betaFeedback

WorksMathematics

Cassini ovals, translated polynomials, and the arithmetic of their discriminants and resultants

Abstract

For a polynomial ff, a>0a>0 and b,ω∈Cb,ω\in\mathbb{C}, consider the translate J(a,b,ω)f(z)=f(z/a)−ωf(b/a)J^{(a,b,ω)}f(z)=f(z/a)-ωf(b/a). We show that for f(x)=(x2−1)nf(x)=(x^{2}-1)^{n} and ω=±1ω=\pm1 every zero of J(a,b,ω)fJ^{(a,b,ω)}f lies on the Cassini oval ∣z2−a2∣=∣b2−a2∣|z^{2}-a^{2}|=|b^{2}-a^{2}|. We call these translates Cassini polynomials and give their zeros explicitly, prove that the zeros for ω=1ω=1 and ω=−1ω=-1 strictly interlace on each component of the oval, and evaluate their discriminants and resultants in closed form. We then prove the reduced monic polynomials Qn=a2nJ(a,b,ω)f/(z2−b2)Q_{n}=a^{2n}J^{(a,b,ω)}f/(z^{2}-b^{2}) factor into homogenized cyclotomic polynomials, satisfying Res⁡(Qm,Qn)=(b2−a2)2(m−1)(n−1)\operatorname{Res}(Q_{m},Q_{n})=(b^{2}-a^{2})^{2(m-1)(n-1)} for coprime m,nm,n, and have discriminants governed by a Lucas sequence. We then prove a comparison principle: if ff is even or odd with all zeros in [−1,1][-1,1], b∈Rb\in\mathbb{R} and ∣ω∣≤1|ω|\le1, then every zero of J(a,ib,ω)fJ^{(a,ib,ω)}f satisfies ∣z2−a2∣≤a2+b2|z^{2}-a^{2}|\le a^{2}+b^{2}, and the symmetry hypothesis cannot be removed. As an application, every solution of det⁡(A(Pn)−λI)=Fn+1\det(A(P_{n})-λI)=F_{n+1}, where A(Pn)A(P_{n}) is the adjacency matrix of the path on nn vertices and Fn+1F_{n+1} is a Fibonacci number, satisfies ∣λ2−4∣≤5|λ^{2}-4|\le5. This disproves, for every nn, a conjecture of Gullerud, Johnson and Mbirika that the real parts of these solutions are unbounded.

Introduction

Zeros of polynomial sequences

The location of the zeros of a sequence of polynomials {Pn(z)}n≥0\{P_n(z)\}_{n\geq0} is a classical theme joining complex analysis, approximation theory, combinatorics and number theory; see [15, 19, 21]. For orthogonal polynomials the zeros are real and interlace in the interval of orthogonality. For the Chebyshev polynomials TnT_n and UnU_n they are explicit and lie in (−1,1)(-1,1) [16, 20]. When the sequence of orthogonal polynomials is perturbed, translated or linearly combined, however, the zeros typically leave the real line and accumulate on curves in the complex plane. Describing these curves remains an active area of study.

One of the key results in this area is the theorem of Beraha, Kahane and Weiss [2, 3]. Suppose Pn(z)=∑j=1kαj(z)λj(z)nP_n(z)=\sum_{j=1}^{k}\alpha_j(z)\lambda_j(z)^n satisfies a linear recurrence with polynomial coefficients and a mild non-degeneracy condition. Then the limit points of the zeros are the points where two dominant characteristic roots have equal modulus, together with the isolated points where the coefficient of a unique dominant root vanishes.

Recent developments involve algebraic approaches to seek curves that contain every zero of every member of the polynomial sequence. Much of the recent progress here is due to Khang Tran and his collaborators. Tran [22] proved that if ∑m≥0Hm(z)tm=1/D(t,z)\sum_{m\ge0} H_m(z)t^m = 1/D(t,z) with D(t,z)=1+B(z)t+A(z)tkD(t,z)=1+B(z)t+A(z)t^k and k∈{2,3,4}k\in\{2,3,4\}, then all zeros of HmH_m lie on an explicit portion of a real algebraic curve. The proof connects root location with the discriminant of DD in tt and with Ismail’s qq-analogue of the discriminant [9]. The three-term recurrence Pn+B(z)Pn−1+A(z)Pn−k=0P_n+B(z)P_{n-1}+A(z)P_{n-k}=0 was treated in [23].

Another approach, closest to this current paper, concerns polynomials whose zeros lie on conics. Dilcher and Stolarsky [5] found polynomials satisfying non-linear recurrences related to Chebyshev polynomials whose zeros lie close to ellipse-like curves. In his thesis, Oyengo [18] showed that for real λ,κ≠0\lambda,\kappa\ne0 with α2=λ2−κ2≠0\alpha^2=\lambda^2-\kappa^2\ne0, the zeros of the translated Chebyshev polynomials

Tn(z/α)−Tn(iκ/α)andTn(z/α)−Tn(λ/α)T_n(z/\alpha)-T_n(i\kappa/\alpha) \qquad\text{and}\qquad T_n(z/\alpha)-T_n(\lambda/\alpha)

lie on the ellipse x2/λ2+y2/κ2=1x^2/\lambda^2+y^2/\kappa^2=1. They are the images of rotated roots of unity under the linear map diag⁡(λ,κ)\operatorname{diag}(\lambda,\kappa), and in suitable cases they interlace on the ellipse. The analogous translates of UnU_n have their zeros close to, and inside, such ellipses. Independently, Kim and Lee [12] proved that the zeros of the difference quotient (Tn(z)−Tn(a))/(z−a)(T_n(z)-T_n(a))/(z-a) lie on the ellipse x2/a2+y2/(a2−1)=1x^2/a^2+y^2/(a^2-1)=1. For the special choice λ=5,κ=1\lambda=\sqrt{5},\kappa=1 the translating values are Lucas and Fibonacci numbers, since Tn(i/2)=in2LnT_n(i/2)=\frac{i^n}{2}L_n.

One might ask the converse problem; given an algebraic curve in the complex plane, can we find a sequence of real polynomials whose zeros lie on this curve? In his thesis, Oyengo showed that for ellipses, one can derive a sequence of translated Chebyshev polynomials whose zeros lie on the ellipses. The same applies for circles which are special ellipses. In this paper, we show that the same is true for Ovals of Cassini, that certain translations of cyclotomic polynomials have their zeros on ovals of Cassini.

Cassini ovals

Let a>0a>0 and R>0R>0. The Cassini oval with foci ±a\pm a and product RR is the locus of points whose distances to the two foci have product RR:

C(a,R)={z∈C:∣z−a∣∣z+a∣=R}={z∈C:∣z2−a2∣=R}.(1)C(a,R)=\{z\in\mathbb{C}:|z-a||z+a|=R\}=\{z\in\mathbb{C}:|z^2-a^2|=R\}. \tag*{(1)}

In Cartesian, polar and complex coordinates it reads

((x−a)2+y2)((x+a)2+y2)=R2,r4−2a2r2cos⁡2θ+a4=R2,∣z2−a2∣=R.(2)\begin{aligned} ((x-a)^2+y^2)((x+a)^2+y^2)&=R^2, \qquad r^4-2a^2r^2\cos2\theta+a^4=R^2, \qquad|z^2-a^2|=R. \tag*{(2)} \end{aligned}

and expanding the first gives the quartic (x2+y2)2−2a2(x2−y2)=R2−a4(x^2+y^2)^2-2a^2(x^2-y^2)=R^2-a^4. The curve is symmetric about both axes and meets the real axis at ±a2+R\pm\sqrt{a^2+R} and, when R<a2R<a^2, also at ±a2−R\pm\sqrt{a^2-R}. Its shape is governed by the ratio R/a2R/a^2 (Figure 1):

Zeros of $g_{12}$ and $h_{12}$ on Cassini ovals

Figure 1. Zeros of g12g_{12} (filled circles) and h12h_{12} (open squares) for a=7a=7 and the indicated values of bb, on the Cassini ovals ∣z2−a2∣=∣b2−a2∣\lvert z^2-a^2\rvert=\lvert b^2-a^2\rvert. Crosses mark the foci ±a\pm a. The zeros are equally spaced in the parameter arg⁡(z2−a2)\arg(z^2-a^2) and interlace on each component (Theorem 3.4). In the lemniscate case n=12n=12 is even and (1−b2/a2)12=1(1-b^2/a^2)^{12}=1, so z=0z=0 is a double zero of g12g_{12} (Theorem 3.1(c)).

  1. R>2a2R>2a^2: a single convex oval;

  2. a2<R<2a2a^2<R<2a^2: a single oval with two “dimples” at the imaginary axis (peanut shape);

  3. R=a2R=a^2: the lemniscate of Bernoulli, with a node at the origin;

  4. 0<R<a20<R<a^2: two disjoint ovals, one around each focus.

Thus C(a,R)C(a,R) is the level set ∣P∣=R|P|=R of the quadratic polynomial P(z)=z2−a2P(z)=z^2-a^2, i.e. a polynomial lemniscate of degree two.

