scipy 1.16.2__cp311-cp311-win_arm64.whl

scipy/sparse/linalg/_special_sparse_arrays.py

@@ -0,0 +1,949 @@
+import numpy as np
+from scipy.sparse.linalg import LinearOperator
+from scipy.sparse import kron, eye_array, dia_array
+
+__all__ = ['LaplacianNd']
+# Sakurai and Mikota classes are intended for tests and benchmarks
+# and explicitly not included in the public API of this module.
+
+
+class LaplacianNd(LinearOperator):
+    """
+    The grid Laplacian in ``N`` dimensions and its eigenvalues/eigenvectors.
+
+    Construct Laplacian on a uniform rectangular grid in `N` dimensions
+    and output its eigenvalues and eigenvectors.
+    The Laplacian ``L`` is square, negative definite, real symmetric array
+    with signed integer entries and zeros otherwise.
+
+    Parameters
+    ----------
+    grid_shape : tuple
+        A tuple of integers of length ``N`` (corresponding to the dimension of
+        the Lapacian), where each entry gives the size of that dimension. The
+        Laplacian matrix is square of the size ``np.prod(grid_shape)``.
+    boundary_conditions : {'neumann', 'dirichlet', 'periodic'}, optional
+        The type of the boundary conditions on the boundaries of the grid.
+        Valid values are ``'dirichlet'`` or ``'neumann'``(default) or
+        ``'periodic'``.
+    dtype : dtype
+        Numerical type of the array. Default is ``np.int8``.
+
+    Methods
+    -------
+    toarray()
+        Construct a dense array from Laplacian data
+    tosparse()
+        Construct a sparse array from Laplacian data
+    eigenvalues(m=None)
+        Construct a 1D array of `m` largest (smallest in absolute value)
+        eigenvalues of the Laplacian matrix in ascending order.
+    eigenvectors(m=None):
+        Construct the array with columns made of `m` eigenvectors (``float``)
+        of the ``Nd`` Laplacian corresponding to the `m` ordered eigenvalues.
+
+    .. versionadded:: 1.12.0
+
+    Notes
+    -----
+    Compared to the MATLAB/Octave implementation [1] of 1-, 2-, and 3-D
+    Laplacian, this code allows the arbitrary N-D case and the matrix-free
+    callable option, but is currently limited to pure Dirichlet, Neumann or
+    Periodic boundary conditions only.
+
+    The Laplacian matrix of a graph (`scipy.sparse.csgraph.laplacian`) of a
+    rectangular grid corresponds to the negative Laplacian with the Neumann
+    conditions, i.e., ``boundary_conditions = 'neumann'``.
+
+    All eigenvalues and eigenvectors of the discrete Laplacian operator for
+    an ``N``-dimensional  regular grid of shape `grid_shape` with the grid
+    step size ``h=1`` are analytically known [2].
+
+    References
+    ----------
+    .. [1] https://github.com/lobpcg/blopex/blob/master/blopex_\
+tools/matlab/laplacian/laplacian.m
+    .. [2] "Eigenvalues and eigenvectors of the second derivative", Wikipedia
+           https://en.wikipedia.org/wiki/Eigenvalues_and_eigenvectors_\
+of_the_second_derivative
+
+    Examples
+    --------
+    >>> import numpy as np
+    >>> from scipy.sparse.linalg import LaplacianNd
+    >>> from scipy.sparse import diags_array, csgraph
+    >>> from scipy.linalg import eigvalsh
+
+    The one-dimensional Laplacian demonstrated below for pure Neumann boundary
+    conditions on a regular grid with ``n=6`` grid points is exactly the
+    negative graph Laplacian for the undirected linear graph with ``n``
+    vertices using the sparse adjacency matrix ``G`` represented by the
+    famous tri-diagonal matrix:
+
+    >>> n = 6
+    >>> G = diags_array(np.ones(n - 1), offsets=1, format='csr')
+    >>> Lf = csgraph.laplacian(G, symmetrized=True, form='function')
+    >>> grid_shape = (n, )
+    >>> lap = LaplacianNd(grid_shape, boundary_conditions='neumann')
+    >>> np.array_equal(lap.matmat(np.eye(n)), -Lf(np.eye(n)))
+    True
+
+    Since all matrix entries of the Laplacian are integers, ``'int8'`` is
+    the default dtype for storing matrix representations.
+
+    >>> lap.tosparse()
+    <DIAgonal sparse array of dtype 'int8'
+        with 16 stored elements (3 diagonals) and shape (6, 6)>
+    >>> lap.toarray()
+    array([[-1,  1,  0,  0,  0,  0],
+           [ 1, -2,  1,  0,  0,  0],
+           [ 0,  1, -2,  1,  0,  0],
+           [ 0,  0,  1, -2,  1,  0],
+           [ 0,  0,  0,  1, -2,  1],
+           [ 0,  0,  0,  0,  1, -1]], dtype=int8)
+    >>> np.array_equal(lap.matmat(np.eye(n)), lap.toarray())
+    True
+    >>> np.array_equal(lap.tosparse().toarray(), lap.toarray())
+    True
+
+    Any number of extreme eigenvalues and/or eigenvectors can be computed.
