betaFeedback

WorksMathematics

Log-concavity for alternating Eulerian polynomials

Abstract

Alternating Eulerian polynomials enumerate permutations by the alternating descent statistic introduced by Chebikin. We prove that their coefficient sequences are log-concave, verifying a conjecture of Lin, Ma, Wang and Wang. Our proof combines a factorization involving reciprocal quadratic factors, estimates for ordinary Eulerian zeros, and a log-concavity criterion for products of such factors. The required finite verifications use exact integer arithmetic. As applications, we obtain log-concavity for the distribution of 33-descents, a local limit theorem for alternating descents, and bounds for the largest coefficient. We also give a Lorentzian reformulation of the main theorem.

Introduction

Let Sn\mathcal{S}_n denote the set of all permutations of [n]={1,2,…,n}[n] = \{1,2,\ldots,n\}. For a permutation π=π1π2⋯πn∈Sn\pi= \pi_1\pi_2\cdots\pi_n \in\mathcal{S}_n, a descent is an index i∈[n−1]i \in[n-1] such that πi>πi+1\pi_i > \pi_{i+1}. As a variation of the classical descent statistic, Chebikin [5] introduced the alternating descent statistic. An index i∈[n−1]i \in[n-1] is an alternating descent of π\pi if πi>πi+1\pi_i > \pi_{i+1} when ii is odd, or πi<πi+1\pi_i < \pi_{i+1} when ii is even. Let altdes⁡(π)\operatorname{altdes}(\pi) denote the number of alternating descents of π\pi.

The alternating Eulerian polynomial is defined by

A^n(t)=∑π∈Sntaltdes⁡(π)=∑k=0n−1an,ktk,\widehat{A}_n(t) = \sum_{\pi\in\mathcal{S}_n} t^{\operatorname{altdes}(\pi)} = \sum_{k=0}^{n-1} a_{n,k}t^k,

where an,ka_{n,k} counts permutations in Sn\mathcal{S}_n with exactly kk alternating descents, and A^n(1)=n!\widehat{A}_n(1) = n!. The first few alternating Eulerian polynomials are A^1(t)=1\widehat{A}_1(t) = 1, A^2(t)=1+t\widehat{A}_2(t) = 1+t, A^3(t)=2+2t+2t2\widehat{A}_3(t) = 2+2t+2t^2.

A polynomial f(t)=∑k=0naktkf(t) = \sum_{k=0}^n a_k t^k is called palindromic if ak=an−ka_k = a_{n-k} for 0≤k≤n0 \leq k \leq n. Replacing each entry πi\pi_i by n+1−πin+1-\pi_i reverses every comparison and changes the number of alternating descents from kk to n−1−kn-1-k. Consequently, A^n(t)\widehat{A}_n(t) is palindromic of degree n−1n-1; equivalently,

an,k=an,n−1−k.(1)a_{n,k} = a_{n,n-1-k}. \tag*{(1)}

Alternating Eulerian polynomials are related to tangent derivative polynomials, and their zeros lie on the unit circle [6, 12]. Lin, Ma, Wang and Wang [10] studied their sign-alternating γ\gamma-expansions, and Zhu [17] proved coefficientwise 2-log-convexity as the order varies. Here we study log-concavity of the coefficients for each fixed order. Further studies include type BB analogues and grammatical descriptions [11], combinatorial formulas and parity refinements [13, 15], and cdcd-indices [14]. Heerten, Sambale and Thäle [8] studied probabilistic limits for alternating descents.

Let a0,a1,…,ana_0,a_1,\ldots,a_n be a sequence of nonnegative real numbers. The sequence is said to be unimodal if there exists an index 0≤m≤n0 \le m \le n such that a0≤⋯≤am−1≤am≥am+1≥⋯≥ana_0 \le\cdots\le a_{m-1} \le a_m \ge a_{m+1} \ge\cdots\ge a_n. The sequence is said to be log-concave if ak2≥ak−1ak+1a_k^2 \ge a_{k-1}a_{k+1} for k=1,…,n−1k=1,\ldots,n-1. For a polynomial f(t)=∑k=0naktkf(t)=\sum_{k=0}^{n}a_k t^k, define

Δk(f)=ak2−ak−1ak+1,\Delta_k(f)=a_k^2-a_{k-1}a_{k+1},

where coefficients outside the range 0≤k≤n0 \le k \le n are taken to be zero. Thus, the coefficient sequence of f(t)f(t) is log-concave if and only if Δk(f)≥0\Delta_k(f) \ge0 for 1≤k≤n−11 \le k \le n-1. A sequence (ak)k=0n(a_k)_{k=0}^n has no internal zeros if there are no indices 0≤i<j<k≤n0 \le i<j<k \le n such that ai≠0a_i \ne0, aj=0a_j=0, and ak≠0a_k \ne0. A nonnegative log-concave sequence with no internal zeros is unimodal (see, e.g., Brenti [4]).

We use two standard facts. The convolution property below follows from Brenti [4 Theorem 2.2.2 and Proposition 2.5.1]. The mixed coefficient inequality is an adjacent-index case of the inequalities of Ellard and Šmigoc [7 Theorem 2.2 and Remark 1].

Lemma 1.1 ([4 Theorem 2.2.2 and Proposition 2.5.1]). Let f(t)f(t) and g(t)g(t) be two nonzero polynomials with nonnegative log-concave coefficient sequences having no internal zeros. Then the coefficient sequence of f(t)g(t)f(t)g(t) is also nonnegative, log-concave, and has no internal zeros.

Lemma 1.2 ([7 Theorem 2.2 and Remark 1]). Let

q(t)=∏i=1m(1+cit+t2)=∑kbktk,ci≥0.q(t)=\prod_{i=1}^{m}(1+c_i t+t^2)=\sum_k b_k t^k,\qquad c_i\ge0.

Then

bkbk−1≥bk+1bk−2(k∈Z),(2)b_kb_{k-1}\ge b_{k+1}b_{k-2}\qquad(k\in\mathbb{Z}), \tag*{(2)}

where coefficients outside the range of q(t)q(t) are taken to be zero.

Since ci≥0c_i\ge0, factoring the quadratics into linear factors with constant term 1 gives a self-conjugate list of parameters in the closed right half-plane. The coefficients bkb_k are the corresponding elementary symmetric functions, so (1.2) follows from the adjacent-index inequalities in the cited results.

