, , ,

Let

x_{0}, x_{1}, x_{2}, \dots

be a sequence of nonnegative numbers and with no internal zeros. By the latter we mean that there are no three indices

i < j < k

such that

x_{i}, x_{k} \neq 0

and

x_{j} = 0

. We say that the sequence is log-concave (LC) if

x_{i - 1} x_{i + 1} \leq x_{i}^{2}

for all

i > 0

. It is well known that the sequence

{x_{k}}

is log-concave if and only if

x_{i - 1} x_{j + 1} \leq x_{i} x_{j}

for all

j \geq i \geq 1

(see [1,Proposition2.5.1] for instance), or equivalently, all minors of order

2

of the infinite Toeplitz matrix

M = (x_{i - j})_{i, j \geq 0}

are nonnegative (where

x_{k} = 0

k < 0

). For this reason a log-concave sequence with no internal zeros is also called PF

_{2}

(the notation actually has a precisely motivation, see [1, 7] ). Log-concave sequences arise often in combinatorics, algebra, geometry, analysis, probability and statistics. There have been many attempts to develop techniques for the log-concavity problems. We refer the reader to Stanley's survey article [16] and Brenti's supplement [2] for details.

Let

{a (n, k)}_{0 \leq k \leq n}

be a triangular array of nonnegative numbers. Define two linear transformations of sequences by

\begin{matrix} z_{n} = \sum_{k = 0}^{n} a (n, k) x_{k}, n = 0, 1, 2, \dots \end{matrix}

(1.1)

and

\begin{matrix} z_{n} = \sum_{k = 0}^{n} a (n, k) x_{k} y_{n - k}, n = 0, 1, 2, \dots \end{matrix}

(1.2)

respectively. We say that the linear transformation ( 1.1 ) has the PLC property if it preserves the log-concavity of sequences, i.e., the log-concavity of

{x_{n}}

implies that of

{z_{n}}

. We say that the linear transformation ( 1.2 ) has the double PLC property if the log-concavity of

{x_{n}}

and

{y_{n}}

implies that of

{z_{n}}

. The corresponding triangle

{a (n, k)}

is also called PLC and double PLC respectively. Clearly, the double PLC property implies the PLC property.

It is well known that the ordinary convolution

\begin{matrix} z_{n} = \sum_{k = 0}^{n} x_{k} y_{n - k}, n = 0, 1, 2, \dots \end{matrix}

is double PLC, which can be obtained as a consequence of the fact that the product of

T P_{2}

matrices is

T P_{2}

(see Karlin [7,p.394] for instance). Using the same fact, Walkup can manage to prove that the binomial convolution

\begin{matrix} z_{n} = \sum_{k = 0}^{n} (\binom{n}{k}) x_{k} y_{n - k}, n = 0, 1, 2, \dots \end{matrix}

is double PLC ([17,Theorem1] ). A more general result is due to Liggett (see [10,Theorem3] or §3 of the present paper). However, there is no systematic study of linear transformations that are double PLC. The possible reason for this is that very few examples of such linear transformations are known. In the present paper we develop techniques to deal with the problems of finding these kind of linear transformations and apply these techniques to generate new log-concave sequences from existing ones.

When the triangle

{a (n, k)}

is PLC, the linear transformation ( 1.1 ) has to send any log-concave sequence

{x_{k}}

to a log-concave sequence

{z_{n}}

. So, by taking the special log-concave sequence

{x_{k}}

, we may obtain certain necessary conditions such that

{a (n, k)}

is PLC from the log-concavity of the associated sequence

{z_{n}}

Example 1.1. Let the triangle

{a (n, k)}

be PLC. Then for

r \in N

and

p > 0

(i) the column sequence ${a (n, r)}_{n \geq r}$ is log-concave;
(ii) the row-sum sequence $a (n) = \sum_{k = 0}^{n} a (n, k)$ is log-concave; and
(iii) the sequence $A_{r} (n; p) = \sum_{k = r}^{n} a (n, k) p^{k}$ is log-concave for $n \geq r$ .

We can view

A_{r} (n; p)

as a polynomial in

p

. By (iii), the polynomial

A_{r}^{2} (n; p) - A_{r} (n - 1; p) A_{r} (n + 1; p)

takes nonnegative values when

p > 0

, and so that its leading coefficient

a^{2} (n, n) - a (n - 1, n - 1) a (n + 1, n + 1)

has to be nonnegative. In other words, the diagonal sequence

{a (n, n)}_{n \geq 0}

is log-concave.

In order to state our sufficient conditions for

{a (n, k)}

to be PLC, we introduce some terminology and notation. Let

q

be an indeterminate and

{f_{n} (q)}_{n \geq 0}

a sequence of polynomials in

q

. We say that the sequence

{f_{n} (q)}_{n \geq 0}

q

-log-concave if for each

n \geq 1

f_{n}^{2} (q) - f_{n - 1} (q) f_{n + 1} (q)

has nonnegative coefficients as a polynomial in

q

. The concept of

q

-log-concavity was first suggested by Stanley (see [15,p.795] ). We refer the reader to [3, 6, 9, 14, 15] for further information about

q

-log-concavity. Now for

0 \leq r \leq n

, define the polynomial

\begin{matrix} A_{r} (n; q) = \sum_{k = r}^{n} a (n, k) q^{k} . \end{matrix}

We say that the triangle

{a (n, k)}

has the LC-positive property if for each

r \geq 0

, the sequence of polynomials

{A_{r} (n; q)}_{n \geq r}

q

-log-concave in

n

. Define the reciprocal triangle

{a^{*} (n, k)}

{a (n . k)}

a^{*} (n, k) = a (n, n - k), 0 \leq k \leq n .

We say that the triangle

{a (n, k)}

has the double LC-positive property if both

{a (n, k)}

and

{a^{*} (n, k)}

have the LC-positive property.

Example 1.2. Consider

a (n, k) \equiv 1

for

0 \leq k \leq n

. Then

A_{r} (n; q) = \sum_{k = r}^{n} q^{k}

for

0 \leq r \leq n

. It immediately follows that

A_{r}^{2} (n; q) - A_{r} (n - 1; q) A_{r} (n + 1; q) = q^{n + r},

and so that

{A_{r} (n; q)}

q

-log-concave in

n

. Thus the constant triangle

{a (n, k)}

is LC-positive and therefore double LC-positive since

a^{*} (n, k) = a (n, k)

Example 1.3. Consider

a (n, k) = (\binom{n}{k})

. Then

A_{r} (n; q) = \sum_{k = r}^{n} (\binom{n}{k}) q^{k}

. We have

A_{r} (n; q) = \sum_{k = r}^{n} [(\binom{n - 1}{k}) + (\binom{n - 1}{k - 1})] q^{k} = (q + 1) A_{r} (n - 1; q) + (\binom{n - 1}{r - 1}) q^{r} .