+
+    >>> lap = LaplacianNd(grid_shape, boundary_conditions='periodic')
+    >>> lap.eigenvalues()
+    array([-4., -3., -3., -1., -1.,  0.])
+    >>> lap.eigenvalues()[-2:]
+    array([-1.,  0.])
+    >>> lap.eigenvalues(2)
+    array([-1.,  0.])
+    >>> lap.eigenvectors(1)
+    array([[0.40824829],
+           [0.40824829],
+           [0.40824829],
+           [0.40824829],
+           [0.40824829],
+           [0.40824829]])
+    >>> lap.eigenvectors(2)
+    array([[ 0.5       ,  0.40824829],
+           [ 0.        ,  0.40824829],
+           [-0.5       ,  0.40824829],
+           [-0.5       ,  0.40824829],
+           [ 0.        ,  0.40824829],
+           [ 0.5       ,  0.40824829]])
+    >>> lap.eigenvectors()
+    array([[ 0.40824829,  0.28867513,  0.28867513,  0.5       ,  0.5       ,
+             0.40824829],
+           [-0.40824829, -0.57735027, -0.57735027,  0.        ,  0.        ,
+             0.40824829],
+           [ 0.40824829,  0.28867513,  0.28867513, -0.5       , -0.5       ,
+             0.40824829],
+           [-0.40824829,  0.28867513,  0.28867513, -0.5       , -0.5       ,
+             0.40824829],
+           [ 0.40824829, -0.57735027, -0.57735027,  0.        ,  0.        ,
+             0.40824829],
+           [-0.40824829,  0.28867513,  0.28867513,  0.5       ,  0.5       ,
+             0.40824829]])
+
+    The two-dimensional Laplacian is illustrated on a regular grid with
+    ``grid_shape = (2, 3)`` points in each dimension.
+
+    >>> grid_shape = (2, 3)
+    >>> n = np.prod(grid_shape)
+
+    Numeration of grid points is as follows:
+
+    >>> np.arange(n).reshape(grid_shape + (-1,))
+    array([[[0],
+            [1],
+            [2]],
+    <BLANKLINE>
+           [[3],
+            [4],
+            [5]]])
+
+    Each of the boundary conditions ``'dirichlet'``, ``'periodic'``, and
+    ``'neumann'`` is illustrated separately; with ``'dirichlet'``
+
+    >>> lap = LaplacianNd(grid_shape, boundary_conditions='dirichlet')
+    >>> lap.tosparse()
+    <Compressed Sparse Row sparse array of dtype 'int8'
+        with 20 stored elements and shape (6, 6)>
+    >>> lap.toarray()
+    array([[-4,  1,  0,  1,  0,  0],
+           [ 1, -4,  1,  0,  1,  0],
+           [ 0,  1, -4,  0,  0,  1],
+           [ 1,  0,  0, -4,  1,  0],
+           [ 0,  1,  0,  1, -4,  1],
+           [ 0,  0,  1,  0,  1, -4]], dtype=int8)
+    >>> np.array_equal(lap.matmat(np.eye(n)), lap.toarray())
+    True
+    >>> np.array_equal(lap.tosparse().toarray(), lap.toarray())
+    True
+    >>> lap.eigenvalues()
+    array([-6.41421356, -5.        , -4.41421356, -3.58578644, -3.        ,
+           -1.58578644])
+    >>> eigvals = eigvalsh(lap.toarray().astype(np.float64))
+    >>> np.allclose(lap.eigenvalues(), eigvals)
+    True
+    >>> np.allclose(lap.toarray() @ lap.eigenvectors(),
+    ...             lap.eigenvectors() @ np.diag(lap.eigenvalues()))
+    True
+
+    with ``'periodic'``
+
+    >>> lap = LaplacianNd(grid_shape, boundary_conditions='periodic')
+    >>> lap.tosparse()
+    <Compressed Sparse Row sparse array of dtype 'int8'
+        with 24 stored elements and shape (6, 6)>
+    >>> lap.toarray()
+        array([[-4,  1,  1,  2,  0,  0],
+               [ 1, -4,  1,  0,  2,  0],
+               [ 1,  1, -4,  0,  0,  2],
+               [ 2,  0,  0, -4,  1,  1],
+               [ 0,  2,  0,  1, -4,  1],
+               [ 0,  0,  2,  1,  1, -4]], dtype=int8)
+    >>> np.array_equal(lap.matmat(np.eye(n)), lap.toarray())
+    True
+    >>> np.array_equal(lap.tosparse().toarray(), lap.toarray())
+    True
+    >>> lap.eigenvalues()
+    array([-7., -7., -4., -3., -3.,  0.])
