Cayley-Newton Iteration on Matrix Manifolds

• Cayley-Newton iteration is a class of Newton-type methods that uses the Cayley transform for efficient optimization on matrix manifolds and solving inverse eigenvalue problems.
• The method pulls back nonlinear equations from Lie groups to Lie algebras, enabling standard Newton updates in a local coordinate chart that preserves quadratic convergence.
• Its application to symmetric inverse eigenvalue problems and implicit dynamical maps demonstrates robust global convergence and computational efficiency with iteration counts as low as 7–10.

The Cayley-Newton iteration is a class of Newton-type methods in which the step update is constructed via the Cayley transform, enabling efficient solutions to nonlinear problems defined on matrix manifolds, including the symmetric inverse eigenvalue problem as well as non-invertible implicit dynamical maps. This approach exploits the local geometry of matrix groups and their Lie algebras, and, through parametrizations such as the Cayley map, achieves computational and convergence advantages over traditional Newton or explicit schemes.

1. Cayley Transform and Matrix Manifold Parametrization

The Cayley transform is a rational map between a matrix Lie algebra $\mathfrak{g}$ and a neighborhood of the identity in a matrix Lie group $G$, typically realized as

$c(X) = (I + \tfrac12 X)(I - \tfrac12 X)^{-1}$

for $X$ sufficiently small. Its local inverse is

$\psi(Y) = 2(Y - I)(Y + I)^{-1}$

for $Y$ near the identity. The Cayley chart provides a $C^2$-diffeomorphic coordinate system on $G$ about $I$, allowing one to pull back problems defined on $G$ to the linear space $G$0, apply standard Newton updates, and then push forward the iterates to the group (Manton, 2012).

This parametrization is particularly valuable when the nonlinear equations or cost functions are invariant under or defined on matrix groups, as in certain inverse eigenvalue problems or optimization problems on Stiefel or orthogonal manifolds.

2. Cayley-Newton Iteration for the Inverse Eigenvalue Problem

The inexact Cayley-Newton method of Ling and Xu (Ling et al., 2013) targets the symmetric inverse eigenvalue problem (IEP):

• Given real symmetric matrices $G$1 and target eigenvalues $G$2,
• Define $G$3 for $G$4.
• Seek $G$5 such that the ordered eigenvalues satisfy $G$6 for all $G$7.

At each iteration $G$8:

• The current modal matrix approximation $G$9 approximates $c(X) = (I + \tfrac12 X)(I - \tfrac12 X)^{-1}$0, the eigenvector matrix of $c(X) = (I + \tfrac12 X)(I - \tfrac12 X)^{-1}$1.
• After computing a Newton direction $c(X) = (I + \tfrac12 X)(I - \tfrac12 X)^{-1}$2 by (possibly inexactly) solving $c(X) = (I + \tfrac12 X)(I - \tfrac12 X)^{-1}$3, where $c(X) = (I + \tfrac12 X)(I - \tfrac12 X)^{-1}$4, a Cayley-type update is applied to the modal matrix.
• The skew-symmetric generator $c(X) = (I + \tfrac12 X)(I - \tfrac12 X)^{-1}$5 is computed so that $c(X) = (I + \tfrac12 X)(I - \tfrac12 X)^{-1}$6 with $c(X) = (I + \tfrac12 X)(I - \tfrac12 X)^{-1}$7, choosing $c(X) = (I + \tfrac12 X)(I - \tfrac12 X)^{-1}$8 to align off-diagonals of $c(X) = (I + \tfrac12 X)(I - \tfrac12 X)^{-1}$9 to second-order.
• Backtracking with an Armijo-type line search on the Rayleigh-quotient residual ensures global convergence.
• The iteration continues until the eigen-residual $X$0 is below a specified threshold (Ling et al., 2013).

3. The Generalized Newton-Cayley Framework on Manifolds

The Cayley-Newton procedure extends Newton's method from Euclidean space to manifolds using local charts:

• One pulls back the cost $X$1 to $X$2.
• Computes the Newton step in the linearized algebra: $X$3.
• Pushes forward to the group: $X$4.
• The group iterate law is $X$5.

