1 Introduction and Motivation

1.1 Why interpolate eigenvalues?

Eigenvalue interpolation studies how spectral quantities evolve as a matrix varies with an interpolation parameter. Given two matrices \(A_0\) and \(A_1\), one often constructs a continuous family \(A(t)\) for \(t\in[0,1]\) and asks how the associated eigenvalues \(\lambda_i(t)\) move. This viewpoint is useful because many problems naturally produce parameter-dependent operators—ranging from discretized differential equations to system models that change smoothly in time or design parameters.

A central motivation is that eigenvalues encode stability, resonance, and mode structure. If a system is represented by a matrix-valued model, then tracking how spectral data changes under interpolation can reveal when modes become dominant, when stability margins shrink, or when qualitative behavior changes.

1.2 Typical interpolation setups (linear, parameterized, and geometric)

Several standard constructions appear repeatedly:

  • Linear interpolation: \(A(t)=(1-t)A_0+tA_1\), which varies coefficients at constant speed in the ambient matrix space.
  • Parameterized interpolation: one applies a scalar reparameterization \(t=\phi(s)\) to control where the matrix moves more quickly or slowly along the path.
  • Geometric or similarity-based interpolation: instead of interpolating entries directly, one interpolates through transformations that preserve structural properties (such as symmetry, conditioning, or certain invariants), often using similarity transforms or unitary factors.

Each setup affects both the geometric path traced by the eigenvalues and the numerical ease with which eigenvalue branches can be followed.

1.3 Applications overview (control, dynamics, model reduction, numerical analysis)

Eigenvalue interpolation appears in diverse contexts:

  • Control and dynamics: parameter changes in feedback gains or physical parameters alter system matrices; eigenvalue tracking supports robustness assessment and design continuation.
  • Model reduction: reduced-order models depend on chosen parameters, and comparing spectral behavior across parameter settings helps validate reduced models.
  • Numerical analysis: iterative methods, continuation algorithms, and regularization schemes often generate families of matrices whose spectra must be monitored.

Although the mathematical details differ across fields, the shared task is to understand how eigenvalues move, whether they remain smooth, and what happens near spectral degeneracies.

2 Mathematical Preliminaries

2.1 Spectral basics (eigenvalues, eigenvectors, algebraic vs. geometric multiplicity)

For a matrix \(A\), eigenvalues are roots of the characteristic polynomial. When an eigenvalue \(\lambda\) has algebraic multiplicity \(m\), it corresponds to a factor \((\lambda-\lambda)^m\) in the polynomial. The geometric multiplicity is the dimension of the eigenspace \(\ker(A-\lambda I)\) and satisfies \(1\le \text{geom}(\lambda)\le m\).

In interpolation problems, multiplicities matter because they influence differentiability and the local structure of eigenvalue branches. In particular, a repeated eigenvalue may or may not split into distinct eigenvalues under perturbation, depending on additional properties of the matrix family.

2.2 Continuity and differentiability of eigenvalues

For general matrices, eigenvalues depend continuously on the entries of \(A(t)\) in the sense that small perturbations cannot create large spectral jumps. However, the assignment of individual eigenvalues to specific indices can fail to be continuous when eigenvalues cross or swap.

For normal matrices (e.g., Hermitian or unitary-normal families), more refined statements hold: eigenvalues can be chosen to depend continuously on \(t\) and, under additional non-degeneracy conditions, can be differentiable. For non-normal matrices, eigenvalues may still be continuous, yet eigenvectors can behave badly, and sensitivity can be much higher.

2.3 Matrix families \(A(t)\) and regularity assumptions

An eigenvalue interpolation problem starts with a parameter-dependent family \(A(t)\). Typical assumptions include:

  • Continuity or smoothness of \(A(t)\) in \(t\) (e.g., \(C^k\) regularity).
  • Structure preservation, such as symmetry (real symmetric) or Hermitian structure, which simplifies spectral behavior.
  • Controlled endpoints \(A(0)=A_0\), \(A(1)=A_1\), often with knowledge that these endpoints have desirable spectral properties (like simple eigenvalues).

These assumptions set expectations for whether eigenvalue curves are smooth, Lipschitz-like, or may exhibit irregular behavior near degeneracies.

