A Demazure crystal construction for Schubert polynomials

Stanley symmetric functions are the stable limits of Schubert polynomials. In this paper, we show that, conversely, Schubert polynomials are Demazure truncations of Stanley symmetric functions. This parallels the relationship between Schur functions and Demazure characters for the general linear group. We establish this connection by imposing a Demazure crystal structure on key tableaux, recently introduced by the first author in connection with Demazure characters and Schubert polynomials, and linking this to the type A crystal structure on reduced word factorizations, recently introduced by Morse and the second author in connection with Stanley symmetric functions.


Introduction
Schubert polynomials S w were first introduced by Bernstein et al. [BGG73] as certain polynomial representatives of cohomology classes of Schubert cycles X w in flag varieties. They were extensively studied by Lascoux and Schützenberger [LS82] using an explicit definition in terms of difference operators ∂ w . Subsequently, a combinatorial expression for Schubert polynomials as the generating polynomial for compatible sequences for reduced expressions of a permutation w was discovered by Billey, Jockusch, and Stanley [BJS93]. In the special case of the Grassmannian subvariety, Schubert polynomials are Schur polynomials, which also arise as the irreducible characters for the general linear group.
The Stanley symmetric functions Edelman and Greene [EG87] showed that the coefficients of the Schur expansion of Stanley symmetric functions are nonnegative integer coefficients.
Demazure modules for the general linear group [Dem74] are closely related to Schubert classes for the cohomology of the flag manifold. In certain cases these modules are irreducible polynomial representations, and so the Demazure characters also contain the Schur polynomials as a special case. Lascoux and Schützenberger [LS85] stated that Schubert polynomials are nonnegative sums of Demazure characters. This was proven by Reiner and Shimozono [RS95a] using the right keys associated to Edelman-Greene insertion. Using a key tableaux interpretation for Demazure characters [Assb], Assaf [Assa] showed that the Edelman and Greene algorithm giving the Schur expansion of a Stanley symmetric function can be modified to a weak Edelman-Greene algorithm which gives the Demazure expansion of a Schubert polynomial.
In this paper, we deepen this connection and provide a converse to (1.1) by showing that Schubert polynomials are Demazure truncations of Stanley symmetric functions. Specifically, we show in Theorem 5.11 that the combinatorial objects underlying the Schubert polynomials, namely the compatible sequences, exhibit a Demazure crystal truncation of the full Stanley crystal of Morse and Schilling [MS16]. We prove this using Theorem 3.14, in which we give an explicit Demazure crystal structure on semi-standard key tableaux, which coincide with semi-skyline augmented fillings of Mason [Mas09]. This, together with Theorem 5.10, in which we show that the crystal operators on reduced factorizations intertwine with (weak) Edelman-Greene insertion, proves our main result.
Lenart [Len04] defined crystal operators on RC graphs [BB93], which are closely related to compatible sequences, though it was not observed there that this structure is a Demazure crystal. Earlier, Reiner and Shimozono [RS95b] defined r-pairings on factorized row-frank words that can now be interpreted as crystal

Crystal structure on tableaux
We begin in Section 2.1 by reviewing the basics of Schur polynomials via the combinatorics of Young tableaux. In Section 2.2, we review the type A crystal structure on semi-standard Young tableaux, and conclude in Section 2.3 with the definition of Demazure crystals.
2.1. Combinatorics of Schur polynomials. Given a partition λ, the Young diagram of shape λ is the array of left-justified cells with λ i boxes in row i. Here we use French notation, where the rows weakly decrease in size from bottom to top in the Young diagram. A Young tableau is a filling of the cells of a Young diagram from some totally ordered alphabet (for example the set of positive integers) such that rows and columns weakly increase. A semi-standard Young tableau is a Young tableau with distinct column entries. Figure 1 provides an example of semi-standard Young tableaux of a fixed shape. The weight of a semi-standard Young tableau T , denoted by wt(T ), is the weak composition whose ith part is the number of occurrences of i in T . The shape λ of T is also denoted by sh(T ).
Definition 2.1. The Schur polynomial in n variables indexed by the partition λ is where SSYT n (λ) is the set of semi-standard Young tableaux of shape λ over the alphabet {1, 2, . . . , n}.
Schur polynomials arise as characters for irreducible highest weight modules for the general linear group with semi-standard Young tableaux giving a natural indexing set for the basis of the module.

2.2.
