pessimistic_pattern_matcher_b_3D.py

Go to the documentation of this file.

 
import numpy as np
from scipy.optimize import least_squares
from scipy.spatial import cKDTree
from scipy.stats import sigmaclip
 
import lsst.pipe.base as pipeBase
 
 
def _rotation_matrix_chi_sq(flattened_rot_matrix,
                            pattern_a,
                            pattern_b,
                            max_dist_rad):
    """Compute the squared differences for least squares fitting.
 
    Given a flattened rotation matrix, one N point pattern and another N point
    pattern to transform into to, compute the squared differences between the
    points in the two patterns after the rotation.
 
    Parameters
    ----------
    flattened_rot_matrix : `numpy.ndarray`, (9, )
        A flattened array representing a 3x3 rotation matrix. The array is
        flattened to comply with the API of scipy.optimize.least_squares.
        Flattened elements are [[0, 0], [0, 1], [0, 2], [1, 0]...]
    pattern_a : `numpy.ndarray`, (N, 3)
        A array containing N, 3 vectors representing the objects we would like
        to transform into the frame of pattern_b.
    pattern_b : `numpy.ndarray`, (N, 3)
        A array containing N, 3 vectors representing the reference frame we
        would like to transform pattern_a into.
    max_dist_rad : `float`
        The maximum distance allowed from the pattern matching. This value is
        used as the standard error for the resultant chi values.
 
    Returns
    -------
    noralized_diff : `numpy.ndarray`, (9,)
        Array of differences between the vectors representing of the source
        pattern rotated into the reference frame and the converse. This is
        used to minimize in a least squares fitter.
    """
    # Unflatten the rotation matrix
    rot_matrix = flattened_rot_matrix.reshape((3, 3))
    # Compare the rotated source pattern to the references.
    rot_pattern_a = np.dot(rot_matrix, pattern_a.transpose()).transpose()
    diff_pattern_a_to_b = rot_pattern_a - pattern_b
    # Return the flattened differences and length tolerances for use in a least
    # squares fitter.
    return diff_pattern_a_to_b.flatten() / max_dist_rad
 
 
class PessimisticPatternMatcherB:
    """Class implementing a pessimistic version of Optimistic Pattern Matcher
    B (OPMb) from Tabur 2007. See `DMTN-031 <http://ls.st/DMTN-031`_
 
    Parameters
    ----------
    reference_array : `numpy.ndarray`, (N, 3)
        spherical points x, y, z of to use as reference objects for
        pattern matching.
    log : `lsst.log.Log`
        Logger for outputting debug info.
 
    Notes
    -----
    The class loads and stores the reference object
    in a convenient data structure for matching any set of source objects that
    are assumed to contain each other. The pessimistic nature of the algorithm
    comes from requiring that it discovers at least two patterns that agree on
    the correct shift and rotation for matching before exiting. The original
    behavior of OPMb can be recovered simply. Patterns matched between the
    input datasets are n-spoked pinwheels created from n+1 points. Refer to
    DMTN #031 for more details. http://github.com/lsst-dm/dmtn-031
    """
 
    def __init__(self, reference_array, log):
        self._reference_array_reference_array = reference_array
        self._n_reference_n_reference = len(self._reference_array_reference_array)
        self.loglog = log
 
        if self._n_reference_n_reference > 0:
            self._build_distances_and_angles_build_distances_and_angles()
        else:
            raise ValueError("No reference objects supplied")
 
    def _build_distances_and_angles(self):
        """Create the data structures we will use to search for our pattern
        match in.
 
        Throughout this function and the rest of the class we use id to
        reference the position in the input reference catalog and index to
        'index' into the arrays sorted on distance.
        """
 
        # Initialize the arrays we will need for quick look up of pairs once
        # have a candidate spoke center.
        self._pair_id_array_pair_id_array = np.empty(
            (self._n_reference_n_reference, self._n_reference_n_reference - 1),
            dtype=np.uint16)
        self._pair_dist_array_pair_dist_array = np.empty(
            (self._n_reference_n_reference, self._n_reference_n_reference - 1),
            dtype=np.float32)
 
        # Create empty lists to temporarily store our pair information per
        # reference object. These will be concatenated into our final arrays.
        sub_id_array_list = []
        sub_dist_array_list = []
 
        # Loop over reference objects storing pair distances and ids.
        for ref_id, ref_obj in enumerate(self._reference_array_reference_array):
 
            # Reserve and fill the ids of each reference object pair.
            # 16 bit is safe for the id array as the catalog input from
            # MatchPessimisticB is limited to a max length of 2 ** 16.
            sub_id_array = np.zeros((self._n_reference_n_reference - 1 - ref_id, 2),
                                    dtype=np.uint16)
            sub_id_array[:, 0] = ref_id
            sub_id_array[:, 1] = np.arange(ref_id + 1, self._n_reference_n_reference,
                                           dtype=np.uint16)
 
            # Compute the vector deltas for each pair of reference objects.
            # Compute and store the distances.
            sub_delta_array = (self._reference_array_reference_array[ref_id + 1:, :]
                               - ref_obj).astype(np.float32)
            sub_dist_array = np.sqrt(sub_delta_array[:, 0] ** 2
                                     + sub_delta_array[:, 1] ** 2
                                     + sub_delta_array[:, 2] ** 2)
 
            # Append to our arrays to the output lists for later
            # concatenation.
            sub_id_array_list.append(sub_id_array)
            sub_dist_array_list.append(sub_dist_array)
 
            # Fill the pair look up arrays row wise and then column wise.
            self._pair_id_array_pair_id_array[ref_id, ref_id:] = sub_id_array[:, 1]
            self._pair_dist_array_pair_dist_array[ref_id, ref_id:] = sub_dist_array
 
            # Don't fill the array column wise if we are on the last as
            # these pairs have already been filled by previous
            # iterations.
            if ref_id < self._n_reference_n_reference - 1:
                self._pair_id_array_pair_id_array[ref_id + 1:, ref_id] = ref_id
                self._pair_dist_array_pair_dist_array[ref_id + 1:, ref_id] = sub_dist_array
 
            # Sort each row on distance for fast look up of pairs given
            # the id of one of the objects in the pair.
            sorted_pair_dist_args = self._pair_dist_array_pair_dist_array[ref_id, :].argsort()
            self._pair_dist_array_pair_dist_array[ref_id, :] = self._pair_dist_array_pair_dist_array[
                ref_id, sorted_pair_dist_args]
            self._pair_id_array_pair_id_array[ref_id, :] = self._pair_id_array_pair_id_array[
                ref_id, sorted_pair_dist_args]
 
        # Concatenate our arrays together.
        unsorted_id_array = np.concatenate(sub_id_array_list)
        unsorted_dist_array = np.concatenate(sub_dist_array_list)
 
        # Sort each array on the pair distances for the initial
        # optimistic pattern matcher lookup.
        sorted_dist_args = unsorted_dist_array.argsort()
        self._dist_array_dist_array = unsorted_dist_array[sorted_dist_args]
        self._id_array_id_array = unsorted_id_array[sorted_dist_args]
 
        return None
 
    def match(self, source_array, n_check, n_match, n_agree,
              max_n_patterns, max_shift, max_rotation, max_dist,
              min_matches, pattern_skip_array=None):
        """Match a given source catalog into the loaded reference catalog.
 