Lin, Ma, Wang and Wang [10 Theorem 1.1 and Conjecture 5.1] proved the unimodality of the alternating Eulerian polynomials and conjectured that their coefficient sequences are log-concave for all n≥1n\ge1. We prove this conjecture in the following theorem.

Theorem 1.3. For every n≥1n\ge1, the coefficient sequence of the alternating Eulerian polynomial A^n(t)\widehat{A}_n(t) is log-concave.

The rest of this paper is organized as follows. In Section 2, we establish the factorization of alternating Eulerian polynomials and a lower bound for the number of factors with parameters exceeding 3/23/2. In Section 3, we prove an endpoint reduction criterion and a smoothing lemma. In Section 4, we establish a block criterion and complete the proof of Theorem 1.3. Finally, Section 5 gives applications to the distribution of 3-descents and local limits, bounds for the largest coefficient, and a Lorentzian reformulation of the main theorem.

Factorization and strong-factor estimates

The reciprocal-quadratic factorization follows from the zero results of Ma and Yeh [12]. We identify its parameters in terms of ordinary Eulerian zeros, which allows us to estimate the number of strong factors needed for the block criterion in § 4.

Let

An(z)=∑π∈Snzdes⁡(π)A_n(z)=\sum_{\pi\in\mathfrak{S}_n}z^{\operatorname{des}(\pi)}

be the ordinary Eulerian polynomial, where des⁡(π)\operatorname{des}(\pi) denotes the number of descents. It is well known that all zeros of An(z)A_n(z) are simple and negative for n≥2n\ge2; see [6]. The complement map also gives the palindromic identity

An(z)=zn−1An(z−1).(3)A_n(z)=z^{n-1}A_n\left(z^{-1}\right). \tag*{(3)}

Hence the roots other than −1-1 occur in reciprocal pairs, with one root less than −1-1 and the other in (−1,0)(-1,0). The root −1-1 occurs precisely when nn is even. We now use these properties to obtain a factorization of the alternating Eulerian polynomials.

Proposition 2.1. For n≥1n\ge1, let m=⌊(n−1)/2⌋m=\lfloor(n-1)/2\rfloor. There exist distinct numbers v1,…,vm>0v_1,\ldots,v_m>0 such that An(−evi)=0A_n(-e^{v_i})=0 and

A^n(t)=an,0(1+t)n−1−2m∏i=1m(1+cit+t2),ci=2sech⁡vi∈(0,2),(4)\widehat{A}_n(t)=a_{n,0}(1+t)^{n-1-2m}\prod_{i=1}^{m}(1+c_i t+t^2),\qquad c_i=2\operatorname{sech}v_i\in(0,2), \tag*{(4)}

where an,0>0a_{n,0}>0.

Proof. The cases n=1,2n=1,2 follow from A^1(t)=1\widehat{A}_1(t)=1 and A^2(t)=1+t\widehat{A}_2(t)=1+t. Suppose that n≥3n\ge3. Let TnT_n be the tangent derivative polynomial, defined by

dndxntan⁡x=Tn(tan⁡x).\frac{\mathrm{d}^n}{\mathrm{d}x^n}\tan x=T_n(\tan x).

Ma and Yeh [12] established the identity

2n(1+t2)A^n(t)=(1−t)n+1Tn(1+t1−t).(5)2^n(1+t^2)\widehat{A}_n(t)=(1-t)^{n+1}T_n\left(\frac{1+t}{1-t}\right). \tag*{(5)}

We first relate the roots of TnT_n to those of the ordinary Eulerian polynomial AnA_n. By the classical Eulerian generating-function identity [16 Proposition 1.4.4], after normalization, we have,

(zddz)n11−z=zAn(z)(1−z)n+1.(6)\left(z\frac{\mathrm{d}}{\mathrm{d}z}\right)^n\frac{1}{1-z}=\frac{zA_n(z)}{(1-z)^{n+1}}. \tag*{(6)}

Consider a pair of reciprocal roots −α-\alpha and −α−1-\alpha^{-1} of An(z)A_n(z), where α>1\alpha>1, and put y=(α−1)/(α+1)y=(\alpha-1)/(\alpha+1). Since tan⁡(i2log⁡α)=iy\tan\left(\frac{i}{2}\log\alpha\right)=iy, the chain rule gives

(i2)nTn(iy)=(αddα)n(iy).\left(\frac{i}{2}\right)^nT_n(iy)=\left(\alpha\frac{\mathrm{d}}{\mathrm{d}\alpha}\right)^n(iy).

Moreover, since y=1−2/(1+α)y=1-2/(1+\alpha), substituting z=−αz=-\alpha into (6) yields

Tn(iy)=−i1−n2n+1(αddα)n11+α=i1−n2n+1αAn(−α)(1+α)n+1.(7)T_n(iy)=-i^{1-n}2^{n+1}\left(\alpha\frac{\mathrm{d}}{\mathrm{d}\alpha}\right)^n\frac{1}{1+\alpha} =i^{1-n}2^{n+1}\frac{\alpha A_n(-\alpha)}{(1+\alpha)^{n+1}}. \tag*{(7)}

Thus An(−α)=0A_n(-\alpha)=0 implies Tn(iy)=0T_n(iy)=0.

For the reciprocal root −α−1-\alpha^{-1}, the corresponding value is (α−1−1)/(α−1+1)=−y(\alpha^{-1}-1)/(\alpha^{-1}+1)=-y. Applying the same identity with α\alpha replaced by α−1\alpha^{-1} gives Tn(−iy)=0T_n(-iy)=0.

Next, solving (1+t)/(1−t)=iy(1+t)/(1-t)=iy gives ty:=(iy−1)/(iy+1)t_y:=(iy-1)/(iy+1). Since α>1\alpha>1, we have 0<y<10<y<1. Consequently,

∣ty∣=1,1+ty2=2(1−y2)(1+iy)2≠0.|t_y|=1,\qquad1+t_y^2=\frac{2(1-y^2)}{(1+iy)^2}\ne0.

As ty≠1t_y\ne1, equation (5) implies that

A^n(ty)=0.\widehat{A}_n(t_y)=0.

The same argument applied to −y-y shows that A^n(t−y)=0\widehat{A}_n(t_{-y})=0.

For each reciprocal pair, write α=evi\alpha=e^{v_i} with vi>0v_i>0. The corresponding roots tyt_y and t−yt_{-y} satisfy t−y=ty‾t_{-y}=\overline{t_y}, tyt−y=1t_yt_{-y}=1. Furthermore,

ty+t−y=2(y2−1)1+y2=−4αα2+1=−2sech⁡vi.t_y+t_{-y}=\frac{2(y^2-1)}{1+y^2}=-\frac{4\alpha}{\alpha^2+1}=-2\operatorname{sech}v_i.

