476 lines
11 KiB
Python
476 lines
11 KiB
Python
from sympy.core.numbers import mod_inverse
|
|
|
|
from .common import MatrixError, NonSquareMatrixError, NonInvertibleMatrixError
|
|
from .utilities import _iszero
|
|
|
|
|
|
def _pinv_full_rank(M):
|
|
"""Subroutine for full row or column rank matrices.
|
|
|
|
For full row rank matrices, inverse of ``A * A.H`` Exists.
|
|
For full column rank matrices, inverse of ``A.H * A`` Exists.
|
|
|
|
This routine can apply for both cases by checking the shape
|
|
and have small decision.
|
|
"""
|
|
|
|
if M.is_zero_matrix:
|
|
return M.H
|
|
|
|
if M.rows >= M.cols:
|
|
return M.H.multiply(M).inv().multiply(M.H)
|
|
else:
|
|
return M.H.multiply(M.multiply(M.H).inv())
|
|
|
|
def _pinv_rank_decomposition(M):
|
|
"""Subroutine for rank decomposition
|
|
|
|
With rank decompositions, `A` can be decomposed into two full-
|
|
rank matrices, and each matrix can take pseudoinverse
|
|
individually.
|
|
"""
|
|
|
|
if M.is_zero_matrix:
|
|
return M.H
|
|
|
|
B, C = M.rank_decomposition()
|
|
|
|
Bp = _pinv_full_rank(B)
|
|
Cp = _pinv_full_rank(C)
|
|
|
|
return Cp.multiply(Bp)
|
|
|
|
def _pinv_diagonalization(M):
|
|
"""Subroutine using diagonalization
|
|
|
|
This routine can sometimes fail if SymPy's eigenvalue
|
|
computation is not reliable.
|
|
"""
|
|
|
|
if M.is_zero_matrix:
|
|
return M.H
|
|
|
|
A = M
|
|
AH = M.H
|
|
|
|
try:
|
|
if M.rows >= M.cols:
|
|
P, D = AH.multiply(A).diagonalize(normalize=True)
|
|
D_pinv = D.applyfunc(lambda x: 0 if _iszero(x) else 1 / x)
|
|
|
|
return P.multiply(D_pinv).multiply(P.H).multiply(AH)
|
|
|
|
else:
|
|
P, D = A.multiply(AH).diagonalize(
|
|
normalize=True)
|
|
D_pinv = D.applyfunc(lambda x: 0 if _iszero(x) else 1 / x)
|
|
|
|
return AH.multiply(P).multiply(D_pinv).multiply(P.H)
|
|
|
|
except MatrixError:
|
|
raise NotImplementedError(
|
|
'pinv for rank-deficient matrices where '
|
|
'diagonalization of A.H*A fails is not supported yet.')
|
|
|
|
def _pinv(M, method='RD'):
|
|
"""Calculate the Moore-Penrose pseudoinverse of the matrix.
|
|
|
|
The Moore-Penrose pseudoinverse exists and is unique for any matrix.
|
|
If the matrix is invertible, the pseudoinverse is the same as the
|
|
inverse.
|
|
|
|
Parameters
|
|
==========
|
|
|
|
method : String, optional
|
|
Specifies the method for computing the pseudoinverse.
|
|
|
|
If ``'RD'``, Rank-Decomposition will be used.
|
|
|
|
If ``'ED'``, Diagonalization will be used.
|
|
|
|
Examples
|
|
========
|
|
|
|
Computing pseudoinverse by rank decomposition :
|
|
|
|
>>> from sympy import Matrix
|
|
>>> A = Matrix([[1, 2, 3], [4, 5, 6]])
|
|
>>> A.pinv()
|
|
Matrix([
|
|
[-17/18, 4/9],
|
|
[ -1/9, 1/9],
|
|
[ 13/18, -2/9]])
|
|
|
|
Computing pseudoinverse by diagonalization :
|
|
|
|
>>> B = A.pinv(method='ED')
|
|
>>> B.simplify()
|
|
>>> B
|
|
Matrix([
|
|
[-17/18, 4/9],
|
|
[ -1/9, 1/9],
|
|
[ 13/18, -2/9]])
|
|
|
|
See Also
|
|
========
|
|
|
|
inv
|
|
pinv_solve
|
|
|
|
References
|
|
==========
|
|
|
|
.. [1] https://en.wikipedia.org/wiki/Moore-Penrose_pseudoinverse
|
|
|
|
"""
