betaFeedback

WorksMathematics

Root separation in one-parameter families of integer polynomials

Abstract

For every n≥2n\ge2 we construct irreducible integer polynomials of degree nn and arbitrarily large height HH with two roots at distance ≪nH−(4n−3)/6\ll_n H^{-(4n-3)/6}. All the roots can be chosen positive, or the two close roots can be chosen nonreal with all the other roots positive. We also construct monic irreducible examples. For a squarefree family F(T,X)∈Z[T,X]F(T,X)\in\mathbb{Z}[T,X] of positive degree in TT, let e(F)e(F) denote the exponent for which the least distance between two roots of F(t,X)F(t,X) is H(F(t,X))−e(F)+o(1)H(F(t,X))^{-e(F)+o(1)} as t→+∞t\to+\infty. Among families of degree nn in XX and degree three in TT, the maximum of e(F)e(F) is (2n−1)/3(2n-1)/3. It is (4n−3)/6(4n-3)/6 for families irreducible over Q(T)\mathbb{Q}(T), and (2n−2)/3(2n-2)/3 for monic families of odd degree n≥3n\ge3. For arbitrary degree in TT, we prove e(F)≤(3n−1)/4e(F)\le(3n-1)/4. Two roots in different irreducible factors over C(T)\mathbb{C}(T) have separation exponent strictly less than 2n/32n/3. If E(n)E(n) denotes the supremum of e(F)e(F) over squarefree families of degree nn in XX, then E(n)/nE(n)/n tends to sup⁡d≥2E(d)/d∈[2/3,3/4]\sup_{d\ge2}E(d)/d\in[2/3,3/4].

Use of artificial intelligence. The research was carried out under human direction from the choice of topic and research questions to the selection of results for the paper. OpenAI’s ChatGPT 6.2 Astra developed the constructions and proofs and drafted the manuscript. Anthropic’s Claude Fable 5.2 reviewed the manuscript, checked examples, and suggested improvements to the proofs and to the comparison with known results. The human review did not include an independent check of every proof.

Introduction

Let P∈Z[X]P \in\mathbb{Z}[X] be squarefree of degree n≥2n \ge2, with complex roots α1,…,αn\alpha_1,\ldots,\alpha_n. Its height H(P)H(P) is the largest absolute value of its coefficients, and its root separation is

sep⁡(P)=min⁡i<j∣αi−αj∣.\operatorname{sep}(P)=\min_{i<j}|\alpha_i-\alpha_j|.

For H(P)>1H(P)>1 define e(P)e(P) by sep⁡(P)=H(P)−e(P)\operatorname{sep}(P)=H(P)^{-e(P)}. Let e(n)e(n) be the upper limit of e(P)e(P) as H(P)→∞H(P)\to\infty over squarefree integer polynomials of degree nn. Write eirr(n)e_{\mathrm{irr}}(n) for the corresponding upper limit over irreducible polynomials and eirr∗(n)e_{\mathrm{irr}}^*(n) for the upper limit over monic irreducible polynomials.

Mahler’s inequality gives e(n)≤n−1e(n)\le n-1 [11]. For n≥4n\ge4, Bugeaud and Dujella constructed reducible polynomials giving e(n)≥(2n−1)/3e(n)\ge(2n-1)/3 [6 Theorem 1]. The two close roots lie in different factors, one of which is linear. For irreducible polynomials they obtained eirr(n)≥n/2+(n−2)/(4(n−1))e_{\mathrm{irr}}(n)\ge n/2+(n-2)/(4(n-1)) for n≥4n\ge4 [5 Theorem 1]. Our construction gives

eirr(n)≥4n−36(n≥2).e_{\mathrm{irr}}(n)\ge\frac{4n-3}{6}\qquad(n\ge2).

For n≥5n\ge5 this improves their bound for irreducible polynomials. It falls short of their reducible bound by 1/61/6.

The construction in § 2 starts with the cubic X=z(z−1)2X=z(z-1)^2. For small positive XX it has two inverse branches near z=1z=1. We construct a rational function T=B(z)/C(z)T=B(z)/C(z) so that the two values of 1/T1/T on these branches agree to high order. At a fixed large real value T=tT=t, the corresponding two values of XX are at distance of order ∣t∣−(2n−3/2)|t|^{-(2n-3/2)}, while the height of the polynomial is of order ∣t∣3|t|^3. One sign of tt gives nn positive roots. The other gives n−2n-2 positive roots and a pair of nonreal conjugates.

Using X=z3−zX=z^3-z in § 3, we obtain

eirr∗(n)≥{(2n−2)/3for odd n≥3,(2n−4)/3for even n≥4.e_{\mathrm{irr}}^*(n) \ge\begin{cases} (2n-2)/3 & \text{for odd } n\ge3, \\ (2n-4)/3 & \text{for even } n\ge4. \end{cases}

These bounds improve Bugeaud and Dujella’s bound eirr∗(n)≥n/2−1/4e_{\mathrm{irr}}^*(n) \ge n/2-1/4 for odd n≥5n\ge5 and even n≥8n\ge8 [6 Theorem 3]. For n=3n=3, Bugeaud and Mignotte proved eirr∗(3)≥3/2e_{\mathrm{irr}}^*(3)\ge3/2 [7 Theorem 4]. For reducible monic polynomials, Bugeaud and Dujella obtained (2n−3)/3(2n-3)/3 in even degrees n≥6n\ge6 and (2n−5)/3(2n-5)/3 in odd degrees n≥7n\ge7 [6 Theorem 2]. Our irreducible construction falls short of the former by 1/31/3 and exceeds the latter by one. In degrees 5, 7, 9, the reducible monic examples of Dujella and Pejković give 7/37/3, 17/517/5, 31/731/7, respectively [8 (3.2)]. Our monic irreducible examples give 8/38/3, 4, 16/316/3 in those degrees.

We now consider F∈Z[T,X]F\in\mathbb{Z}[T,X], squarefree in Q(T)[X]\mathbb{Q}(T)[X], of degree h≥1h\ge1 in TT and degree n≥2n\ge2 in XX. Define its separation exponent by

e(F)=lim⁡t→+∞−log⁡sep⁡(F(t,X))log⁡H(F(t,X)),e(F)=\lim_{t\to+\infty}\frac{-\log\operatorname{sep}(F(t,X))}{\log H(F(t,X))},

where tt runs through the positive integers. For all sufficiently large tt, the specialization is squarefree of degree nn and has height H(F(t,X))≍FthH(F(t,X))\asymp_F t^h. Here we use the height of the specialization itself. If the coefficients of FF in XX are coprime in Q[T]\mathbb{Q}[T], Bézout’s identity bounds their common divisor at every integer tt by a fixed integer, so passing to the primitive part gives the same exponent. To see that the limit exists, put u=1/Tu=1/T. By the Newton–Puiseux theorem [15 §3.1], the roots have convergent Laurent expansions in u1/qu^{1/q} near zero, with coefficients in a number field, for some positive integer qq. The difference of any two distinct roots has a first nonzero term cuscu^s. Their distance at T=tT=t is asymptotic to ∣c∣t−s|c|t^{-s}, so their separation exponent is s/hs/h. The number he(F)he(F) is the largest such order ss.

Let E(n)E(n) be the supremum of e(F)e(F) over these families, and write Ered(n)E_{\mathrm{red}}(n) and Eirr(n)E_{\mathrm{irr}}(n) for the suprema over reducible and irreducible families in Q(T)[X]\mathbb{Q}(T)[X], respectively. We call FF absolutely irreducible when it is irreducible in C[T,X]\mathbb{C}[T,X]. If FF is irreducible over Q(T)\mathbb{Q}(T), Hilbert’s irreducibility theorem, applied to its primitive part in Q[T][X]\mathbb{Q}[T][X], gives arbitrarily large positive integers tt for which F(t,X)F(t,X) is irreducible over Q\mathbb{Q} [18 Lemma 4]. Replacing TT by −T-T gives negative integers of arbitrarily large absolute value.

In the notation of Pejković [14], E(n)=epoly(n)E(n)=e^{\mathrm{poly}}(n). His results give Eirr(3)=E(3)=2E_{\mathrm{irr}}(3)=E(3)=2 and Ered(4)=E(4)=5/2E_{\mathrm{red}}(4)=E(4)=5/2 [14 Theorem 2.1 and Corollary 3.6]. In degree four he constructs an irreducible family with exponent 9/49/4, and proves e(F)≤5/2−1/(2h)e(F)\le5/2-1/(2h) for every squarefree family [14 Proposition 3.1 and Theorem 3.5]. In particular, his irreducible quartics improve on the value 13/613/6 given by Theorem 2.1 in that degree.

When h=3h=3, Theorem 4.3 gives the exact maxima (2n−1)/3(2n-1)/3 for all squarefree families and (4n−3)/6(4n-3)/6 for irreducible families. For odd n≥3n\ge3, Proposition 4.4 gives the monic maximum (2n−2)/3(2n-2)/3, attained by an absolutely irreducible family. The main estimate comes from an auxiliary polynomial whose first coefficients vanish when evaluated at one root. The order of the difference between two roots controls how many of the same coefficients vanish at the second root. The degree of the resultant limits the total order of vanishing.

For arbitrary hh, Zhuang’s bound over C(T)\mathbb{C}(T) gives e(F)≤n−1−n/(40n+2)e(F)\le n-1-n/(40n+2) for n≥4n\ge4 [20 Corollary 8.1.2]. Take the place at infinity and replace TT by TmT^m. This multiplies both the order of the root difference and the degree in TT by mm. Taking mmth roots and letting m→∞m \to\infty removes the multiplicative constant. Theorem 5.1 gives

e(F)≤3n−14−n−34h(n≥3).e(F) \le\frac{3n-1}{4} - \frac{n-3}{4h} \qquad(n \ge3).

If the two roots lie in different irreducible factors over C(T)\mathbb{C}(T), Theorem 5.3 gives s/h≤2n/3−1/hs/h \le2n/3 - 1/h for n≥3n \ge3. In the Bugeaud–Dujella examples h=3h=3 and equality holds. The two proofs use different properties of the curve F(T,X)=0F(T,X)=0. The first bounds how long two branches can agree in terms of the two degrees. For roots of different factors we use the number of intersections of the two curves, together with a theorem of Ruberman.

Finally, Theorem 6.2 shows that E(n)/nE(n)/n and Ered(n)/nE_{\mathrm{red}}(n)/n have the same limit, and that this limit equals lim sup⁡n→∞Eirr(n)/n\limsup_{n\to\infty} E_{\mathrm{irr}}(n)/n. Thus a single family of degree dd with e(F)>2d/3e(F)>2d/3 would improve the coefficient 2/32/3 in every sufficiently large degree. The same improvement would hold for absolutely irreducible families in infinitely many degrees.

Irreducible polynomials of every degree

Theorem 2.1. For every n≥2n \ge2 there is an absolutely irreducible polynomial F∈Z[T,X]F \in\mathbb{Z}[T,X] of degree three in TT and degree nn in XX with the following property. For all real tt of sufficiently large absolute value, two roots of F(t,X)F(t,X) are at distance ≍n∣t∣−(2n−3)/2\asymp_n |t|^{-(2n-3)/2}. The pair is real for one sign of tt and consists of nonreal conjugates for the other. The polynomial F(t,X)F(t,X) is irreducible over Q\mathbb{Q} for infinitely many integers tt of each sign. For integers tt of sufficiently large absolute value, its primitive part has height ≍n∣t∣3\asymp_n |t|^3. In particular,

eirr(n)≥4n−36.e_{\mathrm{irr}}(n) \ge\frac{4n-3}{6}.

