1 Graph Equivalence Classes for Spectral Projector-Based Graph Fourier Transforms Joya A. Deri, Member, IEEE, and Josรฉ M. F. Moura, Fellow, IEEE AbstractโWede๏ฌneanddiscusstheutilityoftwoequiv- Consider a graph ๐ข = ๐บ(๐ด) with adjacency matrix alence graphclasses over whicha spectral projector-based ๐ดโC๐ร๐ with ๐ โค๐ distinct eigenvalues and Jordan graph Fourier transform is equivalent: isomorphic equiv- decomposition ๐ด = ๐๐ฝ๐โ1. The associated Jordan 7 alence classes and Jordan equivalence classes. Isomorphic 1 equivalence classes show that the transform is equivalent subspaces of ๐ด are J๐๐, ๐ = 1,...๐, ๐ = 1,...,๐๐, 0 uptoapermutationonthenodelabels.Jordanequivalence where ๐๐ is the geometric multiplicity of eigenvalue ๐๐, 2 classespermitidenticaltransformsovergraphsofnoniden- or the dimension of the kernel of ๐ดโ๐๐๐ผ. The signal n ticaltopologiesandallowabasis-invariantcharacterization space ๐ฎ can be uniquely decomposed by the Jordan of total variation orderings of the spectral components. a subspaces (see [13], [14] and Section II). For a graph J Methodstoexploittheseclassestoreducecomputationtime signal ๐ โ๐ฎ, the graph Fourier transform (GFT) of [12] of the transform as well as limitations are discussed. 1 is de๏ฌned as 1 Index TermsโJordan decomposition, generalized eigenspaces, directed graphs, graph equivalence classes, โจ๏ธ๐ โจ๏ธ๐๐ ] graphisomorphism,signalprocessingongraphs,networks โฑ :๐ฎ โ J๐๐ I S ๐=1 ๐=1 s. ๐ โ(๐ ฬ๏ธ11,...,๐ ฬ๏ธ1๐1,...,๐ ฬ๏ธ๐1,...,๐ ฬ๏ธ๐๐๐), (1) c [ I. INTRODUCTION where๐ ๐๐ isthe(oblique)projectionof๐ ontotheJordan subspace ๐ฝ parallel to ๐ฎโJ . That is, the Fourier 1 Graph signal processing [1], [2] permits applications transform o๐f๐๐ , is the unique de๐c๐omposition v of digital signal processing concepts to increasingly 4 larger networks. It is based on de๏ฌning a shift ๏ฌlter, โ๏ธ๐ โ๏ธ๐๐ 6 ๐ = ๐ , ๐ โJ . (2) for example, the adjacency matrix in [1], [3], [4] to ฬ๏ธ๐๐ ฬ๏ธ๐๐ ๐๐ 8 analyze undirected and directed graphs, or the graph ๐=1๐=1 2 0 Laplacian [2] that applies to undirected graph struc- ThespectralcomponentsaretheJordansubspacesofthe . tures. The graph Fourier transform is de๏ฌned through adjacency matrix with this formulation. 1 the eigendecomposition of this shift operator, see these 0 This paper presents graph equivalence classes where 7 references. Further developments have been considered equal GFT projections by (1) are the equivalence re- 1 in [5], [6], [7]. In particular, ๏ฌlter design [1], [5], [8] lation. First, the transform (1) is invariant to node : andsampling[9],[10],[11]canbeappliedtoreducethe v permutations, which we formalize with the concept of i computational complexity of graph Fourier transforms. X isomorphic equivalence classes. Furthermore, the GFT With the objective of simplifying graph Fourier trans- permits degrees of freedom in graph topologies, which r formsforlargenetworkapplications,thispaperexplores a we formalize by de๏ฌning Jordan equivalence classes, a methods based on graph equivalence classes to reduce concept that allows graph Fourier transform computa- the computation time of the subspace projector-based tions over graphs of simpler topologies. A frequency- graphFouriertransformproposedin[12].Thistransform like ordering based on total variation of the spectral extendsthegraphsignalprocessingframeworkproposed componentsisalsopresentedtomotivatelow-pass,high- by[1],[3],[4]toconsiderspectralanalysisoverdirected pass, and pass-band graph signals. graphs with potentially non-diagonalizable (defective) adjacency matrices. The graph signal processing frame- Section II provides the graph signal processing and work of [12] allows for a unique, unambiguous signal linear algebra background for the graph Fourier trans- representation over defective adjacency matrices. form (1). Isomorphic equivalence classes are de๏ฌned in SectionIII,andJordanequivalenceclassesarede๏ฌnedin ThisworkwaspartiallysupportedbyNSFgrantsCCF-1011903and SectionIV.TheJordanequivalenceclassesin๏ฌuencethe CCF-1513936andanSYS-CMUgrant. de๏ฌnitionoftotalvariation-basedorderingsoftheJordan The authors are with the Department of Electrical and Computer subspaces, which is discussed in detail in Section V. Engineering,CarnegieMellonUniversity,Pittsburgh,PA15213USA (email:jderi,[email protected]) Section VI illustrates Jordan equivalence classes and total variation orderings. Limitations of the method are Jordan decomposition. Let ๐ denote the ๐ ร๐ ๐๐ ๐๐ discussed in Section VII. matrix whose columns form a Jordan chain of eigen- value ๐ that spans Jordan subspace J . Then the ๐ ๐๐ II. BACKGROUND eigenvector matrix ๐ of ๐ด is This section reviews the concepts of graph signal ๐ =[๏ธ๐ ยทยทยท๐ ยทยทยท ๐ ยทยทยท๐ ]๏ธ, (9) processing and the GFT (1). Background on graphs 11 1๐1 ๐1 ๐๐๐ signal processing, including de๏ฌnitions of graph signals where ๐ is the number of distinct eigenvalues. The and the graph shift, is described in greater detail in [1], columns of ๐ are a Jordan basis of C๐. Then ๐ด [3], [4], [12]. For background on eigendecompositions, has block-diagonal Jordan normal form ๐ฝ consisting of the reader is directed to in [13], [15], [16]. Jordan blocks โก โค ๐ 1 A. Eigendecomposition โขโข ๐ ... โฅโฅ valCuoesnsi๐de,r..m.a,t๐rix, ๐ด๐ โโคC๐๐.ร๐Thewiathlge๐brdaiisctinmctuletiipgleinc-- ๐ฝ(๐)=โขโขโฃ ... 