---
title: Zero Diagonal Spaces in Mathematics
url: https://www.emergentmind.com/topics/zero-diagonal-space
type: topic
---

# Zero Diagonal Spaces in Mathematics

Searching arXiv for the cited papers to ground the response in current records.
“Zero diagonal space” denotes several related constructions in contemporary mathematics, unified by the requirement that a diagonal vanish but differing sharply by context. In operator theory, it can mean the class of bounded operators whose diagonal entries are identically zero in some orthonormal basis, or the closed subspace determined by a fixed basis [1410.7441]. In finite-dimensional matrix theory it often means the linear subspace of real symmetric or Hermitian matrices with diagonal zero [2309.14958], or the subspace of zero-diagonal tridiagonal matrices used to study Leonard pairs [1503.05262]. In the Terwilliger-theoretic setting of Leonard pairs, it can denote a specific subspace of $\mathrm{Span}\{I,A^*,A,AA^*\}$ characterized by vanishing diagonal entries in an $A^*$-eigenbasis [2509.21520]. In topology, diagonal terminology shifts from matrix diagonals to the diagonal subset $\Delta\subseteq X^2$; there, “zero-set diagonal,” “$\mathbb{Q}$-diagonal,” and related notions concern how $\Delta$ or $X^2\setminus\Delta$ is represented or controlled [1112.0883, 1709.06879]. The resulting literature is therefore not a single theory but a collection of structurally analogous theories in operator algebras, matrix analysis, algebraic combinatorics, geometry, integrable systems, and topology.

## 1. Operator-theoretic zero-diagonal classes

In the operator-theoretic sense, the basic notion is due to Fan’s terminology: “Zero-diagonal operators are those whose diagonal entries are identically zero in some basis” [1410.7441]. For a complex Hilbert space $H$ and $T\in B(H)$,
\[
T \text{ is zero-diagonal} \iff \exists \text{ an orthonormal basis } \{e_n\}\subset H \text{ such that } \langle Te_n,e_n\rangle =0 \ \forall n.
\]
For a fixed orthonormal basis $e=\{e_n\}$, the corresponding zero diagonal space is
\[
Z_e:=\{T\in B(H):\langle Te_n,e_n\rangle =0 \ \forall n\},
\]
a closed subspace of $B(H)$ [1410.7441]. By contrast, the intrinsic class
\[
Z:=\{T\in B(H): T \text{ is zero-diagonal in some orthonormal basis}\}
\]
is not a linear subspace, because the property is existential and basis-dependent [1410.7441].

The most detailed structural results in this direction concern idempotents. If $D^2=D$, then relative to
\[
H=\ker(I-D)\oplus \ker D,
\]
one has the canonical decomposition
\[
D=
\begin{pmatrix}
I & T\\
0 & 0
\end{pmatrix},
\]
where $T\in B(\ker(I-D),\ker D)$ is called the nilpotent part [1410.7441]. This block form isolates the deviation from a projection in the off-diagonal corner $T$.

The central characterization is Corollary 2.6 of “Diagonality and idempotents with applications to problems in operator theory and frame theory” [1410.7441]: a nonzero idempotent is zero-diagonal if and only if it is not a Hilbert–Schmidt perturbation of a projection. For infinite-rank idempotents, Theorem 2.5 gives a stronger equivalence: the following are equivalent—being not a Hilbert–Schmidt perturbation of a projection, having nilpotent part not Hilbert–Schmidt, satisfying $R\{\operatorname{Tr}D\}=\mathbb{C}$, being zero-diagonal, and possessing summable or absolutely summable diagonals in some basis [1410.7441]. This identifies the zero-diagonal class of idempotents through a non-Hilbert–Schmidt size condition on the nilpotent corner.