G. D. Cassini introduced these curves in the late seventeenth century as candidate planetary orbits competing with Kepler’s ellipses [8, 11]. They have since found many applications in Physics, Mathematics Engineering and even in Medicine. For more on other properties of ovals of Cassini, refer to [11, 13].

Translated polynomials and the Cassini polynomials

We ask a general question; Let ff be a real polynomial with real zeros and for a,b,ω∈Ca,b,\omega\in\mathbb{C}, a≠0a \ne0 define the operator J(a,b,ω)J^{(a,b,\omega)} as

J(a,b,ω)f=f(z/a)−ωf(b/a).J^{(a,b,\omega)}f=f(z/a)-\omega f(b/a).

When do the zeros of J(a,b,ω)f(z)J^{(a,b,\omega)}f(z) lie on, or are controlled by, a single curve depending only on parameters (a,b,ω)(a,b,\omega)? For Chebyshev polynomials, the curve is an ellipse when ω=±1\omega=\pm1. In this paper, we show that for the family fn(x)=(x2−1)nf_n(x)=(x^2-1)^n, the curve is a Cassini oval when ω=±1\omega=\pm1. We call the resulting translates Cassini polynomials, and in this paper, we study their geometry, arithmetic of their discriminants and their resultants. The Cassini oval then turns out to control the translates of every symmetric real-rooted polynomial with zeros in [−1,1][-1,1]. In particular it controls the Chebyshev, Legendre and Gegenbauer translates, including those arising from path graphs.

Throughout, a>0a>0 and b∈Cb\in\mathbb{C} with b2∉{0,a2}b^2\notin\{0,a^2\}, with β=b2−a2≠0\beta=b^2-a^2\ne0, R=∣β∣R=|\beta|. Define the monic polynomials Gn(z)G_n(z) and Hn(z)H_n(z) as

Gn(z)=(z2−a2)n−βn,Hn(z)=(z2−a2)n+βn,(3)G_n(z)=(z^2-a^2)^n-\beta^n,\qquad H_n(z)=(z^2-a^2)^n+\beta^n, \tag*{(3)}

so that the gn=a−2nGng_n=a^{-2n}G_n and hn=a−2nHnh_n=a^{-2n}H_n are explicitly given by

gn(z)=(z2a2−1)n−(b2a2−1)n,hn(z)=(z2a2−1)n+(b2a2−1)n.(4)g_n(z)=\left(\frac{z^2}{a^2}-1\right)^n-\left(\frac{b^2}{a^2}-1\right)^n,\qquad h_n(z)=\left(\frac{z^2}{a^2}-1\right)^n+\left(\frac{b^2}{a^2}-1\right)^n. \tag*{(4)}

Note that if fn(x)=(x2−1)nf_n(x)=(x^2-1)^n then gn(z)=fn(z/a)−fn(b/a)g_n(z)=f_n(z/a)-f_n(b/a) and hn(z)=fn(z/a)+fn(b/a)h_n(z)=f_n(z/a)+f_n(b/a). Our main results are as follows.

Theorem A (Geometry; Theorems 3.1, 3.4). For every n≥1n\ge1 the zeros of gng_n and hnh_n lie on C(a,R)C(a,R) and are explicitly given by

z=±(a2+βe2πik/n)1/2andz=±(a2+βeπi(2k+1)/n)1/2,0≤k≤n−1,z=\pm\left(a^2+\beta e^{2\pi i k/n}\right)^{1/2}\qquad\text{and}\qquad z=\pm\left(a^2+\beta e^{\pi i(2k+1)/n}\right)^{1/2},\qquad0\le k\le n-1,

respectively. They are simple unless (1−b2/a2)n=1(1-b^2/a^2)^n=1 (resp. =−1=-1), in which case 00 is the only multiple zero and it is double. If R≠a2R\ne a^2 the zeros of gng_n and hnh_n strictly interlace on each component of C(a,R)C(a,R).

Theorem B (Arithmetic; Theorems 5.1, 5.5, 5.7, 5.8). With Qn=Gn/(z2−b2)Q_n=G_n/(z^2-b^2) and En=∑k=0n−1a2k(a2−b2)n−1−kE_n=\sum_{k=0}^{n-1}a^{2k}(a^2-b^2)^{n-1-k} identities for resultants and discriminants are given by

Res⁡z(Gn,Hn)=22nβ2n2,Disc⁡z(Gn)=(2n)2nβ2n(n−1)(a2n−(a2−b2)n),Disc⁡z(Hn)=(2n)2nβ2n(n−1)(a2n+(a2−b2)n),Disc⁡z(Qn)=4n−1n2n−4β2(n−1)(n−2)En,Res⁡z(Qm,Qn)=β2(m−1)(n−1) if gcd⁡(m,n)=1, and 0 otherwise.\begin{aligned} \operatorname{Res}_z(G_n,H_n)&=2^{2n}\beta^{2n^2},\\ \operatorname{Disc}_z(G_n)&=(2n)^{2n}\beta^{2n(n-1)}\left(a^{2n}-(a^2-b^2)^n\right),\\ \operatorname{Disc}_z(H_n)&=(2n)^{2n}\beta^{2n(n-1)}\left(a^{2n}+(a^2-b^2)^n\right),\\ \operatorname{Disc}_z(Q_n)&=4^{n-1}n^{2n-4}\beta^{2(n-1)(n-2)}E_n,\\ \operatorname{Res}_z(Q_m,Q_n)&=\beta^{2(m-1)(n-1)}\text{ if }\gcd(m,n)=1,\text{ and }0\text{ otherwise.} \end{aligned}

If a2,b2∈Za^2,b^2\in\mathbb{Z} are coprime with b2<a2b^2<a^2, then for every n≥3n\ge3, except (a2,b2,n)=(1,−1,6)(a^2,b^2,n)=(1,-1,6), there is a prime p≡1(modn)p\equiv1\pmod n dividing Disc⁡z(Qn)\operatorname{Disc}_z(Q_n) but no Disc⁡z(Qm)\operatorname{Disc}_z(Q_m) with 2≤m<n2\le m<n.

Theorem C (Comparison principle; Theorem 4.1, Corollary 4.5). Let ff be a non-constant polynomial with all zeros in [−1,1][-1,1] and f(−x)=±f(x)f(-x)=\pm f(x), let a,b∈Ra,b\in\mathbb{R} with a>0a>0 and b∈Rb\in\mathbb{R}. Every zero of Jf(a,ib,ω)J_f^{(a,ib,\omega)} with ∣ω∣≤1|\omega|\leq1 lies in K(a,a2+b2)K(a,a^2+b^2). Every zero of Jf(a,b,ω)J_f^{(a,b,\omega)}, ∣b∣>a|b|>a and ∣ω∣≥1|\omega|\geq1 lies outside the interior of K(a,b2−a2)K(a,b^2-a^2). Without the symmetry hypothesis both statements fail.

Theorem A describes the geometry of the zeros of the gng_n and hnh_n polynomials as lying and dense on the ovals of Cassini. The Cassini oval is precisely curve predicted by Tran’s discriminant criterion; see Remarks 3.6 and 3.7. What is special here is that the whole zero set, not just its limit, lies on the curve. Theorem B is the arithmetic counterpart for the polynomials. The resultant formula is a Cassini analogue of Apostol’s cyclotomic resultants and of the Chebyshev resultants of [10, 14]. The discriminant contains a Lucas sequence, which brings in primitive divisors. Theorem C proves the Oyengo’s earlier conjecture for symmetric polynomials and shows by example that it fails in general.

In section 2, we present some well known results about resultants, discriminants and the Cassini region. In section 3, we prove results leading to Theorem A. In section 4, we prove the comparison principle and some of its applications. We conclude this paper in section 5 with identities for discriminants and resultants involving the Cassini polynomials.

Preliminaries

Resultants and discriminants

Let f(x)=am∏i=1m(x−αi)f(x)=a_m\prod_{i=1}^{m}(x-\alpha_i) and g(x)=bn∏j=1n(x−βj)g(x)=b_n\prod_{j=1}^{n}(x-\beta_j) with ambn≠0a_mb_n\ne0. We use the following standard facts; see [6].

