Diagonalizable T-Meshes & Spline Stability

• Diagonalizable T-meshes are a class of structured T-meshes where an ordered sequence of l-edges guarantees independent contributions to the spline space dimension.
• They simplify the algebraic complexity by converting the global rank problem into independent one-dimensional Vandermonde-type subproblems.
• Recent approaches recast diagonalizability in terms of T-connected components and t-partitions, enabling robust basis construction and topological dimension formulas.

Searching arXiv for papers on diagonalizable T-meshes and related spline-space dimension results. Diagonalizable T-meshes are a structural class of T-meshes for which the dimension and, in recent work, the stability classification of polynomial spline spaces become tractable in a particularly explicit way. In the smoothing cofactor-conformality framework, the dimension problem is reduced to the rank behavior of a conformality matrix, and diagonalizability identifies precisely those cases where this algebraic complexity can be controlled by an ordering of $l$-edges that yields independent one-dimensional contributions. In the 2012 formulation, diagonalizable T-meshes are those whose interior $l$-edges can be ordered so that each edge contributes sufficiently many new vertices relative to degree and smoothness thresholds, which forces the conformality matrix to have full column rank regardless of knot intervals (Li, 2012). In the 2025 stability-classification program, the same notion is recast for T-connected components through $t$-partitions and full-row-rank one-edge conformality blocks, and is elevated from a dimension-counting device to the central stable case in a hierarchy based on dimensional stability and dimensional absolute stability (Huang et al., 8 Aug 2025). A subsequent basis-construction paper treats diagonalizable T-meshes as the canonical intermediate objects: arbitrary T-meshes are extended to diagonalizable ones, local tensor-product B-splines are assigned componentwise, and redundant extensions are then eliminated by a dedicated constraint system (Zhong et al., 18 Aug 2025).

1. Definitions and equivalent characterizations

The 2012 paper studies spline spaces

$\mathcal{S}(d_{1}, d_{2}, \alpha, \beta, \mathcal{T}) = \left\{ f(x,y)\in C^{\alpha,\beta}(\Omega)\; \middle|\; f|_{\phi}\in \mathrm{P}_{d_{1}d_{2},\ \forall \phi\in\mathcal{F}\right\},$

for a regular T-mesh without holes, with notation including $T^{h},T^{v}$ for the numbers of horizontal and vertical interior $l$-edges, $C^{h},C^{v}$ for cross-cuts, $V$ for interior vertices, $V^+$ for free-vertices, and

$$N^{h}=\left\lceil \frac{d_{1}+1}{d_{1}-\alpha}\right\rceil,\qquad N^{v}=\left\lceil \frac{d_{2}+1}{d_{2}-\beta}\right\rceil.$$

For an ordering of interior $l$0-edges $l$1, the paper defines the new-vertex-vector

$l$2

where $l$3 counts the vertices on $l$4 not lying on earlier listed $l$5-edges. A T-mesh is diagonalizable if there exists an ordering such that

$l$6

and

$l$7

(Li, 2012).

This criterion is intrinsically sequential: diagonalizability is not merely a local property of individual edges, but a property of the existence of an ordering that progressively exposes enough previously unused vertices. The significance of this ordering is algebraic. It permits the conformality matrix to be reduced to a block upper-triangular form whose diagonal blocks are full rank, which is the mechanism behind coordinate-independent dimension formulas (Li, 2012).

The same paper gives a necessary-and-sufficient subset characterization. A T-mesh is diagonalizable if and only if for every subset $l$8 of interior $l$9-edges, at least one of the following holds: there exists a horizontal $t$0-edge in $t$1 with at least $t$2 vertices not lying on the other $t$3-edges in $t$4, or there exists a vertical $t$5-edge in $t$6 with at least $t$7 such vertices (Li, 2012). This formulation rules out irreducibly entangled edge configurations and is the paper’s principal structural test.

