betaFeedback

WorksMathematics

Contact multiplicity and a sharp six-map bound for collinear self-similar dendrites

Abstract

Let K⊂CK\subset\mathbb{C} be the attractor of n≥2n\ge2 similarities z↦λz+bkz\mapstoλz+b_k, 0<∣λ∣<10<|λ|<1, whose fixed points are evenly spaced on a line, and suppose that KK does not lie in a line. We prove that KK can be a dendrite only if n≤6n\le6, and that such dendrites exist for every 2≤n≤62\le n\le6; no separation or finite-intersection hypothesis is imposed. Hence, for n≥7n\ge7, every connected KK of this kind has interior points or separates the plane. Neighboring pieces of KK meet in similar copies of a single contact set. In a dendrite exactly one first-level piece meets this set, whereas for n≥7n\ge7 at least n−⌈2n−1⌉≥2n-\lceil2\sqrt{n-1}\rceil\ge2 do, a bound attained by explicit quadratic parameters. The proof combines a planar translation obstruction for replicated contact arcs with an estimate based on Jensen's formula. After normalizing to z↦t+z/cz\mapsto t+z/c, t∈{−n+1,−n+3,…,n−1}t\in\{-n+1,-n+3,\ldots,n-1\}, the connectedness locus is the closure, outside the unit disk, of the roots of monic integer polynomials of height at most n−1n-1. The same estimate shows that, for n≥7n\ge7, every such polynomial with a non-real root cc satisfying ∣c∣2≥⌈2n−1⌉+1|c|^2\ge\lceil2\sqrt{n-1}\rceil+1 has subleading coefficient of absolute value at most ⌈2n−1⌉\lceil2\sqrt{n-1}\rceil, and that this bound is sharp. Accordingly, for n≥7n\ge7, at every non-real boundary point cc of the locus with ∣c∣>1|c|>1, each expansion c=∑j≥0djc−jc=\sum_{j\ge0}d_jc^{-j} with integer digits ∣dj∣≤n−1|d_j|\le n-1 has ∣d0∣≤⌈2n−1⌉|d_0|\le\lceil2\sqrt{n-1}\rceil.

Introduction

Let n≥2n \ge2, let λ∈C\lambda\in\mathbb{C} with 0<∣λ∣<10 < |\lambda| < 1, and let K⊂CK \subset\mathbb{C} be the attractor [27] of the similarities

Sk(z)=λz+b0+kδ(0≤k<n, δ≠0).(1)S_k(z)=\lambda z+b_0+k\delta\qquad(0\le k<n,\ \delta\ne0). \tag*{(1)}

The maps share their linear part, and their fixed points (b0+kδ)/(1−λ)(b_0+k\delta)/(1-\lambda) are evenly spaced on a line; conversely, every system of nn orientation-preserving similarities with a common linear part and evenly spaced fixed points has this form. If λ\lambda is real, the line through the fixed points is invariant and contains KK. If λ\lambda is not real, KK lies in no line: a line L⊇KL \supseteq K would give S0(K)⊆L∩S0(L)S_0(K) \subseteq L \cap S_0(L), a single point, whereas S0S_0 is injective and KK contains the distinct fixed points of S0S_0 and S1S_1.

A dendrite is a locally connected continuum containing no simple closed curve; its points are joined by unique arcs [15, 34]. In the plane, a locally connected continuum is a dendrite if and only if it has empty interior and does not separate the plane. Indeed, a dendrite has empty interior, since a disk contains a simple closed curve, and it is contractible, so Alexander duality shows that it does not separate the plane; conversely, if a non-separating continuum contains a simple closed curve, then the bounded complementary domain of that curve lies in the continuum. Planar self-similar dendrites are thus exactly the self-similar continua that are simply connected, in the sense of not separating the plane, and have empty interior. For such a set, the identifications between symbolic addresses must be compatible with the uniqueness of arcs. We ask how many evenly spaced maps can produce one.

Theorem 1.1 (Sharp six-map bound). Let KK be the attractor of (1), and suppose that KK does not lie in a line. If KK is a dendrite, then 2≤n≤62 \le n \le6, and for each 2≤n≤62 \le n \le6 there are such dendrites. In particular, for n≥7n \ge7 every connected KK of this form contains a simple closed curve, and therefore has interior points or separates the plane.

Dendrites and other tree-like sets are basic models in analysis on fractals. Kigami developed harmonic calculus on limits of networks with dendrites as a principal application [29], and self-similar dendrites such as Hata’s tree-like set are standard examples of postcritically finite self-similar sets [30]. Dendrites also arise as Julia sets: if every critical point of a polynomial is strictly preperiodic, as for z2+iz^2+i, its Julia set is a dendrite [16, 33]. Julia sets of postcritically finite rational maps are topologically self-similar [28], and Laplacians have been constructed on several quadratic Julia sets [24, 36]. For binary systems with a common multiplier and singleton overlap, the attractor carries a folded inverse map; in the postcritically finite case, Eroğlu, Rohde, and Solomyak prove a quasisymmetric conjugacy between this map and a quadratic polynomial on its Julia set [17, Theorem 1.1]. Theorem 1.1 shows that evenly spaced systems with a common non-real multiplier produce such tree-like attractors only when there are at most six maps.

The affine map z↦2(z−w)/δz \mapsto2(z-w)/\delta, with w=(b0+(n−1)δ/2)/(1−λ)w=(b_0+(n-1)\delta/2)/(1-\lambda), conjugates (1) to the system ft(z)=t+c−1zf_t(z)=t+c^{-1}z, t∈Ant\in A_n, where c=1/λc=1/\lambda and An={−n+1,−n+3,…,n−1}A_n=\{-n+1,-n+3,\ldots,n-1\}. It carries KK onto

E(c,n)={∑j≥0tjc−j:tj∈An},∣c∣>1,(2)E(c,n)=\left\{\sum_{j\geq0}t_jc^{-j}:t_j\in A_n\right\},\qquad|c|>1, \tag*{(2)}

and KK lies in a line exactly when cc is real. We work in this normalization throughout. Write Mn={c:∣c∣>1, E(c,n) is connected}\mathcal{M}_n=\{c:|c|>1,\ E(c,n)\text{ is connected}\} for its connectedness locus. In these terms, Theorem 1.1 states that E(c,n)E(c,n) is a dendrite for some non-real cc exactly when 2≤n≤62\leq n\leq6; Figure 1 shows one for each such nn.

Figure 1. Non-real dendrites at all five admissible alphabet sizes, with the exact parameters in (6). Colors and labels identify the first-level pieces P0,…,Pn−1P_0,\ldots,P_{n-1} of (3), in increasing digit order; white circles mark the contact points Pi∩Pi+1={ai+1}P_i\cap P_{i+1}=\{a_i+1\}. Panels (a)–(e) are drawn at different scales, and panels (f)–(j) magnify the squares in (a)–(e) five times. The point samples illustrate the sets; exact intersection certificates and proofs are given in § 7.

Barnsley and Harrington initiated the parameter-space study of the binary family [8]. They considered the disconnectedness set for a pair of similarities with a common complex multiplier. The connectedness convention and reciprocal change λ=1/c\lambda=1/c give M2\mathcal{M}_2. Bousch proved connectedness, local connectedness, and Hölder path connectedness of the binary locus [12]. His thesis also treats the consecutive alphabets {0,1,…,n−1}\{0,1,\ldots,n-1\} for every n≥2n\geq2, proving that connectedness is equivalent to overlap between neighboring first-level pieces [11 Chapter 1, p. 41]. An affine change of coordinates identifies these systems with (2). Bandt [3] and Solomyak and Xu [39] developed the study of the binary locus, Shmerkin and Solomyak related zeros of {−1,0,1}\{-1,0,1\} power series to connectedness loci of self-affine sets [37], and Calegari, Koch, and Walker proved Bandt’s conjecture on the binary locus [13]. Espigulé, Juher, and Saldaña [23] investigated the collinear loci uniformly in nn. When cc is a non-real quadratic integer with ∣c∣2=n|c|^2=n, the set E(c,n)E(c,n) is a planar self-affine tile with a collinear digit set; the connectedness, disklikeness, and topology of such tiles were studied by Kirat and Lau [31], Leung and Lau [32], and Akiyama, Loridant, and Thuswaldner [1]. Bandt and Hung studied the analogous family whose fixed points are the vertices of a regular polygon [5].

The non-real restriction is essential: E(2,n)=[−2(n−1),2(n−1)]E(2,n)=[-2(n-1),2(n-1)] is an interval, hence a dendrite, for every n≥2n\geq2. Connected self-similar sets are locally connected [25], so the exclusion amounts to forcing a simple closed curve. Finite overlaps are relevant to this problem: Bandt and Rao prove that finite first-level intersections in a connected planar self-similar set imply the open set condition [7]. Self-similar dendrites can nevertheless have nontrivial subdendrites as intersections, even under that condition [2]. Our argument uses the equal spacing of the digits and applies without a separation or finite-intersection hypothesis.

The exclusion follows from a quantitative statement about contacts. Write

K=E(c,n),ai=−n+1+2i,Pi=ai+c−1K(0≤i<n),Q=K∩(2c+K),(3)\begin{aligned} K&=E(c,n),\qquad a_i=-n+1+2i,\qquad P_i=a_i+c^{-1}K\quad(0\leq i<n),\\ Q&=K\cap(2c+K), \tag*{(3)} \end{aligned}

so that Pi∩Pi+1=ai+c−1QP_i\cap P_{i+1}=a_i+c^{-1}Q for 0≤i<n−10\leq i<n-1. Thus QQ is the common normalized contact between neighboring pieces. Define

k(c,n)=#{i:Pi∩(2c+K)≠∅}.(4)k(c,n)=\#\{i:P_i\cap(2c+K)\ne\varnothing\}. \tag*{(4)}

Since Pi⊆KP_i\subseteq K, equivalently k(c,n)=#{i:Pi∩Q≠∅}k(c,n)=\#\{i:P_i\cap Q\ne\varnothing\}. This counts participating pieces, not points or connected components of QQ.

Theorem 1.2 (Contact multiplicity). Let c∈Mn∖Rc\in\mathcal{M}_n\setminus\mathbb{R}.

(1) If E(c,n)E(c,n) is a dendrite, then exactly one of Pn−1∩(2c+P0)P_{n-1}\cap(2c+P_0) and P0∩(2c+Pn−1)P_0\cap(2c+P_{n-1}) is nonempty, all other sets Pi∩(2c+Pj)P_i\cap(2c+P_j) are empty, and k(c,n)=1k(c,n)=1.

(2) If n≥7n\geq7, then

k(c,n)≥kn:=n−⌈2n−1⌉≥2.(5)k(c,n)\geq k_n:=n-\left\lceil2\sqrt{n-1}\right\rceil\geq2. \tag*{(5)}

For every n≥7n\geq7, equality is attained at a non-real quadratic parameter. Thus knk_n is the exact minimum of k(⋅,n)k(\cdot,n) over Mn∖R\mathcal{M}_n\setminus\mathbb{R}. The sequence (kn)n≥7(k_n)_{n\geq7} is nondecreasing and kn=n−2n+O(1)k_n=n-2\sqrt{n}+O(1).

For example, k7=k8=2k_7=k_8=2, k9=3k_9=3, k10=4k_{10}=4, k20=11k_{20}=11, and k100=80k_{100}=80. On the imaginary axis every piece participates: k(ir,n)=nk(ir,n)=n whenever 1<r≤n1<r\leq\sqrt{n} (Proposition 5.3).