Therefore, the corresponding quadratic factor is

(t−ty)(t−t−y)=1+2sech⁡vi t+t2.(t-t_y)(t-t_{-y})=1+2\operatorname{sech}v_i\,t+t^2.

When nn is even, the additional root −1-1 of AnA_n corresponds to y=0y=0. Since Tn(0)=0T_n(0)=0 for even nn, equation (5) gives A^n(−1)=0\widehat{A}_n(-1)=0. This contributes the factor 1+t1+t.

Since AnA_n has n−1n-1 distinct negative roots and the transformations above are injective, we have obtained all n−1n-1 roots of A^n\widehat{A}_n. Comparing constant terms gives (4). Finally, vi>0v_i>0 implies 0<ci=2sech⁡vi<20<c_i=2\operatorname{sech}v_i<2, completing the proof. □\square

A quadratic factor in (4) is called strong if ci>3/2c_i>3/2. Since sech⁡v\operatorname{sech}v is decreasing for v>0v>0, we estimate the number of strong factors by counting the zeros of An(−ev)A_n(-e^v) in a suitable interval.

Lemma 2.2. For n≥3n\ge3, at least ⌊(n+1)/13⌋−1\lfloor(n+1)/13\rfloor-1 of the quadratic factors in (4) are strong.

Proof. We count zeros of An(−ev)A_n(-e^v) in 0<v<7/90<v<7/9 and then show that every such zero gives ci>3/2c_i>3/2.

Step 1. Expressing the Eulerian polynomial as a sum over poles. Logarithmically differentiating the classical product

cosh⁡v2=∏j=0∞(1+v2(2j+1)2π2)\cosh\frac{v}{2}=\prod_{j=0}^{\infty}\left(1+\frac{v^2}{(2j+1)^2\pi^2}\right)

and using the identity (1+ev)−1=12−12ddvlog⁡cosh⁡(v/2)(1+e^v)^{-1}=\frac{1}{2}-\frac{1}{2}\frac{d}{dv}\log\cosh(v/2) give

11+ev=12−∑j=0∞(1v+(2j+1)πi+1v−(2j+1)πi).(8)\frac{1}{1+e^v}=\frac{1}{2}-\sum_{j=0}^{\infty}\left(\frac{1}{v+(2j+1)\pi i}+\frac{1}{v-(2j+1)\pi i}\right). \tag*{(8)}

The paired terms equal

2vv2+(2j+1)2π2.\frac{2v}{v^2+(2j+1)^2\pi^2}.

The series and its derivatives of every positive order converge locally uniformly on the real axis. Thus, termwise differentiation is justified. Define

Sn(v)=∑j∈Z1(v+(2j+1)πi)n+1.S_n(v)=\sum_{j\in\mathbb{Z}}\frac{1}{(v+(2j+1)\pi i)^{n+1}}.

Differentiating (2.6) nn times gives

dndvn11+ev=(−1)n+1n!Sn(v).\frac{d^n}{dv^n}\frac{1}{1+e^v}=(-1)^{n+1}n!S_n(v).

On the other hand, substituting r=evr=e^v into (2.4) yields

dndvn11+ev=−evAn(−ev)(1+ev)n+1.\frac{d^n}{dv^n}\frac{1}{1+e^v}=-\frac{e^v A_n(-e^v)}{(1+e^v)^{n+1}}.

Comparing these two identities, we obtain

An(−ev)=(−1)nn!e−v(1+ev)n+1Sn(v).(9)A_n(-e^v)=(-1)^n n!e^{-v}(1+e^v)^{n+1}S_n(v). \tag*{(9)}

Since the factor multiplying Sn(v)S_n(v) is nonzero for real vv, the two functions An(−ev)A_n(-e^v) and Sn(v)S_n(v) have the same real zeros.

Step 2. Estimating the contribution of the remaining poles. For 0≤v≤7/90\le v\le7/9, let θ(v)=π/2−arctan⁡(v/π)\theta(v)=\pi/2-\arctan(v/\pi). The two terms in Sn(v)S_n(v) corresponding to ±πi\pm\pi i give

1(v+πi)n+1+1(v−πi)n+1=2cos⁡((n+1)θ(v))(v2+π2)(n+1)/2.\frac{1}{(v+\pi i)^{n+1}}+\frac{1}{(v-\pi i)^{n+1}}=\frac{2\cos((n+1)\theta(v))}{(v^2+\pi^2)^{(n+1)/2}}.

Consequently,

∣(v2+π2)(n+1)/22Sn(v)−cos⁡((n+1)θ(v))∣≤∑j=1∞(v2+π2v2+(2j+1)2π2)n+1.\left|\frac{(v^2+\pi^2)^{(n+1)/2}}{2}S_n(v)-\cos((n+1)\theta(v))\right| \le\sum_{j=1}^{\infty}\left(\frac{\sqrt{v^2+\pi^2}}{\sqrt{v^2+(2j+1)^2\pi^2}}\right)^{n+1}.

Since 0≤v≤7/90\le v\le7/9 and π>3\pi>3,

v2+π2v2+(2j+1)2π2≤1+(7/27)22j+1<2120(2j+1)<1.\frac{\sqrt{v^2+\pi^2}}{\sqrt{v^2+(2j+1)^2\pi^2}} \le\frac{\sqrt{1+(7/27)^2}}{2j+1}<\frac{21}{20(2j+1)}<1.

As n+1≥4n+1\ge4, we have

∑j=1∞(v2+π2v2+(2j+1)2π2)n+1<(2120)4∑j=1∞1(2j+1)4≤(2120)4(181+∫1∞dx(2x+1)4)=(2120)4154<140.\sum_{j=1}^{\infty}\left(\frac{\sqrt{v^2+\pi^2}}{\sqrt{v^2+(2j+1)^2\pi^2}}\right)^{n+1} <\left(\frac{21}{20}\right)^4\sum_{j=1}^{\infty}\frac{1}{(2j+1)^4} \le\left(\frac{21}{20}\right)^4\left(\frac{1}{81}+\int_1^\infty\frac{dx}{(2x+1)^4}\right) =\left(\frac{21}{20}\right)^4\frac{1}{54}<\frac{1}{40}.

