B d!Z?d"Z@eZAGd#ddeBZCdaDd$d%ZEeEd&d'ZFd(d)ZGd*d+ZHd,d-ZId.d/ZJd0d1ZKd2d3ZLeeeeeeeeiZMeeeeeeeeiZNefd4d5ZOefd6d7ZPd8d9ZQeeeeeeeeiZNd:d;ZReZSdd?ZUd@dAZVdBdCZWdDdEZXdFdGZYdHdIZZdJdKZ[dLdMZ\drdNdZ]dOdZ^dsdQdZ_dRdZ`dSdZadtdUdZbdVd ZcdudXd ZddYdZZed[dZfdvd\dZgdwd^dZhdxd_dZidyd`dZjdzdbd Zkdcd Zlddd Zmd{dfdZndgdhZod|djdZpdkdZqdldmZrd}dndoZsdpdqZtdS)~axLite version of scipy.linalg. Notes ----- This module is a lite version of the linalg.py module in SciPy which contains high-level Python interface to the LAPACK library. The lite version only accesses the following LAPACK functions: dgesv, zgesv, dgeev, zgeev, dgesdd, zgesdd, dgelsd, zgelsd, dsyevd, zheevd, dgetrf, zgetrf, dpotrf, zpotrf, dgeqrf, zgeqrf, zungqr, dorgqr. )divisionabsolute_importprint_function matrix_powersolve tensorsolve tensorinvinvcholeskyeigvalseigvalshpinvslogdetdetsvdeigeighlstsqnormqrcond matrix_rank LinAlgError multi_dotN)*arrayasarrayzerosempty empty_like transposeintcsingledoublecsinglecdoubleinexactcomplexfloatingnewaxisravelallInfdotaddmultiplysqrtmaximumfastCopyAndTransposesumisfinitesizefinfoerrstate geterrobj longdoublerollaxisaminamaxproductabs broadcast atleast_2dintp asanyarrayisscalarobject_ones)normalize_axis_index)triuasfarray) lapack_lite _umath_linalg)rNVASLc@seZdZdZdS)ra Generic Python-exception-derived object raised by linalg functions. General purpose exception class, derived from Python's exception.Exception class, programmatically raised in linalg functions when a Linear Algebra-related condition would prevent further correct execution of the function. Parameters ---------- None Examples -------- >>> from numpy import linalg as LA >>> LA.inv(np.zeros((2,2))) Traceback (most recent call last): File "", line 1, in File "...linalg.py", line 350, in inv return wrap(solve(a, identity(a.shape[0], dtype=a.dtype))) File "...linalg.py", line 249, in solve raise LinAlgError('Singular matrix') numpy.linalg.LinAlgError: Singular matrix N)__name__ __module__ __qualname____doc__rRrRF/opt/alt/python37/lib64/python3.7/site-packages/numpy/linalg/linalg.pyr+sc CsBt}|d}tdddddtd}WdQRX||dgadS)NrZcallignore)ZinvalidZoverZdivideZunder)r6r5_linalg_error_extobj)ZerrobjbufsizeZinvalid_call_errmaskrRrRrS_determine_error_statesLs  rXcCs tddS)NzSingular matrix)r)errflagrRrRrS_raise_linalgerror_singularYsr[cCs tddS)NzMatrix is not positive definite)r)rYrZrRrRrS_raise_linalgerror_nonposdef\sr\cCs tddS)NzEigenvalues did not converge)r)rYrZrRrRrS-_raise_linalgerror_eigenvalues_nonconvergence_sr]cCs tddS)NzSVD did not converge)r)rYrZrRrRrS%_raise_linalgerror_svd_nonconvergencebsr^cCstt}||d<|S)N)listrV)callbackextobjrRrRrSget_linalg_error_extobjesrccCst|}t|d|j}||fS)NZ__array_prepare__)rgetattrZ__array_wrap__)anewwraprRrRrS _makearrayjsrhcCs t|tS)N) issubclassr&)trRrRrS isComplexTypeosrkcCs t||S)N)_real_types_mapget)rjdefaultrRrRrS _realType|srocCs t||S)N)_complex_types_maprm)rjrnrRrRrS _complexTypesrqcCstS)z,Cast the type t to either double or cdouble.)r")rjrRrRrS_linalgRealTypesrrcGst}d}xf|D]^}t|jjtr\t|jjr0d}t|jjdd}|dkr`td|jjfnt }|t krt }qW|rt }t |}nt }||fS)NFT)rnz&array type %s is unsupported in linalg) r!ridtypetyper%rkro TypeErrornamer"r$rp)arraysZ result_typeZ is_complexreZrtrjrRrRrS _commonTypes$   rxcGs`g}x>|D]6}|jjdkr6|t||jddq ||q Wt|dkrX|dS|SdS)N)=|ry)rsrUr)rs byteorderappendrZ newbyteorderlen)rwretarrrRrRrS_to_native_byte_orders   rcGs^d}x<|D]4}|jj|kr*|t|f}q |t||f}q Wt|dkrV|dS|SdS)NrRrUr)rsrt_fastCTastyper})rtrwZ cast_arraysrerRrRrS_fastCopyAndTransposes   rcGs*x$|D]}|jdkrtd|jqWdS)Nr_z9%d-dimensional array given. Array must be two-dimensional)ndimr)rwrerRrRrS _assertRank2s  rcGs*x$|D]}|jdkrtd|jqWdS)Nr_zB%d-dimensional array given. Array must be at least two-dimensional)rr)rwrerRrRrS_assertRankAtLeast2s  rcGs.x(|D] }t|jt|jkrtdqWdS)NzArray must be square)maxshapeminr)rwrerRrRrS_assertSquarenesss rcGs>x8|D]0}t|jddt|jddkrtdqWdS)Nz-Last 2 dimensions of the array must be square)rrrr)rwrerRrRrS_assertNdSquarenesss $rcGs&x |D]}t|stdqWdS)Nz#Array must not contain infs or NaNs)r2r)r)rwrerRrRrS _assertFinites  rcCs |jdkot|jdddkS)Nrr)r3r;r)rrRrRrS _isEmpty2dsrcGs"x|D]}t|rtdqWdS)NzArrays cannot be empty)rr)rwrerRrRrS_assertNoEmpty2ds rc Cst|\}}t|}|j}|dk r^ttd|}x"|D]}|||||q6W||}|j||j d}d}x|D] }||9}q~W| d|}| }|t ||} || _| S)a Solve the tensor equation ``a x = b`` for x. It is assumed that all indices of `x` are summed over in the product, together with the rightmost indices of `a`, as is done in, for example, ``tensordot(a, x, axes=b.ndim)``. Parameters ---------- a : array_like Coefficient tensor, of shape ``b.shape + Q``. `Q`, a tuple, equals the shape of that sub-tensor of `a` consisting of the appropriate number of its rightmost indices, and must be such that ``prod(Q) == prod(b.shape)`` (in which sense `a` is said to be 'square'). b : array_like Right-hand tensor, which can be of any shape. axes : tuple of ints, optional Axes in `a` to reorder to the right, before inversion. If None (default), no reordering is done. Returns ------- x : ndarray, shape Q Raises ------ LinAlgError If `a` is singular or not 'square' (in the above sense). See Also -------- numpy.tensordot, tensorinv, numpy.einsum Examples -------- >>> a = np.eye(2*3*4) >>> a.shape = (2*3, 4, 2, 3, 4) >>> b = np.random.randn(2*3, 4) >>> x = np.linalg.tensorsolve(a, b) >>> x.shape (2, 3, 4) >>> np.allclose(np.tensordot(a, x, axes=3), b) True NrrU) rhrrr`rangeremoveinsertrrreshaper(r) rebZaxesrgZanZallaxeskoldshapeprodresrRrRrSrs$/       c Cst|\}}t|t|t|\}}t||\}}|j|jdkrNtj}ntj}t|r`dnd}t t }|||||d} || j |ddS)a Solve a linear matrix equation, or system of linear scalar equations. Computes the "exact" solution, `x`, of the well-determined, i.e., full rank, linear matrix equation `ax = b`. Parameters ---------- a : (..., M, M) array_like Coefficient matrix. b : {(..., M,), (..., M, K)}, array_like Ordinate or "dependent variable" values. Returns ------- x : {(..., M,), (..., M, K)} ndarray Solution