Numerical eigenvalue problems and residual certificates

Assume a diagonalizable matrix has one eigenvalue strictly largest in magnitude and that the starting vector has a nonzero component along its eigenvector. Repeated multiplication makes that component dominate relative to the others. Normalize at each step to avoid uncontrolled growth or underflow. The largest competing eigenvalue magnitude actually present in the initial vector controls the asymptotic directional error. A missing subdominant component can give faster convergence than the full spectral ratio suggests. A negative dominant eigenvalue can reverse the vector's sign each step without preventing convergence of its one-dimensional span.

If dominant magnitudes tie, the basic iteration need not approach one direction. If the starting vector has exactly zero component in the desired eigendirection, exact arithmetic never creates that component. A small spectral gap can make convergence frustratingly slow. Shift-and-invert iteration changes which eigenvalues dominate by solving systems with A minus a shift times the identity, rather than explicitly forming an inverse. That can target an interior eigenvalue, but it transfers work and conditioning concerns into the shifted linear solves.

Given a unit candidate vector, its Rayleigh quotient is a natural scalar estimate. For a real symmetric matrix it lies between the smallest and largest eigenvalues. Subtracting the quotient times the vector from its matrix image gives the eigen-residual. A zero residual means an exact eigenpair; a small residual means an exact eigenpair of a nearby matrix under a suitable perturbation. Neither statement alone identifies which eigenvalue was found or whether the eigenvector direction is accurate.

For symmetric matrices, orthonormal eigenvectors make a strong distance bound possible: the residual norm bounds distance from the candidate scalar to the spectrum. Accurate identification of a particular eigenvector also depends on separation from other eigenvalues. In a tight cluster, a whole invariant subspace may be well determined while individual basis vectors rotate substantially under tiny perturbations. Reporting a subspace can therefore be mathematically more meaningful than attaching excessive digits to a particular mode shape.

Dense eigensolvers use orthogonal or unitary transformations to reduce a matrix to a simpler form before iterative extraction. For general matrices, a Schur decomposition gives an orthonormal basis in which the operator is triangular; it does not demand a potentially ill-conditioned eigenvector basis. Symmetric problems admit stronger structure and specialized methods. Large sparse problems often use Lanczos or Arnoldi subspaces when only a few modes are needed, with reorthogonalization and restart choices affecting accuracy and cost.

Nonnormality means the matrix does not commute with its transpose or adjoint. Such matrices can have sensitive eigenvalues and transient growth that eigenvalues alone do not predict. A small residual can then signal proximity to the pseudospectrum rather than a comparably small distance to the original spectrum. Validate the matrix assembly, symmetry assumptions, normalization, and residuals in original coordinates. Numerical spectral output is evidence about a specified operator; it does not automatically validate the physical discretization or a continuum limit behind that operator.