3 Eigenvalue Curves and Branch Tracking

3.1 Eigenvalue labeling and ordering conventions

Because eigenvalues are not inherently labeled beyond the spectrum set, interpolation requires conventions to track them. Common approaches include:

  • Ordering by size for real spectra, e.g., for Hermitian \(A(t)\) one may label \(\lambda_1(t)\le \lambda_2(t)\le \cdots\).
  • Continuity-based labeling, where the index assignment is chosen to maintain small changes in \(\lambda_i(t)\) between successive parameter samples.
  • Projection- or overlap-based labeling, where one associates eigenvectors (or subspaces) via an inner-product criterion and then labels eigenvalues accordingly.

Different conventions can produce different index sequences even though the *set* of eigenvalues at each \(t\) is the same.

3.2 Handling eigenvalue crossings and near-degeneracies

When two eigenvalues become equal at some \(t^\*\), the behavior depends on whether the family is generic or special. In many contexts:

  • Exact crossings may occur when symmetry or structure enforces it.
  • Near-degeneracies may produce rapid changes that are numerically indistinguishable from crossings unless fine sampling is used.

Near such points, attempting to label eigenvalues solely by size can lead to apparent “swaps” of branches, while continuity-based labeling may require careful tolerance rules.

3.3 Concepts of spectral gaps and avoided crossings

A spectral gap is the separation between eigenvalues (or between groups of eigenvalues) across a parameter interval. Large gaps typically allow stable branch tracking because eigen-subspaces remain well-separated.

In settings where eigenvalues would cross under a simplified model but do not due to perturbations, one observes avoided crossings: eigenvalues come close and then move away, often accompanied by an exchange-like transfer in eigenvector character. Avoided crossings are particularly relevant for robust numerical continuation and for understanding how eigenvectors evolve even when eigenvalue magnitudes do not cross.

4 Theoretical Tools

4.1 Perturbation theory for eigenvalues

Perturbation theory analyzes how eigenvalues change when \(A\) is replaced by \(A+\varepsilon E\). For small \(\varepsilon\), eigenvalues can often be expanded in powers of \(\varepsilon\), at least when eigenvalues are simple.

For repeated eigenvalues, perturbation theory becomes more geometric: the effect of \(E\) depends on how it acts within the invariant subspace associated with the degenerate eigenvalue. In interpolation, one applies these results locally by treating \(A(t)\) as a perturbation of \(A(t_0)\).

4.2 First- and higher-order eigenvalue derivatives

If an eigenvalue branch \(\lambda(t)\) is differentiable and simple near \(t_0\), derivatives can be expressed using the corresponding eigenvectors. For Hermitian families, the first derivative admits a particularly simple form involving the Rayleigh quotient of \(A'(t)\).

Higher-order derivatives are available but depend on contributions from other spectral components. This matters in practice when eigenvalues change slowly or when higher-order terms govern curvature near near-degeneracy regions.

4.3 Eigenvalue sensitivity and conditioning

Eigenvalue sensitivity quantifies how much \(\lambda(t)\) can change under perturbations of \(A(t)\). Conditioning can deteriorate near degeneracies, especially for non-normal matrices or when eigenvalues coalesce.

In numerical interpolation, sensitivity informs step-size choice and error tolerances. Even if the eigenvalues themselves remain continuous, the mapping from matrix perturbations to eigenvalue changes can be steep, causing unstable tracking if sampling is too coarse.

4.4 Lipschitz/Hölder continuity bounds (qualitative discussion)

General results provide that eigenvalues are typically Hölder continuous in the matrix entries for broad classes of families, with stronger Lipschitz behavior under additional regularity and structural assumptions (such as normality and simplicity). Qualitatively, these bounds justify why fine parameter sampling improves accuracy, while also explaining why near degeneracies may still produce visually abrupt eigenvalue motion even when the underlying matrix path is smooth.

5 Specific Interpolation Models

5.1 Linear matrix interpolation \(A(t)=(1-t)A_0+tA_1\)

Linear interpolation is the simplest model: the matrix entries are convex combinations of endpoint entries. Its advantages include ease of implementation and straightforward interpretation as a constant-speed path in the space of matrices.

However, linear paths can behave poorly with respect to conditioning or spectral structure. If \(A(t)\) passes near defective behavior (for non-normal families) or near spectral degeneracies, eigenvalue branches can be difficult to track accurately without adaptive step refinement.

