Lipschitz property for systems of linear mappings and bilinear forms
Abdullah Alazemi, Milica An{\dj}eli\'c, Carlos M. da Fonseca, Vladimir, V. Sergeichuk

TL;DR
This paper establishes a Lipschitz continuity property for the isomorphism classes of representations of graphs with both undirected and directed edges, linking small perturbations to small changes in the isomorphism.
Contribution
It proves that for such graph representations, isomorphisms can be chosen to vary continuously with the representations, providing a Lipschitz-type stability result.
Findings
Isomorphic representations close to each other admit near-identity isomorphisms.
The result applies to systems with bilinear forms and linear maps on graph edges.
Provides a stability guarantee for graph representation isomorphisms.
Abstract
Let G be a graph with undirected and directed edges. Its representation is given by assigning a vector space to each vertex, a bilinear form on the corresponding vector spaces to each directed edge, and a linear map to each directed edge. Two representations A and A' of G are called isomorphic if there is a system of linear bijections between the vector spaces corresponding to the same vertices that transforms A to A'. We prove that if two representations are isomorphic and close to each other, then their isomorphism can be chosen close to the identity.
Peer Reviews
No public reviews on file for this paper yet. If you reviewed it on a platform where reviews are public (OpenReview, ICLR, NeurIPS, ICML), you can paste yours below so the community can read it here.
Videos
No videos yet. Explain this paper in a talk, walkthrough, or lecture? Add one.
Lipschitz property for systems of linear mappings and bilinear forms
Abdullah Alazemi
Department of Mathematics, Kuwait University, Safat 13060, Kuwait
Milica Anđelić
Carlos M. da Fonseca
Kuwait College of Science and Technology, Safat 13133, Kuwait
University of Primorska, FAMNIT, Glagoljsaška 8, 6000 Koper, Slovenia
[email protected], [email protected]
Vladimir V. Sergeichuk
Institute of Mathematics, Tereshchenkivska 3, Kiev, Ukraine
Abstract
Let be a graph with undirected and directed edges. Its representation is given by assigning a vector space to each vertex, a bilinear form on the corresponding vector spaces to each directed edge, and a linear map to each directed edge. Two representations and of are called isomorphic if there is a system of linear bijections between the vector spaces corresponding to the same vertices that transforms to . We prove that if two representations are isomorphic and close to each other, then their isomorphism can be chosen close to the identity.
keywords:
15A21; 15A63; 16G20
MSC:
Systems of operators and forms; tensors of order two; Lipschitz property; representations of quivers and graphs
1 Introduction
If two square complex matrices are similar and close to each other, then they are similar via a matrix that can be chosen close to the identity matrix. This fundamental fact is known as the local Lipschitz property for similarity (Gohberg and Rodman [3]); many mathematical constructions are based on it.
For example, the affine tangent space at the point to the orbit of under similarity is the set since by the Lipschitz property each matrix that is similar to and close to has the following form with a small :
[TABLE]
Pierce and Rodman proved analogous local Lipschitz properties for congruence [7], for simultaneous congruence and simultaneous unitary congruence [8], for matrix group actions that contain simultaneous similarity and simultaneous equivalence [9], and joint similarity and congruence-like actions [11]; see also [10, 12].
The following theorem is a special case of [11, Corollary 1.3(1)].
Theorem 1.1**.**
For each sequence of complex matrices and each , there exist positive real numbers and with the following property. Let be a sequence of complex matrices such that and
[TABLE]
for some nonsingular . Then the equalities (1.1) also hold for some nonsingular satisfying .
Here is any matrix norm, although Rodman [11] uses the operator norm , in which stands for the Euclidean norm of vectors.
The goal of this paper is to give an independent proof of a more general statement about representations of bidirected graphs (Theorem 2.3) using the Lipschitz property for matrix pairs with respect to similarity and two known facts about systems of linear mappings and bilinear forms.
2 Representations of bidirected graphs
We consider systems of linear mappings and bilinear forms as representations of bidirected graphs, which are the graphs with undirected, directed, and bidirected edges; for example,
[TABLE]
The vertices of bidirected graphs that we consider are natural numbers (). Multiple edges and loops are allowed.
Definition 2.2** ([14]).**
Let be a bidirected graph.
A representation of over a field is given by assigning
- (a)
a finite dimensional vector space over to each vertex ,
- (b)
a bilinear form to each undirected edge with ,
- (c)
a linear mapping to each directed edge ,
- (d)
a bilinear form on the dual spaces to each bidirected edge with ( denotes the dual space of all linear forms ).
- 2.
The dimension of a representation is the vector .
- 3.
An isomorphism \varphi:\mathcal{A}\xrightarrow{\text{\raisebox{-3.0pt}{\sim}}}\mathcal{B} of two representations and of of the same dimension is a system of linear bijections that transforms to ; that is, for each (), for each , and for each (.
