nl Dalancea ‘ Matroids Tovnds IBM Almaden 650 Harry Milena FedeF Center Research Road, Bell San Jose, CA 445 South 95120 Abstract We establish strong expansion properties for the basesexchange graph of balanced matroids; consequently, the set of bases of a balanced matroid can be sampled and approximately counted using rapidly mixing Markov chains. Specific classes for which balance is known to hold include graphic and regular matroids. Morristown, the set of vectors. A subclass lar will be of interest Introduction and they matroids: graphic forth matroids denoted nected whose vertex-set S be a finite ground-set, tion of subsets of S. Following the set B is said to form matmid &f(S, is the collection operation the same cardinality element: The bases-exchange has been studied graph combinatorial contexts related I?2 = 131\{e ~~}. properties for any pair exchange property is that terminology, e E B indicating of bases of a matroid of bases J31 and holds: an ~C l?2 such that ample standard for all 131\{ e}U{~} of graphic of G(M). ple is that maximum from the event e is in the chosen basis. the negative The cor- if the inequality linearly ex- and whose bases Another examis and whose bases are independent < Pr[e] Pr[~] exists whose ground-set matmids, Pr[e~] whose ground-set matmids, over some field cardinality there is in t3. A main trees of the graph, of vectorial a set of vectors at random of S, let e denote &f (S, l?) is said to satisfy property relation l?2 the following e E l?l is the set of edges of a given graph are the spanning that before in [12, 20]; here we focus the mnk of the matroid), (called and (ii) of of removing and adding one ground-set 23, and e is an element if (i) all sets in B have hence- by an edge if and only if 132 can be obtained and let B be a collec- the collection 23) if and only M, by Edmonds two bases B1 and 132 con- If B is a basis chosen uniformly Let All [28]. 131 by the fundamental on expansion Preliminaries of regu- over every field. was introduced with matroids is that of a matroid gmph by G(A4), [10] as the graph of vectorial in this paper are regular bases of the matroid, from NJ 07960 are vectorial The bases-exchange various 1 Research Street, which We introduce the notion of ‘balance” , and say that a rnatroid is balanced if the matroid and all its minors satisfy the property that, for a randomly chosen baais, the presence of an element can only make any other element less likely. Mihail Communications subsets of holds for all pairs of distinct elements ize that negative is equivalent Pr[e], correlation thus expressing ence of an element less likely; the intuitive studied before graphic matroids in fact ~ can ordy make the negative correlation [5, 23]. to Pr[el ~] < that the pres- another element property Regular are known e, f in S. Real- and e has been in particular to be negatively corre- lated. *Work done primarily while the author was at Bellcore. Here Permission to copy without fee all or part of this materisl is granted provided that tha copias are not made or distributed for direct commercial advantage, the ACM copyright notice and the title of the publication and its date appaar, and notica is given that copying is by permission of the Association for Computing Machinery. To copy otharwisa, or to rapublish, requiras a fee and/or specific permission. 24th ANNUAL ACM STOC - 5/92/VICTORIA, B. C., CANADA @1992 ACM ()-89791 -51 2-719210004/0026 . ..$1.50 we strengthen lation to that is 6alanced M(S, B) itself, satisfy minor the then or those this selecting operation if all negative notion of negative We its say that minors, correlation is obtained operation that of selected 26 the of a matroid forming and the of “balance”: either do not. those including M property. (A repeatedly an element bases that In the case of graphic corresponds edges from by of choosing the to contraction graph.) corre- a matroid pere E S cent ain matroids, or deletion e We first exchange establish graph (Theorem G(M) both of G(A4 ), the number partition classes is at least as large partition sion properties had been conjectured class. of the to as the size Such strong expan- by Mihail [19], in fact for signific=tly graphs. M of edges incident of the smaller Vazirani the basesmatroid 1, i.e., for any bipartition has cutset expansion vertices 3.4) that of any balanced wider correlation property negative correlation ~ implies and an arbitrary monotone between property and, in view of the fact anced matroids, gorithm e and an element e m on S\ {e}, ● for their of self-reducibility an efficient to approximately Our results troid schemes we obtain of any balanced and classes of for two elements approximation count matroid. of bases of a matroid The core of our proof here is to show that the negative Monte-Carlo the number In general, can be efficiently for bal- randomized al- of bases exact is #P-complete thus show that size, counting [27]. the set of bases of a ma- sampled and approximately counted if balance can be established for the matroid. Balance is known to hold for graphic and more gener- ally regular matroids; some counter-examples to bal- ance are also known. Pr[em] which implies hood an enforcement analogous subgraphs ment to (formalized (this Previous encode paths the discrete geometry neighborfor as a ‘(ratios finally implies for expansion using elements and bound version certain enfome- last step was observed arguments involved We further expansion in [18]), which condition” , of vertex bipartite on G(M) sion for G(M) graphs < Pr[e] Pr[m] expanin [18]). of isoperimetries from and the path proof to bases-exchange known significance graph derives from ideas, see for example of the bases-exchange of the a sequence of well [1, 3, 9,14,15, 19, 25], Consider the natural on G(M): with random is the state If Xt one half Xt+l e from if X’ Xt = Xt\ wise Xt+l chain Xt {e} u {f} = Xt. uniform import oracle), proaching ● In turn, reducible efficient say, an Most of G(A4 ) suggests that Xt and hence possesses the rapid amounts, roughly, distribution [81 ). Therefore, uniform sampling balanced matroid. (given, to the which can it other- the Markov converges its stationary on G(M) efficiently and that has large conductance, for