1 Spectral Radius: Definition and Basic Properties

1.1 Eigenvalues and the Spectrum

For a linear map \(T\) on a finite-dimensional vector space, an eigenvalue \(\lambda\) is a scalar for which there exists a nonzero vector \(v\) satisfying \(Tv=\lambda v\). The collection of all eigenvalues is called the spectrum (more precisely, for matrices it is the set of eigenvalues; in infinite-dimensional settings, the spectrum is typically defined through invertibility properties of \(T-\lambda I\)). The eigenvalues summarize fundamental modes of the operator’s action and often determine long-term behavior of iterated dynamics.

1.2 Definition of Spectral Radius for Matrices

For a square matrix \(A\in \mathbb{C}^{n\times n}\), the spectral radius is \[

\rho(A)=\max\{\lambda:\lambda \text{ is an eigenvalue of }A\}.

\] Thus, \(\rho(A)\) is the largest magnitude among all eigenvalues, capturing the dominant exponential rate that eigenmodes can contribute under repeated multiplication.

1.3 Spectral Radius for Linear Operators

For a bounded linear operator \(T\) on a complex Banach space, \(\rho(T)\) is defined by the spectral radius formula \[

\rho(T)=\sup\{\lambda:\lambda\in\sigma(T)\},

\] where \(\sigma(T)\) denotes the spectrum of \(T\). In this framework, eigenvalues are included in the spectrum but may not exhaust it; the spectrum can contain points for which \(T-\lambda I\) fails to have a bounded inverse.

1.4 Norm Bounds and Basic Inequalities

Spectral radius is controlled by operator norms. For any matrix norm \(\|\cdot\|\),

\[

\rho(A)\le \|A\|.

\]

More strongly, for powers one has \(\rho(A)\le \limsup_{k\to\infty}\|A^k\|^{1/k}\), with equality under suitable limits described by Gelfand’s formula (see Section 3.3). These inequalities connect the eigenvalue-based quantity \(\rho(A)\) to measurable growth of \(A^k\).

1.5 Invariance under Similarity Transformations

If \(A\) and \(B\) are similar, meaning \(B=S^{-1}AS\) for some invertible \(S\), then they share the same eigenvalues. Consequently, \[ \rho(B)=\rho(A). \] This invariance makes the spectral radius a property of the linear transformation itself rather than of a particular choice of coordinates.

2 Eigenvalues: Theory and Computation

2.1 Eigenvalues, Eigenvectors, and Eigenspaces

An eigenvector \(v\neq 0\) for \(\lambda\) satisfies \((A-\lambda I)v=0\). The set of all eigenvectors together with \(0\) forms the eigenspace \[ E_\lambda=\ker(A-\lambda I). \] Eigenspaces can vary in dimension and structure, leading to different notions of multiplicity.

2.1.1 Algebraic Multiplicity

The algebraic multiplicity of \(\lambda\) is its multiplicity as a root of the characteristic polynomial \(\det(A-\lambda I)\). Algebraic multiplicity counts how many times the eigenvalue appears among factors of the characteristic polynomial.

2.1.2 Geometric Multiplicity

The geometric multiplicity is \(\dim(E_\lambda)\), the dimension of the eigenspace. It measures the number of linearly independent eigenvectors associated with \(\lambda\).

In general, geometric multiplicity does not exceed algebraic multiplicity, and equality characterizes a form of “simplicity” for the corresponding eigenspace.

2.2 Characteristic Polynomial and Root Interpretation

The characteristic polynomial \[ p_A(\lambda)=\det(\lambda I-A) \] encodes eigenvalues as its roots. In matrix theory, this polynomial is central for connecting eigenvalue computation to elimination theory and to the algebraic structure of \(A\). Roots can be repeated, and their multiplicities influence how the operator behaves near those eigenvalues under iteration.

2.3 Minimal Polynomial and Its Consequences

The minimal polynomial \(m_A(\lambda)\) is the monic polynomial of smallest degree such that \(m_A(A)=0\). It divides the characteristic polynomial and provides finer information than eigenvalues alone. In particular, the degrees of factors in \(m_A\) indicate how many levels of generalized eigenvectors are required, shaping the size of Jordan blocks and thus the growth behavior of \(A^k\).

