Joint Spectral Radius Overview

• Joint Spectral Radius is a measure in matrix analysis that quantifies the worst-case exponential growth of finite matrix products, crucial for stability analysis in switched systems.
• Algorithmic techniques such as extremal norms, polytope methods, and SOS relaxations serve to approximate and compute this key invariant in both linear and nonlinear settings.
• Its applications span robust control, wavelet regularity, and fractal geometry, linking algebraic, geometric, and computational perspectives in modern research.

The joint spectral radius (JSR) is a key quantity in linear dynamics, matrix analysis, and control theory. It quantifies the maximal exponential rate of growth over all finite products formed from a given set of matrices and provides a unifying framework for stability analysis in switched and time-varying systems, wavelet regularity, and fractal geometry. The theory of the JSR encompasses algebraic, geometric, algorithmic, and functional-analytic methodologies, and continues to generate deep questions regarding extremal norms, spectral finiteness, forbidden patterns, and nonlinear extensions.

1. Definition and Fundamental Properties

Given a finite set $$\mathcal{M} = \{A_1, \ldots, A_m\} \subset \mathbb{K}^{d \times d}$$ ($\mathbb{K} = \mathbb{R}$ or $\mathbb{C}$), the joint spectral radius is

$$\rho(\mathcal{M}) = \lim_{k \to \infty} \sup_{A_{i_1}, \ldots, A_{i_k} \in \mathcal{M}} \|A_{i_k} \cdots A_{i_1}\|^{1/k}.$$

This limit exists and is independent of the matrix norm, due to submultiplicativity and Gelfand-Fekete framework (Guglielmi et al., 2011). The JSR generalizes the spectral radius of a single matrix to sets of matrices, measuring the worst-case exponential growth over all finite switching sequences.

Key properties include:

• Norm invariance: $\rho(\mathcal{M})$ does not depend on the particular induced matrix norm.
• Scaling: For any scalar $\alpha \geq 0$, $$\rho(\alpha \mathcal{M}) = \alpha \rho(\mathcal{M})$$.
• Similarity invariance: For every invertible $T$, $\rho(T^{-1}\mathcal{M}T) = \rho(\mathcal{M})$.
• Berger–Wang formula: The JSR equals the limsup of the spectral radii of products,

$$\rho(\mathcal{M}) = \limsup_{k\to\infty} \max_{A \in \mathcal{M}^k} r(A)^{1/k}$$

where $\mathbb{K} = \mathbb{R}$0 is the spectral radius of $\mathbb{K} = \mathbb{R}$1 (Dai, 2011). This establishes equivalence between asymptotic matrix norm and spectral growth.

• Lower spectral radius: $\mathbb{K} = \mathbb{R}$2, which serves as a dual quantity for characterizing uniform stabilizability (Kozyakin, 2017).

2. Spectral Finiteness, Extremal Norms, and Forbidden Products

Spectral Finiteness Property

A set $\mathbb{K} = \mathbb{R}$3 is said to have the finiteness property if $\mathbb{K} = \mathbb{R}$4 is realized by a finite product, i.e., there exists a word $\mathbb{K} = \mathbb{R}$5 of length $\mathbb{K} = \mathbb{R}$6 such that

$\mathbb{K} = \mathbb{R}$7

In generic cases, the finiteness property fails, and the JSR may not be attained by any finite product (Breuillard et al., 2018). However, it holds in special situations—for instance, if all but one matrix in $\mathbb{K} = \mathbb{R}$8 have rank one, one can always find a finite product realizing the JSR; Dai et al. provide a constructive, algorithmically decidable procedure for this case, which is a sharp decidability boundary (Dai, 2011, Liu et al., 2011).

Extremal and Barabanov Norms

An extremal norm satisfies $\mathbb{K} = \mathbb{R}$9 for all $\mathbb{C}$0 and $\mathbb{C}$1. Under irreducibility, Barabanov constructed a norm with $\mathbb{C}$2 for all $\mathbb{C}$3, critical for achieving tight trajectories and for algorithms (polytope methods) that yield exact JSR values for many practical families (Guglielmi et al., 2011).

Forbidden Products

A forbidden product is a word in the matrix alphabet which, for all choices of matrices, never attains the strict maximum possible normalized spectral radius. Vladimirov established the existence of such forbidden products in dimension two, showing that, e.g., $\mathbb{C}$4 and its isospectral class are never maximizing for any pair of real $\mathbb{C}$5 matrices, answering an open problem on non-attainability in minimal settings (Vladimirov, 2024).

3. Computation, Approximation, and Algorithmic Techniques

Given the undecidability and NP-hardness of the general JSR decision and approximation problems (Ahmadi et al., 2011), multiple computational frameworks have been developed:

Exact Polytope Methods

For irreducible finite families, polytope norm algorithms construct an extremal polytope invariant under $\mathbb{C}$6. If the "dominant product" property holds, these methods terminate in finitely many steps and yield exact values (Guglielmi et al., 2011).

Sum-of-Squares (SOS) Relaxations

Relaxations via SOS Lyapunov functions and semidefinite programming offer systematic upper bounds, with sparse variants (e.g., SparseJSR) exploiting term-sparsity and chordal decomposition to handle high-dimensional, large-scale problems efficiently (Wang et al., 2020).

