Small Dynamical Heights for Quadratic Polynomials and Rational Functions

SMALL DYNAMICAL HEIGHTS FOR QUADRATIC POLYNOMIALS AND RATIONAL FUNCTIONS Abstract. Let φ ∈ Q(z) be a polynomial or rational function of degree 2. A special case of Morton and Silverman’s Dynamical Uniform Boundedness Conjecture states that the number of rational preperiodic points of φ is bounded above by an absolute constant. A related conjecture of Silverman states that the canonical height ĥφ (x) of a non-preperiodic rational point x is bounded below by a uniform multiple of the height of φ itself. We provide support for these conjectures by computing the set of preperiodic and small height rational points for a set of degree 2 maps far beyond the range of previous searches. In this paper, we consider the dynamics of a rational function φ(z) ∈ Q(z) acting on P (Q). The degree of φ = f /g is deg φ := max{deg f, deg g}, where f, g ∈ Q[z] have no common factors. Define φ0 (z) = z, and for every n ≥ 1, let φn (z) = φ ◦ φn−1 (z); that is, φn is the n-th iterate of φ under composition. In this context, the automorphism group PGL(2, Q) of P1 (Q) acts on Q(z) by conjugation. The forward orbit of a point x ∈ P1 (Q) is the set of iterates 1 O(x) = Oφ (x) := {φn (x) : n ≥ 0}. The point x is said to be periodic under φ if there is an integer n ≥ 1 such that φn (x) = x. In that case, we say x is n-periodic, we call the orbit O(x) an n-cycle, and we call n the period of x, or of the cycle. The smallest period n ≥ 1 of a periodic point x is called the minimal period of x, or of the cycle.n More generally, x is preperiodic under φ if there are integers n > m ≥ 0 such that φ (x) = φm (x). Equivalently, φm (x) is periodic for some m ≥ 0; also equivalently, the forward orbit O(x) is finite. We denote the set of preperiodic points of φ in P1 (Q) by Preper(φ, Q). In 1950, using the theory of arithmetic heights, Northcott [23] proved that if deg φ ≥ 2, then φ has only finitely many preperiodic points in P1 (Q). (In fact, he proved a far more general finiteness result, for morphisms of PN over any number field.) In 1994, Morton and Silverman proposed a dynamical Uniform Boundedness Conjecture [20, 21]; for φ ∈ Q(z) acting on P1 (Q), it says the following. Conjecture 1 (Morton-Silverman, 1994). For any d ≥ 2, there is a constant M = M (d) such that for any φ ∈ Q(z) of degree d, # Preper(φ, Q) ≤ M. Only partial results towards Conjecture 1 have been proven, including non-uniform bounds of various strengths, as well as conditions under which certain preperiodic orbit Key words and phrases. canonical height, arithmetic dynamics, preperiodic points. 1 2 structures are possible or impossible. See, for example [2, 5, 8, 16, 18, 19, 20, 21, 22, 24, 25, 29], as well as [28, Section 4.2]. Poonen [25] later stated a sharper version of the conjecture for the special case of quadratic polymials over Q: Conjecture 2 (Poonen, 1998). Let φ ∈ Q[z] be a polynomial of degree 2. Then # Preper(φ, Q) ≤ 9. If true, Conjecture 2 is sharp; for example, z 2 −29/16 and z 2 −21/16 each have exactly 9 rational preperiodic points, including the point at ∞. However, even though it is the simplest case of Conjecture 1, a proof of Conjecture 2 seems to be very far off at this time. As little as we know about the uniform boundedness conjecture for quadratic polynomials, we know even less about rational functions. In her Ph.D. thesis, Manes [16] made the first systematic attack on preperiodic points of quadratic rational functions, including a conjecture that # Preper(φ, Q) ≤ 12 when φ(z) ∈ Q(z) has deg φ = 2. In this paper, we give examples with 14 rational preperiodic points, showing that Manes’ conjecture is false. On the one hand, we found a single map with a rational 7-cycle, along with the immediate preimages of all seven points; see equation (0.1). On the other hand, we found many maps with a rational point x whose sixth iterate φ6 (x) is 2-periodic; the immediate preimages of all those preperiodic points again give a total of 14 points. We also found a single map with a rational point x for which φ5 (x) is 3-periodic, again giving a total of 14 points. See Table 3 for examples. It would appear that there are only finitely many maps with a 7-cycle or with a 3-periodic cycle with a tail of length 5, while there seem to be infinitely many with a 2-periodic cycle having a tail of length 6; for the moment, however, those finiteness questions remain open. Meanwhile, Blanc, Canci, and Elkies, in their study of the space of degree two rational functions with a rational point of period 6, have recently announced an infinite family of quadratic maps with 14 Q-rational preperiodic points; see [4, Lemma 4.7]. Their family uses two separate orbits: a 6-cycle and a fixed point, together with the preimages of all seven periodic points. Besides its preperiodic orbits, any rational function φ ∈ Q(z) of degree d ≥ 2 has an associated canonical height. The canonical height is a function ĥφ : P1 (Q) → [0, ∞) satisfying the functional equation ĥφ (φ(z)) = d · ĥφ (z), and it has the property that ĥφ (x) = 0 if and only if x is a preperiodic point of φ; see Section 1. For a non-preperiodic point y, on the other hand, ĥφ (y) measures how fast the standard Weil height h(φn (y)) of the iterates of y increases with n. By an analogy with Lang’s height lower bound conjecture for elliptic curves, Silverman has asked how small ĥφ (y) can be for nonpreperiodic points y. More precisely, considering φ as a point in the appropriate moduli space, and defining h(φ) to be the Weil height of that point, he stated the following conjecture; see [28, Conjecture 4.98] for a more general version. Conjecture 3 (Silverman, 2007). Let d ≥ 2. Then there is a positive constant M 0 = M 0 (d) > 0 such that for any φ ∈ Q(z) of degree d and any point x ∈ P1 (Q) that is not preperiodic for φ, we have ĥφ (x) ≥ M 0 h(φ). Conjecture 3 essentially says that the height of a non-preperiodic rational point must start to grow rapidly within a bounded number of iterations. Some theoretical evidence DYNAMICS OF QUADRATIC MAPS 3 for Conjecture 3 appears in [1, 14], and computational evidence for polynomials of degree d = 2, 3 appears in [3, 9, 10]. The smallest known value of ĥφ (x)/h(φ) when φ is a 7 181 polynomial of degree 2 occurs for x = 12 and φ(z) = z 2 − 144 . The first few iterates of this pair (x, φ), first discovered in [10], are 11 5 13 1 5 11 377 2445 7 7→ − 7→ 7→ − 7→ − 7→ − 7→ 7→ − 7→ 7→ · · · . 12 12 12 12 12 4 36 324 26244 7 The small canonical height ratio ĥφ ( 12 )/h(φ) ≈ .0066 makes precise the observation that although the numerators and denominators of the iterates eventually explode in size, it takes several iterations for the explosion to get underway. In this paper, we investigate quadratic polynomials and rational functions with coefficients in Q, looking for rational points that either are preperiodic or have small canonical height. More precisely, we search for pairs (x, φ) for which either x is preperiodic under φ or the ratio ĥφ (x)/h(φ) is positive but especially small. Past computational investigations of this type (such as those in [9, 10] for quadratic polynomials, or [3] for cubic polynomials) have started with the map φ and then computed the full set Preper(φ, Q) of rational preperiodic points, or the full set of rational points of small canonical height. This is a slow process, because for any given φ, the region in P1 (Q) that must be exhaustively searched is usually quite large, and the overwhelming majority of points in the region turn out to be false alarms. However, since we are looking for only a single point x with an interesting forward orbit, we can start with the point x and then search for maps φ that give interesting orbits for x. This strategy ends up testing far fewer pairs (x, φ) that never really had a chance of being preperiodic or having small height ratio, and hence we can push our computations much further. Our computations provide further evidence for Conjectures 1, 2, and 3. First, in spite of our very large search region, we found no quadratic polynomials with any Q-rational preperiodic structures not already observed and classified in [25], providing support for Conjecture 1, or more precisely, for Conjecture 2. Second, we also found a number of new pairs (x, φ) with small canonical height ratio for φ a quadratic polynomial, but (7/2, z 2 − 181/144) remains the record-holder. This supports Conjecture 3, as even our much larger search region turned up no points breaking the previously existing record. In particular, we are led to propose the following. Conjecture 4. Let Cpoly,2
ĥz2 −181/144 7/12 .03433 . . .
