Chapter On the Number of Milena 18 Eulerian Orientations MihaiP Peter Abstract orientation We give efficient randomized schemes to sample and approximately count Eulerian orientations of any Eulerian graph. Eulerian orientations are natural flow-like structures, and Welsh has pointed out that computing their number (i) corresponds to evaluating the Tutte polynomial at the point (O, –2) [8,19] and (ii) is equivalent to evaluating “ice-t ype partition functions” in statistical physics [20]. Our algorithms are based on a reduction to sampling and approximately counting perfect matchings for a class of graphs for which the methods of Broder [3,10] and others [4,6] apply. A crucial step of the reduction is the “Monotonicity Lemma” (Lemma 3.3) which is of independent combinatorial interest. Roughly speaking, the Monot onicit y Lemma establishes the intuitive fact that “increasing the number of constraints applied on a flow problem can only decrease the number of solutions”. In turn, the proof of the lemma involves a new decomposition technique which decouples problematically overlapping structures (a recurrent obstacle in handling large combinatorial populations) and allows detailed enumeration arguments. As a byproduct, (i) we exhibit a class of graphs for which perfect and near-perfect mat things are polynomially related, and hence the permanent can be approximated, for reasons other than “short augmenting paths” (previously the only known approach); and (ii) we obtain a further direct sampling scheme for Eulerian orientations which is faster than the one suggested by the reduction to perfect mat things. Finally, with respect to our approximate counting algorithm, we give the complementary hardness result, namely, that counting exactly Eulerian orientations is #P-complete, and provide some connections with Eulerian tours. such that this Research, 445 pwklfl.ash.bellcor South Street, close uniform) to orientations ining counting tion function is equal orientations of some underlying ther see also [2,14]). or hardness ZICE, for one ing scheme the first our #P-completeness result lack of any formula or other ZICE for efficient exactly. to perfect the Ising type ing (The list monomer-dimer model, model which percolation walks, remain In addition, counting for approximation plane. to ations, that trees reliability, the topology in [12]; and the problems and ice- involv- self-avoiding on and the map problems the that Tut t e plane forests, configuration nodes, preof the corre- are, colorings, model, point the ((), –2), results on the of a net work, through that corresponds at the point Ising practical [8,19] hardness translation is a uni-directional flows directly in [3,9,10,18]; orientations and curves one further preserves the to compute contains treated fundamental or spanning explains translates treated polynomial a direct Other points is while problems also observed Eulerian the Tutte have 5.1) polyominoes, therefore here count3.4) open.) to evaluating sented answer- ZICE, method which it has been problem the (Theorem here; fur- evaluating (Theorem of Welsh’s theory, which for Thus, approximate has been (for no reasonable approximate efficient is treated parti- of Eulerian graph graphs. and has been which number our problem matchings observed crucial was known to of exam- Welsh the orientations scheme of Eulerian problems Eulerian questions, Eulerian Eulerian of some Apparently, planar ing the of Welsh’s to the result 4-regular of in the context model”, “ZICE” algorithmic set In particular, “ice-type of arbitrarily the number by Welsh [20]. questions graph. complexity physics even the out finding efficiently; the Eulerian computational in the so-called that a distribution of counting was raised details with from and that tation 138 generation in statistical From e.com concerned edges directed known can be accomplished are significance the of its of edges directed of edges is well of an arbitrary orientations orient Morris- we orientation number It orientation (i.e. The to the dOUt (v). sampling represents *Bell Communications town, NJ 07960. mihaiI@Hkh.bellc ore.com = paper example, Consider an undirected Eulerian graph, that is, a graph in which all vertices have even degree. An Eulerian is an v the number an Eulerian in sp ond Int reduction graph vertex din(v) Tutte 1 of the for every v is equal v: a Graph Winkler* towards of of for acyclic etc. of view, if a graph an Eulerian of the thus, oriennetwork a maximum EULERIAN global ORIENTATIONS flow sampling without inst ante number of Eulerian ing the of such number