ON PARTITIONS AVOIDING 3-CROSSINGS MIREILLEBOUSQUET-MÉLOUANDGUOCEXIN 6 To Xavier Viennot, on the o asion of his 60th birthday 0 0 2 Abstra t. A partition on [n] has a rossing if there exists i1 < i2 < j1 < j2 su h n it2hat i1 and j1 are in the same blo k, i2 and j2 are in the same blo k, but i1 and a are not ikn the same blo k. Re entlky, Chen et al. re(cid:28)ned this lassi al notion by J introdu ing - rossings,foranyinteger .[nIn] thisnew2terminology,a lassi al rossingis a2- rossing. Thenumberof2npartitionsof avoiding - rossingsiswell-knowntobethe 1 nthCatalannumberCn= n /(n+1). Thisraisesthequestionof ountingk-non rossing 2 k≥3 3 partitionsfor . Wepr(cid:0)ove(cid:1)that thesequen e ounting -non rossing partitionsisP- re ursive, that is, satis(cid:28)es a linear re urren e relation with polynomial oe(cid:30) ients. We ] k O giveexpli itlysu hakr≥e u4rsion. However,we onje turethat -non rossingpartitionsare notP-re ursive,for . C 3 Weobtainsimilarresultsforpartitionsavoidingenhan ed - rossings. . h t a 1. Introdu tion m [n]:= 1,2,...,n Apartitionof { }isa olle tionofnonemptyandmutuallydisjointsubsets [ [n] [n] [n] of , alled blo ks, whose union is . The number of partitions of is the Bell number B B 2 n n . A well-known re(cid:28)nement of is given by the Stirling number (of the se ond kind) v S(n,k) [n] k . It ounts partitions of having exa tly blo ks. 1 Re ently another re(cid:28)nement of the Bell numbers by rossings and nestings has attra ted 5 5 some interest [8, 14℄. This re(cid:28)nement is based on the standard representationof a partition P [n] [n] i (i,0) 6 of by a graph, whose vertex set is identi(cid:28)ed with the points ≡ on the 1 i n 0 plane, for ≤ ≤ , and whose edge set onsists of ar s onne ting the elements that 5 o ur onse utivelyin thesameblo k(whenea hblo kistotallyordered). Forexample,the 0 1478 236 5 standard representation of − − is given by the following graph. / h t a m 1 2 3 4 5 6 7 8 : v k P Then rossingsand nestingshaveanaturalde(cid:28)nition. A - rossing of isa olle tionof i k (i ,j ) (i ,j ) (i ,j ) i <i < <i <j < j < < j X edges 1 1 , 2 2 , ..., k k su h that 1 2 ··· k 1 2 ··· k. This P r means a subgraph of as drawn as follows. a i i i j j j 1 2 k 1 2 k k A di(cid:27)erent notion of - rossing is obtained by representing ea h blo k by a omplete graph k P k (i ,j ) (i ,j ) (i ,j ) 1 1 2 2 k k [13, p. 85℄. A -nesting of is a olle tion of edges , , ..., su h that i <i < <i <j <j < <j 1 2 k k k−1 1 ··· ··· , as represented below. Date:January 20,2006. Keywords and phrases. Partitions,Crossings,D-(cid:28)niteseries. 1 2 MIREILLEBOUSQUET-MÉLOUANDGUOCEXIN i i i j j j 1 2 k k 2 1 k k k k Apartitionis -non rossing ifithasno - rossings,and -nonnesting ifithasno -nestings. k k A variation of - rossings (nestings), alled enhan ed - rossings (nestings), was also studied in [8℄. One (cid:28)rst adds a loop to every isolated point in the standard representation i = j k k k 1 of partitions. Then by allowing in a - rossing, we get an enhan ed - rossing; in 2 parti ular,apartitionavoidingenhan ed - rossingshaspartsofsize1and2only. Similarly, i =j k k k k by allowing in the de(cid:28)nition of a -nesting, we get an enhan ed -nesting. [n] Chen et al. gave in [8℄ a bije tion between partitions of and ertain (cid:16)va illating(cid:17) 2n k tableaux of length . Throughthis bije tion, a partition is -non rossingif and only if the k k orrespondingtableau hasheightlessthan , and -nonnestingif the