t= poly-log( = X’, and, be used to Xt arbitrarily the natural as an efficient apclose random almost scheme for the set of bases of any here are of different Diaconis can yield combinatorial uniform populations sampling yields that walk natural random ● c is The cardinality a bound on total a technique O ((n log rn+log than (Theorem 5.1). matroids, of O((n log rn+log version of the significance where the variation of n is the ground-set, distance, In particular, and conduct~ce and we obrandom E–l )mn2 ) convergence this arguments C-* )m2n4) result applies rate to and significantly imbounds. For graphic for the nat Ural random efficient conductance rate for the naturzil algorithmic graphic and regular matroids proves upon previously known pling of [17] we show that rather scheme based on the natural walk with for self- symmetry is as follows: balanced tain a sampling to en- to possess symmetry convergence arbitrary the was Jerrum type; in fact, matroid [7] and Mihail walk. previously strong by extending m is with as well as for a modified improvement For rank, only have been successful appear uses paths a sharper random this do not and Strook these any edge graphs [6, 14, 19]. Our path Furthermore, an analysis matroids, ahnost graphs the and of using the state-space arguments proba- the ratios through The [14]; such arguments graphs from path congestion argument e.g. matching properties. Choose to bound used here matchings, properties, related over 23 (it is symmetric). antly, the expansion property Xt+l ... edge of establishing of paths graph. only for combinatorial t, then and with It is easy to see that distribution mizing at time at random B then c t = 0,1, as follows: S uniformly can be simulated independence = Xt, is determined and ~ from Xt, (basis) probability y one half Xt+l bility walk walk code paths fractional the jlow method and Sinclair’s is derived by means [2, 4]. control any specific The technique congestion of certain turn of these paths 4). This gives an alternative condition differential as follows: ● path existence of the expansion through (Section in arguments the length of cut set expansion. enforcement can be used to define any pair of bases on the bases- such that congestion can be bounded known The algorithmic graph, [6, 14, 19], congestion [9, 16], and coupling paths between exchange to bound of combinatorial of the state-space path random show that balance walk, Broder’s yielded, convergence rate cou- roughly, [4]. For regular matroids, ments yielded Dyer and Frieze’s O ((n log m+log geometric e–l )m4n4) argu- ● probability We analyze a modification rate (Theorem 5.2). for matroids an exponentially ● of the natural and show O ((n log m + log e-l )n3) bound For regular This that bound of this matroids an alternative are tedious e-l)mn4) sampling walk but paper), (details straightforwhich running proportional theorem [28] (involving terminants and arithmetic modulo ning time O (mn4 log m). However, scheme is worse by a factor time. There is evaluation primes) and much easier to implement time braic nature; graphic implement matroids we need satisfy or fail to satisfy torial each step of the modified along the lines of Fredrickson O((n random O(log pling n) scheme .E(C) time graph. O(nm) sented ploits ● For the cover C’ is the scheme However, when in we obtain is the counted (More number only generally, ments for crease the convergence but walk unchanged.) O (m’n4 The the a balanced walk, Kirkhoff’s this leaves tree log m), ability edges without matrix graph with when follows of the further a large introduction matroid can Achieving 2 rates. scheme troids with edges family correlation in the family are 5.3. generally, parallel ele- lated, significantly in- are balanced all regular matroids and the fact that implies that all reguler their a large as “cycles” less efficient, number matroids for sat- all ma- For exam- graphic cannot of parallel imal time makes 28 are negatively minors are balanced. proof We of the fact that and cocircuits and “cut s“, by analogy matroids. corre- are also regulm are balanced. We refer to the circuits on to implement. any increase in the running that Then as well. matroids give here a new combinatorial based of matroids property. regular is both directions Balance random theorem then tighter ple, graphic matroids whose bases are the edge sets of spanning trees in a given graph are balanced. More where Theorem scheme a decom- 7 summarizes and outlines random alternative Section 5 to schemes cases from 2 and implies natural of the Section in certain for the modified rate in 6 introduces a minor-closed isfy the negative number time, of used research. Consider parallel of in the bases-exchange follows work and the ratios Section in Section of this from rates of two sampling that 3 ex- condition 4 uses the existence paths on convergence the context expecta- a large from and more complicated to introduce bounds and we Section one pre- the ) running the bounds The the a sampling of edges once; the construction matroids, derived are then matroids. property that We give a combina- a cert ain ratios mat things paths as follows. classes of matroids to balance. the convergence position up analogously. O ((n log m’ + log e-l )@n3 m’ time proofs for regular to construct for balanced [4] with we introduce case of graphs edges cover than the blows in the sam- Broder were of alge- Section expansion. these bound thus runs in expected faster edges time latter of parallel where alternative can be used Previous balance: to satisfy fractional condition can be implemented is m that for regu- rates for cer- with of balance balance graph; walk, For of balance the convergence counter-examples establish to results time. [2] and is clearly of parallel of the There That and here. number each step to Aldous time, which running [11]. due running underlying tion walk, time time random [11], log rn + log c-l)@n3) natural in proof discuss necessarily O(~) proof of the paper is organized run- certain For con- e.g. see [5] for the special case of graphic 2 is concerned OIU so probability matroids. tight). ● matroids. The remainder (furthermore is not regular Section even though present), with a structure reduce of de- with by