category but paradigm suits, concerns ber of perfect paper the obtaining equivalence and sampling and Vazirani Broder and imately counting Markov chain also [4,6] for further for schemes input which run graph, accuracy, and perfect over graph bilities, can in time ratio perfect while ing edges. #P of the count- the lM._ known and In Section size of the 1 l/lMn small I of the failure in- proba- 3 we show pling and a class graphs obtain t ations. It on I&tn_l treated previously pander graphs arising from by showing for which [ for (dense [10], graphs that perfect remark classes with sequences augmenting path graphs, log n length for (e.g. expander = that large [11] random factors etc.) were matching constant graphs to sam- length etc.) for 0(n4), Eulerian of graphs [3], for any near-perfect “short” for graphs approxi- matchings lMn_ll/lMnl to l/lM. degree and reduces solutions interesting bounds sampling counting efficient is that orientations approximately of we Eulerian orien- all known that were and ex- [4], graphs this large way for Markov of (hence the cuts, to may random rapidly an tions into paths mixing. efficient matching~. and Euler Using sampling a polynomial no of In scheme factor over In Section 5 we show exactly is #P-complete, that has no arguments paths on the orientations facilitates the Hence, Eulerian indirect counting thus that 4 we technique the simulation for to and Section our congestion. directly paths chain applications such how transition many an “expander” y in chain MC states introduction number elements without an is to between too further outline one property populations edge. Markov graphs related example, on such the de- structures that for difficult transition of ver- which mixing” with walk we a technique For such [10] the pair of the Markov see in graphs cardinalities argue bound same decomposition of dense any the Eulerian there Thus and define a natural of and ver- paths. of paths [6,4, 12] for Now decouple use on that, two difficulty any “rapid graph and like elements” congested fast; outgo- overlapping and a random reasoning, saves is a is converges improvements). is “Euler number chain, chain incom- we introduce defined the underlying “small” graph between crucial to orient than arbitrary populations. chains Markov of in in handling the outgo- than The overlapping a substantial the behave problematically establish both 3.3 we show edges. into combinatorial to incoming graphs paths obstacle and ways () d–1 more outgoing problem, graphs two edges. so that of incoming is that problematically Decoupling to a fact disjoint obtained for parallel isolate is a recurrent edge of near- with of so- is when 2d parallel Lemma all favor- number the edges more locally, To bypass couples two solutions constraints case with re- 2d Monotonicity of decomposing this counting In u has two of the simplest are only larger to the car- hence decrease number v has on in our particularly number to orient same such consider approximation numbers exponentially (see randomized in the cannot novel sampling via total of approx- the that ways there in some sense, tices connected tices. Valiant, [7, 11,12, 15] for desired matchings with so that edges, sampling be achieved of the u and () d v have the edges efforts for specifically, suggest are the class using polynomial inverse the (and More 2d There but only connected ing scheme [10] and can based much arise are related the v are u and I is not that exhibit the num- as usual). mately thus the the for which example, can be seminal Jerrum, matchings matchings counting re- the concrete improvements, For polynomithe bound. I is not sets of orientations, problem establishing Sinclair technique results). perfect that lMn_I of graphs “increasing edges, approximate the a flow vertices why problems, behavior: 1I/ lMn cardinalities of certain only yields argument. class these ing directed by by and perfect various approximate for result is that more of the with which on lMn– path the reason [ for finalities able bound of bipartite in Valiant’s established followed [M. to flow-like in approximations. simulation relevant results hardness Jerrum than lutions’). one our augmenting even hardness counting of randomized [13], [3] the matching but is collapses and to be complete was In contrast, outstanding is also waa shown efficient The significantly a perfect of problems mat things time, latter matchings, Roughly, in the solutions, the most which that matchings In turn, towards put known on permanents [18]. harder algorithmic in polynomial falls unexpected the case of