tableau haswidth less k k [n] than . A simple symmetry on tableaux then entails that -non rossing partitions of k [n] k n are equinumerous with -nonnesting partitions of , for all values of and . A se ond bije tion relates partitions to ertain (cid:16)hesitating(cid:17) tableaux, in su h a way the size of the largest enhan ed rossing (nesting) be omes the height (width) of the tableau. This implies [n] k [n] that partitions of avoiding enhan ed - rossingsare equinumerous with partitions of k avoiding enhan ed -nestings. C (n) 2 [n] 2 The number of -non rossing partitions of (Cusua=lly 1alle2dn non rosksi>ng2parti- tions [15, 21℄) is well-known to be the Catalan number n n+1 n . For , the k [n] number of -non rossing partitions of is not known, to the exte(cid:0)nt(cid:1)of our knowledge. k k [n] However, for any , the number of -non rossing mat hings of (that is, partitions in whi h all blo ks have size 2) is known to form a P-re ursive sequen e, that is, to satisfy a linear re urren e relation with polynomial oe(cid:30) ients [11, 8℄. In this paper, we enumer- 3 [n] 3 [n] ate -non rossing partitions of (equivalently, -nonnesting partitions of ). We obtain C (n) 3 for a (not so simple) losed form expression (Proposition 7) and a linear re urren e relation with polynomial oe(cid:30) ients. C (n) 3 C (0)=C (1)= 3 3 3 Proposition1. The number of -non rossingpartitionsis givenby 1, n 0 and for ≥ , 9n(n+3)C (n) 2 5n2+32n+42 C (n+1)+(n+7)(n+6)C (n+2)=0. 3 3 3 − (1) (cid:0) (cid:1) (t)= C (n)tn Equivalently, the asso iated generating fun tion C n≥0 3 satis(cid:28)es d2 d P t2(1 9t)(1 t) (t)+2t 5 27t+18t2 (t)+10(2 3t) (t) 20=0. − − dt2C − dtC − C − (2) n (cid:0) (cid:1) Finally, as goes to in(cid:28)nity, 39 5√3 9n C (n) · . 3 ∼ 25 π n7 C (n) n 0 3 The (cid:28)rst few values of the sequen e , for ≥ , are 1,1,2,5,15,52,202,859,3930,19095,97566,.... A standard study [12, 23℄ of the above di(cid:27)erential equation, whi h an be done automati- C (n) κ9n/n7 3 ally using the Maple pa kage DEtools, suggests that ∼ for some positive κ C (n) 3 onstant . However, one needs the expli it expression of given in Se tion 2.5 to κ prove this statement and (cid:28)nd the value of . The above asymptoti behaviour is on(cid:28)rmed C (n) 3 experimentallybythe omputationofthe(cid:28)rstvaluesof (afewthousandvalues anbe n = 50000 omputed rapidly using the pa kage Gfun of Maple [20℄). For instan e, when , n7C (n)/9n 1694.9 κ 1695.6 3 then ≃ , while ≃ . ON PARTITIONS AVOIDING 3-CROSSINGS 3 As dis ussed in the last se tion of the paper, the above result might remain isolated, as there is no (numeri al) eviden e that the generating fun tion of 4-non rossing partitions should be P-re ursive. 3 We obtain a similar result for partitions avoiding enhan ed - rossings (or enhan ed 3- [n] nestings). The number of partitions of avoidingenhan ed2- rossingsis easilyseen to be n the th Motzkin number [22, Exer ise 6.38℄. E (n) [n] 3 3 Proposition 2. The number of partitions of having no enhan ed -non rossing E (0)=E (1)=1, n 0 3 3 is given by and for ≥ , 8(n+3)(n+1)E (n)+ 7n2+53n+88 E (n+1) (n+8)(n+7)E (n+2)=0. 3 3 3 − Equivalently, the asso iated g(cid:0)enerating fun ti(cid:1)on E(t)= n≥0E3(n)tn satis(cid:28)es d2 d P t2(1+t)(1 8t) (t)+2t 6 23t 20t2 (t)+6 5 7t 4t2 (t) 30=0. − dt2E − − dtE − − E − n (cid:0) (cid:1) (cid:0) (cid:1) Finally, as goes to in(cid:28)nity, 216 5√3 8n E (n) · . 