Crystal operators on semi-standard Young tableaux. A crystal graph is a directed, colored graph with vertex set given by the crystal basis and directed edges given by deformations of the Chevalley generators. For the quantum group U q (sl n ), the crystal basis can be indexed by semi-standard Young tableaux over the alphabet A = {1, 2, . . . , n} and there is an explicit combinatorial construction of the crystal graph on tableaux [KN94,Lit95]. For an introduction to crystals from the quantum group perspective, see [HK02]. For a purely combinatorial introduction to crystals, see [BS17].
For a word w of length k with letters from the alphabet A = {1, 2, . . . , n}, an integer 0 r k, and an integer 1 i < n, define where wt(w) is the weak composition whose jth part is the number of j's in w.
Observe that if M i (w) > 0 and p is the leftmost occurrence of this maximum, then w p = i, and if q is the rightmost occurrence of this maximum, then either q = k or w q+1 = i + 1. For a Young tableau T , the column reading word of T , denoted by w(T ), is the word obtained by reading the entries of T down columns from left to right. For example, the column reading word of the leftmost Young tableau in the top row of Figure 1 is 32131.
Definition 2.2. Given an integer 1 i < n, define the lowering operator f i on semi-standard Young tableaux over the alphabet A as follows: if M i (w(T )) 0, then f i (T ) = 0; otherwise, let p be the smallest index such that M i (w(T ), p) = M i (w(T )), and f i (T ) changes the entry in T corresponding to w(T ) p to i + 1.
An example of the lowering operator f 2 is given in Figure 2. For this example, the column reading word is given below each semi-standard Young tableau with the largest index that attains M 2 (w(T )) > 0 underlined and the corresponding entry in the tableau circled. Definition 2.3. Given an integer 1 i < n, define the raising operator e i on semi-standard Young tableaux over the alphabet A as follows: let q be the largest index such that M i (w(T ), q) = M i (w(T )). If q is the length of w(T ), then e i (T ) = 0; otherwise, e i (T ) changes the entry in T corresponding to w(T ) q+1 to i.
For further examples of raising and lowering operators on semi-standard Young tableaux, see Figure 3. Note that we have drawn the crystal in Figure 3 with lowering operators pointing upward to facilitate the bijection with semi-standard key tableaux as explained in Section 3.
For a partition λ, we may define the highest weight crystal (of type A n ) of highest weight λ, denoted B(λ), as the set SSYT n (λ) together with the operators f i , e i for 1 i < n and the weight function wt. The character of a crystal is defined as which in this case is precisely the Schur polynomial s λ (x 1 , . . . , x n ).
2.3. Demazure crystals. Demazure characters first arose in connection with Schubert classes for the cohomology of the flag manifold in [Dem74].
The divided difference operators ∂ i for 1 i < n act on polynomials by For w ∈ S n , we may define Here s i (1 i < n) is the simple transposition interchanging i and i + 1 and k is the number of inversions (or length) of w. When k is the length of w, the expression s i1 s i2 · · · s i k for w is called a reduced expression. It can be shown that ∂ w is independent of the choice of reduced expression.
There exist degree-preserving divided difference operators π i for 1 i < n, which act on polynomials by As with ∂ i , we extend this definition to w ∈ S n , by π w = π i1 π i2 · · · π i k if w = s i1 s i2 · · · s i k is a reduced expression, and π w is independent of the choice of reduced expression.
Definition 2.4. Given a weak composition a of length n, the Demazure character κ a is defined as where λ is the partition rearrangement of a and w is the shortest permutation that sorts a to λ.
Making this more precise, Demazure crystals are certain subsets of B(λ), which were first conjectured by Littelmann [Lit95] to generalize the Demazure characters. This conjecture was later proven by Kashiwara [Kas93]. Given a subset X ⊆ B(λ), we define D i for 1 i < n as where u λ is the highest weight element in B(λ) satisfying e i (u λ ) = 0 for all 1 i < n.
Let us define the character of a Demazure crystal as It was proven by [Lit95,Kas93] that this character coincides with κ a , where w · a = λ.

Demazure crystal structure on key tableaux
In Section 3.1, we review the combinatorial model of key tableaux [Assb] that is central to our results. In Section 3.2, we introduce a new crystal structure on semi-standard key tableaux and show that this precisely realizes the Demazure character by truncating the crystal structure on semi-standard Young tableaux.