1โฅโฅโฆ. (10) 1 ๐ ๐ ity ๐ of ๐ represents the corresponding exponent of ๐ ๐ the characteristic polynomial of ๐ด. Denote by Ker(๐ด) of size ๐ . The Jordan normal form ๐ฝ of ๐ด is unique ๐๐ the kernel or null space of matrix ๐ด. The geometric up to a permutation of the Jordan blocks. The Jordan multiplicity ๐๐ of eigenvalue ๐๐ equals the dimension of decomposition of ๐ด is ๐ด=๐๐ฝ๐โ1. Ker(๐ดโ๐ ๐ผ),whichistheeigenspaceof๐ where๐ผ is ๐ ๐ the ๐ร๐ identity matrix. The generalized eigenspaces B. Spectral Components G , ๐=1,...,๐, of ๐ด are de๏ฌned as ๐ The spectral components of the Fourier trans- G =Ker(๐ดโ๐ ๐ผ)๐๐, (3) form (1) are expressed in terms of the eigenvector ๐ ๐ basis ๐ฃ ,...,๐ฃ and its dual basis ๐ค ,...,๐ค since 1 ๐ 1 ๐ where๐๐ istheindexofeigenvalue ๐๐.Thegeneralized the Jordan basis may not be orthogonal. Denote the eigenspaces uniquely decompose C๐ as the direct sum basis and dual basis matrices by ๐ = [๐ฃ ยทยทยท๐ฃ ] and 1 ๐ ๐ =[๐ค ยทยทยท ,๐ค ]. The dual basis matrix is the inverse ๐ 1 ๐ C๐ =โจ๏ธG . (4) Hermitian ๐ =๐โ๐ป [14], [17]. ๐ Consider the ๐th spectral component of ๐ ๐=1 ๐ Jordan chains. Let ๐ฃ1 โ Ker(๐ด โ ๐๐๐ผ), ๐ฃ1 ฬธ= 0, J๐๐ =span(๐ฃ1,ยทยทยท๐ฃ๐๐๐). (11) be a proper eigenvector of ๐ด that generates generalized The projection matrix onto J parallel to C๐โJ is eigenvectors by the recursion ๐๐ ๐๐ ๐ =๐ ๐๐ป, (12) ๐ด๐ฃ๐ =๐๐๐ฃ๐+๐ฃ๐โ1, ๐=2,...,๐ (5) ๐๐ ๐๐ ๐๐ where where ๐ is the minimal positive integer such that (๐ดโ๐๐๐ผ)๐๐ฃ๐ = 0 and (๐ดโ๐๐๐ผ)๐โ1๐ฃ๐ ฬธ= 0. A se- ๐๐๐ =[๐ฃ1ยทยทยท๐ฃ๐๐๐] (13) quenceofvectors(๐ฃ1,...,๐ฃ๐)thatsatisfy(5)isaJordan isthecorrespondingsubmatrixof๐ and๐๐ป โC๐๐๐ร๐ ๐๐ chainoflength๐ [13].ThevectorsinaJordanchainare is the corresponding submatrix of ๐ partitioned as linearly independent and generate the Jordan subspace ๐ =[ยทยทยท๐๐ปยทยทยท๐๐ป ยทยทยท]๐. (14) J =span(๐ฃ ,๐ฃ ,...,๐ฃ ). (6) ๐1 ๐๐๐ 1 2 ๐ As shown in [12], the projection of signal ๐ โ C๐ Denote by J๐๐ the ๐th Jordan subspace of ๐๐ with onto Jordan subspace J can be written as ๐๐ dimension ๐ , ๐ = 1,...,๐, ๐ = 1,...,๐ . The Jordan ๐๐ ๐ spaces are disjoint and uniquely decompose the general- ๐ ฬ๏ธ๐๐ =๐ ฬ๏ธ1๐ฃ1+ยทยทยท+๐ ฬ๏ธ๐๐๐๐ฃ๐๐๐ (15) ized eigenspace G๐ (3) of ๐๐ as =๐๐๐๐๐๐ป๐๐ . (16) โจ๏ธ๐๐ The next sections show that invariance of the graph G = J . (7) ๐ ๐๐ Fourier transform (1) is a useful equivalence relation on ๐=1 a set of graphs. Equivalence classes with respect to the ThespaceC๐ canbeexpressedastheuniquedecom- GFT are explored in Sections III and IV. position of Jordan spaces III. ISOMORPHICEQUIVALENCECLASSES C๐ =โจ๏ธ๐ โจ๏ธ๐๐ J . (8) ThissectiondemonstratesthatthegraphFouriertrans- ๐๐ form (1) is invariant up to a permutation of node labels ๐=1 ๐=1 2 andestablishessetsofisomorphicgraphsasequivalence with respect to the invariance of the GFT (1) up to a classes with respect to invariance of the GFT (1). Two permutationofthegraphsignalandinversepermutation graphs ๐ข(๐ด) and ๐ข(๐ต) are isomorphic if their adja- of the graph Fourier transform. cency matrices are similar with respect to a permutation Theorem 1 establishes an invariance of the GFT over matrix ๐, or ๐ต = ๐๐ด๐โ1 [18]. The graphs have graphs that only differ up to a node labeling, and the same Jordan normal form and the same spectra. Theorem 2 follows. Also, if ๐ and ๐ are eigenvector matrices of ๐ด ๐ด ๐ต The isomorphic equivalence of graphs is important and ๐ต, respectively, then ๐ = ๐๐ . We prove that ๐ต ๐ด since it signi๏ฌes that the rows and columns of an adja- the set G๐ผ of all graphs that are isomorphic to ๐ข(๐ด) is ๐ด cencymatrix canbepermuted toacceleratethe eigende- an equivalence class over which the GFT is preserved. composition.Forexample,permutationsofhighlysparse Thenexttheoremshowsthatanappropriatepermutation adjacency matrices can convert an arbitrary matrix to can be imposed on the graph signal and GFT to ensure nearly diagonal forms, such as with the Cuthill-McKee invariance of the GFT over all graphs ๐ข โG๐ผ. ๐ด algorithm [19]. Optimizations for such matrices in this Theorem 1. The graph Fourier transform of a signal form are discussed in [16] and [20], for example. In the ๐ is invariant to the choice of graph ๐ข โ G๐ผ up to a next section, the degrees of freedom in graph topology ๐ด permutationonthegraphsignalandinversepermutation are explored to de๏ฌne another GFT equivalence class. on the graph Fourier transform. IV. JORDANEQUIVALENCECLASSES Proof:For๐ข(๐ด),๐ข(๐ต)โG๐ผ,thereexistsapermu- ๐ด tation matrix ๐ such that ๐ต =๐๐ด๐โ1. For eigenvector Since the Jordan subspaces of defective adjacency matrices ๐ and ๐ of ๐ด and ๐ต, respectively, let ๐ matrices are nontrivial (i.e., they have dimension larger ๐ด ๐ต ๐ด,๐๐ and ๐ denote the ๐ร๐ submatrices of ๐ and ๐ than one), a degree of freedom exists on the graph ๐ต,๐๐ ๐๐ ๐ด ๐ต whose columns span the ๐th Jordan subspaces J structure so that the graph Fourier transform of a signal ๐ด,๐๐ andJ ofthe๐theigenvalueof๐ดand๐ต,respectively. is equal over multiple graphs of different topologies. ๐ต,๐๐ Let ๐ = ๐โ๐ป and ๐ = ๐โ๐ป denote the matrices Thissectionde๏ฌnesJordanequivalenceclassesofgraph ๐ด ๐ด ๐ต ๐ต whose columns form dual bases of ๐ and ๐ . Since structures over which the GFT (1) is equal for a given ๐ด ๐ต ๐ =๐๐ , graph signal. The section proves important properties of ๐ต ๐ด this equivalence class that are used to explore inexact ๐ =(๐๐ )โ๐ป (17) ๐ต ๐ด methods and real-world applications in [21]. =(๐โ1๐โ1)๐ป (18) The intuition behind Jordan equivalence is presented ๐ด =๐โ๐ป๐โ๐ป (19) in Section IV-A, and properties of Jordan equivalence ๐ด are described in Section IV-B. Section IV-C com- =๐๐ , (20) ๐ด pares isomorphic and Jordan equivalent graphs. Sec- where ๐โ๐ป =๐ since ๐ is a permutation