Put f(z)=z(z−1)2f(z)=z(z-1)^2. For small positive xx, the equation f(ξ)=xf(\xi)=x has two roots u,vu,v near 11 and a third root zz near 00. We construct a rational function S=A/BS=A/B, with B(1)≠0B(1)\ne0, for which A(u)B(v)−A(v)B(u)=c(u−v)z2n−2A(u)B(v)-A(v)B(u)=c(u-v)z^{2n-2}. Thus S(u)−S(v)S(u)-S(v) vanishes to higher order than u−vu-v as x→0x\to0. The same identity for f(z)=z(z−1)(z+1)f(z)=z(z-1)(z+1) will be used in the monic case.

Lemma 2.2. Let d≥2d \ge2, N=2d−2N=2d-2 and f(z)=z(z−1)(z−ϵ)f(z)=z(z-1)(z-\epsilon), where ϵ∈{1,−1}\epsilon\in\{1,-1\}. For an indeterminate xx let

Vd={P∈Q[z]:deg⁡P≤d, ∑f(ξ)=xξNP(ξ)f′(ξ)=0}.V_d=\left\{P \in\mathbb{Q}[z]: \deg P \le d,\ \sum_{f(\xi)=x}\frac{\xi^N P(\xi)}{f'(\xi)}=0\right\}.

The space VdV_d has a basis consisting of polynomials of degrees dd and d−1d-1. Any two polynomials AA, BB forming a basis are coprime, and for some c∈Q∖{0}c \in\mathbb{Q}\setminus\{0\} we have

A(u)B(v)−A(v)B(u)=c(u−v)zN(1)A(u)B(v)-A(v)B(u)=c(u-v)z^N \tag*{(1)}

whenever z,u,vz,u,v are the three roots of f(ξ)=xf(\xi)=x.

Proof. Put Hj(x)=∑f(ξ)=xξj/f′(ξ)H_j(x)=\sum_{f(\xi)=x}\xi^j/f'(\xi). Setting z=1/tz=1/t in the partial fraction expansion of 1/(f(z)−x)1/(f(z)-x) gives

∑j≥0Hj(x)tj=t2(1−t)(1−ϵt)−xt3.\sum_{j\ge0} H_j(x)t^j=\frac{t^2}{(1-t)(1-\epsilon t)-xt^3}.

For 0≤j≤d0 \le j \le d the degree of HN+j(x)H_{N+j}(x) is at most d−2d-2. For P∈VdP \in V_d, we show that the coefficients of zd−1z^{d-1} and zdz^d determine the remaining coefficients. Equating the coefficients of 1,x,…,xd−21,x,\ldots,x^{d-2} in the definition of VdV_d gives a linear system whose matrix has entries [xr]HN+j[x^r]H_{N+j} for 0≤r,j≤d−20 \le r,j \le d-2.

For ϵ=1\epsilon=1, expansion of the generating function in powers of xx gives [xr]HN+j=(2d−3+j−r2r+1)[x^r]H_{N+j}=\binom{2d-3+j-r}{2r+1}. For ϵ=−1\epsilon=-1 the entry vanishes unless rr and jj have the same parity. For r=2i+er = 2i + e and j=2l+ej = 2l + e, with e∈{0,1}e \in\{0,1\}, it is

(d−2+l−i2i+e).\binom{d-2+l-i}{2i+e}.

Grouping even and odd indices separately in this case, we obtain in both cases square matrices of the form

((L+l−i2i+e))0≤i,l<k,2(k−1)+e≤L,e∈{0,1}.\left(\binom{L+l-i}{2i+e}\right)_{0\leq i,l<k}, \qquad2(k-1)+e\leq L,\qquad e\in\{0,1\}.

The (i,l)(i,l) entry counts paths from (−l,0)(-l,0) to (L−i,2i+e)(L-i,2i+e) with steps (1,0)(1,0) and (1,1)(1,1). By the Gessel–Viennot lemma [9], the determinant is a signed count of families of disjoint paths. Only the identity permutation of the endpoints can occur. Indeed, of two disjoint paths, the one starting farther left must be above the other wherever both are defined. If it ended at the lower endpoint, which lies farther to the right, this would already fail when the other path reached the higher endpoint. At least one such family exists. The iith path takes 2i+e2i+e diagonal steps followed by horizontal steps. Thus the determinant is positive. The two highest coefficients can now be prescribed arbitrarily, and every nonzero polynomial in VdV_d has degree at least d−1d-1.

Choose AA, BB of degrees dd, d−1d-1, respectively. For the three roots zz, uu, vv put

D(z)=A(u)B(v)−A(v)B(u)u−v.D(z)=\frac{A(u)B(v)-A(v)B(u)}{u-v}.

This is a symmetric polynomial in uu, vv of total degree at most NN. As u+v=1+ϵ−zu+v=1+\epsilon-z and uv=(z−1)(z−ϵ)uv=(z-1)(z-\epsilon), it is a polynomial in zz of degree at most NN. To prove (1), we first compare DD at two roots. In the equation defining VdV_d, take first P=AP=A and then P=BP=B. Multiply the first equation by B(v)B(v) and the second by A(v)A(v) and subtract. The term indexed by vv disappears. The two remaining terms give

zN(z−v)f′(z)D(u)+uN(u−v)f′(u)D(z)=0.\frac{z^N(z-v)}{f'(z)}D(u)+\frac{u^N(u-v)}{f'(u)}D(z)=0.

Here D(u)=(A(z)B(v)−A(v)B(z))/(z−v)D(u)=(A(z)B(v)-A(v)B(z))/(z-v). Cancelling the factors z−vz-v and u−vu-v gives uND(z)=zND(u)u^N D(z)=z^N D(u).

Let z→0z\to0 and choose another root uu of f(ξ)=f(z)f(\xi)=f(z) tending to 1. The equality D(z)/zN=D(u)/uND(z)/z^N=D(u)/u^N shows that D(z)/zND(z)/z^N is bounded at zero. As deg⁡D≤N\deg D\leq N, this gives D(z)=czND(z)=cz^N with c∈Qc\in\mathbb{Q}. To see that c≠0c\ne0, take large positive xx. The three roots of f(ξ)=xf(\xi)=x, divided by x1/3x^{1/3}, tend to the three cube roots of unity. Since A(ξ)/B(ξ)∼kξA(\xi)/B(\xi)\sim k\xi for some k≠0k\ne0, its three values are then distinct.

A common zero β\beta of AA, BB would make D(z)D(z) divisible by

(u−β)(v−β)=z2+(β−1−ϵ)z+(β−1)(β−ϵ).(u-\beta)(v-\beta)=z^2+(\beta-1-\epsilon)z+(\beta-1)(\beta-\epsilon).

If this monic quadratic divided czNcz^N, it would equal z2z^2. Its linear coefficient would then give β=1+ϵ\beta=1+\epsilon, for which the constant term is ϵ≠0\epsilon\ne0. Thus AA, BB are coprime. Changing the basis multiplies the left side of (1) by a nonzero rational constant. □\square

Proof of Theorem 2.1. Apply the lemma with d=nd=n and ϵ=1\epsilon=1. Since the basis polynomials are coprime, we may choose a basis CC, BB with C(1)=0C(1)=0 and B(1)≠0B(1)\ne0. Put S=C/BS=C/B. Near z=1z=1 choose the square root of zz with value 1 there. The function q=(z−1)zq=(z-1)\sqrt{z} has derivative 1 at z=1z=1. By the inverse function theorem we can write z=z(q)=1+q+O(q2)z=z(q)=1+q+O(q^2), and f(z(q))=q2f(z(q))=q^2. Write s(q)=S(z(q))s(q)=S(z(q)). Let z0(x)z_0(x) be the analytic solution of f(z0(x))=xf(z_0(x))=x with z0(0)=0z_0(0)=0. As f′(0)=1f'(0)=1, we have z0(x)=x+O(x2)z_0(x)=x+O(x^2). By (1),

s(q)−s(−q)=c(z(q)−z(−q))z0(q2)NB(z(q))B(z(−q)),N=2n−2.s(q)-s(-q)=\frac{c\bigl(z(q)-z(-q)\bigr)z_0(q^2)^N}{B(z(q))B(z(-q))},\qquad N=2n-2.

The right side has a zero of order 2N+12N+1 at q=0q=0. Since s(0)=0s(0)=0, the first nonzero term of s(q)s(q) cannot have degree one or three. A zero of order at least four would give C=(z−1)4PC=(z-1)^4P, which is impossible for n<4n<4. For n≥4n\ge4 the identity zNC=f(z)2zN−2Pz^N C=f(z)^2z^{N-2}P would give P∈Vn−1P\in V_{n-1} with deg⁡P≤n−4\deg P\le n-4, contrary to the lemma. Hence s(q)=aq2+O(q3)s(q)=aq^2+O(q^3) with a≠0a\ne0.

We now fix y=s(q)y=s(q) and compare the two corresponding values of xx. Put r=qs(q)/(aq2)r=q\sqrt{s(q)/(aq^2)}, taking the square root with value 11 at q=0q=0. Then y=ar2y=ar^2 and r=q+O(q2)r=q+O(q^2). Denote by J(q)J(q) the value of qq corresponding to −r-r. Thus J(q)=−q+O(q2)J(q)=-q+O(q^2) and s(J(q))=s(q)s(J(q))=s(q). For s(w)=aw2+O(w3)s(w)=aw^2+O(w^3) the difference quotient at two nearby points w1,w2w_1,w_2 is a(w1+w2)+O(max⁡(∣w1∣,∣w2∣)2)a(w_1+w_2)+O(\max(|w_1|,|w_2|)^2). Applying this with w1=−qw_1=-q and w2=J(q)w_2=J(q) gives

s(−q)−s(J(q))=(−q−J(q))(−2aq+O(q2)).s(-q)-s(J(q))=(-q-J(q))\left(-2aq+O(q^2)\right).

The left side has a zero of order 2N+12N+1, so q+J(q)q+J(q) has a zero of order 2N2N. The difference q2−J(q)2q^2-J(q)^2 therefore has order 2N+12N+1 in qq. Its absolute value is of order ∣y∣N+1/2|y|^{N+1/2} because y∼aq2y\sim aq^2. For small real yy with y/a>0y/a>0, both values of xx are real. For y/a<0y/a<0, the two values of rr are purely imaginary and conjugate. Since the inverse functions have real Taylor coefficients, the two distinct values of xx are then nonreal conjugates.