:= = ≈ .00660. log 181 h 181/144 For any polynomial φ ∈ Q[z] with deg φ = 2, let h(φ) = h(c), where c ∈ Q is the unique rational number such that φ is conjugate to z 7→ z 2 + c. Then ĥφ (x) ≥ Cpoly,2 · h(c) for any x ∈ P1 (Q). Third, regarding quadratic rational functions, which had not been previously studied at this level of generality, we found several new preperiodic orbit structures. Most notably, we found many degree two maps with 14 Q-rational preperiodic points, including one with a Q-rational periodic cycle of period 7: (0.1) 4655z 2 − 4826z + 171 φ(z) = , 4655z 2 − 8071z + 798 4 with periodic cycle 3 19 1 57 7→ 7→ 7→ 7→ ∞. 14 21 7 35 Each of the seven rational points above also has exactly one rational preimage outside the cycle. Those points are: 2 57 9 563 29 3 27 , , , ,− , , , 19 295 245 665 5 190 10 respectively, and Preper(φ, Q) consists of precisely these fourteen points. Fourth, we found pairs (x, φ) ∈ P1 (Q) × Q(z) with deg(φ) = 2 and with canonical height ratio ĥφ (x)/h(x) much smaller than .0066, the record for quadratic polynomials. But again, the pairs (x, φ) of especially small height ratio, as well as the first maps with 14 rational preperiodic points, turned up early in our large search. This support for Conjectures 1 and 3 inspired the following more specific statements for quadratic rational maps. ∞ 7→ 1 7→ 0 7→ Conjecture 5. Let ψ(z) = (10z 2 − 7z − 3)/(10z 2 + 37z + 9), and let Crat,2 := ĥψ (∞)/h(ψ) ≈ .000466. For any rational function φ ∈ Q(z) with deg φ = 2, and any x ∈ P1 (Q) (a) # Preper(φ, Q) ≤ 14. (b) If x is preperiodic, then #Oφ (x) ≤ 8. (c) If x is periodic, then #Oφ (x) ≤ 7. (d) If x is not preperiodic, then ĥφ (x) ≥ Crat,2 · h(φ). The outline of the paper is as follows. In Section 1 we review some background: heights, canonical heights, multipliers of periodic points, valuations, and good reduction. In Section 2 we recall and then sharpen some known facts about the dynamics of quadratic polynomials over Q, and we describe our search algorithm. We then summarize and discuss our data from that search. Finally, in Section 3, we do a similar analysis for quadratic rational functions over Q. We also state and prove the following result suggested by our data; see Theorem 3.6 for a more precise version. Theorem. Let X5,2 be the parameter space of all pairs (x, φ) with x ∈ P1 and φ a rational function of degree 2, up to coordinate change, for which the forward orbit of x consists of five strictly preperiodic points followed by a periodic cycle of period 2. Then X5,2 is birational over Q to an elliptic surface of positive rank over Q(t) with infinitely many Q-rational points. 1. Background The standard (Weil) height function on Q is the function h : P1 (Q) → R given by h(x) := log max{|m|, |n|}, if we write x = m/n in lowest terms and write ∞ as 1/0. It has a well-known extension to a function h : P1 (Q) → R, although that extension will not be of much concern to us here. The height function satisfies two important properties. First, for any φ(z) ∈ Q(z), there is a constant C = C(φ) such that h(φ(x)) − d · h(x) ≤ C for all x ∈ P1 (Q), DYNAMICS OF QUADRATIC MAPS 5 where d = deg φ. Second, for any bound B ∈ R, (1.1) {x ∈ P1 (Q) : h(x) ≤ B} is a finite set. For any fixed φ ∈ Q(z) of degree d ≥ 2, the canonical height function ĥφ : P1 (Q) → R for φ is given by ĥφ (x) := lim d−n h(φn (x)), n→∞ and it satisfies the functional equation (1.2) ĥφ (φ(x)) = d · ĥφ (x) for all x ∈ P1 (Q). In addition, there is a constant C 0 = C 0 (φ) such that (1.3) ĥφ (x) − h(x) ≤ C 0 for all x ∈ P1 (Q). Northcott’s Theorem [23] that # Preper(φ, Q) < ∞ is immediate from properties (1.1), (1.2), and (1.3), since they imply that for any x ∈ P1 (Q), ĥ(x) = 0 if and only if x is preperiodic under φ. (In fact, Northcott proved this equivalence for any x ∈ P1 (Q)). Meanwhile, our canonical height computations will require the following result. Lemma 1.1. Let φ = f /g ∈ Q(z), where f, g are relatively prime polynomials in Z[z] with d := max{deg f, deg g} ≥ 2. Let R = Res(f, g) ∈ Z be the resultant of f and g, and let max{|f (t)|, |g(t)|} D := min . t∈R∪{∞} max{|t|d , 1} Then D > 0, and for all x ∈ P1 (Q) and all integers i ≥ 0,   
|R| 1 −i i log ĥφ (x) ≥ d h φ (x) − . d−1 D Proof. The function F (t) = max{|f (t)|, |g(t)|}/ max{|t|d , 1} is defined at ∞ because max{deg f, deg g} = d. Moreover, F is real-valued, positive, and continuous on the compact set R ∪ {∞}, and hence the minimum D is indeed both attained and positive. As shown in the proof of Lemma III.30 .(b) of [26],   F (x) h(φ(x)) ≥ dh(x) + log |R| for all x ∈ P1 (Q). Bounding F (x) by D, and applying φ repeatedly to φi (x), we have  

D i+m m i 2 m−1 h φ (x) ≥ d h φ (x) + (1 + d + d + · · · + d ) log |R| for all x ∈ Q and all i, m ≥ 0. Dividing by di+m and taking the limit as m → ∞ gives the desired inequality.  For more background on heights and canonical heights, see [12, Section B.2], [15, Chapter 3], or [28, Chapter 3]. The following notion of good reduction of a dynamical system, which we state here only over Q, was first proposed in [20]. 6 Definition 1.2. Let φ(z) = f (z)/g(z) ∈ Q(z) be a rational function, where f, g ∈ Z[z] have no common factors in Z[z]. Let d = deg φ = max{deg f, deg g}, and let p be a prime number. Let f , g ∈ Fp [z] be the reductions of f and g modulo p. If deg(f /g) = d, we say that φ has good reduction at p; otherwise, we say φ has bad reduction at p. Of course, deg(f /g) ≤ d always; but the degree can drop if either max{deg f , deg g} < d or f and g are no longer relatively prime. If φ = ad z d + · · · + a0 is a polynomial, then φ has good reduction at p if and only if vp (ai ) ≥ 0 for all i, and vp (ad ) = 0; see [21, Example 4.2]. Here, vp (x) denotes the p-adic valuation of x ∈ Q× , given by vp (pr a/b) = r where a, b, r ∈ Z and p - ab. Finally, if φ ∈ Q(z) is a rational function, and x ∈ Q is a periodic point of φ of minimal period n, the multiplier of x is (φn )0 (x). The multiplier is invariant under coordinate change, and therefore one can compute the multiplier of a periodic point at x = ∞ by changing coordinates to move it elsewhere. We will need multipliers to discuss rational functions in Section 3. 