3 ∼ 33 π n7 E (n) n 0 3 The (cid:28)rst few values of the sequen e , for ≥ , are 1,1,2,5,15,51,191,772,3320,15032,71084,.... C (5) E (5) 135 24 3 3 Observethat and di(cid:27)er by 1: this di(cid:27)eren e omes from the partition − 15 whi h has an enhan ed 3- rossing but no 3- rossing (equivalently, from the partition − 24 3 − whi hhasanenhaEn (end)3-nκes8tnin,gbutno3-nesting). Again,κthestudyofthedi(cid:27)erential equation suggests that 3 ∼ n7 for some positive onstant , but we need the expli it E (n) κ 3 expression (33) of to prove this statement and (cid:28)nd the value of . Numeri ally, we n=50000 n7E (n)/8n 6687.3 κ 6691.1 3 have found that for , ≃ , while ≃ . ThestartingpointofourproofofProposition1istheabovementionedbije tionbetween k+1 k partitionsavoiding - rossingsandva illatingtableauxofheightatmost . Asdes ribed k D in [8℄, these tableaux an be easily en oded by ertain -dimensional latti e walks. Let Zk D n be a subset of . A -va illating latti e walk of length is a sequen e of latti e points p ,p ,...,p D i 0 1 n in su h that for all , (i) p =p p =p e e =(0,...,0,1,0,...,0) 2i+1 2i 2i+1 2i j j or − for some oordinateve tor , (ii) p =p p =p +e e 2i 2i−1 2i 2i−1 j j or for some . D Zk Q =Nk k We will be interested in twodi(cid:27)erent domains of : the domain of points with non-negativeinteger oordinatesandtheWeyl hamber(withaslight hangeof oordinates) W = (a ,a ,...,a ) Zk :a >a > >a 0 W k 1 2 k 1 2 k k { ∈ ··· ≥ }. Va illating walks in are related k+1 to -non rossingpartitions as follows. C (n) k k Theorem 3 (Chen et al. [8℄). Let denote the number of -non rossing partitions [n] C (n) W 2n k+1 k of . Then equals the number of -va illating latti e walks of length starting (k 1,k 2,...,0) and ending at − − . The proof of Proposition 1 goes as follows: using the re(cid:29)e tion prin iple, we (cid:28)rst redu e W k the enumerationof va illating walksin the Weyl hamber to that of va illating walksin Q k k the non-negative domain . This redu tion is valid for any . We then fo us on the ase k =2 . We write a fun tional equation satis(cid:28)ed by a 3-variableseries that ounts va illating Q 2 walks in . This equation is based on a simple re ursive onstru tion of the walks. It is solvedusinga2-dimensionalversionoftheso- alledkernelmethod. Thisgivesthegenerating (t) = C (n)tn 3 fun tion C as the onstant term in a ertain algebrai series. We then use C (n) 3 theLagrangeinvPersionformulato(cid:28)ndanexpli itexpressionof ,andapplythe reative teles oping of [19℄ to obtain the re urren e relation. We (cid:28)nally derive from the expression 4 MIREILLEBOUSQUET-MÉLOUANDGUOCEXIN C (n) 3 of the asymptoti behaviour of these numbers. The proof of Proposition 2, given in Se tion 3, is similar. 3 2. Partitions with no - rossing 2.1. The refle tion prin iple δ = (k 1,k 2,...,0) λ µ W k Let − − and let and be two latti e points in . Denote by w (λ,µ,n) q (λ,µ,n) W Q k k k k (resp. ) the number of -va illating (resp. -va illating) latti e n λ µ C (n) k+1 walksoflength startingat andendingat . ThusTheorem3statesthat equals w (δ,δ,2n) k . The re(cid:29)e tion prin iple, in the vein of [10, 24℄, gives the following: λ µ W W k k Proposition 4. For any starting and ending points and in , the number of - λ µ Q k va illating walks going from to an be expressed in terms of the number of -va illating walks as follows: w (λ,µ,n)= ( 1)πq (λ,π(µ),n), k k − (3) πX∈Sk ( 1)π π π(µ ,µ ,...,µ )=(µ ,µ ,...,µ ) 1 2 k π(1) π(2) π(k) where − is the sign of and . = x = x : 1 i < j k i j Proof. Consider