Clear the fixed denominators of B,CB,C. Form the polynomial

F(T,X)=∏f(ξ)=X(B(ξ)−TC(ξ)).(2)F(T,X)=\prod_{f(\xi)=X}(B(\xi)-TC(\xi)). \tag*{(2)}

The coefficients are symmetric polynomials with integer coefficients in the three roots of the monic cubic f(z)−Xf(z)-X, so they belong to Z[X]\mathbb{Z}[X]. The coefficient of T3T^3 is the nonzero product −∏f(ξ)=XC(ξ)-\prod_{f(\xi)=X}C(\xi). To see the degree in XX, factor B(z)−tC(z)=L∏j=1n(z−bj)B(z)-tC(z)=L\prod_{j=1}^{n}(z-b_j) for any tt for which the degree is nn. Since ∏f(ξ)=X(ξ−bj)=X−f(bj)\prod_{f(\xi)=X}(\xi-b_j)=X-f(b_j), we obtain

F(t,X)=L3∏j=1n(X−f(bj)).F(t,X)=L^3\prod_{j=1}^{n}(X-f(b_j)).

This holds for all but at most one tt, since max⁡(deg⁡B,deg⁡C)=n\max(\deg B,\deg C)=n. In particular, deg⁡XF=n\deg_X F=n. Putting t=1/yt=1/y in the preceding calculation now gives the asserted distance between two roots of F(t,X)F(t,X).

Suppose that F=GHF=GH in C[T,X]\mathbb{C}[T,X]. Substituting X=f(z)X=f(z) and T=B(z)/C(z)T=B(z)/C(z) gives zero, so one factor, say GG, vanishes identically after this substitution. For all but finitely many xx, the three roots ξ\xi of f(ξ)=xf(\xi)=x are distinct, and (1) gives three distinct values of B(ξ)/C(ξ)B(\xi)/C(\xi). All three are roots of G(T,x)G(T,x). It follows that deg⁡TG≥3\deg_TG\ge3. Thus HH depends only on XX. For each fixed xx, the product in (2) is a nonzero polynomial in TT, since BB and CC have no common zero. Hence HH has no complex zero and is constant. This proves absolute irreducibility.

In particular the coefficients of FF as a polynomial in XX are coprime in Q[T]\mathbb{Q}[T]. By the Euclidean algorithm, some Z[T]\mathbb{Z}[T]-linear combination of them is a nonzero integer mm. For integer tt the greatest common divisor of the coefficients of F(t,X)F(t,X) divides mm. Since H(F(t,X))∼n∣t∣3H(F(t,X))\sim_n |t|^3, its primitive part has the same order of height.

Hilbert’s irreducibility theorem, applied to F(T,X)F(T,X) and F(−T,X)F(-T,X), gives infinitely many irreducible specializations of each sign. Dividing 2n−3/22n-3/2 by 33 gives the lower bound for eirr(n)e_{\mathrm{irr}}(n).

Corollary 2.3. For all sufficiently large ∣t∣|t|, the polynomial in Theorem 2.1 has nn positive real roots for one sign of tt, and n−2n-2 positive real roots and two nonreal conjugate roots for the other.

Proof. We locate the other n−2n-2 roots by showing that CC has n−2n-2 simple zeros in (4/3,∞)(4/3,\infty). For z>4/3z>4/3, the other two solutions of f(ξ)=f(z)f(\xi)=f(z) are uu and u‾\overline{u}, where

u=1−z2+i2z(3z−4).u=1-\frac{z}{2}+\frac{i}{2}\sqrt{z(3z-4)}.

As zz increases from 4/34/3 to infinity, uu traces an arc in the upper half-plane starting at 1/31/3, with ∣u∣→∞\lvert u\rvert\to\infty and arg⁡u→2π/3\arg u\to2\pi/3. The polynomial CC has no zero on the open arc, since C(u)=0C(u)=0 would also give C(u‾)=0C(\overline{u})=0, contradicting (1). The defining equation for VnV_n gives

zNC(z)f′(z)=−2Re⁡(uNC(u)f′(u)),N=2n−2.\frac{z^N C(z)}{f'(z)}=-2\operatorname{Re}\left(\frac{u^N C(u)}{f'(u)}\right),\qquad N=2n-2.

We count zeros of the real part on the right by estimating the change in the argument of uNC(u)/f′(u)u^N C(u)/f'(u) along this arc.

For a real zero bb of CC, the argument of u−bu-b lies between 00 and π\pi and tends to 2π/32\pi/3. Thus its change is at least −π/3-\pi/3, including the case b=1/3b=1/3. For a nonreal conjugate pair b,b‾b,\overline{b}, the product (u−b)(u−b‾)(u-b)(u-\overline{b}) starts on the positive real axis and has imaginary part 2Im⁡(u)(Re⁡(u)−Re⁡(b))2\operatorname{Im}(u)(\operatorname{Re}(u)-\operatorname{Re}(b)). Since Re⁡(u)\operatorname{Re}(u) decreases strictly, this imaginary part can change sign only once, from positive to negative. The product never vanishes. Its continuous argument starts at zero and tends to −2π/3-2\pi/3 or 4π/34\pi/3. Summing over the zeros of CC with their multiplicities, we find that the change in the argument of C(u)C(u) is at least −nπ/3-n\pi/3.

Finally f′(u)=(u−z)(u−u‾)f'(u)=(u-z)(u-\overline{u}). The second factor has constant argument π/2\pi/2, and the argument of the first changes from π\pi to 5π/65\pi/6. Hence the net change of the argument of uNC(u)/f′(u)u^N C(u)/f'(u) is at least

2Nπ3−nπ3+π6=(n−2)π+5π6.\frac{2N\pi}{3}-\frac{n\pi}{3}+\frac{\pi}{6}=(n-2)\pi+\frac{5\pi}{6}.

By continuity, the argument takes at least n−2n-2 distinct values congruent to π/2\pi/2 modulo π\pi in the interior of the arc. The real part vanishes at each of these points. It follows that CC has at least n−2n-2 distinct zeros in (4/3,∞)(4/3,\infty). Its double zero at 11 accounts for the rest of its degree, so all n−2n-2 zeros are simple. For each such zero bb, the function C/BC/B has a simple zero at bb, since B(b)≠0B(b)\ne0. Its local inverse gives a real solution of C(z)/B(z)=1/tC(z)/B(z)=1/t tending to bb as ∣t∣→∞\lvert t\rvert\to\infty. The corresponding roots of F(t,X)F(t,X) tend to the distinct positive numbers f(b)f(b), as ff is strictly increasing on (4/3,∞)(4/3,\infty). When the remaining two roots are real, they too are positive, since x=q2x=q^2 with qq real and nonzero. □

An irreducible sextic for every odd parameter

For v=z−1v=z-1 we have x=v2(1+v)x=v^2(1+v). When n=6n=6 we may take

C(v)=v2(34−272v+780v2−938v3+391v4),C(v)=v^2(34-272v+780v^2-938v^3+391v^4),
B(v)=19−171v+572v2−852v3+495v4−7v5−46v6.B(v)=19-171v+572v^2-852v^3+495v^4-7v^5-46v^6.

For the three roots v,u,wv,u,w of v2(1+v)=xv^2(1+v)=x, substitution of u+w=−1−vu+w=-1-v and uw=v(v+1)uw=v(v+1) gives

C(u)B(w)−C(w)B(u)u−w=−45885(1+v)10.\frac{C(u)B(w)-C(w)B(u)}{u-w}=-45885(1+v)^{10}.

Since C(v)/B(v)∼(34/19)v2C(v)/B(v)\sim(34/19)v^2 at zero, the proof of Theorem 2.1 gives a sextic of height ∼∣t∣3\sim\lvert t\rvert^3 with two roots at distance ∼∣t∣−21/2\sim\lvert t\rvert^{-21/2}. For large positive tt this pair is positive. The quartic C(v)/v2C(v)/v^2 has alternating signs at 1/3,2/5,1/2,1,21/3,2/5,1/2,1,2, beginning with a positive value. Its four distinct positive zeros give the four remaining roots as in Corollary 2.3. For odd tt the polynomial B(v)−tC(v)B(v)-tC(v) reduces modulo two to g(v)=v6+(v+1)5g(v)=v^6+(v+1)^5. Put b=X2+Xb=X^2+X and a=b+1a=b+1. At each root of v3+v2=Xv^3+v^2=X we have vg(v)=av+bvg(v)=av+b, by direct substitution. Taking the product over the three roots and using their elementary symmetric functions gives

XF(t,X)≡b3+ab2+Xa3=X(X6+X5+1)(mod2).XF(t,X) \equiv b^{3}+ab^{2}+Xa^{3}=X(X^{6}+X^{5}+1)\pmod{2}.

Thus F(t,X)F(t,X) reduces to X6+X5+1X^{6}+X^{5}+1. It has no root in F2\mathbb{F}_{2}. The only remaining possible irreducible factors of degree at most three are X2+X+1X^{2}+X+1, X3+X+1X^{3}+X+1 and X3+X2+1X^{3}+X^{2}+1. Indeed, a cubic over F2\mathbb{F}_{2} is irreducible precisely when it has no root there. Division by the three displayed polynomials gives the respective nonzero remainders X+1X+1, X+1X+1 and X2X^{2}. Thus the sextic is irreducible modulo two. It follows that F(t,X)F(t,X) is irreducible over Q\mathbb{Q} for every odd integer tt.

Monic irreducible polynomials

Theorem 3.1. For every n≥3n\geq3 there is an absolutely irreducible F∈Z[T,X]F\in\mathbb{Z}[T,X], monic of degree nn in XX and of degree three in TT, with the following property. For all sufficiently large positive tt, all roots of F(t,X)F(t,X) are real and the distance between two of them is ≍nt−Mn\asymp_n t^{-M_n}, where

Mn={2n−2,n odd,2n−4,n even.M_n= \begin{cases} 2n-2, & n\text{ odd},\\ 2n-4, & n\text{ even}. \end{cases}

There is also an absolutely irreducible family, again monic in XX and of the same degrees, for which a pair of nonreal conjugate roots is at distance ≍nt−Mn\asymp_n t^{-M_n} for all sufficiently large positive tt. In both families the specialized polynomial is irreducible over Q\mathbb{Q} for infinitely many positive integers tt. Consequently eirr∗(n)≥Mn/3e_{\mathrm{irr}}^{*}(n)\geq M_n/3.

Proof. It suffices to arrange that deg⁡B=n>deg⁡C\deg B=n>\deg C. Clear the denominators of BB and CC. The leading coefficient LL of B−TCB-TC in zz is then a nonzero integer independent of TT. By (2), the leading coefficient of FF in XX is A=L3A=L^{3}. Thus An−1F(T,X/A)A^{n-1}F(T,X/A) is monic in XX and has integer coefficients.

Take f(z)=z3−zf(z)=z^{3}-z and let d≥3d\geq3 be odd. By Lemma 2.2, VdV_d has a basis consisting of an odd polynomial BB of degree dd and an even polynomial CC of degree d−1d-1. Indeed, the even and odd parts of each element of VdV_d belong to VdV_d, and every nonzero element has degree at least d−1d-1. At x=0x=0 the defining equation for CC gives C(1)=C(−1)=0C(1)=C(-1)=0. These zeros are simple. Otherwise C=(z2−1)2PC=(z^{2}-1)^{2}P, which is impossible when d=3d=3. For d≥5d\geq5, the identity z2d−2C=f(z)2z2d−4Pz^{2d-2}C=f(z)^{2}z^{2d-4}P would give P∈Vd−1P\in V_{d-1} with deg⁡P≤d−5\deg P\leq d-5, contrary to the lemma. Since BB and CC are coprime and B(0)=0B(0)=0, we have B(±1)≠0B(\pm1)\neq0 and C(0)≠0C(0)\neq0.