The same paper also gives an unusually broad diagonal realization theorem: every bounded sequence $(d_n)\in \ell^\infty$ is the diagonal of some idempotent on a separable infinite-dimensional Hilbert space [1410.7441]. The zero-diagonal case is the special choice $d_n\equiv 0$. The finite-rank situation is much more rigid: the diagonals of nonzero finite-rank idempotents are precisely the absolutely summable sequences with positive integer sum, so nonzero finite-rank idempotents are never zero-diagonal [1410.7441].

A further motivation comes from frame theory. The same diagonal realization theorem implies that every bounded sequence can appear as $\langle f_n,g_n\rangle$ for some dual frame pair in infinite dimensions, because those inner products are encoded as diagonals of certain idempotents [1410.7441]. In this sense, the zero-diagonal space of idempotents becomes a tool for describing admissible inner-product patterns of dual frames.

## 2. Finite-dimensional matrix spaces and norm geometry

In matrix analysis, “zero diagonal space” is usually literal: the linear subspace of matrices whose main diagonal entries vanish. For real symmetric and Hermitian matrices,
\[
\mathcal S_n^0=\{A\in \mathcal S_n: A_{ii}=0 \text{ for all }i\},\qquad
\mathcal H_n^0=\{A\in \mathcal H_n: A_{ii}=0 \text{ for all }i\},
\]
and these are linear subspaces of $\mathcal S_n$ and $\mathcal H_n$ respectively [2309.14958]. They are also the orthogonal complements of the diagonal subspace $\mathcal D_n$ with respect to the Frobenius inner product [2309.14958].

The paper “Minimum trace norm of real symmetric and Hermitian matrices with zero diagonal” [2309.14958] studies these spaces through extremal norm inequalities. For nonzero real symmetric $A$ with diagonal zero,
\[
\frac{\|A\|_1}{\|A\|_{(1)}} \ge \frac{2}{n},
\]
and the minimum value is exactly $2/n$ [2309.14958]. For nonzero Hermitian $A$ with diagonal zero,
\[
\frac{\|A\|_1}{\|A\|_{(1)}} \ge \tan\left(\frac{\pi}{2n}\right),
\]
and the minimum value is exactly $\tan(\pi/(2n))$ [2309.14958]. These are sharp lower bounds for the trace norm in terms of the entry-wise $L^1$-norm.

The extremal matrices are highly structured. In the real symmetric case, $J_n-I_n$ attains the bound $2/n$ [2309.14958]. In the Hermitian case, extremals are constructed from vectors of $n$-th roots of unity via matrices of the form
\[
A=\mathbbm{1}\mathbbm{1}^*-\alpha\alpha^*,
\]
yielding the sharp constant $\tan(\pi/(2n))$ [2309.14958]. The paper interprets these inequalities dually as spectral-norm bounds for approximation by diagonal matrices:
\[
\min_{D\in \mathcal D_n}\|A-D\|_\infty \le \frac n2\,\|A\|_{(\infty)}
\]
for symmetric matrices, and
\[
\min_{D\in \mathcal D_n}\|A-D\|_\infty \le \cot\left(\frac{\pi}{2n}\right)\|A\|_{(\infty)}
\]
for Hermitian matrices [2309.14958]. Thus the geometry of the zero-diagonal space is linked directly to the distance from the diagonal subspace.

A different finite-dimensional zero-diagonal problem arises inside the orthogonal group. An $\mathrm{OMZD}(n)$ is a real orthogonal $n\times n$ matrix whose diagonal entries are zero and whose off-diagonal entries are all nonzero [1810.08961]. Such matrices exist if and only if $n\neq 1,3$, while symmetric $\mathrm{OMZD}(n)$ exist if and only if $n$ is even and $n\neq 4$ [1810.08961]. The paper also gives a construction from doubly regular tournaments and uses these matrices to determine when certain bipartite graphs satisfy $q(G)=2$ [1810.08961]. Here the zero diagonal space is the hollow-matrix subspace intersected with a nonlinear orthogonality constraint and a full-support condition off the diagonal.

