Introduction
Automatic complexity modulo prime powers
Let p p p be a prime. By Christol’s theorem [8 , 10 ] (see Theorem 1, §5 of [10 ] ), a power series over F p \mathbb{F}_p F p is algebraic over F p ( x ) \mathbb{F}_p(x) F p ( x ) if and only if its coefficient sequence is p p p -automatic. Christol [9 ] and Denef and Lipshitz [11 ] (see also Theorem 1.2 of [18 ] ) extended the automaticity statement to Z p \mathbb{Z}_p Z p : if F = ∑ n ≥ 0 a ( n ) x n ∈ Z p [ [ x ] ] F=\sum_{n\ge0}a(n)x^n\in\mathbb{Z}_p[[x]] F = ∑ n ≥ 0 a ( n ) x n ∈ Z p [[ x ]] is algebraic, then ( a ( n ) m o d p α ) n ≥ 0 (a(n)\bmod p^\alpha)_{n\ge0} ( a ( n ) mod p α ) n ≥ 0 is p p p -automatic for every α ≥ 1 \alpha\ge1 α ≥ 1 ; more generally, the same holds for diagonals of rational functions N / Q N/Q N / Q with Q ( 0 ) ∈ Z p × Q(0)\in\mathbb{Z}_p^\times Q ( 0 ) ∈ Z p × [11 , 18 ] (see Theorem 2.1 of [18 ] ). Rowland and Yassawi [18 ] made the construction of such automata for diagonals explicit and used it to prove many congruences.
Throughout, automata read the base-p p p digits of n n n starting with the least significant digit. The minimal automaton of a sequence is the smallest automaton generating it that is insensitive to leading zeros; by Eilenberg’s theorem its number of states equals the cardinality of the p p p -kernel
ker p ( a ) = { ( a ( p k n + v ) ) n ≥ 0 : k ≥ 0 , 0 ≤ v < p k } \operatorname{ker}_p(a)=\{(a(p^k n+v))_{n\ge0}: k\ge0,\ 0\le v<p^k\} ker p ( a ) = {( a ( p k n + v ) ) n ≥ 0 : k ≥ 0 , 0 ≤ v < p k }
(see [19 §2] , [12 ] , [2 Ch. 6] ). For G = ∑ g ( n ) x n ∈ Z p [ [ x ] ] G=\sum g(n)x^n\in\mathbb{Z}_p[[x]] G = ∑ g ( n ) x n ∈ Z p [[ x ]] we write
K α ( G ) : = ∣ ker p ( ( g ( n ) m o d p α ) n ≥ 0 ) ∣ . K_\alpha(G):=\left|\operatorname{ker}_p\left((g(n)\bmod p^\alpha)_{n\ge0}\right)\right|. K α ( G ) := ∣ ker p ( ( g ( n ) mod p α ) n ≥ 0 ) ∣ .
For α = 1 \alpha=1 α = 1 the size of the minimal automaton is well understood: Bridy [7 ] and Rowland, Stipulanti and Yassawi [17 ] proved bounds of the form ( 1 + o ( 1 ) ) p h d (1+o(1))p^{hd} ( 1 + o ( 1 )) p h d for algebraic series over F p \mathbb{F}_p F p of height h h h and degree d d d , and Adamczewski, Bostan and Caruso [1 ] treated multivariate algebraic series and diagonals. For α ≥ 2 \alpha\ge2 α ≥ 2 , the constructions of Denef and Lipshitz lead to bounds that are doubly exponential in α \alpha α [19 §1] . Rowland and Yassawi [19 ] recently replaced these by polynomial exponents. For the Furstenberg series F F F of P ∈ Z p [ x , y ] P \in\mathbb{Z}_p[x,y] P ∈ Z p [ x , y ] with h = deg x ( P m o d p ) = deg x P h=\deg_x(P \bmod p)=\deg_x P h = deg x ( P mod p ) = deg x P and d = deg y ( P m o d p ) = deg y P d=\deg_y(P \bmod p)=\deg_y P d = deg y ( P mod p ) = deg y P , for nonzero F F F their Theorem 1 bounds the minimal automaton by
( 1 + o ( 1 ) ) p α 6 α ( α + 1 ) ( ( 2 h d − 1 ) α + h d + 1 ) (1+o(1))p^{\frac{\alpha}{6}\alpha(\alpha+1)((2hd-1)\alpha+hd+1)} ( 1 + o ( 1 )) p 6 α α ( α + 1 ) (( 2 h d − 1 ) α + h d + 1 )
and their Theorem 3 bounds it for the diagonal of P / Q P/Q P / Q in m ≥ 2 m\geq2 m ≥ 2 variables, with each reduced coordinate height h i ≥ 1 h_i\geq1 h i ≥ 1 , by p M p^M p M with M = ∑ k = 0 α − 1 ∏ i = 1 m ( ( k + 1 ) h i + 1 ) M=\sum_{k=0}^{\alpha-1}\prod_{i=1}^{m}((k+1)h_i+1) M = ∑ k = 0 α − 1 ∏ i = 1 m (( k + 1 ) h i + 1 ) , whose dominant term is α m + 1 h 1 ⋯ h m / ( m + 1 ) \alpha^{m+1}h_1\cdots h_m/(m+1) α m + 1 h 1 ⋯ h m / ( m + 1 ) . For m = 2 m=2 m = 2 their Theorem 56 improves the leading term to α ( α + 1 ) ( 2 α + 1 ) h d / 6 \alpha(\alpha+1)(2\alpha+1)hd/6 α ( α + 1 ) ( 2 α + 1 ) h d /6 .
Main result
Let m ≥ 1 m\geq1 m ≥ 1 , z = ( z 1 , … , z m ) z=(z_1,\ldots,z_m) z = ( z 1 , … , z m ) , 1 = ( 1 , … , 1 ) \mathbf{1}=(1,\ldots,1) 1 = ( 1 , … , 1 ) , and let
Diag ( ∑ u ∈ Z ≥ 0 m s u z u ) = ∑ n ≥ 0 s n 1 x n . \operatorname{Diag}\left(\sum_{u\in\mathbb{Z}_{\geq0}^{m}}s_u z^u\right)=\sum_{n\geq0}s_{n\mathbf{1}}x^n. Diag u ∈ Z ≥ 0 m ∑ s u z u = n ≥ 0 ∑ s n 1 x n .
Let N , Q ∈ Z p [ z ] N,Q\in\mathbb{Z}_p[z] N , Q ∈ Z p [ z ] with Q ( 0 ) ∈ Z p × Q(0)\in\mathbb{Z}_p^\times Q ( 0 ) ∈ Z p × . Write Δ ( Q ) = conv ( supp Q ) \Delta(Q)=\operatorname{conv}(\operatorname{supp}Q) Δ ( Q ) = conv ( supp Q ) and K ( N ) = conv ( supp N ) K(N)=\operatorname{conv}(\operatorname{supp}N) K ( N ) = conv ( supp N ) for the Newton polytopes (with K ( 0 ) = ∅ K(0)=\varnothing K ( 0 ) = ∅ ), and let h i = max ( deg z i N , deg z i Q ) h_i=\max(\deg_{z_i}N,\deg_{z_i}Q) h i = max ( deg z i N , deg z i Q ) , with deg z i 0 = − ∞ \deg_{z_i}0=-\infty deg z i 0 = − ∞ .
Definition 1.1. The support set of ( N , Q ) (N,Q) ( N , Q ) at the prime p p p is
J = J p ( N , Q ) = { u ∈ Z ≥ 0 m : u + η 1 ∈ ( 1 − ε ) Δ ( Q ) + ε K ( N ) for some η ∈ [ 0 , 1 ) , ε ∈ ( 0 , 1 / p ] } . J=J_p(N,Q)=\left\{u\in\mathbb{Z}_{\geq0}^{m}:u+\eta\mathbf{1}\in(1-\varepsilon)\Delta(Q)+\varepsilon K(N)\text{ for some }\eta\in[0,1),\ \varepsilon\in(0,1/p]\right\}. J = J p ( N , Q ) = { u ∈ Z ≥ 0 m : u + η 1 ∈ ( 1 − ε ) Δ ( Q ) + ε K ( N ) for some η ∈ [ 0 , 1 ) , ε ∈ ( 0 , 1/ p ] } . The set J J J is contained in the box ∏ i { 0 , 1 , … , h i } \prod_i\{0,1,\ldots,h_i\} ∏ i { 0 , 1 , … , h i } (Lemma 4.1 ), so ∣ J ∣ ≤ D : = ∏ i ( h i + 1 ) |J|\leq D:=\prod_i(h_i+1) ∣ J ∣ ≤ D := ∏ i ( h i + 1 ) .
Theorem 1.2. Let p p p be a prime, m ≥ 1 m\geq1 m ≥ 1 , and N , Q ∈ Z p [ z 1 , … , z m ] N,Q\in\mathbb{Z}_p[z_1,\ldots,z_m] N , Q ∈ Z p [ z 1 , … , z m ] with Q ( 0 ) ∈ Z p × Q(0)\in\mathbb{Z}_p^\times Q ( 0 ) ∈ Z p × . Let G = Diag ( N / Q ) G=\operatorname{Diag}(N/Q) G = Diag ( N / Q ) . Then for every α ≥ 1 \alpha\geq1 α ≥ 1
K α ( G ) ≤ p ( ∣ J ∣ + 1 ) α − 1 + p α − 1 p − 1 ≤ p ( D + 1 ) α − 1 + p α − 1 p − 1 . K_\alpha(G)\leq p^{(|J|+1)\alpha-1}+\frac{p^\alpha-1}{p-1}\leq p^{(D+1)\alpha-1}+\frac{p^\alpha-1}{p-1}. K α ( G ) ≤ p ( ∣ J ∣ + 1 ) α − 1 + p − 1 p α − 1 ≤ p ( D + 1 ) α − 1 + p − 1 p α − 1 . In particular log p K α ( G ) ≤ ( ∣ J ∣ + 1 ) α + O ( 1 ) \log_p K_\alpha(G)\leq(|J|+1)\alpha+O(1) log p K α ( G ) ≤ ( ∣ J ∣ + 1 ) α + O ( 1 ) for fixed N N N , Q Q Q , m m m and p p p . No Hasse–Witt invertibility, smoothness or ordinarity hypothesis is used. The proof (Sections 2–5) also produces an explicit automaton with at most this many states (Corollary 5.9): its states are pairs consisting of a window of α − 1 \alpha-1 α − 1 digits and a compatible history, and its transitions are given by explicit matrices over Z p \mathbb{Z}_p Z p , reduced modulo p α p^\alpha p α .
Consequences
Furstenberg’s theorem, in the integral form of Denef and Lipshitz [11 Lemma 6.3] (see [19 Theorem 4] ), represents the Furstenberg series of P P P as Diag ( N F / Q F ) \operatorname{Diag}(N_F/Q_F) Diag ( N F / Q F ) with Q F = P ( x y , y ) / y Q_F=P(xy,y)/y Q F = P ( x y , y ) / y and N F = y ∂ y P ( x y , y ) N_F=y\partial_yP(xy,y) N F = y ∂ y P ( x y , y ) . Applying Theorem 1.2 to this representation gives:
Corollary 1.3. Let P ∈ Z p [ x , y ] P\in\mathbb{Z}_p[x,y] P ∈ Z p [ x , y ] with P ( 0 , 0 ) = 0 P(0,0)=0 P ( 0 , 0 ) = 0 and ∂ y P ( 0 , 0 ) ∈ Z p × \partial_yP(0,0)\in\mathbb{Z}_p^\times ∂ y P ( 0 , 0 ) ∈ Z p × , deg x P ≤ h \deg_xP\leq h deg x P ≤ h and deg y P ≤ d \deg_yP\leq d deg y P ≤ d , and let F F F be its Furstenberg series. Then for every α ≥ 1 \alpha\geq1 α ≥ 1
K α ( F ) ≤ p ( ( h + 1 ) ( h + d + 1 ) + 1 ) α − 1 + p α − 1 p − 1 . K_\alpha(F)\leq p^{((h+1)(h+d+1)+1)\alpha-1}+\frac{p^\alpha-1}{p-1}. K α ( F ) ≤ p (( h + 1 ) ( h + d + 1 ) + 1 ) α − 1 + p − 1 p α − 1 .
For special supports the set J J J is much smaller than the box. For the polynomial x 2 y 2 + ( 1 + x + x 2 ) y + x 2 x^2y^2+(1+x+x^2)y+x^2 x 2 y 2 + ( 1 + x + x 2 ) y + x 2 from [19 Appendix] and its four-parameter deformation, studied in the appendix below, we find ∣ J ∣ ≤ 5 |J|\leq5 ∣ J ∣ ≤ 5 :
Corollary 1.4. Let a , b , c , d ∈ Z 2 a,b,c,d\in\mathbb{Z}_2 a , b , c , d ∈ Z 2 , P = b x 2 y 2 + ( 1 + a x + d x 2 ) y + c x 2 P=bx^2y^2+(1+ax+dx^2)y+cx^2 P = b x 2 y 2 + ( 1 + a x + d x 2 ) y + c x 2 , and let F F F be its Furstenberg series. Then
K α ( F ) ≤ 2 6 α − 1 + 2 α − 1 ( α ≥ 1 ) . K_\alpha(F)\leq2^{6\alpha-1}+2^\alpha-1\qquad(\alpha\geq1). K α ( F ) ≤ 2 6 α − 1 + 2 α − 1 ( α ≥ 1 ) . If a , b , c , d a,b,c,d a , b , c , d are odd, then also K α ( F ) ≥ 2 α + 1 K_\alpha(F)\geq2^{\alpha+1} K α ( F ) ≥ 2 α + 1 .
For odd parameters, § 7 determines the exact exponential rate, strengthening both bounds relevant to that rate. Theorem 7.1 proves
K α ( F ) ≥ 9 ⋅ 8 α 64 α − 32 , lim α → ∞ log 2 K α ( F ) α = 3. K_{\alpha}(F) \ge\frac{9 \cdot8^{\alpha}}{64\alpha- 32}, \qquad\lim_{\alpha\to\infty} \frac{\log_{2} K_{\alpha}(F)}{\alpha} = 3. K α ( F ) ≥ 64 α − 32 9 ⋅ 8 α , α → ∞ lim α log 2 K α ( F ) = 3.
The normalized full-window frames are independent of the odd parameter quadruple. The complete kernel cardinality need not be: for example, K 2 = 53 K_2 = 53 K 2 = 53 for ( a , b , c , d ) = ( 1 , 1 , 1 , 1 ) (a,b,c,d) = (1,1,1,1) ( a , b , c , d ) = ( 1 , 1 , 1 , 1 ) and K 2 = 58 K_2 = 58 K 2 = 58 for ( 3 , 5 , 7 , 9 ) (3,5,7,9) ( 3 , 5 , 7 , 9 ) , as the exact ancillary calculation verifies.
For the diagonal of 1 / ( 1 − z 1 − ⋯ − z m ) 1/(1-z_1-\cdots-z_m) 1/ ( 1 − z 1 − ⋯ − z m ) , that is, for the multinomial numbers ( m n ) ! / ( n ! ) m (mn)!/(n!)^m ( mn )! / ( n ! ) m , one has J = { 0 } J = \{0\} J = { 0 } and hence K α ≤ p 2 α − 1 + ( p α − 1 ) / ( p − 1 ) K_{\alpha} \le p^{2\alpha-1} + (p^\alpha- 1)/(p-1) K α ≤ p 2 α − 1 + ( p α − 1 ) / ( p − 1 ) for every m m m (Example 4.5).
Comparison with earlier work
All comparisons below are for fixed input polynomials, a fixed number of variables and a fixed prime, as α → ∞ \alpha\to\infty α → ∞ . In that regime the exponent in Theorem 1.2 is linear in α \alpha α , while the exponents in [19 Theorems 1, 3 and 56] are polynomials in α \alpha α of degree 3, m + 1 m+1 m + 1 and 3. We do not claim an improvement that is uniform in the degrees, in m m m or in p p p : our slope ∣ J ∣ + 1 |J|+1 ∣ J ∣ + 1 can be much larger than the leading constants of [19 ] , so for small α \alpha α their bounds can be smaller (Remark 6.1). At α = 1 \alpha= 1 α = 1 both bounds reduce to bounds of the classical shape p O ( ∏ ( h i + 1 ) ) p^{O(\prod(h_i+1))} p O ( ∏ ( h i + 1 )) and we claim nothing new there. Also, [19 ] uses the degrees of the reductions modulo p p p , while our J J J is defined by the supports of N N N and Q Q Q ; by Remark 4.4, any lifts of N m o d p α N \bmod p^\alpha N mod p α and Q m o d p α Q \bmod p^\alpha Q mod p α may be used.
Finite fields and constant-term schemes. Bridy [7 ] bounds the minimal automaton size for algebraic series over finite fields using the geometry of the defining curve; Rowland, Stipulanti and Yassawi [17 ] give an elementary proof of the genus-free bound through rational diagonals. Adamczewski, Bostan and Caruso [1 ] establish sharper multivariate bounds in positive characteristic. These results concern the finite-field setting, rather than the growth with the prime-power precision considered here. For constant-term sequences, Henningsen and Straub [13 ] relate generalized Lucas congruences to linear p p p -schemes, and Straub [21 ] proves Gessel–Lucas congruences and supercongruences for sporadic Apéry-like sequences. Such congruence results provide further structure, but do not by themselves count all states in the kernel as the precision increases.
Exact growth for a fixed series. Straub [20 Question 4.5] asks for an exact or asymptotic formula for the minimal automaton sizes of the Motzkin numbers modulo 2 r 2^r 2 r . Theorem 7.1 addresses the analogous exponent question for the odd quadratic family of Corollary 1.4; it does not settle that Motzkin-number question. In the appendix of [19 ] , Rowland and Yassawi consider P = x 2 y 2 + ( 1 + x + x 2 ) y + x 2 P = x^2y^2 + (1+x+x^2)y + x^2 P = x 2 y 2 + ( 1 + x + x 2 ) y + x 2 and observe, for 1 ≤ α ≤ 14 1 \le\alpha\le14 1 ≤ α ≤ 14 , zero-digit transient length α + 1 \alpha+1 α + 1 and eventual period 2 α + 1 2^{\alpha+1} 2 α + 1 . They conjecture the corresponding orbit size 2 α + 1 + α + 1 2^{\alpha+1}+\alpha+1 2 α + 1 + α + 1 and suggest that this may indicate the true growth rate. Their statement concerns the orbit under the zero-digit operator; the full binary kernel also allows arbitrary digit words. Appendix A determines the zero orbit, while Theorem 7.1 determines the exponent of the full kernel for every odd parameter quadruple. We are not aware of an earlier determination of this full-kernel exponent for that family.
Rational special cases. Precision-linear exponents are already elementary for sequences with a fixed finite matrix representation, so the one-variable rational specialization is not claimed as new. If a ( n ) = λ A n v a(n) = \lambda A^n v a ( n ) = λ A n v with A ∈ M r ( Z p ) A \in M_r(\mathbb{Z}_p) A ∈ M r ( Z p ) , then a ( p k n + s ) = ( λ A s ) ( A p k n ) v a(p^k n+s) = (\lambda A^s)(A^{p^k}n)v a ( p k n + s ) = ( λ A s ) ( A p k n ) v . Modulo p α p^\alpha p α there are at most p ( r + r 2 ) α p^{(r+ r^2)\alpha} p ( r + r 2 ) α pairs ( λ A s , A p k ) (\lambda A^s, A^{p^k}) ( λ A s , A p k ) . A rational series has such a representation: normalize Q = 1 + q 1 x + ⋯ + q d x d Q = 1+q_1x+\cdots+q_dx^d Q = 1 + q 1 x + ⋯ + q d x d and put r = max ( d , deg N + 1 , 1 ) r = \max(d,\deg N+1,1) r = max ( d , deg N + 1 , 1 ) . The identity Q ∑ a ( n ) x n = N Q\sum a(n)x^n = N Q ∑ a ( n ) x n = N gives a ( n ) = − ∑ i = 1 d q i a ( n − i ) a(n) = -\sum_{i=1}^d q_i a(n-i) a ( n ) = − ∑ i = 1 d q i a ( n − i ) for n ≥ r n \ge r n ≥ r . Padding the coefficients to order r r r gives a fixed matrix on the vector ( a ( n ) , … , a ( n + r − 1 ) ) T (a(n),\ldots,a(n+r-1))^{\mathsf{T}} ( a ( n ) , … , a ( n + r − 1 ) ) T for every n ≥ 0 n \ge0 n ≥ 0 , with output its first coordinate. The new general conclusion concerns arbitrary rational diagonals, for which a fixed-rank representation need not exist.
Rowland–Yassawi. The states in [19 ] are Laurent polynomials S S S , written in a base-p Q \frac{p}{Q} Q p numeration system as S = ( ∑ k < α T k ( p / Q ) k ) Q p α − 1 S=\left(\sum_{k<\alpha}T_k(p/Q)^k\right)Q^{p^\alpha-1} S = ( ∑ k < α T k ( p / Q ) k ) Q p α − 1 with Laurent-polynomial digits T k T_k T k whose degrees can grow from one digit to the next; bounding these degrees produces the polynomial exponents above. The present argument describes the same kernel elements in different coordinates: by the δ \delta δ -decomposition each kernel element is a combination of numerators supported on the fixed set J J J , whose size does not depend on α \alpha α , and congruences between successive numerators make all but one p p p -adic digit of each coefficient predictable.
Beukers. Beukers [5 ] constructs p p p -linear schemes modulo p r p^r p r , that is, linear representations a k p + ℓ ≡ M ℓ a k ( m o d p r ) a_{kp+\ell}\equiv M_\ell a_k\pmod{p^r} a k p + ℓ ≡ M ℓ a k ( mod p r ) for a vector of sequences. For the coefficients of a rational function in n n n variables with integer coefficients, whose denominator is not divisible by p p p at the origin and whose numerator satisfies a support condition, he obtains a scheme with at most ρ n d 1 ⋯ d n \rho^n d_1\cdots d_n ρ n d 1 ⋯ d n states whenever ρ − ⌈ ρ / p ⌉ ≥ r − 1 \rho-\lceil\rho/p\rceil\ge r-1 ρ − ⌈ ρ / p ⌉ ≥ r − 1 [5 Corollary 4.2] . The number of states of a scheme is the rank of a linear representation, and here it grows polynomially with the precision. Counting state vectors, [19 §1] deduces (when deg x i P ≤ deg x i Q \deg_{x_i}P\le\deg_{x_i}Q deg x i P ≤ deg x i Q ) automata with at most p α ( ρ h 1 + 1 ) ⋯ ( ρ h m + 1 ) p^{\alpha(\rho h_1+1)\cdots(\rho h_m+1)} p α ( ρ h 1 + 1 ) ⋯ ( ρ h m + 1 ) states, again with an exponent of order α m + 1 \alpha^{m+1} α m + 1 . By Proposition 8.1, for every irrational diagonal with integer coefficients the rank of any linear representation modulo p α p^\alpha p α must tend to infinity with α \alpha α , so bounds of the form p α ⋅ rank p^{\alpha\cdot\operatorname{rank}} p α ⋅ rank obtained by counting vectors have superlinear exponents. Our count concerns cardinality, not rank.
Kohen. Kohen [14 ] constructs, for constant-term sequences CT [ P n Q ] \operatorname{CT}[P^nQ] CT [ P n Q ] modulo p a p^a p a , a p p p -linear representation indexed by Laurent polynomials of degree at most max ( deg P ~ − 1 , deg Q 0 ) \max(\deg\widetilde{P}-1,\deg Q_0) max ( deg P − 1 , deg Q 0 ) , where P ~ \widetilde{P} P is defined by P p a − 1 ≡ P ~ ( x p ℓ ) ( m o d p a ) P^{p^{a-1}} \equiv\widetilde{P}(x^{p^\ell})\pmod{p^a} P p a − 1 ≡ P ( x p ℓ ) ( mod p a ) with ℓ \ell ℓ maximal. One may choose deg P ~ ≤ p a − 1 − ℓ deg P \deg\widetilde{P}\le p^{a-1-\ell}\deg P deg P ≤ p a − 1 − ℓ deg P ; equality holds for P = 1 + x P=1+x P = 1 + x , for which ℓ = 0 \ell=0 ℓ = 0 , so the size bound for this representation can grow exponentially in a a a . The paper is concerned with uniform recurrence, not with bounds on kernel size, and it does not give a bound of the type considered here.
Tools. The δ \delta δ -polynomials and their divisibility and support properties are due to Vlasenko [22 Lemmas 3 and 4] , and the congruence closest to our Lemma 3.2 is [22 Lemma 3(ii)] . Expansions with Frobenius-twisted denominators are central in the theory of Dwork crystals [6 ] , where they are used for the Cartier operator of the digit 0. Earlier automaton constructions and Beukers [5 Theorem 4.1] already use sections for arbitrary input digits (indeed, arbitrary residue vectors). The contribution here is the combination of the δ \delta δ -level decomposition along synchronous digit words with the fixed stable support set J J J and the count of compatible histories. We are not aware of such a precision-linear cardinality bound in the literature inspected; this is not a claim of an exhaustive priority search.
Rank versus cardinality
The set of compatible histories with a fixed window is a finite abelian group of order exactly p ∣ J ∣ α p^{|J|\alpha} p ∣ J ∣ α (Lemma 5.7), but we make no claim about its rank. More importantly, for irrational diagonals with integer coefficients no description of the whole kernel by modules of bounded rank can exist (Proposition 8.1). The saving comes from the cardinality of the groups of histories, not from a bounded-dimensional module.
What remains open
For the odd family of Corollary 1.4, Theorem 7.1 determines the growth exponent to be 3. It does not determine a leading constant or prove convergence of K α ( F ) / 8 α K_\alpha(F)/8^\alpha K α ( F ) / 8 α . For a = b = c = d = 1 a=b=c=d=1 a = b = c = d = 1 , the exact data through α = 9 \alpha=9 α = 9 (§ 9 ) give 0.14 < K α ( F ) / 8 α < 0.17 0.14<K_\alpha(F)/8^\alpha<0.17 0.14 < K α ( F ) / 8 α < 0.17 for 6 ≤ α ≤ 9 6\le\alpha\le9 6 ≤ α ≤ 9 , but these finite values do not settle either question. Determining the exponent for general rational diagonals also remains open. Theorem 1.2 requires a representation as a diagonal with unit constant term in the denominator; we do not treat algebraic series without such a representation.
Verification and reproducibility
All proofs are given in full. The ancillary files contain the code, exact data and logs for the computations of § 9 and the finite certificate in Appendix B, with instructions for reproducing them. The latter certificate is supported by an all-precision lattice argument that proves the sufficiency of its finite precision. The comparison with earlier work above is based on the cited sources and on targeted searches; it is not an exhaustive priority search.
Organisation
Section 2 fixes notation for kernels, sections and diagonals. Section 3 proves the δ \delta δ -lemmas and the level formula along a word. Section 4 proves that the support set J J J is stable. Section 5 introduces histories and windows, proves the decoding, compatibility and counting lemmas, and proves Theorem 1.2 together with the explicit automaton. Section 6 proves the corollaries, Section 7 proves the exact exponent for the odd quadratic family, Section 8 discusses rank, Section 9 reports computations, and Section 10 lists open questions. Appendix A proves the exact zero-orbit theorem used in Corollary 1.4, and Appendix B gives the rational lattice and finite certificate for Section 7.