3.1. Combinatorics of Demazure characters. Combinatorial interpretations and definitions for Demazure characters for the general linear group were given by Lascoux and Schützenberger [LS90], Kohnert [Koh91], Reiner and Shimozono [RS95a], and Mason [Mas09], all of whom refer to them as key polynomials. We use an equivalent definition in terms of semi-standard key tableaux due to Assaf [Assb], which is combinatorially equivalent to Mason's semi-skyline augmented fillings but which replaces the triple conditions for more direct row and column conditions (see also [Mon]). Generalizing Young diagrams, given a weak composition a, the key diagram of shape a is the array of left-justified cells with a i boxes in row i.

Definition 3.1 ([Assb]).
A key tableau is a filling of a key diagram with positive integers such that columns have distinct entries, rows weakly decrease, and, if some entry i is above and in the same column as an entry k with i < k, then there is an entry immediately right of k, say j, with i < j.
For the Schur polynomial case, we restrict entries in the Young tableaux globally allowing entries 1 through n to appear anywhere. In the Demazure case, we must restrict the entries in the key tableaux locally allowing entries to appear only in their row and lower. For examples, see Figure 4. The following property of semi-standard key tableaux will be useful. Lemma 3.3. Suppose row r of a semi-standard key tableau has two entries i + 1 in columns c and c + 1. If there is an i above row r in column c, then there cannot be an i below row r in column c + 1.
Proof. If this were the case, then there must be an entry, say k, in column c immediately left of the i in column c+1. By the weakly decreasing rows condition, k i, and so by the distinct column entries condition, and k > i + 1. However, since there is an i + 1 above k, the entry immediately right of k, which is an i, is not larger than i + 1, a contradiction to the key tableaux column inversion condition.
The weight of a semi-standard key tableau T , denoted by wt(T ), is the weak composition whose ith part is the number of occurrences of i in T . The following result is proved in [Assar] by showing that the semistandard key tableaux conditions are equivalent to the triple conditions on Mason's semi-skyline augmented fillings [Mas09]. This more direct characterization facilitates the constructions to follow.

Theorem 3.4 ([Assar]).
The key polynomial κ a (x) is given by where SSKT(a) is the set of semi-standard key tableaux of shape a.
The map from standard key tableaux of shape a to standard Young tableaux of shape λ, where λ is the unique partition rearrangement of a, from [Assb] relates the tableaux models for key polynomials and Schur polynomials. We extend this map to the semi-standard case as follows.
Definition 3.5. Given a weak composition a of length n, define the column sorting map on SSKT(a) by letting cells fall vertically until there are no gaps between rows, sorting the columns to decrease bottom to top, and then replacing all entries by i → n − i + 1.
For example, the semi-standard key tableaux in Figure 4 map to the semi-standard Young tableaux in the first two rows of Figure 1, respectively. The four semi-standard Young tableaux in the bottom row of Figure 1 are not in the image of the column sorting map.
Proposition 3.6. The column sorting map is a well-defined, injective map φ : SSKT(a) → SSYT(λ), where λ is the partition rearrangement of a.
Proof. The column strict condition on semi-standard key tableaux ensures that columns have distinct values. Therefore by construction, a column sorted tableau will have strictly increasing columns. By the column inversion condition for key tableaux, if row j sits above row i and is weakly longer, then column by column the entries in row j must be greater than those in row i. Consider applying the column sorting map by first rearranging rows of longest size at the bottom and reversing the relative order of rows of equal length. Since entries within rows are maintained, the weakly decreasing row condition on semi-standard key tableaux is obviously maintained by this process. The column sorting necessarily brings entries from a strictly shorter row down into a longer row. That is, row values can be increased only when the first k values all increase for some k, and entries decrease only when the entire row is changed, maintaining the weakly decreasing row condition. Hence the image of the map is indeed a semi-standard Young tableau of shape λ.
To see that the map is injective, we can define an inverse map by first applying i → n − i + 1 to all letters in a semi-standard Young tableau. Then fill the shape of a column by column from right to left, and within a column from bottom to top, according to the columns of the given tableau by selecting at each step the smallest available entry that maintains the weakly decreasing row condition. To see that the column inversion condition for key tableaux still holds, suppose j is the smallest label available that can be placed in cell C in order to satisfy weakly decreasing rows. It is easy to see that the column inversion condition for key tableaux is maintained, but it could happen that an entry is placed in a row with smaller index. The tableaux for which this occurs are precisely the ones not in the image of the column sorting map.
3.2. Crystal operators on semi-standard key tableaux. We generalize the crystal structure on semistandard Young tableaux to a Demazure crystal structure on semi-standard key tableaux as follows.