|
|
|
|
# Trivial case: pseudoinverse of all-zero matrix is its transpose.
|
|
if M.is_zero_matrix:
|
|
return M.H
|
|
|
|
if method == 'RD':
|
|
return _pinv_rank_decomposition(M)
|
|
elif method == 'ED':
|
|
return _pinv_diagonalization(M)
|
|
else:
|
|
raise ValueError('invalid pinv method %s' % repr(method))
|
|
|
|
|
|
def _inv_mod(M, m):
|
|
r"""
|
|
Returns the inverse of the matrix `K` (mod `m`), if it exists.
|
|
|
|
Method to find the matrix inverse of `K` (mod `m`) implemented in this function:
|
|
|
|
* Compute `\mathrm{adj}(K) = \mathrm{cof}(K)^t`, the adjoint matrix of `K`.
|
|
|
|
* Compute `r = 1/\mathrm{det}(K) \pmod m`.
|
|
|
|
* `K^{-1} = r\cdot \mathrm{adj}(K) \pmod m`.
|
|
|
|
Examples
|
|
========
|
|
|
|
>>> from sympy import Matrix
|
|
>>> A = Matrix(2, 2, [1, 2, 3, 4])
|
|
>>> A.inv_mod(5)
|
|
Matrix([
|
|
[3, 1],
|
|
[4, 2]])
|
|
>>> A.inv_mod(3)
|
|
Matrix([
|
|
[1, 1],
|
|
[0, 1]])
|
|
|
|
"""
|
|
|
|
if not M.is_square:
|
|
raise NonSquareMatrixError()
|
|
|
|
N = M.cols
|
|
det_K = M.det()
|
|
det_inv = None
|
|
|
|
try:
|
|
det_inv = mod_inverse(det_K, m)
|
|
except ValueError:
|
|
raise NonInvertibleMatrixError('Matrix is not invertible (mod %d)' % m)
|
|
|
|
K_adj = M.adjugate()
|
|
K_inv = M.__class__(N, N,
|
|
[det_inv * K_adj[i, j] % m for i in range(N) for j in range(N)])
|
|
|
|
return K_inv
|
|
|
|
|
|
def _verify_invertible(M, iszerofunc=_iszero):
|
|
"""Initial check to see if a matrix is invertible. Raises or returns
|
|
determinant for use in _inv_ADJ."""
|
|
|
|
if not M.is_square:
|
|
raise NonSquareMatrixError("A Matrix must be square to invert.")
|
|
|
|
d = M.det(method='berkowitz')
|
|
zero = d.equals(0)
|
|
|
|
if zero is None: # if equals() can't decide, will rref be able to?
|
|
ok = M.rref(simplify=True)[0]
|
|
zero = any(iszerofunc(ok[j, j]) for j in range(ok.rows))
|
|
|
|
if zero:
|
|
raise NonInvertibleMatrixError("Matrix det == 0; not invertible.")
|
|
|
|
return d
|
|
|
|
def _inv_ADJ(M, iszerofunc=_iszero):
|
|
"""Calculates the inverse using the adjugate matrix and a determinant.
|
|
|
|
See Also
|
|
========
|
|
|
|
inv
|
|
inverse_GE
|
|
inverse_LU
|
|
inverse_CH
|
|
inverse_LDL
|
|
"""
|
|
|
|
d = _verify_invertible(M, iszerofunc=iszerofunc)
|
|
|
|
return M.adjugate() / d
|
|
|
|
def _inv_GE(M, iszerofunc=_iszero):
|
|
"""Calculates the inverse using Gaussian elimination.
|
|
|
|
See Also
|
|
========
|
|
|
|
inv
|
|
inverse_ADJ
|
|
inverse_LU
|
|
inverse_CH
|
|
inverse_LDL
|
|
"""
|
|
|
|
from .dense import Matrix
|
|
|
|
if not M.is_square:
|
|
raise NonSquareMatrixError("A Matrix must be square to invert.")
|
|
|
|
big = Matrix.hstack(M.as_mutable(), Matrix.eye(M.rows))
|
|
red = big.rref(iszerofunc=iszerofunc, simplify=True)[0]
|
|
|
|
if any(iszerofunc(red[j, j]) for j in range(red.rows)):
|
|
raise NonInvertibleMatrixError("Matrix det == 0; not invertible.")