Kernels, sections and diagonals
Kernels
Let Σ p = { 0 , 1 , … , p − 1 } \Sigma_p = \{0,1,\ldots,p-1\} Σ p = { 0 , 1 , … , p − 1 } . A word w = r 0 r 1 ⋯ r k − 1 ∈ Σ p k w = r_0r_1\cdots r_{k-1} \in\Sigma_p^k w = r 0 r 1 ⋯ r k − 1 ∈ Σ p k is read least significant digit first: v ( w ) = ∑ q = 0 k − 1 r q p q v(w) = \sum_{q=0}^{k-1} r_qp^q v ( w ) = ∑ q = 0 k − 1 r q p q and ∣ w ∣ = k |w| = k ∣ w ∣ = k . Every integer 0 ≤ v < p k 0 \le v < p^k 0 ≤ v < p k equals v ( w ) v(w) v ( w ) for exactly one w ∈ Σ p k w \in\Sigma_p^k w ∈ Σ p k . For a power series g = ∑ n g ( n ) x n g = \sum_n g(n)x^n g = ∑ n g ( n ) x n and r ∈ Σ p r \in\Sigma_p r ∈ Σ p put Λ r g = ∑ n g ( p n + r ) x n \Lambda_r g = \sum_n g(pn+r)x^n Λ r g = ∑ n g ( p n + r ) x n , and Λ w = Λ r k − 1 ∘ ⋯ ∘ Λ r 0 \Lambda_w = \Lambda_{r_{k-1}}\circ\cdots\circ\Lambda_{r_0} Λ w = Λ r k − 1 ∘ ⋯ ∘ Λ r 0 . Induction on k k k gives
Λ w g = ∑ n ≥ 0 g ( p k n + v ( w ) ) x n . (1) \Lambda_w g = \sum_{n\ge0} g(p^k n+v(w))x^n. \tag*{(1)} Λ w g = n ≥ 0 ∑ g ( p k n + v ( w )) x n . (1)
Indeed, Λ r Λ w g = ∑ n g ( p k ( p n + r ) + v ( w ) ) x n \Lambda_r\Lambda_w g = \sum_n g(p^k(pn+r)+v(w))x^n Λ r Λ w g = ∑ n g ( p k ( p n + r ) + v ( w )) x n and p k r + v ( w ) = v ( w r ) p^k r+v(w)=v(wr) p k r + v ( w ) = v ( w r ) . Hence
K α ( g ) = { Λ w g m o d p α : w ∈ Σ p ∗ } , (2) K_\alpha(g)=\{\Lambda_wg\bmod p^\alpha:w\in\Sigma_p^*\}, \tag*{(2)} K α ( g ) = { Λ w g mod p α : w ∈ Σ p ∗ } , (2)
and K α ( g ) K_\alpha(g) K α ( g ) depends only on g m o d p α g\bmod p^\alpha g mod p α .
For completeness, the residual series Λ w g \Lambda_wg Λ w g themselves form an automaton: the digit-r r r transition is S ↦ Λ r S S\mapsto\Lambda_rS S ↦ Λ r S and its output is [ x 0 ] S [x^0]S [ x 0 ] S . Conversely, if two words reach the same state of any automaton generating g g g , then continuing with the digits of any n n n shows that the two residual sequences have the same coefficient at n n n . Thus distinct residuals require distinct reachable states, and the residual automaton attains this lower bound. This proves the minimal-state/p p p -kernel correspondence used above, including leading-zero invariance.
Series in several variables
Write z u = ∏ z i u i z^u=\prod z_i^{u_i} z u = ∏ z i u i for u ∈ Z m u\in\mathbb{Z}^m u ∈ Z m and z ∗ = z 1 = z 1 ⋯ z m z_* = z^1=z_1\cdots z_m z ∗ = z 1 = z 1 ⋯ z m . Let L = ⋃ c ≥ 0 z ∗ − c Z p [ [ z ] ] \mathcal{L}=\bigcup_{c\ge0}z_*^{-c}\mathbb{Z}_p[[z]] L = ⋃ c ≥ 0 z ∗ − c Z p [[ z ]] be the ring of formal Laurent series in z z z whose exponents are bounded below in every coordinate. Define Z p \mathbb{Z}_p Z p -linear maps σ , C : L → L \sigma,C:\mathcal{L}\to\mathcal{L} σ , C : L → L by
σ ( ∑ u s u z u ) = ∑ u s u z p u , C ( ∑ u s u z u ) = ∑ u s p u z u . \sigma\left(\sum_u s_u z^u\right)=\sum_u s_u z^{pu},\qquad C\left(\sum_u s_u z^u\right)=\sum_u s_{pu}z^u. σ ( u ∑ s u z u ) = u ∑ s u z p u , C ( u ∑ s u z u ) = u ∑ s p u z u .
Then σ \sigma σ is a ring endomorphism fixing Z p \mathbb{Z}_p Z p . For r ∈ Σ p r\in\Sigma_p r ∈ Σ p and S ∈ Z p [ [ z ] ] S\in\mathbb{Z}_p[[z]] S ∈ Z p [[ z ]] put
C r ( S ) = C ( z ∗ − r S ) , C_r(S)=C(z_*^{-r}S), C r ( S ) = C ( z ∗ − r S ) ,
and for a word w = r 0 ⋯ r k − 1 w=r_0\cdots r_{k-1} w = r 0 ⋯ r k − 1 put C w = C r k − 1 ∘ ⋯ ∘ C r 0 C_w=C_{r_{k-1}}\circ\cdots\circ C_{r_0} C w = C r k − 1 ∘ ⋯ ∘ C r 0 (with C ∅ = id C_\emptyset=\operatorname{id} C ∅ = id ). We call the C r C_r C r synchronous sections: they select the same residue r r r in every variable.
We use the p p p -adic topology on Z p [ [ z ] ] \mathbb{Z}_p[[z]] Z p [[ z ]] : S n → 0 S_n\to0 S n → 0 if for every e e e we have S n ∈ p e Z p [ [ z ] ] S_n\in p^e\mathbb{Z}_p[[z]] S n ∈ p e Z p [[ z ]] for all large n n n . The ring Z p [ [ z ] ] \mathbb{Z}_p[[z]] Z p [[ z ]] is complete and separated for this topology. All maps used below (σ \sigma σ , C C C , C r C_r C r , multiplication by a fixed series) are Z p \mathbb{Z}_p Z p -linear and map p e Z p [ [ z ] ] p^e\mathbb{Z}_p[[z]] p e Z p [[ z ]] into itself. They are therefore continuous, commute with p p p -adically convergent sums, and are compatible with reduction modulo p α p^\alpha p α .
Lemma 2.1. Let A , Y ∈ L A,Y\in\mathcal{L} A , Y ∈ L , B , S ∈ Z p [ [ z ] ] B,S\in\mathbb{Z}_p[[z]] B , S ∈ Z p [[ z ]] , r ∈ Σ p r\in\Sigma_p r ∈ Σ p and t ≥ 0 t\ge0 t ≥ 0 .
(a) C ( A σ ( B ) ) = C ( A ) B C(A\sigma(B)) = C(A)B C ( A σ ( B )) = C ( A ) B .
(b) C ( z ∗ − r C t ( Y ) ) = C t + 1 ( z ∗ − r p t Y ) C(z_*^{-r}C^t(Y)) = C^{t+1}(z_*^{-rp^t}Y) C ( z ∗ − r C t ( Y )) = C t + 1 ( z ∗ − r p t Y ) .
(c) If t ≥ 1 t \ge1 t ≥ 1 and 0 ≤ v < p t 0 \le v < p^t 0 ≤ v < p t , then C t ( z ∗ − v S ) ∈ Z p [ [ z ] ] C^t(z_*^{-v}S) \in\mathbb{Z}_p[[z]] C t ( z ∗ − v S ) ∈ Z p [[ z ]] , and it is a polynomial if S S S is.
(d) C r C_r C r maps Z p [ [ z ] ] \mathbb{Z}_p[[z]] Z p [[ z ]] to itself, and C w ( S ) = C k ( z ∗ − v ( w ) S ) C_w(S) = C^k(z_*^{-v(w)}S) C w ( S ) = C k ( z ∗ − v ( w ) S ) for every word w w w of length k k k .
Proof. (a) We have A σ ( B ) = ∑ u , w a u b w z u + p w A\sigma(B) = \sum_{u,w}a_ub_wz^{u+pw} A σ ( B ) = ∑ u , w a u b w z u + pw , each coefficient being a finite sum because the exponents of A A A are bounded below and those of B B B are nonnegative. Moreover u + p w ∈ p Z m u+pw \in p\mathbb{Z}^m u + pw ∈ p Z m if and only if u ∈ p Z m u \in p\mathbb{Z}^m u ∈ p Z m . Hence C ( A σ ( B ) ) = ∑ u ′ , w a p u ′ b w z u ′ + w = C ( A ) B C(A\sigma(B)) = \sum_{u',w}a_{pu'}b_wz^{u'+w} = C(A)B C ( A σ ( B )) = ∑ u ′ , w a p u ′ b w z u ′ + w = C ( A ) B .
(b) The coefficient of z u z^u z u on the left is [ z p u + r 1 ] C t ( Y ) = [ z p t ( p u + r 1 ) ] Y = [ z p t + 1 u − r p t ] ( z ∗ − r p t Y ) [z^{pu+r1}]C^t(Y) = [z^{p^t(pu+r1)}]Y = [z^{p^{t+1}u-rp^t}](z_*^{-rp^t}Y) [ z p u + r 1 ] C t ( Y ) = [ z p t ( p u + r 1 ) ] Y = [ z p t + 1 u − r p t ] ( z ∗ − r p t Y ) , which is the coefficient of z u z^u z u on the right.
(c) All exponents of z ∗ − v S z_*^{-v}S z ∗ − v S are ≥ − v > − p t \ge-v > -p^t ≥ − v > − p t in every coordinate. An exponent divisible by p t p^t p t is therefore ≥ 0 \ge0 ≥ 0 , and C t C^t C t divides it by p t p^t p t . Finitely many monomials remain if S S S is a polynomial.
(d) The first claim is (c) with t = 1 t=1 t = 1 . The formula follows from (b) by induction on k k k , since C r ( C k ( z ∗ − v ( w ) S ) ) = C k + 1 ( z ∗ − v ( w ) − r p k S ) C_r(C^k(z_*^{-v(w)}S)) = C^{k+1}(z_*^{-v(w)-rp^k}S) C r ( C k ( z ∗ − v ( w ) S )) = C k + 1 ( z ∗ − v ( w ) − r p k S ) and v ( w ) + r p k = v ( w r ) v(w)+rp^k = v(wr) v ( w ) + r p k = v ( w r ) .
Lemma 2.2. For S ∈ Z p [ [ z ] ] S \in\mathbb{Z}_p[[z]] S ∈ Z p [[ z ]] and r ∈ Σ p r \in\Sigma_p r ∈ Σ p we have Diag ( C r S ) = Λ r Diag ( S ) \operatorname{Diag}(C_rS) = \Lambda_r\operatorname{Diag}(S) Diag ( C r S ) = Λ r Diag ( S ) . Consequently Diag ( C w S ) = Λ w Diag ( S ) \operatorname{Diag}(C_wS) = \Lambda_w\operatorname{Diag}(S) Diag ( C w S ) = Λ w Diag ( S ) for every word w w w , and
K α ( Diag S ) ≤ ∣ { C w S m o d p α : w ∈ Σ p ∗ } ∣ . K_\alpha(\operatorname{Diag} S) \le\left|\{C_wS \bmod p^\alpha: w \in\Sigma_p^*\}\right|. K α ( Diag S ) ≤ { C w S mod p α : w ∈ Σ p ∗ } .
Proof. [ x n ] Diag ( C r S ) = [ z n 1 ] C ( z ∗ − r S ) = [ z p n 1 ] ( z ∗ − r S ) = [ z ( p n + r ) 1 ] S = [ x p n + r ] Diag S [x^n]\operatorname{Diag}(C_rS) = [z^{n1}]C(z_*^{-r}S) = [z^{pn1}](z_*^{-r}S) = [z^{(pn+r)1}]S = [x^{pn+r}]\operatorname{Diag} S [ x n ] Diag ( C r S ) = [ z n 1 ] C ( z ∗ − r S ) = [ z p n 1 ] ( z ∗ − r S ) = [ z ( p n + r ) 1 ] S = [ x p n + r ] Diag S . Iterating gives the second claim. By (2.2), K α ( Diag S ) K_\alpha(\operatorname{Diag} S) K α ( Diag S ) counts the series Λ w Diag S m o d p α = Diag ( C w S m o d p α ) \Lambda_w\operatorname{Diag} S \bmod p^\alpha= \operatorname{Diag}(C_wS \bmod p^\alpha) Λ w Diag S mod p α = Diag ( C w S mod p α ) , and each of them is a function of C w S m o d p α C_wS \bmod p^\alpha C w S mod p α .
Normalisation
Let N , Q ∈ Z p [ z ] N,Q \in\mathbb{Z}_p[z] N , Q ∈ Z p [ z ] with c : = Q ( 0 ) ∈ Z p × c := Q(0) \in\mathbb{Z}_p^\times c := Q ( 0 ) ∈ Z p × , and put R = N / Q ∈ Z p [ [ z ] ] R=N/Q \in\mathbb{Z}_p[[z]] R = N / Q ∈ Z p [[ z ]] and G = Diag R G=\operatorname{Diag} R G = Diag R . Replacing ( N , Q ) (N,Q) ( N , Q ) by ( c − 1 N , c − 1 Q ) (c^{-1}N,c^{-1}Q) ( c − 1 N , c − 1 Q ) changes neither R R R nor the supports, degrees and Newton polytopes. From now on we assume Q ( 0 ) = 1 Q(0)=1 Q ( 0 ) = 1 . We write Δ = Δ ( Q ) \Delta=\Delta(Q) Δ = Δ ( Q ) , which contains 0 0 0 , and K = K ( N ) K=K(N) K = K ( N ) . Both lie in the box ∏ i [ 0 , h i ] \prod_i[0,h_i] ∏ i [ 0 , h i ] .
The delta-decomposition and levels
The delta-polynomials
Define δ s ∈ Z p [ z ] \delta_s \in\mathbb{Z}_p[z] δ s ∈ Z p [ z ] for s ≥ 1 s \ge1 s ≥ 1 recursively by
Q p s − 1 = ∑ t = 1 s δ t σ t ( Q p s − t − 1 ) (3) Q^{p^s-1}=\sum_{t=1}^{s}\delta_t\sigma^t\left(Q^{p^{s-t}-1}\right) \tag*{(3)} Q p s − 1 = t = 1 ∑ s δ t σ t ( Q p s − t − 1 ) (3)
that is, δ 1 = Q p − 1 \delta_1=Q^{p-1} δ 1 = Q p − 1 and δ s = Q p s − 1 − ∑ t = 1 s − 1 δ t σ t ( Q p s − t − 1 ) \delta_s=Q^{p^s-1}-\sum_{t=1}^{s-1}\delta_t\sigma^t\left(Q^{p^{s-t}-1}\right) δ s = Q p s − 1 − ∑ t = 1 s − 1 δ t σ t ( Q p s − t − 1 ) .
We shall use two elementary facts, valid for every prime p p p (including p = 2 p=2 p = 2 ) and for elements of Z p [ z ] \mathbb{Z}_p[z] Z p [ z ] or Z p [ [ z ] ] \mathbb{Z}_p[[z]] Z p [[ z ]] :
(F1) σ ( A ) ≡ A p ( m o d p ) \sigma(A)\equiv A^p\pmod p σ ( A ) ≡ A p ( mod p ) , and hence σ t ( A ) ≡ A p t ( m o d p ) \sigma^t(A)\equiv A^{p^t}\pmod p σ t ( A ) ≡ A p t ( mod p ) for t ≥ 0 t\ge0 t ≥ 0 . Indeed ( ∑ c u z u ) p ≡ ∑ c u p z p u (\sum c_uz^u)^p\equiv\sum c_u^pz^{pu} ( ∑ c u z u ) p ≡ ∑ c u p z p u and c p ≡ c ( m o d p ) c^p\equiv c\pmod p c p ≡ c ( mod p ) for c ∈ Z p c\in\mathbb{Z}_p c ∈ Z p . Note that σ ( A ) \sigma(A) σ ( A ) is again integral, so the congruence can be iterated.
(F2) If A ≡ B ( m o d p e ) A\equiv B\pmod{p^e} A ≡ B ( mod p e ) with e ≥ 1 e\ge1 e ≥ 1 , then A p ≡ B p ( m o d p e + 1 ) A^p\equiv B^p\pmod{p^{e+1}} A p ≡ B p ( mod p e + 1 ) . Indeed, writing A = B + p e X A=B+p^eX A = B + p e X , the term ( p j ) B p − j p e j X j \binom pjB^{p-j}p^{ej}X^j ( j p ) B p − j p e j X j is divisible by p e + 1 p^{e+1} p e + 1 for j = 1 j=1 j = 1 (as p ∣ ( p 1 ) p\mid\binom p1 p ∣ ( 1 p ) ) and for j ≥ 2 j\ge2 j ≥ 2 (as e j ≥ e + 1 ej\ge e+1 e j ≥ e + 1 ). Consequently A ≡ B ( m o d p ) A\equiv B\pmod p A ≡ B ( mod p ) implies A p e ≡ B p e ( m o d p e + 1 ) A^{p^e}\equiv B^{p^e}\pmod{p^{e+1}} A p e ≡ B p e ( mod p e + 1 ) .
Lemma 3.1 (Vlasenko [22 Lemmas 3 and 4] ). For every s ≥ 1 s\ge1 s ≥ 1 : (a) supp δ s ⊂ ( p s − 1 ) Δ \operatorname{supp}\delta_s\subset(p^s-1)\Delta supp δ s ⊂ ( p s − 1 ) Δ ; (b) δ s ∈ p s − 1 Z p [ z ] \delta_s\in p^{s-1}\mathbb{Z}_p[z] δ s ∈ p s − 1 Z p [ z ] .
Proof. (a) For s = 1 s=1 s = 1 , δ 1 = Q p − 1 \delta_1=Q^{p-1} δ 1 = Q p − 1 . For s ≥ 2 s\ge2 s ≥ 2 , supp Q p s − 1 ⊂ ( p s − 1 ) Δ \operatorname{supp}Q^{p^s-1}\subset(p^s-1)\Delta supp Q p s − 1 ⊂ ( p s − 1 ) Δ , and by induction
supp ( δ t σ t ( Q p s − t − 1 ) ) ⊂ ( p t − 1 ) Δ + p t ( p s − t − 1 ) Δ = ( p s − 1 ) Δ , \operatorname{supp}\left(\delta_t\sigma^t\left(Q^{p^{s-t}-1}\right)\right)\subset(p^t-1)\Delta+p^t(p^{s-t}-1)\Delta=(p^s-1)\Delta, supp ( δ t σ t ( Q p s − t − 1 ) ) ⊂ ( p t − 1 ) Δ + p t ( p s − t − 1 ) Δ = ( p s − 1 ) Δ ,
since a Δ + b Δ = ( a + b ) Δ a\Delta+b\Delta=(a+b)\Delta a Δ + b Δ = ( a + b ) Δ for a convex set Δ \Delta Δ and a , b ≥ 0 a,b\geq0 a , b ≥ 0 .
(b) Induction on s s s ; the case s = 1 s=1 s = 1 is trivial. Let s ≥ 2 s\geq2 s ≥ 2 and assume δ t ∈ p t − 1 Z p [ z ] \delta_t\in p^{t-1}\mathbb{Z}_p[z] δ t ∈ p t − 1 Z p [ z ] for t < s t<s t < s . For 1 ≤ t < s 1\leq t<s 1 ≤ t < s ,
σ t ( Q p s − 1 ) = σ t ( Q p s − 1 − t − 1 ) ⋅ σ t ( Q p s − 1 − 1 ) p s − 1 − t . \sigma^t(Q^{p^s-1})=\sigma^t(Q^{p^{s-1-t}-1})\cdot\sigma^t(Q^{p^{s-1}-1})p^{s-1-t}. σ t ( Q p s − 1 ) = σ t ( Q p s − 1 − t − 1 ) ⋅ σ t ( Q p s − 1 − 1 ) p s − 1 − t .
By (F1), σ t ( Q p − 1 ) ≡ Q ( p − 1 ) p t ( m o d p ) \sigma^t(Q^{p-1})\equiv Q^{(p-1)p^t}\pmod p σ t ( Q p − 1 ) ≡ Q ( p − 1 ) p t ( mod p ) , and by (F2), σ t ( Q p s − 1 − t − 1 ) ≡ Q ( p − 1 ) p s − 1 ( m o d p s − t ) \sigma^t(Q^{p^{s-1-t}-1})\equiv Q^{(p-1)p^{s-1}}\pmod{p^{s-t}} σ t ( Q p s − 1 − t − 1 ) ≡ Q ( p − 1 ) p s − 1 ( mod p s − t ) . Multiplying by δ t ∈ p t − 1 \delta_t\in p^{t-1} δ t ∈ p t − 1 we get
δ t σ t ( Q p s − t − 1 ) ≡ δ t σ t ( Q p s − 1 − t − 1 ) Q p s − p s − 1 ( m o d p s − 1 ) . \delta_t\sigma^t(Q^{p^s-t-1})\equiv\delta_t\sigma^t(Q^{p^{s-1-t}-1})Q^{p^s-p^{s-1}}\pmod{p^{s-1}}. δ t σ t ( Q p s − t − 1 ) ≡ δ t σ t ( Q p s − 1 − t − 1 ) Q p s − p s − 1 ( mod p s − 1 ) .
Summing over t < s t<s t < s and using (3) for s − 1 s-1 s − 1 ,
δ s ≡ Q p s − 1 − Q p s − p s − 1 ∑ t = 1 s − 1 δ t σ t ( Q p s − 1 − t − 1 ) = Q p s − 1 − Q p s − p s − 1 Q p s − 1 − 1 = 0 ( m o d p s − 1 ) . \delta_s\equiv Q^{p^s-1}-Q^{p^s-p^{s-1}}\sum_{t=1}^{s-1}\delta_t\sigma^t(Q^{p^{s-1-t}-1})=Q^{p^s-1}-Q^{p^s-p^{s-1}}Q^{p^{s-1}-1}=0\pmod{p^{s-1}}. δ s ≡ Q p s − 1 − Q p s − p s − 1 t = 1 ∑ s − 1 δ t σ t ( Q p s − 1 − t − 1 ) = Q p s − 1 − Q p s − p s − 1 Q p s − 1 − 1 = 0 ( mod p s − 1 ) .
Lemma 3.2. For every s ≥ 1 s\geq1 s ≥ 1 ,
Q − 1 ≡ ∑ t = 1 s δ t σ t ( Q − 1 ) ( m o d p s Z p [ [ z ] ] ) . Q^{-1}\equiv\sum_{t=1}^{s}\delta_t\sigma^t(Q^{-1})\pmod{p^s\mathbb{Z}_p[[z]]}. Q − 1 ≡ t = 1 ∑ s δ t σ t ( Q − 1 ) ( mod p s Z p [[ z ]]) . Consequently Q − 1 = ∑ t ≥ 1 δ t σ t ( Q − 1 ) Q^{-1}=\sum_{t\geq1}\delta_t\sigma^t(Q^{-1}) Q − 1 = ∑ t ≥ 1 δ t σ t ( Q − 1 ) in Z p [ [ z ] ] \mathbb{Z}_p[[z]] Z p [[ z ]] , the series converging p p p -adically.
Proof. Since Q ( 0 ) = 1 Q(0)=1 Q ( 0 ) = 1 , the series Q Q Q and σ t ( Q ) \sigma^t(Q) σ t ( Q ) are units of Z p [ [ z ] ] \mathbb{Z}_p[[z]] Z p [[ z ]] . Put u t = σ t ( Q ) Q − p t u_t=\sigma^t(Q)Q^{-p^t} u t = σ t ( Q ) Q − p t . By (F1), u t ∈ 1 + p Z p [ [ z ] ] u_t\in1+p\mathbb{Z}_p[[z]] u t ∈ 1 + p Z p [[ z ]] , so by (F2), u t p e ∈ 1 + p e + 1 Z p [ [ z ] ] u_t^{p^e}\in1+p^{e+1}\mathbb{Z}_p[[z]] u t p e ∈ 1 + p e + 1 Z p [[ z ]] for e ≥ 0 e\geq0 e ≥ 0 . Since
σ t ( Q p s − t − 1 ) = σ t ( Q ) p s − t σ t ( Q ) − 1 = u t p s − t Q p s σ t ( Q − 1 ) , \sigma^t(Q^{p^{s-t}-1})=\sigma^t(Q)^{p^{s-t}}\sigma^t(Q)^{-1}=u_t^{p^{s-t}}Q^{p^s}\sigma^t(Q^{-1}), σ t ( Q p s − t − 1 ) = σ t ( Q ) p s − t σ t ( Q ) − 1 = u t p s − t Q p s σ t ( Q − 1 ) ,
dividing (3) by Q p s Q^{p^s} Q p s gives Q − 1 = ∑ t = 1 s δ t σ t ( Q − 1 ) u t p s − t Q^{-1}=\sum_{t=1}^{s}\delta_t\sigma^t(Q^{-1})u_t^{p^{s-t}} Q − 1 = ∑ t = 1 s δ t σ t ( Q − 1 ) u t p s − t . Therefore
Q − 1 − ∑ t = 1 s δ t σ t ( Q − 1 ) = ∑ t = 1 s δ t σ t ( Q − 1 ) ( u t p s − t − 1 ) ∈ ∑ t = 1 s p t − 1 p s − t + 1 Z p [ [ z ] ] = p s Z p [ [ z ] ] , Q^{-1}-\sum_{t=1}^{s}\delta_t\sigma^t(Q^{-1})=\sum_{t=1}^{s}\delta_t\sigma^t(Q^{-1})(u_t^{p^{s-t}}-1)\in\sum_{t=1}^{s}p^{t-1}p^{s-t+1}\mathbb{Z}_p[[z]]=p^s\mathbb{Z}_p[[z]], Q − 1 − t = 1 ∑ s δ t σ t ( Q − 1 ) = t = 1 ∑ s δ t σ t ( Q − 1 ) ( u t p s − t − 1 ) ∈ t = 1 ∑ s p t − 1 p s − t + 1 Z p [[ z ]] = p s Z p [[ z ]] ,
using Lemma 3.1 (b). The terms with t > s t>s t > s lie in p s Z p [ [ z ] ] p^s\mathbb{Z}_p[[z]] p s Z p [[ z ]] , which gives the second statement.
Levels
Definition 3.3. A level expansion of S ∈ Z p [ [ z ] ] S\in\mathbb{Z}_p[[z]] S ∈ Z p [[ z ]] is a sequence ( M j ) j ≥ 0 (M_j)_{j\geq0} ( M j ) j ≥ 0 of polynomials in Z p [ z ] \mathbb{Z}_p[z] Z p [ z ] with M j → 0 M_j\to0 M j → 0 p p p -adically and
S = ∑ j ≥ 0 M j σ j ( Q − 1 ) . S=\sum_{j\geq0}M_j\sigma^j(Q^{-1}). S = j ≥ 0 ∑ M j σ j ( Q − 1 ) . The polynomial M j M_j M j is the numerator at level j j j .
Lemma 3.4 (one digit). If ( M j ) j ≥ 0 (M_j)_{j\geq0} ( M j ) j ≥ 0 is a level expansion of S S S and r ∈ Σ p r\in\Sigma_p r ∈ Σ p , then ( C r ( M j + 1 + δ j + 1 M 0 ) ) j ≥ 0 (C_r(M_{j+1}+\delta_{j+1}M_0))_{j\geq0} ( C r ( M j + 1 + δ j + 1 M 0 ) ) j ≥ 0 is a level expansion of C r S C_rS C r S .
Proof. By Lemma 3.2 , M 0 Q − 1 = ∑ j ≥ 0 δ j + 1 M 0 σ j + 1 ( Q − 1 ) M_0Q^{-1}=\sum_{j\geq0}\delta_{j+1}M_0\sigma^{j+1}(Q^{-1}) M 0 Q − 1 = ∑ j ≥ 0 δ j + 1 M 0 σ j + 1 ( Q − 1 ) . Hence S = ∑ j ≥ 0 A j σ ( σ j ( Q − 1 ) ) S=\sum_{j\geq0}A_j\sigma(\sigma^j(Q^{-1})) S = ∑ j ≥ 0 A j σ ( σ j ( Q − 1 )) with A j = M j + 1 + δ j + 1 M 0 A_j=M_{j+1}+\delta_{j+1}M_0 A j = M j + 1 + δ j + 1 M 0 , and A j → 0 A_j\to0 A j → 0 . By continuity of C r C_r C r and Lemma 2.1(a) applied to z ∗ − r A j ∈ L z_*^{-r}A_j\in L z ∗ − r A j ∈ L ,
C r S = ∑ j ≥ 0 C ( z ∗ − r A j σ ( σ j ( Q − 1 ) ) ) = ∑ j ≥ 0 C r ( A j ) σ j ( Q − 1 ) . C_rS=\sum_{j\geq0}C(z_*^{-r}A_j\sigma(\sigma^j(Q^{-1})))=\sum_{j\geq0}C_r(A_j)\sigma^j(Q^{-1}). C r S = j ≥ 0 ∑ C ( z ∗ − r A j σ ( σ j ( Q − 1 ))) = j ≥ 0 ∑ C r ( A j ) σ j ( Q − 1 ) .