to the system a x = b. Returned shape is identical to `b`. Raises ------ LinAlgError If `a` is singular or not square. Notes ----- .. versionadded:: 1.8.0 Broadcasting rules apply, see the `numpy.linalg` documentation for details. The solutions are computed using LAPACK routine _gesv `a` must be square and of full-rank, i.e., all rows (or, equivalently, columns) must be linearly independent; if either is not true, use `lstsq` for the least-squares best "solution" of the system/equation. References ---------- .. [1] G. Strang, *Linear Algebra and Its Applications*, 2nd Ed., Orlando, FL, Academic Press, Inc., 1980, pg. 22. Examples -------- Solve the system of equations ``3 * x0 + x1 = 9`` and ``x0 + 2 * x1 = 8``: >>> a = np.array([[3,1], [1,2]]) >>> b = np.array([9,8]) >>> x = np.linalg.solve(a, b) >>> x array([ 2., 3.]) Check that the solution is correct: >>> np.allclose(np.dot(a, x), b) True rUzDD->Dzdd->d) signaturerbF)copy) rhrrrxrrHZsolve1rrkrcr[r) rer_rgrjresult_tgufuncrrbrrRrRrSr,s<  r_cCsxt|}|j}d}|dkrR||d|d|}x&||dD] }||9}q@Wntd||d}t|}|j|S)a Compute the 'inverse' of an N-dimensional array. The result is an inverse for `a` relative to the tensordot operation ``tensordot(a, b, ind)``, i. e., up to floating-point accuracy, ``tensordot(tensorinv(a), a, ind)`` is the "identity" tensor for the tensordot operation. Parameters ---------- a : array_like Tensor to 'invert'. Its shape must be 'square', i. e., ``prod(a.shape[:ind]) == prod(a.shape[ind:])``. ind : int, optional Number of first indices that are involved in the inverse sum. Must be a positive integer, default is 2. Returns ------- b : ndarray `a`'s tensordot inverse, shape ``a.shape[ind:] + a.shape[:ind]``. Raises ------ LinAlgError If `a` is singular or not 'square' (in the above sense). See Also -------- numpy.tensordot, tensorsolve Examples -------- >>> a = np.eye(4*6) >>> a.shape = (4, 6, 8, 3) >>> ainv = np.linalg.tensorinv(a, ind=2) >>> ainv.shape (8, 3, 4, 6) >>> b = np.random.randn(4, 6) >>> np.allclose(np.tensordot(ainv, b), np.linalg.tensorsolve(a, b)) True >>> a = np.eye(4*6) >>> a.shape = (24, 8, 3) >>> ainv = np.linalg.tensorinv(a, ind=1) >>> ainv.shape (8, 3, 24) >>> b = np.random.randn(24) >>> np.allclose(np.tensordot(ainv, b, 1), np.linalg.tensorsolve(a, b)) True rUrNzInvalid ind argument.r)rr ValueErrorrr )reZindrrZinvshaperZiarRrRrSr|s5 cCsbt|\}}t|t|t|\}}t|r4dnd}tt}tj|||d}||j |ddS)a Compute the (multiplicative) inverse of a matrix. Given a square matrix `a`, return the matrix `ainv` satisfying ``dot(a, ainv) = dot(ainv, a) = eye(a.shape[0])``. Parameters ---------- a : (..., M, M) array_like Matrix to be inverted. Returns ------- ainv : (..., M, M) ndarray or matrix (Multiplicative) inverse of the matrix `a`. Raises ------ LinAlgError If `a` is not square or inversion fails. Notes ----- .. versionadded:: 1.8.0 Broadcasting rules apply, see the `numpy.linalg` documentation for details. Examples -------- >>> from numpy.linalg import inv >>> a = np.array([[1., 2.], [3., 4.]]) >>> ainv = inv(a) >>> np.allclose(np.dot(a, ainv), np.eye(2)) True >>> np.allclose(np.dot(ainv, a), np.eye(2)) True If a is a matrix object, then the return value is a matrix as well: >>> ainv = inv(np.matrix(a)) >>> ainv matrix([[-2. , 1. ], [ 1.5, -0.5]]) Inverses of several matrices can be computed at once: >>> a = np.array([[[1., 2.], [3., 4.]], [[1, 3], [3, 5]]]) >>> inv(a) array([[[-2. , 1. ], [ 1.5, -0.5]], [[-5. , 2. ], [ 3. , -1. ]]]) zD->Dzd->d)rrbF)r) rhrrrxrkrcr[rHr r)rergrjrrrbZainvrRrRrSr s9  cCsftt}tj}t|\}}t|t|t|\}}t|rBdnd}||||d}||j |ddS)a Cholesky decomposition. Return the Cholesky decomposition, `L * L.H`, of the square matrix `a`, where `L` is lower-triangular and .H is the conjugate transpose operator (which is the ordinary transpose if `a` is real-valued). `a` must be Hermitian (symmetric if real-valued) and positive-definite. Only `L` is actually returned. Parameters ---------- a : (..., M, M) array_like Hermitian (symmetric if all elements are real), positive-definite input matrix. Returns ------- L : (..., M, M) array_like Upper or lower-triangular Cholesky factor of `a`. Returns a matrix object if `a` is a matrix object. Raises ------ LinAlgError If the decomposition fails, for example, if `a` is not positive-definite. Notes ----- .. versionadded:: 1.8.0 Broadcasting rules apply, see the `numpy.linalg` documentation for details. The Cholesky decomposition is often used as a fast way of solving .. math:: A \mathbf{x} = \mathbf{b} (when `A` is both Hermitian/symmetric and positive-definite). First, we solve for :math:`\mathbf{y}` in .. math:: L \mathbf{y} = \mathbf{b}, and then for :math:`\mathbf{x}` in .. math:: L.H \mathbf{x} = \mathbf{y}. Examples -------- >>> A = np.array([[1,-2j],[2j,5]]) >>> A array([[ 1.+0.j, 0.-2.j], [ 0.+2.j, 5.+0.j]]) >>> L = np.linalg.cholesky(A) >>> L array([[ 1.+0.j, 0.+0.j], [ 0.+2.j, 1.+0.j]]) >>> np.dot(L, L.T.conj()) # verify that L * L.H = A array([[ 1.+0.j, 0.-2.j], [ 0.+2.j, 5.+0.j]]) >>> A = [[1,-2j],[2j,5]] # what happens if A is only array_like? >>> np.linalg.cholesky(A) # an ndarray object is returned array([[ 1.+0.j, 0.+0.j], [ 0.+2.j, 1.+0.j]]) >>> # But a matrix object is returned if A is a matrix object >>> LA.cholesky(np.matrix(A)) matrix([[ 1.+0.j, 0.+0.j], [ 0.+2.j, 1.+0.j]]) zD->Dzd->d)rrbF)r) rcr\rHZ cholesky_lorhrrrxrkr)rerbrrgrjrrrrRrRrSr sI  reducedc Cs&|dkr^|dkr0dd}tj|tddd}n.