Spectral Favard Theorem

• Spectral Favard Theorem is a generalization of the classical Favard theorem that links banded matrices with total positivity to systems of multiple or mixed-type orthogonal polynomials.
• It leverages positive bidiagonal factorizations to establish biorthogonality and derive (p+2)-term recurrences for both scalar and matrix measures.
• The theorem underpins advanced applications in random walks, Markov chains, and multidimensional orthogonal systems, extending its reach to unit circle and Sobolev cases.

The Spectral Favard Theorem generalizes the classical Favard theorem, establishing a correspondence between banded (potentially non-self-adjoint) infinite matrices with total positivity properties and systems of multiple or mixed-type orthogonal polynomials tied to sets of measures. This framework encodes the spectral structure of banded operators—most notably oscillatory Hessenberg matrices and their higher-band analogues—via their positive bidiagonal factorizations and the associated biorthogonality of the polynomial families. The modern form encompasses scalar, vector, and matrix measures; covers higher-order recurrences; and underpins applications ranging from random walks and Markov chains to multidimensional orthogonal polynomials.

1. Algebraic Setting: Oscillatory Banded Matrices and Total Positivity

Let $T=(T_{i,j})_{i,j\ge0}$ denote a semi-infinite lower Hessenberg matrix of bandwidth $p+1$, i.e., $T_{n,k}=0$ for $k>n+1$ or $k<n-p$, with $T_{n,n+1}=1$ in the monic normalization. Total nonnegativity (TN) means all minors of $T$ are nonnegative; $T$ is oscillatory if it is TN, invertible, and irreducible—equivalently, all first sub- and superdiagonals are strictly positive (Gantmacher–Krein criterion). The structure of such matrices is deeply connected to positive bidiagonal factorizations (PBF), where

$T = L_1 L_2 \cdots L_p U,$

with each $L_k$ unit lower-bidiagonal (with positive subdiagonal entries) and $U$ unit upper-bidiagonal (with positive superdiagonal entries). This product structure ensures strict total positivity.

2. Statement of the Spectral Favard Theorem

Suppose $T$ is a bounded banded lower Hessenberg matrix with PBF, and that initial conditions for type I orthogonal polynomials ensure all Christoffel numbers are positive. Then there exist $p$ nondecreasing functions (Stieltjes–Lebesgue measures) $\mu_1,\ldots,\mu_p$ with compact support $A\subset\mathbb{R}$ such that the sequences $\{B_n\}_{n\ge0}$ (type II) and $\{A_n^{(a)}\}_{n\ge0}$ (type I) are multiple orthogonal with respect to $\{\mu_a\}_{a=1}^p$. The core biorthogonality is expressed as

$$\int_A \left( A_m^{(1)}(x)\,d\mu_1(x) + \cdots + A_m^{(p)}(x)\,d\mu_p(x) \right) B_n(x) = \delta_{m,n}.$$

The polynomials $B_n$ and $A_n^{(a)}$ satisfy dual $(p+2)$-term recurrences explicitly determined by the entries of $T$. The induced spectral (matrix) measure is block-diagonal in the type I basis, and $T$ becomes unitarily equivalent to multiplication by $x$ on $\ell^2$ in this orthogonalization (Branquinho et al., 2022).

The theorem extends to semi-infinite (possibly unbounded) banded matrices with PBF, yielding a matrix-valued spectral measure and an explicit spectral representation for matrix powers in terms of mixed-type multiple orthogonal polynomials. The limiting process employs Gaussian quadrature structure and Helly-type compactness to ensure existence of the limiting measure even when boundedness fails (Branquinho et al., 18 Jan 2026).

3. Multiple Orthogonality, Recurrences, and Mixed Measures

Given measures $\mu_1,\dots,\mu_p$ on $\mathbb{R}$, type II polynomials $B_n(x)$ (deg $B_n = n$) are characterized by $T\,B(x)=xB(x)$ with $B_0(x)=1$, yielding the $(p+2)$-term recurrence: $$B_{n+1}(x) = (x - T_{n,n}) B_n(x) - \sum_{k=1}^p T_{n,n-k} B_{n-k}(x).$$ Dual type I left eigenvectors $A^{(a)}(x)$ satisfy $A^{(a)}(x)\,T = x\,A^{(a)}(x)$, obeying a corresponding recurrence. The resulting system forms an interwoven biorthogonal structure, with precise normalization ensuring all Christoffel numbers are positive.

