WorksMathematics
Dimensionally absolutely stable but non-diagonalizable -meshes
Abstract
We construct an infinite family of rectangular -meshes whose conformality-vector-space dimension is constant over all structurally similar geometric realizations, although the meshes are not diagonalizable. For every integer , the construction has \Tl-edges and intersection graph . Every \Tl-edge contains at least vertices, so no edge is vanishable under the usual vertex-count criterion for bidegree and smoothness . We prove that the conformality-vector space is one-dimensional throughout the structural similarity class. Its generator is the outer product of two vectors of barycentric interpolation weights, extended by four zero coordinates. Non-diagonalizability follows by examining the last edge in any proposed diagonalizing ordering. Thus, for the geometric structural similarity class specified here, diagonalizability is not a necessary condition for dimensional absolute stability.
Introduction
The dimension of a spline space over a -mesh is closely related to the rank of its conformality matrix. For splines of bidegree and smoothness , the constraints associated with a -edge can be written as polynomial moment equations. Their coefficients depend on the coordinates of the vertices on that edge. This is part of a broader literature on dimension instability and stability of spline spaces over -meshes; see Li and Chen [1], Berdinsky et al. [2], and Guo et al. [3]. The smoothing cofactor-conformality method is developed in Li and Deng [4]. Stability properties for highest-order smoothness spaces over -meshes, including the notion of dimensional absolute stability, have been studied by Huang and Chen [5, 6].
Diagonalizability is an ordering property of the -edges. It provides a combinatorial mechanism for controlling these constraints. The converse question is whether coordinate-independent dimension, even throughout a structural similarity class, forces such an ordering to exist.
This paper gives a negative answer for the geometric structural similarity class defined below. For each , we construct a rectangular -mesh with no vanishable -edges whose conformality-vector-space dimension is identically one throughout that class, but for which no diagonalizing ordering exists.
The construction uses a complete bipartite intersection pattern. This pattern forces a shared set of coordinates on all edges of each orientation. The resulting compatibility equations admit a rank-one factorization involving classical barycentric interpolation weights.
To make the scope of the result precise, we explicitly state the structural similarity and stability conventions used here. In particular, the realizations under consideration are geometric collections of horizontal and vertical segments in the plane, not abstract incidence systems with independently assigned local coordinates.
Definitions and conventions
Fix an integer . The associated spline setting is bidegree with smoothness .
A rectangular -mesh is a partition of a rectangle into axis-aligned rectangles, allowing -junctions. An interior -edge is a maximal connected straight union of interior mesh edges. A cross-cut has both endpoints on the boundary of the domain. A -edge has both endpoints at interior -junctions.
For an -edge , let be its set of vertices, including its endpoints, and write
A -edge is called vanishable if
A -connected component is a connected collection of -edges, where connectivity is determined by shared vertices. Within such a component, a vertex belonging to two component edges is called a multi-vertex, and a vertex belonging to exactly one component edge is called a mono-vertex.
Geometric structural similarity
The following two definitions adapt the notion of a generalized -connected component and the structurally similar map of Huang and Chen [6] to the planar geometric setting used in this paper.
Definition 2.1 (Geometric realization). A geometric component consists of finitely many horizontal and vertical closed line segments in , together with a finite set of distinct marked vertices on each segment. Endpoints and intersections of component segments are marked.
Segments of the same orientation do not share a vertex. Segments of opposite orientations meet in at most one vertex. Every marked vertex belongs to one or two component segments. The segment intersection graph is connected.
The segments in Definition 2.1 represent the -edges of a generalized geometric component. Additional marked vertices represent mono-vertices. The common-coordinate requirement in Definition 2.1 (same orientation segments do not share a vertex) is essential for the analysis below; it is precisely the planar geometric counterpart of the common-coordinate property discussed by Huang and Chen [6].
Definition 2.2 (Structural similarity). Two geometric components and are structurally similar if there is a bijection
between their edge sets such that
for every edge , and
for every pair of distinct edges.
This is the geometric counterpart of the structurally similar map of Huang and Chen [6]: the condition plays the role of the incidence-preservation condition in the original definition.
The class of all such geometric realizations is denoted by .
Structural similarity does not prescribe numerical coordinates, distances, or the order of mono-vertices. It also permits an exchange of the two coordinate directions.
Conformality-vector spaces
Assign one real scalar to each distinct vertex of a component . On a horizontal edge, let be the horizontal coordinate of ; on a vertical edge, let be its vertical coordinate.
The moment conditions on an edge are
At a multi-vertex, the same variable is used in the conditions for both incident component edges.
The conformality-vector space is
Equivalently, , where is the matrix of these moment conditions.
Definition 2.3 (Dimensional absolute stability; Huang and Chen [6]). A component is dimensionally absolutely stable if is independent of .
A -mesh with a single -connected component is called dimensionally absolutely stable if that component has this property.
All meshes constructed below have exactly one -connected component.
