Polar-Light Retraction

• Polar-light retraction is a mapping on the Stiefel manifold that uniquely combines second-order accuracy with a closed-form inverse.
• The method employs matrix exponential and logarithm operations to efficiently approximate the Riemannian exponential map.
• Its computational efficiency and scalability make it ideal for replacing traditional geodesic computations in Riemannian optimization.

Polar-light retraction denotes the retraction mapping on the compact Stiefel manifold $St(n, p)$ introduced by Jensen and Zimmermann. It is distinguished by being both second-order accurate under the Euclidean metric and admitting a closed-form inverse. Retractions serve as efficient approximations to the Riemannian exponential map, underpinning the majority of Riemannian optimization and computing applications due to their computational tractability compared to exact geodesics. The polar-light (“PL”) retraction is, to date, the first Stiefel retraction to combine second-order accuracy with a closed-form inverse; previous retractions either lack the required order or their inverses are not directly computable.

1. Formal Definition

Let $\hat U \in St(n,p)$ be a basepoint and $\xi \in T_{\hat U} St(n,p)$ a tangent vector. The tangent vector admits a unique orthogonal decomposition: $$A = \hat U^T\,\xi \in \mathrm{skew}(p), \qquad B = (I-\hat U\hat U^T)\,\xi \in \mathbb{R}^{(n-p)\times p},$$ where $\xi = \hat U A + B$. The polar-light retraction at $\hat U$ is given by: $$R_{\hat U}^{\rm PL}(\xi) = \left(\hat U\,\exp_{m}(A) + (I-\hat U\hat U^T)\,\xi \right) \left(I_p + \xi^T (I - \hat U\hat U^T)\,\xi \right)^{-\frac12},$$ with $\exp_{m}$ denoting the matrix exponential restricted to skew-symmetric matrices ($SO(p)$). An algebraically identical form is: $$R_{\hat U}^{\rm PL}(\xi) = \left(\hat U[\exp_{m}(A)-A] + \xi \right) \left(I_p + \xi^T\xi + A^2\right)^{-\frac12}.$$ The latter form is advantageous in practice due to minor reductions in computational overhead (Jensen et al., 23 Feb 2026).

2. Local Approximation Order and Retractiveness

A fundamental requirement for a retraction is tangent consistency to first order with the Riemannian exponential: $\hat U \in St(n,p)$0 The polar-light retraction additionally achieves second-order accuracy under the Euclidean metric. Explicit Taylor expansion yields: $\hat U \in St(n,p)$1 identical through $\hat U \in St(n,p)$2 to the expansion of the true geodesic ($\hat U \in St(n,p)$3) on $\hat U \in St(n,p)$4. This distinguishes it from quasi-geodesic, QR, and Cayley retractions, which achieve only first-order approximation (Jensen et al., 23 Feb 2026).

3. Closed-Form Inverse Mapping

A central innovation is the closed-form local inverse of the polar-light retraction. For $\hat U \in St(n,p)$5 sufficiently close to $\hat U \in St(n,p)$6, the inverse mapping is: $\hat U \in St(n,p)$7 where $\hat U \in St(n,p)$8 denotes the matrix logarithm on $\hat U \in St(n,p)$9. In practice, computation proceeds via the thin SVD $\xi \in T_{\hat U} St(n,p)$0, with $\xi \in T_{\hat U} St(n,p)$1, so $\xi \in T_{\hat U} St(n,p)$2 and $\xi \in T_{\hat U} St(n,p)$3. This process requires only one $\xi \in T_{\hat U} St(n,p)$4 SVD and logarithm, plus a small number of $\xi \in T_{\hat U} St(n,p)$5 matrix-matrix multiplications, preserving computational efficiency (Jensen et al., 23 Feb 2026).

4. Comparison with Other Stiefel Retractions

The landscape of Stiefel retractions includes several canonical forms, encapsulated below:

Retraction	Order (Euclidean)	Closed-form Inverse	Computational Steps
Polar-light (PL)	Second	Yes	1 SVD + matrix exp/log
Polar-factor (PF)	Second	No	Lyapunov equation
QR	First	No	Thin QR
Cayley	First	No	Linear solve + skew op
Quasi-geodesic	First	Yes	Closed-form formula

The polar-light retraction uniquely combines second-order local accuracy (as with PF retraction) and a closed-form inverse (as with quasi-geodesic retractions), while computational requirements remain practical: both the map and its inverse scale as $\xi \in T_{\hat U} St(n,p)$6 dominated by small SVDs or exponentials/logarithms in $\xi \in T_{\hat U} St(n,p)$7, the frame width (Jensen et al., 23 Feb 2026).

5. Domain of Validity and Smoothness

All expressions presuppose the tangent vector $\xi \in T_{\hat U} St(n,p)$8 is sufficiently small such that

$\xi \in T_{\hat U} St(n,p)$9

Equivalently, the manifold chart remains valid in a neighborhood of $$A = \hat U^T\,\xi \in \mathrm{skew}(p), \qquad B = (I-\hat U\hat U^T)\,\xi \in \mathbb{R}^{(n-p)\times p},$$0 where the upper $$A = \hat U^T\,\xi \in \mathrm{skew}(p), \qquad B = (I-\hat U\hat U^T)\,\xi \in \mathbb{R}^{(n-p)\times p},$$1 block retains full rank. Within this domain, both the polar-light retraction and its inverse are smooth and well-defined (Jensen et al., 23 Feb 2026).

6. Implications and Applications

Polar-light retraction enables efficient, accurate, and invertible local modeling of the Stiefel manifold in Riemannian geometry, with direct advantages for optimization and simulation algorithms where iterative mapping between tangent vectors and manifold points is routine. Its balance of second-order accuracy and closed-form invertibility addresses limitations faced by alternative retractions, notably circumventing the need for Lyapunov/Sylvester equation solves (PF) or non-invertible chart transitions (QR/Cayley). As such, it is an optimal candidate for Riemannian-constrained problems demanding both accuracy and computational tractability (Jensen et al., 23 Feb 2026).