|dkrRd }tj|tddd }n td |t|\}}t|t||j\}}t |\}}t ||}t |}t ||}t |f|} t|rtj} d } n tj} d } d} t | f|} | ||||| | dd}|ddkrtd| |dftt| d} t | f|} | ||||| | | d}|ddkrrtd| |df|dkrt ||ddd|f}|t|S|dkr|| fS|d kr||kr|j|dd}||jS|dkr||kr|}t||f|}n|}t||f|}||d|<t|r>> a = np.random.randn(9, 6) >>> q, r = np.linalg.qr(a) >>> np.allclose(a, np.dot(q, r)) # a does equal qr True >>> r2 = np.linalg.qr(a, mode='r') >>> r3 = np.linalg.qr(a, mode='economic') >>> np.allclose(r, r2) # mode='r' returns the same r as mode='full' True >>> # But only triu parts are guaranteed equal when mode='economic' >>> np.allclose(r, np.triu(r3[:6,:6], k=0)) True Example illustrating a common use of `qr`: solving of least squares problems What are the least-squares-best `m` and `y0` in ``y = y0 + mx`` for the following data: {(0,1), (1,0), (1,2), (2,1)}. (Graph the points and you'll see that it should be y0 = 0, m = 1.) The answer is provided by solving the over-determined matrix equation ``Ax = b``, where:: A = array([[0, 1], [1, 1], [1, 1], [2, 1]]) x = array([[y0], [m]]) b = array([[1], [0], [2], [1]]) If A = qr such that q is orthonormal (which is always possible via Gram-Schmidt), then ``x = inv(r) * (q.T) * b``. (In numpy practice, however, we simply use `lstsq`.) >>> A = np.array([[0, 1], [1, 1], [1, 1], [2, 1]]) >>> A array([[0, 1], [1, 1], [1, 1], [2, 1]]) >>> b = np.array([1, 0, 2, 1]) >>> q, r = LA.qr(A) >>> p = np.dot(q.T, b) >>> np.dot(LA.inv(r), p) array([ 1.1e-16, 1.0e+00]) )rcompleterraw)fZfull)z7The 'full' option is deprecated in favor of 'reduced'. z,For backward compatibility let mode default.r_) stacklevelr)eeconomicz$The 'economic' option is deprecated.rzUnrecognized mode '%s'zgeqrfdgeqrfrUrrinfoz %s returns %drNrF)rrzungqrdorgqr)joinwarningswarnDeprecationWarningrrhrrrrxrrrrrkrGrrrintr<rErTrrr)remodemsgrgmnrjrZmnZtaulapack_routineZ routine_namelworkworkresultsrZmcqrRrRrSr\su                   cCst|\}}t|t|t|t|\}}tt}t|rDdnd}tj |||d}t|st |j dkr~|j }t |}nt|}|j|ddS)a Compute the eigenvalues of a general matrix. Main difference between `eigvals` and `eig`: the eigenvectors aren't returned. Parameters ---------- a : (..., M, M) array_like A complex- or real-valued matrix whose eigenvalues will be computed. Returns ------- w : (..., M,) ndarray The eigenvalues, each repeated according to its multiplicity. They are not necessarily ordered, nor are they necessarily real for real matrices. Raises ------ LinAlgError If the eigenvalue computation does not converge. See Also -------- eig : eigenvalues and right eigenvectors of general arrays eigvalsh : eigenvalues of symmetric or Hermitian arrays. eigh : eigenvalues and eigenvectors of symmetric/Hermitian arrays. Notes ----- .. versionadded:: 1.8.0 Broadcasting rules apply, see the `numpy.linalg` documentation for details. This is implemented using the _geev LAPACK routines which compute the eigenvalues and eigenvectors of general square arrays. Examples -------- Illustration, using the fact that the eigenvalues of a diagonal matrix are its diagonal elements, that multiplying a matrix on the left by an orthogonal matrix, `Q`, and on the right by `Q.T` (the transpose of `Q`), preserves the eigenvalues of the "middle" matrix. In other words, if `Q` is orthogonal, then ``Q * A * Q.T`` has the same eigenvalues as ``A``: >>> from numpy import linalg as LA >>> x = np.random.random() >>> Q = np.array([[np.cos(x), -np.sin(x)], [np.sin(x), np.cos(x)]]) >>> LA.norm(Q[0, :]), LA.norm(Q[1, :]), np.dot(Q[0, :],Q[1, :]) (1.0, 1.0, 0.0) Now multiply a diagonal matrix by Q on one side and by Q.T on the other: >>> D = np.diag((-1,1)) >>> LA.eigvals(D) array([-1., 1.]) >>> A = np.dot(Q, D) >>> A = np.dot(A, Q.T) >>> LA.eigvals(A) array([ 1., -1.]) zD->Dzd->D)rrbrF)r)rhrrrrxrcr]rkrHr r)imagrealrorqr)rergrjrrbrwrRrRrSr 3sC   Lc Cs|}|dkrtdtt}|dkr0tj}ntj}t|\}}t|t |t |\}}t |rjdnd}||||d}|j t |ddS) al Compute the eigenvalues of a Hermitian or real symmetric matrix. Main difference from eigh: the eigenvectors are not computed. Parameters ---------- a : (..., M, M) array_like A complex- or real-valued matrix whose eigenvalues are to be computed. UPLO : {'L', 'U'}, optional Specifies whether the calculation is done with the lower triangular part of `a` ('L', default) or the upper triangular part ('U'). Irrespective of this value only the real parts of the diagonal will be considered in the computation to preserve the notion of a Hermitian matrix. It therefore follows that the imaginary part of the diagonal will always be treated as zero. Returns ------- w : (..., M,) ndarray The eigenvalues in ascending order, each repeated according to its multiplicity. Raises ------ LinAlgError If the eigenvalue computation does not converge. See Also -------- eigh : eigenvalues and eigenvectors of symmetric/Hermitian arrays. eigvals : eigenvalues of general real or complex arrays. eig : eigenvalues and right eigenvectors of general real or complex arrays. Notes ----- .. versionadded:: 1.8.0 Broadcasting rules apply, see the `numpy.linalg` documentation for details. The eigenvalues are computed using LAPACK routines _syevd, _heevd Examples -------- >>> from numpy import linalg as LA >>> a = np.array([[1, -2j], [2j, 5]]) >>> LA.eigvalsh(a) array([ 0.17157288, 5.82842712]) >>> # demonstrate the treatment of the imaginary part of the diagonal >>> a = np.array([[5+2j, 9-2j], [0+2j, 2-1j]]) >>> a array([[ 5.+2.j, 9.-2.j], [ 0.+2.j, 2.-1.j]]) >>> # with UPLO='L' this is numerically equivalent to using LA.eigvals() >>> # with: >>> b = np.array([[5.+0.j, 0.-2.j], [0.+2.j, 2.-0.j]]) >>> b array([[ 5.+0.j, 0.-2.j], [ 0.+2.j, 2.+0.j]]) >>> wa = LA.eigvalsh(a) >>> wb = LA.eigvals(b) >>> wa; wb array([ 1., 6.]) array([ 6.+0.j, 1.