The proof has two independent parts. Equal digit offsets replicate contacts across consecutive pieces. If two such contacts occur inside a dendrite, their unique arcs produce a forbidden translation parallelogram. Analytically, a sufficiently large first digit forces a small real zero of the corresponding power series. Jensen’s formula then excludes a non-real zero at the required modulus, in the spirit of the estimates of Beaucoup, Borwein, Boyd, and Pinner [9]. Connectivity at a smaller alphabet handles the complementary range of moduli.

The same estimate constrains roots of integer polynomials. For n=2n = 2, the description of the connectedness locus through zeros of power series and polynomials with coefficients in {−1,0,1}\{-1, 0, 1\} is classical [13, 37]. For every nn, Mn\mathcal{M}_n is the closure, outside the closed unit disk, of the set of roots of monic integer polynomials of height at most n−1n - 1, that is, with all coefficients in {−n+1,…,n−1}\{-n + 1, \ldots, n - 1\} [22 Proposition 2.3]. The closure in C\mathbb{C} of the roots of all integer polynomials of height at most n−1n-1, monic or not, is Mn∪{∣z∣=1}∪{0}∪{1/c:c∈Mn}\mathcal{M}_n \cup\{|z| = 1\} \cup\{0\} \cup\{1/c : c \in\mathcal{M}_n\} [22 Proposition 1.1]; in particular, non-monic polynomials add nothing outside the closed unit disk. A root cc of zk+a1zk−1+⋯+akz^k + a_1z^{k-1} + \cdots+ a_k satisfies c=−a1−a2c−1−⋯−akc1−kc = -a_1 - a_2c^{-1} - \cdots- a_kc^{1-k}, a finite digit expansion of the kind in Lemma 2.1, whose first digit is −a1-a_1.

Corollary 1.3 (Subleading coefficients of non-real roots). Let n≥7n \ge7, and let c∉Rc \notin\mathbb{R} with ∣c∣2≥⌈2n−1⌉+1|c|^2 \ge\lceil2\sqrt{n-1}\rceil+ 1 be a root of a monic polynomial zk+a1zk−1+⋯+akz^k + a_1z^{k-1} + \cdots+ a_k with integer coefficients of absolute value at most n−1n - 1. Then ∣a1∣≤⌈2n−1⌉|a_1| \le\lceil2\sqrt{n-1}\rceil. The bound is sharp: for every n≥7n \ge7, some such root has ∣c∣2>n|c|^2 > n and ∣a1∣=⌈2n−1⌉|a_1| = \lceil2\sqrt{n-1}\rceil.

The proof is given in § 6, after Theorem 6.2, because the sharpness examples use the quadratic parameters introduced there.

Since ⌈2n−1⌉+1≤n−1\lceil2\sqrt{n-1}\rceil+ 1 \le n - 1 for n≥7n \ge7, the corollary applies in particular to every non-real root with ∣c∣≥n|c| \ge\sqrt{n}; for real coefficients of absolute value at most n−1n - 1, the same argument gives ∣a1∣<⌈2n−1⌉+1|a_1| < \lceil2\sqrt{n-1}\rceil+ 1. For such roots the extreme values a1=±(n−1)a_1 = \pm(n - 1) are excluded from n=7n = 7 on, whereas for 3≤n≤63 \le n \le6 the construction in the proof of Corollary 1.3 gives non-real roots with ∣c∣2>n|c|^2 > n and ∣a1∣=n−1|a_1| = n - 1. Since {1<∣c∣<n}⊆Mn\{1 < |c| < \sqrt{n}\} \subseteq\mathcal{M}_n by Lemma 2.1 and ∣c∣<1+n−1|c| < 1 + \sqrt{n - 1} on Mn∖R\mathcal{M}_n \setminus\mathbb{R} by Proposition 6.1, the non-real points of ∂Mn\partial\mathcal{M}_n with ∣c∣>1|c| > 1, which belong to Mn\mathcal{M}_n because Mn\mathcal{M}_n is closed in {∣c∣>1}\{|c| > 1\} [22 Proposition 2.3], lie in the annulus n≤∣c∣<1+n−1\sqrt{n} \le|c| < 1 + \sqrt{n - 1}. There, by Theorem 6.2, every expansion c=∑j≥0djc−jc = \sum_{j \ge0} d_jc^{-j} with integer digits ∣dj∣≤n−1|d_j| \le n - 1 has ∣d0∣≤⌈2n−1⌉<n−1|d_0| \le\lceil2\sqrt{n-1}\rceil< n - 1 once n≥7n \ge7. Geometrically, the non-real parts of the parameter regions selected by large first digits, and even of their convex relaxations, recede inside the connectedness disk from n=7n = 7 on; Figure 2 and Figure 3 display these vanishing fronts. In § 6 we determine exactly which first digits occur at non-real parameters outside the connectedness disk, give support-function bounds for the corresponding parameter sets, and use a quadratic family to attain the minimum in Theorem 1.2. Figure 4 shows such a parameter, c=3+i2c = 3 + i\sqrt{2} with n=7n = 7 and k(c,7)=2k(c, 7) = 2: the pieces P5P_5 and P6P_6 meet 2c+P02c + P_0 and 2c+P12c + P_1, respectively, while P6P_6 does not meet 2c+P02c + P_0. In a dendrite, by contrast, the only contact is between Pn−1P_{n-1} and 2c+P02c + P_0, or P0P_0 and 2c+Pn−12c + P_{n-1} (Theorem 1.2(1)).

Figure 2. Convex first-digit loci L^7,m\widehat{\mathcal{L}}_{7,m} for m=0,…,6m=0,\ldots,6 in the first quadrant, in blue against their union in gray; panel (a) shows that union, numbered by how many loci contain each point. The dashed arc is ∣c∣=7|c|=\sqrt{7} and the thin gray arc is ∣c∣=1|c|=1; white diamonds mark c6,m∈L7,mc_{6,m}\in\mathcal{L}_{7,m} from (25), 1≤m≤51\le m\le5. The loci overlap near the unit disk, whereas the extreme locus m=6m=6 has no non-real part outside the dashed arc. These are numerical renderings of the convex relaxation.

Figure 3. First-quadrant convex first-digit geometry at six alphabet sizes. The bands are level sets of the largest convex-admissible nonnegative label mmax⁡m_{\max}, whose value is printed in each band; the labeled loci themselves overlap. Dark blue marks mmax⁡=n−1m_{\max}=n-1 and lighter blue ⌈2n−1⌉<mmax⁡<n−1\lceil2\sqrt{n-1}\rceil< m_{\max} < n-1; dashed arcs are ∣c∣=n|c|=\sqrt{n} and thin gray arcs ∣c∣=1|c|=1. The upper right corner of each panel gives the largest convex-admissible label at non-real grid points with ∣c∣≥n|c| \ge\sqrt{n}, namely min⁡{n−1,⌈2n−1⌉}\min\{n-1,\lceil2\sqrt{n-1}\rceil\}, as in Theorem 6.2; white diamonds mark (25) with this mm. Blue extends outside ∣c∣=n|c|=\sqrt{n} only for n≤6n \le6.

The attractor $K$ and its translate $2c+K$, with convex hulls of first-level pieces

Figure 4. The attractor K=E(c,7)K=E(c,7) and its translate 2c+K2c+K at the parameter c=3+i2c=3+i\sqrt{2} of Proposition 6.3, marked by the white diamond in Figure 3(d), with the convex hulls of their first-level pieces; pieces with the same index share a hue. The hulls of PiP_i and 2c+Pj2c+P_j intersect exactly when i−j=5i-j=5, because 55 is the only convex-admissible label at cc (Proposition 6.1 and Theorem 6.2). In particular, the rightmost piece P6P_6 of KK does not meet the leftmost piece 2c+P02c+P_0 of the translate. By (26), c−1∈P5∩(2c+P0)c-1\in P_5\cap(2c+P_0) and c+1∈P6∩(2c+P1)c+1\in P_6\cap(2c+P_1) (white circles), so two pieces of KK meet 2c+K2c+K; these four pieces are drawn in full color, the others in light tints. The symmetry z↦2c−zz\mapsto2c-z about cc (black dot) exchanges KK and 2c+K2c+K, maps PiP_i onto 2c+P6−i2c+P_{6-i}, and swaps c−1c-1 and c+1c+1. The point samples are numerical illustrations.

The examples in Figure 1 establish sharpness at the following exact parameters:

{n=2:c is the root in the upper half-plane of c3−2c2+2=0,3≤n≤6:c=n+i8(n−1)−n22.(6)\begin{cases} n = 2: & c \text{ is the root in the upper half-plane of } c^3 - 2c^2 + 2 = 0,\\ 3 \le n \le6: & c = \dfrac{n + i\sqrt{8(n-1) - n^2}}{2}. \tag*{(6)} \end{cases}

Up to affine normalization and complex conjugation, the binary example is the one in [17 Figure 3]. In each case neighboring pieces meet at one point, non-neighboring pieces are disjoint, and the first-level intersection graph is a path. § 7 proves these assertions by finite exact inverse-state calculations. The computations are included in full and can also be reproduced by the accompanying verification program.

Relation to other work

The present classification concerns the possible alphabet sizes; a classification of all dendrite parameters, even at a fixed size, is a finer problem.

Recent work on the same collinear family concerns the geometry of Mn\mathcal{M}_n. Espigulé and Juher construct semialgebraic interior layers using finite capture and prove uniform bounds on enclosure-tree branching [22]. A separate paper proves that every non-real point of Mn\mathcal{M}_n is a limit of interior points [20], and infinitely many holes of the loci are exhibited in [19]. Here the object of study is the topology of the individual attractor. Our contact count is linked to the inverse search through the replication formula in Lemma 2.2; Remark 5.4 explains the distinction between the number of admissible offsets and the number of pieces they connect. The normalization follows the thesis [18], whose Chapter 3 and Appendix C announce the dendrite classification. The present paper gives the proofs and the sharp minimum contact count.

Connectivity and replicated contacts

We first recall Bousch’s connectedness criterion for consecutive collinear digits [11 Chapter 1, p. 41]. We include a proof together with the separation consequence and the classical connectedness-disk estimate [38 Lemma 2.5], in the form of [23 Proposition 5(i)], needed below. For p≥2p \ge2, put Dp={−p+1,−p+2,…,p−1}D_p=\{-p+1,-p+2,\ldots,p-1\}. Digit subtraction gives

12(E(c,p)−E(c,p))={∑j≥0djc−j:dj∈Dp}.(7)\frac{1}{2}\left(E(c,p)-E(c,p)\right)=\left\{\sum_{j\geq0}d_jc^{-j}:d_j\in D_p\right\}. \tag*{(7)}

Every digit in DpD_p can be lifted to a pair of digits from ApA_p; thus this equality holds in both directions.

Lemma 2.1 (Neighboring contact and separation). For ∣c∣>1|c|>1 and p≥2p\geq2, the following are equivalent:

(1) E(c,p)E(c,p) is connected;

(2) two distinct first-level pieces intersect;

(3) every neighboring pair of first-level pieces intersects;

(4) c=∑j≥0djc−jc=\sum_{j\geq0}d_jc^{-j} for some dj∈Dpd_j\in D_p.

If E(c,p)E(c,p) is disconnected, its first-level pieces are pairwise disjoint. Moreover,

1<∣c∣<p⟹c∈Mp.(8)1<|c|<\sqrt{p}\quad\Longrightarrow\quad c\in\mathcal{M}_p. \tag*{(8)}

Proof. We use the elementary planar crossing fact that a left-to-right continuum and a bottom-to-top continuum in a rectangle must meet; see, for example, [35 Lemma 7]. The argument below is Bousch’s lemme de coinçage [11 Chapter 1, p. 28]. The same overlap criterion can be obtained from Hopf’s horizontal-chord theorem [26]; see [22 Proposition 2.2 and Lemma A.1].