matrix. Thus, tions IV-D, IV-E, IV-F, and IV-G prove properties for Jordan equivalence classes when adjacency matrices ๐๐ป =๐๐ป๐๐ป =๐๐ป๐โ1. (21) ๐ต ๐ด ๐ด have particular Jordan block structures. Considergraphsignal๐ .By(16),thesignalprojection onto J๐ด,๐๐ is A. Intuition ๐ =๐ ๐๐ป ๐ . (22) Consider Figure 1, which shows a basis {๐} = ฬ๏ธ๐ด,๐๐ ๐ด,๐๐ ๐ด,๐๐ {๐ฃ ,๐ฃ ,๐ฃ } of R3 such that ๐ฃ and ๐ฃ span a two- 1 2 3 2 3 Permit a permutation ๐ =๐๐ on the graph signal. Then dimensional Jordan space J of adjacency matrix ๐ด the projection of ๐ onto J๐ต,๐๐ is with Jordan decomposition ๐ด=๐๐ฝ๐โ1. The resulting ฬ๏ธ๐ ๐ต,๐๐ =๐๐๐ด,๐๐๐๐ด๐ป,๐๐๐โ1๐๐ (23) projection of a signal ๐ โR๐ as in (16) is unique. Notethatthede๏ฌnitionofthetwo-dimensionalJordan =๐๐ ๐๐ป ๐ (24) ๐ด,๐๐ ๐ด,๐๐ subspace J in Figure 1 is not basis-dependent because =๐๐ ฬ๏ธ๐ด,๐๐ (25) anyspanningset{๐ค2,๐ค3}couldbechosentode๏ฌneJ. Thiscanbevisualizedbyrotating๐ฃ and๐ฃ onthetwo- by (22). Therefore, the graph Fourier transform (1) is 2 3 dimensional plane. Any choice {๐ค ,๐ค } corresponds to invariant to a choice among isomorphic graphs up to a 2 3 permutationonthegraphsignalandinversepermutation a new basis ๐ฬ๏ธ. Note that ๐ดฬ๏ธ = ๐ฬ๏ธ๐ฝ๐ฬ๏ธโ1 does not equal ๐ด=๐๐ฝ๐โ1 forallchoicesof{๐ค ,๐ค };theunderlying on the Fourier transform. 2 3 graph topologies may be different, or the edge weights Theorem 2. Consider ๐ด โ C๐ร๐. Then the set G๐ผ maybedifferent.Nevertheless,theirspectralcomponents ๐ด of graphs isomorphic to ๐ข(๐ด) is an equivalence class (the Jordan subspaces) are identical, and, consequently, 3 Theorem 4. For ๐ดโC๐ร๐, the set G๐ฝ of all graphs ๐ด that are Jordan equivalent to ๐ข(๐ด) is an equivalence class with respect to invariance of the GFT (1). Jordanequivalentgraphshaveadjacencymatriceswith identical Jordan subspaces and identical Jordan normal forms.Thisimpliesequivalenceofgraphspectra,proven in Theorem 5 below. Theorem 5. Denote by ฮ and ฮ the sets of eigen- ๐ด ๐ต values of ๐ด and ๐ต, respectively. Let ๐ข(๐ด),๐ข(๐ต)โG๐ฝ. Fig. 1: Projections ๐ and ๐ (shown in red) of a signal ๐ด ฬ๏ธ1 ฬ๏ธ2 Thenฮ =ฮ ;thatis,๐ข(๐ด)and๐ข(๐ต)arecospectral. ๐ (shown in black) onto a nontrivial Jordan subspace (span ๐ด ๐ต of ๐ฃ1 and ๐ฃ2) and the span of ๐ฃ3, respectively, in R3. The Proof:Since๐ข(๐ด)and๐ข(๐ต)areJordanequivalent, projection onto the nontrivial subspace is invariant to basis theirJordanformsareequal,sotheirspectra(theunique choices {๐ฃ ,๐ฃ } (in blue) or {๐ฃ ,๐ฃ } (in green). 1 2 ฬ๏ธ1 ฬ๏ธ2 elements on the diagonal of the Jordan form) are equal. the spectral projections of a signal onto these compo- OnceaJordandecompositionforanadjacencymatrix nents are identical; i.e., the GFT (1) is equivalent over is found, it is useful to characterize other graphs in graphs ๐ข(๐ด) and ๐ข(๐ดฬ๏ธ). This observation leads to the the same Jordan equivalence class. To this end, Theo- de๏ฌnition of Jordan equivalence classes which preserve rem6presentsatransformationthatpreservestheJordan the GFT (1) as well as the underlying structure captured equivalence class of a graph. by the Jordan normal form ๐ฝ of ๐ด. These classes are Theorem 6. Consider ๐ด,๐ต โ C๐ร๐ with Jordan formally de๏ฌned in the next section. decompositions ๐ด = ๐๐ฝ๐โ1 and ๐ต = ๐๐ฝ๐โ1 and eigenvector matrices ๐ =[๐ ] and ๐ =[๐ ], respec- ๐๐ ๐๐ B. De๏ฌnition and Properties tively.Then,๐ข(๐ต)โG๐ฝ ifandonlyif๐ตhaseigenvector ๐ด This section de๏ฌnes the Jordan equivalence class of matrix ๐ = ๐๐ for block diagonal ๐ with invertible graphs, over which the graph Fourier transform (1) is submatrices ๐๐๐ โC๐๐๐ร๐๐๐, ๐=1,...,๐, ๐ =1,...,๐๐. invariant. We will show that certain Jordan equivalence Proof: The Jordan normal forms of ๐ด and ๐ต are classes allow the GFT computation to be simpli๏ฌed. equal. By De๏ฌnition 3, it remains to show J =J so ๐ด ๐ต Consider graph ๐ข(๐ด) where ๐ด has a Jordan chain that ๐ข(๐ต) โ G๐ฝ. The identity J = J must be true that spans Jordan subspace J๐๐ of dimension ๐๐๐ > 1. when span{๐ }๐ด= span{๐ } =๐ดJ ,๐ตwhich implies ๐๐ ๐๐ ๐๐ Then(15),andconsequently,(16),wouldholdforanon- that๐ representsaninvertiblelineartransformationof Jordan basis of J ; that is, a basis could be chosen to ๐๐ ๐๐ the columns of ๐ . Thus, ๐ = ๐ ๐ , where ๐ is ๐๐ ๐๐ ๐๐ ๐๐ ๐๐ ๏ฌnd spectral component ๐ such that the basis vectors ฬ๏ธ๐๐ invertible. De๏ฌning ๐ = diag(๐ ,...,๐ ,...,๐ ) do not form a Jordan chain of ๐ด. This highlights that 11 ๐๐ ๐,๐๐ yields ๐ =๐๐. the Fourier transform (1) is characterized not by the Jordan basis of ๐ด but by the set J = {J } of ๐ด ๐๐ ๐๐ Jordan subspaces spanned by the Jordan chains of ๐ด. C. Jordan Equivalent Graphs vs. Isomorphic Graphs Thus, graphs with topologies yielding the same Jordan This section shows that isomorphic graphs do not subspace decomposition of the signal space have the imply Jordan equivalence, and vice versa. First it is same spectral components. Such graphs are termed Jor- shown that isomorphic graphs have isomorphic Jordan dan equivalent with the following formal de๏ฌnition. subspaces. De๏ฌnition 3 (Jordan Equivalent Graphs). Consider Lemma 7. Consider graphs ๐ข(๐ด),๐ข(๐ต)โG๐ผ so that ๐ด graphs๐ข(๐ด)and๐ข(๐ต)withadjacencymatrices๐ด,๐ต โ ๐ต = ๐๐ด๐โ1 for a permutation matrix ๐. Denote by C๐ร๐. Then ๐ข(๐ด) and ๐ข(๐ต) are Jordan equivalent J and J the sets of Jordan subspaces for ๐ด and ๐ต, ๐ด ๐ต graphs if all of the following are true: respectively.If{๐ฃ ,...,๐ฃ }isabasisofJ โJ ,then 1 ๐ ๐ด ๐ด 1) J๐ด =J๐ต; and thereexistsJ๐ต โJ๐ต withbasis{๐ฅ1,...,๐ฅ๐}suchthat 2) ๐ฝ๐ด = ๐ฝ๐ต (with respect to a ๏ฌxed permutation of [๐ฅ1ยทยทยท๐ฅ๐]=๐[๐ฃ1ยทยทยท๐ฃ๐];i.e.,๐ดand๐ต haveisomorphic Jordan blocks). Jordan subspaces. Let G๐ฝ๐ด denote the set of graphs that are Jordan Proof: Consider ๐ด with Jordan decomposition ๐ด= equivalent to ๐ข(๐ด). De๏ฌnition 3 and (1) establish that ๐๐ฝ๐โ1. Since ๐ต =๐๐ด๐โ1, it follows that G๐ฝ is an equivalence class. ๐ด ๐ต =๐๐๐ฝ๐โ1๐โ1 (26) 4 =๐๐ฝ๐โ1 (27) where ๐ = ๐๐ represents an eigenvector matrix of ๐ต that is a permutation of the rows of ๐. (It is clear that the Jordan forms of ๐ด and ๐ต are equivalent.) Let columns๐ฃ ,...,๐ฃ of๐ denoteaJordanchainof๐ดthat 1 ๐ spansJordansubspaceJ .Thecorrespondingcolumns ๐ด in ๐ are ๐ฅ ,...,๐ฅ and span(๐ฅ ,...,๐ฅ ) = J . Fig. 2: Jordan equivalent graph structures with unicellular 1 ๐ 1 ๐ ๐ต Since [๐ฅ ยทยทยท ๐ฅ ] = ๐[๐ฃ ยทยทยท ๐ฃ ], J and J are adjacency matrices. 