ebook img

Graph Equivalence Classes for Spectral Projector-Based Graph Fourier Transforms PDF

0.79 MBยท
by ย Joya A. Deri
Save to my drive
Quick download
Download
Most books are stored in the elastic cloud where traffic is expensive. For this reason, we have a limit on daily download.

Preview Graph Equivalence Classes for Spectral Projector-Based Graph Fourier Transforms

1 Graph Equivalence Classes for Spectral Projector-Based Graph Fourier Transforms Joya A. Deri, Member, IEEE, and Josรฉ M. F. Moura, Fellow, IEEE Abstractโ€”Wede๏ฌneanddiscusstheutilityoftwoequiv- Consider a graph ๐’ข = ๐บ(๐ด) with adjacency matrix alence graphclasses over whicha spectral projector-based ๐ดโˆˆC๐‘ร—๐‘ with ๐‘˜ โ‰ค๐‘ distinct eigenvalues and Jordan graph Fourier transform is equivalent: isomorphic equiv- decomposition ๐ด = ๐‘‰๐ฝ๐‘‰โˆ’1. The associated Jordan 7 alence classes and Jordan equivalence classes. Isomorphic 1 equivalence classes show that the transform is equivalent subspaces of ๐ด are J๐‘–๐‘—, ๐‘– = 1,...๐‘˜, ๐‘— = 1,...,๐‘”๐‘–, 0 uptoapermutationonthenodelabels.Jordanequivalence where ๐‘”๐‘– is the geometric multiplicity of eigenvalue ๐œ†๐‘–, 2 classespermitidenticaltransformsovergraphsofnoniden- or the dimension of the kernel of ๐ดโˆ’๐œ†๐‘–๐ผ. The signal n ticaltopologiesandallowabasis-invariantcharacterization space ๐’ฎ 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‘ โˆˆ๐’ฎ, the graph Fourier transform (GFT) of [12] of the transform as well as limitations are discussed. 1 is de๏ฌned as 1 Index Termsโ€”Jordan decomposition, generalized eigenspaces, directed graphs, graph equivalence classes, โจ๏ธ๐‘˜ โจ๏ธ๐‘”๐‘– ] graphisomorphism,signalprocessingongraphs,networks โ„ฑ :๐’ฎ โ†’ J๐‘–๐‘— I S ๐‘–=1 ๐‘—=1 s. ๐‘ โ†’(๐‘ ฬ‚๏ธ€11,...,๐‘ ฬ‚๏ธ€1๐‘”1,...,๐‘ ฬ‚๏ธ€๐‘˜1,...,๐‘ ฬ‚๏ธ€๐‘˜๐‘”๐‘˜), (1) c [ I. INTRODUCTION where๐‘ ๐‘–๐‘— isthe(oblique)projectionof๐‘ ontotheJordan subspace ๐ฝ parallel to ๐’ฎโˆ–J . That is, the Fourier 1 Graph signal processing [1], [2] permits applications transform o๐‘–f๐‘—๐‘ , is the unique de๐‘–c๐‘—omposition v of digital signal processing concepts to increasingly 4 larger networks. It is based on de๏ฌning a shift ๏ฌlter, โˆ‘๏ธ๐‘˜ โˆ‘๏ธ๐‘”๐‘– 6 ๐‘ = ๐‘  , ๐‘  โˆˆJ . (2) for example, the adjacency matrix in [1], [3], [4] to ฬ‚๏ธ€๐‘–๐‘— ฬ‚๏ธ€๐‘–๐‘— ๐‘–๐‘— 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๏ฌned 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, ๏ฌlter 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๏ฌning 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๏ฌned in SectionIII,andJordanequivalenceclassesarede๏ฌnedin ThisworkwaspartiallysupportedbyNSFgrantsCCF-1011903and SectionIV.TheJordanequivalenceclassesin๏ฌ‚uencethe CCF-1513936andanSYS-CMUgrant. de๏ฌnitionoftotalvariation-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‘ ร—๐‘Ÿ ๐‘–๐‘— ๐‘–๐‘— discussed in Section VII. matrix whose columns form a Jordan chain of eigen- value ๐œ† that spans Jordan subspace J . Then the ๐‘– ๐‘–๐‘— II. BACKGROUND eigenvector matrix ๐‘‰ of ๐ด is This section reviews the concepts of graph signal ๐‘‰ =[๏ธ€๐‘‰ ยทยทยท๐‘‰ ยทยทยท ๐‘‰ ยทยทยท๐‘‰ ]๏ธ€, (9) processing and the GFT (1). Background on graphs 11 1๐‘”1 ๐‘˜1 ๐‘˜๐‘”๐‘˜ signal processing, including de๏ฌnitions 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 ๐ด [3], [4], [12]. For background on eigendecompositions, has block-diagonal Jordan normal form ๐ฝ consisting of the reader is directed to in [13], [15], [16]. Jordan blocks โŽก โŽค ๐œ† 1 A. Eigendecomposition โŽขโŽข ๐œ† ... โŽฅโŽฅ valCuoesnsi๐œ†de,r..m.a,t๐œ†rix, ๐ด๐‘˜ โˆˆโ‰คC๐‘๐‘.ร—๐‘Thewiathlge๐‘˜brdaiisctinmctuletiipgleinc-- ๐ฝ(๐œ†)=โŽขโŽขโŽฃ ... 1โŽฅโŽฅโŽฆ. (10) 1 ๐‘˜ ๐œ† ity ๐‘Ž of ๐œ† represents the corresponding exponent of ๐‘– ๐‘– the characteristic polynomial of ๐ด. Denote by Ker(๐ด) of size ๐‘Ÿ . The Jordan normal form ๐ฝ of ๐ด is unique ๐‘–๐‘— the kernel or null space of matrix ๐ด. The geometric up to a permutation of the Jordan blocks. The Jordan multiplicity ๐‘”๐‘– of eigenvalue ๐œ†๐‘– equals the dimension of decomposition of ๐ด is ๐ด=๐‘‰๐ฝ๐‘‰โˆ’1. Ker(๐ดโˆ’๐œ† ๐ผ),whichistheeigenspaceof๐œ† where๐ผ is ๐‘– ๐‘– the ๐‘ร—๐‘ identity matrix. The generalized eigenspaces B. Spectral Components G , ๐‘–=1,...,๐‘˜, of ๐ด are de๏ฌned as ๐‘– The spectral components of the Fourier trans- G =Ker(๐ดโˆ’๐œ† ๐ผ)๐‘š๐‘–, (3) form (1) are expressed in terms of the eigenvector ๐‘– ๐‘– basis ๐‘ฃ ,...,๐‘ฃ and its dual basis ๐‘ค ,...,๐‘ค since 1 ๐‘ 1 ๐‘ where๐‘š๐‘– istheindexofeigenvalue ๐œ†๐‘–.