|
|
|
|
return M._new(red[:, big.rows:])
|
|
|
|
def _inv_LU(M, iszerofunc=_iszero):
|
|
"""Calculates the inverse using LU decomposition.
|
|
|
|
See Also
|
|
========
|
|
|
|
inv
|
|
inverse_ADJ
|
|
inverse_GE
|
|
inverse_CH
|
|
inverse_LDL
|
|
"""
|
|
|
|
if not M.is_square:
|
|
raise NonSquareMatrixError("A Matrix must be square to invert.")
|
|
if M.free_symbols:
|
|
_verify_invertible(M, iszerofunc=iszerofunc)
|
|
|
|
return M.LUsolve(M.eye(M.rows), iszerofunc=_iszero)
|
|
|
|
def _inv_CH(M, iszerofunc=_iszero):
|
|
"""Calculates the inverse using cholesky decomposition.
|
|
|
|
See Also
|
|
========
|
|
|
|
inv
|
|
inverse_ADJ
|
|
inverse_GE
|
|
inverse_LU
|
|
inverse_LDL
|
|
"""
|
|
|
|
_verify_invertible(M, iszerofunc=iszerofunc)
|
|
|
|
return M.cholesky_solve(M.eye(M.rows))
|
|
|
|
def _inv_LDL(M, iszerofunc=_iszero):
|
|
"""Calculates the inverse using LDL decomposition.
|
|
|
|
See Also
|
|
========
|
|
|
|
inv
|
|
inverse_ADJ
|
|
inverse_GE
|
|
inverse_LU
|
|
inverse_CH
|
|
"""
|
|
|
|
_verify_invertible(M, iszerofunc=iszerofunc)
|
|
|
|
return M.LDLsolve(M.eye(M.rows))
|
|
|
|
def _inv_QR(M, iszerofunc=_iszero):
|
|
"""Calculates the inverse using QR decomposition.
|
|
|
|
See Also
|
|
========
|
|
|
|
inv
|
|
inverse_ADJ
|
|
inverse_GE
|
|
inverse_CH
|
|
inverse_LDL
|
|
"""
|
|
|
|
_verify_invertible(M, iszerofunc=iszerofunc)
|
|
|
|
return M.QRsolve(M.eye(M.rows))
|
|
|
|
def _inv_block(M, iszerofunc=_iszero):
|
|
"""Calculates the inverse using BLOCKWISE inversion.
|
|
|
|
See Also
|
|
========
|
|
|
|
inv
|
|
inverse_ADJ
|
|
inverse_GE
|
|
inverse_CH
|
|
inverse_LDL
|
|
"""
|
|
from sympy.matrices.expressions.blockmatrix import BlockMatrix
|
|
i = M.shape[0]
|
|
if i <= 20 :
|
|
return M.inv(method="LU", iszerofunc=_iszero)
|
|
A = M[:i // 2, :i //2]
|
|
B = M[:i // 2, i // 2:]
|
|
C = M[i // 2:, :i // 2]
|
|
D = M[i // 2:, i // 2:]
|
|
try:
|
|
D_inv = _inv_block(D)
|
|
except NonInvertibleMatrixError:
|
|
return M.inv(method="LU", iszerofunc=_iszero)
|
|
B_D_i = B*D_inv
|
|
BDC = B_D_i*C
|
|
A_n = A - BDC
|
|
try:
|
|
A_n = _inv_block(A_n)
|
|
except NonInvertibleMatrixError:
|
|
return M.inv(method="LU", iszerofunc=_iszero)
|
|
B_n = -A_n*B_D_i
|
|
dc = D_inv*C
|
|
C_n = -dc*A_n
|
|
D_n = D_inv + dc*-B_n
|
|
nn = BlockMatrix([[A_n, B_n], [C_n, D_n]]).as_explicit()
|
|
return nn
|
|
|
|
def _inv(M, method=None, iszerofunc=_iszero, try_block_diag=False):
|
|
"""
|
|
Return the inverse of a matrix using the method indicated. Default for
|
|
dense matrices is is Gauss elimination, default for sparse matrices is LDL.