edges proportional of its presence in a random exhibiting to significantly of n, it is very simple con- of the running of parallel tree is output to that failure of the graph. lar matroids, tain results tree in can be represented of the edge remaining each spanning figuration scheme based on Kirkhoff’s tree matrix analysis even for time to imple- random for the full a number We give a combinatorial we need O (inn) implementation in O((n log m +log our known polynomial that a spanning each edge has an associated to the probability large ground-set. and are left ceptually random to o@Ut (such probabilities introducing convergence is the first remains ment each step of the modified ward for example, where a graph convergence rate [8]. walk it possible, of a matroid with A cycle is thus a minimal be augmented to a basis, while set whose complement the case of set which a cut is a min- does not cent ain a basis. The regular binary matroids [28]. It is possible cycle C and g E C and are known defined matmids, to assign each C(g) to be the orientable by the following (1) values element g, so that = O otherwise; and for each cut 11 and each element if g E D and D(g) C(g) for each C(g) = (2) most &l if D(g) D(g) = +1 and cuts D containing With ~c(9)D(9) 9 for all cycles of graphic to all C and edges each in cycle setting on whether direction; graph, of the two edges the of each graph edges in the D its cut on whether to ~, setting ~, from the sum over all A is zero. by removing can be obtained one element basis IV defines complement cut DN We refer of IV. from as near-bases. a unique Every unicycle U defines is the following. D~(e)D~(~) bases Every near- cent ained a unique relating in the can be one element in U. A useful property and unicycles from to sets which bases by adding 1 ll~(f) = Cu(f) as unicycle Cu near-bases If U = iV U {e, f}, then Intuitively, unicycles tend in opposite containing tions, that are zero, while # O, then is a basis and tential result The equality fact that or equivalently, of the if one of them Aef = non-zero terms matroids, coincides whether a quantity If a unit as a potential from the the the- resistance is asof IBI of e, then the po- before of j is Aef. The for graphs with drop in [5]. We use indices subsets of B satisfying concerning direc- and a current the endpoints has been shown defined cuts tend to be tra- we can show that with networks: drop between below the presence certain con- or absence of certain in the chosen bases. 2.1 The buses of a regular = lBellBfl Proof. is non-zero, is a basis, an important two same or could be + Cu(~)DN(~) # O, giving = ~ llN(e)DN(~) N :from from near-bases on B to indicate rnatroid satisfy -A~f DN(~)/ZIN(e) From quantity, = B’\{e} If e # ~, in the sums follows arise from only (B, B’) we let U {g}, B“ = ~ l?= x I?.f, B U {e}\{ and assign a weight equal to C~uie}(g)Dw\ie](g). B“ zero, then the resulting in Be and 13zt respectively, = - ~ Cu(e)Cu(~). u expressions pairs we ob- tain pairs (1?”, l?’”) ~ 23. x Z3zf, by means of an exchange involving e and an element g + e. More so U\{t} we let A,f Aef e, ~ arising directions, specifically, define the quantity ~ in the enters and leaves at the endpoints -CU(e)/C~(~). We can now b(f). then if all four quantities iV U {e} C’u(~) e and signed to each edge in the graph, lBl”lBe~l say D~(e) ~ U:eGCU matroids, to traverse ory of electrical Theorem involved = e, f arising Aef sum property follows ~N(~) cycles cent aining For graphic quantity both in DN and in Cu are e and ~, so that by the zero O; the above equality the above versed by e and ~ in the same or in opposite elements we have Cu(e)DN(e) ~ for graphic whether ditions the only elements = measures = -Cu(e)Cu(f). To see this, note that then N:eEDN so ~ to A, it follows We refer to sets which to the cycles and IV U {e, f}, = a set A Since a cycle that the value of Aef. simply C(gp(g) A.f the edge is tra- direction. = of the conditions affecting if e belongs U can easily g#e of times g of C(g)D(g) contained separate A and become conventional complement from ~ traverse a cut the same number to ~ as from cycles. cut from = ~ 1 depending obtained then in the violating and without condition, condition case depending edge is traversed This the signs of all elements direction directions, in traverse in the underlying versed in the conventional will Intuitively, a conventional possible the all -D. assign one this equations = +1 for each edge traversed the of vertices traverse the C’ in C(g) D(g) cuts matroids, D(g), cuts involved = o it. by changing chosen cycles or cuts, without on C(g), importantly, For the theorem U = N U {e, f}. below, it will be convenient to select a specific element e and require that D(e) = –C(e) = 1 for all cycles C be enforced values g, so that = O otherwise; pairs iV, U such that property cannot The the with to this above weight e}(9). = exchange If this weight is nonand B’” are indeed bases and a non-zero Wei@ occur here for g = ~ since DB~\fel(.f) %’\{e}(9)cB’’’u{ 29 and B“~ g}, can From also be = 0. expressed the point of view as of a pair (B”, zero for l?’”) some g # B = B“\{e} a pair 6 Be x &f, e, ~, then weight the reverse and 1?’ = B’” U {e}\{g} U {g} (B, B’) if this is non- 3 From exchange Balance pansion, gives back to and Ratios, Fractional Match- ings 6 2.%x Z3~f. Therefore In this section troids. we derive The proof expansion is inductive for balanced on the rank intermediate e step, relies on a certain pei. p=fl = ratios E (z xf%f9#e (B’’,l?’’’)et,% = (x ~ z ~ I&l (~ CBu{e}(g)DB’\{e} g#e ~131/\{e}(f))( “ l~ejl + ~ follows. theorem troids matroids SS that D implies correlation troids not containing Some additional over GF[2] ). There that Ss as a minor matroids property that have to constitute is a binary ma- are balanced [22]. violate the negative been found counter-examples of the graphic spanning subgraphs as well as for transversals. if we consider edges, the path, 5 with add and connected pr[ef] the each an > ~” of the dual of subgraphs ~ and a transversal consisting replaced e joining a self-loop vertex neighbor- graph G(M) if for every element ~atios e is e c S the [18] : f by the with For every balanced graph G(M) matroid enforces M(S, B), ratios. Let AC t?