Matrix Generalizations of Some Theorems on Trees, Cycles and Cocycles in Graphs Author(s): Stephen B. Maurer Source: SIAM Journal on Applied Mathematics, Vol. 30, No. 1 (Jan., 1976), pp. 143-148 Published by: Society for Industrial and Applied Mathematics Stable URL: http://www.jstor.org/stable/2100591 Accessed: 26/08/2010 09:01 Your use of the JSTOR archive indicates your acceptance of JSTOR's Terms and Conditions of Use, available at http://www.jstor.org/page/info/about/policies/terms.jsp. JSTOR's Terms and Conditions of Use provides, in part, that unless you have obtained prior permission, you may not download an entire issue of a journal or multiple copies of articles, and you may use content in the JSTOR archive only for your personal, non-commercial use. Please contact the publisher regarding any further use of this work. Publisher contact information may be obtained at http://www.jstor.org/action/showPublisher?publisherCode=siam. Each copy of any part of a JSTOR transmission must contain the same copyright notice that appears on the screen or printed page of such transmission. JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide range of content in a trusted digital archive. We use information technology and tools to increase productivity and facilitate new forms of scholarship. For more information about JSTOR, please contact [email protected]. Society for Industrial and Applied Mathematics is collaborating with JSTOR to digitize, preserve and extend access to SIAM Journal on Applied Mathematics. http://www.jstor.org SIAM J. APPL. MATH. Vol. 30, No. 1, January1976 MATRIX GENERALIZATIONS OF SOME THEOREMS TREES, CYCLES AND COCYCLES IN GRAPHS* ON STEPHEN B. MAURERt Abstract.We extendto arbitrary matricesfourtheoremsof graphtheory,one about projections onto the cycle and cocyclespaces, one about the intersectionof these spaces, and two matrix-tree theorems.The squares of certaindeterminants, not apparentforgraphs,appear in the extensions. 1. Introduction.We present here four theorems about matrices and associated vectorspaces. When the matricesare chosen in a special way, the theoremsreduceto knownresultsingraphtheory.This paper arose froma desire to generalizethose graphtheoremsto the,generalindependencesystemscalled matroids.Since everyfiniteset of vectors(e.g., the columnsof a matrix)is a matroid,and since manymatroidscan be representedin thisway,we have been partlysuccessful.Our theoremscan also be looked at fromotherpointsof view. For instance,Theorems1 and 2 maybe interpreted intermsofdual linearcodes. We willfirststateour theorems,thenstatethegraphtheoremstheyinclude, thengive proofs. 2. The theorems.To stateour resultsconciselywe need considerablenotation.For easy referencewe introducethisnotationin a list: 9 is a fieldof characteristic p. M is an m x n matrixover S. R and 16are the row and columnspaces of M. X is the nullspace of M, i.e., the orthogonalcomplementof R. , en} is the standardbasis of 3?/. {e1,e2, J={1, 2, . , n}. For KcJ, K=J-K. $Y9 is thecollectionof subsetsofJ indexingcolumnsetsof M whichare bases of T6.Bo is a fixedset in X, B rangesover all of 3. Bo and B willoftenreferto the bases indexedas well as the indexsets. RBej, forj E B, is theunique vectorE a1e, in R such thataj = 1, and ai = 0 for B i,e U {j}. (Uniquenessis well knownand willbe verifiedlater.)For j E B, RBe, is definedto be 0, thezero vectorof RB is the linearmap from 9In to R such thatRB(ej) = RBej forall j E J. R is the matrixwithrowsRBoej. j E Bo. = 0 for NBe,, forJE B, is theunique vectorE