Introduction
For each positive integer h h h , the closure of the nonzero roots of integer polynomials with coefficients in [ − h , h ] [-h,h] [ − h , h ] is connected. What happens if we discard the real roots before taking the closure? Real points may still be limits of nonreal roots, but the real intervals that connect the full locus need not survive. We prove that the resulting set is connected exactly at the pronic heights 2 , 6 , 12 , 20 , … 2,6,12,20,\ldots 2 , 6 , 12 , 20 , … , the products of consecutive positive integers. At a set of heights of natural density one it has infinitely many connected components.
The first pronic height was visible on the real axis twenty years ago. For power series with constant term 1 and coefficients in { − 2 , … , 2 } \{-2,\ldots,2\} { − 2 , … , 2 } , Shmerkin and Solomyak found that the double zero 1 / 2 1/2 1/2 is isolated and observed numerically that the remaining double zeros in ( 0 , 1 ) (0,1) ( 0 , 1 ) appeared connected [28 ] , Proposition 6.3 and Remark 6.4. Proposition 3.2 confirms this: they form the interval [ 1 / ρ , 1 ) [1/\rho,1) [ 1/ ρ , 1 ) , where ρ = 1.839 … \rho=1.839\ldots ρ = 1.839 … is the tribonacci constant, the real root of ρ 3 = ρ 2 + ρ + 1 \rho^3=\rho^2+\rho+1 ρ 3 = ρ 2 + ρ + 1 , and this interval is exactly the set of points of ( 0 , 1 ) (0,1) ( 0 , 1 ) that are limits of nonreal roots of integer polynomials of height at most two. It is the real trace of a connected planar set, and Theorem A below shows that this connectedness occurs exactly at the pronic heights. At height one the same authors conjectured that the double zeros form a set with infinitely many components; Theorem D proves this.
The height parameters below are integers. For h ≥ 1 h\ge1 h ≥ 1 , put
P h = { ∑ j = 0 d a j z j : d ≥ 1 , a d ≠ 0 , a j ∈ Z , ∣ a j ∣ ≤ h } , R h = { z : p ( z ) = 0 for some p ∈ P h } , (1) \mathcal{P}_h=\left\{\sum_{j=0}^{d}a_jz^j:d\ge1,\ a_d\ne0,\ a_j\in\mathbb{Z},\ |a_j|\le h\right\},\qquad\mathcal{R}_h=\{z:p(z)=0\text{ for some }p\in\mathcal{P}_h\}, \tag*{(1)} P h = { j = 0 ∑ d a j z j : d ≥ 1 , a d = 0 , a j ∈ Z , ∣ a j ∣ ≤ h } , R h = { z : p ( z ) = 0 for some p ∈ P h } , (1)
and define
Z h = R h ‾ , C h = R h ∖ R ‾ . (2) \mathcal{Z}_h=\overline{\mathcal{R}_h},\qquad\mathcal{C}_h=\overline{\mathcal{R}_h\setminus\mathbb{R}}. \tag*{(2)} Z h = R h , C h = R h ∖ R . (2)
All closures are in C \mathbb{C} C . Here height refers to the coefficients of a defining polynomial, not necessarily its minimal polynomial. We call C h C_h C h the planar core .
For the full root closure the answer is known and uniform: the nonzero part Z h ∖ { 0 } \mathcal{Z}_h \setminus\{0\} Z h ∖ { 0 } is connected and locally connected at every height, by its interpretation as a connectedness locus for affine iterated function systems (§ 2 ). The planar core behaves differently. Our first result says that its connectedness is an arithmetic property of the height.
Theorem A (Connectedness). For every integer h ≥ 1 h \ge1 h ≥ 1 ,
C h is connected ⟺ h = q ( q − 1 ) for an integer q ≥ 2. C_h\text{ is connected}\quad\Longleftrightarrow\quad h=q(q-1)\quad\text{for an integer }q\ge2. C h is connected ⟺ h = q ( q − 1 ) for an integer q ≥ 2. At these heights C h C_h C h is also locally connected; at every other height it has at least five connected components. The full nonzero root closure Z h ∖ { 0 } \mathcal{Z}_h \setminus\{0\} Z h ∖ { 0 } is connected and locally connected at every height.
The connected heights 2 , 6 , 12 , 20 , 30 , 42 , … 2,6,12,20,30,42,\ldots 2 , 6 , 12 , 20 , 30 , 42 , … are the pronic heights . The cores are nested as h h h increases, but connectedness is not monotone. Real intervals account for the difference between the two sets in the theorem: they can join distinct components of the core. At a pronic height, the remaining real intervals are attached at single endpoints and contract onto the connected core. Figure 1 shows this change from height 19 to 20; Figure 2 places it in the sequence of pronic heights and their successors.
Figure 1. Reconnection at consecutive heights 19 and 20. The exterior parts of the cores (∣ z ∣ > 1 |z| > 1 ∣ z ∣ > 1 ) are shown in the first quadrant, with matched enlargements near their common real endpoint 5. The enlargements straddle the real axis and scale the axes independently. At height 19, infinitely many components accumulate at the singleton component { 5 } \{5\} { 5 } ; at height 20, the core is connected. These assertions follow from Theorems A and B. Shading illustrates finite-depth numerical tests and does not certify component counts.
Figure 2. Connectedness is not monotone in the height. The exterior first-quadrant views show the nested cores at h = q ( q − 1 ) h=q(q-1) h = q ( q − 1 ) (left) and h = q ( q − 1 ) + 1 h=q(q-1)+1 h = q ( q − 1 ) + 1 (right), for 2 ≤ q ≤ 8 2 \le q \le8 2 ≤ q ≤ 8 . The same colour identifies q q q in both panels; smaller-height sets are drawn over larger ones. Every core on the left is connected, whereas every core on the right is disconnected by Theorem A. The coloured regions and outlines are finite-depth numerical illustrations.
Our second result gives infinitely many components throughout both sides of each pronic height, apart from a short interval around the next square.
Theorem B (Infinitely many components). Let p ≥ 3 p \ge3 p ≥ 3 and h h h be integers such that
( p − 1 ) 2 + 1 ≤ h ≤ p 2 − 3 , h ≠ p ( p − 1 ) . (p-1)^2+1\le h\le p^2-3,\qquad h\ne p(p-1). ( p − 1 ) 2 + 1 ≤ h ≤ p 2 − 3 , h = p ( p − 1 ) . Then C h C_h C h has infinitely many connected components, including infinitely many with nonempty planar interior. Every neighbourhood of each of the four points
p , − p , p − 1 , − p − 1 p,\quad-p,\quad p^{-1},\quad-p^{-1} p , − p , p − 1 , − p − 1 contains infinitely many such components. If h < p ( p − 1 ) h<p(p-1) h < p ( p − 1 ) , these four points are singleton components. If h > p ( p − 1 ) h>p(p-1) h > p ( p − 1 ) , they belong to the component containing the unit circle. In both cases the core is not locally connected at these points.
Theorem B is most striking just below a pronic height. At height 19 the point 5 is a connected component of the core by itself, and infinitely many components with interior accumulate on it; at height 20 everything has joined into a single continuum (Figure 1 ). The same happens at ± p \pm p ± p and ± 1 / p \pm1/p ± 1/ p at every height strictly between ( p − 1 ) 2 (p-1)^2 ( p − 1 ) 2 and p ( p − 1 ) p(p-1) p ( p − 1 ) .
Near a square the rational anchors of Theorem B are no longer available. We call q 2 − 2 q^2-2 q 2 − 2 , q 2 − 1 q^2-1 q 2 − 1 , q 2 q^2 q 2 the square-adjacent heights . For at least two of the three we find the same accumulation point in the open interval ( 1 / ( q + 1 ) , 1 / q ) (1/(q+1),1/q) ( 1/ ( q + 1 ) , 1/ q ) between the two rational anchors.
Theorem C (Square-adjacent heights). Let q ≥ 4 q \ge4 q ≥ 4 . At least two of the three cores C q 2 − 2 C_{q^2-2} C q 2 − 2 , C q 2 − 1 C_{q^2-1} C q 2 − 1 , C q 2 C_{q^2} C q 2 have infinitely many connected components, including infinitely many with nonempty planar interior. They accumulate at a common point t ∗ ∈ ( 1 / ( q + 1 ) , 1 / q ) t_\ast\in(1/(q+1),1/q) t ∗ ∈ ( 1/ ( q + 1 ) , 1/ q ) and at its negative and reciprocals, where these cores are not locally connected. All three cores have this property for q ∈ { 4 , 6 , 12 } q\in\{4,6,12\} q ∈ { 4 , 6 , 12 } . At every square-adjacent height with 4 ≤ q ≤ 12 4\le q\le12 4 ≤ q ≤ 12 other than 25, 63 and 81, the core has infinitely many components with nonempty planar interior, accumulating at a point of ( 1 / ( q + 1 ) , 1 / q ) (1/(q+1),1/q) ( 1/ ( q + 1 ) , 1/ q ) ; the same holds at the heights 7 and 8, with q = 3 q=3 q = 3 , and at 194, 195, 223, 224, 225, 254 and 255. At every square-adjacent height h h h with q ≥ 4 q\ge4 q ≥ 4 , consider the series 1 − 2 q z + ∑ j ≥ 2 a j z j 1-2qz+\sum_{j\ge2}a_jz^j 1 − 2 q z + ∑ j ≥ 2 a j z j with integer coefficients h − 2 q + 2 ≤ a j ≤ h h-2q+2\le a_j\le h h − 2 q + 2 ≤ a j ≤ h and a double zero in ( 1 / ( q + 1 ) , 1 / q ) (1/(q+1),1/q) ( 1/ ( q + 1 ) , 1/ q ) . If one of them is not eventually constant with coefficient h − 2 q + 2 h-2q+2 h − 2 q + 2 , h − q + 1 h-q+1 h − q + 1 or h h h , then C h C_h C h has infinitely many components; this is the case unless one of them is eventually constant with coefficient h − q + 1 h-q+1 h − q + 1 .
Corollary 1.1 (Density one). The set of heights at which C h C_h C h has infinitely many connected components has natural density one. Among nonpronic heights h ≤ X h\le X h ≤ X , at most X + 1 \sqrt{X}+1 X + 1 can fail to have this property. Possible exceptions lie in { 1 , 3 , 4 , 9 , 25 , 63 , 81 } \{1,3,4,9,25,63,81\} { 1 , 3 , 4 , 9 , 25 , 63 , 81 } and in the triples { q 2 − 2 , q 2 − 1 , q 2 } \{q^2-2,q^2-1,q^2\} { q 2 − 2 , q 2 − 1 , q 2 } with q ≥ 13 q\ge13 q ≥ 13 , with at most one exception in each triple.
Proof. The nonpronic heights not covered by Theorem B are exactly
{ 1 } ∪ { q 2 − 1 , q 2 : q ≥ 2 } ∪ { q 2 − 2 : q ≥ 3 } . (3) \{1\} \cup\{q^2 - 1, q^2 : q \ge2\} \cup\{q^2 - 2 : q \ge3\}. \tag*{(3)} { 1 } ∪ { q 2 − 1 , q 2 : q ≥ 2 } ∪ { q 2 − 2 : q ≥ 3 } . (3)
Theorem C settles the heights 7 and 8, all but at most one height of each triple with q ≥ 4 q \ge4 q ≥ 4 , and every height of the triples with 4 ≤ q ≤ 12 4 \le q \le12 4 ≤ q ≤ 12 other than 25, 63 and 81. This leaves the seven listed heights and at most one height in each triple with q ≥ 13 q \ge13 q ≥ 13 . Arrange these possible exceptions in a sequence: the seven listed heights, then one height from each triple with q = 13 , 14 , … q = 13, 14, \ldots q = 13 , 14 , … . Its j j j -th term is at least ( j − 1 ) 2 (j - 1)^2 ( j − 1 ) 2 , by inspection for j ≤ 7 j \le7 j ≤ 7 and because q = j + 5 q = j + 5 q = j + 5 for j ≥ 8 j \ge8 j ≥ 8 . Hence at most X + 1 \sqrt{X} + 1 X + 1 exceptions are at most X X X . The pronic heights themselves contribute only O ( X ) O(\sqrt{X}) O ( X ) , so the complement has natural density zero. □
The count bounds actual exceptions, not the heights still undecided individually: at most one of 167, 168, 169 can be an exception, but the argument does not decide which. Figure 3 shows what the theorems prove up to height 49; § 8 discusses the remaining heights and reduces the square-adjacent ones to a question about double zeros of explicit integer polynomials.
Figure 3. What Theorems A–D prove for h ≤ 49 h \le49 h ≤ 49 . Row p p p lists the heights ( p − 1 ) 2 < h ≤ p 2 (p-1)^2 < h \le p^2 ( p − 1 ) 2 < h ≤ p 2 , shifted so that the pronic heights p ( p − 1 ) p(p-1) p ( p − 1 ) form one column. For h < p ( p − 1 ) h < p(p-1) h < p ( p − 1 ) the four points ± p \pm p ± p , ± 1 / p \pm1/p ± 1/ p are singleton components of C h C_h C h ; for h > p ( p − 1 ) h > p(p-1) h > p ( p − 1 ) they lie in the component containing the unit circle. The green cells are settled by the explicit witnesses of Propositions 6.14 , 6.15 and 6.16 . In the white cells, h = 1 , 3 , 4 , 9 , 25 h=1, 3, 4, 9, 25 h = 1 , 3 , 4 , 9 , 25 , infinitude is open (Conjecture 8.1); a dot marks the heights at which the real double-zero set D h D_h D h has infinitely many components (Theorem D), which does not decide the planar question.
The arithmetic mechanism
Taking reciprocals relates the root closure to the restricted series
F h = { 1 + ∑ j ≥ 1 a j z j : a j ∈ [ − h , h ] ∩ Z } . \mathcal{F}_h=\left\{1+\sum_{j\ge1}a_jz^j:a_j\in[-h,h]\cap\mathbb{Z}\right\}. F h = { 1 + j ≥ 1 ∑ a j z j : a j ∈ [ − h , h ] ∩ Z } .
Let D h ⊂ ( 0 , 1 ) \mathcal{D}_h\subset(0,1) D h ⊂ ( 0 , 1 ) be their real multiple-zero set, and let A h = C h ∩ ( 0 , 1 ) \mathcal{A}_h=\mathcal{C}_h\cap(0,1) A h = C h ∩ ( 0 , 1 ) be the active real set. A point is active when nonreal roots approach it. A real collision of conjugate roots gives A h ⊆ D h \mathcal{A}_h\subseteq\mathcal{D}_h A h ⊆ D h , but the inclusion can be strict.
For q ≥ 3 q\ge3 q ≥ 3 , the extremal series
F q ( z ) = ( 1 − ( q + 1 ) z ) 2 1 − z = 1 − ( 2 q + 1 ) z + q 2 ∑ j ≥ 2 z j F_q(z)=\frac{(1-(q+1)z)^2}{1-z}=1-(2q+1)z+q^2\sum_{j\ge2}z^j F q ( z ) = 1 − z ( 1 − ( q + 1 ) z ) 2 = 1 − ( 2 q + 1 ) z + q 2 j ≥ 2 ∑ z j
explains three distinct events. The point 1 / ( q + 1 ) 1/(q+1) 1/ ( q + 1 ) first becomes a double zero at height q 2 q^2 q 2 , when its positive tail is saturated and the zero is inactive. At height q 2 + 1 q^2+1 q 2 + 1 , perturbations F q + z N F_q+z^N F q + z N produce nearby nonreal zeros, so the point becomes active. The core reconnects at height q ( q + 1 ) q(q+1) q ( q + 1 ) , when integer digits allow no real double zero closer to 0 (Lemma 3.1 ) and every point of [ 1 / ( q + 1 ) , 1 ) [1/(q+1),1) [ 1/ ( q + 1 ) , 1 ) is active (Proposition 3.3 ); at the heights from q 2 + 1 q^2+1 q 2 + 1 to q ( q + 1 ) − 1 q(q+1)-1 q ( q + 1 ) − 1 , the point 1 / ( q + 1 ) 1/(q+1) 1/ ( q + 1 ) is a singleton component (Theorem B). Throughout these heights, q + 1 q+1 q + 1 remains the positive outer endpoint of the core.
Perturbing a double zero can produce nearby nonreal roots without producing distinct components. Admissible series may share a zero, and connections can leave the local disk. We fix a reference series with a real double zero, the anchor , and compare every representation in a finite coefficient block with every admissible representation outside it. The first departure from the reference has the form z N C ( z ) z^NC(z) z N C ( z ) . Shared conjugate pairs approaching the anchor would force a coefficient limit of normalized differences to have both value and derivative zero there, contradicting the comparison inequalities. Compactness and a quadratic root bound then make the local root sets clopen in C h \mathcal{C}_h C h (Lemmas 4.2 and 4.3 ). Each such set contains whole global components.
What this argument needs from a single series is local: the comparison inequalities at one double zero, and infinitely many levels at which one coefficient can be lowered and the later ones raised to compensate (Proposition 4.4 ). For the series ( 1 − p z ) 2 / ( 1 − z ) (1-pz)^2/(1-z) ( 1 − p z ) 2 / ( 1 − z ) this compensation is possible exactly when h > p ( p − 1 ) h>p(p-1) h > p ( p − 1 ) , which is how the pronic heights enter above the anchor 1 / p 1/p 1/ p ; the components then accumulate at a point of the component containing the unit circle. At height p ( p − 1 ) − 1 p(p-1)-1 p ( p − 1 ) − 1 , we use instead reference series with double zeros converging to 1 / p 1/p 1/ p . Two coefficient markers in each reference isolate a finite block whose root set contains 1 / p 1/p 1/ p , and the comparison is uniform over this convergent family. The local root sets are clopen and their diameters tend to zero, so { 1 / p } \{1/p\} { 1/ p } is a component.
The comparison also applies when most coefficient choices are removed ((Corollary 4.8)). For polynomials with constant coefficient 1 and all other coefficients in { − 5 , 4 , 5 } \{-5,4,5\} { − 5 , 4 , 5 } , the nonreal-root closure has a nonisolated singleton component at 1 / 3 1/3 1/3 ; with the alphabet { − 7 , 8 , 9 , 13 } \{-7,8,9,13\} { − 7 , 8 , 9 , 13 } , infinitely many components accumulate at 1 / 4 1/4 1/4 . The separating sets were constructed against every representation in the full alphabet, so restricting the alphabet preserves separation, while the explicit perturbations preserve the accumulating roots.
At the three square-adjacent heights next to q 2 q^2 q 2 , neither rational anchor is available: at 1 / q 1/q 1/ q the first comparison inequality fails for the digit range that the construction needs, while 1 / ( q + 1 ) 1/(q+1) 1/ ( q + 1 ) is not a double zero at heights q 2 − 2 q^2-2 q 2 − 2 and q 2 − 1 q^2-1 q 2 − 1 ((Lemma 2.6)) and is the saturated, inactive double zero of F q F_q F q at height q 2 q^2 q 2 ((Proposition 3.4)). We work instead with the series 1 − 2 q z + ∑ j ≥ 2 a j z j 1-2qz+\sum_{j\ge2}a_jz^j 1 − 2 q z + ∑ j ≥ 2 a j z j with digits m ≤ a j ≤ h m\le a_j\le h m ≤ a j ≤ h , m = h − 2 q + 2 m=h-2q+2 m = h − 2 q + 2 , whose double zeros t = 1 / β t=1/\beta t = 1/ β lie in the anchor interval where 2 ( β − q ) ( β − 1 ) 2 ∈ [ m , h ] 2(\beta-q)(\beta-1)^2\in[m,h] 2 ( β − q ) ( β − 1 ) 2 ∈ [ m , h ] , inside ( 1 / ( q + 1 ) , 1 / q ) (1/(q+1),1/q) ( 1/ ( q + 1 ) , 1/ q ) . At width 2 q − 2 2q-2 2 q − 2 , the comparison inequalities hold throughout this interval. The tail sums S N = ∑ i ≥ 1 ( a N + i − m ) t i S_N=\sum_{i\ge1}(a_{N+i}-m)t^i S N = ∑ i ≥ 1 ( a N + i − m ) t i measure the room for changing the coefficient a N a_N a N while keeping a double zero nearby. When this room exists at infinitely many levels N N N , one can change a digit arbitrarily far out and separate components near t t t . A witness without this property is eventually constant with one of three digits, m m m , m + q − 1 m+q-1 m + q − 1 or h h h , and a witness common to all three heights can be exceptional in this way at no more than one of them. This gives the two-of-three conclusion. Of the three exceptional types, only the middle one can make the set of anchors finite, which turns the remaining cases into a question about double zeros of integer polynomials ((Proposition 8.2)). When 1 / t 1/t 1/ t is a Pisot number, a finite graph decides existence and liveness, and every witness is eventually periodic ((Proposition 6.13)); Table 1 and Proposition 6.15 give explicit examples, all of the form g ( z ) 2 / ( 1 − z r ) g(z)^2/(1-z^r) g ( z ) 2 / ( 1 − z r ) with g ∈ Z [ z ] g\in\mathbb{Z}[z] g ∈ Z [ z ] . At the heights 7 and 8 the comparison inequalities fail on part of the anchor interval, but they hold at the double zeros of two such series, to which the local criterion applies ((Proposition 6.16)).
Table 1. Explicit eventually periodic square-adjacent witnesses. The digits a 2 , … , a N − 1 a_2,\ldots,a_{N-1} a 2 , … , a N − 1 are followed by the period c 0 , … , c r − 1 c_0,\ldots,c_{r-1} c 0 , … , c r − 1 repeated indefinitely. The reciprocal anchor β \beta β is the largest real root of the stated polynomial.
Real double zeros
By Lemma 2.7, D h ∖ A h D_h\setminus A_h D h ∖ A h consists of isolated points of D h D_h D h . Shmerkin and Solomyak proved that D 1 D_1 D 1 is disconnected [28 Theorem 2.6] and conjectured that it has infinitely many components [28 Section 1] ; they point out that such results bear on infinite Bernoulli convolutions, where double zeros are the obstruction to the transversality method of [30 , 31 ] . At four small heights the real multiple-zero set breaks into infinitely many pieces. Let Δ h \Delta_h Δ h be the set of λ ∈ ( − 1 , 1 ) \lambda\in(-1,1) λ ∈ ( − 1 , 1 ) such that ( λ , λ ) (\lambda,\lambda) ( λ , λ ) is a limit of pairs ( γ , λ ′ ) ∈ ( − 1 , 1 ) 2 (\gamma,\lambda')\in(-1,1)^2 ( γ , λ ′ ) ∈ ( − 1 , 1 ) 2 , γ ≠ λ ′ \gamma\ne\lambda' γ = λ ′ , of real zeros of a common series in F h \mathcal{F}_h F h . Coefficient compactness and Rolle’s theorem give Δ h ∩ ( 0 , 1 ) ⊆ D h \Delta_h\cap(0,1)\subseteq D_h Δ h ∩ ( 0 , 1 ) ⊆ D h .
Theorem D (Real traces near the minimum). Let h ∈ { 1 , 3 , 4 , 7 } h\in\{1,3,4,7\} h ∈ { 1 , 3 , 4 , 7 } . Then D h D_h D h , its complement ( 0 , 1 ) ∖ D h (0,1)\setminus D_h ( 0 , 1 ) ∖ D h and Δ h \Delta_h Δ h have infinitely many connected components. The gaps of D h D_h D h that separate them accumulate at a point of A h A_h A h within 10 − 87 10^{-87} 1 0 − 87 of min D h \min D_h min D h .
Corollary 1.2 (Solomyak’s diagonal trace). Let N = { ( γ , λ ) ∈ ( − 1 , 1 ) 2 : f ( γ ) = f ( λ ) = 0 for some f ∈ F 1 } \mathcal{N}=\{(\gamma,\lambda)\in(-1,1)^2:f(\gamma)=f(\lambda)=0\text{ for some }f\in\mathcal{F}_1\} N = {( γ , λ ) ∈ ( − 1 , 1 ) 2 : f ( γ ) = f ( λ ) = 0 for some f ∈ F 1 } be the connectedness locus for pairs of affine maps studied in [32 ] . Then N ∖ Diag ( R ) ∩ Diag ( R ) \mathcal{N}\setminus\operatorname{Diag}(\mathbb{R})\cap\operatorname{Diag}(\mathbb{R}) N ∖ Diag ( R ) ∩ Diag ( R ) has infinitely many connected components.
Proof. Inside ( − 1 , 1 ) 2 (-1,1)^2 ( − 1 , 1 ) 2 this set is { ( λ , λ ) : λ ∈ Δ 1 } \{(\lambda,\lambda):\lambda\in\Delta_1\} {( λ , λ ) : λ ∈ Δ 1 } . By Theorem D, infinitely many components of Δ 1 \Delta_1 Δ 1 are separated by gaps of D 1 D_1 D 1 near min D 1 \min D_1 min D 1 ; the closure can add only the points ± ( 1 , 1 ) \pm(1,1) ± ( 1 , 1 ) , which do not join them. □ \square □
For h = 1 h=1 h = 1 , Theorem D proves the conjecture of Shmerkin and Solomyak, and Corollary 1.2 proves the conjecture, recorded by Solomyak [32 Section 1] , that the diagonal trace has infinitely many components. The proof in § 7 is computer-assisted. Near the smallest double zero of series with a prescribed initial block, the other double zeros of such series form scaled slices of a self-affine set, one for each later position at which a corner digit can be lowered. At explicit digit patterns one of these slices lies alone in a short window, which forces a gap; after the last such pattern the remaining digits lie in a set of small dimension, and a finite branch and bound shows that the next block again contains a pattern. The finite statements are checked in bounded integer arithmetic (§ 17 of the supplement). Rigorous computation of this kind has a precedent for Bernoulli convolutions: Kleptsyn, Pollicott and Vytnova proved, with a computer-assisted algorithm, that their Hausdorff dimension is at least 0.96399 for every parameter in (1/2, 1) [20 ] . Theorem D concerns real traces. Real gaps do not separate planar components, so it does not decide Conjecture 8.1 at these heights, nor whether A h \mathcal{A}_h A h itself has infinitely many components.
Context
Restricted polynomial root sets and their power-series limits were studied by Odlyzko and Poonen [24 ] , who proved path-connectedness for the closure of the zeros of polynomials with coefficients 0 and 1 and constant term 1, by Solomyak [29 ] , and by Shmerkin and Solomyak [28 ] . For the alphabet { − 1 , 0 , 1 } \{-1, 0, 1\} { − 1 , 0 , 1 } , the nonzero full locus is related to the connectedness locus for pairs of similarities introduced by Barnsley and Harrington [2 ] ; Bousch proved it connected and locally connected [7 , 8 ] , and for the consecutive integer alphabets considered here these properties hold at every height [15 Theorem 1] , by a theorem of Nakajima [23 ] . Its interior and holes were studied by Bandt [1 ] , Solomyak and Xu [33 ] and Calegari, Koch and Walker [10 ] , who proved Bandt’s conjecture that interior points are dense off the real axis and that the locus has infinitely many holes; [11 ] extends the latter to the collinear loci of every arity, and interior regions and finite-capture methods for collinear systems were developed in [17 , 15 , 14 ] . A related planar set arises from the Galois conjugates of the growth rates of postcritically finite quadratic maps [34 ] ; its three-dimensional extension, Thurston’s Master Teapot, is connected [9 ] . How closely roots of restricted polynomials approach a given point was studied by Borwein and Pinner [6 ] , and real multiple zeros of restricted series in [3 , 4 , 26 , 28 , 32 ] ; Kewarth [18 ] classifies connectedness of the full constant-term-one locus for symmetric integer alphabets. All of these results concern the full locus, real intervals included. Whether those intervals are what joins distinct planar parts is the question answered here: deciding it requires the active real trace A h \mathcal{A}_h A h , which § 3 computes at the pronic heights and whose gaps § 4 shows to separate planar components elsewhere.
Organization of the proofs
§ 2 supplies the root-set dictionary and the analytic bounds used throughout, and proves that the core is a Peano continuum. The remaining dependencies are as follows.
Table 7.
The finite checks are replayed by the computational companion [13 ] . Neither the comparison argument of § 4 nor the sparse-alphabet application requires a finite certificate; density of planar interior is used only for the interior-bearing components in Theorems B and C. Several ingredients do not depend on the classification and may be useful elsewhere: Lemma 4.2 separates global components of a root closure for any coefficient alphabet, and Proposition 4.4 produces infinitely many of them from a single series; Lemma 2.7 and Proposition 2.5 describe the real multiple-zero set and the local structure of the core at every height; and Proposition 6.13 turns the search for double zeros with a Pisot reciprocal anchor into a finite computation. § 8 states the remaining question.
Root closures and double zeros
We first record the passage between polynomials, power series and the connected full root locus. Retain F h \mathcal{F}_h F h , D h \mathcal{D}_h D h , A h \mathcal{A}_h A h from the introduction. All closures are taken in C \mathbb{C} C unless a relative space is specified. Write D = { z : ∣ z ∣ < 1 } \mathbb{D}=\{z:|z|<1\} D = { z : ∣ z ∣ < 1 } , T = { z : ∣ z ∣ = 1 } \mathbb{T}=\{z:|z|=1\} T = { z : ∣ z ∣ = 1 } , X = { z : ∣ z ∣ > 1 } X=\{z:|z|>1\} X = { z : ∣ z ∣ > 1 } , and
B h = [ − h , h ] ∩ Z , Σ h = { c ∈ X : f ( 1 / c ) = 0 for some f ∈ F h } . B_h=[-h,h]\cap\mathbb{Z},\qquad\Sigma_h=\{c\in X:f(1/c)=0\text{ for some }f\in\mathcal{F}_h\}. B h = [ − h , h ] ∩ Z , Σ h = { c ∈ X : f ( 1/ c ) = 0 for some f ∈ F h } .
A series f ∈ F h f\in\mathcal{F}_h f ∈ F h is a witness at w ∈ D w\in\mathbb{D} w ∈ D if f ( w ) = 0 f(w)=0 f ( w ) = 0 . Coefficientwise convergence in F h \mathcal{F}_h F h is locally uniform in D \mathbb{D} D , together with every derivative. This compactness will be used throughout.
Normalization and the full locus
Proposition 2.1. Σ h \Sigma_h Σ h is the relative closure in X X X of the roots of monic integer polynomials of height at most h h h . Every nonreal zero in D \mathbb{D} D of a series in F h \mathcal{F}_h F h belongs to C h \mathcal{C}_h C h .
Proof. Reversal puts each exterior monic root in Σ h \Sigma_h Σ h , and coefficient compactness makes Σ h \Sigma_h Σ h relatively closed. Conversely, polynomial truncations of f ∈ F h f\in\mathcal{F}_h f ∈ F h converge locally uniformly to f f f . Rouché’s theorem on a small circle about a zero gives nearby polynomial zeros. Reversing the truncations gives monic polynomials with the same height bound. At a nonreal zero the circle can be chosen disjoint from R \mathbb{R} R , so the approximating zeros remain nonreal. □ \quad\Box □
We use two established properties of this normalized locus:
{ 1 < ∣ c ∣ < h + 1 } ⊂ int Σ h , Σ h is connected and locally connected . (4) \{1<|c|<\sqrt{h+1}\}\subset\operatorname{int}\Sigma_h,\qquad\Sigma_h\text{ is connected and locally connected}. \tag*{(4)} { 1 < ∣ c ∣ < h + 1 } ⊂ int Σ h , Σ h is connected and locally connected . (4)
For clarity, the dictionary is as follows. If n = h + 1 n=h+1 n = h + 1 and E ( c , n ) E(c,n) E ( c , n ) is the attractor of the maps w ↦ t + c − 1 w w\mapsto t+c^{-1}w w ↦ t + c − 1 w , with t ∈ { − n + 1 , − n + 3 , … , n − 1 } t\in\{-n+1,-n+3,\ldots,n-1\} t ∈ { − n + 1 , − n + 3 , … , n − 1 } , then its connectedness locus is Σ h \Sigma_h Σ h . Consecutive-piece intersection is equivalent to 2 c ∈ E ( c , n ) − E ( c , n ) 2c\in E(c,n)-E(c,n) 2 c ∈ E ( c , n ) − E ( c , n ) ; dividing its difference address by 2 c 2c 2 c gives the series equation. In this normalization, the dictionary, annulus and connectedness assertions are Proposition 2, Proposition 5 and Theorem 1 of [15 ] , respectively. We retain the polynomial height h h h as our parameter.
The following elementary estimate removes the possible nonunit leading coefficient. It will also normalize the witnesses used in Section 4.
Lemma 2.2 (Conjugate-pair bound). If f ( z ) = 1 + ∑ j = 1 d b j z j f(z)=1+\sum_{j=1}^{d}b_jz^j f ( z ) = 1 + ∑ j = 1 d b j z j has real coefficients ∣ b j ∣ ≤ A |b_j|\le A ∣ b j ∣ ≤ A and a nonreal zero λ \lambda λ , then
∣ λ ∣ − 4 ≤ 4 ( 1 + A 2 ) . |\lambda|^{-4}\le4(1+A^2). ∣ λ ∣ − 4 ≤ 4 ( 1 + A 2 ) . Proof. Put r = 1 / 2 r=1/\sqrt{2} r = 1/ 2 and μ = λ / r \mu=\lambda/r μ = λ / r . Factor the conjugate pair from f ( r z ) f(rz) f ( r z ) , then reverse this quadratic factor. Reversal preserves its modulus on T \mathbb{T} T , so Parseval preserves the sum of squared coefficients. The new constant coefficient is ∣ μ ∣ − 2 |\mu|^{-2} ∣ μ ∣ − 2 . Therefore
∣ μ ∣ − 4 ≤ 1 + ∑ j = 1 d b j 2 r 2 j ≤ 1 + A 2 . |\mu|^{-4}\le1+\sum_{j=1}^{d}b_j^2r^{2j}\le1+A^2. ∣ μ ∣ − 4 ≤ 1 + j = 1 ∑ d b j 2 r 2 j ≤ 1 + A 2 . Since ∣ μ ∣ − 4 = ∣ λ ∣ − 4 / 4 |\mu|^{-4}=|\lambda|^{-4}/4 ∣ μ ∣ − 4 = ∣ λ ∣ − 4 /4 , this is the desired estimate. □ \quad\Box □
Proposition 2.3. Monicity does not change Z h \mathcal{Z}_h Z h or C h \mathcal{C}_h C h . Writing ι ( z ) = 1 / z \iota(z)=1/z ι ( z ) = 1/ z and C h e x t = C h ∩ X \mathcal{C}_h^{\mathrm{ext}}=\mathcal{C}_h\cap X C h ext = C h ∩ X , one has
Z h ∖ { 0 } = Σ h ∪ T ∪ ι ( Σ h ) , (5) \mathcal{Z}_h\setminus\{0\}=\Sigma_h\cup\mathbb{T}\cup\iota(\Sigma_h), \tag*{(5)} Z h ∖ { 0 } = Σ h ∪ T ∪ ι ( Σ h ) , (5) C h = C h e x t ∪ T ∪ ι ( C h e x t ) . (6) \mathcal{C}_h=\mathcal{C}_h^{\mathrm{ext}}\cup\mathbb{T}\cup\iota(\mathcal{C}_h^{\mathrm{ext}}). \tag*{(6)} C h = C h ext ∪ T ∪ ι ( C h ext ) . (6) Both sets are invariant under conjugation, negation and inversion. Moreover,
Z h ∩ R = { 0 } ∪ [ − h − 1 , − ( h + 1 ) − 1 ] ∪ [ ( h + 1 ) − 1 , h + 1 ] , (7) \mathcal{Z}_h\cap\mathbb{R}=\{0\}\cup[-h-1,-(h+1)^{-1}]\cup[(h+1)^{-1},h+1], \tag*{(7)} Z h ∩ R = { 0 } ∪ [ − h − 1 , − ( h + 1 ) − 1 ] ∪ [( h + 1 ) − 1 , h + 1 ] , (7) { ( h + 1 ) − 1 / 2 < ∣ z ∣ < h + 1 } ⊂ int C h . (8) \{(h+1)^{-1/2}<|z|<\sqrt{h+1}\}\subset\operatorname{int}\mathcal{C}_h. \tag*{(8)} {( h + 1 ) − 1/2 < ∣ z ∣ < h + 1 } ⊂ int C h . (8) The nonzero full root closure is connected and locally connected.
Proof. For 1 < x ≤ h + 1 1 < x \le h + 1 1 < x ≤ h + 1 , the intervals ( a + J ) / x (a+J)/x ( a + J ) / x , a ∈ B h a \in B_h a ∈ B h , cover J = [ − h / ( x − 1 ) , h / ( x − 1 ) ] J = [-h/(x-1),h/(x-1)] J = [ − h / ( x − 1 ) , h / ( x − 1 )] . Expanding − 1 ∈ J -1 \in J − 1 ∈ J recursively gives a witness in F h F_h F h at 1 / x 1/x 1/ x . The reverse radius bound follows from 1 ≤ h / ( x − 1 ) 1 \le h/(x-1) 1 ≤ h / ( x − 1 ) . Thus Σ h ∩ R = [ − h − 1 , − 1 ) ∪ ( 1 , h + 1 ] \Sigma_h \cap\mathbb{R} = [-h-1,-1) \cup(1,h+1] Σ h ∩ R = [ − h − 1 , − 1 ) ∪ ( 1 , h + 1 ] .
Let c c c be an exterior root of a height-h h h polynomial whose leading coefficient has modulus k ≥ 2 k \ge2 k ≥ 2 . Reversal and division by that coefficient give a witness at 1 / c 1/c 1/ c with coefficient bound h / k h/k h / k . If c c c is nonreal, the preceding lemma gives
∣ c ∣ 4 ≤ 4 ( 1 + ( h / k ) 2 ) ≤ h 2 + 4 < ( h + 1 ) 2 . |c|^4 \le4(1+(h/k)^2) \le h^2+4 < (h+1)^2. ∣ c ∣ 4 ≤ 4 ( 1 + ( h / k ) 2 ) ≤ h 2 + 4 < ( h + 1 ) 2 .
so c ∈ Σ h c \in\Sigma_h c ∈ Σ h by (4) . If c c c is real, the elementary bound ∣ c ∣ ≤ 1 + h / k ≤ h + 1 |c| \le1+h/k \le h+1 ∣ c ∣ ≤ 1 + h / k ≤ h + 1 gives the same conclusion. The case k = 1 k=1 k = 1 is Proposition 2.1. Hence Z h ∩ X = Σ h Z_h \cap X = \Sigma_h Z h ∩ X = Σ h , including equality of the nonreal-root closures there.
Inside D \mathbb{D} D , the monic polynomials p ( z ) + z N p(z)+z^N p ( z ) + z N , N > deg p N>\deg p N > deg p , converge locally uniformly to any fixed p ∈ P h p \in\mathcal{P}_h p ∈ P h ; Rouché gives monic approximations to each of its roots, preserving nonreality. Roots of unity handle T \mathbb{T} T . This proves monic reduction. After removing a maximal power of z z z , polynomial reversal preserves height and inverts every nonzero root. The radius estimate applied in both directions gives ( h + 1 ) − 1 ≤ ∣ z ∣ ≤ h + 1 (h+1)^{-1} \le|z| \le h+1 ( h + 1 ) − 1 ≤ ∣ z ∣ ≤ h + 1 for all nonzero roots. Thus 0 0 0 is isolated, and inversion passes to both closures. This proves (5) –(7) .
The exterior annulus consists of interior points of Σ h \Sigma_h Σ h and hence of its nonreal-root closure. Inversion gives the inner annulus; their closures supply T \mathbb{T} T and a neighbourhood of it. Finally, (5) joins the connected, locally connected exterior locus to its reciprocal across this annular neighbourhood. □
Other restrictions on the defining polynomials are harmless as well: Hokken [16 ] proves that the closure of the zeros of polynomials with coefficients in a fixed integer set is unchanged when the polynomials are required to have square discriminant.
Proposition 2.4 (Density of interior). For every h ≥ 1 h \ge1 h ≥ 1 ,
C h = int Z h = int C h ‾ . C_h = \operatorname{int} Z_h = \operatorname{int} \overline{C_h}. C h = int Z h = int C h . Proof. The nonreal interior-density results of [19 Theorem 1.1] and [12 ] give the first identity off T \mathbb{T} T through the series dictionary and inversion; the annular neighbourhood handles T \mathbb{T} T . Every real interior point of Z h Z_h Z h is also approximated by nonreal points. Finally, int Z h ⊂ C h ⊂ Z h \operatorname{int} Z_h \subset C_h \subset Z_h int Z h ⊂ C h ⊂ Z h gives equality of the two interiors. □
This last input is used to strengthen the separated components in § 4 to components with planar interior. The connectedness classification and the existence of infinitely many components do not require this strengthening.
Although the core is not locally connected at most heights (Theorems B and C), its failure of local connectedness is of the simplest kind. Following Loridant, Luo and Yang [21 ] , a compact metric space is a Peano continuum if every component is locally connected and, for every ε > 0 \varepsilon> 0 ε > 0 , only finitely many components have diameter at least ε \varepsilon ε .
Local connectedness at a point means that the point has arbitrarily small connected neighbourhoods in the set; a compact metric space with this property at every point is locally connected [22 Chapter V] .
Proposition 2.5 (Peano structure of the core). For every h ≥ 1 h \ge1 h ≥ 1 , C h C_h C h is a Peano continuum. The core is not locally connected at a point x x x if and only if a sequence of distinct components of C h C_h C h accumulates at x x x , and every such x x x is real.
Proof. Let N N N be the boundary, relative to R \mathbb{R} R , of the closed set C h ∩ R C_h \cap\mathbb{R} C h ∩ R ; it is a closed nowhere dense subset of R \mathbb{R} R . We first show that C h C_h C h is locally connected at every point of C h ∖ N C_h \setminus N C h ∖ N . If x ∈ C h x \in C_h x ∈ C h is nonreal, choose a disk B B B about x x x disjoint from R \mathbb{R} R . A point of Z h ∩ B Z_h \cap B Z h ∩ B is a limit of roots lying in B B B , hence of nonreal roots, so Z h ∩ B = C h ∩ B Z_h \cap B = C_h \cap B Z h ∩ B = C h ∩ B . If x x x is real and not in N N N , it is an interior point of C h ∩ R C_h \cap\mathbb{R} C h ∩ R relative to R \mathbb{R} R , so for a small disk B B B every real point of Z h ∩ B Z_h \cap B Z h ∩ B lies in C h C_h C h , and the nonreal ones do as before; again Z h ∩ B = C h ∩ B Z_h \cap B = C_h \cap B Z h ∩ B = C h ∩ B . Since 0 ∉ C h 0 \notin C_h 0 ∈ / C h by (7) , Proposition 2.3 gives local connectedness of Z h Z_h Z h at x x x , hence of C h C_h C h .
Next, let K K K be a compact subset of the plane which is not locally connected at a point x x x . We claim that either distinct components of K K K accumulate at x x x , or x x x lies in a nondegenerate continuum at each point of which K K K fails to be locally connected. Choose r > 0 r>0 r > 0 such that no connected subset of K ∩ B ( x , r ) ‾ K\cap\overline{B(x,r)} K ∩ B ( x , r ) contains x x x in its interior relative to K K K , and let Q Q Q be the component of x x x in K ∩ B ( x , r ) ‾ K\cap\overline{B(x,r)} K ∩ B ( x , r ) . Then there are points y n → x y_n\to x y n → x of K K K outside Q Q Q ; let Q n Q_n Q n be the component of y n y_n y n in K ∩ B ( x , r ) ‾ K\cap\overline{B(x,r)} K ∩ B ( x , r ) . A component containing infinitely many y n y_n y n would contain x x x and equal Q Q Q , so after passing to a subsequence the Q n Q_n Q n are pairwise distinct. If Q n Q_n Q n does not meet the circle ∂ B ( x , r ) \partial B(x,r) ∂ B ( x , r ) , then Q n Q_n Q n is a component of K K K : a continuum in K K K containing Q n Q_n Q n and a point outside B ( x , r ) B(x,r) B ( x , r ) would, by the boundary bumping theorem [22 Chapter V] , contain a continuum joining Q n Q_n Q n to the circle inside B ( x , r ) B(x,r) B ( x , r ) , which would then lie in Q n Q_n Q n . So either infinitely many Q n Q_n Q n are components of K K K accumulating at x x x , or infinitely many Q n Q_n Q n meet the circle; in the latter case a subsequence converges in the Hausdorff metric to a continuum C C C containing x x x and a point of the circle. Every point y y y of C C C in the open disk B ( x , r ) B(x,r) B ( x , r ) is a point of non-local-connectedness of K K K : a connected V ⊆ K ∩ B ( x , r ) V\subseteq K\cap B(x,r) V ⊆ K ∩ B ( x , r ) containing y y y in its relative interior would lie in one component of K ∩ B ( x , r ) K\cap B(x,r) K ∩ B ( x , r ) and meet Q n Q_n Q n for all large n n n . By boundary bumping applied to C C C and B ( x , r / 2 ) ‾ \overline{B(x,r/2)} B ( x , r /2 ) , the component of x x x in C ∩ B ( x , r / 2 ) ‾ C\cap\overline{B(x,r/2)} C ∩ B ( x , r /2 ) is a nondegenerate continuum of such points. This proves the claim.
Apply the claim to K = C h K=C_h K = C h at a point x x x where C h C_h C h is not locally connected. By the first paragraph, every point of non-local-connectedness lies in N N N , and a nondegenerate continuum in N ⊂ R N\subset\mathbb{R} N ⊂ R would be an interval, contradicting nowhere density. Hence distinct components of C h C_h C h accumulate at x x x , and x ∈ N x\in N x ∈ N is real. Conversely, if distinct components K n K_n K n accumulate at x x x , then C h C_h C h is not locally connected at x x x : a connected subset of C h C_h C h containing a neighbourhood of x x x would lie in the component of x x x and meet every K n K_n K n with n n n large.
Now let K K K be a component of C h C_h C h and x ∈ K x\in K x ∈ K . If C h C_h C h is locally connected at x x x , so is K K K , because a connected subset of C h C_h C h containing a neighbourhood of x x x lies in K K K . Thus the points of non-local-connectedness of K K K lie in N N N . If K K K is not locally connected at x x x , apply the claim to K K K with r < diam K / 2 r<\operatorname{diam} K/2 r < diam K /2 : the first alternative is impossible for a continuum, and the second gives a nondegenerate continuum inside N N N . So every component is locally connected. Finally, if K n K_n K n were distinct components of diameter at least ε \varepsilon ε , a subsequence would converge to a continuum C C C of diameter at least ε \varepsilon ε each of whose points is an accumulation point of distinct components, hence a point of N N N ; this is again impossible.
By [21 Theorem 3 and Lemma 3.8] , the conclusion is equivalent to the following finiteness property, visible in the figures: for every closed strip bounded by two parallel lines, only finitely many components of its intersection with C h C_h C h meet both lines. At the heights of Theorems B and C, infinitely many components accumulate at real points, and their diameters tend to zero.
Real limits and coefficient saturation
Recall that D h D_h D h is the real multiple-zero set in ( 0 , 1 ) (0,1) ( 0 , 1 ) and A h A_h A h is its active subset. Inversion identifies A h A_h A h with the positive exterior active trace. Both sets are relatively closed, by coefficient compactness and the definition of the core.
Lemma 2.6. One has A h ⊆ D h A_h\subseteq D_h A h ⊆ D h . Every t ∈ D h t\in D_h t ∈ D h satisfies t − 1 ≤ 1 + h t^{-1}\leq1+\sqrt{h} t − 1 ≤ 1 + h . Equality forces
f ( z ) = ( 1 − t − 1 z ) 2 1 − z , a 1 = − ( 1 + 2 h ) , a j = h ( j ≥ 2 ) . f(z)=\frac{(1-t^{-1}z)^2}{1-z},\qquad a_1=-(1+2\sqrt{h}),\qquad a_j=h\quad(j\geq2). f ( z ) = 1 − z ( 1 − t − 1 z ) 2 , a 1 = − ( 1 + 2 h ) , a j = h ( j ≥ 2 ) .
Proof. At a real limit of nonreal roots, take a coefficientwise limit of their normalized exterior witnesses. The conjugate secants tend to f ′ ( t ) f'(t) f ′ ( t ) , so both f ( t ) f(t) f ( t ) and f ′ ( t ) f'(t) f ′ ( t ) vanish. Subtracting the value equation from the weighted derivative equation gives
1 = ∑ j ≥ 2 ( j − 1 ) a j t j ≤ h t 2 ( 1 − t ) 2 . 1=\sum_{j\geq2}(j-1)a_jt^j\leq\frac{ht^2}{(1-t)^2}. 1 = j ≥ 2 ∑ ( j − 1 ) a j t j ≤ ( 1 − t ) 2 h t 2 .
All weights are positive, so equality forces a j = h a_j=h a j = h for j ≥ 2 j\geq2 j ≥ 2 ; the value equation then determines a 1 a_1 a 1 .
For real coefficients in [ − h , h ] [-h,h] [ − h , h ] , the bound t − 1 ≤ 1 + h t^{-1}\leq1+\sqrt{h} t − 1 ≤ 1 + h and its extremal series are due to Beaucouр, Borwein, Boyd and Pinner [4 ] , who proved that for h ≥ 3 + 2 2 h\geq3+2\sqrt{2} h ≥ 3 + 2 2 the smallest disk containing two zeros of such a series with constant term 1 1 1 has radius ( 1 + h ) − 1 (1+\sqrt{h})^{-1} ( 1 + h ) − 1 . The integer alphabet changes what happens at the extremal point. The saturated series with t − 1 = q + 1 t^{-1}=q+1 t − 1 = q + 1 is admissible at height q 2 q^2 q 2 , but its double zero is inactive there (Proposition 3.4 ); nonreal roots reach q + 1 q+1 q + 1 only from height q 2 + 1 q^2+1 q 2 + 1 on, and the core reconnects only at q ( q + 1 ) q(q+1) q ( q + 1 ) (Theorem A and Figure 4 ).
Figure 4. The positive outer active endpoint for 1 ≤ h ≤ 42 1 \le h \le42 1 ≤ h ≤ 42 . Solid points give exact endpoints; circles around them mark connected cores. Open circles mark inactive double zeros. Grey bars give bounds only. For q ≥ 3 q \ge3 q ≥ 3 , each constant segment begins at q 2 + 1 q^2+1 q 2 + 1 and ends at the pronic height q ( q + 1 ) q(q+1) q ( q + 1 ) .
For the binary alphabet, the inclusion A 1 ⊆ D 1 \mathcal{A}_1 \subseteq\mathcal{D}_1 A 1 ⊆ D 1 is Lemma 2.5 of Solomyak [31 ] (see Lemma 2.9 of [28 ] ), and D 1 \mathcal{D}_1 D 1 is the set Ω 2 \Omega_2 Ω 2 of [28 ] ; Calegari, Koch and Walker [10, Section 10] note that it is not known whether some point of Ω 2 \Omega_2 Ω 2 lies outside the closure of the interior of the binary locus. By Proposition 2.4 that question asks whether D 1 ∖ A 1 \mathcal{D}_1 \setminus\mathcal{A}_1 D 1 ∖ A 1 is nonempty. The next lemma shows that this difference consists of isolated points of D h \mathcal{D}_h D h , at every height, so the binary question would be settled by showing that D 1 \mathcal{D}_1 D 1 has no isolated points; Proposition 3.2 identifies the difference as { 1 / 2 } \{1/2\} { 1/2 } at height two, and the binary case remains open.
Lemma 2.7 (Inactive double zeros are isolated). Every point of D h ∖ A h \mathcal{D}_h \setminus\mathcal{A}_h D h ∖ A h is an isolated point of D h \mathcal{D}_h D h , and D h ⊆ A h + 1 \mathcal{D}_h \subseteq\mathcal{A}_{h+1} D h ⊆ A h + 1 . Consequently D h ∖ A h \mathcal{D}_h \setminus\mathcal{A}_h D h ∖ A h is at most countable, and every nondegenerate interval in D h \mathcal{D}_h D h , including its endpoints in ( 0 , 1 ) (0,1) ( 0 , 1 ) , is contained in A h \mathcal{A}_h A h .
Proof. Let f ∈ F h f \in\mathcal{F}_h f ∈ F h have a zero of order ν ≥ 2 \nu\ge2 ν ≥ 2 at t t t , and write f ( t + w ) = b w ν + O ( w ν + 1 ) f(t+w)=bw^\nu+O(w^{\nu+1}) f ( t + w ) = b w ν + O ( w ν + 1 ) , b ≠ 0 b\ne0 b = 0 . Let ϵ ∈ { − 1 , 1 } \epsilon\in\{-1,1\} ϵ ∈ { − 1 , 1 } be such that b w ν + ϵ bw^\nu+\epsilon b w ν + ϵ has a nonreal zero. With r N = t N / ν r_N=t^{N/\nu} r N = t N / ν ,
t − N ( f ( t + r N w ) + ϵ ( t + r N w ) N ) ⟶ b w ν + ϵ t^{-N}\left(f(t+r_Nw)+\epsilon(t+r_Nw)^N\right)\longrightarrow bw^\nu+\epsilon t − N ( f ( t + r N w ) + ϵ ( t + r N w ) N ) ⟶ b w ν + ϵ
locally uniformly, because N r N → 0 Nr_N\to0 N r N → 0 . If f + ϵ z N f+\epsilon z^N f + ϵ z N is admissible for infinitely many N N N , Rouché gives nonreal zeros tending to t t t , so t ∈ A h t\in\mathcal{A}_h t ∈ A h . At height h + 1 h+1 h + 1 every f + ϵ z N f+\epsilon z^N f + ϵ z N is admissible, which proves D h ⊆ A h + 1 \mathcal{D}_h\subseteq\mathcal{A}_{h+1} D h ⊆ A h + 1 .
Now let t ∈ D h ∖ A h t\in\mathcal{D}_h\setminus\mathcal{A}_h t ∈ D h ∖ A h . For every witness f f f of t t t as above, a N = ϵ h a_N=\epsilon h a N = ϵ h for all large N N N . If ν ≥ 3 \nu\ge3 ν ≥ 3 , both signs ϵ \epsilon ϵ give b w ν + ϵ bw^\nu+\epsilon b w ν + ϵ a nonreal zero, so a N = h a_N=h a N = h and a N = − h a_N=-h a N = − h eventually, which is absurd; hence ν = 2 \nu=2 ν = 2 , ϵ \epsilon ϵ is the sign of b = f ′ ′ ( t ) / 2 b=f''(t)/2 b = f ′′ ( t ) /2 , and the coefficients of f f f equal ϵ h \epsilon h ϵ h from some index N 0 N_0 N 0 on. Suppose that distinct points t n ∈ D h t_n\in\mathcal{D}_h t n ∈ D h converge to t t t , with witnesses f n f_n f n having zeros of order at least two at t n t_n t n . Passing to a subsequence, f n f_n f n converges coefficientwise to a series f f f with f ( t ) = f ′ ( t ) = 0 f(t)=f'(t)=0 f ( t ) = f ′ ( t ) = 0 , to which the preceding paragraph applies; say ϵ = 1 \epsilon=1 ϵ = 1 , so f ′ ′ ( t ) > 0 f''(t)>0 f ′′ ( t ) > 0 and a N = h a_N=h a N = h for N ≥ N 0 N\ge N_0 N ≥ N 0 . For large n n n , the coefficients of f n f_n f n of index below N 0 N_0 N 0 agree with those of f f f , and all its coefficients are at most h h h ; hence
f n ( t ) − f ( t ) = ∑ N ≥ N 0 ( a N ( f n ) − h ) t N ≤ 0 , f_n(t)-f(t)=\sum_{N\ge N_0}(a_N(f_n)-h)t^N\le0, f n ( t ) − f ( t ) = N ≥ N 0 ∑ ( a N ( f n ) − h ) t N ≤ 0 ,
that is, f n ( t ) ≤ 0 f_n(t)\le0 f n ( t ) ≤ 0 . On the other hand f ′ ′ > 0 f''>0 f ′′ > 0 on a neighbourhood U U U of t t t , so f n ′ ′ > 0 f_n''>0 f n ′′ > 0 on U U U for large n n n ; then f n f_n f n is strictly convex on U U U with a zero of order at least two at t n ∈ U t_n\in U t n ∈ U , hence nonnegative on U U U with t n t_n t n as its only zero there, and t ≠ t n t\ne t_n t = t n gives f n ( t ) > 0 f_n(t)>0 f n ( t ) > 0 , a contradiction. For ϵ = − 1 \epsilon=-1 ϵ = − 1 all inequalities reverse. Thus t t t is isolated in D h \mathcal{D}_h D h .
A subset of R \mathbb{R} R has at most countably many isolated points, and no point of a nondegenerate interval in D h \mathcal{D}_h D h , nor an endpoint of it in ( 0 , 1 ) (0,1) ( 0 , 1 ) , is isolated in D h \mathcal{D}_h D h . □
The lemma asserts approximation by nonreal zeros, not planar interiority. This distinction is needed at height six, where a bounded value–derivative orbit completes the active trace without a strict planar-interior test at every real parameter.
The connected heights
Connectedness at h = q ( q − 1 ) h=q(q-1) h = q ( q − 1 ) comes from a complete real interval to which all nonreal pieces attach. There are two separate steps: an integer-digit argument bounds the endpoint by q q q , and a coefficient continuation fills the interval up to it. The coefficient construction is given in Sections S1–S2 of the supplement; the endpoint and topological arguments are given here.
The arithmetic endpoint
The universal double-zero bound from Section 2 allows an endpoint as large as 1 + h 1+\sqrt{h} 1 + h . At height h = q ( q − 1 ) h=q(q-1) h = q ( q − 1 ) , integer coefficients give the smaller endpoint q q q .
Lemma 3.1 (Pronic double-zero bound). Let q ≥ 2 q\ge2 q ≥ 2 be an integer and H = q ( q − 1 ) H=q(q-1) H = q ( q − 1 ) . If
f ( z ) = 1 − ∑ k ≥ 1 d k z k , d k ∈ Z , ∣ d k ∣ ≤ H , f(z)=1-\sum_{k\ge1}d_kz^k,\qquad d_k\in\mathbb{Z},\quad|d_k|\le H, f ( z ) = 1 − k ≥ 1 ∑ d k z k , d k ∈ Z , ∣ d k ∣ ≤ H , has a real double zero t ∈ ( − 1 , 1 ) t\in(-1,1) t ∈ ( − 1 , 1 ) , then ∣ t ∣ ≥ q − 1 |t|\ge q^{-1} ∣ t ∣ ≥ q − 1 .
Proof. Replace f ( z ) f(z) f ( z ) by f ( − z ) f(-z) f ( − z ) if necessary and suppose x = t − 1 > q x=t^{-1}>q x = t − 1 > q . The double-zero equations give the tail states
P n = ∑ k ≥ 1 d n + k x − k , V n = ∑ k ≥ 1 ( 1 − k ) d n + k x − k , ( P 0 , V 0 ) = ( 1 , 1 ) . P_n=\sum_{k\ge1}d_{n+k}x^{-k},\qquad V_n=\sum_{k\ge1}(1-k)d_{n+k}x^{-k},\qquad(P_0,V_0)=(1,1). P n = k ≥ 1 ∑ d n + k x − k , V n = k ≥ 1 ∑ ( 1 − k ) d n + k x − k , ( P 0 , V 0 ) = ( 1 , 1 ) .
Absolute convergence yields
P n + 1 = x P n − d n + 1 , V n + 1 = x V n + P n + 1 , ∣ P n ∣ ≤ H x − 1 , ∣ V n ∣ ≤ H ( x − 1 ) 2 . (9) P_{n+1}=xP_n-d_{n+1},\qquad V_{n+1}=xV_n+P_{n+1},\qquad|P_n|\le\frac{H}{x-1},\qquad|V_n|\le\frac{H}{(x-1)^2}. \tag*{(9)} P n + 1 = x P n − d n + 1 , V n + 1 = x V n + P n + 1 , ∣ P n ∣ ≤ x − 1 H , ∣ V n ∣ ≤ ( x − 1 ) 2 H . (9)
Put R = x − 1 R=x-1 R = x − 1 , δ = x − q \delta=x-q δ = x − q , and A = ( q − 1 ) 2 A=(q-1)^2 A = ( q − 1 ) 2 . The bound on V 0 V_0 V 0 gives R 2 ≤ H < q 2 R^2\le H<q^2 R 2 ≤ H < q 2 , hence 0 < δ < 1 0<\delta<1 0 < δ < 1 . Since H < 2 R 2 H<2R^2 H < 2 R 2 , the bound on V 1 = 2 x − d 1 V_1=2x-d_1 V 1 = 2 x − d 1 implies d 1 > 2 q − 2 d_1>2q-2 d 1 > 2 q − 2 . On the other hand,
R d 1 ≤ R x + H < 2 q R , 2 q R − ( R x + H ) = δ ( 1 − δ ) > 0. Rd_1\le Rx+H<2qR,\qquad2qR-(Rx+H)=\delta(1-\delta)>0. R d 1 ≤ R x + H < 2 q R , 2 q R − ( R x + H ) = δ ( 1 − δ ) > 0.
Integrality forces d 1 = 2 q − 1 d_1=2q-1 d 1 = 2 q − 1 . For q = 2 q=2 q = 2 , this already exceeds H H H .
For q ≥ 3 q\ge3 q ≥ 3 , define
U n = R P n + 1 + A , W n = R 2 V n + 1 − A . U_n=RP_{n+1}+A,\qquad W_n=R^2V_{n+1}-A. U n = R P n + 1 + A , W n = R 2 V n + 1 − A .
Their bounds and initial values are
− ( q − 1 ) ≤ U n ≤ H + A , W n ≤ q − 1 , U 0 = δ 2 , W 0 = δ ( 2 R 2 + x + q − 2 ) . -(q-1)\le U_n\le H+A,\qquad W_n\le q-1,\qquad U_0=\delta^2,\qquad W_0=\delta(2R^2+x+q-2). − ( q − 1 ) ≤ U n ≤ H + A , W n ≤ q − 1 , U 0 = δ 2 , W 0 = δ ( 2 R 2 + x + q − 2 ) .
In particular, U 0 ≥ 0 U_0\ge0 U 0 ≥ 0 and x ( q − 1 ) U 0 < δ W 0 x(q-1)U_0<\delta W_0 x ( q − 1 ) U 0 < δ W 0 , since the difference is δ 2 R ( 2 x − q ) > 0 \delta^2R(2x-q)>0 δ 2 R ( 2 x − q ) > 0 . Suppose a state satisfies
U ≥ 0 , x ( q − 1 ) U < δ W . (10) U\ge0,\qquad x(q-1)U<\delta W. \tag*{(10)} U ≥ 0 , x ( q − 1 ) U < δ W . (10)
Then W > 0 W>0 W > 0 and x U < δ xU<\delta xU < δ . Writing the next digit as d = − A + j d=-A+j d = − A + j , j ∈ Z j\in\mathbb{Z} j ∈ Z , gives
U ′ = x U − R j , W ′ = x W + x R U − R 2 j . U'=xU-Rj,\qquad W'=xW+xRU-R^2j. U ′ = xU − R j , W ′ = x W + x R U − R 2 j .
If j ≥ 1 j\ge1 j ≥ 1 , then U ′ < δ − R = − ( q − 1 ) U'<\delta-R=-(q-1) U ′ < δ − R = − ( q − 1 ) ; if j ≤ − 1 j\le-1 j ≤ − 1 , then W ′ > R 2 > q − 1 W'>R^2>q-1 W ′ > R 2 > q − 1 . Both contradict the state bounds, so j = 0 j=0 j = 0 . Moreover,
δ W ′ − x ( q − 1 ) U ′ = x ( δ W − x ( q − 1 ) U ) + δ x R U > 0. \delta W'-x(q-1)U'=x\bigl(\delta W-x(q-1)U\bigr)+\delta xRU>0. δ W ′ − x ( q − 1 ) U ′ = x ( δ W − x ( q − 1 ) U ) + δ x R U > 0.
Thus (3.2) persists, and induction gives U n = x n δ 2 U_n=x^n\delta^2 U n = x n δ 2 , contradicting U n ≤ H + A U_n\le H+A U n ≤ H + A . □
The forced digit is the essential arithmetic step. A continuously variable coefficient could correct the expanding state; an integer coefficient cannot do so without violating one of the two bounds.
Completing the real trace
The same value–derivative coordinates give a constructive way to prove that a whole interval consists of double zeros. For x > 1 x>1 x > 1 , start at ( P 0 , V 0 ) = ( 1 , 1 ) (P_0,V_0)=(1,1) ( P 0 , V 0 ) = ( 1 , 1 ) and choose digits d j ∈ B h d_j\in B_h d j ∈ B h under
P j + 1 = x P j − d j + 1 , V j + 1 = x V j + P j + 1 . (11) P_{j+1}=xP_j-d_{j+1},\qquad V_{j+1}=xV_j+P_{j+1}. \tag*{(11)} P j + 1 = x P j − d j + 1 , V j + 1 = x V j + P j + 1 . (11)
A bounded orbit yields a double zero at 1 / x 1/x 1/ x . Indeed,
x − m P m = 1 − ∑ j = 1 m d j x − j , x − m ( V m − ( m + 1 ) P m ) = ∑ j = 1 m j d j x − j . x^{-m}P_m=1-\sum_{j=1}^{m}d_jx^{-j},\qquad x^{-m}\bigl(V_m-(m+1)P_m\bigr)=\sum_{j=1}^{m}jd_jx^{-j}. x − m P m = 1 − j = 1 ∑ m d j x − j , x − m ( V m − ( m + 1 ) P m ) = j = 1 ∑ m j d j x − j .
Both expressions tend to zero. Thus 1 − ∑ j ≥ 1 d j z j ∈ F h 1-\sum_{j\ge1}d_jz^j\in\mathcal{F}_h 1 − ∑ j ≥ 1 d j z j ∈ F h vanishes with its derivative at 1 / x 1/x 1/ x .
Proposition 3.2 (Pronic continuation). For h = q ( q − 1 ) h=q(q-1) h = q ( q − 1 ) and q ≥ 3 q\ge3 q ≥ 3 , a bounded orbit of (3.3) exists for every 1 < x < q 1<x<q 1 < x < q . At height two, if ρ \rho ρ is the real root of ρ 3 = ρ 2 + ρ + 1 \rho^3=\rho^2+\rho+1 ρ 3 = ρ 2 + ρ + 1 , then
D 2 = { 1 / 2 } ∪ [ 1 / ρ , 1 ) , A 2 = [ 1 / ρ , 1 ) . D_2=\{1/2\}\cup[1/\rho,1),\qquad A_2=[1/\rho,1). D 2 = { 1/2 } ∪ [ 1/ ρ , 1 ) , A 2 = [ 1/ ρ , 1 ) .
The complete coefficient construction is given in Section S2 of the supplement. Here is its organization. Away from q q q , a finite prefix enters a strict invariant rectangle for the two coordinates. Near q q q , long repetitions of the digit − ( q − 1 ) 2 -(q-1)^2 − ( q − 1 ) 2 shadow its fixed state. Their return coordinates are an exponential factor times an affine function of the repetition length. The resulting parameter bands approach q q q and overlap, so there is no omitted interval between scales. Uniform estimates handle large q q q and long repetitions; exact polynomial inequalities handle the finite complement. At height six, bounded continuation replaces strict entry in the last interval. At height two, the continuation is organized by the period-three witness
( 1 − z − z 2 − z 3 ) 2 1 − z 3 = 1 − 2 z − z 2 + z 3 + z 4 + z 5 + 2 z 6 1 − z 3 ∈ F 2 . \frac{(1-z-z^{2}-z^{3})^{2}}{1-z^{3}}=1-2z-z^{2}+z^{3}+\frac{z^{4}+z^{5}+2z^{6}}{1-z^{3}}\in\mathcal{F}_{2}. 1 − z 3 ( 1 − z − z 2 − z 3 ) 2 = 1 − 2 z − z 2 + z 3 + 1 − z 3 z 4 + z 5 + 2 z 6 ∈ F 2 .
Sections S2 and S4 of the supplement include the complete covers, their endpoints and the finite verification statements. In particular, the construction does not infer interval coverage from sampled parameters.
Proposition 3.3 (Connected pronic cores). Let h = q ( q − 1 ) h=q(q-1) h = q ( q − 1 ) , q ≥ 2 q\geq2 q ≥ 2 , and set a = q a=q a = q for q ≥ 3 q\geq3 q ≥ 3 , a = ρ a=\rho a = ρ for q = 2 q=2 q = 2 . Then
C h ∩ R = [ − a , − a − 1 ] ∪ [ a − 1 , a ] . (12) C_h\cap\mathbb{R}=[-a,-a^{-1}]\cup[a^{-1},a]. \tag*{(12)} C h ∩ R = [ − a , − a − 1 ] ∪ [ a − 1 , a ] . (12) The core is connected and locally connected. It is a strong deformation retract of Z h ∖ { 0 } Z_h\setminus\{0\} Z h ∖ { 0 } , and the complement of the core in that set consists of the four intervals
[ − h − 1 , − a ) , ( − a − 1 , − ( h + 1 ) − 1 ] , [ ( h + 1 ) − 1 , a − 1 ) , ( a , h + 1 ] . (13) [-h-1,-a),\quad(-a^{-1},-(h+1)^{-1}],\quad[(h+1)^{-1},a^{-1}),\quad(a,h+1]. \tag*{(13)} [ − h − 1 , − a ) , ( − a − 1 , − ( h + 1 ) − 1 ] , [( h + 1 ) − 1 , a − 1 ) , ( a , h + 1 ] . (13) Proof. For q ≥ 3 q\geq3 q ≥ 3 , continuation and the sharp pronic bound give D h = [ 1 / q , 1 ) D_h=[1/q,1) D h = [ 1/ q , 1 ) . The endpoint is supplied by f q = ( 1 − q z ) 2 / ( 1 − z ) ∈ F h f_q=(1-qz)^2/(1-z)\in\mathcal{F}_h f q = ( 1 − q z ) 2 / ( 1 − z ) ∈ F h . Lemma 2.7 then gives A h = [ 1 / q , 1 ) A_h=[1/q,1) A h = [ 1/ q , 1 ) . For q = 2 q=2 q = 2 , use Proposition 3.2. Negation and inversion prove (3.4).
Every nonreal point of Z h Z_h Z h already belongs to C h C_h C h by definition. Subtracting the active trace from (7) therefore gives (3.5). Fix the core pointwise and contract the closure of each displayed interval linearly onto its endpoint in the core. The definitions agree on their closed overlaps, giving a strong deformation retraction. Connectedness follows from that of the full nonzero locus. For local connectedness, given a neighbourhood U U U of y y y in the core, choose a connected open neighbourhood V V V of y y y in the full locus contained in the inverse image of U U U under the retraction r r r . Then r ( V ) r(V) r ( V ) is a connected neighbourhood of y y y contained in U U U . □
Why the next endpoint appears before reconnection
This short calculation explains the distinction between double-zero existence, activation and reconnection in Figure 4 . For q ≥ 3 q\geq3 q ≥ 3 , set
F q ( z ) = ( 1 − ( q + 1 ) z ) 2 1 − z = 1 − ( 2 q + 1 ) z + q 2 ∑ j ≥ 2 z j . F_q(z)=\frac{(1-(q+1)z)^2}{1-z}=1-(2q+1)z+q^2\sum_{j\geq2}z^j. F q ( z ) = 1 − z ( 1 − ( q + 1 ) z ) 2 = 1 − ( 2 q + 1 ) z + q 2 j ≥ 2 ∑ z j .
Proposition 3.4 (Endpoint activation). At height q 2 q^2 q 2 , the double zero 1 / ( q + 1 ) 1/(q+1) 1/ ( q + 1 ) is isolated in D q 2 D_{q^2} D q 2 and is not active. For q 2 + 1 ≤ h ≤ q ( q + 1 ) q^2+1\leq h\leq q(q+1) q 2 + 1 ≤ h ≤ q ( q + 1 ) , the greatest positive real core point is q + 1 q+1 q + 1 .
Proof. Equality in the universal double-zero bound makes F q F_q F q the unique witness at t = 1 / ( q + 1 ) t=1/(q+1) t = 1/ ( q + 1 ) in height q 2 q^2 q 2 , with F q ′ ′ ( t ) > 0 F_q''(t)>0 F q ′′ ( t ) > 0 . Suppose qualifying zeros z j ≠ t z_j\ne t z j = t tend to t t t , where qualifying means nonreal or real of multiplicity at least two. Compactness forces their witnesses g j g_j g j to converge to F q F_q F q . Factoring their nearby conjugate pair, or a real double zero, gives
g j ( t ) = ∣ t − z j ∣ 2 Q j ( t ) > 0 , Q j ( t ) ⟶ F q ′ ′ ( t ) / 2 > 0. g_j(t)=|t-z_j|^2Q_j(t)>0,\qquad Q_j(t)\longrightarrow F_q''(t)/2>0. g j ( t ) = ∣ t − z j ∣ 2 Q j ( t ) > 0 , Q j ( t ) ⟶ F q ′′ ( t ) /2 > 0. But the first coefficient eventually agrees with that of F q F_q F q , while all later coefficients are at most q 2 q^2 q 2 . Hence g j ( t ) ≤ F q ( t ) = 0 g_j(t)\leq F_q(t)=0 g j ( t ) ≤ F q ( t ) = 0 , a contradiction. This proves isolation and inactivity.
At height q 2 + 1 q^2+1 q 2 + 1 , every F q + z N F_q+z^N F q + z N , N ≥ 2 N\geq2 N ≥ 2 , is admissible. The quadratic scaling used in Lemma 2.7 gives nonreal zeros tending to t t t , so q + 1 q+1 q + 1 is active. Nesting preserves this up to height q ( q + 1 ) q(q+1) q ( q + 1 ) , and the pronic bound at that final height excludes every larger positive real core point. □
Thus an endpoint can be present throughout a long range of heights while its relation to the other components changes. The next section proves that, before the pronic height, this endpoint is itself a singleton component and is approached by infinitely many other components.
Cascades of global components
Perturbing a double zero produces nearby nonreal roots, but does not separate their connected components. A point may be a zero of several admissible series, so disjoint families of coefficient sequences can have overlapping root sets. Even a separation within a small disk could be bypassed by a connection outside it. We address both difficulties by constructing root sets that are open and closed in the entire core. One comparison lemma supplies the required separation throughout both height intervals in Theorem B.
We use the normalized family F H \mathcal{F}_H F H from § 2 . A zero is called qualifying if it is nonreal, or if it is real and has multiplicity at least two. This terminology records a necessary condition at a real limit of nonreal roots; it does not assert that every multiple real zero belongs to the core.
From coefficient blocks to global components
The normalization below ensures that the coefficient comparisons cover every core point in the chosen neighbourhood, including real limits of nonreal roots. Convergence of bounded coefficient sequences is locally uniform in the unit disk, including all derivatives.
Lemma 4.1 (Local witnesses and positive perturbations). Put σ H = ( H 2 + 4 ) − 1 / 4 \sigma_H=(H^2+4)^{-1/4} σ H = ( H 2 + 4 ) − 1/4 . Every z ∈ C H z\in C_H z ∈ C H with 0 < ∣ z ∣ < σ H 0<|z|<\sigma_H 0 < ∣ z ∣ < σ H has a witness in F H \mathcal{F}_H F H for which it is a qualifying zero. The qualifying-zero relation is closed on compact subdisks.
Suppose F ( t ) = F ′ ( t ) = 0 F(t)=F'(t)=0 F ( t ) = F ′ ( t ) = 0 and b = F ′ ′ ( t ) / 2 > 0 b=F''(t)/2>0 b = F ′′ ( t ) /2 > 0 , where 0 < t < 1 0<t<1 0 < t < 1 . Let C N C_N C N be power series whose coefficients are bounded independently of N N N , with C N ( t ) ≥ c > 0 C_N(t)\ge c>0 C N ( t ) ≥ c > 0 . If F + z N C N ∈ F h F+z^N C_N\in\mathcal{F}_h F + z N C N ∈ F h for arbitrarily large N N N , then every sufficiently large such N N N gives a nonreal conjugate pair of zeros of F + z N C N F+z^N C_N F + z N C N in C h C_h C h , and these pairs tend to t t t .
Proof. For a nonreal polynomial root, remove any factor z k z^k z k and make the constant coefficient positive. If its value is an integer d ≥ 2 d\ge2 d ≥ 2 , dividing by d d d and applying Lemma 2.2 gives
∣ z ∣ − 4 ≤ 4 ( 1 + ( H / d ) 2 ) ≤ H 2 + 4. |z|^{-4}\le4(1+(H/d)^2)\le H^2+4. ∣ z ∣ − 4 ≤ 4 ( 1 + ( H / d ) 2 ) ≤ H 2 + 4.
Thus roots in the stated strict disk have constant-term-one witnesses. For a core point, take a convergent sequence of nonreal polynomial roots and a coefficientwise convergent subsequence of these witnesses. At a real limit, the conjugate roots coalesce: their divided difference converges to the derivative, which therefore vanishes. This also proves closedness of the qualifying-zero relation, using derivative convergence when the approximating zeros are already real and multiple.
For the last assertion put ϵ N = t N / 2 \epsilon_N=t^{N/2} ϵ N = t N /2 . Uniformly for w w w in compact sets,
t − N ( F + z N C N ) ( t + ϵ N w ) − ( b w 2 + C N ( t ) ) ⟶ 0 , t^{-N}(F+z^N C_N)(t+\epsilon_N w)-(bw^2+C_N(t))\longrightarrow0, t − N ( F + z N C N ) ( t + ϵ N w ) − ( b w 2 + C N ( t )) ⟶ 0 ,
because t − N F ( t + ϵ N w ) → b w 2 t^{-N}F(t+\epsilon_Nw)\to bw^2 t − N F ( t + ϵ N w ) → b w 2 , ( 1 + ϵ N w / t ) N → 1 (1+\epsilon_Nw/t)^N\to1 ( 1 + ϵ N w / t ) N → 1 since N ϵ N → 0 N\epsilon_N\to0 N ϵ N → 0 , and C N ( t + ϵ N w ) − C N ( t ) = O ( ϵ N ) C_N(t+\epsilon_Nw)-C_N(t)=O(\epsilon_N) C N ( t + ϵ N w ) − C N ( t ) = O ( ϵ N ) by the bound on the coefficients. The zeros ± i C N ( t ) / b \pm i\sqrt{C_N(t)/b} ± i C N ( t ) / b of the comparison function stay in a compact subset of C ∖ R \mathbb{C}\setminus\mathbb{R} C ∖ R , and on circles of a fixed small radius about them it is bounded away from zero uniformly in N N N . Rouché’s theorem gives the required pair. Truncating the series and applying Rouché again near either nonreal zero places the pair in the polynomial-root core. □
If G ≠ F G\ne F G = F , its first departure from F F F is the expression
G − F = z k C , C ( 0 ) ≠ 0 , G-F=z^kC,\qquad C(0)\ne0, G − F = z k C , C ( 0 ) = 0 ,
where k k k is the first differing coefficient. A finite departure block is a finite union of coefficient cylinders specified by such a degree and a set of allowed nonzero first differences. For example, the negative departures with degrees in a given finite interval form one.
The next lemma assumes that no block witness shares a qualifying zero with any outside witness in a fixed neighbourhood. Compactness and rigidity then give relative openness of the block’s nearby root set in the core. A quadratic estimate at the reference double zero places this set strictly inside the chosen disk for sufficiently late blocks. It is therefore clopen in the entire core and contains whole global components, although the set itself need not be connected.
Lemma 4.2 (Global separation of finite blocks). Let F r ∈ F H F_r \in\mathcal{F}_H F r ∈ F H converge locally uniformly to F ∗ F_* F ∗ , and let t r → t ∗ ∈ ( 0 , σ H ) t_r \to t_* \in(0,\sigma_H) t r → t ∗ ∈ ( 0 , σ H ) . Suppose each F r F_r F r has an exact double zero at t r t_r t r , is the unique member of F H \mathcal{F}_H F H with a multiple zero there, and
F ∗ ( t ∗ ) = F ∗ ′ ( t ∗ ) = 0 , F ∗ ′ ′ ( t ∗ ) > 0. F_*(t_*)=F_*'(t_*)=0,\qquad F_*''(t_*)>0. F ∗ ( t ∗ ) = F ∗ ′ ( t ∗ ) = 0 , F ∗ ′′ ( t ∗ ) > 0. For each r r r choose a finite departure block B r \mathcal{B}_r B r relative to F r F_r F r , whose least degree n r n_r n r tends to infinity. Assume that in one fixed neighbourhood of t ∗ t_* t ∗ no witness in B r \mathcal{B}_r B r shares a qualifying zero with any witness outside B r \mathcal{B}_r B r .
In a sufficiently small fixed closed disk D D D about t ∗ t_* t ∗ , let K r K_r K r be the qualifying-root projection of B r \mathcal{B}_r B r , intersected with C H C_H C H . Then K r K_r K r is compact, excludes t r t_r t r , and, for all large r r r , is clopen in the entire core. There are constants c > 0 c>0 c > 0 and ρ < 1 \rho<1 ρ < 1 such that
K r ⊆ { z : ∣ z − t r ∣ 2 ≤ 2 H ρ n r c ( 1 − ρ ) } . (14) K_r\subseteq\left\{z:|z-t_r|^2\leq\frac{2H\rho^{n_r}}{c(1-\rho)}\right\}. \tag*{(14)} K r ⊆ { z : ∣ z − t r ∣ 2 ≤ c ( 1 − ρ ) 2 H ρ n r } . (14) In particular their diameters tend to zero. For any h ∗ ≤ H h_*\leq H h ∗ ≤ H :
(i) If the K r K_r K r are pairwise disjoint and infinitely many meet C h ∗ C_{h_*} C h ∗ , they contain distinct components with nonempty planar interior, nested as the height increases from h ∗ h_* h ∗ to H H H . These components accumulate at t ∗ t_* t ∗ , where none of these cores is locally connected.
(ii) If every K r K_r K r contains t ∗ t_* t ∗ , and t ∗ t_* t ∗ is nonisolated in C h ∗ C_{h_*} C h ∗ , then { t ∗ } \{t_*\} { t ∗ } is a component at every height h ∗ ≤ h ≤ H h_*\leq h\leq H h ∗ ≤ h ≤ H . Every neighbourhood of it contains infinitely many interior-bearing components, again chosen nested across these heights and with their diameters at height H H H tending to zero. The cores are not locally connected at t ∗ t_* t ∗ .
Proof. On a small common disk,
F r ( z ) = ( z − t r ) 2 Q r ( z ) , ∣ Q r ( z ) ∣ ≥ c > 0 , F_r(z)=(z-t_r)^2Q_r(z),\qquad|Q_r(z)|\geq c>0, F r ( z ) = ( z − t r ) 2 Q r ( z ) , ∣ Q r ( z ) ∣ ≥ c > 0 ,
for all large r r r . Indeed the quotients converge to F ∗ ( z ) / ( z − t ∗ ) 2 F_*(z)/(z-t_*)^2 F ∗ ( z ) / ( z − t ∗ ) 2 , which is nonzero near t ∗ t_* t ∗ . Choose D D D within the normalization and separation neighbourhoods, and put ρ = max z ∈ D ∣ z ∣ < 1 \rho=\max_{z\in D}|z|<1 ρ = max z ∈ D ∣ z ∣ < 1 . The coefficient block is compact, as is its qualifying-root projection in the closed disk and closed core. Rigidity excludes t r t_r t r . At a root of a departure of degree k ≥ n r k\geq n_r k ≥ n r ,
c ∣ z − t r ∣ 2 ≤ ∣ F r ( z ) ∣ ≤ 2 H ρ k 1 − ρ , c|z-t_r|^2\leq|F_r(z)|\leq\frac{2H\rho^k}{1-\rho}, c ∣ z − t r ∣ 2 ≤ ∣ F r ( z ) ∣ ≤ 1 − ρ 2 H ρ k ,
which proves (4.1).
Fix r r r and suppose points of ( C H ∩ D ) ∖ K r (C_H\cap D)\setminus K_r ( C H ∩ D ) ∖ K r converge to z ∈ K r z\in K_r z ∈ K r . Their normalized qualifying witnesses lie in F H ∖ B r \mathcal{F}_H\setminus\mathcal{B}_r F H ∖ B r . A finite union of coefficient cylinders is clopen in the compact coefficient space, so this complement is compact. Pass to a convergent subsequence of the witnesses. The limit still lies outside B r \mathcal{B}_r B r and has a qualifying zero at z z z , by closedness of that relation. A witness in B r \mathcal{B}_r B r also has a qualifying zero there, contradicting the separation hypothesis. Thus K r K_r K r is relatively open in C H ∩ D C_H\cap D C H ∩ D . The shrinking estimate eventually places it strictly inside D D D , so it is open in the entire core. Compactness makes it globally closed.
For (i), K r ∩ C h ∗ K_r\cap C_{h_*} K r ∩ C h ∗ is nonempty and relatively open. By Proposition 2.4, it contains a planar interior point. Choose its component at height h ∗ h_* h ∗ and, at each larger height, the component containing it. These components are nested and remain in K r K_r K r , giving distinct components and the asserted shrinking. Their accumulation puts t ∗ t_* t ∗ in every intermediate core. A connected relatively open neighbourhood of t ∗ t_* t ∗ could not meet a clopen K r K_r K r that excludes t ∗ t_* t ∗ , whereas every neighbourhood meets all sufficiently late occupied sets. At most one of the pairwise disjoint K r K_r K r can contain t ∗ t_* t ∗ , so discarding it proves failure of local connectedness.
For (ii), the component of t ∗ t_* t ∗ lies in every K r K_r K r and is therefore a singleton. To obtain disjoint sets meeting the smallest core, form successive finite intersections V j V_j V j of the K r K_r K r . Nonisolation lets us choose x j ∈ V j ∩ C h ∗ x_j\in V_j\cap C_{h_*} x j ∈ V j ∩ C h ∗ different from t ∗ t_* t ∗ ; choose the next intersection to exclude x j x_j x j and to have diameter less than 1 / j 1/j 1/ j . The disjoint clopen differences V j ∖ V j + 1 V_j\setminus V_{j+1} V j ∖ V j + 1 meet C h ∗ C_{h_*} C h ∗ and shrink to t ∗ t_* t ∗ . Apply the component selection just proved. A nonisolated singleton component has no connected relatively open neighbourhood, giving the last assertion. □ \square □
A common comparison for negative departures
The following lemma verifies the no-shared-zero hypothesis by comparing both the values and the derivatives of departure tails. A common nonreal zero brings its conjugate with it; as the pair approaches the real anchor, the limiting difference has a double zero. At a shared real multiple zero, both the difference and its derivative already vanish. The resulting relations contradict three strict coefficient inequalities. Those inequalities also give uniqueness of the reference witness. Their uniformity over a compact family will allow the reference itself to move at the height p ( p − 1 ) − 1 p(p-1)-1 p ( p − 1 ) − 1 just below a pronic height.
Lemma 4.3 (Uniform comparison for a compact family). Fix integers H ≥ 1 H \ge1 H ≥ 1 and 0 ≤ m ≤ H 0 \le m \le H 0 ≤ m ≤ H . Let F \mathcal{F} F be a compact family of pairs ( F , t ) (F,t) ( F , t ) , in coefficientwise and ordinary convergence, where F ∈ F H F \in\mathcal{F}_H F ∈ F H ,
F ( t ) = F ′ ( t ) = 0 , F ′ ′ ( t ) > 0 , 0 < t < 1 , m ≤ a j ( F ) ≤ H ( j ≥ 2 ) . F(t)=F'(t)=0,\qquad F''(t)>0,\qquad0<t<1,\qquad m\le a_j(F)\le H\quad(j\ge2). F ( t ) = F ′ ( t ) = 0 , F ′′ ( t ) > 0 , 0 < t < 1 , m ≤ a j ( F ) ≤ H ( j ≥ 2 ) . Put ℓ = H − m \ell=H-m ℓ = H − m and β = 1 / t \beta=1/t β = 1/ t . Suppose throughout the family that
ℓ β − 1 < 2 , 1 − ℓ ( β − 1 ) 2 > 0 , Δ : = 2 β − ℓ − 2 H ( β − 1 ) 2 > 0. (15) \frac{\ell}{\beta-1}<2,\qquad1-\frac{\ell}{(\beta-1)^2}>0,\qquad\Delta:=2\beta-\ell-\frac{2H}{(\beta-1)^2}>0. \tag*{(15)} β − 1 ℓ < 2 , 1 − ( β − 1 ) 2 ℓ > 0 , Δ := 2 β − ℓ − ( β − 1 ) 2 2 H > 0. (15) Each F F F is the unique normalized height-H H H witness with a multiple zero at t t t . In one neighbourhood of each anchor, of radius independent of the reference, two witnesses cannot share a qualifying zero if their first departures have opposite signs, or if they are both negative and have degrees differing by at least two. Adjacent negative departures at degrees j − 1 , j j-1,j j − 1 , j are also separated whenever a j ( F ) > m a_j(F)>m a j ( F ) > m . Consequently, if N < j N<j N < j and a N ( F ) > m a_N(F)>m a N ( F ) > m , a j ( F ) > m a_j(F)>m a j ( F ) > m , then in that neighbourhood no negative departure from F F F with degree in [ N , j − 1 ] [N,j-1] [ N , j − 1 ] shares a qualifying zero with a witness that is not such a departure.
Proof. For a first departure G − F = z k C G-F=z^kC G − F = z k C , every nonconstant coefficient of C C C lies in [ − 2 H , ℓ ] [-2H,\ell] [ − 2 H , ℓ ] . Write v = C ( t ) v=C(t) v = C ( t ) and w = t C ′ ( t ) w=tC'(t) w = t C ′ ( t ) . The three cases for its nonzero integer constant coefficient give
C ( 0 ) ≥ 1 : v − w ≥ 1 − ℓ ( β − 1 ) 2 > 0 , C ( 0 ) ≤ − 2 : v ≤ − 2 + ℓ β − 1 < 0 , C ( 0 ) = − 1 : w − 2 v ≥ Δ β > 0. (16) \begin{aligned}
C(0)\ge1 &: \quad v-w\ge1-\frac{\ell}{(\beta-1)^2}>0,\\
C(0)\le-2 &: \quad v\le-2+\frac{\ell}{\beta-1}<0,\\
C(0)=-1 &: \quad w-2v\ge\frac{\Delta}{\beta}>0. \tag*{(16)}
\end{aligned} C ( 0 ) ≥ 1 C ( 0 ) ≤ − 2 C ( 0 ) = − 1 : v − w ≥ 1 − ( β − 1 ) 2 ℓ > 0 , : v ≤ − 2 + β − 1 ℓ < 0 , : w − 2 v ≥ β Δ > 0. (16)
For example, in the last line the coefficient of degree two cancels:
t C ′ ( t ) − 2 C ( t ) ≥ 2 − ℓ t − 2 H ∑ j ≥ 3 ( j − 2 ) t j = Δ β . tC'(t)-2C(t)\ge2-\ell t-2H\sum_{j\ge3}(j-2)t^j=\frac{\Delta}{\beta}. t C ′ ( t ) − 2 C ( t ) ≥ 2 − ℓ t − 2 H j ≥ 3 ∑ ( j − 2 ) t j = β Δ .
A second multiple-zero witness would require v = w = 0 v=w=0 v = w = 0 , which proves rigidity. All three margins are uniform by compactness.
Suppose the claimed uniform separation fails. Take reference pairs ( F r , t r ) → ( F ∗ , t ∗ ) (F_r,t_r)\to(F_\ast,t_\ast) ( F r , t r ) → ( F ∗ , t ∗ ) and forbidden shared qualifying zeros tending to t ∗ t_\ast t ∗ . Every coefficient limit of their witnesses has a multiple zero at t ∗ t_\ast t ∗ , by Lemma 4.1. Rigidity makes that limit F ∗ F_\ast F ∗ . Discreteness of the coefficients then forces both first departure degrees to tend to infinity.
The usual two-zero argument is uniform over this compact family. In small real-symmetric disks the reference has just its double zero, is strictly convex on the real diameter, and is positive at its ends. Late departures converge with two derivatives; Rouché gives exactly two zeros, and strict convexity persists. If such a departure has a qualifying zero, its value at t r t_r t r is positive: a nonreal conjugate pair leaves no real zero in the disk, while a real double zero is its unique real minimum; equality at t r t_r t r would contradict rigidity. Thus C ( t r ) > 0 C(t_r)>0 C ( t r ) > 0 , and limiting departure values are nonnegative. By (4.3), every negative departure has constant coefficient − 1 -1 − 1 .
Order the two departure degrees as k r ≤ l r k_r\le l_r k r ≤ l r . First let the earlier one be positive and the later one negative, ordering a mixed pair at equal degrees this way. If l r − k r = s l_r-k_r=s l r − k r = s is fixed, coefficient limits of their tails C , D C,D C , D satisfy
v C = t ∗ s v D , w C = t ∗ s ( w D + s v D ) , v C , v D ≥ 0. v_C=t_\ast^s v_D,\qquad w_C=t_\ast^s(w_D+s v_D),\qquad v_C,v_D\ge0. v C = t ∗ s v D , w C = t ∗ s ( w D + s v D ) , v C , v D ≥ 0.
These equations follow because the difference has coalescing conjugate zeros, or a real multiple zero. The two strict bounds give
v C − w C = t ∗ s ( ( 1 − s ) v D − w D ) ≤ − t ∗ s ( ( s + 1 ) v D + Δ ∗ β ∗ ) < 0 , v_C-w_C=t_\ast^s\bigl((1-s)v_D-w_D\bigr)\le-t_\ast^s\left((s+1)v_D+\frac{\Delta_\ast}{\beta_\ast}\right)<0, v C − w C = t ∗ s ( ( 1 − s ) v D − w D ) ≤ − t ∗ s ( ( s + 1 ) v D + β ∗ Δ ∗ ) < 0 ,
contrary to the positive-departure bound. If instead the delays tend to infinity, C r − z l r − k r D r C_r-z^{l_r-k_r}D_r C r − z l r − k r D r converges locally uniformly to C C C . Its coalescing zeros force C ( t ∗ ) = C ′ ( t ∗ ) = 0 C(t_*)=C'(t_*)=0 C ( t ∗ ) = C ′ ( t ∗ ) = 0 , again impossible.
When the earlier departure is negative, compare the actual witnesses G r G_r G r and G ~ r \widetilde{G}_r G r , which depart at degrees k r k_r k r and l r l_r l r :
E r ( z ) = G r ( z ) − G ~ r ( z ) z k r . E_r(z)=\frac{G_r(z)-\widetilde{G}_r(z)}{z^{k_r}}. E r ( z ) = z k r G r ( z ) − G r ( z ) .
Its constant coefficient is at most − 1 -1 − 1 , and all coefficients of degree at least two lie in [ − 2 H , 2 H ] [-2H,2H] [ − 2 H , 2 H ] . Its degree-one coefficient is at most
l r − k r ≥ 2 H − a k r + 1 ( F r ) ≤ ℓ l r − k r = 1 , later positive H − a k r + 1 ( F r ) − 1 ≤ ℓ − 1 l r − k r = 1 , later negative , a k r + 1 ( F r ) > m H − a k r + 1 ( F r ) + 1 ≤ ℓ . \begin{aligned}
l_r-k_r &\ge2 && H-a_{k_r+1}(F_r)\le\ell\\
l_r-k_r &=1,\ \text{later positive} && H-a_{k_r+1}(F_r)-1\le\ell-1\\
l_r-k_r &=1,\ \text{later negative},\ a_{k_r+1}(F_r)>m && H-a_{k_r+1}(F_r)+1\le\ell.
\end{aligned} l r − k r l r − k r l r − k r ≥ 2 = 1 , later positive = 1 , later negative , a k r + 1 ( F r ) > m H − a k r + 1 ( F r ) ≤ ℓ H − a k r + 1 ( F r ) − 1 ≤ ℓ − 1 H − a k r + 1 ( F r ) + 1 ≤ ℓ .