Let LL be a continuum and suppose two translates ai+La_i+L and aj+La_j+L intersect. Choose the smallest positive index gap ss giving such an intersection. Suppose that LL has positive vertical extent and s≥2s\geq2, and let RR be the bounding rectangle of the continuum (ai+L)∪(ai+s+L)(a_i+L)\cup(a_{i+s}+L). This continuum meets both vertical sides of RR. The intermediate translate ai+1+La_{i+1}+L lies in RR and meets both horizontal sides, because all translates have the vertical extent of LL. An intersection gives a gap 11 or s−1s-1, a contradiction. Thus s=1s=1. If LL has zero vertical extent, it is a horizontal interval and the same conclusion follows from its length. Translation invariance then makes all neighboring translates intersect.

To apply this without assuming the attractor connected, let Kp=E(c,p)K_p=E(c,p), C0=conv⁡KpC_0=\operatorname{conv}K_p, and

Cr+1=⋃t∈Ap(t+c−1Cr).C_{r+1}=\bigcup_{t\in A_p}\left(t+c^{-1}C_r\right).

These compact sets are nested, contain KpK_p, and converge to KpK_p. Suppose two actual first-level pieces intersect. Starting with C0C_0, inductively assume CrC_r connected. The translates of c−1Crc^{-1}C_r inherit that intersection, so the preceding crossing argument shows that all neighboring translates meet. Hence Cr+1C_{r+1} is connected. Their nested intersection KpK_p is therefore connected. Conversely, a connected finite union of nonempty compact pieces cannot have those pieces pairwise disjoint. Applying the crossing argument directly to the connected set c−1Kpc^{-1}K_p proves the equivalence of the first three conditions. The fourth follows from (7), since neighboring contact is equivalent to 2c∈Kp−Kp2c\in K_p-K_p.

If KpK_p is disconnected, the established equivalence gives positive distance between every pair of first-level pieces. For sufficiently small ϵ>0\epsilon>0, let OO be the open ϵ\epsilon-neighborhood of KpK_p. The images ft(O)f_t(O) are pairwise disjoint, and each lies in OO, since its points have distance less than ε/∣c∣\varepsilon/|c| from KpK_p. Taking planar area, which is finite and positive on OO, gives p∣c∣−2area⁡(O)≤area⁡(O)p|c|^{-2}\operatorname{area}(O) \le\operatorname{area}(O). Thus ∣c∣2≥p|c|^2 \ge p, proving (8).

For a connected KK, each consecutive block Pi∪⋯∪PjP_i \cup\cdots\cup P_j is consequently a continuum. This fact will control the arcs used in the topological argument.

Set h=n−1h=n-1, the height bound in Corollary 1.3, so that Dn={−h,…,h}D_n=\{-h,\ldots,h\}, and introduce the admissible offsets

A(c,n)={m∈Z:∣m∣≤h,c=m+∑j≥1djc−j for some dj∈Dn}.(9)\mathcal{A}(c,n)=\left\{m\in\mathbb{Z}: |m|\le h,\quad c=m+\sum_{j\ge1}d_jc^{-j}\text{ for some }d_j\in D_n\right\}. \tag*{(9)}

By Lemma 2.1, this set is nonempty whenever c∈Mnc\in\mathcal{M}_n.

The replication of overlaps under equal digit differences already appears in Bousch’s translation symmetries of the overlap graph [11 Chapter 1, p. 38]. The next lemma records this mechanism for the contact K∩(2c+K)K\cap(2c+K) and counts the resulting pairs.

Lemma 2.2 (Contact replication). For 0≤i,j<n0\le i,j<n,

Pi∩(2c+Pj)≠∅⟺i−j∈A(c,n).(10)P_i\cap(2c+P_j)\ne\varnothing\quad\Longleftrightarrow\quad i-j\in\mathcal{A}(c,n). \tag*{(10)}

Each m∈A(c,n)m\in\mathcal{A}(c,n) gives exactly n−∣m∣n-|m| intersecting ordered pairs, whose intersection sets are successive translates by 22. In particular,

J(c,n):=#{(i,j):Pi∩(2c+Pj)≠∅}=∑m∈A(c,n)(n−∣m∣),(11)J(c,n):=\#\{(i,j):P_i\cap(2c+P_j)\ne\varnothing\}=\sum_{m\in\mathcal{A}(c,n)}(n-|m|), \tag*{(11)}

and

k(c,n)≥n−min⁡m∈A(c,n)∣m∣.(12)k(c,n)\ge n-\min_{m\in\mathcal{A}(c,n)}|m|. \tag*{(12)}

Proof. An equality ai+u/c=2c+aj+v/ca_i+u/c=2c+a_j+v/c, with u,v∈Ku,v\in K, is equivalent to u−v=2c(c−(i−j))u-v=2c(c-(i-j)). Using (7) gives (2.4). For a fixed $m$, there are exactly $n-|m|$ pairs with $i-j=m$. Incrementing both indices by one translates the intersection by $2$. Their left indices are distinct, which proves (2.6); summation gives {xref}(2.5)`.

The topological obstruction

We use standard properties of dendrites: subcontinua are dendrites; nonempty intersections of subcontinua are connected; each pair of points is joined by a unique arc; and a dendrite is contractible [15].

Lemma 3.1 (Translation parallelogram). Let TT be a planar dendrite and let u,v∈Cu,v\in\mathbb{C} be linearly independent over R\mathbb{R}. There are no paths α\alpha from pp to p+up+u and β\beta from pp to p+vp+v such that

α,α+v,β,β+u\alpha,\quad\alpha+v,\quad\beta,\quad\beta+u

all lie in TT.

Proof. These four paths define a continuous map on the boundary of the unit square, using identical parametrizations on opposite sides. Contractibility of TT extends it to a continuous map F:[0,1]2→TF:[0,1]^2\to T. Extend FF over the plane by

H(r+x,s+y)=F(x,y)+ru+sv(r,s∈Z, 0≤x,y≤1).H(r+x,s+y)=F(x,y)+ru+sv\qquad(r,s\in\mathbb{Z},\ 0\le x,y\le1).

The prescribed boundary maps agree on shared edges, so HH is well-defined and continuous. For the invertible real-linear map L(x,y)=xu+yvL(x,y)=xu+yv, the function H−LH-L is periodic and bounded. For any z∈Cz\in\mathbb{C}, the map

x⟼L−1z−L−1(H(x)−Lx)x\longmapsto L^{-1}z-L^{-1}(H(x)-Lx)

takes a sufficiently large closed disk into itself. Brouwer’s fixed-point theorem shows that H(x)=zH(x)=z has a solution. Thus HH is onto. But its image is contained in ⋃r,s∈Z(T+ru+sv)\bigcup_{r,s\in\mathbb{Z}}(T+ru+sv). A planar dendrite has empty interior, since a disk contains a simple closed curve; its translates are compact and nowhere dense. Baire’s theorem rules out this covering of the plane. □

Proposition 3.2 (One participating piece). If c∉Rc\notin\mathbb{R} and K=E(c,n)K=E(c,n) is a dendrite, then precisely one of

Pn−1∩(2c+P0),P0∩(2c+Pn−1)P_{n-1}\cap(2c+P_0),\qquad P_0\cap(2c+P_{n-1})

is nonempty, and all other ordered cross-contacts are empty. Consequently k(c,n)=1k(c,n)=1.

Proof. The contact QQ is nonempty by Lemma 2.1. For adjacent digits, the union of the corresponding pieces is a subcontinuum of KK, and is similar to K∪(2c+K)K\cup(2c+K). The latter union is therefore a dendrite. Its subcontinua KK and 2c+K2c+K have connected intersection QQ.

Suppose m∈A(c,n)m\in A(c,n) and ∣m∣≤n−2|m|\le n-2. Choose i,ji,j with i−j=mi-j=m and i+1,j+1<ni+1,j+1<n, and take x∈Pi∩(2c+Pj)x\in P_i\cap(2c+P_j). Replication gives x+2∈Pi+1∩(2c+Pj+1)x+2\in P_{i+1}\cap(2c+P_{j+1}). The unique arc γ\gamma from xx to x+2x+2 lies in QQ, so both γ\gamma and γ−2c\gamma-2c lie in KK. The unique arc δ\delta from x−2cx-2c to xx lies in the consecutive block

B=Pmin⁡(i,j)∪⋯∪Pmax⁡(i,j),B=P_{\min(i,j)}\cup\cdots\cup P_{\max(i,j)},

because its endpoints lie in PjP_j and PiP_i and BB is a subcontinuum. The shifted block B+2B+2 also lies in KK, giving δ+2⊆K\delta+2\subseteq K. Now apply Lemma 3.1 with

p=x−2c,u=2,v=2c,α=γ−2c,β=δ.p=x-2c,\qquad u=2,\qquad v=2c,\qquad\alpha=\gamma-2c,\qquad\beta=\delta.

The required vectors are independent because cc is non-real. This contradiction excludes every nonextreme offset.

Let Q+=Pn−1∩(2c+P0)Q_+=P_{n-1}\cap(2c+P_0) and Q−=P0∩(2c+Pn−1)Q_-=P_0\cap(2c+P_{n-1}). We have Q=Q+∪Q−Q=Q_+\cup Q_-. If both sets were nonempty, connectedness of QQ would force Q+∩Q−≠∅Q_+\cap Q_-\ne\varnothing. A point of their intersection would lie in P0∩(2c+P0)P_0\cap(2c+P_0), producing the already excluded offset 00. Thus exactly one extreme contact occurs. A point of QQ in any other left piece would produce an additional cross-contact, so exactly one left piece participates. □

The sharp uniform contact bound

Zeros of power series 1+∑i≥1aizi1+\sum_{i\ge1}a_i z^i with real coefficients in [−g,g][-g,g] were studied by Beaucoup, Borwein, Boyd, and Pinner [9, 10]. Among other results, they show that for g≥3+22g\ge3+2\sqrt{2} the smallest disk containing two zeros has radius (g+1)−1(\sqrt{g}+1)^{-1}, the smallest possible double real zero, and they compute the smallest circle containing a real zero and a pair of conjugate zeros of a given argument [9]. The next lemma adapts these ideas to a prescribed large first coefficient, using only a bound on the remaining coefficients, which need not be integers. Its hypothesis h≥6h\ge6 is the integer form of 1+2h≤h1+2\sqrt{h}\le h, that is, h≥3+22h\ge3+2\sqrt{2}. For m=1+2hm=1+2\sqrt{h}, the series 1−mz+h∑j≥2zj1-mz+h\sum_{j\ge2}z^j used for sharpness below has the double zero (1+h)−1(1+\sqrt{h})^{-1}.

Lemma 4.1 (First-coefficient exclusion). Let h≥6h\ge6 be an integer and q=⌊2h⌋+1q=\lfloor2\sqrt{h}\rfloor+1. If

F(z)=1−mz−∑j≥2bjzj,q≤∣m∣≤h,m,bj∈R,∣bj∣≤h,F(z)=1-mz-\sum_{j\ge2}b_jz^j,\qquad q\le|m|\le h,\qquad m,b_j\in\mathbb{R},\qquad|b_j|\le h,

then every zero of FF in ∣z∣≤q−1/2|z|\le q^{-1/2} is real.

Proof. Replacing F(z)F(z) by F(−z)F(-z) if necessary, take m≥qm\ge q. Notice that 6≤q≤h6\le q\le h, since 4<2h≤h−14<2\sqrt{h}\le h-1 for h≥6h\ge6, and m≥1+2hm\ge1+2\sqrt{h}. For 0<t<10<t<1,

F(t)≤U(t):=1−mt+ht21−t=1−(m+1)t+(m+h)t21−t.F(t)\le U(t):=1-mt+\frac{ht^2}{1-t}=\frac{1-(m+1)t+(m+h)t^2}{1-t}.

