Determinants, Paths, and Plane Partitions: k k k i i i i k k π (1) π (k)
Determinants, Paths, and Plane Partitions: k k k i i i i k k π (1) π (k)
Determinants, Paths, and Plane Partitions: k k k i i i i k k π (1) π (k)
(1989 preprint)
Ira M. Gessel1
Department of Mathematics
Brandeis University
Waltham, MA 02254
X. G. Viennot
Departement de Mathematiques
Universite de Bordeaux I
33405 Talence, France
1. Introduction
In studying representability of matroids, Lindstr
om [42] gave a combinatorial interpretation to certain
determinants in terms of disjoint paths in digraphs. In a previous paper [25], the authors applied this theorem
to determinants of binomial coefficients. Here we develop further applications. As in [25], the paths under
consideration are lattice paths in the plane. Our applications may be divided into two classes: first are those
in which a determinant is shown to count some objects of combinatorial interest, and second are those which
give a combinatorial interpretation to some numbers which are of independent interest. In the first class are
formulas for various types of plane partitions, and in the second class are combinatorial interpretations for
Fibonomial coefficients, Bernoulli numbers, and the less-known Salie and Faulhaber numbers (which arise in
formulas for sums of powers, and are closely related to Genocchi and Bernoulli numbers).
Other enumerative applications of disjoint paths and related methods can be found in [14], [26], [19],
[5154], [57], and [67].
2. Lindstr
oms theorem
Let D be an acyclic digraph. D need not be finite, but we assume that there are only finitely many
paths between any two vertices. Let k be a fixed positive integer. A k-vertex is a k-tuple of vertices of
D. If u = (u1 , . . . , uk ) and v = (v1 , v2 , . . . , vk ) are k-vertices of D, a k-path from u to v is a k-tuple
A = (A1 , A2 , . . . , Ak ) such that Ai is a path from ui to vi . The k-path A is disjoint if the paths Ai are
vertex-disjoint.
Let Sk be
the set of permutations of {1, 2, . . . , k}. Then for Sk , by (v) we mean the
Y
P ui , v(i)
P u, (v) =
i=1
We use the notation |mij |sr to denote the determinant of the matrix (mij )i,j=r,...,s .
Theorem 1. (Lindstr
om [42])
X
k
(sgn )N u, (v) = |P (ui , vj )|1 .
Sk
1
(2.1)
Plane Partitions
(sgn )N u, (v) =
(sgn )P u, (v) .
Sk
(2.2)
Sk
P u, (v) N u, (v)
Sk
The conjugate 0 of is the partition whose diagram is the transpose of that of . We write if i i
for each i. If , then the diagram of is obtained from the diagram of by removing the diagram
of .
Plane Partitions
If i = i1 + 1 for i > 1, then the diagram of is called a skew hook (also called rim hook or
border strip). For example,
(3.1)
pij pi+1,j
(3.2)
and
whenever these entries are defined. The integers pij are called the parts of the plane partition. For example,
a plane partition of shape (431) (110) is
3
4
If has no nonzero parts, it is omitted. The plane partition (pij ) is row-strict if (3.2) is replaced by
pij > pi,j+1 and column-strictness is defined similarly.
Reverse plane partitions are defined by reversing all inequalities in the above definitions.
A tableau is a column-strict reverse plane partition. A row-strict tableau is a row-strict (but not necessarily column-strict) reverse plane partition. A standard tableau is a reverse plane partition in which the
parts are 1, 2, . . . , n, without repetitions, for some n.
We shall first apply Theorem 1 to the digraph in which the vertices are lattice points in the plane and
the edges go from (i, j) to (i, j + 1) and (i + 1, j). Thus paths in this digraph are ordinary lattice paths with
unit horizontal and vertical steps. Later, we shall consider some modifications of this digraph.
Correspondences between k-paths and arrays are important in what follows. We restrict ourselves to
k-paths from u to v where ui = (ai , bi ) and vi = (ci , di ), and the parameters satisfy ai+1 < ai , bi+1 bi ,
ci+1 < ci , and di+1 di for all i. These conditions imply that (u, v) is nonpermutable. The strictness
conditions, which are not necessary for nonpermutability, allow shapes to be parametrized by partitions, and
allow a simpler translation of the disjointness condition on k-paths into a condition on arrays.
The correspondences are determined by first choosing a labeling of all the horizontal steps in the digraph
of lattice points. Then we associate to a k-path an array in which row i consists of the labels of the horizontal
steps of path i, with each row shifted one place to the right in relation to the previous row. As an example,
Figure 1
in Figure 1 the horizontal step from (l, h) to (l + 1, h) is assigned the label h and the corresponding array is
1
Plane Partitions
With this labeling, a disjoint k-path corresponds to a tableau. If
4
instead we assign to the horizontal step
Figure 2
from (l, h) to (l + 1, h) the label l + h, as in Figure 2, the corresponding array is a row-strict tableau:
2
To construct the most general labeling, it is convenient to start with the first labeling described above,
and then relabel.
The first correspondence sketched above, which assigns to the k-path A an array T , may be described
more formally as follows: If there is a horizontal step from (l, h) to (l + 1, h) in path i, then we define Ti,l+i
to be h. In other words, Tij is the height of the horizontal step in path i from x = j i to x = j i + 1
if such a step exists, and is undefined otherwise. The essential fact about this correspondence is that A is
disjoint if and only if T is a tableau. (As a technicality, in order to be consistent with the requirement that
j 1 for all parts Tij , we need that ak > k.) Note that a tableau does not uniquely determine a k-path
since the endpoints are not determined. However the correspondence does give a bijection between tableaux
of shape satisfying bi Tij di , where i = ai + i 1 and i = ci + i 1, and k-paths with initial
points (ai , bi ) and endpoints (ci , di ). Theorem 1 allows us then to count these tableaux.
We now relabel the tableau. Let L be a set of labels (which will usually be integers). For each
p L we have a weight w(p), usually an indeterminate. A relabeling function is a sequence f = {fi }
i=
of injective functions from the integers to L. Given a relabeling function f we define the weight of a vertical
step to be 1 and the weight of a horizontal step from (r, s) to (r + 1, s) to be w(fr (s)). Then the sum of the
weights of all paths from (a, b) to (c, d) is
Hf (a, b, c, d) =
(3.3)
(3.4)
1
(Uij ) di .
bi fji
The reader may check, for example, that if we take fi (n) = n + i, then f (T ) is a row-strict tableau and if we
take
Q fi (n) = n, then f (T ) is a column-strict plane partition. Let us define the weight of U = f (T ) to be
u w(u) over all parts u of U . Then by Theorem 1, the sum of the weights of all arrays of the form f (T ),
k
where T is a tableau of shape , and bi Tij di , is the determinant |P (ui , vj )|1 ,where ui = (ai , bi ),
vi = (ci , di ), i = ai + i 1, and i = ci + i 1. Thus we have
Plane Partitions
Theorem 3. Let f be a relabeling function, let and be partitions with k parts, and let the integers
bi and di satisfy bi+1 bi and di+1 di . Then the sum of the weights of f (t) over all tableaux T of shape
satisfying bi Tij di is
k
(3.5)
|Hf (i i + 1, bi , j j + 1, dj )|1 ,
where Hf is defined by (3.3).
In Section 5, we will use Theorem 3 to count tableaux in which an entry l in position (i, j) is assigned
the weight xlij . For now, we consider the case in which fi (n) = n + ti for some number t, with w(l) = xl ,
where the xl are indeterminates. Here Hf (a, b, c, d) becomes
X
(3.6)
where the sum is over b na na+1 nc1 d. If we set nj + tj = mja+1 , we may rewrite (3.6) as
X
xm1 xmca ,
m1 ,m2 ,,mca
where the sum is over all sequences m1 , . . . , mca satisfying m1 b + ta, mca d + t(c 1), and mj+1
mj + t.