        Given array of points on the unit sphere and tolerances, we
        attempt to match a pinwheel like pattern between these input sources
        and the reference objects this class was created with. This pattern
        informs of the shift and rotation needed to align the input source
        objects into the frame of the references.
 
        Parameters
        ----------
        source_array : `numpy.ndarray`, (N, 3)
            An array of spherical x,y,z coordinates and a magnitude in units
            of objects having a lower value for sorting. The array should be
            of shape (N, 4).
        n_check  : `int`
            Number of sources to create a pattern from. Not all objects may be
            checked if n_match criteria is before looping through all n_check
            objects.
        n_match : `int`
            Number of objects to use in constructing a pattern to match.
        n_agree : `int`
            Number of found patterns that must agree on their shift and
            rotation before exiting. Set this value to 1 to recover the
            expected behavior of Optimistic Pattern Matcher B.
        max_n_patters : `int`
            Number of patterns to create from the input source objects to
            attempt to match into the reference objects.
        max_shift : `float`
            Maximum allowed shift to match patterns in arcseconds.
        max_rotation : `float`
            Maximum allowed rotation between patterns in degrees.
        max_dist : `float`
            Maximum distance in arcseconds allowed between candidate spokes in
            the source and reference objects. Also sets that maximum distance
            in the intermediate verify, pattern shift/rotation agreement, and
            final verify steps.
        pattern_skip_array : `int`
            Patterns we would like to skip. This could be due to the pattern
            being matched on a previous iteration that we now consider invalid.
            This assumes the ordering of the source objects is the same
            between different runs of the matcher which, assuming no object
            has been inserted or the magnitudes have changed, it should be.
 
        Returns
        -------
        output_struct : `lsst.pipe.base.Struct`
            Result struct with components
 
            - ``matches`` : (N, 2) array of matched ids for pairs. Empty list if no
              match found (`numpy.ndarray`, (N, 2) or `list`)
            - ``distances_rad`` : Radian distances between the matched objects.
              Empty list if no match found (`numpy.ndarray`, (N,))
            - ``pattern_idx``: Index of matched pattern. None if no match found
              (`int`).
            - ``shift`` : Magnitude for the shift between the source and reference
              objects in arcseconds. None if no match found (`float`).
        """
 
        # Given our input source_array we sort on magnitude.
        sorted_source_array = source_array[source_array[:, -1].argsort(), :3]
        n_source = len(sorted_source_array)
 
        # Initialize output struct.
        output_match_struct = pipeBase.Struct(
            match_ids=[],
            distances_rad=[],
            pattern_idx=None,
            shift=None,
            max_dist_rad=None,)
 
        if n_source <= 0:
            self.loglog.warn("Source object array is empty. Unable to match. "
                          "Exiting matcher.")
            return None
 
        # To test if the shifts and rotations we find agree with each other,
        # we first create two test points situated at the top and bottom of
        # where the z axis on the sphere bisects the source catalog.
        test_vectors = self._compute_test_vectors_compute_test_vectors(source_array[:, :3])
 
        # We now create an empty list of our resultant rotated vectors to
        # compare the different rotations we find.
        rot_vect_list = []
 
        # Convert the tolerances to values we will use in the code.
        max_cos_shift = np.cos(np.radians(max_shift / 3600.))
        max_cos_rot_sq = np.cos(np.radians(max_rotation)) ** 2
        max_dist_rad = np.radians(max_dist / 3600.)
 
        # Loop through the sources from brightest to faintest, grabbing a
        # chunk of n_check each time.
        for pattern_idx in range(np.min((max_n_patterns,
                                         n_source - n_match))):
 
            # If this pattern is one that we matched on the past but we
            # now want to skip, we do so here.
            if pattern_skip_array is not None and \
               np.any(pattern_skip_array == pattern_idx):
                self.loglog.debug(
                    "Skipping previously matched bad pattern %i..." %
                    pattern_idx)
                continue
            # Grab the sources to attempt to create this pattern.
            pattern = sorted_source_array[
                pattern_idx: np.min((pattern_idx + n_check, n_source)), :3]
 
            # Construct a pattern given the number of points defining the
            # pattern complexity. This is the start of the primary tests to
            # match our source pattern into the reference objects.
            construct_return_struct = \
                self._construct_pattern_and_shift_rot_matrix_construct_pattern_and_shift_rot_matrix(
                    pattern, n_match, max_cos_shift, max_cos_rot_sq,
                    max_dist_rad)
 
            # Our struct is None if we could not match the pattern.
            if construct_return_struct.ref_candidates is None or \
               construct_return_struct.shift_rot_matrix is None or \
               construct_return_struct.cos_shift is None or \
               construct_return_struct.sin_rot is None:
                continue
 
            # Grab the output data from the Struct object.
            ref_candidates = construct_return_struct.ref_candidates
            shift_rot_matrix = construct_return_struct.shift_rot_matrix
            cos_shift = construct_return_struct.cos_shift
            sin_rot = construct_return_struct.sin_rot
 
            # If we didn't match enough candidates we continue to the next
            # pattern.
            if len(ref_candidates) < n_match:
                continue
 
            # Now that we know our pattern and shift/rotation are valid we
            # store the the rotated versions of our test points for later
            # use.
            tmp_rot_vect_list = []
            for test_vect in test_vectors:
                tmp_rot_vect_list.append(np.dot(shift_rot_matrix, test_vect))
            # Since our test point are in the source frame, we can test if
            # their lengths are mostly preserved under the transform.
            if not self._test_pattern_lengths_test_pattern_lengths(np.array(tmp_rot_vect_list),
                                              max_dist_rad):
                continue
 
            tmp_rot_vect_list.append(pattern_idx)
            rot_vect_list.append(tmp_rot_vect_list)
 
            # Test if we have enough rotations, which agree, or if we
            # are in optimistic mode.
            if self._test_rotation_agreement_test_rotation_agreement(rot_vect_list, max_dist_rad) < \
               n_agree - 1:
                continue
 
            # Run the final verify step.
            match_struct = self._final_verify_final_verify(source_array[:, :3],
                                              shift_rot_matrix,
                                              max_dist_rad,
                                              min_matches)
            if match_struct.match_ids is None or \
               match_struct.distances_rad is None or \
               match_struct.max_dist_rad is None:
                continue
 
            # Convert the observed shift to arcseconds
            shift = np.degrees(np.arccos(cos_shift)) * 3600.
            # Print information to the logger.
            self.loglog.debug("Succeeded after %i patterns." % pattern_idx)
            self.loglog.debug("\tShift %.4f arcsec" % shift)
            self.loglog.debug("\tRotation: %.4f deg" %
                           np.degrees(np.arcsin(sin_rot)))
 
            # Fill the struct and return.
            output_match_struct.match_ids = \
                match_struct.match_ids
            output_match_struct.distances_rad = \
                match_struct.distances_rad
            output_match_struct.pattern_idx = pattern_idx
            output_match_struct.shift = shift
            output_match_struct.max_dist_rad = match_struct.max_dist_rad
            return output_match_struct
 
        self.loglog.debug("Failed after %i patterns." % (pattern_idx + 1))
        return output_match_struct
 
    def _compute_test_vectors(self, source_array):
        """Compute spherical 3 vectors at the edges of the x, y, z extent
        of the input source catalog.
 