fBiei in .X suchthat1,j= 1, and 13i i/ B U{j}. Forj E B, NBej =0. gn __> is the linearmap withNB(ej) = NBej, j E J. NB: N is the matrixwithrowsNB0eJ, j E Bo. E is element in or an variable over S. Xj,j J, XK =IK Xjforj E K c J. X: gn _>in is the linearmap withwithX(ej) = Xjej. X is the matrixof the mapX. Sn. -f * Received by the editorsDecember 14, 1973, and in revisedformJuly1, 1974. t Departmentof Combinatoricsand Optimization,Universityof Waterloo,Waterloo,Ontario, Canada. Now at MathematicsDepartment,PrincetonUniversity, Princeton,New Jersey08540. 143 144 STEPHEN B. MAURER ofmatrix pt is thetranspose P. ofsquarematrix P. Wedo notorderrowsandcolumns, IPIisthedeterminant so someofourresults arecorrect determinants involving onlyuptosign.Thenwe write = i ratherthan =. M(A, K), whereA is anybasisof16andK c J,is thematrix whosecolumns givethecoordinates relativetoA ofthecolumn-vectors ofM indexedbyK. [A : K]= IM(A,K)J2I assumingM(A, K) is square. Note there is no ambiguity aboutthesignof[A K]. rj F=>ja[Bo: B]. 1. RB isa projection onto24andX ZB[Bo: B]XBRBisself-adjoint. - & [BO: B]RB is theorthogonal In particular, if 1$ 0, qr projection ontoR with X. Similarly, kernel NBprojects ontofN,X & [Bo: B]XBNBis self-adjoint, and if onto with kernel X R. 7 $0, N-1'Z [Bo: B]NB is theorthogonal projection THEOREM 2. R n.A={O} ifand onlyif 1 $ 0. THEOREM 3. JRX Rtl= [Bo: B]XB and IN X Ntl= [Bo: B]]XB THEOREM 4. = ? + R= [Bo: B]XB and NX t>2[BO: B]X . We saidthatthesetheorems includethegraphtheorems tofollow.Actually, twoofthegraphtheorems a somewhat haveas theirbest-known forms anomalous specialcase notincludedin theabove.The above can be statedin evenmore to includethesespecialcases,buttherestatements generality requireevenmore notation.In theprocessof provingtheabove we willin factprovethefurther andpointouthowtheycovertheanomalousgraphcases. generalizations THEOREM RX 3. Graphs.Let graphG have m nodesand n edges.For convenience we assumeG connected. Iftheedgesarenotalreadydirected, directthemarbitrarily. DefineM = [ai,]byai, = 1 ifedgej entersnodei,-1 ifedgej leavesnodei,and0 otherwise.If G was directedto beginwith,the usual fieldis the rationals; otherwise theusualchoiceis {0, 1}. However,whatfollowsis wellknownto be trueforanyi. X is justthespacegenerated of G, i.e.,thecycle bythecircuits space.2 is thespacegenerated bythecut-sets; graph-theorists nowadays usually callthisthecocyclespace,buttraditionally itwasthespaceofcoboundaries, with cocyclesmeaning else.A setofcolumns ofM isa basisof'6 ifandonly something ifitisindexedbya spanning tree,andfortwosuchbases,whichinthissectionwe call TO,T insteadof Bo, B, IM(T0,7T)= ?1. Finally,NTej,j e T, is justwhatis knownas thefundamental ofcyclesforT,and RTe,,jE T,thefundamental system ofcocycles.TheyarebasesofX and2, respectively. system Forgraphsthereis anotherwayto geta basisof 2. Foreachvertextheedgesincident to itforma cut-set, and thecorresponding cocyclesforall butanyone vertexforma basis. Sucha cocycle-basis isalsocalleda fundamental a vertex-isolating system, system. THEOREM 1'. LetTbe inthecollection trees 27ofspanning ofgraphG. ThenRT is a projection ontothecocyclespace,and X E, XTRT iSself-adjoint. Let7 be the number treesmodp, thecharacteristic ofspanning oftheunderlying field.If q $ 0, f'- EgRT is theorthogonal ontothecocyclespacewiththecyclespaceas projection itskernel. ontothecyclespace,X E, XrNTisself-adjoint, Similarly, NTprojects and ontothecyclespacewiththecocycle if 1 $ 0, f-'ZgNTis theorthogonal projection spaceas kernel. MATRIX GENERALIZATIONS OF GRAPHS 145 Theorem1' is due to Nerode and Shank[3]. Theyused theself-adjointness of X E, XtNT to give a short proof of the famous formula,firstestablishedby Kirchhoff [2], forthecurrentsin a resistiveelectricalnetwork.In thatcontextXiis the resistancein wirej. THEOREM 2'. 