The final line uses integrality and the later constant coefficient − 1 -1 − 1 . Passing to a coefficient limit E E E gives E ( t ∗ ) = E ′ ( t ∗ ) = 0 E(t_*)=E'(t_*)=0 E ( t ∗ ) = E ′ ( t ∗ ) = 0 , for bounded or unbounded delays alike. Yet
t ∗ E ′ ( t ∗ ) − 2 E ( t ∗ ) ≥ 2 − ℓ t ∗ − 2 H t ∗ 3 ( 1 − t ∗ ) 2 = Δ ∗ β ∗ > 0. t_*E'(t_*)-2E(t_*)\ge2-\ell t_*-\frac{2Ht_*^3}{(1-t_*)^2}=\frac{\Delta_*}{\beta_*}>0. t ∗ E ′ ( t ∗ ) − 2 E ( t ∗ ) ≥ 2 − ℓ t ∗ − ( 1 − t ∗ ) 2 2 H t ∗ 3 = β ∗ Δ ∗ > 0.
This contradiction proves every remaining asserted separation. The reference coefficients cancel once both witnesses have departed; retaining this cancellation is what yields the third inequality in (15) .
For the final assertion, a witness that is not such a departure is F F F itself, a positive departure, or a negative departure of degree at most N − 1 N-1 N − 1 or at least j j j . In the disks above, F F F has no zero other than t t t , and rigidity excludes t t t from every other witness. The other cases are the separations just proved, with the markers a N ( F ) > m a_N(F)>m a N ( F ) > m and a j ( F ) > m a_j(F)>m a j ( F ) > m for the adjacent degrees. □
A local criterion and the heights above a pronic number
The next proposition isolates what the separation argument needs from a single series: the comparison inequalities at one double zero, and infinitely many levels at which one coefficient can be lowered and the later ones raised to compensate.
Proposition 4.4 (A local criterion). Let 0 ≤ m < h ∗ ≤ H 0\le m<h_*\le H 0 ≤ m < h ∗ ≤ H be integers, put ℓ = H − m \ell=H-m ℓ = H − m , and let F ∈ F h ∗ F\in\mathcal{F}_{h_*} F ∈ F h ∗ have a double zero at t = 1 / β ∈ ( 0 , σ H ) t=1/\beta\in(0,\sigma_H) t = 1/ β ∈ ( 0 , σ H ) and coefficients a j ( F ) ≥ m a_j(F)\ge m a j ( F ) ≥ m for j ≥ 2 j\ge2 j ≥ 2 . Suppose that the three inequalities (15) hold at β \beta β , and that for some η > 0 \eta>0 η > 0 there are infinitely many indices N N N with
a N ( F ) > m , ∑ j ≥ 1 ( h ∗ − a N + j ( F ) ) t j ≥ 1 + η . (17) a_N(F)>m,\qquad\sum_{j\ge1}(h_*-a_{N+j}(F))t^j\ge1+\eta. \tag*{(17)} a N ( F ) > m , j ≥ 1 ∑ ( h ∗ − a N + j ( F )) t j ≥ 1 + η . (17) Then, at every height h ∗ ≤ h ≤ H h_*\le h\le H h ∗ ≤ h ≤ H , infinitely many distinct connected components of C h C_h C h with nonempty planar interior accumulate at each of the four points ± t \pm t ± t and ± 1 / t \pm1/t ± 1/ t , at which C h C_h C h is not locally connected. These components can be chosen nested as h h h increases, with their diameters at height H H H tending to zero. In particular t ∈ A h ∗ t\in A_{h_*} t ∈ A h ∗ .
Proof. Negation and inversion transfer the statements at t t t to the other three points. Since a j ( F ) ≥ m ≥ 0 a_j(F)\ge m\ge0 a j ( F ) ≥ m ≥ 0 for j ≥ 2 j\ge2 j ≥ 2 and some a N ( F ) a_N(F) a N ( F ) is positive, F ′ ′ ( t ) > 0 F''(t)>0 F ′′ ( t ) > 0 . Hence Lemma 4.3 applies at height H H H to the family consisting of the single pair ( F , t ) (F,t) ( F , t ) : the series F F F is the unique member of F H \mathcal{F}_H F H with a multiple zero at t t t , and t t t has a separation neighbourhood.
Every index N N N in (17) is followed by a least index j N > N j_N>N j N > N with a j N ( F ) > m a_{j_N}(F)>m a j N ( F ) > m , since there are infinitely many such indices. Choose indices N 1 < N 2 < ⋯ N_1<N_2<\cdots N 1 < N 2 < ⋯ satisfying (17) with N i + 1 ≥ j N i N_{i+1}\ge j_{N_i} N i + 1 ≥ j N i , and let B i ⊂ F H \mathcal{B}_i\subset\mathcal{F}_H B i ⊂ F H be the block of negative departures from F F F with degrees in [ N i , j N i − 1 ] [N_i,j_{N_i}-1] [ N i , j N i − 1 ] . These blocks are pairwise disjoint, and by the last assertion of Lemma 4.3 no witness in B i \mathcal{B}_i B i shares a qualifying zero in the separation neighbourhood with a witness outside B i \mathcal{B}_i B i . Lemma 4.2 , applied to the constant sequence ( F , t ) (F,t) ( F , t ) , gives globally clopen subsets K i K_i K i of C H C_H C H . They are pairwise disjoint, because a common point would be a shared qualifying zero of two different blocks.
For occupancy, let G i = F + z N i C i G_i=F+z^{N_i}C_i G i = F + z N i C i be the series that agrees with F F F below degree N i N_i N i , has the digit a N i ( F ) − 1 ≥ m a_{N_i}(F)-1\ge m a N i ( F ) − 1 ≥ m in degree N i N_i N i , and has the constant digit h ∗ h_* h ∗ thereafter. Then G i ∈ B i ∩ F h ∗ G_i\in\mathcal{B}_i\cap\mathcal{F}_{h_*} G i ∈ B i ∩ F h ∗ , the coefficients of C i C_i C i are bounded by h ∗ h_* h ∗ , and
C i ( t ) = − 1 + ∑ j ≥ 1 ( h ∗ − a N i + j ( F ) ) t j ≥ η > 0. C_i(t)=-1+\sum_{j\geq1}(h_* - a_{N_i+j}(F))t^j \geq\eta>0. C i ( t ) = − 1 + j ≥ 1 ∑ ( h ∗ − a N i + j ( F )) t j ≥ η > 0.
Lemma 4.1 provides nonreal zeros of G i G_i G i in C h ∗ C_{h_*} C h ∗ tending to t t t ; they lie in K i K_i K i for all large i i i . Part (i) of Lemma 4.2 gives the components, their nesting and the failure of local connectedness, and t ∈ A h ∗ t\in A_{h_*} t ∈ A h ∗ follows from the accumulation.
Corollary 4.5 (Components accumulating at a central anchor). For every integer p ≥ 4 p\geq4 p ≥ 4 and
p ( p − 1 ) + 1 ≤ h ≤ p 2 − 3 , p(p-1)+1\leq h\leq p^2-3, p ( p − 1 ) + 1 ≤ h ≤ p 2 − 3 , infinitely many distinct global components of C h C_h C h with nonempty planar interior accumulate at each of ± p , ± 1 / p \pm p,\pm1/p ± p , ± 1/ p . They can be chosen nested across this entire height interval, with their diameters at its largest height tending to zero. The four anchors belong to the component containing the unit circle, and the cores are not locally connected there.
Proof. Apply Proposition 4.4 to
F ( z ) = ( 1 − p z ) 2 1 − z = 1 − ( 2 p − 1 ) z + ( p − 1 ) 2 ∑ j ≥ 2 z j , t = 1 / p , F(z)=\frac{(1-pz)^2}{1-z}=1-(2p-1)z+(p-1)^2\sum_{j\geq2}z^j,\qquad t=1/p, F ( z ) = 1 − z ( 1 − p z ) 2 = 1 − ( 2 p − 1 ) z + ( p − 1 ) 2 j ≥ 2 ∑ z j , t = 1/ p ,
with h ∗ = p ( p − 1 ) + 1 h_*=p(p-1)+1 h ∗ = p ( p − 1 ) + 1 , H = p 2 − 3 H=p^2-3 H = p 2 − 3 and m = ( p − 1 ) 2 − 1 m=(p-1)^2-1 m = ( p − 1 ) 2 − 1 , so that ℓ = 2 p − 3 \ell=2p-3 ℓ = 2 p − 3 . The three positive margins are
2 − ℓ p − 1 = 1 p − 1 , 1 − ℓ ( p − 1 ) 2 = ( p − 2 ) 2 ( p − 1 ) 2 , 2 p − ℓ − 2 H ( p − 1 ) 2 = ( p − 3 ) 2 ( p − 1 ) 2 , 2-\frac{\ell}{p-1}=\frac{1}{p-1},\qquad1-\frac{\ell}{(p-1)^2}=\frac{(p-2)^2}{(p-1)^2},\qquad2p-\ell-\frac{2H}{(p-1)^2}=\frac{(p-3)^2}{(p-1)^2}, 2 − p − 1 ℓ = p − 1 1 , 1 − ( p − 1 ) 2 ℓ = ( p − 1 ) 2 ( p − 2 ) 2 , 2 p − ℓ − ( p − 1 ) 2 2 H = ( p − 1 ) 2 ( p − 3 ) 2 ,
and t < σ H t<\sigma_H t < σ H since p 4 − H 2 − 4 = 6 p 2 − 13 > 0 p^4-H^2-4=6p^2-13>0 p 4 − H 2 − 4 = 6 p 2 − 13 > 0 . Every coefficient of F F F of degree at least two exceeds m m m , and for every N N N the sum in (17) equals ( h ∗ − ( p − 1 ) 2 ) / ( p − 1 ) = 1 + 1 / ( p − 1 ) (h_*-(p-1)^2)/(p-1)=1+1/(p-1) ( h ∗ − ( p − 1 ) 2 ) / ( p − 1 ) = 1 + 1/ ( p − 1 ) . The connected core at height p ( p − 1 ) p(p-1) p ( p − 1 ) contains the four anchors and the unit circle by Proposition 3.3 ; inclusion in each larger core puts them in its component containing the unit circle.
With an arbitrary height h h h in place of h ∗ h_* h ∗ , the sum in (17) equals ( h − ( p − 1 ) 2 ) / ( p − 1 ) (h-(p-1)^2)/(p-1) ( h − ( p − 1 ) 2 ) / ( p − 1 ) , which exceeds 1 exactly when h > p ( p − 1 ) h>p(p-1) h > p ( p − 1 ) . At the pronic height it equals 1 and the criterion just fails, as Theorem A requires.
Singleton components below a pronic height
At the height p ( p − 1 ) − 1 p(p-1)-1 p ( p − 1 ) − 1 we use nearby reference series. A finite negative block for each reference will contain the same limiting anchor. This produces shrinking clopen neighbourhoods of that anchor without any recurrence assumption on a single coefficient sequence. The references come from an intermediate-value argument, which § 6 uses again.
For integers 0 ≤ m ′ < H ′ 0\leq m'<H' 0 ≤ m ′ < H ′ and a finite integer sequence π = ( a 0 , … , a N − 1 ) \pi=(a_0,\ldots,a_{N-1}) π = ( a 0 , … , a N − 1 ) with N ≥ 2 N\geq2 N ≥ 2 , write
f π − ( z ) = ∑ j < N a j z j + m ′ z N 1 − z , f π + ( z ) = ∑ j < N a j z j + H ′ z N 1 − z , f_\pi^-(z)=\sum_{j<N}a_jz^j+\frac{m'z^N}{1-z},\qquad f_\pi^+(z)=\sum_{j<N}a_jz^j+\frac{H'z^N}{1-z}, f π − ( z ) = j < N ∑ a j z j + 1 − z m ′ z N , f π + ( z ) = j < N ∑ a j z j + 1 − z H ′ z N ,
for the smallest and the largest of the completions ∑ j < N a j z j + ∑ j ≥ N a j z j \sum_{j<N}a_jz^j+\sum_{j\geq N}a_jz^j ∑ j < N a j z j + ∑ j ≥ N a j z j of π \pi π with integer digits a j ∈ [ m ′ , H ′ ] a_j\in[m',H'] a j ∈ [ m ′ , H ′ ] for j ≥ N j\geq N j ≥ N .
Lemma 4.6 (Completion with a prescribed minimum). Let J = [ a , b ] ⊂ ( 0 , 1 ) J=[a,b]\subset(0,1) J = [ a , b ] ⊂ ( 0 , 1 ) satisfy ( H ′ − m ′ ) a ≥ 1 − a (H'-m')a\geq1-a ( H ′ − m ′ ) a ≥ 1 − a , and suppose that
f π − ( a ) ≥ 0 , f π − ( b ) ≥ 0 , min J f π − < 0 ≤ min J f π + . f_\pi^-(a)\geq0,\qquad f_\pi^-(b)\geq0,\qquad\min_J f_\pi^-<0\leq\min_J f_\pi^+. f π − ( a ) ≥ 0 , f π − ( b ) ≥ 0 , J min f π − < 0 ≤ J min f π + . Then π \pi π has a completion F F F with min J F = 0 \min_J F=0 min J F = 0 , and the minimum is attained at an interior point t t t of J J J . If moreover a j ≥ 0 a_j\geq0 a j ≥ 0 for 2 ≤ j < N 2\leq j<N 2 ≤ j < N , then F ′ ′ > 0 F''>0 F ′′ > 0 on ( 0 , 1 ) (0,1) ( 0 , 1 ) , and t t t is the only zero of F F F in ( 0 , 1 ) (0,1) ( 0 , 1 ) and is exactly double.
Proof. Let π ′ \pi' π ′ be an extension of π \pi π by digits in [ m ′ , H ′ ] [m',H'] [ m ′ , H ′ ] , with next free degree n ≥ N n \ge N n ≥ N , and let d < H ′ d < H' d < H ′ be a digit. On J J J ,
f π ′ d + − f π ′ ( d + 1 ) − = z n ( ( H ′ − m ′ ) z 1 − z − 1 ) ≥ 0 , f^+_{\pi'd}-f^-_{\pi'(d+1)}=z^n\left((H'-m')\frac{z}{1-z}-1\right)\ge0, f π ′ d + − f π ′ ( d + 1 ) − = z n ( ( H ′ − m ′ ) 1 − z z − 1 ) ≥ 0 ,
by the hypothesis on a a a and monotonicity of z / ( 1 − z ) z/(1-z) z / ( 1 − z ) . Therefore the intervals [ min J f π ′ d − , min J f π ′ d + ] [\min_J f^-_{\pi'd},\min_J f^+_{\pi'd}] [ min J f π ′ d − , min J f π ′ d + ] , for d = m ′ , … , H ′ d=m',\ldots,H' d = m ′ , … , H ′ , overlap consecutively, and their union is the interval [ min J f π ′ − , min J f π ′ + ] [\min_J f^-_{\pi'},\min_J f^+_{\pi'}] [ min J f π ′ − , min J f π ′ + ] , since the extreme completions of π ′ \pi' π ′ coincide with those of its extreme children. Starting from π \pi π , choose successively a child whose interval contains 0. The resulting completions converge uniformly on J J J with all derivatives, since the tails are bounded by H ′ b n / ( 1 − b ) H'b^n/(1-b) H ′ b n / ( 1 − b ) , and min J F = 0 \min_J F=0 min J F = 0 for the limit F F F . F F F is not the completion f π − f^-_\pi f π − , whose minimum on J J J is negative. Hence some digit of F F F beyond π \pi π exceeds m ′ m' m ′ , and F − f π − = ∑ j ≥ N ( a j ( F ) − m ′ ) z j F-f^-_\pi=\sum_{j\ge N}(a_j(F)-m')z^j F − f π − = ∑ j ≥ N ( a j ( F ) − m ′ ) z j is strictly positive on ( 0 , 1 ) (0,1) ( 0 , 1 ) . Thus F ( a ) > 0 F(a)>0 F ( a ) > 0 and F ( b ) > 0 F(b)>0 F ( b ) > 0 , and the minimum 0 of F F F on J J J is attained at an interior point t t t , where F ′ ( t ) = 0 F'(t)=0 F ′ ( t ) = 0 . Under the last hypothesis the coefficients of F F F of degree at least two are nonnegative and one of them is positive, so F ′ ′ > 0 F''>0 F ′′ > 0 on ( 0 , 1 ) (0,1) ( 0 , 1 ) ; then t t t is the unique zero of F F F there, and the zero is exactly double. □
Theorem 4.7 (Below a pronic height). For every integer p ≥ 3 p \ge3 p ≥ 3 and
( p − 1 ) 2 + 1 ≤ h ≤ p ( p − 1 ) − 1 , (p-1)^2+1\le h\le p(p-1)-1, ( p − 1 ) 2 + 1 ≤ h ≤ p ( p − 1 ) − 1 , the four points ± p \pm p ± p and ± 1 / p \pm1/p ± 1/ p are singleton global components of C h \mathcal{C}_h C h . Every neighbourhood of each of them contains infinitely many distinct global components with nonempty planar interior. These components can be chosen nested as h h h increases through the stated interval, with their diameters at its largest height tending to zero. None of these cores is locally connected at the four anchors.
Proof. Put
H = p ( p − 1 ) − 1 , A = ( p − 1 ) 2 , H 0 = A + 1 , m = ( p − 1 ) ( p − 2 ) , H=p(p-1)-1,\qquad A=(p-1)^2,\qquad H_0=A+1,\qquad m=(p-1)(p-2), H = p ( p − 1 ) − 1 , A = ( p − 1 ) 2 , H 0 = A + 1 , m = ( p − 1 ) ( p − 2 ) ,
t 0 = 1 / p , F 0 ( z ) = ( 1 − p z ) 2 1 − z . t_0=1/p,\qquad F_0(z)=\frac{(1-pz)^2}{1-z}. t 0 = 1/ p , F 0 ( z ) = 1 − z ( 1 − p z ) 2 .
We construct references with a double zero t N > t 0 t_N>t_0 t N > t 0 tending to t 0 t_0 t 0 , all admissible already at height H 0 H_0 H 0 ; note that 2 p − 1 ≤ H 0 ≤ H 2p-1\le H_0\le H 2 p − 1 ≤ H 0 ≤ H for p ≥ 3 p\ge3 p ≥ 3 . For sufficiently large N N N , let r N r_N r N be the first root to the right of t 0 t_0 t 0 of 1 − p z + z N 1-pz+z^N 1 − p z + z N . This root exists and tends to t 0 t_0 t 0 , since 1 − p z + z N 1-pz+z^N 1 − p z + z N converges with its derivative to 1 − p z 1-pz 1 − p z on a fixed small interval about t 0 t_0 t 0 , and its value at t 0 t_0 t 0 is positive. Set I N = [ t 0 , r N ] I_N=[t_0,r_N] I N = [ t 0 , r N ] and consider the integer-tail completions
F N ( z ) = 1 − ( 2 p − 1 ) z + A ∑ j = 2 N − 1 z j + ( A + 1 ) z N + ∑ j > N a j z j , m ≤ a j ≤ A . (18) F_N(z)=1-(2p-1)z+A\sum_{j=2}^{N-1}z^j+(A+1)z^N+\sum_{j>N}a_jz^j,\qquad m\le a_j\le A. \tag*{(18)} F N ( z ) = 1 − ( 2 p − 1 ) z + A j = 2 ∑ N − 1 z j + ( A + 1 ) z N + j > N ∑ a j z j , m ≤ a j ≤ A . (18)
The all-m m m and all-A A A completions are respectively
( 1 − p z ) ( 1 − p z + z N ) 1 − z , F 0 ( z ) + z N . \frac{(1-pz)(1-pz+z^N)}{1-z},\qquad F_0(z)+z^N. 1 − z ( 1 − p z ) ( 1 − p z + z N ) , F 0 ( z ) + z N .
The first vanishes at both endpoints of I N I_N I N and is negative between them, while the second is everywhere positive. Since ( A − m ) t 0 = 1 − t 0 (A-m)t_0=1-t_0 ( A − m ) t 0 = 1 − t 0 , Lemma 4.6 with the digit range [ m , A ] [m,A] [ m , A ] gives a completion F N F_N F N of the form (4.5) whose minimum on I N I_N I N is zero. It is attained at an interior point t N t_N t N , which is the unique zero of F N F_N F N in ( 0 , 1 ) (0,1) ( 0 , 1 ) and is exactly double, and F N ′ ′ > 0 F_N''>0 F N ′′ > 0 there. We have t N → t 0 t_N\to t_0 t N → t 0 and F N → F 0 F_N\to F_0 F N → F 0 locally uniformly.
The compact comparison lemma applies to this family after discarding finitely many initial N N N , together with its limit ( F 0 , t 0 ) (F_0,t_0) ( F 0 , t 0 ) . All coefficients of degree at least two belong to [ m , H ] [m,H] [ m , H ] , including A + 1 ≤ H A+1\le H A + 1 ≤ H . At the limit, ℓ = H − m = 2 p − 3 \ell=H-m=2p-3 ℓ = H − m = 2 p − 3 , and the three margins are
2 − ℓ p − 1 = 1 p − 1 , 1 − ℓ ( p − 1 ) 2 = ( p − 2 ) 2 ( p − 1 ) 2 , 2 p − ℓ − ( p − 2 ) 2 ( p − 1 ) 2 = 2 H ( p − 1 ) 2 . 2-\frac{\ell}{p-1}=\frac{1}{p-1},\qquad1-\frac{\ell}{(p-1)^2}=\frac{(p-2)^2}{(p-1)^2},\qquad2p-\ell-\frac{(p-2)^2}{(p-1)^2}=\frac{2H}{(p-1)^2}. 2 − p − 1 ℓ = p − 1 1 , 1 − ( p − 1 ) 2 ℓ = ( p − 1 ) 2 ( p − 2 ) 2 , 2 p − ℓ − ( p − 1 ) 2 ( p − 2 ) 2 = ( p − 1 ) 2 2 H .
They are strict for every p ≥ 3 p\ge3 p ≥ 3 and persist uniformly by continuity. Also p 2 > H + 1 p^2>H+1 p 2 > H + 1 gives p 4 > H 2 + 4 p^4>H^2+4 p 4 > H 2 + 4 , so a common normalization disk is available. We therefore have full-height rigidity and one common separation neighbourhood for every reference, including F 0 F_0 F 0 . Let j N > N j_N>N j N > N be the first later index with a j N ( F N ) > m a_{j_N}(F_N)>m a j N ( F N ) > m . It exists because F N F_N F N is not the all-m m m completion, and a N ( F N ) = A + 1 > m a_N(F_N)=A+1>m a N ( F N ) = A + 1 > m . Let B N \mathcal{B}_N B N be the block of negative departures from F N F_N F N with degrees in [ N , j N − 1 ] [N,j_N-1] [ N , j N − 1 ] . By the last assertion of Lemma 4.3, all hypotheses of Lemma 4.2 are satisfied. The resulting globally clopen set K N K_N K N contains t 0 t_0 t 0 : the witness F 0 F_0 F 0 first departs negatively from F N F_N F N at degree N N N , and t 0 t_0 t 0 belongs to C H 0 C_{H_0} C H 0 by the admissible perturbations F 0 + z k F_0+z^k F 0 + z k . These same perturbations and Lemma 4.1 show that t 0 t_0 t 0 is nonisolated already at height H 0 H_0 H 0 . The diameters of K N K_N K N tend to zero, regardless of the possibly unbounded lengths of the selected finite blocks. Part (ii) of the transfer lemma now gives the singleton, interior, nesting and local connectedness conclusions at t 0 t_0 t 0 . Negation and inversion give the other three anchors. □
Proof of Theorem B. Corollary 4.5 and Theorem 4.7 cover the two sides of the stated interval; when p = 3 p=3 p = 3 its part above p ( p − 1 ) p(p-1) p ( p − 1 ) is empty. □
Above the pronic height the components accumulate on a component already present at that height; below it, at a singleton component that remains separated until the next pronic height. Both conclusions rest on globally clopen sets, so no reconnection is possible outside the disk used in the proof.
Sparse coefficient alphabets
Separation in the full height-H H H core also controls root sets obtained by deleting coefficient choices. The following consequence needs only three or four nonconstant digits to produce the accumulating components. For a finite integer alphabet S S S , define
C ( S ) = { z ∈ C ∖ R : 1 + ∑ j = 1 d a j z j = 0 , d ≥ 1 , a j ∈ S , a d ≠ 0 } . C(S)=\left\{z\in\mathbb{C}\setminus\mathbb{R}:1+\sum_{j=1}^{d}a_jz^j=0,\quad d\geq1,\quad a_j\in S,\quad a_d\ne0\right\}. C ( S ) = { z ∈ C ∖ R : 1 + j = 1 ∑ d a j z j = 0 , d ≥ 1 , a j ∈ S , a d = 0 } .
The constant coefficient is fixed at 1 1 1 ; it need not belong to S S S .
Corollary 4.8 (Persistence under changes of alphabet). Put A = ( p − 1 ) 2 A=(p-1)^2 A = ( p − 1 ) 2 and t = 1 / p t=1/p t = 1/ p .
(i) Let p ≥ 3 p\geq3 p ≥ 3 , H = p ( p − 1 ) − 1 H=p(p-1)-1 H = p ( p − 1 ) − 1 , and
S p − = { − ( 2 p − 1 ) , A , A + 1 } ⊆ S ⊆ [ − H , H ] ∩ Z . S_p^-=\{-(2p-1),A,A+1\}\subseteq S\subseteq[-H,H]\cap\mathbb{Z}. S p − = { − ( 2 p − 1 ) , A , A + 1 } ⊆ S ⊆ [ − H , H ] ∩ Z . Then { t } \{t\} { t } is a component of C ( S ) C(S) C ( S ) .
(ii) Let p ≥ 4 p\geq4 p ≥ 4 , H = p 2 − 3 H=p^2-3 H = p 2 − 3 , and
S p + = { − ( 2 p − 1 ) , A − 1 , A , p ( p − 1 ) + 1 } ⊆ S ⊆ [ − H , H ] ∩ Z . S_p^+=\{-(2p-1),A-1,A,p(p-1)+1\}\subseteq S\subseteq[-H,H]\cap\mathbb{Z}. S p + = { − ( 2 p − 1 ) , A − 1 , A , p ( p − 1 ) + 1 } ⊆ S ⊆ [ − H , H ] ∩ Z . Then C ( S ) C(S) C ( S ) has infinitely many components accumulating at t t t .
In both cases every neighbourhood of t t t contains infinitely many components containing nonreal points, and C ( S ) C(S) C ( S ) is not locally connected at t t t . These components can be chosen to persist under every allowed enlargement of S S S : if S ⊆ T S\subseteq T S ⊆ T , the component chosen for S S S lies in the corresponding component for T T T . Their diameters tend to zero uniformly over all the allowed alphabets.
Proof. For either value of H H H , the set C ( S ) C(S) C ( S ) is a compact subset of C H C_H C H , and it increases with S S S . Thus the intersection of a globally clopen subset of C H C_H C H with C ( S ) C(S) C ( S ) is clopen in C ( S ) C(S) C ( S ) .
For (ii), use the disjoint shrinking clopen sets constructed in the proof of Proposition 4.4 for the series of Corollary 4.5; write K N K_N K N for the set whose block is the degree N N N . Their occupying witnesses are
( 1 − p z ) 2 1 − z + z N ( − 1 + p z 1 − z ) . \frac{(1-pz)^2}{1-z}+z^N\left(-1+\frac{pz}{1-z}\right). 1 − z ( 1 − p z ) 2 + z N ( − 1 + 1 − z p z ) .
Their nonconstant coefficients are − ( 2 p − 1 ) -(2p-1) − ( 2 p − 1 ) in degree one, A A A in degrees 2 , … , N − 1 2,\ldots,N-1 2 , … , N − 1 , A − 1 A-1 A − 1 in degree N N N , and p ( p − 1 ) + 1 p(p-1)+1 p ( p − 1 ) + 1 thereafter. Their nearby nonreal zeros therefore belong to C ( S p + ) C(S_p^+) C ( S p + ) : truncate the series and apply Rouché’s theorem in a small disk disjoint from the real axis. Choose one such zero z N ∈ K N z_N\in K_N z N ∈ K N . For every allowed S S S , the component of z N z_N z N in C ( S ) C(S) C ( S ) remains in K N K_N K N . These components are distinct, shrink uniformly to t t t , and are nested when the alphabet increases. Their limit t t t belongs to C ( S p + ) C(S_p^+) C ( S p + ) . Since t ∉ K N t\notin K_N t ∈ / K N , no connected relatively open neighbourhood of t t t can meet these sets, proving failure of local connectedness.
For (i), the proof of Theorem 4.7 provides shrinking globally clopen neighbourhoods K N K_N K N of t t t . The witnesses ( 1 − p z ) 2 / ( 1 − z ) + z k (1-pz)^2/(1-z)+z^k ( 1 − p z ) 2 / ( 1 − z ) + z k use only S p − S_p^- S p − , so their nonreal zeros show that t t t is nonisolated in C ( S p − ) C(S_p^{-}) C ( S p − ) . The intersection of the K N K_N K N is { t } \{t\} { t } , proving the singleton assertion for every allowed S S S . Successive finite intersections V j V_j V j of the K N K_N K N can be chosen with diameters tending to zero so that each difference V j ∖ V j + 1 V_j\setminus V_{j+1} V j ∖ V j + 1 contains a nonreal point of C ( S p − ) C(S_p^{-}) C ( S p − ) . Select the component through that point at each alphabet. The same clopen argument proves all the remaining claims, including failure of local connectedness at the nonisolated singleton. □
For example, the alphabet { − 5 , 4 , 5 } \{-5,4,5\} { − 5 , 4 , 5 } gives a singleton component at 1 / 3 1/3 1/3 , and { − 7 , 8 , 9 , 13 } \{-7,8,9,13\} { − 7 , 8 , 9 , 13 } gives infinitely many components accumulating at 1 / 4 1/4 1/4 . No density of planar interior is needed for these conclusions.
Completing the connectedness classification
The preceding section proves more than disconnectedness at every height in Theorem B. The only heights still needed for Theorem A are the three heights next to each square, together with a few small cases. We treat these by planar separators. This avoids repeating the more general separation constructions on the ranges already covered by infinitude. Section 6 proves infinitely many components at most of the square-adjacent heights; the separators below cover all of them uniformly and remain necessary at the heights not reached there.
What the finite exclusions certify
For c = x + i y c=x+iy c = x + i y , x > 1 x>1 x > 1 , put R = ∣ c ∣ 2 R=|c|^2 R = ∣ c ∣ 2 and
T h ( c ) = { ∑ j ≥ 0 d j c − j : d j ∈ B h } . T_h(c)=\left\{\sum_{j\geq0}d_jc^{-j}:d_j\in B_h\right\}. T h ( c ) = { j ≥ 0 ∑ d j c − j : d j ∈ B h } .
The normalized series criterion says c ∈ Σ h c\in\Sigma_h c ∈ Σ h if and only if c ∈ T h ( c ) c\in T_h(c) c ∈ T h ( c ) . Its inverse-digit orbit is z ↦ c ( z − d ) z\mapsto c(z-d) z ↦ c ( z − d ) . For y ≠ 0 y\ne0 y = 0 , use the coordinates
V = Im z y , S = Im ( c z ) y . V=\frac{\operatorname{Im}z}{y},\qquad S=\frac{\operatorname{Im}(cz)}{y}. V = y Im z , S = y Im ( cz ) .
Starting from z 0 = c z_0=c z 0 = c , they satisfy
( V 0 , S 0 ) = ( 1 , 2 x ) , V j + 1 = S j − d j + 1 , S j + 1 = 2 x V j + 1 − R V j . (19) (V_0,S_0)=(1,2x),\qquad V_{j+1}=S_j-d_{j+1},\qquad S_{j+1}=2xV_{j+1}-RV_j. \tag*{(19)} ( V 0 , S 0 ) = ( 1 , 2 x ) , V j + 1 = S j − d j + 1 , S j + 1 = 2 x V j + 1 − R V j . (19)
For a fixed prefix these are polynomials in x , R x,R x , R , so the formulas continue to y = 0 y=0 y = 0 . There they agree with (3.3) under S j = x ( P j + V j ) S_j=x(P_j+V_j) S j = x ( P j + V j ) .
Every point of T h ( c ) T_h(c) T h ( c ) obeys the necessary bounds
∣ V ∣ ≤ h ( ∣ c ∣ − 1 ) 2 , ∣ S ∣ ≤ h + h ( ∣ c ∣ − 1 ) 2 . (20) |V|\leq\frac{h}{(|c|-1)^2},\qquad|S|\leq h+\frac{h}{(|c|-1)^2}. \tag*{(20)} ∣ V ∣ ≤ ( ∣ c ∣ − 1 ) 2 h , ∣ S ∣ ≤ h + ( ∣ c ∣ − 1 ) 2 h . (20)
Indeed ∣ Im ( c − k ) / y ∣ ≤ k ∣ c ∣ − k − 1 |\operatorname{Im}(c^{-k})/y|\leq k|c|^{-k-1} ∣ Im ( c − k ) / y ∣ ≤ k ∣ c ∣ − k − 1 ; summing over the digits gives the first bound, and multiplying by c c c gives the second. Thus every possible next digit lies in a finite interval. A certificate retains all of them and proves that every resulting branch eventually violates (5.2).
Strict exclusion persists on a parameter neighbourhood. It therefore excludes a point from C h C_h C h , including when the point is real: the real full locus is present there, but no nearby nonreal point can belong to it. This is why these tests can certify a separator crossing the real axis.
Proposition 5.1 (Residual planar separation). The exterior core C h e x t C_h^{\mathrm{ext}} C h ext has at least three connected components whenever
q 2 − 2 ≤ h ≤ q 2 ( q ≥ 3 ) , or h ∈ { 1 , 3 , 4 } . q^2-2\leq h\leq q^2\quad(q\geq3),\qquad\text{or}\qquad h\in\{1,3,4\}. q 2 − 2 ≤ h ≤ q 2 ( q ≥ 3 ) , or h ∈ { 1 , 3 , 4 } . The proofs and exact certificate statements are in Sections S3–S4 of the supplement. They consist of two ingredients: a curve disjoint from the core and a core point in each of its separated regions. For a bounded Jordan domain U U U with ∂ U ∩ C h = ∅ \partial U\cap C_h=\varnothing ∂ U ∩ C h = ∅ , the set U ∩ C h U\cap C_h U ∩ C h is relatively clopen. A witness in U U U , its negative in − U -U − U , and the central annulus then give three components. The same reasoning applies to two reflected proper curves on the sphere. No path-connectedness assertion about the core is needed.
For large q q q the boundary inequalities are uniform in q q q . The remaining parameter intervals are covered by rational subintervals. After clearing positive denominators, an inequality is verified by writing its polynomial in Bernstein form on the entire closed interval; strictly positive coefficients
certify positivity. The covers include all joins, and the digit recursion keeps every admissible branch. Interior-entry inequalities or a Rouché enclosure certify the witnesses. The exact-arithmetic checks of the companion [13 ] verify these assertions, rather than sampling the displayed curves.
Figure 5. The two parts of a planar separation certificate at height h = 16 h = 16 h = 16 . In coordinates u = ( Re c − 4.531 ) / 0.009 u = (\operatorname{Re} c - 4.531)/0.009 u = ( Re c − 4.531 ) /0.009 , v = 10 Im c v = 10\operatorname{Im} c v = 10 Im c , the domain is ∣ v ∣ < 1 |v| < 1 ∣ v ∣ < 1 , ( 4 / 9 ) v 2 < u < 1 (4/9)v^2 < u < 1 ( 4/9 ) v 2 < u < 1 . Its blue boundary is disjoint from the core, while the orange point c ∗ = 4.5327 c_* = 4.5327 c ∗ = 4.5327 belongs to its interior and to the planar interior of the root locus. Reflection supplies a second separated region. All decimals specify exact rational numbers. Pale shading is a depth-48 numerical illustration; the certificate, not the shading, proves separation.
Passing from the exterior to the whole plane
Inversion doubles each component away from the central annulus. The following elementary fact makes this precise.
Lemma 5.2 (Reciprocal gluing). Let K K K be bounded and relatively closed in X X X , invariant under conjugation, and contain { 1 < ∣ z ∣ < b } \{1 < |z| < b\} { 1 < ∣ z ∣ < b } for some b > 1 b > 1 b > 1 . Let K 0 K_0 K 0 be the component containing this annulus. The components of K ∪ T ∪ ι ( K ) K \cup\mathbb{T} \cup\iota(K) K ∪ T ∪ ι ( K ) are
K 0 ∪ T ∪ ι ( K 0 ) , B , ι ( B ) ( B a component of K , B ≠ K 0 ) . K_0 \cup\mathbb{T} \cup\iota(K_0), \qquad B,\ \iota(B) \quad(B\ \text{a component of } K,\ B \ne K_0). K 0 ∪ T ∪ ι ( K 0 ) , B , ι ( B ) ( B a component of K , B = K 0 ) . In particular, connectedness is preserved; m < ∞ m < \infty m < ∞ exterior components give 2 m − 1 2m - 1 2 m − 1 components in the whole plane.
Proof. Fix 1 < r 0 < b 1 < r_0 < b 1 < r 0 < b . The radial deformation
z ↦ ( ( 1 − t ) ∣ z ∣ + t max { ∣ z ∣ , r 0 } ) z ∣ z ∣ ( 0 ≤ t ≤ 1 ) z \mapsto\left((1-t)|z| + t\max\{|z|,r_0\}\right)\frac{z}{|z|} \qquad(0 \le t \le1) z ↦ ( ( 1 − t ) ∣ z ∣ + t max { ∣ z ∣ , r 0 } ) ∣ z ∣ z ( 0 ≤ t ≤ 1 )
retracts both K K K and K ∪ T K \cup\mathbb{T} K ∪ T onto K ∩ { ∣ z ∣ ≥ r 0 } K \cap\{|z| \ge r_0\} K ∩ { ∣ z ∣ ≥ r 0 } . Its trajectories stay in the annulus or fix their starting point, so adjoining T \mathbb{T} T changes only K 0 K_0 K 0 . Every other component lies in { ∣ z ∣ ≥ b } \{|z| \ge b\} { ∣ z ∣ ≥ b } . The continuous fold that fixes ∣ z ∣ ≥ 1 |z| \ge1 ∣ z ∣ ≥ 1 and sends z z z to 1 / z ‾ 1/\overline{z} 1/ z on ∣ z ∣ ≤ 1 |z| \le1 ∣ z ∣ ≤ 1 maps a connected subset into one component of K ∪ T K \cup\mathbb{T} K ∪ T . The preimage of the central component is connected through the annulus. The preimage of any other component has an exterior part in ∣ z ∣ ≥ b |z| \ge b ∣ z ∣ ≥ b and an interior part in ∣ z ∣ ≤ b − 1 |z| \le b^{-1} ∣ z ∣ ≤ b − 1 ; they are separated. Conjugation permutes the components of K K K , giving the stated list. □
Proof of Theorem A. Proposition 3.3 proves connectedness and local connectedness at every h = q ( q − 1 ) h=q(q-1) h = q ( q − 1 ) . At all other heights covered by Theorem B, there are infinitely many components. The remaining nonpronic heights are exactly
{ 1 , 3 , 4 } ∪ ⋃ q ≥ 3 { q 2 − 2 , q 2 − 1 , q 2 } . \{1,3,4\}\cup\bigcup_{q\geq3}\{q^2-2,q^2-1,q^2\}. { 1 , 3 , 4 } ∪ q ≥ 3 ⋃ { q 2 − 2 , q 2 − 1 , q 2 } .
Proposition 5.1 gives at least three exterior components at each of them. Apply the gluing lemma with b = h + 1 b=\sqrt{h+1} b = h + 1 and (2.3). The whole-plane core then has at least five components. The assertion about the full nonzero closure was proved in Section 2. □ \square □
The square-adjacent heights
Theorem B leaves open the three heights next to each square, q 2 − 2 q^2-2 q 2 − 2 , q 2 − 1 q^2-1 q 2 − 1 and q 2 q^2 q 2 . There the mechanism of Section 4 cannot be anchored at 1 / q 1/q 1/ q : the comparison margin 2 − ℓ / ( β − 1 ) 2-\ell/(\beta-1) 2 − ℓ / ( β − 1 ) of Lemma 4.3 is not positive at β = q \beta=q β = q for the digit range that the construction needs. Indeed q q q itself soon becomes an interior point of the core: for q ≥ 4 q\geq4 q ≥ 4 this first happens at one of the three heights, since the core is not locally connected at q q q at height q 2 − 3 q^2-3 q 2 − 3 (Theorem B) and contains a neighbourhood of q q q at height q 2 q^2 q 2 by (2.5); which of the three it is remains open (Figure 6). Nor can the mechanism be anchored at 1 / ( q + 1 ) 1/(q+1) 1/ ( q + 1 ) : the saturated series ( 1 − ( q + 1 ) z ) 2 / ( 1 − z ) (1-(q+1)z)^2/(1-z) ( 1 − ( q + 1 ) z ) 2 / ( 1 − z ) is admissible only from height q 2 q^2 q 2 on, and its double zero is inactive there (Proposition 3.4).
We instead use series with a double zero strictly between 1 / ( q + 1 ) 1/(q+1) 1/ ( q + 1 ) and 1 / q 1/q 1/ q , the square-adjacent witnesses defined below. Theorem 6.7 shows that one witness outside three exceptional types suffices for infinitely many components, and Theorem 6.12 finds such a witness common to at least two of the three heights. In the explicit examples of Propositions 6.14 and 6.15 the reciprocal anchor is an algebraic integer of degree at most nine.
Throughout the section, q ≥ 4 q\geq4 q ≥ 4 and k ∈ { 0 , 1 , 2 } k\in\{0,1,2\} k ∈ { 0 , 1 , 2 } are fixed, and
h = q 2 − 2 + k , m = q 2 − 2 q + k , W = h − m = 2 q − 2. (21) h=q^2-2+k,\qquad m=q^2-2q+k,\qquad W=h-m=2q-2. \tag*{(21)} h = q 2 − 2 + k , m = q 2 − 2 q + k , W = h − m = 2 q − 2. (21)
Thus m = ( q − 1 ) 2 − 1 + k m=(q-1)^2-1+k m = ( q − 1 ) 2 − 1 + k ; the three values of k k k give the three heights next to the square q 2 q^2 q 2 .
The square-adjacent family and its anchors
Definition 6.1. The square-adjacent family at height h h h is
Φ h = { 1 − 2 q z + ∑ j ≥ 2 a j z j : a j ∈ Z , m ≤ a j ≤ h } ⊂ F h . \Phi_h=\left\{1-2qz+\sum_{j\geq2}a_jz^j:a_j\in\mathbb{Z},\ m\leq a_j\leq h\right\}\subset\mathcal{F}_h. Φ h = { 1 − 2 q z + j ≥ 2 ∑ a j z j : a j ∈ Z , m ≤ a j ≤ h } ⊂ F h .
A square-adjacent witness is a member of Φ h \Phi_h Φ h with a real double zero in ( 0 , 1 ) (0,1) ( 0 , 1 ) . Its double zero is called its anchor . The set of all anchors of square-adjacent witnesses is the anchor set T h ⊂ ( 0 , 1 ) \mathcal{T}_h\subset(0,1) T h ⊂ ( 0 , 1 ) .
Every F ∈ Φ h F\in\Phi_h F ∈ Φ h satisfies F ′ ′ ( z ) = ∑ j ≥ 2 j ( j − 1 ) a j z j − 2 > 0 F^{\prime\prime}(z)=\sum_{j\geq2}j(j-1)a_jz^{j-2}>0 F ′′ ( z ) = ∑ j ≥ 2 j ( j − 1 ) a j z j − 2 > 0 on ( 0 , 1 ) (0,1) ( 0 , 1 ) . Hence F F F is strictly convex there, has at most one double zero, and a square-adjacent witness is nonnegative on ( 0 , 1 ) (0,1) ( 0 , 1 ) with its anchor as the unique zero and the unique minimum point. The lowest member of the family,
M m ( z ) = 1 − 2 q z + m z 2 1 − z = 1 − ( 2 q + 1 ) z + ( q 2 + k ) z 2 1 − z , (22) M_m(z)=1-2qz+\frac{mz^2}{1-z}=\frac{1-(2q+1)z+(q^2+k)z^2}{1-z}, \tag*{(22)} M m ( z ) = 1 − 2 q z + 1 − z m z 2 = 1 − z 1 − ( 2 q + 1 ) z + ( q 2 + k ) z 2 , (22)
is negative exactly on the interval I 0 = ( t − , t + ) I_0=(t_-,t_+) I 0 = ( t − , t + ) with
t ± = ( 2 q + 1 ) ± 4 q + 1 − 4 k 2 ( q 2 + k ) . (23) t_{\pm}=\frac{(2q+1)\pm\sqrt{4q+1-4k}}{2(q^2+k)}. \tag*{(23)} t ± = 2 ( q 2 + k ) ( 2 q + 1 ) ± 4 q + 1 − 4 k . (23)
Since every F ∈ Φ h F\in\Phi_h F ∈ Φ h satisfies F ≥ M m F\geq M_m F ≥ M m on [ 0 , 1 ) [0,1) [ 0 , 1 ) , all anchors lie in I 0 I_0 I 0 , and every member of Φ h \Phi_h Φ h other than M m M_m M m is strictly positive at t − t_- t − and at t + t_+ t + .
Put
φ ( β ) = 2 ( β − q ) ( β − 1 ) 2 , δ ∗ = q 2 + 4 q − 4 − q 2 , (24) \varphi(\beta)=2(\beta-q)(\beta-1)^2,\qquad\delta^*=\frac{\sqrt{q^2+4q-4}-q}{2}, \tag*{(24)} φ ( β ) = 2 ( β − q ) ( β − 1 ) 2 , δ ∗ = 2 q 2 + 4 q − 4 − q , (24)
so that φ \varphi φ increases on [ q , ∞ ) [q,\infty) [ q , ∞ ) from 0 0 0 , and δ ∗ \delta^* δ ∗ is the positive root of δ 2 + q δ = q − 1 \delta^2+q\delta=q-1 δ 2 + q δ = q − 1 . Let β − < β + \beta_-<\beta_+ β − < β + be the solutions in ( q , ∞ ) (q,\infty) ( q , ∞ ) of φ ( β − ) = m \varphi(\beta_-)=m φ ( β − ) = m and φ ( β + ) = h \varphi(\beta_+)=h φ ( β + ) = h .
!The anchor interval next to the square 2 5 2^5 2 5 , showing panels (a)–(e), known annular interior, numerical finite test, rational anchor 5, explicit Pisot anchor, anchor interval [ β − , β + ] [\beta_-,\beta_+] [ β − , β + ] , and depth-14 tree nodes (reciprocal minimizers) bbox=0.16,0.10,0.89,0.75
Figure 6. The anchor interval next to the square 2 5 2^5 2 5 . Panels (a)–(d) show the exterior core near the real axis at heights 2 2 2^2 2 2 to 2 5 2^5 2 5 . At h = 2 2 = 5 2 − 3 h = 2^2 = 5^2 - 3 h = 2 2 = 5 2 − 3 , components accumulate at the rational anchor 5 (Theorem B). At h = 2 5 h = 2^5 h = 2 5 the point 5 is interior by (8) ; at h = 2 3 , 2 4 h = 2^3, 2^4 h = 2 3 , 2 4 the plots suggest the same. The bracket marks the reciprocal anchor interval [ β − , β + ] [\beta_-,\beta_+] [ β − , β + ] of Lemma 6.2 . At h = 2 3 , 2 4 h = 2^3, 2^4 h = 2 3 , 2 4 the blue point is the explicit Pisot reciprocal anchor of Proposition 6.14 ; no such assertion is made here for h = 2 5 h = 2^5 h = 2 5 . Panel (e) enlarges the interval at h = 2 3 h = 2^3 h = 2 3 . Its strip shows numerical reciprocal minimizers for 583 candidate prefixes at depth 14. The shading and strip are illustrative; the interval bounds and marked anchors have exact definitions.
Lemma 6.2 (Anchors of square-adjacent witnesses). Every square-adjacent witness has anchor t = 1 / β t=1/\beta t = 1/ β with β − ≤ β ≤ β + \beta_{-}\le\beta\le\beta_{+} β − ≤ β ≤ β + , and
q + 1 4 < β − < β + < q + δ ∗ < q + 1. q+\frac{1}{4}<\beta_{-}<\beta_{+}<q+\delta^{*}<q+1. q + 4 1 < β − < β + < q + δ ∗ < q + 1. Proof. Let F = 1 + ∑ j ≥ 1 a j z j F=1+\sum_{j\ge1}a_jz^j F = 1 + ∑ j ≥ 1 a j z j be a square-adjacent witness with anchor t = 1 / β t=1/\beta t = 1/ β . In the notation of (3.1) with d j = − a j d_j=-a_j d j = − a j , the double-zero equations give P 0 = V 0 = 1 P_0=V_0=1 P 0 = V 0 = 1 , P 1 = β − 2 q P_1=\beta-2q P 1 = β − 2 q and V 1 = β V 0 + P 1 = 2 β − 2 q V_1=\beta V_0+P_1=2\beta-2q V 1 = β V 0 + P 1 = 2 β − 2 q , while the series expression V 1 = ∑ i ≥ 1 ( i − 1 ) a i + 1 β − i V_1=\sum_{i\ge1}(i-1)a_{i+1}\beta^{-i} V 1 = ∑ i ≥ 1 ( i − 1 ) a i + 1 β − i and the digit bounds m ≤ a j ≤ h m\le a_j\le h m ≤ a j ≤ h for j ≥ 3 j\ge3 j ≥ 3 give
m ( β − 1 ) 2 ≤ 2 β − 2 q ≤ h ( β − 1 ) 2 , \frac{m}{(\beta-1)^2}\le2\beta-2q\le\frac{h}{(\beta-1)^2}, ( β − 1 ) 2 m ≤ 2 β − 2 q ≤ ( β − 1 ) 2 h , that is m ≤ φ ( β ) ≤ h m\le\varphi(\beta)\le h m ≤ φ ( β ) ≤ h .
Monotonicity of φ \varphi φ gives β − ≤ β ≤ β + \beta_{-}\le\beta\le\beta_{+} β − ≤ β ≤ β + . The lower bound follows from φ ( q + 1 4 ) = 1 2 ( q − 3 4 ) 2 < q 2 − 2 q ≤ m \varphi(q+\frac14)=\frac12(q-\frac34)^2<q^2-2q\le m φ ( q + 4 1 ) = 2 1 ( q − 4 3 ) 2 < q 2 − 2 q ≤ m , valid for q ≥ 3 q\ge3 q ≥ 3 . For the upper bound we show φ ( q + δ ∗ ) > h \varphi(q+\delta^{*})>h φ ( q + δ ∗ ) > h . Since q 2 + 4 q − 4 < q + 2 \sqrt{q^2+4q-4}<q+2 q 2 + 4 q − 4 < q + 2 , we have δ ∗ = 2 ( q − 1 ) / ( q 2 + 4 q − 4 + q ) > ( q − 1 ) / ( q + 1 ) \delta^{*}=2(q-1)/(\sqrt{q^2+4q-4}+q)>(q-1)/(q+1) δ ∗ = 2 ( q − 1 ) / ( q 2 + 4 q − 4 + q ) > ( q − 1 ) / ( q + 1 ) , hence
φ ( q + δ ∗ ) > 2 ( q − 1 ) q + 1 ( ( q − 1 ) ( q + 2 ) q + 1 ) 2 = 2 ( q − 1 ) 3 ( q + 2 ) 2 ( q + 1 ) 3 ≥ q 2 ≥ h ( q ≥ 5 ) , \varphi(q+\delta^{*})>\frac{2(q-1)}{q+1}\left(\frac{(q-1)(q+2)}{q+1}\right)^2=\frac{2(q-1)^3(q+2)^2}{(q+1)^3}\ge q^2\ge h\qquad(q\ge5), φ ( q + δ ∗ ) > q + 1 2 ( q − 1 ) ( q + 1 ( q − 1 ) ( q + 2 ) ) 2 = ( q + 1 ) 3 2 ( q − 1 ) 3 ( q + 2 ) 2 ≥ q 2 ≥ h ( q ≥ 5 ) , because 2 ( q − 1 ) 3 ( q + 2 ) 2 − q 2 ( q + 1 ) 3 2(q-1)^3(q+2)^2-q^2(q+1)^3 2 ( q − 1 ) 3 ( q + 2 ) 2 − q 2 ( q + 1 ) 3 is positive at q = 5 q=5 q = 5 and increasing. For q = 4 q=4 q = 4 one has δ ∗ = 7 − 2 \delta^{*}=\sqrt7-2 δ ∗ = 7 − 2 and φ ( 4 + δ ∗ ) = 2 ( 7 − 2 ) ( 7 + 1 ) 2 = 8 7 − 4 > 16 ≥ h \varphi(4+\delta^{*})=2(\sqrt7-2)(\sqrt7+1)^2=8\sqrt7-4>16\ge h φ ( 4 + δ ∗ ) = 2 ( 7 − 2 ) ( 7 + 1 ) 2 = 8 7 − 4 > 16 ≥ h . Finally δ ∗ < 1 \delta^{*}<1 δ ∗ < 1 because δ 2 + q δ \delta^2+q\delta δ 2 + q δ is increasing and equals q + 1 > q − 1 q+1>q-1 q + 1 > q − 1 at δ = 1 \delta=1 δ = 1 .
The next lemma verifies the three inequalities of Lemma 4.3 at every possible anchor, so that the comparison argument applies to the entire square-adjacent family.
Lemma 6.3 (Comparison at the anchors). Let F \mathcal{F} F be a compact family of pairs ( F , t ) (F,t) ( F , t ) , where F F F is a square-adjacent witness at height h h h with anchor t t t . Then Lemma 4.3 applies to F \mathcal{F} F with H = h H=h H = h , the lower digit bound m m m , and ℓ = W = 2 q − 2 \ell=W=2q-2 ℓ = W = 2 q − 2 . Moreover every anchor satisfies t < σ h = ( h 2 + 4 ) − 1 / 4 t<\sigma_h=(h^2+4)^{-1/4} t < σ h = ( h 2 + 4 ) − 1/4 , so Lemma 4.1 normalizes all witnesses near all anchors.
Proof. Let β = 1 / t ∈ [ β − , β + ] \beta=1/t\in[\beta_{-},\beta_{+}] β = 1/ t ∈ [ β − , β + ] . The first inequality ℓ / ( β − 1 ) < 2 \ell/(\beta-1)<2 ℓ / ( β − 1 ) < 2 is β > q \beta>q β > q . The second, ( β − 1 ) 2 > 2 q − 2 (\beta-1)^2>2q-2 ( β − 1 ) 2 > 2 q − 2 , follows from β > q \beta>q β > q and ( q − 1 ) 2 > 2 q − 2 (q-1)^2>2q-2 ( q − 1 ) 2 > 2 q − 2 , which is ( q − 1 ) ( q − 3 ) > 0 (q-1)(q-3)>0 ( q − 1 ) ( q − 3 ) > 0 . For the third, write Δ = 2 β − ℓ − 2 h / ( β − 1 ) 2 \Delta=2\beta-\ell-2h/(\beta-1)^2 Δ = 2 β − ℓ − 2 h / ( β − 1 ) 2 and observe that Δ > 0 \Delta>0 Δ > 0 is equivalent to ( β − q + 1 ) ( β − 1 ) 2 > h (\beta-q+1)(\beta-1)^2>h ( β − q + 1 ) ( β − 1 ) 2 > h . Since ( β − q ) ( β − 1 ) 2 = φ ( β ) / 2 ≥ m / 2 (\beta-q)(\beta-1)^2=\varphi(\beta)/2\ge m/2 ( β − q ) ( β − 1 ) 2 = φ ( β ) /2 ≥ m /2 , it suffices that ( β − 1 ) 2 > h − m / 2 = ( q 2 + 2 q − 4 + k ) / 2 (\beta-1)^2>h-m/2=(q^2+2q-4+k)/2 ( β − 1 ) 2 > h − m /2 = ( q 2 + 2 q − 4 + k ) /2 . By Lemma 6.2 , β − 1 > q − 3 4 \beta-1>q-\frac34 β − 1 > q − 4 3 , and ( q − 3 4 ) 2 ≥ ( q 2 + 2 q − 2 ) / 2 (q-\frac34)^2\ge(q^2+2q-2)/2 ( q − 4 3 ) 2 ≥ ( q 2 + 2 q − 2 ) /2 is equivalent to q 2 − 5 q + 25 8 ≥ 0 q^2-5q+\frac{25}{8}\ge0 q 2 − 5 q + 8 25 ≥ 0 , which holds for q ≥ 5 q\ge5 q ≥ 5 . For q = 4 q=4 q = 4 one checks directly that φ ( 1 + h − m / 2 ) < m \varphi(1+\sqrt{h-m/2})<m φ ( 1 + h − m /2 ) < m for k = 0 , 1 , 2 k=0,1,2 k = 0 , 1 , 2 , which places 1 + h − m / 2 1+\sqrt{h-m/2} 1 + h − m /2 below β − \beta_{-} β − : the three values are φ ( 1 + 10 ) = 2 ( 10 − 3 ) ⋅ 10 < 8 \varphi(1+\sqrt{10})=2(\sqrt{10}-3)\cdot10<8 φ ( 1 + 10 ) = 2 ( 10 − 3 ) ⋅ 10 < 8 , φ ( 1 + 21 / 2 ) = 21 ( 21 / 2 − 3 ) < 9 \varphi(1+\sqrt{21/2})=21(\sqrt{21/2}-3)<9 φ ( 1 + 21/2 ) = 21 ( 21/2 − 3 ) < 9 , and φ ( 1 + 11 ) = 22 ( 11 − 3 ) < 10 \varphi(1+\sqrt{11})=22(\sqrt{11}-3)<10 φ ( 1 + 11 ) = 22 ( 11 − 3 ) < 10 . All three inequalities hold on the closed interval [ β − , β + ] [\beta_{-},\beta_{+}] [ β − , β + ] , so the margins are uniform over F \mathcal{F} F . For the normalization,
β 4 > ( q + 1 4 ) 4 > q 4 + q 3 > q 4 + 4 ≥ h 2 + 4. \beta^4>\left(q+\frac14\right)^4>q^4+q^3>q^4+4\ge h^2+4. β 4 > ( q + 4 1 ) 4 > q 4 + q 3 > q 4 + 4 ≥ h 2 + 4.
By Lemma 6.3 and the rigidity statement of Lemma 4.3 , a square-adjacent witness is the unique member of F h \mathcal{F}_h F h with a multiple zero at its anchor. In particular distinct square-adjacent witnesses have distinct anchors, and the map from square-adjacent witnesses to T h \mathcal{T}_h T h is a bijection. The set of square-adjacent witnesses is closed in the coefficient topology, because a coefficient limit of square-adjacent witnesses has a double zero at the limit of the anchors by Lemma 4.1 ; hence T h \mathcal{T}_h T h is compact.
Witnesses from intermediate values
Witnesses are produced by Lemma 4.6 . In this section it is applied to prefixes π = ( 1 , − 2 q , a 2 , … , a N − 1 ) \pi=(1,-2q,a_2,\ldots,a_{N-1}) π = ( 1 , − 2 q , a 2 , … , a N − 1 ) with digits a j ∈ [ m , h ] a_j\in[m,h] a j ∈ [ m , h ] and to digit ranges [ m ′ , H ′ ] ⊆ [ m , h ] [m',H']\subseteq[m,h] [ m ′ , H ′ ] ⊆ [ m , h ] . The completion F ′ F' F ′ it provides then belongs to Φ h \Phi_h Φ h , and the interior minimum point is a double zero: F ′ F' F ′ is a square-adjacent witness with anchor in ( a , b ) (a,b) ( a , b ) .
Lemma 6.4 (Existence of square-adjacent witnesses). The square-adjacent family Φ h \Phi_h Φ h contains a square-adjacent witness. For q ≥ 6 q\ge6 q ≥ 6 the family
Ψ q = { 1 − 2 q z + ∑ j ≥ 2 a j z j : q 2 − 2 q + 2 ≤ a j ≤ q 2 − 2 } ⊂ Φ q 2 − 2 ∩ Φ q 2 − 1 ∩ Φ q 2 (25) \Psi_q=\left\{1-2qz+\sum_{j\ge2}a_jz^j:q^2-2q+2\le a_j\le q^2-2\right\}\subset\Phi_{q^2-2}\cap\Phi_{q^2-1}\cap\Phi_{q^2} \tag*{(25)} Ψ q = { 1 − 2 q z + j ≥ 2 ∑ a j z j : q 2 − 2 q + 2 ≤ a j ≤ q 2 − 2 } ⊂ Φ q 2 − 2 ∩ Φ q 2 − 1 ∩ Φ q 2 (25) contains a square-adjacent witness. The same holds for q ∈ { 4 , 5 } q\in\{4,5\} q ∈ { 4 , 5 } .
Proof. Apply Lemma 4.6 to the root prefix π = ( 1 , − 2 q ) \pi=(1,-2q) π = ( 1 , − 2 q ) , the digit range [ m , h ] [m,h] [ m , h ] and J = [ t − , t + ] J=[t_-,t_+] J = [ t − , t + ] . The smallest completion is M m M_m M m , which vanishes at both endpoints and is negative inside. The largest completion is 1 − 2 q z + h z 2 / ( 1 − z ) 1-2qz+hz^2/(1-z) 1 − 2 q z + h z 2 / ( 1 − z ) , whose numerator 1 − ( 2 q + 1 ) z + ( 2 q + h ) z 2 1-(2q+1)z+(2q+h)z^2 1 − ( 2 q + 1 ) z + ( 2 q + h ) z 2 has discriminant ( 2 q − 1 ) 2 − 4 h < 0 (2q-1)^2-4h<0 ( 2 q − 1 ) 2 − 4 h < 0 , so it is positive. The contiguity condition ( 2 q − 2 ) t − ≥ 1 − t − (2q-2)t_-\geq1-t_- ( 2 q − 2 ) t − ≥ 1 − t − is t − ≥ 1 / ( 2 q − 1 ) t_-\geq1/(2q-1) t − ≥ 1/ ( 2 q − 1 ) . Using (6.3), it is equivalent to
2 q 2 − 1 − 2 k ≥ ( 2 q − 1 ) 4 q + 1 − 4 k . 2q^2-1-2k\geq(2q-1)\sqrt{4q+1-4k}. 2 q 2 − 1 − 2 k ≥ ( 2 q − 1 ) 4 q + 1 − 4 k .
and squaring gives
4 q 4 − 16 q 3 + ( 8 + 8 k ) q 2 − 16 k q + 4 k 2 + 8 k ≥ 0. 4q^4-16q^3+(8+8k)q^2-16kq+4k^2+8k\geq0. 4 q 4 − 16 q 3 + ( 8 + 8 k ) q 2 − 16 k q + 4 k 2 + 8 k ≥ 0.
For k = 0 k=0 k = 0 the left side is 4 q 2 ( q 2 − 4 q + 2 ) 4q^2(q^2-4q+2) 4 q 2 ( q 2 − 4 q + 2 ) , for k = 1 k=1 k = 1 it is 4 ( q 2 ( q − 2 ) 2 − 4 q + 3 ) 4(q^2(q-2)^2-4q+3) 4 ( q 2 ( q − 2 ) 2 − 4 q + 3 ) , and for k = 2 k=2 k = 2 it is 4 ( ( q − 1 ) 4 − 4 q + 7 ) 4((q-1)^4-4q+7) 4 (( q − 1 ) 4 − 4 q + 7 ) ; each is positive for q ≥ 4 q\geq4 q ≥ 4 . This proves the first assertion.
For the second, put m ′ = q 2 − 2 q + 2 m'=q^2-2q+2 m ′ = q 2 − 2 q + 2 , H ′ = q 2 − 2 H'=q^2-2 H ′ = q 2 − 2 , and apply the lemma to the root prefix with J = [ a , b ] J=[a,b] J = [ a , b ] , where a = 1 / ( 2 q − 3 ) a=1/(2q-3) a = 1/ ( 2 q − 3 ) and b = t + ′ b=t'_+ b = t + ′ is the larger zero of M m ′ M_{m'} M m ′ ; the contiguity condition holds with equality at a a a . The largest completion 1 − 2 q z + ( q 2 − 2 ) z 2 / ( 1 − z ) 1-2qz+(q^2-2)z^2/(1-z) 1 − 2 q z + ( q 2 − 2 ) z 2 / ( 1 − z ) is positive, as its numerator has discriminant 9 − 4 q < 0 9-4q<0 9 − 4 q < 0 . The smallest completion is M m ′ M_{m'} M m ′ , which vanishes at b b b and is negative on ( t − ′ , b ) (t'_-,b) ( t − ′ , b ) . It is nonnegative at a a a precisely when a ≤ t − ′ a\leq t'_- a ≤ t − ′ , that is,
2 q 2 − 4 q − 7 ≥ ( 2 q − 3 ) 4 q − 7 . 2q^2-4q-7\geq(2q-3)\sqrt{4q-7}. 2 q 2 − 4 q − 7 ≥ ( 2 q − 3 ) 4 q − 7 .
For q ≥ 7 q\geq7 q ≥ 7 this follows from 4 q − 7 ≤ q − 2 \sqrt{4q-7}\leq q-2 4 q − 7 ≤ q − 2 and ( 2 q − 3 ) ( q − 2 ) ≤ 2 q 2 − 4 q − 7 (2q-3)(q-2)\leq2q^2-4q-7 ( 2 q − 3 ) ( q − 2 ) ≤ 2 q 2 − 4 q − 7 , and for q = 6 q=6 q = 6 it reads 41 ≥ 9 17 41\geq9\sqrt{17} 41 ≥ 9 17 . The interval J J J contains points of ( t − ′ , b ) (t'_-,b) ( t − ′ , b ) , so min J M m ′ < 0 \min_J M_{m'}<0 min J M m ′ < 0 .
For q ∈ { 4 , 5 } q\in\{4,5\} q ∈ { 4 , 5 } the root does not satisfy the endpoint condition, and we apply the lemma to a deeper prefix. For q = 4 q=4 q = 4 take J = [ 1 / 5 , 1 / 3 ] J=[1/5,1/3] J = [ 1/5 , 1/3 ] and
π = ( 1 , − 8 , 12 , 13 , 14 , 12 , 13 , 10 , 13 , 13 ) , \pi=(1,-8,12,13,14,12,13,10,13,13), π = ( 1 , − 8 , 12 , 13 , 14 , 12 , 13 , 10 , 13 , 13 ) ,
and for q = 5 q=5 q = 5 take J = [ 1 / 7 , 7 / 25 ] J=[1/7,7/25] J = [ 1/7 , 7/25 ] and π = ( 1 , − 10 , 20 , 22 , 18 , 21 , 22 , 21 , 17 ) \pi=(1,-10,20,22,18,21,22,21,17) π = ( 1 , − 10 , 20 , 22 , 18 , 21 , 22 , 21 , 17 ) . In both cases the contiguity condition holds with equality at the left endpoint, and the four remaining conditions are inequalities between rational numbers, which are verified in exact arithmetic (Proposition S4.13 of the supplement). For the lower bound on f π + f_\pi^+ f π + we use convexity: rationals z 1 < z 2 z_1<z_2 z 1 < z 2 in J J J with ( f π + ) ′ ( z 1 ) ≤ 0 ≤ ( f π + ) ′ ( z 2 ) (f_\pi^+)'(z_1)\leq0\leq(f_\pi^+)'(z_2) ( f π + ) ′ ( z 1 ) ≤ 0 ≤ ( f π + ) ′ ( z 2 ) , f π + ( z 2 ) > 0 f_\pi^+(z_2)>0 f π + ( z 2 ) > 0 and f π + ( z 1 ) + ( f π + ) ′ ( z 1 ) ( z 2 − z 1 ) > 0 f_\pi^+(z_1)+(f_\pi^+)'(z_1)(z_2-z_1)>0 f π + ( z 1 ) + ( f π + ) ′ ( z 1 ) ( z 2 − z 1 ) > 0 certify f π + > 0 f_\pi^+>0 f π + > 0 on J J J . □
Tail sums and the exceptional witnesses
Let F = 1 − 2 q z + ∑ j ≥ 2 a j z j F=1-2qz+\sum_{j\geq2}a_jz^j F = 1 − 2 q z + ∑ j ≥ 2 a j z j be a square-adjacent witness at height h h h with anchor t = 1 / β t=1/\beta t = 1/ β . Write g j = a j − m ∈ [ 0 , W ] g_j=a_j-m\in[0,W] g j = a j − m ∈ [ 0 , W ] for its digit offsets, and put
L = W t 1 − t = W β − 1 , S N = ∑ i ≥ 1 g N + i t i ( N ≥ 1 ) . (26) L=\frac{Wt}{1-t}=\frac{W}{\beta-1},\qquad S_N=\sum_{i\geq1}g_{N+i}t^i\qquad(N\geq1). \tag*{(26)} L = 1 − t W t = β − 1 W , S N = i ≥ 1 ∑ g N + i t i ( N ≥ 1 ) . (26)
Then 0 ≤ S N ≤ L 0\leq S_N\leq L 0 ≤ S N ≤ L and
S N + 1 = β S N − g N + 1 ( N ≥ 1 ) . (27) S_{N+1}=\beta S_N-g_{N+1}\qquad(N\geq1). \tag*{(27)} S N + 1 = β S N − g N + 1 ( N ≥ 1 ) . (27)
By Lemma 6.2, 1 < L < 2 1<L<2 1 < L < 2 . Call a level N ≥ 2 N\geq2 N ≥ 2 ascending if S N ≥ 1 S_N\geq1 S N ≥ 1 and g N ≤ W − 1 g_N\leq W-1 g N ≤ W − 1 , and descending if S N < L − 1 S_N<L-1 S N < L − 1 and g N ≥ 1 g_N\geq1 g N ≥ 1 ; the margin of a descending level is L − 1 − S N L-1-S_N L − 1 − S N . The witness F F F is live if infinitely many of its levels are ascending or descending, and exceptional otherwise.
Lemma 6.5 (Exceptional witnesses are eventually constant). An exceptional square-adjacent witness is eventually constant, with constant digit m m m , m + q − 1 m+q-1 m + q − 1 or h h h .
Proof. Let β \beta β be the reciprocal anchor. By Lemma 6.2, β < q + δ ∗ \beta<q+\delta^* β < q + δ ∗ , that is β 2 − q β − ( q − 1 ) < 0 \beta^2-q\beta-(q-1)<0 β 2 − q β − ( q − 1 ) < 0 . Expanding β ( L − 1 ) = β ( 2 q − 1 − β ) / ( β − 1 ) \beta(L-1)=\beta(2q-1-\beta)/(\beta-1) β ( L − 1 ) = β ( 2 q − 1 − β ) / ( β − 1 ) and ( β − q + 1 ) ( β − 1 ) = β 2 − q β + q − 1 (\beta-q+1)(\beta-1)=\beta^2-q\beta+q-1 ( β − q + 1 ) ( β − 1 ) = β 2 − q β + q − 1 , this single inequality is equivalent to each of
β ( L − 1 ) > q − 1 , β − ( L − 1 ) < q . (28) \beta(L-1)>q-1,\qquad\beta-(L-1)<q. \tag*{(28)} β ( L − 1 ) > q − 1 , β − ( L − 1 ) < q . (28)
Suppose no level N ≥ N 1 N\geq N_1 N ≥ N 1 is ascending or descending. For N ≥ N 1 N\geq N_1 N ≥ N 1 this means: S N < 1 S_N<1 S N < 1 or g N = W g_N=W g N = W , and S N ≥ L − 1 S_N\geq L-1 S N ≥ L − 1 or g N = 0 g_N=0 g N = 0 .
First let N ≥ N 1 N\geq N_1 N ≥ N 1 with 1 ≤ g N ≤ W − 1 1\leq g_N\leq W-1 1 ≤ g N ≤ W − 1 , so that L − 1 ≤ S N < 1 L-1\leq S_N<1 L − 1 ≤ S N < 1 . Then (6.7) and 0 ≤ S N + 1 ≤ L 0\leq S_{N+1}\leq L 0 ≤ S N + 1 ≤ L give
g N + 1 ≥ β S N − L ≥ β ( L − 1 ) − L = W − β > q − 3 , g N + 1 < β S N < β < q + 1 , g_{N+1}\geq\beta S_N-L\geq\beta(L-1)-L=W-\beta>q-3,\qquad g_{N+1}<\beta S_N<\beta<q+1, g N + 1 ≥ β S N − L ≥ β ( L − 1 ) − L = W − β > q − 3 , g N + 1 < β S N < β < q + 1 ,
so q − 2 ≤ g N + 1 ≤ q q-2\leq g_{N+1}\leq q q − 2 ≤ g N + 1 ≤ q , and in particular 1 ≤ g N + 1 ≤ W − 1 1\leq g_{N+1}\leq W-1 1 ≤ g N + 1 ≤ W − 1 . Hence also L − 1 ≤ S N + 1 < 1 L-1\leq S_{N+1}<1 L − 1 ≤ S N + 1 < 1 , and then (6.7) together with (6.8) gives q − 2 < β ( L − 1 ) − 1 < g N + 1 < β − ( L − 1 ) < q q-2<\beta(L-1)-1<g_{N+1}<\beta-(L-1)<q q − 2 < β ( L − 1 ) − 1 < g N + 1 < β − ( L − 1 ) < q , so g N + 1 = q − 1 g_{N+1}=q-1 g N + 1 = q − 1 . By induction g j = q − 1 g_j=q-1 g j = q − 1 and L − 1 ≤ S j < 1 L-1 \leq S_j < 1 L − 1 ≤ S j < 1 for all j > N j>N j > N . The affine recursion S j + 1 − x ∗ = β ( S j − x ∗ ) S_{j+1}-x^*=\beta(S_j-x^*) S j + 1 − x ∗ = β ( S j − x ∗ ) with x ∗ = ( q − 1 ) / ( β − 1 ) x^*=(q-1)/(\beta-1) x ∗ = ( q − 1 ) / ( β − 1 ) then forces S j = x ∗ S_j=x^* S j = x ∗ for all j ≥ N j \geq N j ≥ N , and the digits are eventually m + q − 1 m+q-1 m + q − 1 .
Figure 7. (a) The tail-sum recursion (6.7) at height 23 for the explicit witness of Proposition 6.14: the thin lines are the branches S ↦ β S − g S \mapsto\beta S-g S ↦ β S − g , 0 ≤ g ≤ W 0 \leq g \leq W 0 ≤ g ≤ W , with g = 0 g=0 g = 0 , q − 1 q-1 q − 1 and W W W drawn darker; the shaded band is the dead zone [ L − 1 , 1 ) [L-1,1) [ L − 1 , 1 ) ; the hollow circles are the fixed points 0 0 0 , ( q − 1 ) / ( β − 1 ) (q-1)/(\beta-1) ( q − 1 ) / ( β − 1 ) and L L L , which correspond to the exceptional tails with constant digit m m m , m + q − 1 m+q-1 m + q − 1 and h h h . The cobweb follows the tail sums ( S N , S N + 1 ) (S_N,S_{N+1}) ( S N , S N + 1 ) of the explicit witness from ( S 2 , S 3 ) (S_2,S_3) ( S 2 , S 3 ) on: they end in a two-cycle that alternates between the descending region and the dead zone, so every other level is descending. (b) Numerical counts of the surviving prefixes of the intermediate-value tree of the square-adjacent family, by prefix length, at the six heights next to 16 and 25. The thin lines illustrate growth by factors ( W + 1 ) / 4.47 (W+1)/4.47 ( W + 1 ) /4.47 for q = 4 q=4 q = 4 and ( W + 1 ) / 5.47 (W+1)/5.47 ( W + 1 ) /5.47 for q = 5 q=5 q = 5 per additional digit, not a proved growth rate. Only prefixes passing the exact tests of Section S4 of the supplement are used for the reported lower bounds.
Next let g N = 0 g_N=0 g N = 0 , so S N < 1 S_N<1 S N < 1 . Since β < q + 1 ≤ W \beta<q+1\leq W β < q + 1 ≤ W , the next offset satisfies g N + 1 ≤ β S N < W g_{N+1}\leq\beta S_N<W g N + 1 ≤ β S N < W . If g N + 1 ≥ 1 g_{N+1}\geq1 g N + 1 ≥ 1 , the first case applies at level N + 1 N+1 N + 1 . If g N + 1 = 0 g_{N+1}=0 g N + 1 = 0 , then S N + 1 < 1 S_{N+1}<1 S N + 1 < 1 and S N = S N + 1 / β < β − 1 S_N=S_{N+1}/\beta<\beta^{-1} S N = S N + 1 / β < β − 1 . Either some later offset is nonzero, and the first nonzero offset after N N N lies in [ 1 , W − 1 ] [1,W-1] [ 1 , W − 1 ] by the same bound, so the first case applies; or all later offsets vanish, which means S N = 0 S_N=0 S N = 0 and the digits are eventually m m m .
Finally let g N = W g_N=W g N = W , so S N ≥ L − 1 S_N\geq L-1 S N ≥ L − 1 . Then g N + 1 ≥ β S N − L ≥ W − β > q − 3 g_{N+1}\geq\beta S_N-L\geq W-\beta>q-3 g N + 1 ≥ β S N − L ≥ W − β > q − 3 , so g N + 1 ≥ 1 g_{N+1}\geq1 g N + 1 ≥ 1 . If g N + 1 ≤ W − 1 g_{N+1}\leq W-1 g N + 1 ≤ W − 1 , the first case applies. If g N + 1 = W g_{N+1}=W g N + 1 = W , then S N + 1 = β S N − W ≥ L − 1 S_{N+1}=\beta S_N-W\geq L-1 S N + 1 = β S N − W ≥ L − 1 . Either some later offset is smaller than W W W , and the first such offset is at least q − 2 q-2 q − 2 , so the first case applies; or all later offsets equal W W W , and the recursion S j + 1 − L = β ( S j − L ) S_{j+1}-L=\beta(S_j-L) S j + 1 − L = β ( S j − L ) with S j ≥ L − 1 S_j\geq L-1 S j ≥ L − 1 forces S N = L S_N=L S N = L , so that the digits are eventually h h h . □
The three exceptional types have different meanings. A witness with constant digit m m m or h h h is the smallest or largest completion of its own prefix. A witness with constant digit m + q − 1 m+q-1 m + q − 1 is a polynomial perturbation of the channel series
1 − 2 q z + ( m + q − 1 ) z 2 1 − z = ( 1 − q z ) ( 1 − ( q + 1 ) z ) + ( k − 1 ) z 2 1 − z , (29) 1-2qz+(m+q-1)\frac{z^2}{1-z}=\frac{(1-qz)(1-(q+1)z)+(k-1)z^2}{1-z}, \tag*{(29)} 1 − 2 q z + ( m + q − 1 ) 1 − z z 2 = 1 − z ( 1 − q z ) ( 1 − ( q + 1 ) z ) + ( k − 1 ) z 2 , (29)
and by (6.8) the offset q − 1 q-1 q − 1 is the only one that can keep two successive tail sums in the dead zone [ L − 1 , 1 ) [L-1,1) [ L − 1 , 1 ) ; the corresponding branch of (6.7) has the fixed point ( q − 1 ) / ( β − 1 ) (q-1)/(\beta-1) ( q − 1 ) / ( β − 1 ) .
Lemma 6.6 (Margins). Put y = W / ( β + − 1 ) − 1 − ( β + − q ) > 0 y = W/(\beta_{+} - 1) - 1 - (\beta_{+} - q) > 0 y = W / ( β + − 1 ) − 1 − ( β + − q ) > 0 and η = min { β − − q , y / β + } \eta= \min\{\beta_{-} - q, y/\beta_{+}\} η = min { β − − q , y / β + } . A live square-adjacent witness has infinitely many ascending levels, or infinitely many descending levels of margin at least η \eta η .
Proof. Positivity of y y y is the second inequality in (6.8) at β + \beta_{+} β + . Let F F F be live with reciprocal anchor β \beta β and write δ = β − q \delta= \beta- q δ = β − q and y ( β ) = L − 1 − δ y(\beta) = L - 1 - \delta y ( β ) = L − 1 − δ , which decreases in β \beta β , so y ( β ) ≥ y > 0 y(\beta) \ge y > 0 y ( β ) ≥ y > 0 ; also y ( β ) < 1 y(\beta) < 1 y ( β ) < 1 since L < 2 L < 2 L < 2 . We show that a descending level N N N of margin ε < η \varepsilon< \eta ε < η is followed by an ascending level or by a descending level of margin larger than δ ≥ η \delta\ge\eta δ ≥ η ; this proves the lemma.
Write S N = L − 1 − ε S_N = L - 1 - \varepsilon S N = L − 1 − ε . Then β S N = β ( L − 1 ) − β ε = ( q − 1 ) + y ( β ) − β ε \beta S_N = \beta(L - 1) - \beta\varepsilon= (q - 1) + y(\beta) - \beta\varepsilon β S N = β ( L − 1 ) − β ε = ( q − 1 ) + y ( β ) − β ε , using β ( L − 1 ) = W + L − β \beta(L - 1) = W + L - \beta β ( L − 1 ) = W + L − β . Since β ε < β + η ≤ y ≤ y ( β ) \beta\varepsilon< \beta_{+}\eta\le y \le y(\beta) β ε < β + η ≤ y ≤ y ( β ) , the number β S N \beta S_N β S N lies in ( q − 1 , q ) (q - 1, q) ( q − 1 , q ) with fractional part y ( β ) − β ε y(\beta) - \beta\varepsilon y ( β ) − β ε . By (6.7) and 0 ≤ S N + 1 ≤ L < 2 0 \le S_{N+1} \le L < 2 0 ≤ S N + 1 ≤ L < 2 , the offset g N + 1 g_{N+1} g N + 1 lies in [ β S N − L , β S N ] [\beta S_N - L, \beta S_N] [ β S N − L , β S N ] , so g N + 1 ∈ { q − 2 , q − 1 } g_{N+1} \in\{q - 2, q - 1\} g N + 1 ∈ { q − 2 , q − 1 } . If g N + 1 = q − 1 g_{N+1} = q - 1 g N + 1 = q − 1 , then S N + 1 = y ( β ) − β ε < L − 1 − δ S_{N+1} = y(\beta) - \beta\varepsilon< L - 1 - \delta S N + 1 = y ( β ) − β ε < L − 1 − δ , and level N + 1 N + 1 N + 1 is descending with margin larger than δ \delta δ . If g N + 1 = q − 2 g_{N+1} = q - 2 g N + 1 = q − 2 , then S N + 1 = 1 + y ( β ) − β ε > 1 S_{N+1} = 1 + y(\beta) - \beta\varepsilon> 1 S N + 1 = 1 + y ( β ) − β ε > 1 , and level N + 1 N + 1 N + 1 is ascending because q − 2 ≤ W − 1 q - 2 \le W - 1 q − 2 ≤ W − 1 . □
From a live witness to infinitely many components
Theorem 6.7 (Live square-adjacent witnesses). Let q ≥ 4 q \ge4 q ≥ 4 , let h ∈ { q 2 − 2 , q 2 − 1 , q 2 } h \in\{q^2 - 2, q^2 - 1, q^2\} h ∈ { q 2 − 2 , q 2 − 1 , q 2 } , and let F ∗ F_{*} F ∗ be a live square-adjacent witness at height h h h with anchor t ∗ t_{*} t ∗ . Then t ∗ ∈ A h t_{*} \in A_h t ∗ ∈ A h , and infinitely many distinct connected components of C h C_h C h with nonempty planar interior accumulate at each of the four points ± t ∗ \pm t_{*} ± t ∗ and ± 1 / t ∗ \pm1/t_{*} ± 1/ t ∗ , at which C h C_h C h is not locally connected. If F ∗ F_{*} F ∗ has infinitely many ascending levels, then these four points are singleton components of C h C_h C h .
Proof. Negation and inversion transfer the statements at t ∗ t_{*} t ∗ to the other three points, so we work at t ∗ = 1 / β ∗ t_{*} = 1/\beta_{*} t ∗ = 1/ β ∗ . By Lemma 6.6 there are two cases.
Ascending levels. Suppose infinitely many levels are ascending. For every ascending level N N N the series F ∗ + z N F_{*} + z^N F ∗ + z N is admissible, because g N ≤ W − 1 g_N \le W - 1 g N ≤ W − 1 means a N ( F ∗ ) + 1 ≤ h a_N(F_{*}) + 1 \le h a N ( F ∗ ) + 1 ≤ h , and Lemma 4.1 with C N = 1 C_N = 1 C N = 1 gives nonreal zeros in C h C_h C h converging to t ∗ t_{*} t ∗ . Thus t ∗ ∈ A h t_{*} \in A_h t ∗ ∈ A h and t ∗ t_{*} t ∗ is nonisolated in C h C_h C h .
For an ascending level N N N , let π N = ( 1 , − 2 q , a 2 , … , a N − 1 ) \pi_N = (1, -2q, a_2, \ldots, a_{N-1}) π N = ( 1 , − 2 q , a 2 , … , a N − 1 ) be the prefix of F ∗ F_{*} F ∗ and let π N ′ \pi'_N π N ′ be the prefix π N \pi_N π N followed by the digit a N ( F ∗ ) + 1 ≤ h a_N(F_{*}) + 1 \le h a N ( F ∗ ) + 1 ≤ h . We apply Lemma 4.6 with the digit range [ m , h ] [m,h] [ m , h ] and J = [ t − , t + ] J = [t_{-}, t_{+}] J = [ t − , t + ] ; the contiguity condition was verified in Lemma 6.4. The smallest completion of π N ′ \pi'_N π N ′ is
f π N ′ − = F ∗ + z N − ∑ j > N ( a j ( F ∗ ) − m ) z j , f π N ′ − ( t ∗ ) = t ∗ N ( 1 − S N ) ≤ 0. f^-_{\pi'_N} = F_{*} + z^N - \sum_{j>N}(a_j(F_{*}) - m)z^j,\qquad f^-_{\pi'_N}(t_{*}) = t_{*}^N(1 - S_N) \le0. f π N ′ − = F ∗ + z N − j > N ∑ ( a j ( F ∗ ) − m ) z j , f π N ′ − ( t ∗ ) = t ∗ N ( 1 − S N ) ≤ 0.
If S N > 1 S_N > 1 S N > 1 , its minimum on J J J is negative. If S N = 1 S_N = 1 S N = 1 and the minimum were 0, then f π N ′ − f^-_{\pi'_N} f π N ′ − would be a square-adjacent witness with the anchor t ∗ t_{*} t ∗ , contradicting the rigidity of F ∗ F_{*} F ∗ . The largest completion
f π N ′ + = F ∗ + z N + ∑ j > N ( h − a j ( F ∗ ) ) z j f^+_{\pi'_N} = F_{*} + z^N + \sum_{j>N}(h - a_j(F_{*}))z^j f π N ′ + = F ∗ + z N + j > N ∑ ( h − a j ( F ∗ )) z j
dominates F ∗ ≥ 0 F_{*} \ge0 F ∗ ≥ 0 on J J J , and at the endpoints f π N ′ − f^-_{\pi'_N} f π N ′ − exceeds M m M_m M m , which vanishes there. The lemma gives a square-adjacent witness F N F_N F N with prefix π N ′ \pi'_N π N ′ and anchor t N ∈ ( t − , t + ) t_N \in(t_{-}, t_{+}) t N ∈ ( t − , t + ) . The witnesses F N F_N F N converge to F ∗ F_{*} F ∗ coefficientwise, since they agree with F ∗ F_{*} F ∗ below degree N N N , and t N → t ∗ t_N \to t_{*} t N → t ∗ by strict convexity.
Let j N j_N j N be the least index above N N N with a j N ( F N ) > m a_{j_N}(F_N) > m a j N ( F N ) > m ; it exists because the smallest completion of π N ′ \pi'_N π N ′ has negative minimum. Let B N \mathcal{B}_N B N be the block of all negative departures from F N F_N F N with degrees in [ N , j N − 1 ] [N,j_N - 1] [ N , j N − 1 ] . The witness F ∗ F_{*} F ∗ belongs to B N \mathcal{B}_N B N : its first departure from F N F_N F N has degree N N N and constant coefficient − 1 -1 − 1 . Lemma 6.3 applies to the compact family { ( F N , t N ) } N ∪ { ( F ∗ , t ∗ ) } \{(F_N,t_N)\}_N \cup\{(F_{*},t_{*})\} {( F N , t N ) } N ∪ {( F ∗ , t ∗ )} and yields one common separation neighbourhood. Since a N ( F N ) = a N ( F ∗ ) + 1 > m a_N(F_N) = a_N(F_{*}) + 1 > m a N ( F N ) = a N ( F ∗ ) + 1 > m and a j N ( F N ) > m a_{j_N}(F_N) > m a j N ( F N ) > m , the last assertion of Lemma 4.3 gives the hypotheses of Lemma 4.2 with H = h ∗ = h H = h_{*} = h H = h ∗ = h and n N = N → ∞ n_N = N \to\infty n N = N → ∞ . Every K N K_N K N contains t ∗ t_{*} t ∗ , a qualifying zero of F ∗ ∈ B N F_{*} \in\mathcal{B}_N F ∗ ∈ B N lying in C h C_h C h . Part (ii) of that lemma gives the singleton component { t ∗ } \{t_{*}\} { t ∗ } , the accumulating interior-bearing components, and the failure of local connectedness.
Descending levels. Suppose instead that infinitely many levels are descending with margin at least η \eta η . By Lemma 6.3, the three inequalities (4.2) hold at β ∗ \beta_{*} β ∗ with H = h H = h H = h and ℓ = W \ell= W ℓ = W , and t ∗ < σ h t_{*} < \sigma_h t ∗ < σ h . A descending level N N N of margin at least η \eta η has a N ( F ∗ ) > m a_N(F_{*}) > m a N ( F ∗ ) > m and
∑ j ≥ 1 ( h − a N + j ( F ∗ ) ) t ∗ j = L − S N ≥ 1 + η . (30) \sum_{j \ge1}(h - a_{N+j}(F_{*}))t_{*}^{j} = L - S_N \ge1 + \eta. \tag*{(30)} j ≥ 1 ∑ ( h − a N + j ( F ∗ )) t ∗ j = L − S N ≥ 1 + η . (30)
Proposition 4.4 with h ∗ = H = h h_* = H = h h ∗ = H = h gives the accumulating interior-bearing components, the failure of local connectedness, and t ∗ ∈ A h t_* \in A_h t ∗ ∈ A h .
Lemma 6.8 (Exceptional witnesses are isolated). An exceptional square-adjacent witness is an isolated point of the set of square-adjacent witnesses, and its anchor is an isolated point of T h T_h T h . Consequently T h T_h T h is infinite if and only if some square-adjacent witness is live.
Proof. Let F F F be exceptional with constant digit c ∈ { m , m + q − 1 , h } c \in\{m,m+q-1,h\} c ∈ { m , m + q − 1 , h } from degree N 0 N_0 N 0 on, and anchor t = 1 / β t=1/\beta t = 1/ β . Suppose square-adjacent witnesses G n ≠ F G_n \ne F G n = F converge to F F F ; their first departure degrees M n M_n M n tend to infinity, and their anchors t n t_n t n tend to t t t . Write G n = F + z M n C n G_n=F+z^{M_n}C_n G n = F + z M n C n with C n ( 0 ) = c n ∈ Z ∖ { 0 } C_n(0)=c_n \in\mathbb{Z}\setminus\{0\} C n ( 0 ) = c n ∈ Z ∖ { 0 } ; the other coefficients of C n C_n C n are differences of digits and lie in [ − W , W ] [-W,W] [ − W , W ] , so ∣ C n ( z ) − c n ∣ ≤ W z / ( 1 − z ) |C_n(z)-c_n|\le Wz/(1-z) ∣ C n ( z ) − c n ∣ ≤ W z / ( 1 − z ) for 0 < z < 1 0<z<1 0 < z < 1 . This bound is L < 2 L<2 L < 2 at t t t and tends to L L L at t n t_n t n , hence is also < 2 <2 < 2 there for large n n n . Since G n ≥ 0 G_n\ge0 G n ≥ 0 on ( 0 , 1 ) (0,1) ( 0 , 1 ) and F ( t ) = 0 F(t)=0 F ( t ) = 0 , we have C n ( t ) ≥ 0 C_n(t)\ge0 C n ( t ) ≥ 0 , hence c n ≥ − 1 c_n\ge-1 c n ≥ − 1 ; since F ≥ 0 F\ge0 F ≥ 0 and G n ( t n ) = 0 G_n(t_n)=0 G n ( t n ) = 0 , we have C n ( t n ) ≤ 0 C_n(t_n)\le0 C n ( t n ) ≤ 0 , hence c n ≤ 1 c_n\le1 c n ≤ 1 . Thus G n G_n G n has prefix π \pi π followed by c ± 1 c\pm1 c ± 1 , where π \pi π is the prefix of F F F of length M n M_n M n , and G n G_n G n is a completion of this prefix with minimum 0. We show that for large M n M_n M n no such completion exists.
If c = m + q − 1 c=m+q-1 c = m + q − 1 , the smallest completion of ( π , c + 1 ) (\pi,c+1) ( π , c + 1 ) is
F + z M n ( 1 − ( q − 1 ) z 1 − z ) . F+z^{M_n}\left(1-(q-1)\frac{z}{1-z}\right). F + z M n ( 1 − ( q − 1 ) 1 − z z ) .
Its second factor is positive for z < 1 / q z<1/q z < 1/ q , in particular near t t t , while away from t t t the term F ≥ 0 F\ge0 F ≥ 0 dominates the perturbation for large M n M_n M n ; so this completion is positive on I 0 I_0 I 0 and the prefix ( π , c + 1 ) (\pi,c+1) ( π , c + 1 ) has no completion with minimum 0. The largest completion of ( π , c − 1 ) (\pi,c-1) ( π , c − 1 ) is
F − z M n ( 1 − ( q − 1 ) z 1 − z ) , F-z^{M_n}\left(1-(q-1)\frac{z}{1-z}\right), F − z M n ( 1 − ( q − 1 ) 1 − z z ) ,
which is negative at t t t . If c = m c=m c = m , only the digit c + 1 c+1 c + 1 is available, and the smallest completion
is positive on I 0 I_0 I 0 . If c = h c=h c = h , only c − 1 c-1 c − 1 is available, and the largest completion
is negative at t t t . In all cases G n G_n G n cannot exist for large n n n .
For the last assertion, let F ∗ F_* F ∗ be live. Every ascending level N N N produces a witness F N ≠ F ∗ F_N\ne F_* F N = F ∗ with prefix π N ′ \pi'_N π N ′ , as in the proof of Theorem 6.7. Every sufficiently late descending level N N N of margin at least η \eta η produces a witness with prefix ( π N , a N ( F ∗ ) − 1 ) (\pi_N,a_N(F_*)-1) ( π N , a N ( F ∗ ) − 1 ) by Lemma 4.6: the largest completion of this prefix has value t ∗ N ( − 1 + L − S N ) ≥ η t ∗ N t_*^N(-1+L-S_N)\ge\eta t_*^N t ∗ N ( − 1 + L − S N ) ≥ η t ∗ N at t ∗ t_* t ∗ . Indeed, the largest completion is F ∗ + z N C N ( z ) F_*+z^NC_N(z) F ∗ + z N C N ( z ) , where C N ( t ∗ ) ≥ η C_N(t_*)\ge\eta C N ( t ∗ ) ≥ η and the coefficients of C N C_N C N are uniformly bounded. The derivative bound on a compact interval about t ∗ t_* t ∗ gives C N ≥ η / 2 C_N\ge\eta/2 C N ≥ η /2 on a fixed smaller neighbourhood. Off that neighbourhood in [ t − , t + ] [t_-,t_+] [ t − , t + ] , F ∗ F_* F ∗ has a positive minimum and z N C N z^NC_N z N C N tends uniformly to zero. Thus the largest completion is positive throughout [ t − , t + ] [t_-,t_+] [ t − , t + ] for large N N N ; the smallest is negative at t ∗ t_* t ∗ . By Lemma 6.6, infinitely many levels are of one of these kinds, and distinct levels give distinct witnesses. Conversely, if every square-adjacent witness is exceptional, the compact set of square-adjacent witnesses consists of isolated points and is therefore finite.