Each C r ( A j ) C_r(A_j) C r ( A j ) is a polynomial by Lemma 2.1(c), and C r ( A j ) → 0 C_r(A_j)\to0 C r ( A j ) → 0 .
Definition 3.5. For t ≥ 1 t \ge1 t ≥ 1 , 0 ≤ v < p t 0 \le v < p^t 0 ≤ v < p t and M ∈ Z p [ z ] M \in\mathbb{Z}_p[z] M ∈ Z p [ z ] put
T t , v ( M ) = C t ( z ∗ − v δ t M ) ∈ Z p [ z ] . T_{t,v}(M)=C^t\left(z_*^{-v}\delta_t M\right)\in\mathbb{Z}_p[z]. T t , v ( M ) = C t ( z ∗ − v δ t M ) ∈ Z p [ z ] .
By Lemma 3.1(b), T t , v ( M ) ∈ p t − 1 Z p [ z ] T_{t,v}(M)\in p^{t-1}\mathbb{Z}_p[z] T t , v ( M ) ∈ p t − 1 Z p [ z ] . For a word w = r 0 ⋯ r k − 1 w=r_0\cdots r_{k-1} w = r 0 ⋯ r k − 1 and 1 ≤ t ≤ i ≤ k 1\le t\le i\le k 1 ≤ t ≤ i ≤ k let
v i , t = v i , t ( w ) = ∑ q = 0 t − 1 r i − t + q p q ∈ [ 0 , p t ) , v_{i,t}=v_{i,t}(w)=\sum_{q=0}^{t-1}r_{i-t+q}p^q\in[0,p^t), v i , t = v i , t ( w ) = q = 0 ∑ t − 1 r i − t + q p q ∈ [ 0 , p t ) ,
the integer whose digits (least significant first) are r i − t , … , r i − 1 r_{i-t},\ldots,r_{i-1} r i − t , … , r i − 1 . The zero-level numerators of w w w are
m 0 = N , m i = ∑ t = 1 i T t , v i , t ( m i − t ) ( 1 ≤ i ≤ k ) . (4) m_0=N,\qquad m_i=\sum_{t=1}^{i}T_{t,v_{i,t}}(m_{i-t})\qquad(1\le i\le k). \tag*{(4)} m 0 = N , m i = t = 1 ∑ i T t , v i , t ( m i − t ) ( 1 ≤ i ≤ k ) . (4)
Note that m i m_i m i depends only on r 0 , … , r i − 1 r_0,\ldots,r_{i-1} r 0 , … , r i − 1 .
Proposition 3.6 (levels along a word). Let w = r 0 ⋯ r k − 1 w=r_0\cdots r_{k-1} w = r 0 ⋯ r k − 1 be a word. Then C w ( R ) C_w(R) C w ( R ) has the level expansion
M 0 ( k ) = m k , M j ( k ) = ∑ t = 1 k C t ( z ∗ − v k , t δ j + t m k − t ) ( j ≥ 1 ) . M_0^{(k)}=m_k,\qquad M_j^{(k)}=\sum_{t=1}^{k}C^t\left(z_*^{-v_{k,t}}\delta_{j+t}m_{k-t}\right)\qquad(j\ge1). M 0 ( k ) = m k , M j ( k ) = t = 1 ∑ k C t ( z ∗ − v k , t δ j + t m k − t ) ( j ≥ 1 ) . In particular M j ( k ) ∈ p j Z p [ z ] M_j^{(k)}\in p^j\mathbb{Z}_p[z] M j ( k ) ∈ p j Z p [ z ] for j ≥ 1 j\ge1 j ≥ 1 .
Proof. For k ≥ 1 k\ge1 k ≥ 1 the two displayed formulas are the cases j = 0 j=0 j = 0 and j ≥ 1 j\ge1 j ≥ 1 of the single formula M j ( k ) = ∑ t = 1 k C t ( z ∗ − v k , t δ j + t m k − t ) M_j^{(k)}=\sum_{t=1}^{k}C^t\left(z_*^{-v_{k,t}}\delta_{j+t}m_{k-t}\right) M j ( k ) = ∑ t = 1 k C t ( z ∗ − v k , t δ j + t m k − t ) , by (4) . We argue by induction on k k k . For k = 0 k=0 k = 0 , R = N ⋅ Q − 1 R=N\cdot Q^{-1} R = N ⋅ Q − 1 has the level expansion M 0 = N M_0=N M 0 = N , M j = 0 M_j=0 M j = 0 for j ≥ 1 j\ge1 j ≥ 1 . Suppose the claim holds for w w w of length k k k , and let r = r k r=r_k r = r k . By Lemma 3.4, C w r ( R ) = C r ( C w R ) C_{wr}(R)=C_r(C_wR) C w r ( R ) = C r ( C w R ) has the level expansion M j ( k + 1 ) = C r ( M j + 1 ( k ) + δ j + 1 m k ) M_j^{(k+1)}=C_r(M_{j+1}^{(k)}+\delta_{j+1}m_k) M j ( k + 1 ) = C r ( M j + 1 ( k ) + δ j + 1 m k ) . First,
C r ( δ j + 1 m k ) = C ( z ∗ − v k + 1 , 1 δ j + 1 m ( k + 1 ) − 1 ) , C_r(\delta_{j+1}m_k)=C\left(z_*^{-v_{k+1,1}}\delta_{j+1}m_{(k+1)-1}\right), C r ( δ j + 1 m k ) = C ( z ∗ − v k + 1 , 1 δ j + 1 m ( k + 1 ) − 1 ) ,
since v k + 1 , 1 = r k v_{k+1,1}=r_k v k + 1 , 1 = r k . Second, by Lemma 2.1(b),
C r ( C t ( z ∗ − v k , t δ j + 1 + t m k − t ) ) = C t + 1 ( z ∗ − ( v k , t + r k p t ) δ j + 1 + t m k − t ) , C_r\left(C^t\left(z_*^{-v_{k,t}}\delta_{j+1+t}m_{k-t}\right)\right)
=C^{t+1}\left(z_*^{-(v_{k,t}+r_kp^t)}\delta_{j+1+t}m_{k-t}\right), C r ( C t ( z ∗ − v k , t δ j + 1 + t m k − t ) ) = C t + 1 ( z ∗ − ( v k , t + r k p t ) δ j + 1 + t m k − t ) ,
and v k , t + r k p t = v k + 1 , t + 1 v_{k,t}+r_kp^t=v_{k+1,t+1} v k , t + r k p t = v k + 1 , t + 1 . Renaming t + 1 t+1 t + 1 as t t t , we obtain M j ( k + 1 ) = ∑ t = 1 k + 1 C t ( z ∗ − v k + 1 , t δ j + t m k + 1 − t ) M_j^{(k+1)}=\sum_{t=1}^{k+1}C^t\left(z_*^{-v_{k+1,t}}\delta_{j+t}m_{k+1-t}\right) M j ( k + 1 ) = ∑ t = 1 k + 1 C t ( z ∗ − v k + 1 , t δ j + t m k + 1 − t ) for every j ≥ 0 j\ge0 j ≥ 0 , as claimed. The divisibility follows from δ j + t ∈ p j + t − 1 Z p [ z ] \delta_{j+t}\in p^{j+t-1}\mathbb{Z}_p[z] δ j + t ∈ p j + t − 1 Z p [ z ] , and it also shows M j ( k ) → 0 M_j^{(k)}\to0 M j ( k ) → 0 .
:::
A support set stable under all digit operators
Let J = J p ( N , Q ) J=J_p(N,Q) J = J p ( N , Q ) be as in Definition 1.1, and let L J = ⨁ u ∈ J Z p z u L_J=\bigoplus_{u\in J}\mathbb{Z}_p z^u L J = ⨁ u ∈ J Z p z u be the polynomials supported on J J J .
Lemma 4.1. (a) J ⊂ ∏ i = 1 m { 0 , 1 , … , h i } J\subset\prod_{i=1}^{m}\{0,1,\ldots,h_i\} J ⊂ ∏ i = 1 m { 0 , 1 , … , h i } , so ∣ J ∣ ≤ D = ∏ i ( h i + 1 ) |J|\le D=\prod_i(h_i+1) ∣ J ∣ ≤ D = ∏ i ( h i + 1 ) .
(b) If Δ ⊂ Δ ′ \Delta\subset\Delta' Δ ⊂ Δ ′ and K ⊂ K ′ K\subset K' K ⊂ K ′ , then J J J is contained in the set defined by Δ ′ \Delta' Δ ′ and K ′ K' K ′ .
Proof. (a) If u + η 1 = ( 1 − ε ) a + ε b u+\eta\mathbf{1}=(1-\varepsilon)a+\varepsilon b u + η 1 = ( 1 − ε ) a + ε b with a ∈ Δ a\in\Delta a ∈ Δ , b ∈ K b\in K b ∈ K , then 0 ≤ u i ≤ u i + η ≤ ( 1 − ε ) h i + ε h i = h i 0\le u_i\le u_i+\eta\le(1-\varepsilon)h_i+\varepsilon h_i=h_i 0 ≤ u i ≤ u i + η ≤ ( 1 − ε ) h i + ε h i = h i .
(b) is clear.
:::
Lemma 4.2 (stability). Let t ≥ 1 t\ge1 t ≥ 1 and 0 ≤ v < p t 0\le v<p^t 0 ≤ v < p t . Then T t , v ( L J ) ⊂ L J T_{t,v}(L_J)\subset L_J T t , v ( L J ) ⊂ L J and T t , v ( N ) ∈ L J T_{t,v}(N)\in L_J T t , v ( N ) ∈ L J .
Proof. Let u ∈ J ∪ supp N u \in J \cup\operatorname{supp} N u ∈ J ∪ supp N . If u ∈ J u \in J u ∈ J , choose η ∈ [ 0 , 1 ) \eta\in[0,1) η ∈ [ 0 , 1 ) , ε ∈ ( 0 , 1 / p ] \varepsilon\in(0,1/p] ε ∈ ( 0 , 1/ p ] , a ∈ Δ a \in\Delta a ∈ Δ and b ∈ K b \in K b ∈ K with u + η 1 = ( 1 − ε ) a + ε b u+\eta\mathbf{1}=(1-\varepsilon)a+\varepsilon b u + η 1 = ( 1 − ε ) a + ε b . If u ∈ supp N u \in\operatorname{supp} N u ∈ supp N , put η = 0 \eta=0 η = 0 , ε = 1 \varepsilon=1 ε = 1 , b = u b=u b = u and let a ∈ Δ a \in\Delta a ∈ Δ be arbitrary. The monomials of z ∗ v δ t t z u z_*^v\delta_t^tz^u z ∗ v δ t t z u are z u + e − v 1 z^{u+e-v\mathbf{1}} z u + e − v 1 with e ∈ supp δ t ⊂ ( p t − 1 ) Δ e \in\operatorname{supp}\delta_t \subset(p^t-1)\Delta e ∈ supp δ t ⊂ ( p t − 1 ) Δ , so e = ( p t − 1 ) d e=(p^t-1)d e = ( p t − 1 ) d with d ∈ Δ d \in\Delta d ∈ Δ (Lemma 3.1(a)). Such a monomial survives C t C^t C t only if u + e − v 1 = p t u ′ u+e-v\mathbf{1}=p^tu' u + e − v 1 = p t u ′ with u ′ ∈ Z m u' \in\mathbb{Z}^m u ′ ∈ Z m , and then u ′ ∈ Z ≥ 0 m u' \in\mathbb{Z}_{\geq0}^m u ′ ∈ Z ≥ 0 m by Lemma 2.1(c). Put η ′ = ( η + v ) / p t \eta'=(\eta+v)/p^t η ′ = ( η + v ) / p t . Since 0 ≤ η < 1 0 \leq\eta<1 0 ≤ η < 1 and 0 ≤ v ≤ p t − 1 0 \leq v \leq p^t-1 0 ≤ v ≤ p t − 1 , we have η ′ ∈ [ 0 , 1 ) \eta' \in[0,1) η ′ ∈ [ 0 , 1 ) . Moreover
u ′ + η ′ 1 = p − t ( u + η 1 ) + ( 1 − p − t ) d = ( 1 − ε ′ ) a ′ + ε ′ b , u'+\eta'\mathbf{1}=p^{-t}(u+\eta\mathbf{1})+(1-p^{-t})d=(1-\varepsilon')a'+\varepsilon'b, u ′ + η ′ 1 = p − t ( u + η 1 ) + ( 1 − p − t ) d = ( 1 − ε ′ ) a ′ + ε ′ b ,
where ε ′ = ε p − t ∈ ( 0 , 1 / p ] \varepsilon'=\varepsilon p^{-t}\in(0,1/p] ε ′ = ε p − t ∈ ( 0 , 1/ p ] and
a ′ = p − t ( 1 − ε ) a + ( 1 − p − t ) d 1 − ε ′ a'=\frac{p^{-t}(1-\varepsilon)a+(1-p^{-t})d}{1-\varepsilon'} a ′ = 1 − ε ′ p − t ( 1 − ε ) a + ( 1 − p − t ) d
is a convex combination of a , d ∈ Δ a,d\in\Delta a , d ∈ Δ : its coefficients are nonnegative and add up to ( p − t − ε p − t + 1 − p − t ) / ( 1 − ε ′ ) = 1 (p^{-t}-\varepsilon p^{-t}+1-p^{-t})/(1-\varepsilon')=1 ( p − t − ε p − t + 1 − p − t ) / ( 1 − ε ′ ) = 1 . Hence a ′ ∈ Δ a'\in\Delta a ′ ∈ Δ and u ′ ∈ J u'\in J u ′ ∈ J . Since T t , v T_{t,v} T t , v is Z p \mathbb{Z}_p Z p -linear, both claims follow. □ \square □
Corollary 4.3. For every word w w w and every i ≥ 1 i\geq1 i ≥ 1 , the zero-level numerator m i m_i m i lies in L J L_J L J .
Proof. By (3.2), m i = T i , v i , i ( N ) + ∑ t = 1 i − 1 T t , v i , t ( m i − t ) m_i=T_{i,v_i,i}(N)+\sum_{t=1}^{i-1}T_{t,v_i,t}(m_{i-t}) m i = T i , v i , i ( N ) + ∑ t = 1 i − 1 T t , v i , t ( m i − t ) , and we apply Lemma 4.2 and induction on i i i . □ \square □
Remark 4.4. Only Lemma 4.2 is used about J J J below. Hence Theorem 1.2 holds with ∣ J ∣ |J| ∣ J ∣ replaced by ∣ J ′ ∣ |J'| ∣ J ′ ∣ for any finite J ′ ⊂ Z ≥ 0 m J'\subset\mathbb{Z}_{\geq0}^m J ′ ⊂ Z ≥ 0 m satisfying the conclusion of Lemma 4.2. Moreover K α ( G ) K_\alpha(G) K α ( G ) depends only on N m o d p α N\bmod p^\alpha N mod p α and Q m o d p α Q\bmod p^\alpha Q mod p α . One may therefore compute J J J from any lifts of these reductions with unit constant term; for instance, lifts with coefficients in { 0 , 1 , … , p α − 1 } \{0,1,\ldots,p^\alpha-1\} { 0 , 1 , … , p α − 1 } have supports contained in those of N N N and Q Q Q .
Example 4.5. Let Q = 1 − z 1 − ⋯ − z m Q=1-z_1-\cdots-z_m Q = 1 − z 1 − ⋯ − z m and N = 1 N=1 N = 1 . Then Δ \Delta Δ is the standard simplex and K = { 0 } K=\{0\} K = { 0 } . If u + η 1 ∈ ( 1 − ε ) Δ u+\eta\mathbf{1}\in(1-\varepsilon)\Delta u + η 1 ∈ ( 1 − ε ) Δ with u ∈ Z ≥ 0 m u\in\mathbb{Z}_{\geq0}^m u ∈ Z ≥ 0 m , then ∑ i u i + m η ≤ 1 − ε < 1 \sum_i u_i+m\eta\leq1-\varepsilon<1 ∑ i u i + m η ≤ 1 − ε < 1 , so u = 0 u=0 u = 0 . Thus J = { 0 } J=\{0\} J = { 0 } , while the box has 2 m 2^m 2 m points. Theorem 1.2 gives
K α ( ∑ n ≥ 0 ( m n ) ! n ! m x n ) ≤ p 2 α − 1 + p α − 1 p − 1 . K_\alpha\left(\sum_{n\geq0}\frac{(mn)!}{n!^m}x^n\right)\leq p^{2\alpha-1}+\frac{p^\alpha-1}{p-1}. K α ( n ≥ 0 ∑ n ! m ( mn )! x n ) ≤ p 2 α − 1 + p − 1 p α − 1 . for every m ≥ 1 m\geq1 m ≥ 1 and every prime p p p . Here all zero-level numerators are constants.
Histories, windows and the count
Fix α ≥ 1 \alpha\geq1 α ≥ 1 . For e ≥ 0 e\geq0 e ≥ 0 write ( Z / p e ) J = L J / p e L J (\mathbb{Z}/p^e)^J=L_J/p^eL_J ( Z / p e ) J = L J / p e L J for polynomials supported on J J J with coefficients modulo p e p^e p e .
Precision bookkeeping
Lemma 5.1. Let e ≥ 0 e\geq0 e ≥ 0 , t ≥ 1 t\geq1 t ≥ 1 , j ≥ 0 j\geq0 j ≥ 0 , 0 ≤ v < p t 0\leq v<p^t 0 ≤ v < p t , and let H ∈ ( Z / p e ) J H\in(\mathbb{Z}/p^e)^J H ∈ ( Z / p e ) J have lift H ~ ∈ L J \widetilde{H}\in L_J H ∈ L J . Then
C t ( z ∗ − v δ j + t H ~ ) m o d p e + j + t − 1 C^t\left(z_*^{-v}\delta_{j+t}\widetilde{H}\right)\bmod p^{e+j+t-1} C t ( z ∗ − v δ j + t H ) mod p e + j + t − 1 does not depend on the lift. In particular T t , v T_{t,v} T t , v induces a well-defined map ( Z / p e ) J → ( Z / p e + t − 1 ) J (\mathbb{Z}/p^e)^J\to(\mathbb{Z}/p^{e+t-1})^J ( Z / p e ) J → ( Z / p e + t − 1 ) J , which we denote by the same symbol.
Proof. Two lifts differ by p e Y p^eY p e Y with Y ∈ L J Y\in L_J Y ∈ L J , and C t ( z ∗ − v δ j + t p e Y ) ∈ p e + j + t − 1 Z p [ z ] C^t(z_*^{-v}\delta_{j+t}p^eY)\in p^{e+j+t-1}\mathbb{Z}_p[z] C t ( z ∗ − v δ j + t p e Y ) ∈ p e + j + t − 1 Z p [ z ] by Lemma 3.1(b). The last statement is the case j = 0 j=0 j = 0 , combined with Lemma 4.2. □ \square □
Digit values
For a digit string U = ( u 1 , … , u L ) ∈ Σ p L U=(u_1,\ldots,u_L)\in\Sigma_p^L U = ( u 1 , … , u L ) ∈ Σ p L and integers s ≥ 0 s\ge0 s ≥ 0 , t ≥ 1 t\ge1 t ≥ 1 with s + t ≤ L s+t\le L s + t ≤ L , let
⟨ U ⟩ s , t = ∑ q = 0 t − 1 u L − s − t + 1 + q p q ∈ [ 0 , p t ) . (5) \langle U\rangle_{s,t}=\sum_{q=0}^{t-1}u_{L-s-t+1+q}p^q\in[0,p^t). \tag*{(5)} ⟨ U ⟩ s , t = q = 0 ∑ t − 1 u L − s − t + 1 + q p q ∈ [ 0 , p t ) . (5)
This is the integer whose digits, least significant first, are the t t t digits of U U U that end s s s places before the end of U U U . Directly from the definition, for r ∈ Σ p r\in\Sigma_p r ∈ Σ p and U r = ( u 1 , … , u L , r ) U_r=(u_1,\ldots,u_L,r) U r = ( u 1 , … , u L , r ) : (D1) ⟨ U r ⟩ 0 , 1 = r \langle U_r\rangle_{0,1}=r ⟨ U r ⟩ 0 , 1 = r and ⟨ U r ⟩ 0 , t + 1 = ⟨ U ⟩ 0 , t + r p t \langle U_r\rangle_{0,t+1}=\langle U\rangle_{0,t}+rp^t ⟨ U r ⟩ 0 , t + 1 = ⟨ U ⟩ 0 , t + r p t for 1 ≤ t ≤ L 1\le t\le L 1 ≤ t ≤ L ; (D2) ⟨ U r ⟩ s , t = ⟨ U ⟩ s − 1 , t \langle U_r\rangle_{s,t}=\langle U\rangle_{s-1,t} ⟨ U r ⟩ s , t = ⟨ U ⟩ s − 1 , t for s ≥ 1 s\ge1 s ≥ 1 and s + t ≤ L + 1 s+t\le L+1 s + t ≤ L + 1 ; (D3) if U ′ U' U ′ is obtained from U U U by deleting initial digits, then ⟨ U ′ ⟩ s , t = ⟨ U ⟩ s , t \langle U'\rangle_{s,t}=\langle U\rangle_{s,t} ⟨ U ′ ⟩ s , t = ⟨ U ⟩ s , t whenever s + t ≤ ∣ U ′ ∣ s+t\le|U'| s + t ≤ ∣ U ′ ∣ .
For a word w = r 0 ⋯ r k − 1 w=r_0\cdots r_{k-1} w = r 0 ⋯ r k − 1 , regarded as the string ( r 0 , … , r k − 1 ) (r_0,\ldots,r_{k-1}) ( r 0 , … , r k − 1 ) , we have
v i , t ( w ) = ⟨ w ⟩ k − i , t ( 1 ≤ t ≤ i ≤ k ) . (6) v_{i,t}(w)=\langle w\rangle_{k-i,t}\qquad(1\le t\le i\le k). \tag*{(6)} v i , t ( w ) = ⟨ w ⟩ k − i , t ( 1 ≤ t ≤ i ≤ k ) . (6)
Histories and windows
Definition 5.2. Let w w w be a word of length k ≥ α k\ge\alpha k ≥ α . Its window W ( w ) ∈ Σ p α − 1 W(w)\in\Sigma_p^{\alpha-1} W ( w ) ∈ Σ p α − 1 is the string of its last α − 1 \alpha-1 α − 1 digits. Its history is
H ( w ) = ( H 0 , … , H α − 1 ) , H s = m k − s m o d p α − s ∈ ( Z / p α − s ) J . H(w)=(H_0,\ldots,H_{\alpha-1}),\qquad H_s=m_{k-s}\bmod p^{\alpha-s}\in(\mathbb{Z}/p^{\alpha-s})^J. H ( w ) = ( H 0 , … , H α − 1 ) , H s = m k − s mod p α − s ∈ ( Z / p α − s ) J .
This makes sense by Corollary 4.3, since k − s ≥ k − α + 1 ≥ 1 k-s\ge k-\alpha+1\ge1 k − s ≥ k − α + 1 ≥ 1 .
By (6) and (D3),
v k − s , t ( w ) = ⟨ w ⟩ s , t = ⟨ W ( w ) ⟩ s , t whenever s + t ≤ α − 1. (7) v_{k-s,t}(w)=\langle w\rangle_{s,t}=\langle W(w)\rangle_{s,t}\qquad\text{whenever }s+t\le\alpha-1. \tag*{(7)} v k − s , t ( w ) = ⟨ w ⟩ s , t = ⟨ W ( w ) ⟩ s , t whenever s + t ≤ α − 1. (7)
Definition 5.3 (decoding). For W ∈ Σ p α − 1 W\in\Sigma_p^{\alpha-1} W ∈ Σ p α − 1 and H = ( H 0 , … , H α − 1 ) ∈ ∏ s = 0 α − 1 ( Z / p α − s ) J H=(H_0,\ldots,H_{\alpha-1})\in\prod_{s=0}^{\alpha-1}(\mathbb{Z}/p^{\alpha-s})^J H = ( H 0 , … , H α − 1 ) ∈ ∏ s = 0 α − 1 ( Z / p α − s ) J define Ψ ( H , W ) ∈ ( Z / p α ) [ [ z ] ] \Psi(H,W)\in(\mathbb{Z}/p^\alpha)[[z]] Ψ ( H , W ) ∈ ( Z / p α ) [[ z ]] by
Ψ ( H , W ) = H ~ 0 Q − 1 + ∑ j = 1 α − 1 ∑ t = 1 α − j C ( z ∗ − ⟨ W ⟩ 0 , t δ j + t H ~ t ) t σ j ( Q − 1 ) ( m o d p α ) . (8) \Psi(H,W)=\widetilde H_0Q^{-1}+\sum_{j=1}^{\alpha-1}\sum_{t=1}^{\alpha-j}C^t_{\left(z_*^{-\langle W\rangle_{0,t}}\delta_{j+t}\widetilde H_t\right)}\sigma^j(Q^{-1})\pmod{p^\alpha}. \tag*{(8)} Ψ ( H , W ) = H 0 Q − 1 + j = 1 ∑ α − 1 t = 1 ∑ α − j C ( z ∗ − ⟨ W ⟩ 0 , t δ j + t H t ) t σ j ( Q − 1 ) ( mod p α ) . (8)
where H ~ s ∈ L J \widetilde H_s\in L_J H s ∈ L J are lifts. The value does not depend on the lifts: the first term only involves H 0 H_0 H 0 modulo p α p^\alpha p α , and by Lemma 5.1 with e = α − t e=\alpha-t e = α − t and j ≥ 1 j\ge1 j ≥ 1 , the summand indexed by ( j , t ) (j,t) ( j , t ) is determined modulo p α + j − 1 p^{\alpha+j-1} p α + j − 1 , hence modulo p α p^\alpha p α . Note that t ≤ α − j ≤ α − 1 = ∣ W ∣ t\le\alpha-j\le\alpha-1=|W| t ≤ α − j ≤ α − 1 = ∣ W ∣ .
Lemma 5.4 (decoding). For every word w w w with ∣ w ∣ ≥ α |w|\ge\alpha ∣ w ∣ ≥ α ,
C w ( R ) = Ψ ( H ( w ) , W ( w ) ) ( m o d p α ) . C_w(R)=\Psi(H(w),W(w))\pmod{p^\alpha}. C w ( R ) = Ψ ( H ( w ) , W ( w )) ( mod p α ) .