The 2025 stability paper adopts a related but differently organized definition. Let a T-connected component $t$8 have ordered T $t$9-edges

$\mathcal{S}(d_{1}, d_{2}, \alpha, \beta, \mathcal{T}) = \left\{ f(x,y)\in C^{\alpha,\beta}(\Omega)\; \middle|\; f|_{\phi}\in \mathrm{P}_{d_{1}d_{2},\ \forall \phi\in\mathcal{F}\right\},$0

The associated $\mathcal{S}(d_{1}, d_{2}, \alpha, \beta, \mathcal{T}) = \left\{ f(x,y)\in C^{\alpha,\beta}(\Omega)\; \middle|\; f|_{\phi}\in \mathrm{P}_{d_{1}d_{2},\ \forall \phi\in\mathcal{F}\right\},$1-partition

$\mathcal{S}(d_{1}, d_{2}, \alpha, \beta, \mathcal{T}) = \left\{ f(x,y)\in C^{\alpha,\beta}(\Omega)\; \middle|\; f|_{\phi}\in \mathrm{P}_{d_{1}d_{2},\ \forall \phi\in\mathcal{F}\right\},$2

is defined by $\mathcal{S}(d_{1}, d_{2}, \alpha, \beta, \mathcal{T}) = \left\{ f(x,y)\in C^{\alpha,\beta}(\Omega)\; \middle|\; f|_{\phi}\in \mathrm{P}_{d_{1}d_{2},\ \forall \phi\in\mathcal{F}\right\},$3, while for $\mathcal{S}(d_{1}, d_{2}, \alpha, \beta, \mathcal{T}) = \left\{ f(x,y)\in C^{\alpha,\beta}(\Omega)\; \middle|\; f|_{\phi}\in \mathrm{P}_{d_{1}d_{2},\ \forall \phi\in\mathcal{F}\right\},$4, $\mathcal{S}(d_{1}, d_{2}, \alpha, \beta, \mathcal{T}) = \left\{ f(x,y)\in C^{\alpha,\beta}(\Omega)\; \middle|\; f|_{\phi}\in \mathrm{P}_{d_{1}d_{2},\ \forall \phi\in\mathcal{F}\right\},$5 is obtained from $\mathcal{S}(d_{1}, d_{2}, \alpha, \beta, \mathcal{T}) = \left\{ f(x,y)\in C^{\alpha,\beta}(\Omega)\; \middle|\; f|_{\phi}\in \mathrm{P}_{d_{1}d_{2},\ \forall \phi\in\mathcal{F}\right\},$6 by removing intersection vertices shared with earlier edges. A T-mesh $\mathcal{S}(d_{1}, d_{2}, \alpha, \beta, \mathcal{T}) = \left\{ f(x,y)\in C^{\alpha,\beta}(\Omega)\; \middle|\; f|_{\phi}\in \mathrm{P}_{d_{1}d_{2},\ \forall \phi\in\mathcal{F}\right\},$7 is diagonalizable if there exists such a $\mathcal{S}(d_{1}, d_{2}, \alpha, \beta, \mathcal{T}) = \left\{ f(x,y)\in C^{\alpha,\beta}(\Omega)\; \middle|\; f|_{\phi}\in \mathrm{P}_{d_{1}d_{2},\ \forall \phi\in\mathcal{F}\right\},$8-partition with

$\mathcal{S}(d_{1}, d_{2}, \alpha, \beta, \mathcal{T}) = \left\{ f(x,y)\in C^{\alpha,\beta}(\Omega)\; \middle|\; f|_{\phi}\in \mathrm{P}_{d_{1}d_{2},\ \forall \phi\in\mathcal{F}\right\},$9

The paper then states an equivalent algebraic condition: $T^{h},T^{v}$0 is diagonalizable if and only if there exists a $T^{h},T^{v}$1-partition such that the conformality matrices $T^{h},T^{v}$2 of the one-edge pieces are all full row rank for $T^{h},T^{v}$3. The equivalence is attributed to the fact that each $T^{h},T^{v}$4 is a Vandermonde-type matrix, and the quoted Vandermonde lemma implies full row rank when the number of vertices is at least $T^{h},T^{v}$5 (Huang et al., 8 Aug 2025).