the set of hyperplanes H { ≤ ≤ }. The re(cid:29)e tion (a ,...,a ) x = x 1 k i j of the point with respe t to the hyperplane is simply obtained by a a i j ex hanging the oordinates and . In parti ular, the set of (positive) steps taken by e ,...,e 1 k va illating walks, being { }, is invariant under su h re(cid:29)e tions. The same holds for the negative steps, and of ourse for the (cid:16)stay(cid:17) step, 0. This implies that re(cid:29)e ting a Q x = x Q k i j k -va illating walk with respe t to gives another -va illating walk. Note that x =0 e e i i i this is not true when re(cid:29)e ting with respe t to (sin e is transformed into − ). Q w n λ k De(cid:28)ne a total ordering on H. Take a -va illating walk of length going from to π(µ) m , and assumeit tou hesatleastonehyperplanein H. Let be the(cid:28)rst timeit tou hes x =x m i j ahyperplane. Let bethe smallesthyperplane it tou hesat time . Re(cid:29)e t all steps w m x = x w′ Q i j k of after time a ross ; the resulting walk is a -va illating walk going from λ (i,j)(π(µ)) (i,j) i j to , where denotes the transposition that ex hanges and . Moreover, w′ m also ((cid:28)rst) tou hes H at time , and the smallest hyperplane it tou hes at this time is x =x i j . Q k The abovetransformationisasign-reversinginvolutionon thesetof -va illatingpaths λ π(µ) π that go from to , for some permutation , and hit one of the hyperplanes of H. In the right-hand side of (3), the ontributions of these walks an el out. One is left with the π walksthatstaywithintheWeyl hamber,andthishappensonlywhen istheidentity. The proposition follows. (k + 1) This proposition, ombined with Theorem 3, gives the number of -non rossing n q (δ,π(δ),2n) k partitions of as a linear ombination of the numbers . Hen e, in order to (k+1) q (δ,µ,2n) k ount -non rossingpartitions, it su(cid:30) es to (cid:28)nd a formula for , for ertain µ k =2 ending points . This is what we do below for . 2.2. A fun tional equation k=2 Let us spe ialize Theorem 3 and Proposition 4 to . This gives C (n) = w ((1,0),(1,0),2n) 3 2 = q ((1,0),(1,0),2n) q ((1,0),(0,1),2n). 2 2 − (4) Q (1,0) 2 From now on, a latti e walk always means a -va illating latti e walk starting at , unless spe i(cid:28)ed otherwise. ON PARTITIONS AVOIDING 3-CROSSINGS 5 a (n):=q ((1,0),(i,j),n) n (i,j) i,j 2 Let bethenumberoflatti ewalksoflength endingat . Let F (x,y;t)= a (2n)xiyjt2n e i,j i,j,n≥0 X and F (x,y;t)= a (2n+1)xiyjt2n+1 o i,j i,j,n≥0 X berespe tivelythegeneratingfun tionsoflatti ewalksofevenandoddlength. Theseseries t Q[x,y] are power series in with oe(cid:30) ients in . We will often work in a slightly larger Q[x,1/x,y,1/y][[t]] t ring, namely the ring of power series in whose oe(cid:30) ients are Laurent x y polynomials in and . F (x,y;t) F (x,y;t) e o Now we (cid:28)nd fun tional equations for and . By appending to an even ( 1,0) (0, 1) (0,0) length walk a west step − , or a south step − , or a stay step , we obtain Q x 2 either an odd length walk (in ), or an even length walk ending on the -axis followed by y a south step, or an even length walk ending on the -axis followed by a west step. This orresponden e is easily seen to be a bije tion, and gives the following fun tional equation: F (x,y;t)(1+x¯+y¯)t=F (x,y;t)+y¯tH (x;t)+x¯tV (y;t), e o e e (5) x¯ = 1/x y¯ = 1/y H (x;t) V (y;t) e e where , , and (resp. ) is the generating fun tion of even x y latti e walks ending on the -axis (resp. on the -axis). (1,0) (0,1) Similarly, by