optional, as it is
#         not guaranteed to always improve the accuracy of the reported P_m, although when
#         gauge optimization is performed this limits the possible deviations of the
#         reported P_m from the true P_m.

#     return_error_bounds : bool, optional
#         Sets whether or not to return error bounds for how far the true ASPs can deviate
#         from the values returned by this function.

#     norm : str, optional
#         The norm used in the error bound calculation. Either 'diamond' for the diamond
#         norm (the default) or '1to1' for the Hermitian 1 to 1 norm.

#     Returns
#     -------
#     m : float
#         Array of sequence length values that the ASPs have been calculated for.
#     P_m : float
#         Array containing predicted ASP values for the specified sequence length values.
#     if error_bounds is True :
#         lower_bound: float
#             Array containing lower bounds on the possible ASP values

#         upper_bound: float
#             Array containing upper bounds on the possible ASP values
#     """
#     d = int(round(_np.sqrt(model.dim)))

#     if gauge_optimize:
#         model_go = _algs.gaugeopt_to_target(model, target_model)
#     else:
#         model_go = model.copy()
#     L = L_matrix(model_go, target_model, weights=weights)
#     success_prepLabel = list(model.preps.keys())[0]  # just take first prep
#     success_effectLabel = success_outcomelabel[-1] if isinstance(success_outcomelabel, tuple) \
#         else success_outcomelabel
#     identity_vec = _mtls.vec(_np.identity(d**2, float))

#     if compilation is not None:
#         model_group = _cnst.create_explicit_alias_model(model_go, compilation)
#         model_target_group = _cnst.create_explicit_alias_model(target_model, compilation)
#         delta = gate_dependence_of_errormaps(model_group, model_target_group, norm=norm)
#         emaps = errormaps(model_group, model_target_group)
#         E_eff = _np.dot(model_go.povms['Mdefault'][success_effectLabel].T, emaps.operations['Gavg'])

#         if group_twirled is True:
#             L_group = L_matrix(model_group, model_target_group)

#     if compilation is None:
#         delta = gate_dependence_of_errormaps(model_go, target_model, norm=norm)
#         emaps = errormaps(model_go, target_model)
#         E_eff = _np.dot(model_go.povms['Mdefault'][success_effectLabel].T, emaps.operations['Gavg'])

#     i_max = _np.floor((m_max - m_min) / m_step).astype('int')
#     m = _np.zeros(1 + i_max, int)
#     P_m = _np.zeros(1 + i_max, float)
#     upper_bound = _np.zeros(1 + i_max, float)
#     lower_bound = _np.zeros(1 + i_max, float)

#     Lstep = _np.linalg.matrix_power(L, m_step)
#     Literate = _np.linalg.matrix_power(L, m_min)
#     for i in range(0, 1 + i_max):
#         m[i] = m_min + i * m_step
#         if group_twirled:
#             L_m_rdd = _mtls.unvec(_np.dot(L_group, _np.dot(Literate, identity_vec)))
#         else:
#             L_m_rdd = _mtls.unvec(_np.dot(Literate, identity_vec))
#         P_m[i] = _np.dot(E_eff, _np.dot(L_m_rdd, model_go.preps[success_prepLabel]))
#         Literate = _np.dot(Lstep, Literate)
#         upper_bound[i] = P_m[i] + delta / 2
#         lower_bound[i] = P_m[i] - delta / 2
#         if upper_bound[i] > 1:
#             upper_bound[i] = 1.
#         if lower_bound[i] < 0:
#             lower_bound[i] = 0.
#     if return_error_bounds:
#         return m, P_m, lower_bound, upper_bound
#     else:
#         return m, P_m


def errormaps(model, target_model):
    """
    Computes the 'left-multiplied' error maps associated with a noisy gate set, along with the average error map.

    This is the model [E_1,...] such that

        `G_i = E_iT_i`,

    where `T_i` is the gate which `G_i` is a noisy
    implementation of. There is an additional gate in the set, that has
    the key 'Gavg'. This is the average of the error maps.

    Parameters
    ----------
    model : Model
        The imperfect model.

    target_model : Model
        The target model.

    Returns
    -------
    errormaps : Model
        The left multplied error gates, along with the average error map,
        with the key 'Gavg'.
    """
    errormaps_gate_list = []
    errormaps = model.copy()
    for gate in list(target_model.operations.keys()):
        errormaps.operations[gate] = _np.dot(model.operations[gate],
                                             _np.transpose(target_model.operations[gate]))
        errormaps_gate_list.append(errormaps.operations[gate])

    errormaps.operations['Gavg'] = _np.mean(_np.array([i for i in errormaps_gate_list]),
                                            axis=0, dtype=_np.float64)
    return errormaps


def gate_dependence_of_errormaps(model, target_model, norm='diamond', mx_basis=None):
    """
    Computes the "gate-dependence of errors maps" parameter defined by

    delta_avg = avg_i|| E_i - avg_i(E_i) ||,

    where E_i are the error maps, and the norm is either the diamond norm
    or the 1-to-1 norm. This quantity is defined in Magesan et al PRA 85
    042311 2012.

    Parameters
    ----------
    model : Model
        The actual model

    target_model : Model
        The target model.

    norm : str, optional
        The norm used in the calculation. Can be either 'diamond' for
        the diamond norm, or '1to1' for the Hermitian 1 to 1 norm.

    mx_basis : {"std","gm","pp"}, optional
        The basis of the models. If None, the basis is obtained from
        the model.