2.4 Eigenvalue Localization and Bounds

When exact eigenvalues are hard to obtain, one often seeks regions of the complex plane guaranteed to contain the spectrum. Localization results can use entrywise information, matrix structure, or norms. Such tools are useful in estimating \(\rho(A)\) without full diagonalization.

2.5 Numerical Methods for Eigenvalue Approximation

In computation, eigenvalues are approximated using algorithms such as the QR method and variants for structured matrices. Practical approaches often balance stability, speed, and accuracy, while recognizing that sensitivity can occur when eigenvalues are clustered. Once eigenvalues are approximated, \(\rho(A)\) is typically estimated by the largest computed modulus, sometimes refined by bounds or residual checks.

3 Relationships Between Spectral Radius and Matrix Powers

3.1 Spectral Radius Formula

A foundational link between eigenvalues and iterates is given by \[

\rho(A)=\max\{\lambda:\lambda\in\sigma(A)\}.

\] Beyond this definition, spectral radius also controls the asymptotic size of \(A^k\), as expressed by formulas in the next sections.

3.2 Growth of Iterates \(A^k\)

The behavior of \(A^k\) as \(k\to\infty\) can be summarized by how eigenvalues and Jordan structure interact. In many settings, the leading growth rate is dictated by eigenvalues of maximal modulus, while polynomial factors (arising from nontrivial Jordan blocks) can modify the pure exponential rate. Hence \(\rho(A)\) serves as the “rate exponent” and Jordan data can add additional scaling.

3.3 Gelfand’s Formula for Operators

For a bounded operator \(T\) on a Banach space, \[

\rho(T)=\lim_{k\to\infty}\|T^k\|^{1/k},

\] where the limit exists in the extended sense and equals the spectral radius. In finite-dimensional matrix settings, this provides a powerful conceptual bridge: the eigenvalue-based quantity equals the asymptotic norm growth rate of powers.

3.4 Implications for Stability (Iteration/Recurrence)

In discrete-time linear systems, iteration often takes the form \(x_{k+1}=Ax_k\). If \(\rho(A)<1\), then iterates typically decay to zero (often with an exponential rate in appropriate senses). If \(\rho(A)>1\), growth modes exist and instability is expected. The borderline case \(\rho(A)=1\) is more delicate and depends on Jordan structure and norm growth, not only on the magnitude of eigenvalues.

For maps of the form \(x\mapsto Ax+b\), convergence of the fixed-point iteration \(x_{k+1}=Ax_k+b\) is closely tied to \(\rho(A)\). When \(\rho(A)<1\), the Neumann series for \((I-A)^{-1}\) converges, enabling explicit expressions for the limit and error decay. Thus, spectral radius functions as the criterion distinguishing convergent versus divergent iterative behavior.

4 Special Classes of Matrices and Their Eigenstructure

4.1 Normal, Unitary, and Hermitian Matrices

A matrix \(A\) is normal if \(A^*A=AA^*\). Normal matrices admit an orthonormal eigenbasis, making them unitarily diagonalizable. For Hermitian matrices (\(A=A^*\)), all eigenvalues are real, and the spectral radius becomes \(\rho(A)=\max\{\lambda:\lambda\in\mathbb{R}\}\). For unitary matrices, eigenvalues lie on the complex unit circle, so \(\rho(A)=1\).

Because these classes are diagonalizable with well-behaved eigenvectors, growth and norm estimates often align tightly with \(\rho(A)\).

4.2 Real Symmetric and Positive (Semi)Definite Cases

A real symmetric matrix is Hermitian over \(\mathbb{C}\), hence diagonalizable with real eigenvalues. A matrix is positive semidefinite if \(x^TAx\ge 0\) for all vectors \(x\); in that case eigenvalues are nonnegative and \(\rho(A)=\lambda_{\max}(A)\). For positive definite matrices, the smallest eigenvalue is strictly positive, and invertibility follows directly from spectral data.

4.3 Nonnegative Matrices and Order Preservation

A matrix is nonnegative if all its entries are \(\ge 0\). Such matrices interact with order and positivity properties of linear maps. Under additional irreducibility conditions, the Perron–Frobenius theory implies the existence of a distinguished eigenvalue equal to \(\rho(A)\) with a corresponding nonnegative eigenvector. This makes the spectral radius particularly meaningful in models where states represent quantities that cannot be negative.