Res⁡x(f,g)=amn∏i=1mg(αi)=(−1)mnbnm∏j=1nf(βj),(5)\operatorname{Res}_x(f,g)=a_m^n\prod_{i=1}^{m}g(\alpha_i)=(-1)^{mn}b_n^m\prod_{j=1}^{n}f(\beta_j), \tag*{(5)}
Res⁡x(g,f)=(−1)mnRes⁡x(f,g),Res⁡x(f,g1g2)=Res⁡x(f,g1)Res⁡x(f,g2),(6)\operatorname{Res}_x(g,f)=(-1)^{mn}\operatorname{Res}_x(f,g),\qquad\operatorname{Res}_x(f,g_1g_2)=\operatorname{Res}_x(f,g_1)\operatorname{Res}_x(f,g_2), \tag*{(6)}
Res⁡x(f,g)=amn−ℓRes⁡x(f,r)if g=fq+r, deg⁡r=ℓ,(7)\operatorname{Res}_x(f,g)=a_m^{n-\ell}\operatorname{Res}_x(f,r)\quad\text{if }g=fq+r,\ \deg r=\ell, \tag*{(7)}
Res⁡x(λf,μg)=λnμmRes⁡x(f,g)(λ,μ∈C×),(8)\operatorname{Res}_x(\lambda f,\mu g)=\lambda^n\mu^m\operatorname{Res}_x(f,g)\qquad(\lambda,\mu\in\mathbb{C}^{\times}), \tag*{(8)}
Disc⁡x(f)=(−1)m(m−1)/21amRes⁡x(f,f′)=am2m−2∏i<j(αi−αj)2,(9)\operatorname{Disc}_x(f)=(-1)^{m(m-1)/2}\frac{1}{a_m}\operatorname{Res}_x(f,f')=a_m^{2m-2}\prod_{i<j}(\alpha_i-\alpha_j)^2, \tag*{(9)}
Disc⁡x(λf)=λ2m−2Disc⁡x(f)(λ∈C×).\operatorname{Disc}_x(\lambda f)=\lambda^{2m-2}\operatorname{Disc}_x(f)\qquad(\lambda\in\mathbb{C}^{\times}).

We also use the chain rule of McKay and Wang [17]: if hh has degree kk and leading coefficient h0h_0, then

Res⁡x(f(h(x)),g(h(x)))=[h0deg⁡f⋅deg⁡gRes⁡x(f,g)]k,(10)\operatorname{Res}_x\bigl(f(h(x)),g(h(x))\bigr)=\left[h_0^{\deg f\cdot\deg g}\operatorname{Res}_x(f,g)\right]^k, \tag*{(10)}

unless hh is a constant equal to a common root of ff and gg.

The next lemma uses McKay and Wang’s chainrule for resultants and computes discriminants of even polynomials. It is the tool we will employ in the calculation of the discriminant formulas in Section 5.

Lemma 2.1. Let S∈C[w]S \in\mathbb{C}[w] be monic of degree k≥1k \ge1, let a∈Ca \in\mathbb{C}, and put F(z)=S(z2−a2)F(z)=S(z^2-a^2). Then

Disc⁡z(F)=(−4)kS(−a2)Disc⁡w(S)2.\operatorname{Disc}_{z}(F)=(-4)^k S(-a^2)\operatorname{Disc}_{w}(S)^2.

Proof. FF is monic of degree 2k2k, and F′(z)=2zS′(z2−a2)F'(z)=2zS'(z^2-a^2). Let ρ1,…,ρk\rho_1,\ldots,\rho_k be the zeros of SS. The zeros of FF are the 2k2k numbers ±rj\pm r_j with rj2=a2+ρjr_j^2=a^2+\rho_j. By (5),

Res⁡z(F,F′)=∏F(r)=02rS′(r2−a2)=22k(∏F(r)=0r)∏j=1kS′(ρj)2.\operatorname{Res}_{z}(F,F')=\prod_{F(r)=0}2rS'(r^2-a^2)=2^{2k}\left(\prod_{F(r)=0}r\right)\prod_{j=1}^{k}S'(\rho_j)^2.

Since FF is monic of even degree, we have that F(0)=S(−a2)F(0)=S(-a^2). Also ∏jS′(ρj)=Res⁡w(S,S′)=±Disc⁡w(S)\prod_jS'(\rho_j)=\operatorname{Res}_{w}(S,S')=\pm\operatorname{Disc}_{w}(S). Finally Disc⁡z(F)=(−1)2k(2k−1)/2Res⁡z(F,F′)=(−1)kRes⁡z(F,F′)\operatorname{Disc}_{z}(F)=(-1)^{2k(2k-1)/2}\operatorname{Res}_{z}(F,F')=(-1)^k\operatorname{Res}_{z}(F,F').

Cyclotomic polynomials

For a positive integer nn, the nn-th cyclotomic polynomial is the unique irreducible polynomial with integer coefficients that is a divisor of xn−1x^n-1. Precisely

Φn(x)=∏1≤k≤ngcd⁡(k,n)=1(x−e2iπk/n).\Phi_n(x)=\prod_{\substack{1\le k\le n\\ \gcd(k,n)=1}}\left(x-e^{2i\pi k/n}\right).

Let Φe\Phi_e denote the ee-th cyclotomic polynomial and Φe∗(x,y)=yφ(e)Φe(x/y)\Phi_e^*(x,y)=y^{\varphi(e)}\Phi_e(x/y) its homogenization. Then xn−yn=∏e∣nΦe∗(x,y)x^n-y^n=\prod_{e\mid n}\Phi_e^*(x,y). Put

Ψn(x)=xn−1x−1=∏e∣ne>1Φe(x).\Psi_n(x)=\frac{x^n-1}{x-1}=\prod_{\substack{e\mid n\\e>1}}\Phi_e(x).

We need the classical evaluation

Disc⁡x(Ψn)=(−1)(n−1)(n−2)/2nn−2(n≥2).(11)\operatorname{Disc}_{x}(\Psi_n)=(-1)^{(n-1)(n-2)/2}n^{n-2}\qquad(n\ge2). \tag*{(11)}

and the following observation. If gcd⁡(m,n)=1\gcd(m,n)=1, then ζ↦ζn\zeta\mapsto\zeta^n permutes the non-trivial mm-th roots of unity, so

∏ζm=1ζ≠1Ψn(ζ)=∏ζm=1ζ≠1ζn−1ζ−1=1(gcd⁡(m,n)=1).(12)\prod_{\substack{\zeta^m=1\\\zeta\ne1}}\Psi_n(\zeta)=\prod_{\substack{\zeta^m=1\\\zeta\ne1}}\frac{\zeta^n-1}{\zeta-1}=1\qquad(\gcd(m,n)=1). \tag*{(12)}

The Cassini region

For a,R>0a,R>0 let K(a,R)={z:∣z2−a2∣≤R}K(a,R)=\{z:|z^2-a^2|\le R\}, the compact region bounded by C(a,R)C(a,R).

Lemma 2.2. Let a,R>0a,R>0. For z∈K(a,R)z\in K(a,R) we have ∣z∣≤a2+R|z|\le\sqrt{a^2+R} and

∣Im⁡z∣≤{R/(2a),R≤2a2,R−a2,R≥2a2.|\operatorname{Im}z|\le \begin{cases} R/(2a),&R\le2a^2,\\ \sqrt{R-a^2},&R\ge2a^2. \end{cases}

Both bounds are attained on C(a,R)C(a,R).

Proof. The first bound follows from ∣z∣2=∣z2∣≤∣z2−a2∣+a2|z|^2 = |z^2| \le|z^2-a^2|+a^2. Since Im⁡z\operatorname{Im} z is harmonic, its maximum over K(a,R)K(a,R) is attained on the boundary. Write z2=a2+Reiθz^2=a^2+Re^{i\theta}, from Im⁡z2=∣z2∣−Re⁡z2\operatorname{Im} z^2=|z^2|-\operatorname{Re} z^2 we get Im⁡z2=ϕ(cos⁡θ)\operatorname{Im} z^2=\phi(\cos\theta), where ϕ(c)=a4+R2+2a2Rc−a2−Rc\phi(c)=\sqrt{a^4+R^2+2a^2Rc}-a^2-Rc for c∈[−1,1]c\in[-1,1]. The function ϕ\phi is concave, and ϕ′(c)=0\phi'(c)=0 exactly when a4+R2+2a2Rc=a2\sqrt{a^4+R^2+2a^2Rc}=a^2, i.e. when c=−R/(2a2)c=-R/(2a^2). If R≤2a2R\le2a^2 this point lies in [−1,1][-1,1] and gives ϕ=R2/(2a2)\phi=R^2/(2a^2). Otherwise the maximum is at c=−1c=-1 and equals 2(R−a2)2(R-a^2).

Polynomials with zeros on Cassini ovals

Recall that Gn(z)=(z2−a2)n−βnG_n(z)=(z^2-a^2)^n-\beta^n, and Hn(z)=(z2−a2)n+βnH_n(z)=(z^2-a^2)^n+\beta^n and for z∈Cz\in\mathbb{C} let z\sqrt{z} be the principal square root.

Theorem 3.1. Let n≥1n\ge1, ζk=e2πik/n\zeta_k=e^{2\pi ik/n} and ηk=eπi(2k+1)/n\eta_k=e^{\pi i(2k+1)/n} (0≤k≤n−10\le k\le n-1).

(a) The zeros of Gn(z)G_n(z) (equivalently of gn(z)g_n(z)) are ±a2+βζk\pm\sqrt{a^2+\beta\zeta_k}, and the zeros of Hn(z)H_n(z) (equivalently of hn(z)h_n(z)) are ±a2+βηk\pm\sqrt{a^2+\beta\eta_k}, for $0\le k\le n-1.