1 ๐ 1 ๐ ๐ด ๐ต isomorphic subspaces [13]. Theorem8. AgraphisomorphismdoesnotimplyJordan Proof: A counterexample is provided. The top two equivalence. graphs in Figure 2 correspond to 0/1 adjacency matrices with a single Jordan subspace J = C๐ and eigen- Proof: Consider ๐ข(๐ด),๐ข(๐ต) โ G๐ผ and ๐ต = value 0; therefore, they are Jordan equivalent. On the ๐ด ๐๐ด๐โ1 for permutation matrix ๐. By (27), ๐ฝ๐ด = ๐ฝ๐ต. other hand, they are not isomorphic since the graph on Toshow๐ข(๐ด),๐ข(๐ต)โG๐ฝ,itremainstocheckwhether the right has more edges then the graph on the left. ๐ด J๐ด =J๐ต. Theorem8showsthatchangingthegraphnodelabels By Lemma 7, for any J๐ด โJ๐ด, there exists J๐ต โ may change the Jordan subspaces and the Jordan equiv- J๐ต that is isomorphic to J๐ด. That is, if ๐ฃ1,...,๐ฃ๐ alence class of the graph, while Theorem 9 shows that and ๐ฅ1,...,๐ฅ๐ are bases of J๐ด and J๐ต, respectively, a Jordan equivalence class may include graphs with dif- then [๐ฅ1ยทยทยท๐ฅ๐] = ๐[๐ฃ1ยทยทยท๐ฃ๐]. Checking J๐ด = J๐ต is ferent topologies. Thus, graph isomorphism and Jordan equivalent to checking equivalence are not identical concepts. Nevertheless, the isomorphic and Jordan equivalence classes both imply ๐ผ ๐ฃ +ยทยทยท+๐ผ ๐ฃ =๐ฝ ๐ฅ +ยทยทยท+๐ฝ ๐ฅ (28) 1 1 ๐ ๐ 1 1 ๐ ๐ invariance of the graph Fourier transform with respect =๐ฝ ๐๐ฃ +ยทยทยท+๐ฝ ๐๐ฃ (29) 1 1 ๐ ๐ to equivalence relations as stated in Theorems 1 and 4. Thenexttheoremestablishesanisomorphismbetween for some coef๏ฌcients ๐ผ and ๐ฝ , ๐ = 1,...,๐. How- ๐ ๐ Jordan equivalence classes. ever, (29) does not always hold. Consider matrices ๐ด and ๐ต Theorem 10. If ๐ด,๐ต โ C๐ร๐ and ๐ข(๐ด) and ๐ข(๐ต) [๏ธ2 0 โ1]๏ธ [๏ธ 1 0 0]๏ธ are isomorphic, then their respective Jordan equiva- ๐ด= 0 2 โ1 ,๐ต = โ1 2 0 . (30) lence classes G๐ฝ and G๐ฝ are isomorphic; i.e., any ๐ด ๐ต 0 0 1 โ1 0 2 graph ๐ข(๐ดโฒ) โ G๐ฝ is isomorphic to a graph ๐ข(๐ตโฒ) โ ๐ด These matrices are similar with respect to a permutation G๐ฝ. ๐ต matrix and thus correspond to isomorphic graphs. Their Proof: Let ๐ข(๐ด) and ๐ข(๐ต) be isomorphic by per- Jordan normal forms are both mutation matrix ๐ such that ๐ต = ๐๐ด๐โ1. Consider [๏ธ1 0 0]๏ธ ๐ข(๐ดโฒ) โ G๐ฝ, which implies that Jordan normal forms ๐ด ๐ฝ = 0 2 0 (31) ๐ฝ = ๐ฝ and sets of Jordan subspaces J = J by ๐ดโฒ ๐ด ๐ดโฒ ๐ด 0 0 2 De๏ฌnition 3. Denote by ๐ดโฒ = ๐ ๐ฝ ๐ the Jordan ๐ดโฒ ๐ดโฒ ๐ดโฒ with possible eigenvector matrices ๐ and ๐ given by decomposition of ๐ดโฒ. De๏ฌne ๐ตโฒ = ๐๐ดโฒ๐โ1. It suf๏ฌces ๐ด ๐ต [๏ธ1 1 0]๏ธ [๏ธ1 0 0]๏ธ to show ๐ข(๐ตโฒ)โG๐ฝ๐ต. First simplify: ๐๐ด = 1 0 1 ,๐๐ต = 1 1 0 . (32) ๐ตโฒ =๐๐ดโฒ๐โ1 (33) 1 0 0 1 0 1 =๐๐ ๐ฝ ๐โ1๐โ1 (34) Equation (32) shows that ๐ด and ๐ต both have Jordan ๐ดโฒ ๐ดโฒ ๐ดโฒ subspaces J1 = span([1 1 1]๐) for ๐1 = 1 and =๐๐๐ดโฒ๐ฝ๐ด๐๐ดโโฒ1๐โ1 (since ๐ข(๐ดโฒ)โG๐ฝ๐ด) (35) J21 = span([0 1 0]๐) for one Jordan subspace of =๐๐๐ดโฒ๐ฝ๐ต๐๐ดโโฒ1๐โ1 (since ๐ข(๐ด)โG๐ผ๐ต). (36) ๐ = 2. However, the remaining Jordan subspace is 2 From (36), it follows that ๐ฝ = ๐ฝ . It remains to span([1 0 0]๐) for ๐ด but span([0 0 1]๐) for ๐ต, so (29) ๐ตโฒ ๐ต show that J = J . Choose arbitrary Jordan sub- fails. Thus, ๐ข(๐ด) and ๐ข(๐ต) are not Jordan equivalent. ๐ตโฒ ๐ต space J = span{๐ } of ๐ด. Then J = ๐ด,๐๐ ๐ด,๐๐ ๐ดโฒ,๐๐ span{๐ } = J since ๐ข(๐ดโฒ) โ G๐ฝ. Then the ThenexttheoremshowsthatJordanequivalentgraphs ๐ดโฒ,๐๐ ๐ด,๐๐ ๐ด ๐th Jordan subspace of eigenvalue ๐ for ๐ต is may not be isomorphic. ๐ J =span{๐๐ } (37) Theorem 9. Jordan equivalence does not imply the ๐ต,๐๐ ๐ด,๐๐ =๐span{๐ }. (38) existence of a graph isomorphism. ๐ด,๐๐ 5 For the ๐th Jordan subspace of eigenvalue ๐ for ๐ตโฒ, it Proof: Since the Jordan subspaces of a diagonal- ๐ follows from (36) that izable matrix are one-dimensional, the possible choices of Jordan basis are limited to nonzero scalar multiples J =span{๐๐ } (39) ๐ตโฒ,๐๐ ๐ดโฒ,๐๐ of the eigenvectors. Then, given eigenvector matrix ๐ =๐span{๐๐ดโฒ,๐๐} (40) of ๐ด, all possible eigenvector matrices of ๐ด are given =๐span{๐ } (since ๐ข(๐ดโฒ)โG๐ฝ) by๐ =๐๐,where๐ isadiagonalmatrixwithnonzero ๐ด,๐๐ ๐ด (41) diagonal entries. Let ๐ต = ๐๐ฝ๐โ1, where ๐ฝ is the =J . (by (38)) (42) diagonal canonical Jordan form of ๐ด. Since ๐ and ๐ฝ ๐ต,๐๐ are both diagonal, they commute, yielding Since (42) holds for all ๐ and ๐, the sets of Jordan subspaces J = J . Therefore, ๐ข(๐ตโฒ) and ๐ข(๐ต) are ๐ต =๐๐ฝ๐โ1 (43) ๐ตโฒ ๐ต Jordan equivalent, which proves the theorem. =๐๐๐ฝ๐โ1๐โ1 (44) Theorem10showsthattheJordanequivalenceclasses =๐๐ฝ๐๐โ1๐โ1 (45) of two isomorphic graphs are also isomorphic. This =๐๐ฝ๐โ1 (46) result permits an frequency ordering on the spectral components of a matrix ๐ด that is invariant to both the =๐ด. (47) choice of graph in G๐ฝ and the choice of node labels, as ๐ด Thus, a