P
(t)
Let us now write Hn (a, b) for xi1 xi2 xin over all i1 , . . . , in satisfying i1 a, in b, and il+1 il +t
(t)
for all l. Note that for t = 0, Hn (a, b) is the complete symmetric function of degree n in xa , . . . , xb , and for
(t)
t = 1 it is the elementary symmetric function of degree n in these variables. For other values of t, Hn (a, b)
is not symmetric.
We find that (3.5) reduces to
(t)
Hj i +ij bi + t(i i + 1), dj + t(j j) ,
and simplifying (3.4) we obtain the following:
Corollary 4. Let and be partitions with k parts and let the integers bi and di satisfy bi+1 and
di+1 di . Then the sum of the weights of all arrays U of shape satisfying
is
Uij Ui,j+1 t
(3.7)
(3.8)
(3.9)
(t)
Hj i +ij (bi + t(i i + 1), dj + t(j j)) .
1
Note that (3.9) may be replaced by inequalities on only the first and last element of each row:
bi + t(i + 1 i) Ui,i +1
and
Ui,i di + t(i i).
If we set Ai = bi + t(i i + 1) and Bi = di + t(i i) we may restate this result as follows:
Let and be partitions with k parts and let the integers Ai and Bi satisfy Ai+1 Ai t(i+1 i 1)
and Bi+1 Bi t(i+1 i 1). Then the sum of the weights of all arrays U of shape satisfying
(3.7), (3.8), and
Ai Ui,i +1 , Ui,i Bi
Plane Partitions
(t)
is Hj i +ij (Ai , Bj ) .
1
We obtain further interesting results by substituting for the variables, or removing the part restrictions.
In the cases t = 0 and t = 1, these generating functions are called flagged Schur functions [72].
First we remove the part restrictions in Corollary 4. We can do this most easily by taking the limit as
(t)
bi goes to and di goes to +. (It is easily verified that this is legitimate.) Let us define hn by
X
lim Hn(t) (a, b) =
xi1 xin
h(t)
n = a
b
(1)
Thus hn is the ordinary complete symmetric function hn and hn is the elementary symmetric function
(t)
(t)
en . Let us write s/ for the determinant |hi j +ji |k1 . (To agree with usual practice we have transposed
(t)
the determinant in Corollary 4.) Thus for t = 0, s/ is the ordinary Schur function. Then we have
(t)
(t)
(3.10)
X
n=0
fn
zn
,
n!
z i j +ji
(i j + j i)!
1
for standard tableaux of shape . Since each standard tableau contributes a term z n /n!, where n =
P
i (i i ), the number of standard tableaux of shape is
n!
1
.
(i j + j i)! 1
(t)
An interesting question is whether there is any reasonable expression for the coefficient of x1 x2 xn in s/ .
Plane Partitions
mn
n
Pm (u) =
. Let us set
bm/2c
e(1)
(1, . . . , 1)un
n
| {z }
n=0
m1
Lemma 6.
bm/2c
Pm (u) =
mn n
u .
n
j
).
m+1
(1 + 4u cos2
j=1
m m/2
Pm (u) = (i) u
Um
2 u
where i = 1 and Um (x) is the Chebyshev polynomial determined by Um (cos ) = sin(m + 1)/ sin . Since
Um (x) is a polynomial of degree m with leading coefficient 2m which vanishes at x = cos(j/(m + 1)) for
j = 1, . . . , m, we have
m
Y
j
.
x cos
Um (x) = 2m
m+1
j=1
Since
cos
we have
j
(m + 1 j)
j
= cos
= cos
,
m+1
m+1
m+1
m/2
Y
m
2
2
,
m even;
2
cos
x
m+1
j=1
Um (x) =
(m1)/2
m
2
2
, m odd.
x cos
2 x
m+1
j=1
Pm (u) = (u)
j=1
1
j
cos2
4u
m+1
m/2
j=1
j
1 + 4u cos
m+1
(m1)/2
j=1
1
j
cos2
4u
m+1
(m1)/2
j=1
1 + 4 cos2
j
m+1
Plane Partitions
Theorem 7.
bm/2c
(1)
, . . . , 4 cos2
.
s/ (1, . . . , 1) = s/ 4 cos2
| {z }
m+1
m+1
m1
2
2 bm/2c
(1,
.
.
.
,
1
)
=
e
,
.
.
.
,
4
cos
.
4
cos
e(1)
n
n
| {z }
m+1
m+1
m1
YY
4u cos2
i=1 j=1
i
j
+ 4v cos2
m+1
n+1
(4.1)
We can also interpret the product in (4.1) in terms of tableaux. Macdonald [43, Ex. 5, p. 37] gives the
formula
s
r Y
Y
X
(xi + yj ) =
s (x)s 0 (y),
i=1 j=1
where the sum is over all partitions with at most r parts and largest part at most s, and
0 = (r 0 , . . . , r 0 ),
s
1
when rotated 180 , fits together with the diagram of to form an r s rectangle. It
i.e., the diagram of ,
follows that with m = 2r and n = 2s we have
X
m/2 n/2
YY
i
j
0
(1)
(1)
s (1, . . . , 1)s 0 (1, . . . , 1)u|| v | |
+ 4v cos2
=
4u cos2
| {z }
| {z }
m
+
1
n
+
1
i=1 j=1
m1
n1
0
(1)
(2)
s (1, . . . , 1)s (1, . . . , 1)u|| v | |
| {z }
| {z }
m1
(4.2)
n1
The sum on the right side of (4.2) has a simple combinatorial interpretation: Let us define a dimer
tableau to be an array (pij ) with entries chosen from the alphabet A A0 , where A = {1, . . . , m 1} and
A0 = {10 , . . . , (n 1)0 } such that if pi,j = and pi+1,j = , one of the following holds:
(1) A and A0 ,
(2) , A and + 1,
(3) , A0 and < 10 ,
and if pij = and pi,j+1 = , one of the following holds:
(1) A and A0 ,
(2) , A and < 1,
(3) , A0 and + 10 .
(The total order, addition, and subtraction in A0 have their obvious meaning.) A dimer tableau with i entries
in A and j entries in A0 is assigned the weight ui v j . Then we have
Plane Partitions
Theorem 8. Let m and n be even. Then the number of coverings of an m n rectangle by 2i vertical
dominoes and 2j horizontal dominoes is equal to the number of (m/2) (n/2) dimer tableaux with i entries
in {1, . . . , m 1} and j entries in {10 , . . . , (n 1)0 }.
A simple bijection between these two sets would be interesting. Such a bijection is easy to construct
for m = 2. Here the dimer tableau has only one row, which consists of some number of 1s followed by a
sequence of elements of {10 , 20 , . . . , (n 1)0 }, each at least 20 more than its predecessor. Given such a tableau,
we construct a covering of a 2 n rectangle by putting the left end of a pair of horizontal dominoes at the
positions corresponding to the primed numbers, and fill up the remaining space with vertical dominoes.
Figure 3 illustrates the correspondence for 2 4 rectangles.
Figure 3
|h(i i + 1, bi , j j + 1)|1 ,
where h(a, b, c) is defined by (5.1).
If is empty and the bi are all zero (or equivalently by column-strictness, bi = i 1) then there is an
explicit product expression for the trace generating function, due to Gansner [22], using the Hillman-Grassl
correspondence [31]. To describe Gansners result, we define the hook lengths and contents of a diagram.