|
|
|
|
Parameters
|
|
==========
|
|
|
|
method : ('GE', 'LU', 'ADJ', 'CH', 'LDL')
|
|
|
|
iszerofunc : function, optional
|
|
Zero-testing function to use.
|
|
|
|
try_block_diag : bool, optional
|
|
If True then will try to form block diagonal matrices using the
|
|
method get_diag_blocks(), invert these individually, and then
|
|
reconstruct the full inverse matrix.
|
|
|
|
Examples
|
|
========
|
|
|
|
>>> from sympy import SparseMatrix, Matrix
|
|
>>> A = SparseMatrix([
|
|
... [ 2, -1, 0],
|
|
... [-1, 2, -1],
|
|
... [ 0, 0, 2]])
|
|
>>> A.inv('CH')
|
|
Matrix([
|
|
[2/3, 1/3, 1/6],
|
|
[1/3, 2/3, 1/3],
|
|
[ 0, 0, 1/2]])
|
|
>>> A.inv(method='LDL') # use of 'method=' is optional
|
|
Matrix([
|
|
[2/3, 1/3, 1/6],
|
|
[1/3, 2/3, 1/3],
|
|
[ 0, 0, 1/2]])
|
|
>>> A * _
|
|
Matrix([
|
|
[1, 0, 0],
|
|
[0, 1, 0],
|
|
[0, 0, 1]])
|
|
>>> A = Matrix(A)
|
|
>>> A.inv('CH')
|
|
Matrix([
|
|
[2/3, 1/3, 1/6],
|
|
[1/3, 2/3, 1/3],
|
|
[ 0, 0, 1/2]])
|
|
>>> A.inv('ADJ') == A.inv('GE') == A.inv('LU') == A.inv('CH') == A.inv('LDL') == A.inv('QR')
|
|
True
|
|
|
|
Notes
|
|
=====
|
|
|
|
According to the ``method`` keyword, it calls the appropriate method:
|
|
|
|
GE .... inverse_GE(); default for dense matrices
|
|
LU .... inverse_LU()
|
|
ADJ ... inverse_ADJ()
|
|
CH ... inverse_CH()
|
|
LDL ... inverse_LDL(); default for sparse matrices
|
|
QR ... inverse_QR()
|
|
|
|
Note, the GE and LU methods may require the matrix to be simplified
|
|
before it is inverted in order to properly detect zeros during
|
|
pivoting. In difficult cases a custom zero detection function can
|
|
be provided by setting the ``iszerofunc`` argument to a function that
|
|
should return True if its argument is zero. The ADJ routine computes
|
|
the determinant and uses that to detect singular matrices in addition
|
|
to testing for zeros on the diagonal.
|
|
|
|
See Also
|
|
========
|
|
|
|
inverse_ADJ
|
|
inverse_GE
|
|
inverse_LU
|
|
inverse_CH
|
|
inverse_LDL
|
|
|
|
Raises
|
|
======
|
|
|
|
ValueError
|
|
If the determinant of the matrix is zero.
|
|
"""
|
|
|
|
from sympy.matrices import diag, SparseMatrix
|
|
|
|
if method is None:
|
|
method = 'LDL' if isinstance(M, SparseMatrix) else 'GE'
|
|
|
|
if try_block_diag:
|
|
blocks = M.get_diag_blocks()
|
|
r = []
|
|
|
|
for block in blocks:
|
|
r.append(block.inv(method=method, iszerofunc=iszerofunc))
|
|
|
|
return diag(*r)
|
|
|
|
if method == "GE":
|
|
rv = M.inverse_GE(iszerofunc=iszerofunc)
|
|
elif method == "LU":
|
|
rv = M.inverse_LU(iszerofunc=iszerofunc)
|
|
elif method == "ADJ":
|
|
rv = M.inverse_ADJ(iszerofunc=iszerofunc)
|
|
elif method == "CH":
|
|
rv = M.inverse_CH(iszerofunc=iszerofunc)
|
|
elif method == "LDL":
|
|
rv = M.inverse_LDL(iszerofunc=iszerofunc)
|
|
elif method == "QR":
|
|
rv = M.inverse_QR(iszerofunc=iszerofunc)
|
|
elif method == "BLOCK":
|
|
rv = M.inverse_BLOCK(iszerofunc=iszerofunc)
|
|
else:
|
|
raise ValueError("Inversion method unrecognized")
|
|
|
|
return M._new(rv)
|