For a word w of length k with letters in the alphabet A = {1, 2, . . . , n}, an integer 1 r k, and an integer 1 i < n, define Observe that if m i (w) > 0 and q is the rightmost occurrence of this maximum, then w q = i + 1, and if p is the leftmost occurrence of this maximum, then either p = 1 or w p−1 = i. For T a key tableau, the column reading word of T , denoted by w(T ), is the word obtained by reading the entries of T down columns from right to left. Note that columns for key tableaux are read in the reverse order as columns for Young tableaux. For example, the column reading word of the leftmost key tableau in the top row of Figure 4 is 42432.
Definition 3.7. Given an integer 1 i < n, define the raising operators e i on semi-standard key tableaux of shape a of length n as follows: if m i (w(T )) 0, then e i (T ) = 0; otherwise, let q be the largest index such that m i (w(T ), q) = m i (w(T )), and e i (T ) changes all entries i+1 weakly right of the entry in T corresponding to w(T ) q to i and change all i's in the same columns as these entries to i + 1's.
For an example of the raising operator e 1 , see Figure 5. For this example, the column reading word is given below each key tableau with the largest index that attains m 1 (w(T )) > 0 underlined and the corresponding entry in the tableau circled.
Proof. Let T ∈ SSKT(a), set m = m i (w(T )), and suppose m > 0. Let x, say in row r and column c, be the cell in T that attains m at the rightmost position in column reading order. We claim that every cell weakly right of x in row r with entry i + 1 except for one has an i above it. If the entry immediately right of x is h for some h < i + 1, then the key tableaux conditions ensure that there cannot be an i above x since h i. Suppose, then, that there is an i + 1 immediately right of x. Since x attains the maximum m and there is an i + 1 to its left, we must have an i between them in column reading order. Thus there must be an i either below row r in column c + 1 or above row r in column c. If there is an i in row r ′ < r in column c + 1, then there must be an entry, say k, in row r ′ in column c satisfying k i. Moreover, by the key tableau column inversion condition, we cannot have k > i + 1 since i + 1 > i. Therefore k = i, in which case x cannot be the rightmost position to attain m, a contradiction. Moreover, it now follows by induction from Lemma 3.3 that every i + 1 right of x in row r either has an i above it or no i in its column, and the latter cannot be the case more than once else the rightmost i + 1 would have i-index greater than m. This proves the claim, from which it follows that one more cell changes entry from i + 1 to i than the reverse, thus proving wt(e i (T )) i = wt(T ) i + 1, wt(e i (T )) i+1 = wt(T ) i+1 − 1, and wt(e i (T )) j = wt(T ) j for all j = i, i + 1.
Next we show that rows of e i (T ) are weakly decreasing. This is clear for row r since all i + 1 weakly right of x change to i. If i changes to i + 1 in cell y and the cell immediately left of y also contains an i, then this i also is changed to an i + 1. This is clear from the previous analysis provided y is not in the column of x; if y is in the column of x and has an i immediately to its left, then x cannot be the rightmost cell in column reading order to attain m.
Next we show that columns of e i (T ) have distinct entries. Since x cannot have an i below it and be the leftmost cell in column reading order to attain m, any i + 1 that changes to an i either has no i in the column or an i above it. In the latter case, this i will become an i + 1.
Next we show that e i (T ) if a < c with a above c, then there is an entry b immediately right of c with a < b. If a column contains i and not i + 1, then nothing is changed, and if it has both, then the i + 1 appears above i in e i (T ). Therefore the only potential problem occurs when b = i + 1 in T is changed to i in e i (T ) and a = i. In this case, if the column of a has no i + 1, then b does not attain m and is not changed to i, and otherwise both a and c change removing the inversion triple from consideration.
Finally, decrementing values maintains the property that entries do not exceed their row index, and i changes to i + 1 only when it sits above an i + 1, so these entries lie strictly above row i + 1. Therefore e i (T ) is a semi-standard key tableau.  Proof. Given a word w = w 1 w 2 · · · w k with 1 w j n for all j, let u = (n − w k + 1)(n − w k−1 + 1) · · · (n − w 1 + 1). Then m i (w, r) = M n−i (u, k − r + 1), and q is the index of the rightmost occurrence of m i (w) in w if and only if k − q + 1 is the index of the leftmost occurrence of M n−i (u) in u. If T ∈ SSKT(a) has no column inversions, then since the column reading word of a semi-standard key tableau is right to left and the column reading word of a semi-standard Young tableau is left to right, w(T ) and w(φ(T )) precisely have the relationship of w and u, and the result follows.