Of the three exceptional types, only the middle one can make the anchor set finite: a witness with constant tail m m m or h h h always has a neighbour in the tree of prefixes.
Lemma 6.9 (Extreme tails). Let F F F be a square-adjacent witness that is eventually constant with digit m m m , respectively h h h , and let N N N be the last level at which its digit differs from that constant. Then there is a square-adjacent witness G G G that agrees with F F F below degree N N N , has a N ( G ) = a N ( F ) − 1 a_N(G)=a_N(F)-1 a N ( G ) = a N ( F ) − 1 , respectively a N ( G ) = a N ( F ) + 1 a_N(G)=a_N(F)+1 a N ( G ) = a N ( F ) + 1 , and whose digits beyond level N N N are not all equal.
Proof. The level N ≥ 2 N\ge2 N ≥ 2 exists, because the two constant series are M m M_m M m , which is negative on I 0 I_0 I 0 , and the largest member of Φ h \Phi_h Φ h , which is positive. Let t = 1 / β t=1/\beta t = 1/ β be the anchor of F F F . We apply Lemma 4.6 with the digit range [ m , h ] [m,h] [ m , h ] and J = [ t − , t + ] J=[t_-,t_+] J = [ t − , t + ] , where W z ≥ 1 − z Wz\ge1-z W z ≥ 1 − z , as in Lemma 6.4.
Suppose first that the constant is m m m , and let π \pi π be the prefix of F F F below degree N N N followed by the digit a N ( F ) − 1 ≥ m a_N(F)-1\ge m a N ( F ) − 1 ≥ m . Then
f π − = F − z N , f π + = F + z N ( W z 1 − z − 1 ) ≥ F ≥ 0 on J , f_\pi^- = F-z^N,\qquad
f_\pi^+ = F+z^N\left(\frac{Wz}{1-z}-1\right)\ge F\ge0\quad\text{on }J, f π − = F − z N , f π + = F + z N ( 1 − z W z − 1 ) ≥ F ≥ 0 on J ,
while f π − ( t ) = − t N < 0 f_\pi^-(t)=-t^N<0 f π − ( t ) = − t N < 0 and f π − ≥ M m f_\pi^-\ge M_m f π − ≥ M m , which vanishes at t + t_+ t + . The lemma gives a witness G G G with prefix π \pi π ; let t ′ = 1 / β ′ t'=1/\beta' t ′ = 1/ β ′ be its anchor. If a j ( G ) = c a_j(G)=c a j ( G ) = c for all j > N j>N j > N , then
G = F + z N ( ( c − m ) z 1 − z − 1 ) . G=F+z^N\left(\frac{(c-m)z}{1-z}-1\right). G = F + z N ( 1 − z ( c − m ) z − 1 ) .
By Lemma 6.2 , q − 1 < β − 1 < q q-1<\beta-1<q q − 1 < β − 1 < q and q − 1 < β ′ − 1 < q q-1<\beta'-1<q q − 1 < β ′ − 1 < q . If c − m ≤ q − 1 c-m\le q-1 c − m ≤ q − 1 , then G ( t ) = t N ( ( c − m ) / ( β − 1 ) − 1 ) < 0 G(t)=t^{N((c-m)/(\beta-1)-1)}<0 G ( t ) = t N (( c − m ) / ( β − 1 ) − 1 ) < 0 , which is impossible for a witness. If c − m ≥ q c-m\ge q c − m ≥ q , then G ( t ′ ) ≥ t ′ N ( ( c − m ) / ( β ′ − 1 ) − 1 ) > 0 G(t')\ge t'^{N((c-m)/(\beta'-1)-1)}>0 G ( t ′ ) ≥ t ′ N (( c − m ) / ( β ′ − 1 ) − 1 ) > 0 , contradicting G ( t ′ ) = 0 G(t')=0 G ( t ′ ) = 0 .
If the constant is h h h , let π \pi π be the prefix of F F F below degree N N N followed by the digit a N ( F ) + 1 ≤ h a_N(F)+1\le h a N ( F ) + 1 ≤ h . Now f π + = F + z N > 0 f^+_\pi=F+z^N>0 f π + = F + z N > 0 , and f π − = F + z N ( 1 − W z / ( 1 − z ) ) f^-_\pi=F+z^N(1-Wz/(1-z)) f π − = F + z N ( 1 − W z / ( 1 − z )) has f π − ( t ) = t N ( 1 − L ) < 0 f^-_\pi(t)=t^N(1-L)<0 f π − ( t ) = t N ( 1 − L ) < 0 and f π − ≥ M m f^-_\pi\ge M_m f π − ≥ M m . The lemma gives a witness G G G with prefix π \pi π and anchor t ′ = 1 / β ′ t'=1/\beta' t ′ = 1/ β ′ . If a j ( G ) = c a_j(G)=c a j ( G ) = c for all j > N j>N j > N , then G = F + z N ( 1 − ( h − c ) z / ( 1 − z ) ) G=F+z^N(1-(h-c)z/(1-z)) G = F + z N ( 1 − ( h − c ) z / ( 1 − z )) , and either h − c ≥ q h-c\ge q h − c ≥ q and G ( t ) < 0 G(t)<0 G ( t ) < 0 , or h − c ≤ q − 1 h-c\le q-1 h − c ≤ q − 1 and G ( t ′ ) > 0 G(t')>0 G ( t ′ ) > 0 ; both are impossible.
Corollary 6.10 (A sufficient criterion for infinitude). Let q ≥ 4 q\ge4 q ≥ 4 and h ∈ { q 2 − 2 , q 2 − 1 , q 2 } h\in\{q^2-2,q^2-1,q^2\} h ∈ { q 2 − 2 , q 2 − 1 , q 2 } . If the anchor set T h T_h T h is infinite, or equivalently if some square-adjacent witness is not eventually constant with digit m m m , m + q − 1 m+q-1 m + q − 1 or h h h , then C h C_h C h has infinitely many connected components with nonempty planar interior, accumulating at a point of ( 1 / ( q + 1 ) , 1 / q ) (1/(q+1),1/q) ( 1/ ( q + 1 ) , 1/ q ) at which the core is not locally connected. In particular this holds if no square-adjacent witness at height h h h is eventually constant with digit m + q − 1 m+q-1 m + q − 1 .
Proof. For the first assertion combine Lemmas 6.5 and 6.8 with Theorem 6.7. For the second, suppose that no witness is eventually constant with digit m + q − 1 m+q-1 m + q − 1 , and let F 0 F_0 F 0 be a witness (Lemma 6.4 ). If F i F_i F i is exceptional, it is eventually constant with digit m m m or h h h by Lemma 6.5, and Lemma 6.9 gives a witness F i + 1 F_{i+1} F i + 1 that is live, or exceptional with a larger last nonconstant level than F i F_i F i . Thus some F i F_i F i is live, or the F i F_i F i are pairwise distinct; in both cases T h T_h T h is infinite.
Transfer between neighbouring heights
Shifting the height by one shifts the three exceptional digits by one, so an exceptional witness at one height is usually a live witness at a neighbouring height.
Lemma 6.11 (Transfer). Let q ≥ 4 q\ge4 q ≥ 4 and let F F F be a square-adjacent witness at height h = q 2 − 2 + k h=q^2-2+k h = q 2 − 2 + k .
(i) If k ≤ 1 k\le1 k ≤ 1 , every tail digit a j a_j a j (j ≥ 2 j\ge2 j ≥ 2 ) of F F F is at least m + 1 m+1 m + 1 , and F F F is exceptional at height h h h with constant digit m + q − 1 m+q-1 m + q − 1 or h h h , then F F F is a live square-adjacent witness at height h + 1 h+1 h + 1 .
(ii) If k ≥ 1 k\ge1 k ≥ 1 , every tail digit a j a_j a j (j ≥ 2 j\ge2 j ≥ 2 ) of F F F is at most h − 1 h-1 h − 1 , and F F F is exceptional at height h h h with constant digit m m m or m + q − 1 m+q-1 m + q − 1 , then F F F is a live square-adjacent witness at height h − 1 h-1 h − 1 .
Proof. In case (i), the digits of F F F lie in [ m + 1 , h + 1 ] [m+1,h+1] [ m + 1 , h + 1 ] , which is the digit range of the square-adjacent family at height h + 1 h+1 h + 1 , so F F F is a square-adjacent witness there. Its constant digit is m + q − 1 = ( m + 1 ) + q − 2 m+q-1=(m+1)+q-2 m + q − 1 = ( m + 1 ) + q − 2 or h = ( h + 1 ) − 1 h=(h+1)-1 h = ( h + 1 ) − 1 , neither of which is one of the three exceptional digits m + 1 m+1 m + 1 , ( m + 1 ) + q − 1 (m+1)+q-1 ( m + 1 ) + q − 1 and h + 1 h+1 h + 1 of height h + 1 h+1 h + 1 . By Lemma 6.5, applied at height h + 1 h+1 h + 1 , F F F is live there. Case (ii) is symmetric: the digits lie in [ m − 1 , h − 1 ] [m-1,h-1] [ m − 1 , h − 1 ] , and the constant digit m = ( m − 1 ) + 1 m=(m-1)+1 m = ( m − 1 ) + 1 or m + q − 1 = ( m − 1 ) + q m+q-1=(m-1)+q m + q − 1 = ( m − 1 ) + q is not among m − 1 m-1 m − 1 , ( m − 1 ) + q − 1 (m-1)+q-1 ( m − 1 ) + q − 1 and h − 1 h-1 h − 1 .
Theorem 6.12 (Two of three heights). Let q ≥ 4 q\ge4 q ≥ 4 . At least two of the three cores C q 2 − 2 C_{q^2-2} C q 2 − 2 , C q 2 − 1 C_{q^2-1} C q 2 − 1 and C q 2 C_{q^2} C q 2 have infinitely many connected components, containing infinitely many with nonempty planar interior, accumulating at a common point t ∗ ∈ ( 1 / ( q + 1 ) , 1 / q ) t_*\in(1/(q+1),1/q) t ∗ ∈ ( 1/ ( q + 1 ) , 1/ q ) and at − t ∗ -t_* − t ∗ , ± 1 / t ∗ \pm1/t_* ± 1/ t ∗ , where these cores are not locally connected. More precisely, if F F F is a square-adjacent witness with all digits in [ q 2 − 2 q + 2 , q 2 − 2 ] [q^2-2q+2,q^2-2] [ q 2 − 2 q + 2 , q 2 − 2 ] , then this holds at every height h ∈ { q 2 − 2 , q 2 − 1 , q 2 } h\in\{q^2-2,q^2-1,q^2\} h ∈ { q 2 − 2 , q 2 − 1 , q 2 } at which F F F is not eventually constant with digit m h m_h m h , m h + q − 1 m_h+q-1 m h + q − 1 or h h h , where m h = h − 2 q + 2 m_h=h-2q+2 m h = h − 2 q + 2 , and these three exceptional digits exclude at most one of the three heights.
Proof. Let F F F be a square-adjacent witness in the family Ψ q \Psi_q Ψ q of (25) ; it exists by Lemma 6.4 . Since [ q 2 − 2 q + 2 , q 2 − 2 ] [q^2-2q+2,q^2-2] [ q 2 − 2 q + 2 , q 2 − 2 ] is contained in the digit range of the square-adjacent family at each of the three heights, F F F is a square-adjacent witness at each of them. Within [ q 2 − 2 q + 2 , q 2 − 2 ] [q^2-2q+2,q^2-2] [ q 2 − 2 q + 2 , q 2 − 2 ] , the exceptional digits are
{ q 2 − q − 1 , q 2 − 2 } , { q 2 − q } , { q 2 − 2 q + 2 , q 2 − q + 1 } \{q^2-q-1,\ q^2-2\},\qquad\{q^2-q\},\qquad\{q^2-2q+2,\ q^2-q+1\} { q 2 − q − 1 , q 2 − 2 } , { q 2 − q } , { q 2 − 2 q + 2 , q 2 − q + 1 }
at the heights q 2 − 2 q^2-2 q 2 − 2 , q 2 − 1 q^2-1 q 2 − 1 and q 2 q^2 q 2 respectively. These sets are pairwise disjoint. If F F F is eventually constant, its constant digit belongs to at most one of them, and if it is not eventually constant it belongs to none. By Lemma 6.5, F F F is live at every height not excluded in this way, and Theorem 6.7 gives the conclusions there, with the common anchor t ∗ t_* t ∗ of F F F .
Pisot anchors and explicit witnesses
The witnesses of Lemma 6.4 need not be explicit. At an anchor whose reciprocal is a Pisot number, the witness, if it exists, is a rational function, and a finite computation decides whether it exists and is live. This is the counterpart of the theorem of Bertrand and Schmidt that, for a Pisot number β \beta β , the β \beta β -expansion of every element of Q ( β ) ∩ [ 0 , 1 ) \mathbb{Q}(\beta) \cap[0,1) Q ( β ) ∩ [ 0 , 1 ) is eventually periodic [5 , 27 ] ; here the value and the derivative equations are imposed together, and rigidity replaces the greedy algorithm.
Proposition 6.13 (Witnesses at Pisot anchors). Let β ∈ [ β − , β + ] \beta\in[\beta_{-},\beta_{+}] β ∈ [ β − , β + ] be a Pisot number, an algebraic integer all of whose other conjugates lie in the open unit disk, and let F F F be a square-adjacent witness with anchor 1 / β 1/\beta 1/ β . Then the digit sequence of F F F is eventually periodic, so F F F is a rational function, and F F F is live unless its digits are eventually constant and equal to m m m , m + q − 1 m+q-1 m + q − 1 or h h h . Moreover the digit sequence is the unique infinite path from the initial state in a finite graph determined by β \beta β , m m m and h h h ; in particular, whether a square-adjacent witness with anchor 1 / β 1/\beta 1/ β exists, and whether it is live, is decided by a finite computation.
Proof. Write t = 1 / β t=1/\beta t = 1/ β and a 0 = 1 a_0=1 a 0 = 1 , a 1 = − 2 q a_1=-2q a 1 = − 2 q , and consider the partial sums
P N = ∑ j ≤ N a j β N − j , V N = ∑ j ≤ N ( N + 1 − j ) a j β N − j ( N ≥ 1 ) , P_N=\sum_{j\leq N}a_j\beta^{N-j},\qquad V_N=\sum_{j\leq N}(N+1-j)a_j\beta^{N-j}\qquad(N\geq1), P N = j ≤ N ∑ a j β N − j , V N = j ≤ N ∑ ( N + 1 − j ) a j β N − j ( N ≥ 1 ) ,
which satisfy P N + 1 = β P N + a N + 1 P_{N+1}=\beta P_N+a_{N+1} P N + 1 = β P N + a N + 1 and V N + 1 = β V N + P N + 1 V_{N+1}=\beta V_N+P_{N+1} V N + 1 = β V N + P N + 1 , and lie in Z [ β ] \mathbb{Z}[\beta] Z [ β ] . The equations F ( t ) = 0 F(t)=0 F ( t ) = 0 and t F ′ ( t ) = 0 tF'(t)=0 t F ′ ( t ) = 0 give
P N = − ∑ i ≥ 1 a N + i t i , V N = ∑ i ≥ 1 ( i − 1 ) a N + i t i (31) P_N=-\sum_{i\geq1}a_{N+i}t^i,\qquad V_N=\sum_{i\geq1}(i-1)a_{N+i}t^i \tag*{(31)} P N = − i ≥ 1 ∑ a N + i t i , V N = i ≥ 1 ∑ ( i − 1 ) a N + i t i (31)
so ∣ P N ∣ ≤ h / ( β − 1 ) |P_N|\leq h/(\beta-1) ∣ P N ∣ ≤ h / ( β − 1 ) and ∣ V N ∣ ≤ h / ( β − 1 ) 2 |V_N|\leq h/(\beta-1)^2 ∣ V N ∣ ≤ h / ( β − 1 ) 2 . For an embedding σ \sigma σ of Q ( β ) \mathbb{Q}(\beta) Q ( β ) with ∣ σ ( β ) ∣ < 1 |\sigma(\beta)|<1 ∣ σ ( β ) ∣ < 1 , the defining sums give ∣ σ ( P N ) ∣ ≤ h / ( 1 − ∣ σ ( β ) ∣ ) |\sigma(P_N)|\leq h/(1-|\sigma(\beta)|) ∣ σ ( P N ) ∣ ≤ h / ( 1 − ∣ σ ( β ) ∣ ) and ∣ σ ( V N ) ∣ ≤ h / ( 1 − ∣ σ ( β ) ∣ ) 2 |\sigma(V_N)|\leq h/(1-|\sigma(\beta)|)^2 ∣ σ ( V N ) ∣ ≤ h / ( 1 − ∣ σ ( β ) ∣ ) 2 . Thus P N P_N P N and V N V_N V N are algebraic integers of degree at most [ Q ( β ) : Q ] [\mathbb{Q}(\beta):\mathbb{Q}] [ Q ( β ) : Q ] with all conjugates bounded independently of N N N ; the coefficients of their minimal polynomials are bounded integers, so the pairs ( P N , V N ) (P_N,V_N) ( P N , V N ) range over a finite set, and some state repeats: ( P M , V M ) = ( P N , V N ) (P_M,V_M)=(P_N,V_N) ( P M , V M ) = ( P N , V N ) with N < M N<M N < M .
By (31) , the state at a level determines the two tail sums ∑ i ≥ 1 a N + i t i \sum_{i\geq1}a_{N+i}t^i ∑ i ≥ 1 a N + i t i and ∑ i ≥ 1 ( i − 1 ) a N + i t i \sum_{i\geq1}(i-1)a_{N+i}t^i ∑ i ≥ 1 ( i − 1 ) a N + i t i , and conversely the value and the derivative of a series at t t t depend on its digits beyond level M M M only through these two sums. Hence the series G G G with digits a 1 , … , a M a_1,\ldots,a_M a 1 , … , a M followed by a N + 1 , a N + 2 , … a_{N+1},a_{N+2},\ldots a N + 1 , a N + 2 , … satisfies G ( t ) = F ( t ) = 0 G(t)=F(t)=0 G ( t ) = F ( t ) = 0 and G ′ ( t ) = F ′ ( t ) = 0 G'(t)=F'(t)=0 G ′ ( t ) = F ′ ( t ) = 0 . It belongs to Φ h \Phi_h Φ h , so it is a square-adjacent witness with anchor t t t , and rigidity (Lemma 6.3 ) gives G = F G=F G = F : the digits of F F F are periodic from level N + 1 N+1 N + 1 on, with period M − N M-N M − N . The claim about liveness is Lemma 6.5. For the last assertion, let the states be the pairs ( P , V ) (P,V) ( P , V ) in the finite set just described, with an edge labelled a ∈ [ m , h ] a\in[m,h] a ∈ [ m , h ] from ( P , V ) (P,V) ( P , V ) to ( β P + a , β V + β P + a ) (\beta P+a,\beta V+\beta P+a) ( β P + a , β V + β P + a ) whenever the target satisfies the bounds stated after (31) . The digit sequence of a witness is an infinite path from ( P 1 , V 1 ) = ( β − 2 q , 2 β − 2 q ) (P_1,V_1)=(\beta-2q,2\beta-2q) ( P 1 , V 1 ) = ( β − 2 q , 2 β − 2 q ) . Conversely, along any infinite path the states are bounded, so the series with those digits satisfies ∑ j a j t j = lim N t N P N = 0 \sum_j a_jt^j=\lim_N t^NP_N=0 ∑ j a j t j = lim N t N P N = 0 and ∑ j j a j t j = lim N ( ( N + 1 ) ∑ j ≤ N a j t j − t N V N ) = 0 \sum_j ja_jt^j=\lim_N((N+1)\sum_{j\leq N}a_jt^j-t^NV_N)=0 ∑ j j a j t j = lim N (( N + 1 ) ∑ j ≤ N a j t j − t N V N ) = 0 ; it has a double zero at t t t , so it is a square-adjacent witness with anchor t t t , and rigidity makes it equal to F F F . Hence the path is unique, and a witness with anchor t t t exists if and only if some infinite path leaves the initial state, which is decided by the finite graph. □
The searches in the companion [13 ] enumerate Pisot numbers of bounded degree in [ β − , β + ] [\beta_{-},\beta_{+}] [ β − , β + ] with numerical root filters and follow the paths of the finite graph in floating-point arithmetic. The witnesses of Table 1 are then checked directly in exact arithmetic (Proposition S4.14 of the supplement), independently of the search.
Proposition 6.14 (Explicit eventually periodic witnesses). For each row of Table 1 , the series
F ( z ) = 1 − 2 q z + ∑ j = 2 N − 1 a j z j + z N ∑ i = 0 r − 1 c i z i 1 − z r F(z)=1-2qz+\sum_{j=2}^{N-1}a_jz^j+\frac{z^N\sum_{i=0}^{r-1}c_iz^i}{1-z^r} F ( z ) = 1 − 2 q z + j = 2 ∑ N − 1 a j z j + 1 − z r z N ∑ i = 0 r − 1 c i z i is a live square-adjacent witness at each of the listed heights, with anchor t = 1 / β t=1/\beta t = 1/ β , where β \beta β is the Pisot number with the stated minimal polynomial. Consequently the conclusions of Theorem 6.7 hold at the heights
14 , 15 , 16 , 23 , 24 , 34 , 35 , 36 , 47 , 48 , 62 , 80 , 98 , 119 , 120 , 194. 14,\ 15,\ 16,\ 23,\ 24,\ 34,\ 35,\ 36,\ 47,\ 48,\ 62,\ 80,\ 98,\ 119,\ 120,\ 194. 14 , 15 , 16 , 23 , 24 , 34 , 35 , 36 , 47 , 48 , 62 , 80 , 98 , 119 , 120 , 194.
Proof. Write P P P for the polynomial part and T T T for the period polynomial, so that ( 1 − z r ) F = R (1-z^r)F=R ( 1 − z r ) F = R with R = ( 1 − z r ) P + z N T ∈ Z [ z ] R=(1-z^r)P+z^N T\in\mathbb{Z}[z] R = ( 1 − z r ) P + z N T ∈ Z [ z ] . A double zero of F F F in ( 0 , 1 ) (0,1) ( 0 , 1 ) is the same as a double zero of R R R there. For each row, the greatest common divisor g g g of R R R and R ′ R' R ′ is the reversed form of the stated polynomial, it is irreducible, g 2 g^2 g 2 divides R R R , and g g g has exactly one zero in ( 1 / ( q + 1 ) , 1 / q ) (1/(q+1),1/q) ( 1/ ( q + 1 ) , 1/ q ) ; these are finite exact computations (Proposition S4.14 of the supplement), which also certify that β \beta β is a Pisot number, although the argument does not use this. All digits lie in the digit range of each listed height, and the period is not constant, so the witness is not eventually constant. Lemma 6.5 shows that it is live at each listed height, and Theorem 6.7 applies. □
The table also exhibits explicit irrational singleton components. For the row q = 6 q=6 q = 6 , with β ≈ 6.44513 \beta\approx6.44513 β ≈ 6.44513 and m h = h − 10 m_h=h-10 m h = h − 10 , every late digit 30 is followed by 34, so at that level a N = 30 ≤ h − 1 a_N=30\le h-1 a N = 30 ≤ h − 1 and S N ≥ ( 34 − m h ) / β ≥ 8 / β > 1 S_N\ge(34-m_h)/\beta\ge8/\beta>1 S N ≥ ( 34 − m h ) / β ≥ 8/ β > 1 : infinitely many levels are ascending at each of the heights 34, 35, 36, and by Theorem 6.7 the four points ± β ± 1 \pm\beta^{\pm1} ± β ± 1 are singleton components of C 34 C_{34} C 34 , C 35 C_{35} C 35 and C 36 C_{36} C 36 . Each row lists exactly the heights whose digit range contains all of its digits.
Witnesses of the form g ( z ) 2 / ( 1 − z r ) g(z)^2/(1-z^r) g ( z ) 2 / ( 1 − z r )
Every witness of Table 1 is a quotient g ( z ) 2 / ( 1 − z r ) g(z)^2/(1-z^r) g ( z ) 2 / ( 1 − z r ) with g ∈ Z [ z ] g\in\mathbb{Z}[z] g ∈ Z [ z ] : in the proof of Proposition 6.14, R = g 2 R=g^2 R = g 2 . For such a quotient the double zero is visible, and the verification is a polynomial identity: if g 2 = ∑ j c j z j g^2=\sum_j c_jz^j g 2 = ∑ j c j z j , the coefficients of the quotient are a j = c j + a j − r a_j=c_j+a_{j-r} a j = c j + a j − r , with a j = c j a_j=c_j a j = c j for j < r j<r j < r , and they repeat with period r r r beyond the degree of g 2 g^2 g 2 . The next proposition lists fifteen further series of this form. They come from an exhaustive search over the polynomials g g g of bounded degree, in integer arithmetic: the condition m ≤ a j ≤ h m\le a_j\le h m ≤ a j ≤ h confines each coefficient of g g g in turn to at most q q q consecutive integers. The same search recovers every row of Table 1 .
Proposition 6.15 (Explicit witnesses that are squares). For each row of the following list, let g ( z ) = z d p ( 1 / z ) g(z)=z^dp(1/z) g ( z ) = z d p ( 1/ z ) , where d d d is the degree of p p p . The series F = g 2 / ( 1 − z r ) F=g^2/(1-z^r) F = g 2 / ( 1 − z r ) equals 1 − 2 q z + ∑ j ≥ 2 a j z j 1-2qz+\sum_{j\ge2}a_jz^j 1 − 2 q z + ∑ j ≥ 2 a j z j , where the integers a j a_j a j lie in the stated interval and are eventually periodic with the stated period, and it has a double zero at 1 / β 1/\beta 1/ β , where β \beta β is the only zero of p p p in ( 1 , ∞ ) (1,\infty) ( 1 , ∞ ) and q < β < q + 1 q<\beta<q+1 q < β < q + 1 .
Table 8 (PDF p. 33).
In the rows with q ≥ 7 q \ge7 q ≥ 7 , F F F is a live square-adjacent witness at each listed height. Consequently the conclusions of Theorem 6.7 hold at the heights 49, 64, 79, 98, 99, 100, 121, 142, 143, 144, 195, 223, 224, 225, 254 and 255.
Proof. The digits follow from the recursion for the a j a_j a j . In each row p ( q ) < 0 < p ( q + 1 ) p(q)<0<p(q+1) p ( q ) < 0 < p ( q + 1 ) , so g g g has a zero 1 / β 1/\beta 1/ β with q < β < q + 1 q<\beta<q+1 q < β < q + 1 , which is a double zero of F F F . Since the a j a_j a j are positive, F ′ ′ > 0 F''>0 F ′′ > 0 on ( 0 , 1 ) (0,1) ( 0 , 1 ) , so this is the only zero of F F F , hence of g g g , in ( 0 , 1 ) (0,1) ( 0 , 1 ) . In the rows with q ≥ 7 q\ge7 q ≥ 7 the stated interval is contained in [ h − 2 q + 2 , h ] [h-2q+2,h] [ h − 2 q + 2 , h ] for each listed height h h h , so F F F is a square-adjacent witness there. Its period is not constant, so it is live by Lemma 6.5, and Theorem 6.7 applies. □
The series with q = 12 q=12 q = 12 is a witness common to its three heights, as in Theorem 6.12 .
For q = 3 q=3 q = 3 the inequalities (4.2) hold only on part of the interval of possible anchors, and the results of this section do not apply to the whole family Φ h \Phi_h Φ h . Proposition 4.4 needs them at a single anchor, and the first two series of Proposition 6.15 satisfy all of its hypotheses.
Proposition 6.16 (Heights seven and eight). For h ∈ { 7 , 8 } h\in\{7,8\} h ∈ { 7 , 8 } let Γ h \Gamma_h Γ h be the series of Proposition 6.15 at height h h h ,
Γ 7 ( z ) = ( 1 − 3 z − z 2 − z 3 − z 4 − z 5 + z 6 ) 2 1 − z 5 , Γ 8 ( z ) = ( 1 − 3 z − 2 z 2 + z 3 ) 2 1 − z 2 , \Gamma_7(z)=\frac{(1-3z-z^2-z^3-z^4-z^5+z^6)^2}{1-z^5},\qquad\Gamma_8(z)=\frac{(1-3z-2z^2+z^3)^2}{1-z^2}, Γ 7 ( z ) = 1 − z 5 ( 1 − 3 z − z 2 − z 3 − z 4 − z 5 + z 6 ) 2 , Γ 8 ( z ) = 1 − z 2 ( 1 − 3 z − 2 z 2 + z 3 ) 2 , and 1 / β h 1/\beta_h 1/ β h its double zero, so that β 7 = 3.40974 … \beta_7=3.40974\ldots β 7 = 3.40974 … and β 8 = 3.49086 … \beta_8=3.49086\ldots β 8 = 3.49086 … . Then Γ h ∈ F h \Gamma_h\in\mathcal{F}_h Γ h ∈ F h , and infinitely many connected components of C h C_h C h with nonempty planar interior accumulate at each of the four points ± β h ± 1 \pm\beta_h^{\pm1} ± β h ± 1 , at which C h C_h C h is not locally connected.
Proof. Expanding the squares and dividing gives*
Γ 7 ( z ) = 1 − 6 z + 7 z 2 + 4 z 3 + 5 z 4 + 7 z 5 + 5 z 6 + 5 z 7 + z 8 ( 5 + 5 z + 6 z 2 + 3 z 3 + 6 z 4 ) 1 − z 5 , \Gamma_7(z)=1-6z+7z^2+4z^3+5z^4+7z^5+5z^6+5z^7+\frac{z^8(5+5z+6z^2+3z^3+6z^4)}{1-z^5}, Γ 7 ( z ) = 1 − 6 z + 7 z 2 + 4 z 3 + 5 z 4 + 7 z 5 + 5 z 6 + 5 z 7 + 1 − z 5 z 8 ( 5 + 5 z + 6 z 2 + 3 z 3 + 6 z 4 ) , Γ 8 ( z ) = 1 − 6 z + 6 z 2 + 8 z 3 + 4 z 4 + z 5 ( 4 + 5 z ) 1 − z 2 . \Gamma_8(z)=1-6z+6z^2+8z^3+4z^4+\frac{z^5(4+5z)}{1-z^2}. Γ 8 ( z ) = 1 − 6 z + 6 z 2 + 8 z 3 + 4 z 4 + 1 − z 2 z 5 ( 4 + 5 z ) . Thus Γ 7 ∈ F 7 \Gamma_7\in\mathcal{F}_7 Γ 7 ∈ F 7 with all coefficients of degree at least two in [ 3 , 7 ] [3,7] [ 3 , 7 ] , and Γ 8 ∈ F 8 \Gamma_8\in\mathcal{F}_8 Γ 8 ∈ F 8 with all such coefficients in [ 4 , 8 ] [4,8] [ 4 , 8 ] ; in both cases ℓ = 4 \ell=4 ℓ = 4 . The two polynomials p p p change sign on ( 3.409 , 3.41 ) (3.409,3.41) ( 3.409 , 3.41 ) and on ( 3.49 , 3.491 ) (3.49,3.491) ( 3.49 , 3.491 ) respectively, so
3.409 < β 7 < 3.41 , 3.49 < β 8 < 3.491. 3.409<\beta_7<3.41,\qquad3.49<\beta_8<3.491. 3.409 < β 7 < 3.41 , 3.49 < β 8 < 3.491. For β \beta β in either interval and the corresponding height,
ℓ β − 1 < 1.67 , ( β − 1 ) 2 > 5.8 > ℓ , 2 β − ℓ − 2 h ( β − 1 ) 2 > 0.39 , \frac{\ell}{\beta-1}<1.67,\qquad(\beta-1)^2>5.8>\ell,\qquad2\beta-\ell-\frac{2h}{(\beta-1)^2}>0.39, β − 1 ℓ < 1.67 , ( β − 1 ) 2 > 5.8 > ℓ , 2 β − ℓ − ( β − 1 ) 2 2 h > 0.39 , so the inequalities (4.2) hold, and t h = 1 / β h < 0.294 < σ h t_h=1/\beta_h<0.294<\sigma_h t h = 1/ β h < 0.294 < σ h , because 0.34 4 < 1 / 68 0.34^4<1/68 0.3 4 4 < 1/68 and 0.37 4 < 1 / 53 0.37^4<1/53 0.3 7 4 < 1/53 . It remains to verify (4.4). For Γ 7 \Gamma_7 Γ 7 take the indices N ≥ 10 N\ge10 N ≥ 10 divisible by 5: then a N = 6 > 3 a_N=6>3 a N = 6 > 3 , the next two coefficients are 3 and 6, and the sum in (4.4) is at least 4 t 7 + t 7 2 > 5 / 4 4t_7+t_7^2>5/4 4 t 7 + t 7 2 > 5/4 . For Γ 8 \Gamma_8 Γ 8 take the even indices N ≥ 6 N\ge6 N ≥ 6 : then a N = 5 > 4 a_N=5>4 a N = 5 > 4 , the next two coefficients are 4 and 5, and the sum is at least 4 t 8 + 3 t 8 2 > 5 / 4 4t_8+3t_8^2>5/4 4 t 8 + 3 t 8 2 > 5/4 . Proposition 4.4 applies with h ∗ = H = h h_* = H=h h ∗ = H = h and η = 1 / 4 \eta=1/4 η = 1/4 . □
Some of these series belong to families. The series Γ 8 \Gamma_8 Γ 8 is the case q = 3 q=3 q = 3 of ( 1 − q z − q + 1 2 z 2 + z 3 ) / ( 1 − z 2 ) \left(1-qz-\frac{q+1}{2}z^2+z^3\right)/(1-z^2) ( 1 − q z − 2 q + 1 z 2 + z 3 ) / ( 1 − z 2 ) , q q q odd, whose coefficients lie in the digit range of a square-adjacent height exactly when 3 ≤ q ≤ 9 3 \le q \le9 3 ≤ q ≤ 9 ; for q = 5 , 7 , 9 q=5,7,9 q = 5 , 7 , 9 it gives the first row with q = 5 q=5 q = 5 and the rows with q = 7 q=7 q = 7 and q = 9 q=9 q = 9 of Table 1 . Likewise ( 1 − q z − q 2 z 2 + 3 z 3 − 2 z 4 + z 5 − z 6 + z 7 ) / ( 1 − z 2 ) \left(1-qz-\frac{q}{2}z^2+3z^3-2z^4+z^5-z^6+z^7\right)/(1-z^2) ( 1 − q z − 2 q z 2 + 3 z 3 − 2 z 4 + z 5 − z 6 + z 7 ) / ( 1 − z 2 ) is a live witness common to the three heights next to q q q for q = 22 , 24 , 26 q=22,24,26 q = 22 , 24 , 26 , by the same verification. Such families serve only finitely many q q q . Indeed r ≥ 2 r \ge2 r ≥ 2 , because a 1 = c 1 + 1 a_1=c_1+1 a 1 = c 1 + 1 is odd when r = 1 r=1 r = 1 ; and if g = q u + v g=qu+v g = q u + v , with fixed polynomials u , v u,v u , v and a fixed r ≥ 2 r \ge2 r ≥ 2 , gave square-adjacent witnesses g 2 / ( 1 − z r ) g^2/(1-z^r) g 2 / ( 1 − z r ) for infinitely many q q q , comparing the coefficients of q 2 q^2 q 2 would give u ( z ) 2 = z 2 ( 1 + z + ⋯ + z r − 1 ) u(z)^2=z^2(1+z+\cdots+z^{r-1}) u ( z ) 2 = z 2 ( 1 + z + ⋯ + z r − 1 ) , which is impossible because the zeros of 1 + z + ⋯ + z r − 1 1+z+\cdots+z^{r-1} 1 + z + ⋯ + z r − 1 are simple.
Real double zeros near the minimum
This section proves Theorem D. The real gaps it produces lie inside a window of width at most 10 − 87 10^{-87} 1 0 − 87 just right of the minimum of D h D_h D h . Near a point v v v of D h D_h D h that is the smallest double zero of series with a prescribed initial block P P P , the other double zeros of such series form groups , one for each later position at which a corner digit can be lowered, and each group is a scaled slice of a self-affine set (Theorem 7.7). At certain digit patterns the group that starts the next block lies alone in a window of relative width 2 μ 2\mu 2 μ ; then D h D_h D h has a gap just left of the smallest double zero v ′ v' v ′ of the lowered block, and v ′ v' v ′ has the same structure as v v v (Corollary 7.8). If the tail of v v v has infinitely many such patterns, the gaps accumulate at v v v . Otherwise we move to v ′ v' v ′ from the last one. The digits that follow the last pattern lie in an explicit set of small dimension, and a finite computation shows that the tail of v ′ v' v ′ again contains a pattern (Proposition 7.11). The process therefore never stops. The local analysis holds at every height; the heights 1 , 3 , 4 , 7 1,3,4,7 1 , 3 , 4 , 7 enter only through finitely many inequalities, certified in Section S4 of the supplement (Propositions S4.16–S4.19).
Corner digits, slices and greedy tails
Fix h ≥ 1 h \ge1 h ≥ 1 . For f = 1 + ∑ j ≥ 1 a j z j ∈ F h f=1+\sum_{j\ge1}a_jz^j \in\mathcal{F}_h f = 1 + ∑ j ≥ 1 a j z j ∈ F h write b j = h − a j ∈ { 0 , … , 2 h } b_j=h-a_j \in\{0,\ldots,2h\} b j = h − a j ∈ { 0 , … , 2 h } for its corner digits and p 0 ( t ) = 1 + h t / ( 1 − t ) p_0(t)=1+ht/(1-t) p 0 ( t ) = 1 + h t / ( 1 − t ) . Then
f ( t ) = p 0 ( t ) − ∑ j ≥ 1 b j t j , p k ( t ) = ∑ j > k b j t j − k , q k ( t ) = ∑ j > k ( j − k ) b j t j − k (32) f(t)=p_0(t)-\sum_{j\ge1}b_jt^j,\qquad p_k(t)=\sum_{j>k}b_jt^{j-k},\qquad q_k(t)=\sum_{j>k}(j-k)b_jt^{j-k} \tag*{(32)} f ( t ) = p 0 ( t ) − j ≥ 1 ∑ b j t j , p k ( t ) = j > k ∑ b j t j − k , q k ( t ) = j > k ∑ ( j − k ) b j t j − k (32)
define the remainders and moments of f f f after k k k digits. For t ∈ ( 0 , 1 ) t \in(0,1) t ∈ ( 0 , 1 ) and x ≥ 0 x \ge0 x ≥ 0 let
S t ( x ) = { ∑ r ≥ 1 c r t r : c r ∈ { 0 , … , 2 h } , ∑ r ≥ 1 c r t r = x } S_t(x)=\left\{\sum_{r\ge1}c_rt^r:c_r\in\{0,\ldots,2h\},\ \sum_{r\ge1}c_rt^r=x\right\} S t ( x ) = { r ≥ 1 ∑ c r t r : c r ∈ { 0 , … , 2 h } , r ≥ 1 ∑ c r t r = x }
be the set of moments of the digit sequences of weight x x x ; it is compact, and q k ( t ) ∈ S t ( p k ( t ) ) q_k(t) \in S_t(p_k(t)) q k ( t ) ∈ S t ( p k ( t )) . Splitting off the first digit gives
S t ( x ) = x + t ⋃ c = 0 2 h S t ( x / t − c ) (33) S_t(x)=x+t\bigcup_{c=0}^{2h}S_t(x/t-c) \tag*{(33)} S t ( x ) = x + t c = 0 ⋃ 2 h S t ( x / t − c ) (33)
so S t ( x ) S_t(x) S t ( x ) is the vertical slice over x x x of the attractor of the maps ( x , y ) ↦ ( t ( c + x ) , t ( c + x + y ) ) (x,y)\mapsto(t(c+x),t(c+x+y)) ( x , y ) ↦ ( t ( c + x ) , t ( c + x + y )) , 0 ≤ c ≤ 2 h 0 \le c \le2h 0 ≤ c ≤ 2 h , and S t ( x ) ⊂ [ 0 , 2 h t / ( 1 − t ) 2 ] S_t(x)\subset[0,2ht/(1-t)^2] S t ( x ) ⊂ [ 0 , 2 h t / ( 1 − t ) 2 ] . Put β = 1 / t \beta=1/t β = 1/ t and d ˉ = ⌊ β ⌋ \bar d=\lfloor\beta\rfloor d ˉ = ⌊ β ⌋ . The greedy expansion of x ∈ [ 0 , 2 ) x\in[0,2) x ∈ [ 0 , 2 ) is defined by x 0 = x x_0=x x 0 = x , g k = ⌊ β x k − 1 ⌋ g_k=\lfloor\beta x_{k-1}\rfloor g k = ⌊ β x k − 1 ⌋ , x k = β x k − 1 − g k x_k=\beta x_{k-1}-g_k x k = β x k − 1 − g k . Its remainders satisfy 0 ≤ x k < 1 0\le x_k<1 0 ≤ x k < 1 , so g k ≤ d ˉ g_k\le\bar d g k ≤ d ˉ for k ≥ 2 k\ge2 k ≥ 2 , while g 1 ≤ 2 β ≤ d ˉ + 1 g_1\le2\beta\le\bar d+1 g 1 ≤ 2 β ≤ d ˉ + 1 ; at each height considered here, we write g q ( x ) = ∑ r ≥ 1 g r r t g^q(x)=\sum_{r\ge1}g_rr^t g q ( x ) = ∑ r ≥ 1 g r r t for its moment and d β ( 1 ) d_\beta(1) d β ( 1 ) for the greedy expansion of 1. A finite word over { 0 , … , d ˉ } \{0,\ldots,\bar d\} { 0 , … , d ˉ } is weakly admissible if each of its suffixes is lexicographically at most the prefix of d β ( 1 ) d_\beta(1) d β ( 1 ) of the same length; every window of a greedy expansion of a number in [ 0 , 1 ) [0,1) [ 0 , 1 ) is weakly admissible [25 ] .
The minimum
For 0 < t ≤ 1 / ( h + 1 ) 0<t\le1/(h+1) 0 < t ≤ 1/ ( h + 1 ) every f ∈ F h f\in\mathcal{F}_h f ∈ F h satisfies f ( t ) ≥ 1 − h t / ( 1 − t ) ≥ 0 f(t)\ge1-ht/(1-t)\ge0 f ( t ) ≥ 1 − h t / ( 1 − t ) ≥ 0 , with equality only for t = 1 / ( h + 1 ) t=1/(h+1) t = 1/ ( h + 1 ) and f = 1 − h z / ( 1 − z ) f=1-hz/(1-z) f = 1 − h z / ( 1 − z ) , whose derivative does not vanish there; hence D h ⊂ ( 1 / ( h + 1 ) , 1 ) D_h\subset(1/(h+1),1) D h ⊂ ( 1/ ( h + 1 ) , 1 ) , and D h D_h D h is relatively closed by coefficient compactness. For h ∈ { 1 , 3 , 4 , 7 } h\in\{1,3,4,7\} h ∈ { 1 , 3 , 4 , 7 } it is nonempty, and Proposition S4.16 gives rationals a h < b h a_h<b_h a h < b h , with b h − a h = 2 ⋅ 10 − 150 b_h-a_h=2\cdot10^{-150} b h − a h = 2 ⋅ 1 0 − 150 for h = 1 h=1 h = 1 and 2 ⋅ 10 − 90 2\cdot10^{-90} 2 ⋅ 1 0 − 90 otherwise, such that
t 0 : = min D h ∈ [ a h , b h ] , t_0:=\min D_h\in[a_h,b_h], t 0 := min D h ∈ [ a h , b h ] ,
together with digit words P 0 P_0 P 0 of lengths n 0 = 49 , 24 , 19 , 15 n_0=49,24,19,15 n 0 = 49 , 24 , 19 , 15 and longer words of lengths N h = 850 , 257 , 216 , 162 N_h=850,257,216,162 N h = 850 , 257 , 216 , 162 with the following properties: every series of F h \mathcal{F}_h F h with a double zero in W h = W_h= W h = [ a h − 10 − 12 , b h + 10 − 9 ] [a_h - 10^{-12}, b_h + 10^{-9}] [ a h − 1 0 − 12 , b h + 1 0 − 9 ] begins with P 0 P_0 P 0 , and every series with a double zero in [ a h , b h ] [a_h,b_h] [ a h , b h ] begins with the long word. Table 2 lists the data.
Table 2. Data of the real-trace argument (Propositions S4.16–S4.19). Lower bounds are rounded down and upper bounds up.
For a word P ⊇ P 0 P \supseteq P_0 P ⊇ P 0 of length n n n let F h ( P ) \mathcal{F}_h(P) F h ( P ) be the series of F h \mathcal{F}_h F h that begin with P P P , let D h ( P ) \mathcal{D}_h(P) D h ( P ) be the set of double zeros in W h W_h W h of members of F h ( P ) \mathcal{F}_h(P) F h ( P ) , and, when it is nonempty, m ( P ) = min D h ( P ) m(P)=\min\mathcal{D}_h(P) m ( P ) = min D h ( P ) , a relative tip . Thus t 0 = m ( P 0 ) t_0=m(P_0) t 0 = m ( P 0 ) . Fix v = m ( P ) v=m(P) v = m ( P ) with ∣ v − t 0 ∣ ≤ 10 − 80 |v-t_0|\leq10^{-80} ∣ v − t 0 ∣ ≤ 1 0 − 80 and a series f v ∈ F h ( P ) f_v\in\mathcal{F}_h(P) f v ∈ F h ( P ) with a double zero at v v v , write b j , p k , q k b_j,p_k,q_k b j , p k , q k for its corner digits, remainders and moments at t = v t=v t = v , and call g s = b n + s g_s=b_{n+s} g s = b n + s , y s = p n + s y_s=p_{n+s} y s = p n + s and q s q_s q s (abusing notation) the digits, remainders and moments of its tail at the levels s ≥ 1 s\geq1 s ≥ 1 . Put κ v = f v ′ ′ ( v ) / 2 \kappa_v=f_v^{\prime\prime}(v)/2 κ v = f v ′′ ( v ) /2 and u P = v n − 1 / ( 2 κ v ) u_P=v^{n-1}/(2\kappa_v) u P = v n − 1 / ( 2 κ v ) . The free radius F ( P ) F(P) F ( P ) is the supremum of the F ≥ 0 F\geq0 F ≥ 0 such that ( v , v + F u P ) ⊂ W h (v,v+Fu_P)\subset W_h ( v , v + F u P ) ⊂ W h and every series of F h \mathcal{F}_h F h with a double zero in ( v , v + F u P ) (v,v+Fu_P) ( v , v + F u P ) belongs to F h ( P ) \mathcal{F}_h(P) F h ( P ) . The window property gives F ( P 0 ) ≥ 10 − 9 / u P 0 F(P_0)\geq10^{-9}/u_{P_0} F ( P 0 ) ≥ 1 0 − 9 / u P 0 .
Local structure at a relative tip
Let R = 10 − 6 R=10^{-6} R = 1 0 − 6 , τ = b h + R \tau=b_h+R τ = b h + R , C 3 = h / ( 1 − τ ) 4 C_3=h/(1-\tau)^4 C 3 = h / ( 1 − τ ) 4 , G 1 ( i ) = 2 h ∑ k ≥ i k τ k − 1 G_1(i)=2h\sum_{k\geq i}k\tau^{k-1} G 1 ( i ) = 2 h ∑ k ≥ i k τ k − 1 , G 2 ( i ) = 2 h ∑ k ≥ i k ( k − 1 ) τ k − 2 G_2(i)=2h\sum_{k\geq i}k(k-1)\tau^{k-2} G 2 ( i ) = 2 h ∑ k ≥ i k ( k − 1 ) τ k − 2 and M ‾ = 2 h τ / ( 1 − τ ) 2 \overline{M}=2h\tau/(1-\tau)^2 M = 2 h τ / ( 1 − τ ) 2 . Then ∣ f ′ ′ ′ ∣ / 6 ≤ C 3 |f^{\prime\prime\prime}|/6\leq C_3 ∣ f ′′′ ∣/6 ≤ C 3 on [ v − R / 2 , v + R / 2 ] [v-R/2,v+R/2] [ v − R /2 , v + R /2 ] for every f ∈ F h f\in\mathcal{F}_h f ∈ F h (this interval lies in [ 0 , τ ) [0,\tau) [ 0 , τ ) , since v ≤ b h + 10 − 80 v\leq b_h+10^{-80} v ≤ b h + 1 0 − 80 ); a power series ∑ k ≥ i c k z k \sum_{k\geq i}c_kz^k ∑ k ≥ i c k z k with ∣ c k ∣ ≤ 2 h |c_k|\leq2h ∣ c k ∣ ≤ 2 h has first and second derivatives bounded by G 1 ( i ) G_1(i) G 1 ( i ) and G 2 ( i ) G_2(i) G 2 ( i ) on ∣ z ∣ ≤ τ |z|\leq\tau ∣ z ∣ ≤ τ ; every moment lies in [ 0 , M ‾ ] [0,\overline{M}] [ 0 , M ] . The series f v f_v f v shares P 0 P_0 P 0 with a double-zero series at t 0 t_0 t 0 , so ∣ κ v − κ 0 ∣ ≤ G 2 ( n 0 + 1 ) / 2 + 10 − 70 |\kappa_v-\kappa_0|\leq G_2(n_0+1)/2+10^{-70} ∣ κ v − κ 0 ∣ ≤ G 2 ( n 0 + 1 ) /2 + 1 0 − 70 , where κ 0 \kappa_0 κ 0 is enclosed from the certified digits. With W ( i ) = G 1 ( i ) / ( 2 κ v − 3 C 3 R − G 2 ( i ) ) W(i)=G_1(i)/(2\kappa_v-3C_3R-G_2(i)) W ( i ) = G 1 ( i ) / ( 2 κ v − 3 C 3 R − G 2 ( i )) , ζ i = ( 4 C 3 W ( i ) 3 + 3 2 G 2 ( i ) W ( i ) 2 ) / v i \zeta_i=(4C_3W(i)^3+\frac{3}{2}G_2(i)W(i)^2)/v^i ζ i = ( 4 C 3 W ( i ) 3 + 2 3 G 2 ( i ) W ( i ) 2 ) / v i and
η ‾ i = ( 1 + 10 − 3 ) M ‾ 2 v i − 2 4 κ v , L ∗ ( i ) = i − 4 − c L , \overline{\eta}_i=\frac{(1+10^{-3})\overline{M}^2v^{i-2}}{4\kappa_v},\qquad L^*(i)=i-4-c_L, η i = 4 κ v ( 1 + 1 0 − 3 ) M 2 v i − 2 , L ∗ ( i ) = i − 4 − c L ,
where c L c_L c L is the least integer with β c L ≥ M ‾ 2 / ( 4 κ v ) \beta^{c_L}\geq\overline{M}^2/(4\kappa_v) β c L ≥ M 2 / ( 4 κ v ) (so η ‾ i ≤ 1.001 v L ∗ ( i ) + 2 \overline{\eta}_i\leq1.001v^{L^*(i)+2} η i ≤ 1.001 v L ∗ ( i ) + 2 ), Proposition S4.17 certifies for all i > n 0 i>n_0 i > n 0 the inequalities
(C4) 2 κ v − 3 C 3 R − G 2 ( n 0 + 1 ) > 0 , (C1) W ( i ) ≤ R , (C2) ζ i < 10 − 3 , \text{(C4)}\quad2\kappa_v-3C_3R-G_2(n_0+1)>0,\qquad
\text{(C1)}\quad W(i)\leq R,\qquad
\text{(C2)}\quad\zeta_i<10^{-3}, (C4) 2 κ v − 3 C 3 R − G 2 ( n 0 + 1 ) > 0 , (C1) W ( i ) ≤ R , (C2) ζ i < 1 0 − 3 ,
(C5) η ‾ i ≤ 10 − 7 , (D) 3 C 3 W ( i ) + G 2 ( i ) 2 κ v − 3 C 3 W ( i ) − G 2 ( i ) ≤ 10 − 5 , \text{(C5)}\quad\overline{\eta}_i\leq10^{-7},\qquad
\text{(D)}\quad
\frac{3C_3W(i)+G_2(i)}{2\kappa_v-3C_3W(i)-G_2(i)}\leq10^{-5}, (C5) η i ≤ 1 0 − 7 , (D) 2 κ v − 3 C 3 W ( i ) − G 2 ( i ) 3 C 3 W ( i ) + G 2 ( i ) ≤ 1 0 − 5 ,
and the further conditions (C3), (C6), (C7) and (E1)–(E7) quoted where they are used. If f ∈ F h f\in\mathcal{F}_h f ∈ F h has corner digits b j ′ b'_j b j ′ , and i i i is the first index with b i ′ ≠ b i b'_i\ne b_i b i ′ = b i , then, with ε = f ( v ) \varepsilon=f(v) ε = f ( v ) and ε ′ = f ′ ( v ) \varepsilon'=f'(v) ε ′ = f ′ ( v ) ,
ε = v i [ ( b i − b i ′ ) + ( p i − p i ′ ) ] , ε ′ = v i − 1 [ i ε / v i + q i − q i ′ ] . (34) \varepsilon=v^i[(b_i-b'_i)+(p_i-p'_i)],\qquad
\varepsilon'=v^{i-1}[i\varepsilon/v^i+q_i-q'_i]. \tag*{(34)} ε = v i [( b i − b i ′ ) + ( p i − p i ′ )] , ε ′ = v i − 1 [ i ε / v i + q i − q i ′ ] . (34)
since f v ( v ) = f v ′ ( v ) = 0 f_v(v)=f'_v(v)=0 f v ( v ) = f v ′ ( v ) = 0 .
Lemma 7.1 (Taylor expansion at a tip). Let f f f be as above with i > n 0 i>n_0 i > n 0 , and suppose f f f has a double zero at v + w v+w v + w with ∣ w ∣ ≤ R / 2 |w|\leq R/2 ∣ w ∣ ≤ R /2 . Then ∣ w ∣ ≤ W ( i ) |w|\leq W(i) ∣ w ∣ ≤ W ( i ) , ∣ ε − κ v w 2 ∣ ≤ 4 C 3 ∣ w ∣ 3 + 3 2 G 2 ( i ) w 2 |\varepsilon-\kappa_vw^2|\leq4C_3|w|^3+\frac{3}{2}G_2(i)w^2 ∣ ε − κ v w 2 ∣ ≤ 4 C 3 ∣ w ∣ 3 + 2 3 G 2 ( i ) w 2 , and
w = v i − 1 ( q i ′ − q i − i ε / v i ) 2 κ v + ϑ , ∣ ϑ ∣ ≤ 3 C 3 ∣ w ∣ + G 2 ( i ) . w=\frac{v^{i-1}(q'_i-q_i-i\varepsilon/v^i)}{2\kappa_v+\vartheta},\qquad
|\vartheta|\leq3C_3|w|+G_2(i). w = 2 κ v + ϑ v i − 1 ( q i ′ − q i − i ε / v i ) , ∣ ϑ ∣ ≤ 3 C 3 ∣ w ∣ + G 2 ( i ) . Moreover 0 ≤ ε / v i ≤ η ‾ i 0\leq\varepsilon/v^i\leq\overline{\eta}_i 0 ≤ ε / v i ≤ η i , with equality on the left only for w = 0 w=0 w = 0 .
Proof. Write f = f v + g f=f_v+g f = f v + g with g = ∑ k ≥ i c k z k g=\sum_{k\geq i}c_kz^k g = ∑ k ≥ i c k z k , ∣ c k ∣ ≤ 2 h |c_k|\leq2h ∣ c k ∣ ≤ 2 h . Taylor’s formula for f v f_v f v at v v v and for g g g at v v v gives f ( v + w ) = ε + ε ′ w + κ v w 2 + E f(v+w)=\varepsilon+\varepsilon'w+\kappa_vw^2+E f ( v + w ) = ε + ε ′ w + κ v w 2 + E and f ′ ( v + w ) = ε ′ + 2 κ v w + E ′ f'(v+w)=\varepsilon'+2\kappa_vw+E' f ′ ( v + w ) = ε ′ + 2 κ v w + E ′ with ∣ E ∣ ≤ C 3 ∣ w ∣ 3 + 1 2 G 2 ( i ) w 2 |E|\leq C_3|w|^3+\frac{1}{2}G_2(i)w^2 ∣ E ∣ ≤ C 3 ∣ w ∣ 3 + 2 1 G 2 ( i ) w 2 and ∣ E ′ ∣ ≤ 3 C 3 w 2 + G 2 ( i ) ∣ w ∣ |E'|\leq3C_3w^2+G_2(i)|w| ∣ E ′ ∣ ≤ 3 C 3 w 2 + G 2 ( i ) ∣ w ∣ . At a double zero, ε ′ = − ( 2 κ v + ϑ ) w \varepsilon'=-(2\kappa_v+\vartheta)w ε ′ = − ( 2 κ v + ϑ ) w with ϑ = E ′ / w \vartheta=E'/w ϑ = E ′ / w and ε = κ v w 2 + E ′ w − E \varepsilon=\kappa_vw^2+E'w-E ε = κ v w 2 + E ′ w − E , which gives the second bound. Since ε ′ = g ′ ( v ) \varepsilon'=g'(v) ε ′ = g ′ ( v ) is at most G 1 ( i ) G_1(i) G 1 ( i ) in modulus and 2 κ v + ϑ ≥ 2 κ v − 3 C 3 R − G 2 ( i ) > 0 2\kappa_v+\vartheta\geq2\kappa_v-3C_3R-G_2(i)>0 2 κ v + ϑ ≥ 2 κ v − 3 C 3 R − G 2 ( i ) > 0 , we get ∣ w ∣ ≤ W ( i ) |w|\leq W(i) ∣ w ∣ ≤ W ( i ) . Inserting (7.3) in ε ′ = − ( 2 κ v + ϑ ) w \varepsilon'=-(2\kappa_v+\vartheta)w ε ′ = − ( 2 κ v + ϑ ) w gives the formula for w w w . By (C1) and (C2), ε ≥ w 2 ( κ v − 4 C 3 W ( i ) − 3 2 G 2 ( i ) ) ≥ 0 \varepsilon\geq w^2(\kappa_v-4C_3W(i)-\frac{3}{2}G_2(i))\geq0 ε ≥ w 2 ( κ v − 4 C 3 W ( i ) − 2 3 G 2 ( i )) ≥ 0 , with equality only for w = 0 w=0 w = 0 . For the upper bound, first ε / v i ≤ ( κ v + 4 C 3 W ( i ) + 3 2 G 2 ( i ) ) W ( i ) 2 / v i \varepsilon/v^i\leq(\kappa_v+4C_3W(i)+\frac{3}{2}G_2(i))W(i)^2/v^i ε / v i ≤ ( κ v + 4 C 3 W ( i ) + 2 3 G 2 ( i )) W ( i ) 2 / v i ; inserting this and ∣ q i ′ − q i ∣ ≤ M ‾ |q'_i-q_i|\leq\overline{M} ∣ q i ′ − q i ∣ ≤ M in the formula for w w w and then in ε ≤ ( κ v + 4 C 3 ∣ w ∣ + 3 2 G 2 ( i ) ) w 2 \varepsilon\leq(\kappa_v+4C_3|w|+\frac{3}{2}G_2(i))w^2 ε ≤ ( κ v + 4 C 3 ∣ w ∣ + 2 3 G 2 ( i )) w 2 gives ε / v i ≤ η ‾ i \varepsilon/v^i\leq\overline{\eta}_i ε / v i ≤ η i ; the product of the correction factors is at most 1 + 10 − 3 1+10^{-3} 1 + 1 0 − 3 by (C6). □ \square □
The next lemma is the Shmerkin–Solomyak criterion [28 Lemma 3.1 and Corollary 3.2] , written for height h h h .
Lemma 7.2 (Good prefixes). Let Q = 1 + ∑ j ≤ N c j z j Q=1+\sum_{j\leq N}c_jz^j Q = 1 + ∑ j ≤ N c j z j with c j ∈ [ − h , h ] ∩ Z c_j\in[-h,h]\cap\mathbb{Z} c j ∈ [ − h , h ] ∩ Z , put H ( y ) = h y N + 1 / ( 1 − y ) H(y)=hy^{N+1}/(1-y) H ( y ) = h y N + 1 / ( 1 − y ) , and let 1 / ( 2 h + 1 ) < a < b < 1 1/(2h+1)<a<b<1 1/ ( 2 h + 1 ) < a < b < 1 satisfy Q ( a ) > H ( a ) Q(a)>H(a) Q ( a ) > H ( a ) , Q ( b ) > H ( b ) Q(b)>H(b) Q ( b ) > H ( b ) , Q + H > 0 Q+H>0 Q + H > 0 on [ a , b ] [a,b] [ a , b ] and Q ( x ∗ ) < H ( x ∗ ) Q(x_*) < H(x_*) Q ( x ∗ ) < H ( x ∗ ) for some x ∗ ∈ ( a , b ) x_* \in(a,b) x ∗ ∈ ( a , b ) . Then a series of F h F_h F h with initial part Q Q Q has a double zero in ( a , b ) (a,b) ( a , b ) .
Proof. Choose d N + 1 , d N + 2 , … ∈ [ − h , h ] d_{N+1},d_{N+2},\ldots\in[-h,h] d N + 1 , d N + 2 , … ∈ [ − h , h ] successively, each as small as possible such that g k = Q + ∑ N < j ≤ k d j z j + h ∑ j > k z j g_k=Q+\sum_{N<j\leq k}d_jz^j+h\sum_{j>k}z^j g k = Q + ∑ N < j ≤ k d j z j + h ∑ j > k z j stays positive on [ a , b ] [a,b] [ a , b ] ; the choice d k = h d_k=h d k = h is allowed because it gives g k = g k − 1 g_k=g_{k-1} g k = g k − 1 , and g N = Q + H g_N=Q+H g N = Q + H . The g k g_k g k decrease to f = Q + ∑ j > N d j z j ∈ F h f=Q+\sum_{j>N}d_jz^j\in F_h f = Q + ∑ j > N d j z j ∈ F h , so f ≥ 0 f\geq0 f ≥ 0 on [ a , b ] [a,b] [ a , b ] , and f ( a ) ≥ Q ( a ) − H ( a ) > 0 f(a)\geq Q(a)-H(a)>0 f ( a ) ≥ Q ( a ) − H ( a ) > 0 , f ( b ) > 0 f(b)>0 f ( b ) > 0 . Suppose that m = min [ a , b ] f > 0 m=\min_{[a,b]}f>0 m = min [ a , b ] f > 0 . As g k ≥ f g_k\geq f g k ≥ f , lowering d k d_k d k by one preserves g k ≥ m − b k > 0 g_k\geq m-b^k>0 g k ≥ m − b k > 0 once b k < m b^k<m b k < m , so d k = − h d_k=-h d k = − h for all large $k; not all d j d_j d j equal − h -h − h , since otherwise f ( x ∗ ) = Q ( x ∗ ) − H ( x ∗ ) < 0 f(x_*)=Q(x_*)-H(x_*)<0 f ( x ∗ ) = Q ( x ∗ ) − H ( x ∗ ) < 0 . Let k 0 k_0 k 0 be the last index with d k 0 > − h d_{k_0}>-h d k 0 > − h . Lowering d k 0 d_{k_0} d k 0 was not allowed, so g k 0 ( x ) ≤ x k 0 g_{k_0}(x)\leq x^{k_0} g k 0 ( x ) ≤ x k 0 for some x ∈ [ a , b ] x\in[a,b] x ∈ [ a , b ] , and then f ( x ) = g k 0 ( x ) − 2 h ∑ j > k 0 x j ≤ x k 0 ( 1 − 2 h x / ( 1 − x ) ) < 0 f(x)=g_{k_0}(x)-2h\sum_{j>k_0}x^j\leq x^{k_0}(1-2hx/(1-x))<0 f ( x ) = g k 0 ( x ) − 2 h ∑ j > k 0 x j ≤ x k 0 ( 1 − 2 h x / ( 1 − x )) < 0 , because x > 1 / ( 2 h + 1 ) x>1/(2h+1) x > 1/ ( 2 h + 1 ) . This contradiction shows m = 0 m=0 m = 0 ; the minimum is attained in ( a , b ) (a,b) ( a , b ) , where f ′ f' f ′ vanishes too.
Lemma 7.3 (Realization). Let i > n i>n i > n , let b i ′ ≠ b i b'_i\ne b_i b i ′ = b i be a corner digit and c = b i − b i ′ c=b_i-b'_i c = b i − b i ′ , and let b r ′ ′ , … , b L ′ ′ b''_r,\ldots,b''_L b r ′′ , … , b L ′′ , with 0 ≤ L ≤ L ∗ ( i ) + 1 0\leq L\leq L^*(i)+1 0 ≤ L ≤ L ∗ ( i ) + 1 , be corner digits whose remaining weight r L = ( c + p i − ∑ r ≤ L b r ′ ′ v r ) / v L r_L=(c+p_i-\sum_{r\leq L}b''_rv^r)/v^L r L = ( c + p i − ∑ r ≤ L b r ′′ v r ) / v L lies in [ v + η ˉ i / v L , 1.7 ] [v+\bar{\eta}_i/v^L,1.7] [ v + η ˉ i / v L , 1.7 ] . Then some f ∈ F h ( P ) f\in F_h(P) f ∈ F h ( P ) that agrees with f v f_v f v before position i i i and has the corner digits b i ′ b'_i b i ′ at i i i and b r ′ ′ b''_r b r ′′ at i + r i+r i + r has a double zero at v + w v+w v + w with
w = v i − 1 ( q ′ ′ − q i + ϱ ) 2 κ v ( 1 + θ ′ ) , q ′ ′ = ∑ r ≤ L r b r ′ ′ v r , (35) w=\frac{v^{i-1}(q''-q_i+\varrho)}{2\kappa_v}(1+\theta'),\qquad q''=\sum_{r\leq L}rb''_rv^r, \tag*{(35)} w = 2 κ v v i − 1 ( q ′′ − q i + ϱ ) ( 1 + θ ′ ) , q ′′ = r ≤ L ∑ r b r ′′ v r , (35) where ∣ θ ′ ∣ ≤ 10 − 5 |\theta'|\leq10^{-5} ∣ θ ′ ∣ ≤ 1 0 − 5 and ∣ ϱ ∣ ≤ 2 h v L + 1 ( ( L + 5 ) / ( 1 − v ) + 1 / ( 1 − v ) 2 ) + η ˉ i |\varrho|\leq2hv^{L+1}\left((L+5)/(1-v)+1/(1-v)^2\right)+\bar{\eta}_i ∣ ϱ ∣ ≤ 2 h v L + 1 ( ( L + 5 ) / ( 1 − v ) + 1/ ( 1 − v ) 2 ) + η ˉ i .
Proof. Let N = i + L + 5 N=i+L+5 N = i + L + 5 and let Q Q Q run over the polynomials with constant term 1 1 1 , the coefficients of f v f_v f v below i i i , the corner digits b i ′ b'_i b i ′ and b r ′ ′ b''_r b r ′′ at i i i and i + r i+r i + r , free coefficients at the five positions i + L + 1 , … , N i+L+1,\ldots,N i + L + 1 , … , N , and no further terms. By (7.3) for Q Q Q ,
Q ( v ) = v i + L ( r L − ∑ j = 1 5 b i + L + j Q v j ) − H ( v ) , Q ′ ( v ) = v i − 1 ( i Q ( v ) / v i + q i − q i Q ) . Q(v)=v^{i+L}\left(r_L-\sum_{j=1}^{5}b^Q_{i+L+j}v^j\right)-H(v),\qquad Q'(v)=v^{i-1}\left(iQ(v)/v^i+q_i-q^Q_i\right). Q ( v ) = v i + L ( r L − j = 1 ∑ 5 b i + L + j Q v j ) − H ( v ) , Q ′ ( v ) = v i − 1 ( i Q ( v ) / v i + q i − q i Q ) .
where b Q b^Q b Q are the corner digits of Q Q Q and H H H is as in Lemma 7.2. On [ v − R / 2 , v + R / 2 ] [v-R/2,v+R/2] [ v − R /2 , v + R /2 ] one has ∣ f v ′ ′ − 2 κ v ∣ ≤ 6 C 3 R |f_v''-2\kappa_v|\leq6C_3R ∣ f v ′′ − 2 κ v ∣ ≤ 6 C 3 R and ∣ ( Q − f v ) ′ ′ ∣ ≤ G 2 ( i ) |(Q-f_v)''|\leq G_2(i) ∣ ( Q − f v ) ′′ ∣ ≤ G 2 ( i ) , so Q ′ ′ ≥ 2 κ v ( 1 − 10 − 4 ) Q''\geq2\kappa_v(1-10^{-4}) Q ′′ ≥ 2 κ v ( 1 − 1 0 − 4 ) by (E5); hence Q Q Q has a unique minimum point x Q x_Q x Q there, with ∣ x Q − v ∣ ≤ M ˉ v i − 1 / κ v |x_Q-v|\leq\bar{M}v^{i-1}/\kappa_v ∣ x Q − v ∣ ≤ M ˉ v i − 1 / κ v and Q ( v ) − min Q ≤ Q ′ ( v ) 2 / ( 4 κ v ( 1 − 10 − 4 ) ) ≤ ( 1 + 10 − 4 ) v i η ˉ i Q(v)-\min Q\leq Q'(v)^2/(4\kappa_v(1-10^{-4}))\leq(1+10^{-4})v^i\bar{\eta}_i Q ( v ) − min Q ≤ Q ′ ( v ) 2 / ( 4 κ v ( 1 − 1 0 − 4 )) ≤ ( 1 + 1 0 − 4 ) v i η ˉ i , by (C6). Write M ( c ) = min Q M(c)=\min Q M ( c ) = min Q for the choice c c c of the free coefficients.
By (E3), x k / v k ∈ [ 1 − 2 ⋅ 10 − 3 , 1 + 2 ⋅ 10 − 3 ] x^k/v^k\in[1-2\cdot10^{-3},1+2\cdot10^{-3}] x k / v k ∈ [ 1 − 2 ⋅ 1 0 − 3 , 1 + 2 ⋅ 1 0 − 3 ] for k ≤ N + 1 k\leq N+1 k ≤ N + 1 and x x x between v v v and any of the minimum points, so raising the coefficient at a free position k k k by one raises M M M by an amount in [ v k ( 1 − 2 ⋅ 10 − 3 ) , v k ( 1 + 2 ⋅ 10 − 3 ) ] [v^k(1-2\cdot10^{-3}),v^k(1+2\cdot10^{-3})] [ v k ( 1 − 2 ⋅ 1 0 − 3 ) , v k ( 1 + 2 ⋅ 1 0 − 3 )] ; and, because 2 h v > 1.01 2hv>1.01 2 h v > 1.01 (E7), raising it by one while lowering all later free coefficients from h h h to − h -h − h does not raise M M M . As in the proof of Lemma 4.6, choosing the free coefficients from left to right therefore reaches every interval of length v N ( 1 + 2 ⋅ 10 − 3 ) v^N(1+2\cdot10^{-3}) v N ( 1 + 2 ⋅ 1 0 − 3 ) inside [ M ( − h , … , − h ) , M ( h , … , h ) ] [M(-h,\ldots,-h),M(h,\ldots,h)] [ M ( − h , … , − h ) , M ( h , … , h )] . Here M ( − h , … , − h ) ≤ Q ( v ) < 0 M(-h,\ldots,-h)\leq Q(v)<0 M ( − h , … , − h ) ≤ Q ( v ) < 0 , since 2 h ∑ j ≤ 5 v j > 1.7 2h\sum_{j\leq5}v^j>1.7 2 h ∑ j ≤ 5 v j > 1.7 (E6), and M ( h , … , h ) ≥ v i + L + 1 ( 1 − 3 ⋅ 10 − 4 ) − H ( v ) ≥ 0.9 H ( v ) + v N ( 1 + 2 ⋅ 10 − 3 ) M(h,\ldots,h)\geq v^{i+L+1}(1-3\cdot10^{-4})-H(v)\geq0.9H(v)+v^N(1+2\cdot10^{-3}) M ( h , … , h ) ≥ v i + L + 1 ( 1 − 3 ⋅ 1 0 − 4 ) − H ( v ) ≥ 0.9 H ( v ) + v N ( 1 + 2 ⋅ 1 0 − 3 ) , since r L ≥ v + η ˉ i / v L r_L\geq v+\bar{\eta}_i/v^L r L ≥ v + η ˉ i / v L , v i η ˉ i ≤ 1.001 v i + L + 1 v^i\bar{\eta}_i\leq1.001v^{i+L+1} v i η ˉ i ≤ 1.001 v i + L + 1 and (E4): v 4 ( 1.9 h v / ( 1 − v ) + 1 ) ≤ 0.965 v^4(1.9hv/(1-v)+1)\leq0.965 v 4 ( 1.9 h v / ( 1 − v ) + 1 ) ≤ 0.965 . By (E2′), 0.8 H ( v ) > 1.01 v N 0.8H(v)>1.01v^N 0.8 H ( v ) > 1.01 v N , so some choice gives min Q ∈ [ 0.1 H ( v ) , 0.9 H ( v ) ] \min Q\in[0.1H(v),0.9H(v)] min Q ∈ [ 0.1 H ( v ) , 0.9 H ( v )] . For it, put ρ = c ρ v N + 1 − i \rho=c_\rho v^{N+1-i} ρ = c ρ v N + 1 − i , with c ρ = 1 2 , 2 4 , 3 5 , 4 5 c_\rho=\frac12,\frac24,\frac35,\frac45 c ρ = 2 1 , 4 2 , 5 3 , 5 4 for h = 1 , 3 , 4 , 7 h=1,3,4,7 h = 1 , 3 , 4 , 7 . By (E3), ( N + 1 ) ∣ x Q − v ∣ + ρ ≤ 10 − 3 (N+1)|x_Q-v|+\rho\leq10^{-3} ( N + 1 ) ∣ x Q − v ∣ + ρ ≤ 1 0 − 3 , so H H H varies by at most 1.002 1.002 1.002 on [ x Q − ρ , x Q + ρ ] [x_Q-\rho,x_Q+\rho] [ x Q − ρ , x Q + ρ ] . Then Q ( x Q ) ≤ 0.9 H ( v ) < H ( x Q ) Q(x_Q)\leq0.9H(v)<H(x_Q) Q ( x Q ) ≤ 0.9 H ( v ) < H ( x Q ) , and Q ( x Q ± ρ ) ≥ min Q + κ v ρ 2 ( 1 − 10 − 4 ) > H ( x Q ± ρ ) Q(x_Q\pm\rho)\geq\min Q+\kappa_v\rho^2(1-10^{-4})>H(x_Q\pm\rho) Q ( x Q ± ρ ) ≥ min Q + κ v ρ 2 ( 1 − 1 0 − 4 ) > H ( x Q ± ρ ) by (E1): κ v c ρ 2 > 1.01 h / ( 1 − τ ) \kappa_vc_\rho^2>1.01h/(1-\tau) κ v c ρ 2 > 1.01 h / ( 1 − τ ) . Lemma 7.2 gives f f f with initial part Q Q Q and a double zero in ( x Q − ρ , x Q + ρ ) (x_Q-\rho,x_Q+\rho) ( x Q − ρ , x Q + ρ ) . Its first difference from f v f_v f v is at i > n i>n i > n , so f ∈ F h ( P ) f\in F_h(P) f ∈ F h ( P ) , and Lemma 7.1 gives (7.4): q i ′ q'_i q i ′ exceeds q ′ ′ q'' q ′′ by the moment of the corner digits after i + L i+L i + L , at most 2 h ∑ r > L r v r 2h\sum_{r>L}rv^r 2 h ∑ r > L r v r , the term ε / v i \varepsilon/v^i ε / v i is at most η ˉ i \bar{\eta}_i η ˉ i , and θ ′ \theta' θ ′ is controlled by (D).
The realization lemma is used with long prescriptions only at depths i ≥ 159 i\geq159 i ≥ 159 , where L ∗ ( i ) ≥ 152 L^*(i)\geq152 L ∗ ( i ) ≥ 152 . Put δ k = 2 v L ∗ ( k ) \delta_k=2v^{L^*(k)} δ k = 2 v L ∗ ( k ) . The next lemma uses the slice bounds of Proposition S4.17:
Δ l o w = inf { min S t ( c + p ) − q t g ( p ) : 1 ≤ c ≤ d ˉ , p ∈ [ 0 , p ˉ ) ∪ [ 1 , 1 + 10 − 6 ] } , \Delta_{\mathrm{low}}=\inf\{\min S_t(c+p)-q_t^g(p):\ 1\leq c\leq\bar d,\ p\in[0,\bar p)\cup[1,1+10^{-6}]\}, Δ low = inf { min S t ( c + p ) − q t g ( p ) : 1 ≤ c ≤ d ˉ , p ∈ [ 0 , p ˉ ) ∪ [ 1 , 1 + 1 0 − 6 ]} ,
Δ 1 = inf { min S t ( c + p ) − q t g ( p ) : 1 ≤ c ≤ ⌊ 2 β ⌋ , p ∈ [ 0 , 1 + 10 − 6 ] } , \Delta_1=\inf\{\min S_t(c+p)-q_t^g(p):\ 1\leq c\leq\lfloor2\beta\rfloor,\ p\in[0,1+10^{-6}]\}, Δ 1 = inf { min S t ( c + p ) − q t g ( p ) : 1 ≤ c ≤ ⌊ 2 β ⌋ , p ∈ [ 0 , 1 + 1 0 − 6 ]} ,
q 1 g ∗ = sup { q t g ( x ) : ∣ x − 1 ∣ ≤ 10 − 6 } , q_1^{g*}=\sup\{q_t^g(x):\ |x-1|\leq10^{-6}\}, q 1 g ∗ = sup { q t g ( x ) : ∣ x − 1∣ ≤ 1 0 − 6 } ,
over all t t t within 10 − 15 10^{-15} 1 0 − 15 of the centre of [ a h , b h ] [a_h,b_h] [ a h , b h ] , where p ˉ = 1 \bar p=1 p ˉ = 1 for h ≥ 3 h\ge3 h ≥ 3 and p ˉ = 0.496 > β − 1 \bar p=0.496>\beta-1 p ˉ = 0.496 > β − 1 for h = 1 h=1 h = 1 , and the integers L min L_{\min} L m i n , L G L_G L G and L 2 L_2 L 2 : the least L L L with 2 h v L + 1 ( ( L + 5 ) / ( 1 − v ) + 1 / ( 1 − v ) 2 ) ≤ Δ l o w / 2 − 10 − 6 2hv^{L+1}((L+5)/(1-v)+1/(1-v)^2)\le\Delta_{\mathrm{low}}/2-10^{-6} 2 h v L + 1 (( L + 5 ) / ( 1 − v ) + 1/ ( 1 − v ) 2 ) ≤ Δ low /2 − 1 0 − 6 , the least L L L with v L ( q 1 ∗ + 10 − 6 ) ≤ Δ l o w − 0.0015 v_L(q_1^*+10^{-6})\le\Delta_{\mathrm{low}}-0.0015 v L ( q 1 ∗ + 1 0 − 6 ) ≤ Δ low − 0.0015 , and the least L L L with 2 h v L + 1 ( ( L + 5 ) / ( 1 − v ) + 1 / ( 1 − v ) 2 ) ≤ 10 − 3 2hv^{L+1}((L+5)/(1-v)+1/(1-v)^2)\le10^{-3} 2 h v L + 1 (( L + 5 ) / ( 1 − v ) + 1/ ( 1 − v ) 2 ) ≤ 1 0 − 3 . It also uses two finite checks: (GR) for every weakly admissible word F = F 1 ⋯ F L 1 F=F_1\cdots F_{L_1} F = F 1 ⋯ F L 1 over { 0 , … , d ˉ } \{0,\ldots,\bar d\} { 0 , … , d ˉ } with 1 ≤ L 1 < L G 1\le L_1<L_G 1 ≤ L 1 < L G , F L 1 ≥ 1 F_{L_1}\ge1 F L 1 ≥ 1 and F ( v ) = ∑ r F r v r < 1 + 10 − 6 F(v)=\sum_rF_rv^r<1+10^{-6} F ( v ) = ∑ r F r v r < 1 + 1 0 − 6 , and every 1 ≤ c ≤ ⌊ 2 β ⌋ 1\le c\le\lfloor2\beta\rfloor 1 ≤ c ≤ ⌊ 2 β ⌋ ,
min S t ( c + F ( v ) ) − q ( F ) − v L 1 ( q 1 ∗ + 10 − 5 ) > 0.002 , q ( F ) = ∑ r r F r v r ; \min S_t(c+F(v))-q(F)-v^{L_1}(q_1^*+10^{-5})>0.002,\qquad q(F)=\sum_r rF_rv^r; min S t ( c + F ( v )) − q ( F ) − v L 1 ( q 1 ∗ + 1 0 − 5 ) > 0.002 , q ( F ) = r ∑ r F r v r ;
and (EX) the greedy expansion of every x ∈ [ 1 , 1 + ξ 0 ] x\in[1,1+\xi_0] x ∈ [ 1 , 1 + ξ 0 ] , ξ 0 = 2 v L ∗ ( 159 ) \xi_0=2v^{L^*(159)} ξ 0 = 2 v L ∗ ( 159 ) , is the same for all t ∈ [ t 0 , t 0 + 10 − 80 ] t\in[t_0,t_0+10^{-80}] t ∈ [ t 0 , t 0 + 1 0 − 80 ] (10 − 140 10^{-140} 1 0 − 140 for h = 1 h=1 h = 1 ) through M E ≥ 144 M_E\ge144 M E ≥ 144 digits, with positive remainders, and has no run of ten zeros among its first max ( L min , L 2 ) + 21 \max(L_{\min},L_2)+21 max ( L m i n , L 2 ) + 21 digits.
Lemma 7.4 (Greedy tails). Let v = m ( P ) v=m(P) v = m ( P ) with p n < 2 p_n<2 p n < 2 , and suppose that n ≥ 158 n\ge158 n ≥ 158 , or that P = P 0 P=P_0 P = P 0 and p k < 1 p_k<1 p k < 1 for n 0 < k ≤ 158 n_0<k\le158 n 0 < k ≤ 158 . Then p k < 1 + δ k p_k<1+\delta_k p k < 1 + δ k for every k > n k>n k > n . Consequently the tail of f v f_v f v is the greedy expansion of p n p_n p n , except at restarts: indices k k k with p k ∈ [ 1 , 1 + δ k ) p_k\in[1,1+\delta_k) p k ∈ [ 1 , 1 + δ k ) , where the corner digit is one less than the greedy one and the tail continues as the greedy expansion of p k p_k p k . Restarts occur only at indices k ≥ 159 k\ge159 k ≥ 159 .
Proof. Let k > n k>n k > n be the first index with p k ≥ 1 + δ k p_k\ge1+\delta_k p k ≥ 1 + δ k ; by hypothesis k ≥ 159 k\ge159 k ≥ 159 . Put c = ⌊ p k − δ k ⌋ ≥ 1 c=\lfloor p_k-\delta_k\rfloor\ge1 c = ⌊ p k − δ k ⌋ ≥ 1 and p = p k − c ∈ [ δ k , 1 + δ k ) p=p_k-c\in[\delta_k,1+\delta_k) p = p k − c ∈ [ δ k , 1 + δ k ) . Then b k + c ≤ 2 h b_k+c\le2h b k + c ≤ 2 h : for k ≥ n + 2 k\ge n+2 k ≥ n + 2 because p k − 1 < 1 + δ k − 1 p_{k-1}<1+\delta_{k-1} p k − 1 < 1 + δ k − 1 gives b k , c ≤ d ˉ ≤ h b_k,c\le\bar d\le h b k , c ≤ d ˉ ≤ h , and for k = n + 1 k=n+1 k = n + 1 because b k + c ≤ β p n < 2 β ≤ 2 h + 1 b_k+c\le\beta p_n<2\beta\le2h+1 b k + c ≤ β p n < 2 β ≤ 2 h + 1 . We raise b k b_k b k by c c c and use Lemma 7.3 with digits b ′ ′ b'' b ′′ after k k k to be chosen; its double zero lies left of v v v , contradicting v = min D h ( P ) v=\min D_h(P) v = min D h ( P ) , as soon as q ′ ′ + ϱ < q k q''+\varrho<q_k q ′′ + ϱ < q k , and q k ≥ min S v ( c + p ) q_k\ge\min S_v(c+p) q k ≥ min S v ( c + p ) . Put Δ = Δ l o w \Delta=\Delta_{\mathrm{low}} Δ = Δ low if c ≤ d ˉ c\le\bar d c ≤ d ˉ and Δ = Δ 1 \Delta=\Delta_1 Δ = Δ 1 otherwise (then k = n + 1 k=n+1 k = n + 1 ); both infima include the strip p ∈ [ 1 , 1 + 10 − 6 ] p\in[1,1+10^{-6}] p ∈ [ 1 , 1 + 1 0 − 6 ] , and Δ > Δ l o w / 2 \Delta>\Delta_{\mathrm{low}}/2 Δ > Δ low /2 . Let g g g be the greedy expansion of p p p , with remainders r L r_L r L .
(1) If r L ≥ v + η ˉ k / v L r_L\ge v+\bar\eta_k/v^L r L ≥ v + η ˉ k / v L for some L ∈ [ L min , L ∗ ( k ) ] L\in[L_{\min},L^*(k)] L ∈ [ L m i n , L ∗ ( k )] , take b ′ ′ = g 1 ⋯ g L b''=g_1\cdots g_L b ′′ = g 1 ⋯ g L . Then q ′ ′ ≤ q v g ( p ) q''\le q_v^g(p) q ′′ ≤ q v g ( p ) , ∣ ϱ ∣ < Δ l o w / 2 |\varrho|<\Delta_{\mathrm{low}}/2 ∣ ϱ ∣ < Δ low /2 , and min S v ( c + p ) − q v g ( p ) ≥ Δ \min S_v(c+p)-q_v^g(p)\ge\Delta min S v ( c + p ) − q v g ( p ) ≥ Δ .
(2) Otherwise, writing r L = v ( g L + 1 + r L + 1 ) r_L=v(g_{L+1}+r_{L+1}) r L = v ( g L + 1 + r L + 1 ) shows that among the positions L min + 1 , … , L ∗ ( k ) + 1 L_{\min}+1,\ldots,L^*(k)+1 L m i n + 1 , … , L ∗ ( k ) + 1 the expansion g g g has at most one nonzero digit, a 1 at some position L 1 L_1 L 1 , after which the remainder is below η ˉ k / v L 1 \bar\eta_k/v^{L_1} η ˉ k / v L 1 . (2a) If L 1 ≤ L ∗ ( k ) − 1 L_1\le L^*(k)-1 L 1 ≤ L ∗ ( k ) − 1 (for h ≥ 3 h\ge3 h ≥ 3 : L 1 ≤ L ∗ ( k ) + 1 L_1\le L^*(k)+1 L 1 ≤ L ∗ ( k ) + 1 ), prescribe g 1 ⋯ g L 1 − 1 0 g_1\cdots g_{L_1-1}0 g 1 ⋯ g L 1 − 1 0 . The remaining weight lies in [ 1 , 1 + η ˉ k / v L 1 ] ⊂ [ v + η ˉ k / v L 1 , 1.7 ] [1,1+\bar\eta_k/v^{L_1}]\subset[v+\bar\eta_k/v^{L_1},1.7] [ 1 , 1 + η ˉ k / v L 1 ] ⊂ [ v + η ˉ k / v L 1 , 1.7 ] , because $1.001v^3\le1-v$ for h = 1 h=1 h = 1 and 1.001 v ≤ 1 − v 1.001v\le1-v 1.001 v ≤ 1 − v for $h\ge3; we conclude as in (1). (2b) Otherwise p = F ( v ) + ξ ′ p=F(v)+\xi' p = F ( v ) + ξ ′ with ξ ′ < 2 v L ∗ ( k ) \xi'<2v^{L^*(k)} ξ ′ < 2 v L ∗ ( k ) , where F = g 1 ⋯ g L 1 F=g_1\cdots g_{L_1} F = g 1 ⋯ g L 1 is empty or has its last nonzero digit at L 1 ≤ L min L_1\le L_{\min} L 1 ≤ L m i n . If F F F is empty and c = 1 c=1 c = 1 , then p k < 1 + δ k p_k<1+\delta_k p k < 1 + δ k , contrary to the choice of k k k . If c ≥ 2 c\ge2 c ≥ 2 , use c − 1 c-1 c − 1 and the weight 1 + p 1+p 1 + p , prescribing the greedy expansion of 1 + p 1+p 1 + p up to an L ∈ [ L min , L min + 9 ] L\in[L_{\min},L_{\min}+9] L ∈ [ L m i n , L m i n + 9 ] followed by a nonzero digit, which exists by (EX), and which leaves a remaining weight in [ v + η ˉ k / v L , 1.7 ] [v+\bar\eta_k/v^L,1.7] [ v + η ˉ k / v L , 1.7 ] ; then q ′ ′ ≤ q v g ( 1 + p ) q''\le q_v^g(1+p) q ′′ ≤ q v g ( 1 + p ) , and min S v ( c + p ) − q v g ( 1 + p ) ≥ Δ \min S_v(c+p)-q_v^g(1+p)\ge\Delta min S v ( c + p ) − q v g ( 1 + p ) ≥ Δ by the strip part of the bound. If F F F is nonempty, prescribe F ′ F' F ′ with its last digit lowered by one, followed by the greedy expansion of 1 + ξ ′ / v L 1 1+\xi'/v^{L_1} 1 + ξ ′ / v L 1 up to an L L L with L − L 1 ∈ [ L 2 , L 2 + 9 ] L-L_1\in[L_2,L_2+9] L − L 1 ∈ [ L 2 , L 2 + 9 ] followed by a nonzero digit (again by (EX)); its weight is p p p , the room condition holds, ∣ ϱ ∣ ≤ 0.0011 |\varrho|\le0.0011 ∣ ϱ ∣ ≤ 0.0011 , and q ′ ′ ≤ q ( F ) + v L 1 ( q 1 ∗ + 10 − 6 ) q''\le q(F)+v^{L_1}(q_1^*+10^{-6}) q ′′ ≤ q ( F ) + v L 1 ( q 1 ∗ + 1 0 − 6 ) . If L 1 ≥ L G L_1\ge L_G L 1 ≥ L G , then min S v ( c + p ) − q ( F ) ≥ Δ ≥ v L 1 ( q 1 ∗ + 10 − 6 ) + 0.0015 \min S_v(c+p)-q(F)\ge\Delta\ge v^{L_1}(q_1^*+10^{-6})+0.0015 min S v ( c + p ) − q ( F ) ≥ Δ ≥ v L 1 ( q 1 ∗ + 1 0 − 6 ) + 0.0015 , by the definition of L G L_G L G and by (K4) of Proposition S4.19; if L 1 < L G L_1<L_G L 1 < L G , (GR) applies. In every case q ′ ′ + ϱ < q k q''+\varrho<q_k q ′′ + ϱ < q k .
If p k ∈ [ 1 , 1 + δ k ) p_k\in[1,1+\delta_k) p k ∈ [ 1 , 1 + δ k ) the greedy digit at k k k is b k + 1 b_k+1 b k + 1 , and the lemma applied from k k k on describes the rest of the tail.
The tips of the construction satisfy the hypothesis of Lemma 7.4: t 0 t_0 t 0 by Proposition S4.19, and the later ones because ∣ P ∣ ≥ 158 |P|\ge158 ∣ P ∣ ≥ 158 .
Lemma 7.5 (Restarts). Let v v v satisfy the hypothesis of Lemma 7.4.
After a restart at k k k the next M E M_E M E digits of the tail are those of d β ( 1 ) d_\beta(1) d β ( 1 ) , and two restarts are more than M E M_E M E positions apart.
For every level s ′ s' s ′ let r r r be the distance from s ′ s' s ′ to the first restart after it (r = ∞ r=\infty r = ∞ if there is none). Then q v g ( y s ′ ) ≤ q s ′ ≤ q v g ( y s ′ ) + v r ( q 1 ∗ + 10 − 5 ) q_v^g(y_{s'})\le q_{s'}\le q_v^g(y_{s'})+v^r(q_1^*+10^{-5}) q v g ( y s ′ ) ≤ q s ′ ≤ q v g ( y s ′ ) + v r ( q 1 ∗ + 1 0 − 5 ) .
(iii) For every level s ′ s' s ′ with g s ′ ≥ 1 g_{s'} \ge1 g s ′ ≥ 1 , every 1 ≤ c ≤ g s ′ 1 \le c \le g_{s'} 1 ≤ c ≤ g s ′ and ∣ e ∣ ≤ 10 − 4 |e| \le10^{-4} ∣ e ∣ ≤ 1 0 − 4 , min S v ( c + y s ′ + e ) − q s ′ \min S_v(c+y_{s'}+e)-q_{s'} min S v ( c + y s ′ + e ) − q s ′ is at least Δ − v r ( q 1 ∗ + 10 − 5 ) \Delta-v^r(q_1^*+10^{-5}) Δ − v r ( q 1 ∗ + 1 0 − 5 ) , where Δ = Δ l o w \Delta=\Delta_{\mathrm{low}} Δ = Δ low for s ′ ≥ 2 s'\ge2 s ′ ≥ 2 and Δ = Δ 1 \Delta=\Delta_1 Δ = Δ 1 for s ′ = 1 s'=1 s ′ = 1 , and it always exceeds 10 − 3 10^{-3} 1 0 − 3 .
Proof. At a restart k k k the tail continues as the greedy expansion of p k ∈ [ 1 , 1 + ξ 0 ] p_k\in[1,1+\xi_0] p k ∈ [ 1 , 1 + ξ 0 ] , which by (EX) agrees for M E M_E M E digits with that of 1 1 1 and keeps its remainders in [ 0 , 1 ) [0,1) [ 0 , 1 ) there, so no restart occurs in between. For (ii), let k = n + s ′ + r k=n+s'+r k = n + s ′ + r be the first restart after s ′ s' s ′ . The greedy expansion of y s ′ y_{s'} y s ′ is
F = g s ′ + 1 ⋯ g s ′ + r − 1 ( b k + 1 ) F=g_{s'+1}\cdots g_{s'+r-1}(b_k+1) F = g s ′ + 1 ⋯ g s ′ + r − 1 ( b k + 1 )
followed by the greedy expansion of ξ = p k − 1 < δ k \xi=p_k-1<\delta_k ξ = p k − 1 < δ k , so
q s ′ − q v g ( y s ′ ) = v r ( q k − q v g ( ξ ) ) . q_{s'}-q_v^g(y_{s'})=v^r(q_k-q_v^g(\xi)). q s ′ − q v g ( y s ′ ) = v r ( q k − q v g ( ξ )) .