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‘‰ = [๐‘ฃ ยทยทยท๐‘ฃ ] and 1 ๐‘ ๐‘Š =[๐‘ค ยทยทยท ,๐‘ค ]. The dual basis matrix is the inverse ๐‘˜ 1 ๐‘ C๐‘ =โจ๏ธG . (4) Hermitian ๐‘Š =๐‘‰โˆ’๐ป [14], [17]. ๐‘– Consider the ๐‘—th spectral component of ๐œ† ๐‘–=1 ๐‘– Jordan chains. Let ๐‘ฃ1 โˆˆ Ker(๐ด โˆ’ ๐œ†๐‘–๐ผ), ๐‘ฃ1 ฬธ= 0, J๐‘–๐‘— =span(๐‘ฃ1,ยทยทยท๐‘ฃ๐‘Ÿ๐‘–๐‘—). (11) be a proper eigenvector of ๐ด that generates generalized The projection matrix onto J parallel to C๐‘โˆ–J is eigenvectors by the recursion ๐‘–๐‘— ๐‘–๐‘— ๐‘ƒ =๐‘‰ ๐‘Š๐ป, (12) ๐ด๐‘ฃ๐‘ =๐œ†๐‘–๐‘ฃ๐‘+๐‘ฃ๐‘โˆ’1, ๐‘=2,...,๐‘Ÿ (5) ๐‘–๐‘— ๐‘–๐‘— ๐‘–๐‘— where where ๐‘Ÿ is the minimal positive integer such that (๐ดโˆ’๐œ†๐‘–๐ผ)๐‘Ÿ๐‘ฃ๐‘Ÿ = 0 and (๐ดโˆ’๐œ†๐‘–๐ผ)๐‘Ÿโˆ’1๐‘ฃ๐‘Ÿ ฬธ= 0. A se- ๐‘‰๐‘–๐‘— =[๐‘ฃ1ยทยทยท๐‘ฃ๐‘Ÿ๐‘–๐‘—] (13) quenceofvectors(๐‘ฃ1,...,๐‘ฃ๐‘Ÿ)thatsatisfy(5)isaJordan isthecorrespondingsubmatrixof๐‘‰ and๐‘Š๐ป โˆˆC๐‘Ÿ๐‘–๐‘—ร—๐‘ ๐‘–๐‘— chainoflength๐‘Ÿ [13].ThevectorsinaJordanchainare is the corresponding submatrix of ๐‘Š partitioned as linearly independent and generate the Jordan subspace ๐‘Š =[ยทยทยท๐‘Š๐ปยทยทยท๐‘Š๐ป ยทยทยท]๐‘‡. (14) J =span(๐‘ฃ ,๐‘ฃ ,...,๐‘ฃ ). (6) ๐‘–1 ๐‘–๐‘”๐‘– 1 2 ๐‘Ÿ As shown in [12], the projection of signal ๐‘  โˆˆ C๐‘ Denote by J๐‘–๐‘— the ๐‘—th Jordan subspace of ๐œ†๐‘– with onto Jordan subspace J can be written as ๐‘–๐‘— dimension ๐‘Ÿ , ๐‘– = 1,...,๐‘˜, ๐‘— = 1,...,๐‘” . The Jordan ๐‘–๐‘— ๐‘– spaces are disjoint and uniquely decompose the general- ๐‘ ฬ‚๏ธ€๐‘–๐‘— =๐‘ ฬƒ๏ธ€1๐‘ฃ1+ยทยทยท+๐‘ ฬƒ๏ธ€๐‘Ÿ๐‘–๐‘—๐‘ฃ๐‘Ÿ๐‘–๐‘— (15) ized eigenspace G๐‘– (3) of ๐œ†๐‘– as =๐‘‰๐‘–๐‘—๐‘Š๐‘–๐ป๐‘—๐‘ . (16) โจ๏ธ๐‘”๐‘– The next sections show that invariance of the graph G = J . (7) ๐‘– ๐‘–๐‘— 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๏ธ๐‘˜ โจ๏ธ๐‘”๐‘– J . (8) ThissectiondemonstratesthatthegraphFouriertrans- ๐‘–๐‘— 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 ๐’ข(๐ด) and ๐’ข(๐ต) 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 ๐ต = ๐‘‡๐ด๐‘‡โˆ’1 [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 ๐ด ๐ด ๐ต The isomorphic equivalence of graphs is important and ๐ต, respectively, then ๐‘‰ = ๐‘‡๐‘‰ . We prove that ๐ต ๐ด since it signi๏ฌes that the rows and columns of an adja- the set G๐ผ of all graphs that are isomorphic to ๐’ข(๐ด) is ๐ด 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 ๐’ข โˆˆG๐ผ. ๐ด 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 ๐’ข โˆˆ G๐ผ up to a next section, the degrees of freedom in graph topology ๐ด permutationonthegraphsignalandinversepermutation are explored to de๏ฌne another GFT equivalence class. on the graph Fourier transform. IV. JORDANEQUIVALENCECLASSES Proof:For๐’ข(๐ด),๐’ข(๐ต)โˆˆG๐ผ,thereexistsapermu- ๐ด tation matrix ๐‘‡ such that ๐ต =๐‘‡๐ด๐‘‡โˆ’1. For eigenvector Since the Jordan subspaces of defective adjacency matrices ๐‘‰ and ๐‘‰ of ๐ด and ๐ต, respectively, let ๐‘‰ matrices are nontrivial (i.e., they have dimension larger ๐ด ๐ต ๐ด,๐‘–๐‘— and ๐‘‰ denote the ๐‘ร—๐‘Ÿ submatrices of ๐‘‰ and ๐‘‰ than one), a degree of freedom exists on the graph ๐ต,๐‘–๐‘— ๐‘–๐‘— ๐ด ๐ต whose columns span the ๐‘—th Jordan subspaces J structure so that the graph Fourier transform of a signal ๐ด,๐‘–๐‘— andJ ofthe๐‘–theigenvalueof๐ดand๐ต,respectively. is equal over multiple graphs of different topologies. ๐ต,๐‘–๐‘— Let ๐‘Š = ๐‘‰โˆ’๐ป and ๐‘Š = ๐‘‰โˆ’๐ป denote the matrices Thissectionde๏ฌnesJordanequivalenceclassesofgraph ๐ด ๐ด ๐ต ๐ต whose columns form dual bases of ๐‘‰ and ๐‘‰ . Since structures over which the GFT (1) is equal for a given ๐ด ๐ต ๐‘‰ =๐‘‡๐‘‰ , graph signal. The section proves important properties of ๐ต ๐ด this equivalence class that are used to explore inexact ๐‘Š =(๐‘‡๐‘‰ )โˆ’๐ป (17) ๐ต ๐ด methods and real-world applications in [21]. =(๐‘‰โˆ’1๐‘‡โˆ’1)๐ป (18) The intuition behind Jordan equivalence is presented ๐ด =๐‘‡โˆ’๐ป๐‘‰โˆ’๐ป (19) in Section IV-A, and properties of Jordan equivalence ๐ด are described in Section IV-B. Section IV-C com- =๐‘‡๐‘Š , (20) ๐ด pares isomorphic and Jordan equivalent graphs. Sec- where ๐‘‡โˆ’๐ป =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 ๐‘Š๐ป =๐‘Š๐ป๐‘‡๐ป =๐‘Š๐ป๐‘‡โˆ’1. (21) ๐ต ๐ด ๐ด have particular Jordan block structures. Considergraphsignal๐‘ .By(16),thesignalprojection onto J๐ด,๐‘–๐‘— is A. Intuition ๐‘  =๐‘‰ ๐‘Š๐ป ๐‘ . (22) Consider Figure 1, which shows a basis {๐‘‰} = ฬ‚๏ธ€๐ด,๐‘–๐‘— ๐ด,๐‘–๐‘— ๐ด,๐‘–๐‘— {๐‘ฃ ,๐‘ฃ ,๐‘ฃ } of R3 such that ๐‘ฃ and ๐‘ฃ span a two- 1 2 3 2 3 Permit a permutation ๐‘ =๐‘‡๐‘  on the graph signal. Then dimensional Jordan space J of adjacency matrix ๐ด the projection of ๐‘  onto J๐ต,๐‘–๐‘— is with Jordan decomposition ๐ด=๐‘‰๐ฝ๐‘‰โˆ’1. The resulting ฬ‚๏ธ€๐‘ ๐ต,๐‘–๐‘— =๐‘‡๐‘‰๐ด,๐‘–๐‘—๐‘Š๐ด๐ป,๐‘–๐‘—๐‘‡โˆ’1๐‘‡๐‘  (23) projection of a signal ๐‘ โˆˆR๐‘ as in (16) is unique. Notethatthede๏ฌnitionofthetwo-dimensionalJordan =๐‘‡๐‘‰ ๐‘Š๐ป ๐‘  (24) ๐ด,๐‘–๐‘— ๐ด,๐‘–๐‘— subspace J in Figure 1 is not basis-dependent because =๐‘‡๐‘ ฬ‚๏ธ€๐ด,๐‘–๐‘— (25) anyspanningset{๐‘ค2,๐‘ค3}couldbechosentode๏ฌneJ. Thiscanbevisualizedbyrotating๐‘ฃ and๐‘ฃ onthetwo- by (22). Therefore, the graph Fourier transform (1) is 2 3 dimensional plane. Any choice {๐‘ค ,๐‘ค } corresponds to invariant to a choice among isomorphic graphs up to a 2 3 permutationonthegraphsignalandinversepermutation a new basis ๐‘‰ฬƒ๏ธ€. Note that ๐ดฬƒ๏ธ€ = ๐‘‰ฬƒ๏ธ€๐ฝ๐‘‰ฬƒ๏ธ€โˆ’1 does not equal ๐ด=๐‘‰๐ฝ๐‘‰โˆ’1 forallchoicesof{๐‘ค ,๐‘ค };theunderlying on the Fourier transform. 