It follows that

\begin{matrix} A_{r}^{2} (n; q) - A_{r} (n - 1; q) A_{r} (n + 1; q) \end{matrix}

\begin{matrix} = & A_{r}^{2} (n; q) - A_{r} (n - 1; q) [(q + 1) A_{r} (n; q) + (\binom{n}{r - 1}) q^{r}] \end{matrix}

\begin{matrix} = & [A_{r} (n; q) - (q + 1) A_{r} (n - 1; q)] A_{r} (n; q) - (\binom{n}{r - 1}) q^{r} A_{r} (n - 1; q) \end{matrix}

\begin{matrix} = & (\binom{n - 1}{r - 1}) q^{r} A_{r} (n; q) - (\binom{n}{r - 1}) q^{r} A_{r} (n - 1; q) \end{matrix}

\begin{matrix} = & \sum_{k = r}^{n} [(\binom{n - 1}{r - 1}) (\binom{n}{k}) - (\binom{n}{r - 1}) (\binom{n - 1}{k})] q^{k + r} \end{matrix}

\begin{matrix} = & \sum_{k = r}^{n} [(\binom{n - 1}{r - 1}) (\binom{n - 1}{k - 1}) - (\binom{n - 1}{r - 2}) (\binom{n - 1}{k})] q^{k + r}, \end{matrix}

which has nonnegative coefficients by the log-concavity of the binomial coefficients.

Hence

{A_{r} (n; q)}

q

-log-concave in

n

. Thus the Pascal triangle

{a (n, k)}

is LC-positive and therefore double LC-positive since

a^{*} (n, k) = a (n, k)

The object of this paper is twofold. First, we show that LC-positive triangles are PLC and that double LC-positive triangles are double PLC. Second, we present some examples of PLC and double PLC triangles by showing the LC-positivity.

2 Theorems

In this section we will discuss the LC-positivity in detail and establish the relation between the (double) LC-positivity and the (double) PLC property. The following simple result will be used repeatedly in our discussion.

Lemma 2.1. Suppose that

a_{0}, \dots, a_{s}

and

X_{0}, \dots, X_{s}

satisfy the following conditions:

(a) $\sum_{k = r}^{s} a_{k} \geq 0$ for all $0 \leq r \leq s$ ;
(b) $0 \leq X_{0} \leq X_{1} \leq \dots \leq X_{s}$ .

Then

\sum_{k = 0}^{s} a_{k} X_{k} \geq X_{0} \sum_{k = 0}^{s} a_{k} \geq 0

Proof. The statement follows immediately from the Abel's partial summation $\sum_{k = 0}^{s} a_{k} X_{k} = (a_{0} + a_{1} + \dots + a_{s}) X_{0} + (a_{1} + \dots + a_{s}) (X_{1} - X_{0}) + \dots + a_{s} (X_{s} - X_{s - 1}) .$ □

We first consider the relation between the LC-positivity and the PLC property. Let

{a (n, k)}_{0 \leq k \leq n}

be a triangle of nonnegative numbers and

{x_{k}}_{k \geq 0}

be a log-concave sequence.

It is convenient to extend the definition of

x_{k}

and

a (n, k)

by setting

x_{k} = 0

for

k < 0

and

a (n, k) = 0

for

k < 0

k > n

. Let

{z_{n}}_{n \geq 0}

be the sequence defined by ( 1.1 ) and denote

Δ_{n} = z_{n}^{2} - z_{n - 1} z_{n + 1}

. Then we need that

Δ_{n} \geq 0

for each

n \geq 1

. Note that

\begin{matrix} Δ_{n} = {\sum_{k = 0}^{n} a (n, k) x_{k}}^{2} - {\sum_{k = 0}^{n - 1} a (n - 1, k) x_{k}} {\sum_{k = 0}^{n + 1} a (n + 1, k) x_{k}} \end{matrix}

(2.1)

is a quadratic form in

n + 2

variables

x_{0}, x_{1}, \dots, x_{n + 1}

. Such quadratic forms are not positive semidefinite in general. Hence the log-concavity of

{x_{k}}

is indispensable for our purposes.

To see this let us take

a (n, k) \equiv 1

for

0 \leq k \leq n

as an example. In this case we have

Δ_{2} = (x_{0} + x_{1})^{2} - x_{0} (x_{0} + x_{1} + x_{2}) = x_{1}^{2} + x_{0} x_{1} - x_{0} x_{2} .

Clearly,

Δ_{2}

may take negative values for nonnegative

x_{k}

's, but it must be nonnegative when

x_{0}, x_{1}, x_{2}

is log-concave.

To utilize the assumption for

{x_{k}}

, recall that

{x_{k}}

is log-concave if and only if

x_{i - 1} x_{j + 1} \leq x_{i} x_{j}

for

j \geq i \geq 1

. In other words, the

x_{i} x_{j}

's with the same “weight”

i + j

are comparable.

Collect together those terms in

Δ_{n}

with the same weight

t

and denote their sum by

S_{t}

. For

0 \leq k \leq ⌊ t / 2 ⌋

, let

a_{k} (n, t)

be the coefficient of the term

x_{k} x_{t - k}

Δ_{n}

. Then

Δ_{n} = \sum_{t = 0}^{2 n} S_{t}

and

S_{t} = \sum_{k = 0}^{⌊ t / 2 ⌋} a_{k} (n, t) x_{k} x_{t - k}

. Thus it suffices that

S_{t} \geq 0

for each

0 \leq t \leq 2 n

. Note that

x_{0} x_{t} \leq x_{1} x_{t - 1} \leq x_{2} x_{t - 2} \leq \dots

. Hence by Lemma 2.1 , it suffices that

\sum_{k = r}^{⌊ t / 2 ⌋} a_{k} (n, t) \geq 0

for each

0 \leq r \leq ⌊ t / 2 ⌋

. By ( 2.1 ),

a_{k} (n, t) = 2 a (n, k) a (n, t - k) - a (n - 1, k) a (n + 1, t - k) - a (n + 1, k) a (n - 1, t - k)

for

k < t / 2

, and

a_{k} (n, t) = a^{2} (n, k) - a (n - 1, k) a (n + 1, k)

for

t

even and

k = t / 2

. Denote

\begin{matrix} A_{r} (n, t) = \sum_{k = r}^{⌊ t / 2 ⌋} a_{k} (n, t) . \end{matrix}

(2.2)

Then it is not difficult to see that

A_{r} (n, t)

is precisely the coefficient of

q^{t}

in the polynomial

A_{r}^{2} (n; q) - A_{r} (n - 1; q) A_{r} (n + 1; q)

, i.e.,

\begin{matrix} A_{r}^{2} (n; q) - A_{r} (n - 1; q) A_{r} (n + 1; q) = \sum_{t = 2 r}^{2 n} A_{r} (n, t) q^{t} . \end{matrix}

(2.3)

So the following lemma is immediate.

Lemma 2.2. With the notation above, the triangle

{a (n, k)}_{0 \leq k \leq n}

is LC-positive if and only if

A_{r} (n, t) \geq 0

for all

2 r \leq t \leq 2 n

We can now conclude the first main result of this paper from the discussion above.