        Parameters
        ----------
        source_array : `numpy.ndarray`, (N, 3)
            array of 3 vectors representing positions on the unit
            sphere.
 
        Returns
        -------
        test_vectors : `numpy.ndarray`, (N, 3)
            Array of vectors representing the maximum extents in x, y, z
            of the input source array. These are used with the rotations
            the code finds to test for agreement from different patterns
            when the code is running in pessimistic mode.
        """
 
        # Get the center of source_array.
        if np.any(np.logical_not(np.isfinite(source_array))):
            self.loglog.warn("Input source objects contain non-finite values. "
                          "This could end badly.")
        center_vect = np.nanmean(source_array, axis=0)
 
        # So that our rotation test works over the full sky we compute
        # the max extent in each Cartesian direction x,y,z.
        xbtm_vect = np.array([np.min(source_array[:, 0]), center_vect[1],
                              center_vect[2]], dtype=np.float64)
        xtop_vect = np.array([np.max(source_array[:, 0]), center_vect[1],
                              center_vect[2]], dtype=np.float64)
        xbtm_vect /= np.sqrt(np.dot(xbtm_vect, xbtm_vect))
        xtop_vect /= np.sqrt(np.dot(xtop_vect, xtop_vect))
 
        ybtm_vect = np.array([center_vect[0], np.min(source_array[:, 1]),
                              center_vect[2]], dtype=np.float64)
        ytop_vect = np.array([center_vect[0], np.max(source_array[:, 1]),
                              center_vect[2]], dtype=np.float64)
        ybtm_vect /= np.sqrt(np.dot(ybtm_vect, ybtm_vect))
        ytop_vect /= np.sqrt(np.dot(ytop_vect, ytop_vect))
 
        zbtm_vect = np.array([center_vect[0], center_vect[1],
                              np.min(source_array[:, 2])], dtype=np.float64)
        ztop_vect = np.array([center_vect[0], center_vect[1],
                              np.max(source_array[:, 2])], dtype=np.float64)
        zbtm_vect /= np.sqrt(np.dot(zbtm_vect, zbtm_vect))
        ztop_vect /= np.sqrt(np.dot(ztop_vect, ztop_vect))
 
        # Return our list of vectors for later rotation testing.
        return np.array([xbtm_vect, xtop_vect, ybtm_vect, ytop_vect,
                         zbtm_vect, ztop_vect])
 
    def _construct_pattern_and_shift_rot_matrix(self, src_pattern_array,
                                                n_match, max_cos_theta_shift,
                                                max_cos_rot_sq, max_dist_rad):
        """Test an input source pattern against the reference catalog.
 
        Returns the candidate matched patterns and their
        implied rotation matrices or None.
 
        Parameters
        ----------
        src_pattern_array : `numpy.ndarray`, (N, 3)
            Sub selection of source 3 vectors to create a pattern from
        n_match : `int`
            Number of points to attempt to create a pattern from. Must be
            >= len(src_pattern_array)
        max_cos_theta_shift : `float`
            Maximum shift allowed between two patterns' centers.
        max_cos_rot_sq : `float`
            Maximum rotation between two patterns that have been shifted
            to have their centers on top of each other.
        max_dist_rad : `float`
            Maximum delta distance allowed between the source and reference
            pair distances to consider the reference pair a candidate for
            the source pair. Also sets the tolerance between the opening
            angles of the spokes when compared to the reference.
 
        Return
        -------
        output_matched_pattern : `lsst.pipe.base.Struct`
            Result struct with components:
 
            - ``ref_candidates`` : ids of the matched pattern in the internal
              reference_array object (`list` of `int`).
            - ``src_candidates`` : Pattern ids of the sources matched
              (`list` of `int`).
            - ``shift_rot_matrix_src_to_ref`` : 3x3 matrix specifying the full
              shift and rotation between the reference and source objects.
              Rotates source into reference frame. `None` if match is not
              found. (`numpy.ndarray`, (3, 3))
            - ``shift_rot_matrix_ref_to_src`` : 3x3 matrix specifying the full
              shift and rotation of the reference and source objects. Rotates
              reference into source frame. None if match is not found
              (`numpy.ndarray`, (3, 3)).
            - ``cos_shift`` : Magnitude of the shift found between the two
              patten centers. `None` if match is not found (`float`).
            - ``sin_rot`` : float value of the rotation to align the already
              shifted source pattern to the reference pattern. `None` if no match
              found (`float`).
        """
 
        # Create our place holder variables for the matched sources and
        # references. The source list starts with the 0th and first indexed
        # objects as we are guaranteed to use those and these define both
        # the shift and rotation of the final pattern.
        output_matched_pattern = pipeBase.Struct(
            ref_candidates=[],
            src_candidates=[],
            shift_rot_matrix=None,
            cos_shift=None,
            sin_rot=None)
 
        # Create the delta vectors and distances we will need to assemble the
        # spokes of the pattern.
        src_delta_array = np.empty((len(src_pattern_array) - 1, 3))
        src_delta_array[:, 0] = (src_pattern_array[1:, 0]
                                 - src_pattern_array[0, 0])
        src_delta_array[:, 1] = (src_pattern_array[1:, 1]
                                 - src_pattern_array[0, 1])
        src_delta_array[:, 2] = (src_pattern_array[1:, 2]
                                 - src_pattern_array[0, 2])
        src_dist_array = np.sqrt(src_delta_array[:, 0]**2
                                 + src_delta_array[:, 1]**2
                                 + src_delta_array[:, 2]**2)
 
        # Our first test. We search the reference dataset for pairs
        # that have the same length as our first source pairs to with
        # plus/minus the max_dist tolerance.
        ref_dist_index_array = self._find_candidate_reference_pairs_find_candidate_reference_pairs(
            src_dist_array[0], self._dist_array_dist_array, max_dist_rad)
 
        # Start our loop over the candidate reference objects.
        for ref_dist_idx in ref_dist_index_array:
            # We have two candidates for which reference object corresponds
            # with the source at the center of our pattern. As such we loop
            # over and test both possibilities.
            tmp_ref_pair_list = self._id_array_id_array[ref_dist_idx]
            for pair_idx, ref_id in enumerate(tmp_ref_pair_list):
                src_candidates = [0, 1]
                ref_candidates = []
                shift_rot_matrix = None
                cos_shift = None
                sin_rot = None
                # Test the angle between our candidate ref center and the
                # source center of our pattern. This angular distance also
                # defines the shift we will later use.
                ref_center = self._reference_array_reference_array[ref_id]
                cos_shift = np.dot(src_pattern_array[0], ref_center)
                if cos_shift < max_cos_theta_shift:
                    continue
 
                # We can now append this one as a candidate.
                ref_candidates.append(ref_id)
                # Test to see which reference object to use in the pair.
                if pair_idx == 0:
                    ref_candidates.append(
                        tmp_ref_pair_list[1])
                    ref_delta = (self._reference_array_reference_array[tmp_ref_pair_list[1]]
                                 - ref_center)
                else:
                    ref_candidates.append(
                        tmp_ref_pair_list[0])
                    ref_delta = (self._reference_array_reference_array[tmp_ref_pair_list[0]]
                                 - ref_center)
 