Let z0,z+,z−z_0,z_+,z_- be the three solutions of f(z)=xf(z)=x near x=0x=0, numbered so that z0(0)=0z_0(0)=0, z+(0)=1z_+(0)=1 and z−(0)=−1z_-(0)=-1. Put s±(x)=C(z±(x))/B(z±(x))s_{\pm}(x)=C(z_{\pm}(x))/B(z_{\pm}(x)). By (1),

s+(x)−s−(x)=c(z+(x)−z−(x))z0(x)2d−2B(z+(x))B(z−(x)).s_+(x)-s_-(x)=\frac{c(z_+(x)-z_-(x))z_0(x)^{2d-2}}{B(z_+(x))B(z_-(x))}.

Since z0z_0 has a simple zero, s+−s−s_+-s_- vanishes to order 2d−22d-2 at zero. Both s+s_+ and s−s_- have simple zeros there. Their derivatives at zero are therefore equal and nonzero. This common derivative is rational, so we may multiply CC by a rational constant to make it one. If p=s+−1(y)p=s_+^{-1}(y) and q=s−−1(y)q=s_-^{-1}(y), then s−(p)−s−(q)=(s−−s+)(p)s_-(p)-s_-(q)=(s_--s_+)(p). As p,q→0p,q\to0, the quotient (s−(p)−s−(q))/(p−q)(s_-(p)-s_-(q))/(p-q) tends to s−′(0)=1s_-'(0)=1. Since pp has a simple zero in yy, p−qp-q vanishes to order 2d−22d-2 in yy. Taking y=1/ty=1/t and d=nd=n gives the asserted distance for odd nn.

For even n≥4n\geq4 take d=n−1d=n-1 and replace C,BC,B by

C1(z)=zC(z),B1(z)=zB(z)+C(z)z2−1.C_1(z)=zC(z),\qquad B_1(z)=zB(z)+\frac{C(z)}{z^{2}-1}.

These polynomials have degrees dd and d+1d+1 and are coprime. Indeed, B1(0)=−C(0)≠0B_1(0)=-C(0)\ne0. At z=±1z=\pm1, the equalities s±′(0)=1s_\pm'(0)=1 and f′(±1)=2f'(\pm1)=2 give C′(z)=2B(z)C'(z)=2B(z) and hence B1(z)=2zB(z)≠0B_1(z)=2zB(z)\ne0. At every other zero of C1C_1 we have B1(z)=zB(z)≠0B_1(z)=zB(z)\ne0. On the two inverse branches the values of C1/B1C_1/B_1 are xs±(x)/(x+s±(x))xs_\pm(x)/(x+s_\pm(x)). Both have derivative 1/21/2 at zero, and their difference is

x2(s+(x)−s−(x))(x+s+(x))(x+s−(x)).\frac{x^2\left(s_+(x)-s_-(x)\right)}{\left(x+s_+(x)\right)\left(x+s_-(x)\right)}.

This difference vanishes to order 2d−2=2n−42d-2=2n-4. Since both derivatives are nonzero, the inverse functions give the asserted distance as in the odd case.

We next show that all roots are real for sufficiently large ∣t∣|t|. In both constructions d=2m+1d=2m+1 is odd, so C(z)=P(z2)C(z)=P(z^2) for a polynomial PP of degree mm. For z>2/3z>2/\sqrt{3} the other two solutions of f(u)=f(z)f(u)=f(z) are u,u‾u,\overline{u}, where Im⁡u>0\operatorname{Im}u>0. Set

w=u‾ 2=1−z22+iz3z2−42.w=\overline{u}^{\,2}=1-\frac{z^2}{2}+iz\frac{\sqrt{3z^2-4}}{2}.

As zz increases, ww describes an arc in the upper half-plane from 1/31/3 to infinity. Its real part decreases strictly and its argument tends to 2π/32\pi/3. By (1), P(w)P(w) has no zero on the open arc. The estimate in the proof of Corollary 2.3 therefore gives Δarg⁡P(w)≥−mπ/3\Delta\arg P(w)\ge-m\pi/3. The defining equation for VdV_d gives

z4mC(z)f′(z)=−2Re⁡(w2mP(w)3w−1).\frac{z^{4m}C(z)}{f'(z)}=-2\operatorname{Re}\left(\frac{w^{2m}P(w)}{3w-1}\right).

The argument of 3w−13w-1 changes from π/2\pi/2 to 2π/32\pi/3. Hence the change in the argument of the quotient is at least 2m(2π/3)−mπ/3−π/6=mπ−π/6>(m−1)π2m(2\pi/3)-m\pi/3-\pi/6=m\pi-\pi/6>(m-1)\pi. Its real part vanishes at least m−1m-1 times. Thus CC has at least m−1m-1 distinct zeros in (2/3,∞)(2/\sqrt{3},\infty). Their negatives and ±1\pm1 account for all 2m2m zeros of CC, so each is simple. In even degree C1=zCC_1=zC has the additional simple zero 00. Since the denominator has degree one more than the numerator, C/BC/B (or C1/B1C_1/B_1) also has a simple zero at infinity, as can be seen by putting z=1/wz=1/w and expanding at w=0w=0. For sufficiently small real 1/t1/t, each of these simple real zeros gives one real solution of C/B=1/tC/B=1/t (or C1/B1=1/tC_1/B_1=1/t). Their images under ff are the roots of the polynomial in (2).

Write S=C/BS=C/B in odd degree and S=C1/B1S=C_1/B_1 in even degree. In both cases SS is odd. Hence f~(w)=if(iw)=w3+w\widetilde{f}(w)=if(iw)=w^3+w and S~(w)=iS(iw)\widetilde{S}(w)=iS(iw) have rational coefficients and the same degrees as ff and SS. Indeed, every odd rational function can be written as zR(z2)zR(z^2) for some rational function RR. The two zeros of S~\widetilde{S} at w=±iw=\pm i give inverse branches interchanged by complex conjugation. For real 1/t1/t near zero, their f~\widetilde{f}-values are distinct complex conjugates. The substitutions multiply both coordinates by nonzero constants, so the difference vanishes to the same order as before.

For odd nn, identity (1) shows that SS takes three distinct values at the three roots of f(z)=xf(z)=x for all but finitely many xx. The same holds for even nn, since the map s↦xs/(x+s)s\mapsto xs/(x+s) is one-to-one for x≠0x\ne0 wherever it is defined. The substitutions using ii preserve this property. Thus the factorization argument following (2) proves absolute irreducibility in all cases. Hilbert’s irreducibility theorem gives infinitely many positive integer specializations that are irreducible over Q\mathbb{Q}. The change of scale at the beginning of the proof multiplies all root distances by ∣A∣|A| and preserves the separation exponent. □

Sharp bounds for families of degree three in the parameter

Let FF have degree hh in TT and degree nn in XX. Divide the coefficients of FF, considered as a polynomial in XX, by their greatest common divisor in Q[T]\mathbb{Q}[T], and denote the quotient by F0F_0. Put a=deg⁡TF0≤ha=\deg_T F_0\le h. If a=0a=0, all the roots are constant and e(F)=0e(F)=0. Otherwise, choose an integer cc at which the coefficient of TaT^a in F0F_0 does not vanish. With u=1/Tu=1/T and v=1/(X−c)v=1/(X-c), the polynomial

P(u,v)=uavnF0(1/u,c+1/v)P(u,v)=u^a v^n F_0(1/u,c+1/v)

has degree aa in uu, since F0(0,X)F_0(0,X) is nonzero by primitivity. The coefficient of vnv^n in P(0,v)P(0,v) is the nonzero value used to choose cc, and P(u,0)P(u,0) is nonzero because deg⁡XF0=n\deg_X F_0=n. Thus PP has degree nn in vv, and neither uu nor vv divides it. The substitutions are invertible when uv≠0uv\ne0. Since neither TT nor X−cX-c divides F0F_0, the irreducible factors of F0F_0 correspond to those of PP, over Q\mathbb{Q} as well as over C\mathbb{C}. All roots of PP stay bounded as u→0u\to0, since P(0,v)P(0,v) has degree nn. For two roots we have

v1−v2=X2−X1(X1−c)(X2−c).v_1-v_2=\frac{X_2-X_1}{(X_1-c)(X_2-c)}.

If e(F)>0e(F)>0, take a pair for which the difference in XX has order he(F)he(F). The transformed difference has the same order for a finite common limit, and greater order at a common pole. In either case its order MM satisfies M≥he(F)M\ge he(F).

For later use, let P(u,v)P(u,v) be squarefree of degree dd in uu and b≥2b\ge2 in vv, with deg⁡vP(0,v)=b\deg_v P(0,v)=b. Its discriminant with respect to vv is a nonzero polynomial of degree at most 2d(b−1)2d(b-1) in uu, by homogeneity in the coefficients of PP. All roots v1,…,vbv_1,\ldots,v_b of PP are bounded at zero, where the leading coefficient of PP is nonzero. Hence

∑i<jord⁡u(vi−vj)≤d(b−1).(3)\sum_{i<j}\operatorname{ord}_u(v_i-v_j)\le d(b-1). \tag*{(3)}

We recall the two properties of the resultant used below. If PP has leading coefficient p(u)p(u) and roots v1,…,vbv_1,\ldots,v_b, then

Res⁡v(P,G)=p(u)deg⁡vG∏j=1bG(u,vj).\operatorname{Res}_v(P,G)=p(u)^{\deg_v G}\prod_{j=1}^{b}G(u,v_j).

For nonzero GG, the resultant vanishes identically exactly when P,GP,G have a common factor of positive degree in vv. Otherwise its degree in uu is at most bdeg⁡uG+(deg⁡vG)deg⁡uPb\deg_u G+(\deg_v G)\deg_u P. Indeed, each term of the Sylvester determinant uses deg⁡vG\deg_v G coefficients of PP and bb of GG.

Lemma 4.1. Let P∈C[u,v]P\in\mathbb{C}[u,v] be irreducible, of degree d≥1d\ge1 in uu and b≥2b\ge2 in vv, and suppose that P(0,v)P(0,v) has degree bb. Let MM be the order at u=0u=0 of the difference of two roots. If M>d(b−1)/2M>d(b-1)/2, then 2M2M is an integer and

⌊M⌋≤(d−1)(b−1).\lfloor M\rfloor\le(d-1)(b-1).

Proof. Since M>d(b−1)/2M>d(b-1)/2, (3) shows that only one pair has order MM. The roots have convergent expansions in u1/qu^{1/q} for a common qq. Multiplication of u1/qu^{1/q} by e2πi/qe^{2\pi i/q} permutes the roots and preserves the order of every difference. Hence it fixes or interchanges the two chosen roots.

For an integer mm consider the polynomials G∈C[u,v]G\in\mathbb{C}[u,v] of degree at most mm in uu and at most b−1b-1 in vv. Their coefficients form a vector space of dimension D=b(m+1)D=b(m+1). For every nonzero such GG, irreducibility of PP and deg⁡vG<b\deg_v G<b give

R(u)=Res⁡v(P,G)≠0,deg⁡R≤bm+d(b−1).R(u)=\operatorname{Res}_v(P,G)\ne0,\qquad\deg R\le bm+d(b-1).

