Notes on Schubert Polynomials
Appendix

Arun Ram
Department of Mathematics and Statistics
University of Melbourne
Parkville, VIC 3010 Australia
aram@unimelb.edu.au

Last update: 2 July 2013

A Combinatorial Construction of the Schubert Polynomials

by Nantel Bergeron

In this appendix, we shall give a combinatorial rule based on diagrams for the construction of the Schubert polynomials. A different algorithm had been conjectured (and proved in the case of vexillary permutations) by A. Kohnert. We shall give, at the end of this appendix, a sketch of how one can show the equivalence of the two rules. I wish to acknowledge my indebtedness to Mark Shimozono for the stimulating exchanges regarding this work.

Combinatorial construction

Here a "diagram" will be any finite non empty set of lattice points (i,j) in the positive quadrant (i1,j1). For example the diagram D(w) of a permutation w is a diagram in the above sense. Let D be any diagram. We denote by D(r,r+1) the diagram D restricted to the row r and r+1. Let j(r,D)=(j1,j2,,jk) be the columns of D in which there is exactly one element of D(r,r+1) per column. Choose a column jij(r,D). Assume first that (r+1,ji)D(r,r+1). If i=k or if (r,ji+1)D(r,r+1), let D1 be the diagram obtained from D by replacing the element (r+1,ji) by (r,ji). Now suppose instead that (r,ji)D(r,r+1). We say that the point (r,ji) is r-fixed with respect to D(w) if the number of elements of D in the column ji and in the rows r>r is equal to the number of elements of D(w) in the same area. Now if i=1 (and if there is no r-fixed element with respect to D(w) in D) or if (r+1,ji-1)D(r,r+1), let D1 be the diagram obtained from D by replacing the element (r,ji) by (r+1,ji). In both cases we say that the diagram D1 is obtained from D by a "B-move" (with respect to D(w)). For example let D be such that D(r,r+1) is the following:

1 2 3 4 5 6 7 8 9 r r+1

For this case j(r,D)=(2,5,8,9). We can perform on this diagram a B-move in column 2, 5 or 9 and obtain, respectively, the following diagrams:

The element in column 8 is not allowed to move since (r+1,5)D(r,r+1). Let Ω(w) denote the set of all diagrams (including D(w)) obtainable from D(w) by any sequence of B-moves.

Next for DΩ(w) let xD denote the monomial x1a1 x2a2 x3a3 where ai is the number of elements of D in the ith row. For any permutation w we shall have the following theorem:

(B.1) 𝔖w=DΩ(w) xD.

To prove this we will proceed by reverse induction on (w). If w=w0 (the longest element of Sn) then (B.1) holds since Ω(w0) contains only the element D(w0) and xD(w0)=xδ. On the other hand from (4.3), 𝔖w0=xδ. Now if ww0 then let r=min{i:w(i)<w(i+1)}. From (4.2) we have

(B.2) 𝔖w=r 𝔖wsr.

Let v=wsr. By the induction hypothesis equation (B.1) holds for 𝔖v. The induction step will be to "apply" the operator r to the diagrams in Ω(v). To this end we need more tools.

For the moment let us fix DΩ(v). Let a=ar(D) and b=ar+1(D) be respectively the number of elements of D in the rth and r+1st rows. We have

(B.3) rxD=r xraxr+1b= { k=0a-b-1 xra-r-1 xr+1b+r ifa>b, 0 ifa=b, -k=0a-b-1 xra+r xr+1b-r-1 ifa<b.

This suggests we define the operator r directly on the diagram D. For this we need only to concentrate our attention on the rows r and r+1 of D. Let j(r,D)=(j1,j2,,jp). Notice that in all columns j<w(r) of D(r,r+1) there are exactly two elements and in column w(r)=j1 of D(r,r+1) there is exactly one element in position (r,j1). We shall now reduce the sequence of indices j(r,D) according to the following rule. Let J(0)=(j2,j3,,jp). Remove from J(0) all pairs jk,jk+1 for which (r,jk)D and (r+1,jk+1)D. Let us denote the resulting sequence by J(1). Repeat recursively this process on J(1) until no such pair can be found. Let us denote by f(r,D)=(f1,f2,,fq) the final sequence. From construction, the sequence f(r,D) is such that if (r,fk)D then (r,fk+1)D. Let up(r,D) be the minimal k such that (r,fk)D. If (r+1,fq)D then set up(r,D)=q+1. We are now in a position to define the operation of r on the diagram D. To this end let us first assume that a>b. This means that we have a-b more elements. in row r then in row r+1. Hence q-up(r,D)+1a-b-1 for q the length of f(r,D). The equality holds if and only if up(r,D)=1. In the case a>b the operator r on the diagram D is defined by the map