(b) All zeros of gn(z)g_n(z) and hn(z)h_n(z) lie on the Cassini oval C(a,R)C(a,R), R=∣b2−a2∣R=|b^2-a^2|.

(c) Gn(z)G_n(z) has 2n2n simple zeros unless (1−b2/a2)n=1(1-b^2/a^2)^n=1. In that case R=a2R=a^2, C(a,R)C(a,R) is a lemniscate of Bernoulli, and z=0z=0 (its node) is a double zero while the others are simple. The same holds for Hn(z)H_n(z) with the condition (1−b2/a2)n=−1(1-b^2/a^2)^n=-1.

Proof. (a) Gn(z)=0G_n(z)=0 if and only if ((z2−a2)/β)n=1((z^2-a^2)/\beta)^n=1, i.e. z2=a2+βζkz^2=a^2+\beta\zeta_k for some kk. The case of Hn(z)H_n(z) is the same with ζk\zeta_k replaced by the nn-th roots ηk\eta_k of −1-1.

(b) For every such zero, ∣z−a∣∣z+a∣=∣z2−a2∣=∣β∣∣ζk∣=R|z-a||z+a|=|z^2-a^2|=|\beta||\zeta_k|=R.

(c) The nn values a2+βζka^2+\beta\zeta_k are distinct. Each non-zero value has two distinct square roots, which are simple zeros of Gn(z)G_n(z). The value 00 occurs if and only if ζk=−a2/β\zeta_k=-a^2/\beta for some kk, i.e. if and only if (−a2/β)n=1(-a^2/\beta)^n=1, which is equivalent to (1−b2/a2)n=1(1-b^2/a^2)^n=1. It then occurs for exactly one kk and gives the double zero z=0z=0. Taking absolute values shows that this forces R=∣β∣=a2R=|\beta|=a^2. The argument for the zeros of Hn(z)H_n(z) is identical.

Example 3.2. The operator J(a,b,1)J^{(a,b,1)} applied to fn(x)=(x2−1)nf_n(x)=(x^2-1)^n, gives gn(z)g_n(z). Hence the zeros of gn(z)=a−2n((z2−a2)n−(b2−a2)n)g_n(z)=a^{-2n}((z^2-a^2)^n-(b^2-a^2)^n) lie on ((x−a)2+y2)((x+a)2+y2)=∣b2−a2∣2((x-a)^2+y^2)((x+a)^2+y^2)=|b^2-a^2|^2 for every b∈Cb\in\mathbb{C}. The polynomial J(2,i,1)fn(z)=fn(z/2)−fn(i/2)=4−n((z2−4)n−(−5)n)J^{(2,i,1)}f_n(z)=f_n(z/2)-f_n(i/2)=4^{-n}((z^2-4)^n-(-5)^n) has its zeros on ((x−2)2+y2)((x+2)2+y2)=25((x-2)^2+y^2)((x+2)^2+y^2)=25.

For real a,b>0a,b>0, the four regimes of § 1.2 become: b≥3ab\ge\sqrt{3}a (convex); 2a<b<3a\sqrt{2}a<b<\sqrt{3}a (dimpled); b=2ab=\sqrt{2}a (lemniscate); a<b<2aa<b<\sqrt{2}a (two ovals). Figure 1 illustrates all four with a=7a=7.

Recall that ∣z2−a2∣=R|z^2-a^2|=R with R=∣b2−a2∣R=|b^2-a^2|. When R≠a2R\ne a^2 the oval C(a,R)C(a,R) is naturally parametrized by the angle θ=arg⁡(z2−a2)\theta=\arg(z^2-a^2). In this parameter the zeros of GnG_n and HnH_n are equally spaced.

Lemma 3.3. Let a,R>0a,R>0 with R≠a2R\ne a^2 and put W(θ)=a2+ReiθW(\theta)=a^2+Re^{i\theta}.

(a) If R<a2R<a^2, then C(a,R)C(a,R) consists of two Jordan curves Γ+⊂{Re⁡z>0}\Gamma_+\subset\{\operatorname{Re}z>0\} and Γ−=−Γ+\Gamma_-=-\Gamma_+. The map θ↦W(θ)\theta\mapsto\sqrt{W(\theta)} is a homeomorphism R/2πZ→Γ+\mathbb{R}/2\pi\mathbb{Z}\to\Gamma_+.

(b) If R>a2R>a^2, then C(a,R)C(a,R) is a single Jordan curve. There is a continuous s:R→Cs:\mathbb{R}\to\mathbb{C} with s(θ)2=W(θ)s(\theta)^2=W(\theta) and s(θ+2π)=−s(θ)s(\theta+2\pi)=-s(\theta), and ss induces a homeomorphism R/4πZ→C(a,R)\mathbb{R}/4\pi\mathbb{Z}\to C(a,R).

Proof. (a) If R<a2R<a^2 then Re⁡W(θ)≥a2−R>0\operatorname{Re} W(\theta)\ge a^2-R>0, so W(θ)\sqrt{W(\theta)} is continuous with positive real part. It is injective in θ\theta modulo 2π2\pi, because its square W(θ)W(\theta) is. Its image is the part of C(a,R)\mathcal{C}(a,R) in the right half-plane, and C(a,R)\mathcal{C}(a,R) does not meet the imaginary axis because ∣(iy)2−a2∣=y2+a2>R\lvert(iy)^2-a^2\rvert=y^2+a^2>R.

(b) If R>a2R>a^2, the closed path WW winds once around 00 and never passes through it. Hence a continuous square root ss exists along R\mathbb{R} and changes sign after one turn. If s(θ1)=s(θ2)s(\theta_1)=s(\theta_2), then W(θ1)=W(θ2)W(\theta_1)=W(\theta_2), so θ2−θ1∈2πZ\theta_2-\theta_1\in2\pi\mathbb{Z}. If θ2−θ1\theta_2-\theta_1 is an odd multiple of 2π2\pi, then s(θ1)=−s(θ1)s(\theta_1)=-s(\theta_1), which is impossible since W≠0W\ne0. Hence ss is injective on R/4πZ\mathbb{R}/4\pi\mathbb{Z}. Its image is all of C(a,R)\mathcal{C}(a,R), since both square roots ±s(θ)=s(θ),s(θ+2π)\pm s(\theta)=s(\theta),s(\theta+2\pi) occur. □

Theorem 3.4. Let R≠a2R\ne a^2 and θ0=arg⁡β\theta_0=\arg\beta. In the parameterizations of Lemma 3.3, the zeros of GnG_n correspond to θ≡θ0+2πk/n\theta\equiv\theta_0+2\pi k/n and the zeros of HnH_n to θ≡θ0+π(2k+1)/n\theta\equiv\theta_0+\pi(2k+1)/n. Consequently:

(a) if R>a2R>a^2, the 2n2n zeros of GnG_n and the 2n2n zeros of HnH_n strictly interlace on the single curve C(a,R)\mathcal{C}(a,R);

All the four curves in one graph

Figure 2. All the four curves in one graph.

(b) if R<a2R < a^2, each of Γ±\Gamma_{\pm} carries nn zeros of GnG_n and nn zeros of HnH_n, and they strictly interlace on it.

Proof. W(θ)=a2+βζW(\theta)=a^2+\beta\zeta with ∣ζ∣=1|\zeta|=1 means θ≡θ0+arg⁡ζ(mod2π)\theta\equiv\theta_0+\arg\zeta\pmod{2\pi}. By Theorem 3.1(a) the zeros of GnG_n correspond to arg⁡ζ∈2πnZ\arg\zeta\in\frac{2\pi}{n}\mathbb{Z} and those of HnH_n to arg⁡ζ∈πn+2πnZ\arg\zeta\in\frac{\pi}{n}+\frac{2\pi}{n}\mathbb{Z}. These two arithmetic progressions alternate on R/2πZ\mathbb{R}/2\pi\mathbb{Z}, and also on its double cover R/4πZ\mathbb{R}/4\pi\mathbb{Z}. Homeomorphisms of circles preserve cyclic order, so the claims follow from Lemma 3.3.

Remark 3.5. When R=a2R=a^2 (the lemniscate) the same argument applies to each lobe. The zeros of GnG_n and HnH_n interlace on each lobe, and one of them may have a double zero at the node z=0z=0 (Figure 1). Theorem 3.4 also contains an interlacing statement for a single nn across different translates. The zeros of GnG_n for β\beta and for βeiπ/n\beta e^{i\pi/n} interlace, since the second family is HnH_n for the first.