In the general case, since e i and f i depend only on the letters i, i + 1, we may restrict our attention to the subword on those letters. In doing so, notice that columns with i above i + 1 appear in consecutive runs separated at least by a column immediately right of the run with an i + 1 and no i. In the column reading word, this manifests itself as a string of alternating i's and i + 1's that begins and ends with an i + 1. If we let q ′ denote the leftmost i + 1 in the alternating string that attains m i (w(T )), then k − q ′ + 1 is the smallest index that attains M n−i (w(φ(T ))). That is, the rightmost column of T in which an i + 1 changes to an i without an i also changing to an i + 1 in passing to e i (T ) is precisely the column of φ(T ) in which an n − i changes to an n − i + 1 in passing to f n−i (φ(T )).
For example, the semi-standard key tableaux of shape (0, 5, 3) in Figure 5 map by the column sorting map to the semi-standard Young tableaux of shape (5, 3) in Figure 2, and the raising operator e 1 on the former becomes the lowering operator f 2 on the latter.  Proof. Recall from the analysis in the proof of Proposition 3.8 that when e i (S) = 0, w(S) and w(e i (S)) differ on the restriction to letters i, i + 1 precisely in that an alternating string beginning and ending with i + 1 for which the last entry is the rightmost to attain m i (w(S)) becomes an alternating string beginning and ending with i for which the first entry is immediately left of the leftmost to attain m i (w(e i (S))). Therefore if e i (S) = T , then f i (T ) = S. We have f i (T ) = 0 precisely when there is no place to act (when p = 1) or when acting would violate the semi-standard key tableaux condition that entries cannot exceed their row index. The remainder of the result follows from Proposition 3.8 and Lemma 3.9.
3.3. Demazure crystal on semi-standard key tableaux. To arrive at our main result, that the raising and lowering operators on semi-standard key tableaux give a Demazure crystal, we refine the column sorting map to an injective map between semi-standard key tableaux for different weak compositions.
Lemma 3.12. Given a weak composition a and an index i such that a i < a i+1 , for T ∈ SSKT(a) such that e i (T ) = 0, there exists S ∈ SSKT(s i a) such that φ(T ) = φ(S), where φ is the column sorting map.
Proof. The statement is equivalent to the assertion that there exists S ∈ SSKT(s i a) with the same column sets as T . We may describe the map from T to S explicitly as follows. First, move the a i+1 − a i rightmost cells in row i + 1 down to row i. Since e i (T ) = 0, there cannot be a letter i + 1 that is moved down since if any of these cells contain an i + 1, there will be a positive index allowing e i to act non-trivially. If, after this, row i is not weakly decreasing, then swap the entries in rows i and i + 1 of the offending column. Since e i (T ) = 0, there cannot be any letters i + 1 that are moved down at this step either, so the resulting tableau S has no entry exceeding its row index. Rows clearly maintain their weakly decreasing status, and it is easy to see that no violations of the column inversion condition can arise. Therefore S ∈ SSKT(s i a).
Lemma 3.12 ensures that the following operators are well-defined on semi-standard key tableaux.
Definition 3.13. Given a weak composition a and an index i such that a i < a i+1 , define an operator E i on SSKT(a) by E i (T ) = S, where S ∈ SSKT(s i a) satisfies φ(S) = φ(e k−1 i (T )) for k minimal such that e k i (T ) = 0. For examples of E i , see Figure 7. Similar to π w and ∂ w , we may extend this to define E w = E i1 · · · E i k , where s i1 · · · s i k is any reduced expression for w. It is easy to see that this is well-defined from the local relations of the type A crystal operators on tableaux as characterized by Stembridge [Ste03].  Figure 7. An example of the E i operators on semi-standard key tableaux.
Given a weak composition a, for w the permutation that sorts a to partition shape λ, the operator E w takes T ∈ SSKT(a) to the highest weight element of the crystal along edges specified by w. This is precisely the statement needed to show that the crystal operators defined on semi-standard key tableaux of shape a realize the Demazure crystal for w.
Theorem 3.14. Let a be a weak composition that sorts to the partition λ. The raising and lowering operators on SSKT(a) give the Demazure crystal for highest weight λ truncating with respect to the minimal length permutation w that sorts a to λ.