If both roots are fixed, only integral powers of uu occur, so MM is an integer. Choose mm so large that D−1≥MD-1\ge M. We can take a nonzero GG so that the first D−1D-1 coefficients of G(u,v1)G(u,v_1) vanish. Since G(u,v1)−G(u,v2)G(u,v_1)-G(u,v_2) is divisible by v1−v2v_1-v_2, the order of G(u,v2)G(u,v_2) is at least MM. All other factors in the product for RR are bounded at zero. Thus D−1+M≤bm+d(b−1)D-1+M\le bm+d(b-1), giving M≤(d−1)(b−1)M\le(d-1)(b-1) in this case.

::: If the roots are interchanged, write them as v(t)v(t) and v(−t)v(-t), where t2=ut^2 = u. Indeed, two applications of the substitution fix each root, so only powers of u1/2u^{1/2} occur. The first odd term of v(t)v(t) has degree 2k+1=2M2k+1=2M. In every series G(t2,v(t))G(t^2,v(t)) the coefficients of t,t3,…,t2k−1t,t^3,\ldots,t^{2k-1} vanish. Take mm so large that L=D−1+k>2kL=D-1+k>2k. Requiring the series to vanish to order at least LL therefore imposes at most L−k=D−1L-k=D-1 homogeneous linear conditions. Choose a nonzero solution GG. The product G(t2,v(t))G(t2,v(−t))G(t^2,v(t))G(t^2,v(-t)) then vanishes to order at least LL in uu. Hence D−1+k≤bm+d(b−1)D-1+k\le bm+d(b-1), which gives k≤(d−1)(b−1)k\le(d-1)(b-1).

For every h≥2h\ge2, Lemma 4.1 gives the exact maximum 2−3/(2h)2-3/(2h) for the separation exponent of an absolutely irreducible cubic family of degree hh in the parameter. Indeed, the order MM is at most max⁡{h,2h−3/2}=2h−3/2\max\{h,2h-3/2\}=2h-3/2. To attain it, take FhF_h from Theorem 2.1 with deg⁡XFh=h\deg_X F_h=h and put

Gh(T,X)=ThX3Fh(1/X,1/T).G_h(T,X)=T^hX^3F_h(1/X,1/T).

The polynomial FhF_h is absolutely irreducible and divisible by neither variable. Interchanging the variables and taking reciprocals therefore gives an absolutely irreducible GhG_h of degrees h,3h,3 in T,XT,X, respectively. In the notation of the construction, 1/T=q21/T=q^2, and the two roots s(q),s(−q)s(q),s(-q) of GhG_h have difference of order 4h−34h-3 in qq. Hence e(Gh)=(4h−3)/(2h)e(G_h)=(4h-3)/(2h).

Lemma 4.2. Let F∈Z[T,X]F\in\mathbb{Z}[T,X] be squarefree over Q(T)\mathbb{Q}(T), of degree h≥2h\ge2 in TT and degree n≥2n\ge2 in XX. Then

he(F)≤(h−1)(n−1)+1.he(F)\le(h-1)(n-1)+1.

Proof. We may assume e(F)>0e(F)>0 and use the bounded coordinate vv. For two roots of one irreducible factor, Lemma 4.1 bounds the order of their difference by max⁡{h(n−1)/2,(h−1)(n−1)+1/2}\max\{h(n-1)/2,(h-1)(n-1)+1/2\}. For roots of two distinct factors of degrees (d1,b1)(d_1,b_1) and (d2,b2)(d_2,b_2), the resultant gives M≤d1b2+d2b1M\le d_1b_2+d_2b_1. If one factor is independent of uu, it is linear in vv and M≤hM\le h. Otherwise all four degrees are positive and

(d1+d2−1)(b1+b2−1)+1−(d1b2+d2b1)=(d1−1)(b1−1)+(d2−1)(b2−1)≥0.(d_1+d_2-1)(b_1+b_2-1)+1-(d_1b_2+d_2b_1) = (d_1-1)(b_1-1)+(d_2-1)(b_2-1)\ge0.

Since d1+d2≤hd_1+d_2\le h and b1+b2≤nb_1+b_2\le n, both cases give the asserted bound.

Theorem 4.3. Let F∈Z[T,X]F\in\mathbb{Z}[T,X] be squarefree over Q(T)\mathbb{Q}(T), with deg⁡TF=3\deg_T F=3 and deg⁡XF=n≥2\deg_X F=n\ge2. Then

e(F)≤2n−13.e(F)\le\frac{2n-1}{3}.

If FF is irreducible over Q(T)\mathbb{Q}(T), then

e(F)≤4n−36.e(F)\le\frac{4n-3}{6}.

Both upper bounds are attained for every n≥2n\ge2.

Proof. The first bound is Lemma 4.2 with h=3h=3. For the second assume that FF is irreducible over Q(T)\mathbb{Q}(T) and e(F)>0e(F)>0. Choose two roots in the bounded coordinate vv whose difference has order M≥3e(F)M\ge3e(F). If they lie in the same irreducible factor over C\mathbb{C} of degrees (d,b)(d,b), Lemma 4.1 gives

M≤max⁡{d(b−1)/2,(d−1)(b−1)+1/2}≤2n−3/2.M\le\max\{d(b-1)/2,(d-1)(b-1)+1/2\}\le2n-3/2.

If they lie in distinct factors, recall that the primitive polynomial PP in u,vu,v is irreducible over Q\mathbb{Q}. Its irreducible factors over C\mathbb{C} are conjugate over Q\mathbb{Q}. To see this, substitute v=ua+1v=u^{a+1}, where a=deg⁡uP≤3a=\deg_u P\le3. The monomials in each factor give distinct powers of uu, so the factor can be recovered from the resulting univariate polynomial. Normalize the latter to be monic. Its coefficients are algebraic numbers, being symmetric functions of roots of a polynomial in Q[u]\mathbb{Q}[u]. Thus each normalized bivariate factor has algebraic coefficients. The product of the distinct Galois conjugates of such a factor is rational and hence exhausts the factorization of PP. If there are r>1r>1 factors, each has degrees (a/r,n/r)(a/r,n/r). Since rr divides a≤3a\leq3, we have a/r=1a/r=1. The resultant then gives M≤2n/r≤nM\leq2n/r\leq n.

The general bound is attained by the Bugeaud–Dujella families for n≥4n\geq4 [6 proof of Theorem 1]. Their construction also covers n=2,3n=2,3. Put D=T2+3T+1D=T^2+3T+1 and x=(T+2)/Dx=(T+2)/D. For j=1,2j=1,2 the family of degree j+1j+1 is (DX−T−2)pj(X)(DX-T-2)p_j(X), where

p1(X)=(T+1)X−1,p2(X)=TX2+TX−1.p_1(X)=(T+1)X-1,\qquad p_2(X)=TX^2+TX-1.

In both cases pj(x)=(−1)j−1D−jp_j(x)=(-1)^{j-1}D^{-j}. The positive root yjy_j of pjp_j tends to zero, and pj′(w)∼Tp_j'(w)\sim T uniformly between xx and yjy_j. Thus ∣x−yj∣∼T−(2j+1)|x-y_j|\sim T^{-(2j+1)}, as required. Theorem 2.1 gives equality in the irreducible case for all n≥2n\geq2. □

Proposition 4.4. Let F∈Z[T,X]F\in\mathbb{Z}[T,X] be squarefree over Q(T)\mathbb{Q}(T), monic of degree n≥3n\geq3 in XX, and of degree three in TT. Then

e(F)≤2n−23.e(F)\leq\frac{2n-2}{3}.

For every odd n≥3n\geq3, equality is attained by an absolutely irreducible family.

Proof. Put N=2n−2N=2n-2. Suppose that two roots have difference of order greater than NN in u=1/Tu=1/T. If they lie in one irreducible factor over C[T,X]\mathbb{C}[T,X] of degrees (d,b)(d,b), Lemma 4.1 gives the bound max⁡{d(b−1)/2,(d−1)(b−1)+1/2}\max\{d(b-1)/2,(d-1)(b-1)+1/2\} after the change to vv. It exceeds NN only if (d,b)=(3,n)(d,b)=(3,n). For roots in different factors the resultant bound is at most three if one factor is independent of TT. Otherwise their degrees can be ordered so that d1=1d_1=1 and d2≤2d_2\leq2. Then b2+d2b1≤2n−b2b_2+d_2b_1\leq2n-b_2, which can exceed NN only for (d1,b1)=(1,n−1)(d_1,b_1)=(1,n-1) and (d2,b2)=(2,1)(d_2,b_2)=(2,1). Both factors can be taken monic in XX. The second then has a polynomial root of degree two in TT, whereas every root of the first is O(∣T∣)O(|T|) by Cauchy’s bound. It remains to consider an absolutely irreducible FF.

Let MM be the order after the change to v=1/(X−c)v=1/(X-c). Lemma 4.1 and its proof show that M=N+1/2M=N+1/2 and that the two expansions are v(t),v(−t)v(t),v(-t), with u=t2u=t^2. The first odd term of v(t)v(t) has degree 2N+12N+1. To see that tt is a local parameter on this branch, let p(t2)p(t^2) be the Taylor polynomial of v(t)v(t) through degree 2N2N. Then (v−p(u))/uN(v-p(u))/u^N has order one in tt. We have v(0)≠0v(0)\neq0. Indeed, if v(t)v(t) had a zero of order l≥1l\geq1 in tt, the original difference would have order M−l<NM-l<N in uu.

We show that uu and XX are rational functions of a single parameter. In the function field L=C(u,v)L=\mathbb{C}(u,v), a nonconstant function qq has [L:C(q)][L:\mathbb{C}(q)] zeros and equally many poles, counted with multiplicity on all branches, including those at infinity [19 Theorem 1.4.11]. Thus uu and vv have nn and three poles, respectively, counted with multiplicity. Every nonzero polynomial G(u,v)G(u,v) of degrees at most m,n−1m,n-1 has at most K=nm+3(n−1)K=nm+3(n-1) zeros, since its total pole order is at most KK. The space of these polynomials has dimension D=n(m+1)D=n(m+1), and K=D−1+NK=D-1+N. Take mm with K−1>2NK-1>2N. Requiring G(t2,v(t))G(t^2,v(t)) to vanish to order at least K−1K-1 imposes at most K−1−N=D−2K-1-N=D-2 conditions, since its coefficients of t,t3,…,t2N−1t,t^3,\ldots,t^{2N-1} already vanish. Choose two independent solutions G0,G1G_0,G_1. Their ratio z=G1/G0z=G_1/G_0 is nonconstant because deg⁡vP=m\deg_v P=m. For all but finitely many λ∈C\lambda\in\mathbb{C}, the zeros of z−λz-\lambda lie away from the chosen branch and the poles of G0G_0. They are additional zeros of G1−λG0G_1-\lambda G_0. Hence

K−1+[L:C(z)]≤K,K-1+[L:\mathbb{C}(z)]\leq K,

which proves L=C(z)L=\mathbb{C}(z).