+    >>> eigvals = eigvalsh(lap.toarray().astype(np.float64))
+    >>> np.allclose(lap.eigenvalues(), eigvals)
+    True
+    >>> np.allclose(lap.toarray() @ lap.eigenvectors(),
+    ...             lap.eigenvectors() @ np.diag(lap.eigenvalues()))
+    True
+
+    and with ``'neumann'``
+
+    >>> lap = LaplacianNd(grid_shape, boundary_conditions='neumann')
+    >>> lap.tosparse()
+    <Compressed Sparse Row sparse array of dtype 'int8'
+        with 20 stored elements and shape (6, 6)>
+    >>> lap.toarray()
+    array([[-2,  1,  0,  1,  0,  0],
+           [ 1, -3,  1,  0,  1,  0],
+           [ 0,  1, -2,  0,  0,  1],
+           [ 1,  0,  0, -2,  1,  0],
+           [ 0,  1,  0,  1, -3,  1],
+           [ 0,  0,  1,  0,  1, -2]], dtype=int8)
+    >>> np.array_equal(lap.matmat(np.eye(n)), lap.toarray())
+    True
+    >>> np.array_equal(lap.tosparse().toarray(), lap.toarray())
+    True
+    >>> lap.eigenvalues()
+    array([-5., -3., -3., -2., -1.,  0.])
+    >>> eigvals = eigvalsh(lap.toarray().astype(np.float64))
+    >>> np.allclose(lap.eigenvalues(), eigvals)
+    True
+    >>> np.allclose(lap.toarray() @ lap.eigenvectors(),
+    ...             lap.eigenvectors() @ np.diag(lap.eigenvalues()))
+    True
+
+    """
+
+    def __init__(self, grid_shape, *,
+                 boundary_conditions='neumann',
+                 dtype=np.int8):
+
+        if boundary_conditions not in ('dirichlet', 'neumann', 'periodic'):
+            raise ValueError(
+                f"Unknown value {boundary_conditions!r} is given for "
+                "'boundary_conditions' parameter. The valid options are "
+                "'dirichlet', 'periodic', and 'neumann' (default)."
+            )
+
+        self.grid_shape = grid_shape
+        self.boundary_conditions = boundary_conditions
+        # LaplacianNd folds all dimensions in `grid_shape` into a single one
+        N = np.prod(grid_shape)
+        super().__init__(dtype=dtype, shape=(N, N))
+
+    def _eigenvalue_ordering(self, m):
+        """Compute `m` largest eigenvalues in each of the ``N`` directions,
+        i.e., up to ``m * N`` total, order them and return `m` largest.
+        """
+        grid_shape = self.grid_shape
+        if m is None:
+            indices = np.indices(grid_shape)
+            Leig = np.zeros(grid_shape)
+        else:
+            grid_shape_min = min(grid_shape,
+                                 tuple(np.ones_like(grid_shape) * m))
+            indices = np.indices(grid_shape_min)
+            Leig = np.zeros(grid_shape_min)
+
+        for j, n in zip(indices, grid_shape):
+            if self.boundary_conditions == 'dirichlet':
+                Leig += -4 * np.sin(np.pi * (j + 1) / (2 * (n + 1))) ** 2
+            elif self.boundary_conditions == 'neumann':
+                Leig += -4 * np.sin(np.pi * j / (2 * n)) ** 2
+            else:  # boundary_conditions == 'periodic'
+                Leig += -4 * np.sin(np.pi * np.floor((j + 1) / 2) / n) ** 2
+
+        Leig_ravel = Leig.ravel()
+        ind = np.argsort(Leig_ravel)
+        eigenvalues = Leig_ravel[ind]
+        if m is not None:
+            eigenvalues = eigenvalues[-m:]
+            ind = ind[-m:]
+
+        return eigenvalues, ind
+
+    def eigenvalues(self, m=None):
+        """Return the requested number of eigenvalues.
+
+        Parameters
+        ----------
+        m : int, optional
+            The positive number of smallest eigenvalues to return.
+            If not provided, then all eigenvalues will be returned.
+
+        Returns
+        -------
+        eigenvalues : float array
+            The requested `m` smallest or all eigenvalues, in ascending order.
+        """
+        eigenvalues, _ = self._eigenvalue_ordering(m)
+        return eigenvalues
+
+    def _ev1d(self, j, n):
+        """Return 1 eigenvector in 1d with index `j`
+        and number of grid points `n` where ``j < n``.
+        """
+        if self.boundary_conditions == 'dirichlet':
+            i = np.pi * (np.arange(n) + 1) / (n + 1)
+            ev = np.sqrt(2. / (n + 1.)) * np.sin(i * (j + 1))
+        elif self.boundary_conditions == 'neumann':
+            i = np.pi * (np.arange(n) + 0.5) / n
+            ev = np.sqrt((1. if j == 0 else 2.) / n) * np.cos(i * j)
+        else:  # boundary_conditions == 'periodic'
+            if j == 0:
+                ev = np.sqrt(1. / n) * np.ones(n)
+            elif j + 1 == n and n % 2 == 0:
+                ev = np.sqrt(1. / n) * np.tile([1, -1], n//2)
+            else:
+                i = 2. * np.pi * (np.arange(n) + 0.5) / n
+                ev = np.sqrt(2. / n) * np.cos(i * np.floor((j + 1) / 2))
+        # make small values exact zeros correcting round-off errors
+        # due to symmetry of eigenvectors the exact 0. is correct
+        ev[np.abs(ev) < np.finfo(np.float64).eps] = 0.