= and let m~ = VBCA AeieB,eixe Proof. Note that the set of bases in Be satisfying e;. mA is pre- mA is precisely condition the set I’e(A). is equivalent Hence the first ratios to Pr[mAIZ] Analogously, > l?r[mAle] for A ~%, and note that (1) . = VB6A Aei~B,ei #e ~> let w the set of bases in B= satisfying ~ is the set A, and the set of bases in 23, satisfying ~ is condition is the set I’e(A). equivalent Hence the second ratios to in- of a path two paral- endpoints anywhere, then of In turn, below: the last two conditions follow from the lemma for 6 edges we have The matroid just of the dual of the graphic to be a truncation matroid For of fixed = Pr[e] Pr[f]. as a truncation can also be shown graph edge edge add spanning = ~ described (forests (connected [26], [24] t o nega- y) and for truncations lel 3.1 element corre- all binary recently for truncations ma- the negative of fixed cardinalit of length S of the binary the graphic cardinality) stamce, Lemma I’e denote the corre- cisely the set A, while the set of bases in B? satisfying [23]; it is known tive correlation Pr[e~] and hence balance does not satisfy property and shown holds denote a specific The bases-exchange the bases exchange are a subclass lation correlation following Z%, E) where edges that not involving Let fiuther bases-exchange (9)) matroids. (i.e., vectorial matroid = G(Be, of G(M) of sp ecif- cB’’’U{e}f)))) immediately for regular Regular let Ge(M) subgraph with B) More ‘~f Pr[e] Pr[ f ], so negative follow G(M), enforcement expansion. M(S, ma- and, in an B/ff~~ and the theorem The graph said to enforces B’tEt3e = a matroid are omitted. = +( for hood in Ge(M). X f3~ z (B,Bf)@FX13ef ically, to bipartite spond to exchanges ‘B’’\{e}(9)cB’’’u{e} (9)) DB1/\{e}(f)cBw”{e}(f) (B’’,B’’’)@3e analogous bipartite (B’’,B’’’)623ex17zt i7#e,f + (9)) ‘13’’\{e}(9)cB’’’U{e} Ex- Lemma matroid variables 3.2 (Main M{ S, i?), FOT every Lemma) any monotone in S \ {e} is negatively property correlated of the graphic Pr[me] as well. 30 < Pr[m] Pr[e] . balanced m over with the e: Proof. We show equivalently The reasoning set. The case where (and ity this is inductive M has rank is also the only of m is used). = Pr[~le] and Pr[m] = Pr[~] Note further that (i) that M erty m < Pr[ml~] pothesis on 237 for variable would (iv) by Pr[ml~e] Pr[~] the such that that (iv) Pr[rnl~e] < Pr[m[f] applying to the the 1, Pr[ml~e] Pr[m\~] there holds. by = Pr[m] exists always and with note 3.4 e] In between any number of o), which is equivalent is finally equivalent completes the to Pr[nz[fe] to proof Pr[ml~e] z of Lemma < A: and (iii). element IAI path ~ 3.2, (with Pr[fle] > > Pr[mle], which Pr[ml~e]. This and The proof show that if ml Lemma over disjoint then Pr[mlm2] sets of variables < Pr[ml] that fractional bound A nonnegative each vertex graph if there weights to the in u c U, the on u is IVI, sum of the weights Corollary 3.3 the G(U, to exists every = G(13e, Z?z, E) Proof. from making Consider G. by making IBe I copies edges between in Ge. eve~ Enforcement B), expansion any Thus edge the was obtained and be walk to ~; the on the hence the by in the number of there- 1~1). O that from ?natroid M, exchange variation the gr’aph conductance time byratios in [18]. bases total c of and and directly The each cutset proof balanced mixing: bounded A . [13\/2; previously any at C B, there in inductive enforcement is A expected IC’(A)I and shows expansion random expected of expectations passes paths is rapidly for of paths edge 1~1/113/ 2 m.in(lA/, For means the from some alternative 3.5 that specific constructed z 21A/. by random is @ > distance t = d(t) Q(nlog m + log e-l )m2n2. admits edges in E such a Proof. The bounded of that sum of the weights balanced element the for with in v E V, the from the I*I a fractional bipartite graph copies of each basis of each for graph 4 matching. G* is the the by 1131(1 – @2/2)t product in [25], where transition Hence, of the natural by by l/2mn. has been case of symmetric of the cut set expansion. d(t) NIarkov probabilities from random Theorem walk, 3.4 @ can D From Fractional Path Congestion Matchings to obtained basis of adjacent G(M) above @ in the distance B), in f3=, in Z3z, and including of each pair of ratios M(S, variation and the definition on v is IUI. matroid total conductance chains of edges e E S, the bipartite admits all copies such through 4.2). that, define of bases congestion, An be bounded any and for M(S, has cutset can A is at most Ic(A)I can an assignment of edges incident G.(M) of M by linearity on path l/2mn, ground-set, V, E) and for each vertex For to it, are satisfied. matroid 4 we argue through but, leaving natural properties pair . 1~1 paths Corollary 3.1. can be extended from equal Pr[m2]. a bipartite matching incident of the lemma and 7n2 are two monotone of each correspond matching balanced , we (Corollary C(A); paths that Section leaves G(M) 00 Remark: Say > Pr[~le] every paths lB1/2 are < f#e for some ~, Pr[flme] of P that graph matchings Remark: Hence For Proof. fore f#e of edges for a fractional fractional most lemma (iv), by (ii) some In particular, hy- Pr[m[ (i) the copies 1. it was the principles: By identifying to each edge of E a weight the bases-exchange (iii) m the P. G* has a prop- inductive then and hence c1 fact and property number the conditions by ap- on B~ for condition, assigning Theorem the Hall’s matching and to the that from forced averaging perfect Pr[m\~]. Pr[~] > Pr[nz[~e], +Pr[~] We argue < G* satisfies basis Pr[7nl~e], to O. If, in addition, + Pr[~] Pr[ml~] + Pr~] f that the monotonicstep note monotone Pr[ml~e] follow Pr[~] by ~ forced case that Pr[ml~] variable < Pr[m]. + Pr~\e] hypothesis the Pr[ml~e] the inductive Pr[m\~e] (ii) inductive with where Pr[~\e] is balanced, the Pr[rnle] size of the ground- n = 1 is easy to verify point For the Pr[mle] plying that on the implies In order to obtain are tighter bases the fractional now the previous 31 than bounds those of Corollary matchings section on convergence to defined construct 3.5, in rates we shall Corollary paths that joining use 3.3 of all pairs of bases in the matroid, while anced the number keeping of paths We thus wish exchange path bases-exchange The dom