Remark 3.6. Since Gn=λ1n−λ2nG_n=\lambda_1^n-\lambda_2^n with λ1(z)=z2−a2\lambda_1(z)=z^2-a^2 and λ2=β\lambda_2=\beta, the sequence (Gn)(G_n) satisfies the recurrence Gn+1=(λ1+λ2)Gn−λ1λ2Gn−1G_{n+1}=(\lambda_1+\lambda_2)G_n-\lambda_1\lambda_2G_{n-1} with G0=0G_0=0. Its characteristic roots have equal modulus exactly on C(a,R)C(a,R). The coefficients ±1\pm1 never vanish, so Beraha–Kahane–Weiss Theorem [2, 3] identify C(a,R)C(a,R) as the set of limit points of zeros. Theorem 3.1 says much more: every zero of every GnG_n is on the curve.

Remark 3.7. Let Qn=Gn/(z2−b2)Q_n=G_n/(z^2-b^2) (see Section 5). Then

∑n≥1Qn(z)tn−1=1(1−(z2−a2)t)(1−βt)=11+B(z)t+A(z)t2.\sum_{n\ge1} Q_n(z)t^{n-1}=\frac{1}{(1-(z^2-a^2)t)(1-\beta t)}=\frac{1}{1+B(z)t+A(z)t^2}.

with B(z)=−(z2+b2−2a2)B(z)=-(z^2+b^2-2a^2) and A(z)=β(z2−a2)A(z)=\beta(z^2-a^2). This is exactly the setting of Tran’s discriminant criterion [22], Theorem 1, which places the zeros of QnQ_n on the curve Im⁡(B2/A)=0\operatorname{Im}(B^2/A)=0, 0≤B2/A≤40\le B^2/A\le4. Here B2/A=2+λ+λ−1B^2/A=2+\lambda+\lambda^{-1} with λ=(z2−a2)/β\lambda=(z^2-a^2)/\beta. This quantity is real and lies in [0,4][0,4] if and only if ∣λ∣=1|\lambda|=1, i.e. exactly on C(a,R)C(a,R). The tt-discriminant B2−4A=(z2−a2−β)2B^2-4A=(z^2-a^2-\beta)^2 vanishes only at z=±bz=\pm b, which is the common zero of all GnG_n removed in QnQ_n.

Remark 3.8. The argument of Theorem 3.1 applies verbatim to any non-constant P∈C[z]P\in\mathbb{C}[z]. The zeros of P(z)n−P(b)nP(z)^n-P(b)^n lie on the lemniscate ∣P(z)∣=∣P(b)∣|P(z)|=|P(b)|. Cassini ovals are the case P(z)=z2−a2P(z)=z^2-a^2. The ellipse results for TnT_n recalled in § 1 are of a different nature, because TnT_n is not a power of a fixed polynomial.

A comparison principle and its applications

An outstanding question is to describe the zeros of J(a,b,ω)fJ^{(a,b,\omega)}f for real polynomials ff with zeros in [−1,1][-1,1]. The initial conjecture is that the zeros are either inside or on the Cassini oval C(a,R)C(a,R). The next theorem proves this under a symmetry hypothesis. Example 4.3 shows that the hypothesis cannot be removed.

Theorem 4.1. Let f∈C[x]f\in\mathbb{C}[x] be non-constant with all zeros in [−1,1][-1,1] and f(−x)=±f(x)f(-x)=\pm f(x). Let a>0a>0.

(i) Let b∈Rb\in\mathbb{R} and ω∈C\omega\in\mathbb{C} with ∣ω∣≤1|\omega|\le1. Every zero zz of f(z/a)−ωf(ib/a)f(z/a)-\omega f(ib/a) satisfies ∣z2−a2∣≤a2+b2|z^2-a^2|\le a^2+b^2. If equality holds, then z=±ibz=\pm ib, or ∣ω∣=1|\omega|=1 and f(x)=C(x2−1)mf(x)=C(x^2-1)^m.

(ii) Let c∈Rc\in\mathbb{R} with ∣c∣≥a|c|\ge a and ω∈C\omega\in\mathbb{C} with ∣ω∣≥1|\omega|\ge1. Every zero zz of f(z/a)−ωf(c/a)f(z/a)-\omega f(c/a) satisfies ∣z2−a2∣≥c2−a2|z^2-a^2|\ge c^2-a^2.

Proof. Because the zero set of ff is real, contained in [−1,1][-1,1] and symmetric about 0 with multiplicities, we can write

f(x)=Cxϵ∏j=1m(x2−tj),C≠0, ϵ∈{0,1}, tj∈[0,1].f(x)=Cx^\epsilon\prod_{j=1}^{m}(x^2-t_j),\qquad C\ne0,\ \epsilon\in\{0,1\},\ t_j\in[0,1].

Put u=z2/a2u=z^2/a^2. Then

∣f(z/a)∣2=∣C∣2∣u∣ϵ∏j=1m∣u−tj∣2.(13)|f(z/a)|^2=|C|^2|u|^\epsilon\prod_{j=1}^{m}|u-t_j|^2. \tag*{(13)}

(i) Let s=b2/a2≥0s=b^2/a^2\ge0. Since (ib/a)2=−s(ib/a)^2=-s, formula (13) gives ∣f(ib/a)∣2=∣C∣2sϵ∏j(tj+s)2|f(ib/a)|^2=|C|^2s^\epsilon\prod_j(t_j+s)^2. Suppose ∣z2−a2∣>a2+b2|z^2-a^2|>a^2+b^2, i.e. ∣u−1∣>1+s|u-1|>1+s. For every t∈[0,1]t\in[0,1],

∣u−t∣≥∣u−1∣−(1−t)>(1+s)−(1−t)=t+s≥0.(14)|u-t|\ge|u-1|-(1-t)>(1+s)-(1-t)=t+s\ge0. \tag*{(14)}

Taking t=tjt=t_j, and t=0t=0 for the factor ∣u∣|u|, we obtain ∣f(z/a)∣>∣f(ib/a)∣≥∣ωf(ib/a)∣|f(z/a)|>|f(ib/a)|\ge|\omega f(ib/a)|. So zz is not a zero. If instead ∣u−1∣=1+s|u-1|=1+s, then we do not have a strict inequality in (14). Equality in the triangle inequality for some t<1t<1 forces u−1∈(−∞,0]u-1\in(-\infty,0] with ∣u−1∣=1+s|u-1|=1+s, i.e. u=−su=-s, i.e. z=±ibz=\pm ib. Hence, if z≠±ibz\ne\pm ib is a zero on the oval, then ∣ω∣=1|\omega|=1, ϵ=0\epsilon=0 and all tj=1t_j=1.

(ii) Let d=c2/a2>1d=c^2/a^2>1. Then ∣f(c/a)∣2=∣C∣2dϵ∏j(d−tj)2>0|f(c/a)|^2=|C|^2d^\epsilon\prod_j(d-t_j)^2>0. Suppose ∣z2−a2∣<c2−a2|z^2-a^2|<c^2-a^2, i.e. ∣u−1∣<d−1|u-1|<d-1. For every t∈[0,1]t\in[0,1], ∣u−t∣≤∣u−1∣+(1−t)<d−t|u-t|\le|u-1|+(1-t)<d-t. Hence ∣f(z/a)∣<∣f(c/a)∣≤∣ωf(c/a)∣|f(z/a)|<|f(c/a)|\le|\omega f(c/a)|, and zz is not a zero. ∎

Remark 4.2. Geometrically, {u:∣u−t∣≤t+s}\{u : \lvert u-t\rvert\le t+s\} (0≤t≤10 \le t \le1) is a family of discs that increases with tt and whose members are all internally tangent at −s-s. The largest one, t=1t=1, is the image of K(a,a2+b2)\mathcal{K}(a,a^2+b^2) under z↦z2/a2z \mapsto z^2/a^2. The proof shows that each factor of ff is individually dominated. Part (ii) uses the decreasing family {∣u−t∣≤b−t}\{\lvert u-t\rvert\le b-t\}, whose members are all tangent at bb.

Example 4.3. Symmetry is necessary. Take f(x)=(x−1)2f(x)=(x-1)^2, a=2a=2, b=1b=1. The zeros of f(z/2)−f(i/2)f(z/2)-f(i/2) are z=iz=i and z=4−iz=4-i, and ∣(4−i)2−4∣=∣11−8i∣=185>5\lvert(4-i)^2-4\rvert=\lvert11-8i\rvert=\sqrt{185}>5. So 4−i4-i lies outside K(2,5)\mathcal{K}(2,5).

Corollary 4.4. Let ff be any of TnT_n, UnU_n, the Legendre polynomial PnP_n, or a Gegenbauer polynomial Cn(λ)C_n^{(\lambda)} with λ>−12\lambda>-\frac{1}{2}. For a>0a>0 and b∈Rb\in\mathbb{R}, the zeros of f(z/a)−f(ib/a)f(z/a)-f(ib/a) lie in K(a,a2+b2)\mathcal{K}(a,a^2+b^2). For ∣c∣>a\lvert c\rvert>a the zeros of f(z/a)−f(c/a)f(z/a)-f(c/a) lie outside the interior of K(a,c2−a2)\mathcal{K}(a,c^2-a^2).