+        return ev
+
+    def _one_eve(self, k):
+        """Return 1 eigenvector in Nd with multi-index `j`
+        as a tensor product of the corresponding 1d eigenvectors.
+        """
+        phi = [self._ev1d(j, n) for j, n in zip(k, self.grid_shape)]
+        result = phi[0]
+        for phi in phi[1:]:
+            result = np.tensordot(result, phi, axes=0)
+        return np.asarray(result).ravel()
+
+    def eigenvectors(self, m=None):
+        """Return the requested number of eigenvectors for ordered eigenvalues.
+
+        Parameters
+        ----------
+        m : int, optional
+            The positive number of eigenvectors to return. If not provided,
+            then all eigenvectors will be returned.
+
+        Returns
+        -------
+        eigenvectors : float array
+            An array with columns made of the requested `m` or all eigenvectors.
+            The columns are ordered according to the `m` ordered eigenvalues.
+        """
+        _, ind = self._eigenvalue_ordering(m)
+        if m is None:
+            grid_shape_min = self.grid_shape
+        else:
+            grid_shape_min = min(self.grid_shape,
+                                tuple(np.ones_like(self.grid_shape) * m))
+
+        N_indices = np.unravel_index(ind, grid_shape_min)
+        N_indices = [tuple(x) for x in zip(*N_indices)]
+        eigenvectors_list = [self._one_eve(k) for k in N_indices]
+        return np.column_stack(eigenvectors_list)
+
+    def toarray(self):
+        """
+        Converts the Laplacian data to a dense array.
+
+        Returns
+        -------
+        L : ndarray
+            The shape is ``(N, N)`` where ``N = np.prod(grid_shape)``.
+
+        """
+        grid_shape = self.grid_shape
+        n = np.prod(grid_shape)
+        L = np.zeros([n, n], dtype=np.int8)
+        # Scratch arrays
+        L_i = np.empty_like(L)
+        Ltemp = np.empty_like(L)
+
+        for ind, dim in enumerate(grid_shape):
+            # Start zeroing out L_i
+            L_i[:] = 0
+            # Allocate the top left corner with the kernel of L_i
+            # Einsum returns writable view of arrays
+            np.einsum("ii->i", L_i[:dim, :dim])[:] = -2
+            np.einsum("ii->i", L_i[: dim - 1, 1:dim])[:] = 1
+            np.einsum("ii->i", L_i[1:dim, : dim - 1])[:] = 1
+
+            if self.boundary_conditions == 'neumann':
+                L_i[0, 0] = -1
+                L_i[dim - 1, dim - 1] = -1
+            elif self.boundary_conditions == 'periodic':
+                if dim > 1:
+                    L_i[0, dim - 1] += 1
+                    L_i[dim - 1, 0] += 1
+                else:
+                    L_i[0, 0] += 1
+
+            # kron is too slow for large matrices hence the next two tricks
+            # 1- kron(eye, mat) is block_diag(mat, mat, ...)
+            # 2- kron(mat, eye) can be performed by 4d stride trick
+
+            # 1-
+            new_dim = dim
+            # for block_diag we tile the top left portion on the diagonal
+            if ind > 0:
+                tiles = np.prod(grid_shape[:ind])
+                for j in range(1, tiles):
+                    L_i[j*dim:(j+1)*dim, j*dim:(j+1)*dim] = L_i[:dim, :dim]
+                    new_dim += dim
+            # 2-
+            # we need the keep L_i, but reset the array
+            Ltemp[:new_dim, :new_dim] = L_i[:new_dim, :new_dim]
+            tiles = int(np.prod(grid_shape[ind+1:]))
+            # Zero out the top left, the rest is already 0
+            L_i[:new_dim, :new_dim] = 0
+            idx = [x for x in range(tiles)]
+            L_i.reshape(
+                (new_dim, tiles,
+                 new_dim, tiles)
+                )[:, idx, :, idx] = Ltemp[:new_dim, :new_dim]
+
+            L += L_i
+
+        return L.astype(self.dtype)
+
+    def tosparse(self):
+        """
+        Constructs a sparse array from the Laplacian data. The returned sparse
+        array format is dependent on the selected boundary conditions.
+
+        Returns
+        -------
+        L : scipy.sparse.sparray
+            The shape is ``(N, N)`` where ``N = np.prod(grid_shape)``.