The 2025 paper “Zeroing Diagonals, Conjugate Hollowization, and Characterizing Nondefinite Operators” pushes this orbit-intersection perspective further [2508.00096]. It proves that for any pair of real traceless matrices $L,M$, there exists an orthogonal $V$ such that $V^{-1}LV$ is hollow and $VMV^{-1}$ is almost hollow, thereby proving the conjecture of Damm and Fassbender [2508.00096]. It also characterizes real traceless matrices through orthogonal hollowization and relates nondefiniteness to the possibility of orthogonally zeroing specified diagonal entries [2508.00096]. This suggests a broader viewpoint in which zero-diagonal spaces are studied as target slices intersecting orthogonal similarity orbits.

## 3. Leonard pairs and algebraic zero-diagonal spaces

In algebraic combinatorics, the phrase acquires a more specialized meaning. A Leonard pair is an ordered pair $(A,A^*)$ of linear transformations such that each is diagonalizable and irreducible tridiagonal with respect to an eigenbasis of the other [1503.05262, 2509.21520]. The paper “Leonard pairs having zero-diagonal TD-TD form” studies the case in which both matrices are irreducible tridiagonal with all diagonal entries equal to $0$ in the same basis [1503.05262]. Writing
\[
Z_d=\{M\in \mathrm{Mat}_{d+1}(\mathbb F): M_{ii}=0,\ 0\le i\le d\},
\]
the relevant space is the intersection of $Z_d$ with the tridiagonal locus, and then the subset of pairs that form a Leonard pair [1503.05262].

The main structural criterion is Theorem 1.8: a Leonard pair admits a zero-diagonal TD–TD representation if and only if it is isomorphic to its opposite $(-A,-A^*)$ [1503.05262]. At the parameter-array level, this is equivalent to antisymmetry of eigenvalues,
\[
\theta_i+\theta_{d-i}=0,\qquad \theta_i^*+\theta_{d-i}^*=0,
\]
and palindromy of split sequences,
\[
\varphi_i=\varphi_{d-i+1},\qquad \phi_i=\phi_{d-i+1},
\]
for the indicated ranges [1503.05262]. The classification then reduces the zero-diagonal TD–TD locus to Krawtchouk type, Bannai–Ito type with even diameter, and $q$-Racah type [1503.05262]. Section 7 of that paper gives five explicit matrix families covering all such Leonard pairs up to the stated equivalences [1503.05262].

A second, more intrinsic usage appears in “Spin Leonard pairs and the zero diagonal space” [2509.21520]. For a Leonard pair $(A,A^*)$ with primitive idempotents $\{E_i^*\}_{i=0}^d$, the zero diagonal space is defined by
\[
\mathcal Z(A,A^*)=\{X\in \mathrm{Span}\{I,A^*,A,AA^*\}: E_i^*XE_i^*=0 \text{ for }0\le i\le d\}.
\]
Relative to the $A^*$-eigenbasis, this means precisely that the diagonal entries of $X$ vanish [2509.21520]. Unlike the matrix subspace $Z_d$, this is a small subspace attached to a specific Leonard pair inside its Terwilliger algebra.

Its dimension is at most $2$: if
\[
M=
\begin{pmatrix}
1 & 1 & \cdots & 1\\
\theta_0^* & \theta_1^* & \cdots & \theta_d^*\\
a_0 & a_1 & \cdots & a_d\\
a_0\theta_0^* & a_1\theta_1^* & \cdots & a_d\theta_d^*
\end{pmatrix},
\]
then
\[
\dim \mathcal Z(A,A^*)=4-\operatorname{rank}(M),
\]
so $0\le \dim\mathcal Z(A,A^*)\le 2$ [2509.21520]. Moreover, $\mathcal Z(A,A^*)\neq 0$ if and only if there exist scalars $f_0,f_1,f_2,f_3$, not all zero, such that
\[
f_0+f_1\theta_i^*+f_2 a_i+f_3 a_i\theta_i^*=0,\qquad 0\le i\le d
\]
[2509.21520]. For self-dual Leonard pairs with $d\ge 3$, the paper proves that spin is equivalent to the nontriviality of this zero diagonal space [2509.21520]. It also classifies, across the 13 Leonard-pair types, when $\mathcal Z(A,A^*)$ is nonzero and gives explicit bases in each case [2509.21520].