                # For dense fields it will be faster to compute the absolute
                # rotation this pair suggests first rather than saving it
                # after all the spokes are found. We then compute the cos^2
                # of the rotation and first part of the rotation matrix from
                # source to reference frame.
                test_rot_struct = self._test_rotation_test_rotation(
                    src_pattern_array[0], ref_center, src_delta_array[0],
                    ref_delta, cos_shift, max_cos_rot_sq)
                if test_rot_struct.cos_rot_sq is None or \
                   test_rot_struct.proj_ref_ctr_delta is None or \
                   test_rot_struct.shift_matrix is None:
                    continue
 
                # Get the data from the return struct.
                cos_rot_sq = test_rot_struct.cos_rot_sq
                proj_ref_ctr_delta = test_rot_struct.proj_ref_ctr_delta
                shift_matrix = test_rot_struct.shift_matrix
 
                # Now that we have a candidate first spoke and reference
                # pattern center, we mask our future search to only those
                # pairs that contain our candidate reference center.
                tmp_ref_dist_array = self._pair_dist_array_pair_dist_array[ref_id]
                tmp_ref_id_array = self._pair_id_array_pair_id_array[ref_id]
 
                # Now we feed this sub data to match the spokes of
                # our pattern.
                pattern_spoke_struct = self._create_pattern_spokes_create_pattern_spokes(
                    src_pattern_array[0], src_delta_array, src_dist_array,
                    self._reference_array_reference_array[ref_id], ref_id, proj_ref_ctr_delta,
                    tmp_ref_dist_array, tmp_ref_id_array, max_dist_rad,
                    n_match)
 
                # If we don't find enough candidates we can continue to the
                # next reference center pair.
                if len(pattern_spoke_struct.ref_spoke_list) < n_match - 2 or \
                   len(pattern_spoke_struct.src_spoke_list) < n_match - 2:
                    continue
 
                # If we have the right number of matched ids we store these.
                ref_candidates.extend(pattern_spoke_struct.ref_spoke_list)
                src_candidates.extend(pattern_spoke_struct.src_spoke_list)
 
                # We can now create our full rotation matrix for both the
                # shift and rotation. Reminder shift, aligns the pattern
                # centers, rotation rotates the spokes on top of each other.
                shift_rot_struct = self._create_shift_rot_matrix_create_shift_rot_matrix(
                    cos_rot_sq, shift_matrix, src_delta_array[0],
                    self._reference_array_reference_array[ref_id], ref_delta)
                # If we fail to create the rotation matrix, continue to the
                # next objects.
                if shift_rot_struct.sin_rot is None or \
                   shift_rot_struct.shift_rot_matrix is None:
                    continue
 
                # Get the data from the return struct.
                sin_rot = shift_rot_struct.sin_rot
                shift_rot_matrix = shift_rot_struct.shift_rot_matrix
 
                # Now that we have enough candidates we test to see if it
                # passes intermediate verify. This shifts and rotates the
                # source pattern into the reference frame and then tests that
                # each source/reference object pair is within max_dist. It also
                # tests the opposite rotation that is reference to source
                # frame.
                fit_shift_rot_matrix = self._intermediate_verify_intermediate_verify(
                    src_pattern_array[src_candidates],
                    self._reference_array_reference_array[ref_candidates],
                    shift_rot_matrix, max_dist_rad)
 
                if fit_shift_rot_matrix is not None:
                    # Fill the struct and return.
                    output_matched_pattern.ref_candidates = ref_candidates
                    output_matched_pattern.src_candidates = src_candidates
                    output_matched_pattern.shift_rot_matrix = \
                        fit_shift_rot_matrix
                    output_matched_pattern.cos_shift = cos_shift
                    output_matched_pattern.sin_rot = sin_rot
                    return output_matched_pattern
 
        return output_matched_pattern
 
    def _find_candidate_reference_pairs(self, src_dist, ref_dist_array,
                                        max_dist_rad):
        """Wrap numpy.searchsorted to find the range of reference spokes
        within a spoke distance tolerance of our source spoke.
 
        Returns an array sorted from the smallest absolute delta distance
        between source and reference spoke length. This sorting increases the
        speed for the pattern search greatly.
 
        Parameters
        ----------
        src_dist : `float`
            float value of the distance we would like to search for in
            the reference array in radians.
        ref_dist_array : `numpy.ndarray`, (N,)
            sorted array of distances in radians.
        max_dist_rad : `float`
            maximum plus/minus search to find in the reference array in
            radians.
 
        Return
        ------
        tmp_diff_array : `numpy.ndarray`, (N,)
            indices lookup into the input ref_dist_array sorted by the
            difference in value to the src_dist from absolute value
            smallest to largest.
        """
        # Find the index of the minimum and maximum values that satisfy
        # the tolerance.
        start_idx = np.searchsorted(ref_dist_array, src_dist - max_dist_rad)
        end_idx = np.searchsorted(ref_dist_array, src_dist + max_dist_rad,
                                  side='right')
 
        # If these are equal there are no candidates and we exit.
        if start_idx == end_idx:
            return []
 
        # Make sure the endpoints of the input array are respected.
        if start_idx < 0:
            start_idx = 0
        if end_idx > ref_dist_array.shape[0]:
            end_idx = ref_dist_array.shape[0]
 
        # Now we sort the indices from smallest absolute delta dist difference
        # to the largest and return the vector. This step greatly increases
        # the speed of the algorithm.
        tmp_diff_array = np.fabs(ref_dist_array[start_idx:end_idx] - src_dist)
        return tmp_diff_array.argsort() + start_idx
 
    def _test_rotation(self, src_center, ref_center, src_delta, ref_delta,
                       cos_shift, max_cos_rot_sq):
        """ Test if the rotation implied between the source
        pattern and reference pattern is within tolerance. To test this
        we need to create the first part of our spherical rotation matrix
        which we also return for use later.
 
        Parameters
        ----------
        src_center : `numpy.ndarray`, (N, 3)
            pattern.
        ref_center : `numpy.ndarray`, (N, 3)
            3 vector defining the center of the candidate reference pinwheel
            pattern.
        src_delta : `numpy.ndarray`, (N, 3)
            3 vector delta between the source pattern center and the end of
            the pinwheel spoke.
        ref_delta : `numpy.ndarray`, (N, 3)
            3 vector delta of the candidate matched reference pair
        cos_shift : `float`
            Cosine of the angle between the source and reference candidate
            centers.
        max_cos_rot_sq : `float`
            candidate reference pair after shifting the centers on top of each
            other. The function will return None if the rotation implied is
            greater than max_cos_rot_sq.
 