+
+        """
+        N = len(self.grid_shape)
+        p = np.prod(self.grid_shape)
+        L = dia_array((p, p), dtype=np.int8)
+
+        for i in range(N):
+            dim = self.grid_shape[i]
+            data = np.ones([3, dim], dtype=np.int8)
+            data[1, :] *= -2
+
+            if self.boundary_conditions == 'neumann':
+                data[1, 0] = -1
+                data[1, -1] = -1
+
+            L_i = dia_array((data, [-1, 0, 1]), shape=(dim, dim),
+                            dtype=np.int8
+                            )
+
+            if self.boundary_conditions == 'periodic':
+                t = dia_array((dim, dim), dtype=np.int8)
+                t.setdiag([1], k=-dim+1)
+                t.setdiag([1], k=dim-1)
+                L_i += t
+
+            for j in range(i):
+                L_i = kron(eye_array(self.grid_shape[j], dtype=np.int8), L_i)
+            for j in range(i + 1, N):
+                L_i = kron(L_i, eye_array(self.grid_shape[j], dtype=np.int8))
+            L += L_i
+        return L.astype(self.dtype)
+
+    def _matvec(self, x):
+        grid_shape = self.grid_shape
+        N = len(grid_shape)
+        X = x.reshape(grid_shape + (-1,))
+        Y = -2 * N * X
+        for i in range(N):
+            Y += np.roll(X, 1, axis=i)
+            Y += np.roll(X, -1, axis=i)
+            if self.boundary_conditions in ('neumann', 'dirichlet'):
+                Y[(slice(None),)*i + (0,) + (slice(None),)*(N-i-1)
+                  ] -= np.roll(X, 1, axis=i)[
+                    (slice(None),) * i + (0,) + (slice(None),) * (N-i-1)
+                ]
+                Y[
+                    (slice(None),) * i + (-1,) + (slice(None),) * (N-i-1)
+                ] -= np.roll(X, -1, axis=i)[
+                    (slice(None),) * i + (-1,) + (slice(None),) * (N-i-1)
+                ]
+
+                if self.boundary_conditions == 'neumann':
+                    Y[
+                        (slice(None),) * i + (0,) + (slice(None),) * (N-i-1)
+                    ] += np.roll(X, 0, axis=i)[
+                        (slice(None),) * i + (0,) + (slice(None),) * (N-i-1)
+                    ]
+                    Y[
+                        (slice(None),) * i + (-1,) + (slice(None),) * (N-i-1)
+                    ] += np.roll(X, 0, axis=i)[
+                        (slice(None),) * i + (-1,) + (slice(None),) * (N-i-1)
+                    ]
+
+        return Y.reshape(-1, X.shape[-1])
+
+    def _matmat(self, x):
+        return self._matvec(x)
+
+    def _adjoint(self):
+        return self
+
+    def _transpose(self):
+        return self
+
+
+class Sakurai(LinearOperator):
+    """
+    Construct a Sakurai matrix in various formats and its eigenvalues.
+
+    Constructs the "Sakurai" matrix motivated by reference [1]_:
+    square real symmetric positive definite and 5-diagonal
+    with the main diagonal ``[5, 6, 6, ..., 6, 6, 5], the ``+1`` and ``-1``
+    diagonals filled with ``-4``, and the ``+2`` and ``-2`` diagonals
+    made of ``1``. Its eigenvalues are analytically known to be
+    ``16. * np.power(np.cos(0.5 * k * np.pi / (n + 1)), 4)``.
+    The matrix gets ill-conditioned with its size growing.
+    It is useful for testing and benchmarking sparse eigenvalue solvers
+    especially those taking advantage of its banded 5-diagonal structure.
+    See the notes below for details.
+
+    Parameters
+    ----------
+    n : int
+        The size of the matrix.
+    dtype : dtype
+        Numerical type of the array. Default is ``np.int8``.
+
+    Methods
+    -------
+    toarray()
+        Construct a dense array from Laplacian data
+    tosparse()
+        Construct a sparse array from Laplacian data
+    tobanded()
+        The Sakurai matrix in the format for banded symmetric matrices,
+        i.e., (3, n) ndarray with 3 upper diagonals
+        placing the main diagonal at the bottom.
+    eigenvalues
+        All eigenvalues of the Sakurai matrix ordered ascending.
+
+    Notes
+    -----
+    Reference [1]_ introduces a generalized eigenproblem for the matrix pair
+    `A` and `B` where `A` is the identity so we turn it into an eigenproblem
+    just for the matrix `B` that this function outputs in various formats
+    together with its eigenvalues.
+
+    .. versionadded:: 1.12.0
+
+    References
+    ----------
+    .. [1] T. Sakurai, H. Tadano, Y. Inadomi, and U. Nagashima,
+       "A moment-based method for large-scale generalized
+       eigenvalue problems",
+       Appl. Num. Anal. Comp. Math. Vol. 1 No. 2 (2004).