The hook length of a square in a diagram is the number of squares to its right plus the number of squares
below it plus one. Thus the hook length of square (i, j) is i + 0j i j + 1. The content of square (i, j) is
j i. The hook lengths and contents of the diagram for the partition (431) are as follows:
6
1
Hook lengths
Contents
Since every entry of row i of a tableau of shape with nonnegative parts is at least i 1, we can factor
out
k
Y
i=1
from the trace generating function for tableaux to leave the trace generating function for reverse plane
partitions (with no strictness condition). Gansner [22, p. 132, Theorem 4.1] showed that the trace generating
Plane Partitions
10
function for reverse plane partitions of shape with nonnegative entries may be described as follows: Let
us define the hook of a square s in the diagram of a partition to be the set of all squares to the
Q right of s
or below s, together with s. The content c(s) of square (i, j) is j i. Let X(s) be the product xc(t) over
all squares t in the hook of s. Then Gansners result is that the trace generating function for reverse plane
partitions of shape with nonnegative entries is
Y
s
1
1 X(s)
n
X
hk (x)enk (y).
k=0
Theorem 10.
s/ = hi j +ji .
1
We give here only a sketch of a proof. In the next section we shall consider a generalization (which is
less clear geometrically).
We work with paths with vertical and horizontal steps as before, but we also allow diagonal steps, and
we label them as in Figure 4. One can check that applying Corollary 2 to this digraph yields Theorem 10.
Figure 4
A result equivalent to Theorem 10 was first proved by Stanley [63] using the Littlewood-Richardson
rule. The proof sketched above was given by Remmel [52] in a less general form. A related result was given
by Thomas [68,Theorem 3].
The symmetric functions s are sometimes called hook Schur functions because a tableau counted by
s (x1 , . . . , xm ; y1 , . . . , yn ) lies inside the hook { (i, j) | 1 i m or 1 j n }. They have also been
studied by Berele and Regev [7] and Worley [74].
Plane Partitions
11
sR
/ = hi j +ji .
We now ask, under what conditions is it true that sR
/ counts arrays (pij ) of shape with entries in X
which satisfy
pij R pi,j+1
(7.1)
pij S pi+1,j .
(7.2)
and
R
for some relation S? By looking at sR
(11) and s(21) we infer that the only reasonable choice for S is the relation
{ (a, b) | b 6R a }, where 6R is the negation of R. (The examples mentioned above are of this form.) So we
R=
define an R-tableau to be an array (pij ) which satisfies (7.1) and (7.2) with S =R.
h2 h3
R
There is a simple characterization of relations with this property. We prove the generalization of
Corollary for these relations by restating the proof of Theorem 1 in terms of tableaux.
A relation R on a set X is called semitransitive (see, e.g., [12] or [21]) if it satisfies the following condition:
() For all a, b, c, d in X, if a R b R c then a R d or d R c.
It is easily verified that the following property is equivalent to (), and although less symmetrical, is
more convenient for our proof:
b R c R d then a R d.
For all a, b, c, d in X, if a R
is.
It is easy to show that R is semitransitive if and only if R
Theorem 11. If R is semitransitive then sR
/ counts R-tableaux of shape .
Proof. We work with arrays (aij ) defined for i = 1, . . . , k; i < j mi for some numbers mi , and
satisfying aij R ai,j+1 . A failure of the first kind of such an array is a position (i, j) such that aij R
6 ai+1,j ,
i.e., ai+1,j R aij . A failure of the second kind is a position (i, j) with aij undefined (but i 1) and ai+1,j
Plane Partitions
12
defined. It is clear that an array with no failures is an R-tableau. As in the proof of Theorem 1, we shall
define an involution on the set of arrays with failures that has the right properties.
There is one technical point that we should mention first. In the involution used in the proof of Theorem
1, two nonadjacent paths are sometimes switched. It is more convenient to use a different choice of intersecting
paths as the model for this proof; otherwise we would have to consider failures between nonadjacent rows,
which correspond to intersections of nonadjacent paths. As a first attempt, we might try choosing the least
i such that paths i and i + 1 intersect; however, this does not give an involution. For example, in Figure 5a
Figure 5
paths 2 and 3 would switch at x, giving Figure 5b, but the involution applied to Figure 5b would try to
switch the new paths 1 and 2 at y. Thus a slightly more complicated choice is necessary.
We can modify the involution as follows: call an intersection point x of paths i and j early if x is the
first intersection point on both paths. Then it is easy to see that for the rectangular grid digraphs we are
using, the set of early intersections does not change when a switch is made at an early intersection, every
early intersection involves two consecutive paths (as long as the paths are numbered correctly), and every
nondisjoint k-path has at least one early intersection. (The last statement does not hold in general for
arbitrary acyclic digraphs.) Thus we can choose our switching point to be the least early intersection.
Returning to the proof of Theorem 11, we define an early failure to be a failure (i, j) such that there
are no failures (i 1, j 0 ) with j 0 < j, no failures (i, j 0 ) with j 0 < j, and no failures (i + 1, j 0 ) with j 0 j. It
is clear that every array with a failure has an early failure. We define the earliest failure of an array to be
the early failure (i, j) with i as small as possible. For example, if R is then in the array
1
4
(4, 1) is an early failure, (2, 2) is the earliest failure, and (1, 2) is a failure which is not early.
To construct the involution on arrays with failures, we perform a switching operation on rows i and
i + 1, where i is the row of the earliest failure. The switching is described most easily by (noncommutative!)
diagrams for several cases: First suppose we have a failure of the first kind between b and e where C and G
are the rest of the rows:
R
b C
a x
R
yR
R
R
d e f G
We change this to
yR
R
d
f
x
R
R
e
G
b
d R e R f .) This
which also has a failure of the first kind in the same place. (Here we have a R f since a R
case also applies if a or a and d are absent. If f is absent, we switch the same way, and we obtain a failure
of the second kind.
Similarly, if the earliest failure is of the second kind, above e in
a
yR
R
d
Plane Partitions
we change it to
yR
R
d
f
x
R
e
13
xa xb xc xd ,
Plane Partitions
14
First we prove that p R q. We consider two cases. In the first case, we assume that q 6R q and a R q.
We know that either a R p or p R a, by (i). If p R a then p R a R q, so by (iii), p R q. If a R p then we have
p R p, since otherwise p would be in X2 . Thus by (i), if p 6R q then q R p, so a R q R p, and thus p X2 , a
contradiction.
In the second case, there exists y with y 6R y and a R y R q. Then by the first case, we have p R y, so
p R y R q. Then by (iii), p R q unless p R p. Thus by (i), if p 6R q then q R p and we have q R p R y, so
q R y by (iii). Thus by (ii), y R y, a contradiction.
It remains to prove that q 6R p. But if q R p then since we have just shown that p R q, it follows from
(ii) that q R q. Thus there must exist some y with y 6R y and a R y R q. Therefore we have y R q R p so by
(iii), y R p. But since y X2 , it follows from what we have already proved that p R y, so by (ii) we have
y R y, a contradiction.
It follows that all semitransitive relations can be constructed from partial semiorders.
The following construction shows that partial semiorders are easy to find: Start with a strict total order
< on X. Let f be function from X to X such that for all a, b X, f (a) a and a b implies f (a) f (b).
Then let
R = { (a, b) | f (a) < b }.
(7.3)
To see that R is semitransitive, suppose a R b R c but a 6R d. Then we have f (a) < b, f (b) < c, and
f (a) d. Then d f (a) < b, so f (d) f (b) < c, and thus d R c. The condition f (a) a impies that R
is irreflexive. It is interesting to note that the number of such functions on an n-element set is the Catalan
number Cn . The relation on the integers given by a R b iff a b t, where t 1, comes from the function
f (x) = x + t 1.
