---
title: Leonard Pair in Linear Algebra
url: https://www.emergentmind.com/topics/leonard-pair
type: topic
---

# Leonard Pair in Linear Algebra

A Leonard pair is an ordered pair of diagonalizable linear transformations \(A, A^*\) on a finite-dimensional vector space \(V\) such that there exists a basis in which \(A\) is irreducible tridiagonal and \(A^*\) is diagonal, and there exists a basis in which \(A^*\) is irreducible tridiagonal and \(A\) is diagonal. In this sense, each transformation acts on an eigenbasis for the other one in an irreducible tridiagonal fashion. Leonard pairs form a linear-algebraic abstraction of \(Q\)-polynomial distance-regular graphs and constitute a basic algebraic framework for the bispectrality of finite families of orthogonal polynomials, especially in the Askey and \(q\)-Askey settings [1308.3826, 1201.1645].

## 1. Definition, multiplicity-freeness, and Leonard systems

The defining feature of a Leonard pair is the mutual tridiagonalization/diagonalization property. Here “irreducible tridiagonal” means that all entries immediately above and below the main diagonal are nonzero. A basic consequence is multiplicity-freeness: both \(A\) and \(A^*\) are diagonalizable and each of their eigenspaces is one-dimensional [1201.1645].

A refined object attached to a Leonard pair is a Leonard system
\[
\Phi=(A,\{E_i\}_{i=0}^d,A^*,\{E_i^*\}_{i=0}^d),
\]
where \(\{E_i\}_{i=0}^d\) and \(\{E_i^*\}_{i=0}^d\) are standard orderings of the primitive idempotents of \(A\) and \(A^*\), respectively. The associated parameter array
\[
\left(\{\theta_i\}_{i=0}^d,\{\theta_i^*\}_{i=0}^d,\{\varphi_i\}_{i=1}^d,\{\phi_i\}_{i=1}^d\right)
\]
records the eigenvalue and split-sequence data and determines the Leonard system up to isomorphism. The eigenvalue sequences satisfy a uniform second-order condition: the quantities
\[
\frac{\theta_{i-2}-\theta_{i+1}}{\theta_{i-1}-\theta_i},
\qquad
\frac{\theta^*_{i-2}-\theta^*_{i+1}}{\theta^*_{i-1}-\theta^*_i}
\]
are equal and independent of \(i\) for the indicated ranges, and the common value is written as \(\beta+1\), where \(\beta\) is the fundamental parameter [1408.2180, 1409.4333].

This parameter-array viewpoint is central to classification. It supports both explicit canonical forms and invariant-based descriptions, and it underlies later refinements such as end-parameter theory, self-duality, bipartite conditions, and the passage to more general bispectral structures.

## 2. Recognition criteria and parameter-based characterizations

A substantial part of the theory concerns the inverse problem: given tridiagonal and diagonal data, how can one recognize that a pair is Leonard? One characterization uses the parameter sequence \(\{a_i\}_{i=0}^d\), defined in one standard orientation by
\[
a_i=\operatorname{tr}(E_iA^*),
\]
together with the dual eigenvalues \(\{\theta_i^*\}_{i=0}^d\). In this framework, the key conditions are a leaf condition in an auxiliary graph \(\Delta\), a three-term recurrence
\[
\theta^*_{i-1}-\beta\theta_i^*+\theta^*_{i+1}=\gamma^* \qquad (1\leq i\leq d-1),
\]
a quadratic relation
\[
a_i(\theta_i-\theta_{i-1})(\theta_i-\theta_{i+1})
=
\gamma \theta_i^{*2}+\omega\theta_i^*+\eta^*,
\]
and distinctness of the dual eigenvalues [1205.4368].