permutation Consider el,.. a specific each ~; is either present or absent tination B’ destination in B’. Note are initially path with i steps, arranged with B’ in the first B has exactly number i elements, of paths initially, when that i = after are unifornily O and i steps, over = ’61 ...e~?;+~ and at f3i that are already l)th ing step. The from 23~+1 to given steps, the paths arranged over t3~ = {B’}, Lemma thus with Bi+l. ensuring The B’ number The paths i are those For each tined with such to it leaving B’, before bases wit h destination B’, B’ required. The Note ements gives i, there are in the consider a sequence e2, B2, . . . . em, B~ = it agrees with uniformly B from in the so that 1/2 4.4 The B’ , B~ chosen differ in ei is present from each. such a choice we expect Since to H expected from length of a path from 2n. Lemma = ~ to be fixed. ~ of two bases con- of elements 4.3, is the . . . , em is the random were and the element difference is at most = B’ B;_l path order where from B in whih B to = B’ the edges D uniformly Say that are at a path leaving a Bel ...ei_l. Corollary through so Mutliplying this desti- . . . &_l ] as The By [~~1 a basis all Now Lemma Lemma number Bi in 4,3, where is chosen that E [~~1 Pr~~lel,. 5 From Path 4.1, . , &;-l]], permutations if this basis its destination. expected Pr~~l@l,.. over B and .,,ei., I paths of possible 4.5 is not a basis B = B;l ,,6,,, is at most Proof. [BIE goes through but this the is chosen uniformly from . . ,ii–1]] <2n. Congestion paths expectation the expectation which B. of 2nlt31. ei’s is being are chosen. uniformly from Be, ...~i. implies Cl to Rapid Mixing same ei a basis, Then the expected . Bi_ 1 and B differ in e; is at 2n times. Follows andel, 1) Z?el ~i_lz. ... IB[ paths des- IB[ Pr[~lel of paths el, . . . . ei_l. that n elements, at most BO, B1,... O that than same number Proof. in 1231/lBel Ili?&,.+ej_lzl quantity B’ B, ei is chosen probability Corollary . . . 13~_1]. over in expectation. the each any B to any are all paths the arranged by the number nations step are uniformly at each of these quantity at (i+ with @ has only ~ in B&l ...~i at step some length through so that If ei is chosen happen match...~i. of paths a basis destination paths 2n. t ains the Mel basis B in t3&, ...&i at step i is [13\ Pr[~[41 of 1t31/z. of paths Bo, el, Bl, is chosen other it, then exchange are number expected a basis in B, since the symmetric B’ where after Bi is on the path ezpected i is at most lerm-na. Given of ei such Proof. as desired. 4.1 Proof. most holds weights m steps, step of i matching that destination After the are currently to the associated matching, where number induc- to the minor 4.3 el, . . . . ei and The paths there for the a fractional the the following elements destination to each possible chosen proportionally fractional Such number . . .&I]). is at most number of bases and elements after of a fractional 3.3 applied the probabilities Assume with at Bi that Bi+I. by Corollary B. ~~+1 = B&l ...ei~. at 13i+1 are left paths at expected B’) bound B, agree basis condition to expected one sense that such the The edge (B, order and Lemma after Z?i = 1$+1 u t?&l, are sent to 13i+l by means exists that This L?i = expected an exchange the are uniformly in the the paths arranged ‘;+1 (i+ B’ any we give des- over ensure at each steps is the same for all of them. B~+l with each basis destination the ei is arranged paths with In of S. over Bi = B&l ...ei. the set of bases that expected tively m elements the paths B ‘. We shall that . . . &~_~],Pr[~le~ 4.2 through a ran- on whether destination the choice uniformly set of all bases B, in that such Corollary one = B&, ...&~. where that using [131rnin(Pr[e;l&~ M ( S, B), depending and i, in the bases- the B’ step paths and B and a destination with ., em of the ei or Z, lengths matroid begins during of a bal- small. ]13[2 paths of an origin construction path each basis of a balanced for each choice B’. the through to construct graph graph bound holds a particular and ej with for paths exchange i < j entering involving can be two used ~. el- Consider exchange only 32 the natural graph G(M) random of walk a on balanced the basesmatroid M(S, 23), as defined tion we yield show bounds on ral random icantly better and those of Corollary consider half rate Yt+l (ii) is a near-basis choose = YtU{~}. tionary, the exactly 1/2. sampling Hence We proceed between flow pij In let Let random pair A and B of states paths use edges denote paths an upper path from been defined B. the and suppose that Let rAmB Next the a basis is distance at time an equation chain results along the stating in Markov among hi(t)= Pr[Zt that hi’s at 0: each step of the M.arkov averaging of of (j, i)’s such should bound the discrepancies that pji # O; thus decrease: pij can be obtained (j, i)’s with endpoints we give = i] –Ti, and above length state of the that have one path that assigns A to state (i, j). that is more significantly different of j and an outline we shall the averaging is more convergence of the effective rapid) if it (and involves discrepancies at the i: Now recall pair A and B of states these paths use edges PAB denote event that the Then event A to B. for the variation bound that edges t A to t: of the the (so that # O), let choose edge the stating along any on the probability a specific a bound hence that the PAB that Let B, paths and all paths is left as follows. Let use hi as short for 33 1 denotes one path assigns A to state edge that path O). Let and the E PAB denote in the path from of the path from an upper FinaUy, defined bound recall that bet ween pairs at random according mAZB B. that Recall that (i, j). Then we are using length B probability on the probability a specific maximum was used any (so that p~j # A to and let ij of lAB. between chain that from have been that of [7, 21] (except than such the length that that state bound cent ains for the (i, j) defined Markov path length we choose from (i, j) denote recall among of the was chosen, edge expected have been chosen /Ajj on the upper proof here # Assuming distribution from ji’ow between all paths of states fumlly discrepancies for which chain that And the ergodic flow. defined path matrix