5.2 Affine combinations and parameter re-scaling

More general affine paths replace \(t\) by a shifted parameter: \[ A(t)=A_{\text{ref}} + t(A_1-A_{\text{ref}}) \] or incorporate scaling in the parameter domain. Reparameterization \(t=\phi(s)\) can distribute “effective distance” along the path, making eigenvalue changes more uniform in \(s\).

This is useful when endpoint-to-endpoint interpolation is too aggressive in certain regions. With a suitable \(\phi\), a fixed step size in the new parameter can yield more stable tracking across avoided crossings.

5.3 Unitary-/similarity-based interpolation approaches

Instead of blending matrix entries, one can interpolate through transformations: \[ A(t) = Q(t)^{-1} B Q(t) \] or through unitary/orthogonal factors in settings that preserve Hermitian structure. Similarity-based interpolation can keep eigenvalues closer to a controlled reference set if \(B\) is fixed and only the basis changes. When eigenvalues must change, these methods still provide structural advantages, such as maintaining symmetry and preventing numerical artifacts.

In computational practice, similarity-based paths often yield better-defined eigenvector evolution, though they may require additional effort to construct \(Q(t)\).

5.4 Interpolating via matrix functions (e.g., exponentials)

Matrix-function interpolation uses functional calculus to define a path with desired geometric features. A common example is an exponential-based curve: \[ A(t) = \exp((1-t)X_0 + tX_1) \] or variants that apply exponentials to generate positive-definite families.

Such constructions can preserve invariants like positivity or definiteness. Since many spectral properties are sensitive to definiteness and symmetry, function-based interpolation frequently improves robustness in applications such as dynamics and optimization.

6 Interpolation of Spectral Data

6.1 Interpolating eigenvalues vs. interpolating eigenvectors

Eigenvalues alone form an unordered set at each \(t\), whereas eigenvectors require a choice of phase (complex case) or sign (real case) and are not unique when eigenvalues repeat. Consequently:

  • Eigenvalue interpolation focuses on tracking scalar quantities \(\lambda_i(t)\), using ordering or continuity rules.
  • Eigenvector interpolation involves selecting consistent eigenvectors or bases across \(t\), typically via smoothness criteria or overlap maximization.

When eigenvalues approach degeneracy, eigenvectors can rotate rapidly even if eigenvalues change moderately. This makes eigenvector interpolation more delicate but also more informative about mode identity.

6.2 Reconstructing matrices from partial spectral information

Given some spectral data (e.g., a subset of eigenvalues and associated projectors, or constraints on a dominant spectral interval), one may attempt to construct a matrix family consistent with that information. This arises in inverse spectral problems and in reduced-order modeling where only partial information is retained.

Reconstruction is generally underdetermined: many matrices share the same partial spectral data. Regularization and additional structural constraints (symmetry, sparsity, definiteness) are often used to obtain a stable family.

6.3 Matching eigen-subspaces and spectral projectors

Rather than tracking individual eigenvectors, one can track eigen-subspaces. Spectral projectors associated with a cluster of eigenvalues can be defined via contour integrals (in theory) or via algebraic constructions (in computation). Subspace tracking can remain stable even when individual eigenvectors swap due to crossings, because the projector depends continuously on the parameter as long as the cluster remains isolated by a gap.

This approach is particularly relevant when multiple eigenvalues form a group, as it avoids ambiguity inherent in eigenvector labeling.

7 Numerical Methods and Algorithms

7.1 Direct eigen-computation along a grid in \(t\)

A basic strategy evaluates \(A(t_k)\) on a parameter grid and computes eigenpairs at each point. The computational burden grows with grid density, but the method is straightforward and supports diagnostics: one can visualize how eigenvalues change and detect near-degeneracies.

The main challenge is labeling across points. Since eigenvalue indices can swap near crossings, a post-processing step is needed to match eigenvalues (and possibly eigenvectors) from one grid point to the next.

7.2 Predictor–corrector strategies for eigenvalue tracking

Predictor–corrector methods build a local model of eigenvalue motion. One predicts \(\lambda_i(t_{k+1})\) (or an associated eigenpair) from derivatives or finite differences, then corrects by solving an eigenvalue problem or a constrained residual equation.

This can reduce the need for dense sampling when eigenvalue motion is smooth. Near degeneracies, predictors may lose accuracy, requiring adaptive step sizing and robust correction mechanisms.

7.3 Subspace iteration and projection methods

Subspace methods compute invariant subspaces more directly than full eigen-decompositions. By projecting \(A(t)\) onto a lower-dimensional subspace and iterating, one can approximate spectral clusters efficiently.