2. Quadratic Polynomials It is well known that (except in characteristic 2), any quadratic polynomial is conjugate over the base field to a unique one of the form φc (z) := z 2 + c. (Scaling guarantees the polynomial is monic, and an appropriate translation to complete the square eliminates the linear term.) Thus, the moduli space of quadratic polynomials up to conjugacy is A1 , where the parameter c ∈ A1 corresponds to φc . For the purposes of Conjecture 3, then, the height of φc itself is h(φc ) := h(c). In addition, let us denote the canonical height associated to φc as simply ĥc . However, we are really interested in pairs (x, c) for which the point x ∈ P1 either is preperiodic with a long forward orbit under φc , or else has very small canonical height under φc , as compared with h(c). The appropriate moduli space of such pairs is therefore P1 × A1 . If x = ∞, then x is simply a fixed point with canonical height 0. Thus, we may restrict ourselves to the subspace A2 ∼ = A1 × A1 ⊆ P1 × A1 . When working over Q, we can restrict ourselves even further. The following lemma is well known, but we include the short proof for the convenience of the reader. Lemma 2.1. Let x, c ∈ Q, and suppose that x is preperiodic under φc (z) = z 2 + c. Then writing x = m/n in lowest terms, we must have c = k/n2 , where k ∈ Z satisfies (a) k ≡ −m2 (mod n), (b) the integers n and (k + m2 )/n are relatively prime, 2 (c) k ≤ n( /4, and √ [0, 2] if c ≥ −2, (d) |x| ∈ √ where B := (1 + 1 − 4c)/2. [ −c − B, B] if c < −2, Proof. Write c = k/N , where k, N ∈ Z are relatively prime, and N ≥ 1. If N 6= n2 , then there is a prime p such that vp (N ) 6= 2vp (n). Let r := vp (n) and s := vp (N ). Then (2.1) φc (x) = m2 k m2 N + kn2 + = . n2 N n2 N DYNAMICS OF QUADRATIC MAPS 7 The denominator of (2.1) has vp (n2 N ) = 2r + s. If s > 2r, then the numerator has vp (m2 N + kn2 ) = 2r, and hence vp (φc (x)) = −s. On the other hand, if s < 2r, then the numerator has vp (m2 N + kn2 ) = s, and hence vp (φc (x)) = −2r. Either way, then, vp (φc (x)) < −r = vp (x). Thus, vp (φic (x)) will strictly decrease with i, contradicting the hypothesis that x is preperiodic. Therefore, c = k/n2 , with k ∈ Z relatively prime to n. Applying the previous paragraph to any preperiodic point, not just x, we see that all preperiodic points of φ in Q have denominator exactly n when written in lowest terms. In particular, φc (x) = (k + m2 )/n2 is preperiodic. Hence, n|(k + m2 ), and (k + m2 )/n must be relatively prime to n, giving us statements (a) and (b). Next, if c > 1/4, then for any y ∈ R, we have 1  1 2 φ(y) − y = y 2 − y + c > y 2 − y + = y − ≥ 0. 4 2 Thus, {φnc (x)}n≥0 is a strictly increasing sequence, contradicting the hypothesis that x is preperiodic. Hence, c ≤ 1/4, giving us statement (c). Finally, if c ≥ −2, then for any y ∈ R with |y| > 2, φ(y) − |y| = |y|2 − |y| + c ≥ |y|2 − |y| − 2 = (|y| − 2)(|y| + 1) > 0. Just as in the previous paragraph, then, {φnc (x)}n≥0 is strictly increasing if |x| > 2, contradicting √ the preperiodicity hypothesis.√ Similarly, if c < −2, then noting that A = (1 − 1 − 4c)/2 < −1 and B = (1 + 1 − 4c)/2 > 2 are the two (real) roots of φ(z) − z = 0, we have, for any y ∈ R with |y| > B, φ(y) − |y| = |y|2 − |y| + c = (|y| − A)(|y| − B) > 0. √ √ Meanwhile, φ−1 c ([−B, B]) = [−B, − √−c − B] ∪ [ −c − B, B], and hence for x ∈ Q to be preperiodic, we must have |x| ∈ [ −c − B, B].  We can improve the bound of Lemma 2.1.(c) with a minor extra assumption, as follows. Lemma 2.2. With notation as in Lemma 2.1, let Oc (x) denote the forward orbit of x under φc . If #Oc (x) ≥ 4, then k < −3n2 /4. Proof. Given Lemma 2.1.(c), we need to show that if c ∈ [−3/4, 1/4], then the preperiodic point x has #Oc (x) ≤ 3. First of all, using Lemma 2.1, it is easy to see that n 1o n 1 3o Preper(φ1/4 , Q) = ± , and Preper(φ−3/4 , Q) = ± , ± . 2 2 2 In both cases, it follows immediately that each preperiodic point has forward orbit of length at most 2. Second of all, for any c ∈ (−3/4, 1/4), the dynamics of φc on R has two fixed points, at a < b ∈ R, and every point in R is either attracted to ∞ under iteration, or attracted to a under iteration, or equal to ±b; and φc (±b) = b. This well known fact is easy to check; see, for example, Section VIII.1, and especially Theorem VIII.1.3, of [6]. (Indeed, the c-interval (−3/4, 1/4) is precisely the intersection of the main cardioid of the Mandelbrot set with the real line.) Thus, the only rational preperiodic orbits of φ must end in fixed points. However, in part 6 of Theorem 3 of [25], Poonen proved that for x, c ∈ Q, if x is preperiodic to a fixed point, then #O(x) ≤ 3, as desired.  8 Lemmas 2.1 and 2.2 are only about the case that the point x ∈ Q is preperiodic for φc . The following result, using similar ideas as those in Lemma 2.1, is also relevant to finding non-preperiodic points of small canonical height. Lemma 2.3. Let x, c ∈ Q, let φc (z) = z 2 + s := vp (c), and suppose vp (φi (x)) < min{0, s/2}  2−i−1 (2 − s) log p    2−1 log 2 ĥc (x) ≥  log 2    −1 2 log p c, and let p be a prime number. Set for some i ≥ 0. Then if s ≤ −2 is even, if p = 2 and s = −2, if p = 2 and s = −4, otherwise. Proof. Let r = vp (x). If s ≥ 0, suppose first that r ≥ 0 as well. Then vp (φj (x)) ≥ 0 for all j ≥ 0, contradicting our hypotheses. Thus, we must have r ≤ −1. Applying φ, it is immediate that vp (φj (x)) = 2j r for all j ≥ 1. That is, the denominator of φj (x) features p raised to the power 2j |r|, and hence ĥc (x) = lim 2−j h(φj (x)) ≥ |r| log p ≥ log p. j→∞ If s ≤ −1 is odd, then regardless of the value of r = vp (x), we must have vp (φ(x)) ≤ s, and hence vp (φj (x)) ≤ 2j−1 s for all j ≥ 1. Thus, ĥc (x) = lim 2−j h(φj (x)) ≥ 2−1 log p. j→∞ If s ≤ −2 is even, then the hypothesis that vp (φi (x)) < s/2 gives vp (φi (x)) ≤ (s−2)/2, and hence vp (φj (x)) ≤ 2j−i−1 (s − 2) for all j ≥ i. Thus, ĥc (x) = lim 2−j h(φj (x)) ≥ 2−i−1 (2 − s) log p. j→∞ If p = 2 and s = −2, write c = a/4, where a ∈ Q with v2 (a) = 0. If r ≤ −2, then v2 (φj (x)) = 2j r for all j ≥ 0, and hence ĥc (x) ≥ −r log 2 ≥ 2 log 2. If r ≥ 0, then v2 (φ(x)) = −2, so that ĥc (φ(x)) ≥ 2 log 2, and hence ĥc (x) ≥ log 2. Lastly, if r = −1, suppose first that a ≡ 1 (mod 4). Writing x = b/2 with v2 (b) = 0, we have v2 (b2 +a) = 1, and therefore v2 (φ(x)) = −1 as well. Thus, v2 (φj (x)) = −1 for all j ≥ 0, contradicting the hypotheses. Hence, we must instead have a ≡ 3 (mod 4). Then v2 (b2 + 1) ≥ 2, so that v2 (φ(x)) ≥ 0, implying that ĥc (φ(x)) ≥ log 2, and therefore that ĥc (x) ≥ 2−1 log 2. It remains to consider p = 2 and s = −4. If r ≤ −3, then v2 (φj (x)) = 2j r for all j ≥ 0, and hence ĥc (x) ≥ −r log 2 ≥ 3 log 2. If r ≥ −1, then v2 (φ(x)) = −4, so that ĥc (φ(x)) ≥ 4 log 2, and hence ĥc (x) ≥ 2 log 2. Last, if p = 2, s = −4, and r = −2, write c = a/16 and x = b/4 with v2 (a) = v2 (b) = 0. We consider three cases. First, if a ≡ 3 (mod 8), then v2 (b2 + a) = 2, and hence v2 (φ(x)) = −2; thus, v2 (φj (x)) = −2 for all j ≥ 0, contradicting the hypotheses. Second, if a ≡ 7 (mod 8), then v2 (b2 + a) ≥ 3, and hence v2 (φ(x)) ≥ −1. By the previous paragraph, then, ĥc (φ(x)) ≥ 2 log 2, and therefore ĥc (x) ≥ log 2. Third, if a ≡ 1 (mod 4), then v2 (b2 + a) = 1, and hence v2 (φ(x)) = −3. Again by the previous paragraph, ĥc (φ(x)) ≥ 3 log 2, and therefore ĥc (x) ≥ (3/2) log 2.  