graph with a diagonalizable adjacency matrix is demonstrated in Section V. theoneandonlyelementinitsJordanequivalenceclass. Relation to matrices with the same set of invariant subspaces. Let GInv denote the set of all matrices ๐ด When a matrix has nondefective but repeated eigen- with the same set of invariant subspaces of ๐ด; i.e., values, there are in๏ฌnitely many choices of eigenvec- ๐ข(๐ต) โ GInv if and only if Inv(๐ด) = Inv(๐ต). The ๐ด tors [16]. An illustrative example is the identity matrix, next theorem shows that GInv is a proper subset of the ๐ด which has a single eigenvalue but is diagonalizable. Jordan equivalence class G๐ฝ of ๐ด. ๐ด Since it has in๏ฌnitely many choices of eigenvectors, the Theorem 11. For ๐ดโC๐ร๐, GInv โG๐ฝ. identity matrix corresponds to in๏ฌnitely many Jordan ๐ด ๐ด equivalenceclasses.ByTheorem12,eachoftheseequiv- Proof: If ๐ข(๐ต) โ GInv, then the set of Jordan ๐ด alenceclasseshavesizeone.Thisobservationhighlights subspaces are equal, or J =J . ๐ด ๐ต that the de๏ฌnition of a Jordan equivalence class requires Theorem 11 sets the results of this chapter apart a choice of basis. from analyses such as those in Chapter 10 of [15], which describes structures for matrices with the same E. One Jordan Block invariant spaces, and [22], which describes the eigen- Consider matrix ๐ด with Jordan decomposition ๐ด = decomposition of the discrete Fourier transform matrix ๐๐ฝ๐โ1 where ๐ฝ is a single Jordan block and ๐ = in terms of projections onto invariant spaces. The Jor- [๐ฃ ยทยทยท๐ฃ ] is an eigenvector matrix. Then ๐ด is a repre- 1 ๐ dan equivalence class relaxes the assumption that all sentation of a unicellular transformation ๐ :C๐ โC๐ invariant subspaces of two adjacency matrices must be with respect to Jordan basis ๐ฃ ,...๐ฃ (see [15, Sec- 1 ๐ equal. This translates to more degrees of freedom in the tion 2.5]). In this case the set of Jordan subspaces has graph topology. The following sections present results one element J =C๐. Properties of unicellular Jordan for adjacency matrices with diagonal Jordan forms, one equivalence classes are demonstrated next. Jordan block, and multiple Jordan blocks. Theorem 13. Let๐ข(๐ด)beanelementoftheunicellular D. Diagonalizable Matrices JordanequivalenceclassG๐ฝ.Thenallgraph๏ฌlters๐ป โ ๐ด If the canonical Jordan form ๐ฝ of ๐ด is diagonal (๐ด G๐ฝ are all-pass. ๐ด is diagonalizable), then there are no Jordan chains and Proof: Since ๐ด is unicellular, it has a single Jordan the set of Jordan subspaces J = {J }๐ where ๐ด ๐ ๐=1 chain๐ฃ ,...,๐ฃ oflength๐.Consideragraphsignal๐ J = span(๐ฃ ) and ๐ฃ is the ๐th eigenvector of ๐ด. 1 ๐ ๐ ๐ ๐ over graph ๐ข(๐ด), and let ๐ represent the coordinate ฬ๏ธ Graphs with diagonalizable adjacency matrices include vector of ๐ in terms of the basis {๐ฃ }๐ . Then the ๐ ๐=1 undirected graphs, directed cycles, and other digraphs spectral decomposition of signal ๐ is given by withnormaladjacencymatricessuchasnormallyregular digraphs [23]. A graph with a diagonalizable adjacency ๐ =๐ ฬ๏ธ1๐ฃ1+ยทยทยท๐ ฬ๏ธ๐๐ฃ๐ =๐ ฬ๏ธ; (48) matrixisJordanequivalentonlytoitself,asprovennext. that is, the unique projection of ๐ onto the spectral Theorem 12. A graph ๐ข(๐ด) with diagonalizable adja- component J = C๐ is itself. Therefore, ๐ข(๐ด) acts as cency matrix ๐ด โ C๐ร๐ belongs to a Jordan equiva- an all-pass ๏ฌlter. Moreover, (48) holds for all graphs in lence class of size one. Jordan equivalence class G๐ฝ. ๐ด 6 Inadditiontotheall-passpropertyofunicellulargraph where Inv(ยท) represents the set of invariant spaces of a ๏ฌlters, unicellular isomorphic graphs are also Jordan matrix). If ๐ = ๐, De๏ฌnition 3 can be applied, which equivalent, as proven next. yields ๐ข(๐ด)โG๐ฝ. ๐ฝ On the other hand, consider a unicellular matrix ๐ต Theorem 14. Let ๐ข(๐ด),๐ข(๐ต) โ G๐ผ where ๐ด is a ๐ด suchthatitseigenvectorisnotinthespanofacanonical unicellular matrix. Then ๐ข(๐ด),๐ข(๐ต)โG๐ฝ. ๐ด vector, e.g., Proof: Since ๐ข(๐ด) and ๐ข(๐ต) are isomorphic, Jor- โก1 โ1 1 1 โค dan normal forms ๐ฝ = ๐ฝ . Therefore, ๐ต is also 2 2 2 2 ๐ด ๐ต unicellular, so J๐ด = J๐ต = {C๐}. By De๏ฌnition 3, ๐ต =โขโข12 โ12 โ12 โ21โฅโฅ (49) ๐ข(๐ด),๐ข(๐ต)โG๐ฝ๐ด. โขโฃ0 0 12 โ21โฅโฆ The dual basis of ๐ can also be used to construct 0 0 1 โ1 2 2 graphs in the Jordan equivalence class of unicellular ๐ด. with Jordan normal form ๐ฝ(0). Since the span of Theorem 15. Denote by ๐ an eigenvector matrix of the eigenvectors of ๐ฝ(0) and ๐ต are not identical, unicellular๐ดโC๐ร๐ and๐ =๐โ๐ป isthedualbasis. Inv(๐ฝ(0))ฬธ=Inv(๐ต).However,byDe๏ฌnition3,๐ข(๐ต)is Consider decompositions ๐ด = ๐๐ฝ๐โ1 and ๐ด๐ = inthesameclassofunicellularJordanequivalentgraphs ๐๐ฝ๐โ1. Then ๐ข(๐ด๐)โG๐ฝ๐ด. as those of Figure 2, i.e., ๐ข(๐ต) โ G๐ฝ๐ฝ. In other words, formatrices๐ดand๐ตwiththesameJordannormalforms Proof: Matrices ๐ด and ๐ด have the same Jordan ๐ (๐ฝ =๐ฝ ),Jordanequivalence,i.e.,J =J ,isamore normalformbyde๏ฌnition.SincethereisonlyoneJordan ๐ด ๐ต ๐ด ๐ต block, both matrices have a single Jordan subspace C๐. generalconditionthanInv(๐ด)=Inv(๐ต).Thisillustrates thatgraphshavingadjacencymatriceswithequalJordan ByDe๏ฌnition3,๐ข(๐ด )and๐ข(๐ด)areJordanequivalent. ๐ normalformsandthesamesetsofinvariantspacesform a proper subset of a Jordan equivalence class, as shown The next theorem characterizes the special case of above in Theorem 11. graphsintheJordanequivalenceclassthatcontains๐ข(๐ฝ) Remark on topology. Note that replacing each with adjacency matrix equal to Jordan block ๐ฝ =๐ฝ(๐). nonzero element of (49) with a unit entry results in a Theorem16. Denoteby๐ฝ =๐ฝ(๐)isthe๐ร๐ Jordan matrix that is not unicellular. Therefore, its correspond- block (10) for eigenvalue ๐. Then ๐ข(๐ด) โ G๐ฝ if ๐ด โ inggraphisnotinaunicellularJordanequivalenceclass. ๐ฝ C๐ร๐ is upper triangular with diagonal entries ๐ and This observation demonstrates that topology may not nonzero entries on the ๏ฌrst off-diagonal. determine the Jordan equivalence class of a graph. Proof: Consider upper triangular matrix ๐ด = F. Two Jordan Blocks [๐ ] with diagonal entries ๐ = ยทยทยท = ๐ and ๐๐ 11 ๐๐ nonzero elements on the ๏ฌrst off-diagonal. By [15, Consider ๐ ร ๐ matrix ๐ด with Jordan normal Example 10.2.1], ๐ด has the same invariant subspaces form consisting of two Jordan subspaces J1 = as ๐ฝ = ๐ฝ(๐), which implies J๐ฝ = J๐ด = {C๐}. span(๐ฃ1,...,๐ฃ๐1) and J2 = span(๐ฃ๐1+1,...,๐ฃ๐2) of Therefore, the Jordan normal form of ๐ด is the Jordan dimensions ๐1 >1 and ๐2 =๐ โ๐1 and corresponding block๐ฝ๐ด =๐ฝ(๐11).Restrictthediagonalentriesof๐ดto eigenvalues๐1 and๐2,respectively.Thespectraldecom- ๐so๐ฝ =๐ฝ.Then,๐ข(๐ฝ),๐ข(๐ด)โG๐ฝ byDe๏ฌnition(3). position of signal ๐ over ๐ข(๐ด) yields ๐ด ๐ฝ ๐ =๐ ๐ฃ +ยทยทยท+๐ ๐ฃ +๐ ๐ฃ +ยทยทยท+๐ ๐ฃ Figure 2 shows graph structures that are in the same โฬ๏ธ1 1 โ ฬ๏ธ๐1 ๐ 1 โฬ๏ธ๐1+1 ๐1+1 โ ฬ๏ธ๐ ๐ unicellular Jordan equivalence class by Theorem 16. In ๐ ฬ๏ธ1 ๐ ฬ๏ธ2 (50) addition,thetheoremimpliesthatitissuf๏ฌcienttodeter- minetheGFTofunicellular๐ดbyreplacing๐ข(๐ด)โG๐ฝ =๐ ฬ๏ธ1+๐ ฬ๏ธ2. (51) ๐ฝ with๐ข(๐ฝ),where๐ฝ isasingle๐ร๐ Jordanblock.That Spectral components ๐ and ๐ are the unique pro- ฬ๏ธ1 ฬ๏ธ2 is,withoutlossofgenerality,๐ข(๐ด)canbereplacedwith jections of ๐ onto the respective Jordan subspaces. By a directed chain graph with possible self-edges and the Example 6.5.4 in [13], a Jordan basis matrix ๐ can be eigenvector matrix ๐ = ๐ผ chosen to compute the GFT chosen for ๐ด = ๐๐ฝ๐โ1 such that ๐ = ๐๐, where ๐ of a graph signal. commutes with ๐ฝ and has a particular form as follows. Remarkoninvariantspaces.Example10.2.1of[15] If ๐ ฬธ= ๐ , then ๐ = diag(๐ ,๐ ), where ๐ , shows that a matrix ๐ดโC๐ร๐ having upper triangular 1 2 1 2 ๐ ๐ = 1,2, is an ๐ ร๐ upper triangular Toeplitz matrix; ๐ ๐ entries with constant diagonal entries ๐ and nonzero otherwise, ๐ has form entries on the ๏ฌrst off-diagonal is both necessary and [๏ธ ]๏ธ suf๏ฌcient for ๐ด to have the same invariant subspaces as ๐ =diag(๐ ,๐ )+ 0 ๐12 (52) ๐ ร๐ Jordan block ๐ฝ =๐ฝ(๐) (i.e., Inv(๐ฝ)=Inv(๐ด), 1 2 ๐ 0 21 7 where ๐ is an ๐ ร๐ upper triangular Toeplitz matrix ๐ฝ (๐) is the ๐ ร๐ Jordan block for eigenvalue ๐. ๐ ๐ ๐ ๐๐ ๐ ๐ and ๐ and ๐ are extended upper triangular Toeplitz 12 21 Proof: By Lemma 10.3.3 in [15], ๐ด with structure matrices as in Theorem 12.4.1 in [13]. Thus, all Jordan as described in the theorem have the same invariant bases of ๐ด can be obtained by transforming eigenvector subspaces as ๐ฝ = diag(๐ฝ (๐),๐ฝ (๐)). Therefore, ๐ด matrix ๐ as ๐ =๐๐. ๐1 ๐2 and ๐ฝ have the same Jordan normal form and Jordan A corresponding theorem to Theorem 16 is presented subspaces and so are Jordan equivalent. tocharacterizeJordanequivalentclasseswhentheJordan Theorems 17 and 18 demonstrate two types of Jordan form consists of two Jordan blocks. The reader is di- equivalences that arise from block diagonal matrices rectedtoSections10.2and10.3in[15]formoredetails. with submatrices of form (53) and (54). These theo- Thefollowingde๏ฌnitionsareneeded.Denote๐ร๐upper rems imply that computing the GFT (1) over the block triangular Toeplitz matrices ๐ (๐ ,...,๐ ) of form ๐2 1 ๐2 diagonal matrices can be simpli๏ฌed to computing the โก โค ๐ ๐ ยทยทยท ๐ ๐ transform over the adjacency matrix of a union of 1 2 ๐โ1 ๐ โขโข0 ๐ ... ๐ ๐ โฅโฅ directed chain graphs. That is, the canonical basis can โข 1 ๐โ2 ๐โ1โฅ be chosen for ๐ without loss of generality. ๐๐(๐1,...,๐๐)=โขโข ... ... ... ... ... โฅโฅ, (53) As for the case of unicellular transformations, it is โข โฅ โข โฅ possible to pick bases of J and J that do not form โฃ0 0 ยทยทยท ๐1 ๐2 โฆ 1 2 a Jordan basis of ๐ด. Any two such choices of bases are 0 0 ยทยทยท 0 ๐ 1 relatedbyTheorem6.Concretely,if๐ istheeigenvector andde๏ฌne๐ร๐ uppertriangularmatrixforsome๐ >๐ matrixof๐ดand๐ isthematrixcorrespondingtoanother choice of basis, then Theorem 6 states that a transfor- ๐ (๐ ,...,๐ ;๐น)= ๐ 1 ๐ mation matrix ๐ can be found such that ๐ = ๐๐, โก๐ ยทยทยท ๐ ๐ ๐ ยทยทยท๐ ๐ โค 1 ๐ 11 12 1,๐โ๐โ1 1,๐โ๐ where ๐ is partitioned as ๐ = diag(๐ ,๐ ) with full- 0 ๐ ยทยทยท ๐ ๐ ยทยทยท๐ ๐ 1 2 โขโขโข ... ...1 ... ๐ .2.2. 2,...๐โ๐โ1 2,๐...โ๐ โฅโฅโฅ rank submatrices ๐๐ โC๐๐ร๐๐, ๐=1,2. โข0 0 ยทยทยท ๐ ๐ โฅ โข ๐ ๐โ๐,๐โ๐โฅ (54) โข0 0 ยทยทยท ๐ ๐ โฅ โข ๐โ1 ๐ โฅ G. Multiple Jordan Blocks โข .. .. .. .. .. โฅ โข . . . . . โฅ โฃ โฆ This section brie๏ฌy describes a special case of Jordan 0 0 ยทยทยท 0 ๐ ๐ 0 0 ยทยทยท 0 01 ๐2 equivalence classes whose graphs have adjacency matri- 1 ces ๐ดโC๐ร๐ with ๐ Jordan blocks, 1<๐<๐. where ๐น =[๐ ] is a (๐โ๐)ร(๐โ๐) upper triangular ๐๐ Consider matrix ๐ด with Jordan normal form ๐ฝ com- matrix and ๐ โ C, ๐ = 1,...,๐. The theorems are ๐ prised of ๐ Jordan blocks and eigenvalues ๐ ,...,๐ . 