A later recognition theorem reformulates the problem in terms of the intersection numbers \(\{a_i\}_{i=0}^d\), \(\{b_i\}_{i=0}^{d-1}\), \(\{c_i\}_{i=1}^d\) and the dual eigenvalues \(\{\theta_i^*\}_{i=0}^d\). In a basis where \(A^*\) is diagonal, \(A\) takes the irreducible tridiagonal form with diagonal entries \(a_i\), superdiagonal entries \(b_i\), and subdiagonal entries \(c_i\). This characterization separates the roles of \(b_i\) and \(c_i\) and provides a direct test from explicit matrix data [1901.10659].

These recognition results are purely algebraic. They generalize earlier \(Q\)-polynomial criteria from distance-regular graph theory, but they do so without relying on graph-theoretic hypotheses. A plausible implication is that Leonard-pair recognition is most naturally understood as a problem in the interaction between primitive idempotents, tridiagonal adjacency, and low-order recurrences rather than as a secondary consequence of combinatorial regularity.

## 3. Bipartite, near-bipartite, self-dual, and spin structures

One important special class is the bipartite Leonard pair. In the bipartite setting, the relevant diagonal entries vanish. In one formulation, if \(\{v_i\}_{i=0}^d\) is an eigenbasis for \(A^*\) on which \(A\) is irreducible tridiagonal and \(E_i^*\) denotes the projection onto \(v_i\), then the flat part
\[
F=\sum_{i=0}^d E_i^*AE_i^*
\]
captures the diagonal of \(A\). The pair is bipartite exactly when \(F=0\). It is near-bipartite when \((A-F,A^*)\) is again a Leonard pair; in that case \((A-F,A^*)\) is bipartite and is called the bipartite contraction. Over an algebraically closed field, near-bipartite Leonard pairs are classified up to isomorphism: they are precisely the essentially bipartite, the reinforced dual \(q\)-Krawtchouk type, or the Krawtchouk type [2304.04965].

For bipartite Leonard pairs, the notion of a tail yields a sharp characterization. One forms a graph \(\Delta\) on the primitive idempotents of \(A\), declaring \(i\) adjacent to \(j\) when \(E_iA^*E_j\neq 0\). An ordered pair \((E,F)\) of distinct primitive idempotents is a tail if the vertex corresponding to \(E\) is adjacent to no vertex except \(F\), and the vertex corresponding to \(F\) is adjacent to at most one vertex besides \(E\). In the bipartite case, \((E,F)\) is \(Q\)-polynomial if and only if \(E\) is normalizing, \((E,F)\) is a tail, and the eigenvalues of \(A^*\) are mutually distinct. The proof passes through a three-term recurrence
\[
\theta^*_{i-1}-B\theta_i^*+\theta^*_{i+1}
\]
that is independent of \(i\) [1308.3826].

Another distinguished symmetry is self-duality. A Leonard pair is self-dual when there exists an automorphism of \(\operatorname{End}(V)\) swapping \(A\) and \(A^*\). This automorphism is unique. In the self-dual case there exists an invertible linear map \(T\) such that \(X\mapsto TXT^{-1}\) realizes the duality; moreover, \(T\) can be expressed as a polynomial in \(A\) and \(A^*\), and its action on canonical flags, decompositions, and bases can be described explicitly [1805.02545].

The spin condition strengthens this picture. A Leonard pair has spin when there exist invertible maps \(W,W^*\) such that \(WA=AW\), \(W^*A^*=A^*W^*\), and
\[
WA^*W^{-1}=(W^*)^{-1}AW^*.
\]
Assuming the ground field is algebraically closed and \(d\geq 3\), a Leonard pair has spin if and only if it is self-dual and there exist scalars \(f_0,f_1,f_2,f_3\), not all zero, such that
\[
f_0+f_1\theta_i^*+f_2a_i+f_3a_i\theta_i^*=0
\qquad (0\leq i\leq d).
\]
Equivalently, the zero diagonal space \(\mathcal Z(A,A^*)\subseteq \operatorname{Span}\{I,A^*,A,AA^*\}\) is nonzero [2509.21520].