Here q k ≥ d ˉ v ≥ v q_k\ge\bar d v\ge v q k ≥ d ˉ v ≥ v , since the expansion of p k ≥ 1 p_k\ge1 p k ≥ 1 starts with the digit d ˉ \bar d d ˉ , while q v g ( ξ ) q_v^g(\xi) q v g ( ξ ) is of the order of ξ \xi ξ ; and, by induction over the later restarts, q k ≤ q 1 ∗ + 10 − 6 + v r ( q 1 ∗ + 10 − 5 ) q_k\le q_1^*+10^{-6}+v^r(q_1^*+10^{-5}) q k ≤ q 1 ∗ + 1 0 − 6 + v r ( q 1 ∗ + 1 0 − 5 ) . For (iii), y s ′ ∈ [ 0 , p ˉ ) ∪ [ 1 , 1 + 10 − 6 ] y_{s'}\in[0,\bar p)\cup[1,1+10^{-6}] y s ′ ∈ [ 0 , p ˉ ) ∪ [ 1 , 1 + 1 0 − 6 ] (for h = 1 h=1 h = 1 because g s ′ ≥ 1 g_{s'}\ge1 g s ′ ≥ 1 forces y s ′ < β − 1 y_{s'}<\beta-1 y s ′ < β − 1 ), so (ii) gives the first bound, the table lookups of Proposition S4.17 absorbing ∣ e ∣ ≤ 10 − 4 |e|\le10^{-4} ∣ e ∣ ≤ 1 0 − 4 . If r ≥ L G r\ge L_G r ≥ L G it exceeds 10 − 3 10^{-3} 1 0 − 3 by Proposition S4.19 (K4). If r < L G r<L_G r < L G , the word F F F of the proof of (ii) is a greedy prefix of y s ′ y_{s'} y s ′ , q s ′ ≤ q ( F ) + v r ( q 1 ∗ + 10 − 5 ) q_{s'}\le q(F)+v^r(q_1^*+10^{-5}) q s ′ ≤ q ( F ) + v r ( q 1 ∗ + 1 0 − 5 ) , and (GR) gives a margin above 0.002 0.002 0.002 . □
Lemma 7.6 (Only lowered digits lie to the right). Let f ∈ F h f\in\mathcal{F}_h f ∈ F h have a double zero at v + w v+w v + w with 0 < w < F ( P ) u P 0<w<F(P)u_P 0 < w < F ( P ) u P , and let i i i be the first index where its corner digits differ from those of f v f_v f v . Then f ∈ F h ( P ) f\in\mathcal{F}_h(P) f ∈ F h ( P ) , i > n i>n i > n , b i ′ = b i − c b'_i=b_i-c b i ′ = b i − c with 1 ≤ c ≤ b i 1\le c\le b_i 1 ≤ c ≤ b i , and p i ′ = c + p i − η p'_i=c+p_i-\eta p i ′ = c + p i − η with 0 < η = ε / v i ≤ η ˉ i 0<\eta=\varepsilon/v^i\le\bar\eta_i 0 < η = ε / v i ≤ η ˉ i .
Proof. The first assertion is the definition of F ( P ) F(P) F ( P ) , so i > n i>n i > n . Suppose b i ′ = b i + c b'_i=b_i+c b i ′ = b i + c with c ≥ 1 c\ge1 c ≥ 1 . By (7.3) and Lemma 7.1, − ζ i v i ≤ ε ≤ v i ( p i − 1 ) -\zeta_i v^i\le\varepsilon\le v^i(p_i-1) − ζ i v i ≤ ε ≤ v i ( p i − 1 ) ; hence p i ≥ 1 − ζ i p_i\ge1-\zeta_i p i ≥ 1 − ζ i , while p i ′ ≤ p i − 1 + ζ i < δ i + ζ i = : x i p'_i\le p_i-1+\zeta_i<\delta_i+\zeta_i=:x_i p i ′ ≤ p i − 1 + ζ i < δ i + ζ i =: x i by Lemma 7.4. So the corner digits b i + r ′ b'_{i+r} b i + r ′ vanish for r < r 0 r<r_0 r < r 0 , the least r r r with v r < x n 0 + 1 v^r<x_{n_0+1} v r < x n 0 + 1 , and
q i ′ ≤ 2 h v r 0 ( r 0 1 − v + 1 ( 1 − v ) 2 ) . q'_i\le2h v^{r_0}\left(\frac{r_0}{1-v}+\frac{1}{(1-v)^2}\right). q i ′ ≤ 2 h v r 0 ( 1 − v r 0 + ( 1 − v ) 2 1 ) .
Then ε ′ / v i − 1 ≥ q i − q i ′ − i ζ i > 0 \varepsilon'/v^{i-1}\ge q_i-q'_i-i\zeta_i>0 ε ′ / v i − 1 ≥ q i − q i ′ − i ζ i > 0 by (C3), since q i ≥ p i q_i\ge p_i q i ≥ p i ; but ε ′ \varepsilon' ε ′ has the sign of − w -w − w . Hence b i ′ < b i b'_i<b_i b i ′ < b i , and the bounds on η \eta η are those of Lemma 7.1. □
Theorem 7.7 (Structure at a relative tip). Let x ∈ ( v , v + F ( P ) u P ) x\in(v,v+F(P)u_P) x ∈ ( v , v + F ( P ) u P ) be a double zero of f ∈ F h f\in\mathcal{F}_h f ∈ F h . With i i i and c c c as in Lemma 7.6, put s = i − n s=i-n s = i − n . Then g s ≥ c ≥ 1 g_s\ge c\ge1 g s ≥ c ≥ 1 and
x − v = u P v s ( q ′ − q s − ι ) ( 1 + θ ) , q ′ ∈ S v ( c + y s − η ) , x-v=u_Pv^s(q'-q_s-\iota)(1+\theta),\qquad q'\in S_v(c+y_s-\eta), x − v = u P v s ( q ′ − q s − ι ) ( 1 + θ ) , q ′ ∈ S v ( c + y s − η ) ,
with 0 < η ≤ η ˉ i ≤ 10 − 7 0<\eta\le\bar\eta_i\le10^{-7} 0 < η ≤ η ˉ i ≤ 1 0 − 7 , 0 ≤ ι ≤ i η ˉ i ≤ 10 − 5 0\le\iota\le i\bar\eta_i\le10^{-5} 0 ≤ ι ≤ i η ˉ i ≤ 1 0 − 5 and ∣ θ ∣ ≤ 10 − 5 |\theta|\le10^{-5} ∣ θ ∣ ≤ 1 0 − 5 . We call the set of these values, for fixed s s s and c c c , the group G s , c G_{s,c} G s , c ; its points are the double zeros in ( v , v + F ( P ) u P ) (v,v+F(P)u_P) ( v , v + F ( P ) u P ) of the series f ∈ F h ( P ) f\in\mathcal{F}_h(P) f ∈ F h ( P ) with b n + s ′ = b n + s − c b'_{n+s}=b_{n+s}-c b n + s ′ = b n + s − c that agree with f v f_v f v before n + s n+s n + s .
Proof. Combine Lemmas 7.6 and 7.1: q ′ = q i ′ ∈ S v ( p i ′ ) q'=q'_i\in S_v(p'_i) q ′ = q i ′ ∈ S v ( p i ′ ) , ι = i ε / v i \iota=i\varepsilon/v^i ι = i ε / v i , and ( 1 + θ ) − 1 = 1 + ϑ / ( 2 κ v ) (1+\theta)^{-1}=1+\vartheta/(2\kappa v) ( 1 + θ ) − 1 = 1 + ϑ / ( 2 κ v ) is controlled by (D). □
Moves
Figure 8 shows the groups of t 0 t_0 t 0 at h = 1 h=1 h = 1 and h = 7 h=7 h = 7 . Put μ = 1 / 50 \mu=1/50 μ = 1/50 . At a level s s s with g s ≥ 1 g_s\ge1 g s ≥ 1 let P ′ = P g 1 ⋯ g s − 1 ( g s − 1 ) P'=Pg_1\cdots g_{s-1}(g_s-1) P ′ = P g 1 ⋯ g s − 1 ( g s − 1 ) , of length n + s n+s n + s , and
Figure 8. Groups at the minimum t 0 t_0 t 0 of D h D_h D h for h = 1 h=1 h = 1 (left) and h = 7 h=7 h = 7 (right). Top: row s s s shows, on a logarithmic scale in units of u P 0 u_{P_0} u P 0 , outer enclosures of the groups G s , c G_{s,c} G s , c of the certified tail of t 0 t_0 t 0 , computed from the exact slice tables of Proposition S4.17; shaded rows are the C h C_h C h -patterns at levels s ≥ P 0 ∗ s\geq P_0^* s ≥ P 0 ∗ , and the tick marks the left end v s L s ( 1 − μ ) v^sL_s(1-\mu) v s L s ( 1 − μ ) of the move window. Bottom: the first such pattern, in units of u P 0 v s L s u_{P_0}v^sL_s u P 0 v s L s . The window [ 1 − μ , 1 + μ ] [1-\mu,1+\mu] [ 1 − μ , 1 + μ ] meets no group except G s , 1 G_{s,1} G s , 1 , whose minimum m ( P s ′ ) m(P'_s) m ( P s ′ ) lies at 1; the interval from 1 − μ 1-\mu 1 − μ to m ( P s ′ ) m(P'_s) m ( P s ′ ) is a gap of D h D_h D h . The enclosures contain the true groups; the figure illustrates Propositions S4.18 and S4.19 and certifies nothing beyond them.
L s = { q t g ( 1 + y s + e ) − q s : ∣ e ∣ ≤ 10 − 7 } , \mathcal{L}_s=\{q_t^g(1+y_s+e)-q_s:\ |e|\le10^{-7}\}, L s = { q t g ( 1 + y s + e ) − q s : ∣ e ∣ ≤ 1 0 − 7 } ,
the range of the first-group position at level s s s , taken over t t t within 10 − 15 10^{-15} 1 0 − 15 of the centre of [ a h , b h ] [a_h,b_h] [ a h , b h ] and over all greedy branches.
Corollary 7.8 (Moves). Let s ≥ 2 s\ge2 s ≥ 2 with g s ≥ 1 g_s\ge1 g s ≥ 1 , y s < 1 y_s<1 y s < 1 and n + s ≥ 158 n+s\ge158 n + s ≥ 158 , let L − ≤ L + L_-\le L_+ L − ≤ L + bound L s \mathcal{L}_s L s , and assume:
(a) every group of a level s ′ > s s'>s s ′ > s lies below v s ( L − ( 1 − μ ) − 5 ⋅ 10 − 4 L − ) v^s(L_-(1-\mu)-5\cdot10^{-4}L_-) v s ( L − ( 1 − μ ) − 5 ⋅ 1 0 − 4 L − ) ;
(b) no group G s ′ , c G_{s',c} G s ′ , c with s ′ ≤ s s'\le s s ′ ≤ s and ( s ′ , c ) ≠ ( s , 1 ) (s',c)\ne(s,1) ( s ′ , c ) = ( s , 1 ) meets v s [ L − ( 1 − μ ) − 5 ⋅ 10 − 4 L − , L + ( 1 + μ ) + 5 ⋅ 10 − 4 L + ] v^s[L_-(1-\mu)-5\cdot10^{-4}L_-,L_+(1+\mu)+5\cdot10^{-4}L_+] v s [ L − ( 1 − μ ) − 5 ⋅ 1 0 − 4 L − , L + ( 1 + μ ) + 5 ⋅ 1 0 − 4 L + ] ;
(c) v s L + ( 1 + μ ) < F ( P ) v^sL_+(1+\mu)<F(P) v s L + ( 1 + μ ) < F ( P ) .
Then D h ( P ′ ) ≠ ∅ D_h(P')\ne\varnothing D h ( P ′ ) = ∅ , m ( P ′ ) − v ≥ u P v s L − ( 1 − 10 − 4 ) m(P')-v\ge u_Pv^sL_-(1-10^{-4}) m ( P ′ ) − v ≥ u P v s L − ( 1 − 1 0 − 4 ) , and the interval ( v + u P v s L − ( 1 − μ ) , m ( P ′ ) ) (v+u_Pv^sL_-(1-\mu),m(P')) ( v + u P v s L − ( 1 − μ ) , m ( P ′ )) contains no point of D h D_h D h .
(d) the tail of a double-zero series of P ′ P' P ′ at m ( P ′ ) m(P') m ( P ′ ) has no restart among its first R R R digits, where
v R + 1 ( q 1 ∗ + 10 − 5 ) ≤ 7 ⋅ 10 − 5 L − , v^{R+1}(q_1^*+10^{-5})\le7\cdot10^{-5}L_-, v R + 1 ( q 1 ∗ + 1 0 − 5 ) ≤ 7 ⋅ 1 0 − 5 L − ,
then m ( P ′ ) − v ≤ u P v s L + ( 1 + 10 − 4 ) m(P')-v\le u_Pv^sL_+(1+10^{-4}) m ( P ′ ) − v ≤ u P v s L + ( 1 + 1 0 − 4 ) and F ( P ′ ) ≥ ( μ − 10 − 4 ) L − F(P')\ge(\mu-10^{-4})L_- F ( P ′ ) ≥ ( μ − 1 0 − 4 ) L − .
Proof. Let x = 1 + y s x=1+y_s x = 1 + y s . Apply Lemma 7.3 with c = 1 c=1 c = 1 and the prescription of the proof of Lemma 7.4 for the weight x x x , with L 2 L_2 L 2 in place of L min L_{\min} L m i n ; in case (2b) F F F is nonempty because x ≥ 1 x\geq1 x ≥ 1 , and we prescribe F F F followed by zeros if ξ ′ ≥ 2 η ˉ i \xi'\geq2\bar{\eta}_i ξ ′ ≥ 2 η ˉ i . Then ∣ ϱ ∣ ≤ 0.0011 |\varrho|\leq0.0011 ∣ ϱ ∣ ≤ 0.0011 , and q ′ ′ q'' q ′′ exceeds by less than 10 − 40 10^{-40} 1 0 − 40 the moment of the greedy expansion of a weight in [ x − 10 − 7 , x + 10 − 7 ] [x-10^{-7},x+10^{-7}] [ x − 1 0 − 7 , x + 1 0 − 7 ] : in cases (1) and (2a), and for F F F followed by zeros, that of x x x itself; otherwise that of a weight just below F ( v ) F(v) F ( v ) , whose expansion is F F F lowered followed by the digits of d β ( 1 ) d_\beta(1) d β ( 1 ) for M E M_E M E positions, by (EX). So a member of F h ( P ′ ) \mathcal{F}_h(P') F h ( P ′ ) has a double zero at distance at most u P v s ( L + + 0.0012 ) u_Pv^s(L_++0.0012) u P v s ( L + + 0.0012 ) from v v v , inside ( v , v + F ( P ) u P ) (v,v+F(P)u_P) ( v , v + F ( P ) u P ) by (c). Hence m ( P ′ ) m(P') m ( P ′ ) lies in this interval, and by Theorem 7.7 the points of D h D_h D h there belong to groups, those of D h ( P ′ ) D_h(P') D h ( P ′ ) forming G s , 1 G_{s,1} G s , 1 . A double-zero series of P ′ P' P ′ at m ( P ′ ) m(P') m ( P ′ ) satisfies the hypothesis of Lemma 7.4 (∣ P ′ ∣ ≥ 158 |P'|\geq158 ∣ P ′ ∣ ≥ 158 , and its weight after P ′ P' P ′ is x − η < 2 x-\eta<2 x − η < 2 ), so its tail is the greedy expansion of that weight up to restarts, and its moment is at least the greedy one by Lemma 7.5 (ii). The weight lies within 10 − 7 10^{-7} 1 0 − 7 of x x x , and the parameters v v v and m ( P ′ ) m(P') m ( P ′ ) differ by less than 10 − 80 10^{-80} 1 0 − 80 , which changes moments by less than 10 − 70 10^{-70} 1 0 − 70 . With the factor 1 ± 10 − 5 1 \pm10^{-5} 1 ± 1 0 − 5 and ι ≤ 10 − 5 \iota\le10^{-5} ι ≤ 1 0 − 5 of Theorem 7.7 this gives the lower bound for m ( P ′ ) − v m(P')-v m ( P ′ ) − v . Now (a) and (b) leave no point of D h D_h D h between v + u P v s L − ( 1 − μ ) v+u_Pv^sL_-(1-\mu) v + u P v s L − ( 1 − μ ) and m ( P ′ ) m(P') m ( P ′ ) : points of deeper groups lie below that bound, and the other groups of levels at most s s s avoid the whole window. Under (d), Lemma 7.5 (ii) bounds the excess of the moment over the greedy one by v R + 1 ( q 1 ∗ + 10 − 5 ) ≤ 7 ⋅ 10 − 5 L − v^{R+1}(q_1^*+10^{-5}) \le7\cdot10^{-5}L_- v R + 1 ( q 1 ∗ + 1 0 − 5 ) ≤ 7 ⋅ 1 0 − 5 L − , whence the upper bound for m ( P ′ ) − v m(P')-v m ( P ′ ) − v . For the free radius, a series with a double zero at y > m ( P ′ ) y>m(P') y > m ( P ′ ) , y − m ( P ′ ) < ( μ − 10 − 4 ) L − u P ′ y-m(P')<(\mu-10^{-4})L_-u_{P'} y − m ( P ′ ) < ( μ − 1 0 − 4 ) L − u P ′ , has y − v < F ( P ) u P y-v<F(P)u_P y − v < F ( P ) u P by (c); by Theorem 7.7 and (b) its first difference from f v f_v f v is at n + s n+s n + s with c = 1 c=1 c = 1 , so it belongs to F h ( P ′ ) F_h(P') F h ( P ′ ) . Here u P ′ = u P v s ( 1 + O ( 10 − 9 ) ) u_{P'}=u_Pv^s(1+O(10^{-9})) u P ′ = u P v s ( 1 + O ( 1 0 − 9 )) . □
At a C h C_h C h -pattern (defined next) one has y s < 1 y_s<1 y s < 1 : the digit after the pattern digit is 0, or, at h = 7 h=7 h = 7 where t < 1 / 3 t<1/3 t < 1/3 , at most 2, and all later remainders are below 1 + 10 − 7 1+10^{-7} 1 + 1 0 − 7 .
Patterns, cores and the finite lemmas
For each h ∈ { 1 , 3 , 4 , 7 } h\in\{1,3,4,7\} h ∈ { 1 , 3 , 4 , 7 } , a C h C_h C h -pattern at level s s s of a tail g g g is a digit g s = d ≥ 1 g_s=d\ge1 g s = d ≥ 1 such that g s + 1 g s + 2 ⋯ g_{s+1}g_{s+2}\cdots g s + 1 g s + 2 ⋯ begins with one of the following start words u u u for d d d :
h = 1 h=1 h = 1 : u = 0000 u=0000 u = 0000 (so d = 1 d=1 d = 1 is followed by at least four zeros, or ends the tail);
h = 3 h=3 h = 3 : u = 00 u=00 u = 00 , or u = 01 w u=01w u = 01 w with w ∈ { 0 , 10 , 1100 , 11010 , 110110 , 11011100 , 11011101 } w\in\{0,10,1100,11010,110110,11011100,11011101\} w ∈ { 0 , 10 , 1100 , 11010 , 110110 , 11011100 , 11011101 } ;
h = 4 h=4 h = 4 : u = 0 u=0 u = 0 ;
h = 7 h=7 h = 7 : u = 0 u=0 u = 0 or u = 10 u=10 u = 10 ; or u = w 1 w 2 w 3 w 4 u=w_1w_2w_3w_4 u = w 1 w 2 w 3 w 4 with d w 1 w 2 w 3 w 4 dw_1w_2w_3w_4 d w 1 w 2 w 3 w 4 in the list 11111 , 11112 , 11113 , 11123 , 11211 , 11212 , 11213 , 11223 , 11311 , 11312 , 11313 , 12113 , 21111 , 21112 , 21113 , 21123 , 21211 , 21212 , 21213 , 21223 , 21311 , 21312 , 21313 , 22113 , 31111 , 31112 , 31113 , 31123 , 31211 , 31212 , 31213 , 31223 11111, 11112, 11113, 11123, 11211, 11212, 11213, 11223, 11311, 11312, 11313, 12113, 21111, 21112, 21113, 21123, 21211, 21212, 21213, 21223, 21311, 21312, 21313, 22113, 31111, 31112, 31113, 31123, 31211, 31212, 31213, 31223 11111 , 11112 , 11113 , 11123 , 11211 , 11212 , 11213 , 11223 , 11311 , 11312 , 11313 , 12113 , 21111 , 21112 , 21113 , 21123 , 21211 , 21212 , 21213 , 21223 , 21311 , 21312 , 21313 , 22113 , 31111 , 31112 , 31113 , 31123 , 31211 , 31212 , 31213 , 31223 .
A tail with no pattern from some level on eventually lies in the gapless core K h K_h K h . By Parry’s criterion, and the expansions d β ( 1 ) d_\beta(1) d β ( 1 ) of Table 2 , these cores are: K 1 K_1 K 1 , the concatenations of the blocks 100 and 1000; K 3 K_3 K 3 , the tails over { 0 , 1 } \{0,1\} { 0 , 1 } in which every 0 is followed by at least three 1s and a block of exactly three 1s is followed by a block of at least four; K 4 = { 1 ∞ } K_4=\{1^\infty\} K 4 = { 1 ∞ } ; and K 7 K_7 K 7 , the weakly admissible tails over { 1 , 2 , 3 } \{1,2,3\} { 1 , 2 , 3 } without any of the 32 windows. (At h = 3 h=3 h = 3 a digit 2 forces a pattern within five digits; at h = 4 h=4 h = 4 a 2 is followed by a 0 within three digits; at h = 1 h=1 h = 1 the block 101 forces 10 5 10^5 1 0 5 .)
Proposition 7.9 (Move Lemma). Let h ∈ { 1 , 3 , 4 , 7 } h\in\{1,3,4,7\} h ∈ { 1 , 3 , 4 , 7 } , let v v v be a relative tip within 10 − 15 10^{-15} 1 0 − 15 of the centre of [ a h , b h ] [a_h,b_h] [ a h , b h ] satisfying the hypothesis of Lemma 7.4 , and let its tail have a C h C_h C h -pattern at a level s ≥ J h + 2 s\ge J_h+2 s ≥ J h + 2 , with J h = 10 , 3 , 1 , 1 J_h=10,3,1,1 J h = 10 , 3 , 1 , 1 . Then L s ⊂ [ L min ′ , L max ′ ] L_s\subset[L_{\min}',L_{\max}'] L s ⊂ [ L m i n ′ , L m a x ′ ] (Table 2 ), condition (a) of Corollary 7.8 holds, and so does (b) for the levels s − J h , … , s s-J_h,\ldots,s s − J h , … , s .
This is Proposition S4.18; the certified clearances exceed the margins 5 ⋅ 10 − 4 L ± 5\cdot10^{-4}L_{\pm} 5 ⋅ 1 0 − 4 L ± of the corollary. The levels below s − J h s-J_h s − J h are handled by Lemma 7.5 . Let M h M_h M h be the least M M M with v M L max ′ ( 1 + μ ) ( 1 + 3 ⋅ 10 − 5 ) < 10 − 3 v^{M}L_{\max}'(1+\mu)(1+3\cdot10^{-5})<10^{-3} v M L m a x ′ ( 1 + μ ) ( 1 + 3 ⋅ 1 0 − 5 ) < 1 0 − 3 and call a pattern at level s s s clean if the tail has no restart at the levels s − M h + 1 , … , s s-M_h+1,\ldots,s s − M h + 1 , … , s .
Lemma 7.10 (Beyond the window). At a clean C h C_h C h -pattern of level s ≥ J h + 2 s\ge J_h+2 s ≥ J h + 2 of a tip as in Proposition 7.9 , condition (b) of Corollary 7.8 also holds for the levels s ′ ≤ s − J h − 1 s'\le s-J_h-1 s ′ ≤ s − J h − 1 .
Proof. By Lemma 7.5 (iii) every point of G s ′ , c ′ G_{s',c}' G s ′ , c ′ exceeds u P v s ′ ( 10 − 3 − 2 ⋅ 10 − 5 ) u_Pv^{s'}(10^{-3}-2\cdot10^{-5}) u P v s ′ ( 1 0 − 3 − 2 ⋅ 1 0 − 5 ) , which is above the window when s − s ′ ≥ M h s-s'\ge M_h s − s ′ ≥ M h . If J h + 1 ≤ s − s ′ < M h J_h+1\le s-s'<M_h J h + 1 ≤ s − s ′ < M h , cleanness puts the first restart after s ′ s' s ′ beyond level s s s , at distance r ≥ J h + 2 r\ge J_h+2 r ≥ J h + 2 , so the group lies above u P v s ′ ( Δ − 2 v s ′ ( q 1 ∗ + 10 − 5 ) ) ( 1 − 2 ⋅ 10 − 5 ) u_Pv^{s'}(\Delta-2v^{s'}(q_1^*+10^{-5}))(1-2\cdot10^{-5}) u P v s ′ ( Δ − 2 v s ′ ( q 1 ∗ + 1 0 − 5 )) ( 1 − 2 ⋅ 1 0 − 5 ) , which is above the window by (K1) and (K2) of Proposition S4.19. □
The landing set Λ h l a n d \Lambda_h^{\mathrm{land}} Λ h land consists of the remainders y s y_s y s at the last pattern s s s of tails eventually in the core, evaluated on the whole parameter window.
h = 1 h=1 h = 1 : y s = t k + 1 ( 1 + t e 1 + t e 1 + e 2 + ⋯ ) y_s=t^{k+1}(1+t^{e_1}+t^{e_1+e_2}+\cdots) y s = t k + 1 ( 1 + t e 1 + t e 1 + e 2 + ⋯ ) with k > 4 k>4 k > 4 and e j ∈ { 3 , 4 } e_j\in\{3,4\} e j ∈ { 3 , 4 } , or y s = 0 y_s=0 y s = 0 ;
h = 3 h=3 h = 3 : y s = t k + 1 w ( t ) y_s=t^{k+1}w(t) y s = t k + 1 w ( t ) , where w ( t ) = ∑ i ≥ 0 w i t i w(t)=\sum_{i\ge0}w_it^i w ( t ) = ∑ i ≥ 0 w i t i for an infinite word w = 1 w 1 w 2 ⋯ w=1w_1w_2\cdots w = 1 w 1 w 2 ⋯ over { 0 , 1 } \{0,1\} { 0 , 1 } satisfying the core condition, with k ≥ 2 k\ge2 k ≥ 2 , or with k = 1 k=1 k = 1 and w w w beginning with 10, 110 or 11101110; or y s = 0 y_s=0 y s = 0 ;
h = 4 h=4 h = 4 : y s = t k + 1 / ( 1 − t ) y_s=t^{k+1}/(1-t) y s = t k + 1 / ( 1 − t ) with k ≥ 1 k\ge1 k ≥ 1 , or y s = 0 y_s=0 y s = 0 ;
h = 7 h=7 h = 7 : y s = t k + 1 w ( t ) y_s=t^{k+1}w(t) y s = t k + 1 w ( t ) for a core word w w w (k ≥ 1 k\geq1 k ≥ 1 ), or y s = 0 y_s=0 y s = 0 , or y s = t w ( t ) y_s=tw(t) y s = tw ( t ) for a zero-free, weakly admissible word w w w beginning with one of the 32 windows after d d d and containing no window at the following levels.
Proposition 7.11 (Landing Lemma). Let h ∈ { 1 , 3 , 4 , 7 } h\in\{1,3,4,7\} h ∈ { 1 , 3 , 4 , 7 } , y ∈ Λ h l a n d y\in\Lambda_h^{\mathrm{land}} y ∈ Λ h land , ∣ x − ( 1 + y ) ∣ ≤ ϵ h |x-(1+y)|\leq\epsilon_h ∣ x − ( 1 + y ) ∣ ≤ ϵ h with ϵ h = 10 − 145 , 10 − 80 , 10 − 75 , 10 − 80 \epsilon_h=10^{-145},10^{-80},10^{-75},10^{-80} ϵ h = 1 0 − 145 , 1 0 − 80 , 1 0 − 75 , 1 0 − 80 , and t ∈ [ a h − 10 − 90 , b h + 10 − 86 ] t\in[a_h-10^{-90},b_h+10^{-86}] t ∈ [ a h − 1 0 − 90 , b h + 1 0 − 86 ] ( [ a 1 − 10 − 150 , b 1 + 10 − 140 ] [a_1-10^{-150},b_1+10^{-140}] [ a 1 − 1 0 − 150 , b 1 + 1 0 − 140 ] for h = 1 h=1 h = 1 , b 4 + 10 − 85 b_4+10^{-85} b 4 + 1 0 − 85 for h = 4 h=4 h = 4 ). Then the greedy expansion of x x x has a C h C_h C h -pattern at a level ≥ P 0 ∗ \geq P_0^* ≥ P 0 ∗ , with P 0 ∗ = 13 , 7 , 6 , 4 P_0^*=13,7,6,4 P 0 ∗ = 13 , 7 , 6 , 4 , decided by digits common to all such x x x and t t t ; its first R h = 26 , 12 , 11 , 8 R_h=26,12,11,8 R h = 26 , 12 , 11 , 8 remainders exceed 10 − 20 10^{-20} 1 0 − 20 .
This is Proposition S4.19. Since v R h + 1 ( q 1 ∗ + 10 − 5 ) ≤ 7 ⋅ 10 − 5 L min ∘ v^{R_h+1}(q_1^*+10^{-5})\leq7\cdot10^{-5}L_{\min}^{\circ} v R h + 1 ( q 1 ∗ + 1 0 − 5 ) ≤ 7 ⋅ 1 0 − 5 L m i n ∘ , condition (d) of Corollary 7.8 then holds at the next tip.
Proof of Theorem D
Fix h ∈ { 1 , 3 , 4 , 7 } h\in\{1,3,4,7\} h ∈ { 1 , 3 , 4 , 7 } and the constants of Table 2 . Put F min = ( μ − 10 − 4 ) L min ∘ F_{\min}=(\mu-10^{-4})L_{\min}^{\circ} F m i n = ( μ − 1 0 − 4 ) L m i n ∘ ; Proposition S4.19 checks v P 0 ∗ L max ∘ ( 1 + μ ) < F min v^{P_0^*}L_{\max}^{\circ}(1+\mu)<F_{\min} v P 0 ∗ L m a x ∘ ( 1 + μ ) < F m i n and P 0 ∗ ≥ J h + 2 P_0^*\geq J_h+2 P 0 ∗ ≥ J h + 2 . Call a relative tip v = m ( P ) v=m(P) v = m ( P ) good if v ∈ [ t 0 , t 0 + X h ] v\in[t_0,t_0+X_h] v ∈ [ t 0 , t 0 + X h ] , F ( P ) ≥ F min F(P)\geq F_{\min} F ( P ) ≥ F m i n , P = P 0 P=P_0 P = P 0 or ∣ P ∣ ≥ 158 |P|\geq158 ∣ P ∣ ≥ 158 , and the tail of f v f_v f v has a C h C_h C h -pattern at a level ≥ P 0 ∗ \geq P_0^* ≥ P 0 ∗ .
Claim. A good tip v v v either carries infinitely many gaps of D h \mathcal{D}_h D h accumulating at v v v , or has a gap of D h \mathcal{D}_h D h on its right followed by a good tip v ′ v' v ′ with v < v ′ ≤ v + u P v P 0 ∗ L max ∘ ( 1 + 10 − 4 ) v<v'\leq v+u_Pv^{P_0^*}L_{\max}^{\circ}(1+10^{-4}) v < v ′ ≤ v + u P v P 0 ∗ L m a x ∘ ( 1 + 1 0 − 4 ) .
Proof of the claim. At a clean pattern of level s ≥ P 0 ∗ s\geq P_0^* s ≥ P 0 ∗ with n + s ≥ 158 n+s\geq158 n + s ≥ 158 , the hypotheses (a)–(c) of Corollary 7.8 hold: (a) and (b) by Proposition 7.9 and Lemma 7.10, and (c) because v s L max ∘ ( 1 + μ ) < F min ≤ F ( P ) v^sL_{\max}^{\circ}(1+\mu)<F_{\min}\leq F(P) v s L m a x ∘ ( 1 + μ ) < F m i n ≤ F ( P ) . (i) Suppose that the tail has infinitely many such patterns. Each gives the gap ( v + u P v s L − ( 1 − μ ) , m ( P ′ ) ) (v+u_Pv^sL_-(1-\mu),m(P')) ( v + u P v s L − ( 1 − μ ) , m ( P ′ )) , and for two of them, s < s ′ ′ s<s'' s < s ′′ , the gap and the tip of level s ′ ′ s'' s ′′ lie in groups of levels > s >s > s , hence below the gap of level s s s by the points m ( P ′ ) ∈ D h m(P')\in\mathcal{D}_h m ( P ′ ) ∈ D h and accumulate at v v v . (ii) Otherwise the tail has only finitely many patterns at levels ≥ P 0 ∗ \geq P_0^* ≥ P 0 ∗ : if it had infinitely many, those beyond the last restart, or, if there are infinitely many restarts, those at relative levels in [ M h , M E − 10 ] [M_h,M_E-10] [ M h , M E − 10 ] after each restart (which exist by (EX') of Proposition S4.19), would be clean, and all but finitely many have n + s ≥ 158 n+s\geq158 n + s ≥ 158 . Let s s s be the last pattern at a level ≥ P 0 ∗ \geq P_0^* ≥ P 0 ∗ . Then n + s ≥ 158 n+s\geq158 n + s ≥ 158 , by the choice of P 0 P_0 P 0 and i 1 i_1 i 1 if P = P 0 P=P_0 P = P 0 . The pattern is clean: a restart at a level in ( s − M h , s ] (s-M_h,s] ( s − M h , s ] would be followed by M E M_E M E digits of d β ( 1 ) d_\beta(1) d β ( 1 ) , which contain a pattern at a relative level at least M h M_h M h , hence after s s s . After $s the tail lies in the core, and by the case analysis that defines Λ h l a n d \Lambda_h^{\mathrm{land}} Λ h land its remainder y s y_s y s belongs to the landing set. By Corollary 7.8 and Lemma 7.4, v ′ = m ( P ′ ) v'=m(P') v ′ = m ( P ′ ) has the greedy tail, up to restarts, of a weight x x x with ∣ x − ( 1 + y s ) ∣ ≤ η ˉ n + s + X h ( 2 h + d ˉ ) / ( 1 − τ ) 2 ≤ ϵ h |x-(1+y_s)|\leq\bar\eta_{n+s}+X_h(2h+\bar{d})/(1-\tau)^2\leq\epsilon_h ∣ x − ( 1 + y s ) ∣ ≤ η ˉ n + s + X h ( 2 h + d ˉ ) / ( 1 − τ ) 2 ≤ ϵ h (Proposition S4.19), where y s y_s y s may be evaluated at the parameter v ′ v' v ′ . By Proposition 7.11 the first R h R_h R h remainders of the greedy expansion of x x x exceed $10^{-20}$, which excludes restarts there because δ k < 10 − 20 \delta_k<10^{-20} δ k < 1 0 − 20 ; this is (d), so F ( P ′ ) ≥ F min F(P')\geq F_{\min} F ( P ′ ) ≥ F m i n . The greedy expansion of x x x also has a pattern at a level ≥ P 0 ∗ \geq P_0^* ≥ P 0 ∗ ; it survives in the tail of v ′ v' v ′ unless a restart precedes it, in which case the following M E M_E M E digits of d β ( 1 ) d_\beta(1) d β ( 1 ) contain a pattern at a level ≥ P 0 ∗ \ge P_0^* ≥ P 0 ∗ . Finally
∣ P ′ ∣ = n + s ≥ 158 and v ′ ≤ v + u P v s L max ∘ ( 1 + 10 − 4 ) . |P'|=n+s\ge158\quad\text{and}\quad v'\le v+u_Pv^sL^\circ_{\max}(1+10^{-4}). ∣ P ′ ∣ = n + s ≥ 158 and v ′ ≤ v + u P v s L m a x ∘ ( 1 + 1 0 − 4 ) .
Start. By Propositions S4.16 and S4.19, t 0 = m ( P 0 ) t_0=m(P_0) t 0 = m ( P 0 ) has F ( P 0 ) ≥ F min F(P_0)\ge F_{\min} F ( P 0 ) ≥ F m i n , its remainders satisfy p n 0 < 2 p_{n_0}<2 p n 0 < 2 and p k < 1 p_k<1 p k < 1 for n 0 < k ≤ min ( 200 , N h − 2 ) n_0<k\le\min(200,N_h-2) n 0 < k ≤ min ( 200 , N h − 2 ) , and its certified tail has C h C_h C h -patterns at levels ≥ P 0 ∗ \ge P_0^* ≥ P 0 ∗ , the deepest at absolute position i 1 i_1 i 1 . So t 0 t_0 t 0 is good. Every jump of the iteration is made at an absolute depth at least i 1 i_1 i 1 , the first by at most v i 1 − 1 L max ∘ ( 1 + 10 − 4 ) / ( 2 κ v ) v^{i_1-1}L^\circ_{\max}(1+10^{-4})/(2\kappa_v) v i 1 − 1 L m a x ∘ ( 1 + 1 0 − 4 ) / ( 2 κ v ) and each later one by a factor v P 0 ∗ v^{P_0^*} v P 0 ∗ less, so every tip lies in [ t 0 , t 0 + X h ] [t_0,t_0+X_h] [ t 0 , t 0 + X h ] . This interval is inside the windows of all the finite lemmas, and the weight shifts and parameter changes between consecutive tips are below ϵ h \epsilon_h ϵ h . Hence every tip produced is good.
Conclusion. Iterating the claim from t 0 t_0 t 0 , either some tip carries infinitely many gaps accumulating at it, or there are infinitely many tips v 0 < v 1 < ⋯ v_0<v_1<\cdots v 0 < v 1 < ⋯ with a gap of D h D_h D h between consecutive ones. In both cases the gaps are bounded components of ( 0 , 1 ) ∖ D h (0,1)\setminus D_h ( 0 , 1 ) ∖ D h separated by points of D h D_h D h , so D h D_h D h and ( 0 , 1 ) ∖ D h (0,1)\setminus D_h ( 0 , 1 ) ∖ D h have infinitely many components. The accumulation point t ∞ t_\infty t ∞ of the gaps lies in A h A_h A h : in the first case t ∞ = v t_\infty=v t ∞ = v has a tail with infinitely many nonzero corner digits and f v ′ ′ ( v ) > 0 f_v''(v)>0 f v ′′ ( v ) > 0 , so f v + z N ∈ F h f_v+z^N\in F_h f v + z N ∈ F h for infinitely many N N N and the argument of Lemma 2.7 gives nonreal zeros converging to t ∞ t_\infty t ∞ ; in the second case t ∞ = lim v j t_\infty=\lim v_j t ∞ = lim v j is not isolated in D h D_h D h , and Lemma 2.7 applies.
It remains to treat the diagonal trace. The gaps are separated by tips to which Lemma 7.4 applies: in the second case the v j v_j v j , and in the first case the tips m ( P s ′ ) m(P'_s) m ( P s ′ ) of the clean patterns with n + s ≥ 158 n+s\ge158 n + s ≥ 158 (at a pattern y s < 1 y_s<1 y s < 1 , as noted after Corollary 7.8, so the weight 1 + y s 1+y_s 1 + y s is below 2). Such a tip v v v has f v ′ ′ ( v ) = 2 κ v > 0 f_v''(v)=2\kappa_v>0 f v ′′ ( v ) = 2 κ v > 0 and corner digits at most d ˉ < 2 h \bar d<2h d ˉ < 2 h after its first tail digit, that is, coefficients a N > − h a_N>-h a N > − h . For such N N N the series f v − z N f_v-z^N f v − z N is admissible, negative at v v v and positive at v ± r N v\pm r_N v ± r N for suitable r N → 0 r_N\to0 r N → 0 , so it has two real zeros in ( v − r N , v + r N ) (v-r_N,v+r_N) ( v − r N , v + r N ) , one on each side of v v v . Hence ( v , v ) (v,v) ( v , v ) is a limit of pairs ( γ , λ ) (\gamma,\lambda) ( γ , λ ) , γ ≠ λ \gamma\ne\lambda γ = λ , of zeros of one member of F h F_h F h . Conversely a diagonal limit of such pairs is a double zero, by coefficient compactness and Rolle's theorem. So Δ h ∩ ( 0 , 1 ) ⊆ D h \Delta_h\cap(0,1)\subseteq D_h Δ h ∩ ( 0 , 1 ) ⊆ D h contains all the separating tips, and the gaps of D h D_h D h separate them in Δ h \Delta_h Δ h . This proves Theorem D.
Remark 7.12. Height 8 has β = 3.52706 \beta=3.52706 β = 3.52706 , almost the base of height 7, but its slices are fatter, and a floating-point exploration (not certified) finds moves only at the three zero-free windows. The resulting core then has dimension about 0.63, and so does its landing set, so a landing branch and bound cannot be expected to close; the planar core C 8 C_8 C 8 is settled by Proposition 6.16. The argument of this section also does not decide whether A h A_h A h has infinitely many components: a tip whose tail terminates may be an isolated, inactive double zero.
The remaining question
Theorem A determines every connected height. Theorems B and C show that infinitely many components are the typical alternative. They leave the following precise question.
Conjecture 8.1. For every h ≥ 1 h\ge1 h ≥ 1 that is not a product of consecutive positive integers, C h C_h C h has infinitely many connected components.
Let E E E be the set of nonpronic heights whose cores have only finitely many components, and put E q = E ∩ { q 2 − 2 , q 2 − 1 , q 2 } E_q=E\cap\{q^2-2,q^2-1,q^2\} E q = E ∩ { q 2 − 2 , q 2 − 1 , q 2 } . Then
E ⊂ { 1 , 3 , 4 , 9 } ∪ { 25 , 63 , 81 } ∪ ⋃ q ≥ 13 E q , ∣ E q ∣ ≤ 1 , (36) E\subset\{1,3,4,9\}\cup\{25,63,81\}\cup\bigcup_{q\ge13}E_q,\qquad|E_q|\le1, \tag*{(36)} E ⊂ { 1 , 3 , 4 , 9 } ∪ { 25 , 63 , 81 } ∪ q ≥ 13 ⋃ E q , ∣ E q ∣ ≤ 1 , (36)
where E 14 ⊂ { 196 } E_{14}\subset\{196\} E 14 ⊂ { 196 } , E 15 = ∅ E_{15}=\varnothing E 15 = ∅ and E 16 ⊂ { 256 } E_{16}\subset\{256\} E 16 ⊂ { 256 } , by Propositions 6.14 and 6.15.
A Diophantine reduction
At a square-adjacent height with q ≥ 4 q\ge4 q ≥ 4 , a finite anchor set forces a witness that is eventually constant with the middle exceptional digit (Corollary 6.10).
Proposition 8.2. Let q ≥ 4 q\ge4 q ≥ 4 , k ∈ { 0 , 1 , 2 } k\in\{0,1,2\} k ∈ { 0 , 1 , 2 } and h = q 2 − 2 + k h=q^2-2+k h = q 2 − 2 + k . If C h C_h C h has only finitely many connected components, then there is a polynomial E ( z ) = ∑ j = 2 N e j z j E(z)=\sum_{j=2}^{N}e_jz^j E ( z ) = ∑ j = 2 N e j z j of degree N ≥ 4 N\ge4 N ≥ 4 with integer coefficients ∣ e j ∣ ≤ q − 1 |e_j|\le q-1 ∣ e j ∣ ≤ q − 1 such that
R E ( z ) = ( 1 − q z ) ( 1 − ( q + 1 ) z ) + ( k − 1 ) z 2 + ( 1 − z ) E ( z ) (37) R_E(z)=(1-qz)(1-(q+1)z)+(k-1)z^2+(1-z)E(z) \tag*{(37)} R E ( z ) = ( 1 − q z ) ( 1 − ( q + 1 ) z ) + ( k − 1 ) z 2 + ( 1 − z ) E ( z ) (37) has a double zero 1 / β 1/\beta 1/ β in ( 1 / ( q + 1 ) , 1 / q ) (1/(q+1),1/q) ( 1/ ( q + 1 ) , 1/ q ) . Moreover g ( 1 ) 2 g(1)^2 g ( 1 ) 2 divides q 2 − q − 1 + k q^2-q-1+k q 2 − q − 1 + k , where g g g is the minimal polynomial of β \beta β .
Proof. By Corollary 6.10 , some square-adjacent witness F F F is eventually constant with digit c = m + q − 1 = q 2 − q − 1 + k c=m+q-1=q^{2}-q-1+k c = m + q − 1 = q 2 − q − 1 + k . Let N N N be its last level with a N ( F ) ≠ c a_N(F)\ne c a N ( F ) = c ; it exists, because the numerator in (6.9) has discriminant 5 − 4 k ≠ 0 5-4k\ne0 5 − 4 k = 0 . Put e j = a j ( F ) − c e_j=a_j(F)-c e j = a j ( F ) − c , so that ∣ e j ∣ ≤ q − 1 \lvert e_j\rvert\le q-1 ∣ e j ∣ ≤ q − 1 . Then F F F is the channel series (6.9) plus E E E , so ( 1 − z ) F = R E (1-z)F=R_E ( 1 − z ) F = R E , and the anchor of F F F is a double zero of R E R_E R E in ( 1 / ( q + 1 ) , 1 / q ) (1/(q+1),1/q) ( 1/ ( q + 1 ) , 1/ q ) by Lemma 6.2 . The reversed polynomial x N + 1 R E ( 1 / x ) x^{N+1}R_E(1/x) x N + 1 R E ( 1/ x ) is monic of degree N + 1 N+1 N + 1 with integer coefficients, and β ∈ ( q , q + 1 ) \beta\in(q,q+1) β ∈ ( q , q + 1 ) is a double root of it. As β \beta β is not an integer, g g g has degree at least two and g 2 g^2 g 2 divides the reversed polynomial in Z [ x ] \mathbb{Z}[x] Z [ x ] . Hence N + 1 ≥ 4 N+1\ge4 N + 1 ≥ 4 , and equality would make the reversed polynomial equal to g 2 g^2 g 2 , whose coefficient of x 3 x^3 x 3 is even, whereas it is − ( 2 q + 1 ) -(2q+1) − ( 2 q + 1 ) . Evaluating at x = 1 x=1 x = 1 gives g ( 1 ) 2 ∣ R E ( 1 ) = q 2 − q − 1 + k g(1)^2\mid R_E(1)=q^2-q-1+k g ( 1 ) 2 ∣ R E ( 1 ) = q 2 − q − 1 + k .
Conversely, if R E R_E R E has a double zero in ( 0 , 1 ) (0,1) ( 0 , 1 ) for some E E E with ∣ e j ∣ ≤ q − 1 \lvert e_j\rvert\le q-1 ∣ e j ∣ ≤ q − 1 , then R E / ( 1 − z ) R_E/(1-z) R E / ( 1 − z ) is a square-adjacent witness that is eventually constant with digit m + q − 1 m+q-1 m + q − 1 . Thus Conjecture 8.1 holds at every height other than 1, 3, 4 and 9 if none of the polynomials (8.2), for q ≥ 4 q\ge4 q ≥ 4 and k ∈ { 0 , 1 , 2 } k\in\{0,1,2\} k ∈ { 0 , 1 , 2 } , has a double zero in ( 0 , 1 ) (0,1) ( 0 , 1 ) .
Question 8.3. Can a polynomial (8.2) have a double zero in ( 0 , 1 ) (0,1) ( 0 , 1 ) ? More generally, does an eventually constant square-adjacent witness exist?
We know of none. The searches of the companion [13 ] tested every constant digit after each retained prefix, with prefix lengths 16, 14, 12, 11, 10, 10 for q = 4 , … , 9 q=4,\ldots,9 q = 4 , … , 9 , and found no eventually constant witness; this is a search result, not an exclusion. The criterion of Corollary 6.10 is only sufficient: a finite anchor set would not rule out infinitely many components produced by a different mechanism.
The anchor sets at the open heights are large: T 25 \mathcal{T}_{25} T 25 , T 63 \mathcal{T}_{63} T 63 and T 81 \mathcal{T}_{81} T 81 have at least 384, 332 and 369 points, the numbers of prefixes of length 15, 12 and 12 certified to extend to a square-adjacent witness by Lemma 4.6 (Proposition S4.15 of the supplement). The searches of the companion found no witness with a Pisot reciprocal anchor of degree at most six at 25, or of degree at most five at 63 and 81; they use numerical filters, so these negative results are not nonexistence proofs. The search behind Proposition 6.15 is exhaustive within its bounds and uses integer arithmetic only. At every other square-adjacent height with 4 ≤ q ≤ 12 4\le q\le12 4 ≤ q ≤ 12 it finds a series g 2 / ( 1 − z r ) g^2/(1-z^r) g 2 / ( 1 − z r ) with deg g ≤ 9 \deg g\le9 deg g ≤ 9 and r ≤ 4 r\le4 r ≤ 4 . At the heights 25, 63 and 81 it finds none with r ≤ 6 r\le6 r ≤ 6 and deg g ≤ 16 \deg g\le16 deg g ≤ 16 , 12 and 11 respectively, and none with r ≤ 6 r\le6 r ≤ 6 and deg g ≤ 9 \deg g\le9 deg g ≤ 9 at the three heights with q = 13 q=13 q = 13 .
The heights 1, 3, 4 and 9
At the heights 1 and 3, Proposition 4.4 never applies. Indeed, let F = 1 + ∑ j ≥ 1 a j z j F=1+\sum_{j\ge1}a_jz^j F = 1 + ∑ j ≥ 1 a j z j and t = 1 / β t=1/\beta t = 1/ β be as in that proposition, where we may take h ∗ = H = h h_*=H=h h ∗ = H = h . The equations F ( t ) = t F ′ ( t ) = 0 F(t)=tF'(t)=0 F ( t ) = t F ′ ( t ) = 0 give
∑ j ≥ 2 ( j − 1 ) a j t j = 1 , ∑ j ≥ 2 ( j − 2 ) a j t j = 2 + a 1 t . \sum_{j\ge2}(j-1)a_jt^j=1,\qquad\sum_{j\ge2}(j-2)a_jt^j=2+a_1t. j ≥ 2 ∑ ( j − 1 ) a j t j = 1 , j ≥ 2 ∑ ( j − 2 ) a j t j = 2 + a 1 t .
and the bounds m ≤ a j ≤ h m\le a_j\le h m ≤ a j ≤ h and a 1 ≥ − h a_1\ge-h a 1 ≥ − h turn these into ( β − 1 ) 2 ≥ m (\beta-1)^2\ge m ( β − 1 ) 2 ≥ m and ( 2 β − h ) ( β − 1 ) 2 ≤ h (2\beta-h)(\beta-1)^2\le h ( 2 β − h ) ( β − 1 ) 2 ≤ h . With the second inequality of (4.2), ( β − 1 ) 2 ≥ max { m , ℓ } (\beta-1)^2\ge\max\{m,\ell\} ( β − 1 ) 2 ≥ max { m , ℓ } , where m + ℓ = h m+\ell=h m + ℓ = h . For h = 1 h=1 h = 1 this gives β ≥ 2 \beta\ge2 β ≥ 2 and ( 2 β − 1 ) ( β − 1 ) 2 ≥ 3 (2\beta-1)(\beta-1)^2\ge3 ( 2 β − 1 ) ( β − 1 ) 2 ≥ 3 ; for h = 3 h=3 h = 3 it gives ( β − 1 ) 2 ≥ 2 (\beta-1)^2\ge2 ( β − 1 ) 2 ≥ 2 and ( 2 β − 3 ) ( β − 1 ) 2 > 3 (2\beta-3)(\beta-1)^2>3 ( 2 β − 3 ) ( β − 1 ) 2 > 3 . Both contradict ( 2 β − h ) ( β − 1 ) 2 ≤ h (2\beta-h)(\beta-1)^2\le h ( 2 β − h ) ( β − 1 ) 2 ≤ h . At height 4 these same inequalities, together with Δ > 0 \Delta>0 Δ > 0 , leave only the series 1 − 4 z + ∑ j ≥ 2 a j z j 1-4z+\sum_{j\ge2}a_jz^j 1 − 4 z + ∑ j ≥ 2 a j z j with 2 ≤ a j ≤ 4 2\le a_j\le4 2 ≤ a j ≤ 4 and 2.58 < β < 2.70 2.58<\beta<2.70 2.58 < β < 2.70 , and at height 9 the series 1 − 6 z + ∑ j ≥ 2 a j z j 1-6z+\sum_{j\ge2}a_jz^j 1 − 6 z + ∑ j ≥ 2 a j z j with 5 ≤ a j ≤ 9 5\le a_j\le9 5 ≤ a j ≤ 9 are candidates when β > 3.48 \beta>3.48 β > 3.48 . We know no series satisfying the hypotheses of the proposition at these two heights. The search program behind Table 1 , run for q = 2 q=2 q = 2 and q = 3 q=3 q = 3 , finds the two series of Proposition 6.16 , but none in these two families with a Pisot reciprocal anchor of degree at most six; it uses numerical filters, so this is not a nonexistence proof. The search of Proposition 6.15 finds no series g 2 / ( 1 − z r ) g^2/(1-z^r) g 2 / ( 1 − z r ) in them with r ≤ 12 r\le12 r ≤ 12 and deg g ≤ 36 \deg g\le36 deg g ≤ 36 at height 4, or deg g ≤ 24 \deg g\le24 deg g ≤ 24 at height 9. At the heights 1, 3 and 4 the real gaps of Theorem D do not settle the planar question, as noted in the introduction.
Proofs and computation
The proofs of Theorems A and B are in Sections 2–5, the proof of Theorem C in Section 6, and the proof of Theorem D in Section 7, whose finite inputs are the last four records of the supplement. The continuation estimates behind Proposition 3.3 , the planar separators behind Proposition 5.1 and the acceptance rules of all finite checks are in the supplement. Theorem B and the separation argument of § 4 use no computation. Every finite check is accepted by exact rational, symbolic or bounded-integer arithmetic in the computational companion [13 ] . The search programs that proposed the explicit witnesses of Proposition 6.14 use numerical filters; the certificates that accept these witnesses do not. The witnesses of Proposition 6.15 can be checked by hand: each needs the square of one integer polynomial and its signs at two integers, and Proposition 6.16 a few decimal bounds. No negative search result is used as a proof. Likewise, the programs that proposed the certificates of the minimum of D h D_h D h use floating-point arithmetic with safety margins, and only the exact checks accept them. Figure 3 is drawn from the statements of Theorems A–D and of Propositions 6.14 , 6.15 and 6.16; the other figures are drawn by Python scripts from exact data and from numerical illustrations identified as such in their captions. No figure is an image generated by an AI model.
The author conceived the project, directed the research and checked every argument. AI assistants (Anthropic Claude and OpenAI ChatGPT) were used under the author’s direction to help develop proofs, write and test code, and edit the text. Every finite verification can be replayed from the companion without any AI service.
Data and code availability
The computational companion [13 ] is deposited at Zenodo (doi:10.5281/zenodo.24117729) and supplied as supplementary material; it contains the certificate inputs, the checking programs and the replay logs, the search programs with their logs, and the data and scripts for the figures drawn from computed data. An addendum of 10 November 2026 accompanies it: exact checks of Propositions 6.15 and 6.16, and the program and logs of the searches of § 6.7 and 8. The README of the companion maps each record of § 17 of the supplement to its verifier, under the numbering and the working names explained in Section S4.3 of the supplement.
Acknowledgements
The author thanks David Juher and Joan Saldaña for discussions on collinear connectedness loci and restricted series, Christoph Bandt for suggestions on exposition and terminology, and Pablo Shmerkin for comments on related exposition. This work was supported by grant PID2023-146424NB-I00 (Spain) and the IFUdG 2022–2024 fellowship programme (Universitat de Girona and Banco Santander).
Declaration of generative AI and AI-assisted technologies in the manuscript preparation process
During the preparation of this work the author used Anthropic Claude and OpenAI ChatGPT in order to assist with developing proofs, writing code and editing the text. After using these tools, the author reviewed and edited the content as needed and takes full responsibility for the content of the published article.
Supplement
This supplement contains the technical parts of the proofs of Theorems A, C and D of the paper: the trace coordinates and continuation tests (Section S1), the complete real traces at the pronic heights behind Proposition 3.3 (Section S2), the residual planar separators behind Proposition 5.1 (Section S3), and the exact finite certificates with their acceptance rules, including those used in Sections 6 and 7 (Section S4). Numbers without the prefix S refer to the paper. Theorem B, the density-one statement and the separation argument of Section 4 do not depend on this supplement.
Trace coordinates and continuation tests
The continuation proofs in Section S2 use two finite-word tests developed here. Section S3 gives the remaining planar separators, and Section S4 states and verifies their finite arithmetic obligations.
Retain B h B_h B h , F h F_h F h , D h D_h D h , A h A_h A h and X X X from Section 2, and write Z h e x t = Z h ∩ X = Σ h Z_h^{\mathrm{ext}}=Z_h\cap X=\Sigma_h Z h ext = Z h ∩ X = Σ h . Proposition 2.3 supplies the root correspondence, real slice, symmetries and annulus. In particular,
C h e x t = cl X ( Z h e x t ∖ R ) . (38) \mathcal{C}_h^{\mathrm{ext}}=\operatorname{cl}_X\left(Z_h^{\mathrm{ext}}\setminus\mathbb{R}\right). \tag*{(38)} C h ext = cl X ( Z h ext ∖ R ) . (38)
The other inputs are the inclusion A h ⊆ D h A_h\subseteq D_h A h ⊆ D h from Lemma 2.6, persistence of double-zero intervals from Lemma 2.7, and the pronic bound of Lemma 3.1. These inputs precede the connectedness classification and do not assume it.
Lemma S1.1 (The pronic endpoint). For q ≥ 3 q\ge3 q ≥ 3 and h = q ( q − 1 ) h=q(q-1) h = q ( q − 1 ) , 1 / q ∈ D h 1/q\in D_h 1/ q ∈ D h .
Proof. The series ( 1 − q z ) 2 / ( 1 − z ) = 1 − ( 2 q − 1 ) z + ( q − 1 ) 2 ∑ j ≥ 2 z j (1-qz)^2/(1-z)=1-(2q-1)z+(q-1)^2\sum_{j\ge2}z^j ( 1 − q z ) 2 / ( 1 − z ) = 1 − ( 2 q − 1 ) z + ( q − 1 ) 2 ∑ j ≥ 2 z j belongs to F h F_h F h and has a double zero at 1 / q 1/q 1/ q . □
Lemma S1.2 (Real collision). A real limit of qualifying zeros of coefficientwise convergent bounded real series is a multiple zero of the limiting series. Here qualifying means nonreal or real multiple. If the limiting zero is exactly double with positive second derivative, nearby qualifying witnesses g g g at zeros z ≠ t z\ne t z = t satisfy g ( t ) > 0 g(t)>0 g ( t ) > 0 .
Proof. At a nonreal zero the conjugate zero is also present; the vanishing first divided difference converges to the derivative at the real limit. For real multiple zeros, pass to the derivative directly. For the last assertion, g ( t ) = ∣ z − t ∣ g [ t , z , z ‾ ] g(t)=\lvert z-t\rvert g[t,z,\overline{z}] g ( t ) = ∣ z − t ∣ g [ t , z , z ] , with confluent arguments for a real double zero. The second divided difference converges to f ′ ′ ( t ) / 2 > 0 f''(t)/2>0 f ′′ ( t ) /2 > 0 . □
Trace coordinates and two entry tests
Put N = 2 h + 1 N=2h+1 N = 2 h + 1 . For a finite word u = ( d 1 , … , d m ) ∈ B h m u=(d_1,\ldots,d_m)\in B_h^m u = ( d 1 , … , d m ) ∈ B h m , m ≥ 0 m\ge0 m ≥ 0 , define
P u ( z ) = z m − d 1 z m − 1 − ⋯ − d m , z u ( c ) = 2 c P u ( c ) , P_u(z)=z^m-d_1z^{m-1}-\cdots-d_m,\qquad z_u(c)=2cP_u(c), P u ( z ) = z m − d 1 z m − 1 − ⋯ − d m , z u ( c ) = 2 c P u ( c ) ,
with P ∅ = 1 P_{\varnothing}=1 P ∅ = 1 for the empty word. Appending d d d gives z u ⌢ d = c ( z u − 2 d ) z_{u\smallfrown d}=c(z_u-2d) z u ⌢ d = c ( z u − 2 d ) , so z u ( c ) z_u(c) z u ( c ) is obtained from the initial point 2 c 2c 2 c by successively applying the inverse maps z ↦ c ( z − 2 d 1 ) , … , z ↦ c ( z − 2 d m ) z\mapsto c(z-2d_1),\ldots,z\mapsto c(z-2d_m) z ↦ c ( z − 2 d 1 ) , … , z ↦ c ( z − 2 d m ) .
At x > 1 x>1 x > 1 , set
V u ( x ) = P u ( x ) + x P u ′ ( x ) , S u ( x ) = 2 P u ( x ) + x P u ′ ( x ) . (39) V_u(x)=P_u(x)+xP_u'(x),\qquad S_u(x)=2P_u(x)+xP_u'(x). \tag*{(39)} V u ( x ) = P u ( x ) + x P u ′ ( x ) , S u ( x ) = 2 P u ( x ) + x P u ′ ( x ) . (39)
The empty word gives ( V , S ) = ( 1 , 2 ) (V,S)=(1,2) ( V , S ) = ( 1 , 2 ) , and appending d d d acts by
V u ⌢ d ( x ) = x S u ( x ) − d , S u ⌢ d ( x ) = 2 V u ⌢ d ( x ) − x V u ( x ) . (40) V_{u\smallfrown d}(x)=xS_u(x)-d,\qquad S_{u\smallfrown d}(x)=2V_{u\smallfrown d}(x)-xV_u(x). \tag*{(40)} V u ⌢ d ( x ) = x S u ( x ) − d , S u ⌢ d ( x ) = 2 V u ⌢ d ( x ) − x V u ( x ) . (40)
Put
τ h ( x ) = N − 2 x 2 x 2 , σ h ( x ) = N 2 x . (41) \tau_h(x)=\frac{N-2x}{2x^2},\qquad\sigma_h(x)=\frac{N}{2x}. \tag*{(41)} τ h ( x ) = 2 x 2 N − 2 x , σ h ( x ) = 2 x N . (41)
Then
1 + x τ h ( x ) = σ h ( x ) , τ h ( x ) > 1 2 ⟺ x 2 + 2 x < N . (42) 1+x\tau_h(x)=\sigma_h(x),\qquad\tau_h(x)>\frac{1}{2}\Longleftrightarrow x^2+2x<N. \tag*{(42)} 1 + x τ h ( x ) = σ h ( x ) , τ h ( x ) > 2 1 ⟺ x 2 + 2 x < N . (42)
For c = a + i b c=a+ib c = a + ib with b ≠ 0 b\ne0 b = 0 , the complex coordinates and their inverse are
V c ( z ) = Im z 2 b , S c ( z ) = Im ( c z ) 2 ∣ c ∣ b ; Im z = 2 b V c ( z ) , Re z = 2 ( ∣ c ∣ S c ( z ) − a V c ( z ) ) . (43) V_c(z)=\frac{\operatorname{Im}z}{2b},\qquad S_c(z)=\frac{\operatorname{Im}(cz)}{2\lvert c\rvert b};\qquad\operatorname{Im}z=2bV_c(z),\qquad\operatorname{Re}z=2\left(\lvert c\rvert S_c(z)-aV_c(z)\right). \tag*{(43)} V c ( z ) = 2 b Im z , S c ( z ) = 2 ∣ c ∣ b Im ( cz ) ; Im z = 2 b V c ( z ) , Re z = 2 ( ∣ c ∣ S c ( z ) − a V c ( z ) ) . (43)
Under an inverse digit map z ↦ z ′ = c z − 2 c d z\mapsto z'=cz-2cd z ↦ z ′ = cz − 2 c d , the coordinates V c ′ = V c ( z ′ ) V'_c=V_c(z') V c ′ = V c ( z ′ ) and S c ′ = S c ( z ′ ) S'_c=S_c(z') S c ′ = S c ( z ′ ) satisfy
V c ′ = ∣ c ∣ S c − d , S c ′ = 2 a ∣ c ∣ V c ′ − ∣ c ∣ V c . (44) V'_c=\lvert c\rvert S_c-d,\qquad S'_c=\frac{2a}{\lvert c\rvert}V'_c-\lvert c\rvert V_c. \tag*{(44)} V c ′ = ∣ c ∣ S c − d , S c ′ = ∣ c ∣ 2 a V c ′ − ∣ c ∣ V c . (44)
Lemma S1.3 (Real-axis continuation). For a fixed word u u u , the functions V c ( z u ( c ) ) V_c(z_u(c)) V c ( z u ( c )) and S c ( z u ( c ) ) S_c(z_u(c)) S c ( z u ( c )) extend continuously to c = x > 1 c=x>1 c = x > 1 , with values V u ( x ) V_u(x) V u ( x ) and S u ( x ) S_u(x) S u ( x ) .
Proof. For a real-coefficient polynomial g g g , the quotient Im g ( a + i b ) / b \operatorname{Im}g(a+ib)/b Im g ( a + ib ) / b tends to g ′ ( x ) g'(x) g ′ ( x ) as a + i b → x a+ib\to x a + ib → x . Apply this to g ( c ) = 2 c P u ( c ) g(c)=2cP_u(c) g ( c ) = 2 c P u ( c ) and g ( c ) = 2 c 2 P u ( c ) g(c)=2c^2P_u(c) g ( c ) = 2 c 2 P u ( c ) , and use ∣ c ∣ → x |c|\to x ∣ c ∣ → x in the second coordinate. □
Proposition S1.4 (Interior criterion). Let c = a + i b ∈ X c=a+ib\in X c = a + ib ∈ X with a > 0 a>0 a > 0 , b ≠ 0 b\ne0 b = 0 , and ∣ c ∣ 2 + 2 a < N |c|^2+2a<N ∣ c ∣ 2 + 2 a < N . If some finite word u u u satisfies
∣ V c ( z u ( c ) ) ∣ < N − 2 a 2 ∣ c ∣ 2 , ∣ S c ( z u ( c ) ) ∣ < N 2 ∣ c ∣ |V_c(z_u(c))|<\frac{N-2a}{2|c|^2},\qquad|S_c(z_u(c))|<\frac{N}{2|c|} ∣ V c ( z u ( c )) ∣ < 2∣ c ∣ 2 N − 2 a , ∣ S c ( z u ( c )) ∣ < 2∣ c ∣ N then c ∈ int Z h e x t c\in\operatorname{int} Z_h^{\mathrm{ext}} c ∈ int Z h ext . For a real parameter 1 < x ≤ h + 1 1<x\le h+1 1 < x ≤ h + 1 , the same conclusion follows from
x 2 + 2 x < N , ∣ V u ( x ) ∣ < τ h ( x ) , ∣ S u ( x ) ∣ < σ h ( x ) . x^2+2x<N,\qquad|V_u(x)|<\tau_h(x),\qquad|S_u(x)|<\sigma_h(x). x 2 + 2 x < N , ∣ V u ( x ) ∣ < τ h ( x ) , ∣ S u ( x ) ∣ < σ h ( x ) .
Proof. Idea. Nearest digits keep the inverse orbit of the marked point in the open trace rectangle given by the two inequalities, and a bounded orbit is the series criterion. Choose d ∈ B h d\in B_h d ∈ B h nearest to ∣ c ∣ S c |c|S_c ∣ c ∣ S c ; by the second inequality, ∣ c ∣ S c ∈ ( − h − 1 / 2 , h + 1 / 2 ) |c|S_c\in(-h-1/2,h+1/2) ∣ c ∣ S c ∈ ( − h − 1/2 , h + 1/2 ) , so ∣ ∣ c ∣ S c − d ∣ ≤ 1 / 2 \bigl||c|S_c-d\bigr|\le1/2 ∣ c ∣ S c − d ≤ 1/2 , and (S1.7) gives
∣ V c ′ ∣ ≤ 1 2 < N − 2 a 2 ∣ c ∣ 2 , ∣ S c ′ ∣ ≤ a ∣ c ∣ + ∣ c ∣ ∣ V c ∣ < N 2 ∣ c ∣ . |V'_c|\le\frac{1}{2}<\frac{N-2a}{2|c|^2},\qquad|S'_c|\le\frac{a}{|c|}+|c||V_c|<\frac{N}{2|c|}. ∣ V c ′ ∣ ≤ 2 1 < 2∣ c ∣ 2 N − 2 a , ∣ S c ′ ∣ ≤ ∣ c ∣ a + ∣ c ∣∣ V c ∣ < 2∣ c ∣ N .
So, along suitable inverse digits, the coordinates stay in the rectangle and, by (S1.6), the remainders stay bounded; hence
1 − ∑ k = 1 m d k c − k = z ( d 1 , … , d m ) ( c ) 2 c m + 1 → 0. 1-\sum_{k=1}^{m}d_kc^{-k}=\frac{z_{(d_1,\ldots,d_m)}(c)}{2c^{m+1}}\to0. 1 − k = 1 ∑ m d k c − k = 2 c m + 1 z ( d 1 , … , d m ) ( c ) → 0.
Negating digits gives the criterion of Proposition 2.3, so c ∈ Z h e x t c\in Z_h^{\mathrm{ext}} c ∈ Z h ext . The hypotheses persist near c c c , and the argument applies at nearby parameters, giving interiority. At a real x x x , Lemma S1.3 gives them at all nearby nonreal parameters, and nearby real ones lie in the real slice, since x 2 + 2 x < 2 h + 1 x^2+2x<2h+1 x 2 + 2 x < 2 h + 1 forces x < h + 1 x<h+1 x < h + 1 . □
Corollary S1.5 (Nearest-digit lift). If 1 < x ≤ h + 1 1<x\le h+1 1 < x ≤ h + 1 , x 2 + 2 x < N x^2+2x<N x 2 + 2 x < N , and ∣ V u ( x ) ∣ ≤ τ h ( x ) |V_u(x)|\le\tau_h(x) ∣ V u ( x ) ∣ ≤ τ h ( x ) , ∣ S u ( x ) ∣ ≤ σ h ( x ) |S_u(x)|\le\sigma_h(x) ∣ S u ( x ) ∣ ≤ σ h ( x ) , then x x x is an interior point of Z h e x t Z_h^{\mathrm{ext}} Z h ext .
Proof. Choose d ∈ B h d\in B_h d ∈ B h nearest to x S u xS_u x S u . Then
∣ V u ⌢ d ∣ ≤ 1 2 < τ h , ∣ S u ⌢ d ∣ ≤ 1 + x τ h = σ h . |V_{u\smallfrown d}|\le\frac{1}{2}<\tau_h,\qquad|S_{u\smallfrown d}|\le1+x\tau_h=\sigma_h. ∣ V u ⌢ d ∣ ≤ 2 1 < τ h , ∣ S u ⌢ d ∣ ≤ 1 + x τ h = σ h .
A second nearest digit makes both inequalities strict, the second because ∣ V u ⌢ d ∣ < τ h |V_{u\smallfrown d}|<\tau_h ∣ V u ⌢ d ∣ < τ h . Apply Proposition S1.4. □
Corollary S1.6. If x > 1 x>1 x > 1 is an interior point of Z h e x t Z_h^{\mathrm{ext}} Z h ext , then 1 / x ∈ D h 1/x\in D_h 1/ x ∈ D h . In particular, ( 1 / h + 1 , 1 ) ⊂ D h (1/\sqrt{h+1},1)\subset D_h ( 1/ h + 1 , 1 ) ⊂ D h .
Proof. An interior point x x x is a limit of nonreal points of Z h e x t Z_h^{\mathrm{ext}} Z h ext , which lie in C h C_h C h ; so x ∈ C h x\in C_h x ∈ C h . Apply Lemma 2.6 and Proposition 2.3. □
The same argument shows that Z h e x t Z_h^{\mathrm{ext}} Z h ext and C h e x t C_h^{\mathrm{ext}} C h ext have the same interior points: an open set inside Z h e x t Z_h^{\mathrm{ext}} Z h ext has dense nonreal points, which lie in C h e x t C_h^{\mathrm{ext}} C h ext by (S1.1), so it lies in C h e x t C_h^{\mathrm{ext}} C h ext ; the converse holds as C h e x t ⊂ Z h e x t C_h^{\mathrm{ext}}\subset Z_h^{\mathrm{ext}} C h ext ⊂ Z h ext . Statements about interior points below may therefore refer to either set. In particular, by Proposition 2.3, the central annulus { 1 < ∣ c ∣ < h + 1 } \{1<|c|<\sqrt{h+1}\} { 1 < ∣ c ∣ < h + 1 } lies in C h e x t C_h^{\mathrm{ext}} C h ext .
The second test uses P = P u ( x ) P=P_u(x) P = P u ( x ) and V = V u ( x ) V=V_u(x) V = V u ( x ) , so that S u ( x ) = P + V S_u(x)=P+V S u ( x ) = P + V ; the empty word gives ( P , V ) = ( 1 , 1 ) (P,V)=(1,1) ( P , V ) = ( 1 , 1 ) . By (S1.3), appending a digit d d d acts by
( P , V ) ⟼ ( x P − d , x V + x P − d ) . (45) (P,V)\longmapsto(xP-d,xV+xP-d). \tag*{(45)} ( P , V ) ⟼ ( x P − d , x V + x P − d ) . (45)
Lemma S1.7 (Bounded-orbit test). Let h ≥ 1 h\ge1 h ≥ 1 and x > 1 x>1 x > 1 , and let d 1 , d 2 , … ∈ B h d_1,d_2,\ldots\in B_h d 1 , d 2 , … ∈ B h . Put ( P 0 , V 0 ) = ( 1 , 1 ) (P_0,V_0)=(1,1) ( P 0 , V 0 ) = ( 1 , 1 ) and
( P m , V m ) = ( x P m − 1 − d m , x V m − 1 + x P m − 1 − d m ) ( m ≥ 1 ) . (P_m,V_m)=(xP_{m-1}-d_m,\ xV_{m-1}+xP_{m-1}-d_m)\qquad(m\ge1). ( P m , V m ) = ( x P m − 1 − d m , x V m − 1 + x P m − 1 − d m ) ( m ≥ 1 ) . If the orbit ( P m , V m ) m ≥ 0 (P_m,V_m)_{m\ge0} ( P m , V m ) m ≥ 0 is bounded, then 1 / x ∈ D h 1/x\in D_h 1/ x ∈ D h ; more precisely, the admissible series 1 − ∑ k ≥ 1 d k w k 1-\sum_{k\ge1}d_kw^k 1 − ∑ k ≥ 1 d k w k has a double zero at w = 1 / x w=1/x w = 1/ x .
Proof. Unrolling (S1.8) gives
x − m P m = 1 − ∑ k = 1 m d k x − k , x − m ( V m − ( m + 1 ) P m ) = ∑ k = 1 m k d k x − k . x^{-m}P_m=1-\sum_{k=1}^{m}d_kx^{-k},\qquad x^{-m}\bigl(V_m-(m+1)P_m\bigr)=\sum_{k=1}^{m}kd_kx^{-k}. x − m P m = 1 − k = 1 ∑ m d k x − k , x − m ( V m − ( m + 1 ) P m ) = k = 1 ∑ m k d k x − k .
Boundedness makes both left-hand sides tend to zero. Thus the admissible series f ( w ) = 1 − ∑ k ≥ 1 d k w k f(w)=1-\sum_{k\ge1}d_kw^k f ( w ) = 1 − ∑ k ≥ 1 d k w k satisfies f ( 1 / x ) = f ′ ( 1 / x ) = 0 f(1/x)=f'(1/x)=0 f ( 1/ x ) = f ′ ( 1/ x ) = 0 . □
The all-scale arguments below compare a finite-word state with its limit at an endpoint. The following estimate applies to each choice of trace coordinates and keeps this propagation separate from the estimates for the initial states.
Lemma S1.8 (Perturbation along a finite word). Fix a real square matrix J J J and a vector e e e of matching dimension, and let T x , d Z = x J Z − d e T_{x,d}Z=xJZ-de T x , d Z = x J Z − d e . For a word w = ( d 1 , … , d n ) w=(d_1,\ldots,d_n) w = ( d 1 , … , d n ) , n ≥ 1 n\ge1 n ≥ 1 , put T x , w = T x , d n ∘ ⋯ ∘ T x , d 1 T_{x,w}=T_{x,d_n}\circ\cdots\circ T_{x,d_1} T x , w = T x , d n ∘ ⋯ ∘ T x , d 1 . Suppose 0 < x ≤ q 0<x\le q 0 < x ≤ q , ∣ d i ∣ ≤ h |d_i|\le h ∣ d i ∣ ≤ h , δ = q − x \delta=q-x δ = q − x , ∥ Z − Z 0 ∥ ∞ ≤ E \lVert Z-Z^0\rVert_\infty\le E ∥ Z − Z 0 ∥ ∞ ≤ E and ∥ Z 0 ∥ ∞ ≤ G \lVert Z^0\rVert_\infty\le G ∥ Z 0 ∥ ∞ ≤ G . Write C j = ∥ J j ∥ ∞ C_j=\lVert J^j\rVert_\infty C j = ∥ J j ∥ ∞ for the induced matrix norm and B j = ∥ J j e ∥ ∞ B_j=\lVert J^je\rVert_\infty B j = ∥ J j e ∥ ∞ . Then
T x , w Z = x n J n Z − ∑ i = 1 n d i x n − i J n − i e , (46) T_{x,w}Z=x^nJ^nZ-\sum_{i=1}^{n}d_ix^{n-i}J^{n-i}e, \tag*{(46)} T x , w Z = x n J n Z − i = 1 ∑ n d i x n − i J n − i e , (46) and
∥ T x , w Z − T q , w Z 0 ∥ ∞ ≤ x n C n E + δ ( n q n − 1 C n G + h ∑ j = 1 n − 1 j q j − 1 B j ) . (47) \lVert T_{x,w}Z-T_{q,w}Z^0\rVert_\infty\le x^nC_nE+\delta\left(nq^{n-1}C_nG+h\sum_{j=1}^{n-1}jq^{j-1}B_j\right). \tag*{(47)} ∥ T x , w Z − T q , w Z 0 ∥ ∞ ≤ x n C n E + δ ( n q n − 1 C n G + h j = 1 ∑ n − 1 j q j − 1 B j ) . (47) Alternatively, put Z 0 0 = Z 0 Z^0_0=Z^0 Z 0 0 = Z 0 and Z i 0 = T q , ( d 1 , … , d i ) Z 0 Z^0_i=T_{q,(d_1,\ldots,d_i)}Z^0 Z i 0 = T q , ( d 1 , … , d i ) Z 0 for 1 ≤ i < n 1\le i<n 1 ≤ i < n . If ∥ J Z i 0 ∥ ∞ ≤ D i \lVert JZ^0_i\rVert_\infty\le D_i ∥ J Z i 0 ∥ ∞ ≤ D i for 0 ≤ i < n 0\le i<n 0 ≤ i < n , then
∥ T x , w Z − T q , w Z 0 ∥ ∞ ≤ x n C n E + δ ∑ i = 0 n − 1 x n − 1 − i C n − 1 − i D i . (48) \lVert T_{x,w}Z-T_{q,w}Z^0\rVert_\infty\le x^nC_nE+\delta\sum_{i=0}^{n-1}x^{n-1-i}C_{n-1-i}D_i. \tag*{(48)} ∥ T x , w Z − T q , w Z 0 ∥ ∞ ≤ x n C n E + δ i = 0 ∑ n − 1 x n − 1 − i C n − 1 − i D i . (48)
Proof. Iteration gives (S1.9). Subtract its version at ( q , Z 0 ) (q,Z^0) ( q , Z 0 ) and use ∣ q j − x j ∣ ≤ j q j − 1 δ |q^j-x^j|\le jq^{j-1}\delta ∣ q j − x j ∣ ≤ j q j − 1 δ to obtain (S1.10). For the alternative, let E i E_i E i be the difference of the two states after i i i digits. Then E i + 1 = x J E i − δ J Z i 0 E_{i+1}=xJE_i-\delta JZ^0_i E i + 1 = x J E i − δ J Z i 0 , so iteration and the induced norm give (S1.11). □
S2. Complete real traces at pronic heights
The retraction argument in Section 3 reduces connectedness at a pronic height to a complete active real trace. It suffices to show that C h e x t ∩ ( 1 , ∞ ) C_h^{\mathrm{ext}}\cap(1,\infty) C h ext ∩ ( 1 , ∞ ) is a single interval ( 1 , a ] (1,a] ( 1 , a ] , that is, that every point of [ 1 / a , 1 ) [1/a,1) [ 1/ a , 1 ) is active and no smaller point is. For q ≥ 3 q\ge3 q ≥ 3 , Lemma 3.1 confines the double zeros, hence the active points, to [ 1 / q , 1 ) [1/q,1) [ 1/ q , 1 ) , and an interval of double zeros is active (Lemma 2.7); so it suffices that every point of [ 1 / q , 1 ) [1/q,1) [ 1/ q , 1 ) is a double zero. This section proves this at h = q ( q − 1 ) h=q(q-1) h = q ( q − 1 ) , q ≥ 3 q\ge3 q ≥ 3 , and the corresponding statement with a = ρ a=\rho a = ρ at q = 2 q=2 q = 2 . In the bulk, two digits suffice. Near the endpoint the orbits we construct follow two parabolic channels, and no word of bounded length works there; the terminal window is completed in the all-scale proofs below. At height two, a comparison with two explicit witnesses identifies the endpoint as 1 / ρ 1/\rho 1/ ρ .
S2.1. Bulk capture
In Sections S2.1 to S2.3, q ≥ 3 q\ge3 q ≥ 3 is an integer, h = q ( q − 1 ) h=q(q-1) h = q ( q − 1 ) and N = 2 h + 1 N=2h+1 N = 2 h + 1 , and δ = q − x \delta=q-x δ = q − x for a real parameter x x x . By Proposition 2.3, every x ∈ ( 1 , h + 1 ) x\in(1,\sqrt{h+1}) x ∈ ( 1 , h + 1 ) is an interior point of Z h e x t \mathcal{Z}^{\mathrm{ext}}_h Z h ext . One of q q q two-digit words extends this to q − 1 / q q-1/q q − 1/ q .
Theorem S2.1 (Interior below q − 1 / q q-1/q q − 1/ q ). Let q ≥ 3 q\ge3 q ≥ 3 be an integer and h = q ( q − 1 ) h=q(q-1) h = q ( q − 1 ) . Every x ∈ [ h + 1 , q − 1 / q ] x\in[\sqrt{h+1},q-1/q] x ∈ [ h + 1 , q − 1/ q ] is an interior point of Z h e x t \mathcal{Z}^{\mathrm{ext}}_h Z h ext .
Proof. Idea. The first digit 2 q − 1 2q-1 2 q − 1 cancels the growth of the empty word; the second, chosen among q q q consecutive values, centres the value coordinate.
Since ( q − 1 2 ) 2 = h + 1 4 < h + 1 (q-\frac12)^2=h+\frac14<h+1 ( q − 2 1 ) 2 = h + 4 1 < h + 1 , we have 0 < δ < 1 / 2 0<\delta<1/2 0 < δ < 1/2 , and q δ ≥ 1 q\delta\ge1 q δ ≥ 1 . For the word u r = ( 2 q − 1 , − h + r ) u_r=(2q-1,-h+r) u r = ( 2 q − 1 , − h + r ) , two steps of (S1.3) give
V u r = θ − r , S u r = 2 ( θ − r ) − g , V_{u_r}=\theta-r,\qquad S_{u_r}=2(\theta-r)-g, V u r = θ − r , S u r = 2 ( θ − r ) − g , where
θ = q − 2 ( q + 1 ) δ + 3 δ 2 = ( x 2 − ( h + 1 ) ) + 2 ( x − q + 1 2 ) 2 + 1 2 , g = x ( 2 x − 2 q + 1 ) = ( q − δ ) ( 1 − 2 δ ) . \theta=q-2(q+1)\delta+3\delta^2=(x^2-(h+1))+2\left(x-q+\frac12\right)^2+\frac12,\qquad g=x(2x-2q+1)=(q-\delta)(1-2\delta). θ = q − 2 ( q + 1 ) δ + 3 δ 2 = ( x 2 − ( h + 1 )) + 2 ( x − q + 2 1 ) 2 + 2 1 , g = x ( 2 x − 2 q + 1 ) = ( q − δ ) ( 1 − 2 δ ) . The second expression for θ \theta θ gives θ ≥ 1 / 2 \theta\ge1/2 θ ≥ 1/2 , and g > 0 g>0 g > 0 . Since θ − q + 4 q + 1 2 δ = 3 δ ( δ − 1 2 ) < 0 \theta-q+\frac{4q+1}{2}\delta=3\delta(\delta-\frac12)<0 θ − q + 2 4 q + 1 δ = 3 δ ( δ − 2 1 ) < 0 , g = q − 2 q δ − δ ( 1 − 2 δ ) g=q-2q\delta-\delta(1-2\delta) g = q − 2 q δ − δ ( 1 − 2 δ ) and q δ ≥ 1 q\delta\ge1 q δ ≥ 1 ,
θ < q − 4 q + 1 2 δ ≤ q − 4 q + 1 2 q = q − 2 − 1 2 q , g < q − 2 q δ ≤ q − 2. \theta<q-\frac{4q+1}{2}\delta\le q-\frac{4q+1}{2q}=q-2-\frac{1}{2q},\qquad g<q-2q\delta\le q-2. θ < q − 2 4 q + 1 δ ≤ q − 2 q 4 q + 1 = q − 2 − 2 q 1 , g < q − 2 q δ ≤ q − 2.
Choose an integer 0 ≤ r ≤ q − 1 0 \le r \le q-1 0 ≤ r ≤ q − 1 with ∣ θ − r ∣ ≤ 1 / 2 |\theta-r| \le1/2 ∣ θ − r ∣ ≤ 1/2 . Both digits of u r u_r u r lie in B h B_h B h , because 2 q − 1 ≤ q ( q − 1 ) 2q-1 \le q(q-1) 2 q − 1 ≤ q ( q − 1 ) for q ≥ 3 q \ge3 q ≥ 3 , and
∣ V u r ∣ ≤ 1 2 , ∣ S u r ∣ ≤ 1 + q < q − 1. |V_{u_r}| \le\frac{1}{2}, \qquad|S_{u_r}| \le1+q<q-1. ∣ V u r ∣ ≤ 2 1 , ∣ S u r ∣ ≤ 1 + q < q − 1.
Finally,
N − x 2 − 2 x = ( q − 1 ) ( q − 3 ) + ( 2 q δ − 2 ) + δ ( 2 − δ ) > 0 , N-x^2-2x=(q-1)(q-3)+(2q\delta-2)+\delta(2-\delta)>0, N − x 2 − 2 x = ( q − 1 ) ( q − 3 ) + ( 2 q δ − 2 ) + δ ( 2 − δ ) > 0 ,
so τ h ( x ) > 1 / 2 \tau_h(x)>1/2 τ h ( x ) > 1/2 by (S1.5), while σ h ( x ) = N / ( 2 x ) > N / ( 2 q ) = q − 1 + 1 / ( 2 q ) \sigma_h(x)=N/(2x)>N/(2q)=q-1+1/(2q) σ h ( x ) = N / ( 2 x ) > N / ( 2 q ) = q − 1 + 1/ ( 2 q ) . The real case of Proposition S1.4 applies. □
Two parabolic channels
For q ≥ 4 q\ge4 q ≥ 4 the strip q − 1 / q < x < q q-1/q<x<q q − 1/ q < x < q remains. There x 2 + 2 x < N x^2+2x<N x 2 + 2 x < N , so the interior criterion still applies, but the words it needs become long as x → q x\to q x → q . To see why, use the real-jet coordinates ( P , V ) (P,V) ( P , V ) of (S1.8), write F x , d ( P , V ) = ( x P − d , x V + x P − d ) F_{x,d}(P,V)=(xP-d,xV+xP-d) F x , d ( P , V ) = ( x P − d , x V + x P − d ) for the action of a digit d d d , and put r = x − 1 r=x-1 r = x − 1 . The map F x , d F_{x,d} F x , d fixes z d = ( d / r , − d / r 2 ) z_d=(d/r,-d/r^2) z d = ( d / r , − d / r 2 ) , and repeating the digit acts on deviations from z d z_d z d by a Jordan block:
F x , d ℓ ( z d + ( ξ , ζ ) ) = z d + x ℓ ( ξ , ζ + ℓ ξ ) ( ℓ ≥ 0 ) . (49) F_{x,d}^{\ell}\bigl(z_d+(\xi,\zeta)\bigr)=z_d+x^{\ell}(\xi,\zeta+\ell\xi)\qquad(\ell\ge0). \tag*{(49)} F x , d ℓ ( z d + ( ξ , ζ ) ) = z d + x ℓ ( ξ , ζ + ℓ ξ ) ( ℓ ≥ 0 ) . (49)
Indeed, one step gives z d + x ( ξ , ζ + ξ ) z_d+x(\xi,\zeta+\xi) z d + x ( ξ , ζ + ξ ) , because − x + x r − r 2 = − 1 -x+xr-r^2=-1 − x + x r − r 2 = − 1 . We call the resulting channels parabolic because the linear part is x x x times a unipotent, that is parabolic, matrix; the fixed points themselves are repelling, with multiplier x > 1 x>1 x > 1 . Put
a = q − 1 , b = ( q − 1 ) ( q − 2 ) . a=q-1,\qquad b=(q-1)(q-2). a = q − 1 , b = ( q − 1 ) ( q − 2 ) .
At x = q x=q x = q the digits − a 2 -a^2 − a 2 and − b -b − b have the fixed points
z A = ( − a , 1 ) , z B = ( − ( q − 2 ) , q − 2 q − 1 ) . z_A=(-a,1),\qquad z_B=\left(-(q-2),\frac{q-2}{q-1}\right). z A = ( − a , 1 ) , z B = ( − ( q − 2 ) , q − 1 q − 2 ) .
Both lie just above the trace rectangle. Their value coordinates exceed τ h ( q ) = 1 − 2 / q + 1 / ( 2 q 2 ) \tau_h(q)=1-2/q+1/(2q^2) τ h ( q ) = 1 − 2/ q + 1/ ( 2 q 2 ) by ( 4 q − 1 ) / ( 2 q 2 ) (4q-1)/(2q^2) ( 4 q − 1 ) / ( 2 q 2 ) and ( 2 q 2 − 5 q + 1 ) / ( 2 q 2 ( q − 1 ) ) (2q^2-5q+1)/(2q^2(q-1)) ( 2 q 2 − 5 q + 1 ) / ( 2 q 2 ( q − 1 )) , while their coordinates S = P + V S=P+V S = P + V , namely 2 − q 2-q 2 − q and − ( q − 2 ) 2 / ( q − 1 ) -(q-2)^2/(q-1) − ( q − 2 ) 2 / ( q − 1 ) , lie inside ∣ S ∣ < σ h ( q ) = q − 1 + 1 / ( 2 q ) |S|<\sigma_h(q)=q-1+1/(2q) ∣ S ∣ < σ h ( q ) = q − 1 + 1/ ( 2 q ) . The two channels are joined: the digit − ( a 2 + 1 ) -(a^2+1) − ( a 2 + 1 ) maps the point z A − ( 0 , 1 / a ) z_A-(0,1/a) z A − ( 0 , 1/ a ) exactly onto z B z_B z B , because
F q , − ( a 2 + 1 ) ( − a , 1 − 1 a ) = ( − q a + a 2 + 1 , q − q a − q a + a 2 + 1 ) = ( 2 − q , 1 − 1 a ) F_{q,-(a^2+1)}\left(-a,1-\frac{1}{a}\right)=\left(-qa+a^2+1,q-\frac{q}{a}-qa+a^2+1\right)=\left(2-q,1-\frac{1}{a}\right) F q , − ( a 2 + 1 ) ( − a , 1 − a 1 ) = ( − q a + a 2 + 1 , q − a q − q a + a 2 + 1 ) = ( 2 − q , 1 − a 1 )
when q = a + 1 q=a+1 q = a + 1 .
The words that follow the channels are the parabolic blocks
A m = ( 2 q − 1 , ( − a 2 ) m − 1 ) , B m = ( 2 q − 1 , ( − a 2 ) m − 2 , − ( a 2 + 1 ) ) ( m ≥ 2 ) , A_m=(2q-1,(-a^2)^{m-1}),\qquad B_m=(2q-1,(-a^2)^{m-2},-(a^2+1))\qquad(m\ge2), A m = ( 2 q − 1 , ( − a 2 ) m − 1 ) , B m = ( 2 q − 1 , ( − a 2 ) m − 2 , − ( a 2 + 1 )) ( m ≥ 2 ) ,
where ( d ) k (d)^k ( d ) k denotes k k k repetitions of the digit d d d (words are written in sans-serif), and their extensions
U m , ℓ = B m ( − b ) ℓ ( ℓ ≥ 0 ) . U_{m,\ell}=B_m(-b)^\ell\qquad(\ell\ge0). U m , ℓ = B m ( − b ) ℓ ( ℓ ≥ 0 ) .
The first digit sends ( 1 , 1 ) (1,1) ( 1 , 1 ) to a state whose P P P -coordinate ( δ 2 − a 2 ) / r (\delta^2-a^2)/r ( δ 2 − a 2 ) / r exceeds that of z − a 2 = ( − a 2 / r , a 2 / r 2 ) z_{-a^2}=(-a^2/r,a^2/r^2) z − a 2 = ( − a 2 / r , a 2 / r 2 ) by δ 2 / r \delta^2/r δ 2 / r . By (S2.1), each repetition of − a 2 -a^2 − a 2 multiplies this deviation by x x x , so the orbit stays near the first channel for about m m m steps, where δ x m \delta x^m δ x m is of order 1. The digit − ( a 2 + 1 ) -(a^2+1) − ( a 2 + 1 ) then moves it to the second channel, along which (S2.1) again governs the deviation from z − b z_{-b} z − b . The trace coordinates of the whole family are explicit. Put
V b ♯ ( x ) : = b ( x − 1 ) 2 , S b ♯ ( x ) : = ( 2 − x ) b ( x − 1 ) 2 ; V_b^\sharp(x):=\frac{b}{(x-1)^2},\qquad S_b^\sharp(x):=\frac{(2-x)b}{(x-1)^2}; V b ♯ ( x ) := ( x − 1 ) 2 b , S b ♯ ( x ) := ( x − 1 ) 2 ( 2 − x ) b ;
these are the coordinates ( V , S ) (V,S) ( V , S ) of z − b z_{-b} z − b , the fixed state of the digit − b -b − b in (S1.3). Set
u m ( x ) : = V B m ( x ) − V b ♯ ( x ) , (50) u_m(x):=V_{B_m}(x)-V_b^\sharp(x), \tag*{(50)} u m ( x ) := V B m ( x ) − V b ♯ ( x ) , (50)
c m ( x ) : = ( S B m ( x ) − S b ♯ ( x ) ) − ( V B m ( x ) − V b ♯ ( x ) ) . (51) c_m(x):=\bigl(S_{B_m}(x)-S_b^\sharp(x)\bigr)-\bigl(V_{B_m}(x)-V_b^\sharp(x)\bigr). \tag*{(51)} c m ( x ) := ( S B m ( x ) − S b ♯ ( x ) ) − ( V B m ( x ) − V b ♯ ( x ) ) . (51)
Then
c m ( x ) = δ ( δ x m − 1 − 1 ) x − 1 . (52) c_m(x)=\frac{\delta(\delta x^{m-1}-1)}{x-1}. \tag*{(52)} c m ( x ) = x − 1 δ ( δ x m − 1 − 1 ) . (52)
Proposition S2.2 (Parabolic return formulae). For every m ≥ 2 m \ge2 m ≥ 2 and ℓ ≥ 0 \ell\ge0 ℓ ≥ 0 ,
V U m , ℓ ( x ) = V b ♯ ( x ) + x ℓ ( u m ( x ) + ℓ c m ( x ) ) , (53) V_{U_{m,\ell}}(x)=V_b^\sharp(x)+x^\ell\left(u_m(x)+\ell c_m(x)\right), \tag*{(53)} V U m , ℓ ( x ) = V b ♯ ( x ) + x ℓ ( u m ( x ) + ℓ c m ( x ) ) , (53) S U m , ℓ ( x ) = S b ♯ ( x ) + x ℓ ( u m ( x ) + ( ℓ + 1 ) c m ( x ) ) . (54) S_{U_{m,\ell}}(x)=S_b^\sharp(x)+x^\ell\left(u_m(x)+(\ell+1)c_m(x)\right). \tag*{(54)} S U m , ℓ ( x ) = S b ♯ ( x ) + x ℓ ( u m ( x ) + ( ℓ + 1 ) c m ( x ) ) . (54)
Proof. Centred at ( V b ♯ , S b ♯ ) (V_b^\sharp,S_b^\sharp) ( V b ♯ , S b ♯ ) , appending − b -b − b acts by x J xJ x J , where
J = ( 0 1 − 1 2 ) , J ℓ = I + ℓ ( J − I ) , J=\begin{pmatrix}0&1\\-1&2\end{pmatrix},\qquad J^\ell=I+\ell(J-I), J = ( 0 − 1 1 2 ) , J ℓ = I + ℓ ( J − I ) ,
because ( J − I ) 2 = 0 (J-I)^2=0 ( J − I ) 2 = 0 . Apply this to the centred state ( u m , u m + c m ) (u_m,u_m+c_m) ( u m , u m + c m ) of B m B_m B m . For (S2.4), note that c m = P B m ( x ) + b / r c_m=P_{B_m}(x)+b/r c m = P B m ( x ) + b / r , since S u − V u = P u S_u-V_u=P_u S u − V u = P u by (S1.2). Along B m B_m B m , the deviation δ 2 / r \delta^2/r δ 2 / r from − a 2 / r -a^2/r − a 2 / r is multiplied by x x x at each of the m − 2 m-2 m − 2 digits − a 2 -a^2 − a 2 , and the final digit − ( a 2 + 1 ) -(a^2+1) − ( a 2 + 1 ) adds 1. Hence
c m = − a 2 + x m − 1 δ 2 r + 1 + b r = x m − 1 δ 2 − δ r , c_m=\frac{-a^2+x^{m-1}\delta^2}{r}+1+\frac{b}{r}=\frac{x^{m-1}\delta^2-\delta}{r}, c m = r − a 2 + x m − 1 δ 2 + 1 + r b = r x m − 1 δ 2 − δ ,
because b − a 2 + r = − ( q − 1 ) + x − 1 = − δ b-a^2+r=-(q-1)+x-1=-\delta b − a 2 + r = − ( q − 1 ) + x − 1 = − δ .