2 3 graph topologies may be different, or the edge weights Theorem 2. Consider ๐ด โˆˆ C๐‘ร—๐‘. Then the set G๐ผ maybedifferent.Nevertheless,theirspectralcomponents ๐ด of graphs isomorphic to ๐’ข(๐ด) is an equivalence class (the Jordan subspaces) are identical, and, consequently, 3 Theorem 4. For ๐ดโˆˆC๐‘ร—๐‘, the set G๐ฝ of all graphs ๐ด that are Jordan equivalent to ๐’ข(๐ด) 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- ๐ด ๐ต values of ๐ด and ๐ต, respectively. Let ๐’ข(๐ด),๐’ข(๐ต)โˆˆG๐ฝ. Fig. 1: Projections ๐‘  and ๐‘  (shown in red) of a signal ๐ด ฬ‚๏ธ€1 ฬ‚๏ธ€2 Thenฮ› =ฮ› ;thatis,๐’ข(๐ด)and๐’ข(๐ต)arecospectral. ๐‘  (shown in black) onto a nontrivial Jordan subspace (span ๐ด ๐ต of ๐‘ฃ1 and ๐‘ฃ2) and the span of ๐‘ฃ3, respectively, in R3. The Proof:Since๐’ข(๐ด)and๐’ข(๐ต)areJordanequivalent, projection onto the nontrivial subspace is invariant to basis theirJordanformsareequal,sotheirspectra(theunique choices {๐‘ฃ ,๐‘ฃ } (in blue) or {๐‘ฃ ,๐‘ฃ } (in green). 1 2 ฬƒ๏ธ€1 ฬƒ๏ธ€2 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 ๐’ข(๐ด) and ๐’ข(๐ดฬƒ๏ธ€). This observation leads to the the same Jordan equivalence class. To this end, Theo- de๏ฌnition 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 ๐ฝ of ๐ด. These classes are Theorem 6. Consider ๐ด,๐ต โˆˆ C๐‘ร—๐‘ with Jordan formally de๏ฌned in the next section. decompositions ๐ด = ๐‘‰๐ฝ๐‘‰โˆ’1 and ๐ต = ๐‘‹๐ฝ๐‘‹โˆ’1 and eigenvector matrices ๐‘‰ =[๐‘‰ ] and ๐‘‹ =[๐‘‹ ], respec- ๐‘–๐‘— ๐‘–๐‘— B. De๏ฌnition and Properties tively.Then,๐’ข(๐ต)โˆˆG๐ฝ ifandonlyif๐ตhaseigenvector ๐ด This section de๏ฌnes the Jordan equivalence class of matrix ๐‘‹ = ๐‘‰๐‘Œ for block diagonal ๐‘Œ with invertible graphs, over which the graph Fourier transform (1) is submatrices ๐‘Œ๐‘–๐‘— โˆˆC๐‘Ÿ๐‘–๐‘—ร—๐‘Ÿ๐‘–๐‘—, ๐‘–=1,...,๐‘˜, ๐‘— =1,...,๐‘”๐‘–. invariant. We will show that certain Jordan equivalence Proof: The Jordan normal forms of ๐ด and ๐ต are classes allow the GFT computation to be simpli๏ฌed. equal. By De๏ฌnition 3, it remains to show J =J so ๐ด ๐ต Consider graph ๐’ข(๐ด) where ๐ด has a Jordan chain that ๐’ข(๐ต) โˆˆ G๐ฝ. The identity J = J must be true that spans Jordan subspace J๐‘–๐‘— of dimension ๐‘Ÿ๐‘–๐‘— > 1. when span{๐‘‰ }๐ด= span{๐‘‹ } =๐ดJ ,๐ตwhich implies ๐‘–๐‘— ๐‘–๐‘— ๐‘–๐‘— Then(15),andconsequently,(16),wouldholdforanon- that๐‘‹ representsaninvertiblelineartransformationof Jordan basis of J ; that is, a basis could be chosen to ๐‘–๐‘— ๐‘–๐‘— the columns of ๐‘‰ . Thus, ๐‘‹ = ๐‘‰ ๐‘Œ , where ๐‘Œ is ๐‘–๐‘— ๐‘–๐‘— ๐‘–๐‘— ๐‘–๐‘— ๐‘–๐‘— ๏ฌnd spectral component ๐‘  such that the basis vectors ฬ‚๏ธ€๐‘–๐‘— invertible. De๏ฌning ๐‘Œ = diag(๐‘Œ ,...,๐‘Œ ,...,๐‘Œ ) do not form a Jordan chain of ๐ด. This highlights that 11 ๐‘–๐‘— ๐‘˜,๐‘”๐‘˜ yields ๐‘‹ =๐‘‰๐‘Œ. the Fourier transform (1) is characterized not by the Jordan basis of ๐ด but by the set J = {J } of ๐ด ๐‘–๐‘— ๐‘–๐‘— Jordan subspaces spanned by the Jordan chains of ๐ด. 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๏ฌnition. subspaces. De๏ฌnition 3 (Jordan Equivalent Graphs). Consider Lemma 7. Consider graphs ๐’ข(๐ด),๐’ข(๐ต)โˆˆG๐ผ so that ๐ด graphs๐’ข(๐ด)and๐’ข(๐ต)withadjacencymatrices๐ด,๐ต โˆˆ ๐ต = ๐‘‡๐ด๐‘‡โˆ’1 for a permutation matrix ๐‘‡. Denote by C๐‘ร—๐‘. Then ๐’ข(๐ด) and ๐’ข(๐ต) are Jordan equivalent J and J the sets of Jordan subspaces for ๐ด and ๐ต, ๐ด ๐ต graphs if all of the following are true: respectively.If{๐‘ฃ ,...,๐‘ฃ }isabasisofJ โˆˆJ ,then 1 ๐‘Ÿ ๐ด ๐ด 1) J๐ด =J๐ต; and thereexistsJ๐ต โˆˆJ๐ต withbasis{๐‘ฅ1,...,๐‘ฅ๐‘Ÿ}suchthat 2) ๐ฝ๐ด = ๐ฝ๐ต (with respect to a ๏ฌxed permutation of [๐‘ฅ1ยทยทยท๐‘ฅ๐‘Ÿ]=๐‘‡[๐‘ฃ1ยทยทยท๐‘ฃ๐‘Ÿ];i.e.,๐ดand๐ต haveisomorphic Jordan blocks). Jordan subspaces. Let G๐ฝ๐ด denote the set of graphs that are Jordan Proof: Consider ๐ด with Jordan decomposition ๐ด= equivalent to ๐’ข(๐ด). De๏ฌnition 3 and (1) establish that ๐‘‰๐ฝ๐‘‰โˆ’1. Since ๐ต =๐‘‡๐ด๐‘‡โˆ’1, it follows that G๐ฝ is an equivalence class. ๐ด ๐ต =๐‘‡๐‘‰๐ฝ๐‘‰โˆ’1๐‘‡โˆ’1 (26) 4 =๐‘‹๐ฝ๐‘‹โˆ’1 (27) where ๐‘‹ = ๐‘‡๐‘‰ represents an eigenvector matrix of ๐ต that is a permutation of the rows of ๐‘‰. (It is clear that the Jordan forms of ๐ด and ๐ต are equivalent.) Let columns๐‘ฃ ,...,๐‘ฃ of๐‘‰ denoteaJordanchainof๐ดthat 1 ๐‘Ÿ spansJordansubspaceJ .Thecorrespondingcolumns ๐ด in ๐‘‹ are ๐‘ฅ ,...,๐‘ฅ and span(๐‘ฅ ,...,๐‘ฅ ) = J . Fig. 2: Jordan equivalent graph structures with unicellular 1 ๐‘Ÿ 1 ๐‘Ÿ ๐ต Since [๐‘ฅ ยทยทยท ๐‘ฅ ] = ๐‘‡[๐‘ฃ ยทยทยท ๐‘ฃ ], J and J are adjacency matrices. 