Theorem 2.3. The LC-positive triangles are PLC.

We next relate the double LC-positivity with the double PLC property.

Lemma 2.4. Let

{a (n, k)}_{0 \leq k \leq n}

be a triangle of nonnegative numbers and

{x_{k}}_{k \geq 0}

and

{y_{k}}_{k \geq 0}

be two log-concave sequences. Define three triangles

{b (n, k)}, {c (n, k)}

and

{d (n, k)}

b (n, k) = a (n, k) x_{k}, c (n, k) = a (n, k) y_{n - k}, d (n, k) = a (n, k) x_{k} y_{n - k} .

For

2 r \leq t \leq 2 n

, define

B_{r} (n, t), C_{r} (n, t)

and

D_{r} (n, t)

similar to

A_{r} (n, t)

in ( 2.2 ).

(i) If the triangle ${a (n, k)}$ is LC-positive, then the triangle ${b (n, k)}$ is LC-positive and $B_{r} (n, t) \geq A_{r} (n, t) x_{r} x_{t - r}$ .
(ii) If the triangle ${a (n, k)}$ is double LC-positive, then the triangle ${c (n, k)}$ is LC-positive and $C_{r} (n, t) \geq A_{r} (n, t) y_{n - t + r} y_{n - r}$ for $t \leq n + r$ .
(iii) If the triangle ${a (n, k)}$ is double LC-positive, then the triangle ${d (n, k)}$ is LC-positive and $D_{r} (n, t) \geq A_{r} (n, t) x_{r} x_{t - r} y_{n - t + r} y_{n - r}$ for $t \leq n + r$ .

Proof. Clearly, (iii) follows from (i) and (ii), so it suffices to prove (i) and (ii).

(i) Let

0 \leq t \leq 2 n

. It is easy to see by definition that

b_{k} (n, t) = a_{k} (n, t) x_{k} x_{t - k}

for

0 \leq k \leq ⌊ t / 2 ⌋

. Hence for

0 \leq r \leq ⌊ t / 2 ⌋

\begin{matrix} B_{r} (n, t) = \sum_{k = r}^{⌊ t / 2 ⌋} b_{k} (n, t) = \sum_{k = r}^{⌊ t / 2 ⌋} a_{k} (n, t) x_{k} x_{t - k} . \end{matrix}

Now

{a (n, k)}

is LC-positive and

x_{0} x_{t} \leq x_{1} x_{t - 1} \leq x_{2} x_{t - 2} \leq \dots

by the log-concavity of

{x_{k}}

. From Lemma 2.1 it follows that

B_{r} (n, t) \geq x_{r} x_{t - r} \sum_{k = r}^{⌊ t / 2 ⌋} a_{k} (n, t) = A_{r} (n, t) x_{r} x_{t - r} \geq 0 .

So the triangle

{b (n, k)}

is LC-positive. (ii) Let

2 r \leq t \leq 2 n

. We need to prove

C_{r} (n, t) \geq 0

. For brevity, we do this only for the case

t

odd since the same technique is still valid for the case

t

even.

Let

t = 2 s + 1

. For

0 \leq k \leq s

, denote

\begin{matrix} α_{k} & = & a (n, k) a (n, t - k), \end{matrix}

\begin{matrix} β_{k} & = & a (n - 1, k) a (n + 1, t - k), \end{matrix}

\begin{matrix} γ_{k} & = & a (n + 1, k) a (n - 1, t - k), \end{matrix}

and

Y_{k} = y_{n - t + k} y_{n - k}

. Then

a_{k} (n, t) = 2 α_{k} - β_{k} - γ_{k}

and

\begin{matrix} c_{k} (n, t) = 2 α_{k} Y_{k} - β_{k} Y_{k + 1} - γ_{k} Y_{k - 1} \end{matrix}

by definition. It follows that

\begin{matrix} C_{r} (n, t) & = & \sum_{k = r}^{s} (2 α_{k} Y_{k} - β_{k} Y_{k + 1} - γ_{k} Y_{k - 1}) \end{matrix}

\begin{matrix} = & \sum_{k = r}^{s} (2 α_{k} - β_{k - 1} - γ_{k + 1}) Y_{k} + β_{r - 1} Y_{r} - γ_{r} Y_{r - 1}, \end{matrix}

where we use the fact that

Y_{s + 1} = Y_{s}

and

γ_{s + 1} = β_{s}

. Note that

{Y_{k}}

is nondecreasing by the log-concavity of

{y_{k}}

and

\begin{matrix} 2 α_{k} - β_{k - 1} - γ_{k + 1} = 2 a^{*} (n, n - k) a^{*} (n, n - t + k) \end{matrix}

\begin{matrix} - a^{*} (n - 1, n - k) a^{*} (n + 1, n - t + k) - a^{*} (n + 1, n - k) a^{*} (n - 1, n - t + k) \end{matrix}

\begin{matrix} = & a_{n - t + k}^{*} (n, 2 n - t) . \end{matrix}

Hence by the LC-positivity of

{a^{*} (n, k)}

, we have

\begin{matrix} C_{r} (n, t) & = & \sum_{j = n - t + r}^{⌊ (2 n - t) / 2 ⌋} a_{j}^{*} (n, 2 n - t) Y_{j - n + t} + β_{r - 1} Y_{r} - γ_{r} Y_{r - 1} \end{matrix}

\begin{matrix} \geq & Y_{r} \sum_{j = n - t + r}^{⌊ (2 n - t) / 2 ⌋} a_{j}^{*} (n, 2 n - t) + β_{r - 1} Y_{r} - γ_{r} Y_{r - 1} \end{matrix}

\begin{matrix} = & Y_{r} \sum_{k = r}^{s} (2 α_{k} - β_{k - 1} - γ_{k + 1}) + β_{r - 1} Y_{r} - γ_{r} Y_{r - 1} \end{matrix}

\begin{matrix} = & Y_{r} \sum_{k = r}^{s} (2 α_{k} - β_{k} - γ_{k}) + γ_{r} (Y_{r} - Y_{r - 1}) \end{matrix}

\begin{matrix} = & A_{r} (n, t) Y_{r} + γ_{r} (Y_{r} - Y_{r - 1}) . \end{matrix}

(2.4)

Thus

C_{r} (n, t) \geq A_{r} (n, t) y_{n - t + r} y_{n - r} \geq 0

since

Y_{r} \geq Y_{r - 1}

, as desired. □

Now we present the second main result of this paper.

Theorem 2.5. The double LC-positive triangles are double PLC.