Every pole of XX lies over u=0u=0, since FF is monic in XX. If there were two distinct poles, both would give (u,v)=(0,0)(u,v)=(0,0). Choose a nonzero linear combination of G0,G1G_0,G_1 vanishing at (0,0)(0,0). It would have at least K−1K-1 zeros at the chosen branch and one at each pole, since v(0)≠0v(0) \ne0 on that branch. This exceeds KK. Take the unique pole of XX as z=∞z=\infty. Since [L:C(X)]=3[L:\mathbb{C}(X)]=3, XX is a cubic polynomial in zz. At the chosen branch X(t)=c+1/v(t)X(t)=c+1/v(t) has first odd term of degree 2N+12N+1, and X(t)−X(0)X(t)-X(0) has order at most three. Its order is therefore two. Affine changes of zz and XX now give X=f(z)=z(z−1)2X=f(z)=z(z-1)^2, with the chosen branch at z=1z=1. Write u=C(z)/B(z)u=C(z)/B(z) in lowest terms. This rational function has degree nn, a double zero at 11, and a zero at infinity. In particular deg⁡C≤n−1\deg C\le n-1.

Take the three solutions z,ξ,ηz,\xi,\eta of f(s)=xf(s)=x with z→0z\to0 and ξ,η→1\xi,\eta\to1 as x→0x\to0. The two values of uu near 11 have difference of order N+1/2N+1/2 in xx. Indeed, in the parameter tt we have x=A(t2)+t2N+1E(t2)x=A(t^2)+t^{2N+1}E(t^2) with A′(0)E(0)≠0A'(0)E(0)\ne0. At fixed xx the two parameter values satisfy t2/t1→−1t_2/t_1\to-1. Subtracting the expressions for xx gives t12−t22∼−2E(0)t12N+1/A′(0)t_1^2-t_2^2\sim-2E(0)t_1^{2N+1}/A'(0), as required. As in Lemma 2.2,

D(z)=C(ξ)B(η)−C(η)B(ξ)ξ−ηD(z)=\frac{C(\xi)B(\eta)-C(\eta)B(\xi)}{\xi-\eta}

is a polynomial of degree at most NN. Since ξ−η\xi-\eta has order 1/21/2 in xx, B(1)≠0B(1)\ne0 and z∼xz\sim x, we obtain D(z)=c0zND(z)=c_0z^N with c0≠0c_0\ne0. Multiply the numerator identity by C(z)C(z) and sum cyclically over z,ξ,ηz,\xi,\eta. The left side cancels. Using f′(z)(ξ−η)=f′(ξ)(η−z)=f′(η)(z−ξ)f'(z)(\xi-\eta)=f'(\xi)(\eta-z)=f'(\eta)(z-\xi) on the right gives

∑f(s)=xsNC(s)f′(s)=0.\sum_{f(s)=x}\frac{s^N C(s)}{f'(s)}=0.

Thus CC belongs to the space VnV_n of Lemma 2.2 with complex coefficients and vanishes at 11. Its basis polynomials are coprime, so this condition defines a one-dimensional subspace. The proof of Corollary 2.3 gives a polynomial of degree nn spanning that subspace. This contradicts deg⁡C≤n−1\deg C\le n-1. Theorem 3.1 gives equality for odd nn. □

Upper bounds for arbitrary degrees

Theorem 5.1. Let F∈Z[T,X]F\in\mathbb{Z}[T,X] be squarefree over Q(T)\mathbb{Q}(T), of degree h≥1h\ge1 in TT and degree n≥2n\ge2 in XX. Then e(F)≤(3n−1)/4e(F)\le(3n-1)/4. For n≥3n\ge3 we have

e(F)≤3n−14−n−34h.e(F)\le\frac{3n-1}{4}-\frac{n-3}{4h}.

A curve with local equation y2=zj+1y^2=z^{j+1} has a singularity of type AjA_j for j≥1j\ge1. Let K(D)K(D) be the largest jj that occurs on a plane curve of total degree DD. Gusein-Zade and Nekhoroshev proved 15/28≤lim sup⁡D→∞K(D)/D2≤3/415/28\le\limsup_{D\to\infty}K(D)/D^2\le3/4 [10]. Orevkov raised the lower bound to 7/127/12 [13 §5]. Astashov also constructed examples with prescribed degrees in each variable [1 Theorem 4]. We need an upper bound in terms of those two degrees that also applies when other branches pass through the same point.

Lemma 5.2. Let P∈C[u,v]P\in\mathbb{C}[u,v] be squarefree, with degrees a,n≥1a,n\ge1 in u,vu,v, respectively. Suppose that in analytic coordinates z,yz,y centred at a point of P=0P=0, its equation has a factor y2−zky^2-z^k, where k≥2k\ge2. Then

k−1≤⌊3an−a−n+12⌋.k-1\le\left\lfloor\frac{3an-a-n+1}{2}\right\rfloor.

Proof. Make a fractional linear change in each variable, keeping the given point finite. Choose the new fibres at 00 and ∞\infty in each variable to avoid the singular points, to meet the curve transversely and to have all four of their crossings off the curve. This is possible since a reduced curve has only finitely many exceptional fibres in each projection. After clearing denominators, the two degrees are still a,na,n. All four corner coefficients of the rectangle [0,a]×[0,n][0,a]\times[0,n] are nonzero, and the coefficients of uau^a and of vnv^n have only simple nonzero

roots in the other variable. The polynomial is therefore convenient and nondegenerate at infinity. It follows that it is cohomologically tame [17 p. 199], hence spectrally tame [4 Proposition 3.2.5]. The zero fibre has only isolated singularities because it is reduced. We can thus apply the spectral inequality of [4 Theorem 4.3.1].

For 1≤i<a1 \le i < a and 1≤j<n1 \le j < n put wij=max⁡{i/a,j/n}w_{ij} = \max\{i/a, j/n\}. Form a list SS containing wijw_{ij} and 2−wij2-w_{ij} for each pair, together with a+n−1a+n-1 copies of 11. For any irrational 0<x<10 < x < 1, the spectral inequality gives

#{1≤j<k:x<1/2+j/k<x+1}≤#{s∈S:x<s<x+1}.\#\{1 \le j < k : x < 1/2 + j/k < x+1\} \le\#\{s \in S : x < s < x+1\}.

where entries are counted with multiplicity. On the right, SS is the spectrum at infinity in the convention with entries in (0,2)(0,2), by [12 Theorem 5.16]. The list can also be read directly from [3 Table 1, $e=0$] by counting the pairs with each given value of max⁡{i/a,j/n}\max\{i/a,j/n\}.

On the left, 1/2+j/k1/2+j/k are the entries for y2−zky^2-z^k [2 p. 252]. If the local equation is (y2−zk)H(y^2-z^k)H with H(0,0)=0H(0,0)=0, replace HH by H+εH+\varepsilon. For small ε≠0\varepsilon\ne0 the equation at the origin has just the indicated factor. Local semicontinuity [4 Theorem 4.1.1] therefore bounds its contribution by that of the original equation.

Take x=1/2+ηx=1/2+\eta irrational, with η>0\eta>0 smaller than 1/k1/k and than every nonzero ∣wij−1/2∣\lvert w_{ij}-1/2\rvert. All k−1k-1 local entries lie in (x,x+1)(x,x+1). The global list has 2an−a−n+12an-a-n+1 entries. By its symmetry about 11, we must subtract one entry for each wij≤1/2w_{ij}\le1/2 and one for each wij<1/2w_{ij}<1/2. The total number subtracted is

⌊a2⌋⌊n2⌋+⌊a−12⌋⌊n−12⌋=⌊(a−1)(n−1)2⌋.\left\lfloor\frac{a}{2}\right\rfloor\left\lfloor\frac{n}{2}\right\rfloor+\left\lfloor\frac{a-1}{2}\right\rfloor\left\lfloor\frac{n-1}{2}\right\rfloor=\left\lfloor\frac{(a-1)(n-1)}{2}\right\rfloor.

This gives the asserted bound. □

Proof of Theorem 5.1. Suppose e(F)>0e(F)>0. Use the polynomial P(u,v)P(u,v) of Section 4, with u=1/Tu=1/T and v=1/(X−c)v=1/(X-c). Its degrees are 1≤a≤h1\le a\le h and nn, and its roots are bounded at u=0u=0. Let MM be the largest order of their pairwise differences. Then he(F)≤Mhe(F)\le M. For n=2n=2, the discriminant bound (3) gives M≤aM\le a and hence e(F)≤1e(F)\le1.

Let n≥3n\ge3. If 2M≤a(n−1)2M\le a(n-1), then 4M≤2a(n−1)≤3an−a−n+34M\le2a(n-1)\le3an-a-n+3. Otherwise (3) shows that exactly one unordered pair attains MM. Continuation around u=0u=0 preserves this pair. Its mean m(u)m(u) and its squared difference are therefore convergent power series in uu. Write the latter as 4ukW(u)4u^kW(u) with k=2Mk=2M an integer and W(0)≠0W(0)\ne0. Taking a local square root of WW, the coordinates z=uz=u and y=(v−m(u))/W(u)y=(v-m(u))/\sqrt{W(u)} give the factor y2−zky^2-z^k. Lemma 5.2 gives 4M≤3an−a−n+34M\le3an-a-n+3 in this case as well. Since he(F)≤Mhe(F)\le M and a≤ha\le h, the result follows. □

For roots in different factors, the resultant bounds the order of their difference by the total number of intersections of the two curves, counted with multiplicity. When the order ss exceeds hn/2hn/2, a small translation splits the contact into ss simple intersections. A second bound follows by enclosing these intersections in a single disc and applying a theorem of Ruberman.

Theorem 5.3. Let F∈Z[T,X]F\in\mathbb{Z}[T,X] be squarefree over Q(T)\mathbb{Q}(T), of degree h≥1h\ge1 in TT and n≥2n\ge2 in XX. If two roots belong to different irreducible factors over C(T)\mathbb{C}(T), the order of their difference in u=1/Tu=1/T is strictly less than 2hn/32hn/3. For n≥3n\ge3 it is at most 2hn/3−12hn/3-1.

Proof. We may assume that the order is positive. Pass to the bounded coordinate vv of Section 4, and denote the order of the transformed difference by ss. It is at least the original order. Let p,qp,q be the degrees of the two factors A,BA,B in uu, and let m,ℓm,\ell be their degrees in vv, respectively. Put

I=pℓ+qm,J=pm+qℓ.I=p\ell+qm,\qquad J=pm+q\ell.

Their resultant has degree at most II in uu. Its order at zero is the sum of the orders of all mℓm\ell differences between a root of AA and a root of BB, since both leading coefficients are nonzero there. In particular, s≤Is \le I and I+J=(p+q)(m+ℓ)≤hnI + J = (p + q)(m + \ell) \le hn. If n=2n = 2, then m=ℓ=1m = \ell= 1 and I≤hI \le h. A factor independent of uu is linear in vv and again I≤hI \le h. These cases give both asserted bounds whenever applicable. We may therefore take p,q≥1p,q \ge1 and n≥3n \ge3. Then hn≥6hn \ge6, so s≤hn/2s \le hn/2 suffices.

