Paired Eigenvector Rotation Technique

• The PER technique is a rigorous method for mapping one eigenspace onto another using unitary transformations and spectral theory.
• It employs block-operator and polar decompositions to derive explicit rotation algorithms and sharp perturbative bounds.
• PER is vital for stabilizing eigenvector orientations in numerical computations, quantum mechanics, and dynamic principal component analysis.

The Paired Eigenvector Rotation (PER) technique provides a rigorous, explicit framework for mapping one eigenspace (or subspace) onto another, particularly in the context of perturbations, subspace tracking, and basis orientation problems. By leveraging spectral theory, block operator decomposition, and perturbation analysis, PER yields both practical rotation algorithms and sharp norm bounds for subspace movement in finite- and infinite-dimensional settings. The methodology is central for applications ranging from numerical linear algebra and quantum mechanics to statistical learning and time-evolving principal component analysis.

1. Definitions, Setup, and Motivation

Let $H$ be a real or complex Hilbert space. An operator $P\in B(H)$ is an orthogonal projection if $P^2=P$ and $P=P^*$. Given two such projections $P,Q\in B(H)$, their ranges, $\operatorname{Range}(P)$ and $\operatorname{Range}(Q)$, define the "before" and "after" subspaces, respectively. The aim is to construct a unitary $U$—an intertwining operator—such that $UPU^* = Q$.

Normalizations may be imposed for uniqueness or further structure, such as

$$PU = UQ, \quad UP = QU, \quad U^2 = (Q-Q^\perp)(P-P^\perp), \quad \mathrm{Re}\, U \ge 0,$$

with $Q^\perp=I-Q$, $P^\perp=I-P$.

Motivations for PER arise from:

• Numerical stability in eigenvector tracking where naive eigendecomposition yields sign or handedness flips, introducing discontinuities in time series of bases ($Q_t$) (Damask, 2024).
• Perturbation theory for understanding how spectral subspaces respond to changes or noise in the underlying operator or matrix, formalizing "how much" an eigenspace rotates (Allez et al., 2011).
• Quantum physics, statistics, and data science where the stability of principal subspaces (e.g., PCA) under perturbation/noise is fundamental.

2. Block-Operator Decomposition and Explicit Formulae

A central tool is Halmos’ block-operator decomposition (Dou et al., 2017). The Hilbert space decomposes into six mutually orthogonal subspaces according to intersections and differences of $P$ and $Q$:

• $$H_1 = \operatorname{Range}(P) \cap \operatorname{Range}(Q)$$
• $H_2 = \operatorname{Range}(P) \cap \ker Q$
• $H_3 = \ker P \cap \operatorname{Range}(Q)$
• $H_4 = \ker P \cap \ker Q$
• $H_5, H_6$: generic components with neither total overlap nor total separation

On $H_5\oplus H_6$, projections restrict as follows:

$$P|_{H_5\oplus H_6} = \begin{bmatrix} I_5 & 0 \ 0 & 0 \end{bmatrix}, \quad Q|_{H_5\oplus H_6} = \begin{bmatrix} Q_0 & Q_0^{1/2}(I_5-Q_0)^{1/2} D \ D^* (I_5-Q_0)^{1/2} Q_0^{1/2} & D^* (I_5-Q_0) D \end{bmatrix},$$

with $Q_0 : H_5 \to H_5$ a positive contraction (no eigenvalue $0$ or $1$), and $D: H_6 \to H_5$ a unitary.

The block-matrix form for $P$ and $Q$ thus reduces the general problem to explicit algebraic manipulations on each subspace.

3. Polar Decomposition: The PER Formula

The explicit rotation unitary is constructed via the polar decomposition of the operator

$A = QP + (I-Q)(I-P).$

$A$ is invertible iff $\|P-Q\|<1$. The polar decomposition

$A = U |A|$

yields a unitary $U$ defined by

$$U = \bigl[ QP + (I-Q)(I-P) \bigr] \, \Big| QP + (I-Q)(I-P) \Big|^{-1}.$$

This $U$ satisfies $UPU^*=Q$ and provides the celebrated PER formula. Each term has operational meaning: $QP$ projects vectors into $P$, then into $Q$; $(I-Q)(I-P)$ projects into the orthogonal complements. $|A| = (A^*A)^{1/2}$ ensures unitarity.

When $\|P-Q\|<1$, $A$ is invertible, yielding a unique direct rotation in generic settings. Block-diagonalization via Halmos’ decomposition allows explicit construction even in cases with common or unmatched subspaces (Dou et al., 2017).

4. Spectral Angles, Davis-Kahan Regime, and Perturbative Bounds

The angle $\Theta$ between $\operatorname{Range}(P)$ and $\operatorname{Range}(Q)$ is defined by

$$\cos\Theta = |PQP|^{1/2} \quad \text{on}~\operatorname{Range}(P).$$

The gap $\delta = \|P-Q\| = \sin\|\Theta\|_\infty$ quantifies subspace separation. When $Q$ is a small perturbation of $P$, the rotation operator reduces at leading order to the classical Davis–Kahan sin$\Theta$ result, and the infinitesimal rotation formula

$$U = \exp(iK),\qquad K = \frac{QP-PQ}{|P-Q|} + o(\|P-Q\|),$$

which matches the first-order expansion of the PER operator (Dou et al., 2017).