Therefore,

∣(v2+π2)(n+1)/22Sn(v)−cos⁡((n+1)θ(v))∣<140.(10)\left|\frac{(v^2+\pi^2)^{(n+1)/2}}{2}S_n(v)-\cos((n+1)\theta(v))\right|<\frac{1}{40}. \tag*{(10)}

In particular, whenever cos⁡((n+1)θ(v))=±1\cos((n+1)\theta(v))=\pm1, the function Sn(v)S_n(v) has the same sign as the cosine.

Step 3. Counting the zeros by alternating signs. The function θ(v)\theta(v) is continuous and strictly decreasing on [0,7/9][0,7/9]. Hence (n+1)θ(v)/π(n+1)\theta(v)/\pi runs through the interval

[n+12−n+1πarctan⁡79π,n+12].(11)\left[\frac{n+1}{2}-\frac{n+1}{\pi}\arctan\frac{7}{9\pi},\frac{n+1}{2}\right]. \tag*{(11)}

We claim that this interval has length greater than (n+1)/13(n+1)/13. Using π<22/7\pi<22/7 and arctan⁡x>x−x3/3\arctan x>x-x^3/3 for x>0x>0, we obtain

1πarctan⁡79π>722arctan⁡7198>722(49198−13(49198)3)>113.(12)\frac{1}{\pi}\arctan\frac{7}{9\pi}>\frac{7}{22}\arctan\frac{7}{198}>\frac{7}{22}\left(\frac{49}{198}-\frac{1}{3}\left(\frac{49}{198}\right)^3\right)>\frac{1}{13}. \tag*{(12)}

Thus the interval (11) contains at least ⌊(n+1)/13⌋\lfloor(n+1)/13\rfloor consecutive integers. For each such integer kk, there is a unique v∈[0,7/9]v\in[0,7/9] satisfying (n+1)θ(v)=kπ(n+1)\theta(v)=k\pi. At these points,

cos⁡((n+1)θ(v))=(−1)k.\cos((n+1)\theta(v))=(-1)^k.

By (10), the corresponding values of Sn(v)S_n(v) are nonzero and alternate in sign. The intermediate value theorem therefore gives a zero of Sn(v)S_n(v) between each pair of consecutive points. These zeros are distinct and lie in (0,7/9)(0,7/9). Hence Sn(v)S_n(v) has at least

max⁡{0,⌊n+113⌋−1}\max\left\{0,\left\lfloor\frac{n+1}{13}\right\rfloor-1\right\}

distinct zeros in (0,7/9)(0,7/9). By (9), the same holds for An(−ev)A_n(-e^v).

Step 4. Obtaining strong factors. It remains to show that every zero counted above corresponds to a strong factor. Using the Taylor series for cosh⁡v\cosh v and bounding its terms from v2/2v^2/2 onward by a geometric series, we obtain

cosh⁡79≤1+(7/9)2/21−(7/9)2/12=1217923<43.(13)\cosh\frac{7}{9}\le1+\frac{(7/9)^2/2}{1-(7/9)^2/12}=\frac{1217}{923}<\frac{4}{3}. \tag*{(13)}

Thus 2sech⁡(7/9)>3/22\operatorname{sech}(7/9)>3/2. Since sech⁡v\operatorname{sech}v is strictly decreasing for v>0v>0, every zero vi∈(0,7/9)v_i\in(0,7/9) satisfies

ci=2sech⁡vi>32.c_i=2\operatorname{sech}v_i>\frac{3}{2}.

By Proposition 2.1, distinct positive zeros viv_i correspond to distinct quadratic factors. Therefore at least ⌊(n+1)/13⌋−1\lfloor(n+1)/13\rfloor-1 factors are strong, as required. This proves the lemma. □\square

Log-concavity under parameter variation

To apply the strong-factor estimate from Section 2, we establish two tools for products of reciprocal quadratic factors.

Let q(t)=∑j=0dbjtjq(t)=\sum_{j=0}^{d}b_jt^j with bj=0b_j=0 for j<0j<0 or j>dj>d. For each k∈Zk\in\mathbb{Z}, the identity

[tk](q(t)(1+xt+t2))=bk+xbk−1+bk−2\left[t^k\right]\left(q(t)(1+xt+t^2)\right)=b_k+xb_{k-1}+b_{k-2}

gives

Δk(q(t)(1+xt+t2))=Δk(q(t)(1+t2))+Lk(q)x+Δk−1(q)x2,(14)\Delta_k\left(q(t)(1+xt+t^2)\right)=\Delta_k\left(q(t)(1+t^2)\right)+L_k(q)x+\Delta_{k-1}(q)x^2, \tag*{(14)}

where Lk(q)=(bkbk−1−bk+1bk−2)+(bk−1bk−2−bkbk−3)L_k(q)=(b_kb_{k-1}-b_{k+1}b_{k-2})+(b_{k-1}b_{k-2}-b_kb_{k-3}). By Lemma 1.2, Lk(q)≥0L_k(q)\geq0 whenever q(t)q(t) is a product of factors 1+cit+t21+c_it+t^2 with ci≥0c_i\geq0. The preceding identity leads to the following endpoint reduction criterion for log-concavity.

Proposition 3.1. Let m≥0m \ge0 and 0≤ai≤di0 \le a_i \le d_i for 1≤i≤m1 \le i \le m. For each fixed integer kk, the minimum of

Δk(∏i=1m(1+cit+t2))\Delta_k\left(\prod_{i=1}^{m}(1+c_i t+t^2)\right)

over ai≤ci≤dia_i \le c_i \le d_i is attained when ci∈{ai,di}c_i \in\{a_i,d_i\} for all ii. Consequently, the coefficient sequence of the product is log-concave for all ai≤ci≤dia_i \le c_i \le d_i if and only if it is log-concave whenever ci∈{ai,di}c_i \in\{a_i,d_i\} for all ii.

Proof. The case m=0m=0 is immediate. Fix kk and all parameters except ci=xc_i=x, and let

q(t)=∏j≠i(1+cjt+t2).q(t)=\prod_{j\ne i}(1+c_jt+t^2).

By (3.1), we have Δk(q(t)(1+xt+t2))=Δk(q(t)(1+t2))+Lk(q)x+Δk−1(q)x2\Delta_k(q(t)(1+xt+t^2))=\Delta_k(q(t)(1+t^2))+L_k(q)x+\Delta_{k-1}(q)x^2. Since Lk(q)≥0L_k(q)\ge0 by Lemma 1.2, we consider two cases.