+
+    Examples
+    --------
+    >>> import numpy as np
+    >>> from scipy.sparse.linalg._special_sparse_arrays import Sakurai
+    >>> from scipy.linalg import eig_banded
+    >>> n = 6
+    >>> sak = Sakurai(n)
+
+    Since all matrix entries are small integers, ``'int8'`` is
+    the default dtype for storing matrix representations.
+
+    >>> sak.toarray()
+    array([[ 5, -4,  1,  0,  0,  0],
+           [-4,  6, -4,  1,  0,  0],
+           [ 1, -4,  6, -4,  1,  0],
+           [ 0,  1, -4,  6, -4,  1],
+           [ 0,  0,  1, -4,  6, -4],
+           [ 0,  0,  0,  1, -4,  5]], dtype=int8)
+    >>> sak.tobanded()
+    array([[ 1,  1,  1,  1,  1,  1],
+           [-4, -4, -4, -4, -4, -4],
+           [ 5,  6,  6,  6,  6,  5]], dtype=int8)
+    >>> sak.tosparse()
+    <DIAgonal sparse array of dtype 'int8'
+        with 24 stored elements (5 diagonals) and shape (6, 6)>
+    >>> np.array_equal(sak.dot(np.eye(n)), sak.tosparse().toarray())
+    True
+    >>> sak.eigenvalues()
+    array([0.03922866, 0.56703972, 2.41789479, 5.97822974,
+           10.54287655, 14.45473055])
+    >>> sak.eigenvalues(2)
+    array([0.03922866, 0.56703972])
+
+    The banded form can be used in scipy functions for banded matrices, e.g.,
+
+    >>> e = eig_banded(sak.tobanded(), eigvals_only=True)
+    >>> np.allclose(sak.eigenvalues(), e, atol= n * n * n * np.finfo(float).eps)
+    True
+
+    """
+    def __init__(self, n, dtype=np.int8):
+        self.n = n
+        self.dtype = dtype
+        shape = (n, n)
+        super().__init__(dtype, shape)
+
+    def eigenvalues(self, m=None):
+        """Return the requested number of eigenvalues.
+
+        Parameters
+        ----------
+        m : int, optional
+            The positive number of smallest eigenvalues to return.
+            If not provided, then all eigenvalues will be returned.
+
+        Returns
+        -------
+        eigenvalues : `np.float64` array
+            The requested `m` smallest or all eigenvalues, in ascending order.
+        """
+        if m is None:
+            m = self.n
+        k = np.arange(self.n + 1 -m, self.n + 1)
+        return np.flip(16. * np.power(np.cos(0.5 * k * np.pi / (self.n + 1)), 4))
+
+    def tobanded(self):
+        """
+        Construct the Sakurai matrix as a banded array.
+        """
+        d0 = np.r_[5, 6 * np.ones(self.n - 2, dtype=self.dtype), 5]
+        d1 = -4 * np.ones(self.n, dtype=self.dtype)
+        d2 = np.ones(self.n, dtype=self.dtype)
+        return np.array([d2, d1, d0]).astype(self.dtype)
+
+    def tosparse(self):
+        """
+        Construct the Sakurai matrix in a sparse format.
+        """
+        from scipy.sparse import diags_array
+        d = self.tobanded()
+        # the banded format has the main diagonal at the bottom
+        # `diags_array` inherits dtype from banded
+        return diags_array([d[0], d[1], d[2], d[1], d[0]], offsets=[-2, -1, 0, 1, 2],
+                           shape=(self.n, self.n), dtype=d.dtype)
+
+    def toarray(self):
+        return self.tosparse().toarray()
+
+    def _matvec(self, x):
+        """
+        Construct matrix-free callable banded-matrix-vector multiplication by
+        the Sakurai matrix without constructing or storing the matrix itself
+        using the knowledge of its entries and the 5-diagonal format.
+        """
+        x = x.reshape(self.n, -1)
+        result_dtype = np.promote_types(x.dtype, self.dtype)
+        sx = np.zeros_like(x, dtype=result_dtype)
+        sx[0, :] = 5 * x[0, :] - 4 * x[1, :] + x[2, :]
+        sx[-1, :] = 5 * x[-1, :] - 4 * x[-2, :] + x[-3, :]
+        sx[1: -1, :] = (6 * x[1: -1, :] - 4 * (x[:-2, :] + x[2:, :])
+                      + np.pad(x[:-3, :], ((1, 0), (0, 0)))
+                      + np.pad(x[3:, :], ((0, 1), (0, 0))))
+        return sx
+
+    def _matmat(self, x):
+        """
+        Construct matrix-free callable matrix-matrix multiplication by
+        the Sakurai matrix without constructing or storing the matrix itself
+        by reusing the ``_matvec(x)`` that supports both 1D and 2D arrays ``x``.
+        """
+        return self._matvec(x)
+
+    def _adjoint(self):
+        return self
+
+    def _transpose(self):
+        return self
+
+
+class MikotaM(LinearOperator):
+    """
+    Construct a mass matrix in various formats of Mikota pair.