Proof. Let k = ∣ w ∣ k=|w| k = ∣ w ∣ and use the level expansion of Proposition 3.6 . The level-0 numerator is m k m_k m k , which reduces to H 0 H_0 H 0 . For j ≥ α j\ge\alpha j ≥ α , M j ( k ) ∈ p j Z p [ z ] ⊂ p α Z p [ z ] M_j^{(k)}\in p^j\mathbb{Z}_p[z]\subset p^\alpha\mathbb{Z}_p[z] M j ( k ) ∈ p j Z p [ z ] ⊂ p α Z p [ z ] . For 1 ≤ j ≤ α − 1 1\le j\le\alpha-1 1 ≤ j ≤ α − 1 consider the summand of M j ( k ) M_j^{(k)} M j ( k ) indexed by t t t . If t > α − j t>\alpha-j t > α − j , it lies in p j + t − 1 Z p [ z ] ⊂ p α Z p [ z ] p^{j+t-1}\mathbb{Z}_p[z]\subset p^\alpha\mathbb{Z}_p[z] p j + t − 1 Z p [ z ] ⊂ p α Z p [ z ] . If t ≤ α − j t\le\alpha-j t ≤ α − j , then t ≤ α − 1 ≤ k − 1 t\le\alpha-1\le k-1 t ≤ α − 1 ≤ k − 1 , and v k , t = ⟨ W ( w ) ⟩ 0 , t v_{k,t}=\langle W(w)\rangle_{0,t} v k , t = ⟨ W ( w ) ⟩ 0 , t by (5.3) with s = 0 s=0 s = 0 . Moreover m k − t m_{k-t} m k − t is a lift of H t H_t H t , which is given modulo p α − t p^{\alpha-t} p α − t , so the summand agrees modulo p α + j − 1 p^{\alpha+j-1} p α + j − 1 with the corresponding summand of (5.4) by Lemma 5.1. □ \square □
Compatible histories
Definition 5.5. Let W ∈ Σ p α − 1 W\in\Sigma_p^{\alpha-1} W ∈ Σ p α − 1 . A tuple H ∈ ∏ s = 0 α − 1 ( Z / p α − s ) J H\in\prod_{s=0}^{\alpha-1}(\mathbb{Z}/p^{\alpha-s})^J H ∈ ∏ s = 0 α − 1 ( Z / p α − s ) J is compatible with W W W if, for 0 ≤ s ≤ α − 2 0\le s\le\alpha-2 0 ≤ s ≤ α − 2 ,
H s ≡ ∑ t = 1 α − s − 1 T t , ⟨ W ⟩ s , t ( H s + t ) ( m o d p α − s − 1 ) . (9) H_s\equiv\sum_{t=1}^{\alpha-s-1}T_{t,\langle W\rangle_{s,t}}(H_{s+t})\pmod{p^{\alpha-s-1}}. \tag*{(9)} H s ≡ t = 1 ∑ α − s − 1 T t , ⟨ W ⟩ s , t ( H s + t ) ( mod p α − s − 1 ) . (9)
The right-hand side is well defined in ( Z / p α − s − 1 ) J (\mathbb{Z}/p^{\alpha-s-1})^{J} ( Z / p α − s − 1 ) J by Lemma 5.1, because H s + t H_{s+t} H s + t is given modulo p α − s − t p^{\alpha-s-t} p α − s − t . Let H W \mathcal{H}_{W} H W be the set of histories compatible with W W W . Since the maps T t , v T_{t,v} T t , v are additive, H W \mathcal{H}_{W} H W is a subgroup of ∏ s ( Z / p α − s ) J \prod_{s}(\mathbb{Z}/p^{\alpha-s})^{J} ∏ s ( Z / p α − s ) J .
Lemma 5.6. For every word w w w with ∣ w ∣ ≥ α |w| \ge\alpha ∣ w ∣ ≥ α we have H ( w ) ∈ H W ( w ) H(w) \in\mathcal{H}_{W(w)} H ( w ) ∈ H W ( w ) .
Proof. Let k = ∣ w ∣ k=|w| k = ∣ w ∣ and 0 ≤ s ≤ α − 2 0 \le s \le\alpha-2 0 ≤ s ≤ α − 2 . By (4) , m k − s = ∑ t = 1 k − s T t , v k − s , t ( m k − s − t ) m_{k-s}=\sum_{t=1}^{k-s}T_{t,v_{k-s,t}}(m_{k-s-t}) m k − s = ∑ t = 1 k − s T t , v k − s , t ( m k − s − t ) . The summands with t ≥ α − s t \ge\alpha-s t ≥ α − s lie in p t − 1 Z p [ z ] ⊂ p α − s − 1 Z p [ z ] p^{t-1}\mathbb{Z}_{p}[z]\subset p^{\alpha-s-1}\mathbb{Z}_{p}[z] p t − 1 Z p [ z ] ⊂ p α − s − 1 Z p [ z ] . For t ≤ α − s − 1 t \le\alpha-s-1 t ≤ α − s − 1 (note t < k − s t<k-s t < k − s ) we have v k − s , t = ⟨ W ( w ) ⟩ s , t v_{k-s,t}=\langle W(w)\rangle_{s,t} v k − s , t = ⟨ W ( w ) ⟩ s , t by (7) . By Lemma 5.1, T t , v ( m k − s − t ) m o d p α − s − 1 T_{t,v}(m_{k-s-t})\bmod p^{\alpha-s-1} T t , v ( m k − s − t ) mod p α − s − 1 depends only on m k − s − t m o d p α − s − t m_{k-s-t}\bmod p^{\alpha-s-t} m k − s − t mod p α − s − t , which is H s + t H_{s+t} H s + t . Reducing modulo p α − s − 1 p^{\alpha-s-1} p α − s − 1 gives (9) . □
Lemma 5.7 (exact count). For every W ∈ Σ p α − 1 W\in\Sigma_{p}^{\alpha-1} W ∈ Σ p α − 1 we have ∣ H W ∣ = p ∣ J ∣ α |\mathcal{H}_{W}|=p^{|J|\alpha} ∣ H W ∣ = p ∣ J ∣ α . More precisely, the map sending H ∈ H W H\in\mathcal{H}_{W} H ∈ H W to the top p p p -adic digits of its ∣ J ∣ α |J|\alpha ∣ J ∣ α coefficients is a bijection onto ∏ s = 0 α − 1 ( Z / p ) J \prod_{s=0}^{\alpha-1}(\mathbb{Z}/p)^{J} ∏ s = 0 α − 1 ( Z / p ) J .
Proof. For 0 ≤ s ≤ α − 1 0\le s\le\alpha-1 0 ≤ s ≤ α − 1 let H W ( s ) \mathcal{H}_{W}^{(s)} H W ( s ) be the set of tuples ( H s , … , H α − 1 ) (H_s,\ldots,H_{\alpha-1}) ( H s , … , H α − 1 ) satisfying (9) for all s ′ s' s ′ with s ≤ s ′ ≤ α − 2 s\le s'\le\alpha-2 s ≤ s ′ ≤ α − 2 . Then H W ( α − 1 ) = ( Z / p ) J \mathcal{H}_{W}^{(\alpha-1)}=(\mathbb{Z}/p)^{J} H W ( α − 1 ) = ( Z / p ) J has p ∣ J ∣ p^{|J|} p ∣ J ∣ elements. Let s ≤ α − 2 s\le\alpha-2 s ≤ α − 2 . A tuple in H W ( s ) \mathcal{H}_{W}^{(s)} H W ( s ) consists of a tuple in H W ( s + 1 ) \mathcal{H}_{W}^{(s+1)} H W ( s + 1 ) together with an element H s ∈ ( Z / p α − s ) J H_s\in(\mathbb{Z}/p^{\alpha-s})^{J} H s ∈ ( Z / p α − s ) J whose reduction modulo p α − s − 1 p^{\alpha-s-1} p α − s − 1 is prescribed by (9) . The reduction map ( Z / p α − s ) J → ( Z / p α − s − 1 ) J (\mathbb{Z}/p^{\alpha-s})^{J}\to(\mathbb{Z}/p^{\alpha-s-1})^{J} ( Z / p α − s ) J → ( Z / p α − s − 1 ) J is surjective with fibres of size p ∣ J ∣ p^{|J|} p ∣ J ∣ , and an element of a fibre is determined by the top digits ⌊ H ‾ s / p α − s − 1 ⌋ \lfloor\overline{H}_s/p^{\alpha-s-1}\rfloor ⌊ H s / p α − s − 1 ⌋ of its coefficients, where H ‾ s \overline{H}_s H s has coefficients in [ 0 , p α − s ) [0,p^{\alpha-s}) [ 0 , p α − s ) . Hence ∣ H W ( s ) ∣ = p ∣ J ∣ ∣ H W ( s + 1 ) ∣ |\mathcal{H}_{W}^{(s)}|=p^{|J|}|\mathcal{H}_{W}^{(s+1)}| ∣ H W ( s ) ∣ = p ∣ J ∣ ∣ H W ( s + 1 ) ∣ , and the top digits can be chosen freely at each level. Therefore ∣ H W ∣ = ∣ H W ( 0 ) ∣ = p ∣ J ∣ α |\mathcal{H}_{W}|=|\mathcal{H}_{W}^{(0)}|=p^{|J|\alpha} ∣ H W ∣ = ∣ H W ( 0 ) ∣ = p ∣ J ∣ α . □
Proof of Theorem 1.2
By Lemma 2.2, K α ( G ) ≤ ∣ { C w R m o d p α : w ∈ Σ p ∗ } ∣ K_{\alpha}(G)\le\left|\{C_{w}R\bmod p^{\alpha}:w\in\Sigma_{p}^{*}\}\right| K α ( G ) ≤ { C w R mod p α : w ∈ Σ p ∗ } . Words of length less than α \alpha α contribute at most ∑ k = 0 α − 1 p k = ( p α − 1 ) / ( p − 1 ) \sum_{k=0}^{\alpha-1}p^{k}=(p^{\alpha}-1)/(p-1) ∑ k = 0 α − 1 p k = ( p α − 1 ) / ( p − 1 ) elements. For ∣ w ∣ ≥ α |w|\ge\alpha ∣ w ∣ ≥ α , Lemmas 5.4 and 5.6 give C w R m o d p α = Ψ ( H ( w ) , W ( w ) ) C_{w}R\bmod p^{\alpha}=\Psi(H(w),W(w)) C w R mod p α = Ψ ( H ( w ) , W ( w )) with H ( w ) ∈ H W ( w ) H(w)\in\mathcal{H}_{W(w)} H ( w ) ∈ H W ( w ) . By Lemma 5.7 there are at most
∑ W ∈ Σ p α − 1 ∣ H W ∣ = p α − 1 ⋅ p ∣ J ∣ α = p ( ∣ J ∣ + 1 ) α − 1 \sum_{W\in\Sigma_{p}^{\alpha-1}}|\mathcal{H}_{W}|=p^{\alpha-1}\cdot p^{|J|\alpha}=p^{(|J|+1)\alpha-1} W ∈ Σ p α − 1 ∑ ∣ H W ∣ = p α − 1 ⋅ p ∣ J ∣ α = p ( ∣ J ∣ + 1 ) α − 1
such values. Finally ∣ J ∣ ≤ D |J|\le D ∣ J ∣ ≤ D by Lemma 4.1. □
The explicit automaton
Proposition 5.8 (transitions). Let W ∈ Σ p α − 1 W\in\Sigma_{p}^{\alpha-1} W ∈ Σ p α − 1 , H ∈ ∏ s = 0 α − 1 ( Z / p α − s ) J H\in\prod_{s=0}^{\alpha-1}(\mathbb{Z}/p^{\alpha-s})^{J} H ∈ ∏ s = 0 α − 1 ( Z / p α − s ) J and r ∈ Σ p r\in\Sigma_{p} r ∈ Σ p . Let W ′ W' W ′ be the string of the last α − 1 \alpha-1 α − 1 digits of W r Wr W r , and define H ′ = ( H 0 ′ , … , H α − 1 ′ ) H'=(H'_0,\ldots,H'_{\alpha-1}) H ′ = ( H 0 ′ , … , H α − 1 ′ ) by
H 0 ′ = ∑ t = 1 α T t , ⟨ W r ⟩ 0 , t ( H t − 1 ) ∈ ( Z / p α ) J , H s ′ = H s − 1 m o d p α − s ( 1 ≤ s ≤ α − 1 ) . (10) H'_0=\sum_{t=1}^{\alpha}T_{t,\langle Wr\rangle_{0,t}}(H_{t-1})\in(\mathbb{Z}/p^{\alpha})^{J},\qquad H'_s=H_{s-1}\bmod p^{\alpha-s}\quad(1\le s\le\alpha-1). \tag*{(10)} H 0 ′ = t = 1 ∑ α T t , ⟨ W r ⟩ 0 , t ( H t − 1 ) ∈ ( Z / p α ) J , H s ′ = H s − 1 mod p α − s ( 1 ≤ s ≤ α − 1 ) . (10) Write τ r ( H , W ) = ( H ′ , W ′ ) \tau_r(H,W)=(H',W') τ r ( H , W ) = ( H ′ , W ′ ) . Then: (a) C r Ψ ( H , W ) = Ψ ( τ r ( H , W ) ) C_r\Psi(H,W)=\Psi(\tau_r(H,W)) C r Ψ ( H , W ) = Ψ ( τ r ( H , W )) in ( Z / p α ) [ [ z ] ] (\mathbb{Z}/p^{\alpha})[[z]] ( Z / p α ) [[ z ]] , for every H H H , compatible or not; (b) if H ∈ H W H\in\mathcal{H}_{W} H ∈ H W , then H ′ ∈ H W ′ H'\in\mathcal{H}_{W'} H ′ ∈ H W ′ ; (c) for every word w w w with ∣ w ∣ ≥ α |w|\ge\alpha ∣ w ∣ ≥ α , τ r ( H ( w ) , W ( w ) ) = ( H ( w r ) , W ( w r ) ) \tau_r(H(w),W(w))=(H(wr),W(wr)) τ r ( H ( w ) , W ( w )) = ( H ( w r ) , W ( w r )) .
The summand of H 0 ′ H'_0 H 0 ′ indexed by t t t is well defined in ( Z / p α ) J (\mathbb{Z}/p^{\alpha})^{J} ( Z / p α ) J by Lemma 5.1, since H t − 1 H_{t-1} H t − 1 is given modulo p α − t + 1 p^{\alpha-t+1} p α − t + 1 .
Proof. (a) Choose lifts H ~ s ∈ L J \widetilde{H}_s\in L_J H s ∈ L J and write Ψ ( H , W ) \Psi(H,W) Ψ ( H , W ) as the reduction of ∑ j ≥ 0 M j σ j ( Q − 1 ) \sum_{j\ge0}M_j\sigma^j(Q^{-1}) ∑ j ≥ 0 M j σ j ( Q − 1 ) , where M 0 = H ~ 0 M_0=\widetilde{H}_0 M 0 = H 0 , M j = ∑ t = 1 α − j C t ( z ∗ − ⟨ W ⟩ 0 , t δ j + t H ~ t ) M_j=\sum_{t=1}^{\alpha-j}C^t(z_*\!-\!\langle W\rangle_{0,t}\delta_{j+t}\widetilde{H}_t) M j = ∑ t = 1 α − j C t ( z ∗ − ⟨ W ⟩ 0 , t δ j + t H t ) for 1 ≤ j ≤ α − 1 1\le j\le\alpha-1 1 ≤ j ≤ α − 1 , and M j = 0 M_j=0 M j = 0 for j ≥ α j\ge\alpha j ≥ α . By Lemma 3.4, C r C_r C r of this series has the level expansion N j = C r ( M j + 1 + δ j + 1 H ~ 0 ) N_j=C_r(M_{j+1}+\delta_{j+1}\widetilde{H}_0) N j = C r ( M j + 1 + δ j + 1 H 0 ) . For j ≥ α j\ge\alpha j ≥ α , N j = C r ( δ j + 1 H ~ 0 ) ∈ p j Z p [ z ] N_j=C_r(\delta_{j+1}\widetilde{H}_0)\in p^j\mathbb{Z}_p[z] N j = C r ( δ j + 1 H 0 ) ∈ p j Z p [ z ] vanishes modulo p α p^\alpha p α . For 0 ≤ j ≤ α − 1 0\le j\le\alpha-1 0 ≤ j ≤ α − 1 we compute, using ⟨ W r ⟩ 0 , 1 = r \langle W r\rangle_{0,1}=r ⟨ W r ⟩ 0 , 1 = r , Lemma 2.1(b) and (D1),
N j = C 1 ( z ∗ − ⟨ W r ⟩ 0 , 1 δ j + 1 H ~ 0 ) + ∑ t = 1 α − j − 1 C t + 1 ( z ∗ − ⟨ W r ⟩ 0 , t + 1 δ j + 1 + t H ~ t ) = ∑ t = 1 α − j C t ( z ∗ − ⟨ W r ⟩ 0 , t δ j + t H ~ t − 1 ) . N_j=C^1\left(z_*^{-\langle W r\rangle_{0,1}}\delta_{j+1}\widetilde{H}_0\right)+\sum_{t=1}^{\alpha-j-1}C^{t+1}\left(z_*^{-\langle W r\rangle_{0,t+1}}\delta_{j+1+t}\widetilde{H}_t\right)=\sum_{t=1}^{\alpha-j}C^t\left(z_*^{-\langle W r\rangle_{0,t}}\delta_{j+t}\widetilde{H}_{t-1}\right). N j = C 1 ( z ∗ − ⟨ W r ⟩ 0 , 1 δ j + 1 H 0 ) + t = 1 ∑ α − j − 1 C t + 1 ( z ∗ − ⟨ W r ⟩ 0 , t + 1 δ j + 1 + t H t ) = t = 1 ∑ α − j C t ( z ∗ − ⟨ W r ⟩ 0 , t δ j + t H t − 1 ) .
For j = 0 j=0 j = 0 this is ∑ t = 1 α T t , ⟨ W r ⟩ 0 , t ( H ~ t − 1 ) \sum_{t=1}^{\alpha}T_{t,\langle W r\rangle_{0,t}}(\widetilde{H}_{t-1}) ∑ t = 1 α T t , ⟨ W r ⟩ 0 , t ( H t − 1 ) , a lift of H 0 ′ H'_0 H 0 ′ . Now let 1 ≤ j ≤ α − 1 1\le j\le\alpha-1 1 ≤ j ≤ α − 1 . The level-j j j numerator of Ψ ( H ′ , W ′ ) \Psi(H',W') Ψ ( H ′ , W ′ ) is ∑ t = 1 α − j C t ( z ∗ − ⟨ W ′ ⟩ 0 , t δ j + t H ~ t ′ ) \sum_{t=1}^{\alpha-j}C^t(z_*^{-\langle W'\rangle_{0,t}}\delta_{j+t}\widetilde{H}'_t) ∑ t = 1 α − j C t ( z ∗ − ⟨ W ′ ⟩ 0 , t δ j + t H t ′ ) . Here ⟨ W ′ ⟩ 0 , t = ⟨ W r ⟩ 0 , t \langle W'\rangle_{0,t}=\langle W r\rangle_{0,t} ⟨ W ′ ⟩ 0 , t = ⟨ W r ⟩ 0 , t by (D3), as t ≤ α − 1 t\le\alpha-1 t ≤ α − 1 . Also H ~ t ′ ≡ H ~ t − 1 ( m o d p α − t ) \widetilde{H}'_t\equiv\widetilde{H}_{t-1}\pmod{p^{\alpha-t}} H t ′ ≡ H t − 1 ( mod p α − t ) , so by Lemma 5.1 the two summands agree modulo p α + j − 1 p^{\alpha+j-1} p α + j − 1 . Hence N j ≡ M j ′ ( m o d p α ) N_j\equiv M'_j\pmod{p^\alpha} N j ≡ M j ′ ( mod p α ) for all j j j , and C r Ψ ( H , W ) = Ψ ( H ′ , W ′ ) C_r\Psi(H,W)=\Psi(H',W') C r Ψ ( H , W ) = Ψ ( H ′ , W ′ ) .
(b) First let s = 0 s=0 s = 0 . Reduce (5.6) modulo p α − 1 p^{\alpha-1} p α − 1 . The term t = α t=\alpha t = α lies in p α − 1 p^{\alpha-1} p α − 1 and disappears. For t ≤ α − 1 t\le\alpha-1 t ≤ α − 1 , the term T t , ⟨ W r ⟩ 0 , t ( H t − 1 ) m o d p α − 1 T_{t,\langle W r\rangle_{0,t}}(H_{t-1})\bmod p^{\alpha-1} T t , ⟨ W r ⟩ 0 , t ( H t − 1 ) mod p α − 1 depends only on H t − 1 m o d p α − t = H t ′ H_{t-1}\bmod p^{\alpha-t}=H'_t H t − 1 mod p α − t = H t ′ by Lemma 5.1, and ⟨ W r ⟩ 0 , t = ⟨ W ′ ⟩ 0 , t \langle W r\rangle_{0,t}=\langle W'\rangle_{0,t} ⟨ W r ⟩ 0 , t = ⟨ W ′ ⟩ 0 , t by (D3). This gives (5.5) for ( H ′ , W ′ ) (H',W') ( H ′ , W ′ ) at s = 0 s=0 s = 0 .
Next let 1 ≤ s ≤ α − 2 1\le s\le\alpha-2 1 ≤ s ≤ α − 2 . Relation (5.5) for ( H , W ) (H,W) ( H , W ) at index s − 1 s-1 s − 1 reads
H s − 1 ≡ ∑ t = 1 α − s T t , ⟨ W ⟩ s − 1 , t ( H s − 1 + t ) ( m o d p α − s ) . H_{s-1}\equiv\sum_{t=1}^{\alpha-s}T_{t,\langle W\rangle_{s-1,t}}(H_{s-1+t})\pmod{p^{\alpha-s}}. H s − 1 ≡ t = 1 ∑ α − s T t , ⟨ W ⟩ s − 1 , t ( H s − 1 + t ) ( mod p α − s ) .
Reduce it modulo p α − s − 1 p^{\alpha-s-1} p α − s − 1 . The term t = α − s t=\alpha-s t = α − s disappears. For t ≤ α − s − 1 t\le\alpha-s-1 t ≤ α − s − 1 , the term depends only on H s + t − 1 m o d p α − s − t = H s + t ′ H_{s+t-1}\bmod p^{\alpha-s-t}=H'_{s+t} H s + t − 1 mod p α − s − t = H s + t ′ . Moreover ⟨ W ⟩ s − 1 , t = ⟨ W r ⟩ s , t = ⟨ W ′ ⟩ s , t \langle W\rangle_{s-1,t}=\langle W r\rangle_{s,t}=\langle W'\rangle_{s,t} ⟨ W ⟩ s − 1 , t = ⟨ W r ⟩ s , t = ⟨ W ′ ⟩ s , t by (D2) and (D3). Since H s ′ ≡ H s − 1 ( m o d p α − s ) H'_s\equiv H_{s-1}\pmod{p^{\alpha-s}} H s ′ ≡ H s − 1 ( mod p α − s ) , we obtain (5.5) for ( H ′ , W ′ ) (H',W') ( H ′ , W ′ ) at index s s s .
(c) Let k = ∣ w ∣ k=|w| k = ∣ w ∣ . By (3.2), m k + 1 = ∑ t = 1 k + 1 T t , v k + 1 , t ( w r ) ( m k + 1 − t ) m_{k+1}=\sum_{t=1}^{k+1}T_{t,v_{k+1,t}(wr)}(m_{k+1-t}) m k + 1 = ∑ t = 1 k + 1 T t , v k + 1 , t ( w r ) ( m k + 1 − t ) . Modulo p α p^\alpha p α the terms with t ≥ α + 1 t\ge\alpha+1 t ≥ α + 1 vanish. For t ≤ α t\le\alpha t ≤ α , v k + 1 , t ( w r ) = ⟨ w r ⟩ 0 , t = ⟨ W r ⟩ 0 , t v_{k+1,t}(wr)=\langle wr\rangle_{0,t}=\langle W r\rangle_{0,t} v k + 1 , t ( w r ) = ⟨ w r ⟩ 0 , t = ⟨ W r ⟩ 0 , t by (5.2) and (D3), because W r W r W r is the suffix of w r wr w r of length α \alpha α . The term depends only on m k + 1 − t m o d p α − t + 1 = H t − 1 m_{k+1-t}\bmod p^{\alpha-t+1}=H_{t-1} m k + 1 − t mod p α − t + 1 = H t − 1 . Hence H ( w r ) 0 = H 0 ′ H(wr)_0=H'_0 H ( w r ) 0 = H 0 ′ . For s ≥ 1 s\ge1 s ≥ 1 , H ( w r ) s = m k + 1 − s m o d p α − s = H s − 1 m o d p α − s H(wr)_s=m_{k+1-s}\bmod p^{\alpha-s}=H_{s-1}\bmod p^{\alpha-s} H ( w r ) s = m k + 1 − s mod p α − s = H s − 1 mod p α − s . Finally W ( w r ) = W ′ W(wr)=W' W ( w r ) = W ′ . □
Corollary 5.9 (explicit automaton). Consider the automaton with input alphabet Σ p \Sigma_p Σ p whose states are the words of length less than α \alpha α together with the pairs ( H , W ) (H,W) ( H , W ) with W ∈ Σ p α W\in\Sigma_p^\alpha W ∈ Σ p α and H ∈ H W H\in H_W H ∈ H W . Its initial state is the empty word. Its transitions are w → r w r w\xrightarrow{r}wr w r w r for ∣ w ∣ < α − 1 |w|<\alpha-1 ∣ w ∣ < α − 1 , w → r ( H ( w r ) , W ( w r ) ) w\xrightarrow{r}(H(wr),W(wr)) w r ( H ( w r ) , W ( w r )) for ∣ w ∣ = α − 1 |w|=\alpha-1 ∣ w ∣ = α − 1 , and ( H , W ) → r τ r ( H , W ) (H,W)\xrightarrow{r}\tau_r(H,W) ( H , W ) r τ r ( H , W ) . Its outputs are g ( v ( w ) ) m o d p α g(v(w))\bmod p^\alpha g ( v ( w )) mod p α at a word w w w , and
ω ( H , W ) = [ z 0 ] Ψ ( H , W ) = H 0 ( 0 ) + ∑ j = 1 α − 1 ∑ t = 1 α − j [ z ⟨ W ⟩ 0 , t ] [ δ j + t H ~ t ] ( m o d p α ) \omega(H,W)=[z^0]\Psi(H,W)=H_0(0)+\sum_{j=1}^{\alpha-1}\sum_{t=1}^{\alpha-j}[z^{\langle W\rangle_{0,t}}][\delta_{j+t}\widetilde{H}_t]\pmod{p^\alpha} ω ( H , W ) = [ z 0 ] Ψ ( H , W ) = H 0 ( 0 ) + j = 1 ∑ α − 1 t = 1 ∑ α − j [ z ⟨ W ⟩ 0 , t ] [ δ j + t H t ] ( mod p α ) at ( H , W ) (H,W) ( H , W ) . This automaton generates ( g ( n ) m o d p α ) n ≥ 0 (g(n)\bmod p^\alpha)_{n\ge0} ( g ( n ) mod p α ) n ≥ 0 , reading digits least significant first, and it has at most p ( ∣ J ∣ + 1 ) α − 1 + ( p α − 1 ) / ( p − 1 ) p^{(|J|+1)\alpha-1}+(p^\alpha-1)/(p-1) p ( ∣ J ∣ + 1 ) α − 1 + ( p α − 1 ) / ( p − 1 ) states.
Proof. By Proposition 5.8(b),(c) the transitions are well defined and the state reached after reading w w w with ∣ w ∣ ≥ α |w|\ge\alpha ∣ w ∣ ≥ α is ( H ( w ) , W ( w ) ) (H(w),W(w)) ( H ( w ) , W ( w )) . Its output is [ z 0 ] Ψ ( H ( w ) , W ( w ) ) = [ z 0 ] C w R = [ x 0 ] Λ w G = g ( v ( w ) ) [z^0]\Psi(H(w),W(w))=[z^0]C_wR=[x^0]\Lambda_wG=g(v(w)) [ z 0 ] Ψ ( H ( w ) , W ( w )) = [ z 0 ] C w R = [ x 0 ] Λ w G = g ( v ( w )) modulo p α p^\alpha p α , by Lemmas 5.4 and 2.2 and (2.1). The displayed formula for ω \omega ω holds because [ z 0 ] σ j ( Q − 1 ) = 1 [z^0]\sigma^j(Q^{-1})=1 [ z 0 ] σ j ( Q − 1 ) = 1 and [ z 0 ] C t ( z ∗ − v Y ) = [ z v ] Y [z^0]C^t(z_*^{-v}Y)=[z^{v}]Y [ z 0 ] C t ( z ∗ − v Y ) = [ z v ] Y . The bound on the number of states is Lemma 5.7. □
Remark 5.10. (i) The decoding map Ψ \Psi Ψ need not be injective, and different pairs may represent the same kernel element; the minimal automaton is obtained from this one by Moore minimisation.
(ii) The transitions are given by the reductions modulo p α p^\alpha p α of the ∣ J ∣ × ∣ J ∣ |J|\times|J| ∣ J ∣ × ∣ J ∣ matrices of T t , v T_{t,v} T t , v on L J L_J L J for 1 ≤ t ≤ α 1\le t\le\alpha 1 ≤ t ≤ α and 0 ≤ v < p t 0\le v<p^t 0 ≤ v < p t ; there are about p α + 1 / ( p − 1 ) p^{\alpha+1}/(p-1) p α + 1 / ( p − 1 ) of them. Computing them requires the polynomials δ t \delta_t δ t , whose degrees grow like p t p^t p t . Afterwards each state is stored as α \alpha α vectors of length ∣ J ∣ |J| ∣ J ∣ together with α − 1 \alpha-1 α − 1 digits, and each transition costs O ( α ∣ J ∣ 2 ) O(\alpha|J|^2) O ( α ∣ J ∣ 2 ) operations modulo p α p^\alpha p α (Section 9).