In Algorithm 2.4 below, we will check whether the denominator of φ3 (x) is too large — that is, whether it is more than the square root of the denominator of c. If that DYNAMICS OF QUADRATIC MAPS 9 happens, then some prime p appears in the denominator of φ3 (x) to too large a power. By Lemma 2.3 with i = 3, then, we would have log p log 3 ĥc (x) ≥ min{2−4 · 4, 2−1 } log p = ≥ = 0.275 . . . 4 4 if p ≥ 3, or log 2 ĥc (x) ≥ min{2−4 · 8, 2−1 } log 2 = = 0.347 . . . . 2 if p = 2. Meanwhile, the parameters c in the search range we used had denominator at most 600602 and absolute value |c| ≤ 10, so that h(c) ≤ 2 log(60060) + log(10). Thus, the associated height ratio would be at least log 3 ĥc (x) ≥ = 0.0113 . . . h(c) 4(2 log(60060) + log 10) and usually much larger. In short, if the third iterate φ3 (x) has the wrong denominator, the pair (x, c) will not be of interest to us. Similarly, even though Lemmas 2.1 and 2.2 do not expressly disallow points of small but positive canonical height when c ≥ −3/4, they strongly suggest that such points would be unlikely to show up with denominators in our range of n ≤ 60060. We confirmed the absence of such points for smaller denominators in some preliminary searches. Thus, although we risked missing a very small handful of points of small height and larger denominators, we decided to exclude the parameters c ≥ −3/4 from our search. Our entire search, then, was over pairs (x, c) with x = m/n and c = k/n2 fitting the restrictions of Lemmas 2.1 and 2.2. Meanwhile, since φc (−x) = φc (x), we may assume that the starting point x is positive. Thus, we were led to the following search algorithm, which is guaranteed to find all previously unknown rational preperiodic pairs (x, c) in its search range, as well as nearly all pairs in the same range with especially small canonical height ratio ĥc (x)/h(c). Algorithm 2.4. Fix integers nmax and Nmax as bounds. [We used nmax = 60060 and Nmax = 10.] 1. Let n run through the integers from 1 to nmax . 2. For each such n, let m run through the integers from 1 to n that are relatively prime to n. 3. For each such n and m, let k run through all the integers congruent to m2 modulo n and lying between −3n2 /4 and −n2 Nmax . Let c := k/n2 , and φc (z) := z 2 + c. 4. Let x run through all rational numbers of the form x := (m + nj)/n with j ∈ Z √ and x in the interval [0, 2] or [ −c − B, B] from Lemma 2.1.(d). 5. Compute φi (x) for i = 1, 2, 3, 4. Stop iterating and discard x if φi (x) coincides with an earlier iterate for some i ≤ 4, or has denominator larger than n for some i ≤ 3. 6. Otherwise, compute
2−12 h φ12 (x) . If this value is less than 0.02 times h(c), record it as our approximation for the canonical height ĥc (x), and record ĥc (x)/h(c) as the associated height ratio. In [25], Poonen fully classified all quadratic polynomial preperiodic orbits in Q with #Oc (x) ≤ 4, which is why we discarded them in Step 5 of the algorithm. He also showed that for x, c ∈ Q, if x is preperiodic for φc , then there are only two ways that 10 c x − 181 144 − 1153 576 − 517 144 − 36989 19600 − 31949 19600 − 5149 3600 − 205 144 − 181 144 − 16381 7056 − 931161001 476985600 − 10381 3600 − 9901 3600 − 293749 176400 − 271909 176400 − 1513 576 − 373 144 − 1013082841 476985600 − 160021 63054 7 12 11 24 17 12 153 140 27 140 23 60 1 12 11 12 97 84 30379 21840 121 60 131 60 433 420 43 420 31 24 23 12 10541 21840 181 252 ĥc (x) ĥc (x)/h(c) φc (x), φ2c (x), . . . 11 5 1 .03433 .00660 − 12 , − 12 , − 13 , − 12 , − 54 , 11 , − 377 ,... 12 36 324 43 29 13 41 11 223 8021 .06505 .00923 − 24 , 24 , − 24 , − 24 , 12 , − 192 , − 12288 ,... 13 29 9 53 461 41093 .06885 .01102 − 19 , − , − , , , − , − , ... 12 12 12 4 36 324 26244 97 197 13 263 1609 38821 .12319 .01171 − 140 , − 140 , 140 , − 140 , 980 , 48020 , . . . .12319 .01187 − 223 , 127 , − 113 , − 137 , − 659 , − 56561 ,... 140 140 140 140 980 48020 77 13 83 29 359 13 ,... .10274 .01202 − 60 , 60 , − 60 , 60 , − 300 , 7500 , − 6704413 4687500 17 7 13 1 49 139 32531 .06866 .01290 − 12 , 12 , − 12 , − 4 , − 36 , 324 , − 26244 , . . . 5 1 2545 .06866 .01321 − 12 , − 13 , − 12 , − 54 , 11 , − 377 , − 26244 ... 12 36 324 83 113 43 173 1129 39331 .13059 .01346 − 84 , − 84 , − 84 , − 84 , 588 , 28812 . . . 379 .28548 .01382 − 21840 , − 42629 , 40571 , 32731 , 27809 , − 76737829 ,... 21840 21840 21840 94640 41127840 71 89 41 29 887 43947 147354737 .12758 .01380 , − , − , − , , , , . . . 60 60 60 12 300 7500 4687500 121 79 61 103 59 6779 64722611 .12912 .01403 , , − , − , , − , , . . . 60 60 60 60 300 2500 14062500 547 13 233 6959 630923 .17685 .01405 − 253 , − , , − , , − , . . . 420 420 420 140 6300 1417500 337 377 103 6301 767033 .17685 .01413 − 643 , , − , − , − , − ,... 420 420 420 140 6300 1417500 23 41 7 61 23 2317 1757219 .10590 .01446 − 24 , − 24 , 24 , − 24 , 6 , 192 , 12288 , . . . 13 7 .08640 .01459 , − 17 , − 12 , − 94 , 89 , 1141 , 257491 , . . . 12 12 36 324 26244 349 .30762 .01483 − 41299 , 31709 , − 21840 , − 46381 , 677449 , 5994294739 . . . 21840 21840 21840 283920 1679386800 505 377 71 205 7793 1706041 .17952 .01498 − 252 , 252 , − 252 , − 84 , 2268 , 183708 . . . Table 1. (x, c) ∈ Q2 in the search region with canonical height ratio ĥc (x)/h(c) < 0.015 the forward orbit Oc (x) can have more than four points. Specifically, either c = −29/16 and x = ±3/4, or else the periodic cycle of the orbit has period at least 6; the latter should be impossible, according to Conjecture 2. If such a point did exist, Algorithm 2.4 would simply declare it to be a point of extremely small canonical height. We would have checked any such points by hand for preperiodicity, but our search found none. Our choice of nmax = 60060 = 22 · 3 · 5 · 7 · 11 · 13 is motivated first in order to ensure the computation finished in reasonable time, but also because the results of [20] show that preperiodic orbits in P1 (Q) cannot be very long if there are small primes of good reduction. Moreover, the results of [2] show that there cannot be too many rational preperiodic points unless there are a lot of bad primes. The bad primes of z 2 + c are precisely those dividing the denominator of c, and hence we wanted our search to run at least to n = 60060. In the case of p = 2, we needed