Proof. Let the triangle ${a (n, k)}$ be double LC-positive. Suppose that both ${x_{k}}$ and ${y_{k}}$ are log-concave. Then the triangle ${a (n, k) x_{k} y_{n - k}}$ is LC-positive by Lemma 2.4 (iii) and is therefore PLC by Theorem 2.3 . Thus the row-sum sequence $z_{n} = \sum_{k = 0}^{n} a (n, k) x_{k} y_{n - k}, n = 0, 1, 2, \dots$ is log-concave. In other words, the triangle ${a (n, k)}$ is double PLC. □

We can give some more practical conditions that imply the LC-positivity. We have seen that Lemma 2.1 plays a key role in the proof of the LC-positivity in Lemma 2.4 . Clearly, Condition (a) in Lemma 2.1 is implied by the following two conditions:

(a1) $a_{0}, a_{1}, \dots, a_{s}$ changes from nonpositive to nonnegative values;
(a2) $\sum_{k = 0}^{s} a_{k} \geq 0$ .

These two conditions are easier to check than Condition (a). For example, Condition (a1) can be obtained by showing that the sequence

{a_{k}}

is nondecreasing and eventually nonnegative. In this case the analytic tools are often effective. On the other hand, Condition (a2) is just the simplest one of inequalities appeared in Condition (a) and the methods of generating functions will be useful (see [23] for details). By Lemma 2.2 ,

{a (n, k)}

is LC-positive if and only if the inequality

\sum_{k = r}^{⌊ t / 2 ⌋} a_{k} (n, t) \geq 0

for all

2 r \leq t \leq 2 n

, so the following corollary is immediate.

Corollary 2.6. Suppose that the triangle

{a (n, k)}

satisfies the following two conditions:

(A) There exists an index $m = m (n, t)$ such that $a_{k} (n, t) < 0$ for $k < m$ and $a_{k} (n, t) \geq 0$ for $k \geq m$ ;
(B) The sequence ${A_{0} (n; q)}_{n \geq 0}$ is $q$ -log-concave.

Then the triangle

{a (n, k)}

is LC-positive and therefore PLC.

Corollary 2.7. Suppose that the triangle

{a (n, k)}

satisfies Condition (A) and (B) in Corollary 2.6 and

{a^{*} (n, k)}

satisfies Condition (A). Then

{a (n, k)}

is double LC-positive and therefore double PLC.

Proof. By Theorem 2.5 and Corollary 2.6 , it suffices to show that

{A_{0}^{*} (n; q)}

q

-log-concave.

We have

\begin{matrix} A_{0}^{*} (n; q) = \sum_{k = 0}^{n} a (n, n - k) q^{k} = \sum_{k = 0}^{n} a (n, k) q^{n - k} = q^{n} A_{0} (n; q^{- 1}) . \end{matrix}

It follows that

\begin{matrix} {A_{0}^{*}}^{2} (n; q) - A_{0}^{*} (n - 1; q) A_{0}^{*} (n + 1; q) \end{matrix}

\begin{matrix} = & q^{2 n} [A_{0}^{2} (n; q^{- 1}) - A_{0} (n - 1; q^{- 1}) A_{0} (n + 1; q^{- 1})], \end{matrix}

which has nonnegative coefficients by the

q

-log-concavity of

{A_{0} (n; q)}

, as desired. □

3 Applications

In this section we give some examples of PLC and double PLC triangles by showing their LC-positivity. We have shown that

a (n, k) \equiv 1

and

a (n, k) = (\binom{n}{k})

are double LC-positive in Example 1.2 and 1.3 respectively. So the following two propositions are immediate from Theorem 2.5 .

Proposition 3.1. If the sequences

{x_{n}}

and

{y_{n}}

are log-concave, then so is their ordinary convolution

z_{n} = \sum_{k = 0}^{n} x_{k} y_{n - k}, n = 0, 1, 2, \dots

Proposition 3.2. If the sequences

{x_{n}}

and

{y_{n}}

are log-concave, then so is their binomial convolution

z_{n} = \sum_{k = 0}^{n} (\binom{n}{k}) x_{k} y_{n - k}, n = 0, 1, 2, \dots

It is easy to extend Proposition 3.2 (by induction) to several log-concave sequences.

Corollary 3.3. If the sequences

{x_{k}^{(1)}}, {x_{k}^{(2)}}, \dots, {x_{k}^{(j)}}

are all log-concave, then so is the sequence

X_{n} = \sum (\binom{n}{k_{1}, k_{2}, \dots, k_{j}}) x_{k_{1}}^{(1)} x_{k_{2}}^{(2)} \dots x_{k_{j}}^{(j)}, n = 0, 1, 2, \dots,

where the sum is over all

k_{1}, \dots, k_{j}

such that

k_{1} + k_{2} \dots + k_{j} = n

In [20] , the triangle

a (n, k) = (\binom{a + n}{b + k})

for

0 \leq k \leq n

is shown to satisfy Condition (A) and (B), and is therefore LC-positive. The LC-positivity of

{a (n, k)}

can also be obtained by the same technique used in Example 1.2. Also, note that

a^{*} (n, k) = (\binom{a + n}{b + (n - k)}) = (\binom{a + n}{(a - b) + k}) .

Hence

a^{*} (n, k)

is also LC-positive. Thus the triangle

{a (n, k)}

is double LC-positive and therefore double PLC.

Proposition 3.4. Let

a, b

be two nonnegative integers and

a \geq b

. If the sequences

{x_{n}}

and

{y_{n}}

are log-concave, then so is the sequence

z_{n} = \sum_{k = 0}^{n} (\binom{a + n}{b + k}) x_{k} y_{n - k}, n = 0, 1, 2, \dots .

We can provide a unified setting for the above three propositions. Denote by

S

the set of sequences

{u_{k}}_{k \in Z}

of nonnegative numbers. Given two nonnegative numbers

λ, μ

, define the linear operator

L = L [λ, μ]

S

L (u_{k}) = λ u_{k} + μ u_{k - 1}, k \in Z .

For

n \geq 2

, define

L^{n} = L (L^{n - 1})

by induction. It is convenient to view

L^{0}

as the identity operator. Let the sequence

{u_{k}}_{k \in Z}

be log-concave. Then the sequence

{L (u_{k})}_{k \in Z}

is also log-concave since

\begin{matrix} [L (u_{k})]^{2} - L (u_{k - 1}) L (u_{k + 1}) \end{matrix}

\begin{matrix} = & (λ u_{k} + μ u_{k - 1})^{2} - (λ u_{k - 1} + μ u_{k - 2}) (λ u_{k + 1} + μ u_{k}) \end{matrix}

\begin{matrix} = & λ^{2} (u_{k}^{2} - u_{k - 1} u_{k + 1}) + λ μ (u_{k - 1} u_{k} - u_{k - 2} u_{k + 1}) + μ^{2} (u_{k - 1}^{2} - u_{k - 2} u_{k}) \end{matrix}

\begin{matrix} \geq & 0 . \end{matrix}

By induction, the sequence

{L^{n} (u_{k})}_{k \in Z}

is log-concave for each

n \geq 0

Proposition 3.5. Given two nonnegative numbers

λ, μ

and a log-concave sequence

{u_{k}}_{k \in Z}

, define

a (n, k) = L^{n} [λ, μ] (u_{k})

for

0 \leq k \leq n

. Then the triangle

{a (n, k)}

is double LC-positive and therefore double PLC.