path paper; an endpoints the distribution full of the Pji of of states detailed discrep- chain ergodic ergodic such to the contains bound The such that i)’s can be attributes y to discrepancy of (j, the tot al discrepancy distribu- states on the expected to the the following d(t) of the (i, j) pairs L be an upper path that analogous following of in terms any two been Finally, according chosen this bound between probability the and Yt is sta stationary the have A to E. at random is cut a Markov measure denote w that these an equation st ationarit techniques [17], to use as a rate for the = ~ipij i and j, let so First from using in choose space S and transition w~j is the same between # O, ancy Next, ~. appeared the endpoints D is lD1/2n[13[; convergence % denote wij states when Yt is on particular, on state + 1), that = ~i that derived: Yt+l is 1/21231, and cut that = Yt \ {e}; Yt is also appropriate to bound z~, t=o,l,... with the at random that d’(t) to those y one then Y~+l Let modified ~ S is the of a basis to check for t3. congestion. of Zt. Dfi probability scheme P’ = !j(P and It is easy to verify the The unifornily of a near-basis furthermore tion Dy, probability probability path and ~ from walk one half at random hard of bases probability (i) if Yt is a basis Yt uniforxnly if ~ random probability it is not modification even faster. with as follows: First, natu- uses both With hi(t). are signif- 3.5 in terms the modified becomes = Yt, and Yt, then Yt+l this of the here which Yt is as follows: determined e from walk sec- congestion rate further for In this path derived We random walk on convergence the bounds near-bases; convergence bounds than natural random the the walk; conductance. of the in the introduction. how here) L is an the chosen along expected we get: to to the the path lines rather d’(t) – (i’(t + 1) Theorem 5.2 balanced matmid bounded bye parallel in state the elements collapsing; proportional of elements in it, sum for size of the state due l/mti., all we get the the above and desired noting bound that on the d’(0) ~ variation distance. troid &f(S, lary 4.4, B) we have and definition 5.1 balanced matroid bounded bye the bases on any balanced by Corollary the natural POT the total in time t= modified 4, except The as for that the expectation 4.5. Paths whose by first random, tween bases, paths with thus due to paths the above not hard to verify L = (2n + 1)/lB[. J=4n graph elements whose choosing increase as a starting by associated reduces the differences. been the factor To reduce point weight, for so that collapsed. walk, unlike of parallel time. A basis random the TAD Therefore the natural elements can of largest weight be found weight affect needed by while for random does not of maximum walk elements and on any The bounds for balance matmid number of parallel maintaining The and greedily preserving the modified remain random the same elements after are added balance. algorithmic significance 5.3 was pointed path definition w = of 6 Better out of Theorems 5.1, 5,2, in the introduction. Path Congestion, composition, graph; De- Symmetry, Series Parallel are de- In uniformly path congestion for the modified Furthermore, 5.3 arbitrary by Corollary basis are only the walk while through near-bases to Theorem in number a random path endpoints that and in the obvious paths 2n[Bl include endpoints observations have random the bases, in the may ba- of the weights and n~aX / T~aX .b=i~ is m is the number of elements after where the number selecting an graph of the bases-exchange reducing be between expected a neighboring arbitrary can paths of such by the endpoints then on any are used to simulate is at most choosing and walk bases-exchange near-bases a vertex latter at the c-l)n’m. walk, number the fined and distance m+log random expected through 4.2 random variation Q(nlog an edge can be bounded paths parallel to start ele- of any given independence. the edges of the bases-exchange way. m/n, the running w = l/[ Z3[nzn, 1= 2n by Corol- but number on the resulting by n for weight increase, the transformation space, increase this view of elements This may to it. by m~aX _&& at most walk, ma- Hence are defined Section walk L = l/21B[ of L. Theorem For random a large avoid and a basis of maximum the modified For the natural To multiplied we choose walk element to the product to the be replaced can cause IS I. weights m~.X/rti. quantity, may can be parallel the probability near-bases. random using of the each as a weight sis is now the where on any e-l)n3. of elements elements, of parallel cut case size ment walk distance t = Q(n log m+log elements space random van”ation number all parallel after total the large These this Finally now a very collapse the modified the in time Consider have For be- due to congestion bases. From of L it random l/2nlBl is walk graph a balanced matroid, pair each we have of bases, near-bases constructed and bounded Since each basis has n incident each element on each of the basis, edge to be bounded by for some a, the n3 factor random walk be- load nl 131 on the load one might for paths the sis vertices, for of about and a bound rate for the modified 34 of bases edges using of a IBI holds + 2. Thus consisting tween load and the on of ba- edges, one expect the 1231. If a bound of convergence can be reduced to an2. We shall SUCh bounds. the and a fractional with weight lZ3e1, and note N.. constraint using basis the in B., that a near-basis to for edge each in ~e N U {f} to graph by the above convenience, if weflv. other For ~ U O, wef u = O. The requires wefN try edge U)e~N = 1~1 for wefu by N giving = then exact conservation have thus we now add these each weight basis, from e and for resulting weight cases, the Nit equation. with U with simple matching fractional constraints, namely that assigns and a near-basis. has the fractional exact weight [B!. the = sum weight the 6.1 the If a balanced reduces matching sum of m fractional the One can also consider the which only property always of the basis requires holds for N U {f} is Consider the first the element chosen N U {j} At its The indeed mat things, be for the first to N. matching, e is chosen its path the n3 factor random step, and If e = f, then the load and uniformly the expect ed load joining if the obtained the (i + l)th the i elements then the load This case occurs fractional el, e2,.,. weight ing walk Let e be consider an the load of is determined it is precisely at random Wef N. from all m is by case occurs load for this this 35 quantity if ~ is one of the first to N will contracted occur ~ IBI in i/m. is zero. Otherwise, be involved probability step is again over or deleted, in N or not. case. bounded aIl m steps by ~ depend- This Since (m – i)/m, in the the chosen so the expected this i steps, consideration for some e = e~+l after they with for probability m – i elements, uniform as the with , e; have been bounded paths, e~ used N U {f} matching matroid each element of the for the edge under on whether has only a basis step el, ez, ..., the edge joining matching a balanced one for k of the is a very these property then and step of all paths. edge is ll?