(iii) For α = 1 \alpha=1 α = 1 the state is a single H 0 ∈ ( Z / p ) J H_0\in(\mathbb{Z}/p)^J H 0 ∈ ( Z / p ) J with transition H 0 ↦ T 1 , r ( H 0 ) = C ( z ∗ − r Q p − 1 H 0 ) H_0\mapsto T_{1,r}(H_0)=C(z_*^{-r}Q^{p-1}H_0) H 0 ↦ T 1 , r ( H 0 ) = C ( z ∗ − r Q p − 1 H 0 ) .
This is the classical construction modulo p p p ; Theorem 1.2 extends it to higher precision at the price of only a linear growth of the exponent.
Furstenberg series and a quadratic family
Proof of Corollary 1.3
Write P = ∑ i , j c i j x i y j P=\sum_{i,j}c_{ij}x^iy^j P = ∑ i , j c ij x i y j with c 00 = 0 c_{00}=0 c 00 = 0 and c 01 = ∂ y P ( 0 , 0 ) ∈ Z p × c_{01}=\partial_yP(0,0)\in\mathbb{Z}_p^\times c 01 = ∂ y P ( 0 , 0 ) ∈ Z p × . Then
P ( x y , y ) = ∑ c i j x i y i + j , y ∂ y P ( x y , y ) = ∑ j c i j x i y i + j . P(xy,y)=\sum c_{ij}x^iy^{i+j},\qquad y\partial_yP(xy,y)=\sum j\,c_{ij}x^iy^{i+j}. P ( x y , y ) = ∑ c ij x i y i + j , y ∂ y P ( x y , y ) = ∑ j c ij x i y i + j .
Every monomial with c i j ≠ 0 c_{ij}\ne0 c ij = 0 has i + j ≥ 1 i+j\ge1 i + j ≥ 1 , so Q F = P ( x y , y ) / y = ∑ c i j x i y i + j − 1 Q_F=P(xy,y)/y=\sum c_{ij}x^iy^{i+j-1} Q F = P ( x y , y ) / y = ∑ c ij x i y i + j − 1 and N F = y ∂ y P ( x y , y ) N_F=y\partial_yP(xy,y) N F = y ∂ y P ( x y , y ) are polynomials. The only monomial of Q F Q_F Q F with exponent ( 0 , 0 ) (0,0) ( 0 , 0 ) comes from ( i , j ) = ( 0 , 1 ) (i,j)=(0,1) ( i , j ) = ( 0 , 1 ) , so Q F ( 0 , 0 ) = c 01 Q_F(0,0)=c_{01} Q F ( 0 , 0 ) = c 01 is a unit. By [11 Lemma 6.3] (see [19 Theorem 4] ), F = Diag ( N F / Q F ) F=\operatorname{Diag}(N_F/Q_F) F = Diag ( N F / Q F ) . We have deg x Q F , deg x N F ≤ h \deg_x Q_F,\deg_x N_F\le h deg x Q F , deg x N F ≤ h , deg y Q F ≤ h + d − 1 \deg_y Q_F\le h+d-1 deg y Q F ≤ h + d − 1 and deg y N F ≤ h + d \deg_y N_F\le h+d deg y N F ≤ h + d . Hence the box of Theorem 1.2 has D ≤ ( h + 1 ) ( h + d + 1 ) D\le(h+1)(h+d+1) D ≤ ( h + 1 ) ( h + d + 1 ) points. □
The quadratic family
Let a , b , c , d ∈ Z 2 a,b,c,d\in\mathbb{Z}_2 a , b , c , d ∈ Z 2 and P = b x 2 y 2 + ( 1 + a x + d x 2 ) y + c x 2 P=bx^2y^2+(1+ax+dx^2)y+cx^2 P = b x 2 y 2 + ( 1 + a x + d x 2 ) y + c x 2 , so that ∂ y P ( 0 , 0 ) = 1 \partial_yP(0,0)=1 ∂ y P ( 0 , 0 ) = 1 . Then
Q F = 1 + a x y + c x 2 y + d x 2 y 2 + b x 2 y 2 , N F = y + a x y 2 + d x 2 y 3 + 2 b x 2 y 4 . Q_F=1+axy+cx^2y+dx^2y^2+bx^2y^2,\qquad N_F=y+axy^2+dx^2y^3+2bx^2y^4. Q F = 1 + a x y + c x 2 y + d x 2 y 2 + b x 2 y 2 , N F = y + a x y 2 + d x 2 y 3 + 2 b x 2 y 4 .
Proof of Corollary 1.4. By Lemma 4.1(b) we may enlarge the polytopes to
Δ ′ = conv { ( 0 , 0 ) , ( 2 , 1 ) , ( 2 , 3 ) } = { 0 ≤ X ≤ 2 , X / 2 ≤ Y ≤ 3 X / 2 } , \Delta'=\operatorname{conv}\{(0,0),(2,1),(2,3)\}=\{0\le X\le2,\ X/2\le Y\le3X/2\}, Δ ′ = conv {( 0 , 0 ) , ( 2 , 1 ) , ( 2 , 3 )} = { 0 ≤ X ≤ 2 , X /2 ≤ Y ≤ 3 X /2 } ,
K ′ = conv { ( 0 , 1 ) , ( 2 , 3 ) , ( 2 , 4 ) } = { 0 ≤ X ≤ 2 , 1 + X ≤ Y ≤ 1 + 3 X / 2 } , K'=\operatorname{conv}\{(0,1),(2,3),(2,4)\}=\{0\le X\le2,\ 1+X\le Y\le1+3X/2\}, K ′ = conv {( 0 , 1 ) , ( 2 , 3 ) , ( 2 , 4 )} = { 0 ≤ X ≤ 2 , 1 + X ≤ Y ≤ 1 + 3 X /2 } ,
which contain supp Q F \operatorname{supp}Q_F supp Q F and supp N F \operatorname{supp}N_F supp N F for all parameter values. Let 0 < ε ≤ 1 / 2 0<\varepsilon\le1/2 0 < ε ≤ 1/2 and ( X , Y ) = ( 1 − ε ) ( X 1 , Y 1 ) + ε ( X 2 , Y 2 ) (X,Y)=(1-\varepsilon)(X_1,Y_1)+\varepsilon(X_2,Y_2) ( X , Y ) = ( 1 − ε ) ( X 1 , Y 1 ) + ε ( X 2 , Y 2 ) with ( X 1 , Y 1 ) ∈ Δ ′ (X_1,Y_1)\in\Delta' ( X 1 , Y 1 ) ∈ Δ ′ and ( X 2 , Y 2 ) ∈ K ′ (X_2,Y_2)\in K' ( X 2 , Y 2 ) ∈ K ′ . Then 0 ≤ X ≤ 2 0\le X\le2 0 ≤ X ≤ 2 , and
Y ≥ ( 1 − ε ) X 1 2 + ε ( 1 + X 2 ) ≥ X 2 + ε , Y ≤ ( 1 − ε ) 3 X 1 2 + ε ( 1 + 3 X 2 2 ) = 3 X 2 + ε . Y\ge(1-\varepsilon)\frac{X_1}{2}+\varepsilon(1+X_2)\ge\frac{X}{2}+\varepsilon,\qquad Y\le(1-\varepsilon)\frac{3X_1}{2}+\varepsilon\left(1+\frac{3X_2}{2}\right)=\frac{3X}{2}+\varepsilon. Y ≥ ( 1 − ε ) 2 X 1 + ε ( 1 + X 2 ) ≥ 2 X + ε , Y ≤ ( 1 − ε ) 2 3 X 1 + ε ( 1 + 2 3 X 2 ) = 2 3 X + ε .
If X = 2 X=2 X = 2 , then X 1 = X 2 = 2 X_1=X_2=2 X 1 = X 2 = 2 and Y ≥ ( 1 − ε ) + 3 ε = 1 + 2 ε Y\ge(1-\varepsilon)+3\varepsilon=1+2\varepsilon Y ≥ ( 1 − ε ) + 3 ε = 1 + 2 ε . Now let u = ( u 1 , u 2 ) ∈ Z ≥ 0 2 u=(u_1,u_2)\in\mathbb{Z}_{\ge0}^2 u = ( u 1 , u 2 ) ∈ Z ≥ 0 2 and η ∈ [ 0 , 1 ) \eta\in[0,1) η ∈ [ 0 , 1 ) with ( u 1 + η , u 2 + η ) (u_1+\eta,u_2+\eta) ( u 1 + η , u 2 + η ) of this form. Then:
u 1 + η ≤ 2 u_1+\eta\le2 u 1 + η ≤ 2 , so u 1 ≤ 2 u_1\le2 u 1 ≤ 2 , and η = 0 \eta=0 η = 0 if u 1 = 2 u_1=2 u 1 = 2 ;
if u 1 = 0 u_1=0 u 1 = 0 : u 2 + η ≤ η / 2 + ε u_2+\eta\le\eta/2+\varepsilon u 2 + η ≤ η /2 + ε gives u 2 ≤ η / 2 + ε < 1 u_2\le\eta/2+\varepsilon<1 u 2 ≤ η /2 + ε < 1 , so u = ( 0 , 0 ) u=(0,0) u = ( 0 , 0 ) ;
if u 1 = 1 u_1=1 u 1 = 1 : u 2 + η ≥ 1 / 2 + ε u_2+\eta\ge1/2+\varepsilon u 2 + η ≥ 1/2 + ε gives u 2 ≥ 1 / 2 − η + ε > 0 u_2\ge1/2-\eta+\varepsilon>0 u 2 ≥ 1/2 − η + ε > 0 , and u 2 + η ≤ 3 ( 1 + η ) / 2 + ε u_2+\eta\le3(1+\eta)/2+\varepsilon u 2 + η ≤ 3 ( 1 + η ) /2 + ε gives u 2 ≤ 1 + η / 2 + ε < 5 / 2 u_2\le1+\eta/2+\varepsilon<5/2 u 2 ≤ 1 + η /2 + ε < 5/2 ; so u 2 ∈ { 1 , 2 } u_2\in\{1,2\} u 2 ∈ { 1 , 2 } ;
if u 1 = 2 u_1=2 u 1 = 2 : 1 + 2 ε ≤ u 2 ≤ 3 + ε 1+2\varepsilon\le u_2\le3+\varepsilon 1 + 2 ε ≤ u 2 ≤ 3 + ε , so u 2 ∈ { 2 , 3 } u_2\in\{2,3\} u 2 ∈ { 2 , 3 } .
Hence J 2 ( N F , Q F ) ⊂ { ( 0 , 0 ) , ( 1 , 1 ) , ( 1 , 2 ) , ( 2 , 2 ) , ( 2 , 3 ) } J_2(N_F,Q_F)\subset\{(0,0),(1,1),(1,2),(2,2),(2,3)\} J 2 ( N F , Q F ) ⊂ {( 0 , 0 ) , ( 1 , 1 ) , ( 1 , 2 ) , ( 2 , 2 ) , ( 2 , 3 )} , so ∣ J ∣ ≤ 5 |J|\le5 ∣ J ∣ ≤ 5 , and Theorem 1.2 gives K α ( F ) ≤ 2 6 α − 1 + 2 α − 1 K_\alpha(F)\le2^{6\alpha-1}+2^\alpha-1 K α ( F ) ≤ 2 6 α − 1 + 2 α − 1 . The lower bound for odd parameters is Theorem A.1 (b) in the appendix. □
Remark 6.1. For this family h = d = 2 h=d=2 h = d = 2 . The main term of the explicit bound of [19 Theorem 2] , which also gives the exponent of the asymptotic bound of [19 Theorem 1] , is p N p^N p N with N = 1 6 α ( α + 1 ) ( 7 α + 5 ) N=\frac{1}{6}\alpha(\alpha+1)(7\alpha+5) N = 6 1 α ( α + 1 ) ( 7 α + 5 ) ; this equals 4 , 19 , 52 , 110 , 200 4,19,52,110,200 4 , 19 , 52 , 110 , 200 for α = 1 , … , 5 \alpha=1,\ldots,5 α = 1 , … , 5 . The exponent of Corollary 1.4 is 6 α − 1 6\alpha-1 6 α − 1 , that is 5 , 11 , 17 , 23 , 29 5,11,17,23,29 5 , 11 , 17 , 23 , 29 . The general box bound of Corollary 1.3 has exponent 16 α − 1 16\alpha-1 16 α − 1 , that is 15 , 31 , 47 , 63 , 79 15,31,47,63,79 15 , 31 , 47 , 63 , 79 , and becomes smaller than the cubic exponent only from α = 3 \alpha=3 α = 3 on. This illustrates that all comparisons are asymptotic in α \alpha α for fixed input.
Remark 6.2. In the variables of [19 ] , which use the shear x ↦ x y − 1 x\mapsto xy^{-1} x ↦ x y − 1 and the central row C T y \mathrm{CT}_y CT y instead of the diagonal, the set J J J above becomes { ( 0 , 0 ) , ( 1 , 0 ) , ( 1 , 1 ) , ( 2 , 0 ) , ( 2 , 1 ) } \{(0,0),(1,0),(1,1),(2,0),(2,1)\} {( 0 , 0 ) , ( 1 , 0 ) , ( 1 , 1 ) , ( 2 , 0 ) , ( 2 , 1 )} .
The exact exponent for the odd quadratic family
The lower bound in Corollary 1.4 can be strengthened to a matching exponential rate. Throughout this section the parameters a , b , c , d a,b,c,d a , b , c , d are odd elements of Z 2 \mathbb{Z}_2 Z 2 , and F ( 0 ) = 0 F(0)=0 F ( 0 ) = 0 is the root of
b x 2 F 2 + ( 1 + a x + d x 2 ) F + c x 2 = 0. bx^2F^2+(1+ax+dx^2)F+cx^2=0. b x 2 F 2 + ( 1 + a x + d x 2 ) F + c x 2 = 0.
Theorem 7.1. For every α ≥ 1 \alpha\ge1 α ≥ 1 ,
9 ⋅ 8 α 64 α − 32 ≤ K α ( F ) ≤ 2 3 α − 1 + 3 ( 2 α + 1 + α + 1 ) ( 2 α − 1 ) . (11) \frac{9\cdot8^\alpha}{64\alpha-32}\le K_\alpha(F)\le2^{3\alpha-1}+3(2^{\alpha+1}+\alpha+1)(2^\alpha-1). \tag*{(11)} 64 α − 32 9 ⋅ 8 α ≤ K α ( F ) ≤ 2 3 α − 1 + 3 ( 2 α + 1 + α + 1 ) ( 2 α − 1 ) . (11) Consequently
lim α → ∞ log 2 K α ( F ) α = 3. \lim_{\alpha\to\infty}\frac{\log_2 K_\alpha(F)}{\alpha}=3. α → ∞ lim α log 2 K α ( F ) = 3.
The proof uses the histories of Section 5, first on the two-dimensional space of ordinary numerators and then in normalized series coordinates. The normalized full-window frames are independent of the odd parameters; the entire kernel cardinality need not be. For example, K 2 = 53 K_2=53 K 2 = 53 for ( a , b , c , d ) = ( 1 , 1 , 1 , 1 ) (a,b,c,d)=(1,1,1,1) ( a , b , c , d ) = ( 1 , 1 , 1 , 1 ) , whereas K 2 = 58 K_2=58 K 2 = 58 for ( a , b , c , d ) = ( 3 , 5 , 7 , 9 ) (a,b,c,d)=(3,5,7,9) ( a , b , c , d ) = ( 3 , 5 , 7 , 9 ) , as reproduced by the supplementary exact automaton calculation. The theorem does not assert convergence of K α ( F ) / 8 α K_\alpha(F)/8^\alpha K α ( F ) / 8 α .
Two ordinary coordinates and three exceptional rays
We use the central-row coordinates of Appendix A:
P 1 = 1 + a x + d x 2 , f = P / y = 1 + a x + x 2 ( b y + d + c y − 1 ) , N = P 1 y + 2 b x 2 y 2 , R = N / f . P_1=1+ax+dx^2,\qquad f=P/y=1+ax+x^2(by+d+cy^{-1}),\qquad N=P_1y+2bx^2y^2,\qquad R=N/f. P 1 = 1 + a x + d x 2 , f = P / y = 1 + a x + x 2 ( b y + d + c y − 1 ) , N = P 1 y + 2 b x 2 y 2 , R = N / f .
Thus F = CT y R F=\operatorname{CT}_y R F = CT y R . Write C r = Λ r , 0 C_r=\Lambda_{r,0} C r = Λ r , 0 , C = C 0 C=C_0 C = C 0 and C w C_w C w for the corresponding word operator; then CT y C w R = Λ w F \operatorname{CT}_y C_wR=\Lambda_wF CT y C w R = Λ w F . Under the shear x i y j ↦ x i y i + j x^iy^j\mapsto x^iy^{i+j} x i y j ↦ x i y i + j , these operators and δ t ( f ) \delta_t(f) δ t ( f ) become the operators and δ \delta δ -polynomials of Sections 3–5. In particular their history, compatibility and transition formulas hold with
T t , v ( M ) = C t ( x − v δ t ( f ) M ) , J s h = { ( 0 , 0 ) , ( 1 , 0 ) , ( 1 , 1 ) , ( 2 , 0 ) , ( 2 , 1 ) } . T_{t,v}(M)=C^t(x^{-v}\delta_t(f)M),\qquad J_{\mathrm{sh}}=\{(0,0),(1,0),(1,1),(2,0),(2,1)\}. T t , v ( M ) = C t ( x − v δ t ( f ) M ) , J sh = {( 0 , 0 ) , ( 1 , 0 ) , ( 1 , 1 ) , ( 2 , 0 ) , ( 2 , 1 )} .
The support of δ t ( f ) \delta_t(f) δ t ( f ) is contained in ( 2 t − 1 ) Δ (2^t-1)\Delta ( 2 t − 1 ) Δ , where
Δ = { ( i , j ) : i ≤ 2 , 2 j ≤ i , − 2 j ≤ i } . \Delta=\{(i,j):i\le2,\ 2j\le i,\ -2j\le i\}. Δ = {( i , j ) : i ≤ 2 , 2 j ≤ i , − 2 j ≤ i } .
Put
W 0 = ⟨ 1 , x ⟩ Z 2 , V 0 = ⟨ x 2 , x 2 y ⟩ Z 2 , V 1 = ⟨ x y ⟩ Z 2 . W_0=\langle1,x\rangle_{\mathbb{Z}_2},\qquad V_0=\langle x^2,x^2y\rangle_{\mathbb{Z}_2},\qquad V_1=\langle xy\rangle_{\mathbb{Z}_2}. W 0 = ⟨ 1 , x ⟩ Z 2 , V 0 = ⟨ x 2 , x 2 y ⟩ Z 2 , V 1 = ⟨ x y ⟩ Z 2 .
Here and below a history supported on W 0 W_0 W 0 means that all its entries belong to this rank-two module at their respective precisions.
Lemma 7.2. Every T t , v T_{t,v} T t , v preserves W 0 W_0 W 0 . The zero-level numerator m k ( w ) m_k(w) m k ( w ) belongs to W 0 W_0 W 0 unless
w ∈ E : = { ∅ } ∪ { 0 k , 1 k , 01 k − 1 : k ≥ 1 } . w\in\mathcal{E}:=\{\varnothing\}\cup\{0^k,1^k,01^{k-1}:k\ge1\}. w ∈ E := { ∅ } ∪ { 0 k , 1 k , 0 1 k − 1 : k ≥ 1 } . The language E \mathcal{E} E is closed under taking prefixes.
Proof. Let P = 2 t P=2^t P = 2 t . A source monomial x i y j x^iy^j x i y j can contribute to the target x I y J x^Iy^J x I y J only through the exponent ( P I + v − i , P J − j ) (PI+v-i,PJ-j) ( P I + v − i , P J − j ) of δ t ( f ) \delta_t(f) δ t ( f ) . To reach either target x 2 x^2 x 2 or x 2 y x^2y x 2 y , its first coordinate must be at most 2 P − 2 2P-2 2 P − 2 ; hence i ≥ v + 2 i\ge v+2 i ≥ v + 2 . For a source in J s h J_{\mathrm{sh}} J sh this is possible only if v = 0 v=0 v = 0 and the source belongs to V 0 V_0 V 0 . To reach x y xy x y , the upper facet of ( P − 1 ) Δ (P-1)\Delta ( P − 1 ) Δ gives
2 ( P − j ) ≤ P + v − i , v ≥ P + i − 2 j . 2(P-j)\le P+v-i,\qquad v\ge P+i-2j. 2 ( P − j ) ≤ P + v − i , v ≥ P + i − 2 j . For the sources 1 , x , x y , x 2 , x 2 y 1,x,xy,x^2,x^2y 1 , x , x y , x 2 , x 2 y the thresholds are respectively P , P + 1 , P − 1 , P + 2 , P P,P+1,P-1,P+2,P P , P + 1 , P − 1 , P + 2 , P . Thus only v = P − 1 v=P-1 v = P − 1 and a source in V 1 V_1 V 1 can contribute. Since L J s h = W 0 ⊕ V 0 ⊕ V 1 L_{J_{\mathrm{sh}}}=W_0\oplus V_0\oplus V_1 L J sh = W 0 ⊕ V 0 ⊕ V 1 is stable, these two observations prove the assertion about W 0 W_0 W 0 and show that on the quotient modulo W 0 W_0 W 0 only the
V 0 V_0 V 0 block survives when v = 0 v = 0 v = 0 , only the V 1 V_1 V 1 block survives when v = P − 1 v = P - 1 v = P − 1 , and all other blocks vanish.
For the four possible source monomials of N N N , with exponents ( 0 , 1 ) , ( 1 , 1 ) , ( 2 , 1 ) , ( 2 , 2 ) (0,1),(1,1),(2,1),(2,2) ( 0 , 1 ) , ( 1 , 1 ) , ( 2 , 1 ) , ( 2 , 2 ) , the first-coordinate test permits V 0 V_0 V 0 output only when v = 0 v = 0 v = 0 . The four thresholds for V 1 V_1 V 1 output are P − 2 , P − 1 , P , P − 2 P - 2, P - 1, P, P - 2 P − 2 , P − 1 , P , P − 2 , so only v = P − 2 v = P - 2 v = P − 2 or v = P − 1 v = P - 1 v = P − 1 can contribute. Project the recurrence (4) to the two summands of L J s h / W 0 L_{J_{\mathrm{sh}}}/W_0 L J sh / W 0 . Inductively, a nonzero V 0 V_0 V 0 component requires an all-zero suffix and an all-zero preceding prefix, or the direct source term with v = 0 v = 0 v = 0 ; thus its word is 0 k 0^k 0 k . A nonzero V 1 V_1 V 1 component requires an all-one suffix and a preceding prefix of the form 1 j 1^j 1 j or 01 j − 1 01^{j-1} 0 1 j − 1 , or the direct source term with v = 2 k − 1 v = 2^k - 1 v = 2 k − 1 or 2 k − 2 2^k - 2 2 k − 2 . These are precisely 1 k 1^k 1 k and 01 k − 1 01^{k-1} 0 1 k − 1 . This proves the claim, and prefix closure follows from the displayed list of words.
Lemma 7.3. Modulo 2 α 2^\alpha 2 α , each of the three rays C 0 k R C_{0^k}R C 0 k R , C 1 k R C_{1^k}R C 1 k R and C 01 k − 1 R C_{01^{k-1}}R C 0 1 k − 1 R takes at most L α = 2 α + 1 + α + 1 L_\alpha= 2^{\alpha+1} + \alpha+ 1 L α = 2 α + 1 + α + 1 distinct values, counting the initial state in the union of the rays.
Proof. We extend the annihilation argument of Appendix A to M j = x j N M_j = x^jN M j = x j N , j = 0 , 1 , 2 j = 0,1,2 j = 0 , 1 , 2 . For the all-zero word the support induction in Lemma 4.2 gives
supp m k ( M j ) ⊂ ( 1 − 2 − k ) Δ + 2 − k conv ( supp M j ) . \operatorname{supp} m_k(M_j) \subset(1 - 2^{-k})\Delta+ 2^{-k}\operatorname{conv}(\operatorname{supp} M_j). supp m k ( M j ) ⊂ ( 1 − 2 − k ) Δ + 2 − k conv ( supp M j ) .
On the second polytope the maxima of i i i and 2 v − i 2v - i 2 v − i are j + 2 j + 2 j + 2 and 2 − j 2 - j 2 − j , and the minimum of i + 2 v i + 2v i + 2 v is j + 2 > 0 j + 2 > 0 j + 2 > 0 . Consequently points in the displayed interpolation satisfy
i ≤ 2 + j 2 − k , 2 v − i ≤ ( 2 − j ) 2 − k , i + 2 v ≥ ( j + 2 ) 2 − k > 0. i \le2 + j2^{-k}, \qquad2v - i \le(2-j)2^{-k}, \qquad i + 2v \ge(j+2)2^{-k} > 0. i ≤ 2 + j 2 − k , 2 v − i ≤ ( 2 − j ) 2 − k , i + 2 v ≥ ( j + 2 ) 2 − k > 0.
The remaining facet − 2 v ≤ i -2v \le i − 2 v ≤ i already holds. Taking integer points shows
m k ( M j ) ∈ L J 0 , J 0 = { ( 1 , 0 ) , ( 2 , 0 ) , ( 2 , 1 ) } , k ≥ h j , ( h 0 , h 1 , h 2 ) = ( 2 , 1 , 2 ) . m_k(M_j) \in L_{J_0}, \qquad J_0 = \{(1,0),(2,0),(2,1)\}, \qquad k \ge h_j, \qquad(h_0,h_1,h_2) = (2,1,2). m k ( M j ) ∈ L J 0 , J 0 = {( 1 , 0 ) , ( 2 , 0 ) , ( 2 , 1 )} , k ≥ h j , ( h 0 , h 1 , h 2 ) = ( 2 , 1 , 2 ) .
More generally, suppose an integral numerator M M M has m k ( M ) ∈ L J 0 m_k(M) \in L_{J_0} m k ( M ) ∈ L J 0 for k ≥ h k \ge h k ≥ h , where h ≥ 1 h \ge1 h ≥ 1 . Let Θ s \Theta_s Θ s and χ ( t ) = ∑ r = 0 3 χ r t r = ( t 2 − 1 ) ( t − μ ) \chi(t) = \sum_{r=0}^{3}\chi_rt^r = (t^2 - 1)(t-\mu) χ ( t ) = ∑ r = 0 3 χ r t r = ( t 2 − 1 ) ( t − μ ) , μ \mu μ odd, be as in Appendix A. We claim that
χ ( C ) C k ( M / f ) ∈ 2 k − h + 1 A ^ + ( k ≥ h − 1 ) . (12) \chi(\mathcal{C})\mathcal{C}^k(M/f) \in2^{k-h+1}\widehat{A}_+ \qquad(k \ge h-1). \tag*{(12)} χ ( C ) C k ( M / f ) ∈ 2 k − h + 1 A + ( k ≥ h − 1 ) . (12)
The notation A ^ + \widehat{A}_+ A + denotes the series ring supported on i ≥ 0 i \ge0 i ≥ 0 , i + v ≥ 0 i+v \ge0 i + v ≥ 0 , as defined in that appendix; all series here belong to it. For k ≥ h k \ge h k ≥ h , split the zero-level recurrence as
m k = ∑ t = 1 k − h E t ( m k − t ) + s k , s k = ∑ t = k − h + 1 k E t ( m k − t ) , E t ( M ) = C t ( δ t ( f ) M ) . m_k = \sum_{t=1}^{k-h} E_t(m_{k-t}) + s_k, \qquad s_k = \sum_{t=k-h+1}^{k} E_t(m_{k-t}), \qquad E_t(M) = \mathcal{C}^t(\delta_t(f)M). m k = t = 1 ∑ k − h E t ( m k − t ) + s k , s k = t = k − h + 1 ∑ k E t ( m k − t ) , E t ( M ) = C t ( δ t ( f ) M ) .