adding to an odd length walk an east step , or a north step , or a stay step, we obtain an even length walk of positive length. The above orresponden e is a bije tion and gives another fun tional equation: F (x,y;t)(1+x+y)t=F (x,y;t) x. o e − (6) F (x,y;t) F (x,y;t) e o Solving equations (5) and (6) for and gives x t2y¯(1+x+y)H (x;t) t2x¯(1+x+y)V (y;t) e e F (x,y;t)= − − , e 1 t2(1+x+y)(1+x¯+y¯) − x(1+x¯+y¯) x¯V (y;t) y¯H (x;t) e e F (x,y;t)=t − − . o 1 t2(1+x+y)(1+x¯+y¯) − (1,0) (0,1) Sin e we are mostly interested in even length walks ending at and at , it su(cid:30) es H (x;t) V (y;t) e e to determine the series and , and hen e to solve one of the above fun tional F (x,y;t) o equations. We hoose the one for , whi h is simpler. F (x,y;t) Q o 2 Proposition 5. The generating fun tion of -va illating latti e walks of odd (1,0) length starting from is related to the generating fun tions of even latti e walks of the x y same type ending on the - or -axis by K(x,y;t)F(x,y;t)=xy+x2y+x2 xH(x;t) yV(y;t), − − (7) F (x,y;t)=tF(x,y;t2) V(y;t)=V (y;t2) H(x;t)=H (x;t2) K(x,y;t) o e e where , , , and is the kernel of the equation: K(x,y;t)=xy t(1+x+y)(x+y+xy). − (8) t Fromnowon, wewill veryoftenomitthe variable in ournotation. Forinstan e, wewill H(x) H(x;t) F(x,y) t write insteadof . Observethat(7) de(cid:28)nes uniquely asaseriesin with Q[x,y] y = 0 H(x) = x+t(1+x)F(x,0) oe(cid:30) ients in . Indeed, setting shows that while x=0 V(y)=t(1+y)F(0,y) setting gives . 6 MIREILLEBOUSQUET-MÉLOUANDGUOCEXIN 2.3. The kernel method We are going to apply to (7) the obstinate kernel method that has already been used in [4, 5℄ to solve similar equations. The lassi al kernel method onsists in oupling the x y K(x,y) variables and so as to an el the kernel . This gives some (cid:16)missing(cid:17) information V(y) H(x) about the series and (see for instan e [6, 1℄). In its obstinate version, the kernel methodis ombinedwithapro edurethat onstru tsandexploitsseveral(related) ouplings (x,y) . This pro edure is essentially borrowed from [9℄, where similar fun tional equations o ur in a probabilisti ontext. x Let us start with the standard kernel method. First (cid:28)x , and onsider the kernel as a y Y t 0 quadrati polynomial in . Only one of its roots, denoted below, is a power series in : 1 (x¯+3+x)t (1 (1+x+x¯)t)2 4t Y = − − − − 0 2q(1+x¯)t (1+x)(1+3x+x2) =(1+x)t+ t2+ x ··· x The oe(cid:30) ientsofthisseriesareLaurentpolynomialsin ,asiseasilyseenfromtheequation Y =t(1+x+Y )(1+(1+x¯)Y ). 0 0 0 (9) y =Y Q[x,x¯][[t]] 0 Setting in (7) gives a valid identity between series of , namely xH(x)+Y V(Y )=xY +x2Y +x2. 0 0 0 0 Y =x/Y =O(t−1) F(x,Y ) 1 0 1 The se ond root of the kernel is , so that the expression is not well-de(cid:28)ned. (X,Y) = (0,0) t K Now let 6 be a pair of Laurent series in with oe(cid:30) ients in a (cid:28)eld K(X,Y) = 0 K x y su h that . Re all that is quadrati in and . In parti ular, the equation K(x,Y) = 0 X′ Φ(X,Y) = (X′,Y) admits a se ond solution . De(cid:28)ne . Similarly, de(cid:28)ne Ψ(X,Y)=(X,Y′) Y′ K(X,y)=0 Φ Ψ , where is the se ond solution of . Note that and are Y′ =X/Y X′ =Y/X involutions. Moreover,with the kernelgiven by (8), onehas and . Let Φ Ψ (x,Y ) 6 0 us examine the a tion of and on the pair : we obtain an orbit of ardinality (Figure 1). (x,Y ) 0 Φ Ψ (x¯Y ,Y ) (x,Y ) 0 0 1 Ψ Φ (x¯Y ,x¯) (x¯Y ,Y ) 0 1 1 Φ Ψ (x¯Y ,x¯) 1 (x,Y ) Φ Ψ 0 Figure 1. The orbit of under the a tion of and . The 6 pairs of