In the mixed (non-symmetric) case—where $T$ has bandwidth $(p,q)$—the framework naturally generalizes to a $q\times p$ matrix of measures $\Psi(x) = (\psi_{b,a}(x))$, and left/right families $A^{(a)}(x)$, $B^{(b)}(x)$ satisfying dual recursions (Branquinho et al., 2022, Branquinho et al., 18 Jan 2026). The biorthogonality then reads: $$\sum_{a=1}^p \sum_{b=1}^q \int B_n^{(b)}(x)\,d\psi_{b,a}(x)\,A_m^{(a)}(x) = \delta_{m,n}.$$

4. Spectral Representation and Quadrature Structures

PBF implies all principal truncations $T^{[N]}$ are oscillatory, with simple real spectra. The eigenvalues yield the quadrature nodes, and corresponding left/right eigenvectors expand in terms of $A^{(a)}$ and $B^{(b)}$ polynomials. The Christoffel numbers, constructed from these eigenvectors, supply a basis for discrete measures at the eigenvalues, with sharp positivity determined by total positivity of $T$.

The general (multiple) Gauss quadrature for measures $\mu_a$ employs the zeros $\{x_k\}$ of $B_N(x)$ as nodes and Christoffel numbers $w_k^{(a)}$ as weights, with exactness up to degree

$d_a(N) = (p+1)(N-1) + a, \quad a=1,\dots,p.$

These bounds are sharp, and the quadrature extends to the full matrix-valued measure setting for mixed orthogonality, with degree of precision governed by row and column indices (Branquinho et al., 2022, Branquinho et al., 2022, Branquinho et al., 18 Jan 2026).

5. Generalizations: Unit Circle, Multidimensional, and Sobolev Cases

Unit Circle

The spectral Favard theorem extends to orthogonal polynomials on the unit circle (OPUC), where the measure's discrete mass at a specific spectral endpoint (here $z=1$) becomes a free parameter, and the spectral content of the recurrence is determined via the chain sequence's minimal and maximal parameters. In this regime, uniqueness of the measure and orthogonality properties are captured with the same biorthogonality structure, and explicit examples involve hypergeometric functions (Castillo et al., 2013).

Multidimensional Favard Theorem

For vector-valued orthogonal polynomials (e.g., in several variables), the spectral Favard theorem characterizes compactly supported measures via a block recurrence: $$X_i P_n(x) = B_n P_{n+1}(x) + A_n^{(i)} P_n(x) + B_{n-1}^* P_{n-1}(x)$$ for $i=1,\dots,d$, where $A_n^{(i)}$, $B_n$ are square and rectangular matrix sequences subject to positivity and intricate compatibility constraints. This formalism is framed in terms of symmetric interacting Fock spaces and yields a one-to-one correspondence between such block Jacobi data and measures (Accardi et al., 2014).

Sobolev Orthogonality

For Sobolev orthogonal polynomials, Favard-type theorems involve banded lower–Hessenberg matrices whose structure encodes the order of differentiation in the inner product. The matrix $G$ is called Hessenberg–Sobolev of index $d$ if its associated formal moment matrix is Hankel–Sobolev of the same index, determined by finite-difference vanishing constraints $Y(G,2d+1)=0$, $Y(G,2d)\neq0$. The existence and uniqueness of measures for the associated Sobolev inner product are precisely captured by these algebraic criteria (Pijeira-Cabrera et al., 2022).

6. Applications: Markov Chains and Spectral Classification

The spectral Favard theorem for (bounded) banded matrices with PBF has direct applications to reversible and non-reversible Markov chains with transition matrices beyond the classical tridiagonal (birth-and-death) case. It provides the spectral representation (Karlin–McGregor type) for transition probabilities, characterizes recurrence and stationary distributions in terms of the spectral measure, and relates ergodicity to the presence of a mass at $x=1$, corresponding to the eigenvalue $1$ (the stationary state) (Branquinho et al., 15 Jan 2026, Branquinho et al., 2022). Explicit expressions for stationary distributions and convergence rates are available in terms of the associated multiple orthogonal polynomial families.

7. Structural Significance and Contemporary Directions

The spectral Favard theorem unifies a wide class of operator models (Hessenberg, block matrices, Sobolev-type, mixed orthogonality) via the algebraic device of positive bidiagonal factorization and total positivity. It produces spectral measures—often matrix-valued or vector-valued—that encode finer structural information than scalar spectral theory alone, particularly for non-self-adjoint and non-diagonalizable settings.

Current research extends the theorem to unbounded operators, mixed-type orthogonality, high-bandwidth Markov chains, multidimensional polynomial systems, and further explores connections with integrable systems, spectral theory of random walks, and advanced quadrature methods (Branquinho et al., 2022, Branquinho et al., 2022, Branquinho et al., 18 Jan 2026, Accardi et al., 2014, Castillo et al., 2013, Pijeira-Cabrera et al., 2022). The interplay of algebraic factorization, spectral representation, and orthogonality remains central to the characterization and classification of both classical and modern families of special functions and stochastic models.