Taken together, these formulations present two persistent viewpoints. One is combinatorial or geometric, expressed through vertex counts after successive reductions. The other is algebraic, expressed through full-rank one-edge blocks. The later paper makes this equivalence explicit and uses it to connect diagonalizability to stability classes (Huang et al., 8 Aug 2025).

2. Smoothing cofactor-conformality method and the conformality matrix

The foundational algebraic framework is the smoothing cofactor-conformality method. In the 2012 treatment, local continuity relations around an interior vertex $T^{h},T^{v}$6 are written as

$T^{h},T^{v}$7

$T^{h},T^{v}$8

$T^{h},T^{v}$9

$l$0

From these, one derives

$l$1

and since the factors in $l$2 and $l$3 are relatively prime, there exists a polynomial $l$4 of bi-degree $l$5 such that

$l$6

The coefficients of $l$7 are the vertex cofactors, and collecting edge conformality conditions over all $l$8-edges yields a global linear system

$l$9

where $C^{h},C^{v}$0 is the conformality conditions matrix (Li, 2012).

This reduction makes the dimension problem fundamentally a rank problem. For a general regular T-mesh without holes and without vanished $C^{h},C^{v}$1-edges, the paper gives

$C^{h},C^{v}$2

Thus the unresolved term is $C^{h},C^{v}$3, equivalently the rank behavior of the conformality system (Li, 2012).

The 2025 stability paper specializes this perspective to the scalar-degree setting it considers. For each T $C^{h},C^{v}$4-edge with vertices $C^{h},C^{v}$5, the vertex cofactors $C^{h},C^{v}$6 satisfy

$C^{h},C^{v}$7

equivalently

$C^{h},C^{v}$8

The full family of such relations produces $C^{h},C^{v}$9, and the dimension is written as

$V$0

Here again, the decisive question is rank invariance under changes of geometric realization that preserve the combinatorial structure (Huang et al., 8 Aug 2025).

Diagonalizable T-meshes are important because they convert this global rank problem into a structured assembly of one-dimensional Vandermonde-type problems. In the 2012 formulation, diagonalizability implies that $V$1 has full column rank regardless of knot intervals (Li, 2012). In the 2025 reformulation, each reduced edge block is row-full-rank, and the global conformality vector space splits as a direct sum of edgewise components (Huang et al., 8 Aug 2025). The two viewpoints are technically distinct in presentation but aligned in purpose: both isolate a class where algebraic dependence is fully controlled.

3. Dimension formulas and topological determination

For general T-meshes, the conformality term prevents a purely topological dimension formula. The importance of diagonalizability is that it removes this dependence on knot coordinates or, in the later language, on hidden rank instability inside a structurally fixed class.

The 2012 paper proves that if $V$2 is diagonalizable and has no vanished $V$3-edges or holes, then

$V$4

This expression depends only on counts of cross-cuts, interior $V$5-edges, and interior vertices, and therefore is determined by topology or combinatorics rather than by the knot coordinates themselves (Li, 2012).

The 2025 basis-construction paper quotes, for a diagonalizable T-mesh $V$6 with no vanished $V$7-edges and no holes,

$V$8

where $V$9 are the numbers of horizontal and vertical cross-cuts, $V^+$0 the numbers of horizontal and vertical T $V^+$1-edges, and $V^+$2 the number of interior vertices (Zhong et al., 18 Aug 2025). The paper further rewrites this as

$V^+$3

splitting the dimension into tensor-product, T $V^+$4-edge, and ray components (Zhong et al., 18 Aug 2025). This decomposition is used for basis counting rather than for stability classification, but it presupposes the same stable diagonalizable regime.

The 2025 stability paper gives the univariate-degree analogue for a diagonalizable T-mesh with no holes: $V^+$5 where $V^+$6 is the number of cross-cuts, $V^+$7 is the number of T $V^+$8-edges, and $V^+$9 is the number of interior vertices (Huang et al., 8 Aug 2025). It explicitly interprets this as a topological formula: the right-hand side depends only on combinatorial or topological data, not on geometric coordinates, and therefore the dimension is stable under structural isomorphism.