If Δk−1(q)≥0\Delta_{k-1}(q)\ge0, then

ddxΔk(q(t)(1+xt+t2))=Lk(q)+2xΔk−1(q)≥0\frac{d}{dx}\Delta_k\left(q(t)(1+xt+t^2)\right)=L_k(q)+2x\Delta_{k-1}(q)\ge0

for x≥0x\ge0. Thus the minimum on [ai,di][a_i,d_i] is attained at aia_i.

If Δk−1(q)<0\Delta_{k-1}(q)<0, then (3.1) is a concave quadratic in xx. Its minimum on [ai,di][a_i,d_i] is therefore attained at an endpoint.

Thus, for fixed values of the other parameters, the minimum with respect to cic_i is attained at one of the endpoints of [ai,di][a_i,d_i].

Since the expression is continuous on the compact parameter box, it attains a global minimum. Starting from a minimizer, we can successively replace each parameter by a suitable endpoint without increasing the value. After all parameters have been replaced, we obtain a vertex at which the global minimum is attained.

Consequently, if the coefficient sequence of the product is log-concave at every vertex, then Δk≥0\Delta_k\ge0 throughout the box for every relevant kk. The converse is immediate.

Suppose that the coefficient sequence of q(t)q(t) is nonnegative, log-concave, and has no internal zeros. Then

bkbk−1≥bk+1bk−2(k∈Z).(15)b_kb_{k-1}\ge b_{k+1}b_{k-2}\qquad(k\in\mathbb{Z}). \tag*{(15)}

If bk+1bk−2=0b_{k+1}b_{k-2}=0, the inequality is immediate. Otherwise, the absence of internal zeros implies that bk−1bk>0b_{k-1}b_k>0. Multiplying the two log-concavity inequalities

bk−12≥bk−2bk,bk2≥bk−1bk+1,b_{k-1}^2\ge b_{k-2}b_k,\qquad b_k^2\ge b_{k-1}b_{k+1},

and dividing by bk−1bkb_{k-1}b_k gives (3.2). Applying this inequality at kk and k−1k-1 gives Lk(q)≥0L_k(q)\ge0, as defined in (3.1).

The following lemma allows us to replace factors 1+t21+t^2 by 1+xt+t21+xt+t^2 while preserving log-concavity under suitable assumptions.

Lemma 3.2. Let g(t)g(t) be a nonzero polynomial with nonnegative coefficients and no internal zeros, and let NN be a nonnegative integer. Suppose that the coefficient sequence of

g(t)(1+t2)jg(t)(1+t^2)^j

is log-concave for every 0≤j≤N0 \le j \le N. Then, for every integer 0≤s≤N0 \le s \le N and x1,…,xs≥0x_1,\ldots,x_s \ge0, the coefficient sequence of

g(t)∏i=1s(1+xit+t2)g(t)\prod_{i=1}^{s}(1+x_i t+t^2)

is log-concave.

Proof. First, suppose that q(t)q(t) has nonnegative, log-concave coefficients and no internal zeros, and that q(t)(1+t2)q(t)(1+t^2) also has a log-concave coefficient sequence. By (3.1),

Δk(q(t)(1+xt+t2))=Δk(q(t)(1+t2))+Lk(q)x+Δk−1(q)x2.\Delta_k\bigl(q(t)(1+xt+t^2)\bigr)=\Delta_k\bigl(q(t)(1+t^2)\bigr)+L_k(q)x+\Delta_{k-1}(q)x^2.

All three terms on the right are nonnegative for x≥0x \ge0, since Lk(q)≥0L_k(q) \ge0 by (3.2). Consequently,

q(t)(1+xt+t2)(16)q(t)(1+xt+t^2) \tag*{(16)}

has a log-concave coefficient sequence for every x≥0x \ge0.

If N=0N=0, the conclusion follows directly from the hypothesis. We assume for the remainder of the proof that N≥1N \ge1. Since g(t)(1+t2)g(t)(1+t^2) has a log-concave coefficient sequence, g(t)g(t) cannot be a monomial. Indeed, if g(t)=atrg(t)=at^r with a>0a>0, then Δr+1(g(t)(1+t2))=−a2<0\Delta_{r+1}\bigl(g(t)(1+t^2)\bigr)=-a^2<0, a contradiction. Thus the support of g(t)g(t) contains at least two consecutive indices. Multiplication by 1+xt+t21+xt+t^2, where x≥0x \ge0, preserves this property and introduces no internal zeros. Hence all products considered below have nonnegative coefficients and no internal zeros.

We prove by induction on ℓ\ell that

g(t)(1+t2)j∏i=1ℓ(1+xit+t2)(17)g(t)(1+t^2)^j\prod_{i=1}^{\ell}(1+x_i t+t^2) \tag*{(17)}

has a log-concave coefficient sequence whenever j,ℓ≥0j,\ell\ge0 and j+ℓ≤Nj+\ell\le N.

For ℓ=0\ell=0, the assertion follows from the hypothesis of the lemma. Suppose that it holds for ℓ−1\ell-1, and fix j≥0j \ge0 with j+ℓ≤Nj+\ell\le N.

Since j+ℓ−1≤Nj+\ell-1 \le N, (j+1)+(ℓ−1)=j+ℓ≤N(j+1)+(\ell-1)=j+\ell\le N, the induction hypothesis shows that both

g(t)(1+t2)j∏i=1ℓ−1(1+xit+t2)g(t)(1+t^2)^j\prod_{i=1}^{\ell-1}(1+x_i t+t^2)

and

g(t)(1+t2)j+1∏i=1ℓ−1(1+xit+t2)g(t)(1+t^2)^{j+1}\prod_{i=1}^{\ell-1}(1+x_i t+t^2)

have log-concave coefficient sequences. Moreover, the first product has nonnegative coefficients and no internal zeros, as established above. Since the second product is the first multiplied by 1+t21+t^2, applying (3.3) with x=xℓx=x_\ell shows that

g(t)(1+t2)j∏i=1ℓ(1+xit+t2)g(t)(1+t^2)^j\prod_{i=1}^{\ell}(1+x_i t+t^2)

has a log-concave coefficient sequence. This completes the induction. Taking j=0j=0 and ℓ=s\ell=s gives the desired conclusion.

Proof of Theorem 1.3

We first establish a block criterion for log-concavity and then combine it with Lemma 2.2 to prove Theorem 1.3.