~[. If e # f, then Since ‘ [B[. there fractional edge pToperty, (a + 2) IBI that edge, degree that and all its minom by adding, 1231= ]Drvl to the matroid rate foT the modijied edge joining this e c CU. each satisfying to each can follows to cxn2. Proof. e E ~N, matchings. Note say that decomposition holds, ~ symlnefrom a-decomposition is at most of the convergence is X= [Z3eI = nlZ31, and uniform IB[ We and is equal times D,N. matchings matchings, fractional basis or near-basis e, f, N. property, n, since congestion conser- of e. Suppose is ~eeD~ weight C if symmetry all decomposition this satisfy e. For wefu two fractional near-basis, for each basis for each near-basis both m of the i matchings the ~ e and m fractional each hold for that Lemma basis ,U, we set ?UefN = each choice each N with ~ # e we write in addition, constructed 23 to JV, one for must wfeN a-decomposition elements, from ~ and for some a. Note a= wef N the to the of fractional = 123e] for given edge for we must e f for property by the fractional z We this if wefN holds, weaker IBZ[ to each matching of f,~ given of this e = j’ is /13jl, with = IZ371 for given with an element joining definition ~ O, and the weight weights N U {f} graph mat thing N U {e, f}, choices WefN that first to equations E f z namely} basis associated denote fractional = matching the and n[t3[. of N, the N as representing this bipartite a fractional cut that add to IB[, and since Note is & means edge joining e in Me a weight N in Afe and the assigned vation e 1231. e from this edges joining then bases-exchange belongs weight that this have in Afe does exist. Given that to [13eI to each basis in Ii?z and weight near-basis each weight by removing each near-basis the JVe such of weights, to a particular the fractional can bases, each with assigned weight e, and we can infer weight cut, must joining e, one all at each near-basis to be satisfied a corresponding edges near-bases weight the If we view with the bases in Be can be matched The leaving is simply e. In terms can yield an element in Afe, simply 113.1, satisfying of I&l all that of them. in Be, and that between associated to the near-bases each graph Given matching to their To see this, conditions bipartite near-bases. find belongs examine assigned Consider bases now minor load is now this second the expected ]231. Adding of the paths (plus the . initial load obtain assignment a bound thus the ns factor to bases), for the we congestion, convergence not The main obstacle in obtaining the for fractional bounds matchings plicitly known. explicit construction We now is the that of f. If the direction ing e determines not ex- such an 6.2 D(e)D( then f )Aef a >0 for symmety regular well and matroids elements and exact for hold. the matmid components are cluded (as all trY its K4 or quantities are nonnegative the conservation wefN laws: = - ~ = a DN(e)DN(f)Aef. These by assumption, and satisfy If e E DN, matroids elements D characterithat Fano matroids, for graphs, as well. a complete depending We do not that set wefN in particular binary the f, as before have matroid, satisfies symme- on whether of N form an an exboth a basis with {e, f}). graph. We hold regular one of the hold is not since for in traversing are series-parallel (settingWe f N = 4,6 or only on graphs either minor then for the minors above decomposition, end- f, e, f in travers- used They hold The the with decomposition of the balanced exact The and matmids zation e, f, decomposisition the gmphic Proof. x D sat@es decomposition that holds biconnected series-parallel matmid cuts exact are precisely whose all such as symmetry) minors regular swaps through direction all its minors properties but is an isomor- an endpoint graphs. and since the two can be obtained. If the there e fixed and exact Remark: Theorem leaves e shares for series-parallel K4, because used by a cycle so symmetry a case where vertices, of K. fact are in general give any points for the val- a-decompositions share phism rate O ues of a achievable that near-bases of (a + 2)1231 on the path reducing to cm2. from the anced then know whether an a constant a-decomposition property exists holds for such all bal- matroids. 7 Conclusions and Open Prob- lems DN(e)DN(f)~CU(e)Cu(f) f #e f The = ~ (DN(t?)@(r?))2 =U more, the if e c CU, holds There and wefN thus = Wf .N holds exact are two dual obtained wit h either series. from by spond The als of Ks and Ks,s then in fact be K4, through same tions; otherwise, different but the that isfying position. the conditions In a clique K4 two one cannot regulm cycles f in opposite would f ) cannot different traverse give for symmetry for perfect e in sion 1, and hence and shown matroids. There- of the decom- = O if e, f An do above obtained efficient noting midpoint f(Z) to solutions, for is the exambasesma- cut set expan- sampling the role that and ran- question is = akZk the vertex-pairs in the graph) the proof ~~>~ at for bases of balanced (as used combinatorial meet balanced implies schemes intersect and for bases of graph introduced balance see [19] and interesting pO@IIOdd 36 that It is worth at their expansion; that case, of a graph We have A re- combinatorial is the this is a by remov- necessarily correspond this approximation meeting all edges For matroids, graph. are