The numerator has smaller positive root

t0=2m+1+(m−1)2−4h≤11+h,t0<2m.(13)t_0=\frac{2}{m+1+\sqrt{(m-1)^2-4h}}\le\frac{1}{1+\sqrt{h}},\qquad t_0<\frac{2}{m}. \tag*{(13)}

Since F(0)=1F(0)=1 and F(t0)≤0F(t_0)\le0, there is a positive real zero α≤t0\alpha\le t_0. Suppose it also has a non-real zero λ\lambda with ∣λ∣2≤1/q|\lambda|^2 \le1/q. Its conjugate is another zero. Set ρ=1/2\rho= 1/\sqrt{2}; all three distinct zeros lie in ∣z∣<ρ|z| < \rho. Jensen’s formula, followed by the arithmetic–geometric mean inequality and Parseval’s identity, gives

ρ3α∣λ∣2≤12π∫02πlog⁡∣F(ρeit)∣ dt≤12log⁡(12π∫02π∣F(ρeit)∣2 dt)≤12log⁡(1+m22+h22).(14)\frac{\rho^3}{\alpha|\lambda|^2} \le\frac{1}{2\pi}\int_0^{2\pi}\log|F(\rho e^{it})|\,dt \le\frac{1}{2}\log\left(\frac{1}{2\pi}\int_0^{2\pi}|F(\rho e^{it})|^2\,dt\right) \le\frac{1}{2}\log\left(1+\frac{m^2}{2}+\frac{h^2}{2}\right). \tag*{(14)}

If a zero lies on ∣z∣=ρ|z| = \rho, obtain the same estimate by a limit through radii avoiding the zeros. On the other hand,

ρ3α∣λ∣2>mq42>1+m22+h22.\frac{\rho^3}{\alpha|\lambda|^2} > \frac{mq}{4\sqrt{2}} > \sqrt{1+\frac{m^2}{2}+\frac{h^2}{2}}.

For the last inequality, it suffices that m2(q2−16)>16h2+32m^2(q^2-16)>16h^2+32. Its left side is at least q2(q2−16)q^2(q^2-16). For h=6h=6, subtracting the right side gives 112. For h≥7h\ge7, using q≥2h+1q\ge2\sqrt{h}+1 and h>5/2\sqrt{h}>5/2 gives

q2(q2−16)−16h2−32≥(32h−56)h−40h−47>40h−187>0.q^2(q^2-16)-16h^2-32 \ge(32h-56)\sqrt{h}-40h-47 > 40h-187 > 0.

This contradicts (14) and proves the lemma. □

The middle expression in the last display is positive for every h≥6h\ge6; it equals 46.1…46.1\ldots at h=6h=6 and increases with hh. Hence the same proof applies with qq replaced by the real number 1+2h1+2\sqrt{h}, both in the bound on ∣m∣|m| and in the radius. This threshold cannot be lowered: for real mm slightly below 1+2h1+2\sqrt{h}, the series 1−mz+h∑j≥2zj1-mz+h\sum_{j\ge2}z^j has a pair of non-real zeros of modulus (m+h)−1/2<(1+2h)−1/2(m+h)^{-1/2}<(1+2\sqrt{h})^{-1/2}, which merge at m=1+2hm=1+2\sqrt{h} into the double zero (1+h)−1(1+\sqrt{h})^{-1}.

Proof of Theorem 1.2. Part (1) is Proposition 3.2. For part (2), write h=n−1≥6h=n-1\ge6 and q=⌈2h⌉+1≤hq=\lceil2\sqrt{h}\rceil+1\le h. If ∣c∣2<q|c|^2<q, then E(c,q)E(c,q) is connected by Lemma 2.1. It supplies an expansion of cc with all digits in DqD_q, hence an admissible first offset mm with ∣m∣≤q−1|m|\le q-1 for the nn-map system. By (12), k(c,n)≥n−q+1k(c,n)\ge n-q+1.

If ∣c∣2≥q|c|^2\ge q, take any m∈A(c,n)m\in A(c,n). Its expansion in (9), multiplied by c−1c^{-1}, makes λ=c−1\lambda=c^{-1} a zero of a series of the form in Lemma 4.1, with tail coefficients bounded by hh. As cc is non-real, that lemma excludes ∣m∣≥q|m|\ge q. Thus every admissible offset has ∣m∣≤q−1|m|\le q-1, and (12) again gives k(c,n)≥n−q+1k(c,n)\ge n-q+1.

The inequality q≤hq\le h gives kn≥2k_n\ge2. The increase 2h+1−2h2\sqrt{h+1}-2\sqrt{h} is less than 1, so the corresponding ceiling increases by at most 1. Hence knk_n is nondecreasing. Its asymptotic form follows directly from the ceiling formula. Attainment is proved in Proposition 6.3 below. □

Proof of the exclusion in Theorem 1.1. By the conjugation in § 1, we may assume that K=E(c,n)K=E(c,n) with cc non-real. For n≥7n\ge7 the two conclusions of Theorem 1.2 are incompatible with a dendrite. A connected self-similar set is locally connected [25], so a connected attractor that is not a dendrite contains a simple closed curve Γ\Gamma. If KK contains the bounded complementary domain of Γ\Gamma, it has interior points. Otherwise some point of that domain lies outside KK, and its component in C∖K\mathbb{C}\setminus K lies inside Γ\Gamma; hence KK separates the plane. □

Remark 4.2 (The role of Jensen’s formula). For n≥10n\ge10 the exclusion also follows from elementary estimates. Let cc be a non-real dendrite parameter and h=n−1h=n-1. By Proposition 3.2, A(c,n)A(c,n) is {h}\{h\} or {−h}\{-h\}; replacing cc by −c-c or cˉ\bar{c}, which leaves the attractor unchanged up to a reflection, we may assume that A(c,n)={h}A(c,n)=\{h\} and Im⁡c>0\operatorname{Im}c>0. Then (15) with k(c,n)=1k(c,n)=1 gives r2≥hr^2\ge h for r=∣c∣r=|c|, and Proposition 6.1 gives r<1+hr<1+\sqrt{h} and

h≤2Re⁡c+h(r−1)2≤2r+h(r−1)2.h\le2\operatorname{Re}c+\frac{h}{(r-1)^2}\le2r+\frac{h}{(r-1)^2}.

The convex function f(r)=2r+h(r−1)2−hf(r)=2r+\frac{h}{(r-1)^2}-h satisfies f(1+h)=−(h−3)(h+1)≤0f(1+\sqrt{h})=-(\sqrt{h}-3)(\sqrt{h}+1)\le0 when h≥9h\ge9. Also f(h)<0f(\sqrt{h})<0 for h≥9h\ge9: with s=hs=\sqrt{h}, the value f(h)=2s+s2(s−1)2−s2f(\sqrt{h})=2s+\frac{s^2}{(s-1)^2}-s^2 decreases for s>1s>1 and equals −3/4-3/4 at s=3s=3. Hence f<0f<0 on [h,1+h)[\sqrt{h},1+\sqrt{h}), a contradiction. Lemma 4.1 is therefore needed only for 7≤n≤97\le n\le9 and for the sharp count in Theorem 1.2.

Dependence on the parameter

The uniform estimate can be supplemented by elementary bounds that retain the modulus of cc or the alphabet size.

Proposition 5.1 (Radial bound). Let c∈Mn∖Rc\in\mathcal{M}_n\setminus\mathbb{R} and r=∣c∣r=|c|. Then

k(c,n)≥max⁡{1,n−⌊r2⌋}.(15)k(c,n)\ge\max\{1,n-\lfloor r^2\rfloor\}. \tag*{(15)}

For n≥7n\ge7 one may take the maximum of this bound and knk_n.

Proof. If ⌊r2⌋+1≤n\lfloor r^2\rfloor+1\le n, apply Lemma 2.1 to p=⌊r2⌋+1>r2p=\lfloor r^2\rfloor+1>r^2 and then Lemma 2.2. Otherwise the bound is just k≥1k\ge1. □

Proposition 5.2 (Increasing the alphabet). Fix c∈Mnc\in\mathcal{M}_n. Then

k(c,n+1)≥k(c,n)+1.(16)k(c,n+1)\ge k(c,n)+1. \tag*{(16)}

In particular, writing ν(c)=min⁡{p≥2:c∈Mp}\nu(c)=\min\{p\ge2:c\in\mathcal{M}_p\}, one has k(c,n)≥n−ν(c)+1k(c,n)\ge n-\nu(c)+1 for n≥ν(c)n\ge\nu(c). A non-real dendrite E(c,n)E(c,n) can occur only at the first connected alphabet size: n=ν(c)n=\nu(c).

Proof. The nested balanced digit alphabets give A(c,n)⊆A(c,n+1)\mathcal{A}(c,n)\subseteq\mathcal{A}(c,n+1). An offset mm contributes exactly the left indices

Im(n)={{m,…,n−1},m≥0,{0,…,n−1+m},m<0.I_m^{(n)}= \begin{cases} \{m,\ldots,n-1\}, & m\ge0,\\ \{0,\ldots,n-1+m\}, & m<0. \end{cases}

Their union over A(c,n)\mathcal{A}(c,n) is precisely the participating-index set. Every old interval is contained in its counterpart at alphabet size n+1n+1. If there is a nonnegative offset, the new index nn is added. If all offsets are negative, the union is an initial interval, and its endpoint increases by at least one. This proves (16). Iteration gives the bound involving ν(c)\nu(c); the dendrite assertion then follows from Proposition 3.2. □

Proposition 5.3 (Imaginary-axis count). For r>1r>1, E(ir,n)E(ir,n) is connected if and only if r2≤nr^2\le n. Throughout that range it is a nondegenerate rectangle and k(ir,n)=nk(ir,n)=n.

Proof. Splitting the digit expansion into even and odd positions gives

E(ir,n)=S−iS/r,S={∑j≥0tj(−r−2)j:tj∈An}.E(ir,n)=S-iS/r,\qquad S=\left\{\sum_{j\ge0}t_j(-r^{-2})^j:t_j\in\mathcal{A}_n\right\}.

The convex hull of SS is [−L,L][-L,L], with L=hr2/(r2−1)L=hr^2/(r^2-1). Its nn first-level intervals cover this hull precisely when L/r2≥1L/r^2\ge1, or r2≤nr^2\le n. In this case their invariant set is the whole interval. If r2>nr^2>n, those intervals are pairwise disjoint, so SS and hence E(ir,n)E(ir,n) are disconnected. In the connected case, putting η=hr/(r2−1)\eta=hr/(r^2-1) yields

K=[−rη,rη]+i[−η,η],Pi=[ai−η/r,ai+η/r]+i[−η,η].K=[-r\eta,r\eta]+i[-\eta,\eta],\qquad P_i=[a_i-\eta/r,a_i+\eta/r]+i[-\eta,\eta].

Since η≥r\eta\ge r, every PiP_i meets 2ir+K2ir+K. This includes equality r2=nr^2=n, where K∩(2ir+K)=[−rη,rη]+irK\cap(2ir+K)=[-r\eta,r\eta]+ir is a horizontal segment. □

Remark 5.4 (Branching and contact multiplicity). Put N=2n−1N=2n-1 and gt(z)=c(z−t)g_t(z)=c(z-t). For m∈Dnm\in D_n, (9) gives

m∈A(c,n)⟺g2m(2c)∈E(c,N).m\in\mathcal{A}(c,n)\quad\Longleftrightarrow\quad g_{2m}(2c)\in E(c,N).