Proposition 4.1. For integers 0≤s≤730 \le s \le73 and 0≤j≤80 \le j \le8, let

Ps,j(t)=(1+t2)s(2+3t+2t2)8−j(2+4t+2t2)j.(18)P_{s,j}(t) = (1+t^2)^s(2+3t+2t^2)^{8-j}(2+4t+2t^2)^j. \tag*{(18)}

Then

Δk(Ps,j)≥0,1≤k≤8+s.(19)\Delta_k(P_{s,j}) \ge0,\qquad1 \le k \le8+s. \tag*{(19)}

Equality holds only when (s,j,k)=(73,0,1)(s,j,k)=(73,0,1).

Proof. Each factor in (4.1) is palindromic, so Ps,j(t)P_{s,j}(t) is palindromic of degree 2(8+s)2(8+s) and Δk(Ps,j)=Δ2(8+s)−k(Ps,j)\Delta_k(P_{s,j})=\Delta_{2(8+s)-k}(P_{s,j}). Thus the inequalities in (4.2) imply log-concavity of the entire coefficient sequence.

We verify these inequalities as follows.

For each 0≤j≤80 \le j \le8, we first compute the coefficients of P0,j(t)P_{0,j}(t) by multiplying its eight quadratic factors. The coefficients of Ps+1,j(t)P_{s+1,j}(t) are then obtained recursively from

[tk]Ps+1,j(t)=[tk]Ps,j(t)+[tk−2]Ps,j(t),0≤s<73.(20)[t^k]P_{s+1,j}(t)=[t^k]P_{s,j}(t)+[t^{k-2}]P_{s,j}(t),\qquad0 \le s<73. \tag*{(20)}

Using these coefficients, we evaluate Δk(Ps,j)\Delta_k(P_{s,j}) for 1≤k≤8+s1 \le k \le8+s.

The verification involves 9∑s=073(8+s)=29 6379\sum_{s=0}^{73}(8+s)=29\,637 integer comparisons. All the computed values are nonnegative, with equality occurring only at (s,j,k)=(73,0,1)(s,j,k)=(73,0,1).

The accompanying script verify_log_concavity.py, available in the GitHub repository, reproduces this verification using exact integer arithmetic. This completes the proof.

We now combine Proposition 4.1 with the results of § 3 to obtain the following block criterion.

Proposition 4.2. Suppose c1,…,c8∈[3/2,2]c_1,\ldots,c_8\in[3/2,2], s∈Zs\in\mathbb{Z} with 0≤s≤730\le s\le73, and x1,…,xs≥0x_1,\ldots,x_s\ge0. Then the coefficient sequence of the polynomial

∏i=18(1+cit+t2)∏ℓ=1s(1+xℓt+t2)(21)\prod_{i=1}^{8}(1+c_i t+t^2)\prod_{\ell=1}^{s}(1+x_\ell t+t^2) \tag*{(21)}

is log-concave.

Proof. Let g(t)=∏i=18(1+cit+t2)g(t)=\prod_{i=1}^{8}(1+c_i t+t^2). We first show that g(t)(1+t2)sg(t)(1+t^2)^s has a log-concave coefficient sequence for every integer 0≤s≤730\le s\le73. Fix such an ss. By Proposition 3.1, it suffices to consider ci∈{3/2,2}c_i\in\{3/2,2\} for 1≤i≤81\le i\le8, with the remaining ss parameters fixed at zero. If exactly jj of the eight parameters equal 22, the corresponding product is

g(t)(1+t2)s=(1+t2)s(1+32t+t2)8−j(1+2t+t2)j=2−8Ps,j(t).g(t)(1+t^2)^s=(1+t^2)^s\left(1+\frac{3}{2}t+t^2\right)^{8-j}(1+2t+t^2)^j=2^{-8}P_{s,j}(t).

By Proposition 4.1 and the palindromicity of Ps,j(t)P_{s,j}(t), its coefficient sequence is log-concave for every 0≤j≤80\le j\le8. Since multiplication by a positive constant preserves log-concavity, the same holds for 2−8Ps,j(t)2^{-8}P_{s,j}(t). Proposition 3.1 therefore shows that g(t)(1+t2)sg(t)(1+t^2)^s has a log-concave coefficient sequence for all ci∈[3/2,2]c_i\in[3/2,2].

Moreover, g(t)g(t) has positive coefficients and hence has no internal zeros. Thus all the hypotheses of Lemma 3.2 are satisfied with N=73N=73. It follows that

g(t)∏ℓ=1s(1+xℓt+t2)g(t)\prod_{\ell=1}^{s}(1+x_\ell t+t^2)

has a log-concave coefficient sequence for every 0≤s≤730\le s\le73 and x1,…,xs≥0x_1,\ldots,x_s\ge0, as required.

Remark 4.3. The bound s≤73s \le73 in Proposition 4.2 is sharp. Taking ci=3/2c_i = 3/2 for 1≤i≤81 \le i \le8 and xℓ=0x_\ell= 0 for 1≤ℓ≤s1 \le\ell\le s gives

(1+32t+t2)8(1+t2)s.\left(1+\frac{3}{2}t+t^2\right)^8(1+t^2)^s.

Its first three coefficients are 11, 1212, and 8+(82)(32)2+s=71+s8+\binom{8}{2}\left(\frac{3}{2}\right)^2+s=71+s, respectively. Hence the log-concavity difference at index 11 is 122−(71+s)=73−s12^2-(71+s)=73-s, which is negative when s=74s=74. Thus the bound cannot be improved.

We now complete the proof of Theorem 1.3 using exact integer verification for n≤220n \le220 and the strong-factor estimate together with the block criterion for n≥221n \ge221.

Proof of Theorem 1.3. Case 1. 1≤n≤2201 \le n \le220. We use the recurrence of Lin, Ma, Wang and Wang [10 Theorem 2.1]:

2an+1,k=(k+1)(an,k+1+an,k−1)+(n−k+1)(an,k+an,k−2),a1,0=1,(22)2a_{n+1,k}=(k+1)(a_{n,k+1}+a_{n,k-1})+(n-k+1)(a_{n,k}+a_{n,k-2}),\qquad a_{1,0}=1, \tag*{(22)}

where an,k=0a_{n,k}=0 unless 0≤k≤n−10 \le k \le n-1. Starting from a1,0=1a_{1,0}=1, this recurrence determines the coefficients of A^n(t)\widehat{A}_n(t) for every n≤220n \le220.