+0.j]) )rUz UPLO argument must be 'L' or 'U'rzD->dzd->d)rrbF)r)upperrrcr]rHZ eigvalsh_loZ eigvalsh_uprhrrrxrkrro) reUPLOrbrrgrjrrrrRrRrSr sH  cCs$t|\}}t||}|||fS)N)rxrr)rerjrrRrRrS _convertarrays rcCst|\}}t|t|t|t|\}}tt}t|rDdnd}tj |||d\}}t|st |j dkr|j }|j }t |}nt|}|j|dd}|j|dd||fS)a Compute the eigenvalues and right eigenvectors of a square array. Parameters ---------- a : (..., M, M) array Matrices for which the eigenvalues and right eigenvectors will be computed Returns ------- w : (..., M) array The eigenvalues, each repeated according to its multiplicity. The eigenvalues are not necessarily ordered. The resulting array will be of complex type, unless the imaginary part is zero in which case it will be cast to a real type. When `a` is real the resulting eigenvalues will be real (0 imaginary part) or occur in conjugate pairs v : (..., M, M) array The normalized (unit "length") eigenvectors, such that the column ``v[:,i]`` is the eigenvector corresponding to the eigenvalue ``w[i]``. Raises ------ LinAlgError If the eigenvalue computation does not converge. See Also -------- eigvals : eigenvalues of a non-symmetric array. eigh : eigenvalues and eigenvectors of a symmetric or Hermitian (conjugate symmetric) array. eigvalsh : eigenvalues of a symmetric or Hermitian (conjugate symmetric) array. Notes ----- .. versionadded:: 1.8.0 Broadcasting rules apply, see the `numpy.linalg` documentation for details. This is implemented using the _geev LAPACK routines which compute the eigenvalues and eigenvectors of general square arrays. The number `w` is an eigenvalue of `a` if there exists a vector `v` such that ``dot(a,v) = w * v``. Thus, the arrays `a`, `w`, and `v` satisfy the equations ``dot(a[:,:], v[:,i]) = w[i] * v[:,i]`` for :math:`i \in \{0,...,M-1\}`. The array `v` of eigenvectors may not be of maximum rank, that is, some of the columns may be linearly dependent, although round-off error may obscure that fact. If the eigenvalues are all different, then theoretically the eigenvectors are linearly independent. Likewise, the (complex-valued) matrix of eigenvectors `v` is unitary if the matrix `a` is normal, i.e., if ``dot(a, a.H) = dot(a.H, a)``, where `a.H` denotes the conjugate transpose of `a`. Finally, it is emphasized that `v` consists of the *right* (as in right-hand side) eigenvectors of `a`. A vector `y` satisfying ``dot(y.T, a) = z * y.T`` for some number `z` is called a *left* eigenvector of `a`, and, in general, the left and right eigenvectors of a matrix are not necessarily the (perhaps conjugate) transposes of each other. References ---------- G. Strang, *Linear Algebra and Its Applications*, 2nd Ed., Orlando, FL, Academic Press, Inc., 1980, Various pp. Examples -------- >>> from numpy import linalg as LA (Almost) trivial example with real e-values and e-vectors. >>> w, v = LA.eig(np.diag((1, 2, 3))) >>> w; v array([ 1., 2., 3.]) array([[ 1., 0., 0.], [ 0., 1., 0.], [ 0., 0., 1.]]) Real matrix possessing complex e-values and e-vectors; note that the e-values are complex conjugates of each other. >>> w, v = LA.eig(np.array([[1, -1], [1, 1]])) >>> w; v array([ 1. + 1.j, 1. - 1.j]) array([[ 0.70710678+0.j , 0.70710678+0.j ], [ 0.00000000-0.70710678j, 0.00000000+0.70710678j]]) Complex-valued matrix with real e-values (but complex-valued e-vectors); note that a.conj().T = a, i.e., a is Hermitian. >>> a = np.array([[1, 1j], [-1j, 1]]) >>> w, v = LA.eig(a) >>> w; v array([ 2.00000000e+00+0.j, 5.98651912e-36+0.j]) # i.e., {2, 0} array([[ 0.00000000+0.70710678j, 0.70710678+0.j ], [ 0.70710678+0.j , 0.00000000+0.70710678j]]) Be careful about round-off error! >>> a = np.array([[1 + 1e-9, 0], [0, 1 - 1e-9]]) >>> # Theor. e-values are 1 +/- 1e-9 >>> w, v = LA.eig(a) >>> w; v array([ 1., 1.]) array([[ 1., 0.], [ 0., 1.]]) zD->DDzd->DD)rrbgF)r)rhrrrrxrcr]rkrHrr)rrrorqr)rergrjrrbrrvtrRrRrSrs w   c Cs|}|dkrtdt|\}}t|t|t|\}}tt}|dkrXtj }ntj }t |rjdnd}||||d\}} |j t |dd}| j |dd} ||| fS) aj Return the eigenvalues and eigenvectors of a Hermitian or symmetric matrix. Returns two objects, a 1-D array containing the eigenvalues of `a`, and a 2-D square array or matrix (depending on the input type) of the corresponding eigenvectors (in columns). Parameters ---------- a : (..., M, M) array Hermitian/Symmetric matrices whose eigenvalues and eigenvectors are to be computed. UPLO : {'L', 'U'}, optional Specifies whether the calculation is done with the lower triangular part of `a` ('L', default) or the upper triangular part ('U'). Irrespective of this value only the real parts of the diagonal will be considered in the computation to preserve the notion of a Hermitian matrix. It therefore follows that the imaginary part of the diagonal will always be treated as zero. Returns ------- w : (..., M) ndarray The eigenvalues in ascending order, each repeated according to its multiplicity. v : {(..., M, M) ndarray, (..., M, M) matrix} The column ``v[:, i]`` is the normalized eigenvector corresponding to the eigenvalue ``w[i]``. Will return a matrix object if `a` is a matrix object. Raises ------ LinAlgError If the eigenvalue computation does not converge. See Also -------- eigvalsh : eigenvalues of symmetric or Hermitian arrays. eig : eigenvalues and right eigenvectors for non-symmetric arrays. eigvals : eigenvalues of non-symmetric arrays. Notes ----- .. versionadded:: 1.8.0 Broadcasting rules apply, see the `numpy.linalg` documentation for details. The eigenvalues/eigenvectors are computed using LAPACK routines _syevd, _heevd The eigenvalues of real symmetric or complex Hermitian matrices are always real. [1]_ The array `v` of (column) eigenvectors is unitary and `a`, `w`, and `v` satisfy the equations ``dot(a, v[:, i]) = w[i] * v[:, i]``. References ---------- .. [1] G. Strang, *Linear Algebra and Its Applications*, 2nd Ed., Orlando, FL, Academic Press, Inc., 1980, pg. 222. Examples -------- >>> from numpy import linalg as LA >>> a = np.array([[1, -2j], [2j, 5]]) >>> a array([[ 1.+0.j, 0.-2.j], [ 0.+2.j, 5.+0.j]]) >>> w, v = LA.eigh(a) >>> w; v array([ 0.17157288, 5.82842712]) array([[-0.92387953+0.j , -0.38268343+0.j ], [ 0.00000000+0.38268343j, 0.00000000-0.92387953j]]) >>> np.dot(a, v[:, 0]) - w[0] * v[:, 0] # verify 1st e-val/vec pair array([2.77555756e-17 + 0.j, 0. + 1.38777878e-16j]) >>> np.dot(a, v[:, 1]) - w[1] * v[:, 1] # verify 2nd e-val/vec pair array([ 0.+0.j, 0.+0.j]) >>> A = np.matrix(a) # what happens if input is a matrix object >>> A matrix([[ 1.+0.j, 0.-2.j], [ 0.+2.j, 5.+0.j]]) >>> w, v = LA.eigh(A) >>> w; v array([ 0.17157288, 5.82842712]) matrix([[-0.92387953+0.j , -0.38268343+0.j ], [ 0.00000000+0.38268343j, 0.00000000-0.92387953j]]) >>> # demonstrate the treatment of the imaginary part of the diagonal >>> a = np.array([[5+2j, 9-2j], [0+2j, 2-1j]]) >>> a array([[ 5.