Path-Complete Graph Lyapunov Functions

Hierarchies based on path-complete graphs and multiple Lyapunov functions unify standard common quadratic and min/max-of-quadratic relaxations, providing asymptotically tight upper bounds and converse theorems for stability (Ahmadi et al., 2011).

Nonlinear Eigenproblem Hierarchies

For nonnegative families, hierarchies of nonlinear ergodic eigenproblems provide converging approximations to the JSR, with computational cost essentially independent of the system dimension but exponential in "memory length" (Gaubert et al., 2018).

Operator Space and Noncommutative Function Theory

Spectral radii for matrices over operator spaces generalize the classical JSR and clarify simultaneous similarity to unit balls in matrix function theory, refining simultaneous contraction and unitary similarity characterizations (Shalit et al., 2 Jan 2025).

4. Extensions to Banach and Nonlinear Settings

Banach Algebraic Generalizations and Topological Radicals

The theory extends to precompact sets in Banach algebras via the hypocompact radical. The JSR in such algebras is expressed through a mixed Berger–Wang type formula: $\mathbb{C}$7 where $\mathbb{C}$8 is the essential JSR modulo the hypocompact radical and $\mathbb{C}$9 is the Berger–Wang radius (Shulman et al., 2012). In GCR C*-algebras, the JSR is continuous and coincides with the classical BW radius (Shulman et al., 2012).

Markovian Product Laws

Generalizations to products following Markovian laws yield Markovian analogues of the JSR and the generalized spectral radius. The Berger–Wang formula extends: $$\rho(\mathcal{M}) = \lim_{k \to \infty} \sup_{A_{i_1}, \ldots, A_{i_k} \in \mathcal{M}} \|A_{i_k} \cdots A_{i_1}\|^{1/k}.$$0 holds for any Markov law on the matrix alphabet, following a matrix-theoretic "Ω-lift" construction (Kozyakin, 2014).

Nonlinear Joint Spectral Radius

Recent work extends the JSR to switched families of sub-homogeneous, order-preserving maps on cones. The nonlinear JSR controls asymptotic stability and trajectory growth and is sandwiched between the JSRs of the asymptotically homogeneous families (scaling at $$\rho(\mathcal{M}) = \lim_{k \to \infty} \sup_{A_{i_1}, \ldots, A_{i_k} \in \mathcal{M}} \|A_{i_k} \cdots A_{i_1}\|^{1/k}.$$1 and $$\rho(\mathcal{M}) = \lim_{k \to \infty} \sup_{A_{i_1}, \ldots, A_{i_k} \in \mathcal{M}} \|A_{i_k} \cdots A_{i_1}\|^{1/k}.$$2). Duality extends to generalized extremal prenorms, and a polytopal-type finite algorithm applies in the presence of a dominant product (Deidda et al., 15 Jul 2025).

5. Spectral Bounds, Continuity, and Structural Results

Explicit Bounds and Diagonal Approximations

Bounding the JSR using diagonal entries, one obtains explicit, combinatorial lower and upper bounds for nonnegative matrix families, depending only on principal cycles and the matrix entry range (Bui, 2020). John's ellipsoid and projection constants yield explicit dimension-dependent upper bounds for principal submatrices in the similarity orbit (Epperlein et al., 24 Apr 2025). However, uniform bounds cannot extend to higher-dimensional principal submatrices due to geometric obstructions.

Continuity and Regularity

The JSR is a continuous function of the matrix set with respect to the Hausdorff metric. In fact, it is pointwise Hölder continuous; for finite sets, the exponent improves, and in dimension two with positive JSR, an explicit $$\rho(\mathcal{M}) = \lim_{k \to \infty} \sup_{A_{i_1}, \ldots, A_{i_k} \in \mathcal{M}} \|A_{i_k} \cdots A_{i_1}\|^{1/k}.$$3-Hölder bound holds. Local regularity is closely tied to reducibility structure and block decomposition. Continuity of the spectrum and JSR is further stratified by the structure of topological radicals (e.g., scattered radical and its extensions) (Epperlein et al., 2023, Shulman et al., 2012).

6. Connections, Applications, and Open Problems

The JSR arises in the stability analysis of switched linear and nonlinear systems, design of robust controllers, quantification of wavelet and subdivision-scheme regularity, analysis of iterated function systems in fractals, extremal norm construction in ergodic theory, spectral radii of random products, and spectral theory of Banach algebras. Its role as a unifying, intrinsic asymptotic invariant links disparate areas from rational noncommutative function theory to Lyapunov exponents in random matrix products.

Several open problems remain:

• Optimal continuity exponents for JSR in high dimensions (Epperlein et al., 2023).
• Characterization of all forbidden products and spectral finiteness failure mechanisms (Vladimirov, 2024).
• Extension of spectral finiteness or efficient computability to higher-rank or noncommutative families (Dai, 2011).
• Scaling and practical algorithms for large nonlinear switched systems (Deidda et al., 15 Jul 2025).
• Precise geometric and analytic bounds on submatrix JSRs and projection constants (Epperlein et al., 24 Apr 2025).