symjax.tensor.linalg.eigh

symjax.tensor.linalg.eigh(a, b=None, lower=True, eigvals_only=False, overwrite_a=False, overwrite_b=False, turbo=True, eigvals=None, type=1, check_finite=True)[source]
Solve a standard or generalized eigenvalue problem for a complex
Hermitian or real symmetric matrix.

LAX-backend implementation of eigh(). Original docstring below.

Find eigenvalues array w and optionally eigenvectors array v of array a, where b is positive definite such that for every eigenvalue λ (i-th entry of w) and its eigenvector vi (i-th column of v) satisfies:

              a @ vi = λ * b @ vi
vi.conj().T @ a @ vi = λ
vi.conj().T @ b @ vi = 1

In the standard problem, b is assumed to be the identity matrix.

Parameters:
  • a ((M, M) array_like) – A complex Hermitian or real symmetric matrix whose eigenvalues and eigenvectors will be computed.
  • b ((M, M) array_like, optional) – A complex Hermitian or real symmetric definite positive matrix in. If omitted, identity matrix is assumed.
  • lower (bool, optional) – Whether the pertinent array data is taken from the lower or upper triangle of a and, if applicable, b. (Default: lower)
  • eigvals_only (bool, optional) – Whether to calculate only eigenvalues and no eigenvectors. (Default: both are calculated)
  • type (int, optional) – For the generalized problems, this keyword specifies the problem type to be solved for w and v (only takes 1, 2, 3 as possible inputs):
  • overwrite_a (bool, optional) – Whether to overwrite data in a (may improve performance). Default is False.
  • overwrite_b (bool, optional) – Whether to overwrite data in b (may improve performance). Default is False.
  • check_finite (bool, optional) – Whether to check that the input matrices contain only finite numbers. Disabling may give a performance gain, but may result in problems (crashes, non-termination) if the inputs do contain infinities or NaNs.
  • turbo (bool, optional) – Deprecated since v1.5.0, use ``driver=gvd`` keyword instead. Use divide and conquer algorithm (faster but expensive in memory, only for generalized eigenvalue problem and if full set of eigenvalues are requested.). Has no significant effect if eigenvectors are not requested.
  • eigvals (tuple (lo, hi), optional) – Deprecated since v1.5.0, use ``subset_by_index`` keyword instead. Indexes of the smallest and largest (in ascending order) eigenvalues and corresponding eigenvectors to be returned: 0 <= lo <= hi <= M-1. If omitted, all eigenvalues and eigenvectors are returned.
Returns:

  • w ((N,) ndarray) – The N (1<=N<=M) selected eigenvalues, in ascending order, each repeated according to its multiplicity.
  • v ((M, N) ndarray) – (if eigvals_only == False)
Raises:

LinAlgError – If eigenvalue computation does not converge, an error occurred, or b matrix is not definite positive. Note that if input matrices are not symmetric or Hermitian, no error will be reported but results will be wrong.

See also

eigvalsh()
eigenvalues of symmetric or Hermitian arrays
eig()
eigenvalues and right eigenvectors for non-symmetric arrays
eigh_tridiagonal()
eigenvalues and right eiegenvectors for symmetric/Hermitian tridiagonal matrices

Notes

This function does not check the input array for being hermitian/symmetric in order to allow for representing arrays with only their upper/lower triangular parts. Also, note that even though not taken into account, finiteness check applies to the whole array and unaffected by “lower” keyword.

This function uses LAPACK drivers for computations in all possible keyword combinations, prefixed with sy if arrays are real and he if complex, e.g., a float array with “evr” driver is solved via “syevr”, complex arrays with “gvx” driver problem is solved via “hegvx” etc.

As a brief summary, the slowest and the most robust driver is the classical <sy/he>ev which uses symmetric QR. <sy/he>evr is seen as the optimal choice for the most general cases. However, there are certain occassions that <sy/he>evd computes faster at the expense of more memory usage. <sy/he>evx, while still being faster than <sy/he>ev, often performs worse than the rest except when very few eigenvalues are requested for large arrays though there is still no performance guarantee.

For the generalized problem, normalization with respoect to the given type argument:

type 1 and 3 :      v.conj().T @ a @ v = w
type 2       : inv(v).conj().T @ a @ inv(v) = w

type 1 or 2  :      v.conj().T @ b @ v  = I
type 3       : v.conj().T @ inv(b) @ v  = I

Examples

>>> from scipy.linalg import eigh
>>> A = np.array([[6, 3, 1, 5], [3, 0, 5, 1], [1, 5, 6, 2], [5, 1, 2, 2]])
>>> w, v = eigh(A)
>>> np.allclose(A @ v - v @ np.diag(w), np.zeros((4, 4)))
True

Request only the eigenvalues

>>> w = eigh(A, eigvals_only=True)

Request eigenvalues that are less than 10.

>>> A = np.array([[34, -4, -10, -7, 2],
...               [-4, 7, 2, 12, 0],
...               [-10, 2, 44, 2, -19],
...               [-7, 12, 2, 79, -34],
...               [2, 0, -19, -34, 29]])
>>> eigh(A, eigvals_only=True, subset_by_value=[-np.inf, 10])
array([6.69199443e-07, 9.11938152e+00])

Request the largest second eigenvalue and its eigenvector

>>> w, v = eigh(A, subset_by_index=[1, 1])
>>> w
array([9.11938152])
>>> v.shape  # only a single column is returned
(5, 1)