Not every partial semiorder is of this form; a counterexample is a disjoint union of two 2-element chains.
To see this, note that if R is defined by (7.3), then the sets Rx = { y | x R y } over all x X are totally
ordered by inclusion.
Edrei [18] (see also Karlin [34, p.412]) proved the following theorem, which was conjectured by Schoen
k
berg (see [61]): Let rn be real numbers, with rn = 0 for n < 0. Then all determinants sr = ri j +ji
/
X
n=0
rn un = Cu eu
M
Y
N
Y
(1 + i u)
(1 i u),
i=1
i=1
P
where i , i , and are positive real numbers, C 0, (i + i ) < , and M and N may be infinite. It
P
i
follows that for any finite partial semiorder R, if we set PR (u) = i hR
i u , then for any nonnegative values
of the xi , PR (u) is a polynomial in u with negative roots.
The hook Schur functions give a strengthening of the easy half of Edreis theorem: if the i , i , and
are indeterminates, then sr/ has nonnegative coefficients as a polynomial in these variables. (The parameter
can be accounted for via the homomorphism of Section 3, by introducing a third set of labels, each of
which can appear at most once.)
There is another way of looking at R-Schur functions: we can interpret them as evaluations of ordinary
Schur functions. For the moment, let us think of hn , n 1, as indeterminates, with h0 = 1, and define the
skew Schur functions s/ as polynomials in the hn . Then sR
/ is the image of s/ under the substitution
R
hn 7 hR
n . It follows that any polynomial identity among the s/ remains true when s/ is substituted for
s/ . For example, we have the formulas
X
f s ,
(h1 )n =
over all partitions of n, where f is the number of standard Young tableaux of shape , and
X
c s ,
s s =
Plane Partitions
15
where the integers c are given by the Littlewood-Richardson rule [43, p. 68]. It is reasonable to expect
that these formulas can be explained combinatorially by analogs of the Robinson-Schensted correspondence
[56, 59] and the jeu de taquin of Sch
utzenberger [62]. This has been done in the special case of the hook
Schur functions by Remmel [52] and Worley [74].
8. Stanleys formula
Much of the theory of plane partitions is devoted to counting plane partitions or tableaux by the
sum of their entries. To accomplish this we set xi = q i in the formulas of Section 3. First we evaluate
hn (xa , xa+1 , . . . , xb ) with xi = q i . Using the well-known expansions of (1 z)(1 zq) (1 zq n ) and its
reciprocal [1, p. 19] we have
hn (q a , q a+1 , . . . , q b )z n =
(1
n=0
zq b )
X
1
j aj b a + j
=
z
q
(zq a )ba+1
j
j=0
=
and
zq a ) (1
en (q a , q a+1 , , . . . , q b )z n = (1 + zq a ) (1 + zq b )
n=0
ba
Y
(1 + q zq ) =
i=0
So
a
hn (q , . . . , q ) = q
an
j
+aj b a + 1
(
)
2
z q
.
j
j
ba+n
n
and
n
en (q a , . . . , q b ) = q an+( 2 )
ba+1
.
n
y0 +r
j
| for counting r-tuples of paths between two given
[Kreweras [39, p. 64] gave the determinant | yiij+r
0
paths of heights yi and yj .]
Applying these formulas to Corollary 4, we obtain
Theorem 14. The generating function for tableaux of shape in which parts in row i are at least
ai and at most bi , where ai ai+1 and bi bi+1 , is
hi j +ji (q aj , . . . , q bi )k = q aj (i j +ji) bi aj + i j + j i
(8.1)
1
i j + j i
1
and the corresponding generating function for row-strict tableaux, where ai+1 ai i+1 i 1 and
bi+1 bi i+1 i 1, is
E
bi aj + 1
q ij
,
i j + j i 1
where
Eij =
i j + j i
+ aj (i j + j i).
2
(8.2)
Analogous formulas for column-strict plane partitions are easily obtained from Theorem 14, since replacing each entry with its negative changes a tableau to a column-strict plane partition.
Then replacing
ai
n
n
and bi with their negatives and q by q 1 in Theorem 14, and using the identity
= q k(nk)
, we
k q1
k q
obtain:
Plane Partitions
16
Theorem 15. The generating function for column-strict plane partitions of shape in which parts
in row i are at most ai and at least bi , where ai ai+1 and bi bi+1 is
b ( +ji) aj bi + i j + j i k
.
q i i j
i j + j i
1
(8.3)
We can easily modify formula (8.3) to count (not necessarily column-strict) plane partitions:P
if we add i
to every part in row i of a column-strict plane partition counted by (8.3), we increase the sum by i (i i )
and we obtain a plane partition of shape in which parts in row i are at most ai + i and at least bi + i.
Setting Ai = ai + i and Bi = bi + i, we obtain:
Theorem 16. The generating function for plane partitions of shape in which parts in row i are
at most Ai and at least Bi , where Ai Ai+1 1 and Bi Bi+1 1 is
Aj Bi + i j
.
q i i(i i ) q (Bi i)(i j +ji)
i j + j i 1
There are many determinants of q-binomial coefficients which have explicit formulas as quotients. Some
of these are very difficult to evaluate [24, 6, 46, 47, 49?]. In this section we give an easy evaluation of a
determinant by induction, which yields a result of Stanley counting tableaux of a given shape with a given
maximum part size. (For other work on evaluation of determinants of matrices of binomial and q-binomial
coefficients, see [10,29, 45, 48, and 69]. In the next section we use a simple summation formula to evaluate
a determinant and give two (?) applications which seem to be new.
Let us consider the problem of counting tableaux of shape , with parts 0, 1, , b 1. By formula (8.1)
the generating function for these tableaux is
b + i + j i k
i + j i .
1
However, in such a tableau the smallest part in row i must be at least i 1 (since = 0) so this determinant
is also equal to
i + j i
i + j i
1
(8.4)
1
Hook lengths
2
Contents
We write h(x) and c(x) for the hook length and content of x, and we set H() =
Q
Ca (x) = x (1 q a+c(x) ), where the product is over all squares x of the diagram of .
x (1
q h(x) ) and
Plane Partitions
17
Theorem 17. (Stanley [65]) The generating function for tableaux of shape with parts 0, 1, , b 1
Pk
is q Cb ()/H(), where e = i=1 (i 1)i .
e
Proof. We evaluate the determinant (8.4) by induction. First note that if k = 0 (8.4) reduces to the
formula for 1 , 2 , , k1 .
Now first suppose that b is greater than k, the number of parts of . Note that
Cb () = Cb1 ()
1 q bj
bj
bj1
=
k
Y
(1 q b+i i )
i=1
(1 q bi )
qe
Cb1 ()
Cb ()
= qe
.
H()
H()
Cb ()
Cb ()
= qe
.
H()
H()
and
Cb ()
Cb ()
Cb ()
Ck ()
Ck ()
Cb ()
=
.
=
=
H()
H()
H()
H()
1 n
(i )j =
0
(i j )
0i<jn
n
Y
(i )n
i=0
(i )n n
(ai )j =
0
The left side is
(i j )
0i<jn
Plane Partitions
Lemma 19.
n
Y
(i j )j n
(i )j =
0
i=0
(i )i
18
(j i )
0i<jn
n
Y
(i )n
i=0
1
(i )j
C = AB is
n
X
l=0
(j )i (j)i
. Then the (i, j) entry of
i!
1 (j )l (j)l
(i j )j
=
(i )l
l!