power series given in Figure 1 an el the kernel, and we have framed the (x,y) ones that anbe legally substituted for in the main fun tional equation(7). Denoting Y Y H(x) V(y) 0 ≡ , we thus obtain three equations relating the unknown series and : ON PARTITIONS AVOIDING 3-CROSSINGS 7 xH(x)+YV(Y) = xY +x2Y +x2, (10) x¯YH(x¯Y)+YV(Y) = x¯Y2+x¯2Y3+x¯2Y2, (11) x¯YH(x¯Y)+x¯V(x¯) = x¯2Y +x¯3Y2+x¯2Y2. (12) 2.4. Positive and negative parts of power series + Asimplelinear ombinationoftheabovethreeequations(namely,(10)−(11) (12))allows V(Y) H(x¯Y) us to eliminate the terms and . We are left with: xH(x)+x¯V(x¯)=x2+(x¯2+x+x2)Y +(x¯3 x¯)Y2 x¯2Y3. − − xH(x) x x¯V(x¯) Sin e ontains only positive powers of and ontains only negativepowers of x H(x) V(y) , we have hara terizedthe series and . H(x) V(y) Q 2 Proposition 6. The series and ounting -va illating walks of even length (1,0) x y starting at and ending on the -axis and on the -axis satisfy xH(x)=PT x2+(x¯2+x+x2)Y +(x¯3 x¯)Y2 x¯2Y3 , x − − x¯V(x¯)=NT(cid:0)x2+(x¯2+x+x2)Y +(x¯3 x¯)Y2 x¯2Y3(cid:1), x − − (cid:0) (cid:1) PT NT x x x where the operator (resp. ) extra ts positive (resp. negative) powers of in series Q[x,x¯][[t]] of . F (x,y;t) F (x,y;t) e o One may then go ba k to (5(cid:21)6) to obtain expressionsof the series and . C (n) 3 3 However, our main on ern in this note is the number of -non rossing partitions of [n] C (n) 3 . Going ba k to (4), we (cid:28)nd that is determined by the following three equations: C (n)=q ((1,0),(1,0),2n) q ((1,0),(0,1),2n), 3 2 2 − q ((1,0),(1,0),2n)=[xtn] H(x)=[x2tn] xH(x), 2 q ((1,0),(0,1),2n)=[ytn] V(y)=[x¯2tn] x¯V(x¯). 2 (t) 3 Using Proposition 6, we obtain the generating fun tion C of -non rossingpartitions: (t)=CT(x¯2 x2) x2+(x¯2+x+x2)Y +(x¯3 x¯)Y2 x¯2Y3 , C x − − − (13) (cid:0) (cid:1) CT x Q[x,x¯][[t]] x wheretheoperator extra tsthe onstanttermin ofseriesin . Observethat Y(x)=xY(x¯) k N ℓ Z . This implies that for all ∈ and ∈ , [xℓ]Y(x)k =[xℓ−k]Y(x¯)k =[xk−ℓ]Y(x)k, CT(x−ℓYk)=CT(xℓ−kYk). that is, x x This allows us to rewrite (13) with (cid:16)only(cid:17) six terms: (t)=1+CT (x¯1 x4)Y +(x¯5 x¯)Y2 (x¯4 x0)Y3 . C x − − − − (cid:0) (cid:1) (t) The above equation says that C is the onstant term of an algebrai fun tion. By a very (t) general theory [16℄, C is D-(cid:28)nite. That is, it satis(cid:28)es a linear di(cid:27)erential equation with (t) polynomial oe(cid:30) ients. In the next se tion, we showthat C satis(cid:28)es the equation(2) (or, equivalently, the P-re urren e (1)). Note that this re urren e an be easily guessed using C (n) 3 the Maple pa kage Gfun: indeed, the (cid:28)rst 15 values of already yield the orre t re ursion. 8 MIREILLEBOUSQUET-MÉLOUANDGUOCEXIN 2.5. The Lagrange inversion formula and reative teles oping From now on, several routes lead to the re urren e relation of Proposition 1, depending on how mu h software one is willing to use. We present here the one that we believe to be the shortest. Starting from (9), the Lagrangeinversion formula gives k n n 2j+k [tn]CT(xℓYk)= a (ℓ,k,j) a (ℓ,k,j)= . n n x with n j j+k j ℓ (14) jX∈Z (cid:18) (cid:19)(cid:18) (cid:19)(cid:18) − (cid:19) a 0 b a n 1 By onvention, the binomial oe(cid:30) ient b is zero unless ≤ ≤ . Hen e for ≥ , C (n)= a ( 1,1,j) a (4,1,j)+a(cid:0)((cid:1)5,2,j 1) a ( 1,2,j 1) a ( 4,3,j 1)+a (0,3,j 1) . 3 n n n n n n − − − − − − − − − − − Z jX∈ (cid:0) (cid:1) (15) j 1 j Of ourse, we ould repla e all o urren es of − by in the above expression, but this results in a bigger (cid:28)nal formula. n 1 3 [n] Proposition 7. For ≥ , the number of -non rossing partitions of is n 4(n 1)!n!