        Returns
        -------
        result : `lsst.pipe.base.Struct`
            Result struct with components:
 
            - ``cos_rot_sq`` :  magnitude of the rotation needed to align the
              two patterns after their centers are shifted on top of each
              other. `None` if rotation test fails (`float`).
            - ``shift_matrix`` : 3x3 rotation matrix describing the shift needed to
            align the source and candidate reference center. `None` if rotation
            test fails (`numpy.ndarray`, (N, 3)).
        """
 
        # Make sure the sine is a real number.
        if cos_shift > 1.0:
            cos_shift = 1.
        elif cos_shift < -1.0:
            cos_shift = -1.
        sin_shift = np.sqrt(1 - cos_shift ** 2)
 
        # If the sine of our shift is zero we only need to use the identity
        # matrix for the shift. Else we construct the rotation matrix for
        # shift.
        if sin_shift > 0:
            rot_axis = np.cross(src_center, ref_center)
            rot_axis /= sin_shift
            shift_matrix = self._create_spherical_rotation_matrix_create_spherical_rotation_matrix(
                rot_axis, cos_shift, sin_shift)
        else:
            shift_matrix = np.identity(3)
 
        # Now that we have our shift we apply it to the src delta vector
        # and check the rotation.
        rot_src_delta = np.dot(shift_matrix, src_delta)
        proj_src_delta = (rot_src_delta
                          - np.dot(rot_src_delta, ref_center) * ref_center)
        proj_ref_delta = (ref_delta
                          - np.dot(ref_delta, ref_center) * ref_center)
        cos_rot_sq = (np.dot(proj_src_delta, proj_ref_delta) ** 2
                      / (np.dot(proj_src_delta, proj_src_delta)
                      * np.dot(proj_ref_delta, proj_ref_delta)))
        # If the rotation isn't in tolerance return None.
        if cos_rot_sq < max_cos_rot_sq:
            return pipeBase.Struct(
                cos_rot_sq=None,
                proj_ref_ctr_delta=None,
                shift_matrix=None,)
        # Return the rotation angle, the plane projected reference vector,
        # and the first half of the full shift and rotation matrix.
        return pipeBase.Struct(
            cos_rot_sq=cos_rot_sq,
            proj_ref_ctr_delta=proj_ref_delta,
            shift_matrix=shift_matrix,)
 
    def _create_spherical_rotation_matrix(self, rot_axis, cos_rotation,
                                          sin_rotion):
        """Construct a generalized 3D rotation matrix about a given
        axis.
 
        Parameters
        ----------
        rot_axis : `numpy.ndarray`, (3,)
            3 vector defining the axis to rotate about.
        cos_rotation : `float`
            cosine of the rotation angle.
        sin_rotation : `float`
            sine of the rotation angle.
 
        Return
        ------
        shift_matrix : `numpy.ndarray`, (3, 3)
            3x3 spherical, rotation matrix.
        """
 
        rot_cross_matrix = np.array(
            [[0., -rot_axis[2], rot_axis[1]],
             [rot_axis[2], 0., -rot_axis[0]],
             [-rot_axis[1], rot_axis[0], 0.]], dtype=np.float64)
        shift_matrix = (cos_rotation*np.identity(3)
                        + sin_rotion*rot_cross_matrix
                        + (1. - cos_rotation)*np.outer(rot_axis, rot_axis))
 
        return shift_matrix
 
    def _create_pattern_spokes(self, src_ctr, src_delta_array, src_dist_array,
                               ref_ctr, ref_ctr_id, proj_ref_ctr_delta,
                               ref_dist_array, ref_id_array, max_dist_rad,
                               n_match):
        """ Create the individual spokes that make up the pattern now that the
        shift and rotation are within tolerance.
 
        If we can't create a valid pattern we exit early.
 
        Parameters
        ----------
        src_ctr : `numpy.ndarray`, (3,)
            3 vector of the source pinwheel center
        src_delta_array : `numpy.ndarray`, (N, 3)
            Array of 3 vector deltas between the source center and the pairs
            that make up the remaining spokes of the pinwheel
        src_dist_array : `numpy.ndarray`, (N, 3)
            Array of the distances of each src_delta in the pinwheel
        ref_ctr : `numpy.ndarray`, (3,)
            3 vector of the candidate reference center
        ref_ctr_id : `int`
            id of the ref_ctr in the master reference array
        proj_ref_ctr_delta : `numpy.ndarray`, (3,)
            Plane projected 3 vector formed from the center point of the
            candidate pin-wheel and the second point in the pattern to create
            the first spoke pair. This is the candidate pair that was matched
            in the main _construct_pattern_and_shift_rot_matrix loop
        ref_dist_array : `numpy.ndarray`, (N,)
            Array of vector distances for each of the reference pairs
        ref_id_array : `numpy.ndarray`, (N,)
            Array of id lookups into the master reference array that our
            center id object is paired with.
        max_dist_rad : `float`
            Maximum search distance
        n_match : `int`
            Number of source deltas that must be matched into the reference
            deltas in order to consider this a successful pattern match.
 
        Returns
        -------
        output_spokes : `lsst.pipe.base.Struct`
            Result struct with components:
 
            - ``ref_spoke_list`` : list of ints specifying ids into the master
              reference array (`list` of `int`).
            - ``src_spoke_list`` : list of ints specifying indices into the
              current source pattern that is being tested (`list` of `int`).
        """
        # Struct where we will be putting our results.
        output_spokes = pipeBase.Struct(
            ref_spoke_list=[],
            src_spoke_list=[],)
 
        # Counter for number of spokes we failed to find a reference
        # candidate for. We break the loop if we haven't found enough.
        n_fail = 0
        ref_spoke_list = []
        src_spoke_list = []
 
        # Plane project the center/first spoke of the source pattern using
        # the center vector of the pattern as normal.
        proj_src_ctr_delta = (src_delta_array[0]
                              - np.dot(src_delta_array[0], src_ctr) * src_ctr)
        proj_src_ctr_dist_sq = np.dot(proj_src_ctr_delta, proj_src_ctr_delta)
 
        # Pre-compute the squared length of the projected reference vector.
        proj_ref_ctr_dist_sq = np.dot(proj_ref_ctr_delta, proj_ref_ctr_delta)
 
        # Loop over the source pairs.
        for src_idx in range(1, len(src_dist_array)):
            if n_fail > len(src_dist_array) - (n_match - 1):
                break
 
            # Given our length tolerance we can use it to compute a tolerance
            # on the angle between our spoke.
            src_sin_tol = (max_dist_rad
                           / (src_dist_array[src_idx] + max_dist_rad))
 
            # Test if the small angle approximation will still hold. This is
            # defined as when sin(theta) ~= theta to within 0.1% of each
            # other. If the implied opening angle is too large we set it to
            # the 0.1% threshold.
            max_sin_tol = 0.0447
            if src_sin_tol > max_sin_tol:
                src_sin_tol = max_sin_tol
 
            # Plane project the candidate source spoke and compute the cosine
            # and sine of the opening angle.
            proj_src_delta = (
                src_delta_array[src_idx]
                - np.dot(src_delta_array[src_idx], src_ctr) * src_ctr)
            geom_dist_src = np.sqrt(
                np.dot(proj_src_delta, proj_src_delta)
                * proj_src_ctr_dist_sq)
 
            # Compute cosine and sine of the delta vector opening angle.
            cos_theta_src = (np.dot(proj_src_delta, proj_src_ctr_delta)
                             / geom_dist_src)
            cross_src = (np.cross(proj_src_delta, proj_src_ctr_delta)
                         / geom_dist_src)
            sin_theta_src = np.dot(cross_src, src_ctr)
 