1 ๐‘Ÿ 1 ๐‘Ÿ ๐ด ๐ต 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 ๐’ข(๐ด),๐’ข(๐ต) โˆˆ G๐ผ and ๐ต = value 0; therefore, they are Jordan equivalent. On the ๐ด ๐‘‡๐ด๐‘‡โˆ’1 for permutation matrix ๐‘‡. By (27), ๐ฝ๐ด = ๐ฝ๐ต. other hand, they are not isomorphic since the graph on Toshow๐’ข(๐ด),๐’ข(๐ต)โˆˆG๐ฝ,itremainstocheckwhether the right has more edges then the graph on the left. ๐ด J๐ด =J๐ต. Theorem8showsthatchangingthegraphnodelabels By Lemma 7, for any J๐ด โˆˆJ๐ด, there exists J๐ต โˆˆ may change the Jordan subspaces and the Jordan equiv- J๐ต that is isomorphic to J๐ด. That is, if ๐‘ฃ1,...,๐‘ฃ๐‘Ÿ alence class of the graph, while Theorem 9 shows that and ๐‘ฅ1,...,๐‘ฅ๐‘Ÿ are bases of J๐ด and J๐ต, respectively, a Jordan equivalence class may include graphs with dif- then [๐‘ฅ1ยทยทยท๐‘ฅ๐‘Ÿ] = ๐‘‡[๐‘ฃ1ยทยทยท๐‘ฃ๐‘Ÿ]. Checking J๐ด = J๐ต 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 ๐›ผ ๐‘ฃ +ยทยทยท+๐›ผ ๐‘ฃ =๐›ฝ ๐‘ฅ +ยทยทยท+๐›ฝ ๐‘ฅ (28) 1 1 ๐‘Ÿ ๐‘Ÿ 1 1 ๐‘Ÿ ๐‘Ÿ invariance of the graph Fourier transform with respect =๐›ฝ ๐‘‡๐‘ฃ +ยทยทยท+๐›ฝ ๐‘‡๐‘ฃ (29) 1 1 ๐‘Ÿ ๐‘Ÿ to equivalence relations as stated in Theorems 1 and 4. Thenexttheoremestablishesanisomorphismbetween for some coef๏ฌcients ๐›ผ and ๐›ฝ , ๐‘– = 1,...,๐‘Ÿ. How- ๐‘– ๐‘– Jordan equivalence classes. ever, (29) does not always hold. Consider matrices ๐ด and ๐ต Theorem 10. If ๐ด,๐ต โˆˆ C๐‘ร—๐‘ and ๐’ข(๐ด) and ๐’ข(๐ต) [๏ธƒ2 0 โˆ’1]๏ธƒ [๏ธƒ 1 0 0]๏ธƒ are isomorphic, then their respective Jordan equiva- ๐ด= 0 2 โˆ’1 ,๐ต = โˆ’1 2 0 . (30) lence classes G๐ฝ and G๐ฝ are isomorphic; i.e., any ๐ด ๐ต 0 0 1 โˆ’1 0 2 graph ๐’ข(๐ดโ€ฒ) โˆˆ G๐ฝ is isomorphic to a graph ๐’ข(๐ตโ€ฒ) โˆˆ ๐ด These matrices are similar with respect to a permutation G๐ฝ. ๐ต matrix and thus correspond to isomorphic graphs. Their Proof: Let ๐’ข(๐ด) and ๐’ข(๐ต) be isomorphic by per- Jordan normal forms are both mutation matrix ๐‘‡ such that ๐ต = ๐‘‡๐ด๐‘‡โˆ’1. Consider [๏ธƒ1 0 0]๏ธƒ ๐’ข(๐ดโ€ฒ) โˆˆ G๐ฝ, which implies that Jordan normal forms ๐ด ๐ฝ = 0 2 0 (31) ๐ฝ = ๐ฝ and sets of Jordan subspaces J = J by ๐ดโ€ฒ ๐ด ๐ดโ€ฒ ๐ด 0 0 2 De๏ฌnition 3. Denote by ๐ดโ€ฒ = ๐‘‰ ๐ฝ ๐‘‰ the Jordan ๐ดโ€ฒ ๐ดโ€ฒ ๐ดโ€ฒ with possible eigenvector matrices ๐‘‰ and ๐‘‰ given by decomposition of ๐ดโ€ฒ. De๏ฌne ๐ตโ€ฒ = ๐‘‡๐ดโ€ฒ๐‘‡โˆ’1. It suf๏ฌces ๐ด ๐ต [๏ธƒ1 1 0]๏ธƒ [๏ธƒ1 0 0]๏ธƒ to show ๐’ข(๐ตโ€ฒ)โˆˆG๐ฝ๐ต. First simplify: ๐‘‰๐ด = 1 0 1 ,๐‘‰๐ต = 1 1 0 . (32) ๐ตโ€ฒ =๐‘‡๐ดโ€ฒ๐‘‡โˆ’1 (33) 1 0 0 1 0 1 =๐‘‡๐‘‰ ๐ฝ ๐‘‰โˆ’1๐‘‡โˆ’1 (34) Equation (32) shows that ๐ด and ๐ต both have Jordan ๐ดโ€ฒ ๐ดโ€ฒ ๐ดโ€ฒ subspaces J1 = span([1 1 1]๐‘‡) for ๐œ†1 = 1 and =๐‘‡๐‘‰๐ดโ€ฒ๐ฝ๐ด๐‘‰๐ดโˆ’โ€ฒ1๐‘‡โˆ’1 (since ๐’ข(๐ดโ€ฒ)โˆˆG๐ฝ๐ด) (35) J21 = span([0 1 0]๐‘‡) for one Jordan subspace of =๐‘‡๐‘‰๐ดโ€ฒ๐ฝ๐ต๐‘‰๐ดโˆ’โ€ฒ1๐‘‡โˆ’1 (since ๐’ข(๐ด)โˆˆG๐ผ๐ต). (36) ๐œ† = 2. However, the remaining Jordan subspace is 2 From (36), it follows that ๐ฝ = ๐ฝ . It remains to span([1 0 0]๐‘‡) for ๐ด but span([0 0 1]๐‘‡) for ๐ต, so (29) ๐ตโ€ฒ ๐ต show that J = J . Choose arbitrary Jordan sub- fails. Thus, ๐’ข(๐ด) and ๐’ข(๐ต) are not Jordan equivalent. ๐ตโ€ฒ ๐ต space J = span{๐‘‰ } of ๐ด. Then J = ๐ด,๐‘–๐‘— ๐ด,๐‘–๐‘— ๐ดโ€ฒ,๐‘–๐‘— span{๐‘‰ } = J since ๐’ข(๐ดโ€ฒ) โˆˆ G๐ฝ. Then the ThenexttheoremshowsthatJordanequivalentgraphs ๐ดโ€ฒ,๐‘–๐‘— ๐ด,๐‘–๐‘— ๐ด ๐‘—th Jordan subspace of eigenvalue ๐œ† for ๐ต is may not be isomorphic. ๐‘– J =span{๐‘‡๐‘‰ } (37) Theorem 9. Jordan equivalence does not imply the ๐ต,๐‘–๐‘— ๐ด,๐‘–๐‘— =๐‘‡span{๐‘‰ }. (38) existence of a graph isomorphism. ๐ด,๐‘–๐‘— 5 For the ๐‘—th Jordan subspace of eigenvalue ๐œ† for ๐ตโ€ฒ, 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{๐‘‡๐‘‰ } (39) ๐ตโ€ฒ,๐‘–๐‘— ๐ดโ€ฒ,๐‘–๐‘— of the eigenvectors. Then, given eigenvector matrix ๐‘‰ =๐‘‡span{๐‘‰๐ดโ€ฒ,๐‘–๐‘—} (40) of ๐ด, all possible eigenvector matrices of ๐ด are given =๐‘‡span{๐‘‰ } (since ๐’ข(๐ดโ€ฒ)โˆˆG๐ฝ) by๐‘‹ =๐‘‰๐‘ˆ,where๐‘ˆ isadiagonalmatrixwithnonzero ๐ด,๐‘–๐‘— ๐ด (41) diagonal entries. Let ๐ต = ๐‘‹๐ฝ๐‘‹โˆ’1, where ๐ฝ is the =J . (by (38)) (42) diagonal canonical Jordan form of ๐ด. Since ๐‘ˆ and ๐ฝ ๐ต,๐‘–๐‘— are both diagonal, they commute, yielding Since (42) holds for all ๐‘– and ๐‘—, the sets of Jordan subspaces J = J . Therefore, ๐’ข(๐ตโ€ฒ) and ๐’ข(๐ต) are ๐ต =๐‘‹๐ฝ๐‘‹โˆ’1 (43) ๐ตโ€ฒ ๐ต Jordan equivalent, which proves the theorem. =๐‘‰๐‘ˆ๐ฝ๐‘ˆโˆ’1๐‘‰โˆ’1 (44) Theorem10showsthattheJordanequivalenceclasses =๐‘‰๐ฝ๐‘ˆ๐‘ˆโˆ’1๐‘‰โˆ’1 (45) of two isomorphic graphs are also isomorphic. This =๐‘‰๐ฝ๐‘‰โˆ’1 (46) result permits an frequency ordering on the spectral components of a matrix ๐ด that is invariant to both the =๐ด. (47) choice of graph in G๐ฝ and the choice of node labels, as ๐ด 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 ๐ด When a matrix has nondefective but repeated eigen- with the same set of invariant subspaces of ๐ด; i.e., values, there are in๏ฌnitely many choices of eigenvec- ๐’ข(๐ต) โˆˆ GInv if and only if Inv(๐ด) = Inv(๐ต). The ๐ด tors [16]. An illustrative example is the identity matrix, next theorem shows that GInv is a proper subset of the ๐ด which has a single eigenvalue but is diagonalizable. Jordan equivalence class G๐ฝ of ๐ด. ๐ด Since it has in๏ฌnitely many choices of eigenvectors, the Theorem 11. For ๐ดโˆˆC๐‘ร—๐‘, GInv โŠ‚G๐ฝ. identity matrix corresponds to in๏ฌnitely many Jordan ๐ด ๐ด equivalenceclasses.ByTheorem12,eachoftheseequiv- Proof: If ๐’ข(๐ต) โˆˆ GInv, then the set of Jordan ๐ด alenceclasseshavesizeone.Thisobservationhighlights subspaces are equal, or J =J . ๐ด ๐ต that the de๏ฌnition 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 ๐ด with Jordan decomposition ๐ด = decomposition of the discrete Fourier transform matrix ๐‘‰๐ฝ๐‘‰โˆ’1 where ๐ฝ is a single Jordan block and ๐‘‰ = in terms of projections onto invariant spaces. The Jor- [๐‘ฃ ยทยทยท๐‘ฃ ] is an eigenvector matrix. Then ๐ด is a repre- 1 ๐‘ dan equivalence class relaxes the assumption that all sentation of a unicellular transformation ๐‘‡ :C๐‘ โ†’C๐‘ invariant subspaces of two adjacency matrices must be with respect to Jordan basis ๐‘ฃ ,...