Diagonalizability
Definition 2.4 (Diagonalizability; Li and Deng [4], see also Huang and Chen [6], Definition 2.3). A component with edges is diagonalizable if its edges admit an ordering such that
A -mesh with one -connected component is diagonalizable if its component is diagonalizable.
Remark 2.5 (Scope of the definitions). The common-coordinate property of a planar geometric realization is essential below. The argument does not assert stability for a different model in which each edge is assigned its own independent set of local vertex coordinates. Accordingly, applying the result to an externally formulated conjecture requires matching that conjecture’s definitions with Definitions –.
A classical Vandermonde identity
We use the following standard consequence of the divided difference formula and the Lagrange interpolation formula; see de Boor [7] and Berrut and Trefethen [8]. The explicit vector consists of the barycentric interpolation weights. We state the result without proof.
Lemma 3.1 (One-dimensional Vandermonde null space). Let , and let be pairwise distinct real numbers. Set
Then and is one-dimensional. More explicitly, if
then
defines a generator of . Every component of is nonzero. Consequently, every nonzero vector in has no zero component.
Construction and non-diagonalizability
Let
and consider the rectangle
Insert the two vertical cross-cuts
Insert the vertical segments
the two horizontal segments
and the remaining horizontal segments
Together with the boundary of , these segments define a -mesh denoted by .
Proposition 4.1. The construction defines a rectangular -mesh. Its Tl-edges are precisely
They form one -connected component , whose intersection graph is . Moreover,
In particular, no Tl-edge is vanishable.
Proof. The two cross-cuts partition into three vertical strips. The two outer strips are rectangles. Within the middle strip, and separate a lower rectangle, an upper rectangle, and a central rectangle.
The central rectangle is partitioned by the into two side rectangles and interior columns. The segments subdivide those interior columns into rectangles. Thus all mesh cells are rectangles.
Each and ends at -junctions on the cross-cuts. Each ends at -junctions on and . Each remaining ends at -junctions on and . The listed segments are maximal in their respective directions.
Every meets every , and no two listed edges of the same orientation meet. Hence the intersection graph is .
Every edge contains multi-vertices. The two edges and additionally contain two mono-vertices each, at the cross-cuts. There are no other mono-vertices. The asserted counts follow.
Finally, , so every -edge contains more than vertices.
Proposition 4.2. For every , the -mesh is not diagonalizable.
Proof. Suppose that a diagonalizing ordering exists and consider its last edge . All other component edges precede .
If is a vertical edge or one of , all its vertices belong to preceding edges. Thus it contributes no new vertex.
If is or , all its multi-vertices belong to preceding vertical edges. Only its two mono-vertices remain.
In either case the number of remaining vertices is at most two. Since , this is less than , contradicting (3).
Absolute stability
The essential point is to treat every member of the structural similarity class, not merely coordinate perturbations of the displayed mesh.
Lemma 5.1 (Geometry of the similarity class). Let . After relabeling edges and, if necessary, exchanging the coordinate directions, its edges can be denoted by
with the following properties:
every is horizontal and every is vertical;
every meets every ;
and each contain two mono-vertices;
all remaining vertices are the intersections .
Furthermore, the supporting lines of the have pairwise distinct horizontal coordinates , and the supporting lines of the have pairwise distinct vertical coordinates .
Proof. By Proposition 4.1, the edge-incidence graph of is . Structural similarity preserves the edge-incidence relation, so has the same incidence graph . Since is connected, its bipartition is unique up to exchanging the two parts. In an axis-aligned geometric realization, the horizontal and vertical orientation classes give such a bipartition; hence, after possibly exchanging the coordinate directions, the two partite sets are precisely (horizontal) and (vertical). This proves properties 1 and 2.
Structural similarity also preserves the number of vertices on every corresponding edge. By (8), in the two edges have vertices while all other edges have vertices. Hence the same vertex counts hold in . Every edge meets exactly edges of the opposite orientation, and each such intersection is a distinct vertex, so every edge has exactly multi-vertices. An edge with vertices therefore has no mono-vertex, whereas an edge with vertices has exactly two mono-vertices. This proves properties 3 and 4.
Finally, by Definition 2.1, two same-orientation segments do not share a vertex. If two vertical edges had the same horizontal coordinate, then their intersections with any horizontal edge would coincide, making them share a vertex, a contradiction. Hence the are pairwise distinct. The same argument applies to the . □
The labels and in Lemma 5.1 identify the two special edges; they need not be the geometrically lowest and highest edges of an arbitrary realization.
Theorem 5.2 (Constant conformality-vector-space dimension). For every ,
More explicitly, let
and define
Then every conformality vector has intersection coordinates
for some , and all four mono-vertex coordinates are zero.
Proof. Use the labeling from Lemma 5.1. Set
Since , these are exactly the moment matrices on edges having no mono-vertices.
By Lemma 3.1,
and every component of and is nonzero.
Let denote the coordinate associated with . Each has only its intersection vertices. Its moment conditions therefore give
Consequently, for some ,
Writing , the intersection-coordinate matrix is .