The first sum and m k m_k m k lie in L J 0 L_{J_0} L J 0 , so s k s_k s k does too, and s k ∈ 2 k − h L J 0 s_k \in2^{k-h}L_{J_0} s k ∈ 2 k − h L J 0 . The right convolution identity (A.2) gives by induction
[ m k ] = ∑ l = h k [ s l ] Θ k − l . [m_k] = \sum_{l=h}^{k} [s_l]\Theta_{k-l}. [ m k ] = l = h ∑ k [ s l ] Θ k − l .
In the residual ∑ r χ r m k + r \sum_r \chi_r m_{k+r} ∑ r χ r m k + r the terms with l ≤ k l \le k l ≤ k are divisible by 2 l − h 2 k − l + 1 = 2 k − h + 1 2^{l-h}2^{k-l+1} = 2^{k-h+1} 2 l − h 2 k − l + 1 = 2 k − h + 1 by Lemma A.2; terms with l > k l > k l > k are divisible by 2 l − h 2^{l-h} 2 l − h , which is at least the same power. This proves the required divisibility at level zero. At level j ≥ 1 j \ge1 j ≥ 1 , use (A.1). For t ≤ k − h t \le k-h t ≤ k − h , the inner residual at time k − t k-t k − t is divisible by 2 k − t − h + 1 2^{k-t-h+1} 2 k − t − h + 1 ; multiplication by δ j + t ( f ) \delta_{j+t}(f) δ j + t ( f ) supplies 2 j + t − 1 2^{j+t-1} 2 j + t − 1 , giving 2 k − h + j 2^{k-h+j} 2 k − h + j . For t ≥ k − h + 1 t \ge k-h+1 t ≥ k − h + 1 , the factor δ j + t ( f ) \delta_{j+t}(f) δ j + t ( f ) alone is divisible by 2 k − h + j 2^{k-h+j} 2 k − h + j . These powers are at least 2 k − h + 1 2^{k-h+1} 2 k − h + 1 . The convergent level expansion preserves the divisibility. The case k = h − 1 k = h-1 k = h − 1 has exponent zero and is automatic. This proves (7.2).
At k = α + h − 1 k = \alpha+ h - 1 k = α + h − 1 , the state S = C k ( M / f ) S = \mathcal{C}^k(M/f) S = C k ( M / f ) is killed by χ ( C ) \chi(\mathcal{C}) χ ( C ) modulo 2 α 2^\alpha 2 α . Put g m ( t ) = ∑ i = 0 m − 1 t 2 i g_m(t) = \sum_{i=0}^{m-1} t^{2i} g m ( t ) = ∑ i = 0 m − 1 t 2 i . The polynomial
t 2 m − 1 − g m ( μ ) ( t 2 − 1 ) = ( t 2 − 1 ) ( g m ( t ) − g m ( μ ) ) t^{2m} - 1 - g_m(\mu)(t^2 - 1) = (t^2 - 1)(g_m(t) - g_m(\mu)) t 2 m − 1 − g m ( μ ) ( t 2 − 1 ) = ( t 2 − 1 ) ( g m ( t ) − g m ( μ ))
is divisible by χ ( t ) \chi(t) χ ( t ) . For m = 2 α m=2^\alpha m = 2 α , the factorisation g m ( μ ) = ∏ i = 0 α − 1 ( 1 + μ 2 i + 1 ) g_m(\mu)=\prod_{i=0}^{\alpha-1}(1+\mu^{2^{i+1}}) g m ( μ ) = ∏ i = 0 α − 1 ( 1 + μ 2 i + 1 ) is divisible by 2 α 2^\alpha 2 α , since every factor is even. Thus ( C 2 α + 1 − 1 ) S = 0 (C^{2^{\alpha+1}}-1)S=0 ( C 2 α + 1 − 1 ) S = 0 modulo 2 α 2^\alpha 2 α . The orbit of M / f M/f M / f therefore has preperiod at most α + h − 1 \alpha+h-1 α + h − 1 and period dividing 2 α + 1 2^{\alpha+1} 2 α + 1 .
For a word of length k k k and digit value v v v , its sheared operator is C k ( x − v ⋅ ) C^k(x^{-v}\cdot) C k ( x − v ⋅ ) . Coefficient extraction gives
C 0 k R = C k R , C 1 k R = x − 1 C k ( x R ) , C 01 k − 1 R = x − 1 C k ( x 2 R ) ( k ≥ 1 ) . C_{0^k}R=C^kR,\qquad C_{1^k}R=x^{-1}C^k(xR),\qquad C_{0 1^{k-1}}R=x^{-1}C^k(x^2R)\quad(k\geq1). C 0 k R = C k R , C 1 k R = x − 1 C k ( x R ) , C 0 1 k − 1 R = x − 1 C k ( x 2 R ) ( k ≥ 1 ) .
The division by x x x in the last two identities preserves congruences; the selected numerators are divisible by x x x . Apply the orbit estimate with h = h j ≤ 2 h=h_j\leq2 h = h j ≤ 2 . Each ray has at most 2 α + 1 + α + 1 2^{\alpha+1}+\alpha+1 2 α + 1 + α + 1 values, as required.
Proposition 7.4. For every α ≥ 1 \alpha\geq1 α ≥ 1 ,
K α ( F ) ≤ 2 3 α − 1 + 3 ( 2 α + 1 + α + 1 ) ( 2 α − 1 ) . K_\alpha(F)\leq2^{3\alpha-1}+3(2^{\alpha+1}+\alpha+1)(2^\alpha-1). K α ( F ) ≤ 2 3 α − 1 + 3 ( 2 α + 1 + α + 1 ) ( 2 α − 1 ) .
Proof. Let a word first leave E \mathcal{E} E at its prefix of length e e e . Every later zero-level numerator belongs to W 0 W_0 W 0 , by Lemma 7.2. If its length k k k satisfies k − e ≥ α − 1 k-e\geq\alpha-1 k − e ≥ α − 1 , its entire precision-α \alpha α history is supported on W 0 W_0 W 0 . Since all T t , v T_{t,v} T t , v preserve W 0 W_0 W 0 , the proof of Lemma 5.7 gives exactly 2 2 α 2^{2\alpha} 2 2 α compatible W 0 W_0 W 0 -histories per window. There are 2 α − 1 2^{\alpha-1} 2 α − 1 windows, so these states have at most 2 3 α − 1 2^{3\alpha-1} 2 3 α − 1 distinct decoded values.
Otherwise the word is u r v urv u r v , where u ∈ E u\in\mathcal{E} u ∈ E , r r r is an exit digit and ∣ v ∣ ≤ α − 2 |v|\leq\alpha-2 ∣ v ∣ ≤ α − 2 . By Lemma 7.3 there are at most 3 L α 3L_\alpha 3 L α states after u u u , at most two choices of r r r , and 2 α − 1 − 1 2^{\alpha-1}-1 2 α − 1 − 1 suffixes of length at most α − 2 \alpha-2 α − 2 . Including the exceptional states themselves gives at most
3 L α ( 1 + 2 ( 2 α − 1 − 1 ) ) = 3 L α ( 2 α − 1 ) 3L_\alpha(1+2(2^{\alpha-1}-1))=3L_\alpha(2^\alpha-1) 3 L α ( 1 + 2 ( 2 α − 1 − 1 )) = 3 L α ( 2 α − 1 )
additional states. For α = 1 \alpha=1 α = 1 the suffix sum is empty; the same formula remains valid. Every word is covered by one of these cases, including words too short to fill a history. Taking the central row cannot increase the count.
Window planes and normalized frames
Set R α = Z / 2 α Z R_\alpha=\mathbb{Z}/2^\alpha\mathbb{Z} R α = Z / 2 α Z , P θ = 1 + a x + d x 2 P_\theta=1+ax+dx^2 P θ = 1 + a x + d x 2 , and
B θ = CT y 1 f = ( P θ 2 − 4 b c x 4 ) − 1 / 2 , B θ ( 0 ) = 1 , W 0 = ⟨ 1 , x ⟩ Z 2 . B_\theta=\operatorname{CT}_y\frac{1}{f}=(P_\theta^2-4bcx^4)^{-1/2},\qquad B_\theta(0)=1,\qquad W_0=\langle1,x\rangle_{\mathbb{Z}_2}. B θ = CT y f 1 = ( P θ 2 − 4 b c x 4 ) − 1/2 , B θ ( 0 ) = 1 , W 0 = ⟨ 1 , x ⟩ Z 2 .
Here the constant-term identity follows by expanding 1 / f 1/f 1/ f in powers of x 2 ( b y + c y − 1 ) / P θ x^2(by+cy^{-1})/P_\theta x 2 ( b y + c y − 1 ) / P θ : only the even powers contribute, and their constant terms are ( 2 j j ) ( b c ) j \binom{2j}{j}(bc)^j ( j 2 j ) ( b c ) j . Thus
B θ = P θ − 1 ∑ j ≥ 0 ( 2 j j ) ( b c x 4 P θ 2 ) j ∈ Z 2 [ [ x ] ] . B_\theta=P_\theta^{-1}\sum_{j\geq0}\binom{2j}{j}\left(\frac{bcx^4}{P_\theta^2}\right)^j\in\mathbb{Z}_2[[x]]. B θ = P θ − 1 j ≥ 0 ∑ ( j 2 j ) ( P θ 2 b c x 4 ) j ∈ Z 2 [[ x ]] .
For a binary word w = r 0 ⋯ r d − 1 w=r_0\cdots r_{d-1} w = r 0 ⋯ r d − 1 read in that order, put Q = 2 d Q=2^d Q = 2 d and v ( w ) = ∑ i = 0 d − 1 r i 2 i v(w)=\sum_{i=0}^{d-1}r_i2^i v ( w ) = ∑ i = 0 d − 1 r i 2 i . Let Λ w = Λ r d − 1 ⋯ Λ r 0 = Λ Q , v ( w ) \Lambda_w=\Lambda_{r_{d-1}}\cdots\Lambda_{r_0}=\Lambda_{Q,v(w)} Λ w = Λ r d − 1 ⋯ Λ r 0 = Λ Q , v ( w ) . Write
E ^ w = ( Λ w B θ , Λ w ( x B θ ) ) , L ( h ) = ( [ x 0 ] h [ x 1 ] h ) , M w = L ( E ^ w ) . \widehat{E}_w=(\Lambda_wB_\theta,\Lambda_w(xB_\theta)),\qquad L(h)=\begin{pmatrix}[x^0]h\\ [x^1]h\end{pmatrix},\qquad M_w=L(\widehat{E}_w). E w = ( Λ w B θ , Λ w ( x B θ )) , L ( h ) = ( [ x 0 ] h [ x 1 ] h ) , M w = L ( E w ) .
The map L L L is applied to each column of a row of series. If B θ = ∑ b n x n B_\theta=\sum b_nx^n B θ = ∑ b n x n , its row of index n n n is ( b Q n + v , b Q n + v − 1 ) (b_{Qn+v},b_{Qn+v-1}) ( b Q n + v , b Q n + v − 1 ) , with b j = 0 b_j=0 b j = 0 for j < 0 j<0 j < 0 . Modulo 2 2 2 , B θ = 1 / ( 1 + x + x 2 ) B_\theta=1/(1+x+x^2) B θ = 1/ ( 1 + x + x 2 ) , and sections on the numerator space ⟨ 1 , x ⟩ \langle1,x\rangle ⟨ 1 , x ⟩ have row matrices
( 1 1 0 1 ) , ( 1 0 1 1 ) . \begin{pmatrix}1&1\\0&1\end{pmatrix},\qquad
\begin{pmatrix}1&0\\1&1\end{pmatrix}. ( 1 0 1 1 ) , ( 1 1 0 1 ) .
Both are invertible, as is extraction of the first two coefficients of ( u + v x ) / ( 1 + x + x 2 ) (u+vx)/(1+x+x^2) ( u + v x ) / ( 1 + x + x 2 ) . Hence M w ∈ GL 2 ( Z 2 ) M_w\in\operatorname{GL}_2(\mathbb{Z}_2) M w ∈ GL 2 ( Z 2 ) . Define
E w = E ^ w M w − 1 , G α , w = E w R α 2 ⊂ R α [ [ x ] ] . (13) E_w=\widehat{E}_wM_w^{-1},\qquad G_{\alpha,w}=E_wR_\alpha^2\subset R_\alpha[[x]]. \tag*{(13)} E w = E w M w − 1 , G α , w = E w R α 2 ⊂ R α [[ x ]] . (13)
The first two rows of E w E_w E w form the identity. The plane G α , w G_{\alpha,w} G α , w therefore has exactly 4 α 4^\alpha 4 α elements, each uniquely specified by its first two coefficients.
Lemma 7.5. For ∣ w ∣ = α − 1 |w|=\alpha-1 ∣ w ∣ = α − 1 , the plane G α , w G_{\alpha,w} G α , w is the central-row image of the compatible histories supported on W 0 W_0 W 0 with window w w w . For 1 ≤ s ≤ α 1\le s\le\alpha 1 ≤ s ≤ α and ∣ w ∣ ≥ α − 1 |w|\ge\alpha-1 ∣ w ∣ ≥ α − 1 ,
E w ≡ E w s ( m o d 2 s ) , (14) E_w \equiv E_{w_s} \pmod{2^s}, \tag*{(14)} E w ≡ E w s ( mod 2 s ) , (14) where w s w_s w s consists of the last s − 1 s-1 s − 1 digits of w w w .
Proof. The support calculation in Lemma 7.2 shows that every T t , v T_{t,v} T t , v preserves W 0 W_0 W 0 . The triangular proof of Lemma 5.7 consequently gives 4 α 4^\alpha 4 α compatible W 0 W_0 W 0 -histories for a fixed window. Start with M / f M/f M / f , where M ∈ W 0 M\in W_0 M ∈ W 0 . After d = α − 1 d=\alpha-1 d = α − 1 digits all retained zero-level numerators, including the oldest numerator m 0 = M m_0=M m 0 = M , lie in W 0 W_0 W 0 . The compatibility and decoding arguments of Section 5 apply also at this boundary length: all retained numerator indices are nonnegative, and all omitted terms have the same divisibility as before. The map M ↦ m d M\mapsto m_d M ↦ m d is linear; modulo 2 its matrix is a product of the two displayed invertible matrices. Varying M M M modulo 2 α 2^\alpha 2 α therefore supplies 4 α 4^\alpha 4 α different compatible histories, hence all of them. Their central rows are exactly E ~ w R 2 α \widetilde{E}_wR_2^\alpha E w R 2 α and are distinct because M w M_w M w is invertible.
For a longer word, any initial segment maps W 0 W_0 W 0 to itself. Modulo 2 the induced map on the two coordinates is again invertible, so varying the initial M M M still supplies the entire fixed-window history group. Now reduce a history from precision α \alpha α to precision s s s , discard its older entries, and retain its terminal s − 1 s-1 s − 1 window digits. In a compatibility relation the discarded terms vanish because T t , v T_{t,v} T t , v is divisible by 2 t − 1 2^{t-1} 2 t − 1 . In the decoding formula they vanish by Lemma 5.1. Thus the reduced plane is contained in G s , w s G_{s,w_s} G s , w s . It still has 4 s 4^s 4 s elements, since its first two coefficient map is invertible, so the inclusion is equality. Equal planes with first-two-row normalization have equal frames: the column with initial coefficients ( 1 , 0 ) (1,0) ( 1 , 0 ) , and likewise that with ( 0 , 1 ) (0,1) ( 0 , 1 ) , is unique. This proves (7.4). □
Primitive histories are reachable.
Write W 0 = ⟨ 1 , x ⟩ Z 2 W_0=\langle1,x\rangle_{\mathbb Z_2} W 0 = ⟨ 1 , x ⟩ Z 2 . For a window w ∈ { 0 , 1 } α − 1 w\in\{0,1\}^{\alpha-1} w ∈ { 0 , 1 } α − 1 , let H α , w \mathcal{H}_{\alpha,w} H α , w denote the compatible histories of Definition 5.5 whose entries belong to W 0 W_0 W 0 , using the sheared operators. We call H 0 = u + v x H_0=u+vx H 0 = u + v x primitive if u u u or v v v is odd.
Lemma 7.6. For every window w w w , the map H ↦ H 0 H\mapsto H_0 H ↦ H 0 identifies H α , w \mathcal{H}_{\alpha,w} H α , w with ( Z / 2 α ) 2 (\mathbb{Z}/2^\alpha)^2 ( Z / 2 α ) 2 . Reducing the precision from α \alpha α to β ≤ α \beta\le\alpha β ≤ α reduces H s H_s H s modulo 2 β − s 2^{\beta-s} 2 β − s for s < β s<\beta s < β and keeps the last β − 1 \beta-1 β − 1 digits of the window. These reductions commute with digit transitions. Each transition between two ordinary window fibres is an invertible linear map on H 0 H_0 H 0 .
Proof. The exact matrices of T 1 , r T_{1,r} T 1 , r on W 0 W_0 W 0 , in row coordinates, are
G 1 , 0 = ( 1 d 0 a ) , G 1 , 1 = ( a 0 1 d ) . G_{1,0}=
\begin{pmatrix}
1 & d\\
0 & a
\end{pmatrix},
\qquad
G_{1,1}=
\begin{pmatrix}
a & 0\\
1 & d
\end{pmatrix}. G 1 , 0 = ( 1 0 d a ) , G 1 , 1 = ( a 1 0 d ) .
They are invertible modulo 2. Suppose first that a compatible history has H 0 = 0 H_0=0 H 0 = 0 . Reducing its compatibility equations successively modulo 2 shows that every H s H_s H s is even: at each step only T 1 , r ( H s + 1 ) T_{1,r}(H_{s+1}) T 1 , r ( H s + 1 ) survives, because T t , v T_{t,v} T t , v is divisible by 2 t − 1 2^{t-1} 2 t − 1 . Divide the entries by 2, omit the last entry, and retain the last α − 2 \alpha-2 α − 2 window digits. This gives a compatible history at precision α − 1 \alpha-1 α − 1 with first entry zero. Induction, starting with α = 1 \alpha=1 α = 1 , proves that the original history is zero. The triangular count of Lemma 5.7, applied to the invariant module W 0 W_0 W 0 , gives ∣ H α , w ∣ = 2 2 α |\mathcal{H}_{\alpha,w}|=2^{2\alpha} ∣ H α , w ∣ = 2 2 α , so the injective map to H 0 ∈ ( Z / 2 α ) 2 H_0\in(\mathbb{Z}/2^\alpha)^2 H 0 ∈ ( Z / 2 α ) 2 is bijective.
In a compatibility equation at position s s s , the terms discarded by reduction to precision β \beta β have t ≥ β − s t\ge\beta-s t ≥ β − s and hence are zero modulo 2 β − s − 1 2^{\beta-s-1} 2 β − s − 1 . The retained digit values depend only on the retained suffix. Thus the asserted reduction preserves compatibility and, by uniqueness, sends each canonical history with given H 0 H_0 H 0 to the canonical reduced history. In the transition formula (5.6), the terms t > β t > \beta t > β vanish modulo 2 β 2^{\beta} 2 β ; changing H t − 1 H_{t-1} H t − 1 by a multiple of 2 β − t + 1 2^{\beta-t+1} 2 β − t + 1 also makes no change modulo 2 β 2^{\beta} 2 β . The history shift and window suffix update commute with reduction as well. Finally the transition on H 0 H_0 H 0 is linear, and modulo 2 2 2 is G 1 , r G_{1,r} G 1 , r . Its determinant is odd, proving invertibility at every precision.
Lemma 7.7. Every primitive history in H α , w \mathcal{H}_{\alpha,w} H α , w is reached from R = N / f R = N/f R = N / f by a digit word. Within each window its decoded central-row series are distinct. Consequently each window contributes exactly 3 ⋅ 4 α − 1 3 \cdot4^{\alpha-1} 3 ⋅ 4 α − 1 distinct reachable primitive series.
Proof. At the zero window let A α A_\alpha A α be the digit-0 matrix on H 0 H_0 H 0 . Lemma 7.6 shows that these matrices are coherent, and hence define A ∈ GL 2 ( Z 2 ) A \in\operatorname{GL}_2(\mathbb{Z}_2) A ∈ GL 2 ( Z 2 ) . Modulo 2 2 2 , A A A is ( 1 1 0 1 ) \begin{pmatrix}1 & 1 \\ 0 & 1\end{pmatrix} ( 1 0 1 1 ) . Therefore A 2 = I + 2 D 0 A^2 = I + 2D_0 A 2 = I + 2 D 0 for an integral matrix D 0 D_0 D 0 , and repeated squaring gives
A α 2 α = I ( m o d 2 α ) . A_\alpha^{2^\alpha} = I \pmod{2^\alpha}. A α 2 α = I ( mod 2 α ) .
Put L α = 2 α + 1 L_\alpha= 2^{\alpha+1} L α = 2 α + 1 , and let B α B_\alpha B α be the matrix of the return word 10 L α 10^{L_\alpha} 1 0 L α starting at the zero window. It is invertible. These matrices are coherent: after reduction to precision β \beta β , the first L β L_\beta L β zeros return the window to zero, and the remaining zeros multiply by A β L α − L β = I A_\beta^{L_\alpha-L_\beta} = I A β L α − L β = I , since 2 β ∣ L α − L β 2^\beta\mid L_\alpha- L_\beta 2 β ∣ L α − L β . They define B ∈ GL 2 ( Z 2 ) B \in\operatorname{GL}_2(\mathbb{Z}_2) B ∈ GL 2 ( Z 2 ) . Every positive word in A α , B α A_\alpha,B_\alpha A α , B α is an actual return word. Inverses can also be realized by positive words, since each invertible matrix over the finite ring Z / 2 α \mathbb{Z}/2^\alpha Z / 2 α has finite order.
Here is an algebraic calculation of the return matrices, uniform in all four odd parameters. Set
a = 1 + 2 r , d = 1 + 2 s , b = 1 + 2 u , c = 1 + 2 v , t = u + v . a = 1 + 2r,\qquad d = 1 + 2s,\qquad b = 1 + 2u,\qquad c = 1 + 2v,\qquad t = u + v. a = 1 + 2 r , d = 1 + 2 s , b = 1 + 2 u , c = 1 + 2 v , t = u + v .
All matrices in the next calculation are modulo 8 8 8 . For i , k ∈ { 0 , 1 } i,k \in\{0,1\} i , k ∈ { 0 , 1 } put
( G j , e ) i , k = [ x 2 j k + e − i y i ] δ j ( f ) . (G_{j,e})_{i,k} = [x^{2^j k+e-i}y^i]\delta_j(f). ( G j , e ) i , k = [ x 2 j k + e − i y i ] δ j ( f ) .
The identities
δ 1 = f , δ 2 = f 3 − f σ ( f ) , δ 3 = f 7 − f σ ( f 3 ) − δ 2 σ 2 ( f ) \delta_1 = f,\qquad\delta_2 = f^3-f\sigma(f),\qquad\delta_3 = f^7-f\sigma(f^3)-\delta_2\sigma^2(f) δ 1 = f , δ 2 = f 3 − f σ ( f ) , δ 3 = f 7 − f σ ( f 3 ) − δ 2 σ 2 ( f )
give the following finite coefficient table. Write
g = 2 + 6 r + 4 t + 2 s + 4 r s , h = 6 + 6 r + 4 s , j ∗ = 4 t + 6 s + 4 r s . g = 2 + 6r + 4t + 2s + 4rs,\qquad h = 6 + 6r + 4s,\qquad j_* = 4t + 6s + 4rs. g = 2 + 6 r + 4 t + 2 s + 4 r s , h = 6 + 6 r + 4 s , j ∗ = 4 t + 6 s + 4 r s .
Then
G 1 , 0 = ( 1 1 + 2 s 0 1 + 2 r ) , G 1 , 1 = ( 1 + 2 r 0 1 1 + 2 s ) , G_{1,0}=\begin{pmatrix}1 & 1+2s\\0 & 1+2r\end{pmatrix},\qquad
G_{1,1}=\begin{pmatrix}1+2r & 0\\1 & 1+2s\end{pmatrix}, G 1 , 0 = ( 1 0 1 + 2 s 1 + 2 r ) , G 1 , 1 = ( 1 + 2 r 1 0 1 + 2 s ) ,
G 2 , 0 = ( 0 g 0 h ) , G 2 , 1 = ( 2 + 4 r j ∗ 0 g ) , G 2 , 2 = ( 4 + 6 r + 4 s 6 + 4 t + 6 s 2 + 4 r j ∗ ) , G_{2,0}=\begin{pmatrix}0 & g\\0 & h\end{pmatrix},\qquad
G_{2,1}=\begin{pmatrix}2+4r & j_*\\0 & g\end{pmatrix},\qquad
G_{2,2}=\begin{pmatrix}4+6r+4s & 6+4t+6s\\2+4r & j_*\end{pmatrix}, G 2 , 0 = ( 0 0 g h ) , G 2 , 1 = ( 2 + 4 r 0 j ∗ g ) , G 2 , 2 = ( 4 + 6 r + 4 s 2 + 4 r 6 + 4 t + 6 s j ∗ ) ,
G 3 , 0 = 4 ( 0 1 + r + t + r s 0 r + s ) , G 3 , 1 = 4 ( 1 t + r s 0 1 + r + t + r s ) , G_{3,0}=4\begin{pmatrix}0 & 1+r+t+rs\\0 & r+s\end{pmatrix},\qquad
G_{3,1}=4\begin{pmatrix}1 & t+rs\\0 & 1+r+t+rs\end{pmatrix}, G 3 , 0 = 4 ( 0 0 1 + r + t + r s r + s ) , G 3 , 1 = 4 ( 1 0 t + r s 1 + r + t + r s ) ,
G 3 , 2 = 4 ( 1 + r r + s 1 t + r s ) , G 3 , 4 = 4 ( r + s s + r s 1 + r t + s + r s ) . G_{3,2}=4\begin{pmatrix}1+r & r+s\\1 & t+rs\end{pmatrix},\qquad
G_{3,4}=4\begin{pmatrix}r+s & s+rs\\1+r & t+s+rs\end{pmatrix}. G 3 , 2 = 4 ( 1 + r 1 r + s t + r s ) , G 3 , 4 = 4 ( r + s 1 + r s + r s t + s + r s ) .
For completeness these entries can be obtained directly without parameter specialization: expand the three displayed polynomials, discard terms of degree at least three in 2 r , 2 s , 2 u , 2 v 2r,2s,2u,2v 2 r , 2 s , 2 u , 2 v , and use 4 z 2 = 4 z ( m o d 8 ) 4z^2 = 4z \pmod8 4 z 2 = 4 z ( mod 8 ) . The latter reduction is valid for every z ∈ Z 2 z \in\mathbb{Z}_2 z ∈ Z 2 , since z ( z − 1 ) z(z-1) z ( z − 1 ) is even. For example the coefficients of y 0 y^0 y 0 in δ 2 \delta_2 δ 2 at x x x -degrees 0 , … , 6 0,\ldots,6 0 , … , 6 are
0 , 2 a , 3 a 2 + 2 d − a , a 3 + 6 a d − a 2 , 0,\qquad2a,\qquad3a^2+2d-a,\qquad a^3+6ad-a^2, 0 , 2 a , 3 a 2 + 2 d − a , a 3 + 6 a d − a 2 ,
3 a 2 d + 3 d 2 + 6 b c − a d − d , a ( 3 d 2 + 6 b c − d ) , d 3 + 6 b c d − d 2 , 3a^2d+3d^2+6bc-ad-d,\qquad a(3d^2+6bc-d),\qquad d^3+6bcd-d^2, 3 a 2 d + 3 d 2 + 6 b c − a d − d , a ( 3 d 2 + 6 b c − d ) , d 3 + 6 b c d − d 2 ,
which give the second line of the table by the indicated coefficient extraction. The third-level entries follow by substituting δ 2 = f 3 − f σ ( f ) \delta_2=f^3-f\sigma(f) δ 2 = f 3 − f σ ( f ) into the displayed formula for δ 3 \delta_3 δ 3 and the same degree and square reductions.