Proof. Given T ∈ SSKT(a), for w the permutation that sorts a to partition shape λ, we necessarily have E w (T ) ∈ SSKT(λ). However, the constraint that entries cannot exceed their row index together with distinct column values forces SSKT(λ) to have a single element, the tableau with all entries in row i equal to i. In particular, this element maps via the column sorting map to the highest weight u λ . By Lemma 3.9, this means T ∈ D w {u λ } for every T ∈ SSKT(a), and so φ(SSKT(a)) ⊆ B w (λ). By Theorem 3.4, the sums of the weights on both sides agree, so we must have equality.
Therefore we compute F 143625 using reduced factorizations for 143625 −1 = 153264. Note that reduced factorizations can, in principle, have an arbitrary number of blocks and hence F w (x) is a symmetric function in infinitely many variables x = (x 1 , x 2 , . . .).
We can restrict Stanley symmetric functions to Stanley symmetric polynomials by restricting the number of blocks in the reduced factorizations. Let RF ℓ (w) be the set of reduced factorizations of w with precisely ℓ blocks. Then the Stanley symmetric polynomial in ℓ variables is x wt(r) .

4.2.
Edelman-Greene correspondence. In their study of Stanley symmetric functions, Edelman and Greene [EG87] developed the following insertion algorithm that they used to give a formula for the Schur expansion of Stanley symmetric functions.
Definition 4.4. [EG87, Definition 6.21] Let P be a Young tableau, and let x be a positive integer. Let P i be the ith lowest row of P . Define the Edelman-Greene insertion of x into P , denoted by P ← x, as follows. Set x 0 = x and for i 0, insert x i into P i+1 as follows. If x i z for all z ∈ P i+1 , place x i at the end of P i+1 and stop. Otherwise, let x i+1 denote the smallest element of P i+1 such that x i+1 > x i . If x i+1 = x i + 1 or x i is not already in P i+1 , replace x i+1 by x i in P i+1 and continue (we say that x i bumps x i+1 in row i + 1). Otherwise leave P i+1 unchanged and continue with x i+1 . Given a reduced expression ρ, define the insertion tableau for ρ, denoted by P (ρ), to be the result of inserting the word for ρ letter by letter into the empty tableau. To track the growth of P (ρ), define the recording tableau for ρ, denoted by Q(ρ), to be the result of adding i into the new cell created when inserting the ith letter. For example, Figure 10  Theorem 4.5. [EG87, Theorem 6.25] The Edelman-Greene correspondence ρ → (P (ρ), Q(ρ)) is a bijection between reduced expressions and all pairs of tableaux (P, Q) such that P and Q have the same shape, P is increasing with row(P ) a reduced word, and Q is standard.
We may extend the Edelman-Greene correspondence to a bijection between reduced factorizations and all pairs of tableaux (P, Q) such that P and Q have the same shape, P is increasing with row(P ) a reduced word, and Q is semi-standard. To do so, given a reduced factorization r into ℓ blocks, define P (r) to be P (ρ) where ρ is the underlying reduced expression for r, and define Q(r) to be the result of adding ℓ − i + 1 into each new cell created when inserting a letter from block i (from the right). For example, the recording tableau for the reduced factorization (4)(5)(23)(2) is constructed in Figure 11. Corollary 4.6. The correspondence r → (P (r), Q(r)) is a bijection between reduced factorizations and all pairs of tableaux (P, Q) such that P and Q have the same shape, P is increasing with row(P ) a reduced word, and Q is semi-standard. Moreover, if r has ℓ blocks, then wt(Q(r)) i = wt(r) ℓ−i .
For example, the Edelman-Greene correspondence gives a weight-reversing bijection In particular, by the symmetry of Schur functions, we have the following expansion from [EG87].
Corollary 4.7. The Stanley symmetric function for w may be expressed as where Yam(w −1 ) is the set of insertion tableaux with row(P ) a reduced word for w −1 .
For example, we have

4.3.
Crystal operators on reduced factorizations. Following [MS16], we are going to define an A ℓ−1crystal structure on RF ℓ (w). Let r = r ℓ r ℓ−1 · · · r 1 ∈ RF ℓ (w), where r i is the ith block from the right. The Kashiwara raising and lowering operators e i and f i only act on the blocks r i+1 r i . The action is defined by first bracketing certain letters and then moving an unbracketed letter from one factor to the other. Let us begin by describing the bracketing procedure. Start with the largest letter b in r i and pair it with the smallest a > b in r i+1 . If no such a exists in r i+1 , then b is unpaired. The pairing proceeds in decreasing order on elements of r i , and with each iteration previously paired letters of r i+1 are ignored. Define Then f i (r ℓ · · · r 1 ) is defined by replacing the blocks r i+1 r i by r i+1 r i such that Similarly, e i (r ℓ · · · r 1 ) is defined by replacing the factors r i+1 r i by r i+1 r i such that r i = r i ∪ {a + s} and r i+1 = r i+1 \{a} for a = max(L i (r ℓ · · · r 1 )) and s = min{j 0 | a + j + 1 ∈ r i+1 }. If L i (r ℓ · · · r 1 ) = ∅, then e i (r ℓ · · · r 1 ) = 0.