(i )j
by Vandermondes theorem [5, p. 3]. Thus |C| = |A||B|. Since B is upper triangular,
|B| =
n
n
Y
n Y
(j )j (j)j
= (1)( 2 )
(j )j
j!
j=0
j=0
i +j
We now give two applications of Lemma 19. First let us evaluate the determinant
i +j1 . Since
0
i + j
i + j 1
i ( i + 1)j
i 1
(i )j
= (1)j
(i j)j
( i )!
.
( 2i + 1)! (i 1)!
(i )j
Thus
n
i + j n
n+1 Y
( i )! ( + i)i
= (1)( 2 )
i + j 1
( 2i + 1)! (i 1)! (i )n
0
i=0
=
n
Y
( i )! ( + i)i
( 2i + 1)! (i 1)! (i )n
i=0
n
Y
( i )! ( + i)i
( 2i + 1)! (i + n 1)!
i=0
(j i )
0i<jn
(i j )
0i<jn
(i j ).
(9.1)
0i<jn
This determinant can clearly be interpreted as counting certain tableaux. The value of the determinant
can be simplified and expressed in a form very similar to Stanleys hook length-content formula. The hook
lengths are the same, but the contents are different. If (i, j) is a square of the diagram of we define d(i, j)
as follows:
(i + j 2i + 1) if i j;
d(i, j) =
if i > j.
0j + i 2j
These numbers can be described as follows: if x is a diagonal square, d(x) is one more than the number of
squares to the right of x, and as we move right from a diagonal square d decreases by 1 for each square. If
x is a subdiagonal square, d(x) is 2 more than the number of squares below x, and as we move down from a
Plane Partitions
19
subdiagonal square, d increases by 1 for each square. Here are the values of d(i, j) for the partition (44331):
5 6 7 8 9
5
4 5 6 7
In this section we let H() be the product of the hook lengths of . We will need the following lemma (see
Macdonald [43, p. 9]):
Lemma 20. For any partition ,
Y
(i j i + j)
1i<jk
k
Y
(i i + k)!
1
.
H()
i=1
Theorem 21. The number of tableaux of shape with positive integer parts, in which the largest part
in row i is at most r 2i + 2i 1 is
1 Y
r + d(x) .
H()
x
r i + i + j 2 k
,
i i + j
1
and the combinatorial interpretation follows from Theorem 14. The value of the determinant is
k
Y
(r i + i 1)! (r + i)i1
(r
2i + 2i 2)! (i i + k)!
i=1
(i j + j i)
1i<jk
1 Y (r + 2i 2)(r + 2i 3) (r + 2i 2i 1)
H() i=1
(r + i 1)(r + i 2) (r + i i )
i
(r + 2i 2j)(r + 2i 2j 1)
1 YY
H() i=1 j=1
(r + i j)
Y (r + 2i 2j)(r + 2i 2j 1)
1
H()
(r + i j)
(i,j)
Plane Partitions
20
First we have
k Y
i i
Y (r + 2i 2j)(r + 2i 2j 1) Y
(r 2l)(r 2l 1)
=
(r
+
i
j)
(r l)
(i,j)
i=1
l=0
ij
2iY
2i+1
k
Y
m=0
Y
i i
i=1
(r m)
(r n)
n=0
k
Y
2iY
2i+1
(r m).
(r j i + 2i 1) =
r + d(x) .
i=1 j=i
x+
0 j
j
k
Y (r + 2i 2j)(r + 2i 2j 1)
Y
Y
(r + 2l 1)(r + 2l)
=
(r
+
i
j)
(r + l)
(i,j)
j=1
l=1
i>j
20j 2j
k
Y
j=1
(r + m)
m=1
0j j
(r + n)
n=1
0
k
Y
20j 2j
(r + m).
j
k
Y
Y
(r + i + 0j 2j) =
j=1 i=j+1
r + d(x) ,
Another determinant we can evaluate using Lemma 19 is Ci+j 0 , where Cn is the Catalan number
defined by
(1/2)n
1
2n
.
= 22n
Cn =
n+1 n
(2)n
Since
Ci+j = 22i+2j
(1/2)i+j
(1/2)j (1/2 + j )i
= 22i+2j
,
(2)i+j
(2)j (2 + j )i
Plane Partitions
the determinant is equal to
Q
2
(1/2)j
Qj
j (2)j
n(n+1)+2j j
21
(1/2 + j )i n
(2 + j )i
0
Qn
(1/2 + j )i n
= Qn i=0 (3/2)i
(2 + j )i
j=0 (j + 2)n
0
(j i ).
0i<jn
(j i )
0i<jn
n
Y
n
Y
(2j )!
(2i + 1)!
.
! (j + n + 1)! i=0
i!
j=0 j
See Viennot [71] for an evaluation of a special case of this determinant, using the qd-algorithm of Pade
approximant theory. See also de Sainte-Catherine and Viennot [14].
We recall that Ci is the number of paths from (0, 0) to (2i, 2i) which never go above (but may touch)
the line x = y.
Now assume that 0 0 < 1 < < n and let the points Pi and Qi be defined by Pi (n i, n i)
and Qi = (n + i , n + i ) for i = 0, . . . , n. If we consider the digraph of lattice points with only those steps
that lie below the line x = y, then (P, Q) is nonpermutable. Thus the determinant |Ci+j |n0 is the number
of disjoint (n + 1)-paths from P to Q.
These paths may be expressed as tableaux: we consider separately the cases 0 = 0 and 0 6= 0.
First we consider the case 0 = 0. In Figure 6 an example with 0 = 0, 1 = 2, and 2 = 5 is given.
Figure 6
It is clear that path i begins with 2i horizontal steps, and thus we may remove these steps, and associate
what remains with a tableau in the usual way. Thus Figure 6 corresponds to the tableau
2
In general, the tableau will have shape (n n, n1 n + 1, . . . , 1 1). The requirement that all steps
fall below the main diagonal is equivalent to the condition that pij 2n + j i. This condition is implied
by the inequalities for the last elements in each column.
If we make the substitution i = n+1i (n + 1 i), (so that i = n+1i + i) for i = 1, . . . , n and set
k = n, then we obtain the following:
Theorem 22. The number of tableaux (pij ) of shape with nonnegative entries satisfying pij
2k + j i, or equivalently, p0j ,j 2k + j 0j , is
k
Y
1
(i + k i + 1)!
(k + 1)! i=1
k
Y
(i j i + j)
1i<jk
k
Y
(2i + 2k 2i + 2)!
(2i + 1)!
.
(
+
k
i
+
1)!
(
i
+
2k
+
2)!
i!
i
i
i=1
i=0
Plane Partitions
22
Next we consider the case 0 6= 0. Figure 7 shows an example with 0 = 2, 1 = 4, 2 = 5. Here path
Figure 7
i begins with 2i + 1 horizontal steps, which may be removed as before. Figure 7 corresponds to the tableau
0
3
In general, the tableau will have shape (n n 1, n1 n, . . . , 0 1) and the restriction on the parts is
pij 2n + 1 + j i. Now we make the substitution i = ni+1 n + i 2 for i = 1 to n + 1 and we set
k = n + 1. Then we have the following:
Theorem 23. The number of tableaux (pij ) of shape , with nonnegative parts satisfying pij 2k
1 + j i, or equivalently, p0j ,j 2k 1 + j 0j is
Y
1i<jk
k
Y
k1
Y (2i + 1)!
(2i + 2k 2i + 2)!
.
(i j i + j)
(i + k i + 1)! (i + 2k i + 1)! i=0
i!
i=1
is
n+1
x
where
m
j F
i
nj
i,j=0,...,n
r+1
n+1
)
(
2
+
(1)
xn+1r ,
r
F
r=1
n+1
X
Here Fj is the Fibonacci number (F0 = 0, F1 = 1, and Fj = Fj1 + Fj2 for j 2).