Suppose that s>hn/2s > hn/2. Since 2s>I2s > I, the chosen pair is the only one from the two factors whose difference has order at least ss. Continuation around zero preserves each factor and hence fixes each root of the pair. Write the two roots as convergent power series f(u),g(u)f(u),g(u). In particular, ss is an integer. Also I>JI > J, so we may label the factors so that p>qp > q and ℓ>m\ell> m.

We first separate the chosen pair from all the other roots on a disc containing ss intersections. Write f−g=usE(u)f-g=u^{s}E(u) with E(0)≠0E(0) \ne0. Choosing a holomorphic ss-th root of EE near zero gives a local coordinate z=uE(u)1/sz=uE(u)^{1/s} in which f−g=zsf-g=z^{s}. Translate the curve BB by ε>0\varepsilon> 0 in the vv-coordinate. The graphs of ff and g+εg+\varepsilon now meet at the ss solutions of zs=εz^{s}=\varepsilon. Put ρ=ε1/s\rho=\varepsilon^{1/s}. In the annulus ρ/2<∣z∣<2ρ\rho/2 < |z| < 2\rho, these two graphs are at distance at most (2s+1)ε(2^{s}+1)\varepsilon in each fibre. Every other root of either factor is farther away. Indeed, before translation its difference from ff has order λ<s\lambda< s. For a root of BB this follows from uniqueness of the pair. For another root of AA apply uniqueness to its difference from gg and use f−g=zsf-g=z^{s}. Its distance from ff on the annulus is therefore ≫ρλ\gg\rho^{\lambda}, uniformly on its Puiseux branches. As ε→0\varepsilon\to0, this is arbitrarily large compared with ε\varepsilon, even after the translation.

Choose ε\varepsilon sufficiently small. Cut the annulus along a ray between two successive solutions and take a closed disc DD with smooth boundary in the cut annulus containing all ss solutions in its interior. Over DD the tube ∣v−f∣≤(2s+2)ε|v-f| \le(2^{s}+2)\varepsilon contains just the two chosen graphs, one from each curve. We can now perturb the coefficients of both equations slightly and generically, keeping their bidegrees. In P1(C)×P1(C)\mathbb{P}^{1}(\mathbb{C}) \times\mathbb{P}^{1}(\mathbb{C}) the resulting curves are smooth, connected and transverse to each other by Bertini’s theorem for the two systems of positive bidegree. Inside the tube each still has one simple root over every point of DD. Their difference stays nonzero on ∂D\partial D and, by Rouché’s theorem, still has ss zeros in DD. There are II intersections in all, by the intersection formula for curves of the two bidegrees.

Reverse the orientation of the second curve. Remove the two graph discs over DD and join their boundaries by the straight segment between the two roots in each fibre over ∂D\partial D. These segments form an embedded annulus. The tube is free of other sheets, and the two boundary orientations agree with this gluing. The surface is now connected. At each of the remaining I−sI-s transverse intersections take local complex coordinates x,yx,y in which the original sheets are the axes and the second is the yy-axis. After its orientation is reversed, the local smoothing xy‾=δx\overline{y}=\delta, for small δ≠0\delta\ne0, joins the two sheets by an oriented annulus. Replacing two small discs by this annulus adds one handle to the connected surface. A general curve of bidegree (p,m)(p,m) projects to the uu-line with degree mm and 2p(m−1)2p(m-1) simple branch points, counted by its homogeneous discriminant. Riemann–Hurwitz gives its genus (p−1)(m−1)(p-1)(m-1), and likewise (q−1)(ℓ−1)(q-1)(\ell-1) for the second curve. Thus the resulting closed embedded oriented surface has genus

(p−1)(m−1)+(q−1)(ℓ−1)+I−s(p-1)(m-1)+(q-1)(\ell-1)+I-s

The first replacement takes place in the contractible tube over DD, and each later one in a small ball. Thus the projection degrees of the resulting surface are the differences m−ℓm-\ell and p−qp-q. By Reuberman’s theorem [16 Corollary 1.3], an embedded surface with nonzero projection degrees i,ji,j has genus at least (∣i∣−1)(∣j∣−1)(|i|-1)(|j|-1). Consequently

(p−q−1)(ℓ−m−1)≤(p−1)(m−1)+(q−1)(ℓ−1)+I−s(p-q-1)(\ell-m-1) \le(p-1)(m-1)+(q-1)(\ell-1)+I-s

This gives s≤2J−2m−2q+1≤2J−3s \le2J - 2m - 2q + 1 \le2J - 3. For example, when m=q=1m = q = 1 the two bounds for a contact of order s>hn/2s > hn/2 are pℓ+1p\ell+ 1 and 2(p+ℓ)−32(p+\ell) - 3. Adding 2s≤2I2s \le2I to s≤2J−3s \le2J - 3 yields 3s≤2(I+J)−3≤2hn−33s \le2(I+J) - 3 \le2hn - 3, as required. □

Proposition 5.4. Let FF be as in Theorem 5.1. Suppose two roots belonging to distinct irreducible factors over C(T)\mathbb{C}(T) have difference of order greater than hn/2hn/2 in u=1/Tu = 1/T. Both factors can then be chosen over Q(T)\mathbb{Q}(T), and both roots have convergent Laurent expansions in uu with rational coefficients. Consequently a pair in distinct factors with at least one root nonreal for arbitrarily large positive TT has difference of order at most hn/2hn/2.

If FF is irreducible over Q(T)\mathbb{Q}(T) and has r≥2r \ge2 irreducible factors over C(T)\mathbb{C}(T), then e(F)≤2n/r2e(F) \le2n/r^2.

Proof. Use the bounded coordinate vv of § 4 and let a≤ha \le h be the parameter degree after removing the common factor of the coefficients. The transformed pair has order M>hn/2M > hn/2. The discriminant bound (3) gives a total order at most a(n−1)<hna(n-1) < hn, so every other pair has smaller order. Continuation around zero fixes the unordered pair and preserves each factor. The two roots therefore have ordinary power series in uu.

The coefficients of these series are algebraic numbers by the Newton–Puiseux theorem [15 §3.1]. Every conjugation over Q\mathbb{Q} preserves the order of each difference and thus fixes the unordered pair. If a conjugation interchanged its roots, the two factors would have the same degrees p,dp,d in u,vu,v. Their resultant would give M≤2pd≤an/2M \le2pd \le an/2, since 2p≤a2p \le a and 2d≤n2d \le n. Thus each root and its factor are fixed. The coefficients of the series and of the factors, after one coefficient of each factor is scaled to one, are rational. The formula X=c+1/vX = c+1/v gives the asserted Laurent expansions, which take real values for all sufficiently large positive TT.

For the last assertion the rr factors are conjugate over Q\mathbb{Q}, as in the proof of Theorem 4.3, so each has degrees a/r,n/ra/r,n/r. For roots in different factors the resultant bounds the order by 2an/r22an/r^2. For roots of one factor (3) bounds it by (a/r)(n/r−1)(a/r)(n/r - 1). Division by hh gives the result. □

From one degree to all large degrees

Orevkov used power substitutions and a linear change of variables to produce AjA_j-singularities on plane curves of increasing total degree [13 Proposition 5.2]. Here we increase the degree in XX while keeping the degree in TT fixed. The order of a single root difference gives a lower bound in the new degrees.

Proposition 6.1. Let F∈Z[T,X]F \in\mathbb{Z}[T,X] be squarefree over Q(T)\mathbb{Q}(T), with deg⁡XF=b≥2\deg_X F = b \ge2 and e(F)>b/2e(F) > b/2. For every sufficiently large nn there is a reducible squarefree family of degree nn in XX and degree bb in TT with exponent at least

e(F)bn−OF(1).\frac{e(F)}{b}n - O_F(1).

If FF is irreducible over Q(T)\mathbb{Q}(T), the same estimate holds for absolutely irreducible families of degree bb in TT along an infinite arithmetic progression of degrees nn. If e(F)≥2b/3e(F) \ge2b/3, then FF has an absolutely irreducible factor A∈Z[T,X]A \in\mathbb{Z}[T,X] with e(A)≥e(F)e(A) \ge e(F).

Proof. Put h=deg⁡TFh = \deg_T F and use the bounded coordinate of § 4. The resulting polynomial G(u,v)G(u,v) has degrees a,ba,b, where 1≤a≤h1 \le a \le h, and its leading coefficient in vv is nonzero at u=0u = 0. Its coefficients in vv have no common zero in uu, since the original equation was divided by the common factor of its coefficients. Choose a pair of roots of FF whose difference has order he(F)he(F), and denote the corresponding roots of GG by v1,v2v_1,v_2. Then

M=ord⁡u(v1−v2)≥he(F)>ab/2.M = \operatorname{ord}_u(v_1-v_2) \ge he(F) > ab/2.

By (3), the sum of the orders of all pairs is at most a(b−1)a(b-1). Every other pair therefore has order strictly less than MM. Applying any conjugation over Q\mathbb{Q} to the algebraic coefficients of the Puiseux expansions preserves all orders and thus fixes the unordered pair {v1,v2}\{v_1,v_2\}. In particular its common value θ\theta at u=0u=0 is rational. Continuation around u=0u=0 also fixes the pair, so it either fixes or interchanges its two members. Thus v1v_1 and v2v_2 have power series in u1/2u^{1/2}.

If v1,v2v_1,v_2 belong to the same absolutely irreducible factor, every conjugation fixes that factor once a coefficient is normalized to one. Its coefficients are algebraic, as in the proof of Theorem 4.3, and hence rational. When e(F)≥2b/3e(F) \ge2b/3, Theorem 5.3 shows that the two original roots do belong to the same factor. Its primitive equation A∈Z[T,X]A \in\mathbb{Z}[T,X] divides FF, and 1≤deg⁡TA≤h1 \le\deg_T A \le h. The two roots give e(A)≥he(F)/deg⁡TA≥e(F)e(A) \ge he(F)/\deg_T A \ge e(F).

Choose λ∈Q∖{0}\lambda\in\mathbb{Q} \setminus\{0\} such that G(λ,V)G(\lambda,V) has degree bb and has only simple roots. There are only finitely many excluded values. Put

G~(U,V)=(1+U)aG(λU1+U,V).\widetilde{G}(U,V)=(1+U)^aG\left(\frac{\lambda U}{1+U},V\right).

For any even integer r>br>b with r≥4r\ge4, set

ϕ(x)=λxr1+xr,Pr(T,X)=TbG~(Xr,θ+1T−X).\phi(x)=\frac{\lambda x^r}{1+x^r},\qquad P_r(T,X)=T^b\widetilde{G}\left(X^r,\theta+\frac{1}{T}-X\right).

After choosing a square root of λ/(1+xr)\lambda/(1+x^r) near zero, the two functions fi(x)=x+vi(ϕ(x))−θf_i(x)=x+v_i(\phi(x))-\theta have convergent power series x+O(x2)x+O(x^2). Let xi(y)x_i(y) be their local inverses. Then xi(y)∼yx_i(y)\sim y, and

x1(y)−x2(y)∼f2(x1(y))−f2(x2(y))=(v2−v1)(ϕ(x1(y))).x_1(y)-x_2(y)\sim f_2(x_1(y))-f_2(x_2(y)) =(v_2-v_1)\left(\phi(x_1(y))\right).