+
+    The mass matrix `M` is square real diagonal
+    positive definite with entries that are reciprocal to integers.
+
+    Parameters
+    ----------
+    shape : tuple of int
+        The shape of the matrix.
+    dtype : dtype
+        Numerical type of the array. Default is ``np.float64``.
+
+    Methods
+    -------
+    toarray()
+        Construct a dense array from Mikota data
+    tosparse()
+        Construct a sparse array from Mikota data
+    tobanded()
+        The format for banded symmetric matrices,
+        i.e., (1, n) ndarray with the main diagonal.
+    """
+    def __init__(self, shape, dtype=np.float64):
+        self.shape = shape
+        self.dtype = dtype
+        super().__init__(dtype, shape)
+
+    def _diag(self):
+        # The matrix is constructed from its diagonal 1 / [1, ..., N+1];
+        # compute in a function to avoid duplicated code & storage footprint
+        return (1. / np.arange(1, self.shape[0] + 1)).astype(self.dtype)
+
+    def tobanded(self):
+        return self._diag()
+
+    def tosparse(self):
+        from scipy.sparse import diags_array
+        return diags_array([self._diag()], offsets=[0],
+                           shape=self.shape, dtype=self.dtype)
+
+    def toarray(self):
+        return np.diag(self._diag()).astype(self.dtype)
+
+    def _matvec(self, x):
+        """
+        Construct matrix-free callable banded-matrix-vector multiplication by
+        the Mikota mass matrix without constructing or storing the matrix itself
+        using the knowledge of its entries and the diagonal format.
+        """
+        x = x.reshape(self.shape[0], -1)
+        return self._diag()[:, np.newaxis] * x
+
+    def _matmat(self, x):
+        """
+        Construct matrix-free callable matrix-matrix multiplication by
+        the Mikota mass matrix without constructing or storing the matrix itself
+        by reusing the ``_matvec(x)`` that supports both 1D and 2D arrays ``x``.
+        """
+        return self._matvec(x)
+
+    def _adjoint(self):
+        return self
+
+    def _transpose(self):
+        return self
+
+
+class MikotaK(LinearOperator):
+    """
+    Construct a stiffness matrix in various formats of Mikota pair.
+
+    The stiffness matrix `K` is square real tri-diagonal symmetric
+    positive definite with integer entries.
+
+    Parameters
+    ----------
+    shape : tuple of int
+        The shape of the matrix.
+    dtype : dtype
+        Numerical type of the array. Default is ``np.int32``.
+
+    Methods
+    -------
+    toarray()
+        Construct a dense array from Mikota data
+    tosparse()
+        Construct a sparse array from Mikota data
+    tobanded()
+        The format for banded symmetric matrices,
+        i.e., (2, n) ndarray with 2 upper diagonals
+        placing the main diagonal at the bottom.
+    """
+    def __init__(self, shape, dtype=np.int32):
+        self.shape = shape
+        self.dtype = dtype
+        super().__init__(dtype, shape)
+        # The matrix is constructed from its diagonals;
+        # we precompute these to avoid duplicating the computation
+        n = shape[0]
+        self._diag0 = np.arange(2 * n - 1, 0, -2, dtype=self.dtype)
+        self._diag1 = - np.arange(n - 1, 0, -1, dtype=self.dtype)
+
+    def tobanded(self):
+        return np.array([np.pad(self._diag1, (1, 0), 'constant'), self._diag0])
+
+    def tosparse(self):
+        from scipy.sparse import diags_array
+        return diags_array([self._diag1, self._diag0, self._diag1], offsets=[-1, 0, 1],
+                           shape=self.shape, dtype=self.dtype)
+
+    def toarray(self):
+        return self.tosparse().toarray()
+
+    def _matvec(self, x):
+        """
+        Construct matrix-free callable banded-matrix-vector multiplication by
+        the Mikota stiffness matrix without constructing or storing the matrix
+        itself using the knowledge of its entries and the 3-diagonal format.
+        """
+        x = x.reshape(self.shape[0], -1)
+        result_dtype = np.promote_types(x.dtype, self.dtype)
+        kx = np.zeros_like(x, dtype=result_dtype)
+        d1 = self._diag1
+        d0 = self._diag0
+        kx[0, :] = d0[0] * x[0, :] + d1[0] * x[1, :]
+        kx[-1, :] = d1[-1] * x[-2, :] + d0[-1] * x[-1, :]
+        kx[1: -1, :] = (d1[:-1, None] * x[: -2, :]
+                        + d0[1: -1, None] * x[1: -1, :]
+                        + d1[1:, None] * x[2:, :])
+        return kx
+
+    def _matmat(self, x):
+        """
+        Construct matrix-free callable matrix-matrix multiplication by
+        the Stiffness mass matrix without constructing or storing the matrix itself
+        by reusing the ``_matvec(x)`` that supports both 1D and 2D arrays ``x``.