Proof. Denote

a_{k} = L^{n - 1} [λ, μ] (u_{k})

for

k \in Z

. Then the sequence

{a_{k}}_{k \in Z}

is log-concave and

A_{r} (n - 1; q) = \sum_{k = r}^{n - 1} a_{k} q^{k}

. We have

\begin{matrix} A_{r} (n; q) & = & \sum_{k = r}^{n} L (a_{k}) [λ, μ] q^{k} \end{matrix}

\begin{matrix} = & \sum_{k = r}^{n} (λ a_{k} + μ a_{k - 1}) q^{k} \end{matrix}

\begin{matrix} = & λ \sum_{k = r}^{n} a_{k} q^{k} + μ \sum_{k = r}^{n} a_{k - 1} q^{k} \end{matrix}

\begin{matrix} = & (λ + μ q) A_{r} (n - 1; q) + λ a_{n} q^{n} + μ a_{r - 1} q^{r}, \end{matrix}

and similarly,

\begin{matrix} A_{r} (n + 1; q) = (λ + μ q) A_{r} (n; q) + λ (λ a_{n + 1} + μ a_{n}) q^{n + 1} + μ (λ a_{r - 1} + μ a_{r - 2}) q^{r} . \end{matrix}

It follows that

\begin{matrix} A_{r}^{2} (n; q) - A_{r} (n - 1; q) A_{r} (n + 1; q) \end{matrix}

\begin{matrix} = & A_{r}^{2} (n; q) - (λ + μ q) A_{r} (n - 1; q) A_{r} (n; q) \end{matrix}

\begin{matrix} - [λ (λ a_{n + 1} + μ a_{n}) q^{n + 1} + μ (λ a_{r - 1} + μ a_{r - 2}) q^{r}] A_{r} (n - 1; q) \end{matrix}

\begin{matrix} = & (λ a_{n} q^{n} + μ a_{r - 1} q^{r}) A_{r} (n; q) \end{matrix}

\begin{matrix} - [λ (λ a_{n + 1} + μ a_{n}) q^{n + 1} + μ (λ a_{r - 1} + μ a_{r - 2}) q^{r}] A_{r} (n - 1; q) \end{matrix}

\begin{matrix} = & λ \sum_{k = r}^{n} (λ a_{k} + μ a_{k - 1}) a_{n} q^{n + k} + μ \sum_{k = r}^{n} a_{r - 1} (λ a_{k} + μ a_{k - 1}) q^{k + r} \end{matrix}

\begin{matrix} - λ \sum_{k = r}^{n - 1} a_{k} (λ a_{n + 1} + μ a_{n}) q^{n + k + 1} - μ \sum_{k = r}^{n - 1} (λ a_{r - 1} + μ a_{r - 2}) a_{k} q^{k + r} \end{matrix}

\begin{matrix} = & λ^{2} \sum_{k = r + 1}^{n} (a_{k} a_{n} - a_{k - 1} a_{n + 1}) q^{n + k} + μ^{2} \sum_{k = r}^{n} (a_{r - 1} a_{k - 1} - a_{r - 2} a_{k}) q^{k + r} \end{matrix}

\begin{matrix} + (λ^{2} a_{r} + 2 λ μ a_{r - 1} + μ^{2} a_{r - 2}) a_{n} q^{n + r}, \end{matrix}

(3.1)

which has nonnegative coefficients by the log-concavity of

{a_{k}}

. Hence the triangle

{a (n, k)}

is LC-positive. On the other hand, let

u_{k}^{*} = u_{- k}

for

k \in Z

. Then the sequence

{u_{k}^{*}}_{k \in Z}

is log-concave and

a^{*} (n, k) = L^{n} [μ, λ] (u_{k}^{*})

. Thus the triangle

{a^{*} (n, k)}_{0 \leq k \leq n}

is also LC-positive, and the triangle

{a (n, k)}_{0 \leq k \leq n}

is therefore double LC-positive. □

Remark 3.6. Let the triangle

{a (n, k)}

be the same as Proposition 3.5 . Then by ( 2.3 ) and ( 3.1 ),

A_{r} (n, t) \geq (a_{r - 1} a_{t - r - 1} - a_{r - 2} a_{t - r}) μ^{2}

holds for

t \leq n + r

(The equality holds when

t < n + r

Remark 3.7. In Proposition 3.5 , taking

λ = μ = 1 / 2

and

u_{k} \equiv 1

leads to Proposition 3.1 ; and taking

λ = μ = 1

and

u_{k} = (\binom{a}{b + k})

leads to Proposition 3.4 . In particular, letting

a = b = 0

in the latter leads to Proposition 3.2 .

Proposition 3.8. Let

α, β

be two nonnegative numbers and

{a (n, k)}_{0 \leq k \leq n}

a triangle of nonnegative numbers. Suppose that each row of

{a (n, k)}

is log-concave and satisfies the recurrence relation

\begin{matrix} a (n, k) = α a (n + 1, k) + β a (n + 1, k + 1), k = 0, 1, \dots, n . \end{matrix}

Then the triangle

{a (n, k)}

is double LC-positive and therefore double PLC.

Proof. Denote

a (n + 1, k) = v_{k}

for

0 \leq k \leq n + 1

. Then the sequence

{v_{k}}

is log-concave and

A_{r} (n + 1; q) = \sum_{k = r}^{n + 1} v_{k} q^{k}

. By the recurrence relation we have

\begin{matrix} A_{r} (n; q) & = & \sum_{k = r}^{n} (α v_{k} + β v_{k + 1}) q^{k} \end{matrix}

\begin{matrix} = & α \sum_{k = r}^{n} v_{k} q^{k} + β \sum_{k = r + 1}^{n + 1} v_{k} q^{k - 1} \end{matrix}

\begin{matrix} = & (α + β q^{- 1}) A_{r} (n + 1; q) - α v_{n + 1} q^{n + 1} - β v_{r} q^{r - 1}, \end{matrix}

and similarly,

A_{r} (n - 1; q) = (α + β q^{- 1}) A_{r} (n; q) - α (α v_{n} + β v_{n + 1}) q^{n} - β (α v_{r} + β v_{r + 1}) q^{r - 1} .