The Cayley chart guarantees the preservation of local quadratic convergence, provided the Euclidean Newton conditions are satisfied in the charted coordinates (Manton, 2012). This approach encompasses all coordinate-independent Newton methods on Lie groups.

4. Implicit Cayley–Newton Maps in Dynamical Systems

In discrete dynamical systems, the Cayley-Newton paradigm also appears in the form of implicit (multi-valued) maps resulting from semi-implicit, Cayley-type discretizations of Newton flows. For the cubic equation $X$6, the semi-implicit Euler method with parameters $X$7 and $X$8 produces the implicit Cayley–Newton map: $X$9 This map defines a three-valued correspondence in both forward and backward time, resulting in non-invertible and fractal dynamical behavior. The complex structure of its Julia set and attractors interpolates between classical root-finding fractals, analytic circles, “fat fractals,” and dust-like repellers depending on the choice of parameters $\psi(Y) = 2(Y - I)(Y + I)^{-1}$0 and $\psi(Y) = 2(Y - I)(Y + I)^{-1}$1 (Elistratov et al., 2022).

5. Convergence Theory and Globalization Strategies

For the symmetric IEP, under the assumptions that the target spectrum is simple and that the Jacobian $\psi(Y) = 2(Y - I)(Y + I)^{-1}$2 is nonsingular, the inexact Cayley-Newton scheme satisfies:

• Global convergence from arbitrary initial $\psi(Y) = 2(Y - I)(Y + I)^{-1}$3 with Rayleigh-quotient–based line search.
• Ultimately superlinear or quadratic convergence if the forcing term in the inexact Newton step $\psi(Y) = 2(Y - I)(Y + I)^{-1}$4 decays proportionally to the residual.
• The error satisfies

$\psi(Y) = 2(Y - I)(Y + I)^{-1}$5

with constants $\psi(Y) = 2(Y - I)(Y + I)^{-1}$6, $\psi(Y) = 2(Y - I)(Y + I)^{-1}$7. If $\psi(Y) = 2(Y - I)(Y + I)^{-1}$8 rapidly, local $\psi(Y) = 2(Y - I)(Y + I)^{-1}$9-quadratic convergence is achieved (Ling et al., 2013).

In the context of implicit Cayley–Newton dynamical maps, formal convergence is replaced by the study of invariant sets, fractal basin boundaries (Julia sets), and bifurcation phenomena as discretization parameters vary (Elistratov et al., 2022).

6. Computational Complexity and Numerical Performance

For the IEP, the dominant costs per iteration are:

• Formation of the Jacobian $Y$0, requiring $Y$1.
• The update of the modal matrix via the Cayley transform, $Y$2 per matrix multiplication.
• Solution of the $Y$3 linear system in the inexact Newton step.
• No full eigen-decomposition is required at each step, only Rayleigh-quotient evaluations and updates to the approximate eigenspace.

Numerical tests demonstrate that, even for $Y$4, robust global convergence is observed from distant initial guesses, with convergence in 7–10 iterations and final eigen-residuals below $Y$5 (Ling et al., 2013).

In Newton-Cayley maps for complex dynamics, the computational load stems from solving multivalued algebraic constraints at each step and from the enumeration of branches for fractal structure analysis (Elistratov et al., 2022).

7. Broader Context and Applications

The Cayley-Newton iteration connects classical numerical linear algebra, nonlinear manifold optimization, and discrete dynamical systems:

• It provides a systematic framework for solving inverse eigenvalue and other structured nonlinear problems where constraints are naturally expressed on matrix groups or manifolds.
• Its manifold viewpoint, via the Cayley transform, ensures the preservation of local convergence rates when generalizing from Euclidean to manifold settings (Manton, 2012).
• The implicit Cayley–Newton map extends root-finding iteration into the domain of non-invertible and multivalued dynamical systems, generating parameter-sensitive Julia sets and mixed dissipative–Hamiltonian phenomena (Elistratov et al., 2022).

This synthesis of geometric structure, Newton-type updates, globalization, and dynamical complexity underlies the significance of the Cayley-Newton approach in numerical analysis and dynamical systems theory.