    Returns
    -------
    delta_avg : float
        The value of the parameter defined above.
    """
    error_gs = errormaps(model, target_model)
    delta = []

    if mx_basis is None:
        mx_basis = model.basis.name
    assert(mx_basis == 'pp' or mx_basis == 'gm' or mx_basis == 'std'), "mx_basis must be 'gm', 'pp' or 'std'."

    for gate in list(target_model.operations.keys()):
        if norm == 'diamond':
            print(error_gs.operations[gate])
            print(error_gs.operations['Gavg'])
            delta.append(_optls.diamonddist(error_gs.operations[gate], error_gs.operations['Gavg'],
                                            mx_basis=mx_basis))
        elif norm == '1to1':
            gate_dif = error_gs.operations[gate] - error_gs.operations['Gavg']
            delta.append(_optls.norm1to1(gate_dif, num_samples=1000, mx_basis=mx_basis, return_list=False))
        else:
            raise ValueError("Only diamond or 1to1 norm available.")

    delta_avg = _np.mean(delta)
    return delta_avg

# Future : perhaps put these back in.
#def Magesan_theory_predicted_decay(model, target_model, mlist, success_outcomelabel=('0',),
#                                   norm='1to1', order='zeroth', return_all = False):
#
#    assert(order == 'zeroth' or order == 'first')
#
#    d = int(round(_np.sqrt(model.dim)))
#    MTPs = {}
#    MTPs['r'] = gateset_infidelity(model,target_model,itype='AGI')
#    MTPs['p'] = _analysis.r_to_p(MTPs['r'],d,rtype='AGI')
#    MTPs['delta'] = gate_dependence_of_errormaps(model, target_model, norm)
#    error_gs = errormaps(model, target_model)
#
#    R_list = []
#    Q_list = []
#    for gate in list(target_model.operations.keys()):
#        R_list.append(_np.dot(_np.dot(error_gs.operations[gate],target_model.operations[gate]),
#              _np.dot(error_gs.operations['Gavg'],_np.transpose(target_model.operations[gate]))))
#        Q_list.append(_np.dot(target_model.operations[gate],
#              _np.dot(error_gs.operations[gate],_np.transpose(target_model.operations[gate]))))
#
#    error_gs.operations['GR'] = _np.mean(_np.array([ i for i in R_list]),axis=0)
#    error_gs.operations['GQ'] = _np.mean(_np.array([ i for i in Q_list]),axis=0)
#    error_gs.operations['GQ2'] = _np.dot(error_gs.operations['GQ'],error_gs.operations['Gavg'])
#    error_gs.preps['rhoc_mixed'] = 1./d*_cnst._basis_create_identity_vec(error_gs.basis)#
#
#    #Assumes standard POVM labels
#    povm = _objs.UnconstrainedPOVM( [('0_cm', target_model.povms['Mdefault']['0']),
#                                     ('1_cm', target_model.povms['Mdefault']['1'])] )
#    ave_error_gsl = _cnst.to_circuits([('rho0','Gavg'),('rho0','GR'),('rho0','Gavg','GQ')])
#    data = _cnst.simulate_data(error_gs, ave_error_gsl, num_samples=1, sample_error="none")#

#    pr_L_p = data[('rho0','Gavg')][success_outcomelabel]
#    pr_L_I = data[('rho0','Gavg')][success_outcomelabel_cm]
#    pr_R_p = data[('rho0','GR')][success_outcomelabel]
#    pr_R_I = data[('rho0','GR')][success_outcomelabel_cm]
#    pr_Q_p = data[('rho0','Gavg','GQ')][success_outcomelabel]
#    p = MTPs['p']
#    B_1 = pr_R_I
#    A_1 = (pr_Q_p/p) - pr_L_p + ((p -1)*pr_L_I/p) + ((pr_R_p - pr_R_I)/p)
#    C_1 = pr_L_p - pr_L_I
#    q = _tls.average_gate_infidelity(error_gs.operations['GQ2'],_np.identity(d**2,float))
#    q = _analysis.r_to_p(q,d,rtype='AGI')
#
#    if order  == 'zeroth':
#        MTPs['A'] = pr_L_I
#        MTPs['B'] = pr_L_p - pr_L_I
#    if order == 'first':
#        MTPs['A'] = B_1
#        MTPs['B'] = A_1 - C_1*(q - 1)/p**2
#        MTPs['C'] = C_1*(q- p**2)/p**2
#
#    if order == 'zeroth':
#        Pm = MTPs['A'] +  MTPs['B']*MTPs['p']**_np.array(mlist)
#    if order == 'first':
#        Pm = MTPs['A'] +  (MTPs['B'] + _np.array(mlist)*MTPs['C'])*MTPs['p']**_np.array(mlist)
#
#   sys_eb = (MTPs['delta'] + 1)**(_np.array(mlist)+1) - 1
#    if order == 'first':
#        sys_eb = sys_eb - (_np.array(mlist)+1)*MTPs['delta']
#
#    upper = Pm + sys_eb
#    upper[upper > 1]=1.
#
#    lower = Pm - sys_eb
#    lower[lower < 0]=0.
#
#   return mlist, Pm, upper, lower, MTPs