## 4. Orthogonal polynomials, representation theory, and combinatorial realizations

Leonard pairs are closely tied to finite orthogonal polynomial systems. For Krawtchouk type, both the eigenvalue and dual eigenvalue sequences are
\[
\{N-2i\}_{i=0}^N,
\]
and every Leonard pair of Krawtchouk type arises from finite-dimensional irreducible \(\mathfrak{sl}_2\)-modules. In that realization, Krawtchouk polynomials appear as transition coefficients between eigenbases, and their three-term recurrence, orthogonality, difference equation, and generating function follow from the Leonard-pair and module structure [1201.1645].

At the \(q\)-Racah level, Leonard pairs occur naturally in the representation theory of the universal double affine Hecke algebra \(\hat H_q\) of type \((C_1^\vee,C_1)\). If \(\mathcal V\) is a feasible finite-dimensional irreducible \(\hat H_q\)-module, then \(X+X^{-1}\) and \(Y+Y^{-1}\) act on each eigenspace of \(t_0\) as a Leonard pair of \(q\)-Racah type. The isomorphism class of such a pair is encoded by Huang data \((a,b,c,d)\), and the paper gives necessary and sufficient conditions for a pair of \(q\)-Racah Leonard pairs to arise from this DAHA construction [1701.06089].

In algebraic combinatorics, Leonard pairs govern thin irreducible modules for Terwilliger algebras of distance-regular graphs. They also enter the theory of spin models. For a formally self-dual distance-regular graph of \(q\)-Racah type, if for all base vertices all irreducible Terwilliger modules are thin with equal endpoint and dual endpoint and have the specified intersection numbers, then the graph affords a spin model \(W\) explicitly constructed from Leonard-pair data. This links Leonard pairs, Bose–Mesner algebras, Terwilliger algebras, Nomura algebras, and Jones-type spin models within a single framework [1907.03900].

Leonard pairs also appear in integrable systems. The Heun–Askey–Wilson operator can be diagonalized within the framework of the algebraic Bethe ansatz using Leonard pairs. In this setting, two families of on-shell Bethe states generate explicit bases on which a Leonard pair acts in a tridiagonal fashion, and the associated \(Q\)-polynomials satisfy Baxter \(T\)-\(Q\) relations. For a special case, the \(Q\)-polynomial is identified with the Askey–Wilson polynomial [1909.02464].

## 5. Canonical matrix forms, moduli, and isomorphism invariants

A recurrent theme in the classification theory is the search for canonical matrix realizations. One such realization is LB-TD form: \(A\) is lower bidiagonal with all subdiagonal entries \(1\), and \(A^*\) is irreducible tridiagonal. Over an algebraically closed field and for \(q\) not a root of unity, all Leonard pairs in \(\mathrm{Mat}_{d+1}(\mathbb F)\) with LB-TD form are classified. In that regime, only the \(q\)-Racah, \(q\)-Hahn, and dual \(q\)-Hahn types admit LB-TD forms, and a Leonard pair has LB-TD form if and only if at least two of \(aa'\), \(bb'\), \(\xi\) are nonzero [1404.6794].

Another canonical problem is zero-diagonal TD-TD form, where both \(A\) and \(A^*\) are irreducible tridiagonal with all diagonal entries \(0\). This case is completely classified. The decisive criterion is self-oppositeness: a Leonard pair has zero-diagonal TD-TD form if and only if it is isomorphic to its opposite \((-A,-A^*)\). For \(d\geq 3\), the possible types are Krawtchouk, Bannai–Ito when \(d\) is even, and \(q\)-Racah [1503.05262].