๐‘ฃ (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. Let๐’ข(๐ด)beanelementoftheunicellular D. Diagonalizable Matrices JordanequivalenceclassG๐ฝ.Thenallgraph๏ฌlters๐ป โˆˆ ๐ด If the canonical Jordan form ๐ฝ of ๐ด is diagonal (๐ด G๐ฝ are all-pass. ๐ด is diagonalizable), then there are no Jordan chains and Proof: Since ๐ด is unicellular, it has a single Jordan the set of Jordan subspaces J = {J }๐‘ where ๐ด ๐‘ ๐‘=1 chain๐‘ฃ ,...,๐‘ฃ oflength๐‘.Consideragraphsignal๐‘  J = span(๐‘ฃ ) and ๐‘ฃ is the ๐‘th eigenvector of ๐ด. 1 ๐‘ ๐‘ ๐‘ ๐‘ over graph ๐’ข(๐ด), and let ๐‘  represent the coordinate ฬƒ๏ธ€ Graphs with diagonalizable adjacency matrices include vector of ๐‘  in terms of the basis {๐‘ฃ }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‘ ฬƒ๏ธ€1๐‘ฃ1+ยทยทยท๐‘ ฬƒ๏ธ€๐‘๐‘ฃ๐‘ =๐‘ ฬ‚๏ธ€; (48) matrixisJordanequivalentonlytoitself,asprovennext. that is, the unique projection of ๐‘  onto the spectral Theorem 12. A graph ๐’ข(๐ด) with diagonalizable adja- component J = C๐‘ is itself. Therefore, ๐’ข(๐ด) acts as cency matrix ๐ด โˆˆ C๐‘ร—๐‘ belongs to a Jordan equiva- an all-pass ๏ฌlter. Moreover, (48) holds for all graphs in lence class of size one. Jordan equivalence class G๐ฝ. ๐ด 6 Inadditiontotheall-passpropertyofunicellulargraph where Inv(ยท) represents the set of invariant spaces of a ๏ฌlters, unicellular isomorphic graphs are also Jordan matrix). If ๐‘Ž = ๐œ†, De๏ฌnition 3 can be applied, which equivalent, as proven next. yields ๐’ข(๐ด)โˆˆG๐ฝ. ๐ฝ On the other hand, consider a unicellular matrix ๐ต Theorem 14. Let ๐’ข(๐ด),๐’ข(๐ต) โˆˆ G๐ผ where ๐ด is a ๐ด suchthatitseigenvectorisnotinthespanofacanonical unicellular matrix. Then ๐’ข(๐ด),๐’ข(๐ต)โˆˆG๐ฝ. ๐ด vector, e.g., Proof: Since ๐’ข(๐ด) and ๐’ข(๐ต) are isomorphic, Jor- โŽก1 โˆ’1 1 1 โŽค dan normal forms ๐ฝ = ๐ฝ . Therefore, ๐ต is also 2 2 2 2 ๐ด ๐ต unicellular, so J๐ด = J๐ต = {C๐‘}. By De๏ฌnition 3, ๐ต =โŽขโŽข12 โˆ’12 โˆ’12 โˆ’21โŽฅโŽฅ (49) ๐’ข(๐ด),๐’ข(๐ต)โˆˆG๐ฝ๐ด. โŽขโŽฃ0 0 12 โˆ’21โŽฅโŽฆ The dual basis of ๐‘‰ can also be used to construct 0 0 1 โˆ’1 2 2 graphs in the Jordan equivalence class of unicellular ๐ด. with Jordan normal form ๐ฝ(0). Since the span of Theorem 15. Denote by ๐‘‰ an eigenvector matrix of the eigenvectors of ๐ฝ(0) and ๐ต are not identical, unicellular๐ดโˆˆC๐‘ร—๐‘ and๐‘Š =๐‘‰โˆ’๐ป isthedualbasis. Inv(๐ฝ(0))ฬธ=Inv(๐ต).However,byDe๏ฌnition3,๐’ข(๐ต)is Consider decompositions ๐ด = ๐‘‰๐ฝ๐‘‰โˆ’1 and ๐ด๐‘Š = inthesameclassofunicellularJordanequivalentgraphs ๐‘Š๐ฝ๐‘Šโˆ’1. Then ๐’ข(๐ด๐‘Š)โˆˆG๐ฝ๐ด. as those of Figure 2, i.e., ๐’ข(๐ต) โˆˆ G๐ฝ๐ฝ. In other words, formatrices๐ดand๐ตwiththesameJordannormalforms Proof: Matrices ๐ด and ๐ด have the same Jordan ๐‘Š (๐ฝ =๐ฝ ),Jordanequivalence,i.e.,J =J ,isamore normalformbyde๏ฌnition.SincethereisonlyoneJordan ๐ด ๐ต ๐ด ๐ต block, both matrices have a single Jordan subspace C๐‘. generalconditionthanInv(๐ด)=Inv(๐ต).Thisillustrates thatgraphshavingadjacencymatriceswithequalJordan ByDe๏ฌnition3,๐’ข(๐ด )and๐’ข(๐ด)areJordanequivalent. ๐‘Š normalformsandthesamesetsofinvariantspacesform a proper subset of a Jordan equivalence class, as shown The next theorem characterizes the special case of above in Theorem 11. graphsintheJordanequivalenceclassthatcontains๐’ข(๐ฝ) Remark on topology. Note that replacing each with adjacency matrix equal to Jordan block ๐ฝ =๐ฝ(๐œ†). nonzero element of (49) with a unit entry results in a Theorem16. Denoteby๐ฝ =๐ฝ(๐œ†)isthe๐‘ร—๐‘ Jordan matrix that is not unicellular. Therefore, its correspond- block (10) for eigenvalue ๐œ†. Then ๐’ข(๐ด) โˆˆ G๐ฝ if ๐ด โˆˆ inggraphisnotinaunicellularJordanequivalenceclass. ๐ฝ C๐‘ร—๐‘ is upper triangular with diagonal entries ๐œ† and This observation demonstrates that topology may not nonzero entries on the ๏ฌrst off-diagonal. determine the Jordan equivalence class of a graph. Proof: Consider upper triangular matrix ๐ด = F. Two Jordan Blocks [๐‘Ž ] with diagonal entries ๐‘Ž = ยทยทยท = ๐‘Ž and ๐‘–๐‘— 11 ๐‘๐‘ nonzero elements on the ๏ฌrst off-diagonal. By [15, Consider ๐‘ ร— ๐‘ matrix ๐ด with Jordan normal Example 10.2.1], ๐ด has the same invariant subspaces form consisting of two Jordan subspaces J1 = as ๐ฝ = ๐ฝ(๐œ†), which implies J๐ฝ = J๐ด = {C๐‘}. span(๐‘ฃ1,...,๐‘ฃ๐‘Ÿ1) and J2 = span(๐‘ฃ๐‘Ÿ1+1,...,๐‘ฃ๐‘Ÿ2) of Therefore, the Jordan normal form of ๐ด is the Jordan dimensions ๐‘Ÿ1 >1 and ๐‘Ÿ2 =๐‘ โˆ’๐‘Ÿ1 and corresponding block๐ฝ๐ด =๐ฝ(๐‘Ž11).Restrictthediagonalentriesof๐ดto eigenvalues๐œ†1 and๐œ†2,respectively.Thespectraldecom- ๐œ†so๐ฝ =๐ฝ.Then,๐’ข(๐ฝ),๐’ข(๐ด)โˆˆG๐ฝ byDe๏ฌnition(3). position of signal ๐‘  over ๐’ข(๐ด) yields ๐ด ๐ฝ ๐‘ =๐‘  ๐‘ฃ +ยทยทยท+๐‘  ๐‘ฃ +๐‘  ๐‘ฃ +ยทยทยท+๐‘  ๐‘ฃ Figure 2 shows graph structures that are in the same โŸฬƒ๏ธ€1 1 โž ฬƒ๏ธ€๐‘Ÿ1 ๐‘Ÿ 1 โŸฬƒ๏ธ€๐‘Ÿ1+1 ๐‘Ÿ1+1 โž ฬƒ๏ธ€๐‘ ๐‘ unicellular Jordan equivalence class by Theorem 16. In ๐‘ ฬ‚๏ธ€1 ๐‘ ฬ‚๏ธ€2 (50) addition,thetheoremimpliesthatitissuf๏ฌcienttodeter- minetheGFTofunicellular๐ดbyreplacing๐’ข(๐ด)โˆˆG๐ฝ =๐‘ ฬ‚๏ธ€1+๐‘ ฬ‚๏ธ€2. (51) ๐ฝ with๐’ข(๐ฝ),where๐ฝ isasingle๐‘ร—๐‘ Jordanblock.That Spectral components ๐‘  and ๐‘  are the unique pro- ฬ‚๏ธ€1 ฬ‚๏ธ€2 is,withoutlossofgenerality,๐’ข(๐ด)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‘‰ = ๐ผ chosen to compute the GFT chosen for ๐ด = ๐‘‰๐ฝ๐‘‰โˆ’1 such that ๐‘‹ = ๐‘‰๐‘ˆ, where ๐‘ˆ of a graph signal. commutes with ๐ฝ and has a particular form as follows. Remarkoninvariantspaces.Example10.2.1of[15] If ๐œ† ฬธ= ๐œ† , then ๐‘ˆ = diag(๐‘ˆ ,๐‘ˆ ), where ๐‘ˆ , shows that a matrix ๐ดโˆˆC๐‘ร—๐‘ having upper triangular 1 2 1 2 ๐‘– ๐‘– = 1,2, is an ๐‘Ÿ ร—๐‘Ÿ upper triangular Toeplitz matrix; ๐‘– ๐‘– entries with constant diagonal entries ๐‘Ž and nonzero otherwise, ๐‘ˆ has form entries on the ๏ฌrst off-diagonal is both necessary and [๏ธƒ ]๏ธƒ suf๏ฌcient for ๐ด to have the same invariant subspaces as ๐‘ˆ =diag(๐‘ˆ ,๐‘ˆ )+ 0 ๐‘ˆ12 (52) ๐‘ ร—๐‘ Jordan block ๐ฝ =๐ฝ(๐œ†) (i.e., Inv(๐ฝ)=Inv(๐ด), 1 2 ๐‘ˆ 0 21 7 where ๐‘ˆ is an ๐‘Ÿ ร—๐‘Ÿ upper triangular Toeplitz matrix ๐ฝ (๐œ†) is the ๐‘Ÿ ร—๐‘Ÿ Jordan block for eigenvalue ๐œ†. ๐‘– ๐‘– ๐‘– ๐‘Ÿ๐‘– ๐‘– ๐‘– and ๐‘ˆ and ๐‘ˆ are extended upper triangular Toeplitz 12 21 Proof: By Lemma 10.3.3 in [15], ๐ด 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 ๐ด can be obtained by transforming eigenvector subspaces as ๐ฝ = diag(๐ฝ (๐œ†),๐ฝ (๐œ†)). Therefore, ๐ด matrix ๐‘‰ as ๐‘‹ =๐‘‰๐‘ˆ. ๐‘Ÿ1 ๐‘Ÿ2 and ๐ฝ 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๏ฌnitionsareneeded.Denote๐‘ร—๐‘upper rems imply that computing the GFT (1) over the block triangular Toeplitz matrices ๐‘‡ (๐‘ ,...,๐‘ ) of form ๐‘Ÿ2 1 ๐‘Ÿ2 diagonal matrices can be simpli๏ฌed to computing the โŽก โŽค ๐‘ ๐‘ ยทยทยท ๐‘ ๐‘ transform over the adjacency matrix of a union of 1 2 ๐‘โˆ’1 ๐‘ โŽขโŽข0 ๐‘ ... ๐‘ ๐‘ โŽฅโŽฅ directed chain graphs. That is, the canonical basis can โŽข 1 ๐‘โˆ’2 ๐‘โˆ’1โŽฅ be chosen for ๐‘‰ without loss of generality. ๐‘‡๐‘(๐‘1,...,๐‘๐‘)=โŽขโŽข ... ... ... ... ... โŽฅโŽฅ, (53) As for the case of unicellular transformations, it is โŽข โŽฅ โŽข โŽฅ possible to pick bases of J and J that do not form โŽฃ0 0 ยทยทยท ๐‘1 ๐‘2 โŽฆ 1 2 a Jordan basis of ๐ด. Any two such choices of bases are 0 0 ยทยทยท 0 ๐‘ 1 relatedbyTheorem6.Concretely,if๐‘‰ istheeigenvector andde๏ฌne๐‘žร—๐‘ž uppertriangularmatrixforsome๐‘ž >๐‘ matrixof๐ดand๐‘‹ isthematrixcorrespondingtoanother choice of basis, then Theorem 6 states that a transfor- ๐‘… (๐‘ ,...,๐‘ ;๐น)= ๐‘ž 1 ๐‘ mation matrix ๐‘Œ can be found such that ๐‘‹ = ๐‘‰๐‘Œ, โŽก๐‘ ยทยทยท ๐‘ ๐‘“ ๐‘“ ยทยทยท๐‘“ ๐‘“ โŽค 1 ๐‘ 11 12 1,๐‘žโˆ’๐‘โˆ’1 1,๐‘žโˆ’๐‘ where ๐‘Œ is partitioned as ๐‘Œ = diag(๐‘Œ ,๐‘Œ ) with full- 0 ๐‘ ยทยทยท ๐‘ ๐‘“ ยทยทยท๐‘“ ๐‘“ 1 2 โŽขโŽขโŽข ... ...1 ... ๐‘ .2.2. 2,...๐‘žโˆ’๐‘โˆ’1 2,๐‘ž...โˆ’๐‘ โŽฅโŽฅโŽฅ rank submatrices ๐‘Œ๐‘– โˆˆC๐‘Ÿ๐‘–ร—๐‘Ÿ๐‘–, ๐‘–=1,2. โŽข0 0 ยทยทยท ๐‘ ๐‘“ โŽฅ โŽข ๐‘ ๐‘žโˆ’๐‘,๐‘žโˆ’๐‘โŽฅ (54) โŽข0 0 ยทยทยท ๐‘ ๐‘ โŽฅ โŽข ๐‘โˆ’1 ๐‘ โŽฅ G. Multiple Jordan Blocks โŽข .. .. .. .. .. โŽฅ โŽข . . . . . โŽฅ โŽฃ โŽฆ This section brie๏ฌ‚y describes a special case of Jordan 0 0 ยทยทยท 0 ๐‘ ๐‘ 0 0 ยทยทยท 0 01 ๐‘2 equivalence classes whose graphs have adjacency matri- 1 ces ๐ดโˆˆC๐‘ร—๐‘ with ๐‘ Jordan blocks, 1<๐‘<๐‘. where ๐น =[๐‘“ ] is a (๐‘žโˆ’๐‘)ร—(๐‘žโˆ’๐‘) upper triangular ๐‘–๐‘— Consider matrix ๐ด with Jordan normal form ๐ฝ com- matrix and ๐‘ โˆˆ C, ๐‘– = 1,...,๐‘. The theorems are ๐‘– prised of ๐‘ Jordan blocks and eigenvalues ๐œ† ,...,๐œ† . 1 ๐‘˜ presented below. By Theorem 10.2.1 in [15], there exists an upper tri- Theorem 17. Consider ๐ด = diag(๐ด ,๐ด ) where each angular ๐ด with Jordan decomposition ๐ด = ๐‘‰๐ฝ๐‘‰โˆ’1 1 2 matrix ๐ด , ๐‘– = 1,2, is upper triangular with diag- such that ๐’ข(๐ด) โˆˆ G๐ฝ. Note that the elements in the ๐‘– ๐ฝ onal elements ๐œ† and nonzero elements on the ๏ฌrst Jordan equivalence class G๐ฝ of ๐’ข(๐ฝ) are useful since ๐‘– ๐ฝ off-diagonal. Let ๐œ† ฬธ= ๐œ† . Then ๐’ข(๐ด) is Jordan signals over a graph in this class can be computed 1 2 equivalent to the graph with adjacency matrix ๐ฝ = with respect to the canonical basis with eigenvector diag(๐ฝ (๐œ† ),๐ฝ (๐œ† )) where ๐ฝ (๐œ† ) is the ๐‘Ÿ ร— ๐‘Ÿ matrix ๐‘‰ = ๐ผ. Theorem 19 characterizes the possible ๐‘Ÿ1 1 ๐‘Ÿ2 2 ๐‘Ÿ๐‘– ๐‘– ๐‘– ๐‘– Jordan block for eigenvalue ๐œ† . eigenvector matrices ๐‘‰ such that ๐ด = ๐‘‰๐ฝ๐‘‰โˆ’1 allows ๐‘– ๐’ข(๐ด)โˆˆG๐ฝ. Proof: By Theorem 16, ๐’ข(๐ด ) and ๐’ข(๐ฝ (๐œ† )) are ๐ฝ ๐‘– ๐‘Ÿ๐‘– ๐‘– Jordan equivalent for ๐‘– = 1,2 and ๐ด upper triangular Theorem 19. Let ๐ด = ๐‘‰๐ฝ๐‘‰โˆ’1 be the Jordan decom- ๐‘– with nonzero elements on the ๏ฌrst off-diagonal. There- position of ๐ด โˆˆ C๐‘ร—๐‘ and ๐’ข(๐ด) โˆˆ G๐ฝ. Then ๐‘‰ must ๐ฝ fore, the