At the level of individual eigenpairs, as thoroughly analyzed in "Eigenvector dynamics: theory and some applications" (Allez et al., 2011), the rotation angle between $u_i$ and $u_j$ under a small perturbation $\varepsilon H_1$ is

$$\theta_{ij} \approx \varepsilon\,\frac{|\langle u_i|H_1|u_j\rangle|}{|\lambda_i-\lambda_j|},$$

assuming non-degeneracy and $\varepsilon \ll 1$. The overlap matrix $G_{ij} = \langle u_i(\varepsilon) | u_j \rangle$ provides a matrix characterization of subspace rotation.

5. Algorithmic Realization and Basis Orientation

Algorithmic variants of PER handle evolving orthonormal bases, with a primary focus on removing sign and handedness flips across time-indexed bases ($Q_t$) (Damask, 2024). The PER algorithm in this context involves:

1. Compute the transition matrix $M = Q_{t-1}^T Q_t \in O(N)$;
2. Apply a post-multiplicative diagonal reflection $S_t$ to ensure $\det(M S_t) = +1$ (so $M S_t \in SO(N)$);
3. Factor $M S_t$ into Givens rotations, extracting rotation angles via stable $\arctan2$ computations to maintain continuous orientation in $\mathrm{SO}(N)$ across time;
4. For time series, apply dynamic smoothing or static modal locking to stabilize principal directions and filter noise-driven modal drift.

This procedure replaces sign-flip and reflection ambiguity with continuous, consistently oriented eigenvectors, crucial in time-evolving multivariate analysis.

6. Extensions to Matrix Pairs and Generalized Eigenproblems

PER generalizes to the rotation of invariant subspaces in Hermitian-positive-definite matrix pairs $(A,B)$ under perturbations $(\Delta A,\Delta B)$ (Grubišić et al., 2010). For

$$A x = \lambda B x,\quad (A+\Delta A) \tilde x = \tilde\lambda(B+\Delta B)\tilde x,$$

the angle between reference and perturbed spectral subspaces is bounded by

$$\|\sin\Theta_B(\mathcal E,\widetilde{\mathcal E})\| \leq \frac{\|A^{-1/2}\Delta A A^{-1/2}\|}{\mathrm{RelGap}_A} + \frac{\|B^{-1/2}\Delta B B^{-1/2}\|}{\mathrm{RelGap}_B} \frac{1}{\sqrt{1-2\|B^{-1/2}\Delta B B^{-1/2}\|}},$$

with appropriately defined spectral gaps $\mathrm{RelGap}_A$ and $\mathrm{RelGap}_B$. This provides parameter-exact, norm-sharp estimates for subspace rotation and forms the analytic basis for robust eigensolver practices.

7. Special Cases, Complexity, and Practical Considerations

In the simplest nontrivial setting, $H=\mathbb{C}^2$, PER reduces to an explicit $2\times 2$ rotation matrix:

$P = \begin{bmatrix}1&0\0&0\end{bmatrix}, \quad Q = R(\phi) P R(-\phi),$

with $R(\phi)$ the usual rotation; PER constructs

$U = R(\phi),$

rotating the $x$-axis onto the target axis at angle $\phi$ (Dou et al., 2017).

Computational cost for full-basis rotations is $O(N^3)$ per step (Givens factorization), dominated by the cascade of plane rotations through $\mathrm{SO}(N)$ or by two generalized eigensolves for matrix pairs (Damask, 2024, Grubišić et al., 2010). Practical implementation is available via the thucyd Python package for the time-series orientation use case (Damask, 2024).

Significant limitations include breakdown near degenerate eigenvalues (level repulsion), where first-order formulas fail, and the necessity for small $\|\Delta A\|,\|\Delta B\|$ compared to spectral gaps for sharp bounds (Grubišić et al., 2010, Allez et al., 2011).

8. Connections and Theoretical Advances

PER formalizes and extends classical results by Kato, Davis–Kahan, and Avron–Seiler–Simon by providing fully explicit, blockwise algorithms and expressions that streamline analytical and computational work. The approach generalizes across infinite- and finite-dimensional Hilbert spaces, handles both generic and degenerate (non-generic) cases, and supports extensions to noise filtering in random matrix settings, perturbative analysis in quantum and statistical physics, and real-time signal processing.

Through direct block-operator construction and precise perturbation estimates, the PER technique establishes itself as a foundational tool for analyzing and controlling the geometry of eigenspaces under perturbations of varying magnitude and structure (Dou et al., 2017, Allez et al., 2011, Grubišić et al., 2010, Damask, 2024).