Remark 4.9. In [MS16], the Stanley symmetric function F w is defined using decreasing factorizations of reduced words of w. Here we use increasing factorizations of w −1 . To relate the two, one needs to revert the reduced factorizations. The crystal structures are related by interchanging f i (resp. e i ) with e ℓ−i (resp. f ℓ−i ).
Theorem 4.10. [MS16, Theorem 3.5] The above defined operators f i and e i for 1 i < ℓ and the weight function wt define a A ℓ−1 -crystal structure on RF ℓ (w).
It turns out that this crystal structure on reduced factorizations relates to the crystal structure on semistandard Young tableaux via the Edelman-Greene correspondence.
Theorem 4.12. [MS16,Theorem 4.11] Given r ∈ RF ℓ (w), let P (r) denote its Edelman-Greene insertion tableau and Q(r) its Edelman-Greene semi-standard recording tableau, where letters in block i of r are recorded by the letter i. Then, if e i (r) = 0, we have P (e i (r)) = P (r) and Q(e i (r)) = f ℓ−i (Q(r)).

Demazure crystal structure for Schubert polynomials
We review the combinatorial expression of Billey, Jockusch and Stanley [BJS93] for Schubert polynomials in terms of compatible sequences in Section 5.1 and show that it can be reformulated in terms of reduced factorizations with a cutoff condition. In Section 5.2 we discuss the weak analog of the Edelman-Greene insertion presented in [Assa]. It turns out that the cut-off condition precisely amounts to a Demazure crystal structure as shown in Section 5.3. 5.1. Combinatorics of Schubert polynomials. Schubert polynomials are generalizations of Schur polynomials which represent cohomology classes of Schubert cycles in flag varieties. They were first introduced by Bernstein et al. [BGG73] and extensively studied by Lascoux and Schützenberger [LS82].
Definition 5.1 ( [LS82]). Given a permutation w, the Schubert polynomial for w is given by where w 0 = nn − 1 . . . 21 is the longest permutation of length n 2 . The first proven combinatorial formula for Schubert polynomials, due to Billey, Jockusch and Stanley [BJS93], is in terms of compatible sequences for reduced expressions.
For example, seven of the reduced words for 153264 have compatible sequences as shown in Figure 12. ). The Schubert polynomial S w (x) indexed by a permutation w is given by where x a is the monomial x a1 1 · · · x an n . We may encode compatible sequences for the reduced words as increasing factorizations with an additional cutoff condition.
Definition 5.4. Given a reduced word ρ, an increasing factorization with cutoff is an increasing factorization such that in addition the first entry in block i from the right is at least i.
Given a permutation w, a reduced factorization with cutoff for w is an increasing factorization with cutoff of a reduced word for w.
The set of reduced factorizations with cutoff is denoted by RFC(w). For example, the reduced factorizations with cutoff for 153264 are shown in Figure 13.  Figure 13. The reduced factorizations with cutoff for 153264.
The weight function on reduced factorizations provides a simple bijection between compatible sequences and increasing factorizations with cutoff for a reduced word. For example, compare Figure 13 with Figure 12.
Proposition 5.5. The Schubert polynomial S w (x) is given by x wt(r) .
Proof. To prove that (5.3) is equivalent to (5.2), we show that there is a bijection ρ∈R(w −1 ) RC(ρ) → RFC(w −1 ). Given a compatible sequence α for a reduced word ρ, the letter ρ i belongs to the a-th factor from the right if α i = a. Due to the condition that α j > α j+1 whenever ρ j > ρ j+1 , the letters within each factor are weakly increasing. Since the word ρ is reduced, the letters within each factor must actually be increasing. Furthermore, since α j ρ j , all letters in the a-th factor must be of value at least a. Conversely, given a reduced factorization with cutoff one can immediately construct the compatible sequence α by setting α j = a if ρ j is in factor a.
Reduced factorizations have the advantage of tracking the reduced word along with the weight, making this a more natural indexing set for the crystal structure discussed in the next section.

Weak Edelman-Greene correspondence.