It follows that

\begin{matrix} A_{r}^{2} (n; q) - A_{r} (n - 1; q) A_{r} (n + 1; q) \end{matrix}

\begin{matrix} = & [α (α v_{n} + β v_{n + 1}) q^{n} + β (α v_{r} + β v_{r + 1}) q^{r - 1}] A_{r} (n + 1; q) \end{matrix}

\begin{matrix} - (α v_{n + 1} q^{n + 1} + β v_{r} q^{r - 1}) A_{r} (n; q) \end{matrix}

\begin{matrix} = & α^{2} \sum_{k = r + 1}^{n} (v_{k} v_{n} - v_{k - 1} v_{n + 1}) q^{n + k} + β^{2} \sum_{k = r}^{n} (v_{r + 1} v_{k + 1} - v_{r} v_{k + 2}) q^{r + k} \end{matrix}

\begin{matrix} + v_{r} (α^{2} v_{n} + 2 α β v_{n + 1} + β^{2} v_{n + 2}) q^{n + r}, \end{matrix}

which has nonnegative coefficients by the log-concavity of

{v_{k}}

. So the triangle

{a (n, k)}

is LC-positive. Clearly, the reciprocal triangle

{a^{*} (n, k)}

possesses the same property as the triangle

{a (n, k)}

does. Hence

{a^{*} (n, k)}

is also LC-positive. Thus the triangle

{a (n, k)}

is double LC-positive. □

In Proposition 3.8 , taking

α = β = 1 / 2

and

a (n, k) \equiv 1

for

0 \leq k \leq n

leads to Proposition 3.1 ; and taking

α = β = 1

and

a (n, k) = (\binom{a - n}{b - k})

for

0 \leq k \leq n

leads to the following.

Proposition 3.9. Let

a, b \in N

and

a \geq b

. If the sequences

{x_{k}}

and

{y_{k}}

are log-concave, then so is the sequence

z_{n} = \sum_{k = 0}^{n} (\binom{a - n}{b - k}) x_{k} y_{n - k}, n = 0, 1, 2, \dots .

In what follows we generalize a result of Liggett. Let

{x_{k}}_{k \geq 0}

be a sequence of nonnegative numbers and with no internal zeros. Following Pemantle [12] and Liggett [10] , the sequence is ultra-log-concave of order

m

(ULC(

m

)) if

x_{k} = 0

for

k > m

and the sequence

{x_{k} / (\binom{m}{k})}_{k = 0}^{m}

is log-concave. The sequence

{x_{k}}_{k \geq 0}

is ULC(

\infty

) if the sequence

{k! x_{k}}_{k \geq 0}

is log-concave.

It is clear from definitions that ULC(

m

) implies ULC(

ℓ

) for

0 \leq m \leq ℓ \leq \infty

. The concept of ultra-log-concavity is closely related to negatively dependent Bernoulli sequences (see [12] for details). Pemantle speculates that ultra-log-concavity is characteristic of negative dependence in the exchangeable case. This leads to a conjecture that the ordinary convolution of a ULC(

m

) sequence and a ULC(

ℓ

) sequence is ULC(

m + ℓ

) where

m

and

ℓ

may be infinity ([12,Conjecture7] ). It is not difficult to see that the conjecture actually consists of two parts:

(i) The Pascal triangle ${(\binom{n}{k})}$ is double PLC;
(ii) The triangle ${(\binom{n}{k}) (\binom{a - n}{b - k})}$ is double PLC.

Liggett verified the conjecture by established a stronger result ([10,Theorem3] ).

Liggett Theorem. Given three log-concave sequences

{v_{k}}

{x_{k}}

and

{y_{k}}

, let

\begin{matrix} z_{n - 1} & = & \sum_{k = 0}^{n - 1} (\binom{n - 1}{k}) (v_{k} + 2 v_{k + 1} + v_{k + 2}) x_{k} y_{n - 1 - k}, \end{matrix}

\begin{matrix} z_{n} & = & \sum_{k = 0}^{n} (\binom{n}{k}) (v_{k} + v_{k + 1}) x_{k} y_{n - k}, \end{matrix}

\begin{matrix} z_{n + 1} & = & \sum_{k = 0}^{n + 1} (\binom{n + 1}{k}) v_{k} x_{k} y_{n + 1 - k} . \end{matrix}

Then

z_{n - 1} z_{n + 1} \leq z_{n}^{2}

Liggett's proof for his theorem, essentially using the double LC-positivity of the Pascal triangle, is not simple. To see his idea more clearly, we show the following more general result.

Proposition 3.10. Given four nonnegative numbers

α, β, λ, μ

and four log-concave sequences

{u_{k}}_{k \in Z}

{v_{k}}_{k \geq 0}

{x_{k}}_{k \geq 0}

and

{y_{k}}_{k \geq 0}

, let

a (n, k) = L^{n} [λ, μ] (u_{k})

and

\begin{matrix} z_{n - 1} & = & \sum_{k = 0}^{n - 1} a (n - 1, k) (α^{2} v_{k} + 2 α β v_{k + 1} + β^{2} v_{k + 2}) x_{k} y_{n - 1 - k}, \end{matrix}

\begin{matrix} z_{n} & = & \sum_{k = 0}^{n} a (n, k) (α v_{k} + β v_{k + 1}) x_{k} y_{n - k}, \end{matrix}

\begin{matrix} z_{n + 1} & = & \sum_{k = 0}^{n + 1} a (n + 1, k) v_{k} x_{k} y_{n + 1 - k} . \end{matrix}

Then

z_{n - 1} z_{n + 1} \leq z_{n}^{2}

Proof. Clearly,

z_{n}^{2} - z_{n - 1} z_{n + 1}

can be viewed as a quadratic form in

n + 2

variables

v_{0}

v_{1}

, . . . ,

v_{n + 1}

. Let

z_{n}^{2} - z_{n - 1} z_{n + 1} = \sum_{t = 0}^{2 (n + 1)} \sum_{k = 0}^{⌊ t / 2 ⌋} e_{k} (n, t) v_{k} v_{t - k} .

Then we need to show that

\sum_{k = r}^{⌊ t / 2 ⌋} e_{k} (n, t) \geq 0

for

2 r \leq t \leq 2 (n + 1)

. For brevity, we do this only for the case

t

odd. Let

t = 2 s + 1

Define

d (n, k) = a (n, k) x_{k} y_{n - k}

for

0 \leq k \leq n

. For convenience, set

x_{k} = y_{k} = 0

for

k < 0

and

d (n, k) = 0

for

k < 0

k > n

. The triangle

{a (n, k)}

is double LC-positive by Proposition 3.5 , and so is the triangle

{d (n, k)}

by Lemma 2.4 . Rewrite

\begin{matrix} z_{n - 1} & = & \sum_{k = 0}^{n + 1} [α^{2} d (n - 1, k) + 2 α β d (n - 1, k - 1) + β^{2} d (n - 1, k - 2)] v_{k}, \end{matrix}

\begin{matrix} z_{n} & = & \sum_{k = 0}^{n + 1} [α d (n, k) + β d (n, k - 1)] v_{k}, \end{matrix}

\begin{matrix} z_{n + 1} & = & \sum_{k = 0}^{n + 1} d (n + 1, k) v_{k} . \end{matrix}