4.4 Companion Matrices and Structured Eigenvalues

A companion matrix encodes the coefficients of a polynomial so that the eigenvalues of the matrix match the roots of that polynomial. This provides a structured way to translate between polynomial root problems and linear algebra. Since \(\rho(A)\) depends directly on eigenvalues, it also translates polynomial root magnitude into an operator growth measure.

4.5 Diagonalizable vs. Defective Matrices

A matrix is diagonalizable if it has a basis of eigenvectors, which occurs when its Jordan form has only size-one blocks. Defective matrices lack enough eigenvectors, leading to nontrivial Jordan blocks. For spectral radius alone, the exponential rate is still controlled by eigenvalues of maximal modulus, but defective matrices can introduce polynomial growth factors that alter exact asymptotics of \(\|A^k\|\).

5 Jordan Form, Decomposition, and Spectral Radius

5.1 Jordan Canonical Form (Overview)

Over the complex field, every square matrix is similar to a Jordan matrix consisting of Jordan blocks along the diagonal. Each Jordan block corresponds to an eigenvalue and has a particular size determining how many times generalized eigenvectors chain together. This decomposition refines the spectral picture by revealing not only which eigenvalues occur but also how they influence iteration.

5.2 Jordan Blocks and Their Growth Behavior

A Jordan block associated with eigenvalue \(\lambda\) of size \(m\) contributes terms to \(A^k\) that scale like \(\lambda^k\) multiplied by a polynomial in \(k\) of degree at most \(m-1\). Therefore, while \(\rho(A)\) dictates the dominant exponential factor, Jordan block sizes specify how large the polynomial prefactors can become.

5.3 Generalized Eigenvectors

A generalized eigenvector for \(\lambda\) satisfies \((A-\lambda I)^r v=0\) for some positive integer \(r\). The set of generalized eigenvectors forms generalized eigenspaces that assemble into the Jordan decomposition. These vectors are essential for describing the full operator action when \(A\) is not diagonalizable.

5.4 Semisimple vs. Non-Semisimple Spectral Effects

An eigenvalue is semisimple if its algebraic and geometric multiplicities coincide, meaning there are no nontrivial Jordan blocks for that eigenvalue. When the eigenvalues of maximal modulus are semisimple, iteration often behaves close to pure exponential decay or growth. If they are non-semisimple, polynomial corrections can slow convergence even when \(\rho(A)&lt;1\) and can accelerate or complicate growth when \(\rho(A)\ge 1\).

5.5 Computing Spectral Radius from Jordan Data

Given a Jordan form, the eigenvalues are read directly from the diagonal entries of the Jordan blocks. Hence \(\rho(A)\) is obtained as the maximum modulus among these eigenvalues. The Jordan data also clarifies which eigenvalues dominate iteration and whether polynomial factors accompany their contribution, enabling sharper asymptotic descriptions of \(A^k\).

6 Spectral Theorems and Operator-Theoretic Views

6.1 Spectral Decomposition for Diagonalizable Operators

For diagonalizable operators (finite-dimensional diagonalizable matrices or, in some settings, operators admitting a basis decomposition), the action can be written in terms of eigenprojections. In such cases, growth and norms of iterates can be estimated by combining contributions from each spectral component. The spectral radius corresponds to the spectral component(s) with largest modulus.

6.2 Functional Calculus High-Level

Functional calculus allows one to define expressions like \(f(A)\) using the spectral information of \(A\), typically through polynomials, power series, or contour integrals. Spectral radius enters naturally when bounding \(\|f(A)\|\) and when understanding how operations on spectra (such as applying \(f\) to eigenvalues) change the dominant magnitude.

6.3 Resolvent and Spectrum of Operators

For operators \(T\), the resolvent set consists of \(\lambda\) values where \(T-\lambda I\) is bijective with a bounded inverse. The complement is the spectrum \(\sigma(T)\). This framework provides geometric and analytic tools for studying spectral radius: \(\rho(T)\) is the largest modulus in the spectral set, even when eigenvalues are not explicitly known.

6.4 Spectral Measures for Self-Adjoint Operators

Self-adjoint operators (in Hilbert space settings) admit a spectral theorem that represents them through an integral with respect to a spectral measure. The spectrum lies on the real line, and spectral radius relates to the extreme points of that spectrum. This makes \(\rho(T)\) interpretable in terms of operator bounds and quadratic form behavior.