            # Find the reference pairs that include our candidate pattern
            # center and sort them in increasing delta
            ref_dist_idx_array = self._find_candidate_reference_pairs_find_candidate_reference_pairs(
                src_dist_array[src_idx], ref_dist_array, max_dist_rad)
 
            # Test the spokes and return the id of the reference object.
            # Return None if no match is found.
            ref_id = self._test_spoke_test_spoke(
                cos_theta_src,
                sin_theta_src,
                ref_ctr,
                ref_ctr_id,
                proj_ref_ctr_delta,
                proj_ref_ctr_dist_sq,
                ref_dist_idx_array,
                ref_id_array,
                src_sin_tol)
            if ref_id is None:
                n_fail += 1
                continue
 
            # Append the successful indices to our list. The src_idx needs
            # an extra iteration to skip the first and second source objects.
            ref_spoke_list.append(ref_id)
            src_spoke_list.append(src_idx + 1)
            # If we found enough reference objects we can return early. This is
            # n_match - 2 as we already have 2 source objects matched into the
            # reference data.
            if len(ref_spoke_list) >= n_match - 2:
                # Set the struct data and return the struct.
                output_spokes.ref_spoke_list = ref_spoke_list
                output_spokes.src_spoke_list = src_spoke_list
                return output_spokes
 
        return output_spokes
 
    def _test_spoke(self, cos_theta_src, sin_theta_src, ref_ctr, ref_ctr_id,
                    proj_ref_ctr_delta, proj_ref_ctr_dist_sq,
                    ref_dist_idx_array, ref_id_array, src_sin_tol):
        """Test the opening angle between the first spoke of our pattern
        for the source object against the reference object.
 
        This method makes heavy use of the small angle approximation to perform
        the comparison.
 
        Parameters
        ----------
        cos_theta_src : `float`
            Cosine of the angle between the current candidate source spoke and
            the first spoke.
        sin_theta_src : `float`
            Sine of the angle between the current candidate source spoke and
            the first spoke.
        ref_ctr : `numpy.ndarray`, (3,)
            3 vector of the candidate reference center
        ref_ctr_id : `int`
            id lookup of the ref_ctr into the master reference array
        proj_ref_ctr_delta : `float`
            Plane projected first spoke in the reference pattern using the
            pattern center as normal.
        proj_ref_ctr_dist_sq : `float`
            Squared length of the projected vector.
        ref_dist_idx_array : `numpy.ndarray`, (N,)
            Indices sorted by the delta distance between the source
            spoke we are trying to test and the candidate reference
            spokes.
        ref_id_array : `numpy.ndarray`, (N,)
            Array of id lookups into the master reference array that our
            center id object is paired with.
        src_sin_tol : `float`
            Sine of tolerance allowed between source and reference spoke
            opening angles.
 
        Returns
        -------
        id : `int`
            If we can not find a candidate spoke we return `None` else we
            return an int id into the master reference array.
        """
 
        # Loop over our candidate reference objects.
        for ref_dist_idx in ref_dist_idx_array:
            # Compute the delta vector from the pattern center.
            ref_delta = (self._reference_array_reference_array[ref_id_array[ref_dist_idx]]
                         - ref_ctr)
            # Compute the cos between our "center" reference vector and the
            # current reference candidate.
            proj_ref_delta = ref_delta - np.dot(ref_delta, ref_ctr) * ref_ctr
            geom_dist_ref = np.sqrt(proj_ref_ctr_dist_sq
                                    * np.dot(proj_ref_delta, proj_ref_delta))
            cos_theta_ref = (np.dot(proj_ref_delta, proj_ref_ctr_delta)
                             / geom_dist_ref)
 
            # Make sure we can safely make the comparison in case
            # our "center" and candidate vectors are mostly aligned.
            if cos_theta_ref ** 2 < (1 - src_sin_tol ** 2):
                cos_sq_comparison = ((cos_theta_src - cos_theta_ref) ** 2
                                     / (1 - cos_theta_ref ** 2))
            else:
                cos_sq_comparison = ((cos_theta_src - cos_theta_ref) ** 2
                                     / src_sin_tol ** 2)
            # Test the difference of the cosine of the reference angle against
            # the source angle. Assumes that the delta between the two is
            # small.
            if cos_sq_comparison > src_sin_tol ** 2:
                continue
 
            # The cosine tests the magnitude of the angle but not
            # its direction. To do that we need to know the sine as well.
            # This cross product calculation does that.
            cross_ref = (np.cross(proj_ref_delta, proj_ref_ctr_delta)
                         / geom_dist_ref)
            sin_theta_ref = np.dot(cross_ref, ref_ctr)
 
            # Check the value of the cos again to make sure that it is not
            # near zero.
            if abs(cos_theta_src) < src_sin_tol:
                sin_comparison = (sin_theta_src - sin_theta_ref) / src_sin_tol
            else:
                sin_comparison = \
                    (sin_theta_src - sin_theta_ref) / cos_theta_ref
 
            if abs(sin_comparison) > src_sin_tol:
                continue
 
            # Return the correct id of the candidate we found.
            return ref_id_array[ref_dist_idx]
 
        return None
 
    def _create_shift_rot_matrix(self, cos_rot_sq, shift_matrix, src_delta,
                                 ref_ctr, ref_delta):
        """ Create the final part of our spherical rotation matrix.
 
        Parameters
        ----------
        cos_rot_sq : `float`
            cosine of the rotation needed to align our source and reference
            candidate patterns.
        shift_matrix : `numpy.ndarray`, (3, 3)
            3x3 rotation matrix for shifting the source pattern center on top
            of the candidate reference pattern center.
        src_delta : `numpy.ndarray`, (3,)
            3 vector delta of representing the first spoke of the source
            pattern
        ref_ctr : `numpy.ndarray`, (3,)
            3 vector on the unit-sphere representing the center of our
            reference pattern.
        ref_delta : `numpy.ndarray`, (3,)
            3 vector delta made by the first pair of the reference pattern.
 
        Returns
        -------
        result : `lsst.pipe.base.Struct`
            Result struct with components:
 
            - ``sin_rot`` : float sine of the amount of rotation between the
              source and reference pattern. We use sine here as it is
              signed and tells us the chirality of the rotation (`float`).
            - ``shift_rot_matrix`` : float array representing the 3x3 rotation
              matrix that takes the source pattern and shifts and rotates
              it to align with the reference pattern (`numpy.ndarray`, (3,3)).
        """
        cos_rot = np.sqrt(cos_rot_sq)
        rot_src_delta = np.dot(shift_matrix, src_delta)
        delta_dot_cross = np.dot(np.cross(rot_src_delta, ref_delta), ref_ctr)
 
        sin_rot = np.sign(delta_dot_cross) * np.sqrt(1 - cos_rot_sq)
        rot_matrix = self._create_spherical_rotation_matrix_create_spherical_rotation_matrix(
            ref_ctr, cos_rot, sin_rot)
 
        shift_rot_matrix = np.dot(rot_matrix, shift_matrix)
 
        return pipeBase.Struct(
            sin_rot=sin_rot,
            shift_rot_matrix=shift_rot_matrix,)
 
    def _intermediate_verify(self, src_pattern, ref_pattern, shift_rot_matrix,
                             max_dist_rad):
        """ Perform an intermediate verify step.
 