The consistency across these papers is substantial. Although they work with somewhat different settings and notational conventions, all treat diagonalizability as the exact condition under which spline-space dimension ceases to be sensitive to geometric perturbation and becomes computable from discrete structural counts.

4. T-connected components, $$N^{h}=\left\lceil \frac{d_{1}+1}{d_{1}-\alpha}\right\rceil,\qquad N^{v}=\left\lceil \frac{d_{2}+1}{d_{2}-\beta}\right\rceil.$$0-partitions, and decomposition by independent edges

A major conceptual development in the 2025 stability paper is the repositioning of diagonalizability within the structure of the T-connected component $$N^{h}=\left\lceil \frac{d_{1}+1}{d_{1}-\alpha}\right\rceil,\qquad N^{v}=\left\lceil \frac{d_{2}+1}{d_{2}-\beta}\right\rceil.$$1. A T-connected component is the connected union of all T $$N^{h}=\left\lceil \frac{d_{1}+1}{d_{1}-\alpha}\right\rceil,\qquad N^{v}=\left\lceil \frac{d_{2}+1}{d_{2}-\beta}\right\rceil.$$2-edges. The paper emphasizes that the spline-space dimension is controlled not merely by the existence of this component, but by its associated cross-cuts, rays, and vertex arrangement (Huang et al., 8 Aug 2025).

To analyze this structure, the paper introduces a general $$N^{h}=\left\lceil \frac{d_{1}+1}{d_{1}-\alpha}\right\rceil,\qquad N^{v}=\left\lceil \frac{d_{2}+1}{d_{2}-\beta}\right\rceil.$$3-partition

$$N^{h}=\left\lceil \frac{d_{1}+1}{d_{1}-\alpha}\right\rceil,\qquad N^{v}=\left\lceil \frac{d_{2}+1}{d_{2}-\beta}\right\rceil.$$4

with reduced components

$$N^{h}=\left\lceil \frac{d_{1}+1}{d_{1}-\alpha}\right\rceil,\qquad N^{v}=\left\lceil \frac{d_{2}+1}{d_{2}-\beta}\right\rceil.$$5

formed by subtracting intersections with earlier parts. The $$N^{h}=\left\lceil \frac{d_{1}+1}{d_{1}-\alpha}\right\rceil,\qquad N^{v}=\left\lceil \frac{d_{2}+1}{d_{2}-\beta}\right\rceil.$$6-partition is the special case in which each part is a single T $$N^{h}=\left\lceil \frac{d_{1}+1}{d_{1}-\alpha}\right\rceil,\qquad N^{v}=\left\lceil \frac{d_{2}+1}{d_{2}-\beta}\right\rceil.$$7-edge (Huang et al., 8 Aug 2025). This language generalizes the sequential ordering idea already present in the 2012 definition and places it inside a decomposition theory for the entire T-connected component.

For a diagonalizable T-mesh, the paper proves

$$N^{h}=\left\lceil \frac{d_{1}+1}{d_{1}-\alpha}\right\rceil,\qquad N^{v}=\left\lceil \frac{d_{2}+1}{d_{2}-\beta}\right\rceil.$$8

This states that the conformality vector space splits as a direct sum of independent one-dimensional edge contributions (Huang et al., 8 Aug 2025). The result is central because it formally captures the sense in which diagonalizable meshes are “decoupled”: the constraints contributed by different T $$N^{h}=\left\lceil \frac{d_{1}+1}{d_{1}-\alpha}\right\rceil,\qquad N^{v}=\left\lceil \frac{d_{2}+1}{d_{2}-\beta}\right\rceil.$$9-edges do not generate hidden inter-edge rank interactions once the correct partition has been chosen.

The paper also gives a more general $l$00-partition version,

$l$01

when the partition pieces are diagonalizable in the relevant sense (Huang et al., 8 Aug 2025). This extends the edgewise decomposition to more general reduced subcomponents and supplies the formal setting for separating diagonalizable regions from non-diagonalizable ones.

