Exact SVD and Polar Decomposition

• Exact SVD and polar decomposition are methods that factor matrices into isometries and positive semidefinite parts, providing well-defined spectral forms.
• They utilize techniques such as Löwdin orthogonalizations, Riemannian gradient descent, and block reductions in semi-simple and Clifford algebras to ensure unique decomposition properties.
• These decompositions find practical applications in quantum computing, signal processing, and numerical linear algebra with robust convergence guarantees and high computational accuracy.

Exact singular value decomposition (SVD) and polar decomposition constitute central pillars in matrix analysis and multilinear algebra. These decompositions provide canonical forms for linear operators in both real/complex and noncommutative settings. The SVD factors a matrix as a product of two isometries and a positive semidefinite diagonal (or block-diagonal) matrix, while the polar decomposition generalizes the factorization of a linear operator as an isometry followed by a positive-semidefinite operator. Both decompositions have natural algebraic and geometric interpretations and admit explicit constructions via Löwdin orthogonalizations, optimization on Lie groups, and generalizations to *-algebras and Clifford geometric algebras.

1. Löwdin Orthogonalizations and Canonical Matrix Decompositions

Löwdin's symmetric and canonical orthogonalization procedures furnish a unified analytic framework for both the reduced SVD and polar decomposition of a full-rank matrix $A \in \mathbb{R}^{m \times n}$, $\operatorname{rank}(A)= n$ ($m \geq n$). The symmetric Löwdin orthogonalization is given by

$A_S = A (A^T A)^{-\frac{1}{2}},$

yielding columns that are orthonormal: $A_S^T A_S = I_n$. The canonical Löwdin orthogonalization uses the Gram spectral decomposition $A^T A = U \Lambda U^T$ with $$\Lambda = \operatorname{diag}(\lambda_1, \dots, \lambda_n)$$, $\lambda_i > 0$, $U \in O(n)$:

$A_C = A U \Lambda^{-\frac{1}{2}}.$

Both $\operatorname{rank}(A)= n$0 and $\operatorname{rank}(A)= n$1 form orthonormal bases of the column space of $\operatorname{rank}(A)= n$2.

The polar decomposition is then

$\operatorname{rank}(A)= n$3

with $\operatorname{rank}(A)= n$4 and $\operatorname{rank}(A)= n$5 symmetric positive definite. The reduced SVD,

$\operatorname{rank}(A)= n$6

arises naturally from the canonical Löwdin construction. Analytic relations connect these decompositions: $\operatorname{rank}(A)= n$7, $\operatorname{rank}(A)= n$8, making explicit the transformation between orthonormal bases and SVD factors (Naidu, 2011).

2. Exact Polar Decomposition: Optimization and Riemannian Geometry

For square matrices $\operatorname{rank}(A)= n$9, the exact polar decomposition $m \geq n$0 with $m \geq n$1 and $m \geq n$2 has a unique solution when $m \geq n$3 is invertible. The polar factor $m \geq n$4 can be characterized as

$m \geq n$5

with $m \geq n$6. Endowing $m \geq n$7 with its natural Riemannian metric, the gradient and Hessian of $m \geq n$8 can be computed explicitly, facilitating Riemannian-gradient-based algorithms. Despite the nonconvexity of $m \geq n$9, a geodesic weak-quasi-convexity and quadratic growth property ensure global landscape tractability: there are no spurious local minima.

Riemannian gradient descent (RGD) on $A_S = A (A^T A)^{-\frac{1}{2}},$0, with iterates

$A_S = A (A^T A)^{-\frac{1}{2}},$1

converges linearly in squared distance to $A_S = A (A^T A)^{-\frac{1}{2}},$2 for invertible $A_S = A (A^T A)^{-\frac{1}{2}},$3, at a rate dependent on the condition number $A_S = A (A^T A)^{-\frac{1}{2}},$4:

$A_S = A (A^T A)^{-\frac{1}{2}},$5

and algebraically in the singular case ($A_S = A (A^T A)^{-\frac{1}{2}},$6) (Alimisis et al., 2024).

Classical matrix SVD and Newton/Padé iterations remain $A_S = A (A^T A)^{-\frac{1}{2}},$7. RGD, requiring matrix exponentials per step, maintains this cost per iteration, yielding an iterative alternative particularly relevant for large-scale or distributed settings.

3. SVD and Polar Decomposition in Real Semi-simple *-Algebras

Every matrix over a finite-dimensional real semi-simple *-algebra $A_S = A (A^T A)^{-\frac{1}{2}},$8 admits an exact SVD and polar decomposition, generalizing the classical theory. The general SVD theorem asserts that for $A_S = A (A^T A)^{-\frac{1}{2}},$9, there exist unitary $A_S^T A_S = I_n$0, $A_S^T A_S = I_n$1, and a block-diagonal $A_S^T A_S = I_n$2 (with block entries in the centers of the simple components) so that $A_S^T A_S = I_n$3, where diagonal blocks generalize singular values (Ginzberg et al., 2015).

Two main computational strategies are described:

• Jacobi-Type Algorithms: Utilize generalized Givens rotations defined via “decent” involutive maps $A_S^T A_S = I_n$4, enabling the sequential zeroing of subdiagonal elements. Exact triangularization in finitely many steps is guaranteed in division algebra cases.
• Artin–Wedderburn Block Reduction: Decompose $A_S^T A_S = I_n$5 as an $A_S^T A_S = I_n$6-algebra into a direct sum of full matrix rings over $A_S^T A_S = I_n$7, $A_S^T A_S = I_n$8, or $A_S^T A_S = I_n$9. Extending this decomposition to matrix entries allows SVD and polar decomposition to be performed blockwise using the standard classical SVD. The polar decomposition then takes the form $A^T A = U \Lambda U^T$0, with $A^T A = U \Lambda U^T$1 unitary and $A^T A = U \Lambda U^T$2 Hermitian positive semidefinite.