There is an ordinary horizontal edge, for example . Its moment conditions yield
Since , it follows that . Hence for some , which proves (9). All remaining ordinary horizontal equations are then automatically satisfied.
Consider either special horizontal edge , where . Let its mono-vertices have horizontal coordinates and , with corresponding conformality coordinates and . These are distinct vertices on a horizontal line, so . For every , the contribution of its intersection vertices is
Thus the remaining conditions are
Taking gives
The matrix is nonsingular because , so .
We have shown that every conformality vector has the asserted form. Conversely, (9), extended by four zero mono-vertex coordinates, satisfies every moment condition because is not required: the relevant identities are and . In particular, the horizontal equations use and the vertical equations use .
Taking gives a nonzero vector, and the solution space is therefore exactly one-dimensional.
Remark 5.3. With a fixed vectorization convention, the full space is
The four final entries correspond to the mono-vertices.
Theorem 5.4 (An infinite counterexample family). For every , the rectangular -mesh has no vanishable -edges, is dimensionally absolutely stable in the sense of Definition 2.3, and is not diagonalizable.
Proof. Proposition 4.1 gives the vertex-count condition. Theorem 5.2 gives dimensional absolute stability, and Proposition 4.2 gives non-diagonalizability.
Corollary 5.5 (Rank of the conformality matrix). For every , the matrix has rows, columns, and rank
Its row-rank deficiency is
Proof. There are edges, each contributing moment equations. There are multi-vertices and four mono-vertices. The rank follows from Theorem 5.2 and the rank-nullity theorem. □
The lowest-degree example in the family
For , we have and
The cross-cuts are and . The vertical -edges are
and the horizontal -edges are
The relevant counts are
Thus there are eight -edges and twenty component vertices. The conformality matrix has size and rank 19.
Figure 1. The -mesh . Filled circles are multi-vertices of its -connected component. Open circles are its four mono-vertices. The two full-height interior lines are cross-cuts.
Since both coordinate sets are equally spaced with unit spacing, one may use
as a generator of each relevant Vandermonde null space. Accordingly, an intersection-coordinate generator is
extended by four zero mono-vertex coordinates.
This is the lowest-degree member of the present family. No claim is made that it minimizes the number of edges or vertices among all possible counterexamples.
Discussion
The example distinguishes constant rank from the existence of a diagonalizing ordering. Its conformality matrix has coordinate-independent rank throughout the specified structural similarity class, despite having unavoidable row dependencies.
The complete bipartite intersection pattern is central. It forces every vertical edge to meet the same set of horizontal supporting lines and every horizontal edge to meet the same set of vertical supporting lines. The vertical constraints first force the intersection matrix to have the form . One ordinary horizontal edge then forces . The remaining four mono-vertex variables vanish.
The failure of diagonalizability has a separate combinatorial cause. Every component edge has fewer than mono-vertices. Thus no edge can occupy the last position of a diagonalizing ordering.
The restriction is necessary for this last-edge argument. For , two remaining mono-vertices meet the threshold , and the argument no longer applies.
Under Definitions 2.1–2.4, the implication
is therefore false. Any application to another notion of generalized component must first verify the common-coordinate requirement described in Remark 2.5.
Declaration of competing interest
The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.
Data availability
No external datasets were used. All constructions and calculations needed to establish the results are included in the article.
References
- [1]X. Li, F. Chen, On the instability in the dimension of spline spaces over T-meshes, Computer Aided Geometric Design 28 (7) (2011) 420–426.DOI
- [2]D. Berdinsky, M. Oh, T. Kim, B. Mourrain, On the problem of instability in the dimension of a spline space over a T-mesh, Computers & Graphics 36 (5) (2012) 507–513.DOI
- [3]Q. Guo, R. Wang, C. Li, On the problem of instability in the dimensions of spline spaces over T-meshes with T-cycles, Journal of Computational Mathematics 33 (3) (2015) 248–262.
- [4]X. Li, J. Deng, On the dimension of spline spaces over T-meshes with smoothing cofactor-conformality method, Computer Aided Geometric Design 41 (2016) 76–86. doi:10.1016/j.cagd.2015.12.002.DOI
- [5]B. Huang, F. Chen, On the stability of the dimensions of spline spaces with highest order of smoothness over T-meshes, Journal of Computational and Applied Mathematics 441 (2024) 115681. doi:10.1016/j.cam.2023.115681.DOI
- [6]B. Huang, F. Chen, A preliminary study on the dimensional stability classification of polynomial spline spaces over T-meshes, Computer Aided Geometric Design 130 (2026) 102609. arXiv:2508.06217v2, doi:10.1016/j.cagd.2026.102609.DOI
- [7]C. de Boor, Divided differences, Surveys in Approximation Theory 1 (2005) 46–69.arxiv.org/abs/math/0502036
- [8]J.-P. Berrut, L. N. Trefethen. Barycentric Lagrange interpolation. SIAM Review 46 (3) (2004) 501–517. doi:10.1137/S0036144502417715.DOI