A plausible implication is that the later theory reinterprets diagonalizability less as a global binary label and more as the canonical local building block in a decomposition of arbitrary T-connected components. That interpretation is directly supported by the use of complete partitions and by the basis-construction strategy in the later basis paper, where diagonalizable pieces support local tensor-product B-splines associated to cross-cuts, T $l$02-edges, and rays (Zhong et al., 18 Aug 2025).

5. CNDC, stability hierarchy, and concentration of instability

The 2025 stability paper integrates diagonalizable T-meshes into a broader classification based on dimensional stability. It defines dimensional stability as invariance of $l$03 across the structurally isomorphic class $l$04, and dimensional absolute stability as invariance of the conformality-vector-space dimension across a structurally similar class (Huang et al., 8 Aug 2025). The underlying equivalence relations are carefully distinguished.

A structurally isomorphic map on T-meshes preserves boundary edges, cross-cuts, rays, T-connected components, intersection relations, and the order of horizontal and vertical $l$05-edges. This induces the structurally isomorphic class $l$06. A structurally similar map on generalized T-connected components preserves only connectivity and intersection pattern, not coordinate order, and induces the structurally similar class $l$07 (Huang et al., 8 Aug 2025).

Within this framework, the paper recalls the machinery of regular partitions, complete partitions, and the completely non-diagonalizable component (CNDC). A regular partition splits a T-connected component into

$l$08

where one part is non-diagonalizable and the other is diagonalizable. A complete partition is a maximal such split, and the non-diagonalizable part in the complete partition is the CNDC (Huang et al., 8 Aug 2025).

The key rank-reduction theorem is

$l$09

where $l$10 is the CNDC and $l$11 is the number of T $l$12-edges in the CNDC. Consequently,

$l$13

where $l$14 is the conformality matrix of the CNDC (Huang et al., 8 Aug 2025). The interpretation given in the paper is precise: the diagonalizable part contributes a predictable stable term, and all difficult instability is concentrated in the CNDC.

This reframes diagonalizable T-meshes as the degenerate case of the general classification in which the CNDC is empty. Equivalently, the entire T-connected component splits into independent one-dimensional pieces (Huang et al., 8 Aug 2025). That interpretation also clarifies why diagonalizable meshes occupy the strongest stable position. The paper states explicitly that if $l$15 is diagonalizable, then $l$16 is dimensional stable and dimensional absolute stable. It also proves

$l$17

and conjectures

$l$18

The conjecture is presented as one of the paper’s main classification hypotheses (Huang et al., 8 Aug 2025).

This hierarchy directly addresses a common misconception: diagonalizability is not merely a convenient sufficient condition for closed-form dimension counting. In the 2025 program it is treated as the canonical absolutely stable class, with non-diagonalizable behavior localized to the CNDC. What remains conjectural is whether every absolutely stable case is in fact diagonalizable (Huang et al., 8 Aug 2025).

6. Basis construction and extension to arbitrary T-meshes

The 2025 basis-construction paper places diagonalizable T-meshes at the center of a constructive theory for polynomial spline spaces over arbitrary T-meshes. Its main strategy is to begin with an arbitrary T-mesh $l$19, extend some rays or T $l$20-edges, and obtain an extended mesh $l$21 that is diagonalizable: $l$22 where $l$23 is the full tensor-product mesh (Zhong et al., 18 Aug 2025).

The paper states that for any T-mesh $l$24, there exists some extended mesh $l$25 with $l$26 that is diagonalizable and satisfies the needed local B-spline counting property on each T $l$27-edge. The proof idea described in the paper is orientation-based: extend all horizontal T $l$28-edges and horizontal rays into cross-cuts; then the remaining vertical $l$29-edges are non-vanishing and non-intersecting, so they satisfy the diagonalizability vertex-count condition, with a symmetric argument in the other orientation (Zhong et al., 18 Aug 2025).