        Rotate the matches references into the source frame and test their
        distances against tolerance. Only return true if all points are within
        tolerance.
 
        Parameters
        ----------
        src_pattern : `numpy.ndarray`, (N,3)
            Array of 3 vectors representing the points that make up our source
            pinwheel pattern.
        ref_pattern : `numpy.ndarray`, (N,3)
            Array of 3 vectors representing our candidate reference pinwheel
            pattern.
        shift_rot_matrix : `numpy.ndarray`, (3,3)
            3x3 rotation matrix that takes the source objects and rotates them
            onto the frame of the reference objects
        max_dist_rad : `float`
            Maximum distance allowed to consider two objects the same.
 
        Returns
        -------
        fit_shift_rot_matrix : `numpy.ndarray`, (3,3)
           Return the fitted shift/rotation matrix if all of the points in our
           source pattern are within max_dist_rad of their matched reference
           objects. Returns None if this criteria is not satisfied.
        """
        if len(src_pattern) != len(ref_pattern):
            raise ValueError(
                "Source pattern length does not match ref pattern.\n"
                "\t source pattern len=%i, reference pattern len=%i" %
                (len(src_pattern), len(ref_pattern)))
 
        if self._intermediate_verify_comparison_intermediate_verify_comparison(
                src_pattern, ref_pattern, shift_rot_matrix, max_dist_rad):
            # Now that we know our initial shift and rot matrix is valid we
            # want to fit the implied transform using all points from
            # our pattern. This is a more robust rotation matrix as our
            # initial matrix only used the first 2 points from the source
            # pattern to estimate the shift and rotation. The matrix we fit
            # are allowed to be non unitary but need to preserve the length of
            # the rotated vectors to within the match tolerance.
            fit_shift_rot_matrix = least_squares(
                _rotation_matrix_chi_sq,
                x0=shift_rot_matrix.flatten(),
                args=(src_pattern, ref_pattern, max_dist_rad)
            ).x.reshape((3, 3))
            # Do another verify in case the fits have wondered off.
            if self._intermediate_verify_comparison_intermediate_verify_comparison(
                    src_pattern, ref_pattern, fit_shift_rot_matrix,
                    max_dist_rad):
                return fit_shift_rot_matrix
 
        return None
 
    def _intermediate_verify_comparison(self, pattern_a, pattern_b,
                                        shift_rot_matrix, max_dist_rad):
        """Test the input rotation matrix against one input pattern and
        a second one.
 
        If every point in the pattern after rotation is within a distance of
        max_dist_rad to its candidate point in the other pattern, we return
        True.
 
        Parameters
        ----------
        pattern_a :  `numpy.ndarray`, (N,3)
            Array of 3 vectors representing the points that make up our source
            pinwheel pattern.
        pattern_b :  `numpy.ndarray`, (N,3)
            Array of 3 vectors representing our candidate reference pinwheel
            pattern.
        shift_rot_matrix :  `numpy.ndarray`, (3,3)
            3x3 rotation matrix that takes the source objects and rotates them
            onto the frame of the reference objects
        max_dist_rad : `float`
            Maximum distance allowed to consider two objects the same.
 
 
        Returns
        -------
        bool
            True if all rotated source points are within max_dist_rad of
            the candidate references matches.
        """
        shifted_pattern_a = np.dot(shift_rot_matrix,
                                   pattern_a.transpose()).transpose()
        tmp_delta_array = shifted_pattern_a - pattern_b
        tmp_dist_array = (tmp_delta_array[:, 0] ** 2
                          + tmp_delta_array[:, 1] ** 2
                          + tmp_delta_array[:, 2] ** 2)
        return np.all(tmp_dist_array < max_dist_rad ** 2)
 
    def _test_pattern_lengths(self, test_pattern, max_dist_rad):
        """ Test that the all vectors in a pattern are unit length within
        tolerance.
 
        This is useful for assuring the non unitary transforms do not contain
        too much distortion.
 
        Parameters
        ----------
        test_pattern : `numpy.ndarray`, (N, 3)
            Test vectors at the maximum and minimum x, y, z extents.
        max_dist_rad : `float`
            maximum distance in radians to consider two points "agreeing" on
            a rotation
 
        Returns
        -------
        test : `bool`
            Length tests pass.
        """
        dists = (test_pattern[:, 0] ** 2
                 + test_pattern[:, 1] ** 2
                 + test_pattern[:, 2] ** 2)
        return np.all(
            np.logical_and((1 - max_dist_rad) ** 2 < dists,
                           dists < (1 + max_dist_rad) ** 2))
 
    def _test_rotation_agreement(self, rot_vects, max_dist_rad):
        """ Test this rotation against the previous N found and return
        the number that a agree within tolerance to where our test
        points are.
 
        Parameters
        ----------
        rot_vects : `numpy.ndarray`, (N, 3)
            Arrays of rotated 3 vectors representing the maximum x, y,
            z extent on the unit sphere of the input source objects rotated by
            the candidate rotations into the reference frame.
        max_dist_rad : `float`
            maximum distance in radians to consider two points "agreeing" on
            a rotation
 
        Returns
        -------
        tot_consent : `int`
            Number of candidate rotations that agree for all of the rotated
            test 3 vectors.
        """
 
        self.loglog.debug("Comparing pattern %i to previous %i rotations..." %
                       (rot_vects[-1][-1], len(rot_vects) - 1))
 
        tot_consent = 0
        for rot_idx in range(max((len(rot_vects) - 1), 0)):
            tmp_dist_list = []
            for vect_idx in range(len(rot_vects[rot_idx]) - 1):
                tmp_delta_vect = (rot_vects[rot_idx][vect_idx]
                                  - rot_vects[-1][vect_idx])
                tmp_dist_list.append(
                    np.dot(tmp_delta_vect, tmp_delta_vect))
            if np.all(np.array(tmp_dist_list) < max_dist_rad ** 2):
                tot_consent += 1
        return tot_consent
 
    def _final_verify(self,
                      source_array,
                      shift_rot_matrix,
                      max_dist_rad,
                      min_matches):
        """Match the all sources into the reference catalog using the shift/rot
        matrix.
 
        After the initial shift/rot matrix is found, we refit the shift/rot
        matrix using the matches the initial matrix produces to find a more
        stable solution.
 
        Parameters
        ----------
        source_array : `numpy.ndarray` (N, 3)
            3-vector positions on the unit-sphere representing the sources to
            match
        shift_rot_matrix : `numpy.ndarray` (3, 3)
            Rotation matrix representing inferred shift/rotation of the
            sources onto the reference catalog. Matrix need not be unitary.
        max_dist_rad : `float`
            Maximum distance allowed for a match.
        min_matches : `int`
            Minimum number of matched objects required to consider the
            match good.
 