The last expression has order rMrM in yy. Taking y=1/ty=1/t gives two roots of Pr(t,X)P_r(t,X) at distance ≍F,rt−rM\asymp_{F,r}t^{-rM}.

To compute the degrees, put Hr(X,S)=G~(Xr,θ+S−X)H_r(X,S)=\widetilde{G}(X^r,\theta+S-X). The coefficient of UaU^a in G~(U,V)\widetilde{G}(U,V) is G(λ,V)G(\lambda,V), of degree bb. The highest power of XX in HrH_r is therefore Xar+bX^{ar+b}, with nonzero constant coefficient. In particular Hr(X,0)≠0H_r(X,0)\ne0. Also deg⁡SHr=b\deg_S H_r=b, by the substitution for VV. It follows that PrP_r has degrees bb in TT and ar+bar+b in XX, and that Pr(0,X)≠0P_r(0,X)\ne0.

The polynomial G~(Xr,V)\widetilde{G}(X^r,V) is squarefree over C(X)\mathbb{C}(X) because the discriminant of G(z,V)G(z,V) is nonzero at the transcendental element z=ϕ(X)z=\phi(X). Its coefficients have no common zero in XX. For 1+Xr≠01+X^r\ne0 this follows from the corresponding property of GG. At Xr=−1X^r=-1 the polynomial specializes to (−λ)a(-\lambda)^a times the nonzero coefficient of uau^a in G(u,V)G(u,V). Thus it is squarefree in C[X,V]\mathbb{C}[X,V]. The substitution V=θ+S−XV=\theta+S-X is invertible, and replacing SS by 1/T1/T preserves squarefreeness after denominators are cleared since Pr(0,X)≠0P_r(0,X)\ne0. After multiplication by a nonzero integer, PrP_r belongs to Z[T,X]\mathbb{Z}[T,X] and has exponent at least rM/brM/b.

Given a sufficiently large nn, take the largest even rr for which ar+b<nar+b<n. Then k=n−ar−bk=n-ar-b satisfies 1≤k≤2a1\le k\le2a. Multiplication by kk distinct factors X−ciX-c_i, with integers cic_i chosen outside the constant roots of PrP_r, produces a reducible squarefree family of the required degrees. Its exponent is at least rM/b≥(n−b−k)e(F)/brM/b\ge(n-b-k)e(F)/b.

Finally suppose FF is irreducible over Q(T)\mathbb{Q}(T). Since e(F)>b/2e(F)>b/2, Proposition 5.4 shows that GG is absolutely irreducible. Take a root vv of G(z,V)G(z,V) and put L=C(z,v)L=\mathbb{C}(z,v). At a simple root of G(λ,V)G(\lambda,V) the implicit function theorem gives an embedding L⊂C((z−λ))L\subset\mathbb{C}((z-\lambda)). The polynomial Wr−(λ−z)/zW^r-(\lambda-z)/z is Eisenstein at z−λz-\lambda, so it is irreducible over C((z−λ))\mathbb{C}((z-\lambda)) and hence over LL. If wr=z/(λ−z)w^r=z/(\lambda-z), it follows that [L(w):L]=[C(w):C(z)]=r[L(w):L]=[\mathbb{C}(w):\mathbb{C}(z)]=r. As [L:C(z)]=b[L:\mathbb{C}(z)]=b, we obtain [L(w):C(w)]=b[L(w):\mathbb{C}(w)]=b. Thus G~(wr,V)\widetilde{G}(w^r,V) is irreducible over C(w)\mathbb{C}(w). Its coefficients have no common zero, so it is irreducible in C[w,V]\mathbb{C}[w,V]. The same changes of variables as above now give the absolute irreducibility of PrP_r. Taking n=ar+bn=ar+b proves the estimate along the stated progression. □

Theorem 6.2. We have

lim⁡n→∞E(n)n=lim⁡n→∞Ered(n)n=lim sup⁡n→∞Eirr(n)n=sup⁡d≥2E(d)d.(4)\lim_{n\to\infty}\frac{E(n)}{n}=\lim_{n\to\infty}\frac{E_{\mathrm{red}}(n)}{n}=\limsup_{n\to\infty}\frac{E_{\mathrm{irr}}(n)}{n}=\sup_{d\ge2}\frac{E(d)}{d}. \tag*{(4)}

The common value belongs to [2/3,3/4][2/3,3/4]. It exceeds 2/32/3 if and only if some absolutely irreducible family A∈Z[T,X]A\in\mathbb{Z}[T,X] satisfies e(A)>2deg⁡X(A)/3e(A)>2\deg_X(A)/3. Moreover, E(n)≤2n/3+O(1)E(n)\le2n/3+O(1) as n→∞n\to\infty if and only if E(d)≤2d/3E(d)\le2d/3 for every d≥2d\ge2.

Proof. Write C=sup⁡d≥2E(d)/dC=\sup_{d\ge2}E(d)/d. Since E(3)/3=2/3E(3)/3=2/3 and Theorem 5.1 gives E(d)/d≤3/4E(d)/d\le3/4, we have 2/3≤C≤3/42/3\le C\le3/4. If 1/2<c<C1/2<c<C, choose a family FF of some degree dd with e(F)>cde(F)>cd. Proposition 6.1 gives lim inf⁡n→∞Ered(n)/n≥e(F)/d>c\liminf_{n\to\infty}E_{\mathrm{red}}(n)/n\ge e(F)/d>c. Since Ered(n)≤E(n)≤CnE_{\mathrm{red}}(n)\le E(n)\le Cn, the first two limits equal CC.

The absolutely irreducible examples of Theorem 2.1 give lim sup⁡n→∞Eirr(n)/n≥2/3\limsup_{n\to\infty}E_{\mathrm{irr}}(n)/n\ge2/3. If C>2/3C>2/3, choose 2/3<c<C2/3<c<C and a family FF of degree dd with e(F)>cde(F)>cd. Proposition 6.1 gives an absolutely irreducible factor AA with e(A)>cd≥cdeg⁡XAe(A)>cd\ge c\deg_X A. Applying the same proposition to AA gives absolutely irreducible families in infinitely many degrees with lim sup⁡n→∞Eirr(n)/n≥e(A)/deg⁡XA>c\limsup_{n\to\infty}E_{\mathrm{irr}}(n)/n\ge e(A)/\deg_X A>c. Letting cc tend to CC proves (4). Conversely, every absolutely irreducible family with e(A)>2deg⁡X(A)/3e(A)>2\deg_X(A)/3 gives C>2/3C>2/3 by definition.

If E(n)≤2n/3+O(1)E(n)\le2n/3+O(1), the first limit in (4) gives C≤2/3C\le2/3. The definition of CC then gives E(d)≤2d/3E(d)\le2d/3 for all d≥2d\ge2, and the converse follows at once. □

Funding. T. P. acknowledges support from the Croatian Science Foundation under project IP-2022-10-5008 (TEBAG). He also acknowledges support from the project “Implementation of cutting-edge research and its application as part of the Scientific Center of Excellence for Quantum and Complex Systems, and Representations of Lie Algebras” (Grant No. PK.1.1.10.0004), co-financed by the European Union through the European Regional Development Fund under the Competitiveness and Cohesion Programme 2021–2027. This research was funded by the European Union through NextGenerationEU under the National Recovery and Resilience Plan 2021–2026 (institutional grant of the Faculty of Science, University of Zagreb, IK IA 1.1.3. Impact4Math).

References

References

  1. [1]E. A. Astashov, A_k-singularities of curves and surfaces of fixed degree, quasi-degree, or multi-degree, Moscow Univ. Math. Bull. 70 (2015), 241–247.
  2. [2]M. Borodzik, M. Hedden and C. Livingston, Plane algebraic curves of arbitrary genus via Heegaard Floer homology, Comment. Math. Helv. 92 (2017), 215–256.
  3. [3]M. Borodzik and T. K. Moe, Topological obstructions for rational cuspidal curves in Hirzebruch surfaces, Michigan Math. J. 65 (2016), 761–797.
  4. [4]M. Borodzik, A. Nemethi and A. Ranicki, On the semicontinuity of the mod 2 spectrum of hypersurface singularities, J. Algebraic Geom. 24 (2015), 379–398.
  5. [5]Y. Bugeaud and A. Dujella, Root separation for irreducible integer polynomials, Bull. Lond. Math. Soc. 43 (2011), 1239–1244.
  6. [6]Y. Bugeaud and A. Dujella, Root separation for reducible integer polynomials, Acta Arith. 162 (2014), 393–403.
  7. [7]Y. Bugeaud and M. Mignotte, Polynomial root separation, Int. J. Number Theory 6 (2010), 587–602.DOI
  8. [8]A. Dujella and T. Pejković, Root separation for reducible monic polynomials of odd degree, Rad Hrv. Akad. Znan. Umjet. Mat. Znan. 21 (2017), 21–27.
  9. [9]I. Gessel and G. Viennot, Binomial determinants, paths, and hook length formulae, Adv. Math. 58 (1985), 300–321.
  10. [10]S. M. Gusein-Zade and N. N. Nekhoroshev, Singularities of type A_k on plane curves of a chosen degree, Funct. Anal. Appl. 34 (2000), 214–215.
  11. [11]K. Mahler, An inequality for the discriminant of a polynomial, Michigan Math. J. 11 (1964), 257–262.DOI
  12. [12]Y. Matsui and K. Takeuchi, Monodromy at infinity of polynomial maps and Newton polyhedra (with an appendix by C. Sabbah), Int. Math. Res. Not. IMRN (2013), no. 8, 1691–1746.
  13. [13]S. Yu. Orevkov, Some examples of real algebraic and real pseudoholomorphic curves, in Perspectives in Analysis, Geometry, and Topology, Progress in Mathematics, vol. 296, Birkhäuser, 2012, 355–387.DOI
  14. [14]T. Pejkovic, Real and p-adic root separation for cubic and quartic polynomials, preprint, https://web.mathematik.uni-kiel.de/~pejkovic/parsep/rootsep34-v1.pdf.
  15. [15]A. Poteaux and M. Rybowicz, Good reduction of Puiseux series and applications, J. Symbolic Comput. 47 (2012), 32–63.DOI
  16. [16]D. Ruberman, The minimal genus of an embedded surface of non-negative square in a rational surface, Turkish J. Math. 20 (1996), 129–133.
  17. [17]C. Sabbah, Hypergeometric periods for a tame polynomial, Port. Math. (N.S.) 63 (2006), 173–226.
  18. [18]A. Schinzel and U. Zannier, The least admissible value of the parameter in Hilbert’s irreducibility theorem, Acta Arith. 69 (1995), 293–302.DOI
  19. [19]H. Stichtenoth, Algebraic Function Fields and Codes, 2nd ed., Graduate Texts in Mathematics, vol. 254, Springer, Berlin, 2009.DOI
  20. [20]W. Zhang, Symmetric Diophantine approximation over function fields, Ph.D. thesis, Leiden University, 2015.

Paper details

vlnc:W_qgxh6873
Version
v1 · latest
Published
Fresh — published within the last 30 days
License
CC BY 4.0
Usage

views · 1

downloads · 0

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

  • Mathematics

OpenAI ChatGPT, Anthropic Claude Fable

View statement
Looking for related papers…
Contents