Jordan normal forms of ๐ฝ and ๐ด are the same. be an invertible block diagonal matrix. Moreover, the set of irreducible subspaces of ๐ฝ is the Proof: Consider ๐’ข(๐ฝ) with eigenvector matrix ๐ผ. unionoftheirreduciblesubspacesof[๐ฝ 0]๐‘‡ and[0๐ฝ ]๐‘‡, 1 2 By Theorem 6, ๐’ข(๐ด)โˆˆG๐ฝ implies which are the same as the irreducible subspaces of ๐ฝ [๐ด1 0]๐‘‡ and [0 ๐ด2]๐‘‡, respectively. Therefore, J๐ด =J๐ฝ, ๐‘‰ =๐ผ๐‘Œ =๐‘Œ (55) so ๐’ข(๐ด) and ๐’ข(๐ฝ) are Jordan equivalent. where ๐‘Œ is an invertible block diagonal matrix. Theorem 18. Consider ๐ด=diag(๐ด ,๐ด ) where ๐ด = 1 2 1 Thestructureof๐‘‰ giveninTheorem19allowsachar- ๐‘ˆ๐‘Ÿ๐‘Ÿ1โ‰ฅ(๐œ†๐‘Ÿ,๐‘.1,T.h.e.n,๐‘๐’ข๐‘Ÿ2(,๐ด๐น))isanJodrd๐ดan2 e=quiv๐‘‡a๐‘Ÿl2e(n๐œ†t,t๐‘o1,th.e..g,๐‘ra๐‘Ÿ2p)h, acterizationofgraphsintheJordanequivalenceclassG๐ฝ๐ฝ 1 2 with the dual basis of ๐‘‰ as proved in Theorem 20. with adjacency matrix ๐ฝ = diag(๐ฝ (๐œ†),๐ฝ (๐œ†)) where ๐‘Ÿ1 ๐‘Ÿ2 8 Theorem 20. Let ๐’ข(๐ด) โˆˆ G๐ฝ, where ๐ด has Jordan can be chosen without loss of generality. ๐ฝ decomposition ๐ด=๐‘‰๐ฝ๐‘‰โˆ’1 and ๐‘Š =๐‘‰โˆ’๐ป is the dual Jordan equivalence classes show that the GFT (1) basis of ๐‘‰. If ๐ด =๐‘Š๐ฝ๐‘Šโˆ’1, then ๐’ข(๐ด )โˆˆG๐ฝ. permits degrees of freedom in graph topologies. This ๐‘Š ๐‘Š ๐ฝ has rami๏ฌcations for total variation-based orderings of Proof: By Theorem 19, ๐‘‰ is block diagonal with thespectralcomponents,asdiscussedinthenextsection. invertible submatrices ๐‘‰ . Thus, ๐‘Š = ๐‘‰โˆ’๐ป is block ๐‘– diagonal with submatrices ๐‘Š = ๐‘‰โˆ’๐ป. By Theorem 6, ๐‘– ๐‘– V. FREQUENCYORDERINGOFSPECTRAL ๐‘Š is an appropriate eigenvector matrix such that, for COMPONENTS ๐ด =๐‘Š๐ฝ๐‘Šโˆ’1, ๐’ข(๐ด )โˆˆG๐ฝ. ๐‘Š ๐‘Š ๐ฝ Thissectionde๏ฌnesamappingofspectralcomponents 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‘  โˆˆ C๐‘ equivalence classes (assuming the eigenvalues can be is de๏ฌned as [4] computed) and potentially choose a different matrix in the class for which the GFT can be computed more TV (๐‘ )=โ€–๐‘ โˆ’๐ด๐‘ โ€– . (56) ๐บ 1 easily. Sections IV-E and IV-G show that working with Matrix ๐ด can be replaced by ๐ดnorm = 1 ๐ด when unicellular matrices and matrices in Jordan normal form |๐œ†max| the maximum eigenvalue satis๏ฌes |๐œ† |>0. permits the choice of the canonical basis. In this way, max Equation (56) can be generalized to de๏ฌne the total for matrices with Jordan blocks of size greater than variation of the Jordan subspaces of the graph shift ๐ด one, ๏ฌnding a spanning set for each Jordan subspace as described in [12]. Choose a Jordan basis of ๐ด so may be more ef๏ฌcient than attempting to compute the that๐‘‰ istheeigenvectormatrixof๐ด,i.e.,๐ด=๐‘‰๐ฝ๐‘‰โˆ’1, Jordan chains. Nevertheless, relying on graph topology where ๐ฝ is the Jordan form of ๐ด. Partition ๐‘‰ into is not always possible. Such an example was presented ๐‘ ร—๐‘Ÿ submatrices ๐‘‰ whose columns are a Jordan in Section IV-E with adjacency matrix (49). ๐‘–๐‘— ๐‘–๐‘— chain of (and thus span) the ๐‘—th Jordan subspace J ๐‘–๐‘— Relation to algebraic signal processing. The emer- of eigenvalue ๐œ† , ๐‘–=1,...,๐‘˜ โ‰ค๐‘, ๐‘— =1,...,๐‘” . Then ๐‘– ๐‘– gence of Jordan equivalence from the graph Fourier the (graph) total variation of ๐‘‰ is de๏ฌned as [12] ๐‘–๐‘— transform(1)isrelatedtoalgebraicsignalprocessingand thesignalmodel(๐’œ,โ„ณ,ฮฆ),where๐’œisasignalalgebra TV๐บ(๐‘‰๐‘–๐‘—)=โ€–๐‘‰๐‘–๐‘— โˆ’๐ด๐‘‰๐‘–๐‘—โ€–1, (57) corresponding to the ๏ฌlter space, โ„ณ is a module of ๐’œ whereโ€–ยทโ€– representstheinducedL1matrixnorm(equal corresponding to the signal space, and ฮฆ : ๐‘‰ โ†’ โ„ณ 1 to the maximum absolute column sum). is a bijective linear mapping that generalizes the ๐‘ง- 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 diagonaladjacencymatrix๐ด=๐œ†๐ผ,whereanybasisthat spectral components for graphs in the same isomorphic spansC๐‘ de๏ฌnestheeigenvectors(theJordansubspaces equivalence class as de๏ฌned 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 ๐ด,๐ต โˆˆC๐‘ร—๐‘ and ๐’ข(๐ต)โˆˆG๐ผ, i.e., ๐ด distinctspectralcomponents.Similarly,thesignalmodel ๐’ข(๐ต) is isomorphic to ๐’ข(๐ด). Let ๐‘‰๐ด,๐‘–๐‘— โˆˆ C๐‘ร—๐‘Ÿ๐‘–๐‘— be (๐’œ,โ„ณ,ฮฆ) requires a choice of basis for module (signal a Jordan chain of matrix ๐ด and ๐‘‰๐ต,๐‘–๐‘— โˆˆ C๐‘ร—๐‘Ÿ๐‘–๐‘— the space) โ„ณ in order to de๏ฌne the frequency response corresponding Jordan chain of ๐ต. Then (irreducible representation) of a signal [26]. On the TV (๐‘‰ )=TV (๐‘‰ ). (58) other hand, this section demonstrated the equivalence of ๐บ ๐ด,๐‘–๐‘— ๐บ ๐ต,๐‘–๐‘— the GFT (1) over graphs in Jordan equivalence classes, Proof: Since ๐’ข(๐ด) and ๐’ข(๐ต) are isomorphic, there which implies an equivalence of certain bases. This exists a permutation matrix ๐‘‡ such that ๐ต = ๐‘‡๐ด๐‘‡โˆ’1 observation suggests the concept of equivalent signal and the eigenvector matrices ๐‘‰ and ๐‘‰ of ๐ด and ๐ต, ๐ด ๐ต models in the algebraic signal processing framework. respectively,arerelatedby๐‘‰ =๐‘‡๐‘‰ .Thus,theJordan ๐ต ๐ด Just as working with graphs that are Jordan equivalent