The invariant theory of Leonard systems admits a particularly sharp boundary description. The end-parameters
\[
\theta_0,\theta_d,\theta_0^*,\theta_d^*,\varphi_1,\varphi_d,\phi_1,\phi_d
\]
together with the fundamental parameter \(\beta\) determine a Leonard system up to isomorphism. Moreover, for specified end-parameters, there are up to inverse at most \(\lfloor (d-1)/2\rfloor\) Leonard systems, and this upper bound is best possible [1408.2180]. A related refinement replaces end-parameters by end-entries
\[
\theta_0,\theta_d,\theta_0^*,\theta_d^*,a_0,a_d,a_0^*,a_d^*.
\]
In this formulation, a Leonard system is determined up to isomorphism by its end-entries and \(\beta\) precisely in the generic cases \(\beta\neq \pm 2\) with \(q^{d-1}\neq -1\), or \(\beta=\pm 2\) with \(\operatorname{Char}(\mathbb F)\neq 2\); otherwise there are infinitely many non-isomorphic systems with the same data [1409.4333].

Compatibility theory adds another layer to the moduli problem. Two Leonard pairs \(A,A^*\) and \(B,B^*\) on the same space are compatible when \(A^*=B^*\) and \([A,A^*]=[B,B^*]\). A companion of \(A,A^*\) is a polynomial \(K\in \mathbb F[A^*]\) such that \(A-K,A^*\) is again a Leonard pair. These notions are equivalent under \(K=A-B\), and the compatible pairs are classified through equalities involving the invariant value \(\kappa\) and the endpoint split-sequence products [2112.13326].

## 6. Generalizations, failures of Leonard duality, and higher-rank extensions

Leonard duality does not exhaust finite bispectrality. The dual \(-1\) Hahn polynomials provide a basic counterexample: they are orthogonal on a finite discrete set and satisfy a three-term recurrence, but they do not satisfy the Leonard duality property. Instead of a second-order difference equation, they satisfy a fourth-order difference eigenvalue equation. The corresponding generalized Leonard pair consists of matrices \(A,B\) such that, in the eigenbasis where \(A\) is diagonal, \(B\) is 3-diagonal, while in the eigenbasis where \(B\) is diagonal, \(A\) is 5-diagonal [1108.0132].

A direct extension of the Leonard-pair paradigm is the Leonard trio \((V,\overline V,Z)\). In an irreducible Leonard trio, both \((V,Z)\) and \((\overline V,Z)\) are Leonard pairs, while the overlap coefficients between eigenbases are in general rational rather than polynomial. These coefficients satisfy generalized eigenvalue problems and biorthogonality relations. For \(q\)-Racah-type irreducible Leonard trios, Wilson’s rational functions appear as overlap coefficients, and they admit recurrence, difference, and biorthogonality relations [2601.15052].

A higher-rank analogue is the factorized \(A_2\)-Leonard pair. It is introduced as a rank \(2\) Leonard pair whose actions in certain bases correspond to the root system of the Weyl group \(A_2\), with additional factorization properties. The transition-matrix entries are bivariate orthogonal polynomials of Tratnik type with bispectral properties, and the most general examples are associated with an intricate product of univariate \((q\)-)Hahn and dual \((q\)-)Hahn polynomials [2312.08312].

Recent work also shows that Leonard-pair operations can change the polynomial family while preserving the same inner product. For dual Hahn polynomials with parameters \(r\neq 0\) and \(r+s=0\), if \(L,L^*\) denote the corresponding Leonard pair, then
\[
L,\left(L^*+\frac{r-d}{2}\right)^2
\]
is again a Leonard pair. In that regime, the same orthogonal system can be treated not only as dual Hahn polynomials but also as Racah polynomials with respect to the same inner product [2508.02032].

These developments indicate that the classical Leonard-pair framework is both rigid and extensible: rigid enough to admit exact classifications, explicit canonical forms, and strong invariant theory, yet extensible enough to accommodate generalized dualities, higher-order difference operators, rational overlap coefficients, and higher-rank bispectral structures.

Source: https://www.emergentmind.com/topics/leonard-pair