Completion at heights q ( q − 1 ) q(q-1) q ( q − 1 )
For q ≥ 4 q\ge4 q ≥ 4 the strip ( q − 1 / q , q ) (q-1/q,q) ( q − 1/ q , q ) splits into intervals captured by six short words, two compact windows and a terminal window. The three statements below are proved in Section S2.5; each uses finitely many exact checks.
Theorem S2.3 (Refinement at scale q − 3 q^{-3} q − 3 ). Let q ≥ 4 q\ge4 q ≥ 4 and h = q ( q − 1 ) h=q(q-1) h = q ( q − 1 ) . Every point of ( 1 , q ) ∖ R q ( 3 ) (1,q)\setminus R_q^{(3)} ( 1 , q ) ∖ R q ( 3 ) is an interior point of Z h e x t \mathcal{Z}_h^{\mathrm{ext}} Z h ext , where
R q ( 3 ) = ( q − 1 q 3 , q ) ∪ ( q − 1 2 q 2 − 1 2 q 3 , q − 1 2 q 2 + 1 3 q 3 ) R_q^{(3)}=\left(q-\frac{1}{q^3},q\right)\cup\left(q-\frac{1}{2q^2}-\frac{1}{2q^3},q-\frac{1}{2q^2}+\frac{1}{3q^3}\right) R q ( 3 ) = ( q − q 3 1 , q ) ∪ ( q − 2 q 2 1 − 2 q 3 1 , q − 2 q 2 1 + 3 q 3 1 ) ∪ ( q − 1 2 q − 3 4 q 3 , q − 1 2 q + 3 4 q 3 ) . {}\cup\left(q-\frac{1}{2q}-\frac{3}{4q^3},q-\frac{1}{2q}+\frac{3}{4q^3}\right). ∪ ( q − 2 q 1 − 4 q 3 3 , q − 2 q 1 + 4 q 3 3 ) . In particular,
∣ R q ( 3 ) ∣ = 10 3 q 3 . \lvert R_q^{(3)}\rvert=\frac{10}{3q^3}. ∣ R q ( 3 ) ∣ = 3 q 3 10 .
Proposition S2.4 (The compact windows). Let q ≥ 4 q\ge4 q ≥ 4 and h = q ( q − 1 ) h=q(q-1) h = q ( q − 1 ) , and put
J q , 2 = [ q − 1 2 q − 3 4 q 3 , q − 1 2 q + 3 4 q 3 ] , J_{q,2}=\left[q-\frac{1}{2q}-\frac{3}{4q^3},q-\frac{1}{2q}+\frac{3}{4q^3}\right], J q , 2 = [ q − 2 q 1 − 4 q 3 3 , q − 2 q 1 + 4 q 3 3 ] , J q , 3 = [ q − 1 2 q 2 − 1 2 q 3 , q − 1 2 q 2 + 1 3 q 3 ] . J_{q,3}=\left[q-\frac{1}{2q^2}-\frac{1}{2q^3},q-\frac{1}{2q^2}+\frac{1}{3q^3}\right]. J q , 3 = [ q − 2 q 2 1 − 2 q 3 1 , q − 2 q 2 1 + 3 q 3 1 ] . Every point of J q , 2 ∪ J q , 3 J_{q,2}\cup J_{q,3} J q , 2 ∪ J q , 3 is an interior point of Z h e x t \mathcal{Z}_h^{\mathrm{ext}} Z h ext .
Theorem S2.5 (Terminal window). For every integer q ≥ 4 q\ge4 q ≥ 4 and h = q ( q − 1 ) h=q(q-1) h = q ( q − 1 ) , every point of
[ q − 1 q 3 , q ) \left[q-\frac{1}{q^3},q\right) [ q − q 3 1 , q ) is an interior point of Z h e x t \mathcal{Z}_h^{\mathrm{ext}} Z h ext .
The short words are A2 _2 2 , B2 _2 2 , A3 _3 3 , B3 _3 3 , A2 ( − h ) a n d B _2(-h) and B 2 ( − h ) an d B _2(-b)$; a single coefficient test proves all their bounds. In the terminal window, the parabolic blocks supply words of unbounded length. The window is cut into the bands B q , m = { x : ( q − x ) x m − 1 ≤ 1 2 ≤ ( q − x ) x m } B_{q,m}=\{x:(q-x)x^{m-1}\le\frac{1}{2}\le(q-x)x^m\} B q , m = { x : ( q − x ) x m − 1 ≤ 2 1 ≤ ( q − x ) x m } , which accumulate at q q q . On a band, most phases are captured by two families of overlapping layers, B m ( − b ) ℓ ( − ( b − 1 ) ) B_m(-b)^\ell(-(b-1)) B m ( − b ) ℓ ( − ( b − 1 )) and B m + 1 ( − b ) ℓ B_{m+1}(-b)^\ell B m + 1 ( − b ) ℓ , whose trace coordinates are given exactly by Proposition S2.2. In a central window of phases the capturing words leave the second channel and follow the fixed state of the top digit h h h before returning; there are about q / 4 q/4 q /4 of them, and their final digits depend on q q q modulo 4. The limiting overlaps are rational in q q q and are checked symbolically on residue classes of q q q modulo 4; explicit error bounds carry them to finite x x x , except for finitely many pairs ( q , m ) (q,m) ( q , m ) , which are covered by exact rational cells.
One way to see why no gap is expected near the endpoint is that the witness of the endpoint itself, ( 1 − q w ) 2 / ( 1 − w ) = 1 − ( 2 q − 1 ) w + a 2 ∑ j ≥ 2 w j (1-qw)^2/(1-w)=1-(2q-1)w+a^2\sum_{j\ge2}w^j ( 1 − q w ) 2 / ( 1 − w ) = 1 − ( 2 q − 1 ) w + a 2 ∑ j ≥ 2 w j , has the digits 2 q − 1 , − a 2 , − a 2 , … 2q-1,-a^2,-a^2,\ldots 2 q − 1 , − a 2 , − a 2 , … : it is the first channel. Orbits for x x x near q q q shadow it and then correct with the digits − a 2 − k -a^2-k − a 2 − k , $0\le k\le a, which lie in B h B_h B h exactly because h − a 2 = a h-a^2=a h − a 2 = a . At h = 2 h=2 h = 2 the analogous series ( 1 − 2 w ) 2 / ( 1 − w ) (1-2w)^2/(1-w) ( 1 − 2 w ) 2 / ( 1 − w ) has the coefficient − 3 -3 − 3 , outside the alphabet, and indeed a gap ( 1 / 2 , 1 / ρ ) (1/2,1/\rho) ( 1/2 , 1/ ρ ) appears (Section S2.4).
At q = 3 q=3 q = 3 the trace rectangle does not reach the endpoint: x 2 + 2 x < 13 x^2+2x<13 x 2 + 2 x < 13 only for x < 14 − 1 < 2.75 x<\sqrt{14}-1<2.75 x < 14 − 1 < 2.75 . The height-six completion below therefore uses the bounded-orbit test of Lemma S1.7 instead, in three steps, for 2999 / 1000 ≤ x < 3 2999/1000\le x<3 2999/1000 ≤ x < 3 . First, entry: a finite word brings ( 1 , 1 ) (1,1) ( 1 , 1 ) into the rectangle R x = [ − 3 / r , 3 / r ] × [ − 2 / r 2 , 2 / r 2 ] R_x=[-3/r,3/r]\times[-2/r^2,2/r^2] R x = [ − 3/ r , 3/ r ] × [ − 2/ r 2 , 2/ r 2 ] , with r = x − 1 r=x-1 r = x − 1 , which moves with x x x ; the word is selected by the first crossing ( 3 − x ) x M − 1 ≥ 1 / 5 (3-x)x^{M-1}\ge1/5 ( 3 − x ) x M − 1 ≥ 1/5 (Theorem S2.25). Second, renewal: from every point of R x R_x R x a finite word returns to R x R_x R x ; when a one-step return fails, two adjacent digit choices remain, whose paired excursions either return or enter a fixed compact cone, and all intermediate states stay in one bounded set (Theorem S2.23). Third, concatenating the returns gives an infinite bounded orbit, hence a double zero at 1 / x 1/x 1/ x by Lemma S1.7. On ( 1 , 2999 / 1000 ] (1,2999/1000] ( 1 , 2999/1000 ] a finite buffered cover gives interior points instead (Theorem S2.22).
Theorem S2.6 (Pronic trace completion). Let q ≥ 3 q\ge3 q ≥ 3 be an integer and h = q ( q − 1 ) h=q(q-1) h = q ( q − 1 ) . Then
D h = A h = [ 1 / q , 1 ) , C h e x t ∩ ( 1 , ∞ ) = ( 1 , q ] . D_h=A_h=[1/q,1),\qquad C_h^{\mathrm{ext}}\cap(1,\infty)=(1,q]. D h = A h = [ 1/ q , 1 ) , C h ext ∩ ( 1 , ∞ ) = ( 1 , q ] . Proof. By Lemma 3.1, D h ⊆ [ 1 / q , 1 ) D_h\subseteq[1/q,1) D h ⊆ [ 1/ q , 1 ) . Let 1 < x < q 1<x<q 1 < x < q . If q ≥ 4 q\ge4 q ≥ 4 , then x x x is an interior point of Z h e x t Z_h^{\mathrm{ext}} Z h ext : by Proposition 2.3 if x < h + 1 x<\sqrt{h+1} x < h + 1 , by Theorem S2.1 if h + 1 ≤ x ≤ q − 1 / q \sqrt{h+1}\le x\le q-1/q h + 1 ≤ x ≤ q − 1/ q , and otherwise by Theorems S2.3 and S2.5 and Proposition S2.4, whose sets cover ( q − 1 / q , q ) (q-1/q,q) ( q − 1/ q , q ) . Then 1 / x ∈ D h 1/x\in D_h 1/ x ∈ D h by Corollary S1.6. If q = 3 q=3 q = 3 , then x x x is an interior point for x ≤ 2999 / 1000 x\le2999/1000 x ≤ 2999/1000 (Theorem S2.22), so again 1 / x ∈ D 6 1/x\in D_6 1/ x ∈ D 6 , and 1 / x ∈ D 6 1/x\in D_6 1/ x ∈ D 6 for 2999 / 1000 ≤ x < 3 2999/1000\le x<3 2999/1000 ≤ x < 3 by Theorem S2.25. Thus ( 1 / q , 1 ) ⊆ D h (1/q,1)\subseteq D_h ( 1/ q , 1 ) ⊆ D h , and 1 / q ∈ D h 1/q\in D_h 1/ q ∈ D h by Lemma S1.1, so D h = [ 1 / q , 1 ) D_h=[1/q,1) D h = [ 1/ q , 1 ) . By Lemma 2.7 the whole interval, endpoint included, is active, so A h = D h A_h=D_h A h = D h . Finally, A h = { t ∈ ( 0 , 1 ) : 1 / t ∈ C h } A_h=\{t\in(0,1):1/t\in C_h\} A h = { t ∈ ( 0 , 1 ) : 1/ t ∈ C h } gives the positive trace. □ \square □
Height-two exclusion
At h = 2 h=2 h = 2 the pronic endpoint q = 2 q=2 q = 2 is not attained. The attachment point is the tribonacci constant ρ = 1.8392 … \rho=1.8392\ldots ρ = 1.8392 … , the real root of ρ 3 = ρ 2 + ρ + 1 \rho^3=\rho^2+\rho+1 ρ 3 = ρ 2 + ρ + 1 . Two admissible series explain the answer. Put G ( w ) = 1 − w − w 2 − w 3 G(w)=1-w-w^2-w^3 G ( w ) = 1 − w − w 2 − w 3 , so that G ( 1 / ρ ) = 0 G(1/\rho)=0 G ( 1/ ρ ) = 0 . The series
f ρ ( w ) = G ( w ) 2 1 − w 3 = 1 − 2 w − w 2 + w 3 + w 4 + w 5 + 2 w 6 1 − w 3 (55) f_\rho(w)=\frac{G(w)^2}{1-w^3}=1-2w-w^2+w^3+\frac{w^4+w^5+2w^6}{1-w^3} \tag*{(55)} f ρ ( w ) = 1 − w 3 G ( w ) 2 = 1 − 2 w − w 2 + w 3 + 1 − w 3 w 4 + w 5 + 2 w 6 (55)
has coefficients 1 , − 2 , − 1 , 1 1,-2,-1,1 1 , − 2 , − 1 , 1 followed by the periodic block 1 , 1 , 2 1,1,2 1 , 1 , 2 , so it belongs to F 2 \mathcal{F}_2 F 2 , and it has a double zero at 1 / ρ 1/\rho 1/ ρ . The series
f 1 / 2 ( w ) = ( 1 − 2 w ) 2 ( 1 + w ) 1 − w = 1 − 2 w − 2 w 2 + 2 w 3 1 − w f_{1/2}(w)=\frac{(1-2w)^2(1+w)}{1-w}=1-2w-2w^2+\frac{2w^3}{1-w} f 1/2 ( w ) = 1 − w ( 1 − 2 w ) 2 ( 1 + w ) = 1 − 2 w − 2 w 2 + 1 − w 2 w 3
belongs to F 2 \mathcal{F}_2 F 2 and has a double zero at 1 / 2 1/2 1/2 , with f 1 / 2 ′ ′ ( 1 / 2 ) = 24 f''_{1/2}(1/2)=24 f 1/2 ′′ ( 1/2 ) = 24 . Shmerkin and Solomyak proved that 1 / 2 1/2 1/2 is an isolated double zero at this height and that the next one lies in ( 0.5436 , 0.5438 ) (0.5436,0.5438) ( 0.5436 , 0.5438 ) [28 Proposition 6.3] . The next theorem identifies it as 1 / ρ = 0.54368 … 1/\rho=0.54368\ldots 1/ ρ = 0.54368 … and shows that all larger parameters are double zeros.
Theorem S2.7 (Height-two trace completion). One has
D 2 = { 1 / 2 } ∪ [ 1 / ρ , 1 ) , A 2 = [ 1 / ρ , 1 ) . D_2=\{1/2\}\cup[1/\rho,1),\qquad A_2=[1/\rho,1). D 2 = { 1/2 } ∪ [ 1/ ρ , 1 ) , A 2 = [ 1/ ρ , 1 ) .
The inclusion [ 1 / ρ , 1 ) ⊆ A 2 [1/\rho,1)\subseteq A_2 [ 1/ ρ , 1 ) ⊆ A 2 is proved in Section S2.7: a return set around the periodic configuration of f ρ f_\rho f ρ is entered along its digit block. The exclusion of all other double zeros is elementary.
Lemma S2.8 (Below one half). If f ∈ F 2 f\in\mathcal{F}_2 f ∈ F 2 has a double zero λ ∈ ( 0 , 1 / 2 ] \lambda\in(0,1/2] λ ∈ ( 0 , 1/2 ] , then λ = 1 / 2 \lambda=1/2 λ = 1/2 and f = f 1 / 2 f=f_{1/2} f = f 1/2 .
Proof. Write f ( w ) = 1 + ∑ k ≥ 1 a k w k f(w)=1+\sum_{k\ge1}a_kw^k f ( w ) = 1 + ∑ k ≥ 1 a k w k . Since a 1 ≥ − 2 a_1\ge-2 a 1 ≥ − 2 , and a k ≤ 2 a_k\le2 a k ≤ 2 with 1 − k / 2 < 0 1-k/2<0 1 − k /2 < 0 for k ≥ 3 k\ge3 k ≥ 3 ,
0 = f ( λ ) − λ 2 f ′ ( λ ) = 1 + a 1 λ 2 + ∑ k ≥ 3 ( 1 − k 2 ) a k λ k ≥ 1 − λ − λ 3 ( 1 − λ ) 2 = ( 1 − 2 λ ) ( 1 − λ + λ 2 ) ( 1 − λ ) 2 . 0=f(\lambda)-\frac{\lambda}{2}f'(\lambda)=1+\frac{a_1\lambda}{2}+\sum_{k\ge3}\left(1-\frac{k}{2}\right)a_k\lambda^k\ge1-\lambda-\frac{\lambda^3}{(1-\lambda)^2}=\frac{(1-2\lambda)(1-\lambda+\lambda^2)}{(1-\lambda)^2}. 0 = f ( λ ) − 2 λ f ′ ( λ ) = 1 + 2 a 1 λ + k ≥ 3 ∑ ( 1 − 2 k ) a k λ k ≥ 1 − λ − ( 1 − λ ) 2 λ 3 = ( 1 − λ ) 2 ( 1 − 2 λ ) ( 1 − λ + λ 2 ) .
The right-hand side is positive for λ < 1 / 2 \lambda<1/2 λ < 1/2 . At λ = 1 / 2 \lambda=1/2 λ = 1/2 , equality forces a 1 = − 2 a_1=-2 a 1 = − 2 and a k = 2 a_k=2 a k = 2 for k ≥ 3 k\ge3 k ≥ 3 , and then f ( 1 / 2 ) = 0 f(1/2)=0 f ( 1/2 ) = 0 forces a 2 = − 2 a_2=-2 a 2 = − 2 . □ \square □
Lemma S2.9 (Between one half and 1 / ρ 1/\rho 1/ ρ ). No f ∈ F 2 f\in\mathcal{F}_2 f ∈ F 2 has a double zero in ( 1 / 2 , 1 / ρ ) (1/2,1/\rho) ( 1/2 , 1/ ρ ) .
Proof. Idea. Compare f f f with the two witnesses. If f f f does not begin like f 1 / 2 f_{1/2} f 1/2 , its first deviation from f ρ f_{\rho} f ρ must be negative at λ \lambda λ and increasing there, which the coefficient bounds forbid. If f f f begins like f 1 / 2 f_{1/2} f 1/2 , its first deviation from f 1 / 2 f_{1/2} f 1/2 must come so late that the value equation fails.
Fix λ ∈ ( 1 / 2 , 1 / ρ ) \lambda\in(1/2,1/\rho) λ ∈ ( 1/2 , 1/ ρ ) and f ( w ) = 1 + ∑ k ≥ 1 a k w k ∈ F 2 f(w)=1+\sum_{k\geq1}a_k w^k\in\mathbb{F}_2 f ( w ) = 1 + ∑ k ≥ 1 a k w k ∈ F 2 with f ( λ ) = f ′ ( λ ) = 0 f(\lambda)=f'(\lambda)=0 f ( λ ) = f ′ ( λ ) = 0 , and put x = 1 / λ x=1/\lambda x = 1/ λ and λ + = 5437 / 10000 \lambda_+=5437/10000 λ + = 5437/10000 . Since G G G decreases on ( 0 , 1 ) (0,1) ( 0 , 1 ) and G ( λ + ) = − 32678453 ⋅ 10 − 12 < 0 = G ( 1 / ρ ) G(\lambda_+)=-32678453\cdot10^{-12}<0=G(1/\rho) G ( λ + ) = − 32678453 ⋅ 1 0 − 12 < 0 = G ( 1/ ρ ) , we have λ < λ + \lambda<\lambda_+ λ < λ + and G ( λ ) > 0 G(\lambda)>0 G ( λ ) > 0 . We use the series
Ψ 0 = w + w 2 1 − w 3 , Ψ 1 = w + w 3 1 − w 3 , Ψ 2 = w 2 + w 3 1 − w 3 . \Psi_0=\frac{w+w^2}{1-w^3},\qquad\Psi_1=\frac{w+w^3}{1-w^3},\qquad\Psi_2=\frac{w^2+w^3}{1-w^3}. Ψ 0 = 1 − w 3 w + w 2 , Ψ 1 = 1 − w 3 w + w 3 , Ψ 2 = 1 − w 3 w 2 + w 3 .
whose coefficients lie in { 0 , 1 } \{0,1\} { 0 , 1 } , and four facts.
(i) 1 − Ψ 0 = G / ( 1 − w 3 ) 1-\Psi_0=G/(1-w^3) 1 − Ψ 0 = G / ( 1 − w 3 ) .
(ii) The coefficients e n e_n e n of f ρ f_{\rho} f ρ satisfy e n ≥ 1 e_n\geq1 e n ≥ 1 for n ≥ 3 n\geq3 n ≥ 3 , and for n ≥ 3 n\geq3 n ≥ 3 the series ∑ k ≥ 1 ( 2 − e n + k ) w k \sum_{k\geq1}(2-e_{n+k})w^k ∑ k ≥ 1 ( 2 − e n + k ) w k equals Ψ s \Psi_s Ψ s with s ≡ n ( m o d 3 ) s\equiv n\pmod3 s ≡ n ( mod 3 ) .
(iii) On ( 0 , 1 ) (0,1) ( 0 , 1 ) , Ψ s ≤ Ψ 0 \Psi_s\leq\Psi_0 Ψ s ≤ Ψ 0 and Ψ s ′ ≤ Ψ 0 ′ \Psi'_s\leq\Psi'_0 Ψ s ′ ≤ Ψ 0 ′ for s = 1 , 2 s=1,2 s = 1 , 2 , and ( Ψ 0 − w ) ′ ≤ Ψ 2 ′ (\Psi_0-w)'\leq\Psi'_2 ( Ψ 0 − w ) ′ ≤ Ψ 2 ′ and ( Ψ 1 − w ) ′ ≤ Ψ 2 ′ (\Psi_1-w)'\leq\Psi'_2 ( Ψ 1 − w ) ′ ≤ Ψ 2 ′ . Indeed, the differences
w 2 1 + w + w 2 , w ( 1 + w ) 1 + w + w 2 , w 3 1 + w + w 2 , w 2 ( 1 + w ) 1 + w + w 2 \frac{w^2}{1+w+w^2},\qquad\frac{w(1+w)}{1+w+w^2},\qquad\frac{w^3}{1+w+w^2},\qquad\frac{w^2(1+w)}{1+w+w^2} 1 + w + w 2 w 2 , 1 + w + w 2 w ( 1 + w ) , 1 + w + w 2 w 3 , 1 + w + w 2 w 2 ( 1 + w )
are nonnegative and increasing.
(iv) For 0 < λ ≤ λ + 0<\lambda\leq\lambda_+ 0 < λ ≤ λ + ,
λ 3 < 2 ( 1 − λ ) 3 , Ψ 2 ′ ( λ ) = 2 λ + 3 λ 2 + λ 4 ( 1 − λ 3 ) 2 < 3. \lambda^3<2(1-\lambda)^3,\qquad\Psi'_2(\lambda)=\frac{2\lambda+3\lambda^2+\lambda^4}{(1-\lambda^3)^2}<3. λ 3 < 2 ( 1 − λ ) 3 , Ψ 2 ′ ( λ ) = ( 1 − λ 3 ) 2 2 λ + 3 λ 2 + λ 4 < 3.
The two sides of the first are monotone in opposite directions, and Ψ 2 ′ \Psi'_2 Ψ 2 ′ increases, so it is enough to check both at λ + \lambda_+ λ + , where λ + 3 − 2 ( 1 − λ + ) 3 = − 29289174641 ⋅ 10 − 12 \lambda_+^3-2(1-\lambda_+)^3=-29289174641\cdot10^{-12} λ + 3 − 2 ( 1 − λ + ) 3 = − 29289174641 ⋅ 1 0 − 12 and Ψ 2 ′ ( λ + ) < 2.927 \Psi'_2(\lambda_+)<2.927 Ψ 2 ′ ( λ + ) < 2.927 .
Case ( a 1 , a 2 ) ≠ ( − 2 , − 2 ) (a_1,a_2)\ne(-2,-2) ( a 1 , a 2 ) = ( − 2 , − 2 ) . Since f ρ ( λ ) > 0 f_{\rho}(\lambda)>0 f ρ ( λ ) > 0 , f ≠ f ρ f\ne f_{\rho} f = f ρ . Let ν \nu ν be the first index with a ν ≠ e ν a_{\nu}\ne e_{\nu} a ν = e ν , and write f = f ρ + w ν C f=f_{\rho}+w^{\nu}C f = f ρ + w ν C with C = ∑ k ≥ 0 c k w k C=\sum_{k\geq0}c_k w^k C = ∑ k ≥ 0 c k w k and c 0 ≠ 0 c_0\ne0 c 0 = 0 . Then c k ≤ 2 − e ν + k ≤ 1 c_k\leq2-e_{\nu+k}\leq1 c k ≤ 2 − e ν + k ≤ 1 for k ≥ 2 k\geq2 k ≥ 2 . Put ϖ = x ν G ( λ ) > 0 \varpi=x^{\nu}G(\lambda)>0 ϖ = x ν G ( λ ) > 0 and ζ = − f ′ ( λ ) / G ( λ ) = 2 Ψ 0 ′ ( λ ) + 3 λ 2 G ( λ ) / ( 1 − λ 3 ) 2 \zeta=-f'(\lambda)/G(\lambda)=2\Psi'_0(\lambda)+3\lambda^2G(\lambda)/(1-\lambda^3)^2 ζ = − f ′ ( λ ) / G ( λ ) = 2 Ψ 0 ′ ( λ ) + 3 λ 2 G ( λ ) / ( 1 − λ 3 ) 2 , so that ζ > 2 Ψ 0 ′ ( λ ) \zeta>2\Psi'_0(\lambda) ζ > 2 Ψ 0 ′ ( λ ) . The equations f ( λ ) = f ′ ( λ ) = 0 f(\lambda)=f'(\lambda)=0 f ( λ ) = f ′ ( λ ) = 0 give
C ( λ ) = − ϖ G ( λ ) 1 − λ 3 < 0 , C ′ ( λ ) = x ν ( x ν f ρ ( λ ) − f ρ ′ ( λ ) ) ≥ ϖ ζ > 0. (56) C(\lambda)=-\frac{\varpi G(\lambda)}{1-\lambda^3}<0,\qquad C'(\lambda)=x^{\nu}(x\nu f_{\rho}(\lambda)-f'_{\rho}(\lambda))\geq\varpi\zeta>0. \tag*{(56)} C ( λ ) = − 1 − λ 3 ϖ G ( λ ) < 0 , C ′ ( λ ) = x ν ( xν f ρ ( λ ) − f ρ ′ ( λ )) ≥ ϖ ζ > 0. (56)
If ν = 1 \nu=1 ν = 1 , then c 0 ≥ 1 c_0\geq1 c 0 ≥ 1 and c 1 ≥ − 1 c_1\geq-1 c 1 ≥ − 1 . If ν = 2 \nu=2 ν = 2 , then c 0 = a 2 + 1 c_0=a_2+1 c 0 = a 2 + 1 , which is − 1 -1 − 1 only when a 1 = a 2 = − 2 a_1=a_2=-2 a 1 = a 2 = − 2 . If ν ≥ 3 \nu\geq3 ν ≥ 3 , then c 0 ≤ 1 c_0\leq1 c 0 ≤ 1 . So one of three cases occurs.
(A) c 0 ≤ − 1 c_0\leq-1 c 0 ≤ − 1 and ν ≥ 3 \nu\geq3 ν ≥ 3 . By facts (ii) and (iii), C ( λ ) ≤ − 1 + Ψ s ( λ ) ≤ − 1 + Ψ 0 ( λ ) C(\lambda)\leq-1+\Psi_s(\lambda)\leq-1+\Psi_0(\lambda) C ( λ ) ≤ − 1 + Ψ s ( λ ) ≤ − 1 + Ψ 0 ( λ ) and C ′ ( λ ) ≤ Ψ 0 ′ ( λ ) C'(\lambda)\leq\Psi'_0(\lambda) C ′ ( λ ) ≤ Ψ 0 ′ ( λ ) . With fact (i) and (S2.8), the first gives ϖ ≥ 1 \varpi\geq1 ϖ ≥ 1 , and the second gives 2 ϖ Ψ 0 ′ ( λ ) < ϖ ζ ≤ Ψ 0 ′ ( λ ) 2\varpi\Psi'_0(\lambda)<\varpi\zeta\leq\Psi'_0(\lambda) 2 ϖ Ψ 0 ′ ( λ ) < ϖ ζ ≤ Ψ 0 ′ ( λ ) , so ϖ < 1 / 2 \varpi<1/2 ϖ < 1/2 .
(B) c 0 ≥ 2 c_0\geq2 c 0 ≥ 2 , or c 0 = 1 c_0=1 c 0 = 1 and c 1 ≥ − 2 c_1\geq-2 c 1 ≥ − 2 . Then 2 c 0 / λ + c 1 ≥ 2 / λ − 2 2c_0/\lambda+c_1\geq2/\lambda-2 2 c 0 / λ + c 1 ≥ 2/ λ − 2 ; for c 0 ≥ 2 c_0\geq2 c 0 ≥ 2 use c 1 ≥ − 4 c_1\geq-4 c 1 ≥ − 4 and λ < 1 \lambda<1 λ < 1 . Since c k ≤ 1 c_k\leq1 c k ≤ 1 for k ≥ 2 k\geq2 k ≥ 2 and C ( λ ) < 0 C(\lambda)<0 C ( λ ) < 0 ,
C ′ ( λ ) − 2 C ( λ ) λ = ∑ k ≥ 0 ( k − 2 ) c k λ k − 1 ≤ − 2 c 0 λ − c 1 + ∑ k ≥ 2 ( k − 2 ) λ k − 1 , C'(\lambda)-\frac{2C(\lambda)}{\lambda}=\sum_{k\geq0}(k-2)c_k\lambda^{k-1}\leq-\frac{2c_0}{\lambda}-c_1+\sum_{k\geq2}(k-2)\lambda^{k-1}, C ′ ( λ ) − λ 2 C ( λ ) = k ≥ 0 ∑ ( k − 2 ) c k λ k − 1 ≤ − λ 2 c 0 − c 1 + k ≥ 2 ∑ ( k − 2 ) λ k − 1 ,
and the last sum is λ 2 / ( 1 − λ ) 2 \lambda^2/(1-\lambda)^2 λ 2 / ( 1 − λ ) 2 . Hence
C ′ ( λ ) < λ 2 ( 1 − λ ) 2 − 2 λ + 2 = λ 3 − 2 ( 1 − λ ) 3 λ ( 1 − λ ) 2 < 0 C'(\lambda)<\frac{\lambda^2}{(1-\lambda)^2}-\frac{2}{\lambda}+2=\frac{\lambda^3-2(1-\lambda)^3}{\lambda(1-\lambda)^2}<0 C ′ ( λ ) < ( 1 − λ ) 2 λ 2 − λ 2 + 2 = λ ( 1 − λ ) 2 λ 3 − 2 ( 1 − λ ) 3 < 0
by fact (iv), contrary to (S2.8).
(C) c 0 = 1 c_0=1 c 0 = 1 and c 1 ≤ − 3 c_1\leq-3 c 1 ≤ − 3 . Here ν ≥ 2 \nu\geq2 ν ≥ 2 . The series Ξ = ∑ k ≥ 2 ( 2 − e ν + k ) w k \Xi=\sum_{k\geq2}(2-e_{\nu+k})w^k Ξ = ∑ k ≥ 2 ( 2 − e ν + k ) w k dominates ∑ k ≥ 2 c k w k \sum_{k\geq2}c_k w^k ∑ k ≥ 2 c k w k coefficientwise. It equals Ψ 2 \Psi_2 Ψ 2 if ν = 2 \nu=2 ν = 2 , and if ν ≥ 3 \nu\geq3 ν ≥ 3 it is Ψ s \Psi_s Ψ s minus its linear term, namely Ψ 0 − w \Psi_0-w Ψ 0 − w , Ψ 1 − w \Psi_1-w Ψ 1 − w or Ψ 2 \Psi_2 Ψ 2 . By fact (iii), Ξ ′ ≤ Ψ 2 ′ \Xi'\leq\Psi'_2 Ξ ′ ≤ Ψ 2 ′ , so C ′ ( λ ) ≤ − 3 + Ψ 2 ′ ( λ ) < 0 C'(\lambda)\leq-3+\Psi'_2(\lambda)<0 C ′ ( λ ) ≤ − 3 + Ψ 2 ′ ( λ ) < 0 by fact (iv), contrary to (S2.8).
Case a 1 = a 2 = − 2 a_1=a_2=-2 a 1 = a 2 = − 2 . Since f 1 / 2 ( λ ) > 0 f_{1/2}(\lambda)>0 f 1/2 ( λ ) > 0 , f ≠ f 1 / 2 f\ne f_{1/2} f = f 1/2 . Let ν ≥ 3 \nu\geq3 ν ≥ 3 be the first index with a ν ≠ 2 a_{\nu}\ne2 a ν = 2 , and write f 1 / 2 − f = w ν D f_{1/2}-f=w^{\nu}D f 1/2 − f = w ν D with D = ∑ k ≥ 0 ( 2 − a ν + k ) w k D=\sum_{k\geq0}(2-a_{\nu+k})w^k D = ∑ k ≥ 0 ( 2 − a ν + k ) w k . Its coefficients lie in { 0 , … , 4 } \{0,\ldots,4\} { 0 , … , 4 } and the first is positive, so
1 ≤ D ( λ ) ≤ 4 1 − λ , 0 ≤ λ D ′ ( λ ) ≤ 4 λ ( 1 − λ ) 2 . 1 \le D(\lambda) \le\frac{4}{1-\lambda}, \qquad0 \le\lambda D'(\lambda) \le\frac{4\lambda}{(1-\lambda)^2}. 1 ≤ D ( λ ) ≤ 1 − λ 4 , 0 ≤ λ D ′ ( λ ) ≤ ( 1 − λ ) 2 4 λ .
The equations become λ n D ( λ ) = f 1 / 2 ( λ ) \lambda^nD(\lambda)=f_{1/2}(\lambda) λ n D ( λ ) = f 1/2 ( λ ) and, after division,
n + λ D ′ ( λ ) D ( λ ) = λ f 1 / 2 ′ ( λ ) f 1 / 2 ( λ ) = 4 λ 2 λ − 1 + 2 λ 1 − λ 2 . n+\frac{\lambda D'(\lambda)}{D(\lambda)}=\frac{\lambda f'_{1/2}(\lambda)}{f_{1/2}(\lambda)}=\frac{4\lambda}{2\lambda-1}+\frac{2\lambda}{1-\lambda^2}. n + D ( λ ) λ D ′ ( λ ) = f 1/2 ( λ ) λ f 1/2 ′ ( λ ) = 2 λ − 1 4 λ + 1 − λ 2 2 λ .
Put ε = 2 λ − 1 ∈ ( 0 , 2 λ + − 1 ) \varepsilon=2\lambda-1\in(0,2\lambda_+-1) ε = 2 λ − 1 ∈ ( 0 , 2 λ + − 1 ) . Since 4 λ / ( 2 λ − 1 ) = 2 + 2 / ε 4\lambda/(2\lambda-1)=2+2/\varepsilon 4 λ / ( 2 λ − 1 ) = 2 + 2/ ε , 2 λ / ( 1 − λ 2 ) > 0 2\lambda/(1-\lambda^2)>0 2 λ / ( 1 − λ 2 ) > 0 and 4 λ ( 1 − λ ) 2 < 4 λ + ( 1 − λ + ) 2 = 217480000 20820969 < 21 2 \frac{4\lambda}{(1-\lambda)^2}<\frac{4\lambda_+}{(1-\lambda_+)^2}=\frac{217480000}{20820969}<\frac{21}{2} ( 1 − λ ) 2 4 λ < ( 1 − λ + ) 2 4 λ + = 20820969 217480000 < 2 21 ,
n > 2 ε − 17 2 > 2 2 λ + − 1 − 17 2 = 12571 874 > 14 , n>\frac{2}{\varepsilon}-\frac{17}{2}>\frac{2}{2\lambda_+-1}-\frac{17}{2}=\frac{12571}{874}>14, n > ε 2 − 2 17 > 2 λ + − 1 2 − 2 17 = 874 12571 > 14 ,
so n ≥ 15 n\ge15 n ≥ 15 . On the other hand, since 2 / ε < n + 17 / 2 2/\varepsilon<n+17/2 2/ ε < n + 17/2 ,
x n = D ( λ ) f 1 / 2 ( λ ) ≤ 4 ε 2 ( 1 + λ ) < ( 2 ε ) 2 < ( n + 17 2 ) 2 . x^n=\frac{D(\lambda)}{f_{1/2}(\lambda)}\le\frac{4}{\varepsilon^2(1+\lambda)}<\left(\frac{2}{\varepsilon}\right)^2<\left(n+\frac{17}{2}\right)^2. x n = f 1/2 ( λ ) D ( λ ) ≤ ε 2 ( 1 + λ ) 4 < ( ε 2 ) 2 < ( n + 2 17 ) 2 .
But x > 1 / λ + > 9 / 5 x>1/\lambda_+>9/5 x > 1/ λ + > 9/5 , and ( 9 / 5 ) n > ( n + 17 / 2 ) 2 (9/5)^n>(n+17/2)^2 ( 9/5 ) n > ( n + 17/2 ) 2 for every n ≥ 15 n\ge15 n ≥ 15 : it holds at n = 15 n=15 n = 15 , as ( 9 / 5 ) 15 > 6746 > 2209 4 (9/5)^{15}>6746>\frac{2209}{4} ( 9/5 ) 15 > 6746 > 4 2209 , and each step multiplies the left side by 9 / 5 9/5 9/5 and the right side by at most ( 49 / 47 ) 2 < 9 / 5 (49/47)^2<9/5 ( 49/47 ) 2 < 9/5 . This contradiction completes the proof.
Lemma S2.10 (One half is inactive). One has 1 / 2 ∉ A 2 1/2\notin A_2 1/2 ∈ / A 2 ; equivalently, 2 ∉ C 2 2\notin C_2 2 ∈ / C 2 .
Proof. Suppose that f j ∈ F 2 f_j\in\mathcal{F}_2 f j ∈ F 2 have nonreal zeros z j → 1 / 2 z_j\to1/2 z j → 1/2 . By Lemma S1.2, every coefficientwise limit has a double zero at $1/2. It is f 1 / 2 f_{1/2} f 1/2 by Lemma S2.8, so f j → f 1 / 2 f_j\to f_{1/2} f j → f 1/2 , and since f 1 / 2 ′ ′ ( 1 / 2 ) > 0 f''_{1/2}(1/2)>0 f 1/2 ′′ ( 1/2 ) > 0 , Lemma S1.2 gives f j ( 1 / 2 ) > 0 f_j(1/2)>0 f j ( 1/2 ) > 0 for large j j j . But for large j j j the first two coefficients of f j f_j f j equal − 2 -2 − 2 and all others are at most 2 2 2 , so
f j ( 1 / 2 ) ≤ 1 − 1 − 1 2 + ∑ k ≥ 3 2 2 k = 0 , f_j(1/2)\le1-1-\frac{1}{2}+\sum_{k\ge3}\frac{2}{2^k}=0, f j ( 1/2 ) ≤ 1 − 1 − 2 1 + k ≥ 3 ∑ 2 k 2 = 0 ,
a contradiction. □
Proof of Theorem S2.7. By Lemmas S2.8 and S2.9, D 2 ∩ ( 0 , 1 / ρ ) = { 1 / 2 } D_2\cap(0,1/\rho)=\{1/2\} D 2 ∩ ( 0 , 1/ ρ ) = { 1/2 } , attained by f 1 / 2 f_{1/2} f 1/2 . Proposition S2.33 gives [ 1 / ρ , 1 ) ⊆ A 2 ⊆ D 2 [1/\rho,1)\subseteq A_2\subseteq D_2 [ 1/ ρ , 1 ) ⊆ A 2 ⊆ D 2 . This proves the formula for D 2 D_2 D 2 , and Lemma S2.10 gives 1 / 2 ∉ A 2 1/2\notin A_2 1/2 ∈ / A 2 . □
All-scale completion. Every estimate used at infinitely many scales is proved below. Only the exact finite obligations in § 17 are delegated to code.
Heights q ( q − 1 ) q(q-1) q ( q − 1 ) with q ≥ 4 q\ge4 q ≥ 4
Throughout, q ≥ 4 q\ge4 q ≥ 4 , h = q ( q − 1 ) h=q(q-1) h = q ( q − 1 ) , N = 2 h + 1 N=2h+1 N = 2 h + 1 , and a a a , b b b , δ = q − x \delta=q-x δ = q − x , r = x − 1 r=x-1 r = x − 1 are as in § 15.2 ; interior means an interior point of Z h e x t \mathbb{Z}^{\mathrm{ext}}_h Z h ext , and τ ∗ = τ h ( q ) = 1 − q 2 q 2 \tau_*=\tau_h(q)=1-\frac{q}{2q^2} τ ∗ = τ h ( q ) = 1 − 2 q 2 q , σ ∗ = σ h ( q ) = a + 1 2 q \sigma_*=\sigma_h(q)=a+\frac{1}{2q} σ ∗ = σ h ( q ) = a + 2 q 1 . We also write q ‾ = q − 1 2 \overline{q}=q-\frac{1}{2} q = q − 2 1 , and j j j for the offset of an outer layer; words are named by their digits or by the blocks of § 15.2 .
This subsection proves Theorems S2.3 and S2.5 and Proposition S2.4, which with Proposition 2.3 and Theorem S2.1 make every x ∈ ( 1 , q ) x\in(1,q) x ∈ ( 1 , q ) interior, as Theorem S2.6 requires. Six short words leave three windows; finitely many words cover the two compact ones. In the terminal window, phase bands accumulate at q q q ; on each band two families of outer layers and a central hierarchy of returning words overlap with explicit error bounds, and finitely many pairs ( q , m ) (q,m) ( q , m ) are checked exactly.
Short words
For 0 < δ < 1 / q 0<\delta<1/q 0 < δ < 1/ q we use A 2 A_2 A 2 , B 2 B_2 B 2 , A 3 A_3 A 3 , B 3 B_3 B 3 and A 2 ( − h ) = ( 2 q − 1 , − a 2 , − h ) A_2(-h)=(2q-1,-a^2,-h) A 2 ( − h ) = ( 2 q − 1 , − a 2 , − h ) and B 2 ( − b ) = ( 2 q − 1 , − ( a 2 + 1 ) , − b ) = U 2 , 1 B_2(-b)=(2q-1,-(a^2+1),-b)=U_{2,1} B 2 ( − b ) = ( 2 q − 1 , − ( a 2 + 1 ) , − b ) = U 2 , 1 . They capture the following δ \delta δ -intervals, all nondegenerate for q ≥ 4 q\ge4 q ≥ 4 ; the degree columns are explained in the proof below.
Table S2.11.
Lemma S2.11 (Direct interval positivity). For every word and interval in the table, every real q ≥ 4 q \ge4 q ≥ 4 , and δ ∈ [ δ − ( q ) , δ + ( q ) ] \delta\in[\delta_{-}(q),\delta_{+}(q)] δ ∈ [ δ − ( q ) , δ + ( q )] , the four margins
N − 2 x ± 2 x 2 V u ( x ) , N ± 2 x S u ( x ) , N = 2 q ( q − 1 ) + 1 , (57) N-2x\pm2x^{2}V_u(x),\qquad N\pm2xS_u(x),\qquad N=2q(q-1)+1, \tag*{(57)} N − 2 x ± 2 x 2 V u ( x ) , N ± 2 x S u ( x ) , N = 2 q ( q − 1 ) + 1 , (57) are strictly positive, giving the interval captures used in Theorem S2.3.
Proof. Idea. After rescaling, each margin is a polynomial with nonnegative Bernstein coefficients on a half-line times an interval. Set δ = δ − + ( δ + − δ − ) t \delta=\delta_{-}+(\delta_{+}-\delta_{-})t δ = δ − + ( δ + − δ − ) t with 0 ≤ t ≤ 1 0\le t\le1 0 ≤ t ≤ 1 , form the margins from the literal words by (S1.3), multiply by a positive constant times a power of q q q , and put q = 4 + s q=4+s q = 4 + s . The result p ( s , t ) p(s,t) p ( s , t ) is a polynomial whose degrees in ( s , t ) (s,t) ( s , t ) are at most the degree pair of its coordinate. If p = ∑ a k e s k t e p=\sum a_{ke}s^{k}t^{e} p = ∑ a k e s k t e , then
p ( s , t ) = ∑ k = 0 d s s k ∑ i = 0 d t b k i ( d t i ) t i ( 1 − t ) d t − i , b k i = ∑ e ≤ i a k e ( i e ) / ( d t e ) . (58) p(s,t)=\sum_{k=0}^{d_s}s^k\sum_{i=0}^{d_t}b_{ki}\binom{d_t}{i}t^i(1-t)^{d_t-i},\qquad b_{ki}=\sum_{e\le i}a_{ke}\binom{i}{e}\big/\binom{d_t}{e}. \tag*{(58)} p ( s , t ) = k = 0 ∑ d s s k i = 0 ∑ d t b k i ( i d t ) t i ( 1 − t ) d t − i , b k i = e ≤ i ∑ a k e ( e i ) / ( e d t ) . (58)
If all b k i ≥ 0 b_{ki}\ge0 b k i ≥ 0 and all b 0 i > 0 b_{0i}>0 b 0 i > 0 , then p > 0 p>0 p > 0 for s ≥ 0 s\ge0 s ≥ 0 and 0 ≤ t ≤ 1 0\le t\le1 0 ≤ t ≤ 1 , since the Bernstein basis is nonnegative and sums to one. The positivity of the twenty-four margins is Proposition S4.4; its certificate is this coefficient test. □
Proof of Theorem S2.3. Idea. The six capture intervals tile ( 0 , 1 / q ] (0,1/q] ( 0 , 1/ q ] in δ \delta δ except for three windows. Proposition 2.3 and Theorem S2.1 cover ( 1 , q − 1 / q ] (1,q-1/q] ( 1 , q − 1/ q ] . For 0 < δ ≤ 1 / q 0<\delta\le1/q 0 < δ ≤ 1/ q , Lemma S2.11 gives strict capture on the six closed δ \delta δ -intervals of the table. In increasing δ \delta δ they belong to A3 _3 3 , B3 _3 3 , A2 _2 2 , A2 ( − h ) , B _2(-h), B 2 ( − h ) , B _2(-b), B2 _2 2 ; consecutive ones overlap or share an endpoint except for two gaps, and the first starts at q − 3 q^{-3} q − 3 . The omitted parts are exactly the three windows of R q ( 3 ) R_q^{(3)} R q ( 3 ) , of lengths q − 3 q^{-3} q − 3 , 5 / ( 6 q 3 ) 5/(6q^3) 5/ ( 6 q 3 ) and 3 / ( 2 q 3 ) 3/(2q^3) 3/ ( 2 q 3 ) , whose sum is 10 / ( 3 q 3 ) 10/(3q^3) 10/ ( 3 q 3 ) . The six words have digits in B h B_h B h for every integer q ≥ 4 q\ge4 q ≥ 4 , and x < q x<q x < q gives x 2 + 2 x < q 2 + 2 q < N x^2+2x<q^2+2q<N x 2 + 2 x < q 2 + 2 q < N , so Proposition S1.4 gives interiority on the union. □
The compact windows
Proof of Proposition S2.4. Idea. Finitely many words per window, with margins positive uniformly in q q q , as for the short words. A word with digits in B h B_h B h whose four margins (S2.9) are positive at x x x captures x x x , and then x x x is interior by Proposition S1.4, as x 2 + 2 x < N x^2+2x<N x 2 + 2 x < N . The statement is Proposition S4.5 (interior points of C h e x t C_h^{\mathrm{ext}} C h ext and of Z h e x t Z_h^{\mathrm{ext}} Z h ext coincide, by the remark after Corollary S1.6). Its certificate covers J q , 2 J_{q,2} J q , 2 and J q , 3 J_{q,3} J q , 3 by consecutive closed cells, each with such a word: rational cells for finitely many q q q and, on residue classes of q q q modulo 4 4 4 , cells whose endpoints and words depend polynomially on q q q , with margins positive on each cell by Lemma S4.1. □
Terminal capture at every scale
Write Z u = ( V u , S u ) Z_u=(V_u,S_u) Z u = ( V u , S u ) ; a word u u u captures x x x if ∣ V u ( x ) ∣ < τ h ( x ) |V_u(x)|<\tau_h(x) ∣ V u ( x ) ∣ < τ h ( x ) and ∣ S u ( x ) ∣ < σ h ( x ) |S_u(x)|<\sigma_h(x) ∣ S u ( x ) ∣ < σ h ( x ) . For x < q x<q x < q , x 2 + 2 x < N x^2+2x<N x 2 + 2 x < N , τ h ( x ) > τ ∗ \tau_h(x)>\tau_* τ h ( x ) > τ ∗ and σ h ( x ) > σ ∗ \sigma_h(x)>\sigma_* σ h ( x ) > σ ∗ , so capture gives interiority (Proposition S1.4). The estimates below cover all large scales and, at each remaining scale, all large q q q ; (S2.26) lists the finite complement.
Definition S2.12 (Trace cells and linked covers). In a common real coordinate T T T , a trace cell has four strict affine endpoint inequalities
L V < T < U V , L S < T < U S . L_V<T<U_V,\qquad L_S<T<U_S. L V < T < U V , L S < T < U S .
They express the two strict trace bounds of its assigned word. An ordered list of cells is linked if each lower endpoint of one cell is strictly below both upper endpoints of its predecessor.
Lemma S2.13 (Strict chain principle). Let C 0 , … , C k C_0,\ldots,C_k C 0 , … , C k be a linked list of trace cells in a common coordinate T T T . If T T T exceeds both lower endpoints of C 0 C_0 C 0 and is less than both upper endpoints of C k C_k C k , then one of the cells contains T T T . Whenever its four inequalities imply strict capture at a parameter x x x , the point x x x is an interior point of Z h x Z_h^x Z h x .
Proof. Suppose both lower inequalities hold at one cell. Either its two upper inequalities also hold, or a failed upper inequality, together with the link inequalities, implies both lower inequalities of the next cell. Induction reaches a capturing cell, since both upper inequalities hold at the final cell. Apply Proposition S1.4. □
A cell whose estimates hold only for T T T above a constant carries that constant as a fifth, constant lower endpoint; linked lists and Lemma S2.13, with the same proof, apply to such cells.
All chains below are evaluated at a fixed parameter x x x , so their endpoints are numbers and the comparisons give pointwise capture.
Phase bands. For m ≥ 4 m\ge4 m ≥ 4 put
B q , m = { x ∈ [ q − q − 3 , q ) : ( q − x ) x m − 1 ≤ 1 2 ≤ ( q − x ) x m } . (59) B_{q,m}=\left\{x\in[q-q^{-3},q):(q-x)x^{m-1}\le\frac{1}{2}\le(q-x)x^m\right\}. \tag*{(59)} B q , m = { x ∈ [ q − q − 3 , q ) : ( q − x ) x m − 1 ≤ 2 1 ≤ ( q − x ) x m } . (59)
For x ∈ B q , m x\in B_{q,m} x ∈ B q , m define the phases s = 1 − 2 δ x m − 1 s=1-2\delta x^{m-1} s = 1 − 2 δ x m − 1 and t = 2 δ x m − 1 t=2\delta x^m-1 t = 2 δ x m − 1 and the centred coordinates T m = − x m s ≤ 0 T_m=-x^m s\le0 T m = − x m s ≤ 0 and T m + 1 = x m + 1 t ≥ 0 T_{m+1}=x^{m+1}t\ge0 T m + 1 = x m + 1 t ≥ 0 , so that
t = x ( 1 − s ) − 1 = r − x s . (60) t=x(1-s)-1=r-xs. \tag*{(60)} t = x ( 1 − s ) − 1 = r − x s . (60)
With q ˉ = q − 1 2 \bar q=q-\frac{1}{2} q ˉ = q − 2 1 and D m = 1 / ( 2 q ˉ m − 1 ) D_m=1/(2\bar q^{m-1}) D m = 1/ ( 2 q ˉ m − 1 ) we shall repeatedly use 0 < δ ≤ D m 0<\delta\le D_m 0 < δ ≤ D m , x ≥ q ˉ x\ge\bar q x ≥ q ˉ and r ≥ q − 3 2 r\ge q-\frac{3}{2} r ≥ q − 2 3 . The two outer families are W m , ℓ = B m ( − b ) ℓ ( − ( b − 1 ) ) W_{m,\ell}=B_m(-b)^\ell(-(b-1)) W m , ℓ = B m ( − b ) ℓ ( − ( b − 1 )) and U m + 1 , ℓ = B m + 1 ( − b ) ℓ U_{m+1,\ell}=B_{m+1}(-b)^\ell U m + 1 , ℓ = B m + 1 ( − b ) ℓ , with W m , − 1 = A m W_{m,-1}=A_m W m , − 1 = A m . A W-layer of offset j j j has ℓ = m − j − 2 \ell=m-j-2 ℓ = m − j − 2 and coordinate z W = x ℓ + 1 s = − T m / x j z_W=x^{\ell+1}s=-T_m/x^j z W = x ℓ + 1 s = − T m / x j ; a U-layer has ℓ = m − j \ell=m-j ℓ = m − j and z U = x ℓ + 1 t = T m + 1 / x j z_U=x^{\ell+1}t=T_{m+1}/x^j z U = x ℓ + 1 t = T m + 1 / x j . The variable layers are 1 ≤ j ≤ m − 1 1\le j\le m-1 1 ≤ j ≤ m − 1 (W) and 1 ≤ j ≤ m 1\le j\le m 1 ≤ j ≤ m (U); the fixed cells have j ∈ { 0 , − 1 } j\in\{0,-1\} j ∈ { 0 , − 1 } . Table 3 lists the cells of one band in the order in which they are linked below, with the coordinate in which each is used and the results that control it.
Table 3. The linked chain on one band B q , m B_{q,m} B q , m , listed from the centre T m = 0 T_m=0 T m = 0 to the centre T m + 1 = 0 T_{m+1}=0 T m + 1 = 0 ; read in the opposite order, it is linked in the sense of Definition S2.12 (proof of Theorem S2.5). For 4 ≤ q ≤ 7 4\le q\le7 4 ≤ q ≤ 7 the chains of Table 4 replace the central hierarchy from P m P_m P m to R m , 1 , 0 R_{m,1,0} R m , 1 , 0 , and the proof of Proposition S2.19 bounds the fixed cells. The pairs ( q , m ) (q,m) ( q , m ) of (S2.26) and the shoulder of Proposition S2.20 are covered instead by the rational and symbolic cells of Proposition S4.6.
Lemma S2.14 (Phase selection). Every x ∈ [ q − q − 3 , q ) x\in[q-q^{-3},q) x ∈ [ q − q − 3 , q ) satisfies
x ≤ q − ( 1 2 + 3 4 q 4 ) q − 3 or x ∈ B q , m for some m ≥ 4. x\le q-\left(\frac{1}{2}+\frac{3}{4q^4}\right)q^{-3}\quad\text{or}\quad x\in B_{q,m}\text{ for some }m\ge4. x ≤ q − ( 2 1 + 4 q 4 3 ) q − 3 or x ∈ B q , m for some m ≥ 4. Throughout this interval, x ≥ q − 1 / 2 > 1 x\ge q-1/2>1 x ≥ q − 1/2 > 1 and x ∈ [ q − 1 / q , q ] x\in[q-1/q,q] x ∈ [ q − 1/ q , q ] . On each selected band, 0 ≤ s < 1 0\le s<1 0 ≤ s < 1 , 0 ≤ t ≤ x − 1 0\le t\le x-1 0 ≤ t ≤ x − 1 , and 0 < δ ≤ D m 0<\delta\le D_m 0 < δ ≤ D m .
Proof. Choose the least m m m with δ x m ≥ 1 / 2 \delta x^m\ge1/2 δ x m ≥ 1/2 . Since δ x 2 < q − 1 < 1 / 2 \delta x^2<q^{-1}<1/2 δ x 2 < q − 1 < 1/2 , m ≥ 3 m\ge3 m ≥ 3 , and minimality gives δ x m − 1 < 1 / 2 \delta x^{m-1}<1/2 δ x m − 1 < 1/2 . If m = 3 m=3 m = 3 , then δ ≥ 1 / ( 2 x 3 ) ≥ 1 / ( 2 q 3 ) \delta\ge1/(2x^3)\ge1/(2q^3) δ ≥ 1/ ( 2 x 3 ) ≥ 1/ ( 2 q 3 ) , and x = q ( 1 − δ / q ) x=q(1-\delta/q) x = q ( 1 − δ / q ) gives δ ≥ ( 1 − δ / q ) − 3 2 q 3 ≥ \delta\ge\frac{(1-\delta/q)^{-3}}{2q^3}\ge δ ≥ 2 q 3 ( 1 − δ / q ) − 3 ≥ @page 55/80 starts_mid=0 ends_mid=1
Table 3 (PDF p. 55). The linked chain on one band B q , m B_{q,m} B q , m , listed from the centre T m = 0 T_m=0 T m = 0 to the centre T m + 1 = 0 T_{m+1}=0 T m + 1 = 0 ; read in the opposite order, it is linked in the sense of Definition S2.12 (proof of Theorem S2.5). For 4 ≤ q ≤ 7 4\le q\le7 4 ≤ q ≤ 7 the chains of Table 4 replace the central hierarchy from P m P_m P m to R m , 1 , 0 R_{m,1,0} R m , 1 , 0 , and the proof of Proposition S2.19 bounds the fixed cells. The pairs ( q , m ) (q,m) ( q , m ) of (S2.26) and the shoulder of Proposition S2.20 are covered instead by the rational and symbolic cells of Proposition S4.6.
Lemma S2.13 (Strict chain principle). Let C 0 , … , C k C_0,\ldots,C_k C 0 , … , C k be a linked list of trace cells in a common coordinate T T T . If T T T exceeds both lower endpoints of C 0 C_0 C 0 and is less than both upper endpoints of C k C_k C k , then one of the cells contains T T T . Whenever its four inequalities imply strict capture at a parameter x x x , the point x x x is an interior point of Z h x Z_h^x Z h x .
Proof. Suppose both lower inequalities hold at one cell. Either its two upper inequalities also hold, or a failed upper inequality, together with the link inequalities, implies both lower inequalities of the next cell. Induction reaches a capturing cell, since both upper inequalities hold at the final cell. Apply Proposition S1.4. □
A cell whose estimates hold only for T T T above a constant carries that constant as a fifth, constant lower endpoint; linked lists and Lemma S2.13, with the same proof, apply to such cells.
All chains below are evaluated at a fixed parameter x x x , so their endpoints are numbers and the comparisons give pointwise capture.
Phase bands. For m ≥ 4 m\ge4 m ≥ 4 put
B q , m = { x ∈ [ q − q − 3 , q ) : ( q − x ) x m − 1 ≤ 1 2 ≤ ( q − x ) x m } . (61) B_{q,m}=\left\{x\in[q-q^{-3},q):(q-x)x^{m-1}\le\frac{1}{2}\le(q-x)x^m\right\}. \tag*{(61)} B q , m = { x ∈ [ q − q − 3 , q ) : ( q − x ) x m − 1 ≤ 2 1 ≤ ( q − x ) x m } . (61) For x ∈ B q , m x\in B_{q,m} x ∈ B q , m define the phases s = 1 − 2 δ x m − 1 s=1-2\delta x^{m-1} s = 1 − 2 δ x m − 1 and t = 2 δ x m − 1 t=2\delta x^m-1 t = 2 δ x m − 1 and the centred coordinates T m = − x m s ≤ 0 T_m=-x^m s\le0 T m = − x m s ≤ 0 and T m + 1 = x m + 1 t ≥ 0 T_{m+1}=x^{m+1}t\ge0 T m + 1 = x m + 1 t ≥ 0 , so that
t = x ( 1 − s ) − 1 = r − x s . (62) t=x(1-s)-1=r-xs. \tag*{(62)} t = x ( 1 − s ) − 1 = r − x s . (62) With q ˉ = q − 1 2 \bar q=q-\frac{1}{2} q ˉ = q − 2 1 and D m = 1 / ( 2 q ˉ m − 1 ) D_m=1/(2\bar q^{m-1}) D m = 1/ ( 2 q ˉ m − 1 ) we shall repeatedly use 0 < δ ≤ D m 0<\delta\le D_m 0 < δ ≤ D m , x ≥ q ˉ x\ge\bar q x ≥ q ˉ and r ≥ q − 3 2 r\ge q-\frac{3}{2} r ≥ q − 2 3 . The two outer families are W m , ℓ = B m ( − b ) ℓ ( − ( b − 1 ) ) W_{m,\ell}=B_m(-b)^\ell(-(b-1)) W m , ℓ = B m ( − b ) ℓ ( − ( b − 1 )) and U m + 1 , ℓ = B m + 1 ( − b ) ℓ U_{m+1,\ell}=B_{m+1}(-b)^\ell U m + 1 , ℓ = B m + 1 ( − b ) ℓ , with W m , − 1 = A m W_{m,-1}=A_m W m , − 1 = A m . A W-layer of offset j j j has ℓ = m − j − 2 \ell=m-j-2 ℓ = m − j − 2 and coordinate z W = x ℓ + 1 s = − T m / x j z_W=x^{\ell+1}s=-T_m/x^j z W = x ℓ + 1 s = − T m / x j ; a U-layer has ℓ = m − j \ell=m-j ℓ = m − j and z U = x ℓ + 1 t = T m + 1 / x j z_U=x^{\ell+1}t=T_{m+1}/x^j z U = x ℓ + 1 t = T m + 1 / x j . The variable layers are 1 ≤ j ≤ m − 1 1\le j\le m-1 1 ≤ j ≤ m − 1 (W) and 1 ≤ j ≤ m 1\le j\le m 1 ≤ j ≤ m (U); the fixed cells have j ∈ { 0 , − 1 } j\in\{0,-1\} j ∈ { 0 , − 1 } . Table 3 lists the cells of one band in the order in which they are linked below, with the coordinate in which each is used and the results that control it.
Lemma S2.14 (Phase selection). Every x ∈ [ q − q − 3 , q ) x\in[q-q^{-3},q) x ∈ [ q − q − 3 , q ) satisfies
x ≤ q − ( 1 2 + 3 4 q 4 ) q − 3 or x ∈ B q , m for some m ≥ 4. x\le q-\left(\frac{1}{2}+\frac{3}{4q^4}\right)q^{-3}\quad\text{or}\quad x\in B_{q,m}\text{ for some }m\ge4. x ≤ q − ( 2 1 + 4 q 4 3 ) q − 3 or x ∈ B q , m for some m ≥ 4. Throughout this interval, x ≥ q − 1 / 2 > 1 x\ge q-1/2>1 x ≥ q − 1/2 > 1 and x ∈ [ q − 1 / q , q ] x\in[q-1/q,q] x ∈ [ q − 1/ q , q ] . On each selected band, 0 ≤ s < 1 0\le s<1 0 ≤ s < 1 , 0 ≤ t ≤ x − 1 0\le t\le x-1 0 ≤ t ≤ x − 1 , and 0 < δ ≤ D m 0<\delta\le D_m 0 < δ ≤ D m .
Proof. Choose the least m m m with δ x m ≥ 1 / 2 \delta x^m\ge1/2 δ x m ≥ 1/2 . Since δ x 2 < q − 1 < 1 / 2 \delta x^2<q^{-1}<1/2 δ x 2 < q − 1 < 1/2 , m ≥ 3 m\ge3 m ≥ 3 , and minimality gives δ x m − 1 < 1 / 2 \delta x^{m-1}<1/2 δ x m − 1 < 1/2 . If m = 3 m=3 m = 3 , then δ ≥ 1 / ( 2 x 3 ) ≥ 1 / ( 2 q 3 ) \delta\ge1/(2x^3)\ge1/(2q^3) δ ≥ 1/ ( 2 x 3 ) ≥ 1/ ( 2 q 3 ) , and x = q ( 1 − δ / q ) x=q(1-\delta/q) x = q ( 1 − δ / q ) gives δ ≥ ( 1 − δ / q ) − 3 2 q 3 ≥ \delta\ge\frac{(1-\delta/q)^{-3}}{2q^3}\ge δ ≥ 2 q 3 ( 1 − δ / q ) − 3 ≥
1 + 3 δ / q 2 q 3 ≥ 1 2 q 3 ( 1 + 3 2 q ) \frac{1+3\delta/q}{2q^3} \ge\frac{1}{2q^3}\left(1+\frac{3}{2q}\right) 2 q 3 1 + 3 δ / q ≥ 2 q 3 1 ( 1 + 2 q 3 ) , which is the first alternative. Otherwise x ∈ B q , m x \in B_{q,m} x ∈ B q , m with m ≥ 4 m \ge4 m ≥ 4 . Then s ∈ [ 0 , 1 ) s \in[0,1) s ∈ [ 0 , 1 ) and t = x ( 1 − s ) − 1 ∈ [ 0 , x − 1 ] t=x(1-s)-1 \in[0,x-1] t = x ( 1 − s ) − 1 ∈ [ 0 , x − 1 ] . Finally x ≥ q − q − 3 ≥ q ˉ x \ge q-q^{-3} \ge\bar q x ≥ q − q − 3 ≥ q ˉ , and δ x m − 1 ≤ 1 / 2 \delta x^{m-1} \le1/2 δ x m − 1 ≤ 1/2 gives δ ≤ D m \delta\le D_m δ ≤ D m . □
Lemma S2.15 (Exact outer trace formulae). For the words above,
V W = b r 2 − 1 + z W r + δ x ℓ + 1 2 ( 1 + s r 2 + j − ( 2 m − j − 2 ) s r ) , (63) V_W=\frac{b}{r^2}-1+\frac{z_W}{r}+\frac{\delta x^{\ell+1}}{2}\left(\frac{1+s}{r^2}+\frac{j-(2m-j-2)s}{r}\right), \tag*{(63)} V W = r 2 b − 1 + r z W + 2 δ x ℓ + 1 ( r 2 1 + s + r j − ( 2 m − j − 2 ) s ) , (63) S W = ( 2 − x ) b r 2 − 2 + z W r + δ x ℓ + 1 2 ( 1 + s r 2 + j − 1 − ( 2 m − j − 1 ) s r ) , (64) S_W=\frac{(2-x)b}{r^2}-2+\frac{z_W}{r}+\frac{\delta x^{\ell+1}}{2}\left(\frac{1+s}{r^2}+\frac{j-1-(2m-j-1)s}{r}\right), \tag*{(64)} S W = r 2 ( 2 − x ) b − 2 + r z W + 2 δ x ℓ + 1 ( r 2 1 + s + r j − 1 − ( 2 m − j − 1 ) s ) , (64) V U = b r 2 − z U r + δ x ℓ 2 ( 1 − t r 2 + j + ( 2 m − j ) t r ) , (65) V_U=\frac{b}{r^2}-\frac{z_U}{r}+\frac{\delta x^\ell}{2}\left(\frac{1-t}{r^2}+\frac{j+(2m-j)t}{r}\right), \tag*{(65)} V U = r 2 b − r z U + 2 δ x ℓ ( r 2 1 − t + r j + ( 2 m − j ) t ) , (65) S U = ( 2 − x ) b r 2 − z U r + δ x ℓ 2 ( 1 − t r 2 + j − 1 + ( 2 m + 1 − j ) t r ) . (66) S_U=\frac{(2-x)b}{r^2}-\frac{z_U}{r}+\frac{\delta x^\ell}{2}\left(\frac{1-t}{r^2}+\frac{j-1+(2m+1-j)t}{r}\right). \tag*{(66)} S U = r 2 ( 2 − x ) b − r z U + 2 δ x ℓ ( r 2 1 − t + r j − 1 + ( 2 m + 1 − j ) t ) . (66) The same identities hold at the fixed offsets j = 0 , − 1 j=0,-1 j = 0 , − 1 whenever the corresponding word is defined.
Proof. Use (S2.5)–(S2.6) for the block ( − b ) ℓ (-b)^\ell ( − b ) ℓ , (S1.3) for the final digit of W, and the formula u m = 1 + a / r 2 − 2 x y / r − δ y / r 2 + ( m − 1 ) δ y / r u_m=1+a/r^2-2xy/r-\delta y/r^2+(m-1)\delta y/r u m = 1 + a / r 2 − 2 x y / r − δ y / r 2 + ( m − 1 ) δ y / r with y = δ x m − 1 y=\delta x^{m-1} y = δ x m − 1 , which follows from (S2.1) along B m B_m B m as (S2.4) does: explicitly, V B m = 1 + a 2 r 2 + x m − 1 ( 1 − 2 δ − a 2 r 2 + ( m − 1 ) δ 2 r ) V_{B_m}=1+\frac{a^2}{r^2}+x^{m-1}\left(1-2\delta-\frac{a^2}{r^2}+(m-1)\frac{\delta^2}{r}\right) V B m = 1 + r 2 a 2 + x m − 1 ( 1 − 2 δ − r 2 a 2 + ( m − 1 ) r δ 2 ) , where 1 − 2 δ − a 2 r 2 = − δ ( 2 x r + δ r 2 ) 1-2\delta-\frac{a^2}{r^2}=-\delta\left(\frac{2x}{r}+\frac{\delta}{r^2}\right) 1 − 2 δ − r 2 a 2 = − δ ( r 2 x + r 2 δ ) and a 2 − b = a a^2-b=a a 2 − b = a . Substituting s s s and t t t gives the four identities. □
Since δ x m − 1 = 1 − s 2 \delta x^{m-1}=\frac{1-s}{2} δ x m − 1 = 2 1 − s and δ x m = 1 + t 2 \delta x^m=\frac{1+t}{2} δ x m = 2 1 + t , the prefactors are δ x ℓ + 1 = 1 − s 2 x j \delta x^{\ell+1}=\frac{1-s}{2x^j} δ x ℓ + 1 = 2 x j 1 − s and δ x ℓ = 1 + t 2 x j \delta x^\ell=\frac{1+t}{2x^j} δ x ℓ = 2 x j 1 + t . Hence the last line of (S2.13) is L j ( x ) − s 4 x j ( s r 2 + 2 m − 2 − ( 2 m − j − 2 ) s r ) L_j(x)-\frac{s}{4x^j}\left(\frac{s}{r^2}+\frac{2m-2-(2m-j-2)s}{r}\right) L j ( x ) − 4 x j s ( r 2 s + r 2 m − 2 − ( 2 m − j − 2 ) s ) and that of (S2.15) is L j ( x ) + t 4 x j ( 2 m + ( 2 m − j ) t r − t r 2 ) L_j(x)+\frac{t}{4x^j}\left(\frac{2m+(2m-j)t}{r}-\frac{t}{r^2}\right) L j ( x ) + 4 x j t ( r 2 m + ( 2 m − j ) t − r 2 t ) , where L j ( x ) = 1 4 x j ( 1 r 2 + 1 r ) L_j(x)=\frac{1}{4x^j}\left(\frac{1}{r^2}+\frac{1}{r}\right) L j ( x ) = 4 x j 1 ( r 2 1 + r 1 ) ; for S replace the numerator j j j of L j L_j L j by j − 1 j-1 j − 1 (keeping the factor x − j x^{-j} x − j ), and 2 m − j − 2 , 2 m − j 2m-j-2,2m-j 2 m − j − 2 , 2 m − j by 2 m − j − 1 , 2 m + 1 − j 2m-j-1,2m+1-j 2 m − j − 1 , 2 m + 1 − j .
Limiting states
Dropping the last lines and putting x = q x=q x = q gives the affine states
Z W 0 ( z ) = ( z − 1 a , z − ( a 2 + 1 ) a ) , Z U 0 ( z ) = ( q − 2 − z a , − ( q − 2 ) 2 + z a ) . Z_W^0(z)=\left(\frac{z-1}{a},\frac{z-(a^2+1)}{a}\right),\qquad Z_U^0(z)=\left(\frac{q-2-z}{a},-\frac{(q-2)^2+z}{a}\right). Z W 0 ( z ) = ( a z − 1 , a z − ( a 2 + 1 ) ) , Z U 0 ( z ) = ( a q − 2 − z , − a ( q − 2 ) 2 + z ) .
Since a σ ∗ = a 2 + a 2 q a\sigma_* = a^2+\frac{a}{2q} a σ ∗ = a 2 + 2 q a , their capture intervals are
λ W − < z < λ W + ( W ) , λ U − < z < λ U + ( U ) , λ W − = q + 1 2 q , λ W + = 1 + a τ ∗ , λ U ∓ = q − 2 ∓ a τ ∗ . (67) \lambda_W^-<z<\lambda_W^+\quad(\mathrm{W}),\qquad\lambda_U^-<z<\lambda_U^+\quad(\mathrm{U}),\qquad\lambda_W^-=\frac{q+1}{2q},\quad\lambda_W^+=1+a\tau_*,\quad\lambda_U^\mp=q-2\mp a\tau_*. \tag*{(67)} λ W − < z < λ W + ( W ) , λ U − < z < λ U + ( U ) , λ W − = 2 q q + 1 , λ W + = 1 + a τ ∗ , λ U ∓ = q − 2 ∓ a τ ∗ . (67)
At x = q x=q x = q the zero-phase layer terms are α j = L j ( q ) = j a + 1 4 a 2 q j \alpha_j=L_j(q)=\frac{ja+1}{4a^2q^j} α j = L j ( q ) = 4 a 2 q j j a + 1 (V) and β j = ( j − 1 ) a + 1 4 a 2 q j \beta_j=\frac{(j-1)a+1}{4a^2q^j} β j = 4 a 2 q j ( j − 1 ) a + 1 (S). The corrected states Z ‾ W , j = Z W 0 + ( α j , β j ) \overline Z_{W,j}=Z_W^0+(\alpha_j,\beta_j) Z W , j = Z W 0 + ( α j , β j ) and Z ‾ U , j = Z U 0 + ( α j , β j ) \overline Z_{U,j}=Z_U^0+(\alpha_j,\beta_j) Z U , j = Z U 0 + ( α j , β j ) have the capture intervals I ‾ W , j = ( λ W − − a β j , λ W + − a α j ) \overline I_{W,j}=(\lambda_W^- -a\beta_j,\lambda_W^+-a\alpha_j) I W , j = ( λ W − − a β j , λ W + − a α j ) and, for j ≥ 0 j\ge0 j ≥ 0 , I ‾ U , j = ( λ U − + a α j , λ U + + a α j ) \overline I_{U,j}=(\lambda_U^-+a\alpha_j,\lambda_U^++a\alpha_j) I U , j = ( λ U − + a α j , λ U + + a α j ) ; for j = − 1 j=-1 j = − 1 the capture interval of Z ‾ U , − 1 \overline Z_{U,-1} Z U , − 1 lies in I ‾ U , − 1 \overline I_{U,-1} I U , − 1 , the S-bound being binding above when q ≥ 9 q\ge9 q ≥ 9 . All these states have z z z -slope ± 1 / a \pm1/a ± 1/ a in both coordinates, so a state error ϵ \epsilon ϵ is absorbed by shrinking a capture interval by a ϵ a\epsilon a ϵ .
Proposition S2.16 (Outer-annulus cover). Suppose either q ≥ 8 q\ge8 q ≥ 8 , m ≥ 4 m\ge4 m ≥ 4 , or q ∈ { 4 , 5 , 6 , 7 } q\in\{4,5,6,7\} q ∈ { 4 , 5 , 6 , 7 } , m ≥ 11 m\ge11 m ≥ 11 . Then all variable W- and U-layers in B q , m B_{q,m} B q , m form a linked strict trace chain between their fixed handoff cells. The final W-layer and first U-layer overlap under the transition (S2.12).
Proof. Idea. At x = q x=q x = q the layer states are affine in z z z , and consecutive layers differ by the factor x x x in z z z ; the finite-x x x errors are smaller than the overlaps.
Let q ≥ 8 q\ge8 q ≥ 8 . On (S2.17), z W < λ W + < q z_W<\lambda_W^+<q z W < λ W + < q and z U < λ U + < 2 q z_U<\lambda_U^+<2q z U < λ U + < 2 q , so s / x j = z W x − m < 2 q D m / x s/x^j=z_Wx^{-m}<2qD_m/x s / x j = z W x − m < 2 q D m / x , t / x j = z U x − m − 1 < 4 q D m / x 2 t/x^j=z_Ux^{-m-1}<4qD_m/x^2 t / x j = z U x − m − 1 < 4 q D m / x 2 and t ≤ z U / x < 3 t\le z_U/x<3 t ≤ z U / x < 3 . With ∣ b ( r − 2 − a − 2 ) ∣ ≤ 2 a δ / r 2 |b(r^{-2}-a^{-2})|\le2a\delta/r^2 ∣ b ( r − 2 − a − 2 ) ∣ ≤ 2 a δ / r 2 , b ( r − 1 − a − 1 ) ≤ a δ / r b(r^{-1}-a^{-1})\le a\delta/r b ( r − 1 − a − 1 ) ≤ a δ / r , z ( r − 1 − a − 1 ) = z δ / ( a r ) z(r^{-1}-a^{-1})=z\delta/(ar) z ( r − 1 − a − 1 ) = z δ / ( a r ) , 0 ≤ 2 m − 2 − ( 2 m − j − 2 ) s ≤ 2 m 0\le2m-2-(2m-j-2)s\le2m 0 ≤ 2 m − 2 − ( 2 m − j − 2 ) s ≤ 2 m , and j / x j ≤ 1 / x j/x^j\le1/x j / x j ≤ 1/ x , ( j − 1 ) / x j ≤ 1 / x 2 (j-1)/x^j\le1/x^2 ( j − 1 ) / x j ≤ 1/ x 2 for j ≥ 1 j\ge1 j ≥ 1 , the four coordinate errors of Z W − Z ‾ W 0 ( z W ) Z_W-\overline Z_W^0(z_W) Z W − Z W 0 ( z W ) and Z U − Z ‾ U 0 ( z U ) Z_U-\overline Z_U^0(z_U) Z U − Z U 0 ( z U ) are at most
E W , V = D m ( 2 a r 2 + q a r ) + 1 4 x ( 1 r 2 + 1 r ) + q D m 2 x ( 1 r 2 + 2 m r ) , \mathcal E_{W,V}=D_m\left(\frac{2a}{r^2}+\frac{q}{ar}\right)+\frac{1}{4x}\left(\frac{1}{r^2}+\frac{1}{r}\right)+\frac{qD_m}{2x}\left(\frac{1}{r^2}+\frac{2m}{r}\right), E W , V = D m ( r 2 2 a + a r q ) + 4 x 1 ( r 2 1 + r 1 ) + 2 x q D m ( r 2 1 + r 2 m ) ,
E W , S = D m ( 2 a r 2 + a r + q a r ) + 1 4 ( 1 x r 2 + 1 x 2 r ) + q D m 2 x ( 1 r 2 + 2 m r ) , \mathcal E_{W,S}=D_m\left(\frac{2a}{r^2}+\frac{a}{r}+\frac{q}{ar}\right)+\frac{1}{4}\left(\frac{1}{xr^2}+\frac{1}{x^2r}\right)+\frac{qD_m}{2x}\left(\frac{1}{r^2}+\frac{2m}{r}\right), E W , S = D m ( r 2 2 a + r a + a r q ) + 4 1 ( x r 2 1 + x 2 r 1 ) + 2 x q D m ( r 2 1 + r 2 m ) ,
E U , V = D m ( 2 a r 2 + 2 q a r ) + 1 4 x ( 1 r 2 + 1 r ) + q D m x 2 ( 3 r 2 + 8 m r ) , \mathcal E_{U,V}=D_m\left(\frac{2a}{r^2}+\frac{2q}{ar}\right)+\frac{1}{4x}\left(\frac{1}{r^2}+\frac{1}{r}\right)+\frac{qD_m}{x^2}\left(\frac{3}{r^2}+\frac{8m}{r}\right), E U , V = D m ( r 2 2 a + a r 2 q ) + 4 x 1 ( r 2 1 + r 1 ) + x 2 q D m ( r 2 3 + r 8 m ) ,
ε U , S = D m ( 2 a r 2 + a r + 2 q a r ) + 1 4 ( 1 x r 2 + 1 x 2 r ) + q D m x 2 ( 3 r 2 + 8 m r ) . \varepsilon_{U,S}=D_m\left(\frac{2a}{r^2}+\frac{a}{r}+\frac{2q}{ar}\right)+\frac{1}{4}\left(\frac{1}{xr^2}+\frac{1}{x^2r}\right)+\frac{qD_m}{x^2}\left(\frac{3}{r^2}+\frac{8m}{r}\right). ε U , S = D m ( r 2 2 a + r a + a r 2 q ) + 4 1 ( x r 2 1 + x 2 r 1 ) + x 2 q D m ( r 2 3 + r 8 m ) .
Each term decreases in x x x and, through D m D_m D m and m D m mD_m m D m , in m m m ; so put x = q ˉ x=\bar{q} x = q ˉ , r = q − 3 2 r=q-\frac{3}{2} r = q − 2 3 and m = 4 m=4 m = 4 . Then each term of q E q\mathcal{E} q E is a product of the factors q / q ˉ q/\bar{q} q / q ˉ , a / r a/r a / r , q / r q/r q / r , 1 / a 1/a 1/ a , 1 / r 1/r 1/ r and 1 / q ˉ 1/\bar{q} 1/ q ˉ , all decreasing in q q q . At q = 8 q=8 q = 8 the four values of E \mathcal{E} E are 437863 59889375 \frac{437863}{59889375} 59889375 437863 , 248128 59889375 \frac{248128}{59889375} 59889375 248128 , 6799049 898340625 \frac{6799049}{898340625} 898340625 6799049 and 3953024 898340625 \frac{3953024}{898340625} 898340625 3953024 , all below 1 40 \frac{1}{40} 40 1 , so q E ≤ 8 E ∣ q = 8 < 1 5 q\mathcal{E}\leq8\mathcal{E}|_{q=8}<\frac{1}{5} q E ≤ 8 E ∣ q = 8 < 5 1 . Hence
∥ Z W − Z W 0 ( z W ) ∥ ∞ < 1 5 q , ∥ Z U − Z U 0 ( z U ) ∥ ∞ < 1 5 q , (68) \lVert Z_W-Z^0_W(z_W)\rVert_\infty<\frac{1}{5q},\qquad\lVert Z_U-Z^0_U(z_U)\rVert_\infty<\frac{1}{5q}, \tag*{(68)} ∥ Z W − Z W 0 ( z W ) ∥ ∞ < 5 q 1 , ∥ Z U − Z U 0 ( z U ) ∥ ∞ < 5 q 1 , (68)
and the intervals I W = ( λ W − + a 5 q , λ W + − a 5 q ) I_W=\left(\lambda_W^-+\frac{a}{5q},\lambda_W^+-\frac{a}{5q}\right) I W = ( λ W − + 5 q a , λ W + − 5 q a ) and I U = ( λ U − + a 5 q , λ U + − a 5 q ) I_U=\left(\lambda_U^-+\frac{a}{5q},\lambda_U^+-\frac{a}{5q}\right) I U = ( λ U − + 5 q a , λ U + − 5 q a ) suffice. Layer j j j covers − T m ∈ x j I W -T_m\in x^jI_W − T m ∈ x j I W , resp. T m + 1 ∈ x j I U T_{m+1}\in x^jI_U T m + 1 ∈ x j I U ; since x < q x<q x < q , successive layers overlap by at least
( λ W + − a 5 q ) − q ( λ W − + a 5 q ) = a ( 3 q 2 − 22 q + 5 ) 10 q 2 > 0 , (69) \left(\lambda_W^+-\frac{a}{5q}\right)-q\left(\lambda_W^-+\frac{a}{5q}\right)=\frac{a(3q^2-22q+5)}{10q^2}>0, \tag*{(69)} ( λ W + − 5 q a ) − q ( λ W − + 5 q a ) = 10 q 2 a ( 3 q 2 − 22 q + 5 ) > 0 , (69)
( λ U + − a 5 q ) − q ( λ U − + a 5 q ) = a ( 8 q 2 − 17 q + 5 ) 10 q 2 > 0. (70) \left(\lambda_U^+-\frac{a}{5q}\right)-q\left(\lambda_U^-+\frac{a}{5q}\right)=\frac{a(8q^2-17q+5)}{10q^2}>0. \tag*{(70)} ( λ U + − 5 q a ) − q ( λ U − + 5 q a ) = 10 q 2 a ( 8 q 2 − 17 q + 5 ) > 0. (70)
At the middle seam ( A m , j = m − 1 (A_{m},j=m-1 ( A m , j = m − 1 to B m + 1 , j = m ) B_{m+1},j=m) B m + 1 , j = m ) , (S2.12) reads z U = x r − x z W z_U=xr-xz_W z U = x r − x z W , and the overlap is at least λ U + − a 5 q + x ( λ W + − a 5 q ) − x r \lambda_U^+-\frac{a}{5q}+x\left(\lambda_W^+-\frac{a}{5q}\right)-xr λ U + − 5 q a + x ( λ W + − 5 q a ) − x r . This decreases in x x x and equals (S2.20) at x = q x=q x = q , because λ W + + λ U − = a \lambda_W^++\lambda_U^-=a λ W + + λ U − = a .
Now let 4 ≤ q ≤ 7 4\leq q\leq7 4 ≤ q ≤ 7 and m ≥ 11 m\geq11 m ≥ 11 , and compare with the corrected states. On I ‾ W , j \overline{I}_{W,j} I W , j and I ‾ U , j \overline{I}_{U,j} I U , j (j ≥ 1 j\geq1 j ≥ 1 ),
∥ Z W − Z ‾ W , j ( z W ) ∥ ∞ < 2 m D m , ∥ Z U − Z ‾ U , j ( z U ) ∥ ∞ < 2 m D m . (71) \lVert Z_W-\overline{Z}_{W,j}(z_W)\rVert_\infty<2mD_m,\qquad\lVert Z_U-\overline{Z}_{U,j}(z_U)\rVert_\infty<2mD_m. \tag*{(71)} ∥ Z W − Z W , j ( z W ) ∥ ∞ < 2 m D m , ∥ Z U − Z U , j ( z U ) ∥ ∞ < 2 m D m . (71)
Indeed, the fixed rational part costs at most ( 2 a r 2 + a r + 2 q a r ) δ ≤ 242 75 δ < 4 D m \left(\frac{2a}{r^2}+\frac{a}{r}+\frac{2q}{ar}\right)\delta\leq\frac{242}{75}\delta<4D_m ( r 2 2 a + r a + a r 2 q ) δ ≤ 75 242 δ < 4 D m (each summand decreases in q q q , and 242 75 \frac{242}{75} 75 242 is the value at q = 4 q=4 q = 4 ); moving L j L_j L j from x x x to q q q costs less than D m D_m D m , as ∣ L j ′ ∣ ≤ x − j ( j 2 + 2 j + 2 ) / ( 4 r ) ≤ 1 7 \lvert L'_j\rvert\leq x^{-j}(j^2+2j+2)/(4r)\leq\frac{1}{7} ∣ L j ′ ∣ ≤ x − j ( j 2 + 2 j + 2 ) / ( 4 r ) ≤ 7 1 for x ≥ q 2 x\geq\frac{q}{2} x ≥ 2 q ; and the remainders are at most q D m 2 x ( 1 r 2 + 2 m r ) < m D m \frac{qD_m}{2x}\left(\frac{1}{r^2}+\frac{2m}{r}\right)<mD_m 2 x q D m ( r 2 1 + r 2 m ) < m D m (W) and, as t ≤ z U / x < ( λ U + + a α 1 ) / q ˉ ≤ 11041 7644 < 3 2 t\leq z_U/x<(\lambda_U^++a\alpha_1)/\bar{q}\leq\frac{11041}{7644}<\frac{3}{2} t ≤ z U / x < ( λ U + + a α 1 ) / q ˉ ≤ 7644 11041 < 2 3 , at most q D m x 2 ( 3 2 r 2 + 5 m r ) < m D m \frac{qD_m}{x^2}\left(\frac{3}{2r^2}+\frac{5m}{r}\right)<mD_m x 2 q D m ( 2 r 2 3 + r 5 m ) < m D m (U), both largest at q = 4 q=4 q = 4 . The sum is ( m + 5 ) D m < 2 m D m (m+5)D_m<2mD_m ( m + 5 ) D m < 2 m D m . With ε m = 2 m D m \varepsilon_m=2mD_m ε m = 2 m D m the shrunken intervals are ( λ W − − a β j + a ε m , λ W + − a α j − a ε m ) (\lambda_W^--a\beta_j+a\varepsilon_m,\lambda_W^+-a\alpha_j-a\varepsilon_m) ( λ W − − a β j + a ε m , λ W + − a α j − a ε m ) and ( λ U − + a α j + a ε m , λ U + + a α j − a ε m ) (\lambda_U^-+a\alpha_j+a\varepsilon_m,\lambda_U^++a\alpha_j-a\varepsilon_m) ( λ U − + a α j + a ε m , λ U + + a α j − a ε m ) . By α j = q β j + 1 \alpha_j=q\beta_{j+1} α j = q β j + 1 and α j − q α j + 1 = − 1 / ( 4 a q j ) \alpha_j-q\alpha_{j+1}=-1/(4aq^j) α j − q α j + 1 = − 1/ ( 4 a q j ) , the successive overlaps and the middle seam are at least
λ W + − q λ W − − 1 4 q j − a ( q + 1 ) ε m , λ U + − q λ U − − 1 4 q j − a ( q + 1 ) ε m , λ U + − q λ U − − a q α m − 1 − a ( q + 1 ) ε m . \lambda_W^+-q\lambda_W^--\frac{1}{4q^j}-a(q+1)\varepsilon_m,\qquad\lambda_U^+-q\lambda_U^--\frac{1}{4q^j}-a(q+1)\varepsilon_m,\qquad\lambda_U^+-q\lambda_U^--aq\alpha_{m-1}-a(q+1)\varepsilon_m. λ W + − q λ W − − 4 q j 1 − a ( q + 1 ) ε m , λ U + − q λ U − − 4 q j 1 − a ( q + 1 ) ε m , λ U + − q λ U − − a q α m − 1 − a ( q + 1 ) ε m .
Here λ W + − q λ W − = q − 5 2 + 5 2 q − 1 2 q 2 ≥ 3 32 \lambda_W^+-q\lambda_W^-=\frac{q-5}{2}+\frac{5}{2q}-\frac{1}{2q^2}\geq\frac{3}{32} λ W + − q λ W − = 2 q − 5 + 2 q 5 − 2 q 2 1 ≥ 32 3 and λ U + − q λ U − = a ( 1 − 3 2 q + 1 2 q 2 ) ≥ 63 32 \lambda_U^+-q\lambda_U^-=a\left(1-\frac{3}{2q}+\frac{1}{2q^2}\right)\geq\frac{63}{32} λ U + − q λ U − = a ( 1 − 2 q 3 + 2 q 2 1 ) ≥ 32 63 increase in q q q , while 1 4 q j ≤ 1 16 \frac{1}{4q^j}\leq\frac{1}{16} 4 q j 1 ≤ 16 1 , a q α m − 1 < 10 − 3 aq\alpha_{m-1}<10^{-3} a q α m − 1 < 1 0 − 3 and a ( q + 1 ) ε m ≤ 168960 282475249 < 10 − 3 a(q+1)\varepsilon_m\leq\frac{168960}{282475249}<10^{-3} a ( q + 1 ) ε m ≤ 282475249 168960 < 1 0 − 3 , the value at q = 4 q=4 q = 4 , m = 11 m=11 m = 11 . (For the seam, as for q ≥ 8 q\geq8 q ≥ 8 , the bound decreases in x x x ; we put x = q x=q x = q and drop the term a α m > 0 a\alpha_m>0 a α m > 0 .) All three bounds are positive. □
Fixed cells
With P m = U m , m P_m=U_{m,m} P m = U m , m , the four fixed cells are W m , m − 2 W_{m,m-2} W m , m − 2 (j = 0 j=0 j = 0 , z = − T m z=-T_m z = − T m ), W m , m − 1 W_{m,m-1} W m , m − 1 (j = − 1 j=-1 j = − 1 , z = − x T m z=-xT_m z = − x T m ), U m + 1 , m U_{m+1,m} U m + 1 , m (j = 0 j=0 j = 0 , z = T m + 1 z=T_{m+1} z = T m + 1 ) and P m + 1 P_{m+1} P m + 1 (j = − 1 j=-1 j = − 1 , z = x T m + 1 z=xT_{m+1} z = x T m + 1 ). The limiting T T T -slope of offset j j j has absolute value 1 / ( a q j ) 1/(aq^j) 1/ ( a q j ) .
Lemma S2.17 (Fixed-offset estimate). If q ≥ 8 q\geq8 q ≥ 8 and m ≥ 4 m\geq4 m ≥ 4 , restrict each of the four fixed cells listed above to the following domain, which contains its corrected limiting capture interval: for j ∈ { − 1 , 0 } j\in\{-1,0\} j ∈ { − 1 , 0 } , require
λ W − − a β j < z W < λ W + − a α j or λ U − + a α j < z U < λ U + + a α j , \lambda_W^--a\beta_j<z_W<\lambda_W^+-a\alpha_j\quad\text{or}\quad\lambda_U^-+a\alpha_j<z_U<\lambda_U^++a\alpha_j, λ W − − a β j < z W < λ W + − a α j or λ U − + a α j < z U < λ U + + a α j , according to the side, W or U. On these domains, in particular, z W < 5 q / 4 z_W<5q/4 z W < 5 q /4 and z U < 2 q z_U<2q z U < 2 q . At the same T T T , the finite state then differs from its corrected limiting state by less than m D m mD_m m D m in sup norm.