By (1), we have Δk(A^n)=Δn−1−k(A^n)\Delta_k(\widehat{A}_n)=\Delta_{n-1-k}(\widehat{A}_n). Thus it suffices to verify Δk(A^n)≥0\Delta_k(\widehat{A}_n)\ge0 for 1≤k≤⌊(n−1)/2⌋1 \le k \le\lfloor(n-1)/2\rfloor. The number of required comparisons is ∑n=1220⌊n−12⌋=11 990\sum_{n=1}^{220}\left\lfloor\frac{n-1}{2}\right\rfloor=11\,990. An exact integer computation using (22) shows that all these differences are nonnegative, with equality only at (n,k)=(3,1)(n,k)=(3,1). The accompanying script verify_log_concavity.py reproduces this verification. Hence the theorem holds for n≤220n \le220.

Case 2. n≥221n \ge221. Let m=⌊(n−1)/2⌋m=\lfloor(n-1)/2\rfloor. By Proposition 2.1, we have

A^n(t)=an,0(1+t)n−1−2m∏i=1m(1+cit+t2),\widehat{A}_n(t)=a_{n,0}(1+t)^{n-1-2m}\prod_{i=1}^{m}(1+c_i t+t^2),

where an,0>0a_{n,0}>0 and 0<ci<20<c_i<2.

Let b=⌊(2m−11)/104⌋b=\lfloor(2m-11)/104\rfloor. Since m≥110m\ge110, we have b≥2b\ge2. By Lemma 2.2, the number of strong factors is at least

⌊n+113⌋−1≥⌊2m−1113⌋≥8b.(23)\left\lfloor\frac{n+1}{13}\right\rfloor-1\ge\left\lfloor\frac{2m-11}{13}\right\rfloor\ge8b. \tag*{(23)}

Moreover, the definition of bb gives 2m−11<104(b+1)2m-11<104(b+1), and hence

m≤52b+57≤81b,(24)m\le52b+57\le81b, \tag*{(24)}

where the last inequality follows from b≥2b\ge2.

We can therefore choose 8b8b strong factors and divide them into bb groups, each containing eight strong factors. By (24), the number of remaining factors satisfies

m−8b≤73b.m-8b\le73b.

These factors can be distributed among the bb groups so that each group contains at most 73 additional factors.

The eight strong factors in each group have parameters in [3/2,2][3/2,2], while all remaining parameters are nonnegative. Thus, by Proposition 4.2, every group has a log-concave coefficient sequence. Since all the factors have positive coefficients, each group also has no internal zeros. Applying Lemma 1.1 repeatedly, we conclude that

∏i=1m(1+cit+t2)\prod_{i=1}^{m}(1+c_i t+t^2)

has a log-concave coefficient sequence with no internal zeros.

Finally, n−1−2m∈{0,1}n-1-2m\in\{0,1\}. Since the coefficient sequence of 1+t1+t is log-concave, another application of Lemma 1.1 shows that multiplying by the possible factor 1+t1+t preserves log-concavity. Multiplication by an,0>0a_{n,0}>0 also preserves log-concavity. Therefore, the coefficient sequence of A^n(t)\widehat{A}_n(t) is log-concave for all n≥221n\ge221, which completes the proof.

As a consequence of Theorem 1.3, we recover the unimodality established by Lin, Ma, Wang and Wang [10].

Corollary 4.4. For every n≥1n\ge1, the coefficient sequence of A^n(t)\widehat{A}_n(t) is unimodal.

Applications

We give four consequences of Theorem 1.3. We first transfer log-concavity to the distribution of 3-descents and derive a local limit theorem. We then bound the largest coefficient and obtain a Lorentzian reformulation of the main theorem.

For π∈Sn+1\pi\in\mathfrak{S}_{n+1}, let d3(π)d_3(\pi) denote the number of indices i∈[n−1]i\in[n-1] for which the consecutive triple πiπi+1πi+2\pi_i\pi_{i+1}\pi_{i+2} has relative order 132, 213 or 321. Chebikin’s equidistribution theorem [5 Theorem 2.3 and Corollary 2.2] relates this statistic to alternating descents.

Corollary 5.1. For every n≥1n\ge1,

∑π∈Sn+1td3(π)=(n+1)A^n(t).(25)\sum_{\pi\in\mathfrak{S}_{n+1}} t^{d_3(\pi)}=(n+1)\widehat{A}_n(t). \tag*{(25)}

Consequently, the coefficient sequence on the left is log-concave for every n≥1n\ge1.

Normal approximation for alternating descents was studied by Heerten, Sambale and Thäle [8]. Theorem 1.3 supplies the log-concavity hypothesis in Bender’s local limit theorem [1]. We give the resulting coefficient estimate below.

Corollary 5.2. As n→∞n\to\infty,

sup⁡k∈Z∣5n12an,kn!−12πexp⁡(−65n(k−n−12)2)∣⟶0,(26)\sup_{k\in\mathbb{Z}}\left|\sqrt{\frac{5n}{12}}\frac{a_{n,k}}{n!}-\frac{1}{\sqrt{2\pi}}\exp\left(-\frac{6}{5n}\left(k-\frac{n-1}{2}\right)^2\right)\right|\longrightarrow0, \tag*{(26)}

where an,k=0a_{n,k}=0 outside 0≤k≤n−10\le k\le n-1. Consequently,

an,⌊(n−1)/2⌋∼n!65πn.(27)a_{n,\lfloor(n-1)/2\rfloor}\sim n!\sqrt{\frac{6}{5\pi n}}. \tag*{(27)}

Proof. Hwang, Chern and Duh [9 Section 5.4.2] established asymptotic normality with centering n/2n/2 and scale 5n/12\sqrt{5n/12}. The centering may be replaced by (n−1)/2(n-1)/2, since the difference divided by 5n/12\sqrt{5n/12} tends to zero. By Proposition 2.1, the coefficients are positive, and by Theorem 1.3, they form a log-concave sequence. Since 5n/12→∞\sqrt{5n/12}\to\infty, Bender’s local limit theorem [1] gives (5.2).

Taking k=⌊(n−1)/2⌋k=\lfloor(n-1)/2\rfloor in (5.2), the exponential term tends to 1, giving (5.3).

The main theorem also yields bounds for the largest coefficient through a concentration inequality for symmetric log-concave distributions.