(2j)! C (n)= − P(j,n) 3 (j 1)!j!(j+1)!(j+4)!(n j)!(n j+2)! j=1 − − − X with P(j,n)=24+18n+(5 13n)j+(11n+20)j2+(10n 2)j3+(4n 11)j4 6j5. − − − − Proof of the re urren e relation of Proposition 1. We (cid:28)nally apply to the above expression Zeilberger's algorithm for reative teles oping [19, Ch. 6℄ (we used the Maple C (n) n 2 3 pa kage Ekhad). This gives a re urren e relation for the sequen e : for ≥ , 9n(n 2)(n 3) 4n2+15n+17 C (n 3) (n 2) 76n4+373n3+572n2+203n 144 C (n 2) 3 3 − − − − − − − +(n+3) 44n4+(cid:0)189n3+227n2(cid:1)+30n 160 C (n(cid:0) 1)=(n+5)(n+4)(n+3) 4n2+7n+(cid:1) 6 C (n). 3 3 − − C (0)=C (1)=1 (cid:0) 3 3 (cid:1) (cid:0) (cid:1) The initial onditions are . It is then very simple to he k that the three term re ursion given in Proposition 1 satis(cid:28)es also the above four term re ursion. This an be rephrased by saying that (1) is a right fa tor of the four term re ursion obtained via reative teles oping. More pre isely, applying to (1) the operator (n+1) 4n2+39n+98 (n+6) 4n2+31n+63 N, − N (cid:0) n(cid:1) n+1 (cid:0) (cid:1) n where is the shift operator repla ing by , gives the four term re ursion (with n+3 repla ed by ). w ((1,0),(1,0),2n) : n N 2 It is worth noting that though the sequen e { ∈ } satis(cid:28)es a 2 q ((1,0),(1,0),2n):n N 2 simplere urren eoforder ,Mapletellsusthesequen es{ ∈ }and q ((1,0),(0,1),2n):n N 3 2 { ∈ } satisfy more ompli ated re urren es of order . 2.6. Asymptoti s C (n) 3 Finally, wewill derivefrom the expression(15)of the asymptoti behaviourofthis k ℓ k 1 sequen e,asstatedinProposition1. Forany(cid:28)xedvaluesof and ,with ≥ ,we onsider a (ℓ,k,j) a (j) n n the numbers de(cid:28)ned by (14). For the sake of simpli ity, we denote them , and introdu e the numbers n n 2j+k b (j)= , n j j+k j ℓ (cid:18) (cid:19)(cid:18) (cid:19)(cid:18) − (cid:19) a (j) = kb (j)/n. B = b (j) j [ℓ ,n k] so that n n Let n j n . The sum runs over ∈ + − , where ℓ = max(ℓ,0) j + . Below, we often onPsider as a real variable running in that interval. ON PARTITIONS AVOIDING 3-CROSSINGS 9 b (j) n The number is then de(cid:28)ned in terms of the Gamma fun tion rather than in terms of B N n fa torials. We will show that admits, for any , an expansion of the form N c B =9n i +O(n−N−1) , n ni ! (16) i=1 X c k ℓ i where the oe(cid:30) ients depend on and , and explain how to obtain these oe(cid:30) ients. We follow the standards steps for estimating sums of positive terms that are des ribed, for b (j) n instan e, in [2, Se tion 3℄. We begin with a unimodality property of the numbers . n b (j) j [ℓ ,n k] n + Lemma 8 (Unimodality). For large enough, the sequen e , for ∈ − , is [2n/3 k/2 1/2,2n/3 k/2 1/3] unimodal, and its maximum is rea hed in the interval − − − − . Proof. One has b (j) 1 (j+1)(j+k+1)(j ℓ+1)(j+k+ℓ+1) n q (j):= = − . n b (j+1) (n j)(n j k) 2j+k+1 2j+k+2 n − − − j Ea h of these three fa tors is easily seen to be an in reasing fun tion of on the relevant n q (ℓ ) < 1 q (n k 1) > 1 j n + n 0 interval. Moreover, for large enough, while − − . Let be the j q (j) 1 b (j)<b (j+1) j <j b (j) b (j+1) n n n 0 n n smallestvalueof su hthat ≥ . Then for and ≥ j j q (x) = 1 x 0 n for ≥ . We have thus proved unimodality. Solving for helps to lo ate the mode: 2n 5+6k x= +O(1/n). 