(B.4a) rD { D0,D1,D2, ,Da-b-1 }

where D0 is identical to D except that we remove the element in position (r,w(r)) and for k=1,2,,a-b-1 we successively set Dk to be identical to Dk-1 except that the element (r,fup(r,D)+k-1) is replaced by (r+1,fup(r,D)+k-1). Now if a<b we have up(r,D)-1b-a+1 (with equality iff up(r,D)=q+1). So up(r,D)-1>b-a. In this case the operator r on the diagram D is defined by the map

(B.4b) rD { D0,D1,D2, ,Db-a-1 }

where D0 is identical to D except that we remove the element in position (r,wr) and the element (r+1,fup(r,D)-1) is replaced by (r,fup(r,D)-1). For k=1,2,,b-a-1 we successively set Dk to be identical to Dk-1 except that the element (r+1,fup(r,D)-k-1) is replaced by (r,fup(r,D)-k-1). Finally if a=b then

(B.4c) rD{}.

With this definition of r we have that

(B.5) rxD=± DirD xDi,

with the positive sign in case (B.4a), and the negative sign in case (B.4b). For (B.4c) the result of (B.5) is zero.

We shall now show that

(B.6) rmaps Ω(v)into Ω(w).

Proof.

The reader will notice that in D(v) the rectangle defined by the rows 1,2,,r+1 and the columns 1,2,,w(r)-1 is filled with elements. None of these elements can B-move. Hence these elements are fixed in any diagram DΩ(v). The same applies to all elements in column w(r); they are packed in the smallest rows and there are no elements in the rows strictly greater than r. Now let D be a diagram in Ω(v) and assume that rD={D0,D1,,Dm} is non-empty. The remark above implies that the element in position (r,w(r)) does not affect the sequence of B-moves from D(v) to D. Hence we can apply the same sequence of B-moves to D(v)-{(r,w(r))} and obtain D0. Moreover D(v)-{(r,w(r))} is obtainable from D(w) by a simple sequence of B-moves in rows r,r+1, for this one successively B-moves all the elements in row r+1 and columns given by j(r,D(w)). This gives that D0 is obtainable from D(w) by a sequence of B-moves, that is D0Ω(w). Now from the construction of rD, Dk (k>0) is obtained from Dk-1 by exactly one B-move. Hence rDΩ(w).

It is appropriate at this point to give an example. Let w=(6,3,9,5,1,2,11,8,4,7,10). Hence r=2 and v=(6,9,3,5,1,2,11,8,4,7,10). We have depicted below the diagrams D(w) and D(v). In our example the fixed elements described above are colored in grey and the element in position (r,w(r)) is colored black.

1 2 3 4 5 6 7 8 9 10 11 1 2 3 4 5 6 7 8 9 10 11 1 2 3 4 5 6 7 8 9 10 11 1 2 3 4 5 6 7 8 9 10 11 D(w)D(v)

Now let D be the following diagram of Ω(v).

D=

Here, ar(D)=7, ar+1(D)=4 and j(r,D)=(3,5,7,8,10). The reduced sequence f(r,D) is (8,10) and up(r,D)=1. Hence rD={D0,D1,D2} where

D0D1D2

To prove (B.1) the first step is to find a subset of Ω(v) such that when we operate with r we obtain Ω(w). To this end let

Ω0(v)= { DΩ(v): ar(D)> ar+1(D) andup(r,D) =1 } .

We have

(B.7) Ω(w)= DΩ0(v) rD(disjoint union).

Proof.

It is clear from construction that the subsets rD are disjoint when DΩ0(v). From (B.6) we only have to prove that for any DΩ(w) there is a DΩ0(v) such that DrD. To see that, reduce the sequence j(r,D)=(j1,,jp) by removing recursively all pairs jk,jk+1 for which (r,jk)D and (r+1,jk+1)D. Denote the final sequence by f(r,D). Let D be the bubble diagram obtained from D by adding an element in position (r,w(r)) and successively B-moving all elements in positions (r+1,fi)D. We have that DΩ(v). To see this one applies to D(v) the sequence of B-moves from D(w) to D-{(r,w(r))}. Of course one should ignore any B-move in rows r,r+1 performed on the original elements of D(v) in row r. But by the choice of r, the other B-moves apply almost directly and the resulting diagram is precisely D. Moreover since f(r,D)=f(r,D) and up(r,D)=1 we have DΩ0(v) and DrD.