perfect it is well constructed can be associated a short of max- a solution set of all Perhaps our matching few near-perfect network. question constructing are either behavior, both the orientations unless occur. of such use for graphs: ing sense, classes we shall of view, the or intractable, complexity the the count- number the apparently countin$ in an approximation involved, to ally duction, for which time, and counting around total Eulerian of problems of sampling flows point counting in polynomial while larger perfect a behaved. a theoretical and Consequently to observing is equivalent of maximum are better From sampling sinks. a network, with 139 amounts orientations networks flows and orientations random imum sources Eulerian Presumably, OF A GRAPH choice MC of A4C is as orientations reduction Eulerian justifying of to orientathe neces- MIHAIL 140 sit y of our well efiicient aa the lack counting method. positions into The the known some terms 3 we a class Section can the chain defined faster sampling such an approach how the tours. rapid mixing directly on result and Summary property We faster 5 open and the the a connection and a In Markov ratio potential of counting lMnJ–l for algorithm numbers Let with probability outputs scheme for y at least for It$l for aim is to obtain approximation tations efficient, c-l, logd–l, Let IV(I of perfect graph running log6-1, = Let matchings respectively be algorithm that ● and where and near-perfect matching IVI = n. in n, where be the [10], there in n’, The log$-l, matchings of G’ is a mat thing 2.2. set of perfect scheme in is an runs the [?] and For scheme [Broder there Jerrum [6].] and improvements as above, the in M.,. logd-l, FACT [3], is a d-sampling matchings [10], [Broder improvements above, I{u – ~ any for in ratio For [4,6].] of matchings M~J. Sinclair < V- ={v the last graph G’ ● if out Eulerian, regard {u, v} P as c E towards then v), or of v). and a vertex v c V, the : (U, V) E F}l for ~, where d(v) zero charge is called of a positively directed P= degree balanced. vertex of v that must be balanced. (V, F), let : q(v) <O}. is the charged out v to become orientations V+= The {V E V : q(v)> of P is charge graph scheme where ?k = G’ runs for 1)I?%-11. the ratio k there verify from first matchings, one :dP=k} > IE I/2. we near-Eulerian. some in V+ sense, expect Eulerian This lMn/_ behavior last orientations bound 1 l/ lMn~ upon the I for for any cardinalities and k, that 3.3 show we that l~k I ~ n(n is polynomially graphs “more ?k _ 1, and Lemma generality, and flows in Monotonicity G, flow that in V–. in ‘Pk are in full graph by to verify one set of 2k edge-disjoint to reflect In hard to vertices than this ~h_~. this Furthermore, orientations constrained)) , it is not is at least vertices might any to charge considerations, from In easy near-perfect with 0 for preservation formalize for is to let Of Qk and Sinclair scheme equality analogy P~ ={P one any charge EV By Finally, as polynomial and < with of edges in order call time The may is directed ● F}l-l{u q(v) the principles. paths lM~J_ll/lM~~l. For we need necessarily We where edge : (v, u) number where severely (~, b)-approximation not P = (V, F) an orientation the set of perfect , Jerrum cardinalities terminology. is directed A vertex for any P G ~k 2.1. (the sets size n’–l). FACT the reduction, (c, 6)- graph, Mnl-l (a near-perfect in the P = (V, F), edge example, shall and problems. about orien- polynomial a bipartite M., by a efficient of v is reversed as usual. V~, E’) 2n’. (V, E), 2.2 imply 2 For S is an times which above outputs set ‘PO of Eulerian G= arise S is an 1 –&, d-sampling the Eulerian = (V{, lV~l for we mean and G’ = eficient schemes of any By (the of v in G. ISI such 2.1 and argue countfor bounded G = (V, E). an orientation v is the scheme of graphs an orientation, graph graph Clearly, an estimate approximately orientation to further (u, v) E F O}, and Our some a directed c For 6. that (c, &)-approximation approxi- of Euler ~lpr[s=z]-&l<d An and and Facts – 1 that P denote either XEs that M.) ~(v)= that, ratio is polynomial- a class is always Eulerian of an Eulerian Preliminaries a s G S such for in order to introduce charge 2 the sampling orientations l/lMn,l the and that sampling in n’, hence (v, u) E F are in Section For some set S # 0, a d-sampling to mat things However, ● a complementary with problems the of A4n, hence approximate contains and technique for discuss perfect schemes matchings decomposition f– 1, logd–l, n’, of Eulerian reducible ing to efficiently. orientations, to yield sampling perfect cen be handled Section hardness counting scheme. algorithms. from in 3 we show counting polynomial orientations that 4 we outline and of orientations. Eulerian Section mate follows: matchings, reduction counting of graphs yield for as polynomial In a connection organized results and approximately time WINKLER lA4nJ-11/lA4mIl. time is the in of decom- we observe for the treatment give approximately application as exact tours. 