Let P 0 = ( 0 − 1 1 0 ) P_0=\begin{pmatrix}0&-1\\1&0\end{pmatrix} P 0 = ( 0 1 − 1 0 ) and J 0 ′ = ( 1 1 0 1 ) J'_0=\begin{pmatrix}1&1\\0&1\end{pmatrix} J 0 ′ = ( 1 0 1 1 ) . At precision 3, the canonical zero-window history with H 0 = z = ( z 0 , z 1 ) H_0=z=(z_0,z_1) H 0 = z = ( z 0 , z 1 ) has
H 1 = z P 0 ( m o d 4 ) , H 2 = z ( m o d 2 ) . H_1=zP_0\pmod4,\qquad H_2=z\pmod2. H 1 = z P 0 ( mod 4 ) , H 2 = z ( mod 2 ) .
Indeed compatibility says H 1 = H 2 G 1 , 0 ( m o d 2 ) H_1=H_2G_{1,0}\pmod2 H 1 = H 2 G 1 , 0 ( mod 2 ) and H 0 = H 1 G 1 , 0 + H 2 G 2 , 0 ( m o d 4 ) H_0=H_1G_{1,0}+H_2G_{2,0}\pmod4 H 0 = H 1 G 1 , 0 + H 2 G 2 , 0 ( mod 4 ) . Eliminating H 2 H_2 H 2 makes the latter matrix G 1 , 0 + G 1 , 0 − 1 G 2 , 0 = P 0 ( m o d 4 ) G_{1,0}+G_{1,0}^{-1}G_{2,0}=P_0\pmod4 G 1 , 0 + G 1 , 0 − 1 G 2 , 0 = P 0 ( mod 4 ) , whose inverse is itself. The transition formula and the table now give the following four explicit matrix identities:
A = G 1 , 0 + P 0 G 2 , 0 + G 3 , 0 = ( 1 5 0 3 ) , A=G_{1,0}+P_0G_{2,0}+G_{3,0}=\begin{pmatrix}1&5\\0&3\end{pmatrix}, A = G 1 , 0 + P 0 G 2 , 0 + G 3 , 0 = ( 1 0 5 3 ) ,
C 1 = G 1 , 1 + P 0 G 2 , 2 + G 3 , 4 = ( 7 6 3 1 ) , C_1=G_{1,1}+P_0G_{2,2}+G_{3,4}=\begin{pmatrix}7&6\\3&1\end{pmatrix}, C 1 = G 1 , 1 + P 0 G 2 , 2 + G 3 , 4 = ( 7 3 6 1 ) ,
C 2 = C 1 G 1 , 0 + G 2 , 1 + J 0 ′ G 3 , 2 = ( 1 5 7 6 ) , C_2=C_1G_{1,0}+G_{2,1}+J'_0G_{3,2}=\begin{pmatrix}1&5\\7&6\end{pmatrix}, C 2 = C 1 G 1 , 0 + G 2 , 1 + J 0 ′ G 3 , 2 = ( 1 7 5 6 ) ,
C 3 = C 2 G 1 , 0 + C 1 G 2 , 0 + G 3 , 1 = ( 5 0 7 5 ) . C_3=C_2G_{1,0}+C_1G_{2,0}+G_{3,1}=\begin{pmatrix}5&0\\7&5\end{pmatrix}. C 3 = C 2 G 1 , 0 + C 1 G 2 , 0 + G 3 , 1 = ( 5 7 0 5 ) .
Here C 1 , C 2 , C 3 C_1,C_2,C_3 C 1 , C 2 , C 3 are the successive matrices after the digits 1, 0, 0; the indices of their ghost terms are respectively (1, 2, 4), (0, 1, 2) and (0, 0, 1). Every parameter term cancels in the four displayed products. Since A 2 = ( 1 4 0 1 ) A^2=\begin{pmatrix}1&4\\0&1\end{pmatrix} A 2 = ( 1 0 4 1 ) and A 4 = I ( m o d 8 ) A^4=I\pmod8 A 4 = I ( mod 8 ) , the remaining fourteen zeros in 10 16 10^{16} 1 0 16 give
A = ( 1 5 0 3 ) , B = C 3 A 14 = ( 5 4 7 1 ) ( m o d 8 ) . (15) A=\begin{pmatrix}1&5\\0&3\end{pmatrix},\qquad B=C_3A^{14}=\begin{pmatrix}5&4\\7&1\end{pmatrix}\pmod8. \tag*{(15)} A = ( 1 0 5 3 ) , B = C 3 A 14 = ( 5 7 4 1 ) ( mod 8 ) . (15)
Modulo 4 the orbit of ( 1 , 0 ) (1,0) ( 1 , 0 ) contains all twelve primitive vectors. The following table gives an actual producing word for each vector; e e e denotes the empty word.
Table 3.
Put X = A 2 X=A^2 X = A 2 and Y = A B 4 A Y=AB^4A Y = A B 4 A . Multiplication in (15) yields
X = I + 4 E ( m o d 8 ) , Y = I + 4 D ( m o d 8 ) , E = ( 0 1 0 0 ) , D = ( 1 0 1 1 ) . X=I+4E\pmod8,\qquad Y=I+4D\pmod8,\qquad E=\begin{pmatrix}0&1\\0&0\end{pmatrix},\qquad D=\begin{pmatrix}1&0\\1&1\end{pmatrix}. X = I + 4 E ( mod 8 ) , Y = I + 4 D ( mod 8 ) , E = ( 0 0 1 0 ) , D = ( 1 1 0 1 ) .
Repeated squaring gives, for m ≥ 2 m\ge2 m ≥ 2 ,
X 2 m − 2 = I + 2 m E ( m o d 2 m + 1 ) , Y 2 m − 2 = I + 2 m D ( m o d 2 m + 1 ) . X^{2^{m-2}}=I+2^mE\pmod{2^{m+1}},\qquad Y^{2^{m-2}}=I+2^mD\pmod{2^{m+1}}. X 2 m − 2 = I + 2 m E ( mod 2 m + 1 ) , Y 2 m − 2 = I + 2 m D ( mod 2 m + 1 ) .
Indeed a perturbation divisible by 4 acquires exactly one extra factor 2 upon squaring at the precision in question; its square has higher valuation. For a row z = ( 1 , 0 ) ( m o d 2 ) z=(1,0)\pmod2 z = ( 1 , 0 ) ( mod 2 ) , the two rows z E = ( 0 , 1 ) zE=(0,1) z E = ( 0 , 1 ) and z D = ( 1 , 0 ) zD=(1,0) z D = ( 1 , 0 ) modulo 2 are independent. Thus these powers correct either coordinate bit at level m m m without changing lower bits. Induction shows that the orbit of ( 1 , 0 ) (1,0) ( 1 , 0 ) contains every vector congruent to ( 1 , 0 ) (1,0) ( 1 , 0 ) modulo 4. For any primitive target z ∗ z^* z ∗ , choose g g g from the table with ( 1 , 0 ) g = z ∗ ( m o d 4 ) (1,0)g=z^*\pmod4 ( 1 , 0 ) g = z ∗ ( mod 4 ) . The vector z ∗ g − 1 z^*g^{-1} z ∗ g − 1 is congruent to ( 1 , 0 ) (1,0) ( 1 , 0 ) modulo 4, so it is reachable by the lifting argument; applying g g g reaches z ∗ z^* z ∗ . This proves transitivity on primitive rows modulo every 2 α 2^\alpha 2 α (and modulo 2 by reduction).
There is an actual primitive starting history at the zero window. The word 10 leaves the exceptional language, so all subsequent zero-level numerators lie in W 0 W_0 W 0 . Modulo 2, the first digit gives x + x y x+xy x + x y , and digit 0 sends this to x x x : it kills x y xy x y and fixes x x x . Each further zero digit fixes x x x modulo 2. Therefore 10 α + 1 10^{\alpha+1} 1 0 α + 1 has a wholly ordinary history at the zero window with primitive H 0 H_0 H 0 . Applying actual return words realizes the whole primitive set there by transitivity. To reach another prescribed window, append its α − 1 \alpha-1 α − 1 digits. Each transition is invertible on H 0 H_0 H 0 by Lemma 7.6, so the image of the primitive set is exactly the primitive set in the new window. Uniqueness of the compatible history with a given H 0 H_0 H 0 proves the claimed reachability.
Finally decoding is linear on each window. Modulo 2 it is CT y ( ( u + v x ) / f ) \operatorname{CT}_y((u+vx)/f) CT y (( u + v x ) / f ) , whose constant and x x x coefficients are u u u and v − a u v-au v − a u . Their matrix has determinant one. Hence these same two coefficient functionals at precision α \alpha α give an invertible matrix over Z / 2 α \mathbb{Z}/2^\alpha Z / 2 α , and the decoded series are distinct. There are 4 α − 4 α − 1 = 3 ⋅ 4 α − 1 4^\alpha-4^{\alpha-1}=3\cdot4^{\alpha-1} 4 α − 4 α − 1 = 3 ⋅ 4 α − 1 primitive coordinate rows.
Universality of the normalized window frames
The following argument proves universality for arbitrary odd 2-adic parameters, rather than by checking finitely many residue classes.
Proposition 7.8. Let E w ( θ ) E_w^{(\theta)} E w ( θ ) denote (7.3) for an odd quadruple θ = ( a , b , c , d ) \theta=(a,b,c,d) θ = ( a , b , c , d ) , and let E w ( 1 ) E_w^{(1)} E w ( 1 ) denote it for ( 1 , 1 , 1 , 1 ) (1,1,1,1) ( 1 , 1 , 1 , 1 ) . If ∣ w ∣ = k |w|=k ∣ w ∣ = k , then
E w ( θ ) ≡ E w ( 1 ) ( m o d 2 k + 1 ) . E_w^{(\theta)}\equiv E_w^{(1)}\pmod{2^{k+1}}. E w ( θ ) ≡ E w ( 1 ) ( mod 2 k + 1 ) . In particular, the full-window frames at precision α \alpha α , and their planes, are independent of θ \theta θ .
Proof. Put q = 1 + x + x 2 q=1+x+x^2 q = 1 + x + x 2 , D = q 2 − 4 x 4 D=q^2-4x^4 D = q 2 − 4 x 4 , B = D − 1 / 2 B=D^{-1/2} B = D − 1/2 and
G = ( ( 1 + x ) B , x B ) , G θ = ( ( 1 + a x ) B θ , x B θ ) . G=((1+x)B,xB),\qquad G_\theta=((1+ax)B_\theta,xB_\theta). G = (( 1 + x ) B , x B ) , G θ = (( 1 + a x ) B θ , x B θ ) .
These rows have the same section planes as ( B , x B ) (B,xB) ( B , x B ) and ( B θ , x B θ ) (B_\theta,xB_\theta) ( B θ , x B θ ) , respectively. We work in the 2-adic completion of Z 2 [ x , q − 1 ] \mathbb{Z}_2[x,q^{-1}] Z 2 [ x , q − 1 ] ; all rational denominators have constant term one. For m ≥ 0 m\geq0 m ≥ 0 let
R m = { N / q m : deg N ≤ 2 m + 1 } , R m 0 = { N / q m : deg N ≤ 2 m } , \mathcal{R}_m=\{N/q^m:\deg N\leq2m+1\},\qquad\mathcal{R}_m^0=\{N/q^m:\deg N\leq2m\}, R m = { N / q m : deg N ≤ 2 m + 1 } , R m 0 = { N / q m : deg N ≤ 2 m } ,
with integral polynomial numerators. They are nested in m m m , and R m 0 R n ⊂ R m + n \mathcal{R}_m^0\mathcal{R}_n\subset\mathcal{R}_{m+n} R m 0 R n ⊂ R m + n . Congruences in this completed rational ring imply coefficientwise series congruences.
Write P θ = q + 2 H P_\theta=q+2H P θ = q + 2 H , b c = 1 + 2 v 0 bc=1+2v_0 b c = 1 + 2 v 0 , and Z = q H + H 2 − b c x 4 Z=qH+H^2-bcx^4 Z = q H + H 2 − b c x 4 . Thus D θ = P θ 2 − 4 b c x 4 = q 2 + 4 Z D_\theta=P_\theta^2-4bcx^4=q^2+4Z D θ = P θ 2 − 4 b c x 4 = q 2 + 4 Z . Set
T = x 4 ( v 0 − 2 H / q − 2 H 2 / q 2 ) D θ . T=\frac{x^4(v_0-2H/q-2H^2/q^2)}{D_\theta}. T = D θ x 4 ( v 0 − 2 H / q − 2 H 2 / q 2 ) .
A direct calculation gives
( B θ B ) 2 = ( q P θ ) 2 ( 1 + 8 T ) , B θ B = ( q P θ ) ∑ r ≥ 0 ( 1 / 2 r ) 8 r T r . (16) \left(\frac{B_\theta}{B}\right)^2=\left(\frac{q}{P_\theta}\right)^2(1+8T),\qquad\frac{B_\theta}{B}=\left(\frac{q}{P_\theta}\right)\sum_{r\geq0}\binom{1/2}{r}8^rT^r. \tag*{(16)} ( B B θ ) 2 = ( P θ q ) 2 ( 1 + 8 T ) , B B θ = ( P θ q ) r ≥ 0 ∑ ( r 1/2 ) 8 r T r . (16)
The second identity uses the square root with constant term one. For r ≥ 1 r\geq1 r ≥ 1 ,
v 2 ( 8 r ( 1 / 2 r ) ) = r + s 2 ( r ) , v_2\left(8^r\binom{1/2}{r}\right)=r+s_2(r), v 2 ( 8 r ( r 1/2 ) ) = r + s 2 ( r ) ,
where s 2 ( r ) s_2(r) s 2 ( r ) counts the binary ones of r r r . Indeed ( 1 / 2 r ) = ( − 1 ) r − 1 ( 2 r ) ! / ( 4 r r ( 2 r − 1 ) ) \binom{1/2}{r}=(-1)^{r-1}(2r)!/(4^r r(2r-1)) ( r 1/2 ) = ( − 1 ) r − 1 ( 2 r )! / ( 4 r r ( 2 r − 1 )) , and counting multiples of powers of two gives v 2 ( n ! ) = n − s 2 ( n ) v_2(n!)=n-s_2(n) v 2 ( n !) = n − s 2 ( n ) . This also proves convergence of (7.6).
The expansion
D θ − 1 = q − 2 ∑ h ≥ 0 ( − 4 Z / q 2 ) h D_\theta^{-1}=q^{-2}\sum_{h\geq0}(-4Z/q^2)^h D θ − 1 = q − 2 h ≥ 0 ∑ ( − 4 Z / q 2 ) h
shows that T ∈ ∑ j ≥ 0 2 2 j R 2 j + 2 0 T\in\sum_{j\geq0}2^{2j}\mathcal{R}_{2j+2}^0 T ∈ ∑ j ≥ 0 2 2 j R 2 j + 2 0 . To check the indices, the three terms in its numerator have (precision, pole) pairs ( 0 , 2 ) (0,2) ( 0 , 2 ) , ( 1 , 3 ) (1,3) ( 1 , 3 ) and ( 1 , 4 ) (1,4) ( 1 , 4 ) , while the h h h th geometric term adds ( 2 h , 2 h ) (2h,2h) ( 2 h , 2 h ) . Thus
T r ∈ ∑ j ≥ 0 2 2 j R 2 j + 2 r 0 . T^r\in\sum_{j\geq0}2^{2j}\mathcal{R}_{2j+2r}^0. T r ∈ j ≥ 0 ∑ 2 2 j R 2 j + 2 r 0 .
A nonconstant square-root term of total precision n n n in (7.6) consequently has pole at most 2 n − 2 s 2 ( r ) ≤ 2 n − 2 2n-2s_2(r)\leq2n-2 2 n − 2 s 2 ( r ) ≤ 2 n − 2 . Combining this with q / P θ = ∑ h ≥ 0 ( − 2 H / q ) h q/P_\theta=\sum_{h\geq0}(-2H/q)^h q / P θ = ∑ h ≥ 0 ( − 2 H / q ) h gives
B θ / B − 1 ∈ 2 R 1 0 + ∑ n ≥ 2 2 n R 2 n − 2 0 , ( G θ − G ) / B ∈ 2 R 1 + ∑ n ≥ 2 2 n R 2 n − 2 . B_\theta/B-1\in2\mathcal{R}_1^0+\sum_{n\geq2}2^n\mathcal{R}_{2n-2}^0,\qquad
(G_\theta-G)/B\in2\mathcal{R}_1+\sum_{n\geq2}2^n\mathcal{R}_{2n-2}. B θ / B − 1 ∈ 2 R 1 0 + n ≥ 2 ∑ 2 n R 2 n − 2 0 , ( G θ − G ) / B ∈ 2 R 1 + n ≥ 2 ∑ 2 n R 2 n − 2 .
For the second inclusion multiply by ( 1 + x , x ) (1+x,x) ( 1 + x , x ) and add ( ( a − 1 ) x B θ / B , 0 ) ((a-1)xB_{\theta}/B,0) (( a − 1 ) x B θ / B , 0 ) ; the extra factor a − 1 a-1 a − 1 is even and obeys the same bounds.
Let H j = x j ∂ x [ j ] H_j=x^j\partial_x^{[j]} H j = x j ∂ x [ j ] , where ∂ x [ j ] x n = ( n j ) x n − j \partial_x^{[j]}x^n=\binom{n}{j}x^{n-j} ∂ x [ j ] x n = ( j n ) x n − j is the j j j th Hasse derivative, and put V j = ( H j G ) / B V_j=(H_jG)/B V j = ( H j G ) / B . Differentiation shows
V j = N j D j , deg N j ≤ 4 j + 1 , V_j=\frac{N_j}{D^j},\qquad\deg N_j\leq4j+1, V j = D j N j , deg N j ≤ 4 j + 1 ,
with integral polynomial numerators. For integrality, the series H j G H_jG H j G , B − 1 B^{-1} B − 1 and D j D^j D j are integral, and their product is the polynomial N j N_j N j ; no division by j ! j! j ! in Z 2 \mathbb{Z}_2 Z 2 is used. Modulo 2 2 2 , the columns of V j V_j V j have pole at most j j j at q q q . At a root ρ \rho ρ of q q q in F 4 \mathbb{F}_4 F 4 , their leading principal parts for j > 0 j>0 j > 0 are a common nonzero multiple of ( 1 + ρ , ρ ) (1+\rho,\rho) ( 1 + ρ , ρ ) . The latter two entries are independent over F 2 \mathbb{F}_2 F 2 : their sum is 1 1 1 , and neither is in F 2 \mathbb{F}_2 F 2 . Hasse differentiation of a simple pole has a nonzero leading term at every order. At j = 0 j=0 j = 0 the polynomials 1 + x , x 1+x,x 1 + x , x are a basis. It follows successively by pole order that the 2 m + 2 2m+2 2 m + 2 columns of V 0 , … , V m V_0,\ldots,V_m V 0 , … , V m are a basis of R m / 2 R m \mathcal{R}_m/2\mathcal{R}_m R m /2 R m .
For j > 0 j>0 j > 0 choose any lift v j ∈ R j v_j\in\mathcal{R}_j v j ∈ R j of V j V_j V j mod 2 2 2 . Since
V j = N j q 2 j ∑ h ≥ 0 ( j + h − 1 h ) ( 4 x 4 q 2 ) h , V_j=\frac{N_j}{q^{2j}}\sum_{h\geq0}\binom{j+h-1}{h}\left(\frac{4x^4}{q^2}\right)^h, V j = q 2 j N j h ≥ 0 ∑ ( h j + h − 1 ) ( q 2 4 x 4 ) h ,
the leading difference has even numerator of pole at most 2 j 2j 2 j , and the remaining terms have precision at least 2 h 2h 2 h . Therefore
V j − v j ∈ ∑ t ≥ 1 2 t R 2 j + 2 ⌊ t / 2 ⌋ . (17) V_j-v_j\in\sum_{t\geq1}2^t\mathcal{R}_{2j+2\lfloor t/2\rfloor}. \tag*{(17)} V j − v j ∈ t ≥ 1 ∑ 2 t R 2 j + 2 ⌊ t /2 ⌋ . (17)
For j = 0 j=0 j = 0 take v 0 = V 0 v_0=V_0 v 0 = V 0 , so that its tail is zero.
Define filtered tails
D ≥ n = ∑ m ≥ n 2 m R 2 m − 1 ( n ≥ 1 ) . D_{\geq n}=\sum_{m\geq n}2^m\mathcal{R}_{2m-1}\qquad(n\geq1). D ≥ n = m ≥ n ∑ 2 m R 2 m − 1 ( n ≥ 1 ) .
The initial deformation lies in D ≥ 1 D_{\geq1} D ≥ 1 , because 2 m − 2 ≤ 2 m − 1 2m-2\leq2^{m-1} 2 m − 2 ≤ 2 m − 1 for m ≥ 2 m\geq2 m ≥ 2 . At stage n n n , expand its coefficient of 2 n 2^n 2 n modulo 2 2 2 in the basis given by V 0 , … , V 2 n − 1 V_0,\ldots,V_{2^{n-1}} V 0 , … , V 2 n − 1 , and subtract the corresponding 2 n 2^n 2 n multiples with constant binary matrix coefficients. The residual lies in D ≥ n + 1 D_{\geq n+1} D ≥ n + 1 : the leading difference is even, and (7.7) fits the later bounds since, for j ≤ 2 n − 1 j\leq2^{n-1} j ≤ 2 n − 1 and t ≥ 1 t\geq1 t ≥ 1 ,
2 j + 2 ⌊ t / 2 ⌋ ≤ 2 n + 2 ⌊ t / 2 ⌋ ≤ 2 n + t − 1 . 2j+2\lfloor t/2\rfloor\leq2^n+2\lfloor t/2\rfloor\leq2^{n+t-1}. 2 j + 2 ⌊ t /2 ⌋ ≤ 2 n + 2 ⌊ t /2 ⌋ ≤ 2 n + t − 1 .
For t = 1 t=1 t = 1 this is equality at the largest j j j ; for t ≥ 2 t\geq2 t ≥ 2 it follows from 2 ⌊ t / 2 ⌋ ≤ 2 t − 1 − 1 2\lfloor t/2\rfloor\leq2^{t-1}-1 2 ⌊ t /2 ⌋ ≤ 2 t − 1 − 1 . Iteration yields convergent constant 2 × 2 2\times2 2 × 2 matrices A j A_j A j such that
G θ = ∑ j ≥ 0 ( H j G ) A j , A 0 ≡ I ( m o d 2 ) , v 2 ( A j ) ≥ 1 + ⌈ log 2 j ⌉ ( j ≥ 1 ) . (18) G_{\theta}=\sum_{j\geq0}(H_jG)A_j,\qquad A_0\equiv I\pmod2,\qquad v_2(A_j)\geq1+\lceil\log_2j\rceil\quad(j\geq1). \tag*{(18)} G θ = j ≥ 0 ∑ ( H j G ) A j , A 0 ≡ I ( mod 2 ) , v 2 ( A j ) ≥ 1 + ⌈ log 2 j ⌉ ( j ≥ 1 ) . (18)
The matrices tend to zero 2 2 2 -adically and the V j V_j V j are integral, so ∑ j V j A j = G θ / B \sum_jV_jA_j=G_{\theta}/B ∑ j V j A j = G θ / B converges in the rational completion. Multiplication by B B B gives the displayed identity in Z 2 [ [ x ] ] \mathbb{Z}_2[[x]] Z 2 [[ x ]] , with convergence in the 2 2 2 -adic topology.
For Q = 2 k Q=2^k Q = 2 k , coefficient extraction gives the exact operator identity
Λ Q , v H j = ( Q x ∂ x + v j ) Λ Q , v . \Lambda_{Q,v}H_j=\binom{Qx\partial_x+v}{j}\Lambda_{Q,v}. Λ Q , v H j = ( j Q x ∂ x + v ) Λ Q , v .
Vandermonde’s identity gives, for n ≥ 0 n\geq0 n ≥ 0 ,
( Q n + v j ) − ( v j ) = ∑ i = 1 j ( Q n i ) ( v j − i ) . \binom{Qn+v}{j}-\binom{v}{j}=\sum_{i=1}^{j}\binom{Qn}{i}\binom{v}{j-i}. ( j Q n + v ) − ( j v ) = i = 1 ∑ j ( i Q n ) ( j − i v ) .
If ⌊ log 2 j ⌋ ≤ k \lfloor\log_{2} j \rfloor\le k ⌊ log 2 j ⌋ ≤ k , each summand is divisible by 2 k − ⌊ log 2 j ⌋ 2^{k-\lfloor\log_{2}j \rfloor} 2 k − ⌊ l o g 2 j ⌋ , since ( Q n i ) = ( Q n / i ( Q n − 1 ) / ( i − 1 ) ) \binom{Qn}{i}=\binom{Qn/i}{(Qn-1)/(i-1)} ( i Q n ) = ( ( Q n − 1 ) / ( i − 1 ) Q n / i ) ; for n = 0 n=0 n = 0 it is zero. If ⌊ log 2 j ⌋ > k \lfloor\log_{2}j \rfloor> k ⌊ log 2 j ⌋ > k , integrality suffices. After multiplication by A j A_j A j in (7.8), the nonconstant-in-n n n part thus vanishes modulo 2 k + 1 2^{k+1} 2 k + 1 . We obtain
Λ Q , v G θ ≡ ( Λ Q , v G ) C v ( m o d 2 k + 1 ) , C v = ∑ j = 0 v ( v j ) A j ≡ I ( m o d 2 ) . \Lambda_{Q,v}G_{\theta}\equiv(\Lambda_{Q,v}G)C_v\pmod{2^{k+1}},\qquad C_v=\sum_{j=0}^{v}\binom{v}{j}A_j\equiv I\pmod{2}. Λ Q , v G θ ≡ ( Λ Q , v G ) C v ( mod 2 k + 1 ) , C v = j = 0 ∑ v ( j v ) A j ≡ I ( mod 2 ) .
The constant matrix C v C_v C v is invertible and disappears on first-two-coefficient normalization. The proposition follows. □
Weak contact and the energy bound
For two windows of length α − 1 \alpha-1 α − 1 , define their clipped contact by
r α ( w , v ) = max { 0 ≤ s ≤ α : E w ≡ E v ( m o d 2 s ) } . r_{\alpha}(w,v)=\max\{0\le s\le\alpha:E_w\equiv E_v\pmod{2^s}\}. r α ( w , v ) = max { 0 ≤ s ≤ α : E w ≡ E v ( mod 2 s )} .
For distinct windows let ℓ ( w , v ) \ell(w,v) ℓ ( w , v ) be the length of their longest common terminal suffix; this counts the last applied digits.
Lemma 7.9. For distinct full windows,
r α ( w , v ) ≤ min { α , ℓ ( w , v ) + 3 } . r_{\alpha}(w,v)\le\min\{\alpha,\ell(w,v)+3\}. r α ( w , v ) ≤ min { α , ℓ ( w , v ) + 3 } . For every s ≥ 1 s\ge1 s ≥ 1 , an equivalence class of length-( s − 1 ) (s-1) ( s − 1 ) windows with equal normalized frames modulo 2 s 2^s 2 s has at most four members.
Proof. By Proposition 7.8 it suffices to use the all-one parameters. Proposition B.6 in Appendix B proves the asserted contact bound, by a normal rational lattice and a ten-dimensional quotient on which the normalized difference cannot vanish. Its proof includes the finite matrices, seeds, invariant quotient and sufficient-precision justification. For the second statement let M s M_s M s be the largest equivalence-class size. If s ≤ 3 s\le3 s ≤ 3 there are at most 2 s − 1 ≤ 4 2^{s-1}\le4 2 s − 1 ≤ 4 windows in total. If s ≥ 4 s\ge4 s ≥ 4 , any two different members of a class have contact s s s , so their common terminal suffix has length at least s − 3 s-3 s − 3 . Once a member is fixed, only its two older digits can vary, leaving at most four members. □
Proof of Theorem 7.1. The upper bound is Proposition 7.4. For the lower bound, put d = α − 1 d=\alpha-1 d = α − 1 , n = 2 d n=2^d n = 2 d , and let P w P_w P w be the primitive part of G α , w G_{\alpha,w} G α , w , namely E w z E_wz E w z for coordinate columns z ∈ R α 2 α z\in R_{\alpha}^{2^\alpha} z ∈ R α 2 α with at least one odd entry. By Lemma 7.7, all these states belong to the kernel of F F F mod 2 α 2^\alpha 2 α , and
∣ P w ∣ = 4 α − 4 α − 1 = 3 ⋅ 4 α − 1 . |P_w|=4^\alpha-4^{\alpha-1}=3\cdot4^{\alpha-1}. ∣ P w ∣ = 4 α − 4 α − 1 = 3 ⋅ 4 α − 1 .