the 2 in the denominator of x to appear to at least the power 2, for the following reason. Although any map φq/4 (z) = z 2 + q/4 with q ∈ Q and q ≡ 1 (mod 4) has bad reduction as written, the conjugate φq/4 (z + 1/2) − 1/2 = z 2 + z + (q − 1)/4 would have good reduction at 2; by the results of [29] (summarized in [28], Theorem 2.28), the forward orbit of a preperiodic point x ∈ Q could then have length at most 4. (On the other hand, if q ≡ 3 (mod 4), then all points of Q would have canonical height at least log 2, as shown in the proof of Lemma 2.3.) The results of Algorithm 2.4 are summarized in Table 1, which lists all pairs (x, c) we found with height ratio at most 0.015. For the sake of accuracy, we manually computed the canonical heights and height ratios in the table to many more than the 12 iterates originally computed by the algorithm. Note that although there continue to be pairs with canonical height ratio about .013 for c up to fairly large height (e.g., two on the DYNAMICS OF QUADRATIC MAPS c x − 9142351 5832225 − 28236091 12744900 − 322457899 143280900 − 136643866 79655625 1394 2415 5981 3570 24347 11970 3854 8925 11 ĥc (x) ĥc (x)/h(c) φc (x), φ2c (x), . . . 2981 106 2134 .27774 .01733 − 2415 , − 2415 , − 3781 , − 43699 , − 139366718 ,... 2415 2415 55545 146916525 2111 6661 4519 2189 260493 21085814279 .35856 .02090 , − 3570 , 3570 , − 3570 , − 141610 , 18048052890 , . . . 3570 22583 15667 919 .48733 .02487 , , − 1710 , − 164371 , 124637651 , . . . 11970 11970 83790 78008490 11854 62 9097034 3820075881834 13646 5554 .46675 .02492 − 8925 , 8925 , − 8925 , 1275 , − 5310375 , 3133342515625 , . . . Table 2. (x, c) ∈ Q2 in the search region with v2 (x) ≥ −1 and canonical height ratio ĥc (x)/h(c) < 0.025 list with n = 21840 = 24 · 3 · 5 · 7 · 13), none comes close to the ratio of less than .007 attained by the already-known pair (7/12, −181/144). Given the large size of this search (far larger than any prior search for such pairs), we view this as strong computational support for Conjecture 3, at least for quadratic polynomials over Q. In addition, as noted above, we found no preperiodic orbits outside of those classified in [25], thus providing further computational evidence for Conjecture 2, and hence for Conjecture 1 in the case of quadratic polynomials over Q. Our data suggests that just as a small prime of good reduction limits the length of a Q-rational preperiodic orbit, such a prime also appears to make it difficult for the canonical height ĥc (x) to be particularly small. Indeed, even though our search included pairs (x, c) for which the denominator n of x was not divisible by 4, the reader may have noticed that almost all of the pairs in Table 1 have n divisible by 12, and usually by 60. Motivated by this observation, we did another search allowing height ratios up to 0.03 when 4 - n. Table 2 lists all pairs (x, c) from that search with height ratio ĥc (x)/h(c) less than .025. There were only four, only one of which even came close to the cutoff of .015 used for Table 1. In addition, all four had n divisible by 15. The full data from both searches may be found at http://www3.amherst.edu/~rlbenedetto/quadpolydata/. 3. Quadratic Rational Functions The space M2 of conjugacy classes of degree-2 self-morphisms of P1 is known, by work of Milnor [17] and Silverman [27], to be isomorphic to A2 . In particular, they proved the following result, which appears as Theorem 4.56 in [28]. Lemma 3.1. Let φ ∈ Q(z) be a rational function of degree 2. Then φ has three fixed points in P1 (Q), counting multiplicity, and the multipliers of the fixed points are precisely the roots of a cubic polynomial T 3 + σ1 (φ)T 2 + σ2 (φ)T + (σ1 (φ) − 2), where σ1 (φ), σ2 (φ) ∈ Q are certain explicit rational functions of the coefficients of φ. Moreover, the function M2 → A2 given by φ 7→ (σ1 (φ), σ2 (φ)) is an isomorphism of algebraic varieties defined over Q. The precise formulas for σ1 and σ2 in terms of the coefficients of φ are not important here, but the interested reader can find them on page 189 of [28]. Instead, what is important is that Lemma 3.1 allows us to define a height function on M2 , given by (3.1) h(φ) = h(σ1 (φ), σ2 (φ)) := max{|a|, |b|, |c|}, where σ1 (φ) = a/c and σ2 (φ) = b/c, with a, b, c ∈ Z and gcd(a, b, c) = 1. 12 However, we are interested instead in the moduli space not of morphisms φ, but rather of pairs (x, φ) consisting of a point x and a morphism φ. More precisely, given a field K, we define PM2 (K) := {(x, φ) ∈ P1 (K) × K(z) : deg φ = 2}/ ∼, where the equivalence relation ∼ is given by
(x, φ) ∼ η(x), η ◦ φ ◦ η −1 for any η ∈ PGL(2, K). Of course, to verify that this quotient is well-behaved from the perspective of algebraic geometry, the machinery of geometric invariant theory is required. However, given the parametrization we are about to construct, and given that we are really interested in PM2 merely as a set, we can safely sidestep these issues. Our parametrization, based on an idea suggested by Elkies [7], begins with an informal dimension count. The set PM2 should be a space of dimension 3, because there are five dimensions of choices for φ (the six coefficients of the rational function, minus one for multiplying both top and bottom by an element of K) and one for x, but then the quotient by the action of PGL(2, K) subtracts three dimensions. Since we will be interested in both infinite orbits of small canonical height, and in preperiodic points with long forward orbits, we will restrict our attention to pairs (x, φ) for which the the first six iterates {φi (x) : 0 ≤ i ≤ 5} are all distinct. (Incidentally, the moduli spaces corresponding to pairs (x, φ) where x is periodic of period 5 or less are all birational over Q to P2 , as shown in Theorem 1(1) of [4].) Call the subset of PM2 consisting of (equivalence classes of) such pairs PM02 . Given (x, φ) ∈ PM02 , and writing xi := φi (x), then, there is a unique η ∈ PGL(2, K) such that η(x0 ) = ∞, η(x1 ) = 1, and η(x2 ) = 0. Hence, the equivalence class of (x, φ) under ∼ contains a unique element whose orbit begins in the form (3.2) ∞ 7→ 1 7→ 0 7→ x3 7→ x4 7→ x5 . Thus, we can view PM02 as coinciding, as a set, with the Zariski open subset of A3 (K) consisting of triples (x3 , x4 , x5 ) for which the map φ giving the partial orbit (3.2) actually has degree 2. (In particular, the three coordinates must be distinct, with none equal to 0 or 1, and φ must not degenerate to a lower-degree map.) The restrictions φ(∞) = 1 and φ(1) = 0 dictate that the associated map φ is of the form (a1 z + a0 )(z − 1) (3.3) φ(z) = . a1 z 