Taken together, these two papers show that in Leonard theory “zero diagonal space” can mean either a large ambient subspace of zero-diagonal tridiagonal matrices [1503.05262] or a canonical low-dimensional subspace inside the Terwilliger algebra of a fixed Leonard pair [2509.21520].

## 4. Geometry, flag manifolds, and diagonal loci

In geometry, the diagonal is not the main diagonal of a matrix but the subset
\[
\Delta=\{(x,x):x\in X\}\subseteq X\times X.
\]
The paper “Representing a point and the diagonal as zero loci in flag manifolds” studies when a point in a generalized flag manifold $X=G/P$, or the diagonal in $X\times X$, can be realized as the zero locus of a section of a complex vector bundle [1807.09375]. For a rank-$m$ bundle $\mathcal E\to N$ and a generic section $s$, the zero locus $Z(s)$ is a smooth submanifold of codimension $2m$, and its class is the top Chern class $c_m(\mathcal E)$ [1807.09375].

The diagonal problem asks for a rank $\dim_{\mathbb C}X$ bundle $\mathcal E\to X\times X$ with a generic section whose zero locus is exactly $\Delta$ [1807.09375]. In type $A$, Fulton’s construction yields representability of the diagonal for generalized flag manifolds $SL(k+1)/P$ [1807.09375]. In type $C$, the paper proves representability of points for $\mathrm{Lag}_\omega(\mathbb C^{2k})\simeq Sp(k)/U(k)$, but shows that the diagonal is not representable when $k\equiv 2\pmod 4$ [1807.09375]. The diagonal class is tied to top equivariant Schubert classes and double Schubert polynomials [1807.09375].

This geometric use is formally different from matrix zero-diagonal spaces, but the analogy is explicit: a distinguished diagonal object is characterized by vanishing conditions and by being the zero locus of a section. A plausible implication is that the shared terminology marks a common structural theme rather than a common algebraic formalism.

A still different geometric-spectral role is played by the zero diagonal entry in the AKNS/Zakharov–Shabat system studied in “Matrix Zakharov-Shabat Systems with Zero Diagonal Entry” [2511.15348]. There,
\[
\Sigma=\operatorname{diag}(I_{m_+},0,-I_{m_-}),
\]
so the spectral matrix has a single zero diagonal entry [2511.15348]. The anticommutation condition $\Sigma Q=-Q\Sigma$ forces the potential to be off-diagonal relative to the $\Sigma$-eigenspace decomposition [2511.15348]. The zero eigenvalue creates a distinguished neutral mode, leading to a nonstandard direct and inverse scattering theory that requires dual Jost solutions and a wedge-product construction to recover the missing analytic solution [2511.15348]. Through the inverse scattering transform, this yields the initial-value problem for a system of long-wave–short-wave equations [2511.15348]. Here the “zero diagonal” feature is a spectral singularity built into the Lax pair, rather than a vanishing main diagonal in the usual linear-algebraic sense.

## 5. Topological diagonal properties

Topology uses diagonal language in yet another way. A space $X$ has a zero-set diagonal if the diagonal
\[
\Delta_X=\{(x,x):x\in X\}\subseteq X^2
\]
is the zero-set of a continuous map $f:X^2\to [0,1]$, that is,
\[
\Delta_X=f^{-1}(0)
\]
[1112.0883]. Basile, Bella, and Ridderbos show that if $X$ has a zero-set diagonal and $X^2$ has countable weak extent, then $X$ is submetrizable [1112.0883]. They also prove that if $X$ has a regular $G_\delta$-diagonal and $X^2$ has countable weak extent, then $X$ condenses onto a second countable Hausdorff space [1112.0883]. In this setting the diagonal is a zero locus in the topological sense, and the relevant separation properties are measured by diagonal degrees such as $\Delta_n(X)$ and $s\Delta_n(X)$ [1112.0883].