- 4.
A representation of over a field in which all vector spaces are of the form is called a matrix representation. (All forms and linear mappings of a matrix representation are given by matrices.)
For example, a representation of (2.1) is a system
[TABLE]
consisting of vector spaces , bilinear forms and , linear mappings and , and bilinear forms on the dual spaces and .
The forms and mappings in (2.2) can be considered as elements of tensor products , , , , and , ; that is, as tensors of types , , and . Therefore, each representation of a bidirected graph is a system of vector spaces and tensors of order 2.
If all edges of are directed, then is a quiver and its representations are quiver representations. If all edges are directed and undirected, then is a mixed graph; its representations are considered in [5].
Each representation of a bidirected graph is isomorphic to a matrix representation; therefore, it is enough to study only matrix representations; they are given by matrices assigned to all edges. The norm of a matrix representation is the sum of norms of its matrices. For definiteness, we use the Frobenius norm
[TABLE]
of a complex matrix , although we could use an arbitrary matrix norm.
The main result of the paper is the following theorem, which is the local Lipschitz property for representations of bidirected graphs.
Theorem 2.3**.**
Let be a complex matrix representation of dimension of a bidirected graph . There exist positive real numbers and with the following property: for every complex matrix representation that is isomorphic to and satisfies there exists an isomorphism \varphi=(S_{1},\dots,S_{t}):\mathcal{B}\xrightarrow{\text{\raisebox{-3.0pt}{\sim}}}\mathcal{A} such that
[TABLE]
We show in Section 3 that Theorem 2.3 follows easily from the Lipschitz property for complex matrix pairs with respect to similarity (which is proved in [9]) and the following known facts:
the problem of classifying complex matrix pairs with respect to similarity transformations
[TABLE]
contains the problem of classifying complex representations of an arbitrary quiver (which is proved in [1, 2]), and
- 2.
the problem of classifying complex representations of a bidirected graph is reduced to the problem of classifying complex representations of some quiver (which is proved in [14]).
3 Proof of Theorem 2.3
We first prove the Lipschitz property for quiver representations and then extend it to representations of bidirected graphs.
3.1 From matrix pairs to quiver representations
Let us prove the following theorem, which is the global Lipschitz property for representations of quivers.
Theorem 3.4**.**
Let be a complex matrix representation of dimension of a quiver . There exists a positive real number such that for every complex matrix representation that is isomorphic to there exists an isomorphism \varphi=(S_{1},\dots,S_{t}):\mathcal{B}\xrightarrow{\text{\raisebox{-3.0pt}{\sim}}}\mathcal{A} satisfying
[TABLE]
Let us consider complex matrix representations
[TABLE]
(where with summands) of the quiver
[TABLE]
Define the matrix pair
[TABLE]
by the representation .
Lemma 3.5** ([1, Theorem 2.1]).**
The matrix representations and in (3.1) are isomorphic if and only if the pairs and are similar.
Proof.
Each matrix representation that is isomorphic to has the form
[TABLE]
where are nonsingular matrices, which can be considered as the change of basis matrices.
. Let
[TABLE]
for some nonsingular . The equality implies that . We conclude from the second equality in (3.4) that acts on by similarity transformations as follows:
[TABLE]
Since these transformations preserve the th block , . We have that is of the form (3.3), and so is isomorphic to .
. Let be isomorphic to . Then is of the form (3.3), and so with . ∎
Proof of Theorem 3.4.
For the sake of clarity, we prove Theorem 3.4 for quiver representations (3.1); its proof for representations of an arbitrary quiver is analogous.
By [9, Theorem 3.1], the global Lipschitz property holds for complex matrix pairs with respect to similarity. Applying it to the matrix pair (3.2), we find that there is a positive real number with the following property: for every complex matrix pair that is similar to there exists a nonsingular matrix satisfying
[TABLE]
Let a representation of the form (3.1) be isomorphic to . By Lemma 3.5, the pairs and are similar, which allows us to take in (3.5) and (3.6). Since , . The equality ensures that . Hence is the representation (3.3) and \varphi:=(S_{1},S_{2},S_{3}):\mathcal{B}\xrightarrow{\text{\raisebox{-3.0pt}{\sim}}}\mathcal{A}. By (2.3) and (3.6),
[TABLE]
3.2 From quiver representations to representations of bidirected graphs
Let us recall the method that reduces the problem of classifying systems of linear mappings and forms to the problem of classifying systems of linear mappings. This method was developed by Roiter and Sergeichuk [13, 14]; it is used in [4, 15].
For every bidirected graph , we denote by the quiver obtained from by replacing
each vertex of by the vertices and ,
- 2.
each arrow by the arrows and ,
- 3.
each edge by and ,
- 4.
each edge () by and .
We put and for all vertices and arrows. The mappings and are involutions on the sets of vertices and arrows of the quiver .