2 + b1 z + b0 Moreover, it is straightforward to check that (3.2) stipulates a1 = x4 (x23 x4 − x23 x5 − x23 + 2x3 x5 − x4 x5 ), (3.4) a0 = −x3 b0 = x23 x4 (x4 − 1)(x5 − x4 + x4 x5 − x3 x5 ), b1 = x23 x24 x5 − x33 x24 + 2x33 x4 − x23 x24 + x34 x5 − x23 x4 x5 − x3 x24 x5 − x33 + x3 x24 − x34 + x23 x5 + x23 − x3 x4 + x24 − x3 x5 . Of course, the Zariski closed subset of A3 on which φ does not have degree 2 is precisely the zero locus of the resultant of the two polynomials (a1 z +a0 )(z −1) and a1 z 2 +b1 z +b0 . DYNAMICS OF QUADRATIC MAPS 13 Definition 3.2. For any integers m ≥ 0 and n ≥ 1 with m + n ≥ 6, let Xm,n denote the space of triples (x3 , x4 , x5 ) ∈ PM02 ⊆ A3 for which the map φ given by equations (3.3) and (3.4) satisfies φm+n (∞) = φm (∞), but φi (∞) 6= φj (∞) for any other i 6= j between 0 and m + n. As a set, then, Xm,n is the moduli space of pairs (x, φ) up to coordinate change for which φm (x) is periodic of minimal period n, but φi (x) is not periodic for 0 ≤ i ≤ m − 1. (Presumably it is in fact birational to the actual geometric moduli space of such pairs up to coordinate change, but as we noted before, we will not need to know that here.) Our preliminary searches showed an abundance of orbits with φ6 (∞) = φ4 (∞). This observation led us to the following result. Lemma 3.3. X4,2 is birational to P2 . Proof. X4,2 is the locus of triples for which φ(x5 ) = x4 . Solving this equation using the formula for φ from equations (3.3) and (3.4) gives x4 (x4 − 1)(x3 − x5 )(x4 − x5 )(x4 − x4 x5 − x23 x5 + 2x3 x5 − x3 ) = 0. Dividing out by the first four factors, which correspond to parameter choices outside PM02 , we get x4 − x3 (3.5) x5 = 2 . x3 − 2x3 + x4 Thus, X4,2 is parametrized by (x3 , x4 ), with x5 given by equation (3.5).  Remark 3.4. With x5 given by equation (3.5), it is easy to check with computational software that the resultant of the numerator and denominator of the corresponding map φ from equations (3.3) and (3.4) is (3.6) x23 x24 (x3 − x4 )(x3 − 1)8 (x4 − 1)2 (x23 − 2x3 + x4 )(x3 x4 − x3 + x4 )5 It is fairly clear why most of the terms in expression (3.6) appear, when we recall that none of x3 , x4 , x5 can equal 0, 1, or ∞. The term x3 x4 − x3 + x4 , meanwhile, appears because it is zero precisely when the degree 1 map taking ∞ 7→ 1 7→ 0 7→ x3 already maps x3 to x4 , and hence φ degenerates. Bearing in mind everything so far in this section, we are led to the following algorithm. Algorithm 3.5. Fix a height bound B > 0 and a threshold r > 0 for small height ratios. [We used B = log(100) and r = .002.] 1. Let (x3 , x4 , x5 ) run through all triples in Q3 with h(x3 ), h(x4 ), h(x5 ) ≤ B and with 0, 1, x3 , x4 , x5 all distinct. 2. For each triple, discard it if the resulting map φ of equations (3.3) and (3.4) degenerates to degree less than 2. In light of Lemma 3.3, also discard the triple if φ(x5 ) = x4 . 3. Compute φi (∞) = φi−5 (x5 ) for i = 6, . . . , 10. For each such i, stop iterating and record (∞, φ) as preperiodic if φi (∞) coincides with an earlier iterate. 4. Let C = log(|R|/D) > 0 be the constant of Lemma 1.1 for φ, and let h(φ) ≥ 0 be the height of φ from Lemma 3.1 and equation (3.1). 5. Discard the triple (x3 , x4 , x5 ) if

h φ8 (∞) ≥ 28 r · h(φ) + C or h φ10 (∞) ≥ 210 r · h(φ) + C, 14 since either one forces the height ratio ĥφ (∞)/h(φ) to be larger than the threshold r. 6. Otherwise, compute
2−15 h φ15 (∞) . If this value is less than r · h(φ) + 2−15 C, record it as our approximation for the canonical height ĥφ (∞), and record ĥφ (x)/h(φ) as the associated height ratio. The most noticeable feature of the data was the huge numberof points on X5,2 . We therefore analyzed X5,2 , leading us to the following result, whichis the precise form of the Theorem stated in the introduction. Theorem 3.6. X5,2 is birational to the elliptic surface E : y 2 = 4x3 + (4t4 + 4t3 + 1)x2 − 2t3 (t + 1)2 (2t2 + 2t + 1)x + t6 (t + 1)4 . (3.7) Moreover, E has positive rank over Q(t), including the non-torsion point P given by (x, y) = (0, t3 (t + 1)2 ). Proof. There is a morphism X5,2 → X4,2 taking a pair (x, φ) to the pair (φ(x), φ). Parametrizing X4,2 by (x3 , x4 ) via equation (3.5) as in the proof of Lemma 3.3, and letting ψ denote the morphism giving the orbit ∞, 1, 0, x3 , x4 , x5 , x4 , we can parametrize X5,2 by triples (x3 , x4 , w) for which ψ(w) = ∞. That is, X5,2 is the subvariety of A3 defined by g(x3 , x4 , w) = 0, where g(x3 , x4 , w) = w(1 − x4 )x33 + (1 − w)x3 x24 + w(x4 w + x4 − 2)x23 − (2w2 − 2w + 1)x3 x4 + wx3 + (w2 − w)x24 is the denominator of ψ(w). This variety is singular along the line x3 = w = 1. Blowing up along this line via (x3 − 1)s = w − 1 gives the surface s(s − 1)x23 x4 + sx23 − (2s2 − 2s + 1)x3 x4 − (s − 1)x3 + s(s − 1)x24 = 0, which is cubic in (x3 , x4 ). Standard manipulations, along with the substitution s = t + 1, give equation (3.7), with the birationality given by w − x3 (w − x3 )2 (w − 1) , x= , x3 − 1 x3 (x3 − 1)3
 (w − x3 )2 (w − 1)  2 2 2 y= (w − x )(w − 1)(2x + x ) − x (w − x ) + (w − 1) 3 4 3 3 3 2 x3 (x3 − 1)5 t= and
t t2 (t + 1) t(t + 1)2 − x , x3 = , x x
t x4 = 2 y + (2t2 + 2t + 1)x − t3 (t + 1)2 . 2x The point P in the statement of the Theorem obviously lies on the Q(t)-curve E of equation (3.7). Finally, specializing E at t = 1 gives an isomorphic copy of curve 142a1 in Cremona’s tables, which has trivial torsion and rank 1 over Q. The point P specializes to the generator, and hence P must have infinite order in E(Q(t)).  w= DYNAMICS OF QUADRATIC MAPS 15 Obviously the point P itself lies in the degeneracy locus of the birationality of Theorem 3.6, given the formula x3 = t2 (t + 1)/x. Meanwhile, it is easy to compute the first few multiples of P on E:

[2]P = t(t + 1)2 , t(t + 1)4 , P = 0, t3 (t + 1)2 ,

[3]P = − t2 (t + 1)2 , t2 (t + 1)3 , [4]P = t2 (t + 1), −t2 (t + 1)(t2 + t − 1) ,
[5]P = − t(t + 1), −t(t + 1)(t3 + 3t2 + 2t − 1) ,