Then

\begin{matrix} e_{k} (n, t) = 2 [α d (n, k) + β d (n, k - 1)] [α d (n, t - k) + β d (n, t - k - 1)] \end{matrix}

\begin{matrix} - [α^{2} d (n - 1, k) + 2 α β d (n - 1, k - 1) + β^{2} d (n - 1, k - 2)] d (n + 1, t - k) \end{matrix}

\begin{matrix} - d (n + 1, k) [α^{2} d (n - 1, t - k) + 2 α β d (n - 1, t - k - 1) + β^{2} d (n - 1, t - k - 2)] \end{matrix}

\begin{matrix} = α^{2} P_{k} + 2 α β Q_{k} + β^{2} R_{k}, \end{matrix}

where

\begin{matrix} P_{k} = 2 d (n, k) d (n, t - k) - d (n - 1, k) d (n + 1, t - k) - d (n + 1, k) d (n - 1, t - k), \end{matrix}

\begin{matrix} Q_{k} = d (n, k) d (n, t - k - 1) + d (n, k - 1) d (n, t - k) \end{matrix}

\begin{matrix} - d (n - 1, k - 1) d (n + 1, t - k) - d (n + 1, k) d (n - 1, t - k - 1), \end{matrix}

\begin{matrix} R_{k} = 2 d (n, k - 1) d (n, t - k - 1) \end{matrix}

\begin{matrix} - d (n - 1, k - 2) d (n + 1, t - k) - d (n + 1, k) d (n - 1, t - k - 2) . \end{matrix}

Thus it suffices to show that the inequality

\begin{matrix} α^{2} \sum_{k = r}^{s} P_{k} + 2 α β \sum_{k = r}^{s} Q_{k} + β^{2} \sum_{k = r}^{s} R_{k} \geq 0 . \end{matrix}

(3.2)

Note that

P_{k} = d_{k} (n, t)

and

R_{k} = d_{n - t + k + 1}^{*} (n, 2 n - t + 2)

. Hence both

\begin{matrix} \sum_{k = r}^{s} P_{k} = D_{r} (n, t) \end{matrix}

(3.3)

and

\begin{matrix} \sum_{k = r}^{s} R_{k} = D_{n - t + r + 1}^{*} (n, 2 n - t + 2) \end{matrix}

(3.4)

are nonnegative by the double LC-positivity of the triangle

{d (n, k)}

. Also,

\begin{matrix} \sum_{k = r}^{s} Q_{k} = \sum_{k = r}^{s} [d (n, k) d (n, t - k - 1) + d (n, k - 1) d (n, t - k) \end{matrix}

\begin{matrix} - d (n - 1, k - 1) d (n + 1, t - k) - d (n + 1, k) d (n - 1, t - k - 1)] \end{matrix}

\begin{matrix} = [d^{2} (n, s) - d (n - 1, s) d (n + 1, s)] + \sum_{k = r - 1}^{s - 1} [2 d (n, k) d (n, t - 1 - k) \end{matrix}

\begin{matrix} - d (n - 1, k) d (n + 1, t - 1 - k) - d (n + 1, k) d (n - 1, t - 1 - k)] \end{matrix}

\begin{matrix} + [d (n + 1, r - 1) d (n - 1, t - r) - d (n, r - 1) d (n, t - r)] \end{matrix}

\begin{matrix} = D_{r - 1} (n, t - 1) + [d (n + 1, r - 1) d (n - 1, t - r) - d (n, r - 1) d (n, t - r)] . \end{matrix}

(3.5)

Assume that

r = 0

t > n + r

. Then

\sum_{k = r}^{s} Q_{k} = D_{r - 1} (n, t - 1) \geq 0

. Thus the inequality ( 3.2 ) is trivial. So let

r \geq 1

and

t \leq n + r

If we can show that there exists a nonnegative number

E = E (n, t, r)

such that

\begin{matrix} {\begin{matrix} \sum_{k = r}^{s} P_{k} & \geq & μ^{2} & E x_{r} x_{t - r} y_{n - t + r} y_{n - r}, \\ \sum_{k = r}^{s} Q_{k} & \geq & - λ μ & E x_{r - 1} x_{t - r} y_{n - t + r} y_{n - r + 1}, \\ \sum_{k = r}^{s} R_{k} & \geq & λ^{2} & E x_{r - 1} x_{t - r - 1} y_{n - t + r + 1} y_{n - r + 1}, \end{matrix} \end{matrix}

(3.6)

then the arithmetic-geometric mean inequality and the log-concavity of

{x_{k}}

and

{y_{k}}

will give

α^{2} \sum_{k = r}^{s} P_{k} + β^{2} \sum_{k = r}^{s} R_{k} \geq - 2 α β \sum_{k = r}^{s} Q_{k},

the required inequality. So, to prove ( 3.2 ), it suffices to prove ( 3.6 ).

We use Lemma 2.4 to estimate the lower bounds for

\sum_{k = r}^{s} P_{k}

\sum_{k = r}^{s} Q_{k}

and

\sum_{k = r}^{s} R_{k}

From ( 3.3 ) and Lemma 2.4 (iii) it is immediate that

\begin{matrix} \sum_{k = r}^{s} P_{k} \geq A_{r} (n, t) x_{r} x_{t - r} y_{n - t + r} y_{n - r} . \end{matrix}

(3.7)

Similarly, note that

d^{*} (n, k) = a^{*} (n, k) y_{k} x_{n - k}

, it follows from ( 3.4 ) and Lemma 2.4 (iii) that

\begin{matrix} \sum_{k = r}^{s} R_{k} \geq A_{n - t + r + 1}^{*} (n, 2 n - t + 2) y_{n - t + r + 1} y_{n - r + 1} x_{t - r - 1} x_{r - 1} . \end{matrix}

(3.8)

To get an analogous lower bound for

\sum_{k = r}^{s} Q_{k}

, let

c (n, k) = a (n, k) y_{n - k}

. Then

d (n, k) = c (n, k) x_{k}

and so

D_{r - 1} (n, t - 1) \geq C_{r - 1} (n, t - 1) x_{r - 1} x_{t - r}

by Lemma 2.4 (i). However,

\begin{matrix} C_{r - 1} (n, t - 1) \geq A_{r - 1} (n, t - 1) y_{n - t + r} y_{n - r + 1} \end{matrix}

\begin{matrix} + a (n + 1, r - 1) a (n - 1, t - r) (y_{n - t + r} y_{n - r + 1} - y_{n - t + r - 1} y_{n - r + 2}) \end{matrix}

by the inequality ( 2.4 ). Hence we have by ( 3.5 )

\begin{matrix} \sum_{k = r}^{s} Q_{k} \geq [A_{r - 1} (n, t - 1) y_{n - t + r} y_{n - r + 1} \end{matrix}

\begin{matrix} + a (n + 1, r - 1) a (n - 1, t - r) (y_{n - t + r} y_{n - r + 1} - y_{n - t + r - 1} y_{n - r + 2})] x_{r - 1} x_{t - r} \end{matrix}

\begin{matrix} + [a (n + 1, r - 1) x_{r - 1} y_{n - r + 2} a (n - 1, t - r) x_{t - r} y_{n - t + r - 1} \end{matrix}

\begin{matrix} - a (n, r - 1) x_{r - 1} y_{n - r + 1} a (n, t - r) x_{t - r} y_{n - t + r}] \end{matrix}

\begin{matrix} = Q x_{r - 1} x_{t - r} y_{n - t + r} y_{n - r + 1}, \end{matrix}

(3.9)

where

\begin{matrix} Q = A_{r - 1} (n, t - 1) + a (n + 1, r - 1) a (n - 1, t - r) - a (n, r - 1) a (n, t - r) . \end{matrix}

(3.10)

It remains to show that three coefficients

A_{r} (n, t)

A_{n - t + r + 1}^{*} (n, 2 n - t + 2)

and

Q

in inequalities ( 3.7 ), ( 3.8 ) and ( 3.9 ) have the lower bounds of the forms in ( 3.6 ). We do this using Remark 3.6 .