Corollary 5.3. For every n≥2n \ge2, the maximum coefficient of A^n(t)\hat{A}_n(t) is an,⌊(n−1)/2⌋a_{n,\lfloor(n-1)/2\rfloor}, and

n!5n−6≤an,⌊(n−1)/2⌋≤n!65n−1.(28)\frac{n!}{\sqrt{5n-6}} \le a_{n,\lfloor(n-1)/2\rfloor} \le n!\sqrt{\frac{6}{5n-1}}. \tag*{(28)}

Proof. Palindromicity and Corollary 4.4 show that the central coefficient is a maximum coefficient. Let XnX_n be the number of alternating descents in a uniformly random permutation in Sn\mathfrak{S}_n. Then P(Xn=k)=an,k/n!\mathbb{P}(X_n=k)=a_{n,k}/n!.

Each alternating descent indicator has variance 1/41/4. Two adjacent indicators are simultaneously 11 with probability 1/31/3, so their covariance is 1/121/12; nonadjacent indicators are independent. Hence

Var⁡(Xn)=n−14+n−26=5n−712.\operatorname{Var}(X_n)=\frac{n-1}{4}+\frac{n-2}{6}=\frac{5n-7}{12}.

By Theorem 1.3 and Proposition 2.1, the distribution of XnX_n is log-concave with no internal zeros, and palindromicity makes it symmetric about (n−1)/2(n-1)/2. Thus Bobkov, Marsiglietti and Melbourne [2 Theorem 1.1] give

11+12Var⁡(Xn)≤max⁡kan,kn!≤11+2Var⁡(Xn).\frac{1}{\sqrt{1+12\operatorname{Var}(X_n)}}\le\max_k\frac{a_{n,k}}{n!}\le\frac{1}{\sqrt{1+2\operatorname{Var}(X_n)}}.

Substituting the variance above and multiplying by n!n! gives (5.4). □

We finish with a reformulation of Theorem 1.3 using the characterization of bivariate Lorentzian polynomials.

Corollary 5.4. For every n≥1n \ge1, the homogeneous polynomial

∑k=0n−1(n−1k)an,kxkyn−1−k(29)\sum_{k=0}^{n-1}\binom{n-1}{k}a_{n,k}x^k y^{n-1-k} \tag*{(29)}

is Lorentzian and completely log-concave.

Proof. The cases n=1,2n=1,2 are immediate. For n≥3n \ge3, the bivariate characterization in [3 Example 2.26] reduces the Lorentzian property to log-concavity and the absence of internal zeros of the coefficients after division by (n−1k)\binom{n-1}{k}. For (5.5), these normalized coefficients are exactly an,ka_{n,k}. They are positive by Proposition 2.1 and log-concave by Theorem 1.3. Complete log-concavity follows from [3 Theorem 2.30]. □

Conflict of Interest

The authors declare no conflict of interest.

Acknowledgements

This work was supported in part by the National Natural Science Foundation of China (Grant Nos. 12271078, 12101110) and the Scientific Research Fund of Liaoning Provincial Education Department (Grant No. LJ212410173071).

References

  1. [1]E.A. Bender, Central and local limit theorems applied to asymptotic enumeration, J. Combin. Theory Ser. A 15(1) (1973) 91–111.DOI
  2. [2]S.G. Bobkov, A. Marsiglietti, J. Melbourne, Concentration functions and entropy bounds for discrete log-concave distributions, Combin. Probab. Comput. 31(1) (2022) 54–72.arxiv.org/abs/2007.11030
  3. [3]P. Brändén, J. Huh, Lorentzian polynomials, Ann. of Math. (2) 192(3) (2020) 821–891.arxiv.org/abs/1902.03719
  4. [4]F. Brenti, Unimodal, log-concave and Pólya frequency sequences in combinatorics, Mem. Amer. Math. Soc. 81 (1989) no. 413.DOI
  5. [5]D. Chebikin, Variations on descents and inversions in permutations, Electron. J. Combin. 15(1) (2008) Paper No. R132.arxiv.org/abs/0804.1935
  6. [6]C.O. Chow, New proofs of interlacing of zeros of Eulerian polynomials. III, Enumer. Combin. Appl. 4(2) (2024) Paper No. S2R14.DOI
  7. [7]R. Ellard, H. Šmigoc, Families of Newton-like inequalities for sets of self-conjugate complex numbers, Linear Algebra Appl. 597 (2020) 46–68.arxiv.org/abs/1604.05148
  8. [8]N. Heerten, H. Sambale, C. Thäle, Probabilistic limit theorems induced by the zeros of polynomials, Math. Nachr. 297(5) (2024) 1772–1792.arxiv.org/abs/2207.08469
  9. [9]H.-K. Hwang, H.-H. Chern, G.-H. Duh, An asymptotic distribution theory for Eulerian recurrences with applications, Adv. in Appl. Math. 112 (2020) Paper No. 101960.arxiv.org/abs/1807.01412
  10. [10]Z. Lin, S.-M. Ma, D.G.L. Wang, L. Wang, Positivity and divisibility of enumerators of alternating descents, Ramanujan J. 58(1) (2022) 203–228.DOI
  11. [11]S.-M. Ma, Q. Fang, T. Mansour, Y.-N. Yeh, Alternating Eulerian polynomials and left peak polynomials, Discrete Math. 345(3) (2022) Paper No. 112714.arxiv.org/abs/2104.09374
  12. [12]S.-M. Ma, Y.-N. Yeh, Enumeration of permutations by number of alternating descents, Discrete Math. 339(4) (2016) 1362–1367.DOI
  13. [13]Q. Pan, A new combinatorial formula for alternating descent polynomials, preprint (2022), arXiv:2207.06212.arxiv.org/abs/2207.06212
  14. [14]Q. Pan, C. Qian, On the $cd$-index for alternating descents, Electron. J. Combin. 31(4) (2024) Paper No. P4.29.DOI
  15. [15]Q. Pan, J. Zeng, Enumeration of permutations by the parity of descent positions, Discrete Math. 346(10) (2023) Paper No. 113575.arxiv.org/abs/2209.15302
  16. [16]R.P. Stanley, Enumerative Combinatorics, Vol. 1, 2nd ed., Cambridge University Press, Cambridge, 2012.
  17. [17]B.-X. Zhu, Positivity of iterated sequences of polynomials, SIAM J. Discrete Math. 32(3) (2018) 1993–2010.arxiv.org/abs/1807.01062

Paper details

vlnc:W_pe37zdmq
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC0-1.0
  • Mathematics

Not recorded for this paper

View statement
Looking for related papers…
Contents