First we generalize Carlitzs result. Let be the linear transformation on the vector space of polynomials
in x of degree at most n defined by
1
(A(x)) = xn A s(1 + )
x
for any such polynomial A(x), where s is arbitrary.
i i
Pn
s , the matrix of with respect to the basis {xi }i=0,...,n
Since (xi ) = xn si (1 + x1 )i = j=ni xj nj
is
i
.
si
nj
i,j=0,...,n
Plane Partitions
23
Now we have
k
nk
1 + as
1 + bs
k
nk
k
nk
x
x
1+
1+
.
= (as) (bs)
(1 + ax) (1 + bx)
as
bs
Thus if (1 + as)/as = a and (1 + bs)/bs = b, i.e.,
a, b =
s2 + 4s
2s
then (1 + ax)k (1 + bx)nk will be an eigenvector for with eigenvalue (as)k (bs)nk . As long as s2 + 4s 6= 0
these will be distinct for k = 0, . . . , n, and thus give all the eigenvalues of .
By continuity, we may remove the restriction on s and we obtain:
i
are k nk , k = 0, . . . , n,
Theorem 24. The eigenvalues of the matrix M =
si
nj
i,j=0,...,n
where
s s2 + 4s
,
, =
2
and thus the characteristic polynomial of M is
n
Y
|(xI M )| =
(x k nk )
(10.1)
k=0
k nk
(x + p q
)=
k=0
n+1
X
n+1k
k=0
n+1
X
k=0
k
n+1
)
(
2
(pq)
.
k p,q
k+1
k
n+1
xn+1k (1)( 2 ) s(2)
,
k ,
since = s.
Now let Gn = (n n )/( ). The following facts about Gn are easily verified:
G0 = 0,
X
u
Gn un =
1
s(u
+ u2 )
n=0
n
X
i
i
s
.
Gn =
ni
i=dn/2e
(10.2)
Plane Partitions
24
to be m
Note that for s = 1, Gn reduces to Fn . Let us define m
j
j , , so that
m
Gm Gm1 Gmj+1
.
=
G 1 G 2 Gj
j
We note that from the easily proved formula Gm+n = Gm Gn+1 + sGm1 Gn we obtain the recurrence
m1
m1
m
+ sGj1
(10.3)
= Gmj+1
j1
j
j
We now turn to the combinatorial interpretation. We know that the coefficient of xn+1k in |(xI M )|
is (1)k times the sum of all the k k principal minors of M , i.e., the minors obtained by choosing k rows
and the same k columns from M . Such a minor is a determinant
ri
ri
(10.4)
n rj s .
1
To get a determinant in the form we can interpret, we reverse the order of the columns in (10.4); then (10.4)
k
is equal to (1)(2) sri times the determinant
ri
(10.5)
n rk+1j
1
The determinant (10.5) is easily interpreted by our theory. It is in fact a binomial determinant, as studied
in Gessel and Viennot [25], and it has the following interpretation: define points Pi and Qi by Pi = (0, i)
and Qi = (n + i, n + i). For any subset R = {r1 < r2 < < rk } of {0,1, . . . , n}, let N (R) be the
i
number of nonintersecting k-paths from (Pr1 , , Prk ) to (Qrk , , Qr1 ). Then nrrk+1j
= N (R).
P
Theorem 25. For any subset R of {0, 1, . . . , n} let kRk = iR i and let N (R) be defined as above.
Then
X
k
n+1
kRk
)
(
2
,
s N (R) = s
k
R
(11.1).
In many cases, especially when M is a triangular matrix with 1s on the diagonal, (11.1) can be interpreted combinatorially. See, for example, Gessel and Viennot [25, Section 4].
One of the simplest examples is the case in which M is the matrix (hij ). It is easily verified that
M = (eij ). Then Jacobis theorem gives the expression of a skew Schur function in terms of the ei .
The following theorem gives a pair of sign-inverse matrices to which we can apply Jacobis theorem.
Plane Partitions
25
pi u =
i=0
(1) ai u
i=0
and
!1
i
qi u =
i=0
(1) bi u
!
X
i=0
!
i
(1) ai u
i=0
Then
(i) The matrices (aji ) and (pji ) are sign-inverse (where ak = pk = 0 for k < 0),
(ii) The matrices
b0
a0
0
..
.
b1
a1
a0
..
.
0
V =
.
.
.
q0
p0
0
..
.
q1
p1
p0
..
.
0
U =
.
.
.
0
and
bn
an
an1
..
a0
qn
pn
pn1
..
p0
are sign-inverse.
The proof is straightforward.
Now let us take ai = him+r for fixed m and r, where 0 r < m, and bi = him+s . Then
pi u =
i=0
i=0
him+r u
i=0
and
!1
i
qi ui =
(1)i him+s ui
i=0
.
i
(1) him+r u
i=0
(The variable u is actually redundant in these formulas.) The significance of the restriction 0 r < m is
that it implies that the sum could as well start at instead of at 0.[why?]
Before we apply Jacobis theorem we need to relate complements of sets to conjugate partitions. By
reasoning as in Macdonald [43, p. 15; see also (1.7), p. 3] it follows from Jacobis theorem that if and
are two partitions with at most s parts and largest part at most t, where s + t = w, and if M and N are
sign-inverse matrices with rows and columns indexed 0, 1, . . . , w, then we have
M [{i + s i}1is |{j + s j}1js ] = |M | N [{s 1 + i 0i }1it |{s 1 + j 0j }1jt ]
Plane Partitions
where
26
i = m0 (m 1)i + r + C
i = m0i (m 1)i + C
and
where C is large enough to make
nonnegative.
The simplest case is that in which = (n), = . So 0i = 1 for i = 1, . . . , n. We may take s = 1, t = n.
We get
i = m (m 1)i + r + C, i = 1, . . . , n
i = (m 1)i + C, i = 1, . . . , n
We may take C = (m 1)n, and we obtain
i = (m 1)(n i) + m + r
i = (m 1)(n i).
(11.2)
i+1
i
i = m + r and
i = r + 1.
Thus
Figure 8
Applying the homomorphism of Section 2 and setting u = 1 we get
n=0
(1)n
!1
xmn+r
n1
(11.3)
X
x2n
x
dn
=
,
sin x n=0 (2n)!
where dn is related to the Bernoulli number by dn = (1)n+1 (22n 2)B2n , it follows that
X
(3n)! x2n
x3/2
=
dn
3/2
sin(x ) n=0 (2n)! (2n)!
(3n)!
dn is the number of Young tableaux of shape (n + 2, n + 1, . . . , 3) (n 1, n 2, . . . , 0).
(2n)!
This result may be compared with a combinatorial interpretation of 1 3 (2n + 1)d2n given in Gessel and
Viennot [25, p. 315].
and thus
12. Sali
e numbers and Faulhaber numbers
In view of Jacobis theorem, it is natural to ask when matrices of binomial coefficients to which our
theory applies have inverses which can be expressed in some kind of explicit form. In this section we discuss
a closely related pair of such matrices. The entries of their inverses are numbers which have arisen in other
contexts, but are not well known. One of them, according to Edwards [17], was studied by Faulhaber [20]
in the seventeenth century in connection with formulas for sums of powers. The other was apparently first
considered by Salie [58] in 1963 in connection with the number-theoretic properties of the coefficients of
cosh z/ cos z. (See also Hammersley [ ] and Dumont and Zeng [16].)
Both arrays of numbers can be defined in three ways: as entries of the inverse of a matrix of binomial
coefficients, by generating functions, and by formulas for sums of powers. We shall start with the generating
functions.