Proof. These polynomials have parity (−1)n(-1)^n and all their zeros in (−1,1)(-1,1) [21]. Apply Theorem 4.1 with ω=1\omega=1. □

For f=Tnf=T_n this is consistent with the ellipse theorem of [18]. The ellipse x2/(a2+b2)+y2/b2=1x^2/(a^2+b^2)+y^2/b^2=1 has foci ±a\pm a and lies in K(a,a2+b2)\mathcal{K}(a,a^2+b^2), touching C(a,a2+b2)\mathcal{C}(a,a^2+b^2) at ±ib\pm ib.

Corollary 4.5. Let fn(λ)=det⁡(A(Pn)−λI)f_n(\lambda)=\det(A(P_n)-\lambda I) be the characteristic polynomial of the path graph on nn vertices, and let FkF_k denote the Fibonacci numbers. For every n≥1n\ge1, every complex solution λ\lambda of fn(λ)=Fn+1f_n(\lambda)=F_{n+1} satisfies

∣λ2−4∣≤5,hence∣λ∣≤3and∣Im⁡λ∣≤54.\lvert\lambda^2-4\rvert\le5,\qquad\text{hence}\qquad\lvert\lambda\rvert\le3\quad\text{and}\quad\lvert\operatorname{Im}\lambda\rvert\le\frac{5}{4}.

In particular the real parts of the solutions are bounded independently of nn. This disproves, for every nn, the conjecture of [7] that these real parts are unbounded (Conjecture 1.1(2) in the numbering of [4]); the case n≡0(mod4)n\equiv0\pmod{4} was obtained in [4].

Proof. It is well known that fn(λ)=Un(−λ/2)f_n(\lambda)=U_n(-\lambda/2) [7]. From Un+1(x)=2xUn(x)−Un−1(x)U_{n+1}(x)=2xU_n(x)-U_{n-1}(x) one checks by induction that Un(i/2)=inFn+1U_n(i/2)=i^nF_{n+1}. Hence Fn+1=ωUn(i/2)F_{n+1}=\omega U_n(i/2) with ω=i−n\omega=i^{-n}. The substitution z=−λz=-\lambda turns fn(λ)=Fn+1f_n(\lambda)=F_{n+1} into Un(z/2)−ωUn(i/2)=0U_n(z/2)-\omega U_n(i/2)=0. Since UnU_n has parity (−1)n(-1)^n and its zeros lie in (−1,1)(-1,1), Theorem 4.1 with f=Unf=U_n, a=2a=2, b=1b=1 and the unimodular constant ω\omega gives ∣z2−4∣≤5\lvert z^2-4\rvert\le5, and λ2=z2\lambda^2=z^2. The bounds on ∣λ∣\lvert\lambda\rvert and ∣Im⁡λ∣\lvert\operatorname{Im}\lambda\rvert follow from Lemma 2.2 with a=2a=2 and R=5≤2a2R=5\le2a^2. □

Remark 4.6. Numerically the solutions satisfy ∣Im⁡λ∣≤1\lvert\operatorname{Im}\lambda\rvert\le1 and ∣Re⁡λ∣<5\lvert\operatorname{Re}\lambda\rvert<\sqrt{5} and approach the ellipse x2/5+y2=1x^2/5+y^2=1 (Figure 3), as conjectured in [7, 4]. Corollary 4.5 gives the uniform but weaker bounds ∣Re⁡λ∣≤3\lvert\operatorname{Re}\lambda\rvert\le3 and ∣Im⁡λ∣≤5/4\lvert\operatorname{Im}\lambda\rvert\le5/4. Closing this gap, i.e. proving an ellipse inclusion for the translates Un(z/α)−ωUn(iκ/α)U_n(z/\alpha)-\omega U_n(i\kappa/\alpha) with ∣ω∣=1\lvert\omega\rvert=1, remains open.

Discriminants and resultants {#sec:5}

Theorem 5.1. For n≥1n\ge1,

Disc⁡z(Gn)=(2n)2nβ2n(n−1)(a2n−(a2−b2)n)(15)\operatorname{Disc}_{z}(G_n)=(2n)^{2n}\beta^{2n(n-1)}\left(a^{2n}-(a^2-b^2)^n\right) \tag*{(15)}
Disc⁡z(Hn)=(2n)2nβ2n(n−1)(a2n+(a2−b2)n)(16)\operatorname{Disc}_{z}(H_n)=(2n)^{2n}\beta^{2n(n-1)}\left(a^{2n}+(a^2-b^2)^n\right) \tag*{(16)}

Figure 3. Solutions of fPn(λ)=Fn+1f_{P_n}(\lambda)=F_{n+1} for n=12,21,30n=12,21,30. They lie in the Cassini region ∣λ2−4∣≤5\lvert\lambda^2-4\rvert\le5 for every nn (Corollary 4.5) and cluster near the ellipse x2/5+y2=1x^2/5+y^2=1 conjectured in [7, 4].

and consequently

Disc⁡z(gn)=a−4n(2n−1)Disc⁡z(Gn)andDisc⁡z(hn)=a−4n(2n−1)Disc⁡z(Hn).\operatorname{Disc}_{z}(g_n)=a^{-4n(2n-1)}\operatorname{Disc}_{z}(G_n) \qquad\text{and}\qquad \operatorname{Disc}_{z}(h_n)=a^{-4n(2n-1)}\operatorname{Disc}_{z}(H_n).

Proof. Apply Lemma 2.1 with S(w)=wn−βnS(w)=w^n-\beta^n and k=nk=n. The zeros of SS are βζk\beta\zeta_k, so S′(βζk)=nβn−1ζkn−1S'(\beta\zeta_k)=n\beta^{n-1}\zeta_k^{n-1} and

Disc⁡w(S)2=(∏kS′(βζk))2=n2nβ2n(n−1)(∏kζk)2(n−1)=n2nβ2n(n−1),\operatorname{Disc}_{w}(S)^2=\left(\prod_k S'(\beta\zeta_k)\right)^2=n^{2n}\beta^{2n(n-1)}\left(\prod_k\zeta_k\right)^{2(n-1)}=n^{2n}\beta^{2n(n-1)},

because ∏kζk=±1\prod_k\zeta_k=\pm1. Moreover (−4)nS(−a2)=4n(a2n−(−β)n)=4n(a2n−(a2−b2)n)(-4)^nS(-a^2)=4^n(a^{2n}-(-\beta)^n)=4^n(a^{2n}-(a^2-b^2)^n). This proves (15). For (16) take S(w)=wn+βnS(w)=w^n+\beta^n; the same computation gives (−4)nS(−a2)=4n(a2n+(a2−b2)n)(-4)^nS(-a^2)=4^n(a^{2n}+(a^2-b^2)^n). The formula for gn=a−2nGng_n=a^{-2n}G_n and hn=a−2nHnh_n=a^{-2n}H_n follow from (9) with λ=a−2n\lambda=a^{-2n} and deg⁡Gn=deg⁡Hn=2n\deg G_n=\deg H_n=2n respectively. □

The vanishing of (15) is exactly the condition of Theorem 3.1(c). For real a,ba,b it happens only when b2=2a2b^2=2a^2 and nn is even, i.e. on the lemniscate, when the node z=0z=0 is a zero. This fits the geometry.

Proposition 5.2. Res⁡z(Gn,Hn)=22nβ2n2\operatorname{Res}_{z}(G_n,H_n)=2^{2n}\beta^{2n^2}, and hence Res⁡z(gn,hn)=22nβ2n2a−8n2\operatorname{Res}_{z}(g_n,h_n)=2^{2n}\beta^{2n^2}a^{-8n^2}.

Proof. Hn=Gn+2βnH_n=G_n+2\beta^n, so by (5) Res⁡z(Gn,Hn)=∏Gn(r)=02βn=(2βn)2n\operatorname{Res}_{z}(G_n,H_n)=\prod_{G_n(r)=0}2\beta^n=(2\beta^n)^{2n}. The normalized version follows from (8). □

Every GnG_n vanishes at z=±bz=\pm b, so resultants Res⁡(Gm,Gn)\operatorname{Res}(G_m,G_n) are identically zero. The natural objects are the reduced polynomials

Qn(z)=Gn(z)z2−b2=Sn(z2−a2),Sn(w)=wn−βnw−β=βn−1Ψn(w/β),(17)Q_n(z)=\frac{G_n(z)}{z^2-b^2}=S_n(z^2-a^2),\qquad S_n(w)=\frac{w^n-\beta^n}{w-\beta}=\beta^{n-1}\Psi_n(w/\beta), \tag*{(17)}

which are monic of degree 2(n−1)2(n-1). They are Cassini analogues of the Chebyshev difference quotients of Kim and Lee [12]. Their generating function was given in Remark 3.7, so they satisfy

Qn+1(z)=(z2+b2−2a2)Qn(z)−β(z2−a2)Qn−1(z),Q0=0, Q1=1.Q_{n+1}(z) = \left(z^2+b^2-2a^2\right)Q_n(z)-\beta\left(z^2-a^2\right)Q_{n-1}(z), \qquad Q_0=0,\ Q_1=1.