Thus admissible offsets are precisely the first inverse branches of the marked point 2c2c that remain in the difference attractor. When A(c,n)≠∅\mathcal{A}(c,n)\ne\varnothing, also 2c∈E(c,N)2c\in E(c,N) by Lemma 2.1. Since E(c,N)E(c,N) lies in the canonical enclosure of [22 Section 6.1], each digit 2m2m with m∈A(c,n)m\in\mathcal{A}(c,n) then labels a child of the root 2c2c in the exact enclosure tree of [22 Definition 6.3]. Theorem E of that paper therefore gives #A(c,n)≤4\#\mathcal{A}(c,n) \le4 for non-real cc with ∣c∣≥n|c| \ge\sqrt{n}, and #A(c,n)≤3\#\mathcal{A}(c,n) \le3 if n=2n = 2 or n≥25n \ge25. For n≥10n \ge10 and n≥26n \ge26, respectively, these two bounds also follow from Proposition 6.1: when ∣c∣≥n|c| \ge\sqrt{n}, the admissible offsets lie in a closed interval of length at most 2+4/(n−1)2 + 4/(\sqrt{n} - 1). These bounds count offsets. By Lemma 2.2, one offset mm produces n−∣m∣n - |m| intersecting pairs of original pieces. For example, the even–odd decomposition in Proposition 5.3 gives A(in,n)={−1,0,1}\mathcal{A}(i\sqrt{n},n) = \{-1,0,1\}, while k(in,n)=nk(i\sqrt{n},n) = n and J(in,n)=3n−2J(i\sqrt{n},n) = 3n - 2. Bounded branching is therefore compatible with a contact count that grows with the alphabet size.

First-digit loci and convex parameter geometry

We now express the contact condition in parameter space. Fix n≥2n \ge2, write h=n−1h = n - 1 and N=2n−1N = 2n - 1, and let −h≤m≤h-h \le m \le h be an integer. Recall from (7) that E(c,N)=K−KE(c,N) = K - K for K=E(c,n)K = E(c,n). The first-digit locus Ln,m\mathcal{L}_{n,m} consists of parameters for which some overlap expansion (9) begins with mm. Replacing the remaining digit choices by their convex hull gives the convex locus L^n,m\widehat{\mathcal{L}}_{n,m}:

Ln,m={c:∣c∣>1, 2c∈2m+c−1E(c,N)}.(17)\mathcal{L}_{n,m} = \{c : |c| > 1,\ 2c \in2m + c^{-1}E(c,N)\}. \tag*{(17)}
L^n,m={c:∣c∣>1, 2c∈2m+c−1conv⁡E(c,N)}.(18)\widehat{\mathcal{L}}_{n,m} = \{c : |c| > 1,\ 2c \in2m + c^{-1}\operatorname{conv} E(c,N)\}. \tag*{(18)}

Equivalently, c∈Ln,mc \in\mathcal{L}_{n,m} if and only if

c=m+∑j≥1djc−j,dj∈[−h,h]∩Z.(19)c = m + \sum_{j \ge1} d_j c^{-j}, \qquad d_j \in[-h,h] \cap\mathbb{Z}. \tag*{(19)}

The convexification identity

conv⁡E(c,N)=∑j≥0[−2h,2h]c−j\operatorname{conv} E(c,N) = \sum_{j \ge0} [-2h,2h]c^{-j}

follows by taking convex hulls of finite digit sums and passing to the Hausdorff limit; see also [11 Chapter 1, p. 42]. Thus L^n,m\widehat{\mathcal{L}}_{n,m} is described by (19) with real tail digits dj∈[−h,h]d_j \in[-h,h], while the first digit mm remains an integer. By Lemma 2.1,

Mn=⋃m=−hhLn,m,A(c,n)={m:c∈Ln,m},Ln,m⊆L^n,m.(20)\mathcal{M}_n = \bigcup_{m=-h}^{h} \mathcal{L}_{n,m}, \qquad\mathcal{A}(c,n) = \{m : c \in\mathcal{L}_{n,m}\}, \qquad\mathcal{L}_{n,m} \subseteq\widehat{\mathcal{L}}_{n,m}. \tag*{(20)}

These sets generally overlap. A disjoint decomposition is obtained by recording the full set A(c,n)\mathcal{A}(c,n) of admissible labels: for each nonempty S⊆[−h,h]∩ZS \subseteq[-h,h] \cap\mathbb{Z}, take the parameters belonging to precisely the loci labeled by SS. Complex conjugation preserves each label, while c↦−cc \mapsto-c exchanges mm and −m-m. Hence there are nn representative labels up to symmetry in the whole plane; negative labels may still occur in the first quadrant. Figure 2 displays the nonnegative convex loci separately.

Support functions and localization

The convex relaxation has an explicit support description. Put

Z(c)=∑j≥1[−1,1]c−j,σc(ϕ)=∑j≥1∣Re⁡(e−iϕc−j)∣.Z(c) = \sum_{j \ge1} [-1,1]c^{-j}, \qquad\sigma_c(\phi) = \sum_{j \ge1} \left|\operatorname{Re}(e^{-i\phi}c^{-j})\right|.

The sums converge for ∣c∣>1|c| > 1, and σc\sigma_c is the support function of the centrally symmetric compact convex set Z(c)Z(c). Consequently

c∈L^n,m⟺∣Re⁡(e−iϕ(c−m))∣≤hσc(ϕ)  for every ϕ∈[0,π].(21)c \in\widehat{\mathcal{L}}_{n,m} \quad\Longleftrightarrow\quad\left|\operatorname{Re}(e^{-i\phi}(c-m))\right| \le h\sigma_c(\phi)\ \text{ for every }\phi\in[0,\pi]. \tag*{(21)}

At a fixed parameter, the eligible real first digits form an interval: they are the horizontal section of c−hZ(c)c - hZ(c) on the real axis. The convex-admissible integer labels therefore form a consecutive interval, possibly empty. No such interval assertion is needed for the exact labels.

For n≥3n \ge3, the radial enclosure for Mn∖R\mathcal{M}_n \setminus\mathbb{R}, and for the large set of expansions with real digits, is a theorem of Beaucou​p, Borwein, Boyd, and Pinner [9 Theorem 2], stated there for coefficient bounds g>1g > 1; it is applied to these loci in [23 Proposition 4]. An elementary proof by vertical support, valid for every n≥2n\ge2, is also given in [22 Proposition A.2]. That elementary estimate applies verbatim to the convex relaxation and also localizes each first-digit locus in a vertical strip.

Proposition 6.1 (Radial and strip bounds). If c=a+ib∈L^n,m∖Rc=a+ib\in\widehat{\mathcal{L}}_{n,m}\setminus\mathbb{R} and r=∣c∣r=|c|, then

r<1+h,∣2a−m∣≤h(r−1)2.(22)r<1+\sqrt{h},\qquad|2a-m|\le\frac{h}{(r-1)^2}. \tag*{(22)}

In particular, the non-real part of L^n,m\widehat{\mathcal{L}}_{n,m} with r2≥nr^2\ge n lies in the vertical strip

∣2a−m∣≤1+2n−1.|2a-m|\le1+\frac{2}{\sqrt{n}-1}.

Proof. Write c=reiθc=re^{i\theta} and use the real-tail expansion in (19). The estimate ∣sin⁡(jθ)∣≤j∣sin⁡θ∣|\sin(j\theta)|\le j|\sin\theta| gives

∣b∣≤h∑j≥1∣Im⁡(c−j)∣≤h∣b∣∑j≥1jr−j−1=h∣b∣(r−1)2.|b|\le h\sum_{j\ge1}|\operatorname{Im}(c^{-j})|\le h|b|\sum_{j\ge1}jr^{-j-1}=\frac{h|b|}{(r-1)^2}.

The second inequality is strict at a non-real parameter, because its j=2j=2 sine estimate is strict. Dividing by ∣b∣|b| proves the radial bound. Multiplying the expansion by cc gives c2−mc=d1+∑j≥1dj+1c−jc^2-mc=d_1+\sum_{j\ge1}d_{j+1}c^{-j}. Its imaginary part gives ∣2a−m∣≤h/(r−1)2|2a-m|\le h/(r-1)^2. Finally, h/(n−1)2=1+2/(n−1)h/(\sqrt{n}-1)^2=1+2/(\sqrt{n}-1).

For numerical rendering, truncate Z(c)Z(c) to ZJ(c)=∑j=1J[−1,1]c−jZ_J(c)=\sum_{j=1}^{J}[-1,1]c^{-j}. The geometric tail gives

ZJ(c)⊆Z(c)⊆ZJ(c)+B‾(0,∣c∣−J∣c∣−1).(23)Z_J(c)\subseteq Z(c)\subseteq Z_J(c)+\overline{B}\left(0,\frac{|c|^{-J}}{|c|-1}\right). \tag*{(23)}

Finite zonotopes and their enlarged supporting half-planes thus give inner and outer membership bounds for the convex relaxation. The figures use this construction on sampled parameter grids; the tail bound is explicit, but the floating-point renderings are not interval certificates. The source code and grid data are among the ancillary files described at the end of the paper.

First digits outside the connectedness disk

The disk estimate in Lemma 2.1 singles out the range ∣c∣≥n|c| \ge\sqrt{n} for the following classification. For each fixed first digit mm, both Ln,m\mathcal{L}_{n,m} and L^n,m\widehat{\mathcal{L}}_{n,m} increase with nn. The restriction below concerns the portions outside this growing disk and first digits whose magnitude increases with the alphabet size. Since c↦c‾c \mapsto\overline{c} preserves every label and c↦−cc \mapsto-c exchanges mm and −m-m, positive labels in the first quadrant are representative; the last assertion of the theorem records the symmetric form.

Theorem 6.2 (First digits outside the connectedness disk). Let h=n−1h=n-1. The positive integers mm for which Ln,m\mathcal{L}_{n,m} contains a parameter with Re⁡c>0\operatorname{Re} c>0, Im⁡c>0\operatorname{Im} c>0, and ∣c∣2≥n|c|^2\ge n are precisely

1≤m≤min⁡{h,⌈2h⌉}.(24)1 \le m \le\min\{h,\lceil2\sqrt{h}\rceil\}. \tag*{(24)}

The same classification holds for L^n,m\widehat{\mathcal{L}}_{n,m}. Moreover, if n≥7n\ge7 and c∈L^n,m∖Rc\in\widehat{\mathcal{L}}_{n,m}\setminus\mathbb{R} satisfies ∣c∣2≥⌈2h⌉+1|c|^2\ge\lceil2\sqrt{h}\rceil+1, in particular if ∣c∣2≥n|c|^2\ge n, then ∣m∣≤⌈2h⌉|m|\le\lceil2\sqrt{h}\rceil. For every mm in (24), the exact locus Ln,m\mathcal{L}_{n,m} contains the parameter

ch,m=m+1+i4h−(m−1)22,∣ch,m∣2=h+m.(25)c_{h,m}=\frac{m+1+i\sqrt{4h-(m-1)^2}}{2}, \qquad|c_{h,m}|^2=h+m. \tag*{(25)}

Consequently, for every integer mm with ∣m∣≤h|m|\le h, each of the loci Ln,m\mathcal{L}_{n,m} and L^n,m\widehat{\mathcal{L}}_{n,m} contains a non-real parameter with ∣c∣2≥n|c|^2\ge n if and only if ∣m∣≤min⁡{h,⌈2h⌉}|m|\le\min\{h,\lceil2\sqrt{h}\rceil\}.