2 reviews sampling for of e counting, reasonable one further paper Section approximate elements the introduces and for of Eulerian rest Section By Euler wit h numbers In scheme of any computationally AND arise – ‘related to from the EULERIAN ORIENTATIONS reduction of counting perfect mat things efficient sampling Eulerian In this and section imate counting of any and [E I = m. Let The = where Yv {y~,i and v; = fiecE{w.} , 1 < i = each partition any Eulerian ( bY I-I.ev PROOF. can class G as is a pair exactly one cover G, first that with each so with and !. of divide each must matching Eulerian edge {xv, ~, we } of which be in of {Zu,e, the of G’ orientation {u, v} and direct notice be that matched forces the G, we} perfect orientation there in G’, fact, 11.<v that (~)!, balances the each of partition ‘ince ‘rice that XV and (~) 3.3. thaf Let gives class has ‘he ‘e’s perfect rise each that cardinality ‘atched of G, and matchings 2.1 and 2.2 of Eulerian be upperbounded by a reasoning lM~~-11 ! 1711, as in < 2~lM~/1+ and the proof follows ❑ fullpaper). where Lemma.) For any IVI = n, and for any lE1/2, q be a realizable vertex exists . q(v). Clearly and V-, respect decreasing by one We ~~1~ shall exactly ‘ith q is fixed, and q. v c V-, show and that leaving the K%(!7)I s I be the same V+ the few next V+ let u E V+, such %(q) v has charge and V- Q~ be the can be obtained some from increasing q by q(v) q otherwise the holds any following ‘: e Finally, for %-l(q’)1 u q’cQ~ To prove (1) we shall relate orientations orientations number each of in orient ation orientations tion Ug,e ~, ~fi-l(q’) of orientations in in ~k (q) is ~k (q) in Ug, c Q, ?k_l(q’). relate orientations and in ~fi (q) to so that, roughly, in Uq,c Qq Tk_ 1(q’) of in is by v E V-. what related than with course, Tk (q) reversing In larger related a path follows sition in Euler elements the with the number each orienta- the with called “Euler elements” so that the be reversed can be e~ily accounted a nonand For that by one for some Let q(v). fixed functions q(v). in ‘Pk (q) have that on V, orientation each vertex k = ~Ucv+ to the q(u) for same. clearly function” one charge such that we assume with least all orientations and paragraphs are “charge at v 6 V haa the set of orientations u c V+ re~oning to G, natural waY orientations in with we endpoints introduce into techniqueto decompose an Eulerianorientation Eulerian mat things, are can (Monotonicity there U g, ~ Q, Tk_ l(q’) with. perfect orientations and near-perfect G = (V, E), l<k< suppose to outgoing to the to see by similar orientation XV, associated Hence according in YV in in their incoming are associated class vertices orientation. it is easy Eulerian to is indeed vertices vertices matched corresponding partition ~ d(v) ~ be they v, all of the can be partitioned that each each a unique this •l graph straightforward forthe graph PROOF. i.e. mat thing Thus To see that ~ to Furthermore, empty as (v, u). remaining clearly of the mat things that v} for to unmatched we ‘s, which edges {u, of G is obtained. Eulerian, which but we. orientation remain LEMMA that approx- of G’ of YV ‘s. set ~ () number efficiently perfect a unique For edges otherwise must tedious are left k such Further- the ways ! the , where PO and T, near-Eulerian of G if I?l l/l’Pol in n.] By (details the the G. ~V~V approximate matchings 4jQ with Eulerian are perfect -1- 1)2 ~ucv Eulerian If {XV,., w,} is in the perfect matching of G’ then direct {u, v} as (u, v) in the orientation of to and set of charge follows: of (%Y = hence G, of of G if we can XV’s any For 3.1, it can be shown x Y.. relation has cardinality of perfect associated of E’ x. !$!l)1!$ Note be = of G’ can be partitioned orientations of the 1~~1< n(n - l)lp~-11 graph are in one-to-one efficiently the number edges = 2m, are = O (n21?