Plane Partitions
27
X
n,k=0
cosh 1 + 4t x2
x2n
s(n, k)t
=
(2n)!
cosh x2
k
(12.1)
cosh
x2n+1
=
f (n, k)t
(2n + 1)!
k
n,k=0
1 + 4t x2 cosh x2
.
t sinh x2
(12.2)
It is easily seen that s(n, k) = f (n, k) = 0 for k > n. As we shall see later, |s(n, k)| = (1)nk s(n, k)
and |f (n, k)| = (1)nk f (n, k). The first few values of these numbers are as follows (zeros are omitted):
s(n, k) :
f (n, k) :
n\k
0
1
2
3
4
5
6
7
n\k
0
1
2
3
4
5
6
7
0
1
1
1
1
3
3
1
17
17
6
1
155
155
55
10
1
2073
2073 736
135 15
1
38227 38227 13573 2492 280 21 1
0
1
1/2
1/6
1/6
3/10
5/6
691/210
35/2
1/3
1/3
3/5
5/3
691/105
35
1/4
1/2
1/5
17/12
2/3
118/21 41/15
359/12 44/3
1/6
5/6
14/3
Pm
i=0
1/7
1 1/8
k+1
1X
f (n, k) m(m + 1)
.
2
k=0
Proof. We have
X
r=0
Sr (m)
xr
e(m+1)x 1
= 1 + ex + + emx =
r!
ex 1
sinh(m + 12 )x + cosh(m + 12 )x + sinh x2 cosh x2
e(m+ 2 )x ex/2
=
.
2 sinh x2
ex/2 ex/2
1
=
Thus
X
n=0
S2n+1 (m)
cosh(m + 12 )x cosh x2
x2n+1
=
.
(2n + 1)!
2 sinh x2
(12.3)
Plane Partitions
28
1 + 4t = 2m + 1. Then
cosh 1 + 4t x2 cosh x2
x2n+1
=
S2n+1 (m)
(2n + 1)!
2 sinh x2
n=0
x2n+1
t X
f (n, k)tk
2
(2n + 1)!
n,k=0
k+1
1 X x2n+1 X
f (n, k) m(m + 1)
,
2 n=0 (2n + 1)!
k=0
2n+1
x
.
(2n + 1)!
Formula (12.3) was first studied by Faulhaber [20] in the seventeenth century. Faulhabers work is
described in Edwards [17] Schneider [60], and Knuth . Formula (12.3) was rediscovered by Jacobi [33], who
gave the recurrence (in our notation)
k
1 X
f (n, k) m(m + 1) ,
S2n (m) = (m + )
2
k=0
n
X
sinh 1 + 4t x2
x2n X
,
f (n, k)tk =
(2n)!
1 + 4t sinh x2
n=1
k=0
and moreover,
k+1
f (n, k).
f(n, k) =
2n + 1
Next we consider the analogous formula for Salie numbers. Let
Tr (m) =
m
1 X
(1)mi ir ,
2 i=m
m
X
(1)mi i2n .
i=1
Theorem 28.
k
1X
s(n, k) m(m + 1) .
T2n (m) =
2
k=0
Proof. We have
m
!
1
1
m
ex(m+ 2 ) + ex(m+ 2 )
1 X
xr
1X X
mi r
(1)
i
(1)mi eix =
=
2 r=0 i=m
r!
2 i=m
2(ex/2 + ex/2 )
cosh 1 + 4t x2
cosh(m + 12 )x
1 X
x2n k
=
=
s(n, k)
t
=
x
x
2 cosh 2
2 cosh 2
2
(2n)!
n,k=0
Plane Partitions
and the theorem follows by equating coefficients of
There is a companion formula for
P2n+1
i=1
29
x2n
.
(2n)!
Next we relate the Faulhaber and Salie numbers to inverses of matrices of binomial coefficients.
i+1
is the matrix f (i, j) i,j=0,m , and
Theorem 29. The inverse of the matrix
2i 2j + 1 i,j=0,m
i
is the matrix s(i, j) i,j=0,m .
the inverse of the matrix
2i 2j
i,j=0,m
Proof. We have
X r + 1
r+1
r+1
i2r+2l .
(i 1)i
=2
i(i + 1)
l
l odd
r
X r + 1
X
r+1
r+1
=2
S2r+2l (m) = 2
S2j+1 (m)
m(m + 1)
l
2r 2j + 1
j=0
l odd
X r
r
r
i2rl .
i(i + 1) + (i 1)i = 2
l
l even
r
X r
X
r
r
T2rl (m) = 2
T2j (m)
m(m + 1) = 2
l
2r 2j
j=0
l even
i
Note that since the matrix
is lower triangular with 1s on the diagonal, its inverse has integer
2i 2j
entries, so the Salie numbers are integers. We can now explain Salies interest in these numbers [58]. (See
also Comtet [13, pp. 8687].) Salie was studying the numbers S2n defined by
x2n
cosh x X
S2n
=
.
cos x
(2n)!
n=0
n
x2n X
1
cosh x X 2n
=
2 (1)n
s(n, k)( )k
cos x
(2n)!
2
n=0
k=0
and thus
S2n = 22n (1)n
n
X
k=0
k X
n
1
s(n, k)
=
s(n, k)(1)nk 22nk .
2
k=0
(i)
Thus S2n is divisible by 2n , as shown by Salie. (Salie used the notation cij for our s(i, j).)
Plane Partitions
30
n
It follows from equation (12.4) below that s(n, n) = 1, s(n, n 1) =
, and s(n, n 2) =
2
n1 n
n
n
n1 n
n
+4
4
(mod 8).
2n S2n 1 + 2
2
2
2
4
Salie obtained a congruence (mod 16) for the numbers 2n S2n by a different method. See also Carlitz [8].
Next we consider the combinatorial interpretation of the Salie numbers and Faulhaber numbers. First
we note the following lemma, which follows easily from the formula for the inverse of a matrix.
Lemma 30. Let Aij i,j=0,...,m be an invertible lower triangular matrix, and let (Bij ) = (Aij )1 . Then
for 0 k n m, we have
Bn,k =
(1)nk
|Ak+i+1,k+j |i,j=0,...,nk1 .
Ak,k Ak+1,k+1 An,n
1
i
, we have
Since the Salie numbers are entries of
2i 2j
k + i + 1 nk1
s(n, k) = (1)nk
2i 2j + 2 0
1
i+1
, we have
and since the Faulhaber numbers are entries of
2i 2j + 1
f (n, k) = (1)
nk
k + i + 2 nk1
k!
.
(n + 1)! 2i 2j + 3 0
(12.4)
(12.5)
Although we can give a combinatorial interpretation of these determinants using Theorem 14, the
simplest interpretations are most easily derived directly from the paths.
For the Salie numbers, we define the points Pm = (2m, 2m) and Qm = (2m, m). Then it follows from (12.4) that (1)nk s(n, k) is the number of disjoint (n k)-paths from (Pk , Pk+1 , . . . , Pn1 ) to
(Qk+1 , Qk+2 , . . . , Qn ). Figure 9 illustrates a 3-path counted by |s(5, 2)|. We note that it is immediate from
Figure 9
this combinatorial interpretation that |s(n, 1)| = |s(n, 2)|. To represent these (n k)-paths most simply as
tableaux, we assign to the horizontal segment from (i, j) to (i + 1, j) the label i j + 1 and make the
labels on each path into a row of the tableau, shifting as usual. Thus the 3-path of Figure 9 becomes the
tableau
3
2
1
(12.6)
Converting the conditions on the paths into conditions on the tableau, we have
Theorem 31. |s(n, k)| = (1)nk s(n, k) is the number of row-strict tableaux of shape (n k + 1, n
k, . . . , 2) (n k 1, n k 2, . . . , 0) with positive integer entries in which the largest entry in row i is at
most n + 1 i.