For eigenvalue interpolation, projection techniques can couple naturally with subspace tracking: one computes a basis for the relevant eigenspace at each \(t_k\) and updates it smoothly, improving stability when eigenvectors change rapidly.

7.4 Robustness considerations near degeneracies

Near crossings, the naive assumption that eigenvalues can be continuously labeled by size often breaks down. Robust approaches include:

  • detecting when eigenvalues are closer than a tolerance,
  • switching from eigenvalue tracking to cluster/subspace tracking,
  • using overlap measures to associate eigenvectors between steps.

Robustness is also influenced by arithmetic precision, scaling, and the conditioning of the eigenvalue problem. Practical algorithms often use normalization and gap-based adaptivity to prevent mislabeling.

8 Error Analysis and Stability

8.1 Discretization error in parameter sampling

When \(t\) is sampled on a grid, the interpolation error depends on how rapidly eigenvalues change with \(t\). Smooth eigenvalue branches permit accurate approximations with moderate grid density, while near avoided crossings the rate of change can increase, requiring finer steps.

Error estimation can be done empirically by comparing results at two grid resolutions, or theoretically using regularity assumptions such as Hölder or Lipschitz continuity.

8.2 Sensitivity to perturbations in matrix endpoints

Even if the interpolation path is fixed, uncertainties in \(A_0\) or \(A_1\) produce perturbations in \(A(t)\) throughout the interval. Sensitivity analysis determines how such uncertainties amplify in the eigenvalues.

The worst-case impact often occurs near points where eigenvalues become ill-conditioned with respect to perturbations, which is why stability checks should consider the smallest spectral gaps observed along the path.

8.3 Practical criteria for detecting branch swapping

Branch swapping is the numerical manifestation of labeling ambiguity near eigenvalue crossings. Common detection criteria include:

  • sudden changes in the estimated eigenvector overlap,
  • abrupt changes in eigenvalue derivatives inferred from finite differences,
  • inconsistencies between eigenvalue ordering rules across neighboring points.

Once detected, algorithms may reinitialize branch labels, switch to cluster tracking, or refine sampling locally.

9 Computational Examples (Conceptual)

9.1 Toy examples in low-dimensional matrices

Low-dimensional examples (such as \(2\times2\) or \(3\times3\) Hermitian matrices) illustrate how eigenvalues move under interpolation. These models allow one to compute eigenvalues explicitly and to observe how crossings depend on symmetry or parameter choices.

Such examples also clarify how eigenvalue ordering conventions behave when eigenvalues approach each other closely but do not truly coincide.

9.2 Illustrating eigenvalue crossings vs. avoided crossings

A conceptual demonstration compares two scenarios:

  • Crossing: eigenvalues coincide exactly at a parameter value, and eigenvectors span a consistent eigenspace structure.
  • Avoided crossing: eigenvalues approach and then separate due to a perturbation that breaks the structure enforcing a crossing.

Even without heavy computation, these cases show why branch tracking based solely on continuity can be misleading unless the relevant eigenspace is tracked robustly.

9.3 Comparing different interpolation strategies

Comparing linear interpolation with similarity-based or exponential-based interpolation highlights how the path affects spectral evolution. In some cases, linear interpolation may pass through regions where gaps shrink, causing rapid eigenvalue changes, while alternative constructions maintain healthier separations and smoother branch behavior.

Conceptual comparisons therefore emphasize that interpolation is not only a numerical convenience; it shapes the spectral geometry of the problem.

10.1 Spectral perturbation theory

Spectral perturbation theory provides the core analytical framework underlying eigenvalue interpolation. It supplies local expansions, stability estimates, and guidance on what happens near degeneracies.

10.2 Matrix functions and functional calculus

Functional calculus defines functions of matrices, enabling interpolation paths that preserve positivity, symmetry, or other invariants. Understanding matrix functions helps in designing spectral-preserving interpolation schemes.

10.3 Control of parameter-dependent systems

In control, eigenvalues determine stability and dynamic response. Parameter-dependent system matrices arise naturally when controllers or plant parameters vary, so eigenvalue interpolation supports continuation and robustness assessment.

10.4 Connections to interpolation in reduced-order modeling

Reduced-order modeling often interpolates or compares reduced operators across parameter settings. Eigenvalue interpolation helps evaluate whether the reduced model preserves modal structure, spectral gaps, and qualitative stability properties.