~\/\%l) in the size of G. orientations imate d(v) there with Xv Uvg” half 3.2. orientations polynomial PROOF. U (JVEVYU, u LEMMA Lemma and ‘we }} IVJI = ~VeV classes can ~}, < We}, {Z”,., the set PO of Eulerian Eulerian n to a perfect graph V, of G’. [Hence, by Lemma 3.1 and Facts we ~an efic2entlY Upproxamate the number [VI= where each other are Eulerian orien- where be a bipartite “ Xv, the MO I and MmI_l of Eulerian problem for [A4~-1[/lM~~l and approx- (V, E), of this of perfect matchings we G= Uv {{ZU,~, the partition more for results. sampling Xu’s () Counting = e}, n’ = IV(I [Hence, approximation V( LEMMA 3.1. For that schemes of the matching graph. the set P. V;, E’) size of G1 is polynomial M~I half we obtain is as follows: (V{, = Ue~E:e={u,u} Clearly for graph bipartition UeCE:e={u,v}{~v, the reduction problem G’ vertex Eulerian efficient schemes Eulerian to counting thus counting Approximate we present tation mat things of any to establish Sampling orientations 3.1 and 3.2), and approximate We proceed 3 Eulerian (Lemmas orientations 141 OF A GRAPH a so- number of paths to for. The decompo- is as follows: in (I). A circuit (w1, w~), (w’, of an orientation w~), . . .. (wl_l, P= Wl), (V, F) is a sequence (WI) Wl), where all 142 MIHAIL the (wi, (wl, Wi+l)’s are distinct WJ2), (wz, WI)} as (~z, W), (W, a generalized (Is). A circuit (lb). (w3, A circuit in F. there (w, is no starting circuit that, i.e., w2), e.g. circuit there are ~W ~vlv+uv. ~ for a circuit is each Now point.] Pin ~ () U Ug,cQ, ‘P~(q) to prove ! distinct WINKLER sets of pairings ‘P~-1(/). (1) it suffices to show if no edge is incident (w,vguv-(%)!)l~k(q),~ of P is a V+-circuit one incoming [Note is the same U V-. in the to a vertex w), of P is a free in V+ edges edges W3), (W3, Wl) w3), cycle: to a vertex two (wl, AND circuit if there incident and one outgoing, in V-. Analogously, to some and (wJ!uv_(%) ,,y’,-(q’), are exactly vertex in V+: no edge is incident define V--circuits. Consider ~wEv\v+uv- ~ copies ! of each ori- () (II). A (V+, (w~,lo2), (tom, w3), and v), moreover, as in V-)-path (w~, all the circuits, has point: (V-, (III). a V+-circuit, a (V–, v), all point: u, V+)-paths, or or or a (V+, entation of pairing wi’s in F; (which, are tained )- in a free a (V+, circuit, or V-)-path, V+)-path, or or a (V–, V–)- to V+)-paths, lated they P = p’, that U~~~Q, %-1(~’) tinct each orientation for of the ~ F) haa exactly nW6v\V+Uv. partitionings see this, P = (~ of its each edges vertex into ~ () elements. Euler U V– w c V \ V+ c ‘P~(q)U ! disTo consider each (P’, Now k >1 counting completes ! pairings () in w. edges pairing. For The suggested of the some vertex partition by some incoming w, of F into fixed let and q = {pW elements : w E V \ V+ proof p’) are re- that and is related by related a (V-, to 2k+c is related Vf)- (P, p) ’s, to at most (2) holds c+l true, be- by elementary and hence (1) also holds step where we allow of Lemma is in suggested path to see that It true. q to vary 3.3: outgoing pW denote Euler the )-paths also be c (P’, 2k + c ~ c+ 1, and final V- must and in P’ ~’) P’), V– )-paths there reversed (P, p)’s hence following then if (P, p) (P’, @ = p. p’, V-)-path if (P’, it is easy and (V+, moreover 2k + c (V+, one considerations, The one is a (V+, Hence p’, to decompositions the one of these ~’)’s. cause for orientation. if P can be ob- are that identical except in P. then We argue have which by and only one of the by if there are suggested (V-, them by reversing each waY P of the is related if and are suggested for one edges P E %(q), T,_l(q’), see that that ~k–l(q’)j the where P’ that path path. (P, p), from P’ P’ U.Uq,cQ~ partitioning P’ E Ug,~Q, easy V+ )-paths, and that in V- an ending (V+, Say in ?,(q) where not a (V+, and WJJ, defined. of P is either a V--circuit, V+)-path, (u, Wl), (u, edges the distinct) are analogously element where to circuits, starting (V-, sequence are distinct and necessarily Finally, Euler a (w, in contrast V-)-paths An is w), W;+l)’s not that, a specified v).] and P v c V–, are [Note path (wi, u E V+, V+UV-. of . . . . (w-l, such a that is U V– } is as follows: . Free circuits are $7W2(W, W2)) PW(W~, w1)(w1, ● V+-circuits W1),. ... and the wi’s E V\ are of analogous ● (V+, the W) form: = of (w,, PW, (W1, WZ), where are PW, (W, of (W> ..., the the W2), (W2, w~) = W), = (W, Wi’S E V \ V+ form: Wz) U V-. (u, WI)) (WI, (w~, ~) = P~,(w/-l, V+ U V-. And, W2) = w~), where u c V+ of course, V–-circuits form. V-)-paths where are of the form: (u, WI), (WI, W2) . . . , (w1, V) = PW (w1-1, wl), = Pwl(%wl), where the u~V+, VGV–, and the wi’s ~ V\ V+ UV–. And, clearly, (V–, of analogous It of F of pairings (V+, V+), and (V–, V–)-paths are form. is not into V+), hard Euler •1 THEOREM IVI to see that q suggest different this and is indeed that partitions. two a partition distinct Furthermore, sets n, For there approximate counting orientations of G. n, E‘1, logd–l, PROOF. 