3 − 12 q (2n/3 k/2 1/2)<1 q (2n/3 k/2 1/3)>1 n n n It is then easy to he k that − − and − − for large enough. The se ond step of the proof redu es the range of summation. ǫ (0,1/6) m Lemma 9 (A smaller range). Let ∈ . Then for all , B = b (j)+o(9nn−m). n n |j−2n/X3|≤n1/2+ǫ j =2n/3 n1/2+ǫ Proof. Let ± . The Stirling formula gives b (j)= 3 5/2 9n e−9n2ǫ/2(1+o(1))=o(9nn−m) n 2 (πn)3/2 (17) (cid:18) (cid:19) m for all (the details of the al ulation are given in greater detail below, in the proof of Lemma 10). Thus by Lemma 8, b (j) n b (2n/3 n1/2+ǫ)+b (2n/3+n1/2+ǫ) =o(9nn−m). n n n ≤ − |j−2n/X3|>n1/2+ǫ (cid:16) (cid:17) The result follows. b (j) n A (cid:28)rst order estimate of , generalizing (17), su(cid:30) es to obtain, upon summing over j B N = 1 n , an estimate of of the form (16) with . However, numerous an ellations o ur a (ℓ,k,j) C (n) n 3 when summing the 6terms in theexpression(15) of . Thisexplains why we b (n) B j n have to work out a longer expansion of the numbers and . b (n) ǫ (0,1/6) j =2n/3+r r=s√n j Lemma 10 (Expansion of ). Let ∈ . Write with and s nǫ N 1 | |≤ . Then for all ≥ , b (j)= n n 2j+k = 3 5/2 9n e−9r2/(2n) N−1ci(s) +O(nN(3ǫ−1/2)) n (cid:18)j(cid:19)(cid:18)j+k(cid:19)(cid:18) j−ℓ (cid:19) (cid:18)2(cid:19) (πn)3/2 i=0 ni/2 ! X c (s) k ℓ s 3i s c i i where is a polynomial in , and , of degree in . Moreover, is an even [odd℄ s i c (s)=1. s 0 fun tion of if is even [odd℄. In parti ular, This expansion is uniform in . 10 MIREILLEBOUSQUET-MÉLOUANDGUOCEXIN Proof. Note that we simply want to prove the existen e of an expansion of the aboveform. The oe(cid:30) ients anbe obtained routinely using(preferably) a omputeralgebrasystem. In (c ) c = 1 i i≥0 0 what follows, stands for a sequen e of real numbers su h that . The a tual c (c (s)) i i i≥0 valueof may hangefromoneformulatoanother. Similarly, denotesasequen e s c (s)=1 0 of polynomials in su h that , having the parity property stated in the lemma. N 1 We start from the Stirling expansion of the Gamma fun tion: for all ≥ , N−1 c Γ(n+1)=nn√2πn e−n i +O(n−N) . ni ! (18) i=0 X j =2n/3+r r =s√n N This gives, for , with , and for any , N−1 πn c (s) Γ(j+1)=2jj e−j i +O(nN(ǫ−1/2)) r 3 i=0 ni/2 ! (19) X c (s) i s s i for some polynomials of degree in . This estimate is uniform in . Similarly, N−1 2πn c (s) Γ(n j+1)=(n j)n−j ej−n i +O(nN(ǫ−1/2)) , − − r 3 i=0 ni/2 ! (20) X c (s) i s i for some polynomials of degree in . Now nn 3n 9r2 3iri+1 log = log 1 ( 1/2)i jj(n j)n−j 2j − 4n − i(i+1)ni − − − i≥2 X (cid:0) (cid:1) 3n 9r2 N−1c (s) = log i +O(n2ǫ+N(ǫ−1/2)), 2j − 4n − ni/2 i=1 X c (s) i+2 s (i+2)/(i/2) 6 i 1 i for some polynomials of degree in . Observe that ≤ for ≥ . Hen e nn = 3ne−9r2/(4n)exp N−1ci(s) +O(n2ǫ+N(ǫ−1/2)) jj(n j)n−j 2j − ni/2 ! − i=1 X = 3ne−9r2/(4n) N−1ci(s) +O(nN(3ǫ−1/2)) , 2j ni/2 ! (21) i=0 X c (s) 3i s i for polynomials of degree in . Putting together (18(cid:21)21), one obtains, uniformly in s : n = 3 3ne−9r2/(4n) N−1ci(s) +O(nN(3ǫ−1/2)) , (cid:18)j(cid:19) 2√πn 2j i=0 ni/2 ! (22) X c (s) 3i s i for polynomials of degree in . Similarly, n = 3 3n e−9r2/(4n) N−1ci(s) +O(nN(3ǫ−1/2)) , (cid:18)j+k(cid:19) 2√πn2j+k i=0 ni/2 ! (23) X c (s) i with the same degree ondition on the polynomials . Finally, N−1 2j+k 3 c (s) = 22j+k i +O(nN(ǫ−1/2)) (cid:18) j−l (cid:19) r2πn i=0 ni/2 ! (24) X c (s) i s i for polynomials of degree in . Putting together (22(cid:21)24), we obtain the estimate of b (j) n given in the lemma. b (j) j j 2n/3 n1/2+ǫ n It remains to sum our estimates of for values of su h that | − |≤ .