6.5 Compact Operators and Eigenvalue Accumulation

For compact operators, the spectrum is discrete except possibly for accumulation at \(0\). Nonzero spectral values are eigenvalues of finite multiplicity. In such cases, the spectral radius equals the largest modulus among nonzero eigenvalues, and the structure of eigenvalue accumulation near \(0\) can be described using compactness-based arguments.

7 Applications and Motivating Examples

7.1 Linear Dynamical Systems and Stability Criteria

In discrete and continuous linear dynamics, eigenvalues determine whether trajectories grow, decay, or oscillate. In discrete-time systems, the condition \(\rho(A)<1\) is a common stability criterion for decay toward the origin. Even when exact solutions are unavailable, spectral radius provides a reliable indicator of which modes dominate the asymptotic regime.

Numerical iterative solvers often involve repeatedly applying a matrix to approximate a solution. Convergence hinges on whether the iteration matrix has spectral radius below \(1\). When \(\rho(A)\) is small, errors contract more quickly; when it is close to \(1\), convergence can be slow or sensitive to perturbations. The link between spectral radius and error reduction underpins many conceptual discussions of iterative performance.

7.3 Markov Chains and Long-Run Behavior General Ideas

In a finite-state Markov chain, the update rule is governed by a stochastic matrix. While the transition structure has probabilistic meaning, its linear-algebraic skeleton is the spectrum of the transition matrix. The eigenvalue at \(1\) corresponds to steady behavior, while other eigenvalues—through their moduli—govern how quickly distributions converge to equilibrium.

7.4 Markov-like Matrices and Spectral Quantities

More generally, matrices arising from averaging or mixing processes often have spectral properties that encode “memory loss.” Even when not strictly Markov, positivity and normalization constraints can make the spectral radius and the subdominant eigenvalues central to understanding long-run behavior and rate of homogenization.

7.5 Worked Examples with Small Matrices

Consider a \(2\times 2\) matrix whose characteristic polynomial can be computed explicitly. One can find eigenvalues by solving a quadratic equation, then take the maximum modulus to obtain \(\rho(A)\). Comparing cases such as a diagonalizable matrix versus one with a Jordan block illustrates how \(\rho(A)\) may predict exponential behavior while Jordan structure determines whether powers \(A^k\) include additional polynomial factors.

8 Common Theorems and Useful Lemmas

8.1 Cayley–Hamilton Theorem Connection

The Cayley–Hamilton theorem states that every square matrix satisfies its own characteristic polynomial: \[ p_A(A)=0. \] This implies that the minimal polynomial divides the characteristic polynomial and that powers of \(A\) can be expressed as linear combinations of lower powers. While this does not directly give \(\rho(A)\), it ties spectral quantities to algebraic identities and often supports derivations of bounds on \(A^k\).

8.2 Inequalities Involving Norms and Spectral Radius

Because eigenvalues are controlled by norms, one can derive inequalities that relate \(\rho(A)\) to \(\|A^k\|\) and to various induced matrix norms. These bounds are frequently used to establish convergence or stability without computing eigenvalues exactly. In practice, selecting a convenient norm can tighten the inequality and improve the usefulness of the estimate.

8.3 Interlacing and Perturbation Themes At a Glance

Eigenvalues of related matrices (such as principal submatrices) can exhibit interlacing patterns, especially for Hermitian cases. Separately, perturbation principles describe how eigenvalues change under small modifications of the matrix. Together, these themes support reasoning about how \(\rho(A)\) might change when parameters are adjusted, even if exact spectra are hard to recompute.

8.4 Continuity Properties Under Matrix Perturbations

The spectral radius is continuous as a function of the matrix in finite-dimensional settings. More generally, spectral quantities exhibit stability under small perturbations, though the sensitivity of individual eigenvalues can vary, particularly near defective cases. Continuity ensures that approximate computations of eigenvalues and hence \(\rho(A)\) can be meaningfully interpreted.

8.5 Bounds for Gershgorin-Type Regions

Gershgorin’s circle theorem provides regions in the complex plane where eigenvalues must lie, based on row or column sums of magnitudes of off-diagonal entries. These regions yield computable upper bounds on \(\rho(A)\) by bounding eigenvalue magnitudes without solving the characteristic polynomial. Such estimates are particularly helpful for large matrices where direct eigenvalue computation is expensive.