Denote

a_{k} = L^{n - 1} [λ, μ] (u_{k})

. It follows from Remark 3.6 that

A_{r} (n, t) \geq (a_{r - 1} a_{t - r - 1} - a_{r - 2} a_{t - r}) μ^{2}

and that

\begin{matrix} Q & \geq & (a_{r - 2} a_{t - r - 1} - a_{r - 3} a_{t - r}) μ^{2} + (λ^{2} a_{r - 1} + 2 λ μ a_{r - 2} + μ^{2} a_{r - 3}) a_{t - r} \end{matrix}

\begin{matrix} - (λ a_{r - 1} + μ a_{r - 2}) (λ a_{t - r} + μ a_{t - r - 1}) \end{matrix}

\begin{matrix} = & - (a_{r - 1} a_{t - r - 1} - a_{r - 2} a_{t - r}) λ μ \end{matrix}

by ( 3.10 ). Also, note that

a^{*} (n, k) = L^{n} [μ, λ] (u_{- k})

. Again by Remark 3.6 ,

\begin{matrix} A_{n - t + r + 1}^{*} (n, 2 n - t + 2) & \geq & [a^{*} (n - 1, n - t + r) a^{*} (n - 1, n - r) \end{matrix}

\begin{matrix} - a^{*} (n - 1, n - t + r - 1) a^{*} (n - 1, n - r + 1)] λ^{2} \end{matrix}

\begin{matrix} = & (a_{r - 1} a_{t - r - 1} - a_{r - 2} a_{t - r}) λ^{2} . \end{matrix}

Finally, recall that the sequence

{a_{k}}_{k \in Z}

is log-concave, so for

r \leq ⌊ t / 2 ⌋

E = a_{r - 1} a_{t - r - 1} - a_{r - 2} a_{t - r} \geq 0,

as required. This completes our proof. □

Remark 3.11. Let

{a (n, k)}

and

{a^{'} (n, k)}

be the double LC-positive triangles in Proposition 3.5 and 3.8 respectively. Although the triangle

{a (n, k) a^{'} (n, k)}

is not double LC-positive in general, it is double PLC from Proposition 3.10 .

References

F. Brenti, Unimodal, log-concave, and Pólya frequency sequences in combinatorics, Mem. Amer. Math. Soc. no. 413 (1989). MR 0963833 (90d:05014)
F. Brenti, Log-concave and unimodal sequences in algebra, combinatorics, and geometry: an update, Contemp. Math. 178 (1994), 71-89. MR 1310575 (95j:05026)
L. M. Butler, The $q$ -log-concavity of $q$ -binomial coefficients, J. Combin. Theory Ser. A 54 (1990), 54–63. MR 1051778 (91c:05023)
H. Davenport and G. Pólya, On the product of two power series, Canadian J. Math. 1 (1949), 1–5. MR 0027306 (10,286b)
R. Ehrenborg and E. Steingrímsson, The excedance set of a permutation, Adv. Appl. Math. 24 (2000), 284–299. MR 1751147 (2001e:05003)
C. Krattenthaler, On the $q$ -log-concavity of Gaussian binomial coefficients, Monatsh. Math. 107 (1989), 333–339. MR 1012464 (90j:05027)
S. Karlin, Total Positivity, Vol.I, Stanford University Press, 1968. MR 0230102 (37 #5667)
D. C. Kurtz, A note on concavity properties of triangular arrays of numbers, J. Combin. Theory Ser. A 13 (1972), 135-139. MR 0304296 (46 #3431)
P. Leroux, Reduced matrices and $q$ -log-concavity properties of $q$ -Stirling numbers, J. Combin. Theory Ser. A 54 (1990), 64–84. MR 1051779 (91c:05024)
T. M. Liggett, Ultra logconcave sequence and negative dependence, J. Combin. Theory Ser. A 79 (1997), 315-325. MR 1462561 (98j:60018)
K. V. Menon, On the convolution of logarithmically concave sequences, Proc. Amer. Math. Soc. 23 (1969), 439–441. MR 0246012 (39 #7318)
R. Pemantle, Towards a theory of negative dependence, J. Math. Phys. 41 (2000), 1371–1390. MR 1757964 (2001g:62039)
B. E. Sagan, Inductive and injective proofs of log concavity results, Discrete Math. 68 (1988), 281-292. MR 0926131 (89b:05009)
B. E. Sagan, Inductive proofs of $q$ -log concavity, Discrete Math. 99 (1992), 289–306. MR 1158792 (93e:05004)
B. E. Sagan, Log concave sequences of symmetric functions and analogs of the Jacobi-Trudi determinants, Trans. Amer. Math. Soc. 329 (1992), 795–811. MR 1066448 (92e:05121)
R. P. Stanley, Log-concave and unimodal sequences in algebra, combinatorics, and geometry, Ann. New York Acad. Sci. 576 (1989), 500–534. MR 1110850 (92e:05124)
D. W. Walkup, Pólya sequences, binomial convolution and the union of random sets, J. Appl. Probability 13 (1976), 76–85. MR 0494391 (58 #13258)
Y. Wang, A simple proof of a conjecture of Simion, J. Combin. Theory Ser. A 100 (2002), 399–402. MR 1940344 (2003i:05016)
Y. Wang, Proof of a conjecture of Ehrenborg and Steingrímsson on excedance statistic, European J. Combin. 23 (2002), 355–365. MR 1908656 (2004c:05016)
Y. Wang, Linear transformations preserving log-concavity, Linear Algebra Appl. 359 (2003), 162–167. MR 1948441 (2003m:15009)
Y. Wang and Y.-N. Yeh, Polynomials with real zeros and Pólya frequency sequences, J. Combin. Theory Ser. A 109 (2005), 63–74. MR 2110198
Y. Wang and Y.-N. Yeh, Proof of a conjecture on unimodality, European J. Combin. 26 (2005), 617–627.
H. S. Wilf, Generatingfunctionology, 2nd ed., Academic Press, Boston, 1994. MR 1277813 (95a:05002)

Department of Applied Mathematics, Dalian University of Technology, Dalian 116024, China Current address : Institute of Mathematics, Academia Sinica, Taipei 11529, Taiwan E-mail address : wangyi@dlut.edu.cn Institute of Mathematics, Academia Sinica, Taipei 11529, Taiwan E-mail address : mayeh@math.sinica.edu.tw

Yi Wang

Yeong-Nan Yeh