For example,
[TABLE]
For every complex matrix representation of a bidirected graph , we define the complex matrix representation of with the same and with for each edge of . For example,
[TABLE]
for complex matrix representations of (3.7).
The following lemma is a special case of [13, Theorem] or [14, Theorem 2].
Lemma 3.6**.**
Let and be two complex matrix representations of a bidirected graph . Then and are isomorphic if and only if and are isomorphic.
Proof.
For clarity, we prove Lemma 3.6 for matrix representations of the bidirected graph given in (3.7); its proof for matrix representations of an arbitrary bidirected graph is analogous.
Let be obtained from in (3.8) by replacing with .
. Let \varphi=(\Phi_{1},\Phi_{2}):\mathcal{B}\xrightarrow{\text{\raisebox{-3.0pt}{\sim}}}\mathcal{A}. Then
[TABLE]
is the isomorphism \underline{\varphi}=(\Phi_{1},\Phi_{2},\Phi_{1^{*}},\Phi_{2^{*}}):=(\Phi_{1},\Phi_{2},\Phi_{1}^{-T},\Phi_{2}^{-T}):\underline{\mathcal{B}}\xrightarrow{\text{\raisebox{-3.0pt}{\sim}}}\underline{\mathcal{A}}.
. Let
[TABLE]
Then \psi^{\circ}:=(R^{T},S^{T},P^{T},Q^{T}):\underline{\mathcal{A}}\xrightarrow{\text{\raisebox{-3.0pt}{\sim}}}\underline{\mathcal{B}} and
[TABLE]
Take a nonzero polynomial and consider the morphisms of quiver representations
[TABLE]
We must choose the polynomial such that has the form (3.9); that is,
[TABLE]
The first equality is equivalent to , and so the equalities (3.11) are equivalent to , which are equivalent to
[TABLE]
Such an exists since for each nonsingular complex matrix there is a polynomial in whose square is ; see Kaplansky [6, Theorem 68]. We obtain of the form (3.9); its first two matrices define the isomorphism
[TABLE]
which proves Lemma 3.6. ∎
Proof of Theorem 2.3.
For clarity, we prove Theorem 2.3 for matrix representations of the bidirected graph in (3.7).
Let be a complex matrix representation of . We must prove that there exist positive numbers and with the following property: for every matrix representation that is isomorphic to and satisfies there exists an isomorphism \varphi=(\Phi_{1},\Phi_{2}):\mathcal{B}\xrightarrow{\text{\raisebox{-3.0pt}{\sim}}}\mathcal{A} such that
[TABLE]
Let be any matrix representation of that is isomorphic to . Then and are isomorphic representations of and
[TABLE]
By Theorem 3.4, there exists (the same for all ) and an isomorphism (3.10) satisfying
[TABLE]
Let us prove that (3.12) is a desired isomorphism; that is, there exists (the same for all that are sufficiently close to ) such that
[TABLE]
Let us find for the first inequality:
(i) Write and , in which and are sufficiently small matrices. Then
[TABLE]
Since for a sufficiently small , we have . Hence
[TABLE]
(ii) For each , we have
[TABLE]
Taking and , we get for a sufficiently small .
(iii) Write , in which is a sufficiently small matrix. Since , we have , and so
[TABLE]
Hence, .
By (3.14), (i), and (iii),
[TABLE]
Therefore, we can take in the first inequality of (3.13). Analogously, we can take in the second inequality. ∎
The reference list from the paper itself. Each links out to its DOI / PubMed record.
- 1[1] G.R. Belitskii, V.V. Sergeichuk, Complexity of matrix problems, Linear Algebra Appl. 361 (2003) 203–222.
- 2[2] I.M. Gelfand, V.A. Ponomarev, Remarks on the classification of a pair of commuting linear transformations in a finite dimensional vector space, Funct. Anal. Appl. 3 (1969) 325–326.
- 3[3] I. Gohberg, L. Rodman, On the distance between lattices of invariant subspaces of matrices, Linear Algebra Appl. 76 (1986) 85–120.
- 4[4] R.A. Horn, V.V. Sergeichuk, Canonical matrices of bilinear and sesquilinear forms, Linear Algebra Appl. 428 (2008) 193–223.
- 5[5] R.A. Horn, V.V. Sergeichuk, Representations of quivers and mixed graphs. Chapter 34 in: L. Hogben (Ed.), Handbook of Linear Algebra, 2nd ed., CRC Press, 2014.
- 6[6] I. Kaplansky, Linear Algebra and Geometry; a Second Course, Chelsea Publishing Company, New York, 1974.
- 7[7] S. Pierce, L. Rodman, Congruences and norms of Hermitian matrices, Canad. J. Math. 39 (1987) 1446–1458.
- 8[8] S. Pierce, L. Rodman, On simultaneous congruence and norms of Hermitian matrices, SIAM J. Matrix Anal. Appl. 12 (1991) 70–83.