2.1 and elements, = 3.1. (d, $)-sampling schemes The for G= (V, E), and (c, S)- the set ‘PO of Eulerian running time is polynomial in Follows by Lemmas 3.1, 3.2, an 3.3, and Facts ❑ 2.2. REMARK. We prove of Theorem is needed. are graph and log8-1. the proof any Eulerian In the Lemma 3.4 following 3.3 for only the section general ii, while special case we need both k = in 1 cases EULERIAN ORIENTATIONS k = 1 and that estimates k = reduction full 2, and in to perfect 3 at the end general rather matchings We shall 143 a potential ITO I directly generality. Remark OF A GRAPH than Lemma have more of Section scheme through Let the 3.3 is needed to say about this with PA= q(u) in ence” in symmetric 4. PAB V\{u, Euler Decompositions Arguments Theorem 3.4 suggests by simulating latter UMnJ_ Mnl property (this attention; etc. obtain a direct rapidly mixing the is the elements for so that bounded [10]. the the by the Sinclair “path chain on received [1,3,10,16,17] we outline how PO, by simulating to a on ‘POU’P1. Except is rather use obvious) for the new of decompositions definition of “random path encoding” method chain can of Jerrum probabilities on state defined be and U ‘PI 7’o as follows: be the state = (V, F(t)) space on time If “heads” i!; Pick (u, v) E F(t) uniformly F(t) \ {(u, v)} u {(v, u)}) G 70 UP1 action circuits and to another have is thus to move walking by reversing and to the check aperiodic, distribution involved unit charges from edges that and one Define the unit to show MC that with from a ({u}, on the therefore it converges U ?1. It irre- to the uni- is significantly over PO that MC has the rapid mixing prop- THEOREM 4.1. For any Eulerian graph G and any < 2°(m)e-o(*) (Outline). of Eulerian we shall bound edge. standard This, path in arguments turn, and {u}- v. by its {v})- restriction reversing for {v}-circuits, a free circuit. its from u. A ({v}, edges successively So a ~A&canonical c P. edges {u})one path of steps 1) !c! distinct its a (u, u’) reversing + as then starting from for of a sequence c’!(c and by edge PB of ({u}, no other edge starting to order); from on states PAB-canOnical pair PA from uniformly at path P~/) random, uniformly path E(PM, &) < bound and on d(t) with of pairs from the 6 PI, by first pick a choosing then chosing a at random. number along combined PB to PB Let E(P~, be an edge in MC. the expected It can be argued and PA (PA, PB) Pjwl) such PA to PB uses (PM, lines [Po of the a bound of [4,5] U that PMJ ). that pll~(~4) , theorem in follows by the [6]: . We shall any pair {u})- -paths, (v, v’) reversed indeed path (PM, the c ({v}, PA starting first from representative and = PA to PB. pAB above, pAB circuits, any reverses with by reversing are each t: pAB(W)’S, from is reversed a is starting from denote edges the terms and induces a decom- (in {u}) (and one at a time There Let ({v}, circuit {v})-path ~AB-CanOniCal more let and outgoing free path one at a time, is reversed some let distinct circuits lexicographically For charges of {u]-circuits, ({u}, w in PAB, of w, {v})-paths, by alternating A for the incomincoming pAB ‘s, where include the {V})-PATH order). edge the PROOF. paths successively paths ({u}, which reverses edges path c+l reverses the paths ation is symmetric, with dAB(w)! a “~AB-canonical” first reverses along orient one more vertex distinct u must than {v}-circuits. PA to PB consists ert y: time circuits, in c’ circuits “representative” is trivial ducible, and of MC. traversed. It form MC of paths, clearly d AB (w)! of PAB paths, at a time then P(i! + 1) :=P’ else P(i! + 1):= P(t); So the position successively at random; have in (same while outgoing of incoming (B denotes all vertices edges); more “differ- are balanced {PAB(W) : w C V \ {u, v}}. Following the reasoning of Lemma 3.3, each pAB A else (where in-degree=out-degree are HWev\{U,UI and then P(t + 1):= P(t) lJ’=(v, be a pairing one that canonical the There outgoing c PI the in PAB For a balanced pAB w. PB = (V, FB) Consider that v must edge. denote (w) let and they one similarly outgoing into congestion and and with dA~ (w) and much even degrees degree edge, and = —1. Notice of incoming odd than in mixing fair coin; Toss a If proposed In particular, transition Let P(t) for MC expected MC be a Markov Let and (which here have set To rapid is referred scheme chain the Mm. As now section of MC point paths”, this sampling Path is a Markov by reader In for the so-called has Markov definition technical Euler property details). for