Once the extension has been made, basis functions on the diagonalizable mesh are constructed from local tensor-product B-splines assigned to three components determined by the dimension formula. The tensor-product component contributes

$l$30

the T $l$31-edge component contributes

$l$32

and the ray component contributes

$l$33

(Zhong et al., 18 Aug 2025). For each T $l$34-edge $l$35, the paper states that there are at least

$l$36

local tensor-product B-splines associated with it, with $l$37 for horizontal and $l$38 for vertical edges, and these may be selected so that the corresponding one-dimensional B-splines are linearly independent (Zhong et al., 18 Aug 2025). For each ray $l$39, exactly $l$40 such local tensor-product B-splines can be chosen (Zhong et al., 18 Aug 2025).

Theorem $l$41 in that paper states that the union of cross-cut, T $l$42-edge, and ray B-splines forms a basis for $l$43 (Zhong et al., 18 Aug 2025). The linear-independence proof is derivative-based: across a horizontal T $l$44-edge, apply $l$45, which annihilates all basis functions not associated with that edge and reduces the remaining relation to a one-dimensional combination of B-splines, whose linear independence is already established. The same argument is then applied to rays, and the residual tensor-product functions are linearly independent (Zhong et al., 18 Aug 2025).

The final step is Extended Edge Elimination (EEE), which removes redundant edges introduced by extension. For an added horizontal edge $l$46, the EEE condition is

$l$47

with the analogous $l$48-derivative condition for added vertical edges (Zhong et al., 18 Aug 2025). Collecting these over all added edges yields a homogeneous system

$l$49

The paper proves that if $l$50 is a basis of the solution space of this EEE system, then

$l$51

is a basis for $l$52, with

$l$53

(Zhong et al., 18 Aug 2025).

In this construction, diagonalizable T-meshes function as the stable intermediate objects that support a local, componentwise basis. Arbitrary T-meshes are handled by passing through this stable class and then enforcing compatibility with the original unextended geometry.

7. Instability, correction results, and broader significance

The 2012 paper emphasizes that dimension can be instable under previously proposed conditions. It presents a counterexample for the bi-cubic space $l$54: one choice of knots gives dimension $l$55, while a small perturbation changes the dimension to $l$56 (Li, 2012). This example is used to show that the earlier condition of $l$57 is insufficient to guarantee topological stability.

The paper then provides a correction theorem: if a regular T-mesh without holes satisfies that every horizontal $l$58-edge has at least $l$59 mono-vertices and every vertical $l$60-edge has at least $l$61 mono-vertices except the two end vertices, then the dimension formula of the diagonalizable case holds (Li, 2012). The role of the theorem is not to redefine diagonalizability, but to supply a stronger easily checked condition ensuring that the mesh is diagonalizable and hence stable in the desired sense.

The 2025 basis paper reinterprets instability through EEE. It states that when the same topological T-mesh has different geometric coordinates, the number of solutions of the EEE condition can change, causing the dimension of the spline space to change (Zhong et al., 18 Aug 2025). This suggests a direct computational manifestation of the instability detected abstractly in the conformality-matrix language: geometric perturbation changes the number of admissible basis combinations after elimination.

The practical and conceptual importance of diagonalizable T-meshes is therefore consistent across the cited works. They are the class for which spline-space dimension is predictable from topological counts, the class for which conformality constraints decompose into independent one-dimensional pieces, and the class that enables basis construction by local tensor-product building blocks [(Li, 2012); (Huang et al., 8 Aug 2025); (Zhong et al., 18 Aug 2025)]. In the 2025 stability framework, they also anchor a broader taxonomy of stable and unstable meshes, with the CNDC singled out as the sole source of residual coupled instability (Huang et al., 8 Aug 2025).

A plausible implication is that diagonalizable T-meshes serve as the structural analogue of well-posedness in this area: once diagonalizability is present, dimension, rank, and basis construction align. When it fails, the obstruction is no longer diffuse but can be localized to the non-diagonalizable core of the T-connected component. That localization, rather than the original definition alone, is the main conceptual advance of the recent literature (Huang et al., 8 Aug 2025).