Proof. For h≥6h\ge6, put q=⌈2h⌉+1≤hq=\lceil2\sqrt{h}\rceil+1\le h. If c∈L^n,mc\in\widehat{\mathcal{L}}_{n,m} with ∣m∣≥q|m|\ge q, its expansion makes c−1c^{-1} a zero of a series as in Lemma 4.1, with real tail coefficients. When ∣c∣2≥q|c|^2\ge q and cc is non-real, this contradicts that lemma. Hence ∣m∣≤q−1=⌈2h⌉|m|\le q-1=\lceil2\sqrt{h}\rceil for every non-real c∈L^n,mc\in\widehat{\mathcal{L}}_{n,m} with ∣c∣2≥q|c|^2\ge q; since q≤h<nq\le h<n, this includes the range ∣c∣2≥n|c|^2\ge n. For h≤5h\le5, all positive allowed labels satisfy m≤min⁡{h,⌈2h⌉}m\le\min\{h,\lceil2\sqrt{h}\rceil\}.

For every label in (24), one has m−1<2hm-1<2\sqrt{h}, so (25) is non-real. Its equation c2−(m+1)c+h+m=0c^2-(m+1)c+h+m=0 gives

c=m−hc−1=m+∑j≥1(−h)c−j.(26)c=m-\frac{h}{c-1}=m+\sum_{j\ge1}(-h)c^{-j}. \tag*{(26)}

These integer tail digits are admissible. Thus ch,m∈Ln,mc_{h,m}\in\mathcal{L}_{n,m}, and ∣ch,m∣2=h+m≥n|c_{h,m}|^2=h+m\ge n.

For the final assertion, the exclusion above gives ∣m∣≤⌈2h⌉|m|\le\lceil2\sqrt{h}\rceil whenever n≥7n\ge7, while min⁡{h,⌈2h⌉}=h\min\{h,\lceil2\sqrt{h}\rceil\}=h for n≤6n\le6. Conversely, ch,mc_{h,m} serves for m>0m>0; for m<0m<0, the map c↦−cc\mapsto-c carries ch,∣m∣c_{h,|m|} into Ln,m\mathcal{L}_{n,m}; and in∈Ln,0i\sqrt{n}\in\mathcal{L}_{n,0}, because the proof of Proposition 5.3 with r2=nr^2=n, so that η=r\eta=r, gives P0∩(2in+P0)=[a0−1,a0+1]+in≠∅P_0\cap(2i\sqrt{n}+P_0)=[a_0-1,a_0+1]+i\sqrt{n}\ne\varnothing, whence 0∈A(in,n)0\in\mathcal{A}(i\sqrt{n},n) by Lemma 2.2. Since Ln,m⊆L^n,m\mathcal{L}_{n,m}\subseteq\widehat{\mathcal{L}}_{n,m}, the same parameters serve for the convex loci. □

Proof of Corollary 1.3. Put h=n−1≥6h=n-1\ge6 and q=⌈2h⌉+1≤hq=\lceil2\sqrt{h}\rceil+1\le h. The point λ=1/c\lambda=1/c is a non-real zero of F(z)=1+a1z+⋯+akzkF(z)=1+a_1z+\cdots+a_kz^k, where ∣aj∣≤1/q|a_j|\le1/q. If ∣a1∣≥q|a_1|\ge q, by Lemma 4.1, applied with m=−a1m=-a_1 and bj=−ajb_j=-a_j, shows that every zero of FF in ∣z∣≤q−1/2|z|\le q^{-1/2} is real, a contradiction. Hence ∣a1∣≤⌈2h⌉|a_1|\le\lceil2\sqrt{h}\rceil.

For the constructions, let h≥1h\ge1 and let mm be an integer with 2≤m≤min⁡{h,⌈2h⌉}2\le m\le\min\{h,\lceil2\sqrt{h}\rceil\}. By (26), λ0=1/ch,m\lambda_0=1/c_{h,m} is a non-real zero of G(z)=1−mz+h∑j≥2zjG(z)=1-mz+h\sum_{j\ge2}z^j, which converges in the unit disk. The partial sums GkG_k of GG converge to GG locally uniformly there, so Hurwitz’s theorem gives zeros λk→λ0\lambda_k\to\lambda_0 of GkG_k. The polynomials zkGk(1/z)=zk−mzk−1+h(zk−2+⋯+1)z^kG_k(1/z)=z^k-mz^{k-1}+h(z^{k-2}+\cdots+1) are monic of height hh, with subleading coefficient −m-m, and have the roots 1/λk→ch,m1/\lambda_k\to c_{h,m}. Since ch,mc_{h,m} is non-real and ∣ch,m∣2=h+m>n|c_{h,m}|^2=h+m>n, the same holds for 1/λk1/\lambda_k when kk is large. Taking m=⌈2h⌉m = \lceil2\sqrt{h}\rceil for n≥7n \ge7 proves sharpness, and taking m=hm = h for 3≤n≤63 \le n \le6 gives non-real roots with ∣c∣2>n|c|^2 > n and ∣a1∣=n−1|a_1| = n - 1. □

The extreme locus Ln,n−1\mathcal{L}_{n,n-1} therefore meets the non-real range ∣c∣≥n|c| \ge\sqrt{n} exactly when 2≤n≤62 \le n \le6. For n≥7n \ge7, the number of positive first digits whose convex loci have no non-real parameter in that range is

n−1−⌈2n−1⌉.n - 1 - \lceil2\sqrt{n-1}\rceil.

This number is nondecreasing; it equals 1, 4, and 11 at n=7n = 7, 12, and 21, respectively. Figure 3 shows the transition and the subsequent loss of outer first-digit loci. For 3≤n≤63 \le n \le6, the extreme-label parameter in (25) is exactly the quadratic dendrite example in (6).

Proposition 6.3 (Attainment of the contact bound). For h≥4h \ge4, put p=⌈2h⌉p=\lceil2\sqrt{h}\rceil and n=h+1n=h+1. At the parameter ch=ch,pc_h=c_{h,p} in (25),

k(ch,n)=n−p.(27)k(c_h,n)=n-p. \tag*{(27)}

In particular, for every n≥7n \ge7,

min⁡c∈Mn∖Rk(c,n)=n−⌈2n−1⌉.\min_{c\in\mathcal{M}_n\setminus\mathbb{R}} k(c,n)=n-\lceil2\sqrt{n-1}\rceil.

Proof. Since p≤hp \le h, (26) shows that pp is an admissible label. Also 2Re⁡ch=p+12\operatorname{Re} c_h = p+1 and ∣ch∣2=h+p|c_h|^2=h+p. The strict estimate

h(h+p−1)2<2\frac{h}{\left(\sqrt{h+p}-1\right)^2}<2

is equivalent to h/2+p−1−2h>0h/2+p-1-\sqrt{2h}>0, which follows from p≥2hp \ge2\sqrt{h} and h≥4h \ge4. By Proposition 6.1, every admissible integer label mm therefore satisfies m>p−1m>p-1, hence m≥pm\ge p. For a positive label mm, the participating left indices are exactly {m,…,n−1}\{m,\ldots,n-1\} by (10). These intervals are nested, and pp itself occurs. Their union has n−pn-p indices. Combining this equality with Theorem 1.2 proves the minimum.

Figure 4 illustrates Proposition 6.3 for h=6h=6, at c6=3+i2c_6=3+i\sqrt{2}, the white diamond in Figure 3(d).

Confinement and rescaled geometry

All first-digit loci retain a common part near the unit disk. For c=irc=ir with 1<r2≤n1<r^2\le n, splitting the balanced digits into even and odd positions as in Proposition 5.3 gives

A(ir,n)={m∈Z:∣m∣≤min⁡{h,⌊hr2−1⌋}}.\mathcal{A}(ir,n)=\left\{m\in\mathbb{Z}: |m|\le\min\left\{h,\left\lfloor\frac{h}{r^2-1}\right\rfloor\right\}\right\}.

Indeed, since r2≤n≤2n−1r^2\le n\le2n-1, the digits in even and in odd positions each fill an interval, so the tails ∑j≥1dj(ir)−j\sum_{j\ge1}d_j(ir)^{-j} fill the rectangle hZ(ir)hZ(ir) displayed below, and m∈A(ir,n)m\in\mathcal{A}(ir,n) exactly when ir−mir-m lies in it. In particular, every label is exactly admissible when 1<r2≤21<r^2\le2. For 1<r2<21<r^2<2, the point ir−mir-m lies strictly inside the tail rectangle

hZ(ir)=[−hr2−1,hr2−1]+i[−hrr2−1,hrr2−1].hZ(ir)=\left[-\frac{h}{r^2-1},\frac{h}{r^2-1}\right]+i\left[-\frac{hr}{r^2-1},\frac{hr}{r^2-1}\right].

for every label mm. Hausdorff continuity of c↦Z(c)c\mapsto Z(c), which follows from (23), then gives a common open intersection of all convex loci, extending into the first quadrant.

First digits comparable to the alphabet size satisfy a stronger radial bound.

Proposition 6.4 (Confinement of large first digits). Let h=n−1≥6h=n-1\ge6, and let mm be an integer with 1+2h≤m≤h1+2\sqrt{h}\le m\le h. Set

Rh(m)=21+2h2(2/3)3/2(m+1+(m−1)2−4h).(28)R_h(m)=\frac{2\sqrt{1+2h^2}}{(2/3)^{3/2}\left(m+1+\sqrt{(m-1)^2-4h}\right)}. \tag*{(28)}

Every non-real c∈L^n,mc\in\widehat{\mathcal{L}}_{n,m} satisfies ∣c∣2≤Rh(m)|c|^2\le R_h(m). For any sequence of integers mh∈[1,h]m_h\in[1,h] such that mh/h→τ∈(0,1]m_h/h\to\tau\in(0,1], the bound applies for all sufficiently large hh, and

Rh(mh)⟶332τ.(29)R_h(m_h)\longrightarrow\frac{3\sqrt{3}}{2\tau}. \tag*{(29)}

Proof. The majorant in (13) provides a real zero α≤t0\alpha\le t_0. If ∣c∣2>3/2|c|^2>3/2, the zeros α\alpha, c−1c^{-1}, and c‾−1\overline{c}^{-1} lie in ∣z∣<ρ=2/3|z|<\rho=\sqrt{2/3}, and Jensen and Parseval give ρ3∣c∣2/α≤(1+23m2+43h2)1/2≤1+2h2\rho^3|c|^2/\alpha\le(1+\frac{2}{3}m^2+\frac{4}{3}h^2)^{1/2}\le\sqrt{1+2h^2}, yielding the bound. The remaining case is automatic because t0>1/m≥1/ht_0>1/m\ge1/h makes Rh(m)>33/2>3/2R_h(m)>3\sqrt{3}/2>3/2. If mh/h→τ>0m_h/h\to\tau>0, then mh≥1+2hm_h\ge1+2\sqrt{h} for all sufficiently large hh. Dividing the numerator and denominator in (28) by hh gives (29). □

Thus, when mh/h→τ>0m_h/h\to\tau>0, the non-real portions of the convex loci remain in a bounded disk while the connectedness disk has radius n\sqrt{n}. Their distances from the origin, after division by n\sqrt{n}, tend uniformly to zero. The constant in (29) is not optimal; a choice of ρ\rho depending on τ\tau improves it. For τ<0.37\tau<0.37, Proposition 6.1 alone gives the better asymptotic bound ∣c∣≤1+τ−1/2+o(1)|c|\le1+\tau^{-1/2}+o(1), because mh≤2∣c∣+h/(∣c∣−1)2m_h\le2|c|+h/(|c|-1)^2 and ∣c∣<1+h|c|<1+\sqrt{h}.

For the portions outside the connectedness disk, shown in Figure 5, set ζ=c/h\zeta=c/\sqrt{h}. If c∈L^n,m∖Rc\in\widehat{\mathcal{L}}_{n,m}\setminus\mathbb{R} and ∣c∣2≥h+1|c|^2\ge h+1, then Proposition 6.1 gives the explicit inequalities

Figure 5. Convex first-digit loci outside the connectedness disk for n=7n=7, 21, and 100, in the coordinate ζ=c/n−1\zeta=c/\sqrt{n-1}, shaded by mmax⁡/n−1m_{\max}/\sqrt{n-1} (as in Figure 3) on a common scale. The thin gray arc is the unit circle; the dashed arc ∣c∣=n|c|=\sqrt{n} and the dotted arc ∣ζ∣=1+1/n−1|\zeta|=1+1/\sqrt{n-1} bound the band of (30). The localization (31) explains the narrowing band and the relation between the first digit and the horizontal position.