If a2,b2∈Za^2,b^2\in\mathbb{Z}, then Qn∈Z[z]Q_n\in\mathbb{Z}[z].

Proposition 5.3 (Cyclotomic factorization). For n≥2n\ge2, QnQ_n has the cyclotomic factorization;

Qn(z)=∏e∣ne>1Φe∗(z2−a2,β),Q_n(z)=\prod_{\substack{e\mid n\\e>1}}\Phi_e^*\left(z^2-a^2,\beta\right),

and the factor with index ee has degree 2φ(e)2\varphi(e) in zz. Its zeros are the points of C(a,R)C(a,R) with arg⁡(z2−a2)−arg⁡β∈2πe(Z/eZ)×\arg(z^2-a^2)-\arg\beta\in\frac{2\pi}{e}(\mathbb{Z}/e\mathbb{Z})^\times.

Proof. Substitute x=z2−a2x=z^2-a^2 and y=βy=\beta in xn−yn=∏e∣nΦe∗(x,y)x^n-y^n=\prod_{e\mid n}\Phi_e^*(x,y) and cancel the factor Φ1∗=x−y=z2−b2\Phi_1^*=x-y=z^2-b^2. The description of the zeros follows from Theorem 3.1(a). □

Remark 5.4. The factor with e=2e=2 is Φ2∗(z2−a2,β)=z2−(2a2−b2)\Phi_2^*(z^2-a^2,\beta)=z^2-(2a^2-b^2). Its zeros ±2a2−b2\pm\sqrt{2a^2-b^2} are points where the oval meets a coordinate axis: the real axis if b2<2a2b^2<2a^2, the imaginary axis if b2>2a2b^2>2a^2. For a,b∈Za,b\in\mathbb{Z} it splits over Q\mathbb{Q} if and only if b2+c2=2a2b^2+c^2=2a^2 for some c∈Zc\in\mathbb{Z}. This conic has infinitely many integer solutions, e.g. (a,b,c,a)=(5,1,7),(13,7,17),(17,7,23),(25,31,17)(a,b,c,a)=(5,1,7),(13,7,17),(17,7,23),(25,31,17). The ratio R/a2=∣b2−a2∣/a2R/a^2=|b^2-a^2|/a^2 then records which regime of §1.2 the oval is in. In general, the factorization of Φe∗(z2−a2,β)\Phi_e^*(z^2-a^2,\beta) over Q\mathbb{Q} is a Diophantine question.

Theorem 5.5. For m,n≥2m,n\ge2,

Res⁡z(Qm,Qn)={β2(m−1)(n−1)=(b2−a2)2(m−1)(n−1),gcd⁡(m,n)=1,0,gcd⁡(m,n)>1.\operatorname{Res}_z(Q_m,Q_n)= \begin{cases} \beta^{2(m-1)(n-1)}=(b^2-a^2)^{2(m-1)(n-1)}, & \gcd(m,n)=1,\\ 0, & \gcd(m,n)>1. \end{cases}

Proof. By (17) and the chain rule (10) with h(z)=z2−a2h(z)=z^2-a^2, h0=1h_0=1, deg⁡h=2\deg h=2, we have Res⁡z(Qm,Qn)=Res⁡w(Sm,Sn)2\operatorname{Res}_z(Q_m,Q_n)=\operatorname{Res}_w(S_m,S_n)^2. The zeros of SmS_m are βζ\beta\zeta with ζm=1\zeta^m=1, ζ≠1\zeta\ne1. Since SmS_m is monic,

Res⁡w(Sm,Sn)=∏ζm=1ζ≠1Sn(βζ)=β(m−1)(n−1)∏ζm=1ζ≠1Ψn(ζ).\operatorname{Res}_w(S_m,S_n)=\prod_{\substack{\zeta^m=1\\\zeta\ne1}}S_n(\beta\zeta)=\beta^{(m-1)(n-1)}\prod_{\substack{\zeta^m=1\\\zeta\ne1}}\Psi_n(\zeta).

If e=gcd⁡(m,n)>1e=\gcd(m,n)>1, the factor with ζ=e2πi/e\zeta=e^{2\pi i/e} vanishes. If gcd⁡(m,n)=1\gcd(m,n)=1, the last product equals 11 by (12). □

Remark 5.6. Geometrically, QmQ_m and QnQ_n share a zero if and only if gcd⁡(m,n)>1\gcd(m,n)>1. In that case, by proposition 5.3, they share exactly the 2∑e∣gcd⁡(m,n)e>1φ(e)=2(gcd⁡(m,n)−1)2\sum_{\substack{e\mid\gcd(m,n)\\e>1}}\varphi(e)=2(\gcd(m,n)-1) zeros of Qgcd⁡(m,n)Q_{\gcd(m,n)}. When a2,b2∈Za^2,b^2\in\mathbb{Z} and gcd⁡(m,n)=1\gcd(m,n)=1, the resultant is the 2(m−1)(n−1)2(m-1)(n-1)-th power of the integer β\beta. This parallels Apostol’s theorem that Res⁡(Φm,Φn)=1\operatorname{Res}(\Phi_m,\Phi_n)=1 unless m/nm/n is a prime power [1], and the Jacobi-symbol formulas for Chebyshev resultants [10, 14].

Theorem 5.7. For n≥2n\ge2,

Disc⁡z(Qn)=4n−1n2n−4β2(n−1)(n−2)En,En=a2n−(a2−b2)nb2=∑k=0n−1a2k(a2−b2)n−1−k.\operatorname{Disc}_z(Q_n)=4^{n-1}n^{2n-4}\beta^{2(n-1)(n-2)}E_n,\qquad E_n=\frac{a^{2n}-(a^2-b^2)^n}{b^2}=\sum_{k=0}^{n-1}a^{2k}(a^2-b^2)^{n-1-k}.

Proof. Apply Lemma 2.1 with S=SnS=S_n and k=n−1k=n-1. By (17) and (9), Disc⁡w(Sn)=β(n−1)(n−2)Disc⁡x(Ψn)\operatorname{Disc}_w(S_n)=\beta^{(n-1)(n-2)}\operatorname{Disc}_x(\Psi_n). By (11), Disc⁡w(Sn)2=β2(n−1)(n−2)n2(n−2)\operatorname{Disc}_w(S_n)^2=\beta^{2(n-1)(n-2)}n^{2(n-2)}. Finally

(−4)n−1Sn(−a2)=(−4)n−1(−a2)n−βn−a2−β=4n−1a2n−(−β)nb2=4n−1En,(-4)^{n-1}S_n(-a^2)=(-4)^{n-1}\frac{(-a^2)^n-\beta^n}{-a^2-\beta}=4^{n-1}\frac{a^{2n}-(-\beta)^n}{b^2}=4^{n-1}E_n,

since β=b2−a2\beta=b^2-a^2.

For example, Disc⁡z(Q2)=4(2a2−b2)\operatorname{Disc}_z(Q_2)=4(2a^2-b^2), the discriminant of the quadratic in Remark 5.4.

Primitive prime divisors

Put A=a2A=a^2 and B=a2−b2B=a^2-b^2. Then En=(An−Bn)/(A−B)E_n=(A^n-B^n)/(A-B) is the Lucas sequence attached to the pair (A,B)(A,B). We recall Zsigmondy’s theorem [24]. If A>B≥1A>B\geq1 are coprime integers and n≥2n\geq2, then An−BnA^n-B^n has a prime divisor pp that divides no Ak−BkA^k-B^k with 1≤k<n1\leq k<n (a primitive divisor), except when n=2n=2 and A+BA+B is a power of 22, or (A,B,n)=(2,1,6)(A,B,n)=(2,1,6). A primitive divisor satisfies p≡1(modn)p\equiv1\pmod n, since the order of A/BA/B modulo pp is exactly nn.

Theorem 5.8. Let a2,b2∈Za^2,b^2\in\mathbb{Z} with b2<a2b^2<a^2, b2≠0b^2\neq0 and gcd⁡(a2,b2)=1\gcd(a^2,b^2)=1. Then for every n≥3n\geq3 with (a2,b2,n)≠(1,−1,6)(a^2,b^2,n)\neq(1,-1,6) there is a prime p≡1(modn)p\equiv1\pmod n such that

p∣Disc⁡z(Qn)andp∤Disc⁡z(Qm)for all 2≤m<n.p\mid\operatorname{Disc}_z(Q_n)\qquad\text{and}\qquad p\nmid\operatorname{Disc}_z(Q_m)\quad\text{for all }2\leq m<n.

Equivalently, modulo pp two zeros of QnQ_n collide, whereas for m<nm<n the zeros of QmQ_m remain distinct modulo pp.