Proof. Idea. As in Proposition S2.16, but the corrected limits absorb the layer terms, so only O ( m δ ) O(m\delta) O ( m δ ) remains. Put i = − j ∈ { 0 , 1 } i=-j\in\{0,1\} i = − j ∈ { 0 , 1 } . The bounds on z z z follow from the upper endpoints, for instance 5 q 4 − ( λ W + − a α − 1 ) = q 3 + 8 q 2 q − 2 a 2 4 a q 2 \frac{5q}{4}-(\lambda_W^+-a\alpha_{-1})=\frac{q^3+8q^2q-2a^2}{4aq^2} 4 5 q − ( λ W + − a α − 1 ) = 4 a q 2 q 3 + 8 q 2 q − 2 a 2 . They give 0 ≤ s , t ≤ 1 2 0\leq s,t\leq\frac{1}{2} 0 ≤ s , t ≤ 2 1 , since the corresponding powers of x ≥ q ˉ x\geq\bar{q} x ≥ q ˉ have exponent at least four. In (S2.13)–(S2.16) the rational part differs from its value at q q q by less than 2 δ 2\delta 2 δ , its larger coefficient being 2 a r 2 + a r + 2 q a r < 2 \frac{2a}{r^2}+\frac{a}{r}+\frac{2q}{ar}<2 r 2 2 a + r a + a r 2 q < 2 . At zero phase the layer terms are L i W = 1 4 r 2 − 1 4 L_i^W=\frac{1}{4r^2}-\frac{1}{4} L i W = 4 r 2 1 − 4 1 and L i U = 1 4 r 2 − 1 4 r − 1 2 L_i^U=\frac{1}{4r^2}-\frac{1}{4r}-\frac{1}{2} L i U = 4 r 2 1 − 4 r 1 − 2 1 , with derivatives − r − 2 2 r 3 -\frac{r-2}{2r^3} − 2 r 3 r − 2 and r − 2 4 r 3 \frac{r-2}{4r^3} 4 r 3 r − 2 , so their drift from x x x to q q q is at most δ / 16 \delta/16 δ /16 . For the remainders put u = ( 1 − s ) x i / 4 u=(1-s)x^{i}/4 u = ( 1 − s ) x i /4 (W) or u = ( 1 + t ) x i / 4 u=(1+t)x^{i}/4 u = ( 1 + t ) x i /4 (U), so that u s = δ z W / ( 2 x ) us=\delta z_W/(2x) u s = δ z W / ( 2 x ) and u t = δ z U / ( 2 x ) ut=\delta z_U/(2x) u t = δ z U / ( 2 x ) . The remainders then equal − u s ( 2 m − 2 ( 1 − s ) r 2 + i r − i ( 1 − s ) r ) -us\left(\frac{2m-2}{(1-s)r^2}+\frac{i}{r}-\frac{i}{(1-s)r}\right) − u s ( ( 1 − s ) r 2 2 m − 2 + r i − ( 1 − s ) r i ) and u t ( 2 m r + i ( 1 + t ) r − i ( 1 + t ) r 2 ) ut\left(\frac{2m}{r}+\frac{i}{(1+t)r}-\frac{i}{(1+t)r^2}\right) u t ( r 2 m + ( 1 + t ) r i − ( 1 + t ) r 2 i ) for the V V V coordinates; for the S S S coordinates the middle terms carry i + 1 i+1 i + 1 in place of i i i . By s , t ≤ 1 2 s,t\leq\frac{1}{2} s , t ≤ 2 1 all of them are at most 5 q δ 8 x ( 1 r 2 + 2 m − 2 r ) \frac{5q\delta}{8x}\left(\frac{1}{r^2}+\frac{2m-2}{r}\right) 8 x 5 q δ ( r 2 1 + r 2 m − 2 ) and q δ x ( 1 3 r 2 + 2 m + 2 / 3 r ) \frac{q\delta}{x}\left(\frac{1}{3r^2}+\frac{2m+2/3}{r}\right) x q δ ( 3 r 2 1 + r 2 m + 2/3 ) , and with q / x ≤ 16 15 q/x\leq\frac{16}{15} q / x ≤ 15 16 and 1 / r ≤ 2 13 1/r\leq\frac{2}{13} 1/ r ≤ 13 2 each is at most ( m 3 + 1 8 ) δ \left(\frac{m}{3}+\frac{1}{8}\right)\delta ( 3 m + 8 1 ) δ . Finally, comparison at the same T T T rather than the same z z z changes z z z by δ ∣ T ∣ \delta|T| δ ∣ T ∣ when i = 1 i=1 i = 1 and costs at most δ ∣ T ∣ / a < δ / 3 \delta|T|/a<\delta/3 δ ∣ T ∣/ a < δ /3 ; it vanishes when i = 0 i=0 i = 0 . Thus each coordinate error is less than
( 2 + m 16 + m 3 + 1 8 + 1 3 ) δ = ( m 3 + 121 48 ) δ < m D m . \left(2+\frac{m}{16}+\frac{m}{3}+\frac{1}{8}+\frac{1}{3}\right)\delta=\left(\frac{m}{3}+\frac{121}{48}\right)\delta<mD_m. ( 2 + 16 m + 3 m + 8 1 + 3 1 ) δ = ( 3 m + 48 121 ) δ < m D m .
Central words. The central cells use P m P_m P m and the returns
R m , ℓ , k = B m ( − b ) ℓ ( − ( b − 1 ) , − h , − q ) ( h ) m − ℓ − 3 ( h − k ) , Φ k = R m , 0 , k , R_{m,\ell,k}=B_m(-b)^\ell\bigl(-(b-1),-h,-q\bigr)(h)^{m-\ell-3}(h-k),\qquad\Phi_k=R_{m,0,k}, R m , ℓ , k = B m ( − b ) ℓ ( − ( b − 1 ) , − h , − q ) ( h ) m − ℓ − 3 ( h − k ) , Φ k = R m , 0 , k ,
on a central window − ϑ ≤ T ≤ 0 -\vartheta\leq T\leq0 − ϑ ≤ T ≤ 0 with T = T m T=T_m T = T m and ϑ ∈ { 1 2 , 2 3 } \vartheta\in\left\{\frac{1}{2},\frac{2}{3}\right\} ϑ ∈ { 2 1 , 3 2 } . Put y = δ x m − 1 = 1 2 ( 1 + T / x m ) y=\delta x^{m-1}=\frac{1}{2}(1+T/x^m) y = δ x m − 1 = 2 1 ( 1 + T / x m ) and κ = y − 1 2 = T / ( 2 x m ) \kappa=y-\frac{1}{2}=T/(2x^m) κ = y − 2 1 = T / ( 2 x m ) . For ℓ ∈ { 0 , 1 } \ell\in\{0,1\} ℓ ∈ { 0 , 1 } , (S1.3) for the digits − ( b − 1 ) , − h , − q -(b-1),-h,-q − ( b − 1 ) , − h , − q after U m , ℓ U_{m,\ell} U m , ℓ , and (S2.1) for the run ( h ) m − ℓ − 3 (h)^{m-\ell-3} ( h ) m − ℓ − 3 about z h z_h z h (the last digit adds ( k , k ) (k,k) ( k , k ) to the ( P , V ) (P,V) ( P , V ) ), give the closed formulae
V R m , ℓ , k = k − 2 ℓ y ( x + 1 ) x + − h − r x 2 T + y g 0 + y 2 g 1 + δ y f ℓ r 2 , (72) V_{R_{m,\ell,k}}=k-\frac{2\ell y(x+1)}{x}+\frac{-h-rx^2T+yg_0+y^2g_1+\delta yf_\ell}{r^2}, \tag*{(72)} V R m , ℓ , k = k − x 2 ℓ y ( x + 1 ) + r 2 − h − r x 2 T + y g 0 + y 2 g 1 + δ y f ℓ , (72)
S R m , ℓ , k = V R m , ℓ , k + h r + k − δ y ( x + 1 ) x ℓ + 1 r + x 2 y ( y − 1 ) r , S_{R_{m,\ell,k}}=V_{R_{m,\ell,k}}+\frac{h}{r}+k-\frac{\delta y(x+1)}{x^{\ell+1}r}+\frac{x^2y(y-1)}{r}, S R m , ℓ , k = V R m , ℓ , k + r h + k − x ℓ + 1 r δ y ( x + 1 ) + r x 2 y ( y − 1 ) ,
where g 0 = − 2 + ( m + 4 ) x 2 − ( m + 1 ) x 3 g_0=-2+(m+4)x^2-(m+1)x^3 g 0 = − 2 + ( m + 4 ) x 2 − ( m + 1 ) x 3 , g 1 = x 2 ( 2 m r − 1 ) g_1=x^2(2mr-1) g 1 = x 2 ( 2 m r − 1 ) and f ℓ = ( ( m − 1 − ℓ ) ( 1 − x 2 ) + 2 x ) / x ℓ + 1 f_\ell=((m-1-\ell)(1-x^2)+2x)/x^{\ell+1} f ℓ = (( m − 1 − ℓ ) ( 1 − x 2 ) + 2 x ) / x ℓ + 1 . At x = q x=q x = q , y = 1 2 y=\frac{1}{2} y = 2 1 , the limiting states are
Z P 0 ( T ) = ( ( q − 2 ) ( 3 q − 4 ) 4 a 2 , − ( q − 2 ) 2 a − q ( 2 q − 3 ) 4 a 2 ) − q a T ( 1 , 1 ) , Z_P^0(T)=\left(\frac{(q-2)(3q-4)}{4a^2},-\frac{(q-2)^2}{a}-\frac{q(2q-3)}{4a^2}\right)-\frac{q}{a}T(1,1), Z P 0 ( T ) = ( 4 a 2 ( q − 2 ) ( 3 q − 4 ) , − a ( q − 2 ) 2 − 4 a 2 q ( 2 q − 3 ) ) − a q T ( 1 , 1 ) ,
Z ℓ , k 0 ( T ) = ( φ ℓ , k , φ ℓ , k + q + k − q 2 4 a ) − q 2 a T ( 1 , 1 ) , φ ℓ , k = k − ℓ ( 1 + 1 q ) + 1 a − q 2 ( 2 q − 3 ) 4 a 2 . (73) Z_{\ell,k}^0(T)=\left(\varphi_{\ell,k},\varphi_{\ell,k}+q+k-\frac{q^2}{4a}\right)-\frac{q^2}{a}T(1,1),\qquad\varphi_{\ell,k}=k-\ell\left(1+\frac{1}{q}\right)+\frac{1}{a}-\frac{q^2(2q-3)}{4a^2}. \tag*{(73)} Z ℓ , k 0 ( T ) = ( φ ℓ , k , φ ℓ , k + q + k − 4 a q 2 ) − a q 2 T ( 1 , 1 ) , φ ℓ , k = k − ℓ ( 1 + q 1 ) + a 1 − 4 a 2 q 2 ( 2 q − 3 ) . (73)
Appending a digit d d d acts by Z ↦ x J Z − d ( 1 , 2 ) Z\mapsto xJZ-d(1,2) Z ↦ x J Z − d ( 1 , 2 ) , with J J J as in the proof of Proposition S2.2; since J ( 1 , 1 ) = ( 1 , 1 ) J(1,1)=(1,1) J ( 1 , 1 ) = ( 1 , 1 ) , each appended digit multiplies the limiting T T T -slopes q / a q/a q / a (P) and q 2 / a q^2/a q 2 / a (R) by q q q .
Lemma S2.18 (Base errors). Let m ≥ 4 m\geq4 m ≥ 4 , ℓ ∈ { 0 , 1 } \ell\in\{0,1\} ℓ ∈ { 0 , 1 } , $0\leq k\leq q/4$ and x ∈ B q , m x\in B_{q,m} x ∈ B q , m with $-\vartheta\leq T_m\leq 0. The limiting states (S2.23) have sup norm at most 2 q 2q 2 q , and:
(i) ∥ Z R m , ℓ , k − Z ℓ , k 0 ( T m ) ∥ ∞ < ( m + 3 ) δ \|Z_{R_{m,\ell,k}}-Z_{\ell,k}^0(T_m)\|_\infty<(m+3)\delta ∥ Z R m , ℓ , k − Z ℓ , k 0 ( T m ) ∥ ∞ < ( m + 3 ) δ if ϑ = 1 2 \vartheta=\frac{1}{2} ϑ = 2 1 or q ≥ 8 q\geq8 q ≥ 8 , and < ( 2 m + 8 ) δ <(2m+8)\delta < ( 2 m + 8 ) δ if ϑ = 2 3 \vartheta=\frac{2}{3} ϑ = 3 2 ;
(ii) ∥ Z P m − Z P 0 ( T m ) ∥ ∞ < ( m 5 + 6 5 ) δ \|Z_{P_m}-Z_P^0(T_m)\|_\infty<\left(\frac{m}{5}+\frac{6}{5}\right)\delta ∥ Z P m − Z P 0 ( T m ) ∥ ∞ < ( 5 m + 5 6 ) δ if ϑ = 1 2 \vartheta=\frac{1}{2} ϑ = 2 1 , and < 2 D m <2D_m < 2 D m if moreover q ≥ 8 q\geq8 q ≥ 8 .
Proof. Idea. After freezing y = 1 2 y=\frac{1}{2} y = 2 1 , the closed formulae are smooth in x x x ; the rest is of order δ \delta δ with a coefficient linear in m m m , because ∣ κ ∣ = ∣ T ∣ δ / ( 2 x y ) |\kappa|=|T|\delta/(2xy) ∣ κ ∣ = ∣ T ∣ δ / ( 2 x y ) .
Since 2 y ≥ 1 − 2 3 x − 4 > 99 100 2y\geq1-\frac{2}{3}x^{-4}>\frac{99}{100} 2 y ≥ 1 − 3 2 x − 4 > 100 99 , we have ∣ κ ∣ ≤ ϑ ′ δ / x |\kappa|\leq\vartheta'\delta/x ∣ κ ∣ ≤ ϑ ′ δ / x with ϑ ′ = 100 99 ϑ \vartheta'=\frac{100}{99}\vartheta ϑ ′ = 99 100 ϑ . The coefficient bounds below decrease in x x x , and then in q q q ; we evaluate them at x = q ˉ x=\bar q x = q ˉ , r = q − 3 2 r=q-\frac{3}{2} r = q − 2 3 and q = 4 q=4 q = 4 or q = 8 q=8 q = 8 .
(i) Write y g 0 + y 2 g 1 = ( g 0 2 + g 1 4 ) + κ H yg_0+y^2g_1=\left(\frac{g_0}{2}+\frac{g_1}{4}\right)+\kappa H y g 0 + y 2 g 1 = ( 2 g 0 + 4 g 1 ) + κ H , where H = g 0 + ( y + 1 2 ) g 1 H=g_0+\left(y+\frac{1}{2}\right)g_1 H = g 0 + ( y + 2 1 ) g 1 and g 0 2 + g 1 4 = − 1 + 7 4 x 2 − 1 2 x 3 \frac{g_0}{2}+\frac{g_1}{4}=-1+\frac{7}{4}x^2-\frac{1}{2}x^3 2 g 0 + 4 g 1 = − 1 + 4 7 x 2 − 2 1 x 3 . Then V R = F ℓ ( x ) + Δ V_R=F_\ell(x)+\Delta V R = F ℓ ( x ) + Δ , where
F ℓ ( x ) = k − ℓ x + 1 x + − h − 1 + 7 4 x 2 − 1 2 x 3 − r x 2 T r 2 F_\ell(x)=k-\ell\frac{x+1}{x}+\frac{-h-1+\frac{7}{4}x^2-\frac{1}{2}x^3-rx^2T}{r^2} F ℓ ( x ) = k − ℓ x x + 1 + r 2 − h − 1 + 4 7 x 2 − 2 1 x 3 − r x 2 T
satisfies F ℓ ( q ) = φ ℓ , k − q 2 4 a T F_\ell(q)=\varphi_{\ell,k}-\frac{q^2}{4a}T F ℓ ( q ) = φ ℓ , k − 4 a q 2 T , and Δ = δ y f ℓ + κ H r 2 − 2 ℓ κ ( x + 1 ) x \Delta=\frac{\delta yf_\ell+\kappa H}{r^2}-\frac{2\ell\kappa(x+1)}{x} Δ = r 2 δ y f ℓ + κ H − x 2 ℓ κ ( x + 1 ) . For t ′ ∈ [ x , q ] t'\in[x,q] t ′ ∈ [ x , q ] and ρ ′ = t ′ − 1 \rho'=t'-1 ρ ′ = t ′ − 1 ,
∂ t ′ F ℓ = N 2 ρ ′ 3 + ℓ t ′ 2 , N = − ρ ′ 3 + 4 ρ ′ 2 − 1 + ( q − t ′ ) ( 8 ρ ′ + 4 + 4 ( q − t ′ ) ) + 2 ∣ T ∣ ( ρ ′ 3 − ρ ′ ) , \partial_{t'}F_\ell=\frac{N}{2\rho'^3}+\frac{\ell}{t'^2},\qquad N=-\rho'^3+4\rho'^2-1+(q-t')(8\rho'+4+4(q-t'))+2|T|(\rho'^3-\rho'), ∂ t ′ F ℓ = 2 ρ ′3 N + t ′2 ℓ , N = − ρ ′3 + 4 ρ ′2 − 1 + ( q − t ′ ) ( 8 ρ ′ + 4 + 4 ( q − t ′ )) + 2∣ T ∣ ( ρ ′3 − ρ ′ ) ,
and q − t ′ ≤ D 4 ≤ 4 343 q-t'\leq D_4\leq\frac{4}{343} q − t ′ ≤ D 4 ≤ 343 4 gives − ρ ′ 3 < N < ( 2 ϑ − 1 ) ρ ′ 3 + 4 ρ ′ 2 − ( 2 ϑ − 1 5 ) ρ ′ -\rho'^3<N<(2\vartheta-1)\rho'^3+4\rho'^2-\left(2\vartheta-\frac{1}{5}\right)\rho' − ρ ′3 < N < ( 2 ϑ − 1 ) ρ ′3 + 4 ρ ′2 − ( 2 ϑ − 5 1 ) ρ ′ ; at ρ ′ = 5 2 \rho'=\frac{5}{2} ρ ′ = 2 5 , t ′ = 7 2 t'=\frac{7}{2} t ′ = 2 7 this yields ∣ ∂ t ′ F ℓ ∣ < 1 |\partial_{t'}F_\ell|<1 ∣ ∂ t ′ F ℓ ∣ < 1 . Next, 0 < − x ℓ + 1 f ℓ < ( m − 1 − ℓ ) ( x 2 − 1 ) 0<-x^{\ell+1}f_\ell<(m-1-\ell)(x^2-1) 0 < − x ℓ + 1 f ℓ < ( m − 1 − ℓ ) ( x 2 − 1 ) for m ≥ 4 m\geq4 m ≥ 4 , and 0 < H < x 2 ( ( m − 1 ) r + 2 ) 0<H<x^2((m-1)r+2) 0 < H < x 2 (( m − 1 ) r + 2 ) , because g 1 > 0 g_1>0 g 1 > 0 , y + 1 2 ∈ [ 7 8 , 1 ] y+\frac{1}{2}\in\left[\frac{7}{8},1\right] y + 2 1 ∈ [ 8 7 , 1 ] and g 0 + 7 8 g 1 = x 2 ( ( 3 m 4 − 1 ) x + 25 8 − 3 m 4 ) − 2 > 0 g_0+\frac{7}{8}g_1=x^2\left(\left(\frac{3m}{4}-1\right)x+\frac{25}{8}-\frac{3m}{4}\right)-2>0 g 0 + 8 7 g 1 = x 2 ( ( 4 3 m − 1 ) x + 8 25 − 4 3 m ) − 2 > 0 . Hence
∣ V R − V ℓ , k 0 ∣ δ < 1 + ( m − 1 ) μ + 2 ϑ ′ ( x r 2 + x + 1 x 2 ) , μ = x 2 x r + ϑ ′ x r . \frac{|V_R-V_{\ell,k}^0|}{\delta}<1+(m-1)\mu+2\vartheta'\left(\frac{x}{r^2}+\frac{x+1}{x^2}\right),\qquad\mu=\frac{x}{2xr}+\vartheta'\frac{x}{r}. δ ∣ V R − V ℓ , k 0 ∣ < 1 + ( m − 1 ) μ + 2 ϑ ′ ( r 2 x + x 2 x + 1 ) , μ = 2 x r x + ϑ ′ r x .
By (S2.22), S R − S ℓ , k 0 S_{\mathrm R}-S_{\ell,k}^{0} S R − S ℓ , k 0 adds to this h ( 1 r − 1 a ) = q δ 4 h\left(\frac{1}{r}-\frac{1}{a}\right)=\frac{q\delta}{4} h ( r 1 − a 1 ) = 4 q δ , at most ( x + 1 ) δ 2 x r \frac{(x+1)\delta}{2xr} 2 x r ( x + 1 ) δ , and x 2 y ( y − 1 ) 4 a + q 2 4 a 2 = δ 4 ( 1 − 1 a ) + x 2 r 2 4 a 2 < ( 1 4 + 1 400 ) δ \frac{x^{2}y(y-1)}{4a}+\frac{q^{2}}{4a^{2}}=\frac{\delta}{4}\left(1-\frac{1}{a}\right)+\frac{x^{2}r^{2}}{4a^{2}}<\left(\frac{1}{4}+\frac{1}{400}\right)\delta 4 a x 2 y ( y − 1 ) + 4 a 2 q 2 = 4 δ ( 1 − a 1 ) + 4 a 2 x 2 r 2 < ( 4 1 + 400 1 ) δ . Writing the total as ( m − 1 ) μ + ν (m-1)\mu+\nu ( m − 1 ) μ + ν , in the three cases ( ϑ = 1 2 , q ≥ 4 ) (\vartheta=\frac{1}{2},q\geq4) ( ϑ = 2 1 , q ≥ 4 ) , ( ϑ = 2 3 , q ≥ 4 ) (\vartheta=\frac{2}{3},q\geq4) ( ϑ = 3 2 , q ≥ 4 ) and ( ϑ = 2 3 , q ≥ 8 ) (\vartheta=\frac{2}{3},q\geq8) ( ϑ = 3 2 , q ≥ 8 ) the values of ( μ , ν ) (\mu,\nu) ( μ , ν ) are ( 3341 3465 , 983263 242550 ) \left(\frac{3341}{3465},\frac{983263}{242550}\right) ( 3465 3341 , 242550 983263 ) , ( 12473 10395 , 3176989 727650 ) \left(\frac{12473}{10395},\frac{3176989}{727650}\right) ( 10395 12473 , 727650 3176989 ) and ( 5561 6435 , 6202231 2053350 ) \left(\frac{5561}{6435},\frac{6202231}{2053350}\right) ( 6435 5561 , 2053350 6202231 ) , below ( 0.9643 , 4.054 ) (0.9643,4.054) ( 0.9643 , 4.054 ) , ( 1.2 , 4.367 ) (1.2,4.367) ( 1.2 , 4.367 ) and ( 0.8642 , 3.021 ) (0.8642,3.021) ( 0.8642 , 3.021 ) . So the error is below ( 0.9643 m + 3.09 ) δ < ( m + 3 ) δ (0.9643m+3.09)\delta<(m+3)\delta ( 0.9643 m + 3.09 ) δ < ( m + 3 ) δ , ( 1.2 m + 3.17 ) δ < ( 2 m + 8 ) δ (1.2m+3.17)\delta<(2m+8)\delta ( 1.2 m + 3.17 ) δ < ( 2 m + 8 ) δ and ( 0.8642 m + 2.16 ) δ < ( m + 3 ) δ (0.8642m+2.16)\delta<(m+3)\delta ( 0.8642 m + 2.16 ) δ < ( m + 3 ) δ .
(ii) By (S2.5)–(S2.6) with ℓ = m \ell=m ℓ = m , (S2.4), the formula for u m u_m u m in the proof of Lemma S2.15, and x 1 − m = δ / y x^{1-m}=\delta/y x 1 − m = δ / y ,
V P − V P 0 = ( b + 1 4 ) ( 1 r 2 − 1 a 2 ) + T ( 2 − x ) δ 4 r 2 y + T δ ( m − 1 / 2 r − 1 a r − 1 2 r 2 ) , V_{\mathrm P}-V_{\mathrm P}^{0}=\left(b+\frac{1}{4}\right)\left(\frac{1}{r^{2}}-\frac{1}{a^{2}}\right)+\frac{T(2-x)\delta}{4r^{2}y}+T\delta\left(\frac{m^{-1/2}}{r}-\frac{1}{ar}-\frac{1}{2r^{2}}\right), V P − V P 0 = ( b + 4 1 ) ( r 2 1 − a 2 1 ) + 4 r 2 y T ( 2 − x ) δ + T δ ( r m − 1/2 − a r 1 − 2 r 2 1 ) ,
S P − S P 0 = ( b + 1 4 ) [ ( 1 r 2 − 1 a 2 ) − ( 1 r − 1 a ) ] + T ( 3 − 2 x ) δ 4 r 2 y + T δ ( m r − 1 a r − 1 2 r 2 ) . S_{\mathrm P}-S_{\mathrm P}^{0}=\left(b+\frac{1}{4}\right)\left[\left(\frac{1}{r^{2}}-\frac{1}{a^{2}}\right)-\left(\frac{1}{r}-\frac{1}{a}\right)\right]+\frac{T(3-2x)\delta}{4r^{2}y}+T\delta\left(\frac{m}{r}-\frac{1}{ar}-\frac{1}{2r^{2}}\right). S P − S P 0 = ( b + 4 1 ) [ ( r 2 1 − a 2 1 ) − ( r 1 − a 1 ) ] + 4 r 2 y T ( 3 − 2 x ) δ + T δ ( r m − a r 1 − 2 r 2 1 ) .
As b + 1 4 = ( a − 1 2 ) 2 ≤ r 2 b+\frac{1}{4}=\left(a-\frac{1}{2}\right)^{2}\leq r^{2} b + 4 1 = ( a − 2 1 ) 2 ≤ r 2 , the first terms are at most 2 δ a \frac{2\delta}{a} a 2 δ and max ( 2 a , 1 a ) δ ≤ δ \max\left(\frac{2}{a},\frac{1}{a}\right)\delta\leq\delta max ( a 2 , a 1 ) δ ≤ δ ; for ∣ T ∣ ≤ 1 2 |T|\leq\frac{1}{2} ∣ T ∣ ≤ 2 1 the second terms are at most 0.07 δ 0.07\delta 0.07 δ and 0.17 δ 0.17\delta 0.17 δ , and the third at most m δ 2 r ≤ m δ 5 \frac{m\delta}{2r}\leq\frac{m\delta}{5} 2 r m δ ≤ 5 m δ . This gives the first bound. If q ≥ 8 q\geq8 q ≥ 8 , then δ = y x 1 − m ≤ D m θ m − 1 \delta=yx^{1-m}\leq D_m\theta^{m-1} δ = y x 1 − m ≤ D m θ m − 1 with θ = q ˉ / x ∈ [ 15 16 , 1 − 1 3 q ] \theta=\bar q/x\in\left[\frac{15}{16},1-\frac{1}{3q}\right] θ = q ˉ / x ∈ [ 16 15 , 1 − 3 q 1 ] , since x − q ˉ ≥ 1 2 − D 4 > 1 3 x-\bar q\geq\frac{1}{2}-D_4>\frac{1}{3} x − q ˉ ≥ 2 1 − D 4 > 3 1 . As max n n θ n ≤ 1 e ( 1 − θ ) \max_n n\theta^n\leq\frac{1}{e(1-\theta)} max n n θ n ≤ e ( 1 − θ ) 1 , we get m θ m − 1 ≤ 16 15 ⋅ 3 q e m\theta^{m-1}\leq\frac{16}{15}\cdot\frac{3q}{e} m θ m − 1 ≤ 15 16 ⋅ e 3 q , and the third terms are at most 16 15 ⋅ 3 q 2 e r D m < 0.725 D m \frac{16}{15}\cdot\frac{3q}{2er}D_m<0.725D_m 15 16 ⋅ 2 er 3 q D m < 0.725 D m , while the others are at most 1.08 D m 1.08D_m 1.08 D m . So both errors are below 2 D m 2D_m 2 D m .
Finally, the limiting coordinates (S2.23) are affine in k k k and T T T , so it suffices to bound them at the corners of 0 ≤ k ≤ q 4 0\leq k\leq\frac{q}{4} 0 ≤ k ≤ 4 q , ℓ ∈ { 0 , 1 } \ell\in\{0,1\} ℓ ∈ { 0 , 1 } , − 2 3 ≤ T ≤ 0 -\frac{2}{3}\leq T\leq0 − 3 2 ≤ T ≤ 0 . The largest is the S S S coordinate of Z 0 , q / 4 0 ( − 2 3 ) Z_{0,q/4}^{0}\left(-\frac{2}{3}\right) Z 0 , q /4 0 ( − 3 2 ) , which equals 17 12 q + 1 6 + 17 12 a + 1 4 a 2 \frac{17}{12}q+\frac{1}{6}+\frac{17}{12a}+\frac{1}{4a^{2}} 12 17 q + 6 1 + 12 a 17 + 4 a 2 1 , as q 2 a 2 = q + 1 + 1 a \frac{q^{2}}{a^{2}}=q+1+\frac{1}{a} a 2 q 2 = q + 1 + a 1 and 2 q − 3 4 a = 1 2 − 1 4 a \frac{2q-3}{4a}=\frac{1}{2}-\frac{1}{4a} 4 a 2 q − 3 = 2 1 − 4 a 1 ; it is below 2 q 2q 2 q by 7 12 q − 1 6 − 17 12 a − 1 4 a 2 ≥ 5 3 \frac{7}{12}q-\frac{1}{6}-\frac{17}{12a}-\frac{1}{4a^{2}}\geq\frac{5}{3} 12 7 q − 6 1 − 12 a 17 − 4 a 2 1 ≥ 3 5 . At each of the other corners, 2 q ∓ ( coordinate ) 2q\mp(\text{coordinate}) 2 q ∓ ( coordinate ) is likewise a rational function of q q q with nonnegative coefficients after q = 4 + s q=4+s q = 4 + s . □
Suffixes and endpoint movements. Let Z Z Z and Z 0 Z^{0} Z 0 be a finite state and its limit at the same T T T , with ∥ Z − Z 0 ∥ ∞ ≤ E 0 \lVert Z-Z^{0}\rVert_{\infty}\leq E_{0} ∥ Z − Z 0 ∥ ∞ ≤ E 0 and ∥ Z 0 ∥ ∞ ≤ G 0 \lVert Z^{0}\rVert_{\infty}\leq G_{0} ∥ Z 0 ∥ ∞ ≤ G 0 . Apply Lemma S1.8 with the matrix J J J of Proposition S2.2 and e = ( 1 , 2 ) e=(1,2) e = ( 1 , 2 ) . Since ∥ J ∥ ∞ = 3 \lVert J\rVert_{\infty}=3 ∥ J ∥ ∞ = 3 , ∥ J 2 ∥ ∞ = 5 \lVert J^{2}\rVert_{\infty}=5 ∥ J 2 ∥ ∞ = 5 and ∥ J e ∥ ∞ = 3 \lVert Je\rVert_{\infty}=3 ∥ J e ∥ ∞ = 3 , one and two appended digits give, respectively,
E i + 1 ≤ 3 q E i + 3 δ G i , G i + 1 ≤ 3 q G i + 2 h , (74) E_{i+1}\leq3qE_i+3\delta G_i,\qquad G_{i+1}\leq3qG_i+2h, \tag*{(74)} E i + 1 ≤ 3 q E i + 3 δ G i , G i + 1 ≤ 3 q G i + 2 h , (74)
E 2 ≤ 5 q 2 E 0 + δ ( 10 q G 0 + 3 h ) . E_2\leq5q^{2}E_0+\delta(10qG_0+3h). E 2 ≤ 5 q 2 E 0 + δ ( 10 q G 0 + 3 h ) .
Dividing by the T T T -slope gives the normalized endpoint movement. With the base errors of Lemmas S2.17 and S2.18, written as E 0 ≤ c D m E_0\leq cD_m E 0 ≤ c D m , and with a < q a<q a < q , G 0 = 4 q G_0=4q G 0 = 4 q , G 1 ≤ 14 q 2 G_1\leq14q^{2} G 1 ≤ 14 q 2 , G 2 ≤ 44 q 3 G_2\leq44q^{3} G 2 ≤ 44 q 3 (or G 0 = 2 q G_0=2q G 0 = 2 q , G 1 ≤ 8 q 2 G_1\leq8q^{2} G 1 ≤ 8 q 2 where marked ∘ \circ ∘ ), the movements of cells based on the listed words are as follows. The fixed cells and the returns carry no suffix; the central cells carry suffixes of at most two digits, or three where stated. For 4 ≤ q ≤ 7 4\leq q\leq7 4 ≤ q ≤ 7 the two R rows give the allowances inserted in the record: on − 2 3 ≤ T ≤ 0 -\frac{2}{3}\leq T\leq0 − 3 2 ≤ T ≤ 0 the base error proved for these returns is ( 2 m + 8 ) δ (2m+8)\delta ( 2 m + 8 ) δ (row R ∘ \mathrm R^{\circ} R ∘ ); its movement ( 10 m + 63 ) D m / q (10m+63)D_m/q ( 10 m + 63 ) D m / q decreases in m m m and is at most 173 D 11 / q 173D_{11}/q 173 D 11 / q , below the values η R 1 ( 11 ) = 204 D 11 / q \eta_{\mathrm R_1}(11)=204D_{11}/q η R 1 ( 11 ) = 204 D 11 / q and η R 0 ( 11 ) = 285 D 11 / q \eta_{\mathrm R_0}(11)=285D_{11}/q η R 0 ( 11 ) = 285 D 11 / q that the record inserts.
Table S2.25.
For instance, E 0 = ( m + 12 ) D m E_0=(m+12)D_m E 0 = ( m + 12 ) D m and G 0 = 4 q G_0=4q G 0 = 4 q give E 2 ≤ q 2 ( 9 m + 186 ) D m E_2\leq q^{2}(9m+186)D_m E 2 ≤ q 2 ( 9 m + 186 ) D m , and the two-digit bound with E 0 = ( 2 m + 8 ) D m E_0=(2m+8)D_m E 0 = ( 2 m + 8 ) D m , G 0 = 2 q G_0=2q G 0 = 2 q gives the last row. Each allowance ( α m + β ) D m (\alpha m+\beta)D_m ( α m + β ) D m in the table decreases for m ≥ 11 m\geq11 m ≥ 11 , since consecutive values have ratio ( 1 + α α m + β ) / q ˉ < 1 (1+\frac{\alpha}{\alpha m+\beta})/\bar q<1 ( 1 + α m + β α ) / q ˉ < 1 ; thus η P ≤ 285 D 11 \eta_P\leq285D_{11} η P ≤ 285 D 11 and η Φ ≤ 285 D 11 / q \eta_\Phi\leq285D_{11}/q η Φ ≤ 285 D 11 / q . The return estimates come from the direct closed formulae (S2.22), not from removing digits from another state.
The central hierarchy. If q = 4 K + ν q=4K+\nu q = 4 K + ν with 0 ≤ ν ≤ 3 0\le\nu\le3 0 ≤ ν ≤ 3 , put d = ( − 4 q 2 + 11 q − 8 + ν ) / 4 d=(-4q^2+11q-8+\nu)/4 d = ( − 4 q 2 + 11 q − 8 + ν ) /4 , and write Φ k ( e ) \Phi_k(e) Φ k ( e ) and P ( e ) P(e) P ( e ) for appended suffixes, with P = P m P=P_m P = P m . After the consecutive cells Φ 0 , … , Φ K − 1 \Phi_0,\ldots,\Phi_{K-1} Φ 0 , … , Φ K − 1 use the residue chains
ν = 0 : Φ K − 1 ( h − q − K + 2 ) ⇝ Φ K − 1 ( h − q − K + 1 ) ⇝ P ( d , d ) ⇝ P ( d ) ⇝ P ; \nu=0:\ \Phi_{K-1}(h-q-K+2)\rightsquigarrow\Phi_{K-1}(h-q-K+1)\rightsquigarrow P(d,d)\rightsquigarrow P(d)\rightsquigarrow P; ν = 0 : Φ K − 1 ( h − q − K + 2 ) ⇝ Φ K − 1 ( h − q − K + 1 ) ⇝ P ( d , d ) ⇝ P ( d ) ⇝ P ;
ν = 1 : Φ K − 1 ( h − 6 K ) ⇝ P ( d , − b − 2 K ) ⇝ P ( d , − b − 2 K − 1 ) ⇝ P ( d ) ⇝ P ; \nu=1:\ \Phi_{K-1}(h-6K)\rightsquigarrow P(d,-b-2K)\rightsquigarrow P(d,-b-2K-1)\rightsquigarrow P(d)\rightsquigarrow P; ν = 1 : Φ K − 1 ( h − 6 K ) ⇝ P ( d , − b − 2 K ) ⇝ P ( d , − b − 2 K − 1 ) ⇝ P ( d ) ⇝ P ;
ν = 2 : Φ K − 1 ( h − q − 3 K ) ⇝ P ( d , − b − 3 K − 1 ) ⇝ P ( d , − b − 3 K − 2 ) ⇝ P ( d ) ⇝ P ( d − 1 ) ⇝ P ; \nu=2:\ \Phi_{K-1}(h-q-3K)\rightsquigarrow P(d,-b-3K-1)\rightsquigarrow P(d,-b-3K-2)\rightsquigarrow P(d)\rightsquigarrow P(d-1)\rightsquigarrow P; ν = 2 : Φ K − 1 ( h − q − 3 K ) ⇝ P ( d , − b − 3 K − 1 ) ⇝ P ( d , − b − 3 K − 2 ) ⇝ P ( d ) ⇝ P ( d − 1 ) ⇝ P ;
ν = 3 : Φ K − 1 ( b ) ⇝ Φ K − 1 ( b − 1 ) ⇝ Φ K ⇝ Φ K ( q ( q − 2 ) + 2 ) ⇝ Φ K ( q ( q − 2 ) + 1 ) ⇝ P ( d − 1 , − b ) ⇝ P ( d − 1 ) ⇝ P . \nu=3:\ \Phi_{K-1}(b)\rightsquigarrow\Phi_{K-1}(b-1)\rightsquigarrow\Phi_K\rightsquigarrow\Phi_K(q(q-2)+2)\rightsquigarrow\Phi_K(q(q-2)+1)\rightsquigarrow P(d-1,-b)\rightsquigarrow P(d-1)\rightsquigarrow P. ν = 3 : Φ K − 1 ( b ) ⇝ Φ K − 1 ( b − 1 ) ⇝ Φ K ⇝ Φ K ( q ( q − 2 ) + 2 ) ⇝ Φ K ( q ( q − 2 ) + 1 ) ⇝ P ( d − 1 , − b ) ⇝ P ( d − 1 ) ⇝ P .
Proposition S2.19 (Central hierarchy and fixed seams). The ordered central hierarchy, together with
W m , m − 2 , W m , m − 1 , R m , 1 , 0 , R m , 0 , 0 and P m + 1 , U m + 1 , m , W_{m,m-2}, W_{m,m-1}, R_{m,1,0}, R_{m,0,0}\quad\text{and}\quad P_{m+1}, U_{m+1,m}, W m , m − 2 , W m , m − 1 , R m , 1 , 0 , R m , 0 , 0 and P m + 1 , U m + 1 , m , has consecutive strict overlaps in each of the following ranges:
For q = 4 , 5 , 6 , 7 q=4,5,6,7 q = 4 , 5 , 6 , 7 and m ≥ 11 m\ge11 m ≥ 11 , explicit replacement chains have strict overlaps. Their ordered words and rational limiting endpoints are listed in Table 4 ; their finite-m m m handoffs are checked in exact arithmetic. In every listed range, the first central cell strictly contains T = − 1 / 2 T=-1/2 T = − 1/2 and the final P-based cell strictly contains T = 0 T=0 T = 0 . All overlaps remain strict after the finite-x x x endpoint movements.
Table 4. Replacement chains for q = 4 , 5 , 6 , 7 q=4,5,6,7 q = 4 , 5 , 6 , 7 and m ≥ 11 m\ge11 m ≥ 11 in Proposition S2.19, with m m m omitted: R ℓ , k = R m , ℓ , k R_{\ell,k}=R_{m,\ell,k} R ℓ , k = R m , ℓ , k and P = P m P=P_m P = P m , followed by the displayed suffix digits. The endpoints bound the open interval of T T T on which the limiting state satisfies ∣ V ∣ < τ ∗ |V|<\tau_* ∣ V ∣ < τ ∗ and ∣ S ∣ < σ ∗ |S|<\sigma_* ∣ S ∣ < σ ∗ . In each chain the first interval contains T = − 1 / 2 T=-1/2 T = − 1/2 , the last contains T = 0 T=0 T = 0 , and consecutive intervals overlap, by at least 3 / 8192 3/8192 3/8192 , 24 / 15625 24/15625 24/15625 , 65 / 46656 65/46656 65/46656 and 103 / 67228 103/67228 103/67228 for q = 4 , 5 , 6 , 7 q=4,5,6,7 q = 4 , 5 , 6 , 7 ; all suffix digits lie in B h B_h B h .
Proof. Idea. At x = q x=q x = q every cell is an explicit rational interval in T T T , and the finite-x x x movements (S2.25) are smaller than all overlaps.
By (S2.23) every limiting trace coordinate in the chains is affine in T T T , so the four trace inequalities of a cell give a rational open interval. Consecutive Φ k \Phi_k Φ k -intervals overlap, the residue chains close the interval from Φ K − 1 \Phi_{K-1} Φ K − 1 to P P P , and every displayed digit lies in B h B_h B h . On each progression q = q 0 + 4 i q=q_0+4i q = q 0 + 4 i the overlaps and the two end containments, after the movements (S2.25) are inserted, are bounded below by positive coefficients (Proposition S4.6). Each cell is used only where its estimate holds: P- and Φ \Phi Φ -based cells on − 1 2 ≤ T ≤ 0 -\frac{1}{2} \le T \le0 − 2 1 ≤ T ≤ 0 , R m , 1 , 0 R_{m,1,0} R m , 1 , 0 on − 2 3 ≤ T ≤ 0 -\frac{2}{3} \le T \le0 − 3 2 ≤ T ≤ 0 ; for q ≥ 8 q \ge8 q ≥ 8 the record shows that R m , 1 , 0 R_{m,1,0} R m , 1 , 0 and R m , 0 , 0 R_{m,0,0} R m , 0 , 0 both strictly capture T = − 1 2 T=-\frac{1}{2} T = − 2 1 and that the W m , m − 1 / R m , 1 , 0 W_{m,m-1}/R_{m,1,0} W m , m − 1 / R m , 1 , 0 overlap meets T > − 2 3 T>-\frac{2}{3} T > − 3 2 .
For q ≥ 8 q \ge8 q ≥ 8 the variable annuli join the fixed cells directly. Put L W = λ W + + a 5 q L_W=\lambda_W^+ + \frac{a}{5q} L W = λ W + + 5 q a and L U = λ U − + a 5 q L_U=\lambda_U^- + \frac{a}{5q} L U = λ U − + 5 q a . The j = 1 j=1 j = 1 W-annular interval has boundary T = − x L W T=-xL_W T = − x L W , and W m , m − 2 W_{m,m-2} W m , m − 2 strictly contains the limiting target − q L W -qL_W − q L W after the additional shrink D m > δ L W D_m>\delta L_W D m > δ L W . Likewise the j = 1 j=1 j = 1 U-annular interval has boundary T = x L U T=xL_U T = x L U , and U m + 1 , m U_{m+1,m} U m + 1 , m contains q L U qL_U q L U after the additional shrink q D m > δ L U qD_m>\delta L_U q D m > δ L U . As these containments are strict, each fixed cell overlaps the adjacent open annular interval; the record checks them with the movements (S2.25).
For q = 4 , 5 , 6 , 7 q=4,5,6,7 q = 4 , 5 , 6 , 7 and m ≥ 11 m\ge11 m ≥ 11 the chains of Table 4 replace the central hierarchy from P m P_m P m to R m , 1 , 0 R_{m,1,0} R m , 1 , 0 . Their R-based cells are used on − 2 3 ≤ T ≤ 0 -\frac{2}{3}\le T\le0 − 3 2 ≤ T ≤ 0 , with movement ( 10 m + 63 ) D m / q ≤ 173 D 11 / q (10m+63)D_m/q\le173D_{11}/q ( 10 m + 63 ) D m / q ≤ 173 D 11 / q , below η R 1 ( 11 ) = 204 D 11 / q \eta_{R_1}(11)=204D_{11}/q η R 1 ( 11 ) = 204 D 11 / q and η R 0 ( 11 ) = 285 D 11 / q \eta_{R_0}(11)=285D_{11}/q η R 0 ( 11 ) = 285 D 11 / q . Next to the chain lie W m , m − 3 W_{m,m-3} W m , m − 3 and U m + 1 , m − 1 U_{m+1,m-1} U m + 1 , m − 1 (j = 1 j=1 j = 1 ) and the four fixed cells. At the same T T T their errors are below 2 m D m 2mD_m 2 m D m for j = 1 j=1 j = 1 (the proof of (S2.21) gives ( m + 5 ) D m (m+5)D_m ( m + 5 ) D m , passing from z z z to T T T adds less than D m D_m D m , and m + 6 < 2 m m+6<2m m + 6 < 2 m ) and below ( m + 6 ) D m < 3 m D m (m+6)D_m<3mD_m ( m + 6 ) D m < 3 m D m for j ∈ { 0 , − 1 } j\in\{0,-1\} j ∈ { 0 , − 1 } , since with q / x ≤ 8 7 q/x\le\frac{8}{7} q / x ≤ 7 8 , 1 / r ≤ 5 2 1/r\le\frac{5}{2} 1/ r ≤ 2 5 and a ≥ 3 a\ge3 a ≥ 3 the four contributions in the proof of Lemma S2.17 become 242 75 δ \frac{242}{75}\delta 75 242 δ , δ 16 \frac{\delta}{16} 16 δ , ( 32 35 m + 175 175 ) δ (\frac{32}{35}m+\frac{175}{175})\delta ( 35 32 m + 175 175 ) δ and 1621 21 δ \frac{1621}{21}\delta 21 1621 δ , whose sum ( 32 35 m + 12367 2800 ) δ (\frac{32}{35}m+\frac{12367}{2800})\delta ( 35 32 m + 2800 12367 ) δ is less than ( m + 6 ) δ (m+6)\delta ( m + 6 ) δ . Divided by the slopes 1 / ( a q j ) 1/(aq^j) 1/ ( a q j ) , the movements are at most 2 m a q D m 2maqD_m 2 ma q D m , 3 m a D m 3maD_m 3 ma D m and 3 m a D m / q 3maD_m/q 3 ma D m / q ; they decrease in m m m , and at m = 11 m=11 m = 11 they are at most 2 ⋅ 11 ⋅ 6 ⋅ 7 D 11 = 924 D 11 < 1000 D 11 2\cdot11\cdot6\cdot7D_{11}=924D_{11}<1000D_{11} 2 ⋅ 11 ⋅ 6 ⋅ 7 D 11 = 924 D 11 < 1000 D 11 . With this single allowance for the W and U cells of W m , m − 3 ⇝ W m , m − 2 ⇝ W m , m − 1 W_{m,m-3}\rightsquigarrow W_{m,m-2}\rightsquigarrow W_{m,m-1} W m , m − 3 ⇝ W m , m − 2 ⇝ W m , m − 1 and P m + 1 ⇝ U m + 1 , m ⇝ U m + 1 , m − 1 P_{m+1}\rightsquigarrow U_{m+1,m}\rightsquigarrow U_{m+1,m-1} P m + 1 ⇝ U m + 1 , m ⇝ U m + 1 , m − 1 , the allowance η P , ≤ 2 \eta_{P,\le2} η P , ≤ 2 for P m + 1 P_{m+1} P m + 1 , and the decreasing R- and P-allowances of the chains at m = 11 m=11 m = 11 , every consecutive overlap of the four chains stays strict, and the W m , m − 1 / R m , 1 , 0 W_{m,m-1}/R_{m,1,0} W m , m − 1 / R m , 1 , 0 overlap meets T > − 2 3 T>-\frac{2}{3} T > − 3 2 (Proposition S4.6). □
The shoulder and the finite range.
Proposition S2.20 (Initial shoulder). For every q ≥ 4 q\ge4 q ≥ 4 , every point of [ q − 1 q 3 , q − ( 1 2 + 3 4 q 4 ) q − 3 ] \left[q-\frac{1}{q^3},q-\left(\frac{1}{2}+\frac{3}{4q^4}\right)q^{-3}\right] [ q − q 3 1 , q − ( 2 1 + 4 q 4 3 ) q − 3 ] is an interior point of Z q ( q − 1 ) e x t Z^{\mathrm{ext}}_{q(q-1)} Z q ( q − 1 ) ext .
Proof. Idea. The words B 4 ( − b ) i B_4(-b)^i B 4 ( − b ) i follow the second channel across the shoulder. For q ≥ 12 q\ge12 q ≥ 12 , this rational interval is covered by five closed cells of the phase q 3 ( q − x ) q^3(q-x) q 3 ( q − x ) carrying the words B 4 B_4 B 4 , B 4 ( − b ) B_4(-b) B 4 ( − b ) , B 4 ( − b ) 2 B_4(-b)^2 B 4 ( − b ) 2 , B 4 ( − b ) 3 B_4(-b)^3 B 4 ( − b ) 3 , B 4 ( − b ) 4 B_4(-b)^4 B 4 ( − b ) 4 , whose four strict trace margins, cleared of positive factors, are positive for all q ≥ 12 q\ge12 q ≥ 12 by Lemma S4.1. For 4 ≤ q ≤ 11 4\le q\le11 4 ≤ q ≤ 11 , exact univariate Bernstein subdivision gives rational pieces; only q = 4 q=4 q = 4 needs the additional word B 4 ( − ( a 2 + 1 ) ) B_4(-(a^2+1)) B 4 ( − ( a 2 + 1 )) . Every digit lies in B h B_h B h . The statement is part (ii) of Proposition S4.6, whose certificate consists of these checks. □
The low-height chains cover 4 ≤ q ≤ 7 4\le q\le7 4 ≤ q ≤ 7 at all depths m ≥ 11 m\ge11 m ≥ 11 . The finite complement of these ranges and of those in Proposition S2.19 is
m = 4 , 4 ≤ q ≤ 274 ; m = 5 , 4 ≤ q ≤ 22 ; m = 6 , 4 ≤ q ≤ 10 ; 7 ≤ m ≤ 10 , 4 ≤ q ≤ 7. (75) m=4,\ 4\le q\le274;\qquad m=5,\ 4\le q\le22;\qquad m=6,\ 4\le q\le10;\qquad7\le m\le10,\ 4\le q\le7. \tag*{(75)} m = 4 , 4 ≤ q ≤ 274 ; m = 5 , 4 ≤ q ≤ 22 ; m = 6 , 4 ≤ q ≤ 10 ; 7 ≤ m ≤ 10 , 4 ≤ q ≤ 7. (75)
For each of these pairs, exact endpoint phase tests enclose B q , m B_{q,m} B q , m in a rational interval, and consecutive rational cells cover it by words with digits in B h B_h B h whose four strict trace margins are positive by rational Bernstein certificates (Proposition S4.6).
Proof of Theorem S2.5. Idea. On each band the cells above form one linked chain from one centre to the next. Fix x ∈ [ q − q − 3 , q ) x\in[q-q^{-3},q) x ∈ [ q − q − 3 , q ) . By Lemma S2.14, either Proposition S2.20 applies or x ∈ B q , m x\in B_{q,m} x ∈ B q , m for some m ≥ 4 m\ge4 m ≥ 4 . If ( q , m ) (q,m) ( q , m ) lies in (S2.26), the finite certificate applies. Otherwise, at the fixed x x x , T m T_m T m and T m + 1 = x m + 1 r + x 2 T m T_{m+1}=x^{m+1}r+x^2T_m T m + 1 = x m + 1 r + x 2 T m are increasing affine functions of each other, and by Propositions S2.16 and S2.19 and Lemma S2.17 the following list, read in the opposite order, is linked after all endpoint movements (for 4 ≤ q ≤ 7 4\le q\le7 4 ≤ q ≤ 7 the chains of Table 4 replace the central hierarchy, joined as in the proof of Proposition S2.19): the central hierarchy from P m P_m P m to R m , 0 , 0 R_{m,0,0} R m , 0 , 0 , then R m , 1 , 0 R_{m,1,0} R m , 1 , 0 , W m , m − 1 W_{m,m-1} W m , m − 1 , W m , m − 2 W_{m,m-2} W m , m − 2 , the variable layers across the middle seam (S2.12), U m + 1 , m U_{m+1,m} U m + 1 , m and P m + 1 P_{m+1} P m + 1 . Its end cells contain the ends T m = 0 T_m=0 T m = 0 and T m + 1 = 0 T_{m+1}=0 T m + 1 = 0 of B q , m B_{q,m} B q , m : for P m P_m P m by Proposition S2.19; and P m + 1 P_{m+1} P m + 1 , in the coordinate T m + 1 T_{m+1} T m + 1 , has the same limiting state Z P 0 Z_P^0 Z P 0 of (S2.23). This affine state satisfies ∣ V ∣ < τ ∗ |V|<\tau_* ∣ V ∣ < τ ∗ and ∣ S ∣ < σ ∗ |S|<\sigma_* ∣ S ∣ < σ ∗ exactly for max ( ℓ V , ℓ S ) < T < min ( u V , u S ) \max(\ell_V,\ell_S)<T<\min(u_V,u_S) max ( ℓ V , ℓ S ) < T < min ( u V , u S ) , where
ℓ V = − q 4 − 6 q 3 + 14 q 2 − 12 q + 2 4 q 3 a , u V = 7 q 4 − 26 q 3 + 30 q 2 − 12 q + 2 4 q 3 a , \ell_V=-\frac{q^4-6q^3+14q^2-12q+2}{4q^3a},\qquad u_V=\frac{7q^4-26q^3+30q^2-12q+2}{4q^3a}, ℓ V = − 4 q 3 a q 4 − 6 q 3 + 14 q 2 − 12 q + 2 , u V = 4 q 3 a 7 q 4 − 26 q 3 + 30 q 2 − 12 q + 2 ,
ℓ S = − 8 q 4 − 30 q 3 + 43 q 2 − 24 q + 2 4 q 2 a , u S = 6 q 3 − 15 q 2 + 8 q + 2 4 q 2 a . \ell_S=-\frac{8q^4-30q^3+43q^2-24q+2}{4q^2a},\qquad u_S=\frac{6q^3-15q^2+8q+2}{4q^2a}. ℓ S = − 4 q 2 a 8 q 4 − 30 q 3 + 43 q 2 − 24 q + 2 , u S = 4 q 2 a 6 q 3 − 15 q 2 + 8 q + 2 .
The numerators of − 1 16 − ℓ V -\frac{1}{16}-\ell_V − 16 1 − ℓ V , u V − 1 16 u_V-\frac{1}{16} u V − 16 1 , − 1 16 − ℓ S -\frac{1}{16}-\ell_S − 16 1 − ℓ S and u S − 1 16 u_S-\frac{1}{16} u S − 16 1 over the positive denominators 16 q 3 a 16q^3a 16 q 3 a , 16 q 3 a 16q^3a 16 q 3 a , 16 q 2 a 16q^2a 16 q 2 a , 16 q 2 a 16q^2a 16 q 2 a are ( q − 2 ) ( 3 q 3 − 17 q 2 + 22 q − 4 ) (q-2)(3q^3-17q^2+22q-4) ( q − 2 ) ( 3 q 3 − 17 q 2 + 22 q − 4 ) , ( q − 2 ) ( 27 q 3 − 49 q 2 + 22 q − 4 ) (q-2)(27q^3-49q^2+22q-4) ( q − 2 ) ( 27 q 3 − 49 q 2 + 22 q − 4 ) , 32 q 4 − 121 q 3 + 173 q 2 − 96 q + 8 32q^4-121q^3+173q^2-96q+8 32 q 4 − 121 q 3 + 173 q 2 − 96 q + 8 and 23 q 3 − 59 q 2 + 32 q + 8 23q^3-59q^2+32q+8 23 q 3 − 59 q 2 + 32 q + 8 , whose coefficients in q − 4 q-4 q − 4 are positive; the smallest margin is 1 384 \frac{1}{384} 384 1 , at q = 4 q=4 q = 4 . So the limiting interval contains [ − 1 16 , 1 16 ] [-\frac{1}{16},\frac{1}{16}] [ − 16 1 , 16 1 ] , while the movement m a D m / q maD_m/q ma D m / q (or 3 m a D m / q 3maD_m/q 3 ma D m / q ) is below 1 16 \frac{1}{16} 16 1 . Lemma S2.13 gives capture. □
S2.5.4. Finite checks delegated to the records. The records enter this section only through their statements; Table 6 lists their uses. For Proposition S4.4 the certificate consists of the coefficient conditions (S2.10) for the margins of Lemma S2.11; for Proposition S4.5, of covers of the two windows by rational cells for finitely many q q q and by symbolic cells on residue classes of q q q modulo 4; for Proposition S4.6, of rational cells for the pairs of (S2.26), rational and symbolic cells for the shoulder, the coefficient positivity, on residue classes of q q q , of the limiting overlaps and end containments of Proposition S2.19 with the allowances stated in the record (the fixed seams are checked with the larger movements 3 m a D m 3maD_m 3 ma D m and 3 m a D m / q 3maD_m/q 3 ma D m / q ), and the rational comparisons of the chains of Table 4 at m = 11 m=11 m = 11 . All monotonicity statements are proved above.
Table 6. Proof obligations and their uses.
Height six
We prove Theorems S2.22 and S2.25, the case q = 3 q=3 q = 3 of Theorem S2.6, by the scheme of Section S2.3: buffered returns give interior points up to 2999 / 1000 2999/1000 2999/1000 , and bounded orbits in a moving rectangle give double zeros beyond. We use F x , d F_{x,d} F x , d , r = x − 1 r=x-1 r = x − 1 , δ = 3 − x \delta=3-x δ = 3 − x , z d z_d z d and (S2.1) from Section S2.2, with h = 6 h=6 h = 6 , and F x , d + 1 = F x , d − ( 1 , 1 ) F_{x,d+1}=F_{x,d}-(1,1) F x , d + 1 = F x , d − ( 1 , 1 ) . Words and orbits are admissible if their digits lie in B 6 B_6 B 6 ; F x , u = F x , d m ∘ ⋯ ∘ F x , d 1 F_{x,u}=F_{x,d_m}\circ\cdots\circ F_{x,d_1} F x , u = F x , d m ∘ ⋯ ∘ F x , d 1 for u = ( d 1 , … , d m ) u=(d_1,\ldots,d_m) u = ( d 1 , … , d m ) , so F x , u ( 1 , 1 ) = ( P u ( x ) , V u ( x ) ) F_{x,u}(1,1)=(P_u(x),V_u(x)) F x , u ( 1 , 1 ) = ( P u ( x ) , V u ( x )) .
The compact interval
Lemma S2.21 (Buffered control). Let x > 1 x>1 x > 1 and η > 0 \eta>0 η > 0 , and let Y ⊂ R 2 Y\subset\mathbb{R}^2 Y ⊂ R 2 be compact. Suppose that Y Y Y is the union of finitely many closed sets Y i Y_i Y i , each with an assigned admissible nonempty word u i u_i u i such that
F x , u i ( Y i ) + [ − η , η ] 2 ⊂ Y , (76) F_{x,u_i}(Y_i)+[-\eta,\eta]^2\subset Y, \tag*{(76)} F x , u i ( Y i ) + [ − η , η ] 2 ⊂ Y , (76)
and that some admissible word v v v satisfies
F x , v ( 1 , 1 ) + [ − η , η ] 2 ⊂ Y . (77) F_{x,v}(1,1)+[-\eta,\eta]^2\subset Y. \tag*{(77)} F x , v ( 1 , 1 ) + [ − η , η ] 2 ⊂ Y . (77)
Then x x x is an interior point of Z 6 e x t Z_6^{\mathrm{ext}} Z 6 ext .
Proof. Idea. The buffers absorb the change from the real map to the remainder map at nearby complex parameters. For c = x ′ + i y c=x'+iy c = x ′ + i y , x ′ > 0 x'>0 x ′ > 0 , y ≠ 0 y\ne0 y = 0 , write a remainder Z Z Z as P = Re Z P=\operatorname{Re}Z P = Re Z , V = Re Z + x ′ Im Z / y V=\operatorname{Re}Z+x'\operatorname{Im}Z/y V = Re Z + x ′ Im Z / y ; then Z ↦ c Z − d Z\mapsto cZ-d Z ↦ c Z − d becomes
P + = x ′ P − d + y 2 x ′ ( P − V ) , V + = x ′ V + P + , (78) P^+=x'P-d+\frac{y^2}{x'}(P-V),\qquad V^+=x'V+P^+, \tag*{(78)} P + = x ′ P − d + x ′ y 2 ( P − V ) , V + = x ′ V + P + , (78)
and Z = 1 Z=1 Z = 1 has state ( 1 , 1 ) (1,1) ( 1 , 1 ) . This map F c , d F_{c,d} F c , d is F x ′ , d F_{x',d} F x ′ , d at y = 0 y=0 y = 0 , so F c , u ( z ) F_{c,u}(z) F c , u ( z ) is continuous in ( c , z ) (c,z) ( c , z ) . By compactness of the Y i Y_i Y i and finiteness of the words, some open disc U ∋ x U\ni x U ∋ x in { c : ∣ c ∣ > 1 , Re c > 0 } \{c\colon|c|>1,\operatorname{Re}c>0\} { c : ∣ c ∣ > 1 , Re c > 0 } has, for c ∈ U c\in U c ∈ U , F c , u i F_{c,u_i} F c , u i within η \eta η of F x , u i F_{x,u_i} F x , u i on Y i Y_i Y i and F c , v ( 1 , 1 ) F_{c,v}(1,1) F c , v ( 1 , 1 ) within η \eta η of F x , v ( 1 , 1 ) F_{x,v}(1,1) F x , v ( 1 , 1 ) in the sup norm; so F c , u i ( Y i ) ⊂ Y F_{c,u_i}(Y_i)\subset Y F c , u i ( Y i ) ⊂ Y and F c , v ( 1 , 1 ) ∈ Y F_{c,v}(1,1)\in Y F c , v ( 1 , 1 ) ∈ Y by (S2.27)–(S2.28). For c ∈ U c\in U c ∈ U , apply v v v , then at each state in Y Y Y the word of some Y i Y_i Y i containing it. States ending a word lie in Y Y Y , and the others are images of ( 1 , 1 ) (1,1) ( 1 , 1 ) or of the Y i Y_i Y i under finitely many continuous maps; so the remainders Z k = P k + i ( y / x ′ ) ( V k − P k ) Z_k=P_k+i(y/x')(V_k-P_k) Z k = P k + i ( y / x ′ ) ( V k − P k ) (Z k = P k Z_k=P_k Z k = P k if y = 0 y=0 y = 0 ) are bounded. As Z k = c k ( 1 − ∑ j ≤ k d j c − j ) Z_k=c^k(1-\sum_{j\le k}d_jc^{-j}) Z k = c k ( 1 − ∑ j ≤ k d j c − j ) and ∣ c ∣ > 1 |c|>1 ∣ c ∣ > 1 , 1 − ∑ j ≥ 1 d j c − j = 0 1-\sum_{j\ge1}d_jc^{-j}=0 1 − ∑ j ≥ 1 d j c − j = 0 , and Proposition 2.3 (digits − d j -d_j − d j ) with (S1.1) gives U ⊂ Z 6 e x t U\subset Z_6^{\mathrm{ext}} U ⊂ Z 6 ext . □
Theorem S2.22 (Compact height-six interval). One has [ 1 , 2999 / 1000 ] ⊂ int R Z 6 e x t [1,2999/1000]\subset\operatorname{int}_{\mathbb{R}}Z_6^{\mathrm{ext}} [ 1 , 2999/1000 ] ⊂ int R Z 6 ext .
Proof. Idea. The annulus, one word and buffered control cover three ranges. By Proposition 2.3 , ( 1 , 7 ) ⊂ int Z 6 ext (1,\sqrt{7}) \subset\operatorname{int} Z_{6}^{\operatorname{ext}} ( 1 , 7 ) ⊂ int Z 6 ext . On [ 7 , 137 / 50 ] [\sqrt{7},137/50] [ 7 , 137/50 ] the word ( 5 , − 5 ) (5,-5) ( 5 , − 5 ) has V = 3 x 2 − 10 x + 5 V=3x^{2}-10x+5 V = 3 x 2 − 10 x + 5 and S = 4 x 2 − 15 x + 10 S=4x^{2}-15x+10 S = 4 x 2 − 15 x + 10 by (S1.3), increasing for x ≥ 15 / 8 x \ge15/8 x ≥ 15/8 , with endpoint values 26 − 10 7 > − 2674 2500 26-10\sqrt{7}>-\frac{2674}{2500} 26 − 10 7 > − 2500 2674 and 38 − 15 7 > − 2 38-15\sqrt{7}>-2 38 − 15 7 > − 2 . As ( 137 / 50 + 1 ) 2 = 34969 / 2500 < 14 (137/50+1)^2=34969/2500<14 ( 137/50 + 1 ) 2 = 34969/2500 < 14 , x 2 + 2 x < 13 x^2+2x<13 x 2 + 2 x < 13 , so ∣ V ∣ < 2 < τ 6 ( x ) |V|<2<\tau_6(x) ∣ V ∣ < 2 < τ 6 ( x ) by (S1.5) and ∣ S ∣ < 2 < 13 / ( 2 x ) = σ 6 ( x ) |S|<2<13/(2x)=\sigma_6(x) ∣ S ∣ < 2 < 13/ ( 2 x ) = σ 6 ( x ) ; Proposition S1.4 applies. For x ∈ [ 137 / 50 , 2999 / 1000 ] x\in[137/50,2999/1000] x ∈ [ 137/50 , 2999/1000 ] , choose a band I I I containing x x x . Part (i) of Proposition S4.7 supplies the compact set Y I Y_I Y I , its finite closed cell cover, buffered one-digit returns, and a buffered entry from ( 1 , 1 ) (1,1) ( 1 , 1 ) . Hence Lemma S2.21 applies, proving that x x x is an interior point of Z 6 ext Z_{6}^{\operatorname{ext}} Z 6 ext .
By Corollary S1.6, [ 1000 / 2999 , 1 ) ⊂ D 6 [1000/2999,1)\subset D_6 [ 1000/2999 , 1 ) ⊂ D 6 .
A moving rectangle and paired renewal
For 2999 / 1000 ≤ x ≤ 3 2999/1000\le x\le3 2999/1000 ≤ x ≤ 3 let R x = [ − p ˉ , p ˉ ] × [ − v ˉ , v ˉ ] R_x=[-\bar p,\bar p]\times[-\bar v,\bar v] R x = [ − p ˉ , p ˉ ] × [ − v ˉ , v ˉ ] , with p ˉ = 3 / r \bar p=3/r p ˉ = 3/ r and v ˉ = 2 / r 2 \bar v=2/r^2 v ˉ = 2/ r 2 , be the rectangle of Section S2.3, and p 0 = − 2 / r p_0=-2/r p 0 = − 2/ r , so z − 2 = ( p 0 , v ˉ ) z_{-2}=(p_0,\bar v) z − 2 = ( p 0 , v ˉ ) . Put
λ − = ( 2 x − 4 − 2 v ˉ ) / x 2 , κ = x λ − , λ + = ( 2 x − 3 ) / x 2 , ξ ∗ = ( x − 2 − 1 / r ) / x 2 , W = 1 − 1 / r ; (79) \lambda_-=(2x-4-2\bar v)/x^2,\quad\kappa=x\lambda_-,\quad\lambda_+=(2x-3)/x^2,\quad\xi_*=(x-2-1/r)/x^2,\quad W=1-1/r; \tag*{(79)} λ − = ( 2 x − 4 − 2 v ˉ ) / x 2 , κ = x λ − , λ + = ( 2 x − 3 ) / x 2 , ξ ∗ = ( x − 2 − 1/ r ) / x 2 , W = 1 − 1/ r ; (79)
these are positive, and 1.999 ≤ r ≤ 2 1.999\le r\le2 1.999 ≤ r ≤ 2 gives 1 ≤ 2 v ˉ < 1.002 1\le2\bar v<1.002 1 ≤ 2 v ˉ < 1.002 . By (S2.1) and F x , d + 1 = F x , d − ( 1 , 1 ) F_{x,d+1}=F_{x,d}-(1,1) F x , d + 1 = F x , d − ( 1 , 1 ) ,
F x , − 6 ( z − 2 + ( ξ − 1 , ζ − 1 ) ) = z − 2 + ( 4 − x + x ξ , 4 − 2 x + x ( ξ + ζ ) ) , (80) F_{x,-6}(z_{-2}+(\xi-1,\zeta-1))=z_{-2}+(4-x+x\xi,4-2x+x(\xi+\zeta)), \tag*{(80)} F x , − 6 ( z − 2 + ( ξ − 1 , ζ − 1 )) = z − 2 + ( 4 − x + x ξ , 4 − 2 x + x ( ξ + ζ )) , (80)
F x , ( − 1 , − j − 2 ) ( z − 2 + ( ξ , ζ ) ) = z − 2 + ( x 2 ξ − x + j , x 2 ( ζ + 2 ξ ) − 2 x + j ) , (81) F_{x,(-1,-j-2)}(z_{-2}+(\xi,\zeta))=z_{-2}+(x^2\xi-x+j,x^2(\zeta+2\xi)-2x+j), \tag*{(81)} F x , ( − 1 , − j − 2 ) ( z − 2 + ( ξ , ζ )) = z − 2 + ( x 2 ξ − x + j , x 2 ( ζ + 2 ξ ) − 2 x + j ) , (81)
and F x , − 5 F_{x,-5} F x , − 5 gives the first image minus ( 1 , 1 ) (1,1) ( 1 , 1 ) . Two captures recur.
(a) If − δ − 1 / r ≤ x ξ ≤ x W -\delta-1/r\le x\xi\le xW − δ − 1/ r ≤ x ξ ≤ x W and κ ≤ ξ + ζ ≤ ( 2 x − 3 ) / x \kappa\le\xi+\zeta\le(2x-3)/x κ ≤ ξ + ζ ≤ ( 2 x − 3 ) / x , then − 6 -6 − 6 or − 5 -5 − 5 sends z − 2 + ( ξ − 1 , ζ − 1 ) z_{-2}+(\xi-1,\zeta-1) z − 2 + ( ξ − 1 , ζ − 1 ) into R x R_x R x . The vertical constraints hold for ξ + ζ \xi+\zeta ξ + ζ in [ κ , ( 2 x − 4 ) / x ] [\kappa,(2x-4)/x] [ κ , ( 2 x − 4 ) / x ] and [ ( 2 x − 3 − 2 v ˉ ) / x , ( 2 x − 3 ) / x ] [(2x-3-2\bar v)/x,(2x-3)/x] [( 2 x − 3 − 2 v ˉ ) / x , ( 2 x − 3 ) / x ] , which overlap as 2 v ˉ ≥ 1 2\bar v\ge1 2 v ˉ ≥ 1 ; the horizontal ones are − δ − 1 / r ≤ x ξ ≤ x − 4 + 5 / r -\delta-1/r\le x\xi\le x-4+5/r − δ − 1/ r ≤ x ξ ≤ x − 4 + 5/ r , and x W ≤ x − 4 + 5 / r xW\le x-4+5/r x W ≤ x − 4 + 5/ r is x ≤ 3 x\le3 x ≤ 3 .
(b) If − x / ( 4 r ) ≤ x 2 ξ ≤ x − 2 -x/(4r)\le x^2\xi\le x-2 − x / ( 4 r ) ≤ x 2 ξ ≤ x − 2 and λ − ≤ ζ + 2 ξ ≤ λ + \lambda_-\le\zeta+2\xi\le\lambda_+ λ − ≤ ζ + 2 ξ ≤ λ + , then ( − 1 , − 6 ) (-1,-6) ( − 1 , − 6 ) or ( − 1 , − 5 ) (-1,-5) ( − 1 , − 5 ) sends z − 2 + ( ξ , ζ ) z_{-2}+(\xi,\zeta) z − 2 + ( ξ , ζ ) into R x R_x R x . The vertical constraints hold for ζ + 2 ξ \zeta+2\xi ζ + 2 ξ in [ λ − , ( 2 x − 4 ) / x 2 ] [\lambda_-,(2x-4)/x^2] [ λ − , ( 2 x − 4 ) / x 2 ] and [ ( 2 x − 3 − 2 v ˉ ) / x 2 , λ + ] [(2x-3-2\bar v)/x^2,\lambda_+] [( 2 x − 3 − 2 v ˉ ) / x 2 , λ + ] , which overlap likewise; the horizontal ones reduce to 1 − x / 4 + δ r ≥ 0 1-x/4+\delta r\ge0 1 − x /4 + δ r ≥ 0 , 4 − x + ( 1 − x / 4 ) / r ≥ 0 4-x+(1-x/4)/r\ge0 4 − x + ( 1 − x /4 ) / r ≥ 0 and 2 ≤ 5 / r 2\le5/r 2 ≤ 5/ r (which also gives the bound 1 ≤ 5 / r 1\le5/r 1 ≤ 5/ r for the digit − 5 -5 − 5 ).
Theorem S2.23 (Uniform rectangle continuation). Every point of R x R_x R x admits an infinite bounded admissible orbit under (S1.8), for every 2999 / 1000 ≤ x ≤ 3 2999/1000\le x\le3 2999/1000 ≤ x ≤ 3 . More precisely, every point has a nonempty finite return to R x R_x R x . Return lengths need not be uniformly bounded, but all intermediate states in the construction lie in one bounded set, independent of x x x and of the return length.
Lemma S2.24 (Compact paired return). Let
Q = { ( u , w ) : u ≥ 1 / 28 , 0 ≤ w ≤ u , u + w ≤ 1 / 3 } . Q=\{(u,w):u\ge1/28,\ 0\le w\le u,\ u+w\le1/3\}. Q = {( u , w ) : u ≥ 1/28 , 0 ≤ w ≤ u , u + w ≤ 1/3 } . For every ( u , w ) ∈ Q (u,w)\in Q ( u , w ) ∈ Q and 2999 / 1000 ≤ x ≤ 3 2999/1000\le x\le3 2999/1000 ≤ x ≤ 3 , at least one of
( p 0 + u , v ˉ + w ) , ( p 0 + u − 1 , v ˉ + w − 1 ) (p_0+u,\bar v+w),\qquad(p_0+u-1,\bar v+w-1) ( p 0 + u , v ˉ + w ) , ( p 0 + u − 1 , v ˉ + w − 1 ) has an admissible word of length at most three whose image satisfies
∣ P ∣ ≤ p ˉ − δ / 100 , ∣ V ∣ ≤ v ˉ − δ / 100. |P|\le\bar p-\delta/100,\qquad|V|\le\bar v-\delta/100. ∣ P ∣ ≤ p ˉ − δ /100 , ∣ V ∣ ≤ v ˉ − δ /100.
Proof. This is part (ii) of Proposition S4.7.
Proof of Theorem S2.23. Idea. A failed one-step return leaves two adjacent states; a fast return, the cone Q Q Q or an excursion along z − 2 z_{-2} z − 2 takes over, and an excursion returns or renews the pair with u u u multiplied by at least x 2 x^2 x 2 .
Pairs. By the symmetry ( P , V , d ) ↦ ( − P , − V , − d ) (P,V,d)\mapsto(-P,-V,-d) ( P , V , d ) ↦ ( − P , − V , − d ) let 0 < V < v ˉ 0<V<\bar v 0 < V < v ˉ . Then F x , d ( P , V ) ∈ R x F_{x,d}(P,V)\in R_x F x , d ( P , V ) ∈ R x if and only if x P − d xP-d x P − d lies in J = [ max ( − p ˉ , − x V − v ˉ ) , − x V + v ˉ ] \mathcal{J}=[\max(-\bar p,-xV-\bar v),-xV+\bar v] J = [ max ( − p ˉ , − x V − v ˉ ) , − x V + v ˉ ] . As − x p ˉ − v ˉ > − 6 -x\bar p-\bar v>-6 − x p ˉ − v ˉ > − 6 and x p ˉ + r v ˉ = ( 3 x + 2 ) / r < 6 x\bar p+r\bar v=(3x+2)/r<6 x p ˉ + r v ˉ = ( 3 x + 2 ) / r < 6 , the digit d 1 = ⌊ x P + x V − v ˉ ⌋ d_1=\lfloor xP+xV-\bar v\rfloor d 1 = ⌊ x P + x V − v ˉ ⌋ is legal, and x P − d 1 xP-d_1 x P − d 1 , the largest point of x P − Z xP-\mathbb{Z} x P − Z not above J \mathcal{J} J , is a return if J \mathcal{J} J has length at least one. As 2 v ˉ ≥ 1 2\bar v\ge1 2 v ˉ ≥ 1 , a failure forces J \mathcal{J} J to start at − p ˉ -\bar p − p ˉ and x P − d 1 < − p ˉ xP-d_1<-\bar p x P − d 1 < − p ˉ ; then the digits d 1 − 1 d_1-1 d 1 − 1 and d 1 d_1 d 1 give
z = ( p 0 + u , v ˉ + w ) , z − ( 1 , 1 ) , 0 < w ≤ u < W , (82) z=(p_0+u,\bar v+w),\qquad z-(1,1),\qquad0<w\le u<W, \tag*{(82)} z = ( p 0 + u , v ˉ + w ) , z − ( 1 , 1 ) , 0 < w ≤ u < W , (82)
with u = x P − d 1 + 1 − p 0 < 1 − p ˉ − p 0 = W u=xP-d_1+1-p_0<1-\bar{p}-p_0=W u = x P − d 1 + 1 − p 0 < 1 − p ˉ − p 0 = W and w = x ( V − v ˉ ) + u ∈ ( 0 , u ] w=x(V-\bar{v})+u\in(0,u] w = x ( V − v ˉ ) + u ∈ ( 0 , u ] , positive as x P − d 1 + 1 xP-d_1+1 x P − d 1 + 1 lies above J \mathcal{J} J . Both digits are legal, since d 1 − 1 = x P − p 0 − u ∈ ( − x p ˉ − p 0 − W , x p ˉ − p 0 ) = ( − 4 , ( 3 x + 2 ) / r ) d_1-1=xP-p_0-u\in(-x\bar{p}-p_0-W,x\bar{p}-p_0)=(-4,(3x+2)/r) d 1 − 1 = x P − p 0 − u ∈ ( − x p ˉ − p 0 − W , x p ˉ − p 0 ) = ( − 4 , ( 3 x + 2 ) / r ) . The upper state alone need not return.
Fast return and cone. If u + w ≥ κ u+w\ge\kappa u + w ≥ κ , capture (a) applies to the lower state with ( ξ , ζ ) = ( u , w ) (\xi,\zeta)=(u,w) ( ξ , ζ ) = ( u , w ) , whose horizontal condition holds as 0 < x u < x W 0<xu<xW 0 < xu < x W , and as u + w < 2 W ≤ ( 2 x − 3 ) / x u+w<2W\le(2x-3)/x u + w < 2 W ≤ ( 2 x − 3 ) / x , which is x ≤ 3 x\le3 x ≤ 3 . If u + w < κ u+w<\kappa u + w < κ , then 2 v ˉ < 1.002 2\bar{v}<1.002 2 v ˉ < 1.002 , 2 x − 4 ≥ 1.998 2x-4\ge1.998 2 x − 4 ≥ 1.998 , x 2 ≤ 9 x^2\le9 x 2 ≤ 9 and 5 x − 12 ≤ 3 ≤ 6 v ˉ 5x-12\le3\le6\bar{v} 5 x − 12 ≤ 3 ≤ 6 v ˉ give λ − / 3 ≥ 83 / 2250 > 1 / 28 \lambda_-/3\ge83/2250>1/28 λ − /3 ≥ 83/2250 > 1/28 and κ ≤ 1 / 3 \kappa\le1/3 κ ≤ 1/3 ; so Lemma S2.24 gives a return if u ≥ λ − / 3 u\ge\lambda_-/3 u ≥ λ − /3 .
Excursions. Let 0 < u < λ − / 3 0<u<\lambda_-/3 0 < u < λ − /3 . By (S2.1), m m m digits − 2 -2 − 2 take the upper state to z − 2 + ( ξ , ζ ) z_{-2}+(\xi,\zeta) z − 2 + ( ξ , ζ ) with ( ξ , ζ ) = x m ( u , w + m u ) (\xi,\zeta)=x^m(u,w+mu) ( ξ , ζ ) = x m ( u , w + m u ) ; put γ = ζ + 2 ξ = x m ( w + ( m + 2 ) u ) \gamma=\zeta+2\xi=x^m(w+(m+2)u) γ = ζ + 2 ξ = x m ( w + ( m + 2 ) u ) , which grows without bound. Let m m m be first with γ ≥ λ − \gamma\ge\lambda_- γ ≥ λ − . Then m ≥ 1 m\ge1 m ≥ 1 , as w + 2 u ≤ 3 u < λ − w+2u\le3u<\lambda_- w + 2 u ≤ 3 u < λ − , and γ = x γ ′ + ξ \gamma=x\gamma'+\xi γ = x γ ′ + ξ with γ ′ = ( ξ / x ) ( m + 1 + w / u ) < λ − \gamma'=(\xi/x)(m+1+w/u)<\lambda_- γ ′ = ( ξ / x ) ( m + 1 + w / u ) < λ − the preceding value, so
ξ < x λ − / 3 ≤ ( x − 2 ) / x 2 , λ − ≤ γ < x λ − + ξ ≤ 1 / x + ξ . (83) \xi<x\lambda_-/3\le(x-2)/x^2,\qquad\lambda_-\le\gamma<x\lambda_-+\xi\le1/x+\xi. \tag*{(83)} ξ < x λ − /3 ≤ ( x − 2 ) / x 2 , λ − ≤ γ < x λ − + ξ ≤ 1/ x + ξ . (83)
The first bound is u < λ − / 3 u<\lambda_-/3 u < λ − /3 if m = 1 m=1 m = 1 and follows from γ ′ < λ − \gamma'<\lambda_- γ ′ < λ − if m ≥ 2 m\ge2 m ≥ 2 ; the second is ( x − 3 ) ( 2 x 3 − 5 x 2 + 7 x − 2 ) ≤ 0 (x-3)(2x^3-5x^2+7x-2)\le0 ( x − 3 ) ( 2 x 3 − 5 x 2 + 7 x − 2 ) ≤ 0 , and the last is 2 x − 5 ≤ 2 v ˉ 2x-5\le2\bar{v} 2 x − 5 ≤ 2 v ˉ . If γ ≤ λ + \gamma\le\lambda_+ γ ≤ λ + , capture (b) applies, as 0 < x 2 ξ < x − 2 0<x^2\xi<x-2 0 < x 2 ξ < x − 2 by (S2.34). If γ > λ + \gamma>\lambda_+ γ > λ + and ξ ≥ ξ ∗ \xi\ge\xi_* ξ ≥ ξ ∗ , take j = 2 j=2 j = 2 in (S2.32): P ≥ − p ˉ P\ge-\bar{p} P ≥ − p ˉ is exactly ξ ≥ ξ ∗ \xi\ge\xi_* ξ ≥ ξ ∗ , P ≤ p 0 P\le p_0 P ≤ p 0 , and λ + ≥ ( 2 x − 2 − 2 v ˉ ) / x 2 \lambda_+\ge(2x-2-2\bar{v})/x^2 λ + ≥ ( 2 x − 2 − 2 v ˉ ) / x 2 , γ < 1 / x + ξ ≤ ( 2 x − 2 ) / x 2 \gamma<1/x+\xi\le(2x-2)/x^2 γ < 1/ x + ξ ≤ ( 2 x − 2 ) / x 2 give the vertical bounds. If γ > λ + \gamma>\lambda_+ γ > λ + and ξ < ξ ∗ \xi<\xi_* ξ < ξ ∗ , then j = 3 j=3 j = 3 gives the renewed upper state z − 2 + ( u ′ , w ′ ) z_{-2}+(u',w') z − 2 + ( u ′ , w ′ ) with u ′ = x 2 ξ + δ u'=x^2\xi+\delta u ′ = x 2 ξ + δ , w ′ = x 2 γ + 3 − 2 x w'=x^2\gamma+3-2x w ′ = x 2 γ + 3 − 2 x , and j = 2 j=2 j = 2 the renewed lower state. Here w ′ > 0 w'>0 w ′ > 0 is γ > λ + \gamma>\lambda_+ γ > λ + , w ′ ≤ u ′ w'\le u' w ′ ≤ u ′ is γ ≤ ξ + 1 / x \gamma\le\xi+1/x γ ≤ ξ + 1/ x , and u ′ < x 2 ξ ∗ + δ = W u'<x^2\xi_*+\delta=W u ′ < x 2 ξ ∗ + δ = W ; so (S2.33) holds again, with u ′ ≥ x 2 u u'\ge x^2u u ′ ≥ x 2 u . After finitely many renewals, a fast return, the cone or an excursion returns.
Boundedness. Concatenating returns gives an infinite admissible orbit. Each state is a point of R x R_x R x , a pair state or an excursion state z − 2 + ( ξ ′ , ζ ′ ) z_{-2}+(\xi',\zeta') z − 2 + ( ξ ′ , ζ ′ ) with 0 ≤ ξ ′ < 1 / 9 0\le\xi'<1/9 0 ≤ ξ ′ < 1/9 , 0 ≤ ζ ′ < 1 / 2 0\le\zeta'<1/2 0 ≤ ζ ′ < 1/2 (by (S2.34), as ( x − 2 ) / x 2 ≤ 1 / 9 (x-2)/x^2\le1/9 ( x − 2 ) / x 2 ≤ 1/9 and 1 / x + 1 / 9 < 1 / 2 1/x+1/9<1/2 1/ x + 1/9 < 1/2 ), all of sup norm at most 5 / 2 5/2 5/2 , or the image of one of these under at most three digits; as ∥ F x , d ( z ) ∥ ∞ ≤ 6 ∥ z ∥ ∞ + 6 \lVert F_{x,d}(z)\rVert_\infty\le6\lVert z\rVert_\infty+6 ∥ F x , d ( z ) ∥ ∞ ≤ 6 ∥ z ∥ ∞ + 6 , every state has sup norm at most 6 ( 6 ( 6 ⋅ 5 2 + 6 ) + 6 ) + 6 = 798 6(6(6\cdot\frac{5}{2}+6)+6)+6=798 6 ( 6 ( 6 ⋅ 2 5 + 6 ) + 6 ) + 6 = 798 . □
Entry at every terminal scale.
Theorem S2.25 (Terminal bounded orbits). For every 2999 / 1000 ≤ x < 3 2999/1000\le x<3 2999/1000 ≤ x < 3 , the state ( 1 , 1 ) (1,1) ( 1 , 1 ) has a finite admissible word entering R x R_x R x , and hence an infinite bounded admissible orbit. Consequently ( 1 / 3 , 1000 / 2999 ] ⊂ D 6 (1/3,1000/2999]\subset D_6 ( 1/3 , 1000/2999 ] ⊂ D 6 .
First let 0 < δ ≤ δ 0 : = 1 / ( 5 ⋅ 3 15 ) 0<\delta\le\delta_0:=1/(5\cdot3^{15}) 0 < δ ≤ δ 0 := 1/ ( 5 ⋅ 3 15 ) , and let M M M be the first positive integer with
α = δ x M − 1 ≥ 1 5 . (84) \alpha=\delta x^{M-1}\ge\frac{1}{5}. \tag*{(84)} α = δ x M − 1 ≥ 5 1 . (84)
As δ x 15 < δ 0 3 15 = 1 / 5 \delta x^{15}<\delta_0 3^{15}=1/5 δ x 15 < δ 0 3 15 = 1/5 , M ≥ 17 M\ge17 M ≥ 17 . Minimality gives δ x M − 2 < 1 / 5 \delta x^{M-2}<1/5 δ x M − 2 < 1/5 , so, with x ≥ 2 x\ge2 x ≥ 2 , r ≥ 1 r\ge1 r ≥ 1 and the decrease of M / 2 M M/2^M M / 2 M ,
1 5 ≤ α < x / 5 < 3 / 5 , M δ / r < M / ( 5 x M − 2 ) ≤ 4 5 M / 2 M ≤ 4 5 ⋅ 17 2 17 < 1 1000 . (85) \frac{1}{5}\le\alpha<x/5<3/5,\qquad M\delta/r<M/(5x^{M-2})\le\frac{4}{5}M/2^M\le\frac{4}{5}\cdot\frac{17}{2^{17}}<\frac{1}{1000}. \tag*{(85)} 5 1 ≤ α < x /5 < 3/5 , M δ / r < M / ( 5 x M − 2 ) ≤ 5 4 M / 2 M ≤ 5 4 ⋅ 2 17 17 < 1000 1 . (85)
At q = 3 q=3 q = 3 the blocks of Section S2.2 are A M = 5 ( − 4 ) M − 1 A_M=5(-4)^{M-1} A M = 5 ( − 4 ) M − 1 , B M = 5 ( − 4 ) M − 2 ( − 5 ) B_M=5(-4)^{M-2}(-5) B M = 5 ( − 4 ) M − 2 ( − 5 ) and U M , k = B M ( − 2 ) k U_{M,k}=B_M(-2)^k U M , k = B M ( − 2 ) k ; the B M B_M B M state, z − 2 + ( ξ , ζ ) z_{-2}+(\xi,\zeta) z − 2 + ( ξ , ζ ) say, is the A M A_M A M state plus ( 1 , 1 ) (1,1) ( 1 , 1 ) , and ξ = ξ M \xi=\xi_M ξ = ξ M by (S2.4). As F x , 5 ( 1 , 1 ) = z − 4 + ( δ 2 / r , − δ ( 2 x − 3 x + 3 ) / r 2 ) F_{x,5}(1,1)=z_{-4}+(\delta^2/r,-\delta(2x-3x+3)/r^2) F x , 5 ( 1 , 1 ) = z − 4 + ( δ 2 / r , − δ ( 2 x − 3 x + 3 ) / r 2 ) and z − 4 + ( 1 , 1 ) = ( 1 − 2 / r , 1 + 2 / r ) z_{-4}+(1,1)=(1-2/r,1+2/r) z − 4 + ( 1 , 1 ) = ( 1 − 2/ r , 1 + 2/ r ) , (S2.1) at z − 4 z_{-4} z − 4 for M − 1 M-1 M − 1 steps, the shift of the last digit and δ x M − 1 = α \delta x^{M-1}=\alpha δ x M − 1 = α give
ξ = δ ( α − 1 ) r < 0 , ζ = x ( 1 − 2 α ) r + δ ( ( M − 1 ) α r + 1 − α r 2 ) . (86) \xi=\frac{\delta(\alpha-1)}{r}<0,\qquad\zeta=\frac{x(1-2\alpha)}{r}+\delta\left(\frac{(M-1)\alpha}{r}+\frac{1-\alpha}{r^2}\right). \tag*{(86)} ξ = r δ ( α − 1 ) < 0 , ζ = r x ( 1 − 2 α ) + δ ( r ( M − 1 ) α + r 2 1 − α ) . (86)
By (S2.1), the U M , k U_{M,k} U M , k state is z − 2 + x k ( ξ , ζ + k ξ ) = : z − 2 + ( ξ k , ζ k ) z_{-2}+x^k(\xi,\zeta+k\xi)=:z_{-2}+(\xi_k,\zeta_k) z − 2 + x k ( ξ , ζ + k ξ ) =: z − 2 + ( ξ k , ζ k ) .
Lemma S2.26 (The descending half of the first phase). *If α ≥ 1 / 2 \alpha\ge1/2 α ≥ 1/2 , an admissible continuation of B M B_M B M enters R x R_x R x after at most M M M further digits − 2 -2 − 2 .
Proof. Idea. Along the − 2 -2 − 2 excursion the vertical displacement turns negative within M M M steps, while the horizontal one stays small. By (S2.37), ζ + M ξ = x ( 1 − 2 α ) / r + δ ( ( ( 2 M − 1 ) α − M ) / r + ( 1 − α ) / r 2 ) \zeta+M\xi=x(1-2\alpha)/r+\delta(((2M-1)\alpha-M)/r+(1-\alpha)/r^2) ζ + M ξ = x ( 1 − 2 α ) / r + δ ((( 2 M − 1 ) α − M ) / r + ( 1 − α ) / r 2 ) . As a function of α \alpha α this has derivative [ − 2 x + δ ( 2 M − 1 ) − δ ] / r < 0 [-2x+\delta(2M-1)-\delta]/r<0 [ − 2 x + δ ( 2 M − 1 ) − δ ] / r < 0 , as M δ < r / 1000 < 1 M\delta<r/1000<1 M δ < r /1000 < 1 by (S2.36), and value δ ( 2 − x ) / ( 2 r 2 ) < 0 \delta(2-x)/(2r^2)<0 δ ( 2 − x ) / ( 2 r 2 ) < 0 at α = 1 / 2 \alpha=1/2 α = 1/2 ; so ζ M < 0 \zeta_M<0 ζ M < 0 . If ζ ≤ 0 \zeta\le0 ζ ≤ 0 , B M B_M B M already enters R x R_x R x : ζ > − x / ( 5 r ) ≥ − 2 v ˉ \zeta>-x/(5r)\ge-2\bar{v} ζ > − x / ( 5 r ) ≥ − 2 v ˉ as α < 3 / 5 \alpha<3/5 α < 3/5 , and ∣ ξ ∣ ≤ δ / ( 2 r ) |\xi|\le\delta/(2r) ∣ ξ ∣ ≤ δ / ( 2 r ) gives ∣ P ∣ < p ˉ |P|<\bar{p} ∣ P ∣ < p ˉ . Otherwise let k ≤ M k\le M k ≤ M be first with ζ k ≤ 0 \zeta_k\le0 ζ k ≤ 0 . Then ζ k − ξ k = x ζ k − 1 > 0 \zeta_k-\xi_k=x\zeta_{k-1}>0 ζ k − ξ k = x ζ k − 1 > 0 , so ξ k < ζ k ≤ 0 \xi_k<\zeta_k\le0 ξ k < ζ k ≤ 0 , and ∣ ξ k ∣ ≤ x M δ ( 1 − α ) / r = x α ( 1 − α ) / r ≤ x / ( 4 r ) |\xi_k|\le x^M\delta(1-\alpha)/r=x\alpha(1-\alpha)/r\le x/(4r) ∣ ξ k ∣ ≤ x M δ ( 1 − α ) / r = xα ( 1 − α ) / r ≤ x / ( 4 r ) . Thus − p ˉ ≤ − ( 2 + x / 4 ) / r ≤ P ≤ p 0 -\bar{p}\le-(2+x/4)/r\le P\le p_0 − p ˉ ≤ − ( 2 + x /4 ) / r ≤ P ≤ p 0 and − v ˉ ≤ v ˉ − x / ( 4 r ) ≤ V ≤ v ˉ -\bar{v}\le\bar{v}-x/(4r)\le V\le\bar{v} − v ˉ ≤ v ˉ − x / ( 4 r ) ≤ V ≤ v ˉ , by x < 4 x<4 x < 4 and x r ≤ 6 < 16 xr\le6<16 x r ≤ 6 < 16 .
Suppose henceforth that α ≤ 1 / 2 \alpha\le1/2 α ≤ 1/2 , and set
Θ = x M ( 2 α − 1 ) ≤ 0 , γ k = ζ k + 2 ξ k = x k ( ζ + ( k + 2 ) ξ ) . (87) \Theta=x^M(2\alpha-1)\le0,\qquad\gamma_k=\zeta_k+2\xi_k=x^k(\zeta+(k+2)\xi). \tag*{(87)} Θ = x M ( 2 α − 1 ) ≤ 0 , γ k = ζ k + 2 ξ k = x k ( ζ + ( k + 2 ) ξ ) . (87)
If k ≤ M − 2 k\le M-2 k ≤ M − 2 and λ − ≤ γ k ≤ λ + \lambda_-\le\gamma_k\le\lambda_+ λ − ≤ γ k ≤ λ + , capture (b) applies to the U M , k U_{M,k} U M , k state, since 0 > x 2 ξ k ≥ x M ξ = − x α ( 1 − α ) / r ≥ − x / ( 4 r ) 0>x^2\xi_k\ge x^M\xi=-x\alpha(1-\alpha)/r\ge-x/(4r) 0 > x 2 ξ k ≥ x M ξ = − xα ( 1 − α ) / r ≥ − x / ( 4 r ) . Also
( x λ − − λ + ) x 2 r 2 = ( x − 3 ) ( 2 x 3 − 4 x 2 + 5 x − 1 ) ≤ 0 , so κ = x λ − ≤ λ + . (88) (x\lambda_- - \lambda_+)x^2r^2=(x-3)(2x^3-4x^2+5x-1)\le0,\qquad\text{so }\kappa=x\lambda_-\le\lambda_+. \tag*{(88)} ( x λ − − λ + ) x 2 r 2 = ( x − 3 ) ( 2 x 3 − 4 x 2 + 5 x − 1 ) ≤ 0 , so κ = x λ − ≤ λ + . (88)
Lemma S2.27 (The outer part of the second phase). Suppose α ≤ 1 / 2 \alpha\le1/2 α ≤ 1/2 . There is finite entry into R x R_x R x if either γ 0 > λ + \gamma_0>\lambda_+ γ 0 > λ + , or γ 0 ≤ λ + \gamma_0\le\lambda_+ γ 0 ≤ λ + and Θ ≤ − 4 / 5 \Theta\le-4/5 Θ ≤ − 4/5 .