[6]P = t3 (t + 1)2 (t + 2), −t3 (t + 1)3 (2t3 + 6t2 + 4t − 1) , and it turns out that all of the points [n]P , for n ∈ {0, ±1, ±2, ±3, ±4, ±5}, also lie in the degeneracy locus. However, all but finitely many [n]P must lie off it, and in particular, both [6]P and [−6]P are nondegenerate. [6]P , for example, corresponds to x3 = 1 , t(t + 1)(t + 2) x4 = − (t2 + t − 1) , t2 (t + 2)2 x5 = t+1 , t+2 with pole (t + 1)(t2 + t − 1) . t(t + 2) Meanwhile, given the results of [4] on the infinitude of conjugacy classes admitting a rational 6-periodic point, there are also certainly infinitely many rational functions with a length 7 preperiodic orbit. Indeed, given any non-critical point x in the 6-cycle (which will be most, and usually all, of them), the other preimage y of φ(x) is necessarily rational and strictly preperiodic with orbit length 7. Thus, (y, φ) gives a point in X1,6 (Q). We also found numerous rational points (though not nearly as many as for X5,2 ) on the other moduli surfaces corresponding to strictly preperiodic orbits of length 7: X6,1 , X4,3 , X3,4 , and X2,5 It would be interesting to find descriptions of these other moduli surfaces, some of which probably have infinitely many rational points. By contrast, the only rational points we found on X0,7 , the moduli space of maps with 7-cycles, corresponded to conjugates of the map φ of equation (0.1). As for preperiodic orbits of length 8, certainly X1,7 (Q) is nonempty. After all, each of the seven strictly preperiodic points of the map φ of equation (0.1) has forward orbit of length 8. The resulting elements of X1,7 did not show up directly in our search, however, because the coordinate changes required (to move any such preperiodic point x to ∞, φ(x) to 1, and φ2 (x) to 0) result in triples (x3 , x4 , x5 ) outside our search region. We found several other triples (x3 , x4 , x5 ) corresponding to rational forward orbits of length 8; they are listed in Table 3. We found 26 points in X6,2 , two in X5,3 , and none in any other surface Xm,n with m + n = 8. (It should be expected that 2-cycles are the easiest to realize, as a degree-two map has three fixed points, six 3-periodic points, and many more of any higher period, but only two 2-periodic points. Thus, the 2-periodic points are roots of a quadratic polynomial, while all the other periodic points are roots of higher-degree polynomials.) Note, however, that for m ≥ 1, points in the moduli space Xm,n generally come in sets of two. After all, if (x, φ) is a point in the moduli space, then unless x is a critical point, φ(x) has two rational preimages: x, and some other point y. Thus, (y, φ) is also a point in the same moduli space. The two points we found in X5,3 were related to each other in this way, as were 11 sets of two points we found in X6,2 . In Table 3, therefore, w=− 16 φ 330z 2 −187z−143 330z 2 +1217z+429 21z 2 −84z+63 21z 2 −16z−21 52z 2 −30z−22 52z 2 +245z+88 120z 2 −98z−22 120z 2 +749z+132 30z 2 −10z−20 30z 2 +7z−30 33z 2 −429z+396 33z 2 −197z+132 176z 2 +1397z−1573 176z 2 +500z−1144 1350z 2 −837z−513 1350z 2 +5585z+1710 700z 2 −95z−605 700z 2 +1336z+880 784z 2 −416z−368 784z 2 +3885z+644 1428z 2 −1668z+240 1428z 2 −1723z+900 308z 2 +19292z−19600 308z 2 +1937z+7700 9009z 2 −17094z+8085 9009z 2 −18454z−10395 5712z 2 −5937z+225 5712z 2 −137612z+5400 51480z 2 +910z−52390 51480z 2 +275477z−120900 24255z 2 −277830z+253575 24255z 2 +314788z+65205 orbit tail period total length 3 55 13 3 ∞, 1, 0, − 13 , − 11 , − , − , − , − 5 3 8 15 5 114 44 5 ∞, 1, 0, −3, 73 , − 13 , −7, − 32 , −7 6 2 8 1 3 4 9 4 ∞, 1, 0, − 4 , − 8 , −1, − 7 , − 26 , − 7 6 2 8 1 12 ∞, 1, 0, − 16 , − 29 , − 15 , − 12 , − , − 6 2 8 65 12 65 2 6 10 6 ∞, 1, 0, 23 , 10 , , , , 6 2 8 9 5 7 3 7 11 3 ∞, 1, 0, 3, 3 , 5, 33, 4 , 33 6 2 8 11 11 55 55 ∞, 1, 0, 11 , − , − , , 2, 6 2 8 8 2 4 16 16 3 9 72 72 ∞, 1, 0, − 10 , − 10 , − 35 , − 175 , − 16 , − 175 6 2 8 11 5 7 5 11 5 ∞, 1, 0, − 16 , − 7 , − 11 , − 6 , − 70 , − 6 6 2 8 2 1 20 ∞, 1, 0, − 47 , − 21 , − 87 , − 20 , , − 6 2 8 49 12 49 4 4 10 4 ∞, 1, 0, 15 , − 21 , 21 , − 47 , 12 , − 6 2 8 17 7 28 28 308 40 308 ∞, 1, 0, − 11 , −14, − 5 , − 17 , − 11 , − 17 6 2 8 35 63 77 63 ∞, 1, 0, − 79 , 77 , , , − , 6 2 8 27 11 31 78 31 1 1 3 117 2 117 ∞, 1, 0, 24 , 26 , 68 , 2992 , 35 , 2992 6 2 8 13 26 91 780 13 780 ∞, 1, 0, 30 , − 5 , − 11 , 253 , 36 , 253 6 2 8 5 49 105 15 105 ∞, 1, 0, 35 , − , − , , − , 6 2 8 9 18 3 13 154 13 Table 3. (∞, φ) ∈ PM02 (Q) in the search region with preperiodic orbit of length (at least) 8 we have listed only one point in X5,3 , and fifteen in X6,2 . The four unpaired points we found in X6,2 appear last in the table; the points they would be paired with lie outside our search region. Accounting for both preimages of each point in the strict forward orbit of ∞, each of the maps listed in Table 3 comes with 14 rational preperiodic points, assuming none of the points involved are critical images. And a simple but tedious computation shows that this assumption is accurate. (In fact, none of the maps in Table 3 has any Q-rational critical points at all.) Moreover, a longer computation — using Hutz’s algorithm [13] for computing Preper(φ, Q) — shows that in each case, Preper(φ, Q) is precisely this set of 14 points. The foregoing evidence from our data led us to propose parts (a)–(c) of Conjecture 5. The abundance of points in X6,2 suggests that X6,2 may be infinite, although we have not yet found a proof or disproof of this statement. Meanwhile, the surfaces Xm,n for other m, n with m + n = 8 all appear to have at most finitely many rational points. However, here the limitations of our search region, even though it is quite large, should inspire some caution. In particular, as noted above, we know that X1,7 must have some Q-rational points, but they all lie outside our region. That is, all such points (x3 , x4 , x5 ) have max{h(x3 ), h(x4 ), h(x5 )} > log 100. We close with Table 4, which gives the pairs (x, φ) in our search region with the smallest positive height ratios ĥφ (x)/h(φ) that we found — specifically, with ratio less than .0012. (Recall that h(φ) is the height of φ itself as given by Lemma 3.1 and equation (3.1). As in Tables 1 and 2, we manually computed the canonical heights and height ratios in the table to many more than the 15 iterates originally computed by the algorithm.) In addition, as with the preperiodic points in Table 3, pairs (x, φ) with small height ratio come in sets of two; Table 4 only lists one of the two pairs in this situation. DYNAMICS OF QUADRATIC MAPS φ 10z 2 −7z−3 10z 2 +37z+9 48165z 2 −54663z+6498 48165z 2 −49361z+1482 91z 2 +399z−490 91z 2 −16z−350 1701z 2 −427z−1274 1701z 2 −3222z+546 7z 2 −6z−1 7z 2 +20z−3 60z 2 −24z−36 60z 2 −143z+6 42z 2 −67z+25 42z 2 −75z+30 845z 2 −20z−825 845z 2 +3302z+2145 8190z 2 −10983z+2793 8190z 2 −23941z+7182 1625z 2 −1417z−208 1625z 2 +11075z+260 14z 2 −9z−5 14z 2 +76z+20 945z 2 −980z+35 945z 2 −3372z+315 90z 2 −20826z+20736 90z 2 −533z+1080 1375z 2 −847z−528 1375z 2 +4025z+1320 1400z 2 −2005z+605 1400z 2 −3148z+1100 290z 2 −966z+676 290z 2 −1077z+754 17 ĥφ (∞) ĥφ (∞)/h(φ) φ3 (x), φ4 (x), . . . 1 1 3 9 41 .00360 .000466 − 3 , − 5 , − 5 , − 12 , − 37 , − 15 , − 32 , − 105 ,... 41 57 38 76 57 12 2109 12084 .01425 .000747 , , , , , , ,... 