We may convert the tableau into a sequence of integers by reading each row from left to right, starting
with the last row, so that (12.6) becomes 1 2 2 4 3 5. Checking the conditions on this sequence, we find that we
Plane Partitions
31
are counting sequences a1 a2 a2n2k of positive integers satisfying a2i1 < a2i , a2i a2i+1 , and a2i k + i
for each i. This combinatorial interpretation is closely related to the combinatorial interpretation of the
Genocchi numbers given by Dumont and Viennot [15]. See also Viennot [70].
Next we move on the Faulhaber numbers. Let Pm = (2m, 2m) as before, and let Rm = (2m + 1, m).
(n + 1)!
f (n, k) is the number of nonintersecting (n k)-paths from (Pk , Pk+1 , . . . , Pn1 ) to
Then (1)nk
k!
(Rk+1 , Rk+2 , . . . , Rn ). As in the case of the Salie numbers, we represent these (n k)-paths as tableaux,
Figure 10
so the 4-path of Figure 10 becomes
5
(12.7)
and we have
(n + 1)!
f (n, k) is the number of row-strict tableaux of shape (n k + 2, n k +
k!
1, . . . , 2) (n k 1, n k 2, . . . , 0) with positive integer entries in which the largest entry in row i is at
most n + 2 i.
Theorem 32. (1)nk
We can also represent these tableaux as sequences, so for example, (12.7) becomes 1 2 3 1 3 4 3 4 5 1 3 5,
and we have
(n + 1)!
f (n, k) is the number of sequences a1 a2 a3n3k of positive integers
Theorem 33. (1)nk
k!
satisfying a3i2 < a3i1 < a3i , a3i1 a3i+1 , a3i a3i+2 , and a3i k + i + 1 for all i.
We now derive some explicit formulas for the Salie numbers in terms of the Genocchi numbers, and for
the Faulhaber numbers in terms of the Bernoulli numbers.
We first consider the Salie numbers. We shall need the formula [55, Ex. 2, pp. 153154]
Now
cosh
so
1 + 4t
2
X
k=i
2k i
i
(t)k .
2k i
k
x
1 + 4t
x
2
2
x
1 1 + 4t
x
1 1 + 4t
x cosh sinh
x sinh
= cosh
2
2
2
2
x
1 + 4t = cosh
2
X
cosh 1 + 4t x2
2k 2j
x2j X
2j
=
(t)k
cosh x2
(2j)!
2k
2j
k
j=0
k=2j
tanh
X
x X x2j+1
2k 2j 1
2j + 1
(t)k .
2 j=0 (2j + 1)!
2k 2j 1
k
k=2j+1
X
x
xn
2x
=
x(1
tanh
)
=
.
G
n
ex + 1
2
n!
n=1
(12.8)
Plane Partitions
32
(Thus G2n+1 = 0 for n > 0 and G2n = 2(1 22n B2n ), where B2n is the Bernoulli number.) Then we have
x2n
= (1)k
s(n, k)
(2n)!
n=0
bk/2c
X
j=0
2k 2j x2j
2j
2k 2j
k
(2j)!
2k 2j 1 x2j
1
xX
x tanh
2 j=0 2k 2j 1
k
(2j)!
and thus
s(n, k) = (1)k
Since
2k2n
k
b(k1)/2c
2k 2n
2n
+
2k 2n
k
X
j=0
2k 2j 1
1
2k 2j 1
k
2n
G2n2j .
2j
s(n, k) = (1)
X
j=0
2k 2j 1 2n
1
G2n2j
2k 2j 1
k
2j
(12.9)
for n > k/2, and since s(n, k) = 0 for n < k, we also obtain the identity for Genocchi numbers
b(k1)/2c
X
j=0
2k 2j 1 2n
2k 2n
1
2n
G2n2j =
,
2k 2j 1
k
2j
2k 2n
k
n < k.
n1
n2
1 2n
G2n2 , n 2
s(n, 3) = 2G2n
3 2
2n
G2n2 , n 3
s(n, 4) = 5G2n +
2
s(n, 2) =
G2n ,
cosh 1 + 4t x2 cosh x2
x
1 1 + 4t
1
= t coth
cosh
x1
t sinh x2
2
2
1 1 + 4t
1
x
t sinh
2
2k 2j
2j
x X x2j1 X
(1)k tk1
= x coth
2 j=1 (2j)!
2k 2j
k
k=2j
X x2j+1
X
2k 2j 1
2j + 1
(1)k tk1
(2j
+
1)!
2k
2j
1
k
j=0
k=2j+1
X
x X x2j+1
2k 2j
1
= x coth
(1)k+1 tk
2 j=0 (2j + 1)!
2k 2j k + 1
k=2j+1
X x2j+1 X
2j + 1
2k 2j + 1
+
(1)k tk .
(2j
+
1)!
2k
2j
+
1
k
+
1
j=0
k=2j
Plane Partitions
Now since
33
X
x2n
x
,
B2n
x coth = 2
2
(2n)!
n=0
2k 2j
2n + 1
1
B2n2j
kj k+1
2j + 1
j=0
2k 2n + 1
2n + 1
k
.
+ (1)
2k 2n + 1
k+1
b(k1)/2c
f (n, k) = (1)k+1
f (n, k) = (1)
X
j=0
2k 2j
2n + 1
1
B2n2j
kj k+1
2j + 1
(12.10)
X
j=0
2k 2j
2n + 1
2k 2n + 1
1
2n + 1
B2n2j =
,
kj k+1
2j + 1
2k 2n + 1
k+1
n < k.
n1
n2
1 2n + 1
B2n2 ,
f (n, 3) = 5(2n + 1)B2n +
2
3
f (n, 2) = 2(2n + 1)B2n ,
.
n2
References
1. G. E. Andrews, The Theory of Partitions (Encyclopedia of Mathematics and its Applications, Vol. 2),
Addison-Wesley, Reading, 1976.
2. G. E. Andrews, Plane partitions (I): The MacMahon conjecture, Studies in Foundations and Combinatorics,
Adv. in Math. Supplementary Studies, Vol. 1, 1978, 131150.
3. G. E. Andrews, Plane partitions (II): The equivalence of the Bender-Knuth and MacMahon conjectures,
Pacific J. Math. 72 (1977), 283291.
4. G. E. Andrews, Plane partitions (III): The weak Macdonald conjecture, Inventiones Math. 53 (1979), 193
225.
5. W. N. Bailey, Generalized Hypergeometric Series, Hafner, 1972. Originally published by Cambridge University Press, 1935.
6. E. A. Bender and D. E. Knuth, Enumeration of plane partitions, J. Combin. Theory Ser. B 13 (1972), 4054.
7. A. Berele and A. Regev, Hook Young diagrams with applications to combinatorics and to representations of
Lie superalgebras, Adv. in Math. 64 (1987), 118175.
Plane Partitions
34
Plane Partitions
35
37. P. W. Kasteleyn, The statistics of dimers on a lattice I. The number of dimer arrangements on a quadratic
lattice, Physica 27 (1961), 12091225.
38. D. E. Knuth, Faulhaber polynomials, preprint.
39. G. Kreweras, Sur une classe de probl`emes de denombrement lies au treillis des partitions des entiers, Cahiers
du Bureau Universitaire de Recherche Operationelle, no. 6, Institut de Statistique des Universitees de Paris,
1965.
Plane Partitions
36