scheme possesses the unfamiliar for scheme scheme sampling I which further q(v) : PAB = (V, FA @ FB \ FA) v} have ing a sampling a sampling in [3], the and ~ ?., and difference). number 4 (V, FA) = +1 near congestion implies [4,5,6,10]. first define Eulerian through rapid paths beween orientations any mixing and particular following where d-1 is the transition is rn- 1, 1 is the our length case is m. and probability of the c is E(PM. y which canonical PMI). paths in our which case in ❑ 144 MIHAIL REMARK by 1. A sampling simulating ciently MC small, claim fail tions yields scheme for and if I’(to and P(to) so that ) E 70 output repeat. Roughly, E 7’0, the d(to) is suffi- P(to), one and oriented for TO can be obtained to steps, out otherwise of nz simula- whole scheme is effi- receives receive exactly d(v)–1 of the edges out REMARK 2. reduction to perfect Broder’s The Markov of Section sampling only for through ‘Po is obtained the by simulating on A4nf UA4n/-1 [3] for the graph The convergence of this and, since MC reduction to perfect REMARK 3. chain = yields mat thing. for sampling mat things a perfect To is faster by a factor O(n’4), than defined these for ~~ chains = l$=op~ analogous for are rapidly approximate chains chains all mixing, estimates 1< rl & by (i) lPo1/1~11, m ‘111/2 ~ the ri’s l~IEI/2-1[/l@I121, are fl(n2)) (ii) s to V1 are oriented out V2 to t are oriented towards Xj VI Eulerian) are PROOF. We IVII = IV21 = perfect that n, argue = of lMn that (Vl, IEI matchings computing we a reduction G = oriented m, and G. It has with lMn that counts Eulerian orientations. Assume without loss of generality at least a graph E U E$ U Et, edges VI, so that parallel edges connect Hence are ~V~v, the graph Gm, connecting together with k parallel Now realize that edges symmetry l~kl (5.2) t to s, all edges connecting edges edges between ml edges t. We have lMnl implies that vertex t}, and d(v) for edges. – 2~ = together A’j s and with m’ where if m’ is odd, . set of Eulerian . . . . m’ then if the in terms m! orienta- is even, cardinalities of the and of the cardinalities of follows: as be oriented t must (5) holds be since i parallel edges remaining k<” from that () ways between V1 to be directed from there k are i of directing s to t and the s and ‘tfrbm t to s, and this forces exactly ~a” between s and s to VI (so s is balanced). calls to (one the an query (5) yield l~k, for Finally tours: graph. Let v into P = each {$% counts graph and time) lMml, from system (3), (4), in and unknowns lMml can be can be solved cl the following orient Let G= (V, E) p. be a partition :VE which the lm–2nj+l orientations . of Eulerian pairs hence equations is easy to see that by Eulerian G,), 2 independent we provide number to I‘s can be inferred that . . .,m’, (it in polynomial Euler l’Po(G~) oracle M’+ k=O, deduced the of G~t t are directed V2 to ‘S the k = 0,2, can be written (Gk)’s Gk is G’ s to t. orientation denotes where ., m’ Finally – 2 t to V2. Finally in general if ‘PO(Gk) k=l,3.. m’~k>[~~ l~m,-kl in each similarly s to V1 must connecting of the towards of oracle are and = of G~, connecting there connecting in an Eulerian that all is G! j V2 connecting [18] an d(v) – 2 parallel s to t, and = edges follows – 2) > ‘~u~~,(d(v) which such s, while to of edges s to VI and from of all parallel set in achieved that v E VI t with (d(v) edges be s to v in G’, connecting parallel shown set VI UVZU{S, consists of s, while lX~/1 edges edge v c Vz to = m’ edges consider E. each m! (not in VI and graph, the two. where for be + 1 calls G’ on vertex connecting each there m–2n Lrn – 2nj out (5.1) the perfect In what I can time Consider Mn I is #P-complete. computing of G has degree counting been polynomial s to orien- a bipartite let j of the t. G’ lL?ll/1~21, R Eulerian from be exactly of all vertices that tions for V2, E) that of G where of s, while orientations established PO give Let such of to is #P-complete. matchings. out and set of the by since counting the all Counting THEOREM 5.1. Exact tations Et denote necessarily Now of Exact mat thing connecting If R Hardness of u. So are oriented Lemma 3.3 l~lE112] = 2m, use (since and, V2 that in Gm, scheme 2m Hi ri. However, establishing the rapid mixing property for all these chains remains open. IPoI is exactly out connecting Furthermore . . .. u E V2 there orientation Eulerian t to V2 are oriented these the rest be oriented mat thing connecting simulating V2, while a perfect v c V1 each perfect be lE1/2. a d~rect define can MC k < then To can be obtained to obtain to s must all edges