Proof. Put A=max⁡(a2,a2−b2)A=\max(a^2,a^2-b^2) and B=min⁡(a2,a2−b2)B=\min(a^2,a^2-b^2). Both are positive integers, since b2<a2b^2<a^2 and a2≥1a^2\geq1. They are distinct, and gcd⁡(A,B)=gcd⁡(a2,b2)=1\gcd(A,B)=\gcd(a^2,b^2)=1. Also An−Bn=±(a2n−(a2−b2)n)A^n-B^n=\pm(a^{2n}-(a^2-b^2)^n), and (A,B)=(2,1)(A,B)=(2,1) exactly when (a2,b2)=(1,−1)(a^2,b^2)=(1,-1). By Zsigmondy’s theorem there is a primitive prime divisor p≡1(modn)p\equiv1\pmod n of An−BnA^n-B^n. It does not divide A−B=±b2A-B=\pm b^2, so p∣Enp\mid E_n, and therefore p∣Disc⁡z(Qn)p\mid\operatorname{Disc}_z(Q_n) by Theorem 5.7.

Let 2≤m<n2\leq m<n. We check the factors of Disc⁡z(Qm)=4m−1m2m−4β2(m−1)(m−2)Em\operatorname{Disc}_z(Q_m)=4^{m-1}m^{2m-4}\beta^{2(m-1)(m-2)}E_m one by one. Since p≡1(modn)p\equiv1\pmod n and n≥3n\geq3, pp is odd and p>n>mp>n>m, so p∤4m−1m2m−4p\nmid4^{m-1}m^{2m-4}. If p∣β=−(a2−b2)p\mid\beta=-(a^2-b^2), then p∣Bp\mid B or p∣Ap\mid A, and p∣An−Bnp\mid A^n-B^n forces p∣gcd⁡(A,B)=1p\mid\gcd(A,B)=1, which is impossible. By primitivity p∤Am−Bm=±b2Emp\nmid A^m-B^m=\pm b^2E_m, so p∤Emp\nmid E_m.

The final sentence holds because the discriminant of a polynomial with integer coefficients vanishes modulo pp if and only if its reduction modulo pp has a repeated factor, or its degree drops. The leading coefficient of QnQ_n is 11, so the degree does not drop. □

Remark 5.9. The exceptional case (a2,b2)=(1,−1)(a^2,b^2)=(1,-1), i.e. a=1a=1, b=ib=i, has En=2n−1E_n=2^n-1. Here Disc⁡z(Q6)\operatorname{Disc}_z(Q_6) has no new prime, because 26−1=32⋅72^6-1=3^2\cdot7 and 3∣22−13\mid2^2-1, 7∣23−17\mid2^3-1. The case b2>a2b^2>a^2 reduces to primitive divisors of a2n±(b2−a2)na^{2n}\pm(b^2-a^2)^n and can be treated in the same way with the Zsigmondy theorem for An+BnA^n+B^n.

Concluding remarks and open problems

In this paper, we have demonstrated that certain translates of real polynomials of the type fn(x)=(x2−1)nf_n(x)=(x^2-1)^n have their zeros on Cassini ovals. The results above show that Cassini ovals play for the family fn(x)=(x2−1)nf_n(x)=(x^2-1)^n the role that ellipses play for Chebyshev polynomials. Certain questions remain outstanding, and these come from the work of Oyengo [18, 7, 4]. That is: for λ2−κ2=α2>0\lambda^2-\kappa^2=\alpha^2>0 and ∣ω∣=1|\omega|=1 is there an explicitly defined ellipse on which the zeros of Un(z/α)−ωUn(iκ/α)U_n(z/\alpha)-\omega U_n(i\kappa/\alpha) lie? Oyengo [18] conjectured that then lie close to the ellipse x2/λ2+y2/κ2≤1x^{2}/\lambda^{2}+y^{2}/\kappa^{2}\le1. Corollary 4.5 settles the conjecture ∣Im⁡λ∣≤1|\operatorname{Im}\lambda|\le1 of [7] (Conjecture 1.1(3) in [4]) and the bound ∣Re⁡λ∣≤5|\operatorname{Re}\lambda|\le\sqrt{5} conjectured in [4].

For symmetric real-rooted ff with zeros in [−1,1][-1,1], Theorem 4.1 is sharp for fn(x)=(x2−1)nf_n(x)=(x^2-1)^n. What is the smallest region containing the zeros of f(z/a)−f(ib/a)f(z/a)-f(ib/a) for all such polynomials ff of a fixed degree. What is the analogous region for non-symmetric ff (Example 4.3).

References

Department of Mathematics, Maseno University, Private Bag, Maseno, Kenya

Email address: [email protected]

References

  1. [1]T. M. Apostol. Resultants of cyclotomic polynomials. Proceedings of the American Mathematical Society, 24:457–462, 1970.DOI
  2. [2]S. Beraha, J. Kahane, and N. J. Weiss. Limits of zeroes of recursively defined polynomials. Proceedings of the National Academy of Sciences of the United States of America, 72(11):4209, 1975.pubmed.ncbi.nlm.nih.gov/16592282
  3. [3]S. Beraha, J. Kahane, and N. J. Weiss. Limits of zeroes of recursively defined families of polynomials. In Studies in Foundations and Combinatorics, volume 1 of Advances in Mathematics Supplementary Studies, pages 213–232. Academic Press, New York, 1978.
  4. [4]R. Davis and N. Farnsworth. A note on conjectures of gullerud, johnson, and mbirika. arXiv preprint arXiv:2510.02008, 2025.arxiv.org/abs/2510.02008
  5. [5]K. Dilcher and K. B. Stolarsky. Nonlinear recurrences related to chebyshev polynomials. Ramanujan Journal, 2014. Published online.DOI
  6. [6]I. M. Gelfand, M. M. K 복ranov, and A. V. Zelevinsky. Discriminants, Resultants, and Multidimensional Determinants. Birkhäuser, Boston, 1994.DOI
  7. [7]E. Gullerud, R. Johnson, and aBa Mbirika. Tridiagonal real symmetric matrices with a connection to pascal’s triangle and the fibonacci sequence. arXiv preprint arXiv:2201.08490, 2023.arxiv.org/abs/2201.08490
  8. [8]A. E. Hirst and E. K. Lloyd. Cassini, his ovals, and a space probe to saturn. The Mathematical Gazette, 81(492):409–421, 1997.DOI
  9. [9]M. E. H. Ismail. Difference equations and quantized discriminants for q-orthogonal polynomials. Advances in Applied Mathematics, 30:562–589, 2003.DOI
  10. [10]D. P. Jacobs, M. O. Rayes, and V. Trevisan. The resultant of chebyshev polynomials. Canadian Mathematical Bulletin, 54(2):288–296, 2011.DOI
  11. [11]M. Karatas. A multi foci closed curve: Cassini oval, its properties and applications. Doğuş Üniversitesi Dergisi, 14(2):231–248, 2013.DOI
  12. [12]S.-H. Kim and J. H. Lee. On difference quotients of chebyshev polynomials. Bulletin of the Korean Mathematical Society, 53(2):373–, 2016.DOI
  13. [13]J. D. Lawrence. A Catalog of Special Plane Curves. Dover, New York, 1972.
  14. [14]S. R. Louboutin. Resultants of chebyshev polynomials: a short proof. Canadian Mathematical Bulletin, 56(3):602–605, 2013.DOI
  15. [15]M. Marden. Geometry of Polynomials, volume 3 of Mathematical Surveys. American Mathematical Society, Providence, RI, 2 edition, 1966.
  16. [16]J. C. Mason and D. C. Handscomb. Chebyshev Polynomials. Chapman and Hall/CRC, Boca Raton, FL, 2003.
  17. [17]J. H. McKay and S. S.-S. Wang. A chain rule for the resultant of two polynomials. Archiv der Mathematik, 53:347–351, 1989.DOI
  18. [18]M. O. Oyengo. Chebyshev-like polynomials, conic distribution of roots, and continued fractions. PhD thesis, University of Illinois at Urbana-Champaign, 2018.
  19. [19]Q. I. Rahman and G. Schmeisser. Analytic Theory of Polynomials, volume 26 of London Mathematical Society Monographs. New Series. Oxford University Press, Oxford, 2002.
  20. [20]T. J. Rivlin. Chebyshev Polynomials: From Approximation Theory to Algebra and Number Theory. Wiley, New York, 2 edition, 1990.
  21. [21]G. Szegő. Orthogonal Polynomials, volume 23 of American Mathematical Society Colloquium Publications. American Mathematical Society, Providence, RI, 4 edition, 1975.
  22. [22]K. Tran. Connections between discriminants and the root distribution of polynomials with rational generating function. Journal of Mathematical Analysis and Applications, 410:330–340, 2014.DOI
  23. [23]K. Tran. The root distribution of polynomials with a three-term recurrence. Journal of Mathematical Analysis and Applications, 421:878–892, 2015.DOI
  24. [24]K. Zsigmondy. Zur theorie der potenzreste. Monatshefte für Mathematik und Physik, 3:265–284, 1892.DOI

Paper details

vlnc:W_dfqmcs3v
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY-SA 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Mathematics

Not recorded for this paper

View statement
Looking for related papers…
Contents