We recall a generalization of the Edelman-Greene correspondence [Assa] that gives the Demazure expansion of a Schubert polynomial, parallel to the Schur expansion of a Stanley symmetric function.
Following [Assa], for P a semi-standard Young tableau with strictly increasing rows, define the lift of P , denoted by lift(P ), to be the tableau of key shape obtained by raising each entry in the first column of P until it equals its row index, and, once columns 1 through c − 1 have been lifted, raising entries in column c from top to bottom, maintaining their relative order, placing each entry in the highest available row such that there is an entry in column c − 1 that is strictly smaller.
Definition 5.6 ( [Assa]). For ρ a reduced expression, define the weak insertion tableau P (ρ) by P (ρ) = lift(P (ρ)), where P (ρ) is the insertion tableau under the Edelman-Greene insertion. In addition, define the weak recording tableau Q(ρ) to be the unique standard key tableau of the same key shape as P (ρ) such that φ( Q(ρ)) = Q(ρ), where Q(ρ) is the Edelman-Greene recording tableau and φ is the column sorting map. For P a key tableau, define the drop of P , denoted by drop(P ), to be the Young tableau defined by letting the entries of P fall in their columns while maintaining their relative order. It is clear that drop(lift(P )) = P for any P of partition shape.
Theorem 5.7 ( [Assa]). The weak Edelman-Greene correspondence ρ → ( P (ρ), Q(ρ)) is a bijection between reduced expressions and all pairs of tableaux (P, Q) such that P and Q have the same key shape, P has increasing rows and columns with row(P ) a reduced word and lift(drop(P )) = P , and Q is a standard key tableau.
Analogous to the Edelman-Greene correspondence, this extends to a bijection between reduced factorizations with cutoff and all pairs of tableaux (P, Q) such that P and Q have the same key shape, P is increasing with row(P ) a reduced word and lift(drop(P )) = P , and Q is a semi-standard key tableau. For example, the recording tableau for the reduced factorization (4)(5)(23)(2) is constructed in Figure 15.
Corollary 5.8. The correspondence r → ( P (r), Q(r)) is a weight-preserving bijection between reduced factorizations and all pairs of tableaux (P, Q) such that P and Q have the same key shape, P is increasing with row(P ) a reduced word and lift(drop(P )) = P , and Q is a semi-standard key tableau.
Proof. Theorem 5.7 is proved in [Assa,Theorem 5.16] using the standard key tableau. To get the semistandard case, we appeal to [Assar, Proposition 2.6] where it is shown that the fundamental slide polynomial, defined in [AS17], associated to a standard key tableau is the sum of monomials associated to the semistandard key tableaux that standardize to it. As shown in [Assa, Theorem 2.4], the fundamental slide polynomial associated to a reduced expression is the sum of monomials associated to the corresponding compatible sequences. The result follows from the bijection between compatible sequences and increasing factorizations. In particular, we have the following expansion from [Assa]. where Yam(w −1 ) is the set of increasing tableaux of key shape with row(P ) a reduced word for w −1 and lift(drop(P )) = P .

5.3.
Demazure crystal operators on reduced factorizations with cutoff. Since RFC(w) ⊆ RF n (w) for w ∈ S n , we can restrict the crystal operators f i and e i on reduced factorizations to RFC(w) by defining f i (r) as in Section 4.3 if f i (r) ∈ RFC(w) and f i (r) = 0 otherwise and similarly for e i . An example is given in Figure 16. We will show in this section that this amounts to a union of Demazure crystal structures. We begin with an analog of Theorem 4.12.
Theorem 5.10. Given r ∈ RFC(w) for w ∈ S n , denote by P (r) the weak Edelman-Greene insertion tableau and by Q(r) the weak Edelman-Greene recording tableau, where letters in block i of r are recorded by the letter i. Then, if e i (r) = 0, we have P (e i (r)) = P (r) and Q(e i (r)) = e i ( Q(r)) for 1 i < n.
Proof. By Theorem 4.12 we have P (e i (r)) = P (r) and Q(e i (r)) = f n−i (Q(r)), where P and Q are the Edelman-Greene insertion and recording tableaux, respectively. By Definition 5.6, we have P (r) = lift(P (r)), which proves P (e i (r)) = P (r). Again by Definition 5.6, we have φ( Q(r)) = Q(r). By Lemma 3.9, we have φ(e i Q(r)) = f n−i φ( Q(r)) = f n−i Q(r), proving that Q(e i (r)) = e i ( Q(r)).