jax.scipy.linalg.polar(a, side='right', *, method='qdwh', eps=None, max_iterations=None)[source]#

Computes the polar decomposition.

Given the \(m \times n\) matrix \(a\), returns the factors of the polar decomposition \(u\) (also \(m \times n\)) and \(p\) such that \(a = up\) (if side is "right"; \(p\) is \(n \times n\)) or \(a = pu\) (if side is "left"; \(p\) is \(m \times m\)), where \(p\) is positive semidefinite. If \(a\) is nonsingular, \(p\) is positive definite and the decomposition is unique. \(u\) has orthonormal columns unless \(n > m\), in which case it has orthonormal rows.

Writing the SVD of \(a\) as \(a = u_\mathit{svd} \cdot s_\mathit{svd} \cdot v^h_\mathit{svd}\), we have \(u = u_\mathit{svd} \cdot v^h_\mathit{svd}\). Thus the unitary factor \(u\) can be constructed as the application of the sign function to the singular values of \(a\); or, if \(a\) is Hermitian, the eigenvalues.

Several methods exist to compute the polar decomposition. Currently two are supported:

  • method="svd":

    Computes the SVD of \(a\) and then forms \(u = u_\mathit{svd} \cdot v^h_\mathit{svd}\).

  • method="qdwh":

    Applies the QDWH (QR-based Dynamically Weighted Halley) algorithm.

  • a (Union[Array, ndarray, bool_, number, bool, int, float, complex]) – The \(m \times n\) input matrix.

  • side (str) – Determines whether a right or left polar decomposition is computed. If side is "right" then \(a = up\). If side is "left" then \(a = pu\). The default is "right".

  • method (str) – Determines the algorithm used, as described above.

  • precisionPrecision object specifying the matmul precision.

  • eps (Optional[float]) – The final result will satisfy \(\left|x_k - x_{k-1}\right| < \left|x_k\right| (4\epsilon)^{\frac{1}{3}}\), where \(x_k\) are the QDWH iterates. Ignored if method is not "qdwh".

  • max_iterations (Optional[int]) – Iterations will terminate after this many steps even if the above is unsatisfied. Ignored if method is not "qdwh".

Return type:

Tuple[Array, Array]


A (unitary, posdef) tuple, where unitary is the unitary factor (\(m \times n\)), and posdef is the positive-semidefinite factor. posdef is either \(n \times n\) or \(m \times m\) depending on whether side is "right" or "left", respectively.