ex- O-1 coordinates). graph matchings [19]. troids, ~ O, sat- and exact one has Aef ple, corre- of the polytope. can be obtained O-1 vertices two and have coincide; is the set edges 1 (a O-1 polytope such edges must graphs [19] of a con- O-1 polytopes expansion vertices where two a K4 mi- whose ). For cert ain structured the exchange all Vazirani vertex-set and that complete midpoint domized C. the matroid whose on O-1 vertices edges (two direc- be nonzero cycles horn problems, must find whose their component = -CU(e)CU(f)Aef Kn, du- a bicon- as a minor polytope graph cutset and I-skeleton faces (edges) states with other the polytope to l-dimensional panding, ing if it is a series-parallel traverse C(e)C( the so the edges e, f that signs for two DN(e)DN(f)Ae~ then furthermore, graph, two given direction It follows have if and only edge of the conjecture lated or (2) two edges in a biconnected In a series-parallel cycles nor. and K~ a single are also excluded, [28]; do not namely as minors, cent aining has no K4 as minor graph. edges, be graphic nect ed component that K4 by replacing are excluded must matroids of Mihail Define as the graph of vertices decomposition decomposition, (1) two parallel If both matroid graphic or exact conjecture unresolved. vex polytope 123,1. Further- Wef u = property of Aef, ymmetry and K~, fore ~f following remains as well. satisfys the then the symrnetry definition 1 = Iq. UeeCu u Similarly, ~ play in proofs of Theorem evaluation at definition Wme of 2.1. of the point ~, where ak n + k. is For the number graphic the number of spanning the number of connected (1 - p)np~-n~(~-l) where pressions graphic t ained The quantities eration for eigenvalues value f(1) gives plied subgraphs, f(z) reliability generally sets (forests the of joining two property that does not ficients ak; for instance, balance for spanning graphs hold for ex- case of the individual subgraphs grateful comments; Dyer, A. Paul Seymour Paul property. istair Sinclair, Madhu Peter Winkler for for coef- [1o] to pointed out the nega- Sudan, several to thank Umesh valuable Vazirani, and [12] “On and simulated and Inf. [2] D. Aldous, “The of a matroid, 193. “How dom? (On nent),” STOC [4] A.Z. the Broder, trees? hard [14] Eng, F. for trees,” SIAIW [15] the pp 375– matroids CombinatoTial l?roc. stmc- Calgary In- 1969. FOCS 1991, data structures and k smallest pp 632-641. On the tree graph J. Applied is it to marry of the Math. 25, pp 187– M.R. Jerrum rapid mixing the approximation 1988, Jerrum, perma- computing of the L.G. Science M arkov property Valiant, generation a uniform L. Lov&sz D.J .A. Welsh, of Jones Proc. Carob. and Phil. “Conductance for Markov permanent and chains: resolved,” pp 235-243, puter pp 442-447. Math. and A. Sinclair, the M.R. and complexity 108, pp 35-53. from spanning Vertigan, polynomials,” at ran- [16] random D .L. computational “Random pp 50-58. 1989, ran- estimating 1989, functions, and F. Harary, Jaeger, STOC pp450-465. “Generating FOCS in construction labeled approximation 1986, simulation distributions walk 3, 1990, “A for STOC connectivity SIAikl SOC. 1990, pp 33–46. random Math. Broder, chain Probability trees and uniform J. of Disc. [3] A.Z. Markov combinatorial 1, 1987, Carnegie discussions. annealing,” Sci. spanning the uniform algorithm “Ambivalent 2-edge C. Holtzmann, Tutte for to- randomized 1992, Kannan, applications, trees,” and References method R. bodies,” Conference, dynamic “On Aldous, and polyhedra,” theiT Al- [13] [1] D. walks, and “Submodular certain tures spanning insightful “Random LPCO time of convex J. Edmonds, and in Section his We also wish of Ap- 1991. zonotopes, Frieze, polynomial for in particular, correlation Frieze, [11] G .N. Fredrickson, to chains ~ Annals 1, 36-61, algorithm”, ternational are bounds 381. Acknowledgements tive M. dom 2. We Vol. “Geometric to appear. volumes a counter-example connected [9] op- vertex, simplex Mellon, of T [13, 29]. a single the see the dual under at dual ~(z – 1) = in the are multiplicative Strock, of Markov Probability, and A. [81 M. Dyer tally unimodular sets can be ob- arguments D. and for fail- polynomial; of spanning by interchanging and the of the last two quan- Tutte instead [7] P. Diaconis gives More for independent matroids) size ~(0) spanning T is the of value trees, [27]. sets the p; the evaluation is #P-complete 2’(1, z), spanning gives network ure probability tities of matroids, and of combinat distribution,” 43, 1986, the Vazhani, strut tures Theoretical Com- pp 169-188. and M. Simonovis, chains, V.V. orial “The mixing An isoperimetric inequalit volume 1990, ,“ FOCS rate of y, and pp 346– 354. [5] R.L. Brooks, W.T. Tutte, squares? C.A.B. “The Duke Smith, dissection Math. A.H. Stone, of rectangles J. 7, 1940, and into [17] M. Markov 312-340. expanders [6] P. Dagum permanent Journal and M. of graphs on Computing, Luby, with “Approximating large factors,” “Conduct Mihail, chains ),” (A ante and combinatorial FOG’S 1989, convergence of Treatment of Pp 526-531. the SIAM [18] M. Mihail and M. ties of matroids to appear. 37 Sudan, ,“ TR-89, “Connectivity U.C. Berkeley. proper- [19] M. Mihail, and U. Vazirani, Theory, [20] Series Naddef [21] and ity in O-1 polyhedra,” Series A. Sinclair, “Improved of Markov chains [22] P. Seymour, [23] W. Pulleyblank, Journal B 37, 1984, PTOC. of Latin expan- “Hamiltonicof Combinatom”al pp 41-52. bounds and for mixing multicomodit rates y flows? 1992, to appear, personal communication. P. Seymour and applications of an inequalit chanics,” the of Combinatorial B, to appear. D. Theory, “On Journal sion of O-1 polytopes,” ikiath. D.J.A. Proc. Welsh, “Combinatorial y from C’amb. statistical Phil. Sot. me- 1975, 77, pp 485-495. [24] P. Seymour and P. Winkler, personal communic- ation. [25] A. Sincl@ counting, and M.R. uniform Maxkov chains? Jerrum, generation “Approximate and rapidly Information and mixing Computing, to appear. [26] M. Sudan, [27] L.G. Valiant, and reliability Computing [28] D. Welsh, personal “The y communication. complexity problems 8, 1979, Matroid ,“ of enumeration SIAM Journal on pp 410-421. Theory, Academic Press, 1976. [29] D.J.A. Notes, Welsh, DIMACS “On May the Tutte 91/Dagstuhl map,” June Lecture 91. 38