s to Markov from all edges are balanced, of n4. VI vertex is oriented V1 and G, and clearly rise to a unique Let the into graph each v to V2 must u to VI that between WINKLER edges edge from connecting for each vertex edges of V2 and gives is lMn,_l]/lMn,l of n4 simulations simulating rate out the original chain 3. one out Hence scheme matchings 2@I~je-ni&), 5 – 2 incoming one incoming of v. Similarly, clearly, all d(v) one edge connecting cient. G’ t. Consequently towards which AND (notice V}. connection ations and the be an Eulerian that of G Realize is the between number edges incident undirected). that of (undirected) p imposes Let a EULERIAN ORIENTATIONS OF A GRAPH decomposition/partition partition Let class Tk = that of consists I{p : p induces if p induces exactly tours hard k partition for to verify edges of G {u, w}, pu(e), one partition 2 Euler It is not the of {e= 145 classes class, then G, so 71 = where a [8] F. p. (e), . . .}. } 1. Notice this tours) /2. Proc. Math. Physics Hence, ’lPol 5.1, It might to specify . which some M.R. me worth ~k’s pp of the Tk’s must investigating are provably’ 6 be hard [11] (6) further [12] hard. approximately count the complementary tions we for treat hardness complexity solve case Remark The capacities M.R. of the on the results 3 in Section for [14] Extend edges; here (ii) Classify the Euler tours. parallelizable? M.R. Special very thanks us to Madhu due to Dominic discussions, to Welsh’s Sudan, Prasad Vijay Vazirani for stages of this work. and work. for to Madhu Thanks Tetali, many Welsh [16] Umesh discussions for also due Vazirani, during early Jerrum, D. Aldous, for On the Uniform lated Markov CombinatorisJ Annealing, [3] A.Z. Broder, in Eng. Solved Academic Simulation Distributions Probability 1987, pp 33-46. [2] R. Baxter, Exactly chanics, chain Methods Method and and Inf. SimuSci. 1, in Statistical Me- at Random? (On Press, London. How Hard the Approximation is it to Marry of the Permanent), STOC 1986, pp 50-58. [4] P. Dagum of Graphs and M. Luby, with Large Approximating Factors, the Permanent Siam J. on Comp, to appear. [5] P. Dagum, topes, FOCS M. Luby, M. Mihail, Permanents , and 1988, pp 412-421. [6] P. Diaconis and D. Strock, and U. Vazirani, Graphs with Geometric Large Bounds Chains, preprint. values of Matkov [7] M. Dyer, A. Frieze, and R. Kannan, Polynomial of Convex Time Algorithm Bodies,STOC for Estimating 1989, pp 375-381. PolyFactors, for EigenA Random Volumes chains: resolved, Fast ICALP for and the the Ap- STOC L.G. 1988, Generation 1989. Polynomial-time Ising Approx- Model, Tech. Report, V.V. Vazirani, (1990). Valiant, of Uniform Proceedings the and Combinatorial Theoretical Distribution, E.H. Lieb, L. ResiduaJ 162, Lovisz Ran- Structures from Computer Science a M. and the [18] FOCS Simonovis, Isoperimetric 1989, pp Square Ice, Physical The 1990, and Rate to and of Com- appear. Convergence Treatment M.R. Jerrum, and of Markov of Expanders), Approximate rapidly mixing and Computing, Valiant, Mixing Inequality, 526-531. and generation The Theoretical manent, FOCS Combinatorial Sinclair L.G. M. Conductance (A of An Volume, Mihail, Chains A. Entropy 162-171. Chains, counting, markov chains, to appear. Complexity of Computing Computer Science, the Map”, the Per- 1979, pp 189- 201. [19] D. J.A. Welsh, [20] D.J.A. [1] Systems of Statistical Conductance Markov A. Sinclair, Generation DIMACS References and Information and the M.R. uniform some Sudan are for of Edinburgh, Markov [17] are enlightening directing [15] 4, Acknowledgements Sinclair, A. Sinclair, Algorithms puting Re- and Jerrum Review (iii) (iv) 108, 43, 1986, pp 169-188. ques- (i) (1990), Journal Permanent Graphs, Uniform as A. property Jerrum dom and as well following 1. problems are our sample further: capacities of counting extend result. with with to orientations, to investigate graphs the To what 7 Eulerian are interesting the results schemes Sot. the Poly- 235-243. imation [13] randomized On Monomer-Dimer Intractable, and Mixing of Regular Summary efficient Welsh, and Tutte 48, pp 121-134. Jerrum University We gave PM. 2 Dimensional proximation k=l by Theorem to compute. = ‘y2’7i D.J.A. of Jones Carob. are Computationally that (y) and nomials, Rapid H Vcv Vertigan, Complexity [9] M.R. Jerrum, [10] (5.4) D.L. 35, pp 35-53. suggests (#Euler Jaeger, Computational May Welsh, “On The Comput Classical Problems order Physical (1990), in pp Tutte ’91 and Dagstuhl 307-321. from Systems, June ational Lecture Notes, ’91. Complexit Statistical Physics, Oxford University y of Some in DisPress,