        Returns
        -------
        output_struct : `lsst.pipe.base.Struct`
            Result struct with components:
 
            - ``match_ids`` : Pairs of indexes into the source and reference
              data respectively defining a match (`numpy.ndarray`, (N, 2)).
            - ``distances_rad`` : distances to between the matched objects in
              the shift/rotated frame. (`numpy.ndarray`, (N,)).
            - ``max_dist_rad`` : Value of the max matched distance. Either
              returning the input value of the 2 sigma clipped value of the
              shift/rotated distances. (`float`).
        """
        output_struct = pipeBase.Struct(
            match_ids=None,
            distances_rad=None,
            max_dist_rad=None,
        )
 
        # Perform an iterative final verify.
        match_sources_struct = self._match_sources_match_sources(source_array,
                                                   shift_rot_matrix)
        cut_ids = match_sources_struct.match_ids[
            match_sources_struct.distances_rad < max_dist_rad]
 
        n_matched = len(cut_ids)
        clipped_struct = self._clip_distances_clip_distances(
            match_sources_struct.distances_rad)
        n_matched_clipped = clipped_struct.n_matched_clipped
 
        if n_matched < min_matches or n_matched_clipped < min_matches:
            return output_struct
 
        # The shift/rotation matrix returned by
        # ``_construct_pattern_and_shift_rot_matrix``, above, was
        # based on only six points. Here, we refine that result by
        # using all of the good matches from the “final verification”
        # step, above. This will produce a more consistent result.
        fit_shift_rot_matrix = least_squares(
            _rotation_matrix_chi_sq,
            x0=shift_rot_matrix.flatten(),
            args=(source_array[cut_ids[:, 0], :3],
                  self._reference_array_reference_array[cut_ids[:, 1], :3],
                  max_dist_rad)
        ).x.reshape((3, 3))
 
        # Redo the matching using the newly fit shift/rotation matrix.
        match_sources_struct = self._match_sources_match_sources(
            source_array, fit_shift_rot_matrix)
 
        # Double check the match distances to make sure enough matches
        # survive still. We'll just overwrite the previously used variables.
        n_matched = np.sum(
            match_sources_struct.distances_rad < max_dist_rad)
        clipped_struct = self._clip_distances_clip_distances(
            match_sources_struct.distances_rad)
        n_matched_clipped = clipped_struct.n_matched_clipped
        clipped_max_dist = clipped_struct.clipped_max_dist
 
        if n_matched < min_matches or n_matched_clipped < min_matches:
            return output_struct
 
        # Store our matches in the output struct. Decide between the clipped
        # distance and the input max dist based on which is smaller.
        output_struct.match_ids = match_sources_struct.match_ids
        output_struct.distances_rad = match_sources_struct.distances_rad
        if clipped_max_dist < max_dist_rad:
            output_struct.max_dist_rad = clipped_max_dist
        else:
            output_struct.max_dist_rad = max_dist_rad
 
        return output_struct
 
    def _clip_distances(self, distances_rad):
        """Compute a clipped max distance and calculate the number of pairs
        that pass the clipped dist.
 
        Parameters
        ----------
        distances_rad : `numpy.ndarray`, (N,)
            Distances between pairs.
 
        Returns
        -------
        output_struct : `lsst.pipe.base.Struct`
            Result struct with components:
 
            - ``n_matched_clipped`` : Number of pairs that survive the
              clipping on distance. (`float`)
            - ``clipped_max_dist`` : Maximum distance after clipping.
              (`float`).
        """
        clipped_dists, _, clipped_max_dist = sigmaclip(
            distances_rad,
            low=100,
            high=2)
        # Check clipped distances. The minimum value here
        # prevents over convergence on perfect test data.
        if clipped_max_dist < 1e-16:
            clipped_max_dist = 1e-16
            n_matched_clipped = np.sum(distances_rad < clipped_max_dist)
        else:
            n_matched_clipped = len(clipped_dists)
 
        return pipeBase.Struct(n_matched_clipped=n_matched_clipped,
                               clipped_max_dist=clipped_max_dist)
 
    def _match_sources(self,
                       source_array,
                       shift_rot_matrix):
        """ Shift both the reference and source catalog to the the respective
        frames and find their nearest neighbor using a kdTree.
 
        Removes all matches who do not agree when either the reference or
        source catalog is rotated. Cuts on a maximum distance are left to an
        external function.
 
        Parameters
        ----------
        source_array : `numpy.ndarray`, (N, 3)
            array of 3 vectors representing the source objects we are trying
            to match into the source catalog.
        shift_rot_matrix : `numpy.ndarray`, (3, 3)
            3x3 rotation matrix that performs the spherical rotation from the
            source frame into the reference frame.
 
        Returns
        -------
        results : `lsst.pipe.base.Struct`
            Result struct with components:
 
            - ``matches`` : array of integer ids into the source and
              reference arrays. Matches are only returned for those that
              satisfy the distance and handshake criteria
              (`numpy.ndarray`, (N, 2)).
            - ``distances`` : Distances between each match in radians after
              the shift and rotation is applied (`numpy.ndarray`, (N)).
        """
        shifted_references = np.dot(
            np.linalg.inv(shift_rot_matrix),
            self._reference_array_reference_array.transpose()).transpose()
        shifted_sources = np.dot(
            shift_rot_matrix,
            source_array.transpose()).transpose()
 
        ref_matches = np.empty((len(shifted_references), 2),
                               dtype=np.uint16)
        src_matches = np.empty((len(shifted_sources), 2),
                               dtype=np.uint16)
 
        ref_matches[:, 1] = np.arange(len(shifted_references),
                                      dtype=np.uint16)
        src_matches[:, 0] = np.arange(len(shifted_sources),
                                      dtype=np.uint16)
 
        ref_kdtree = cKDTree(self._reference_array_reference_array)
        src_kdtree = cKDTree(source_array)
 
        ref_to_src_dist, tmp_ref_to_src_idx = \
            src_kdtree.query(shifted_references)
        src_to_ref_dist, tmp_src_to_ref_idx = \
            ref_kdtree.query(shifted_sources)
 
        ref_matches[:, 0] = tmp_ref_to_src_idx
        src_matches[:, 1] = tmp_src_to_ref_idx
 
        handshake_mask = self._handshake_match_handshake_match(src_matches, ref_matches)
        return pipeBase.Struct(
            match_ids=src_matches[handshake_mask],
            distances_rad=src_to_ref_dist[handshake_mask],)
 
    def _handshake_match(self, matches_src, matches_ref):
        """Return only those matches where both the source
        and reference objects agree they they are each others'
        nearest neighbor.
 
        Parameters
        ----------
        matches_src : `numpy.ndarray`, (N, 2)
            int array of nearest neighbor matches between shifted and
            rotated reference objects matched into the sources.
        matches_ref : `numpy.ndarray`, (N, 2)
            int array of nearest neighbor matches between shifted and
            rotated source objects matched into the references.
        Return
        ------
        handshake_mask_array : `numpy.ndarray`, (N,)
           Return the array positions where the two match catalogs agree.
        """
        handshake_mask_array = np.zeros(len(matches_src), dtype=bool)
 
        for src_match_idx, match in enumerate(matches_src):
            ref_match_idx = np.searchsorted(matches_ref[:, 1], match[1])
            if match[0] == matches_ref[ref_match_idx, 0]:
                handshake_mask_array[src_match_idx] = True
        return handshake_mask_array