chains are related by ๐‘‰ =๐‘‡๐‘‰ . By (57), ๐ต,๐‘–๐‘— ๐ด,๐‘–๐‘— to those with adjacency matrices in Jordan normal form TV(๐‘‰ )=โ€–๐‘‰ โˆ’๐ต๐‘‰ โ€– (59) simpli๏ฌes GFT computation, we expect similar classes ๐ต ๐ต,๐‘–๐‘— ๐ต,๐‘–๐‘— 1 ofequivalentsignalmodelsforwhichthecanonicalbasis =โƒฆโƒฆ๐‘‡๐‘‰๐ด,๐‘–๐‘— โˆ’(๏ธ€๐‘‡๐ด๐‘‡โˆ’1)๏ธ€๐‘‡๐‘‰๐ด,๐‘–๐‘—โƒฆโƒฆ1 (60) 9 =โ€–๐‘‡๐‘‰ โˆ’๐‘‡๐ด๐‘‰ โ€– (61) composition ๐ต = ๐‘‰๐ฝ๐‘‰โˆ’1. Let the columns of eigen- ๐ด,๐‘–๐‘— ๐ด,๐‘–๐‘— 1 =โ€–๐‘‡ (๐‘‰๐ด,๐‘–๐‘— โˆ’๐ด๐‘‰๐ด,๐‘–๐‘—)โ€–1 (62) vector submatrix ๐‘‰๐‘–๐‘— span the Jordan subspace J๐‘–๐‘— of ๐ด. Then the class total variation of spectral component =โ€–๐‘‰ โˆ’๐ด๐‘‰ โ€– (63) ๐ด,๐‘–๐‘— ๐ด,๐‘–๐‘— 1 J is de๏ฌned as the supremum of the graph total ๐‘–๐‘— =TV(๐‘‰๐ด), (64) variation of ๐‘‰๐‘–๐‘— over the Jordan equivalence class (for all ๐’ข(๐ต)โˆˆG๐ฝ): where(63)holdsbecausethemaximumabsolutecolumn ๐ด sumofamatrixisinvarianttoapermutationonitsrows. TV (J )= sup TV (๐‘‰ ). (68) G๐ฝ ๐‘–๐‘— ๐บ ๐‘–๐‘— ๐ด ๐’ข(๐ต)โˆˆG๐ฝ ๐ด Theorem 21 shows that the graph total variation is ๐ต=๐‘‰๐ฝ๐‘‰โˆ’1 invariant to a node relabeling, which implies that an span{๐‘‰๐‘–๐‘—}=J๐‘–๐‘— โ€–๐‘‰๐‘–๐‘—โ€–1=1 ordering of the total variations of the frequency com- Theorem 23. Let ๐ด,๐ต โˆˆC๐‘ร—๐‘ and ๐’ข(๐ต)โˆˆG๐ผ. Let ponents is also invariant. ๐ด ๐‘‰ and ๐‘‰ be the respective eigenvector matrices with Reference [12] demonstrates that each eigenvector ๐ด ๐ต Jordan subspaces J = span{๐‘‰ } and J = submatrix corresponding to a Jordan chain can be nor- ๐ด,๐‘–๐‘— ๐ด,๐‘–๐‘— ๐ต,๐‘–๐‘— span{๐‘‰ } spanned by the ๐‘—th Jordan chain of eigen- malized. This is stated as a property below: ๐ต,๐‘–๐‘— value ๐œ† . Then TV (J )=TV (J ). ๐‘– G๐ฝ ๐ด,๐‘–๐‘— G๐ฝ ๐ต,๐‘–๐‘— Property 22. The eigenvector matrix ๐‘‰ of adjacency ๐ด ๐ต matrix ๐ด โˆˆ C๐‘ร—๐‘ can be chosen so that each Jordan Proof: Let ๐‘‰๐ด* denote the eigenvector matrix corre- sponding to ๐’ข(๐ด*)โˆˆG๐ฝ that maximizes the class total chain represented by the eigenvector submatrix ๐‘‰ โˆˆ ๐ด C๐‘ร—๐‘Ÿ๐‘–๐‘— satis๏ฌes โ€–๐‘‰๐‘–๐‘—โ€–1 = 1; i.e., โ€–๐‘‰โ€–1 = 1 wit๐‘–h๐‘—out variation of Jordan subspace J๐ด,๐‘–๐‘—; i.e., loss of generality. TV (J )=TV (๏ธ€๐‘‰* )๏ธ€. (69) G๐ฝ ๐ด,๐‘–๐‘— ๐บ ๐ด,๐‘–๐‘— ๐ด Itisassumedthattheeigenvectormatricesarenormal- Similarly, let ๐‘‰* denote the eigenvector matrix corre- ๐ต ized as in Property 22 for the remainder of the section. sponding to ๐’ข(๐ต*)โˆˆG๐ฝ that maximizes the class total ๐ต Furthermore, [12] shows that (57) can be written as variation of Jordan subspace J , or ๐ต,๐‘–๐‘— TV๐บ(๐‘‰๐‘–๐‘—)=โƒฆโƒฆ๐‘‰๐‘–๐‘—(๏ธ€๐ผ๐‘Ÿ๐‘–๐‘— โˆ’๐ฝ๐‘–๐‘—)๏ธ€โƒฆโƒฆ1 (65) TVG๐ฝ (J๐ต,๐‘–๐‘—)=TV๐บ(๏ธ€๐‘‰๐ต*,๐‘–๐‘—)๏ธ€. (70) ๐ต = max {|1โˆ’๐œ†|โ€–๐‘ฃ โ€– ,โ€–(1โˆ’๐œ†)๐‘ฃ โˆ’๐‘ฃ โ€– }. ๐‘–=2,...,๐‘Ÿ๐‘–๐‘— 1 1 ๐‘– ๐‘–โˆ’1 1 Since ๐’ข(๐ด) and ๐’ข(๐ต) are isomorphic, Theorem 10 (66) implies that there exists ๐’ข(๐ตโ€ฒ) โˆˆ G๐ฝ such that ๐ตโ€ฒ = ๐ต ๐‘‡๐ด*๐‘‡โˆ’1; i.e., ๐‘‰ =๐‘‡๐‘‰* where ๐‘‰ is an eigenvector and establishes the upper bound for the total variation ๐ตโ€ฒ ๐ด ๐ตโ€ฒ matrixof ๐ตโ€ฒ. Bythe classtotal variationde๏ฌnition (68), of spectral components as TV (๐‘‰ ) โ‰ค TV (๐‘‰* ). Applying Theorem 21 to ๐บ ๐ตโ€ฒ,๐‘–๐‘— ๐บ ๐ต,๐‘–๐‘— TV๐บ(๐‘‰๐‘–๐‘—)โ‰ค|1โˆ’๐œ†๐‘–|+1. (67) isomorphic graphs ๐’ข(๐ด*) and ๐’ข(๐ตโ€ฒ) yields Equations(65),(66),and(67)characterizethe(graph) TV (๏ธ€๐‘‰* )๏ธ€=TV (๐‘‰ )โ‰คTV (๏ธ€๐‘‰* )๏ธ€. (71) ๐บ ๐ด,๐‘–๐‘— ๐บ ๐ตโ€ฒ,๐‘–๐‘— ๐บ ๐ต,๐‘–๐‘— total variation of a Jordan chain by quantifying the Similarly, by Theorem 10, there exists ๐’ข(๐ดโ€ฒ) โˆˆ G๐ฝ change in a set of vectors that spans the Jordan sub- ๐ด space J๐‘–๐‘— when they are transformed by the graph asuncehigtehnavte๐ตct*or=m๐‘‡at๐ดrixโ€ฒ๐‘‡oโˆ’f1๐ด,โ€ฒo.rA๐‘‰p๐ตp*ly=(6๐‘‡8๐‘‰)๐ดanโ€ฒdwThheereor๐‘‰em๐ดโ€ฒ2i1s shift ๐ด. 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 (๏ธ€๐‘‰* )๏ธ€โ‰ฅTV (๐‘‰ )=TV (๏ธ€๐‘‰* )๏ธ€. (72) ๐บ ๐ด,๐‘–๐‘— ๐บ ๐ดโ€ฒ,๐‘–๐‘— ๐บ ๐ต,๐‘–๐‘— 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 TV๐บ(๏ธ€๐‘‰๐ด*,๐‘–๐‘—)๏ธ€ = over any graph in a given Jordan equivalence class. TV๐บ(๏ธ€๐‘‰๐ต*,๐‘–๐‘—)๏ธ€, or Furthermore, for any two graphs ๐’ข(๐ด),๐’ข(๐ต) โˆˆ G๐ฝ, ๐ด ๐ด TV (J )=TV (J ). (73) and๐ต haveJordanbasesforthesameJordansubspaces, G๐ฝ๐ด ๐ด,๐‘–๐‘— G๐ฝ๐ต ๐ต,๐‘–๐‘— 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๏ฌcant because it means that an derivehereade๏ฌnitionofthetotalvariationofaspectral ordering of the spectral components by their class total component of ๐ด in relation to the Jordan equivalence variations is invariant to node labels. class G๐ฝ. Next, the signi๏ฌcance of the class total variation (68) ๐ด Class total variation. Let ๐’ข(๐ต) be an element in isillustratedforadjacencymatriceswithdiagonalJordan Jordan equivalence class G๐ฝ where ๐ต has Jordan de- form, one Jordan block, and multiple Jordan blocks. ๐ด 10

See more

The list of books you might like

Most books are stored in the elastic cloud where traffic is expensive. For this reason, we have a limit on daily download.