7he cycleand cocyclespaces of G overa fieldof characteristic p have nontrivial intersection ifand onlyifthenumberof spanningtreesequals 0 mod p. Theorem2' is due to Shank (unpublished)butthespecial case p = 2 has been publishedindependently by Shank [4] and Chen [1]. Actually,Chen's argument worksin general,as we show below. THEOREM 3'. Let R (resp. N) have some fundamentalsystemof cocycles (cycles) of G foritsrows.Then IRXRtl=ZE.XT, INXNNI= ,XTr. Theorem 3' is a generalmatrix-tree theoremappearingin Trent[5], When X = I and p = 0, eitherdeterminant givesthenumberofspanningtrees.Whenthe rowsof R are a vertex-isolating fundamentalsystem,RR' is a principalminorof the degree-minus-adjacency matrix.This is thebest-knowncase of thetheorem, but also the case not coveredby Theorem 3 as stated. NowadaysTheorem3' is oftenattributedto Kirchhoff, forit is claimedto be setdowna systemoflinearequationsand showed, implicitin hispaper. Kirchhoff withoutusing determinants, that the solutionscould be writtenwithcommon denominator ,jXt.This certainlysuggeststhat the matrixof coefficients has determinantZ3XT,but thiscoefficient matrixis not any of those in Theorem 3'. The following,whichwe have not seen statedin the literature,is morerightfully called Kirchhoff's matrix-tree theorem. THEOREM 4'. WithR and N as in Theorem3, N RX NXXT, R NX matrixwas of the formon the right,withR arising Specifically,Kirchhoff's froma vertex-isolating fundamentalsystemand N fromsome tree. 4. Proofs.Theorems 1-3 can all be proved by a few modificationsin the proofscited forTheorems 1'-3', but forclarityand completenesswe give full argumentsforall of Theorems 1-4. We presentthe proofsin severalsteps.The purposeof the firstthreestepsis to cut the remainingworkin half. One morebatchofnotation:M* is anymatrixwithrowspace X and thusnull space R. (From the matroidpoint of view, M and M* representduals). The columnsof M*, like thoseof M, are indexedbyJ.Let c6*be thecolumnspace of M*. If A* is a basis of V, and Kc J, M*(A*, K) and [A* : K]* are defined analogouslyto M(A, K) and [A : K]. Finally,( ,.) is the usual innerproductin if u = E aej and v = Z3jej,then(u, v) = E a,31j. 9n 146 STEPHEN B. MAURER (a) K c Jindexesa basisoft* ifand onlyifK E $@.(Thiswas first observedby matroids[6].) Any two matriceswiththe same Whitneyin his paper introducing rowspace can be obtainedone fromtheotherbyelementaryrowoperations,and theseoperationsdo not affectwhichsets of columnsare columnspace bases. So foranyB E X3,we mayreplaceM with[I, P], whereI is theidentityindexedby B. Indeed, [I, P] = M(B, J). It is theneasy to show that[-Pt, I'] is a choice forM*, where1*is theidentity indexedbyB. Thus B indexesa basis of c(*. Indeed,forany choice of M*, [-PI' I'] = M*(B, J). Conversely,to show thateverybasis of 10 is some B, repeat the above argumentwiththe roles of M and M* reversed. (b) ForK c J,IM*(B, K)J= ?IM(B, K)j, so [B: K]* = [B: K]. We have just shown that M(B, J) = [I, P], M*(B, J) = [-P', F]. Thus if Q is the minorof P indexedby rowsB n K and columnsB n K, thenIM(B, K)I = ? IQI.Likewise,if Q*is the minorof -Pt indexedby (B n K, B n K), thenIM*(B, K)| = ?IQ*I.But (c) Let A be any basis of '6 (notnecessarilyone in 2%7). Let A * be any basis of 10. Foreveryresultbelowinvolving[A: B], XB,RB,X, .X,R and/orN, thereis an analogous resultinvolving[A* B]*, XB,NB,X, X, N and/orR. Simplyapplythe statedresultto M* insteadof M. (d) RBej is well-defined.Let u, v be two vectors satisfyingthe defining conditionsforRBej forsome j E B. Then u - v = E ate,E R, where ai = 0 forall i E B. Since B indexesa columnbasis forM, all theothera, dependon thosewith i E B. Thus u = v. (e) RB is a projection on R. Clearlytheimageof RB is in R, so we mustshow RB is the identityon R. WritingRBu - u = E ajej, we have forany u E i9n that ai = 0 forall i c B. If u e , thenRBU - u E R and the argumentin (d) applies. (f) The vectors RBeJ, j E B, are therowsofM(B, J). In particularR = M(Bo, J). Since M(B, J) = [I, P], with I indexed by B, row j meets the two conditions definingRBej. In the nextfewparagraphswe abbreviateB U {j} - {i} as B - i + j. (g) (RBei, e1) ?0 if and only if B - i +j e 2. (RBei,e,) is the (i, j) entryof M(B, J) = [I, P]. Clearly,ifthisentryis 0, columnj dependson columnsB - i,and conversely. (h) IfB' = B-i +j,then (RBej, ei) = (RBei, ej)-,. First,RB'ej = (RBei, ej)-'RBei, fortherightside meetsthetwoconditionsdefining theleft.Second,(RBei, ei) = 1. (i) [A: B][B :K] = [A :K]. This follows because M(A, B) x M(B, K) -M(A, K). (j) X > [A: B]XBRBis self-adjoint.Call thisoperatorH forshort.We must show (Hei, ei) = (ei,Hei) for all i, j. By (g) the only summandsin H that give = {BIB - i + j E 2}, and on therightonly nonzerotermson theleftare thosein /3ij ,. ClearlyB->B'= B-i +j is a bijectionbetween 4'ijand s?I,,so it sufficesto prove (X[A: B]XBRBei,e) = (ei,X[A: B']XBRBej). The left simplifiesto XjXB[A: B](RBei,e,), the rightto XiXB'[A: B'](ei, RBe). Since XjXB= XiXB,by (h) and (i) it sufficesto prove (RBei,ej)2=[B: B']. But M(B, B') is an identitymatrix except for one column, so in light of (f), IM(B, B')I = ?(RBei, e1). MATRIX OF GRAPHS GENERALIZATIONS 147 (k) Define, = [A : B] = [A : Bo]r1,and H = E[A : B]RB. If , #0 , then ontoR withkernelfV. the -'1His orthogonal projection By (e), IJJ-His a projection, A and by (j), it is self-adjoint.A projectionis self-adjointifand onlyifitskernelis the orthogonalcomplementof its image, and by definitionX is the orthogonal complementof R. We have now provedTheorem 1. ChoosingBo forA, (e), (j) and (k) givethe firsttwosentencesof thattheorem.Applying(b) and (c), we getthelastsentence. A to Bo originally.Because of (b) it was There were two reasons forrestricting thenunnecessaryto introduceM* and relatednotationat thebeginning.Of more mathematicalinterest,onlywhenA E X13is A a basis of the associatedmatroid. Next we prove Theorem 3. It followsfrom(1) below in the same way that Theorem 1 followedfrom(e), (j), (k). (1) Let R be any matrixwhoserowsare a basis forR. Then R = M[A, J] for somebasisA ofX, and IRX RI| = Z [A: B]XB. We have alreadynotedthatiftwo matriceshave the same row space, e.g., R and R, theydifferonlyby elementary rowoperations.If in additiontheybothhave k rows,theircolumnsetsrepresent bases of 3;k, whichis the column the same vectorswithrespectto two different space whenevertherowsofeitherare independent.ThisprovesR = M(A, J). The secondclaimis an immediateconsequenceof theChauchy-Binettheorem:ifP is m x n and Q is n x m, then IPQI= |Pcl IQK|, wherePK rangesover all m x m m x m minorsofQ. (See Trent minorsofP, and QK rangesoverthecorresponding [5] fora completedescription.)Now take P = RX, Q = Rt, and recall thatX is diagonal and IM(A, K)I #0 ifand onlyifK E A. (m) R n X = {0} ifand onlyif7 #0 (Theorem2). We followChen [1], using Theorem 3 wherehe used Theorem 3' restrictedto the field{0, 1}. Considerthe square matrixP = [NjIt sufficesto show IPI= ?77,forclearlyR n X = {O}ifand onlyifIPI #0 . Now t= [RRt RNt]RRt NNt 0 forcingIPI= q. NRt Thus by Theorem 3, IPPtl (n) = 72, 01 NN'J = Z [BoB]XB (Theorem4). Let P= RNX RRt 0 1 VN =[NXRt Then [N]Q= [NX] NXNNtJ so IQIIPI= 7q [Bo: B]XB by Theorem 3. Since IPI= ?, Theorem 4 follows wheneverq #0 . To cover the case q = 0, recall thatR = M(BO, J) is of the form [I, R'], and temporarilylet the entriesin R' be independentvariables over 9 ratherthanelementsin i. Then surelyq #0 and thetheoremholds,providinga wheneverR' is made up of elementsof i.. formulawhichholds by substitution (o) Let R = M(A, J). Then NX = from(n) and the factRt= M(A, Bo) X Rl. ? IM(A, Bo)IlZ [Bo: B]XB. This follows 148 STEPHEN B. MAURER The specialcase ofTheorem3' followsfrom(1).Recallthatinthiscase R of fundamental cocyclesystem. thetheorem (RIin(1))comesfroma vertex-isolating Forgraphs, [A: B] = ?1 whenbasisA comesfroma treeandinthisspecialcase. Likewise(o) coversthespecialcasesofTheorem4', ofwhichtherearetwo:when andwhenitcomesfroma ofR isvertex-isolating, thefundamental cocyclesystem treethanthefundamental ofN. different cyclesystem We havean alternate here,which proofofTheorem2, omitted 5. Remarks. overthenumber of ofinduction technique usesthestandard graph(andmatroid) and deletions.The statement of 2' does not edgeswiththeaid of contractions proof. andneither, inthisgraphcase,doesouralternate involvematrices, withsomecoefficients to arbitrary matrices for ThatTheorem3' generalizes toanyonewhohasstudiedtheproof.Whatmaybe beenevident thex's hassurely andthefact ofthesecoefficients description newinourTheorem3 is thespecific and INX N'l. thattheyarethesameforJRX RWj to thenumberof bases in 0Yis a problemof interest Finally,determining when matroidtheorists. Theorems3 and 4 are mostinteresting Consequently M forwhichthe p = 0, X = 1,and[Bo: B] = 1 forall B E $. The classofmatrices matrices unimodular holdsin justthewell-known lastcondition (equivalently, ofbasesforanyM by thenumber In theory wemaydetermine regular matroids). thenumber eachXia different computing making variable, JRX R'J,andcounting of distinctproductsXB whichoccur,ignoringtheircoefficients. However, variablesis,toputitmildly, determinants time-consuming. containing evaluating to Ifp = 2, Shankhasshown[4,Lemma5] thattofindallthedistinct XBitsuffices butitisverydoubtful theproduct ofthediagonalentries, anysuchresult compute holdsforotherp. us in these We thankProfessor Shankforinteresting Acknowledgment. the for several valuable We the referees concerning matters. thank suggestions ofthispaper. organization REFERENCES 20 (1971),pp.526-529. thisJournal, witha graph, spacesassociated [1] W. K. CHEN, On vector der derGleichungen, aufwelchemanbeiderUntersuchung [2] G. KIRCHHOFF, UberdieAufl6sung Ann.Phys.Chem.,72 gerfuhrt Poggendorf's Strome galvanischer Verteilung linearen wird, 5 (1958),pp.4-8. IRE Trans.Circuit Theory, (1847),pp. 497-508;Englishtranslation, Amer.Math. network ofKirchhoff's proof [3] A. NERODE AND H. SHANK, An algebraic theorem, 68 (1961),pp. 244-247. Monthly, 5 (1971),pp.45-49. Math.Systems machines, recognition [43 H. SHANK, Graphproperty Theory, linear and listing ofall maximaltreesofa connected [5] H. M. TRENT, A noteon theenumeration Proc.Nat.Acad.Sci. U.S.A.,40 (1954),pp. 1004-1007. graph, Amer.J. Math.,57 (1935),pp. of linearindependence, properties [6] H. WHITNEY, The abstract 509-533.
© Copyright 2026 Paperzz