+2.j, 9.-2.j], [ 0.+2.j, 2.-1.j]]) >>> # with UPLO='L' this is numerically equivalent to using LA.eig() with: >>> b = np.array([[5.+0.j, 0.-2.j], [0.+2.j, 2.-0.j]]) >>> b array([[ 5.+0.j, 0.-2.j], [ 0.+2.j, 2.+0.j]]) >>> wa, va = LA.eigh(a) >>> wb, vb = LA.eig(b) >>> wa; wb array([ 1., 6.]) array([ 6.+0.j, 1.+0.j]) >>> va; vb array([[-0.44721360-0.j , -0.89442719+0.j ], [ 0.00000000+0.89442719j, 0.00000000-0.4472136j ]]) array([[ 0.89442719+0.j , 0.00000000-0.4472136j], [ 0.00000000-0.4472136j, 0.89442719+0.j ]]) )rrz UPLO argument must be 'L' or 'U'rzD->dDzd->dd)rrbF)r)rrrhrrrxrcr]rHZeigh_loZeigh_uprkrro) rerrgrjrrbrrrrrRrRrSr{s"p  rUcCs.t|\}}t|t|t|\}}tt}|jd}|jd}|r|rd||kr\tj} qztj } n||krttj } ntj } t |rdnd} | || |d\} } } | j |dd} | j t|dd} | j |dd} || | || fS||krtj} ntj} t |rdnd } | || |d} | j t|dd} | Sd S) a Singular Value Decomposition. Factors the matrix `a` as ``u * np.diag(s) * v``, where `u` and `v` are unitary and `s` is a 1-d array of `a`'s singular values. Parameters ---------- a : (..., M, N) array_like A real or complex matrix of shape (`M`, `N`) . full_matrices : bool, optional If True (default), `u` and `v` have the shapes (`M`, `M`) and (`N`, `N`), respectively. Otherwise, the shapes are (`M`, `K`) and (`K`, `N`), respectively, where `K` = min(`M`, `N`). compute_uv : bool, optional Whether or not to compute `u` and `v` in addition to `s`. True by default. Returns ------- u : { (..., M, M), (..., M, K) } array Unitary matrices. The actual shape depends on the value of ``full_matrices``. Only returned when ``compute_uv`` is True. s : (..., K) array The singular values for every matrix, sorted in descending order. v : { (..., N, N), (..., K, N) } array Unitary matrices. The actual shape depends on the value of ``full_matrices``. Only returned when ``compute_uv`` is True. Raises ------ LinAlgError If SVD computation does not converge. Notes ----- .. versionadded:: 1.8.0 Broadcasting rules apply, see the `numpy.linalg` documentation for details. The decomposition is performed using LAPACK routine _gesdd The SVD is commonly written as ``a = U S V.H``. The `v` returned by this function is ``V.H`` and ``u = U``. If ``U`` is a unitary matrix, it means that it satisfies ``U.H = inv(U)``. The rows of `v` are the eigenvectors of ``a.H a``. The columns of `u` are the eigenvectors of ``a a.H``. For row ``i`` in `v` and column ``i`` in `u`, the corresponding eigenvalue is ``s[i]**2``. If `a` is a `matrix` object (as opposed to an `ndarray`), then so are all the return values. Examples -------- >>> a = np.random.randn(9, 6) + 1j*np.random.randn(9, 6) Reconstruction based on full SVD: >>> U, s, V = np.linalg.svd(a, full_matrices=True) >>> U.shape, V.shape, s.shape ((9, 9), (6, 6), (6,)) >>> S = np.zeros((9, 6), dtype=complex) >>> S[:6, :6] = np.diag(s) >>> np.allclose(a, np.dot(U, np.dot(S, V))) True Reconstruction based on reduced SVD: >>> U, s, V = np.linalg.svd(a, full_matrices=False) >>> U.shape, V.shape, s.shape ((9, 6), (6, 6), (6,)) >>> S = np.diag(s) >>> np.allclose(a, np.dot(U, np.dot(S, V))) True rrzD->DdDzd->ddd)rrbF)rzD->dzd->dN)rhrrrxrcr^rrHZsvd_m_fZsvd_n_fZsvd_m_sZsvd_n_srkrroZsvd_mZsvd_n)reZ full_matrices compute_uvrgrjrrbrrrrusrrRrRrSrs8S    cCsPt|}|dkr,t|dd}|d|dSt||ddtt||ddSdS)a Compute the condition number of a matrix. This function is capable of returning the condition number using one of seven different norms, depending on the value of `p` (see Parameters below). Parameters ---------- x : (..., M, N) array_like The matrix whose condition number is sought. p : {None, 1, -1, 2, -2, inf, -inf, 'fro'}, optional Order of the norm: ===== ============================ p norm for matrices ===== ============================ None 2-norm, computed directly using the ``SVD`` 'fro' Frobenius norm inf max(sum(abs(x), axis=1)) -inf min(sum(abs(x), axis=1)) 1 max(sum(abs(x), axis=0)) -1 min(sum(abs(x), axis=0)) 2 2-norm (largest sing. value) -2 smallest singular value ===== ============================ inf means the numpy.inf object, and the Frobenius norm is the root-of-sum-of-squares norm. Returns ------- c : {float, inf} The condition number of the matrix. May be infinite. See Also -------- numpy.linalg.norm Notes ----- The condition number of `x` is defined as the norm of `x` times the norm of the inverse of `x` [1]_; the norm can be the usual L2-norm (root-of-sum-of-squares) or one of a number of other matrix norms. References ---------- .. [1] G. Strang, *Linear Algebra and Its Applications*, Orlando, FL, Academic Press, Inc., 1980, pg. 285. Examples -------- >>> from numpy import linalg as LA >>> a = np.array([[1, 0, -1], [0, 1, 0], [1, 0, 1]]) >>> a array([[ 1, 0, -1], [ 0, 1, 0], [ 1, 0, 1]]) >>> LA.cond(a) 1.4142135623730951 >>> LA.cond(a, 'fro') 3.1622776601683795 >>> LA.cond(a, np.inf) 2.0 >>> LA.cond(a, -np.inf) 1.0 >>> LA.cond(a, 1) 2.0 >>> LA.cond(a, -1) 1.0 >>> LA.cond(a, 2) 1.4142135623730951 >>> LA.cond(a, -2) 0.70710678118654746 >>> min(LA.svd(a, compute_uv=0))*min(LA.svd(LA.inv(a), compute_uv=0)) 0.70710678118654746 NF)r).r).r)rr)axis)rrrr )xprrRrRrSr}s O cCstt|}|jdkr$tt|dk St|dd}|dkrd|jdddt|jd dt|jj }||kj dd S) a Return matrix rank of array using SVD method Rank of the array is the number of SVD singular values of the array that are greater than `tol`. Parameters ---------- M : {(M,), (..., M, N)} array_like input vector or stack of matrices tol : {None, float}, optional threshold below which SVD values are considered zero. If `tol` is None, and ``S`` is an array with singular values for `M`, and ``eps`` is the epsilon value for datatype of ``S``, then `tol` is set to ``S.max() * max(M.shape) * eps``. Notes ----- The default threshold to detect rank deficiency is a test on the magnitude of the singular values of `M`. By default, we identify singular