The block approach is well suited for high-accuracy or symbolic computation and easily parallelizable.

4. SVD and Polar Decomposition in Clifford Geometric Algebras

In Clifford algebras $A^T A = U \Lambda U^T$3, a canonical exact SVD and polar decomposition are defined entirely within the geometry of multivectors. Hermitian conjugation $A^T A = U \Lambda U^T$4 is extended:

$A^T A = U \Lambda U^T$5

and the Euclidean product $A^T A = U \Lambda U^T$6 is positive definite. The unitary-orthogonal group in $A^T A = U \Lambda U^T$7 is

$A^T A = U \Lambda U^T$8

The SVD of an arbitrary multivector $A^T A = U \Lambda U^T$9 is realized by constructing a faithful matrix representation $$\Lambda = \operatorname{diag}(\lambda_1, \dots, \lambda_n)$$0 with $$\Lambda = \operatorname{diag}(\lambda_1, \dots, \lambda_n)$$1. Classical SVD on the image, combined with the algebraic isomorphism, yields

$$\Lambda = \operatorname{diag}(\lambda_1, \dots, \lambda_n)$$2

with $$\Lambda = \operatorname{diag}(\lambda_1, \dots, \lambda_n)$$3 and $$\Lambda = \operatorname{diag}(\lambda_1, \dots, \lambda_n)$$4 a positive linear combination of pairwise orthogonal idempotents (the "diagonal blades").

The polar decomposition follows as

$$\Lambda = \operatorname{diag}(\lambda_1, \dots, \lambda_n)$$5

Equivalently, factor $$\Lambda = \operatorname{diag}(\lambda_1, \dots, \lambda_n)$$6, set $$\Lambda = \operatorname{diag}(\lambda_1, \dots, \lambda_n)$$7, $$\Lambda = \operatorname{diag}(\lambda_1, \dots, \lambda_n)$$8. These decompositions remain valid in complexified Clifford algebras, with conjugation extended to the scalars (Shirokov, 2024).

Algorithmic descriptions in geometric algebra solely utilize multivector operations: Hermitian conjugation, GA-spectral factorization, projectors, and grade projections. Complexity remains comparable—$$\Lambda = \operatorname{diag}(\lambda_1, \dots, \lambda_n)$$9—to that of standard matrix factorizations, respecting the algebraic dimension.

5. Analytic and Algebraic Relationships Between Decompositions

Löwdin symmetric and canonical orthogonalizations provide the algebraic bridge between polar decomposition and SVD. Specifically, for $\lambda_i > 0$0 full column rank,

• Symmetric Löwdin yields the polar decomposition $\lambda_i > 0$1 with $\lambda_i > 0$2, $\lambda_i > 0$3.
• Canonical Löwdin gives the reduced SVD, $\lambda_i > 0$4, with $\lambda_i > 0$5, $\lambda_i > 0$6 from the eigendecomposition of $\lambda_i > 0$7.

The relationships

$\lambda_i > 0$8

and

$\lambda_i > 0$9

explicitly connect the polar and SVD factors.

In Clifford and real semi-simple *-algebras, the pullback from faithful matrix representations and the use of central block-diagonals or idempotent projectors make these relationships explicit and algorithmically consistent across different algebraic frameworks (Naidu, 2011, Ginzberg et al., 2015, Shirokov, 2024).

6. Computational and Practical Considerations

A summary of computational approaches:

Approach/Setting	Main Algorithmic Techniques	Complexity
Classical real, complex matrices	SVD, polar via SVD, Newton–Padé	$U \in O(n)$0
Riemannian gradient descent ($U \in O(n)$1)	Iterative optimization on $U \in O(n)$2	$U \in O(n)$3 per iteration
Semi-simple real *-algebras	Jacobi–Givens, Artin–Wedderburn block	$U \in O(n)$4–$U \in O(n)$5
Clifford geometric algebras	Matrix embedding / GA-spectral	$U \in O(n)$6

Classical SVD and block algorithms achieve high-accuracy and are suitable for symbolic computation. Iterative methods via Riemannian optimization offer scalability and robust convergence properties with theoretical guarantees—linear for full-rank, algebraic for singular cases (Alimisis et al., 2024). In noncommutative and geometric algebra contexts, exact decompositions preserve native algebraic structure, with cost scaling dictated by the underlying faithful representations.

Caution is required in floating-point arithmetic to maintain stability and preserve structural properties such as unitarity and positive semidefiniteness. Pre/post-normalization and adaptive line/spectral scaling are commonly applied (Ginzberg et al., 2015).

7. Extensions and Applications

The exact SVD and polar decomposition framework extends to distributed optimization, quantized algorithms, and blockwise parallel computation in structured algebras. In Clifford and geometric algebra, these decompositions underpin structure-preserving algorithms in quantum computing, signal processing, and theoretical physics. The ability to work natively in geometric algebra or over general *-algebras allows for algebra-intrinsic treatments of spectral theory, higher-dimensional rotations, and canonical transformations, linking linear algebraic decompositions with group and module actions (Ginzberg et al., 2015, Shirokov, 2024).

This suggests a universal applicability of SVD/polar paradigms across associative algebraic systems, with technical realization dependent on the algebra's decomposition into simple components and the explicit construction of orthogonal/unitary group actions.