1+1h≤∣ζ∣<1+1h,∣mh−2Re⁡ζ∣≤1+2/(h+1−1)h.(30)\sqrt{1+\frac{1}{h}}\le|\zeta|<1+\frac{1}{\sqrt{h}},\qquad\left|\frac{m}{\sqrt{h}}-2\operatorname{Re}\zeta\right|\le\frac{1+2/(\sqrt{h+1}-1)}{\sqrt{h}}. \tag*{(30)}

In particular, uniformly over these parameters and first digits,

∣ζ∣=1+O(h−1/2),mh=2Re⁡ζ+O(h−1/2).(31)|\zeta|=1+O(h^{-1/2}),\qquad\frac{m}{\sqrt{h}}=2\operatorname{Re}\zeta+O(h^{-1/2}). \tag*{(31)}

The rescaled non-real outer portions therefore concentrate near the unit circle, with their first digits ordered approximately by horizontal position. The estimates describe their location without determining the topology of the individual loci.

Examples at every alphabet size from two to six

We complete the proof of the sharp bound by exhibiting a dendrite at each of the five admissible alphabet sizes. Up to affine normalization and complex conjugation, the binary example is the one in [17 Figure 3]. The calculations below give a self-contained verification of all five contact structures, using integer arithmetic and one rational interval for the binary case.

Proposition 7.1 (Dendrites for 2≤n≤62 \le n \le6). For each 2≤n≤62 \le n \le6, there is a non-real parameter cc for which E(c,n)E(c,n) is a dendrite. One may take the upper-half-plane root of

{c3−2c2+2=0,n=2,c2−nc+2(n−1)=0,3≤n≤6.\begin{cases} c^{3} - 2c^{2} + 2 = 0, & n = 2, \\ c^{2} - nc + 2(n - 1) = 0, & 3 \le n \le6. \end{cases}

At each of these parameters,

E(c,n)∩(2c+E(c,n))={c},E(c,n) \cap(2c + E(c,n)) = \{c\},

and non-neighboring first-level pieces are disjoint. Thus the first-level intersection graph is a path, with neighboring pieces meeting at the points t+1t + 1, t∈{−n+1,−n+3,…,n−3}t \in\{-n + 1, -n + 3, \ldots, n - 3\}.

Proof. Difference expansions and inverse states. Write h=n−1h = n - 1, Dn={−h,…,h}D_n = \{-h,\ldots,h\}, and K=E(c,n)K = E(c,n). An intersection between pieces separated by 2j2j gives an expansion

jc=∑ℓ≥0eℓc−ℓ,eℓ∈Dn.jc = \sum_{\ell\ge0} e_{\ell}c^{-\ell}, \qquad e_{\ell} \in D_n.

Conversely, every such expansion can be lifted coordinatewise to a pair of addresses in AnA_n, and hence gives an intersection. Its inverse states satisfy

z0=jc,zℓ+1=c(zℓ−eℓ),∣zℓ∣≤R:=h1−∣c∣−1.z_0 = jc, \qquad z_{\ell+1} = c(z_{\ell} - e_{\ell}), \qquad|z_{\ell}| \le R := \frac{h}{1 - |c|^{-1}}.

We enumerate the states reachable from jcjc while satisfying a strict upper bound on the squared norm; these states play the role of the neighbor maps of Bandt and Graf [4]. In the five cases below, the reachable graphs are finite. Repeatedly deleting vertices with no outgoing edge identifies all infinite admissible paths.

The quadratic examples. Suppose 3≤n≤63 \le n \le6. Represent a state as z=a+bcz = a + bc, where a,b∈Za,b \in\mathbb{Z}. Since c+cˉ=nc + \bar{c} = n and ccˉ=2hc\bar{c} = 2h, its transition and squared norm are

(a,b)→e(−2hb,a+nb−e),Hn(a,b)=a2+nab+2hb2.(32)(a,b) \xrightarrow{e} (-2hb, a + nb - e), \qquad H_n(a,b) = a^{2} + nab + 2hb^{2}. \tag*{(32)}

For n=3,4,5,6n = 3,4,5,6, respectively, choose Bn=17,26,39,54B_n = 17,26,39,54. Then R2<BnR^{2} < B_n. Indeed, if wn=Bn(2h+1)−2h3w_n = B_n(2h + 1) - 2h^{3}, the required inequality follows from wn>0w_n > 0 and wn2>8hBn2w_n^{2} > 8hB_n^{2}, both direct integer inequalities. Thus every bounded inverse path has Hn(a,b)<BnH_n(a,b) < B_n. The following table lists every transition reachable from c=(0,1)c = (0,1) and from 2c=(0,2)2c = (0,2) that satisfies this norm bound. Entries e:(a,b)e : (a,b) specify the digit and successor. A dash means that no successor satisfies the bound. After repeatedly deleting vertices with no outgoing edge, the unique infinite path is

nnState (a,b)(a,b)Successors with Hn<BnH_n < B_n
3(0,1)(0,1)0:(−4,3), 1:(−4,2), 2:(−4,1)0:(-4,3),\ 1:(-4,2),\ 2:(-4,1)
(−4,1)(-4,1)−2:(−4,1), −1:(−4,0)-2:(-4,1),\ -1:(-4,0)
(−4,0)(-4,0)−2:(0,−2)-2:(0,-2)
(−4,2),(−4,3),(0,−2),(0,2)(-4,2),(-4,3),(0,-2),(0,2)—
4(0,1)(0,1)1:(−6,3), 2:(−6,2), 3:(−6,1)1:(-6,3),\ 2:(-6,2),\ 3:(-6,1)
(−6,1)(-6,1)−3:(−6,1)-3:(-6,1)
(−6,2),(−6,3),(0,2)(-6,2),(-6,3),(0,2)—
5(0,1)(0,1)1:(−8,4), 2:(−8,3), 3:(−8,2), 4:(−8,1)1:(-8,4),\ 2:(-8,3),\ 3:(-8,2),\ 4:(-8,1)
(−8,1)(-8,1)−4:(−8,1)-4:(-8,1)
(−8,2),(−8,3),(−8,4),(0,2)(-8,2),(-8,3),(-8,4),(0,2)—
6(0,1)(0,1)1:(−10,5), 2:(−10,4), 3:(−10,3), 4:(−10,2), 5:(−10,1)1:(-10,5),\ 2:(-10,4),\ 3:(-10,3),\ 4:(-10,2),\ 5:(-10,1)
(−10,1)(-10,1)−5:(−10,1)-5:(-10,1)
(−10,2)(-10,2)−5:(−20,7), −4:(−20,6), −3:(−20,5)-5:(-20,7),\ -4:(-20,6),\ -3:(-20,5)
(0,2)(0,2)5:(−20,7)5:(-20,7)
(−10,3),(−10,4),(−10,5)(-10,3),(-10,4),(-10,5)—
(−20,5),(−20,6),(−20,7)(-20,5),(-20,6),(-20,7)—

Table 1.

c→hc−2h→−hc−2h→−h⋯c \xrightarrow{h} c-2h \xrightarrow{-h} c-2h \xrightarrow{-h} \cdots

The state 2c2c has no successor satisfying the bound for n=3,4,5n=3,4,5; for n=6n=6, its only such successor is (−20,7)(-20,7), which has no successor satisfying the bound. For every j≥3j \ge3, Hn(0,j)=2hj2≥BnH_n(0,j)=2hj^2 \ge B_n. It follows that cc has the unique expansion

c=h−hc−1−hc−2−⋯c=h-hc^{-1}-hc^{-2}-\cdots

whereas jcjc has none for 2≤j≤n−12 \le j \le n-1.

The binary example. For n=2n=2, both critical values of P(x)=x3−2x2+2P(x)=x^3-2x^2+2 are positive, so PP has a unique real root β\beta. Direct rational evaluation gives −84/100<β<−83/100-84/100<\beta<-83/100. Its other roots are cc and cˉ\bar{c}, and ccˉ=−2/β>9/4c\bar{c}=-2/\beta>9/4. Hence ∣c∣>3/2|c|>3/2 and R<3R<3. Represent a state as a+bc+dc2a+bc+dc^2. Its transition is

(a,b,d)→e(−2d,a−e,b+2d).(a,b,d) \xrightarrow{e} (-2d,a-e,b+2d).

Its squared modulus is v0+v1β+v2β2v_0+v_1\beta+v_2\beta^2, where

v0=a2+2ab+4ad+2bd+4d2,v1=−ab−2b2−4bd−2d2,v2=−ad+b2+2bd.\begin{aligned} v_0 &= a^2+2ab+4ad+2bd+4d^2,\\ v_1 &= -ab-2b^2-4bd-2d^2,\\ v_2 &= -ad+b^2+2bd. \end{aligned}

Substitution of the rational interval for β\beta gives the following complete list of successors with squared modulus less than 9. Every excluded successor has squared modulus strictly greater than 9.

State (a,b,d)(a,b,d)Successors with squared modulus less than 9
(0,1,0)(0,1,0)0:(0,0,1), 1:(0,−1,1)0:(0,0,1),\ 1:(0,-1,1)
(0,0,1)(0,0,1)1:(−2,−1,2)1:(-2,-1,2)
(0,−1,1)(0,-1,1)−1:(−2,1,1), 0:(−2,0,1), 1:(−2,−1,1)-1:(-2,1,1),\ 0:(-2,0,1),\ 1:(-2,-1,1)
(−2,−1,1)(-2,-1,1)−1:(−2,−1,1)-1:(-2,-1,1)
(−2,0,1)(-2,0,1)−1:(−2,−1,2), 0:(−2,−2,2)-1:(-2,-1,2),\ 0:(-2,-2,2)
(−2,1,1),(−2,−1,2),(−2,−2,2)(-2,1,1),(-2,-1,2),(-2,-2,2)—

Table 2.

Thus the unique infinite path from cc has digits 1,1,−1,−1,…1,1,-1,-1,\ldots, giving

c=1+c−1−c−2−c−3−⋯ .c=1+c^{-1}-c^{-2}-c^{-3}-\cdots.