values less than ``S.max() * max(M.shape) * eps`` as indicating rank deficiency (with the symbols defined above). This is the algorithm MATLAB uses [1]. It also appears in *Numerical recipes* in the discussion of SVD solutions for linear least squares [2]. This default threshold is designed to detect rank deficiency accounting for the numerical errors of the SVD computation. Imagine that there is a column in `M` that is an exact (in floating point) linear combination of other columns in `M`. Computing the SVD on `M` will not produce a singular value exactly equal to 0 in general: any difference of the smallest SVD value from 0 will be caused by numerical imprecision in the calculation of the SVD. Our threshold for small SVD values takes this numerical imprecision into account, and the default threshold will detect such numerical rank deficiency. The threshold may declare a matrix `M` rank deficient even if the linear combination of some columns of `M` is not exactly equal to another column of `M` but only numerically very close to another column of `M`. We chose our default threshold because it is in wide use. Other thresholds are possible. For example, elsewhere in the 2007 edition of *Numerical recipes* there is an alternative threshold of ``S.max() * np.finfo(M.dtype).eps / 2. * np.sqrt(m + n + 1.)``. The authors describe this threshold as being based on "expected roundoff error" (p 71). The thresholds above deal with floating point roundoff error in the calculation of the SVD. However, you may have more information about the sources of error in `M` that would make you consider other tolerance values to detect *effective* rank deficiency. The most useful measure of the tolerance depends on the operations you intend to use on your matrix. For example, if your data come from uncertain measurements with uncertainties greater than floating point epsilon, choosing a tolerance near that uncertainty may be preferable. The tolerance may be absolute if the uncertainties are absolute rather than relative. References ---------- .. [1] MATLAB reference documention, "Rank" http://www.mathworks.com/help/techdoc/ref/rank.html .. [2] W. H. Press, S. A. Teukolsky, W. T. Vetterling and B. P. Flannery, "Numerical Recipes (3rd edition)", Cambridge University Press, 2007, page 795. Examples -------- >>> from numpy.linalg import matrix_rank >>> matrix_rank(np.eye(4)) # Full rank matrix 4 >>> I=np.eye(4); I[-1,-1] = 0. # rank deficient matrix >>> matrix_rank(I) 3 >>> matrix_rank(np.ones((4,))) # 1 dimension - rank 1 unless all 0 1 >>> matrix_rank(np.zeros((4,))) 0 r_rF)rNrT)rkeepdimsr)r) rrrr)rrrr4rsZepsr1)MZtolSrRrRrSrsL  ,V瞯>> a = np.random.randn(9, 6) >>> B = np.linalg.pinv(a) >>> np.allclose(a, np.dot(a, np.dot(B, a))) True >>> np.allclose(B, np.dot(B, np.dot(a, B))) True Nrr)rsrrUg?g)rhrrrrs conjugaterr/reducerrr+rr-r') rercondrgrrrrrrcutoffirRrRrSr +s? ,    $cCst|}t|t|t|\}}t|}t|r8dnd}tj||d\}}t|rb| |}n|j |dd}t|r| |}n|j |dd}||fS)a< Compute the sign and (natural) logarithm of the determinant of an array. If an array has a very small or very large determinant, then a call to `det` may overflow or underflow. This routine is more robust against such issues, because it computes the logarithm of the determinant rather than the determinant itself. Parameters ---------- a : (..., M, M) array_like Input array, has to be a square 2-D array. Returns ------- sign : (...) array_like A number representing the sign of the determinant. For a real matrix, this is 1, 0, or -1. For a complex matrix, this is a complex number with absolute value 1 (i.e., it is on the unit circle), or else 0. logdet : (...) array_like The natural log of the absolute value of the determinant. If the determinant is zero, then `sign` will be 0 and `logdet` will be -Inf. In all cases, the determinant is equal to ``sign * np.exp(logdet)``. See Also -------- det Notes ----- .. versionadded:: 1.8.0 Broadcasting rules apply, see the `numpy.linalg` documentation for details. .. versionadded:: 1.6.0 The determinant is computed via LU factorization using the LAPACK routine z/dgetrf. Examples -------- The determinant of a 2-D array ``[[a, b], [c, d]]`` is ``ad - bc``: >>> a = np.array([[1, 2], [3, 4]]) >>> (sign, logdet) = np.linalg.slogdet(a) >>> (sign, logdet) (-1, 0.69314718055994529) >>> sign * np.exp(logdet) -2.0 Computing log-determinants for a stack of matrices: >>> a = np.array([ [[1, 2], [3, 4]], [[1, 2], [2, 1]], [[1, 3], [3, 1]] ]) >>> a.shape (3, 2, 2) >>> sign, logdet = np.linalg.slogdet(a) >>> (sign, logdet) (array([-1., -1., -1.]), array([ 0.69314718, 1.09861229, 2.07944154])) >>> sign * np.exp(logdet) array([-2., -3., -8.]) This routine succeeds where ordinary `det` does not: >>> np.linalg.det(np.eye(500) * 0.1) 0.0 >>> np.linalg.slogdet(np.eye(500) * 0.1) (1, -1151.2925464970228) zD->Ddzd->dd)rF)r) rrrrxrorkrHrrAr)rerjrreal_trZsignZlogdetrRrRrSr}sJ   cCsht|}t|t|t|\}}t|r0dnd}tj||d}t|rV||}n|j|dd}|S)a Compute the determinant of an array. Parameters ---------- a : (..., M, M) array_like Input array to compute determinants for. Returns ------- det : (...) array_like Determinant of `a`. See Also -------- slogdet : Another way to representing the determinant, more suitable for large matrices where underflow/overflow may occur. Notes ----- .. versionadded:: 1.8.0 Broadcasting rules apply, see the `numpy.linalg` documentation for details. The determinant is computed via LU factorization using the LAPACK routine z/dgetrf. Examples -------- The determinant of a 2-D array [[a, b], [c, d]] is ad - bc: >>> a = np.array([[1, 2], [3, 4]]) >>> np.linalg.det(a) -2.0 Computing determinants for a stack of matrices: >>> a = np.array([ [[1, 2], [3, 4]], [[1, 2], [2, 1]], [[1, 3], [3, 1]] ]) >>> a.shape (3, 2, 2) >>> np.linalg.det(a) array([-2., -3., -8.]) zD->Dzd->d)rF)r) rrrrxrkrHrrAr)rerjrrrrRrRrSrs/  rcCsHddl}t|\}}t|\}}|jdk}|r>|ddtf}t||t|||jd}|jd}|jd} t||} ||jdkrtdt ||\} } t | } t | }t | | f| }| |d|jdd| f<t| ||\}}t||\}}t t||f|}tdt|tt||dd}t dt|||dt||ft}t| rNtj}d}t |f|}t |f| }|||| |||| ||d|d||d}tt|d}t |f|}t ||f|}t | | f|}t||| |||| ||d|d|d}t|d}t |f| }t |f|}|||| |||| ||d||||d}nrtj}d}t |f| }|||| |||| ||d|d|d}t|d}t |f| }|||| |||| ||d|||d}|d dkrtd tg| }|rntt|d|| d d }|d |kr||krt| rJtttt||ddg| d}n"ttt||ddg| d}ntt|d|ddf| d d }|d |kr||krt| rttt||dddfdddj| dd}n.tt||dddfdddj| dd}|dt||j| d d}|||||d |fS)a Return the least-squares solution to a linear matrix equation. Solves the equation `a x = b` by computing a vector `x` that minimizes the Euclidean 2-norm `|| b - a x ||^2`. The equation may be under-, well-, or over- determined (i.e., the number of linearly independent rows of `a` can be less than, equal to, or greater than its number of linearly independent columns). If `a` is square and of full rank, then `x` (but for round-off error) is the "exact" solution of the equation. Parameters ---------- a : (M, N) array_like "Coefficient" matrix. b : {(M,), (M, K)} array_like Ordinate or "dependent variable" values. If `b` is two-dimensional, the least-squares solution is calculated for each of the `K` columns of `b`. rcond : float, optional Cut-off ratio for small singular values of `a`. For the purposes of rank determination, singular values are treated as zero if they are smaller than `rcond` times the largest singular value of `a`. Returns ------- x : {(N,), (N, K)} ndarray Least-squares solution. If `b` is two-dimensional, the solutions are in the `K` columns of `x`. residuals : {(), (1,), (K,)} ndarray Sums of residuals; squared Euclidean 2-norm for each column in ``b - a*x``. If the rank of `a` is < N or M <= N, this is an empty array. If `b` is 1-dimensional, this is a (1,) shape array. Otherwise the shape is (K,). rank : int Rank of matrix `a`. s : (min(M, N),) ndarray Singular values of `a`. Raises ------ LinAlgError If computation does not converge. Notes ----- If `b` is a matrix, then all array results are returned as matrices. Examples -------- Fit a line, ``y = mx + c``, through some noisy data-points: >>> x = np.array([0, 1, 2, 3]) >>> y = np.array([-1, 0.2, 0.9, 2.1]) By examining the coefficients, we see that the line should have a gradient of roughly 1 and cut the y-axis at, more or less, -1. We can rewrite the line equation as ``y = Ap``, where ``A = [[x 1]]`` and ``p = [[m], [c]]``. Now use `lstsq` to solve for `p`: >>> A = np.vstack([x, np.ones(len(x))]).T >>> A array([[ 0., 1.], [ 1., 1.], [ 2., 1.], [ 3., 1.]]) >>> m, c = np.linalg.lstsq(A, y)[0] >>> print(m, c) 1.0 -0.95 Plot the data along with the fitted line: >>> import matplotlib.pyplot as plt >>> plt.plot(x, y, 'o', label='Original data', markersize=10) >>> plt.plot(x, m*x + c, 'r', label='Fitted line') >>> plt.legend() >>> plt.show() rNrUzIncompatible dimensionsg@ rrz,SVD did not converge in Linear Least SquaresT)rsrZrankr_)rs)rF)r)mathrhrr'rrrrrrxrorrrrrrrrlogfloat fortran_intrkrGZzgelsdr<Zdgelsdrr(r1rr)rerrrrrgZis_1drrZn_rhsZldbrjrZ result_real_trZbstarrZnlvlZiworkrrZrworkrrZa_realZ bstar_realZlrworkZresidsrstrRrRrSrsT          &(              " ( $ cCs>||kr|d8}tt|||j|d}|t|dddd}|S)aCompute a function of the singular values of the 2-D matrices in `x`. This is a private utility function used by numpy.linalg.norm(). Parameters ---------- x : ndarray row_axis, col_axis : int The axes of `x` that hold the 2-D matrices. op : callable This should be either numpy.amin or numpy.amax or numpy.sum. Returns ------- result : float or ndarray If `x` is 2-D, the return values is a float. Otherwise, it is an array with ``x.ndim - 2`` dimensions. The return values are either the minimum or maximum or sum of the singular values of the matrices, depending on whether `op` is `numpy.amin` or `numpy.amax` or `numpy.sum`. rUrr)r)r)r8rr)rrow_axiscol_axisopyresultrRrRrS_multi_svd_norms rFc CsPt|}t|jjttfs$|t}|dkr|j}|dksZ|dkrJ|dksZ|dkr|dkr|j dd}t |jjrt |j |j t |j |j }n t ||}t|}|r||dg}|S|j}|dkrtt|}n4t|ts y t|}WntdYnX|f}t|dkrn|tkr4t|j||dS|t krRt|j||dS|d krt|d ktj||dS|dkrtjt|||dS|dks|dkr||j }ttj|||dSy |dWntk rtd YnX|jjtkr t|} n>t |jjr|nt |} | j|jkr>t| } n t| | d | |C} tj| ||dd |Snt|dkrD|\} } t!| |} t!| |} | | krtd |dkrt"|| | t#}nH|dkrt"|| | t$}n,|dkr| | kr| d8} tjt|| dj| d}n|tkrR| | kr6| d8} tjt|| dj| d}n|dkr| | krn| d8} tjt|| dj| d}n|t kr| | kr| d8} tjt|| dj| d}nJ|dkrttj||j |d}n"|dkrt"|| | t}ntd|r@t%|j&} d| |d <d| |d<|| }|StddS)aY Matrix or vector norm. This function is able to return one of eight different matrix norms, or one of an infinite number of vector norms (described below), depending on the value of the ``ord`` parameter. Parameters ---------- x : array_like Input array. If `axis` is None, `x` must be 1-D or 2-D. ord : {non-zero int, inf, -inf, 'fro', 'nuc'}, optional Order of the norm (see table under ``Notes``). inf means numpy's `inf` object. axis : {int, 2-tuple of ints, None}, optional If `axis` is an integer, it specifies the axis of `x` along which to compute the vector norms. If `axis` is a 2-tuple, it specifies the axes that hold 2-D matrices, and the matrix norms of these matrices are computed. If `axis` is None then either a vector norm (when `x` is 1-D) or a matrix norm (when `x` is 2-D) is returned. keepdims : bool, optional If this is set to True, the axes which are normed over are left in the result as dimensions with size one. With this option the result will broadcast correctly against the original `x`. .. versionadded:: 1.10.0 Returns ------- n : float or ndarray Norm of the matrix or vector(s). Notes ----- For values of ``ord <= 0``, the result is, strictly speaking, not a mathematical 'norm', but it may still be useful for various numerical purposes. The following norms can be calculated: ===== ============================ ========================== ord norm for matrices norm for vectors ===== ============================ ========================== None Frobenius norm 2-norm 'fro' Frobenius norm -- 