Proof. Idea. Either A M A_M A M is one digit from R x R_x R x , or the − 2 -2 − 2 excursion meets [ λ − , λ + ] [\lambda_-,\lambda_+] [ λ − , λ + ] by step M − 2 M-2 M − 2 . If γ 0 > λ + \gamma_0>\lambda_+ γ 0 > λ + , then β 0 = ζ + ξ = γ 0 − ξ > λ + ≥ κ \beta_0=\zeta+\xi=\gamma_0-\xi>\lambda_+\ge\kappa β 0 = ζ + ξ = γ 0 − ξ > λ + ≥ κ by (S2.39), and by (S2.36)–(S2.37) and v ˉ < 0.501 \bar{v}<0.501 v ˉ < 0.501 ,
β 0 = 1 + v ˉ − 3 α + δ ( M α − 1 ) / r − α δ x / r 2 ≤ 2 5 + v ˉ + M δ / ( 2 r ) < 1803 2000 < 99 100 < 2 x − 3 x . \beta_0=1+\bar{v}-3\alpha+\delta(M\alpha-1)/r-\alpha\delta x/r^2\le\frac25+\bar{v}+M\delta/(2r)<\frac{1803}{2000}<\frac{99}{100}<\frac{2x-3}{x}. β 0 = 1 + v ˉ − 3 α + δ ( M α − 1 ) / r − α δ x / r 2 ≤ 5 2 + v ˉ + M δ / ( 2 r ) < 2000 1803 < 100 99 < x 2 x − 3 . As 0 > x ξ ≥ − δ − δ / r ≥ − δ − 1 / r 0>x\xi\ge-\delta-\delta/r\ge-\delta-1/r 0 > x ξ ≥ − δ − δ / r ≥ − δ − 1/ r , capture (a) applies to the A M A_M A M state z − 2 + ( ξ − 1 , ζ − 1 ) z_{-2}+(\xi-1,\zeta-1) z − 2 + ( ξ − 1 , ζ − 1 ) .
Now let γ 0 ≤ λ + \gamma_0\le\lambda_+ γ 0 ≤ λ + and Θ ≤ − 4 / 5 \Theta\le-4/5 Θ ≤ − 4/5 . Substituting x M − 2 δ = α / x x^{M-2}\delta=\alpha/x x M − 2 δ = α / x , x M − 1 ( 1 − 2 α ) = − Θ / x x^{M-1}(1-2\alpha)=-\Theta/x x M − 1 ( 1 − 2 α ) = − Θ/ x and x α ( 2 α − 1 ) = δ Θ x\alpha(2\alpha-1)=\delta\Theta xα ( 2 α − 1 ) = δ Θ , and using x / r = 3 / 2 + δ / ( 2 r ) x/r=3/2+\delta/(2r) x / r = 3/2 + δ / ( 2 r ) , into γ M − 2 = x M − 2 ( ζ + M ξ ) \gamma_{M-2}=x^{M-2}(\zeta+M\xi) γ M − 2 = x M − 2 ( ζ + M ξ ) gives
γ M − 2 = α ( 1 − α x ) x r 2 − Θ x 2 ( 3 2 − δ ( M − 1 2 ) r ) ≥ − 21 1000 + 4 5 ⋅ 1499 1000 ⋅ 9 = 5051 45000 > 1 9 ≥ λ − . (89) \gamma_{M-2}=\frac{\alpha(1-\alpha x)}{xr^2}-\frac{\Theta}{x^2}\left(\frac32-\frac{\delta(M-\frac12)}{r}\right)\ge-\frac{21}{1000}+\frac45\cdot\frac{1499}{1000\cdot9}=\frac{5051}{45000}>\frac19\ge\lambda_-. \tag*{(89)} γ M − 2 = x r 2 α ( 1 − α x ) − x 2 Θ ( 2 3 − r δ ( M − 2 1 ) ) ≥ − 1000 21 + 5 4 ⋅ 1000 ⋅ 9 1499 = 45000 5051 > 9 1 ≥ λ − . (89) Here the first term decreases for α ≥ 1 / 5 > 1 / ( 2 x ) \alpha\ge1/5>1/(2x) α ≥ 1/5 > 1/ ( 2 x ) , so it is at least its value ( 2 − x ) / ( 4 x r 2 ) > − 21 / 1000 (2-x)/(4xr^2)>-21/1000 ( 2 − x ) / ( 4 x r 2 ) > − 21/1000 at α = 1 / 2 \alpha=1/2 α = 1/2 ; the second is bounded by − Θ ≥ 4 / 5 -\Theta\ge4/5 − Θ ≥ 4/5 , (S2.36) and x 2 ≤ 9 x^2\le9 x 2 ≤ 9 ; and λ − ≤ ( 2 x − 5 ) / x 2 ≤ 1 / 9 \lambda_-\le(2x-5)/x^2\le1/9 λ − ≤ ( 2 x − 5 ) / x 2 ≤ 1/9 as x 2 − 18 x + 45 = ( x − 3 ) ( x − 15 ) ≥ 0 x^2-18x+45=(x-3)(x-15)\ge0 x 2 − 18 x + 45 = ( x − 3 ) ( x − 15 ) ≥ 0 . If γ 0 ≥ λ − \gamma_0\ge\lambda_- γ 0 ≥ λ − , capture (b) applies with k = 0 k=0 k = 0 . Otherwise, for the first k ≤ M − 2 k\le M-2 k ≤ M − 2 with γ k ≥ λ − \gamma_k\ge\lambda_- γ k ≥ λ − , ξ k − 1 < 0 \xi_{k-1}<0 ξ k − 1 < 0 and (S2.39) give γ k = x ( γ k − 1 + ξ k − 1 ) < x λ − ≤ λ + \gamma_k=x(\gamma_{k-1}+\xi_{k-1})<x\lambda_-\le\lambda_+ γ k = x ( γ k − 1 + ξ k − 1 ) < x λ − ≤ λ + , and capture (b) applies. This is uniform even when − Θ -\Theta − Θ is unbounded. □
The central second phase
Only − 4 / 5 ≤ Θ ≤ 0 -4/5\le\Theta\le0 − 4/5 ≤ Θ ≤ 0 remains. For ℓ = 0 , 1 \ell=0,1 ℓ = 0 , 1 let P M = U M , M P_M=U_{M,M} P M = U M , M , C M = U M , M − 1 ( − 1 , − 6 ) C_M=U_{M,M-1}(-1,-6) C M = U M , M − 1 ( − 1 , − 6 ) and R M , ℓ = U M , ℓ ( − 1 , − 6 , − 3 ) 6 M − ℓ − 2 R_{M,\ell}=U_{M,\ell}(-1,-6,-3)6^{M-\ell-2} R M , ℓ = U M , ℓ ( − 1 , − 6 , − 3 ) 6 M − ℓ − 2 , an excursion around z 6 = ( 6 / r , − 6 / r 2 ) z_6=(6/r,-6/r^2) z 6 = ( 6/ r , − 6/ r 2 ) . With G ℓ = ( x + 1 ) / ( r x ℓ + 1 ) G_\ell=(x+1)/(rx^{\ell+1}) G ℓ = ( x + 1 ) / ( r x ℓ + 1 ) , their exact coordinates are
P P = − 2 + x α ( α − 1 ) r , V P = 2 + x α ( 1 − α ) r 2 − x α 2 r − ( x − M δ ) Θ r , (90) P_P=\frac{-2+x\alpha(\alpha-1)}{r},\qquad V_P=\frac{2+x\alpha(1-\alpha)}{r^2}-\frac{x\alpha^2}{r}-\frac{(x-M\delta)\Theta}{r}, \tag*{(90)} P P = r − 2 + xα ( α − 1 ) , V P = r 2 2 + xα ( 1 − α ) − r x α 2 − r ( x − M δ ) Θ , (90)
P R ℓ = 6 + x 2 α ( α − 1 ) r − δ α G ℓ , (91) P_{R_\ell}=\frac{6+x^2\alpha(\alpha-1)}{r}-\delta\alpha G_\ell, \tag*{(91)} P R ℓ = r 6 + x 2 α ( α − 1 ) − δ α G ℓ , (91)
V R ℓ = − 6 r 2 + x 2 α ( α − 1 ) ( 2 x − 3 ) r 2 − x 2 Θ r + 2 x α G ℓ V_{R_\ell}=-\frac{6}{r^2}+\frac{x^2\alpha(\alpha-1)(2x-3)}{r^2}-\frac{x^2\Theta}{r}+2x\alpha G_\ell V R ℓ = − r 2 6 + r 2 x 2 α ( α − 1 ) ( 2 x − 3 ) − r x 2 Θ + 2 xα G ℓ
+ δ [ ( M − 1 ) x Θ / r + α G ℓ ( ℓ + 1 − M + x / r − x / ( x + 1 ) ) ] , (92) \qquad+\delta\left[(M-1)x\Theta/r+\alpha G_\ell(\ell+1-M+x/r-x/(x+1))\right], \tag*{(92)} + δ [ ( M − 1 ) x Θ/ r + α G ℓ ( ℓ + 1 − M + x / r − x / ( x + 1 )) ] , (92)
P C = x P P − x + 6 , V C = x ( V P + P P ) − 2 x + 6. (93) P_C=xP_P-x+6,\qquad V_C=x(V_P+P_P)-2x+6. \tag*{(93)} P C = x P P − x + 6 , V C = x ( V P + P P ) − 2 x + 6. (93)
For P M P_M P M take k = M k=M k = M above, with x M ξ = x α ( α − 1 ) / r x^M\xi=x\alpha(\alpha-1)/r x M ξ = xα ( α − 1 ) / r and x α ( 2 α − 1 ) = δ Θ x\alpha(2\alpha-1)=\delta\Theta xα ( 2 α − 1 ) = δ Θ . For R M , ℓ R_{M,\ell} R M , ℓ , the map F x , ( − 1 , − 6 , − 3 ) ( P , V ) = ( x 3 P + x 2 + 6 x + 3 , x 3 ( V + 3 P ) + 3 x 2 + 12 x + 3 ) F_{x,(-1,-6,-3)}(P,V)=(x^3P+x^2+6x+3,\ x^3(V+3P)+3x^2+12x+3) F x , ( − 1 , − 6 , − 3 ) ( P , V ) = ( x 3 P + x 2 + 6 x + 3 , x 3 ( V + 3 P ) + 3 x 2 + 12 x + 3 ) sends z − 2 z_{-2} z − 2 to z 6 + ( − δ 2 ( x + 1 ) / r , δ ( 3 x 3 − 5 x 2 + 3 x + 3 ) / r 2 ) z_6+(-\delta^2(x+1)/r,\ \delta(3x^3-5x^2+3x+3)/r^2) z 6 + ( − δ 2 ( x + 1 ) / r , δ ( 3 x 3 − 5 x 2 + 3 x + 3 ) / r 2 ) ; apply it to z − 2 + ( ξ ℓ , ζ ℓ ) z_{-2}+(\xi_\ell,\zeta_\ell) z − 2 + ( ξ ℓ , ζ ℓ ) and then (S2.1) at z 6 z_6 z 6 for M − ℓ − 2 M-\ell-2 M − ℓ − 2 steps, using x M − 1 δ = α x^{M-1}\delta=\alpha x M − 1 δ = α and x α ( 2 α − 1 ) = δ Θ x\alpha(2\alpha-1)=\delta\Theta xα ( 2 α − 1 ) = δ Θ . Finally, C M C_M C M is P M P_M P M with its last digit raised to − 1 -1 − 1 , which subtracts ( 1 , 1 ) (1,1) ( 1 , 1 ) , followed by − 6 -6 − 6 . At x = 3 x=3 x = 3 , α = 1 / 2 \alpha=1/2 α = 1/2 , keeping Θ \Theta Θ , the limiting states are
P : ( − 11 8 , 5 6 − 3 2 Θ ) , R ℓ : ( 15 8 , 2 3 − 51 16 − 9 2 Θ ) , C : ( − 9 8 , − 51 16 − 9 2 Θ ) . (94) P:\ \left(-\frac{11}{8},\frac56-\frac32\Theta\right),\qquad R_\ell:\ \left(\frac{15}{8},\frac23-\frac{51}{16}-\frac92\Theta\right),\qquad C:\ \left(-\frac98,-\frac{51}{16}-\frac92\Theta\right). \tag*{(94)} P : ( − 8 11 , 6 5 − 2 3 Θ ) , R ℓ : ( 8 15 , 3 2 − 16 51 − 2 9 Θ ) , C : ( − 8 9 , − 16 51 − 2 9 Θ ) . (94)
Lemma S2.28 (Eighteen central entry cells). For every scale and parameter in (S2.35) with − 4 / 5 ≤ Θ ≤ 0 -4/5\le\Theta\le0 − 4/5 ≤ Θ ≤ 0 , one of the words in Table 5 sends ( 1 , 1 ) (1,1) ( 1 , 1 ) strictly into R 0 = [ − 3 / 2 , 3 / 2 ] × [ − 1 / 2 , 1 / 2 ] ⊂ R x R_0=[-3/2,3/2]\times[-1/2,1/2]\subset R_x R 0 = [ − 3/2 , 3/2 ] × [ − 1/2 , 1/2 ] ⊂ R x .
Proof. Idea. At x = 3 x=3 x = 3 , α = 1 / 2 \alpha=1/2 α = 1/2 the images are affine in Θ \Theta Θ and lie in R 0 R_0 R 0 with margins; the actual images differ by O ( M δ ) O(M\delta) O ( M δ ) . By Proposition S4.7(iii), with suffixes of length at most four, every limiting image lies in [ − 11 / 8 , 11 / 8 ] × [ − 39 / 80 , 39 / 80 ] [-11/8,11/8]\times[-39/80,39/80] [ − 11/8 , 11/8 ] × [ − 39/80 , 39/80 ] , so with margins 1 / 8 1/8 1/8 and 1 / 80 1/80 1/80 in R 0 R_0 R 0 , and every limiting partial state, the base included, lies in [ − 3 , 3 ] × [ − 3 / 2 , 3 / 2 ] [-3,3]\times[-3/2,3/2] [ − 3 , 3 ] × [ − 3/2 , 3/2 ] .
Table 5. Central cells: a closed partition of [ − 4 / 5 , 0 ] [-4/5,0] [ − 4/5 , 0 ] in Θ \Theta Θ .
To transfer this uniformly in M M M , note that, by α ≤ 1 2 \alpha\le\frac{1}{2} α ≤ 2 1 , ∣ Θ ∣ ≤ 4 5 |\Theta| \le\frac{4}{5} ∣Θ∣ ≤ 5 4 and x M ≥ ( 2999 / 1000 ) 17 x^M \ge(2999/1000)^{17} x M ≥ ( 2999/1000 ) 17 , ∣ α − 1 2 ∣ = ∣ Θ ∣ / ( 2 x M ) |\alpha-\frac{1}{2}|=|\Theta|/(2x^M) ∣ α − 2 1 ∣ = ∣Θ∣/ ( 2 x M ) gives α ≥ 49 100 \alpha\ge\frac{49}{100} α ≥ 100 49 , and then ∣ α − 1 2 ∣ = ∣ Θ ∣ δ / ( 2 α x ) < δ / 3 |\alpha-\frac{1}{2}|=|\Theta|\delta/(2\alpha x)<\delta/3 ∣ α − 2 1 ∣ = ∣Θ∣ δ / ( 2 α x ) < δ /3 . Write each base coordinate as f ( x , α , Θ ) + δ g f(x,\alpha,\Theta)+\delta g f ( x , α , Θ ) + δ g , where δ g \delta g δ g is the displayed term with the factor δ \delta δ : none in P P P_P P P , M δ Θ / r M\delta\Theta/r M δ Θ/ r in V P V_P V P , − δ α G ℓ -\delta\alpha G_\ell − δ α G ℓ in P R ℓ P_{R_\ell} P R ℓ , the bracket in V R ℓ V_{R_\ell} V R ℓ ; treat x , α , Θ x,\alpha,\Theta x , α , Θ as independent. On the box 2999 / 1000 ≤ x ≤ 3 2999/1000\le x\le3 2999/1000 ≤ x ≤ 3 , 49 100 ≤ α ≤ 1 2 \frac{49}{100}\le\alpha\le\frac{1}{2} 100 49 ≤ α ≤ 2 1 we have 1.999 ≤ r ≤ 2 1.999\le r\le2 1.999 ≤ r ≤ 2 , x ( x − 2 ) ≤ 3 x(x-2)\le3 x ( x − 2 ) ≤ 3 , α ( 1 − α ) ≤ 1 4 \alpha(1-\alpha)\le\frac14 α ( 1 − α ) ≤ 4 1 , ∣ 2 α − 1 ∣ ≤ 1 50 |2\alpha-1|\le\frac{1}{50} ∣2 α − 1∣ ≤ 50 1 , and ∣ Θ ∣ ≤ 4 5 |\Theta|\le\frac45 ∣Θ∣ ≤ 5 4 , or ∣ Θ ∣ ≤ 3 5 |\Theta|\le\frac35 ∣Θ∣ ≤ 5 3 for the R R R words. Hence
∣ ∂ x P P ∣ = ( 2 + α ( 1 − α ) ) / r 2 < 1 , ∣ ∂ α P P ∣ = x ∣ 2 α − 1 ∣ / r < 1 , ∣ ∂ x V P ∣ = ∣ ( α + Θ ) / r 2 − 2 ( 2 + x α ( 1 − α ) ) / r 3 ∣ < 2 , ∣ ∂ α V P ∣ = ∣ x ( 1 − 2 α ) / r 2 − 2 x α / r ∣ < 2 , ∣ ∂ x P R ℓ ∣ = ( 6 + α ( 1 − α ) x ( x − 2 ) ) / r 2 < 2 , ∣ ∂ α P R ℓ ∣ = x 2 ∣ 2 α − 1 ∣ / r < 1 , ∣ ∂ x V R ℓ ∣ ≤ 12 / r 3 + 3 4 + 9 5 / r 2 + 1 < 5 , ∣ ∂ α V R ℓ ∣ ≤ 7 50 + 8 / r < 6. \begin{aligned}
|\partial_xP_P|&=(2+\alpha(1-\alpha))/r^2<1, &
|\partial_\alpha P_P|&=x|2\alpha-1|/r<1,\\
|\partial_xV_P|&=|(\alpha+\Theta)/r^2-2(2+x\alpha(1-\alpha))/r^3|<2, &
|\partial_\alpha V_P|&=|x(1-2\alpha)/r^2-2x\alpha/r|<2,\\
|\partial_xP_{R_\ell}|&=(6+\alpha(1-\alpha)x(x-2))/r^2<2, &
|\partial_\alpha P_{R_\ell}|&=x^2|2\alpha-1|/r<1,\\
|\partial_xV_{R_\ell}|&\le12/r^3+\frac34+\frac95/r^2+1<5, &
|\partial_\alpha V_{R_\ell}|&\le\frac{7}{50}+8/r<6.
\end{aligned} ∣ ∂ x P P ∣ ∣ ∂ x V P ∣ ∣ ∂ x P R ℓ ∣ ∣ ∂ x V R ℓ ∣ = ( 2 + α ( 1 − α )) / r 2 < 1 , = ∣ ( α + Θ ) / r 2 − 2 ( 2 + xα ( 1 − α )) / r 3 ∣ < 2 , = ( 6 + α ( 1 − α ) x ( x − 2 )) / r 2 < 2 , ≤ 12/ r 3 + 4 3 + 5 9 / r 2 + 1 < 5 , ∣ ∂ α P P ∣ ∣ ∂ α V P ∣ ∣ ∂ α P R ℓ ∣ ∣ ∂ α V R ℓ ∣ = x ∣2 α − 1∣/ r < 1 , = ∣ x ( 1 − 2 α ) / r 2 − 2 xα / r ∣ < 2 , = x 2 ∣2 α − 1∣/ r < 1 , ≤ 50 7 + 8/ r < 6.
The last row uses ∂ x V R ℓ = 12 / r 3 + α ( α − 1 ) Q ′ − Θ x ( x − 2 ) / r 2 + 2 α J ℓ ′ \partial_xV_{R_\ell}=12/r^3+\alpha(\alpha-1)Q'-\Theta x(x-2)/r^2+2\alpha J_\ell' ∂ x V R ℓ = 12/ r 3 + α ( α − 1 ) Q ′ − Θ x ( x − 2 ) / r 2 + 2 α J ℓ ′ and ∂ α V R ℓ = ( 2 α − 1 ) Q + 2 J ℓ \partial_\alpha V_{R_\ell}=(2\alpha-1)Q+2J_\ell ∂ α V R ℓ = ( 2 α − 1 ) Q + 2 J ℓ , where Q = x 2 ( 2 x − 3 ) / r 2 ∈ ( 0 , 7 ] Q=x^2(2x-3)/r^2\in(0,7] Q = x 2 ( 2 x − 3 ) / r 2 ∈ ( 0 , 7 ] , Q ′ = 2 x ( x 2 − 3 x + 3 ) / r 3 ∈ ( 0 , 3 ] Q'=2x(x^2-3x+3)/r^3\in(0,3] Q ′ = 2 x ( x 2 − 3 x + 3 ) / r 3 ∈ ( 0 , 3 ] , J ℓ = x G ℓ ∈ ( 0 , 4 / r ] J_\ell=xG_\ell\in(0,4/r] J ℓ = x G ℓ ∈ ( 0 , 4/ r ] and ∣ J ℓ ′ ∣ ≤ 1 |J_\ell'|\le1 ∣ J ℓ ′ ∣ ≤ 1 . Also ∣ g ∣ ≤ M / 2 |g|\le M/2 ∣ g ∣ ≤ M /2 , 3 2 + 3 2 M \frac{3}{2}+\frac{3}{2}M 2 3 + 2 3 M for V P V_P V P , P R ℓ P_{R_\ell} P R ℓ , V R ℓ V_{R_\ell} V R ℓ , by ∣ Θ ∣ / r ≤ 1 2 |\Theta|/r\le\frac12 ∣Θ∣/ r ≤ 2 1 , α G ℓ ≤ 1 2 \alpha G_\ell\le\frac12 α G ℓ ≤ 2 1 , x ∣ Θ ∣ / r < 1 x|\Theta|/r<1 x ∣Θ∣/ r < 1 for R R R , and − M ≤ ℓ + 1 − M + x / r − x / ( x + 1 ) ≤ 0 -M\le\ell+1-M+x/r-x/(x+1)\le0 − M ≤ ℓ + 1 − M + x / r − x / ( x + 1 ) ≤ 0 . By the mean value theorem, a non-bridge base coordinate is within ( 5 + 6 3 + 3 2 M + 3 ) δ ≤ ϵ M : = ( 2 M + 10 ) δ (5+\frac{6}{3}+\frac{3}{2}M+3)\delta\le\epsilon_M:=(2M+10)\delta ( 5 + 3 6 + 2 3 M + 3 ) δ ≤ ϵ M := ( 2 M + 10 ) δ of its limit, the V R ℓ V_{R_\ell} V R ℓ row being the worst; for P P P the errors are e P ≤ 4 3 δ e_P\le\frac43\delta e P ≤ 3 4 δ and e V ≤ ( 3 8 + M 2 ) δ e_V\le(\frac38+\frac{M}{2})\delta e V ≤ ( 8 3 + 2 M ) δ , also when ∣ Θ ∣ ≤ 4 5 |\Theta|\le\frac45 ∣Θ∣ ≤ 5 4 . By (S2.44), P C P_C P C and V C V_C V C differ from their limits by x ( P P + 11 8 ) + 19 8 δ x(P_P+\frac{11}{8})+\frac{19}{8}\delta x ( P P + 8 11 ) + 8 19 δ and x ( V P + P P − 5 16 + 3 2 Θ + 11 8 ) + δ ( 49 16 + 3 2 Θ ) x(V_P+P_P-\frac{5}{16}+\frac32\Theta+\frac{11}{8})+\delta(\frac{49}{16}+\frac32\Theta) x ( V P + P P − 16 5 + 2 3 Θ + 8 11 ) + δ ( 16 49 + 2 3 Θ ) , so by at most 3 e P + 19 8 δ 3e_P+\frac{19}{8}\delta 3 e P + 8 19 δ and 3 ( e P + e V ) + 49 16 δ 3(e_P+e_V)+\frac{49}{16}\delta 3 ( e P + e V ) + 16 49 δ , both below ( 2 M + 25 ) δ (2M+25)\delta ( 2 M + 25 ) δ .
For a nonempty suffix of length k ≤ 4 k\le4 k ≤ 4 , apply (S1.11) with J = ( 1 1 1 0 ) J=\begin{pmatrix}1&1\\1&0\end{pmatrix} J = ( 1 1 1 0 ) and q = 3 q=3 q = 3 . Here ∥ J j ∥ ∞ ≤ j + 1 \|J^j\|_\infty\le j+1 ∥ J j ∥ ∞ ≤ j + 1 , while each limiting partial state has ∥ J ( P , V ) ∥ ∞ = max { ∣ P ∣ , ∣ P + V ∣ } ≤ 9 / 2 \|J(P,V)\|_\infty=\max\{|P|,|P+V|\}\le9/2 ∥ J ( P , V ) ∥ ∞ = max { ∣ P ∣ , ∣ P + V ∣ } ≤ 9/2 . Thus the propagated base error is at most 405 ϵ M 405\epsilon_M 405 ϵ M , and the parameter error is at most 9 2 ∑ j = 0 3 3 j ( j + 1 ) δ = 639 δ < 1166 δ \frac92\sum_{j=0}^{3}3^j(j+1)\delta=639\delta<1166\delta 2 9 ∑ j = 0 3 3 j ( j + 1 ) δ = 639 δ < 1166 δ . The bridge cell has an empty suffix. Keeping the common allowance, every final coordinate is therefore within 405 ( 2 M + 10 ) δ + 1166 δ = ( 810 M + 5216 ) δ 405(2M+10)\delta+1166\delta=(810M+5216)\delta 405 ( 2 M + 10 ) δ + 1166 δ = ( 810 M + 5216 ) δ of its limit. As α ≤ 1 2 \alpha\le\frac12 α ≤ 2 1 , δ ≤ 1 / ( 2 x M − 1 ) ≤ δ M : = 1 2 ( 2999 / 1000 ) 1 − M \delta\le1/(2x^{M-1})\le\delta_M:=\frac12(2999/1000)^{1-M} δ ≤ 1/ ( 2 x M − 1 ) ≤ δ M := 2 1 ( 2999/1000 ) 1 − M , and ( 810 M + 5216 ) δ M (810M+5216)\delta_M ( 810 M + 5216 ) δ M decreases in M M M , with ratio 810 M + 6026 810 M + 5216 ⋅ 1000 2999 < 6026 5216 ⋅ 1000 2999 < 1 \frac{810M+6026}{810M+5216}\cdot\frac{1000}{2999}<\frac{6026}{5216}\cdot\frac{1000}{2999}<1 810 M + 5216 810 M + 6026 ⋅ 2999 1000 < 5216 6026 ⋅ 2999 1000 < 1 , from 18986 δ 17 < 2.3 ⋅ 10 − 4 < 1 80 18986\delta_{17}<2.3\cdot10^{-4}<\frac1{80} 18986 δ 17 < 2.3 ⋅ 1 0 − 4 < 80 1 . So all eighteen images lie strictly inside R 0 R_0 R 0 , and R 0 ⊂ R x R_0\subset R_x R 0 ⊂ R x as p ˉ ≥ 3 2 \bar p\ge\frac32 p ˉ ≥ 2 3 and v ˉ ≥ 1 2 \bar v\ge\frac12 v ˉ ≥ 2 1 .
The finite join.
Lemma S2.29 (Compact entry into the moving rectangle). For every
2999 1000 ≤ x ≤ 3 − 1 24 ⋅ 3 14 = 344373767 114791256 \frac{2999}{1000}\le x\le3-\frac{1}{24\cdot3^{14}}=\frac{344373767}{114791256} 1000 2999 ≤ x ≤ 3 − 24 ⋅ 3 14 1 = 114791256 344373767 a finite admissible word sends ( 1 , 1 ) (1,1) ( 1 , 1 ) strictly into R x R_x R x .
Proof. This is part (iv) of Proposition S4.7.
Proof of Theorem S2.25. For 0 < 3 − x ≤ δ 0 0<3-x\le\delta_0 0 < 3 − x ≤ δ 0 choose M M M by (S2.35). Lemma S2.26 treats α ≥ 1 / 2 \alpha\ge1/2 α ≥ 1/2 , Lemma S2.27 treats α ≤ 1 / 2 \alpha\le1/2 α ≤ 1/2 with γ 0 > λ + \gamma_0>\lambda_+ γ 0 > λ + or Θ ≤ − 4 / 5 \Theta\le-4/5 Θ ≤ − 4/5 , and Lemma S2.28 treats − 4 / 5 ≤ Θ ≤ 0 -4/5\le\Theta\le0 − 4/5 ≤ Θ ≤ 0 ; so ( 1 , 1 ) (1,1) ( 1 , 1 ) enters R x R_x R x . Lemma S2.29 supplies the remaining parameters, with overlap since 1 / ( 24 ⋅ 3 14 ) = 5 8 δ 0 < δ 0 1/(24\cdot3^{14})=\frac58\delta_0<\delta_0 1/ ( 24 ⋅ 3 14 ) = 8 5 δ 0 < δ 0 . After entry, Theorem S2.23 continues the orbit; with its finite prefix it is infinite and bounded. By Lemma S1.7, 1 / x ∈ D 6 1/x\in D_6 1/ x ∈ D 6 .
Height two
This subsection proves the inclusion [ 1 / ρ , 1 ) ⊆ A 2 [1/\rho,1) \subseteq A_2 [ 1/ ρ , 1 ) ⊆ A 2 of Theorem S2.7; the exclusion of the other double zeros is proved in § 15.4 . For x ∈ [ 433 / 250 , ρ ) x \in[433/250,\rho) x ∈ [ 433/250 , ρ ) we build a bounded real orbit and apply Lemma S1.7. The orbit enters a fixed diamond Q Q Q along the periodic digit block of the tribonacci witness f ρ f_\rho f ρ of (S2.7), followed by a short suffix, and a finite return cover then brings it back to Q Q Q again and again.
Trace states and the return diamond
Fix x > 1 x>1 x > 1 . A state is a pair Y = ( S , V ) ∈ R 2 Y=(S,V)\in\mathbb{R}^2 Y = ( S , V ) ∈ R 2 . For a digit d ∈ B 2 d\in B_2 d ∈ B 2 and a word w = ( d 1 , … , d n ) w=(d_1,\ldots,d_n) w = ( d 1 , … , d n ) over B 2 B_2 B 2 put
Θ x , d ( S , V ) = ( 2 x S − x V − 2 d , x S − d ) , Θ x , w = Θ x , d n ∘ ⋯ ∘ Θ x , d 1 . (95) \Theta_{x,d}(S,V)=(2xS-xV-2d,xS-d),\qquad\Theta_{x,w}=\Theta_{x,d_n}\circ\cdots\circ\Theta_{x,d_1}. \tag*{(95)} Θ x , d ( S , V ) = ( 2 x S − x V − 2 d , x S − d ) , Θ x , w = Θ x , d n ∘ ⋯ ∘ Θ x , d 1 . (95)
By (S1.3), Θ x , w ( 2 , 1 ) = ( S w ( x ) , V w ( x ) ) \Theta_{x,w}(2,1)=(S_w(x),V_w(x)) Θ x , w ( 2 , 1 ) = ( S w ( x ) , V w ( x )) . In the coordinates ( P , V ) = ( S − V , V ) (P,V)=(S-V,V) ( P , V ) = ( S − V , V ) the map Θ x , d \Theta_{x,d} Θ x , d is (S1.8), and ( 2 , 1 ) (2,1) ( 2 , 1 ) becomes ( 1 , 1 ) (1,1) ( 1 , 1 ) . Hence, by Lemma S1.7, if a digit sequence d 1 , d 2 , … d_1,d_2,\ldots d 1 , d 2 , … in B 2 B_2 B 2 moves ( 2 , 1 ) (2,1) ( 2 , 1 ) along a bounded orbit, then 1 − ∑ k d k z k ∈ F 2 1-\sum_k d_kz^k\in\mathcal{F}_2 1 − ∑ k d k z k ∈ F 2 has a double zero at z = 1 / x z=1/x z = 1/ x . The linear part L ( P , V ) = ( x P , x V + x P ) L(P,V)=(xP,xV+xP) L ( P , V ) = ( x P , x V + x P ) of (S1.8) satisfies L n ( P , V ) = x n ( P , V + n P ) L^n(P,V)=x^n(P,V+nP) L n ( P , V ) = x n ( P , V + n P ) . Rewritten in ( S , V ) (S,V) ( S , V ) and unrolled, this gives
Θ x , w Y = x n J n Y − ∑ k = 1 n d k x n − k ( n − k + 2 , n − k + 1 ) , J n = ( n + 1 − n n 1 − n ) . (96) \Theta_{x,w}Y=x^nJ^nY-\sum_{k=1}^{n}d_kx^{n-k}(n-k+2,n-k+1),\qquad J^n=\begin{pmatrix}n+1&-n\\n&1-n\end{pmatrix}. \tag*{(96)} Θ x , w Y = x n J n Y − k = 1 ∑ n d k x n − k ( n − k + 2 , n − k + 1 ) , J n = ( n + 1 n − n 1 − n ) . (96)
with states read as column vectors. For n ≥ 1 n\geq1 n ≥ 1 the absolute row sums of J n J^n J n are 2 n + 1 2n+1 2 n + 1 and 2 n − 1 2n-1 2 n − 1 .
In the coordinates
u 1 = 5 V 7 , u 2 = 2 S 3 − V 5 = 2 P 3 + 7 V 15 , u_1=\frac{5V}{7},\qquad u_2=\frac{2S}{3}-\frac{V}{5}=\frac{2P}{3}+\frac{7V}{15}, u 1 = 7 5 V , u 2 = 3 2 S − 5 V = 3 2 P + 15 7 V ,
put ∥ Y ∥ Q = ∣ u 1 ∣ + ∣ u 2 ∣ \lVert Y\rVert_Q=|u_1|+|u_2| ∥ Y ∥ Q = ∣ u 1 ∣ + ∣ u 2 ∣ and
Q = { Y ∈ R 2 : ∥ Y ∥ Q ≤ 1 } . (97) Q=\{Y\in\mathbb{R}^2:\lVert Y\rVert_Q\leq1\}. \tag*{(97)} Q = { Y ∈ R 2 : ∥ Y ∥ Q ≤ 1 } . (97)
Thus Y ∈ Q Y\in Q Y ∈ Q if and only if its four signed margins 1 − ϵ 1 u 1 − ϵ 2 u 2 1-\epsilon_1u_1-\epsilon_2u_2 1 − ϵ 1 u 1 − ϵ 2 u 2 , ϵ 1 , ϵ 2 ∈ { − 1 , 1 } \epsilon_1,\epsilon_2\in\{-1,1\} ϵ 1 , ϵ 2 ∈ { − 1 , 1 } , are nonnegative, and Y ∈ int Q Y\in\operatorname{int}Q Y ∈ int Q if and only if they are positive. Since
( S , V ) = ( 21 50 u 1 + 3 2 u 2 , 7 5 u 1 ) , (98) (S,V)=\left(\frac{21}{50}u_1+\frac{3}{2}u_2,\frac{7}{5}u_1\right), \tag*{(98)} ( S , V ) = ( 50 21 u 1 + 2 3 u 2 , 5 7 u 1 ) , (98)
every Y ∈ Q Y\in Q Y ∈ Q has ∥ Y ∥ ∞ ≤ 3 / 2 \lVert Y\rVert_\infty\leq3/2 ∥ Y ∥ ∞ ≤ 3/2 , where ∥ Y ∥ ∞ = max { ∣ S ∣ , ∣ V ∣ } \lVert Y\rVert_\infty=\max\{|S|,|V|\} ∥ Y ∥ ∞ = max { ∣ S ∣ , ∣ V ∣ } . Conversely,
∥ Y ∥ Q ≤ 2 3 ∣ S ∣ + ( 5 7 + 1 5 ) ∣ V ∣ ≤ 166 105 ∥ Y ∥ ∞ . (99) \lVert Y\rVert_Q\leq\frac{2}{3}|S|+\left(\frac{5}{7}+\frac{1}{5}\right)|V|\leq\frac{166}{105}\lVert Y\rVert_\infty. \tag*{(99)} ∥ Y ∥ Q ≤ 3 2 ∣ S ∣ + ( 7 5 + 5 1 ) ∣ V ∣ ≤ 105 166 ∥ Y ∥ ∞ . (99)
Finally, with G G G as in (S2.7), the cubic χ ( x ) = x 3 G ( 1 / x ) = x 3 − x 2 − x − 1 \chi(x)=x^3G(1/x)=x^3-x^2-x-1 χ ( x ) = x 3 G ( 1/ x ) = x 3 − x 2 − x − 1 increases on [ 1 , ∞ ) [1,\infty) [ 1 , ∞ ) , and
χ ( 183 100 ) = − 50413 10 6 < 0 < 3828037384313 15625 ⋅ 10 15 = χ ( 4598217 2500000 ) , \chi\left(\frac{183}{100}\right)=-\frac{50413}{10^6}<0<\frac{3828037384313}{15625\cdot10^{15}}=\chi\left(\frac{4598217}{2500000}\right), χ ( 100 183 ) = − 1 0 6 50413 < 0 < 15625 ⋅ 1 0 15 3828037384313 = χ ( 2500000 4598217 ) ,
so that 183 / 100 < ρ < 4598217 / 2500000 < 46 / 25 183/100<\rho<4598217/2500000<46/25 183/100 < ρ < 4598217/2500000 < 46/25 .
Proposition S2.30 (Closed return cover). For every x ∈ [ 433 / 250 , 4598217 / 2500000 ] x\in[433/250,4598217/2500000] x ∈ [ 433/250 , 4598217/2500000 ] and Y ∈ Q Y\in Q Y ∈ Q , some digit word w w w , with 1 ≤ ∣ w ∣ ≤ 10 1\leq|w|\leq10 1 ≤ ∣ w ∣ ≤ 10 , satisfies Θ x , w Y ∈ Q \Theta_{x,w}Y\in Q Θ x , w Y ∈ Q . The parameter interval contains ρ \rho ρ .
Proof. For a fixed word w w w , (S2.47) shows that the four signed margins of Θ x , w Y \Theta_{x,w}Y Θ x , w Y are affine in ( u 1 , u 2 ) (u_1,u_2) ( u 1 , u 2 ) , with coefficients that are polynomials in x x x . On a closed rational parameter band, exact rational enclosures of these polynomials over the band define a polygon of states, the word region of w w w , on which all four margins are nonnegative for every x x x in the band; thus Θ x , w \Theta_{x,w} Θ x , w maps the word region into Q Q Q . Part (i) of Proposition S4.8 splits [ 433 / 250 , 4598217 / 2500000 ] [433/250,4598217/2500000] [ 433/250 , 4598217/2500000 ] into consecutive closed rational bands and covers Q Q Q , on each band, by word regions of words of length between 1 and 10. This proves the first assertion. The second follows from the bracket for ρ \rho ρ displayed above. □ \square □
Tribonacci entry
Let
a = ( 2 , 1 , − 1 ) , b = ( − 1 , − 1 , − 2 ) , Y m ( x ) = Θ x , a b m ( 2 , 1 ) ( m ≥ 0 ) , a=(2,1,-1),\qquad b=(-1,-1,-2),\qquad Y_m(x)=\Theta_{x,ab^m}(2,1)\quad(m\geq0), a = ( 2 , 1 , − 1 ) , b = ( − 1 , − 1 , − 2 ) , Y m ( x ) = Θ x , a b m ( 2 , 1 ) ( m ≥ 0 ) ,
where b m b^m b m is the m m m -fold concatenation. In the convention of Lemma S1.7, the digits of f ρ f_\rho f ρ in (S2.7) are a a a followed by b , b , … b,b,\ldots b , b , … . Appending a digit d d d to a word u u u sends P u ( x ) P_u(x) P u ( x ) to x P u ( x ) − d xP_u(x)-d x P u ( x ) − d , so appending b b b acts by P ↦ x 3 P + x 2 + x + 2 P\mapsto x^3P+x^2+x+2 P ↦ x 3 P + x 2 + x + 2 . This affine map fixes P ˉ ( x ) = − ( x 2 + x + 2 ) / ( x 3 − 1 ) = − 1 / ( x − 1 ) − 1 / ( x 3 − 1 ) \bar P(x)=-(x^2+x+2)/(x^3-1)=-1/(x-1)-1/(x^3-1) P ˉ ( x ) = − ( x 2 + x + 2 ) / ( x 3 − 1 ) = − 1/ ( x − 1 ) − 1/ ( x 3 − 1 ) , and
Δ ( x ) : = P a ( x ) − P ˉ ( x ) = x 3 − 2 x 2 − x + 1 + x 2 + x + 2 x 3 − 1 = x χ ( x ) 2 x 3 − 1 . (100) \Delta(x):=P_a(x)-\bar P(x)=x^3-2x^2-x+1+\frac{x^2+x+2}{x^3-1}=\frac{x\chi(x)^2}{x^3-1}. \tag*{(100)} Δ ( x ) := P a ( x ) − P ˉ ( x ) = x 3 − 2 x 2 − x + 1 + x 3 − 1 x 2 + x + 2 = x 3 − 1 x χ ( x ) 2 . (100)
Hence P a b m = P ˉ + x 3 m Δ P_{ab^m}=\bar P+x^{3m}\Delta P a b m = P ˉ + x 3 m Δ as rational functions of x x x . Differentiating, and using S = 2 P + x P ′ S=2P+xP' S = 2 P + x P ′ and V = P + x P ′ V=P+xP' V = P + x P ′ from (S1.2),
Y m ( x ) = Y ˉ ( x ) + x 3 m ( ω + ( 3 m + 1 ) Δ , ω + 3 m Δ ) ( x ) , ω = Δ + x Δ ′ , (101) Y_m(x)=\bar Y(x)+x^{3m}\bigl(\omega+(3m+1)\Delta,\ \omega+3m\Delta\bigr)(x),\qquad\omega=\Delta+x\Delta', \tag*{(101)} Y m ( x ) = Y ˉ ( x ) + x 3 m ( ω + ( 3 m + 1 ) Δ , ω + 3 m Δ ) ( x ) , ω = Δ + x Δ ′ , (101)
where Y ˉ = ( S ˉ , V ˉ ) = ( 2 P ˉ + x P ˉ ′ , P ˉ + x P ˉ ′ ) \bar Y=(\bar S,\bar V)=(2\bar P+x\bar P',\bar P+x\bar P') Y ˉ = ( S ˉ , V ˉ ) = ( 2 P ˉ + x P ˉ ′ , P ˉ + x P ˉ ′ ) is the state fixed by the block:
S ˉ ( x ) = 2 − x ( x − 1 ) 2 + x 3 + 2 ( x 3 − 1 ) 2 , V ˉ ( x ) = 1 ( x − 1 ) 2 + 2 x 3 + 1 ( x 3 − 1 ) 2 . (102) \bar S(x)=\frac{2-x}{(x-1)^2}+\frac{x^3+2}{(x^3-1)^2},\qquad\bar V(x)=\frac{1}{(x-1)^2}+\frac{2x^3+1}{(x^3-1)^2}. \tag*{(102)} S ˉ ( x ) = ( x − 1 ) 2 2 − x + ( x 3 − 1 ) 2 x 3 + 2 , V ˉ ( x ) = ( x − 1 ) 2 1 + ( x 3 − 1 ) 2 2 x 3 + 1 . (102)
Put ε = ρ − x \varepsilon=\rho-x ε = ρ − x . As ρ 3 = ρ 2 + ρ + 1 \rho^3=\rho^2+\rho+1 ρ 3 = ρ 2 + ρ + 1 , we have χ ( x ) = − ε ψ ( x ) \chi(x)=-\varepsilon\psi(x) χ ( x ) = − ε ψ ( x ) with ψ ( x ) = x 2 + ( ρ − 1 ) x + ρ 2 − ρ − 1 \psi(x)=x^2+(\rho-1)x+\rho^2-\rho-1 ψ ( x ) = x 2 + ( ρ − 1 ) x + ρ 2 − ρ − 1 , so that ψ ( ρ ) = χ ′ ( ρ ) \psi(\rho)=\chi'(\rho) ψ ( ρ ) = χ ′ ( ρ ) . With η = ψ 2 / ( x 3 − 1 ) \eta=\psi^2/(x^3-1) η = ψ 2 / ( x 3 − 1 ) ,
Δ = ε 2 η , ω = − 2 x ε η + ε 2 ( η + x η ′ ) . (103) \Delta=\varepsilon^2\eta,\qquad\omega=-2x\varepsilon\eta+\varepsilon^2(\eta+x\eta'). \tag*{(103)} Δ = ε 2 η , ω = − 2 x ε η + ε 2 ( η + x η ′ ) . (103)
In particular Δ ( ρ ) = ω ( ρ ) = 0 \Delta(\rho)=\omega(\rho)=0 Δ ( ρ ) = ω ( ρ ) = 0 , so Y m ( ρ ) = Y ˉ ( ρ ) Y_m(\rho)=\bar Y(\rho) Y m ( ρ ) = Y ˉ ( ρ ) for every m m m . Put κ = 2 η ( ρ ) / ρ 2 \kappa=2\eta(\rho)/\rho^2 κ = 2 η ( ρ ) / ρ 2 , so that ω ( x ) = − κ ρ 3 ε + O ( ε 2 ) \omega(x)=-\kappa\rho^3\varepsilon+O(\varepsilon^2) ω ( x ) = − κ ρ 3 ε + O ( ε 2 ) . Multiply (S2.52) by ( x 3 − 1 ) 2 = ( x − 1 ) 2 ( x 2 + x + 1 ) 2 (x^3-1)^2=(x-1)^2(x^2+x+1)^2 ( x 3 − 1 ) 2 = ( x − 1 ) 2 ( x 2 + x + 1 ) 2 ; the polynomial identities
( 2 − x ) ( x 2 + x + 1 ) 2 + x 3 + 2 − ( 7 + 2 x − 3 x 2 ) ( x 3 − 1 ) 2 = ( 3 x 5 + x 4 − 3 x 3 − 6 x 2 − 4 x + 3 ) χ ( x ) , (2-x)(x^2+x+1)^2+x^3+2-(7+2x-3x^2)(x^3-1)^2=(3x^5+x^4-3x^3-6x^2-4x+3)\chi(x), ( 2 − x ) ( x 2 + x + 1 ) 2 + x 3 + 2 − ( 7 + 2 x − 3 x 2 ) ( x 3 − 1 ) 2 = ( 3 x 5 + x 4 − 3 x 3 − 6 x 2 − 4 x + 3 ) χ ( x ) ,
( x 2 + x + 1 ) 2 + 2 x 3 + 1 − ( 5 + 2 x − 2 x 2 ) ( x 3 − 1 ) 2 = ( 2 x 5 − 3 x 3 − 5 x 2 − 3 x + 3 ) χ ( x ) , (x^2+x+1)^2+2x^3+1-(5+2x-2x^2)(x^3-1)^2=(2x^5-3x^3-5x^2-3x+3)\chi(x), ( x 2 + x + 1 ) 2 + 2 x 3 + 1 − ( 5 + 2 x − 2 x 2 ) ( x 3 − 1 ) 2 = ( 2 x 5 − 3 x 3 − 5 x 2 − 3 x + 3 ) χ ( x ) ,
2 χ ′ ( x ) 2 − ( 16 + 52 x − 32 x 2 ) x 2 ( x 3 − 1 ) = ( 32 x 4 − 20 x 3 − 4 x 2 − 6 x − 2 ) χ ( x ) , 2\chi'(x)^2-(16+52x-32x^2)x^2(x^3-1)=(32x^4-20x^3-4x^2-6x-2)\chi(x), 2 χ ′ ( x ) 2 − ( 16 + 52 x − 32 x 2 ) x 2 ( x 3 − 1 ) = ( 32 x 4 − 20 x 3 − 4 x 2 − 6 x − 2 ) χ ( x ) ,
taken at x = ρ x=\rho x = ρ , give
Y ˉ ( ρ ) = Y ∗ : = ( 7 + 2 ρ − 3 ρ 2 , 5 + 2 ρ − 2 ρ 2 ) , κ = 16 + 52 ρ − 32 ρ 2 . (104) \bar Y(\rho)=Y_*:=(7+2\rho-3\rho^2,\ 5+2\rho-2\rho^2),\qquad\kappa=16+52\rho-32\rho^2. \tag*{(104)} Y ˉ ( ρ ) = Y ∗ := ( 7 + 2 ρ − 3 ρ 2 , 5 + 2 ρ − 2 ρ 2 ) , κ = 16 + 52 ρ − 32 ρ 2 . (104)
For x < ρ x<\rho x < ρ and m ≥ 0 m\ge0 m ≥ 0 let τ m ( x ) = ( ρ − x ) ρ 3 m + 3 \tau_m(x)=(\rho-x)\rho^{3m+3} τ m ( x ) = ( ρ − x ) ρ 3 m + 3 be the phase of x x x at scale m m m (it is unrelated to τ h \tau_h τ h in (S1.4)), and put
τ − = 9 80 , τ + = τ − ρ 3 , Y ∞ ( τ ) = Y ∗ − κ τ ( 1 , 1 ) . \tau_-=\frac{9}{80},\qquad\tau_+=\tau_-\rho^3,\qquad Y_\infty(\tau)=Y_*-\kappa\tau(1,1). τ − = 80 9 , τ + = τ − ρ 3 , Y ∞ ( τ ) = Y ∗ − κ τ ( 1 , 1 ) .
For x ∈ [ ρ − τ − , ρ ) x\in[\rho-\tau_-,\rho) x ∈ [ ρ − τ − , ρ ) , the least m ≥ 0 m\ge0 m ≥ 0 with τ m ( x ) ≥ τ − \tau_m(x)\ge\tau_- τ m ( x ) ≥ τ − is the entry scale of x x x ; since τ 0 ( x ) ≤ τ + \tau_0(x)\le\tau_+ τ 0 ( x ) ≤ τ + and τ m + 1 = ρ 3 τ m \tau_{m+1}=\rho^3\tau_m τ m + 1 = ρ 3 τ m , it satisfies τ m ( x ) ∈ [ τ − , τ + ] \tau_m(x)\in[\tau_-,\tau_+] τ m ( x ) ∈ [ τ − , τ + ] . By (S2.57) below, Y m ( ρ − τ ρ − 3 m − 3 ) → Y ∞ ( τ ) Y_m(\rho-\tau\rho^{-3m-3})\to Y_\infty(\tau) Y m ( ρ − τ ρ − 3 m − 3 ) → Y ∞ ( τ ) as m → ∞ m\to\infty m → ∞ , for each τ ∈ [ τ − , τ + ] \tau\in[\tau_-,\tau_+] τ ∈ [ τ − , τ + ] : after the blocks, the state lies near the segment Y ∞ ( [ τ − , τ + ] ) Y_\infty([\tau_-,\tau_+]) Y ∞ ([ τ − , τ + ]) , at a point fixed by the phase. One of the eight suffixes
s 1 = ( − 1 , − 2 , 0 , − 2 , 0 , − 2 ) , s 2 = ( − 1 , − 2 , 0 , − 2 ) , s 3 = ( − 1 , − 2 ) , s 4 = ( − 2 , 0 ) , s_1=(-1,-2,0,-2,0,-2),\quad s_2=(-1,-2,0,-2),\quad s_3=(-1,-2),\quad s_4=(-2,0), s 1 = ( − 1 , − 2 , 0 , − 2 , 0 , − 2 ) , s 2 = ( − 1 , − 2 , 0 , − 2 ) , s 3 = ( − 1 , − 2 ) , s 4 = ( − 2 , 0 ) ,
s 5 = ( − 2 ) , s 6 = ( − 2 , − 2 ) , s 7 = ( − 2 , − 2 , 0 , 2 , 2 , 2 ) , s 8 = ( − 2 , − 2 , 0 , 2 , 2 ) s_5=(-2),\quad s_6=(-2,-2),\quad s_7=(-2,-2,0,2,2,2),\quad s_8=(-2,-2,0,2,2) s 5 = ( − 2 ) , s 6 = ( − 2 , − 2 ) , s 7 = ( − 2 , − 2 , 0 , 2 , 2 , 2 ) , s 8 = ( − 2 , − 2 , 0 , 2 , 2 )
moves each point of this segment well inside Q Q Q : by part (iii) of Proposition S4.8, for every τ ∈ [ τ − , τ + ] \tau\in[\tau_-,\tau_+] τ ∈ [ τ − , τ + ] some j j j satisfies
∥ Θ ρ , s j Y ∞ ( τ ) ∥ Q ≤ 1 − μ , μ = 223 46500 . (105) \left\|\Theta_{\rho,s_j}Y_\infty(\tau)\right\|_Q\le1-\mu,\qquad\mu=\frac{223}{46500}. \tag*{(105)} Θ ρ , s j Y ∞ ( τ ) Q ≤ 1 − μ , μ = 46500 223 . (105)
The record assigns s j s_j s j to the j j j th of eight consecutive closed phase cells. As Θ ρ , s j Y ∞ ( τ ) \Theta_{\rho,s_j}Y_\infty(\tau) Θ ρ , s j Y ∞ ( τ ) is affine in τ \tau τ , so are its signed margins, and it suffices to check them at the two ends of each cell. In the coordinates ( P , V ) (P,V) ( P , V ) , Y ∞ ( τ ) = ( 2 − ρ 2 , 5 + 2 ρ − 2 ρ 2 − κ τ ) Y_\infty(\tau)=(2-\rho^2,\ 5+2\rho-2\rho^2-\kappa\tau) Y ∞ ( τ ) = ( 2 − ρ 2 , 5 + 2 ρ − 2 ρ 2 − κ τ ) , the form used in the record.
Uniform control of all entry scales.
Lemma S2.31 (Entry at every scale). Let m ≥ 0 m\ge0 m ≥ 0 , τ ∈ [ τ − , τ + ] \tau\in[\tau_-,\tau_+] τ ∈ [ τ − , τ + ] and x = ρ − τ ρ − 3 m − 3 x=\rho-\tau\rho^{-3m-3} x = ρ − τ ρ − 3 m − 3 .
(i) If m ≤ 8 m\le8 m ≤ 8 , some word s s s over B 2 B_2 B 2 satisfies Θ x , s Y m ( x ) ∈ int Q \Theta_{x,s}Y_m(x)\in\operatorname{int}Q Θ x , s Y m ( x ) ∈ int Q .
(ii) If m ≥ 9 m\ge9 m ≥ 9 , then for j = 1 , … , 8 j=1,\ldots,8 j = 1 , … , 8 ,
∥ Θ x , s j Y m ( x ) − Θ ρ , s j Y ∞ ( τ ) ∥ Q < 9 2000 < μ , \left\|\Theta_{x,s_j}Y_m(x)-\Theta_{\rho,s_j}Y_\infty(\tau)\right\|_Q<\frac{9}{2000}<\mu, Θ x , s j Y m ( x ) − Θ ρ , s j Y ∞ ( τ ) Q < 2000 9 < μ , so Θ x , s j Y m ( x ) ∈ int Q \Theta_{x,s_j}Y_m(x)\in\operatorname{int}Q Θ x , s j Y m ( x ) ∈ int Q whenever s j s_j s j satisfies (S2.55).
Proof. Part (i) is part (ii) of Proposition S4.8.
Idea for (ii). By (S2.51), the entry state lies within O ( m ρ − 3 m ) O(m\rho^{-3m}) O ( m ρ − 3 m ) of the limiting segment, a suffix of length at most six enlarges errors by a bounded factor, and the gain ρ − 3 \rho^{-3} ρ − 3 per scale beats the factor m m m . The constants below are crude.
Write ε = ρ − x \varepsilon=\rho-x ε = ρ − x , so that ε ρ 3 m = τ ρ − 3 ≤ 9 / 80 \varepsilon\rho^{3m}=\tau\rho^{-3}\le9/80 ε ρ 3 m = τ ρ − 3 ≤ 9/80 . As 183 / 100 < ρ < 46 / 25 183/100<\rho<46/25 183/100 < ρ < 46/25 , we have ρ 3 < ( 46 / 25 ) 3 < 25 / 4 \rho^3<(46/25)^3<25/4 ρ 3 < ( 46/25 ) 3 < 25/4 ; as m ≥ 9 m\ge9 m ≥ 9 , also ε ≤ 9 80 ρ − 27 < 10 − 6 \varepsilon\le\frac{9}{80}\rho^{-27}<10^{-6} ε ≤ 80 9 ρ − 27 < 1 0 − 6 , and thus 9 / 5 ≤ x ≤ ρ 9/5\le x\le\rho 9/5 ≤ x ≤ ρ . The block. For 9 / 5 ≤ ξ ≤ 46 / 25 9/5 \le\xi\le46/25 9/5 ≤ ξ ≤ 46/25 ,
0 ≤ ψ ( ξ ) ≤ 11 2 , 0 ≤ ψ ′ ( ξ ) = 2 ξ + ρ − 1 ≤ 23 5 , 24 5 ≤ ξ 3 − 1 ≤ 21 4 , 3 ξ 2 ≤ 21 2 . 0 \le\psi(\xi) \le\frac{11}{2}, \qquad0 \le\psi'(\xi)=2\xi+\rho-1\le\frac{23}{5}, \qquad\frac{24}{5}\le\xi^3-1\le\frac{21}{4}, \qquad3\xi^2\le\frac{21}{2}. 0 ≤ ψ ( ξ ) ≤ 2 11 , 0 ≤ ψ ′ ( ξ ) = 2 ξ + ρ − 1 ≤ 5 23 , 5 24 ≤ ξ 3 − 1 ≤ 4 21 , 3 ξ 2 ≤ 2 21 .
Hence 0 ≤ η ≤ ( 11 / 2 ) 2 / ( 24 / 5 ) < 7 0\le\eta\le(11/2)^2/(24/5)<7 0 ≤ η ≤ ( 11/2 ) 2 / ( 24/5 ) < 7 , and η ′ = 2 ψ ψ ′ / ( ξ 3 − 1 ) − 3 ξ 2 ψ 2 / ( ξ 3 − 1 ) 2 \eta'=2\psi\psi'/(\xi^3-1)-3\xi^2\psi^2/(\xi^3-1)^2 η ′ = 2 ψ ψ ′ / ( ξ 3 − 1 ) − 3 ξ 2 ψ 2 / ( ξ 3 − 1 ) 2 is a difference of numbers in [ 0 , 11 ] [0,11] [ 0 , 11 ] and [ 0 , 14 ] [0,14] [ 0 , 14 ] , so ∣ η ′ ∣ ≤ 14 |\eta'|\le14 ∣ η ′ ∣ ≤ 14 . By the mean value theorem ∣ η ( ρ ) − η ( x ) ∣ ≤ 14 ε |\eta(\rho)-\eta(x)|\le14\varepsilon ∣ η ( ρ ) − η ( x ) ∣ ≤ 14 ε , so ∣ ρ η ( ρ ) − x η ( x ) ∣ ≤ ρ ∣ η ( ρ ) − η ( x ) ∣ + ε η ( x ) ≤ 33 ε |\rho\eta(\rho)-x\eta(x)|\le\rho|\eta(\rho)-\eta(x)|+\varepsilon\eta(x)\le33\varepsilon ∣ ρ η ( ρ ) − x η ( x ) ∣ ≤ ρ ∣ η ( ρ ) − η ( x ) ∣ + ε η ( x ) ≤ 33 ε ; also ∣ η + x η ′ ∣ ≤ 7 + 46 25 ⋅ 14 < 33 |\eta+x\eta'|\le7+\frac{46}{25}\cdot14<33 ∣ η + x η ′ ∣ ≤ 7 + 25 46 ⋅ 14 < 33 . As κ ρ 3 = 2 ρ η ( ρ ) \kappa\rho^3=2\rho\eta(\rho) κ ρ 3 = 2 ρ η ( ρ ) , (S2.53) gives ω + κ ρ 3 ε = 2 ε ( ρ η ( ρ ) − x η ( x ) ) + ε 2 ( η + x η ′ ) ( x ) \omega+\kappa\rho^3\varepsilon=2\varepsilon(\rho\eta(\rho)-x\eta(x))+\varepsilon^2(\eta+x\eta')(x) ω + κ ρ 3 ε = 2 ε ( ρ η ( ρ ) − x η ( x )) + ε 2 ( η + x η ′ ) ( x ) , and therefore
∣ Δ ( x ) ∣ ≤ 7 ε 2 , ∣ ω ( x ) + κ ρ 3 ε ∣ ≤ 99 ε 2 . (106) |\Delta(x)|\le7\varepsilon^2,\qquad|\omega(x)+\kappa\rho^3\varepsilon|\le99\varepsilon^2. \tag*{(106)} ∣Δ ( x ) ∣ ≤ 7 ε 2 , ∣ ω ( x ) + κ ρ 3 ε ∣ ≤ 99 ε 2 . (106)
Differentiating (S2.52),
S ˉ ′ = x − 3 ( x − 1 ) 3 − 3 x 2 ( x 3 + 5 ) ( x 3 − 1 ) 3 , V ˉ ′ = − 2 ( x − 1 ) 3 − 6 x 2 ( x 3 + 2 ) ( x 3 − 1 ) 3 . \bar{S}'=\frac{x-3}{(x-1)^3}-\frac{3x^2(x^3+5)}{(x^3-1)^3},\qquad\bar{V}'=-\frac{2}{(x-1)^3}-\frac{6x^2(x^3+2)}{(x^3-1)^3}. S ˉ ′ = ( x − 1 ) 3 x − 3 − ( x 3 − 1 ) 3 3 x 2 ( x 3 + 5 ) , V ˉ ′ = − ( x − 1 ) 3 2 − ( x 3 − 1 ) 3 6 x 2 ( x 3 + 2 ) .
On [ 9 / 5 , 46 / 25 ] [9/5,46/25] [ 9/5 , 46/25 ] we have ∣ x − 3 ∣ ≤ 6 / 5 |x-3|\le6/5 ∣ x − 3∣ ≤ 6/5 , ( x − 1 ) 3 ≥ 64 / 125 (x-1)^3\ge64/125 ( x − 1 ) 3 ≥ 64/125 , ( x 3 − 1 ) 3 > 110 (x^3-1)^3>110 ( x 3 − 1 ) 3 > 110 , 3 x 2 ≤ 11 3x^2\le11 3 x 2 ≤ 11 , x 3 + 5 ≤ 12 x^3+5\le12 x 3 + 5 ≤ 12 , 6 x 2 ≤ 21 6x^2\le21 6 x 2 ≤ 21 and x 3 + 2 ≤ 9 x^3+2\le9 x 3 + 2 ≤ 9 , so ∣ S ˉ ′ ∣ < 75 32 + 132 110 < 4 |\bar{S}'|<\frac{75}{32}+\frac{132}{110}<4 ∣ S ˉ ′ ∣ < 32 75 + 110 132 < 4 and ∣ V ˉ ′ ∣ < 125 32 + 189 110 < 6 |\bar{V}'|<\frac{125}{32}+\frac{189}{110}<6 ∣ V ˉ ′ ∣ < 32 125 + 110 189 < 6 . By the mean value theorem, ∥ Y ˉ ( x ) − Y ˉ ( ρ ) ∥ ∞ ≤ 6 ε \|\bar{Y}(x)-\bar{Y}(\rho)\|_\infty\le6\varepsilon ∥ Y ˉ ( x ) − Y ˉ ( ρ ) ∥ ∞ ≤ 6 ε .
The entry error. Since κ τ = κ ρ 3 ε ρ 3 m \kappa\tau=\kappa\rho^3\varepsilon\rho^{3m} κ τ = κ ρ 3 ε ρ 3 m , for i ∈ { 3 m , 3 m + 1 } i\in\{3m,3m+1\} i ∈ { 3 m , 3 m + 1 }
x 3 m ( ω + i Δ ) + κ τ = x 3 m ( ω + κ ρ 3 ε ) + κ ρ 3 ε ( ρ 3 m − x 3 m ) + i x 3 m Δ . x^{3m}(\omega+i\Delta)+\kappa\tau=x^{3m}(\omega+\kappa\rho^3\varepsilon)+\kappa\rho^3\varepsilon(\rho^{3m}-x^{3m})+ix^{3m}\Delta. x 3 m ( ω + i Δ ) + κ τ = x 3 m ( ω + κ ρ 3 ε ) + κ ρ 3 ε ( ρ 3 m − x 3 m ) + i x 3 m Δ.
Here 0 < x 3 m ≤ ρ 3 m 0<x^{3m}\le\rho^{3m} 0 < x 3 m ≤ ρ 3 m , 0 ≤ ρ 3 m − x 3 m ≤ 3 m ρ 3 m ε 0\le\rho^{3m}-x^{3m}\le3m\rho^{3m}\varepsilon 0 ≤ ρ 3 m − x 3 m ≤ 3 m ρ 3 m ε , and 0 < κ < 4 0<\kappa<4 0 < κ < 4 , because κ = 16 + 52 ρ − 32 ρ 2 \kappa=16+52\rho-32\rho^2 κ = 16 + 52 ρ − 32 ρ 2 decreases in ρ \rho ρ from 2497 / 625 2497/625 2497/625 at 183 / 100 183/100 183/100 to 2088 / 625 2088/625 2088/625 at 46 / 25 46/25 46/25 . By (S2.51), (S2.56), ε ρ 3 m ≤ 9 / 80 \varepsilon\rho^{3m}\le9/80 ε ρ 3 m ≤ 9/80 and ρ 3 < 7 \rho^3<7 ρ 3 < 7 ,
∥ Y m ( x ) − Y ∞ ( τ ) ∥ ∞ ≤ 6 ε + ρ 3 m ( 106 + 21 m + 12 m ρ 3 ) ε 2 ≤ 3 + 2 m ρ 3 m . (107) \|Y_m(x)-Y_\infty(\tau)\|_\infty\le6\varepsilon+\rho^{3m}(106+21m+12m\rho^3)\varepsilon^2\le\frac{3+2m}{\rho^{3m}}. \tag*{(107)} ∥ Y m ( x ) − Y ∞ ( τ ) ∥ ∞ ≤ 6 ε + ρ 3 m ( 106 + 21 m + 12 m ρ 3 ) ε 2 ≤ ρ 3 m 3 + 2 m . (107)
because 6 ⋅ 9 80 + ( 106 + 105 m ) ( 9 80 ) 2 < 3 + 2 m 6\cdot\frac{9}{80}+(106+105m)(\frac{9}{80})^2<3+2m 6 ⋅ 80 9 + ( 106 + 105 m ) ( 80 9 ) 2 < 3 + 2 m .
The suffix. Let s = s j s=s_j s = s j and 1 ≤ ℓ = ∣ s ∣ ≤ 6 1\le\ell=|s|\le6 1 ≤ ℓ = ∣ s ∣ ≤ 6 . We apply Lemma S1.8 with the matrix J J J of (S2.47), e = ( 2 , 1 ) e=(2,1) e = ( 2 , 1 ) , h = 2 h=2 h = 2 and q = ρ q=\rho q = ρ . The row sums give C ℓ ≤ 13 C_\ell\le13 C ℓ ≤ 13 and B j = j + 2 B_j=j+2 B j = j + 2 , hence ℓ C ℓ ≤ 13 ( 46 / 25 ) 6 < 505 \ell C_\ell\le13(46/25)^6<505 ℓ C ℓ ≤ 13 ( 46/25 ) 6 < 505 . For the limiting initial state put Y = Y ∞ ( τ ) Y=Y_\infty(\tau) Y = Y ∞ ( τ ) . On [ 183 / 100 , 46 / 25 ] [183/100,46/25] [ 183/100 , 46/25 ] both coordinates of Y ∗ Y_* Y ∗ decrease in ρ \rho ρ , from 6133 / 10 4 6133/10^4 6133/1 0 4 and 19622 / 10 4 19622/10^4 19622/1 0 4 to 327 / 625 327/625 327/625 and 1193 / 625 1193/625 1193/625 , so Y ∗ ∈ ( 0 , 1 ) × ( 0 , 2 ) Y_*\in(0,1)\times(0,2) Y ∗ ∈ ( 0 , 1 ) × ( 0 , 2 ) ; with 0 ≤ κ τ ≤ 4 τ + < 3 0\le\kappa\tau\le4\tau_+<3 0 ≤ κ τ ≤ 4 τ + < 3 this gives ∥ Y ∥ ∞ ≤ 3 \|Y\|_\infty\le3 ∥ Y ∥ ∞ ≤ 3 . Since j ρ j − 1 < 130 j\rho^{j-1}<130 j ρ j − 1 < 130 for 1 ≤ j ≤ 6 1\le j\le6 1 ≤ j ≤ 6 , the parameter term in (S1.10) is bounded by
130 ( 13 ⋅ 3 + 2 ∑ j = 1 5 ( j + 2 ) ) ε ≤ 93 ⋅ 130 ε = 12090 ε < 1361 ρ 3 m . 130\left(13\cdot3+2\sum_{j=1}^{5}(j+2)\right)\varepsilon\le93\cdot130\varepsilon=12090\varepsilon<\frac{1361}{\rho^{3m}}. 130 ( 13 ⋅ 3 + 2 j = 1 ∑ 5 ( j + 2 ) ) ε ≤ 93 ⋅ 130 ε = 12090 ε < ρ 3 m 1361 .
Conclusion. Substituting these bounds and (S2.57) into (S1.10), then using (S2.50), gives
∥ Θ x , s Y m ( x ) − Θ ρ , s Y ∞ ( τ ) ∥ Q ≤ 166 105 ⋅ 505 ( 3 + 2 m ) + 1361 ρ 3 m . \|\Theta_{x,s}Y_m(x)-\Theta_{\rho,s}Y_\infty(\tau)\|_Q\le\frac{166}{105}\cdot\frac{505(3+2m)+1361}{\rho^{3m}}. ∥ Θ x , s Y m ( x ) − Θ ρ , s Y ∞ ( τ ) ∥ Q ≤ 105 166 ⋅ ρ 3 m 505 ( 3 + 2 m ) + 1361 .
At m = 9 m=9 m = 9 , as ρ > 9 / 5 \rho>9/5 ρ > 9/5 , this is less than 166 105 ⋅ 11966 ⋅ ( 5 / 9 ) 27 < 0.0025 < 9 / 2000 \frac{166}{105}\cdot11966\cdot(5/9)^{27}<0.0025<9/2000 105 166 ⋅ 11966 ⋅ ( 5/9 ) 27 < 0.0025 < 9/2000 . Passing from m m m to m + 1 m+1 m + 1 multiplies the numerator by less than 2 and the denominator by ρ 3 > 2 \rho^3>2 ρ 3 > 2 , so the bound holds for every m ≥ 9 m\ge9 m ≥ 9 . Finally 9 / 2000 < μ 9/2000<\mu 9/2000 < μ , so under (S2.55) we get ∥ Θ x , s j Y m ( x ) ∥ Q < 1 − μ + μ = 1 \|\Theta_{x,s_j}Y_m(x)\|_Q<1-\mu+\mu=1 ∥ Θ x , s j Y m ( x ) ∥ Q < 1 − μ + μ = 1 . □
Proposition S2.32 (Tribonacci entry). For every
x ∈ [ ρ − 9 80 ρ − m , ρ ) , x\in\left[\rho-\frac{9}{80}\rho^{-m},\rho\right), x ∈ [ ρ − 80 9 ρ − m , ρ ) , there is a finite digit word w w w such that Θ x , w ( 2 , 1 ) ∈ Q \Theta_{x,w}(2,1)\in Q Θ x , w ( 2 , 1 ) ∈ Q .
Proof. Let m m m be the entry scale of x x x and τ = τ m ( x ) ∈ [ τ − , τ + ] \tau=\tau_m(x)\in[\tau_-,\tau_+] τ = τ m ( x ) ∈ [ τ − , τ + ] , so that x = ρ − τ ρ − 3 m − 3 x=\rho-\tau\rho^{-3m-3} x = ρ − τ ρ − 3 m − 3 . If m ≤ 8 m\le8 m ≤ 8 , take w = a b m s w=ab^ms w = a b m s with s s s from Lemma S2.31(i). If m ≥ 9 m\ge9 m ≥ 9 , choose j j j by (S2.55) and take w = a b m s j w=ab^ms_j w = a b m s j ; then Lemma S2.31(ii) gives Θ x , w ( 2 , 1 ) ∈ Q \Theta_{x,w}(2,1)\in Q Θ x , w ( 2 , 1 ) ∈ Q . □
:::
Completion
Proposition S2.33 (The height-two active trace). One has [ 1 / ρ , 1 ) ⊆ A 2 ⊆ D 2 [1/\rho,1) \subseteq A_2 \subseteq D_2 [ 1/ ρ , 1 ) ⊆ A 2 ⊆ D 2 . Equivalently, ( 1 , ρ ] ⊆ C 2 e x t (1,\rho] \subseteq C^{\mathrm{ext}}_2 ( 1 , ρ ] ⊆ C 2 ext .
Proof. Idea. An entry word followed by return words gives a bounded orbit, hence a double zero; an interval of double zeros is active.
Let 433 / 250 ≤ x < ρ 433/250 \le x < \rho 433/250 ≤ x < ρ . Then x ≥ ρ − 9 / 80 x \ge\rho- 9/80 x ≥ ρ − 9/80 , as ρ < 46 / 25 \rho< 46/25 ρ < 46/25 , and Proposition S2.32 gives a word w 0 w_0 w 0 with Y ( 0 ) = Θ x , w 0 ( 2 , 1 ) ∈ Q Y^{(0)}=\Theta_{x,w_0}(2,1)\in Q Y ( 0 ) = Θ x , w 0 ( 2 , 1 ) ∈ Q . As x < ρ < 4598217 / 2500000 x < \rho< 4598217/2500000 x < ρ < 4598217/2500000 , Proposition S2.30 assigns to each Y ∈ Q Y\in Q Y ∈ Q a word w ( Y ) w(Y) w ( Y ) with 1 ≤ ∣ w ( Y ) ∣ ≤ 10 1\le|w(Y)|\le10 1 ≤ ∣ w ( Y ) ∣ ≤ 10 and Θ x , w ( Y ) Y ∈ Q \Theta_{x,w(Y)}Y\in Q Θ x , w ( Y ) Y ∈ Q . Put Y ( k + 1 ) = Θ x , w ( Y ( k ) ) Y ( k ) Y^{(k+1)}=\Theta_{x,w(Y^{(k)})}Y^{(k)} Y ( k + 1 ) = Θ x , w ( Y ( k ) ) Y ( k ) and follow the infinite digit sequence w 0 w ( Y ( 0 ) ) w ( Y ( 1 ) ) ⋯ w_0w(Y^{(0)})w(Y^{(1)})\cdots w 0 w ( Y ( 0 ) ) w ( Y ( 1 ) ) ⋯ . Its orbit from ( 2 , 1 ) (2,1) ( 2 , 1 ) consists of the states along w 0 w_0 w 0 , the return states Y ( k ) ∈ Q Y^{(k)}\in Q Y ( k ) ∈ Q , and states Θ x , v Y ( k ) \Theta_{x,v}Y^{(k)} Θ x , v Y ( k ) with ∣ v ∣ ≤ 9 |v|\le9 ∣ v ∣ ≤ 9 . On Q Q Q , ∥ Y ∥ ∞ ≤ 3 / 2 \|Y\|_\infty\le3/2 ∥ Y ∥ ∞ ≤ 3/2 , and if ∥ Y ∥ ∞ ≤ M \|Y\|_\infty\le M ∥ Y ∥ ∞ ≤ M with M ≥ 1 M\ge1 M ≥ 1 , then ∥ Θ x , d Y ∥ ∞ ≤ 3 x M + 4 ≤ 10 M \|\Theta_{x,d}Y\|_\infty\le3xM+4\le10M ∥ Θ x , d Y ∥ ∞ ≤ 3 x M + 4 ≤ 10 M , as x < 2 x<2 x < 2 . So every state after w 0 w_0 w 0 has ∥ ⋅ ∥ ∞ ≤ 10 9 ⋅ 3 / 2 \|\cdot\|_\infty\le10^9\cdot3/2 ∥ ⋅ ∥ ∞ ≤ 1 0 9 ⋅ 3/2 . The orbit is bounded, and so is its ( P , V ) (P,V) ( P , V ) -orbit from ( 1 , 1 ) (1,1) ( 1 , 1 ) ; by Lemma S1.7, 1 / x ∈ D 2 1/x\in D_2 1/ x ∈ D 2 .
At x = ρ x=\rho x = ρ , the digits a , b , b , … a,b,b,\ldots a , b , b , … of f ρ f_\rho f ρ give the state Y m ( ρ ) = Y ‾ ( ρ ) Y_m(\rho)=\overline{Y}(\rho) Y m ( ρ ) = Y ( ρ ) after every block, and two further states inside each block. This orbit is bounded, so 1 / ρ ∈ D 2 1/\rho\in D_2 1/ ρ ∈ D 2 , with witness f ρ f_\rho f ρ . For 1 < x < 433 / 250 1<x<433/250 1 < x < 433/250 we have x < 3 x<\sqrt{3} x < 3 , since 433 2 = 187489 < 187500 = 3 ⋅ 250 2 433^2=187489<187500=3\cdot250^2 43 3 2 = 187489 < 187500 = 3 ⋅ 25 0 2 , and 1 / x ∈ D 2 1/x\in D_2 1/ x ∈ D 2 by Corollary S1.6.
Thus [ 1 / ρ , 1 ) ⊆ D 2 [1/\rho,1)\subseteq D_2 [ 1/ ρ , 1 ) ⊆ D 2 . This interval is nondegenerate, so Lemma 2.7 gives [ 1 / ρ , 1 ) ⊆ A 2 [1/\rho,1)\subseteq A_2 [ 1/ ρ , 1 ) ⊆ A 2 , including the endpoint 1 / ρ 1/\rho 1/ ρ , and A 2 ⊆ D 2 A_2\subseteq D_2 A 2 ⊆ D 2 . By the definition of A 2 A_2 A 2 , ( 1 , ρ ] ⊆ C 2 (1,\rho]\subseteq C_2 ( 1 , ρ ] ⊆ C 2 , and these points lie in X X X . □
Theorem S2.7 combines this inclusion with the exclusion proved in Section S2.4.
Finite checks delegated to Proposition S4.8. The closed return cover, finite entry scales 0 ≤ m ≤ 8 0\le m\le8 0 ≤ m ≤ 8 , limiting phase cover and algebraic bracket for ρ \rho ρ are the four exact obligations in that record. The entry and phase checks are made at r = ρ r=\rho r = ρ over Q ( ρ ) \mathbb{Q}(\rho) Q ( ρ ) . The estimates for all scales m ≥ 9 m\ge9 m ≥ 9 are proved above.
Residual planar separators
This section proves Proposition 5.1, which supplies the disconnectedness cases outside the ranges of Theorem B. The required heights are q 2 − 2 ≤ h ≤ q 2 q^2-2\le h\le q^2 q 2 − 2 ≤ h ≤ q 2 for q ≥ 3 q\ge3 q ≥ 3 , together with h = 1 , 3 , 4 h=1,3,4 h = 1 , 3 , 4 .
Necessary bounds and exact exclusion
By Proposition 2.3, c ∈ X c\in X c ∈ X lies in Z h e x t Z_h^{\mathrm{ext}} Z h ext if and only if c ∈ T h ( c ) = { ∑ j ≥ 0 b j c − j : b j ∈ B h } c\in T_h(c)=\left\{\sum_{j\ge0}b_jc^{-j}:b_j\in B_h\right\} c ∈ T h ( c ) = { ∑ j ≥ 0 b j c − j : b j ∈ B h } , a bounded set. An inverse orbit z 0 = c z_0=c z 0 = c , z j = c ( z j − 1 − d j ) z_j=c(z_{j-1}-d_j) z j = c ( z j − 1 − d j ) , d j ∈ B h d_j\in B_h d j ∈ B h , is admissible if all z j z_j z j lie in T h ( c ) T_h(c) T h ( c ) . As c ( z − b 0 ) = ∑ j ≥ 0 b j + 1 c − j c(z-b_0)=\sum_{j\ge0}b_{j+1}c^{-j} c ( z − b 0 ) = ∑ j ≥ 0 b j + 1 c − j for z = ∑ j ≥ 0 b j c − j z=\sum_{j\ge0}b_jc^{-j} z = ∑ j ≥ 0 b j c − j , and c = ∑ j = 1 m d j c 1 − j + c − m z m c=\sum_{j=1}^{m}d_jc^{1-j}+c^{-m}z_m c = ∑ j = 1 m d j c 1 − j + c − m z m , the point c c c lies in Z h e x t Z_h^{\mathrm{ext}} Z h ext exactly when it has an admissible inverse orbit.
Let c = x + i y c=x+iy c = x + i y with x > 1 x>1 x > 1 , y ≠ 0 y\ne0 y = 0 and R = ∣ c ∣ 2 R=|c|^2 R = ∣ c ∣ 2 , and put V c ( z ) = Im z / y V_c(z)=\operatorname{Im}z/y V c ( z ) = Im z / y and S c ( z ) = Im ( c z ) / y S_c(z)=\operatorname{Im}(cz)/y S c ( z ) = Im ( cz ) / y in the dynamical plane; the marked point z = c z=c z = c has ( S 0 , V 0 ) = ( 2 x , 1 ) (S_0,V_0)=(2x,1) ( S 0 , V 0 ) = ( 2 x , 1 ) . For each admissible orbit, let z u ( c ) z_u(c) z u ( c ) be the half point z u ( c ) z_u(c) z u ( c ) defined in Section S1 for u = ( d 1 , … , d j ) u=(d_1,\ldots,d_j) u = ( d 1 , … , d j ) , so V c ( z j ) = V c ( z u ( c ) ) V_c(z_j)=V_c(z_u(c)) V c ( z j ) = V c ( z u ( c )) and S c ( z j ) = ∣ c ∣ S c ( z u ( c ) ) S_c(z_j)=|c|S_c(z_u(c)) S c ( z j ) = ∣ c ∣ S c ( z u ( c )) in the notation of (S1.6). If z ′ = c ( z − d ) z'=c(z-d) z ′ = c ( z − d ) , then Im z ′ = Im ( c z ) − d y \operatorname{Im}z'=\operatorname{Im}(cz)-dy Im z ′ = Im ( cz ) − d y and c z ′ = c 2 ( z − d ) = 2 x z ′ − R ( z − d ) cz'=c^2(z-d)=2xz'-R(z-d) c z ′ = c 2 ( z − d ) = 2 x z ′ − R ( z − d ) , as c 2 = 2 x c − R c^2=2xc-R c 2 = 2 x c − R . Hence the inverse digit map acts by
V j + 1 = S j − d j + 1 , S j + 1 = 2 x V j + 1 − R V j . (108) V_{j+1}=S_j-d_{j+1},\qquad S_{j+1}=2xV_{j+1}-RV_j. \tag*{(108)} V j + 1 = S j − d j + 1 , S j + 1 = 2 x V j + 1 − R V j . (108)
For a fixed word these are polynomials in ( x , R ) (x,R) ( x , R ) , which extend to y = 0 y=0 y = 0 with R = x 2 R=x^2 R = x 2 .
Lemma S3.1 (Outer support bounds). For z ∈ T h ( c ) z\in T_h(c) z ∈ T h ( c ) ,
∣ V c ( z ) ∣ ≤ β c , ∣ S c ( z ) ∣ ≤ h + β c , β c = h ( ∣ c ∣ − 1 ) 2 . |V_c(z)|\le\beta_c,\qquad|S_c(z)|\le h+\beta_c,\qquad\beta_c=\frac{h}{(|c|-1)^2}. ∣ V c ( z ) ∣ ≤ β c , ∣ S c ( z ) ∣ ≤ h + β c , β c = ( ∣ c ∣ − 1 ) 2 h . Consequently an admissible inverse orbit satisfies
∣ V j ∣ ≤ β : = h ( x − 1 ) 2 , ∣ S j ∣ ≤ h + β . (109) |V_j|\le\beta:=\frac{h}{(x-1)^2},\qquad|S_j|\le h+\beta. \tag*{(109)} ∣ V j ∣ ≤ β := ( x − 1 ) 2 h , ∣ S j ∣ ≤ h + β . (109) Moreover, c ∈ Z h e x t ∖ R c\in Z_h^{\mathrm{ext}}\setminus\mathbb{R} c ∈ Z h ext ∖ R implies ∣ c ∣ ≤ 1 + h |c|\le1+\sqrt{h} ∣ c ∣ ≤ 1 + h .
Proof. With w = 1 / c w=1/c w = 1/ c , the identity w k − w ‾ k = ( w − w ‾ ) ∑ ℓ = 0 k − 1 w ℓ w ‾ k − 1 − ℓ w^k-\overline{w}^k=(w-\overline{w})\sum_{\ell=0}^{k-1}w^\ell\overline{w}^{k-1-\ell} w k − w k = ( w − w ) ∑ ℓ = 0 k − 1 w ℓ w k − 1 − ℓ gives ∣ Im ( c − k ) ∣ ≤ k ∣ y ∣ ∣ c ∣ − k − 1 |\operatorname{Im}(c^{-k})|\le k|y||c|^{-k-1} ∣ Im ( c − k ) ∣ ≤ k ∣ y ∣∣ c ∣ − k − 1 . Summing over the digits of z z z , with ∑ k ≥ 1 k ∣ c ∣ − k − 1 = ( ∣ c ∣ − 1 ) − 2 \sum_{k\ge1}k|c|^{-k-1}=(|c|-1)^{-2} ∑ k ≥ 1 k ∣ c ∣ − k − 1 = ( ∣ c ∣ − 1 ) − 2 , gives the V V V bound and the S S S bound, in which b 0 b_0 b 0 adds b 0 Im c / y = b 0 b_0\operatorname{Im}c/y=b_0 b 0 Im c / y = b 0 . As ∣ c ∣ ≥ x |c|\ge x ∣ c ∣ ≥ x , (S3.2) follows. Finally, V 0 = 1 V_0=1 V 0 = 1 in an admissible orbit, so 1 ≤ h / ( ∣ c ∣ − 1 ) 2 1\le h/(|c|-1)^2 1 ≤ h / ( ∣ c ∣ − 1 ) 2 .
Strict inequalities exclude all admissible continuations through m m m steps at c 0 c_0 c 0 if each ( d 1 , … , d m ) ∈ B h m (d_1,\ldots,d_m)\in B_h^m ( d 1 , … , d m ) ∈ B h m has some 0 ≤ j ≤ m 0\le j\le m 0 ≤ j ≤ m with ∣ V j ∣ > β c |V_j|>\beta_c ∣ V j ∣ > β c or ∣ S j ∣ > h + β |S_j|>h+\beta ∣ S j ∣ > h + β , the coordinates (S3.1) being taken at ( x , R ) = ( Re c 0 , ∣ c 0 ∣ 2 ) (x,R)=(\operatorname{Re}c_0,|c_0|^2) ( x , R ) = ( Re c 0 , ∣ c 0 ∣ 2 ) ; β \beta β may replace β c \beta_c β c in the first test.
Lemma S3.2 (Strict finite exclusion and real contacts). Suppose strict inequalities in the continuous variables ( x , R ) (x,R) ( x , R ) exclude all admissible continuations of (S3.1) through finitely many steps at a parameter c 0 c_0 c 0 , with Re c 0 > 1 \operatorname{Re}c_0>1 Re c 0 > 1 . Then a neighbourhood of c 0 c_0 c 0 contains no nonreal points of Z h e x t \mathcal{Z}^{\mathrm{ext}}_h Z h ext . In particular, c 0 ∉ C h e x t c_0\notin\mathcal{C}^{\mathrm{ext}}_h c 0 ∈ / C h ext . This remains true when c 0 c_0 c 0 is real.
Proof. Finitely many words are involved, and the excluding inequalities are strict and continuous in ( x , R ) (x,R) ( x , R ) for x > 1 x>1 x > 1 , so they persist at ( Re c , ∣ c ∣ 2 ) (\operatorname{Re}c,|c|^2) ( Re c , ∣ c ∣ 2 ) for c c c near c 0 c_0 c 0 , with R = x 2 R=x^2 R = x 2 if c 0 c_0 c 0 is real. At nonreal such c c c , Lemma S3.1 is necessary for an admissible inverse orbit, so none exists. As C h e x t = cl X ( Z h e x t ∖ R ) \mathcal{C}^{\mathrm{ext}}_h=\operatorname{cl}_X(\mathcal{Z}^{\mathrm{ext}}_h\setminus\mathbb{R}) C h ext = cl X ( Z h ext ∖ R ) by (S1.1), c 0 ∉ C h e x t c_0\notin\mathcal{C}^{\mathrm{ext}}_h c 0 ∈ / C h ext .
We call R + 2 x < N R+2x<N R + 2 x < N the lens condition ; in these coordinates Proposition S1.4 reads as follows.
Lemma S3.3 (Strict inner trap). Put N = 2 h + 1 N=2h+1 N = 2 h + 1 . Assume R + 2 x < N R+2x<N R + 2 x < N , and define
τ h ( c ) = N − 2 x 2 R > 1 2 . \tau_h(c)=\frac{N-2x}{2R}>\frac{1}{2}. τ h ( c ) = 2 R N − 2 x > 2 1 . If a finite word brings the initial state ( 2 x , 1 ) (2x,1) ( 2 x , 1 ) into
∣ V ∣ < τ h ( c ) , ∣ S ∣ < N / 2 (110) |V|<\tau_h(c),\qquad|S|<N/2 \tag*{(110)} ∣ V ∣ < τ h ( c ) , ∣ S ∣ < N /2 (110) then c c c is an interior point of Z h e x t \mathcal{Z}^{\mathrm{ext}}_h Z h ext . The conclusion also applies to a real c = x > 1 c=x>1 c = x > 1 when the continued finite-word coordinates satisfy the strict inequalities.
Proof. After a word u u u the orbit point is c P u ( c ) = 1 2 z u ( c ) cP_u(c)=\frac{1}{2}z_u(c) c P u ( c ) = 2 1 z u ( c ) , so, in the coordinates of Lemma S1.3, V = V c ( z u ( c ) ) V=V_c(z_u(c)) V = V c ( z u ( c )) and S = ∣ c ∣ S c ( z u ( c ) ) S=|c|S_c(z_u(c)) S = ∣ c ∣ S c ( z u ( c )) , and these are the inequalities of Proposition S1.4. At a real x x x the continued values are V u ( x ) V_u(x) V u ( x ) and x S u ( x ) xS_u(x) x S u ( x ) (Lemma S1.3), τ h ( x ) \tau_h(x) τ h ( x ) is (S1.4), and ∣ S ∣ < N / 2 |S|<N/2 ∣ S ∣ < N /2 reads ∣ S u ( x ) ∣ < σ h ( x ) |S_u(x)|<\sigma_h(x) ∣ S u ( x ) ∣ < σ h ( x ) : this is the real case.
Lemma S3.4 (Separation by an excluded boundary). Let K = C h e x t K=\mathcal{C}^{\mathrm{ext}}_h K = C h ext . If an open set U U U meets K K K , ∂ U ∩ K = ∅ \partial U\cap K=\varnothing ∂ U ∩ K = ∅ , and U ‾ ⊂ { Re c > 0 } \overline{U}\subset\{\operatorname{Re}c>0\} U ⊂ { Re c > 0 } , then K K K has at least three components.
Proof. The sets K ∩ U K\cap U K ∩ U and K ∩ ( − U ) K\cap(-U) K ∩ ( − U ) are disjoint nonempty relatively clopen sets, since K ∩ U = K ∩ U ‾ K\cap U=K\cap\overline{U} K ∩ U = K ∩ U . A point i y iy i y of the central annulus, 1 < y < h + 1 1<y<\sqrt{h+1} 1 < y < h + 1 , lies in neither. Every component lies in one of these three disjoint relatively clopen sets.
An elementary separator at height seven.
Proposition S3.5 (Height seven). The lines Re c = ± 16 / 5 \operatorname{Re}c=\pm16/5 Re c = ± 16/5 miss C 7 e x t \mathcal{C}^{\mathrm{ext}}_7 C 7 ext . The exterior core C 7 e x t \mathcal{C}^{\mathrm{ext}}_7 C 7 ext has at least three connected components.
Proof. Idea. Only the first digits 5 and 6 survive, and the second digit fails a bound, except off the axis, where β c \beta_c β c rejects 5.
Set x = 16 / 5 x=16/5 x = 16/5 , so R = 256 / 25 + y 2 R=256/25+y^2 R = 256/25 + y 2 , β = 7 / ( x − 1 ) 2 = 175 / 121 \beta=7/(x-1)^2=175/121 β = 7/ ( x − 1 ) 2 = 175/121 and 7 + β = 1022 / 121 7+\beta=1022/121 7 + β = 1022/121 ; the bound β c = 7 / ( ∣ c ∣ − 1 ) 2 \beta_c=7/(|c|-1)^2 β c = 7/ ( ∣ c ∣ − 1 ) 2 of Lemma S3.1 satisfies β c ≤ β \beta_c\le\beta β c ≤ β . By (S3.1), V 1 = 32 / 5 − d 1 V_1=32/5-d_1 V 1 = 32/5 − d 1 and S 1 = 768 / 25 − 32 d 1 / 5 − y 2 S_1=768/25-32d_1/5-y^2 S 1 = 768/25 − 32 d 1 /5 − y 2 ; digits d 1 ≤ 4 d_1\le4 d 1 ≤ 4 give V 1 ≥ 12 / 5 > 175 / 121 V_1\ge12/5>175/121 V 1 ≥ 12/5 > 175/121 , and d 1 = 5 d_1=5 d 1 = 5 gives S 1 ≤ − 352 / 25 − y 2 < − 1022 / 121 S_1\le-352/25-y^2<-1022/121 S 1 ≤ − 352/25 − y 2 < − 1022/121 . If d 1 = 6 d_1=6 d 1 = 6 , then V 1 = 2 / 5 V_1=2/5 V 1 = 2/5 , S 1 = − 192 / 25 − y 2 S_1=-192/25-y^2 S 1 = − 192/25 − y 2 , and d 2 ≥ − 7 d_2\ge-7 d 2 ≥ − 7 ,
S 2 = − 6656 125 − 32 5 d 2 − 34 5 y 2 ≤ − 1056 125 − 34 5 y 2 , 1056 125 − 1022 121 = 26 15125 > 0. S_2=-\frac{6656}{125}-\frac{32}{5}d_2-\frac{34}{5}y^2
\le-\frac{1056}{125}-\frac{34}{5}y^2,\qquad
\frac{1056}{125}-\frac{1022}{121}=\frac{26}{15125}>0. S 2 = − 125 6656 − 5 32 d 2 − 5 34 y 2 ≤ − 125 1056 − 5 34 y 2 , 125 1056 − 121 1022 = 15125 26 > 0.