1 ๐ presented below. By Theorem 10.2.1 in [15], there exists an upper tri- Theorem 17. Consider ๐ด = diag(๐ด ,๐ด ) where each angular ๐ด with Jordan decomposition ๐ด = ๐๐ฝ๐โ1 1 2 matrix ๐ด , ๐ = 1,2, is upper triangular with diag- such that ๐ข(๐ด) โ G๐ฝ. Note that the elements in the ๐ ๐ฝ onal elements ๐ and nonzero elements on the ๏ฌrst Jordan equivalence class G๐ฝ of ๐ข(๐ฝ) are useful since ๐ ๐ฝ off-diagonal. Let ๐ ฬธ= ๐ . Then ๐ข(๐ด) is Jordan signals over a graph in this class can be computed 1 2 equivalent to the graph with adjacency matrix ๐ฝ = with respect to the canonical basis with eigenvector diag(๐ฝ (๐ ),๐ฝ (๐ )) where ๐ฝ (๐ ) is the ๐ ร ๐ matrix ๐ = ๐ผ. Theorem 19 characterizes the possible ๐1 1 ๐2 2 ๐๐ ๐ ๐ ๐ Jordan block for eigenvalue ๐ . eigenvector matrices ๐ such that ๐ด = ๐๐ฝ๐โ1 allows ๐ ๐ข(๐ด)โG๐ฝ. Proof: By Theorem 16, ๐ข(๐ด ) and ๐ข(๐ฝ (๐ )) are ๐ฝ ๐ ๐๐ ๐ Jordan equivalent for ๐ = 1,2 and ๐ด upper triangular Theorem 19. Let ๐ด = ๐๐ฝ๐โ1 be the Jordan decom- ๐ with nonzero elements on the ๏ฌrst off-diagonal. There- position of ๐ด โ C๐ร๐ and ๐ข(๐ด) โ G๐ฝ. Then ๐ must ๐ฝ fore, the Jordan normal forms of ๐ฝ and ๐ด are the same. be an invertible block diagonal matrix. Moreover, the set of irreducible subspaces of ๐ฝ is the Proof: Consider ๐ข(๐ฝ) with eigenvector matrix ๐ผ. unionoftheirreduciblesubspacesof[๐ฝ 0]๐ and[0๐ฝ ]๐, 1 2 By Theorem 6, ๐ข(๐ด)โG๐ฝ implies which are the same as the irreducible subspaces of ๐ฝ [๐ด1 0]๐ and [0 ๐ด2]๐, respectively. Therefore, J๐ด =J๐ฝ, ๐ =๐ผ๐ =๐ (55) so ๐ข(๐ด) and ๐ข(๐ฝ) are Jordan equivalent. where ๐ is an invertible block diagonal matrix. Theorem 18. Consider ๐ด=diag(๐ด ,๐ด ) where ๐ด = 1 2 1 Thestructureof๐ giveninTheorem19allowsachar- ๐๐๐1โฅ(๐๐,๐.1,T.h.e.n,๐๐ข๐2(,๐ด๐น))isanJodrd๐ดan2 e=quiv๐a๐l2e(n๐t,t๐o1,th.e..g,๐ra๐2p)h, acterizationofgraphsintheJordanequivalenceclassG๐ฝ๐ฝ 1 2 with the dual basis of ๐ as proved in Theorem 20. with adjacency matrix ๐ฝ = diag(๐ฝ (๐),๐ฝ (๐)) where ๐1 ๐2 8 Theorem 20. Let ๐ข(๐ด) โ G๐ฝ, where ๐ด has Jordan can be chosen without loss of generality. ๐ฝ decomposition ๐ด=๐๐ฝ๐โ1 and ๐ =๐โ๐ป is the dual Jordan equivalence classes show that the GFT (1) basis of ๐. If ๐ด =๐๐ฝ๐โ1, then ๐ข(๐ด )โG๐ฝ. permits degrees of freedom in graph topologies. This ๐ ๐ ๐ฝ has rami๏ฌcations for total variation-based orderings of Proof: By Theorem 19, ๐ is block diagonal with thespectralcomponents,asdiscussedinthenextsection. invertible submatrices ๐ . Thus, ๐ = ๐โ๐ป is block ๐ diagonal with submatrices ๐ = ๐โ๐ป. By Theorem 6, ๐ ๐ V. FREQUENCYORDERINGOFSPECTRAL ๐ is an appropriate eigenvector matrix such that, for COMPONENTS ๐ด =๐๐ฝ๐โ1, ๐ข(๐ด )โG๐ฝ. ๐ ๐ ๐ฝ Thissectionde๏ฌnesamappingofspectralcomponents Relation to graph topology. Certain types of ma- to the real line to achieve an ordering of the spectral trices have Jordan forms that can be deduced from components. This ordering can be used to distinguish their graph structure. For example, [24] and [25] relate generalizedlowandhighfrequenciesasin[4].Anupper the Jordan blocks of certain adjacency matrices to a bound for a total-variation based mapping of a spectral decomposition of their graph structures into unions of component (Jordan subspace) is derived and generalized cycles and chains. Applications where such graphs are to Jordan equivalence classes. inusewouldallowapractitionertodeterminetheJordan The graph total variation of a graph signal ๐ โ C๐ equivalence classes (assuming the eigenvalues can be is de๏ฌned as [4] computed) and potentially choose a different matrix in the class for which the GFT can be computed more TV (๐ )=โ๐ โ๐ด๐ โ . (56) ๐บ 1 easily. Sections IV-E and IV-G show that working with Matrix ๐ด can be replaced by ๐ดnorm = 1 ๐ด when unicellular matrices and matrices in Jordan normal form |๐max| the maximum eigenvalue satis๏ฌes |๐ |>0. permits the choice of the canonical basis. In this way, max Equation (56) can be generalized to de๏ฌne the total for matrices with Jordan blocks of size greater than variation of the Jordan subspaces of the graph shift ๐ด one, ๏ฌnding a spanning set for each Jordan subspace as described in [12]. Choose a Jordan basis of ๐ด so may be more ef๏ฌcient than attempting to compute the that๐ istheeigenvectormatrixof๐ด,i.e.,๐ด=๐๐ฝ๐โ1, Jordan chains. Nevertheless, relying on graph topology where ๐ฝ is the Jordan form of ๐ด. Partition ๐ into is not always possible. Such an example was presented ๐ ร๐ submatrices ๐ whose columns are a Jordan in Section IV-E with adjacency matrix (49). ๐๐ ๐๐ chain of (and thus span) the ๐th Jordan subspace J ๐๐ Relation to algebraic signal processing. The emer- of eigenvalue ๐ , ๐=1,...,๐ โค๐, ๐ =1,...,๐ . Then ๐ ๐ gence of Jordan equivalence from the graph Fourier the (graph) total variation of ๐ is de๏ฌned as [12] ๐๐ transform(1)isrelatedtoalgebraicsignalprocessingand thesignalmodel(๐,โณ,ฮฆ),where๐isasignalalgebra TV๐บ(๐๐๐)=โ๐๐๐ โ๐ด๐๐๐โ1, (57) corresponding to the ๏ฌlter space, โณ is a module of ๐ whereโยทโ representstheinducedL1matrixnorm(equal corresponding to the signal space, and ฮฆ : ๐ โ โณ 1 to the maximum absolute column sum). is a bijective linear mapping that generalizes the ๐ง- Theorem 21 shows that the graph total variation of transform [26], [27]. We emphasize that the GFT (1) is a spectral component is invariant to a relabeling of tied to a basis. This is most readily seen by considering the graph nodes; that is, the total variations of the diagonaladjacencymatrix๐ด=๐๐ผ,whereanybasisthat spectral components for graphs in the same isomorphic spansC๐ de๏ฌnestheeigenvectors(theJordansubspaces equivalence class as de๏ฌned in Section III are equal. and spectral components) of a graph signal; that is, a matrix, even a diagonalizable matrix, may not have Theorem 21. Let ๐ด,๐ต โC๐ร๐ and ๐ข(๐ต)โG๐ผ, i.e., ๐ด distinctspectralcomponents.Similarly,thesignalmodel ๐ข(๐ต) is isomorphic to ๐ข(๐ด). Let ๐๐ด,๐๐ โ C๐ร๐๐๐ be (๐,โณ,ฮฆ) requires a choice of basis for module (signal a