We shall now investigate the effect of r on Ω1(v)=Ω(v)-Ω0(v). More precisely we have

(B.8) DΩ1(v) rxD=0.

Proof.

There are two classes of diagrams in Ω1(v). The first class contains the diagrams D for which ar(D)=ar+1(D). In this case it is trivial that rxD=0. The other class is formed by the diagrams D such that ar(D)ar+1(D) and up(r,D)>1. In this case we shall construct an involution, DD, such that rxD+rxD=0. Let f(r,D)=(f1,f2,,fq), a=ar(D) and b=ar+1(D). We first define the involution for the case a>b. Since up(r,D)>1 we must have q-up(r,D)+1a-b. So let D be identical to D except that the elements in positions (r,fup(r,D)), (f,fup(r,D)+1), , (r,fup(r,D)+a-b-1) are B-moved to the positions (r+1,fup(r,D)), (r+1,fup(r,D)+1), , (r,fup(r,D)+a-b-1). It is clear that DΩ(v). But f(r,D)=f(r,D) and up(r,D)>up(r,D)>1, hence DΩ1(v). Moreover we have ar(D)=b and ar+1(D)=a, hence rxD+rxD=0. The case a<b is similar to the previous one.

A proof of (B.1) is now completed combining (B.2), (B.5), (B.7) and (B.8). More precisely using the induction hypothesis, we have

𝔖w = r𝔖v (B.2) = DΩ(v) rxD = DΩ0(v) rxD (B.8) = DΩ0(v) DirD xDi (B.5) = DΩ(w) xD. (B.7)

Kohnert's construction

Let D be any diagram. Choose (i,j)D such that (i,j)D for all j<j. Let us suppose that there is a point (i,j)D with i<i. Then let h<i be the largest integer such that (h,j)D and let D1 denote the diagram obtained from D by replacing (i,j) by (h,j). We say that D1 is obtained from D by a "K-move". Now let K(D(w)) denote the set of all diagrams (including D itself) obtainable from D by any sequence of K-moves. Kohnert's conjecture states that for any permutation w we have

(B.9) 𝔖w=DK(D(w)) xD.

A. Kohnert has proved (B.9) for the case where w is a vexillary permutation but the general case was still open. For the interested reader here is a sketch of how one may prove (B.9).

We have noticed by computer that Ω(w)=K(D(w)). The idea then is to show both inclusions by induction. The inclusion K(D(w))Ω(w) is the easiest one. We only have to show that any K-move of an element (i,j) to (h,j) can be simulated using B-moves. For this we proceed by induction on i-h. If i-h=1 then the K-move is simply one B-move. Now if i-h>1, we first perform the sequence of B-moves in row h,h+1 necessary to B-move the element (h+1,j) to (h,j). Then using the induction hypothesis we can K-move (i,j) to (h+1,j). Finally we reverse the first sequence of B-moves in rows h,h+1. That shows K(D(w))Ω(w).

The other inclusion needs a lot more work. For DK(D(w)) and i any row of D let Bi(D) denote the set of all diagrams (including D) obtainable from D by any sequence of B-moves in the rows i,i+1 only. It is clear that if i is big enough then Bi(D)K(D(w)). We may then proceed by reverse induction on i. Now for a fixed i, notice that Bi(D(w)) is obtainable from D(w) using only K-moves. Let Ω0 denote the set of all diagrams obtainable from Bi(D(w)) by any sequence of K-moves for which no elements crosses the border between the rows i+1 and i+2. A simple inductive algorithm may be used here to show that for any DΩ0 we have Bi(D)Ω0. Next let Ωk denote the set of all diagrams of K(D(w)) which have k more elements than D(w) in the rows 1,2,,i+1. For almost all the cases it is fairly easy to show (using induction on k and the induction hypothesis on i) that for DΩk we have Bi(D)Ωk. But some of the cases are really hard to formalize! Now this completed would show that Ω(w)K(D(w)) since K(D(w))=Ωk.

Notes and References

This is a typed excerpt of the book Notes on Schubert Polynomials by I. G. Macdonald.

page history