If d 1 = 5 d_1=5 d 1 = 5 , then V 1 = 7 / 5 ≤ β c V_1=7/5\le\beta_c V 1 = 7/5 ≤ β c forces ( ∣ c ∣ − 1 ) 2 ≤ 5 (|c|-1)^2\le5 ( ∣ c ∣ − 1 ) 2 ≤ 5 , which fails strictly if y 2 ≥ 161 / 625 y^2\ge161/625 y 2 ≥ 161/625 : then ∣ c ∣ 2 ≥ 6561 / 625 |c|^2\ge6561/625 ∣ c ∣ 2 ≥ 6561/625 , so ∣ c ∣ − 1 ≥ 56 / 25 |c|-1\ge56/25 ∣ c ∣ − 1 ≥ 56/25 and ( 56 / 25 ) 2 = 3136 / 625 > 5 (56/25)^2=3136/625>5 ( 56/25 ) 2 = 3136/625 > 5 . Hence y 2 < 161 / 625 y^2<161/625 y 2 < 161/625 , and as S 1 = − 32 / 25 − y 2 S_1=-32/25-y^2 S 1 = − 32/25 − y 2 , digits d 2 ≤ − 3 d_2\le-3 d 2 ≤ − 3 give V 2 = − 32 / 25 − y 2 − d 2 > 914 / 625 > 175 / 121 V_2=-32/25-y^2-d_2>914/625>175/121 V 2 = − 32/25 − y 2 − d 2 > 914/625 > 175/121 , and d 2 ≥ − 2 d_2\ge-2 d 2 ≥ − 2 give
S 2 = − 2816 125 − 32 5 d 2 − 39 5 y 2 ≤ − 1216 125 < − 1022 121 . S_2=-\frac{2816}{125}-\frac{32}{5}d_2-\frac{39}{5}y^2\le-\frac{1216}{125}<-\frac{1022}{121}. S 2 = − 125 2816 − 5 32 d 2 − 5 39 y 2 ≤ − 125 1216 < − 121 1022 .
So a strict inequality excludes every two-digit word, and by Lemma S3.2 the line misses C 7 e x t C^{\mathrm{ext}}_7 C 7 ext ; negation gives Re c = − 16 / 5 \operatorname{Re}c=-16/5 Re c = − 16/5 .
The normalized series
1 − 6 z + 7 z 2 1 − z = 1 − 7 z + 13 z 2 1 − z 1-6z+\frac{7z^2}{1-z}=\frac{1-7z+13z^2}{1-z} 1 − 6 z + 1 − z 7 z 2 = 1 − z 1 − 7 z + 13 z 2
has height seven and the nonreal zero γ 3 − 1 \gamma_3^{-1} γ 3 − 1 , where γ 3 = ( 7 + i 3 ) / 2 \gamma_3=(7+i\sqrt{3})/2 γ 3 = ( 7 + i 3 ) /2 . Polynomial truncation and inversion give γ 3 ∈ C 7 e x t \gamma_3\in C^{\mathrm{ext}}_7 γ 3 ∈ C 7 ext . The two excluded vertical lines leave nonempty clopen portions containing γ 3 \gamma_3 γ 3 , − γ 3 -\gamma_3 − γ 3 , and the central annulus, respectively. □
Uniform and finite near-square domains
Proof of Proposition 5.1. For q ≥ 22 q\ge22 q ≥ 22 , Proposition S4.9 gives the domain Ω q = Φ ( D q ) \Omega_q=\Phi(D_q) Ω q = Φ ( D q ) , where Φ ( v , Y ) = X q ( v , Y 2 ) + i Y / q \Phi(v,Y)=X_q(v,Y^2)+iY/\sqrt{q} Φ ( v , Y ) = X q ( v , Y 2 ) + iY / q . Its boundary misses C q 2 C_{q^2} C q 2 and its interior meets C q 2 − 2 C_{q^2-2} C q 2 − 2 . Since ∂ X q / ∂ v = ( q − 2 ) / ( 2 q 3 ) > 0 \partial X_q/\partial v=(q-2)/(2q^3)>0 ∂ X q / ∂ v = ( q − 2 ) / ( 2 q 3 ) > 0 , the map is a homeomorphism of the plane and takes ∂ D q \partial D_q ∂ D q to ∂ Ω q \partial\Omega_q ∂ Ω q . The defining bounds give Ω ‾ q ⊂ { Re c > q } \overline{\Omega}_q\subset\{\operatorname{Re}c>q\} Ω q ⊂ { Re c > q } . Nesting and the separation lemma prove all three heights. For 5 ≤ q ≤ 21 5\le q\le21 5 ≤ q ≤ 21 , Proposition S4.10 gives the corresponding bounded domains with closure in the positive half-plane. For q = 3 , 4 q=3,4 q = 3 , 4 , height seven was proved above and Proposition S4.11 supplies heights 8 , 9 , 14 , 15 , 16 8,9,14,15,16 8 , 9 , 14 , 15 , 16 . The same proposition supplies 1 , 3 , 4 1,3,4 1 , 3 , 4 . Every supplied domain has an excluded whole boundary and a certified core point inside it, so the separation lemma applies in each case. □
Exact finite certificates
Every record below states a finite exact obligation used in the preceding proofs. The acceptance rules of the checks are proved here. The integer and rational coefficients, cell covers and finite word lists that the checks consume are supplied in the computational companion [13 ] , described in Section S4.3. All estimates used at infinitely many scales are proved in Section S2; a finite numerical cutoff does not replace them. Every finite acceptance decision uses exact arithmetic.
Reduction lemmas
Lemma S4.1 (Positivity on a half-line times a box). Let P ( q , t 1 , … , t r ) P(q,t_1,\ldots,t_r) P ( q , t 1 , … , t r ) be a rational polynomial. After setting q = q 0 + s q=q_0+s q = q 0 + s , write
P ( q 0 + s , t ) = ∑ k = 0 d s k ∑ α ≤ m b k , α ∏ j = 1 r ( m j α j ) t j α j ( 1 − t j ) m j − α j . P(q_0+s,t)=\sum_{k=0}^{d}s^k\sum_{\alpha\le m}b_{k,\alpha}\prod_{j=1}^{r}\binom{m_j}{\alpha_j}t_j^{\alpha_j}(1-t_j)^{m_j-\alpha_j}. P ( q 0 + s , t ) = k = 0 ∑ d s k α ≤ m ∑ b k , α j = 1 ∏ r ( α j m j ) t j α j ( 1 − t j ) m j − α j . If every b k , α ≥ 0 b_{k,\alpha}\ge0 b k , α ≥ 0 and every b 0 , α > 0 b_{0,\alpha}>0 b 0 , α > 0 , then P > 0 P>0 P > 0 for all q ≥ q 0 q\ge q_0 q ≥ q 0 and t ∈ [ 0 , 1 ] r t\in[0,1]^r t ∈ [ 0 , 1 ] r .
Proof. The Bernstein basis functions are nonnegative on the box and sum to one, so the s 0 s^0 s 0 term is at least min α b 0 , α > 0 \min_\alpha b_{0,\alpha}>0 min α b 0 , α > 0 ; the other terms are nonnegative for s ≥ 0 s\ge0 s ≥ 0 . □
In one variable, p ( t ) = ∑ j a j t j p(t)=\sum_j a_jt^j p ( t ) = ∑ j a j t j may instead be bounded through its coefficients:
p ( t ) ≥ a 0 + ∑ j ≥ 1 min { a j , 0 } ( 0 ≤ t ≤ 1 ) , p ( t ) ≥ a 0 − ∑ j ≥ 1 ∣ a j ∣ ( − 1 ≤ t ≤ 1 ) . (111) p(t)\ge a_0+\sum_{j\ge1}\min\{a_j,0\}\quad(0\le t\le1),\qquad p(t)\ge a_0-\sum_{j\ge1}|a_j|\quad(-1\le t\le1). \tag*{(111)} p ( t ) ≥ a 0 + j ≥ 1 ∑ min { a j , 0 } ( 0 ≤ t ≤ 1 ) , p ( t ) ≥ a 0 − j ≥ 1 ∑ ∣ a j ∣ ( − 1 ≤ t ≤ 1 ) . (111)
Before either test, margins are multiplied only by explicitly positive factors.
For a word u u u over B h B_h B h , let ( S u , V u ) (S_u,V_u) ( S u , V u ) be the coordinates (S3.1) after the digits of u u u , starting from ( 2 x , 1 ) (2x,1) ( 2 x , 1 ) ; the extensions of u u u are the words u d ud u d , d ∈ B h d\in B_h d ∈ B h .
Lemma S4.2 (Digit-tree exclusion). Let h ≥ 1 h\ge1 h ≥ 1 , let Λ ⊂ { Re c > 1 } \Lambda\subset\{\operatorname{Re}c>1\} Λ ⊂ { Re c > 1 } , and let T T T be a finite set of words over B h B_h B h that contains the empty word and every prefix of its members. Suppose that at every c ∈ Λ c\in\Lambda c ∈ Λ , with ( x , R ) = ( Re c , ∣ c ∣ 2 ) (x,R)=(\operatorname{Re}c,|c|^2) ( x , R ) = ( Re c , ∣ c ∣ 2 ) and β = h / ( x − 1 ) 2 \beta=h/(x-1)^2 β = h / ( x − 1 ) 2 , every u ∈ T u\in T u ∈ T satisfies:
(i) if no extension of u u u lies in T T T , then ∣ V u ∣ > β |V_u|>\beta ∣ V u ∣ > β or ∣ S u ∣ > h + β |S_u|>h+\beta ∣ S u ∣ > h + β ;
(ii) otherwise, ∣ S u − d ∣ > β |S_u-d|>\beta ∣ S u − d ∣ > β for every d ∈ B h d\in B_h d ∈ B h with u d ∉ T ud\notin T u d ∈ / T .
Then Λ ∩ C h ext = ∅ \Lambda\cap C_h^{\operatorname{ext}}=\varnothing Λ ∩ C h ext = ∅ .
Proof. Fix c ∈ Λ c\in\Lambda c ∈ Λ and ( d 1 , … , d m ) ∈ B h m (d_1,\ldots,d_m)\in B_h^m ( d 1 , … , d m ) ∈ B h m , where m m m is the largest length of a word in T T T , and let u u u be its longest prefix in T T T . If u u u has no extension in T T T , (i) violates (S3.2) strictly; otherwise the next digit d d d has u d ∉ T ud\notin T u d ∈ / T , and (ii) gives ∣ V u d ∣ = ∣ S u − d ∣ > β |V_{ud}|=|S_u-d|>\beta ∣ V u d ∣ = ∣ S u − d ∣ > β . So strict inequalities exclude every admissible continuation through m m m steps, and Lemma S3.2 gives c ∉ C h ext c\notin C_h^{\operatorname{ext}} c ∈ / C h ext . □
In the records, Λ \Lambda Λ is a closed cell of an arc t ↦ x ( t ) + i y ( t ) t\mapsto x(t)+iy(t) t ↦ x ( t ) + i y ( t ) , t ∈ I t\in I t ∈ I , on which x x x and y 2 y^2 y 2 , hence R R R and ( x − 1 ) 2 (x-1)^2 ( x − 1 ) 2 , are rational polynomials. After multiplication by ( x − 1 ) 2 (x-1)^2 ( x − 1 ) 2 , (i) is the positivity on I I I of one of the exit polynomials ± V u ( x − 1 ) 2 − h \pm V_u(x-1)^2-h ± V u ( x − 1 ) 2 − h and ± S u ( x − 1 ) 2 − h ( ( x − 1 ) 2 + 1 ) \pm S_u(x-1)^2-h((x-1)^2+1) ± S u ( x − 1 ) 2 − h (( x − 1 ) 2 + 1 ) . For (ii), let the extensions of u u u in T T T be the u d ud u d with lo ≤ d < up \operatorname{lo}\le d<\operatorname{up} lo ≤ d < up . One may check ( S u − lo + 1 ) ( x − 1 ) 2 − h > 0 (S_u-\operatorname{lo}+1)(x-1)^2-h>0 ( S u − lo + 1 ) ( x − 1 ) 2 − h > 0 if lo > − h \operatorname{lo}>-h lo > − h and ( up − S u ) ( x − 1 ) 2 − h > 0 (\operatorname{up}-S_u)(x-1)^2-h>0 ( up − S u ) ( x − 1 ) 2 − h > 0 if up ≤ h \operatorname{up}\le h up ≤ h . The checks take [ lo , up − 1 ] = [ − h , h ] ∩ [ ⌈ S ∗ − h / H ∗ ⌉ , ⌊ S ∗ + h / H ∗ ⌋ ] [\operatorname{lo},\operatorname{up}-1]=[-h,h]\cap[\lceil S_*-h/H_*\rceil,\lfloor S^*+h/H_*\rfloor] [ lo , up − 1 ] = [ − h , h ] ∩ [⌈ S ∗ − h / H ∗ ⌉ , ⌊ S ∗ + h / H ∗ ⌋] , where S ∗ ≤ S u ≤ S ∗ S_*\le S_u\le S^* S ∗ ≤ S u ≤ S ∗ and ( x − 1 ) 2 ≥ H ∗ > 0 (x-1)^2\ge H_*>0 ( x − 1 ) 2 ≥ H ∗ > 0 on I I I . Positivity on I I I is certified by Lemma S4.1 or (S4.1), after an affine change of the cell variable.
Lemma S4.3 (Rouché discs). Let P P P be a polynomial, z 0 ∈ C z_0\in\mathbb{C} z 0 ∈ C , ε > 0 \varepsilon>0 ε > 0 , P ( z 0 + w ) = ∑ k A k w k P(z_0+w)=\sum_k A_kw^k P ( z 0 + w ) = ∑ k A k w k and ∣ a ∣ 1 = ∣ Re a ∣ + ∣ Im a ∣ |a|_1=|\operatorname{Re}a|+|\operatorname{Im}a| ∣ a ∣ 1 = ∣ Re a ∣ + ∣ Im a ∣ . If ∣ A 0 ∣ 1 + ∑ k ≥ 2 ∣ A k ∣ 1 ε k < ε max { ∣ Re A 1 ∣ , ∣ Im A 1 ∣ } |A_0|_1+\sum_{k\ge2}|A_k|_1\varepsilon^k<\varepsilon\max\{|\operatorname{Re}A_1|,|\operatorname{Im}A_1|\} ∣ A 0 ∣ 1 + ∑ k ≥ 2 ∣ A k ∣ 1 ε k < ε max { ∣ Re A 1 ∣ , ∣ Im A 1 ∣ } , then P P P has exactly one zero in ∣ z − z 0 ∣ < ε |z-z_0|<\varepsilon ∣ z − z 0 ∣ < ε , and it is simple; if moreover P ∈ P h P\in P_h P ∈ P h and the closed disc lies in { Im c > 0 } \{\operatorname{Im}c>0\} { Im c > 0 } , this zero belongs to C h C_h C h .
Proof. As ∣ a ∣ ≤ ∣ a ∣ 1 |a|\le|a|_1 ∣ a ∣ ≤ ∣ a ∣ 1 and ∣ A 1 ∣ ≥ max { ∣ Re A 1 ∣ , ∣ Im A 1 ∣ } |A_1|\ge\max\{|\operatorname{Re}A_1|,|\operatorname{Im}A_1|\} ∣ A 1 ∣ ≥ max { ∣ Re A 1 ∣ , ∣ Im A 1 ∣ } , the hypothesis gives ∣ P ( z 0 + w ) − A 1 w ∣ < ∣ A 1 w ∣ |P(z_0+w)-A_1w|<|A_1w| ∣ P ( z 0 + w ) − A 1 w ∣ < ∣ A 1 w ∣ on ∣ w ∣ = ε |w|=\varepsilon ∣ w ∣ = ε ; apply Rouché’s theorem. A zero in the upper half-plane lies in R h ∖ R ⊂ C h R_h\setminus\mathbb{R}\subset C_h R h ∖ R ⊂ C h . □
The records.
Proposition S4.4 (Certified: short-word strip). Let q ≥ 4 q\ge4 q ≥ 4 be real, read the words A m , B m , U m , ℓ A_m,B_m,U_{m,\ell} A m , B m , U m , ℓ of § 15.2 with real digits, and put a = q − 1 a=q-1 a = q − 1 , b = ( q − 1 ) ( q − 2 ) b=(q-1)(q-2) b = ( q − 1 ) ( q − 2 ) , h = q ( q − 1 ) h=q(q-1) h = q ( q − 1 ) and N = 2 h + 1 N=2h+1 N = 2 h + 1 . For each word u u u below, each δ \delta δ in its interval and x = q − δ x=q-\delta x = q − δ , the four margins
N − 2 x ± 2 x 2 V u ( x ) , N ± 2 x S u ( x ) (112) N-2x\pm2x^2V_u(x),\qquad N\pm2xS_u(x) \tag*{(112)} N − 2 x ± 2 x 2 V u ( x ) , N ± 2 x S u ( x ) (112) are positive, with V u , S u V_u,S_u V u , S u as in (S1.2):
Verification. Put δ = δ − + ( δ + − δ − ) t \delta=\delta_-+(\delta_+-\delta_-)t δ = δ − + ( δ + − δ − ) t with 0 ≤ t ≤ 1 0\le t\le1 0 ≤ t ≤ 1 . Each margin, multiplied by a positive constant times a power of q q q and written in q = 4 + s q=4+s q = 4 + s , is a polynomial in ( s , t ) (s,t) ( s , t ) , positive by Lemma S4.1 with q 0 = 4 q_0=4 q 0 = 4 and r = 1 r=1 r = 1 . The checker computes its Bernstein coefficients, whose constant row satisfies b 0 i ≥ p ( 0 , 0 ) ⋅ 4231364608 / 127302260011 > p ( 0 , 0 ) / 32 b_{0i}\ge p(0,0)\cdot4231364608/127302260011>p(0,0)/32 b 0 i ≥ p ( 0 , 0 ) ⋅ 4231364608/127302260011 > p ( 0 , 0 ) /32 . Spot check: for q = 4 q=4 q = 4 , u = B 2 = ( 7 , − 10 ) u=B_2=(7,-10) u = B 2 = ( 7 , − 10 ) and δ = 1 / 4 \delta=1/4 δ = 1/4 , so x = 15 / 4 x=15/4 x = 15/4 : P u ( x ) = − 35 / 16 P_u(x)=-35/16 P u ( x ) = − 35/16 , V u ( x ) = − 5 / 16 V_u(x)=-5/16 V u ( x ) = − 5/16 , S u ( x ) = − 5 / 2 S_u(x)=-5/2 S u ( x ) = − 5/2 and N = 25 N=25 N = 25 , and the four margins are 1115 / 128 1115/128 1115/128 , 3365 / 128 3365/128 3365/128 , 25 / 4 25/4 25/4 and 175 / 4 175/4 175/4 . Two independent programs compute
the Bernstein coefficients, one from the stored polynomials and one by rebuilding them from the recurrence.
Companion: record S4.4 of [13 ] .
Proposition S4.5 (Certified: compact windows). Let q ≥ 4 q \ge4 q ≥ 4 be an integer. Every point of the windows J q , 2 J_{q,2} J q , 2 and J q , 3 J_{q,3} J q , 3 of Proposition S2.4 is an interior point of C q ( q − 1 ) C_{q(q-1)} C q ( q − 1 ) .
Verification. Each window is covered by consecutive closed cells, each with a word over B q ( q − 1 ) B_{q(q-1)} B q ( q − 1 ) whose margins (S4.2) are positive on the cell, where also x 2 + 2 x < N x^2+2x<N x 2 + 2 x < N . Proposition S1.4 gives an interior point of Z q ( q − 1 ) e x t Z_{q(q-1)}^{\mathrm{ext}} Z q ( q − 1 ) ext , and int Z h = int C h \operatorname{int} Z_h=\operatorname{int} C_h int Z h = int C h , since an open disc in Z h Z_h Z h minus R \mathbb{R} R lies in the closed set C h C_h C h . For J q , 2 J_{q,2} J q , 2 with q ≤ 43 q \le43 q ≤ 43 and J q , 3 J_{q,3} J q , 3 with q ∈ { 4 , … , 11 , 13 , 15 , … , 23 } q \in\{4,\ldots,11,13,15,\ldots,23\} q ∈ { 4 , … , 11 , 13 , 15 , … , 23 } the cells are rational intervals, positive by exact Bernstein coefficients. For the remaining q q q , rational cells handle 44 ≤ q ≤ 1000 44 \le q \le1000 44 ≤ q ≤ 1000 for J q , 2 J_{q,2} J q , 2 , followed by symbolic cells on the residue classes q = q 0 + 4 k q=q_0+4k q = q 0 + 4 k , q 0 = 1001 , … , 1004 q_0=1001,\ldots,1004 q 0 = 1001 , … , 1004 . For J q , 3 J_{q,3} J q , 3 , symbolic cells start at q 0 = 12 , 25 , 14 , 27 q_0=12,25,14,27 q 0 = 12 , 25 , 14 , 27 . Their endpoints and admissible digits depend polynomially on q q q . Substituting k = y / ( 1 − y ) k=y/(1-y) k = y / ( 1 − y ) and clearing positive denominators gives nonnegative tensor Bernstein coefficients with positive edge coefficients, hence strict positivity for every finite k ≥ 0 k \ge0 k ≥ 0 .
Spot check: for q = 4 q=4 q = 4 , J 4 , 2 = [ 989 / 256 , 995 / 256 ] J_{4,2}=[989/256,995/256] J 4 , 2 = [ 989/256 , 995/256 ] , and its last cell [ 7939 / 2048 , 995 / 256 ] [7939/2048,995/256] [ 7939/2048 , 995/256 ] carries the word ( 7 , − 10 , − 5 ) (7,-10,-5) ( 7 , − 10 , − 5 ) . At x = 97 / 25 x=97/25 x = 97/25 , ( P u , V u ) = ( − 49527 / 15625 , 1592 / 15625 ) (P_u,V_u)=(-49527/15625,1592/15625) ( P u , V u ) = ( − 49527/15625 , 1592/15625 ) , S u = − 9587 / 3125 S_u=-9587/3125 S u = − 9587/3125 and x 2 + 2 x = 14259 / 625 < 25 = N x^2+2x=14259/625<25=N x 2 + 2 x = 14259/625 < 25 = N ; the margins are 198317631 / 9765625 198317631/9765625 198317631/9765625 , 138401119 / 9765625 138401119/9765625 138401119/9765625 , 93247 / 78125 93247/78125 93247/78125 and 3813003 / 78125 3813003/78125 3813003/78125 . The cells, words and Bernstein coefficients of every rational and symbolic case are stored and rechecked.
Companion: record S4.5 of [13 ] .
Proposition S4.6 (Certified: finite terminal checks). Let q ≥ 4 q \ge4 q ≥ 4 be an integer, h = q ( q − 1 ) h=q(q-1) h = q ( q − 1 ) , a = q − 1 a=q-1 a = q − 1 , b = ( q − 1 ) ( q − 2 ) b=(q-1)(q-2) b = ( q − 1 ) ( q − 2 ) , and let B q , m B_{q,m} B q , m be the bands of (S2.11).
(i) For every pair ( q , m ) (q,m) ( q , m ) of the finite range (S2.26), every point of B q , m B_{q,m} B q , m is an interior point of C h C_h C h .
(ii) Every x x x with q − q − 3 ≤ x ≤ q − ( 1 / 2 + 3 / ( 4 q 4 ) ) q − 3 q-q^{-3}\le x\le q-(1/2+3/(4q^4))q^{-3} q − q − 3 ≤ x ≤ q − ( 1/2 + 3/ ( 4 q 4 )) q − 3 is an interior point of C h C_h C h .
(iii) With the movements (S2.25) and the further allowances below, the overlap and containment inequalities in the proof of Proposition S2.19 hold: for q ≥ 8 q\ge8 q ≥ 8 (at m = 11 m=11 m = 11 and the small-depth thresholds) those of the residue chains, fixed seams and annulus targets, the targets of W m , m − 2 W_{m,m-2} W m , m − 2 and U m + 1 , m U_{m+1,m} U m + 1 , m shrunk further by D m D_m D m and q D m qD_m q D m , with R m , 1 , 0 R_{m,1,0} R m , 1 , 0 and R m , 0 , 0 R_{m,0,0} R m , 0 , 0 strictly capturing T = − 1 / 2 T=-1/2 T = − 1/2 after their movements; for 4 ≤ q ≤ 7 4\le q\le7 4 ≤ q ≤ 7 (at m = 11 m=11 m = 11 ) those of Table 4 and of the seams W m , m − 3 ⇝ W m , m − 2 ⇝ W m , m − 1 W_{m,m-3}\rightsquigarrow W_{m,m-2}\rightsquigarrow W_{m,m-1} W m , m − 3 ⇝ W m , m − 2 ⇝ W m , m − 1 and P m + 1 ⇝ U m + 1 , m ⇝ U m + 1 , m − 1 P_{m+1}\rightsquigarrow U_{m+1,m}\rightsquigarrow U_{m+1,m-1} P m + 1 ⇝ U m + 1 , m ⇝ U m + 1 , m − 1 , whose W W W and U U U cells take the single allowance 1000 D 11 1000D_{11} 1000 D 11 and whose P P P cell takes η P , ≤ 2 \eta_{P,\le2} η P , ≤ 2 of (S2.25). In both cases the W m , m − 1 / R m , 1 , 0 W_{m,m-1}/R_{m,1,0} W m , m − 1 / R m , 1 , 0 overlap, so moved, meets T > − 2 / 3 T>-2/3 T > − 2/3 .
Verification. (i) In the finite range m ≤ 10 m\le10 m ≤ 10 , so ( q − x ) x m − 1 (q-x)x^{m-1} ( q − x ) x m − 1 and ( q − x ) x m (q-x)x^m ( q − x ) x m decrease in x x x on [ q − q − 3 , q ) [q-q^{-3},q) [ q − q − 3 , q ) , with derivatives x m − 2 ( ( m − 1 ) ( q − x ) − x ) < 0 x^{m-2}((m-1)(q-x)-x)<0 x m − 2 (( m − 1 ) ( q − x ) − x ) < 0 and x m − 1 ( m ( q − x ) − x ) < 0 x^{m-1}(m(q-x)-x)<0 x m − 1 ( m ( q − x ) − x ) < 0 ; exact endpoint tests enclose B q , m B_{q,m} B q , m in a rational interval, covered by consecutive closed cells, each with a word over B h B_h B h whose margins (S4.2) are positive, and Proposition S1.4 applies as for Proposition S4.5. (ii) The same for 4 ≤ q ≤ 11 4\le q\le11 4 ≤ q ≤ 11 , with the extra word B 4 ( − ( a 2 + 1 ) ) B_4(-(a^2+1)) B 4 ( − ( a 2 + 1 )) at q = 4 q=4 q = 4 only; for q ≥ 12 q\ge12 q ≥ 12 , the words U 4 , j U_{4,j} U 4 , j , 0 ≤ j ≤ 4 0\le j\le4 0 ≤ j ≤ 4 , on five closed cells of the phase q 3 ( q − x ) q^3(q-x) q 3 ( q − x ) , whose cleared margins are positive for all q ≥ 12 q\ge12 q ≥ 12 by Lemma S4.1. (iii) Each inequality compares endpoints of limiting intervals, rational in q q q , moved by the allowances; after clearing positive denominators it has nonnegative coefficients in q − q 0 q-q_0 q − q 0 and a positive constant term on each residue class of q q q modulo 4 from the thresholds of Proposition S2.19; for 4 ≤ q ≤ 7 4\le q\le7 4 ≤ q ≤ 7 the comparisons are rational, at m = 11 m=11 m = 11 . Spot check (iii): at q = 4 q=4 q = 4 the limiting intervals of W m , m − 1 W_{m,m-1} W m , m − 1 and R m , 1 , 0 R_{m,1,0} R m , 1 , 0 are ( − 313 / 384 , − 55 / 96 ) (-313/384,-55/96) ( − 313/384 , − 55/96 ) and ( − 259 / 384 , − 751 / 1536 ) (-259/384,-751/1536) ( − 259/384 , − 751/1536 ) , so their overlap ( − 259 / 384 , − 220 / 384 ) (-259/384,-220/384) ( − 259/384 , − 220/384 ) meets T > − 2 / 3 = − 256 / 384 T>-2/3=-256/384 T > − 2/3 = − 256/384 and lies below − 1 / 2 -1/2 − 1/2 . The checks of (i) run over all pairs of (S2.26), those of (ii) over rational cells for q ≤ 11 q\le11 q ≤ 11 and symbolic cells for q ≥ 12 q\ge12 q ≥ 12 , and those of (iii) inequality by inequality on each residue class. For 4 ≤ q ≤ 7 4\le q\le7 4 ≤ q ≤ 7 the checker derives all 42 printed intervals and tests the four W m , m − 1 / R m , 1 , 0 W_{m,m-1}/R_{m,1,0} W m , m − 1 / R m , 1 , 0 domain seams, 60 formal and 58 domain-clipped consecutive overlaps, and the digit, endpoint, allowance and monotonicity conditions: 389 exact rational comparisons in all.
Companion: record S4.6 of [13 ] .
Proposition S4.7 (Certified: finite checks at height six). Let the digits d ∈ B 6 d \in B_6 d ∈ B 6 act on states by the maps F x , d F_{x,d} F x , d of (S1.8), write F x , u F_{x,u} F x , u for the action of a word u u u , and put r = x − 1 r=x-1 r = x − 1 .
(i) There is a finite cover of [ 137 / 50 , 2999 / 1000 ] [137/50,2999/1000] [ 137/50 , 2999/1000 ] by consecutive closed rational parameter bands I I I . For each band there are a nonempty finite union Y I ⊂ [ − 4 , 4 ] × [ − 2 , 2 ] Y_I \subset[-4,4] \times[-2,2] Y I ⊂ [ − 4 , 4 ] × [ − 2 , 2 ] of closed rational grid cells, a positive rational number η I \eta_I η I , and a digit d C ∈ B 6 d_C \in B_6 d C ∈ B 6 for each selected cell C C C , with
F x , d C ( C ) + [ − η I , η I ] 2 ⊂ Y I ( x ∈ I ) . F_{x,d_C}(C)+[-\eta_I,\eta_I]^2 \subset Y_I \qquad(x \in I). F x , d C ( C ) + [ − η I , η I ] 2 ⊂ Y I ( x ∈ I ) . Moreover, each I I I has a finite partition into closed rational subintervals, each supplied with a word v v v of length at most 11 satisfying
F x , v ( 1 , 1 ) + [ − η I , η I ] 2 ⊂ Y I F_{x,v}(1,1)+[-\eta_I,\eta_I]^2 \subset Y_I F x , v ( 1 , 1 ) + [ − η I , η I ] 2 ⊂ Y I throughout that subinterval.
(ii) For every ( u , w ) (u,w) ( u , w ) in the set Q Q Q of Lemma S2.24 and 2999 / 1000 ≤ x ≤ 3 2999/1000 \le x \le3 2999/1000 ≤ x ≤ 3 , one of the two states there has a word of length at most three whose image satisfies the two bounds displayed there.
(iii) Let x = 3 x=3 x = 3 . The eighteen cells of Table 5 cover [ − 4 / 5 , 0 ] [-4/5,0] [ − 4/5 , 0 ] . Let a cell have the word P M P_M P M , R M , ℓ R_{M,\ell} R M , ℓ or C M C_M C M followed by s s s , and let Y ( Θ ) Y(\Theta) Y ( Θ ) be the limiting state P P P , R ℓ R_\ell R ℓ or C C C of (S2.45). For Θ \Theta Θ in the cell, the images of Y ( Θ ) Y(\Theta) Y ( Θ ) under all prefixes of s s s , the empty one included, lie in [ − 3 , 3 ] × [ − 3 / 2 , 3 / 2 ] [-3,3] \times[-3/2,3/2] [ − 3 , 3 ] × [ − 3/2 , 3/2 ] , and F 3 , s ( Y ( Θ ) ) F_{3,s}(Y(\Theta)) F 3 , s ( Y ( Θ )) in [ − 11 / 8 , 11 / 8 ] × [ − 39 / 80 , 39 / 80 ] [-11/8,11/8] \times[-39/80,39/80] [ − 11/8 , 11/8 ] × [ − 39/80 , 39/80 ] .
(iv) For 2999 / 1000 ≤ x ≤ 344373767 / 114791256 2999/1000 \le x \le344373767/114791256 2999/1000 ≤ x ≤ 344373767/114791256 , some word maps ( 1 , 1 ) (1,1) ( 1 , 1 ) to a state with ∣ P ∣ < 3 / r |P|<3/r ∣ P ∣ < 3/ r and ∣ V ∣ < 2 / r 2 |V|<2/r^2 ∣ V ∣ < 2/ r 2 .
Verification. (i) For the lower bands, a cell indexed by ( i , j ) (i,j) ( i , j ) is [ ( i − 4 N ) / N , ( i + 1 − 4 N ) / N ] × [ ( j − 2 N ) / N , ( j + 1 − 2 N ) / N ] [(i-4N)/N,(i+1-4N)/N] \times[(j-2N)/N,(j+1-2N)/N] [( i − 4 N ) / N , ( i + 1 − 4 N ) / N ] × [( j − 2 N ) / N , ( j + 1 − 2 N ) / N ] ; the stored integer array selects the cells and their digits. For the upper bands the same grid is encoded by consecutive vertical strips, with signed indices. The two coordinates of F x , d ( P , V ) = ( x P − d , x ( P + V ) − d ) F_{x,d}(P,V)=(xP-d,x(P+V)-d) F x , d ( P , V ) = ( x P − d , x ( P + V ) − d ) have extrema at the endpoints of the parameter band and the source rectangle. The verifier rounds these extrema outwards, adds one grid unit in each direction, and checks that every resulting target cell is selected. Thus the displayed inclusion holds with η I = 1 / N \eta_I=1/N η I = 1/ N . For the entry words, direct rational Bernstein bounds enclose P v ( x ) P_v(x) P v ( x ) and V v ( x ) V_v(x) V v ( x ) on each closed entry interval; the same grid-unit enlargement is contained in the selected union. The entry intervals cover their parameter bands, whose exact endpoints join from 137 / 50 137/50 137/50 to 2999 / 1000 2999/1000 2999/1000 . All integer operations are checked against their storage bounds before fixed-width arithmetic is used.
(ii) The paired-cone certificate consists of sixty branch–word regions, each defined by four affine halfspaces in ( u , w ) (u,w) ( u , w ) . The checker reconstructs each halfspace from its branch and its word of length at most three. After multiplication by x 3 ( x − 1 ) 2 x^3(x-1)^2 x 3 ( x − 1 ) 2 , these are polynomial identities of degree at most six; agreement at seven distinct rational parameters verifies them identically. At the rational reference parameter x = 5999 / 2000 x=5999/2000 x = 5999/2000 , it reconstructs the polygons by intersecting the halfspaces and checks their union by a horizontal sweep. The exact signs of all parallel-line comparisons and triple-line determinants remain unchanged for 0 < 3 − x < 1 / 1000 0<3-x<1/1000 0 < 3 − x < 1/1000 : the substitution 3 − x = t / ( 1000 ( 1 + t ) ) 3-x=t/(1000(1+t)) 3 − x = t / ( 1000 ( 1 + t )) , followed by clearing the positive denominator, gives coefficients all of one sign, with at least one nonzero coefficient. The face incidences and halfspace signs of the line arrangement are therefore constant: a change would force one of these determinants or parallel-line comparisons to vanish. Coverage by the same fixed combinations of halfspaces thus persists throughout the open interval. The two endpoints follow by taking limits in the finite collection of closed regions.
(iii) States and images are affine in Θ \Theta Θ , and each inequality is checked at both ends of its cell.
(iv) On each closed rational entry interval the checker reconstructs P v P_v P v from the digit word and V v = P v + x P v ′ V_v=P_v+xP_v' V v = P v + x P v ′ . Rational Bernstein coefficients establish the strict inequalities
3 ( x − 1 ) − ∣ ( x − 1 ) 2 P v ( x ) ∣ > 0 , 2 − ∣ ( x − 1 ) 2 V v ( x ) ∣ > 0. 3(x-1)-|(x-1)^2P_v(x)|>0,\qquad2-|(x-1)^2V_v(x)|>0. 3 ( x − 1 ) − ∣ ( x − 1 ) 2 P v ( x ) ∣ > 0 , 2 − ∣ ( x − 1 ) 2 V v ( x ) ∣ > 0.
Since x > 1 x>1 x > 1 , these are precisely ∣ P v ( x ) ∣ < 3 / ( x − 1 ) |P_v(x)|<3/(x-1) ∣ P v ( x ) ∣ < 3/ ( x − 1 ) and ∣ V v ( x ) ∣ < 2 / ( x − 1 ) 2 |V_v(x)|<2/(x-1)^2 ∣ V v ( x ) ∣ < 2/ ( x − 1 ) 2 . The entry intervals have exact consecutive endpoints and cover [ 2999 / 1000 , 344373767 / 114791256 ] [2999/1000,344373767/114791256] [ 2999/1000 , 344373767/114791256 ] .
Spot check (iii): on the first cell, C ( Θ ) = ( − 9 / 8 , − 51 / 16 − 9 Θ / 2 ) C(\Theta)=(-9/8,-51/16-9\Theta/2) C ( Θ ) = ( − 9/8 , − 51/16 − 9Θ/2 ) has V = 33 / 80 V=33/80 V = 33/80 at Θ = − 4 / 5 \Theta=-4/5 Θ = − 4/5 and V = − 39 / 80 V=-39/80 V = − 39/80 at Θ = − 3 / 5 \Theta=-3/5 Θ = − 3/5 . Being affine in Θ \Theta Θ , it lies in [ − 11 / 8 , 11 / 8 ] × [ − 39 / 80 , 39 / 80 ] [-11/8,11/8] \times[-39/80,39/80] [ − 11/8 , 11/8 ] × [ − 39/80 , 39/80 ] on the whole cell, with margins 3 / 8 3/8 3/8 and 1 / 80 1/80 1/80 in R 0 = [ − 3 / 2 , 3 / 2 ] × [ − 1 / 2 , 1 / 2 ] R_0=[-3/2,3/2] \times[-1/2,1/2] R 0 = [ − 3/2 , 3/2 ] × [ − 1/2 , 1/2 ] . The band covers of (i), the sixty word regions of (ii), the eighteen cells of (iii) and the entry cells of (iv) with their Bernstein coefficients are stored and rechecked.
Companion: record S4.7 of [13 ] .
Proposition S4.8 (Certified: return and entry at height two). Let Q Q Q , ∥ ⋅ ∥ Q \|\cdot\|_{Q} ∥ ⋅ ∥ Q , Θ x , w \Theta_{x,w} Θ x , w , Y m Y_m Y m , s 1 , … , s 8 s_1,\ldots,s_8 s 1 , … , s 8 and μ \mu μ be as in § 15.12 . Put ρ − = 1839286755214161 ⋅ 10 − 15 \rho_{-}=1839286755214161\cdot10^{-15} ρ − = 1839286755214161 ⋅ 1 0 − 15 ; in (ii) and (iii) let r = ρ r=\rho r = ρ and 9 / 80 ≤ τ ≤ 9 r 3 / 80 9/80\leq\tau\leq9r^3/80 9/80 ≤ τ ≤ 9 r 3 /80 .
(i) For x ∈ [ 433 / 250 , 4598217 / 2500000 ] x\in[433/250,4598217/2500000] x ∈ [ 433/250 , 4598217/2500000 ] and Y ∈ Q Y\in Q Y ∈ Q , some word w w w over B 2 B_2 B 2 with 1 ≤ ∣ w ∣ ≤ 10 1\leq|w|\leq10 1 ≤ ∣ w ∣ ≤ 10 has Θ x , w Y ∈ Q \Theta_{x,w}Y\in Q Θ x , w Y ∈ Q .
(ii) If 0 ≤ m ≤ 8 0\leq m\leq8 0 ≤ m ≤ 8 and x = r − τ r − 3 m − 3 x=r-\tau r^{-3m-3} x = r − τ r − 3 m − 3 , some word s s s over B 2 B_2 B 2 has Θ x , s Y m ( x ) ∈ int Q \Theta_{x,s}Y_m(x)\in\operatorname{int}Q Θ x , s Y m ( x ) ∈ int Q .
(iii) If ( S − V , V ) = ( 2 − r 2 , 5 + 2 r − 2 r 2 − ( 16 + 52 r − 32 r 2 ) τ ) (S-V,V)=(2-r^2,5+2r-2r^2-(16+52r-32r^2)\tau) ( S − V , V ) = ( 2 − r 2 , 5 + 2 r − 2 r 2 − ( 16 + 52 r − 32 r 2 ) τ ) for a state Y Y Y , some j j j has ∥ Θ r , s j Y ∥ Q ≤ 1 − μ \|\Theta_{r,s_j}Y\|_{Q}\leq1-\mu ∥ Θ r , s j Y ∥ Q ≤ 1 − μ .
(iv) The polynomial x 3 − x 2 − x − 1 x^3-x^2-x-1 x 3 − x 2 − x − 1 has a zero in [ ρ − , ρ − + 10 − 15 ] [\rho_{-},\rho_{-}+10^{-15}] [ ρ − , ρ − + 1 0 − 15 ] .
Verification. The reductions are printed in § 15.12 : rational enclosures of word-image margins on closed parameter bands, with triangulations of the return diamond, for (i); words a b m s ab^{m}s a b m s on exact algebraic phase cells for (ii); margins affine in τ \tau τ , checked at both ends of eight phase cells, for (iii). Positivity is certified by exact Bernstein bounds over Q ( ρ ) \mathbb{Q}(\rho) Q ( ρ ) in (ii) and (iii); (iv) is the sign of x 3 − x 2 − x − 1 x^3-x^2-x-1 x 3 − x 2 − x − 1 at the two ends. Spot check, in the coordinates ( P , V ) = ( S − V , V ) (P,V)=(S-V,V) ( P , V ) = ( S − V , V ) at x = 9 / 5 x=9/5 x = 9/5 : (i) the vertex ( 3 / 2 , 0 ) (3/2,0) ( 3/2 , 0 ) of Q Q Q is mapped by ( 2 , 2 ) (2,2) ( 2 , 2 ) to ( − 37 / 50 , 13 / 25 ) (-37/50,13/25) ( − 37/50 , 13/25 ) , where ∣ u 1 ∣ + ∣ u 2 ∣ = 13 / 35 + 94 / 375 = 1633 / 2625 < 1 |u_1|+|u_2|=13/35+94/375=1633/2625<1 ∣ u 1 ∣ + ∣ u 2 ∣ = 13/35 + 94/375 = 1633/2625 < 1 ; (ii) here m = 0 m=0 m = 0 , τ = ( ρ − 9 / 5 ) ρ 3 ∈ [ 3 / 20 , 3 / 10 ] \tau=(\rho-9/5)\rho^3\in[3/20,3/10] τ = ( ρ − 9/5 ) ρ 3 ∈ [ 3/20 , 3/10 ] , and the word ( 2 , 1 , − 1 , − 1 , − 2 ) (2,1,-1,-1,-2) ( 2 , 1 , − 1 , − 1 , − 2 ) of the cell containing 9 / 5 9/5 9/5 maps ( 1 , 1 ) (1,1) ( 1 , 1 ) to ( − 2786 / 3125 , 1219 / 3125 ) (-2786/3125,1219/3125) ( − 2786/3125 , 1219/3125 ) , where ∣ u 1 ∣ + ∣ u 2 ∣ = 1219 / 4375 + 19327 / 46875 = 226714 / 328125 < 1 |u_1|+|u_2|=1219/4375+19327/46875=226714/328125<1 ∣ u 1 ∣ + ∣ u 2 ∣ = 1219/4375 + 19327/46875 = 226714/328125 < 1 . The return cover of (i) consists of 135 parameter bands with their triangulations; (ii) and (iv) are checked over Q ( ρ ) \mathbb{Q}(\rho) Q ( ρ ) on closed phase cells, and (iii) by exact capture intervals whose smallest overlap margin, 16347138486163573 / ( 3255 ⋅ 10 15 ) 16347138486163573/(3255\cdot10^{15}) 16347138486163573/ ( 3255 ⋅ 1 0 15 ) , exceeds μ \mu μ .
Companion: record S4.8 of [13 ] .
Proposition S4.9 (Certified: the lower block for q ≥ 22 q\geq22 q ≥ 22 ). For an integer q ≥ 22 q\geq22 q ≥ 22 put r q = ⌊ 3 q / 4 + 5 / 8 ⌋ r_q=\lfloor3q/4+5/8\rfloor r q = ⌊ 3 q /4 + 5/8 ⌋ , C q = 3 q / 4 + 5 / 8 − r q ∈ { 1 / 8 , 3 / 8 , 5 / 8 , 7 / 8 } C_q=3q/4+5/8-r_q\in\{1/8,3/8,5/8,7/8\} C q = 3 q /4 + 5/8 − r q ∈ { 1/8 , 3/8 , 5/8 , 7/8 } ,
X q ( v , w ) = q + 1 2 + 1 8 q + v + w − 3 / 8 2 q 2 + 9 / 64 − v q 3 , Ω q = { X q ( v , Y 2 ) + i Y q : ( v , Y ) ∈ D q } , X_q(v,w)=q+\frac{1}{2}+\frac{1}{8q}+\frac{v+w-3/8}{2q^2}+\frac{9/64-v}{q^3},\qquad\Omega_q=\left\{X_q(v,Y^2)+\frac{iY}{\sqrt{q}}:(v,Y)\in D_q\right\}, X q ( v , w ) = q + 2 1 + 8 q 1 + 2 q 2 v + w − 3/8 + q 3 9/64 − v , Ω q = { X q ( v , Y 2 ) + q iY : ( v , Y ) ∈ D q } , where D q = { ( v , Y ) ∈ R 2 : Y 2 < C q + 1 / 8 , Y 2 − C q + 1 / 16 < v < 17 / 16 } D_q=\{(v,Y)\in\mathbb{R}^2:Y^2<C_q+1/8,\ Y^2-C_q+1/16<v<17/16\} D q = {( v , Y ) ∈ R 2 : Y 2 < C q + 1/8 , Y 2 − C q + 1/16 < v < 17/16 } . Then ∂ Ω q ∩ C q 2 = ∅ \partial\Omega_q\cap C_{q^2}=\varnothing ∂ Ω q ∩ C q 2 = ∅ and Ω q ∩ int C q 2 − 2 ≠ ∅ \Omega_q\cap\operatorname{int}C_{q^2-2}\neq\varnothing Ω q ∩ int C q 2 − 2 = ∅ .
Verification. Here guards force the first inverse digits d 1 , d 2 , … d_1,d_2,\ldots d 1 , d 2 , … along the curve, and a final margin excludes every continuation (a tree of depth at most four). With the states ( S j , V j ) (S_j,V_j) ( S j , V j ) along the forced digits and the height h h h of the exclusion, the guards are
L j = ( S j − 1 − d j + 1 ) ( x − 1 ) 2 − h , L j + = ( S j − 1 − d j ) ( x − 1 ) 2 − h , L_j=(S_{j-1}-d_j+1)(x-1)^2-h,\qquad L_j^{+}=(S_{j-1}-d_j)(x-1)^2-h, L j = ( S j − 1 − d j + 1 ) ( x − 1 ) 2 − h , L j + = ( S j − 1 − d j ) ( x − 1 ) 2 − h ,
U j = ( 2 x − h − S j ) ( x − 1 ) 2 − h , Z j = ( − S j − h ) ( x − 1 ) 2 − h , W j = V j ( x − 1 ) 2 − h . (113) U_j=(2x-h-S_j)(x-1)^2-h,\qquad Z_j=(-S_j-h)(x-1)^2-h,\qquad W_j=V_j(x-1)^2-h. \tag*{(113)} U j = ( 2 x − h − S j ) ( x − 1 ) 2 − h , Z j = ( − S j − h ) ( x − 1 ) 2 − h , W j = V j ( x − 1 ) 2 − h . (113)
Given the earlier digits, L j > 0 L_j>0 L j > 0 excludes j j j th digits below d j d_j d j and L j + > 0 L_j^{+}>0 L j + > 0 those below d j + 1 d_j+1 d j + 1 , by the V bound. Raising the j j j th digit by k k k lowers V j V_j V j by k k k and S j S_j S j by 2 x k 2xk 2 x k ; so, by the S bound, U j > 0 U_j>0 U j > 0 excludes the digits above d j d_j d j and Z j > 0 Z_j>0 Z j > 0 excludes d j d_j d j and all larger ones, and by the V bound W j > 0 W_j>0 W j > 0 excludes d j d_j d j and all smaller ones. On the boundary, at h = q 2 h=q^2 h = q 2 , with x = X q ( v , w ) x=X_q(v,w) x = X q ( v , w ) , Y 2 = w / q Y^2=w/q Y 2 = w / q and ( d 1 , d 2 , d 3 ) = ( 2 q , − q 2 + q , − q 2 + r q ) (d_1,d_2,d_3)=(2q,-q^2+q,-q^2+r_q) ( d 1 , d 2 , d 3 ) = ( 2 q , − q 2 + q , − q 2 + r q ) : L 1 , U 1 , L 2 , U 2 > 0 L_1,U_1,L_2,U_2>0 L 1 , U 1 , L 2 , U 2 > 0 on ∂ D q \partial D_q ∂ D q once the forced d 1 , d 2 d_1,d_2 d 1 , d 2 ; then W 2 > 0 W_2>0 W 2 > 0 on the right side v = 17 / 16 v=17/16 v = 17/16 , and L 3 , Z 3 > 0 L_3,Z_3>0 L 3 , Z 3 > 0 on the left side and the caps. By Lemma S4.2, ∂ Ω q ∩ C q e x t = ∅ \partial\Omega_q\cap C_q^{\mathrm{ext}}=\varnothing ∂ Ω q ∩ C q ext = ∅ , and ∂ Ω q ⊂ X \partial\Omega_q\subset X ∂ Ω q ⊂ X . Each guard, cleared of positive factors, is positive by Lemma S4.1 for q ≥ 24 q\geq24 q ≥ 24 and each value of C q C_q C q , with s = q − 24 s=q-24 s = q − 24 and the boundary parameter in [ 0 , 1 ] [0,1] [ 0 , 1 ] (the guards of the first two digits on the whole rectangle − 13 / 16 ≤ v ≤ 17 / 16 -13/16\leq v\leq17/16 − 13/16 ≤ v ≤ 17/16 , 0 ≤ w ≤ 1 0\leq w\leq1 0 ≤ w ≤ 1 ), and directly at q = 22 , 23 q=22,23 q = 22 , 23 . Interior point: for q ≥ 40 q\geq40 q ≥ 40 , c ∗ = X q ( 1 − 3 C q / 2 , 0 ) ∈ Ω q c_*=X_q(1-3C_q/2,0)\in\Omega_q c ∗ = X q ( 1 − 3 C q /2 , 0 ) ∈ Ω q and the word ( d 1 , d 2 , d 3 ) (d_1,d_2,d_3) ( d 1 , d 2 , d 3 ) , whose digits lie in B q − 3 B_{q-3} B q − 3 , satisfies Lemma S3.3 at h = q 2 − 3 h=q^2-3 h = q 2 − 3 , by five margins positive by Lemma S4.1 with s = q − 40 s=q-40 s = q − 40 ; for 22 ≤ q ≤ 39 22\leq q\leq39 22 ≤ q ≤ 39 an explicit word does so at c ∗ = X q ( v ∗ , 0 ) ∈ Ω q c_*=X_q(v_*,0)\in\Omega_q c ∗ = X q ( v ∗ , 0 ) ∈ Ω q and h = q 2 − 2 h=q^2-2 h = q 2 − 2 . Nesting and ∂ Z b = int C b \partial Z_b=\operatorname{int}C_b ∂ Z b = int C b give the containment. Spot check: for q = 23 q=23 q = 23 , r 23 = 17 r_{23}=17 r 23 = 17 , C 23 = 7 / 8 C_{23}=7/8 C 23 = 7/8 , and v ∗ = − 1051 / 1600 v_*=-1051/1600 v ∗ = − 1051/1600 lies in ( − 13 / 16 , 17 / 16 ) (-13/16,17/16) ( − 13/16 , 17/16 ) . At x = X 23 ( v ∗ , 0 ) = 915134579 / 38934400 x=X_{23}(v_*,0)=915134579/38934400 x = X 23 ( v ∗ , 0 ) = 915134579/38934400 the word ( 46 , − 506 , − 512 , − 526 ) (46,-506,-512,-526) ( 46 , − 506 , − 512 , − 526 ) has digits in B 527 B_{527} B 527 , and the lens, V and S margins of Lemma S3.3 at h = 527 h=527 h = 527 are, rounded, 455.5 455.5 455.5 , 0.392 0.392 0.392 and 381.3 381.3 381.3 . The guard polynomials are generated from the recurrence and their Bernstein coefficients tested; a second program rechecks the exported coefficients, and the witness words are stored.
Companion: record S4.9 of [13 ] .
Proposition S4.10 (Certified: twenty domains, 5 ≤ q ≤ 21 5 \le q \le21 5 ≤ q ≤ 21 ). For all integers q , h q,h q , h with 5 ≤ q ≤ 21 5 \le q \le21 5 ≤ q ≤ 21 and q 2 − 2 ≤ h ≤ q 2 q^{2}-2 \le h \le q^{2} q 2 − 2 ≤ h ≤ q 2 , there is an open set U ⊂ { Re c > 1 } U \subset\{\operatorname{Re} c > 1\} U ⊂ { Re c > 1 } such that ∂ U ∩ C h = ∅ \partial U \cap C_h = \varnothing ∂ U ∩ C h = ∅ and U ∩ C h ≠ ∅ U \cap C_h \ne\varnothing U ∩ C h = ∅ .
Verification. Here U U U is one of twenty bounded Jordan domains with closure in { Re c > 2 } \{\operatorname{Re} c > 2\} { Re c > 2 } , each with a range h min ≤ h ≤ h max h_{\min} \le h \le h_{\max} h m i n ≤ h ≤ h m a x within q 2 − 2 ≤ h ≤ q 2 q^{2}-2 \le h \le q^{2} q 2 − 2 ≤ h ≤ q 2 . Its upper boundary consists of three or four arcs that are straight in ( x , y 2 ) (x,y^{2}) ( x , y 2 ) or in the chart ( v , w ) (v,w) ( v , w ) of Proposition S4.9, so x x x and y 2 y^{2} y 2 are affine on each arc; each arc is covered by closed cells excluded at h max h_{\max} h m a x by Lemma S4.2, and conjugation gives the lower half. A real point x ∗ x_{*} x ∗ of the domain and a word satisfy Lemma S3.3 at h min h_{\min} h m i n , so x ∗ ∈ int Z h min e x t ⊂ C h min x_{*} \in\operatorname{int} Z_{h_{\min}}^{\mathrm{ext}} \subset C_{h_{\min}} x ∗ ∈ int Z h m i n ext ⊂ C h m i n ; nesting gives every h h h of the range. Spot check: for q = 6 q=6 q = 6 the domain with range h = 34 , 35 h=34,35 h = 34 , 35 has the witness x ∗ = 162 / 25 x_{*}=162/25 x ∗ = 162/25 with the word ( 12 , − 30 ) (12,-30) ( 12 , − 30 ) . At h = 34 h=34 h = 34 it gives ( S 2 , V 2 ) = ( − 538488 / 15625 , 282 / 625 ) (S_{2},V_{2})=(-538488/15625,282/625) ( S 2 , V 2 ) = ( − 538488/15625 , 282/625 ) , and the lens, V and S margins are 8781 / 625 8781/625 8781/625 , 2363003 / 10935000 2363003/10935000 2363003/10935000 and 1149 / 31250 1149/31250 1149/31250 . Every exclusion tree is rebuilt from the stored cells by the exit and branching rules of § 17.1 with Lemma S4.1, and a second program reconstructs the Bernstein coefficients by Taylor–Bernstein conversion.
Companion: record S4.10 of [13 ] .
Proposition S4.11 (Certified: small heights). For every h ∈ { 1 , 3 , 4 , 8 , 9 , 14 , 15 , 16 } h \in\{1,3,4,8,9,14,15,16\} h ∈ { 1 , 3 , 4 , 8 , 9 , 14 , 15 , 16 } , there is an open set U ⊂ { Re c > 1 } U \subset\{\operatorname{Re} c > 1\} U ⊂ { Re c > 1 } such that ∂ U ∩ C h = ∅ \partial U \cap C_h = \varnothing ∂ U ∩ C h = ∅ and U ∩ C h ≠ ∅ U \cap C_h \ne\varnothing U ∩ C h = ∅ .
Verification. Here U U U is a bounded Jordan domain with closure in the positive half-plane. The boundary arcs are covered by closed cells excluded by Lemma S4.2; at h = 1 h=1 h = 1 the cell variable runs over [ − 1 , 1 ] [-1,1] [ − 1 , 1 ] and the check uses the second bound in (S4.1) with the constant ( x min − 1 ) − 2 ≥ β (x_{\min}-1)^{-2} \ge\beta ( x m i n − 1 ) − 2 ≥ β .
Core points: for h = 1 , 3 , 4 , 8 , 9 h=1,3,4,8,9 h = 1 , 3 , 4 , 8 , 9 a nonreal root of an explicit polynomial in P h \mathcal{P}_h P h , isolated by Lemma S4.3 in a disc inside U U U ; for h = 14 , 15 , 16 h=14,15,16 h = 14 , 15 , 16 the real points 221 / 50 221/50 221/50 , 112 / 25 112/25 112/25 and 45327 / 10000 45327/10000 45327/10000 , where the words ( 8 , − 13 , − 9 ) (8,-13,-9) ( 8 , − 13 , − 9 ) , ( 8 , − 12 ) (8,-12) ( 8 , − 12 ) and ( 8 , − 12 , − 13 , − 13 , − 16 ) (8,-12,-13,-13,-16) ( 8 , − 12 , − 13 , − 13 , − 16 ) satisfy Lemma S3.3. Spot check: at h = 15 h=15 h = 15 , x = 112 / 25 x=112/25 x = 112/25 and ( 8 , − 12 ) (8,-12) ( 8 , − 12 ) give ( S 2 , V 2 ) = ( − 226688 / 15625 , 332 / 625 ) (S_{2},V_{2})=(-226688/15625,332/625) ( S 2 , V 2 ) = ( − 226688/15625 , 332/625 ) , with lens, V and S margins 1231 / 625 1231/625 1231/625 , 280159 / 15680000 280159/15680000 280159/15680000 and 30999 / 31250 30999/31250 30999/31250 . For h = 3 , 4 , 8 , 9 h=3,4,8,9 h = 3 , 4 , 8 , 9 the exclusion trees have depths 7 , 5 , 4 , 6 7,5,4,6 7 , 5 , 4 , 6 and the witness polynomials degrees 7 , 6 , 4 , 6 7,6,4,6 7 , 6 , 4 , 6 , each isolated with Rouché radius 10 − 7 10^{-7} 1 0 − 7 ; for h = 14 , 15 , 16 h=14,15,16 h = 14 , 15 , 16 the cells and words are rechecked by the two programs of Proposition S4.10; for h = 1 h=1 h = 1 the exclusion propagates ( Re z , Im z / y ) (\operatorname{Re} z,\operatorname{Im} z/y) ( Re z , Im z / y ) instead of ( S , V ) (S,V) ( S , V ) and the witness root is isolated exactly.
Companion: record S4.11 of [13 ] .
Proposition S4.12 (Certified: inequalities of the square-adjacent family). For 4 ≤ q ≤ 60 4 \le q \le60 4 ≤ q ≤ 60 and k ∈ { 0 , 1 , 2 } k \in\{0,1,2\} k ∈ { 0 , 1 , 2 } , with h = q 2 − 2 + k h=q^{2}-2+k h = q 2 − 2 + k and m = q 2 − 2 q + k m=q^{2}-2q+k m = q 2 − 2 q + k , the following hold, where φ ( β ) = 2 ( β − q ) ( β − 1 ) 2 \varphi(\beta)=2(\beta-q)(\beta-1)^{2} φ ( β ) = 2 ( β − q ) ( β − 1 ) 2 and t − ( m ′ ) t_{-}(m') t − ( m ′ ) denotes the smaller zero of 1 − ( 2 q + 1 ) z + ( 2 q + m ′ ) z 2 1-(2q+1)z+(2q+m')z^{2} 1 − ( 2 q + 1 ) z + ( 2 q + m ′ ) z 2 :
φ ( q + δ ∗ ) > h \varphi(q+\delta^{*})>h φ ( q + δ ∗ ) > h with δ ∗ = ( q 2 + 4 q − 4 − q ) / 2 \delta^{*}=(\sqrt{q^{2}+4q-4}-q)/2 δ ∗ = ( q 2 + 4 q − 4 − q ) /2 , and φ ( 1 + h − m / 2 ) < m \varphi(1+\sqrt{h-m/2})<m φ ( 1 + h − m /2 ) < m ;
t − ( m ) ≥ 1 / ( 2 q − 1 ) t_{-}(m)\ge1/(2q-1) t − ( m ) ≥ 1/ ( 2 q − 1 ) ;
t − ( q 2 − 2 q + 2 ) ≥ 1 / ( 2 q − 3 ) t_{-}(q^{2}-2q+2)\ge1/(2q-3) t − ( q 2 − 2 q + 2 ) ≥ 1/ ( 2 q − 3 ) for q ≥ 6 q\ge6 q ≥ 6 .
Verification. Items (i)–(iii) are finite instances of the inequalities proved by hand, for every q ≥ 4 q \ge4 q ≥ 4 , in Lemmas 6.2, 6.3 and 6.4; the record checks the instances 4 ≤ q ≤ 60 4 \le q \le60 4 ≤ q ≤ 60 in exact arithmetic as a control on those proofs, which do not depend on it. Each inequality compares a rational number with a number of the form a + b D a+b\sqrt{D} a + b D , a , b ∈ Q a,b \in\mathbb{Q} a , b ∈ Q , D ∈ Q > 0 D \in\mathbb{Q}_{>0} D ∈ Q > 0 ; its sign is decided exactly by comparing a 2 a^{2} a 2 with b 2 D b^{2}D b 2 D when a a a and b D b\sqrt{D} b D have opposite signs.
Companion: record S4.12 of [13 ] .
Proposition S4.13 (Certified: inner prefixes). Let ( q , m ′ , H ′ , J , π ) (q,m',H',J,\pi) ( q , m ′ , H ′ , J , π ) be one of
( 4 , 10 , 14 , [ 1 5 , 1 3 ] , ( 1 , − 8 , 12 , 13 , 14 , 12 , 13 , 10 , 13 , 13 ) ) , (4,10,14,[\tfrac{1}{5},\tfrac{1}{3}],(1,-8,12,13,14,12,13,10,13,13)), ( 4 , 10 , 14 , [ 5 1 , 3 1 ] , ( 1 , − 8 , 12 , 13 , 14 , 12 , 13 , 10 , 13 , 13 )) , ( 5 , 17 , 23 , [ 1 7 , 7 25 ] , ( 1 , − 10 , 20 , 22 , 18 , 21 , 22 , 21 , 17 ) ) . (5,17,23,[\tfrac{1}{7},\tfrac{7}{25}],(1,-10,20,22,18,21,22,21,17)). ( 5 , 17 , 23 , [ 7 1 , 25 7 ] , ( 1 , − 10 , 20 , 22 , 18 , 21 , 22 , 21 , 17 )) .
Then ( H ′ − m ′ ) a = 1 − a (H^{\prime}-m^{\prime})a=1-a ( H ′ − m ′ ) a = 1 − a at the left endpoint a a a of J J J ; f π − > 0 f_{\pi}^{-}>0 f π − > 0 at both endpoints of J J J ; f π − ( z 0 ) < 0 f_{\pi}^{-}(z_{0})<0 f π − ( z 0 ) < 0 for z 0 = 138 / 625 z_{0}=138/625 z 0 = 138/625 , respectively z 0 = 907 / 5000 z_{0}=907/5000 z 0 = 907/5000 ; and there are rationals z 1 < z 2 z_{1}<z_{2} z 1 < z 2 in J J J with ( f π + ) ′ ( z 1 ) ≤ 0 ≤ ( f π + ) ′ ( z 2 ) (f_{\pi}^{+})^{\prime}(z_{1})\leq0\leq(f_{\pi}^{+})^{\prime}(z_{2}) ( f π + ) ′ ( z 1 ) ≤ 0 ≤ ( f π + ) ′ ( z 2 ) , f π + ( z 1 ) > 0 f_{\pi}^{+}(z_{1})>0 f π + ( z 1 ) > 0 , f π + ( z 2 ) > 0 f_{\pi}^{+}(z_{2})>0 f π + ( z 2 ) > 0 and f π + ( z 1 ) + ( f π + ) ′ ( z 1 ) ( z 2 − z 1 ) > 0 f_{\pi}^{+}(z_{1})+(f_{\pi}^{+})^{\prime}(z_{1})(z_{2}-z_{1})>0 f π + ( z 1 ) + ( f π + ) ′ ( z 1 ) ( z 2 − z 1 ) > 0 .
Verification. By convexity of f π + f_{\pi}^{+} f π + on ( 0 , 1 ) (0,1) ( 0 , 1 ) , the last three inequalities give f π + > 0 f_{\pi}^{+}>0 f π + > 0 on J J J : on [ z 1 , z 2 ] [z_{1},z_{2}] [ z 1 , z 2 ] the function lies above its tangent at z 1 z_{1} z 1 , and outside this interval it is monotone away from the minimum. All quantities are rational and the inequalities are checked exactly from the stored prefixes and points.
Companion: record S4.13 of [13 ] .
Proposition S4.14 (Certified: explicit witnesses). For each row of Table 1 , let P P P be the polynomial 1 − 2 q z + ∑ j = 2 N − 1 a j z j 1-2qz+\sum_{j=2}^{N-1}a_{j}z^{j} 1 − 2 q z + ∑ j = 2 N − 1 a j z j , T = ∑ i < r c i z i T=\sum_{i<r}c_{i}z^{i} T = ∑ i < r c i z i and R = ( 1 − z r ) P + z N T R=(1-z^{r})P+z^{N}T R = ( 1 − z r ) P + z N T . Then g = gcd ( R , R ′ ) g=\gcd(R,R^{\prime}) g = g cd( R , R ′ ) is irreducible of degree three, four or five, g 2 g^{2} g 2 divides R R R , the reversed polynomial of g g g is the stated minimal polynomial, g g g has exactly one zero in ( 1 / ( q + 1 ) , 1 / q ) (1/(q+1),1/q) ( 1/ ( q + 1 ) , 1/ q ) , all digits lie in [ m , h ] [m,h] [ m , h ] for each listed height, and the period is not constant. Moreover the stated polynomial is monic and all its zeros other than β \beta β lie strictly inside the unit disk, so that β \beta β is a Pisot number.
Verification. Exact polynomial arithmetic over Z \mathbb{Z} Z ; the zero count in ( 1 / ( q + 1 ) , 1 / q ) (1/(q+1),1/q) ( 1/ ( q + 1 ) , 1/ q ) is a Sturm count with rational endpoints. The Pisot property, which the proof of Proposition 6.14 does not use, is certified by exact isolating intervals and rectangles for all zeros of the stated polynomial. For a rectangle I × J I\times J I × J , the minimum of ∣ z ∣ 2 |z|^{2} ∣ z ∣ 2 is dist ( 0 , I ) 2 + dist ( 0 , J ) 2 \operatorname{dist}(0,I)^{2}+\operatorname{dist}(0,J)^{2} dist ( 0 , I ) 2 + dist ( 0 , J ) 2 ; the maximum is attained at a corner. The verifier requires maximum < 1 <1 < 1 or minimum > 1 >1 > 1 , refining any undecided box. Real intervals use the same test with J = { 0 } J=\{0\} J = { 0 } .
Companion: record S4.14 of [13 ] .
Proposition S4.15 (Certified: surviving prefixes). For h ∈ { 25 , 49 , 62 , 63 , 64 , 79 , 81 } h\in\{25,49,62,63,64,79,81\} h ∈ { 25 , 49 , 62 , 63 , 64 , 79 , 81 } , with q q q , k k k and m m m as in (6.1), the companion lists n h n_{h} n h distinct prefixes π = ( 1 , − 2 q , a 2 , … , a N − 1 ) \pi=(1,-2q,a_{2},\ldots,a_{N-1}) π = ( 1 , − 2 q , a 2 , … , a N − 1 ) with digits in [ m , h ] [m,h] [ m , h ] , of a common length N N N , where
( h , N , n h ) ∈ { ( 25 , 15 , 384 ) , ( 49 , 12 , 171 ) , ( 62 , 12 , 314 ) , ( 63 , 12 , 332 ) , ( 64 , 12 , 299 ) , ( 79 , 12 , 463 ) , ( 81 , 12 , 369 ) } , (h,N,n_{h})\in\{(25,15,384),(49,12,171),(62,12,314),(63,12,332),
(64,12,299),(79,12,463),(81,12,369)\}, ( h , N , n h ) ∈ {( 25 , 15 , 384 ) , ( 49 , 12 , 171 ) , ( 62 , 12 , 314 ) , ( 63 , 12 , 332 ) , ( 64 , 12 , 299 ) , ( 79 , 12 , 463 ) , ( 81 , 12 , 369 )} , each with rationals z 0 , z 1 , z 2 ∈ ( 0 , 1 ) z_{0},z_{1},z_{2}\in(0,1) z 0 , z 1 , z 2 ∈ ( 0 , 1 ) such that 1 − ( 2 q + 1 ) z 0 + ( q 2 + k ) z 0 2 < 0 1-(2q+1)z_{0}+(q^{2}+k)z_{0}^{2}<0 1 − ( 2 q + 1 ) z 0 + ( q 2 + k ) z 0 2 < 0 , f π − ( z 0 ) < 0 f_{\pi}^{-}(z_{0})<0 f π − ( z 0 ) < 0 , z 1 < z 2 z_{1}<z_{2} z 1 < z 2 , ( f π + ) ′ ( z 1 ) ≤ 0 ≤ ( f π + ) ′ ( z 2 ) (f_{\pi}^{+})^{\prime}(z_{1})\leq0\leq(f_{\pi}^{+})^{\prime}(z_{2}) ( f π + ) ′ ( z 1 ) ≤ 0 ≤ ( f π + ) ′ ( z 2 ) , f π + ( z 1 ) > 0 f_{\pi}^{+}(z_{1})>0 f π + ( z 1 ) > 0 , f π + ( z 2 ) > 0 f_{\pi}^{+}(z_{2})>0 f π + ( z 2 ) > 0 and f π + ( z 1 ) + ( f π + ) ′ ( z 1 ) ( z 2 − z 1 ) > 0 f_{\pi}^{+}(z_{1})+(f_{\pi}^{+})^{\prime}(z_{1})(z_{2}-z_{1})>0 f π + ( z 1 ) + ( f π + ) ′ ( z 1 ) ( z 2 − z 1 ) > 0 . Consequently the anchor set T h T_{h} T h has at least n h n_{h} n h points.
Verification. The first inequality places z 0 z_{0} z 0 in I 0 I_{0} I 0 , where f π − f_{\pi}^{-} f π − is then negative; the remaining ones give f π + > 0 f_{\pi}^{+}>0 f π + > 0 on ( 0 , 1 ) (0,1) ( 0 , 1 ) by convexity, as in Proposition S4.13. Since f π − ( t ± ) ≥ M m ( t ± ) = 0 f_{\pi}^{-}(t_{\pm})\geq M_{m}(t_{\pm})=0 f π − ( t ± ) ≥ M m ( t ± ) = 0 and the contiguity inequality t − ≥ 1 / ( 2 q − 1 ) t_{-}\geq1/(2q-1) t − ≥ 1/ ( 2 q − 1 ) holds (Proposition S4.12), Lemma 4.6 extends each prefix to a square-adjacent witness, and distinct prefixes of equal length extend to distinct witnesses. All quantities are rational.
Companion: record S4.15 of [13 ] .
The last four records concern Section 7 and the heights h ∈ { 1 , 3 , 4 , 7 } h\in\{1,3,4,7\} h ∈ { 1 , 3 , 4 , 7 } . Their programs represent every real quantity by an interval [ ℓ , u ] 2 − P [\ell,u]2^{-P} [ ℓ , u ] 2 − P with integers ℓ ≤ u \ell\leq u ℓ ≤ u ; sums are exact, and every product and quotient is rounded outward by exact floor division, so the computed interval contains the exact one. The parameter t t t always runs over an interval, and each inequality is decided by comparing integers for all t t t in it at once.
Proposition S4.16 (Certified: the minimum of D h D_{h} D h ). For h ∈ { 1 , 3 , 4 , 7 } h\in\{1,3,4,7\} h ∈ { 1 , 3 , 4 , 7 } the companion lists rationals a h < b h a_{h}<b_{h} a h < b h , with b h − a h = 2 ⋅ 10 − 150 b_{h}-a_{h}=2\cdot10^{-150} b h − a h = 2 ⋅ 1 0 − 150 for h = 1 h=1 h = 1 and 2 ⋅ 10 − 90 2\cdot10^{-90} 2 ⋅ 1 0 − 90 otherwise, a tiling of [ 1 / ( h + 1 ) , a h ] [1/(h+1),a_{h}] [ 1/ ( h + 1 ) , a h ] by closed intervals, a polynomial Q h Q_{h} Q h , and words of lengths N h = 850 , 257 , 216 , 162 N_{h}=850,257,216,162 N h = 850 , 257 , 216 , 162 and n 0 = 49 , 24 , 19 , 15 n_{0}=49,24,19,15 n 0 = 49 , 24 , 19 , 15 , such that: no series of F h \mathcal{F}_{h} F h has a double zero in any tile; Q h Q_{h} Q h satisfies the hypotheses of Lemma 7.2 on an interval inside ( a h , b h ) (a_{h},b_{h}) ( a h , b h ) ; every series with a double zero in [ a h , b h ] [a_{h},b_{h}] [ a h , b h ] begins with the word of length N h N_{h} N h , and every series with a double zero in [ a h − 10 − 12 , b h + 10 − 9 ] [a_{h}-10^{-12},b_{h}+10^{-9}] [ a h − 1 0 − 12 , b h + 1 0 − 9 ] begins with the word of length n 0 n_{0} n 0 .
Verification. On a tile [ x , y ] [x,y] [ x , y ] put β ∈ [ 1 / y , 1 / x ] \beta\in[1/y,1/x] β ∈ [ 1/ y , 1/ x ] , enclosed by [ β 0 − δ , β 0 + δ ] [\beta_0-\delta,\beta_0+\delta] [ β 0 − δ , β 0 + δ ] with β 0 \beta_0 β 0 and δ \delta δ dyadic. A double zero of 1 + ∑ a j z j 1+\sum a_jz^j 1 + ∑ a j z j at t = 1 / β t=1/\beta t = 1/ β forces, for every N N N ,
∣ P N ( β ) ∣ ≤ h β − 1 , ∣ V N ( β ) ∣ ≤ h ( β − 1 ) 2 , |P_N(\beta)|\leq\frac{h}{\beta-1},\qquad|V_N(\beta)|\leq\frac{h}{(\beta-1)^2}, ∣ P N ( β ) ∣ ≤ β − 1 h , ∣ V N ( β ) ∣ ≤ ( β − 1 ) 2 h ,
where P N = β N + ∑ j ≤ N a j β N − j P_N=\beta^N+\sum_{j\leq N}a_j\beta^{N-j} P N = β N + ∑ j ≤ N a j β N − j and V N = ( N + 1 ) β N + ∑ j ≤ N ( N + 1 − j ) a j β N − j V_N=(N+1)\beta^N+\sum_{j\leq N}(N+1-j)a_j\beta^{N-j} V N = ( N + 1 ) β N + ∑ j ≤ N ( N + 1 − j ) a j β N − j , because P N ( β ) = − β N ∑ j > N a j t j P_N(\beta)=-\beta^N\sum_{j>N}a_jt^j P N ( β ) = − β N ∑ j > N a j t j , and similarly for the derivative. The check runs the tree of all digit words, carrying P N ( β 0 ) P_N(\beta_0) P N ( β 0 ) , V N ( β 0 ) V_N(\beta_0) V N ( β 0 ) and their β \beta β -derivatives as integers at scale 2 K 2^K 2 K , with Taylor remainders 4 h N 2 β ˉ N δ 2 4hN^2\bar{\beta}^N\delta^2 4 h N 2 β ˉ N δ 2 and 4 h N 3 β ˉ N δ 2 4hN^3\bar{\beta}^N\delta^2 4 h N 3 β ˉ N δ 2 (β ˉ = β 0 + δ \bar{\beta}=\beta_0+\delta β ˉ = β 0 + δ ; valid since ∣ P N ′ ′ ∣ ≤ 1.37 h N 2 β N |P_N''|\leq1.37hN^2\beta^N ∣ P N ′′ ∣ ≤ 1.37 h N 2 β N for β ≥ 1.49 \beta\geq1.49 β ≥ 1.49 ), and deletes a word when its enclosure misses the box. A tile is accepted when no word survives; the tiles abut exactly. At h = 1 h=1 h = 1 , 469 tiles cover [ 1 / 2 , a 1 ] [1/2,a_1] [ 1/2 , a 1 ] (depth at most 850); at h = 3 , 4 , 7 h=3,4,7 h = 3 , 4 , 7 , 291, 295 and 308 tiles. The good-prefix conditions are integer inequalities over 10 D n 10^{Dn} 1 0 D n , positivity of Q h Q_h Q h on the interval following from a lower bound for Q h ′ ′ Q_h'' Q h ′′ on a grid and the Lipschitz constant 6 h / ( 1 − b ) 4 6h/(1-b)^4 6 h / ( 1 − b ) 4 . The same tree on [ a h , b h ] [a_h,b_h] [ a h , b h ] and on the window gives the common prefixes of the surviving words.
Companion: record S4.16 of [13 ] .
Proposition S4.17 (Certified: slices and local constants). For h ∈ { 1 , 3 , 4 , 7 } h\in\{1,3,4,7\} h ∈ { 1 , 3 , 4 , 7 } , with t t t within 10 − 15 10^{-15} 1 0 − 15 of the centre T 0 T_0 T 0 of [ a h , b h ] [a_h,b_h] [ a h , b h ] , the quantities Δ l o w \Delta_{\mathrm{low}} Δ low , Δ 1 \Delta_1 Δ 1 and q 1 ∗ q_1^* q 1 ∗ of § 7 satisfy the bounds of Table 2 ; the conditions (C1)–(C7), (D), (E1)–(E7) hold for all i > n 0 i>n_0 i > n 0 and all relative tips within 10 − 80 10^{-80} 1 0 − 80 of t 0 t_0 t 0 ; and the checks (GR) and (EX) hold, together with L min + L 2 + 20 ≤ L ( 159 ) ∗ L_{\min}+L_2+20\leq L^*_{(159)} L m i n + L 2 + 20 ≤ L ( 159 ) ∗ , 1.001 v 3 ≤ 1 − v 1.001v^3\leq1-v 1.001 v 3 ≤ 1 − v (h = 1 h=1 h = 1 ) and 1.001 v ≤ 1 − v 1.001v\leq1-v 1.001 v ≤ 1 − v (h ≥ 3 h\geq3 h ≥ 3 ). Here (C3) is v − ( n 0 + 1 ) ζ n 0 + 1 − 2 h v r 0 ( r 0 / ( 1 − v ) + 1 / ( 1 − v ) 2 ) > 0 v-(n_0+1)\zeta_{n_0+1}-2hvr_0\left(r_0/(1-v)+1/(1-v)^2\right)>0 v − ( n 0 + 1 ) ζ n 0 + 1 − 2 h v r 0 ( r 0 / ( 1 − v ) + 1/ ( 1 − v ) 2 ) > 0 with r 0 r_0 r 0 the least r r r with v r < δ n 0 + 1 + ζ n 0 + 1 v^r<\delta_{n_0+1}+\zeta_{n_0+1} v r < δ n 0 + 1 + ζ n 0 + 1 ; (C6), the product of the correction factors in the proof of Lemma 7.1, is ( 1 + ( 4 C 3 W ( i ) + 3 2 G 2 ( i ) ) / κ v ) ( 1 + e i / M ˉ 2 ) ( 1 − ( 3 C 3 W ( i ) + G 2 ( i ) ) / ( 2 κ v ) ) − 2 ≤ 1 + 10 − 3 \left(1+(4C_3W(i)+\frac{3}{2}G_2(i))/\kappa_v\right)\left(1+e_i/\bar{M}^2\right)\left(1-(3C_3W(i)+G_2(i))/(2\kappa_v)\right)^{-2}\leq1+10^{-3} ( 1 + ( 4 C 3 W ( i ) + 2 3 G 2 ( i )) / κ v ) ( 1 + e i / M ˉ 2 ) ( 1 − ( 3 C 3 W ( i ) + G 2 ( i )) / ( 2 κ v ) ) − 2 ≤ 1 + 1 0 − 3 with e i = ( κ v + 4 C 3 W ( i ) + 3 2 G 2 ( i ) ) W ( i ) 2 / v i e_i=(\kappa_v+4C_3W(i)+\frac{3}{2}G_2(i))W(i)^2/v^i e i = ( κ v + 4 C 3 W ( i ) + 2 3 G 2 ( i )) W ( i ) 2 / v i ; (C7) is i η ˉ i ≤ 10 − 5 i\bar{\eta}_i\leq10^{-5} i η ˉ i ≤ 1 0 − 5 ; (E3) is ( N + 1 ) ( M ˉ v i − 1 / κ v + c ρ v ( N + 1 ) / 2 ) / v ≤ 10 − 3 (N+1)(\bar{M}v^{i-1}/\kappa_v+c_\rho v^{(N+1)/2})/v\leq10^{-3} ( N + 1 ) ( M ˉ v i − 1 / κ v + c ρ v ( N + 1 ) /2 ) / v ≤ 1 0 − 3 for i + 5 ≤ N ≤ 2 i + 2 i+5\leq N\leq2i+2 i + 5 ≤ N ≤ 2 i + 2 , a range that contains the values N = i + L + 5 N=i+L+5 N = i + L + 5 of Lemma 7.3; (E5) is ( 6 C 3 R + G 2 ( n 0 + 1 ) ) / ( 2 κ v ) ≤ 10 − 4 (6C_3R+G_2(n_0+1))/(2\kappa_v)\leq10^{-4} ( 6 C 3 R + G 2 ( n 0 + 1 )) / ( 2 κ v ) ≤ 1 0 − 4 ; (E6) is 2 h ( v + ⋯ + v 5 ) > 1.7 2h(v+\cdots+v^5)>1.7 2 h ( v + ⋯ + v 5 ) > 1.7 ; and (E7) is 2 h v > 1.01 2hv>1.01 2 h v > 1.01 . Conditions (E1), (E2′) and (E4) are κ v c ρ 2 > 1.01 h / ( 1 − τ ) \kappa_vc_\rho^2>1.01h/(1-\tau) κ v c ρ 2 > 1.01 h / ( 1 − τ ) , 0.8 h v / ( 1 − v ) > 1.01 0.8hv/(1-v)>1.01 0.8 h v / ( 1 − v ) > 1.01 and v 4 ( 1.9 h v / ( 1 − v ) + 1 ) ≤ 0.965 v^4(1.9hv/(1-v)+1)\leq0.965 v 4 ( 1.9 h v / ( 1 − v ) + 1 ) ≤ 0.965 .
Verification. The slices are taken from an outer table: for x x x in cells of width 10 − 4 10^{-4} 1 0 − 4 , a finite union of intervals containing S t ( x ) S_t(x) S t ( x ) for every t t t within 10 − 12 10^{-12} 1 0 − 12 of T 0 T_0 T 0 , obtained from the trivial table [ 0 , 2 h t / ( 1 − t ) 2 ] [0,2ht/(1-t)^2] [ 0 , 2 h t / ( 1 − t ) 2 ] by iterating (7.2) with outward rounding (each iterate is again an outer table; 45 iterations, 43 at h = 7 h=7 h = 7 , where the table is stationary). A lookup also includes the neighbouring cells, which absorbs weight shifts below 10 − 4 10^{-4} 1 0 − 4 . The greedy moment q t g q_t^g q t g is enclosed over an interval of weights by following every branch of the greedy expansion, and the moment of the unknown remainder after n n n digits by t n ( n r + [ 0 , d ˉ t / ( 1 − t ) 2 ] ) t^n(nr+[0,\bar{d}t/(1-t)^2]) t n ( n r + [ 0 , d ˉ t / ( 1 − t ) 2 ]) with r ∈ [ 0 , 1 + 10 − 7 ] r\in[0,1+10^{-7}] r ∈ [ 0 , 1 + 1 0 − 7 ] . Δ l o w \Delta_{\mathrm{low}} Δ low and Δ 1 \Delta_1 Δ 1 are minima over 4000 (h = 1 h=1 h = 1 ) or 2000 cells of p p p , and q 1 ∗ q_1^* q 1 ∗ is a maximum over all branches. Each condition indexed by i > n 0 i>n_0 i > n 0 is reduced to i = n 0 + 1 i=n_0+1 i = n 0 + 1 by monotone majorants, for instance W ( i ) ≤ A i r − 1 W(i)\leq Ai^{r-1} W ( i ) ≤ A i r − 1 and G 2 ( i ) ≤ B i 2 − i − 2 G_2(i)\leq Bi^{2-i-2} G 2 ( i ) ≤ B i 2 − i − 2 , and κ v \kappa_v κ v is enclosed from the N h N_h N h certified digits with the unknown tail bounded by G 2 ( N h + 1 ) / 2 G_2(N_h+1)/2 G 2 ( N h + 1 ) /2 . (GR) runs over the 118 (h = 1 h=1 h = 1 ) and 7 (h = 3 h=3 h = 3 ) admissible pairs ( c , F ) (c,F) ( c , F ) , with worst margins 0.0708 and 0.685; it is empty at h = 4 , 7 h=4,7 h = 4 , 7 , where L G = 1 L_G=1 L G = 1 .
Companion: record S4.17 of [13 ] .
Proposition S4.18 (Certified: Move Lemma). Proposition 7.9 holds, with the constants of Table 2 , and the clearances in (a) and (b) of Corollary 7.8, that is the distances of the groups from v s L − ( 1 − μ ) v^sL_-(1-\mu) v s L − ( 1 − μ ) in (a) and from v s [ L − ( 1 − μ ) , L + ( 1 + μ ) ] v^s[L_-(1-\mu),L_+(1+\mu)] v s [ L − ( 1 − μ ) , L + ( 1 + μ )] in (b), are at least 8 ⋅ 10 − 4 v s L − 8\cdot10^{-4}v^sL_- 8 ⋅ 1 0 − 4 v s L − .
Verification. Fix a start word d u du d u of C h C_h C h . The digits after the pattern are enumerated by cylinders u w uw u w of weakly admissible words over { 0 , … , d ˉ } \{0,\ldots,\bar{d}\} { 0 , … , d ˉ } (∣ w ∣ = 8 , 2 , 2 , 2 |w|=8,2,2,2 ∣ w ∣ = 8 , 2 , 2 , 2 initially, refined on failure up to ∣ w ∣ = 40 , 24 , 24 , 14 |w|=40,24,24,14 ∣ w ∣ = 40 , 24 , 24 , 14 ), the remainder after them in [ 0 , 1 + 10 − 7 ] [0,1+10^{-7}] [ 0 , 1 + 1 0 − 7 ] . The groups of the levels inside u w uw u w are computed from the known digits; beyond them each group lies below t ∣ u w ∣ + 1 D ∗ t^{|uw|+1}D^* t ∣ u w ∣ + 1 D ∗ , where D ∗ = sup { max S t ( c + y ) − q t g ( y ) } D^*=\sup\{\max S_t(c+y)-q_t^g(y)\} D ∗ = sup { max S t ( c + y ) − q t g ( y )} over 1 ≤ c ≤ d ˉ 1\leq c\leq\bar{d} 1 ≤ c ≤ d ˉ and y ∈ [ 0 , 1 + 10 − 7 ] y\in[0,1+10^{-7}] y ∈ [ 0 , 1 + 1 0 − 7 ] is at most 7.71, 5.56, 4.75, 4.89; this bound is valid for tails with restarts by Lemma 7.5 (ii). The J h J_h J h levels before the pattern run over all weakly admissible context words. For each cylinder, each branch of the greedy expansion of 1 + y s + [ − 10 − 7 , 10 − 7 ] 1+y_s+[-10^{-7},10^{-7}] 1 + y s + [ − 1 0 − 7 , 1 0 − 7 ] gives an enclosure of the first-group position, and the check verifies (a) and (b) against it with the stated clearance. In all, 71, 65, 66 and 512 cylinder checks pass for h = 1 , 3 , 4 , 7 h = 1, 3, 4, 7 h = 1 , 3 , 4 , 7 ; the smallest clearances, in units of v L − s v^{s}_{L-} v L − s , are 1.5 ⋅ 10 − 3 1.5 \cdot10^{-3} 1.5 ⋅ 1 0 − 3 and 4.2 ⋅ 10 − 3 4.2 \cdot10^{-3} 4.2 ⋅ 1 0 − 3 for h = 1 h = 1 h = 1 , 0.063 0.063 0.063 and 0.41 0.41 0.41 for h = 3 h = 3 h = 3 , 0.014 0.014 0.014 and 0.23 0.23 0.23 for h = 4 h = 4 h = 4 , and 8.9 ⋅ 10 − 4 8.9 \cdot10^{-4} 8.9 ⋅ 1 0 − 4 and 0.54 0.54 0.54 for h = 7 h = 7 h = 7 .
Companion: record S4.18 of [13 ] .
Proposition S4.19 (Certified: Landing Lemma and assembly). Proposition 7.11 holds. Moreover, with the constants of Table 2 and F min = ( μ − 10 − 4 ) L min ∘ F_{\min} = (\mu- 10^{-4})L_{\min}^{\circ} F m i n = ( μ − 1 0 − 4 ) L m i n ∘ , v 0 P ∗ L max ∘ ( 1 + μ ) < F min ≤ 10 − 9 / u P 0 v_{0}^{P^\ast}L_{\max}^{\circ}(1+\mu) < F_{\min} \le10^{-9}/u_{P_0} v 0 P ∗ L m a x ∘ ( 1 + μ ) < F m i n ≤ 1 0 − 9 / u P 0 and P ∗ ≥ J h + 2 P^\ast\ge J_h + 2 P ∗ ≥ J h + 2 ;
(K1) ( Δ l o w − v J h + 2 ( q 1 ∗ + 10 − 5 ) ) β J h + 1 > L max ∘ ( 1 + μ ) ( 1 + 3 ⋅ 10 − 5 ) ; \text{(K1)}\quad(\Delta_{\mathrm{low}} - v^{J_h+2}(q_1^\ast+ 10^{-5}))\beta^{J_h+1} > L_{\max}^{\circ}(1+\mu)(1+3\cdot10^{-5}); (K1) ( Δ low − v J h + 2 ( q 1 ∗ + 1 0 − 5 )) β J h + 1 > L m a x ∘ ( 1 + μ ) ( 1 + 3 ⋅ 1 0 − 5 ) ; (K2) the same with Δ 1 in place of Δ l o w ; \text{(K2)}\quad\text{the same with }\Delta_1\text{ in place of }\Delta_{\mathrm{low}}; (K2) the same with Δ 1 in place of Δ low ; (K4) min ( Δ l o w , Δ 1 ) − v L G ( q 1 ∗ + 10 − 5 ) > 0.002 ; \text{(K4)}\quad\min(\Delta_{\mathrm{low}},\Delta_1) - v^{L_G}(q_1^\ast+ 10^{-5}) > 0.002; (K4) min ( Δ low , Δ 1 ) − v L G ( q 1 ∗ + 1 0 − 5 ) > 0.002 ; (K5) 1.0001 v , R h + 1 ( q 1 ∗ + 10 − 5 ) ≤ 7 ⋅ 10 − 5 L min ∘ ; \text{(K5)}\quad1.0001^{v,R_h+1}(q_1^\ast+ 10^{-5}) \le7\cdot10^{-5}L_{\min}^{\circ}; (K5) 1.000 1 v , R h + 1 ( q 1 ∗ + 1 0 − 5 ) ≤ 7 ⋅ 1 0 − 5 L m i n ∘ ; (EX′) the M E ≥ 144 common digits of (EX) contain a C h -pattern at a level at least max ( M h , P ∗ ) ; \text{(EX′)}\quad\text{the }M_E \ge144\text{ common digits of (EX) contain a }C_h\text{-pattern at a level at least }\max(M_h,P^\ast); (EX′) the M E ≥ 144 common digits of (EX) contain a C h -pattern at a level at least max ( M h , P ∗ ) ; the remainders of t 0 t_0 t 0 satisfy p n 0 < 2 p_{n_0}<2 p n 0 < 2 and 0 ≤ p k < 1 0\le p_k<1 0 ≤ p k < 1 for n 0 < k ≤ min ( 200 , N h − 2 ) n_0<k\le\min(200,N_h-2) n 0 < k ≤ min ( 200 , N h − 2 ) ; its certified tail has C h C_h C h -patterns at levels ≥ P 0 ∗ \ge P_0^\ast ≥ P 0 ∗ , the deepest at absolute position i 1 ≥ 158 i_1\ge158 i 1 ≥ 158 ; and X h = v i 1 − 1 L max ∘ ( 1 + 10 − 4 ) / ( 2 κ v ( 1 − v 0 ∗ ) ) X_h=v^{i_1-1}L_{\max}^{\circ}(1+10^{-4})/(2\kappa v(1-v_0^\ast)) X h = v i 1 − 1 L m a x ∘ ( 1 + 1 0 − 4 ) / ( 2 κ v ( 1 − v 0 ∗ )) and η ˉ i 1 + X h ( 2 h + d ˉ ) / ( 1 − τ ) 2 ≤ ϵ h \bar{\eta}_{i_1}+X_h(2h+\bar d)/(1-\tau)^2\le\epsilon_h η ˉ i 1 + X h ( 2 h + d ˉ ) / ( 1 − τ ) 2 ≤ ϵ h are inside the windows of Propositions 7.9 and 7.11 .
Verification. The landing set is covered by a branch and bound over the core language: a node fixes a finite core word, encloses y s y_s y s over all its continuations (the constraints of the core are dropped for the unknown part, which only enlarges the enclosure), and is closed when the digits of the greedy expansion that are common to the whole node determine a C h C_h C h -pattern at a level ≥ P 0 ∗ \ge P_0^\ast ≥ P 0 ∗ and its first R h R_h R h remainders exceed 10 − 20 10^{-20} 1 0 − 20 ; otherwise it is split. The zero runs 0 k 0^k 0 k are treated one by one for k ≤ 40 , 30 , 40 , 30 k\le40,30,40,30 k ≤ 40 , 30 , 40 , 30 , and the longer ones with the terminating tails as one node. All 6263, 1211, 40 and 8678 nodes close, at depth at most 119, 25, 0 and 23; the patterns lie at levels 13–421, 7–26, 6–10 and 4–18. The remainders of t 0 t_0 t 0 are enclosed by a centred form at the midpoint of [ a h , b h ] [a_h,b_h] [ a h , b h ] , using f ( t 0 ) = f ′ ( t 0 ) = 0 f(t_0)=f'(t_0)=0 f ( t 0 ) = f ′ ( t 0 ) = 0 for the derivative bound. The remaining items are rational inequalities between the outputs of the previous records.
Companion: record S4.19 of [13 ] .
Supplementary material
The computational companion [13 ] contains, for each of Propositions S4.4–S4.19, the frozen inputs (integer and rational coefficients, cell covers, word lists and witness data) and a program that accepts them only by exact rational, symbolic or bounded-integer arithmetic; decimal output is diagnostic and never decides acceptance. Its documentation lists, record by record, the inputs and checking programs, and a complete run of all eighty-four checks on the supplied inputs is recorded there together with the software versions used; the complete runs of the forty-eight checks of the earlier records made on 5 and 3 November, the second before the unit-disk test of record S4.14 was tightened, are preserved alongside it. The documentation numbers the records D.4–D.19, as in an earlier version of this paper in which this section was Appendix D: Proposition S4.k k k here is record D.k k k there. It also keeps the working name “finger” for the square-adjacent family of Section 6 (directory certificates/finger), and its figure finger_region.pdf is Figure 6 with “finger interval” for the anchor interval. The companion also contains the search programs and logs behind Proposition 6.14, the generators that proposed the certificates of Proposition S4.16, and the data and scripts of the figures drawn from computed data.
Department of Computer Science, Applied Mathematics and Statistics, Universitat de Girona, Catalonia, Spain
Email address: bernat@espigule.com