The displayed identity also follows directly from P(c)=0P(c)=0. Contact structure and dendricity. In all five cases the unique difference expansion uses only extreme digits ±h\pm h. Each such digit has a unique lift to a pair from AnA_n, namely (h,−h)(h,-h) or (−h,h)(-h,h). Therefore the neighboring contact is exactly {c}\{c\}. The absence of expansions of jcjc for j≥2j \ge2 proves that non-neighboring pieces are disjoint. Neighboring contact gives connectedness of KK by Lemma 2.1, and Hata’s theorem gives local connectedness [25]. The first-level intersection graph is a path. Every first-level junction is a cut point separating the pieces on its two sides. A simple closed curve cannot cross such a junction, since a circle minus one point remains connected. Consequently every simple closed curve in KK would lie in a single first-level piece. Repeating inside that piece traps the curve in nested cylinders of diameter tending to zero, a contradiction. Hence KK is a dendrite; compare the criterion of Bandt and Keller [6]. □

Further questions

For 2≤n≤62 \le n \le6, it remains to classify the dendrite parameters and their possible contact sets. With h=n−1h=n-1, Proposition 3.2 locates every non-real dendrite parameter in

(Ln,h∪Ln,−h)∖⋃∣m∣<hLn,m\left(\mathcal{L}_{n,h}\cup\mathcal{L}_{n,-h}\right)\setminus\bigcup_{\lvert m\rvert<h}\mathcal{L}_{n,m}

with the first-digit loci of § 6. The condition k(c,n)=1k(c,n)=1 counts participating pieces; it does not assert that the contact is a singleton. For n=2n=2, a singleton contact makes a connected attractor a dendrite, by the argument at the end of § 7 (compare [6]). The converse fails, so the equivalence stated in [17 Section 1] holds only in the forward direction. For real cc with 1<∣c∣<21<\lvert c\rvert<2, the attractor E(c,2)E(c,2) is an arc whose two pieces overlap in an interval. For a non-real example, take c=1/s≈0.9334−1.1325ic=1/s\approx0.9334-1.1325i, where s≈0.4334+0.5258is\approx0.4334+0.5258i is the root of z4+z3+z2−z+1z^4+z^3+z^2-z+1 in [14 Example 3.8]. The equation for ss is equivalent to f−1f1f−1f−1f−1=f1f−1f1f1f1f_{-1}f_1f_{-1}f_{-1}f_{-1}=f_1f_{-1}f_1f_1f_1, so the two first-level pieces share a fifth-level piece. Calegari and Walker note that this piece is their entire intersection and that the attractor is a dendrite. It would be useful to determine which non-singleton contacts occur at the other admissible alphabet sizes.

For n≥7n\ge7, the exact minimum is known, but its full set of attaining parameters remains to be described. The first-digit loci also raise questions about the topology of their incidence strata and the difference between exact membership and convex relaxation. The quadratic parameters and support estimates in § 6 give explicit starting points.

The proofs use both the common linear part and the equal spacing of the fixed points; equal spacing drives the replication of contacts in Lemma 2.2. It would be interesting to know whether collinear systems with a common non-real multiplier but unevenly spaced fixed points can produce dendrites with seven or more maps.

Acknowledgments

The author thanks David Juher and Joan Saldaña for discussions on collinear self-similar sets. The research program was supported by project PID2023-146424NB-I00, funded by MICIU/AEI/10.13039/501100011033 and ERDF/EU, and by the Universitat de Girona and Banco Santander Grant Programme for Researchers in Training (IFUdG 2022–2024).

Computational materials

The exact-arithmetic verifier for the tables in § 7 (which uses only the Python standard library), the figure programs, and the sampled parameter data are available as ancillary files with the arXiv version of this article. The Collinear Fractals explorer [21] provides interactive views of the attractors and connectedness loci:

https://complextrees.com/collinear-fractals-gpu/.

Use of artificial intelligence

ChatGPT (OpenAI) assisted in developing and checking the proofs in § 3, § 4, and § 6, preparing the verification and visualization code, and drafting and revising the manuscript. Claude (Anthropic) assisted in checking the proofs, computations, and references, in drafting and revising parts of the text, including Corollary 1.3 and Remark 4.2, and in preparing Figure 4. The author is responsible for the mathematical content.

References

DEPARTMENT D’INFORMÀTICA, MATEMÀTICA APLICADA I ESTADÍSTICA, UNIVERSITAT DE GIRONA, 17003 GIRONA, CATALONIA, SPAIN

Email address: [email protected]

References

  1. [1]S. Akiyama, B. Loridant, and J. M. Thuswaldner, Topology of planar self-affine tiles with collinear digit set, J. Fractal Geom. 8 (2021), 53–93. doi: 10.4171/JFG/98.DOI
  2. [2]K. Allabergenova, M. Samuel, and A. Tetenov, Intersections of the pieces of self-similar dendrites in the plane, Chaos Solitons Fractals 182 (2024), 114805. doi: 10.1016/j.chaos.2024.114805.DOI
  3. [3]C. Bandt, On the Mandelbrot set for pairs of linear maps, Nonlinearity 15 (2002), 1127–1147. doi: 10.1088/0951-7715/15/4/309.DOI
  4. [4]C. Bandt and S. Graf, Self-similar sets 7. A characterization of self-similar fractals with positive Hausdorff measure, Proc. Amer. Math. Soc. 114 (1992), 995–1001. doi: 10.1090/S0002-9939-1992-1100644-3.DOI
  5. [5]C. Bandt and N. V. Hung, Fractal n-gons and their Mandelbrot sets, Nonlinearity 21 (2008), 2653–2670. doi: 10.1088/0951-7715/21/11/009.DOI
  6. [6]C. Bandt and K. Keller, Self-similar sets 2. A simple approach to the topological structure of fractals, Math. Nachr. 154 (1991), 27–39. doi: 10.1002/mana.19911540104.DOI
  7. [7]C. Bandt and H. Rao, Topology and Hausdorff measure of self-similar fractals in the plane, Nonlinearity 20 (2007), 1463–1474. doi: 10.1088/0951-7715/20/6/008.DOI
  8. [8]M. F. Barnsley and A. N. Harrington, A Mandelbrot set for pairs of linear maps, Physica D 15 (1985), 421–432. doi: 10.1016/S0167-2789(85)80008-7.DOI
  9. [9]F. BeaucouP, P. Borwein, D. W. Boyd, and C. Pinner, Power series with restricted coefficients and a root on a given ray, Math. Comp. 67 (1998), 715–736. doi: 10.1090/S0025-5718-98-00960-0.DOI
  10. [10]F. BeaucouP, P. Borwein, D. W. Boyd, and C. Pinner, Multiple roots of [−1, 1] power series, J. London Math. Soc. (2) 57 (1998), 135–147. doi: 10.1112/S0024610798005857.DOI
  11. [11]T. Bousch, Sur quelques problèmes de dynamique holomorphe, doctoral thesis, Université Paris-Sud, Orsay, 1992. Author’s electronic edition.
  12. [12]T. Bousch, Connexité locale et par chemins höldériens pour les systèmes itérés de fonctions, unpublished manuscript, 1993; version of 5 January 1996. Author’s electronic edition.
  13. [13]D. Calegari, S. Koch, and A. Walker, Roots, Schottky semigroups, and a proof of Bandt’s conjecture, Ergodic Theory Dynam. Systems 37 (2017), 2487–2555. doi: 10.1017/etds.2016.17.DOI
  14. [14]D. Calegari and A. Walker, Laminations and external angles for similarity pairs, preprint, 2026. arXiv:2608.12774v1.arxiv.org/abs/2608.12774
  15. [15]J. J. Charatonik and W. J. Charatonik, Dendrites, Aportaciones Mat. Comun. 22 (1998), 227–253.
  16. [16]A. Douady and J. H. Hubbard, Étude dynamique des polynômes complexes. Partie I, Publications Mathématiques d’Orsay 84-2, Université Paris-Sud, Orsay, 1984; Partie II, Publications Mathématiques d’Orsay 85-4, 1985.
  17. [17]K. I. Eroğlu, S. Rohde, and B. Solomyak, Quasisymmetric conjugacy between quadratic dynamics and iterated function systems, Ergodic Theory Dynam. Systems 30 (2010), 1665–1684. doi: 10.1017/S0143385709000789.DOI
  18. [18]B. Espigulé Pons, Collinear fractals and connectedness loci: topology, finite capture, and restricted polynomial roots, doctoral thesis, Universitat de Girona, 2026.
  19. [19]B. Espigulé, Infinitely many holes in connectedness loci for collinear affine iterated function systems, preprint, 2026. arXiv:2606.01467v1.arxiv.org/abs/2606.01467
  20. [20]B. Espigulé, Regular closedness off the real axis for root closures of bounded-height integer polynomials, preprint, 2026. arXiv:2610.38971v1.arxiv.org/abs/2610.38971
  21. [21]B. Espigulé, Collinear Fractals GPU: companion software for collinear fractals, finite capture, and restricted polynomial roots, version 0.2.0-alpha, 2026 [software]. Software release.
  22. [22]B. Espigulé and D. Juher, Semialgebraic interior layers in root closures of bounded-height integer polynomials, preprint, 2026. arXiv:2603.07397v2.arxiv.org/abs/2603.07397
  23. [23]B. Espigulé, D. Juher, and J. Saldaña, Collinear fractals and Bandt’s conjecture, Fractal Fract. 8 (2024), 725. doi: 10.3390/fractalfract8120725.DOI
  24. [24]T. C. Flock and R. S. Strichartz, Laplacians on a family of quadratic Julia sets I, Trans. Amer. Math. Soc. 364 (2012), 3915–3965. doi: 10.1090/S0002-9947-2012-05398-0.DOI
  25. [25]M. Hata, On the structure of self-similar sets, Japan J. Appl. Math. 2 (1985), 381–414. doi: 10.1007/BF03167083.
  26. [26]H. Hopf, Über die Sehnen ebener Kontinuen und die Schleifen geschlossener Wege, Comment. Math. Helv. 9 (1936), 303–319. doi: 10.1007/BF01258195.DOI
  27. [27]J. E. Hutchinson, Fractals and self-similarity, Indiana Univ. Math. J. 30 (1981), 713–747. doi: 10.1512/iumj.1981.30.30055.DOI
  28. [28]A. Kameyama, Julia sets of postcritically finite rational maps and topological self-similar sets, Nonlinearity 13 (2000), 165–188. doi: 10.1088/0951-7715/13/1/308.DOI
  29. [29]J. Kigami, Harmonic calculus on limits of networks and its application to dendrites, J. Funct. Anal. 128 (1995), 48–86. doi: 10.1006/jfan.1995.1023.DOI
  30. [30]J. Kigami, Analysis on Fractals, Cambridge Tracts in Mathematics 143, Cambridge University Press, Cambridge, 2001. doi: 10.1017/CBO9780511470943.DOI
  31. [31]I. Kirat and K.-S. Lau, On the connectedness of self-affine tiles, J. London Math. Soc. (2) 62 (2000), 291–304. doi: 10.1112/S002461070000106X.DOI
  32. [32]K.-S. Leung and K.-S. Lau, Disjointness of planar self-affine tiles, Trans. Amer. Math. Soc. 359 (2007), 3337–3355. doi: 10.1090/S0002-9947-07-04106-2.DOI
  33. [33]J. Milnor, Dynamics in One Complex Variable, 3rd ed., Annals of Mathematics Studies 160, Princeton University Press, Princeton, NJ, 2006. doi: 10.1515/9781400835539.
  34. [34]S. B. Nadler, Jr., Continuum Theory: An Introduction, Monographs and Textbooks in Pure and Applied Mathematics 158, Marcel Dekker, New York, 1992.
  35. [35]A. Pascaletti and F. Zanolin, A crossing lemma for annular regions and invariant sets with an application to planar dynamical systems, J. Math. 2013 (2013), 267393. doi: 10.1155/2013/267393.
  36. [36]L. G. Rogers and A. Teplyaev, Laplacians on the basilica Julia set, Commun. Pure Appl. Anal. 9 (2010), 211–231. doi: 10.3934/cpaa.2010.9.211.DOI
  37. [37]P. Shmerkin and B. Solomyak, Zeros of {−1, 0, 1} power series and connectedness loci for self-affine sets, Experiment. Math. 15 (2006), 499–511. doi: 10.1080/10586458.2006.10128977.DOI
  38. [38]B. Solomyak, Measure and dimension for some fractal families, Math. Proc. Cambridge Philos. Soc. 124 (1998), 531–546. doi: 10.1017/S0305004198002680.DOI
  39. [39]B. Solomyak and H. Xu, On the “Mandelbrot set” for a pair of linear maps and complex Bernoulli convolutions, Nonlinearity 16 (2003), 1733–1749. doi: 10.1088/0951-7715/16/5/311.DOI

Paper details

vlnc:W_c6n4czec
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

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

  • Mathematics

ChatGPT (OpenAI), Claude (Anthropic)

View statement
Looking for related papers…
Contents