Jordan chain of matrix ๐ด and ๐๐ต,๐๐ โ C๐ร๐๐๐ the space) โณ in order to de๏ฌne the frequency response corresponding Jordan chain of ๐ต. Then (irreducible representation) of a signal [26]. On the TV (๐ )=TV (๐ ). (58) other hand, this section demonstrated the equivalence of ๐บ ๐ด,๐๐ ๐บ ๐ต,๐๐ the GFT (1) over graphs in Jordan equivalence classes, Proof: Since ๐ข(๐ด) and ๐ข(๐ต) are isomorphic, there which implies an equivalence of certain bases. This exists a permutation matrix ๐ such that ๐ต = ๐๐ด๐โ1 observation suggests the concept of equivalent signal and the eigenvector matrices ๐ and ๐ of ๐ด and ๐ต, ๐ด ๐ต models in the algebraic signal processing framework. respectively,arerelatedby๐ =๐๐ .Thus,theJordan ๐ต ๐ด Just as working with graphs that are Jordan equivalent chains are related by ๐ =๐๐ . By (57), ๐ต,๐๐ ๐ด,๐๐ to those with adjacency matrices in Jordan normal form TV(๐ )=โ๐ โ๐ต๐ โ (59) simpli๏ฌes GFT computation, we expect similar classes ๐ต ๐ต,๐๐ ๐ต,๐๐ 1 ofequivalentsignalmodelsforwhichthecanonicalbasis =โฆโฆ๐๐๐ด,๐๐ โ(๏ธ๐๐ด๐โ1)๏ธ๐๐๐ด,๐๐โฆโฆ1 (60) 9 =โ๐๐ โ๐๐ด๐ โ (61) composition ๐ต = ๐๐ฝ๐โ1. Let the columns of eigen- ๐ด,๐๐ ๐ด,๐๐ 1 =โ๐ (๐๐ด,๐๐ โ๐ด๐๐ด,๐๐)โ1 (62) vector submatrix ๐๐๐ span the Jordan subspace J๐๐ of ๐ด. Then the class total variation of spectral component =โ๐ โ๐ด๐ โ (63) ๐ด,๐๐ ๐ด,๐๐ 1 J is de๏ฌned as the supremum of the graph total ๐๐ =TV(๐๐ด), (64) variation of ๐๐๐ over the Jordan equivalence class (for all ๐ข(๐ต)โG๐ฝ): where(63)holdsbecausethemaximumabsolutecolumn ๐ด sumofamatrixisinvarianttoapermutationonitsrows. TV (J )= sup TV (๐ ). (68) G๐ฝ ๐๐ ๐บ ๐๐ ๐ด ๐ข(๐ต)โG๐ฝ ๐ด Theorem 21 shows that the graph total variation is ๐ต=๐๐ฝ๐โ1 invariant to a node relabeling, which implies that an span{๐๐๐}=J๐๐ โ๐๐๐โ1=1 ordering of the total variations of the frequency com- Theorem 23. Let ๐ด,๐ต โC๐ร๐ and ๐ข(๐ต)โG๐ผ. Let ponents is also invariant. ๐ด ๐ and ๐ be the respective eigenvector matrices with Reference [12] demonstrates that each eigenvector ๐ด ๐ต Jordan subspaces J = span{๐ } and J = submatrix corresponding to a Jordan chain can be nor- ๐ด,๐๐ ๐ด,๐๐ ๐ต,๐๐ span{๐ } spanned by the ๐th Jordan chain of eigen- malized. This is stated as a property below: ๐ต,๐๐ value ๐ . Then TV (J )=TV (J ). ๐ G๐ฝ ๐ด,๐๐ G๐ฝ ๐ต,๐๐ Property 22. The eigenvector matrix ๐ of adjacency ๐ด ๐ต matrix ๐ด โ C๐ร๐ can be chosen so that each Jordan Proof: Let ๐๐ด* denote the eigenvector matrix corre- sponding to ๐ข(๐ด*)โG๐ฝ that maximizes the class total chain represented by the eigenvector submatrix ๐ โ ๐ด C๐ร๐๐๐ satis๏ฌes โ๐๐๐โ1 = 1; i.e., โ๐โ1 = 1 wit๐h๐out variation of Jordan subspace J๐ด,๐๐; i.e., loss of generality. TV (J )=TV (๏ธ๐* )๏ธ. (69) G๐ฝ ๐ด,๐๐ ๐บ ๐ด,๐๐ ๐ด Itisassumedthattheeigenvectormatricesarenormal- Similarly, let ๐* denote the eigenvector matrix corre- ๐ต ized as in Property 22 for the remainder of the section. sponding to ๐ข(๐ต*)โG๐ฝ that maximizes the class total ๐ต Furthermore, [12] shows that (57) can be written as variation of Jordan subspace J , or ๐ต,๐๐ TV๐บ(๐๐๐)=โฆโฆ๐๐๐(๏ธ๐ผ๐๐๐ โ๐ฝ๐๐)๏ธโฆโฆ1 (65) TVG๐ฝ (J๐ต,๐๐)=TV๐บ(๏ธ๐๐ต*,๐๐)๏ธ. (70) ๐ต = max {|1โ๐|โ๐ฃ โ ,โ(1โ๐)๐ฃ โ๐ฃ โ }. ๐=2,...,๐๐๐ 1 1 ๐ ๐โ1 1 Since ๐ข(๐ด) and ๐ข(๐ต) are isomorphic, Theorem 10 (66) implies that there exists ๐ข(๐ตโฒ) โ G๐ฝ such that ๐ตโฒ = ๐ต ๐๐ด*๐โ1; i.e., ๐ =๐๐* where ๐ is an eigenvector and establishes the upper bound for the total variation ๐ตโฒ ๐ด ๐ตโฒ matrixof ๐ตโฒ. Bythe classtotal variationde๏ฌnition (68), of spectral components as TV (๐ ) โค TV (๐* ). Applying Theorem 21 to ๐บ ๐ตโฒ,๐๐ ๐บ ๐ต,๐๐ TV๐บ(๐๐๐)โค|1โ๐๐|+1. (67) isomorphic graphs ๐ข(๐ด*) and ๐ข(๐ตโฒ) yields Equations(65),(66),and(67)characterizethe(graph) TV (๏ธ๐* )๏ธ=TV (๐ )โคTV (๏ธ๐* )๏ธ. (71) ๐บ ๐ด,๐๐ ๐บ ๐ตโฒ,๐๐ ๐บ ๐ต,๐๐ total variation of a Jordan chain by quantifying the Similarly, by Theorem 10, there exists ๐ข(๐ดโฒ) โ G๐ฝ change in a set of vectors that spans the Jordan sub- ๐ด space J๐๐ when they are transformed by the graph asuncehigtehnavte๐ตct*or=m๐at๐ดrixโฒ๐oโf1๐ด,โฒo.rA๐p๐ตp*ly=(6๐8๐)๐ดanโฒdwThheereor๐em๐ดโฒ2i1s shift ๐ด. These equations, however, are dependent on again to obtain a particular choice of Jordan basis. As seen in Sec- tionsIV-E,IV-F,andIV-G,defectivegraphshiftmatrices TV (๏ธ๐* )๏ธโฅTV (๐ )=TV (๏ธ๐* )๏ธ. (72) ๐บ ๐ด,๐๐ ๐บ ๐ดโฒ,๐๐ ๐บ ๐ต,๐๐ belong to Jordan equivalence classes that contain more than one element, and the GFT of a signal is the same Equations (71) and (72) imply that TV๐บ(๏ธ๐๐ด*,๐๐)๏ธ = over any graph in a given Jordan equivalence class. TV๐บ(๏ธ๐๐ต*,๐๐)๏ธ, or Furthermore, for any two graphs ๐ข(๐ด),๐ข(๐ต) โ G๐ฝ, ๐ด ๐ด TV (J )=TV (J ). (73) and๐ต haveJordanbasesforthesameJordansubspaces, G๐ฝ๐ด ๐ด,๐๐ G๐ฝ๐ต ๐ต,๐๐ buttherespectivetotalvariationsofthespanningJordan chains as computed by (65) may be different. Since it Theorem 23 shows that the class total variation of is desirable to be able to order spectral components in a a spectral component is invariant to a relabeling of mannerthatisinvarianttothechoiceofJordanbasis,we the nodes. This is signi๏ฌcant because it means that an derivehereade๏ฌnitionofthetotalvariationofaspectral ordering of the spectral components by their class total component of ๐ด in relation to the Jordan equivalence variations is invariant to node labels. class G๐ฝ. Next, the signi๏ฌcance of the class total variation (68) ๐ด Class total variation. Let ๐ข(๐ต) be an element in isillustratedforadjacencymatriceswithdiagonalJordan Jordan equivalence class G๐ฝ where ๐ต has Jordan de- form, one Jordan block, and multiple Jordan blocks. ๐ด 10