A related but distinct notion is that of a $\mathbb Q$-diagonal. A space $X$ has a $\mathbb Q$-diagonal if $X^2\setminus \Delta$ has a $K(\mathbb Q)$-directed compact cover [1709.06879]. Feng proves that any compact space with a $\mathbb Q$-diagonal is metrizable, and any Tychonoff space with a $\mathbb Q$-diagonal is cosmic [1709.06879]. The proof relies on the structure of $K(\mathbb Q)$ under Tukey order and on “BIG” subsets of $2^{\omega_1}$ [1709.06879]. This is a diagonal-complement property rather than a matrix-like vanishing condition, but it again treats the diagonal as the central organizing object.

The distinction between zero-dimensionality and stronger diagonal-type properties appears in the 2021 counterexample “A zero-dimensional F-space that is not strongly zero-dimensional” [2108.12903]. That paper constructs a space $X$ that is zero-dimensional and an $F$-space but not strongly zero-dimensional [2108.12903]. Since a Tychonoff space is strongly zero-dimensional if and only if $\beta X$ is zero-dimensional, the example shows that local clopen structure does not force global diagonal-separation behavior in the Čech–Stone compactification [2108.12903]. This topological line of work is conceptually close to diagonal control, but the underlying diagonal is a subset of $X^2$, not a matrix diagonal.

## 6. Comparative perspective and recurring themes

Across these literatures, “zero diagonal space” is best understood as a family of context-dependent constructions linked by a common vanishing motif. Some versions are genuine linear subspaces, such as $Z_e\subset B(H)$ [1410.7441], $Z_d\subset \mathrm{Mat}_{d+1}(\mathbb F)$ [1503.05262], $\mathcal S_n^0$ and $\mathcal H_n^0$ [2309.14958], or the hollow subspace underlying OMZD theory [1810.08961]. Others are orbit-intersection problems, such as orthogonal hollowization [2508.00096] or the classification of zero-diagonal idempotents by Hilbert–Schmidt perturbation theory [1410.7441]. Still others are canonical subspaces attached to structured objects, such as $\mathcal Z(A,A^*)$ for Leonard pairs [2509.21520]. In topology and geometry, the term shifts from vanishing matrix entries to the geometry of the diagonal subset in a product space [1112.0883, 1709.06879, 1807.09375].

Several recurring features nevertheless stand out. First, basis dependence versus intrinsic structure is central: operator zero-diagonality is basis-dependent [1410.7441], whereas $Z_e$ for fixed $e$ is intrinsic to the chosen coordinate system; Leonard-pair zero diagonal spaces are intrinsic once the pair and its idempotents are fixed [2509.21520]. Second, zero-diagonal conditions are frequently linked to strong orbit constraints: not being a Hilbert–Schmidt perturbation of a projection [1410.7441], being isomorphic to the opposite [1503.05262], or being traceless under orthogonal similarity [2508.00096]. Third, extremal and reconstruction problems recur: which diagonal sequences are realizable [1410.7441], what norm inequalities are forced by zero diagonal [2309.14958], which graph patterns admit orthogonal hollow matrices [1810.08961], or when a diagonal subset is representable as a zero locus [1807.09375].

This suggests that “zero diagonal space” functions less as a single technical term than as a reusable structural template. In each domain, a diagonal object is singled out, vanishing on that object defines a class or subspace, and the main questions concern realization, classification, rigidity, and approximation. The differences among these settings are substantial, but the shared formal emphasis on annihilating a diagonal remains the unifying principle.

Source: https://www.emergentmind.com/topics/zero-diagonal-space