'nuc' nuclear norm -- inf max(sum(abs(x), axis=1)) max(abs(x)) -inf min(sum(abs(x), axis=1)) min(abs(x)) 0 -- sum(x != 0) 1 max(sum(abs(x), axis=0)) as below -1 min(sum(abs(x), axis=0)) as below 2 2-norm (largest sing. value) as below -2 smallest singular value as below other -- sum(abs(x)**ord)**(1./ord) ===== ============================ ========================== The Frobenius norm is given by [1]_: :math:`||A||_F = [\sum_{i,j} abs(a_{i,j})^2]^{1/2}` The nuclear norm is the sum of the singular values. References ---------- .. [1] G. H. Golub and C. F. Van Loan, *Matrix Computations*, Baltimore, MD, Johns Hopkins University Press, 1985, pg. 15 Examples -------- >>> from numpy import linalg as LA >>> a = np.arange(9) - 4 >>> a array([-4, -3, -2, -1, 0, 1, 2, 3, 4]) >>> b = a.reshape((3, 3)) >>> b array([[-4, -3, -2], [-1, 0, 1], [ 2, 3, 4]]) >>> LA.norm(a) 7.745966692414834 >>> LA.norm(b) 7.745966692414834 >>> LA.norm(b, 'fro') 7.745966692414834 >>> LA.norm(a, np.inf) 4.0 >>> LA.norm(b, np.inf) 9.0 >>> LA.norm(a, -np.inf) 0.0 >>> LA.norm(b, -np.inf) 2.0 >>> LA.norm(a, 1) 20.0 >>> LA.norm(b, 1) 7.0 >>> LA.norm(a, -1) -4.6566128774142013e-010 >>> LA.norm(b, -1) 6.0 >>> LA.norm(a, 2) 7.745966692414834 >>> LA.norm(b, 2) 7.3484692283495345 >>> LA.norm(a, -2) nan >>> LA.norm(b, -2) 1.8570331885190563e-016 >>> LA.norm(a, 3) 5.8480354764257312 >>> LA.norm(a, -3) nan Using the `axis` argument to compute vector norms: >>> c = np.array([[ 1, 2, 3], ... [-1, 1, 4]]) >>> LA.norm(c, axis=0) array([ 1.41421356, 2.23606798, 5. ]) >>> LA.norm(c, axis=1) array([ 3.74165739, 4.24264069]) >>> LA.norm(c, ord=1, axis=1) array([ 6., 6.]) Using the `axis` argument to compute matrix norms: >>> m = np.arange(8).reshape(2,2,2) >>> LA.norm(m, axis=(1,2)) array([ 3.74165739, 11.22497216]) >>> LA.norm(m[0, :, :]), LA.norm(m[1, :, :]) (3.7416573867739413, 11.224972160321824) N)rfror_rUK)orderz6'axis' must be None, an integer or a tuple of integers)rrrzInvalid norm order for vectors.)outg?zDuplicate axes given.r)rr)NrrZnucz Invalid norm order for matrices.z&Improper number of dimensions to norm.)'rrirsrtr%rBrrrr(rkr+rrr.rtupler isinstancerrur}r*r<rrr1r,rZconjrr7rFrDrr:r9r`r) rordrrrZsqnormr~ZndrZabsxrrZ ret_shaperRrRrSrs                                  cCst|}|dkrtdn|dkr4t|d|dSdd|D}|dj|dj}}|djdkrvt|d|d<|djdkrt|dj|d<t||dkrt|d|d|d}nt|}t ||d|d}|dkr|dkr|d S|dks|dkr| S|Sd S) aq Compute the dot product of two or more arrays in a single function call, while automatically selecting the fastest evaluation order. `multi_dot` chains `numpy.dot` and uses optimal parenthesization of the matrices [1]_ [2]_. Depending on the shapes of the matrices, this can speed up the multiplication a lot. If the first argument is 1-D it is treated as a row vector. If the last argument is 1-D it is treated as a column vector. The other arguments must be 2-D. Think of `multi_dot` as:: def multi_dot(arrays): return functools.reduce(np.dot, arrays) Parameters ---------- arrays : sequence of array_like If the first argument is 1-D it is treated as row vector. If the last argument is 1-D it is treated as column vector. The other arguments must be 2-D. Returns ------- output : ndarray Returns the dot product of the supplied arrays. See Also -------- dot : dot multiplication with two arguments. References ---------- .. [1] Cormen, "Introduction to Algorithms", Chapter 15.2, p. 370-378 .. [2] http://en.wikipedia.org/wiki/Matrix_chain_multiplication Examples -------- `multi_dot` allows you to write:: >>> from numpy.linalg import multi_dot >>> # Prepare some data >>> A = np.random.random(10000, 100) >>> B = np.random.random(100, 1000) >>> C = np.random.random(1000, 5) >>> D = np.random.random(5, 333) >>> # the actual dot multiplication >>> multi_dot([A, B, C, D]) instead of:: >>> np.dot(np.dot(np.dot(A, B), C), D) >>> # or >>> A.dot(B).dot(C).dot(D) Notes ----- The cost for a matrix multiplication can be calculated with the following function:: def cost(A, B): return A.shape[0] * A.shape[1] * B.shape[1] Let's assume we have three matrices :math:`A_{10x100}, B_{100x5}, C_{5x50}`. The costs for the two different parenthesizations are as follows:: cost((AB)C) = 10*100*5 + 10*5*50 = 5000 + 2500 = 7500 cost(A(BC)) = 10*100*50 + 100*5*50 = 50000 + 25000 = 75000 r_zExpecting at least two arrays.rrUcSsg|] }t|qSrR)r@).0rerRrRrS ) szmulti_dot..rr)rrN) r}rr+rr>rr_multi_dot_three_multi_dot_matrix_chain_order _multi_dotr()rwrZ ndim_firstZ ndim_lastrrrRrRrSrs*L c Cs`|j\}}|j\}}||||}||||}||krLtt|||St|t||SdS)z Find the best order for three arrays and do the multiplication. For three arguments `_multi_dot_three` is approximately 15 times faster than `_multi_dot_matrix_chain_order` N)rr+) ABCZa0Za1b0Zb1c0Zc1Zcost1Zcost2rRrRrSrE s  rc Cst|}dd|D|djdg}t||ftd}t||ftd}xtd|D]}xt||D]}||}t|||f<xvt||D]h} ||| f|| d|f|||| d||d} | |||fkr| |||f<| |||f<qWqdWqRW|r ||fS|S)a Return a np.array that encodes the optimal order of mutiplications. The optimal order array is then used by `_multi_dot()` to do the multiplication. Also return the cost matrix if `return_costs` is `True` The implementation CLOSELY follows Cormen, "Introduction to Algorithms", Chapter 15.2, p. 370-378. Note that Cormen uses 1-based indices. cost[i, j] = min([ cost[prefix] + cost[suffix] + cost_mult(prefix, suffix) for k in range(i, j)]) cSsg|]}|jdqS)r)r)rrerRrRrSrn sz1_multi_dot_matrix_chain_order..rrU)rs)r}rrr"rr?rr*) rwZ return_costsrrrrlrjrrrRrRrSrZ s < rcCsF||kr||Stt||||||ft|||||fd|SdS)z4Actually do the multiplication with the given order.rUN)r+r)rwrrrrRrRrSr sr)N)r_)r)r)r)rUrU)N)N)r)r)NNF)F)urQZ __future__rrr__all__rZ numpy.corerrrrrrr r!r"r#r$r%r&r'r(r)r*r+r,r-r.r/r0r1r2r3r4r5r6r7r8r9r:r;r<r=r>r?r@rArBrCZnumpy.core.multiarrayrDZ numpy.librErFZ numpy.linalgrGrHZnumpy.matrixlib.defmatrixrZ_NZ_VZ_AZ_SZ_Lr ExceptionrrVrXr[r\r]r^rcrhrkrlrprorqrrrxrrrrrrrrrrrrrr r rr r rrrrrrr rrrrrrrrrrRrRrRrS s          EP EFU XW [   y W W R[= , yo )