13 39 65 65 13 2197 11525 7 14 14 28 28 329 975016 .01221 .000867 , − , , , 21, , − , ,... 5 11 3 13 23 591 446071 14 56 7 106 4424 6245167 .01553 .000919 − 73 , 14 , , − , , , , − ,... 27 9 9 9 171 3987 16848423 1 1 1 1 1 5 57 3403 .00721 .000931 , − 2 , − 3 , − 5 , − 14 , 41 , 16 , 8043 , . . . 3 3 6 3 78 .01128 .000935 −6, 34 , 10 , 5 , − 11 , − 48 , − 853 , − 96159 ,... 95 56528 5 1 2 3 13 8 23 2697 .00829 .000958 , , , , , , − , , . . . 6 2 3 4 21 7 6 2804 5 9 285 345 510 1395 829730 .01243 .001028 − 13 , − 13 , − 20 , − , − , , , − 3831763 ,... 13 221 403 169 4057 7 7 21 28 9 119 50022 .02187 .001040 , , , , , , , . . . 18 26 65 85 26 537 149725 7 1 71 1183 113139661 .01780 .001101 − 45 , − 13 , − 65 , 25 , − 13 , − 475 , − 40085 , 44931025 , . . . 50 35 1 3 1 4 3 611 .01192 .001125 − 4 , −1, − 7 , − 7 , − 11 , − 140 , − 2339 , − 308865 ,... 201037 1 35 1 5 1 35 26287 86934355 .02104 .001145 , , − , , , , , , . . . 9 27 7 21 3 132 85071 311687886 96 .02642 .001161 , − 72 , 99 , −27, 8, − 1251 , 131004 , − 532888651446 ,... 5 5 8 23 27343 4100060723 2 11 3 99 121 19017 36280937 .01778 .001183 − 5 , − 25 , − 5 , − 125 , −1, − 95 , − 10775 , − 10371365 , . . . 11 5 7 5 23 110 2809 23315575 .01621 .001186 , , , , , , , ,... 20 14 16 11 56 269 6844 56982922 26 2 10 18 122 190862 26084280098 .02290 .001198 , 2, , , , − , , ,... 29 5 11 7 109 209065 9085163135 Table 4. (∞, φ) ∈ PM02 (Q) in the search region with positive canonical height ratio less than .0012. Incidentally, the fifth map listed in Table 4 is simply a coordinate change of the first! More precisely, calling the first pair (∞, ψ ), the fifth is the equivalence class of (ψ (∞), ψ ), and the canonical height ratio of the latter pair is exactly double that of (∞, ψ ). This pair (∞, ψ ) provides further evidence for Conjecture 3. First discovered by Elkies in a similar but smaller-scale search [7], it showed up very early in our search, because the point (∞, ψ) ∈ PM02 is represented by the small-height triple (x3 , x4 , x5 ) = (−1/3, −1/5, −3/5). It is telling that the second-place candidate had height ratio one and a half times as large, even though our search extended to points in PM02 of far larger height. This observation led us to pose part (d) of Conjecture 5. After all, if there were pairs (x, φ) ∈ PM02 with even smaller positive canonical height ratio, presumably at least one of them would have shown up. Remark 3.7. The decision to check the height of φi (∞) for iterates i = 8, 10, 15 in Algorithm 3.5 was somewhat arbitrary, but the goal was to minimize the algorithm’s run time. In practice, most points tested tend to blow up in height rapidly under iteration, and therefore we did a preliminary check only a few iterations after x5 and another two iterations later, to avoid computing ten iterates of x5 for every candidate. On the other hand, the computation time required to test all the iterates i = 6, 7, . . . for each candidate also slowed the program down. Thus, we ultimately settled on testing only at i = 8, 10, 15 by trial and error on relatively small search regions. Similarly, the threshold of .002 as an upper bound for height ratios of interest was chosen by trial and error — small preliminary searches suggested there were plenty of interesting points with ratio below .002, but simply too many larger than that. 18 References [1] M. Baker, A lower bound for average values of dynamical Green’s functions, Math. Res. Lett. 13 (2006), 245–257. [2] R. Benedetto, Preperiodic points of polynomials over global fields, J. Reine Angew. Math. 608 (2007), 123–153. [3] R. Benedetto, B. Dickman, S. Joseph, B. Krause, D. Rubin, and X. Zhou, Computing points of small height for cubic polynomials, Involve 2 (2009), 37–64. [4] J. Blanc, J. K. Canci, and N. D. Elkies, Moduli spaces of quadratic rational maps with a marked periodic point of small order, preprint, 2013; available at arXiv:1305.1054. [5] G. Call and S. Goldstine, Canonical heights on projective space, J. Number Theory 63 (1997), 211–243. [6] L. Carleson and T. Gamelin, Complex Dynamics, Springer-Verlag, New York 1991. [7] N. Elkies, personal communication, 2012–13. [8] E. V. Flynn, B. Poonen, and E. Schaefer, Cycles of quadratic polynomials and rational points on a genus-2 curve, Duke Math. J. 90 (1997), 435–463. [9] J. R. Doyle, X. Faber, and D. Krumm, Preperiodic points for quadratic polynomials over quadratic fields, preprint, 2013; available at arXiv:1309.6401. [10] A. Gillette, Searching for Small Canonical Heights: Finding a Minumum Speed Limit on the Road to Infinity, undergraduate thesis, Amherst College, 2004. [11] F. Gouvêa, p-adic Numbers. An Introduction, 2nd ed., Springer-Verlag, Berlin, 1997. [12] M. Hindry, J. Silverman, Diophantine Geometry: an Introduction, Springer-Verlag, New York, 2000. [13] B. Hutz, Determination of all rational preperiodic points for morphisms of PN , Math. Comp., to appear. [14] P. Ingram, Lower bounds on the canonical height associated to the morphism φ(z) = z d + c, Monatsh. Math. 157 (2009), 69–89. [15] S. Lang, Fundamentals of Diophantine Geometry, Springer-Verlag, New York, 1983. [16] M. Manes, Arithmetic Dynamics of Rational Maps, Ph.D. thesis, Brown University, 2007. [17] J. Milnor, Geometry and dynamics of quadratic rational maps, with an appendix by the author and Lei Tan, Experiment. Math. 2 (1993), 37-83. [18] P. Morton, Arithmetic properties of periodic points of quadratic maps, Acta Arith. 62 (1992), 343–372. [19] P. Morton, Arithmetic properties of periodic points of quadratic maps, II, Acta Arith. 87 (1998), 89–102. [20] P. Morton and J. Silverman, Rational periodic points of rational functions, Inter. Math. Res. Notices 2 (1994), 97–110. [21] P. Morton and J. Silverman, Periodic points, multiplicities, and dynamical units, J. Reine Angew. Math. 461 (1995), 81–122. [22] W. Narkiewicz, Polynomial cycles in algebraic number fields, Coll. Math. 58 (1989), 151–155. [23] D. Northcott, Periodic points of an algebraic variety, Ann. Math. 51 (1950), 167–177. [24] T. Pezda, Polynomial cycles in certain local domains, Acta Arith. 66 (1994), 11–22. [25] B. Poonen, The classification of rational preperiodic points of quadratic polynomials over Q: a refined conjecture, Math. Z. 228 (1998), 11–29. DYNAMICS OF QUADRATIC MAPS [26] [27] [28] [29] J. Silverman and J. Tate, Rational Points on Elliptic Curves, Springer, New York, 1992. J. Silverman, The space of rational maps on P1 , Duke Math. J. 94 (1998), 41-77. J. Silverman, The Arithmetic of Dynamical Systems, Springer, New York, 2007. M. Zieve, Cycles of Polynomial Mappings, Ph.D. thesis, UC Berkeley, 1996. 19