+        """
+        return self._matvec(x)
+
+    def _adjoint(self):
+        return self
+
+    def _transpose(self):
+        return self
+
+
+class MikotaPair:
+    """
+    Construct the Mikota pair of matrices in various formats and
+    eigenvalues of the generalized eigenproblem with them.
+
+    The Mikota pair of matrices [1, 2]_ models a vibration problem
+    of a linear mass-spring system with the ends attached where
+    the stiffness of the springs and the masses increase along
+    the system length such that vibration frequencies are subsequent
+    integers 1, 2, ..., `n` where `n` is the number of the masses. Thus,
+    eigenvalues of the generalized eigenvalue problem for
+    the matrix pair `K` and `M` where `K` is the system stiffness matrix
+    and `M` is the system mass matrix are the squares of the integers,
+    i.e., 1, 4, 9, ..., ``n * n``.
+
+    The stiffness matrix `K` is square real tri-diagonal symmetric
+    positive definite. The mass matrix `M` is diagonal with diagonal
+    entries 1, 1/2, 1/3, ...., ``1/n``. Both matrices get
+    ill-conditioned with `n` growing.
+
+    Parameters
+    ----------
+    n : int
+        The size of the matrices of the Mikota pair.
+    dtype : dtype
+        Numerical type of the array. Default is ``np.float64``.
+
+    Attributes
+    ----------
+    eigenvalues : 1D ndarray, ``np.uint64``
+        All eigenvalues of the Mikota pair ordered ascending.
+
+    Methods
+    -------
+    MikotaK()
+        A `LinearOperator` custom object for the stiffness matrix.
+    MikotaM()
+        A `LinearOperator` custom object for the mass matrix.
+
+    .. versionadded:: 1.12.0
+
+    References
+    ----------
+    .. [1] J. Mikota, "Frequency tuning of chain structure multibody oscillators
+       to place the natural frequencies at omega1 and N-1 integer multiples
+       omega2,..., omegaN", Z. Angew. Math. Mech. 81 (2001), S2, S201-S202.
+       Appl. Num. Anal. Comp. Math. Vol. 1 No. 2 (2004).
+    .. [2] Peter C. Muller and Metin Gurgoze,
+       "Natural frequencies of a multi-degree-of-freedom vibration system",
+       Proc. Appl. Math. Mech. 6, 319-320 (2006).
+       http://dx.doi.org/10.1002/pamm.200610141.
+
+    Examples
+    --------
+    >>> import numpy as np
+    >>> from scipy.sparse.linalg._special_sparse_arrays import MikotaPair
+    >>> n = 6
+    >>> mik = MikotaPair(n)
+    >>> mik_k = mik.k
+    >>> mik_m = mik.m
+    >>> mik_k.toarray()
+    array([[11., -5.,  0.,  0.,  0.,  0.],
+           [-5.,  9., -4.,  0.,  0.,  0.],
+           [ 0., -4.,  7., -3.,  0.,  0.],
+           [ 0.,  0., -3.,  5., -2.,  0.],
+           [ 0.,  0.,  0., -2.,  3., -1.],
+           [ 0.,  0.,  0.,  0., -1.,  1.]])
+    >>> mik_k.tobanded()
+    array([[ 0., -5., -4., -3., -2., -1.],
+           [11.,  9.,  7.,  5.,  3.,  1.]])
+    >>> mik_m.tobanded()
+    array([1.        , 0.5       , 0.33333333, 0.25      , 0.2       ,
+        0.16666667])
+    >>> mik_k.tosparse()
+    <DIAgonal sparse array of dtype 'float64'
+        with 16 stored elements (3 diagonals) and shape (6, 6)>
+    >>> mik_m.tosparse()
+    <DIAgonal sparse array of dtype 'float64'
+        with 6 stored elements (1 diagonals) and shape (6, 6)>
+    >>> np.array_equal(mik_k(np.eye(n)), mik_k.toarray())
+    True
+    >>> np.array_equal(mik_m(np.eye(n)), mik_m.toarray())
+    True
+    >>> mik.eigenvalues()
+    array([ 1,  4,  9, 16, 25, 36])
+    >>> mik.eigenvalues(2)
+    array([ 1,  4])
+
+    """
+    def __init__(self, n, dtype=np.float64):
+        self.n = n
+        self.dtype = dtype
+        self.shape = (n, n)
+        self.m = MikotaM(self.shape, self.dtype)
+        self.k = MikotaK(self.shape, self.dtype)
+
+    def eigenvalues(self, m=None):
+        """Return the requested number of eigenvalues.
+
+        Parameters
+        ----------
+        m : int, optional
+            The positive number of smallest eigenvalues to return.
+            If not provided, then all eigenvalues will be returned.
+
+        Returns
+        -------
+        eigenvalues : `np.uint64` array
+            The requested `m` smallest or all eigenvalues, in ascending order.
+        """
+        if m is None:
+            m = self.n
+        arange_plus1 = np.arange(1, m + 1, dtype=np.uint64)
+        return arange_plus1 * arange_plus1