If a state belongs to both planes, its coordinate column in both is the same, since the first two frame rows are the identity. Hence ( E w − E v ) z = 0 (E_w-E_v)z=0 ( E w − E v ) z = 0 . If r = r α ( w , v ) < α r=r_\alpha(w,v)<\alpha r = r α ( w , v ) < α , a row attaining the smallest valuation gives one congruence a z 0 + b z 1 = 0 ( m o d 2 α ) az_0+bz_1=0\pmod{2^\alpha} a z 0 + b z 1 = 0 ( mod 2 α ) with min ( v 2 ( a ) , v 2 ( b ) ) = r \min(v_2(a),v_2(b))=r min ( v 2 ( a ) , v 2 ( b )) = r . Dividing by 2 r 2^r 2 r leaves a unit coefficient modulo 2 α − r 2^{\alpha-r} 2 α − r , so there are exactly 2 α + r 2^{\alpha+r} 2 α + r solutions in R α 2 R_\alpha^2 R α 2 . If r = α r=\alpha r = α , the same bound is the full-plane size 4 α 4^\alpha 4 α . Consequently
∣ P w ∩ P v ∣ ≤ 2 α + r α ( w , v ) . (19) |P_w\cap P_v|\le2^{\alpha+r_\alpha(w,v)}. \tag*{(19)} ∣ P w ∩ P v ∣ ≤ 2 α + r α ( w , v ) . (19)
At precision s s s , let m s , 1 , … , m s , t s m_{s,1},\ldots,m_{s,t_s} m s , 1 , … , m s , t s be the sizes of the frame equivalence classes among length-( s − 1 ) (s-1) ( s − 1 ) windows; their sum is 2 s − 1 2^{s-1} 2 s − 1 and their maximum is M s M_s M s . Coherence (7.4) gives the exact ordered-pair count
Q s : = # { ( w , v ) : r α ( w , v ) ≥ s } = 2 2 ( α − s ) ∑ j m s , j 2 ≤ M s 2 2 α − s − 1 . Q_s:=\#\{(w,v):r_\alpha(w,v)\ge s\}=2^{2(\alpha-s)}\sum_jm_{s,j}^2\le M_s2^{2\alpha-s-1}. Q s := # {( w , v ) : r α ( w , v ) ≥ s } = 2 2 ( α − s ) j ∑ m s , j 2 ≤ M s 2 2 α − s − 1 .
Since 2 r = 1 + ∑ s = 1 α 2 s − 1 1 r ≥ s 2^r=1+\sum_{s=1}^{\alpha}2^{s-1}\mathbf{1}_{r\ge s} 2 r = 1 + ∑ s = 1 α 2 s − 1 1 r ≥ s ,
∑ w , v 2 r α ( w , v ) ≤ n 2 ( 1 + ∑ s = 1 α M s ) . (20) \sum_{w,v}2^{r_\alpha(w,v)}\le n^2\left(1+\sum_{s=1}^{\alpha}M_s\right). \tag*{(20)} w , v ∑ 2 r α ( w , v ) ≤ n 2 ( 1 + s = 1 ∑ α M s ) . (20)
Let h ( g ) = # { w : g ∈ P w } h(g)=\#\{w:g\in\mathcal{P}_w\} h ( g ) = # { w : g ∈ P w } . Equations (19) –(20) imply
∑ g h ( g ) = n 3 4 α − 1 , ∑ g h ( g ) 2 ≤ 2 α n 2 ( 1 + ∑ s = 1 α M s ) . \sum_g h(g)=n\,3^{4\alpha-1},\qquad\sum_g h(g)^2\leq2^{\alpha}n^2\left(1+\sum_{s=1}^{\alpha}M_s\right). g ∑ h ( g ) = n 3 4 α − 1 , g ∑ h ( g ) 2 ≤ 2 α n 2 ( 1 + s = 1 ∑ α M s ) .
Cauchy–Schwarz therefore gives
K α ( F ) ≥ ∣ ⋃ w P w ∣ ≥ 9 2 3 α − 4 1 + ∑ s = 1 α M s . K_{\alpha}(F)\geq\left|\bigcup_w\mathcal{P}_w\right|\geq\frac{9\,2^{3\alpha-4}}{1+\sum_{s=1}^{\alpha}M_s}. K α ( F ) ≥ w ⋃ P w ≥ 1 + ∑ s = 1 α M s 9 2 3 α − 4 .
There is one empty window at s = 1 s=1 s = 1 , so M 1 = 1 M_1=1 M 1 = 1 ; Lemma 7.9 gives M s ≤ 4 M_s\leq4 M s ≤ 4 . Thus the denominator is at most 4 α − 2 4\alpha-2 4 α − 2 , yielding the stated lower bound, including α = 1 \alpha=1 α = 1 . Taking logarithms of the two bounds gives liminf at least 3 and limsup at most 3, since the logarithmic loss log 2 α \log_2\alpha log 2 α is o ( α ) o(\alpha) o ( α ) . □
Rank versus cardinality
Proposition 8.1. Let F ∈ Z [ [ x ] ] F\in\mathbb{Z}[[x]] F ∈ Z [[ x ]] be D D D -finite but not rational; for example, F F F may be an irrational algebraic series, or an irrational diagonal of a rational function with integer coefficients [15 Theorem 1] . Let p p p be a prime. For α ≥ 1 \alpha\geq1 α ≥ 1 let ρ α \rho_\alpha ρ α be the least r r r such that ker p ( F m o d p α ) \ker_p(F\bmod p^\alpha) ker p ( F mod p α ) is contained in a Z / p α \mathbb{Z}/p^\alpha Z / p α -submodule of ( Z / p α ) Z ≥ 0 (\mathbb{Z}/p^\alpha)^{\mathbb{Z}_{\geq0}} ( Z / p α ) Z ≥ 0 generated by r r r elements. Set ρ α = + ∞ \rho_\alpha=+\infty ρ α = + ∞ if there is no such finite r r r . Then ρ α → ∞ \rho_\alpha\to\infty ρ α → ∞ . In particular, the minimal number of states of a p p p -linear scheme generating F m o d p α F\bmod p^\alpha F mod p α tends to infinity with α \alpha α .
Proof. If ker p ( F m o d p α + 1 ) \ker_p(F\bmod p^{\alpha+1}) ker p ( F mod p α + 1 ) lies in a submodule generated by r r r elements, then reducing modulo p α p^\alpha p α shows that ker p ( F m o d p α ) \ker_p(F\bmod p^\alpha) ker p ( F mod p α ) lies in a submodule generated by the r r r reductions. Hence ρ α ≤ ρ α + 1 \rho_\alpha\leq\rho_{\alpha+1} ρ α ≤ ρ α + 1 . Suppose ρ α ≤ r \rho_\alpha\leq r ρ α ≤ r for all α \alpha α . Let w 0 , … , w r w_0,\ldots,w_r w 0 , … , w r be words and n 0 , … , n r n_0,\ldots,n_r n 0 , … , n r indices, and let μ = det [ x n l Λ w i F ] i , l ∈ Z \mu=\det\left[x^{n_l}\Lambda_{w_i}F\right]_{i,l}\in\mathbb{Z} μ = det [ x n l Λ w i F ] i , l ∈ Z . Modulo p α p^\alpha p α the rows lie in a module generated by r r r sequences e 1 , … , e r e_1,\ldots,e_r e 1 , … , e r , so the matrix factors as X Y XY X Y with X X X of size ( r + 1 ) × r (r+1)\times r ( r + 1 ) × r and Y Y Y of size r × ( r + 1 ) r\times(r+1) r × ( r + 1 ) over Z / p α \mathbb{Z}/p^\alpha Z / p α . Expand the determinant multilinearly in the rows of X Y XY X Y , each a combination of the r r r rows of Y Y Y . Every choice of r + 1 r+1 r + 1 rows repeats one, so alternation gives det ( X Y ) = 0 \det(XY)=0 det ( X Y ) = 0 . Hence p α ∣ μ p^\alpha\mid\mu p α ∣ μ for every α \alpha α , so μ = 0 \mu=0 μ = 0 . All ( r + 1 ) (r+1) ( r + 1 ) -minors of the matrix with rows Λ w F \Lambda_wF Λ w F therefore vanish. If r + 1 r+1 r + 1 such sequences were linearly independent, their coordinate functionals would span the dual of their ( r + 1 ) (r+1) ( r + 1 ) -dimensional span, since a sequence killed by every coordinate is zero. Choosing a basis of these functionals would give a nonzero finite minor, a contradiction. Thus the Q \mathbb{Q} Q -span of { Λ w F } \{\Lambda_wF\} { Λ w F } has dimension at most r r r . That is, F F F is p p p -regular over Q \mathbb{Q} Q . A p p p -regular power series is p p p -Mahler [3 ] ; here is the needed argument. Choose a Q \mathbb{Q} Q -basis H = ( H 1 = F , … , H d ) T H=(H_1=F,\ldots,H_d)^{T} H = ( H 1 = F , … , H d ) T of the Cartier span. The identity H i ( x ) = ∑ r = 0 p − 1 x r ( Λ r H i ) ( x p ) H_i(x)=\sum_{r=0}^{p-1}x^r(\Lambda_rH_i)(x^p) H i ( x ) = ∑ r = 0 p − 1 x r ( Λ r H i ) ( x p ) gives H ( x ) = A ( x ) H ( x p ) H(x)=A(x)H(x^p) H ( x ) = A ( x ) H ( x p ) for a polynomial matrix A A A . Iteration expresses each F ( x p i ) F(x^{p^i}) F ( x p i ) , 0 ≤ i ≤ d 0\leq i\leq d 0 ≤ i ≤ d , as a polynomial row vector times H ( x p d ) H(x^{p^d}) H ( x p d ) . Linear dependence of these d + 1 d+1 d + 1 row vectors over Q ( x ) \mathbb{Q}(x) Q ( x ) , followed by clearing denominators, gives ∑ i = 0 d a i ( x ) F ( x p i ) = 0 \sum_{i=0}^{d}a_i(x)F(x^{p^i})=0 ∑ i = 0 d a i ( x ) F ( x p i ) = 0 with a j ≠ 0 a_j\neq0 a j = 0 . Write a i ( x ) = ∑ v = 0 p j − 1 x v b i , v ( x p j ) a_i(x)=\sum_{v=0}^{p^j-1}x^v b_{i,v}(x^{p^j}) a i ( x ) = ∑ v = 0 p j − 1 x v b i , v ( x p j ) . Separating the residue classes of exponents gives ∑ i = 0 d b i , v ( x ) F ( x p i − j ) = 0 \sum_{i=0}^{d}b_{i,v}(x)F(x^{p^{i-j}})=0 ∑ i = 0 d b i , v ( x ) F ( x p i − j ) = 0 for every v v v . Choose v v v with b j , v ≠ 0 b_{j,v}\neq0 b j , v = 0 and omit any zero final coefficients. Since F ≠ 0 F\neq0 F = 0 , this is a p p p -Mahler equation with both endpoint coefficients nonzero. By a theorem of Bézivin (see [4 Theorem 1] ), a D D D -finite p p p -Mahler power series in C [ [ x ] ] \mathbb{C}[[x]] C [[ x ]] is a rational function. (For algebraic F F F one may instead use the rational–transcendental dichotomy for Mahler functions [16 Theorem 5.1.7, 4] .) A series with rational coefficients that is rational over C \mathbb{C} C is rational over Q \mathbb{Q} Q : normalize a complex denominator to have constant term 1. Its finitely many unknown coefficients solve rational linear equations expressing the vanishing of the product beyond the numerator degree. A finite set of rows spans this system (including its constant terms), and Gaussian elimination gives a rational solution. The resulting numerator is also rational. This contradicts the hypotheses on F F F . The last statement holds because the kernel of a sequence generated by a scheme with s s s states lies in a module generated by s s s elements. □
For example, the series of Corollary 1.4 with a = b = c = d = 1 a=b=c=d=1 a = b = c = d = 1 is irrational: its discriminant ( 1 + x + x 2 ) 2 − 4 x 4 = ( 1 + x − x 2 ) ( 1 + x + 3 x 2 ) (1+x+x^2)^2-4x^4=(1+x-x^2)(1+x+3x^2) ( 1 + x + x 2 ) 2 − 4 x 4 = ( 1 + x − x 2 ) ( 1 + x + 3 x 2 ) is not a square: both factors are squarefree, with discriminants 5 5 5 and − 11 -11 − 11 , and their difference is 4 x 2 4x^2 4 x 2 , so they have no common root. Their nonconstant product is squarefree and cannot be a square in Q ( x ) \mathbb{Q}(x) Q ( x ) . Therefore no description of its kernels by modules of bounded rank exists.
By contrast, for each window W W W the set H W \mathcal{H}_W H W is a group of order exactly p ∣ J ∣ α p^{|J|\alpha} p ∣ J ∣ α (Lemma 5.7), and Ψ ( ⋅ , W ) \Psi(\cdot,W) Ψ ( ⋅ , W ) is additive. The multivariate sections C w R C_wR C w R for ∣ w ∣ ≥ α |w|\geq\alpha ∣ w ∣ ≥ α thus lie in a union of p α − 1 p^{\alpha-1} p α − 1 subgroups of ( Z / p α ) [ [ z ] ] (\mathbb{Z}/p^\alpha)[[z]] ( Z / p α ) [[ z ]] , each of order at most p ∣ J ∣ α p^{|J|\alpha} p ∣ J ∣ α . Applying Diag gives the corresponding statement for the univariate kernel elements in ( Z / p α ) [ [ x ] ] (\mathbb{Z}/p^\alpha)[[x]] ( Z / p α ) [[ x ]] . We make no claim about the ranks of these groups. Even in the trivial case Q = N = 1 Q=N=1 Q = N = 1 (so J = { 0 } J=\{0\} J = { 0 } , δ 1 = 1 \delta_1=1 δ 1 = 1 and δ t = 0 \delta_t=0 δ t = 0 for t ≥ 2 t\geq2 t ≥ 2 ), the group H W \mathcal{H}_W H W for W = 0 α − 1 W=0^{\alpha-1} W = 0 α − 1 is cyclic of order p α p^\alpha p α , so its rank is ∣ J ∣ |J| ∣ J ∣ rather than ∣ J ∣ α |J|\alpha ∣ J ∣ α . The bound of Theorem 1.2 is a statement about cardinality.
Computations
The computations in this section are exact finite checks and do not establish asymptotic laws. The finite proof certificate for Theorem 7.1 is given separately in Appendix B.
The family with a = b = c = d = 1 a=b=c=d=1 a = b = c = d = 1
Let F F F be the Furstenberg series of P = x 2 y 2 + ( 1 + x + x 2 ) y + x 2 P=x^2y^2+(1+x+x^2)y+x^2 P = x 2 y 2 + ( 1 + x + x 2 ) y + x 2 and p = 2 p=2 p = 2 . Table 1 reports exact finite computations of its minimal kernel size; no asymptotic statement is inferred from them. All state identities, polynomial identities and partition refinements in these computations use exact arithmetic and full equality checks.
Table 1. Exact finite kernel computations for the quadratic family at p = 2 p=2 p = 2 . “Rational states” refers to the independent construction below, not the Rowland–Yassawi polynomial-state count. The history construction was run through α = 8 \alpha=8 α = 8 ; the rational construction through α = 9 \alpha=9 α = 9 . Theorem 7.1 determines the asymptotic exponent.
Two constructions. The first is the history automaton of Corollary 5.9. Our implementation builds its transition matrices by exact integer polynomial multiplication, uses full equality to identify states, and minimizes the complete graph. It gives the history-state counts and K α K_\alpha K α through α = 8 \alpha=8 α = 8 .
The independent rational construction does not use histories or the δ \delta δ -decomposition. Put f = 1 + x + x 2 ( y + 1 + y − 1 ) f=1+x+x^2(y+1+y^{-1}) f = 1 + x + x 2 ( y + 1 + y − 1 ) , N = y + x y + x 2 y + 2 x 2 y 2 N=y+xy+x^2y+2x^2y^2 N = y + x y + x 2 y + 2 x 2 y 2 , E = ( f 2 − σ ( f ) ) / 2 E=(f^2-\sigma(f))/2 E = ( f 2 − σ ( f )) /2 , and e = ⌈ α / 2 ⌉ e=\lceil\alpha/2\rceil e = ⌈ α /2 ⌉ . Represent a state by V / f α V/f^\alpha V / f α . The binomial expansion, modulo 2 α 2^\alpha 2 α , gives
C r ( V / f α ) = ∑ j = 0 α − 1 c j C r ( V f α − 2 j m o d 2 E j ) f e + j , c j = ( − 2 ) j ( e + j − 1 j ) . C_r(V/f^\alpha)=\sum_{j=0}^{\alpha-1}c_j\frac{C_r(Vf^{\alpha-2j}\bmod2E^j)}{f^{e+j}},\qquad c_j=(-2)^j\binom{e+j-1}{j}. C r ( V / f α ) = j = 0 ∑ α − 1 c j f e + j C r ( V f α − 2 j mod 2 E j ) , c j = ( − 2 ) j ( j e + j − 1 ) .
where here C r = C ∘ x − r C_r=C\circ x^{-r} C r = C ∘ x − r in the sheared variables of the appendix. Indeed, write f α = f 2 e / f α m o d 2 f^\alpha=f^{2e}/f^\alpha\bmod2 f α = f 2 e / f α mod 2 , expand ( σ ( f ) + 2 E ) − e (\sigma(f)+2E)^{-e} ( σ ( f ) + 2 E ) − e , discard terms divisible by 2 α 2^\alpha 2 α , and use the projection formula for C C C . To return to denominator f α f^\alpha f α , form
W = ∑ j = 0 α − 1 c j C r ( V f α m o d 2 E j ) f α − 1 − j W=\sum_{j=0}^{\alpha-1}c_jC_r(Vf^\alpha\bmod2E^j)f^{\alpha-1-j} W = j = 0 ∑ α − 1 c j C r ( V f α mod 2 E j ) f α − 1 − j
and compute a candidate polynomial V ′ V' V ′ with W = f e − 1 V ′ W=f^{e-1}V' W = f e − 1 V ′ modulo 2 α 2^\alpha 2 α . The complete polynomial residual, including coefficients outside the stored frame, is checked on every reached state; thus no truncated-series equality is assumed. The initial numerator is N f α − 1 Nf^{\alpha-1} N f α − 1 , and its two outgoing transitions are checked separately.
Both f f f and E E E are invariant under y ↦ y − 1 y\mapsto y^{-1} y ↦ y − 1 . The Cartier sections commute with this involution, and the central-row output kills every y j − y − j y^j-y^{-j} y j − y − j . Therefore one may quotient by their span, storing the sum of the coefficients at j j j and − j -j − j for j > 0 j>0 j > 0 , with the coefficient at j = 0 j=0 j = 0 unchanged. No division by 2 is used. Multiplication is always by a symmetric polynomial. A finite frame 0 ≤ i ≤ 2 α 0\leq i\leq2\alpha 0 ≤ i ≤ 2 α , 0 ≤ j ≤ ⌈ i / 2 ⌉ 0\leq j\leq\lceil i/2\rceil 0 ≤ j ≤ ⌈ i /2 ⌉ supplies the candidate numerator; its validity follows from the full residual checks, not from an assumed closure assertion. All residual constraints are obtained by exact polynomial operations; invertible row operations reduce them before the graph is built.
At α = 8 \alpha=8 α = 8 , the coefficients c 5 , c 6 , c 7 c_5,c_6,c_7 c 5 , c 6 , c 7 vanish modulo 256. The transition is consequently the polynomial
V ′ = ∑ j = 0 4 c j C r ( V E j ) f 4 − j ( m o d 256 ) , V'=\sum_{j=0}^{4}c_jC_r(VE^j)f^{4-j}\pmod{256}, V ′ = j = 0 ∑ 4 c j C r ( V E j ) f 4 − j ( mod 256 ) ,
and every full residual vanishes identically. A separate implementation independently recomputed all 2 ⋅ 89 2 2\cdot89^2 2 ⋅ 8 9 2 matrix entries and both initial transition vectors. The resulting complete graph has 2 991 449 reachable states. Independent sorting-based partition refinement gives 65 536, 2 369 812, 2 747 455, 2 747 455 classes, confirming K 8 K_8 K 8 without using a history graph.
The complete synchronous product with the independently rebuilt history graph at α = 8 \alpha=8 α = 8 has 8 587 651 reachable pairs, and all paired outputs agree. This checks equality for every input word, beyond the agreement of the two minimum counts.
At α = 9 \alpha=9 α = 9 the graph has 23 136 226 states, and all 1 503 854 625 reduced residual constraints passed; the minimum has 21 151 568 states. A separate implementation reconstructed both transition matrices, the complete residual matrices and initial vectors by direct polynomial division, and independently checked graph reachability and the minimum by sorting-based refinement. This is one mathematical construction with separately checked matrices and minimization, rather than agreement of two constructions at that precision. Building and minimizing the graphs took 6.70 seconds at α = 8 \alpha=8 α = 8 and 85.39 seconds at α = 9 \alpha=9 α = 9 on a single core; matrix generation is logged separately. A third, separately written minimizer reproduces both minima and the final partitions. Direct coefficient recurrence checks agree for n ≤ 4096 n\leq4096 n ≤ 4096 at every α ≤ 9 \alpha\leq9 α ≤ 9 , and for n ≤ 2 18 n\leq2^{18} n ≤ 2 18 at α = 8 \alpha=8 α = 8 and α = 9 \alpha=9 α = 9 ; these prefix checks are regression tests, not the justification of the kernel cardinalities. Source code, complete logs, matrix certificates, graph checksums and regeneration instructions are in anc/.
Other series
Table 2 collects further minimal kernel sizes. All 57 entries were reproduced by new exact code: the non-multinomial entries by an independent implementation of the history construction, and the multinomial row by a separate fixed-denominator polynomial Cartier construction. The latter construction also independently reproduces every Catalan and Example 13 entry, all listed entries at p = 3 , 5 p=3,5 p = 3 , 5 , and the first three entries in each other odd-parameter row.
Table 2. Minimal kernel sizes for further series; ∣ J ∣ = ∣ J p ( N F , Q F ) ∣ |J|=|J_p(N_F,Q_F)| ∣ J ∣ = ∣ J p ( N F , Q F ) ∣ is computed from the integral polynomials. Every listed entry was independently reproduced; the multinomial row is confirmed through α = 4 \alpha=4 α = 4 .
Support-set membership is certified by rational primal/dual or infeasibility certificates; floating-point linear programming is used only to propose candidates. Exact random tests in two and three variables at p = 2 , 3 , 5 p=2,3,5 p = 2 , 3 , 5 are logged separately.
Finite ratios and the proved exponent
For the family with a = b = c = d = 1 a=b=c=d=1 a = b = c = d = 1 , the ratio K α / 8 α K_\alpha/8^\alpha K α / 8 α equals 0.148, 0.169, 0.164 and 0.158 for α = 6 , 7 , 8 , 9 \alpha=6,7,8,9 α = 6 , 7 , 8 , 9 , and the ratios of successive history-state counts are 7.3, 7.6 and 7.8. Theorem 7.1 proves the exponent 3 for the entire odd family independently of these data. It does not establish convergence of the displayed ratios or identify a leading constant.
Open questions
The general exponent. Determine whether lim α α − 1 log p K α ( G ) \lim_{\alpha}\alpha^{-1}\log_p K_\alpha(G) lim α α − 1 log p K α ( G ) exists for a general rational diagonal, and compute it in terms of N N N , Q Q Q and p p p . For the odd quadratic family the limit is 3. Is the general slope governed by ∣ J ∣ |J| ∣ J ∣ , or by a unit-root part of the operators T t , v T_{t,v} T t , v ?
The leading constant. Does K α ( F ) / 8 α K_\alpha(F)/8^\alpha K α ( F ) / 8 α converge for the odd quadratic family? Determine its limit, if it exists, and its dependence on the parameters. Can the factor α \alpha α in the lower bound of Theorem 7.1 be removed?
Lower bounds beyond the odd family. Find sharp exponential lower bounds for other natural classes of rational diagonals, and identify conditions under which their upper and lower exponents agree.
Beyond a chosen unit-denominator representation. Treat algebraic series whose given annihilating polynomial has nonunit derivative at the origin, distinguishing failure of the Furstenberg conversion from nonexistence of another unit-denominator diagonal representation. The central trinomial coefficients, although singular in that algebraic presentation at p = 2 p=2 p = 2 [19 §10] , are already covered here:
∑ n ≥ 0 CT y ( y − 1 + 1 + y ) n x n = Diag 1 1 − x ( 1 + y + y 2 ) . \sum_{n\geq0}\operatorname{CT}_y(y^{-1}+1+y)^n x^n=\operatorname{Diag}\frac{1}{1-x(1+y+y^2)}. n ≥ 0 ∑ CT y ( y − 1 + 1 + y ) n x n = Diag 1 − x ( 1 + y + y 2 ) 1 .
Indeed, the diagonal coefficient is [ y n ] ( 1 + y + y 2 ) n [y^n](1+y+y^2)^n [ y n ] ( 1 + y + y 2 ) n . For this representation J = { 0 } J=\{0\} J = { 0 } : membership implies u 1 + η ≤ 1 − ϵ < 1 u_1+\eta\leq1-\epsilon<1 u 1 + η ≤ 1 − ϵ < 1 , hence u 1 = 0 u_1=0 u 1 = 0 , and u 2 + η ≤ 2 η u_2+\eta\leq2\eta u 2 + η ≤ 2 η , hence u 2 = 0 u_2=0 u 2 = 0 ; conversely 0 ∈ J 0\in J 0 ∈ J . Theorem 1.2 therefore gives K α ≤ p 2 α − 1 + ( p α − 1 ) / ( p − 1 ) K_\alpha\leq p^{2\alpha-1}+(p^\alpha-1)/(p-1) K α ≤ p 2 α − 1 + ( p α − 1 ) / ( p − 1 ) .
Sharper support sets. Find stable sets smaller than J J J , for instance by using the Hasse–Witt rank or the structure of the operators modulo p p p , and decide whether such refinements reach the true slope.
Uniformity. Prove bounds that are good simultaneously in α \alpha α , in the degrees and in p p p . Can the methods of [19 ] and of this paper be combined to give a bound that is never worse than either?
Structure. Describe the groups H W \mathcal{H}_W H W (their Smith forms and ranks) and the fibres of Ψ \Psi Ψ . How much of the gap between Table 1 and the bound comes from non-injectivity of Ψ \Psi Ψ , and how much from unreachable histories?
The zero orbit. Determine the exact preperiod of the observed sequences Λ 0 k F \Lambda_0^k F Λ 0 k F in Theorem A.1 . Extend the exact orbit formula to other polynomials, including non-ordinary ones.
Verification. Recompute K 9 K_9 K 9 by a second mathematical construction, extend the computations to α = 10 \alpha=10 α = 10 , broaden the literature search beyond the accessible sources, and obtain an independent human verification of the proofs.
The author directed the preparation and verification of this manuscript and selected the results included in it. Claude (Anthropic) assisted in verifying the zero-orbit proof, finding and proving the linear kernel bound for the two-variable family, independently checking the proof of Theorem 7.1 , searching the literature, performing computations, and drafting the text. ChatGPT (OpenAI), including Codex, was used to independently verify the proofs, extend the bound to diagonals in any number of variables, audit the literature, perform independent exact computations, and prepare the submission files. Under the author’s direction, ChatGPT was also used to derive the proof of Theorem 7.1 and to draft Section 7 and Appendix B; a separate ChatGPT session audited this proof independently. The author checked all results and takes full responsibility for the content of this paper.
Appendix A. The exact zero orbit for the quadratic family
In this appendix p = 2 p=2 p = 2 , a , b , c , d ∈ Z 2 a,b,c,d\in\mathbb{Z}_2 a , b , c , d ∈ Z 2 , and
P = b x 2 y 2 + P 1 ( x ) y + c x 2 , P 1 ( x ) = 1 + a x + d x 2 , f = P / y = 1 + a x + x 2 ( b y + d + c y − 1 ) , P=bx^2y^2+P_1(x)y+cx^2,\qquad P_1(x)=1+ax+dx^2,\qquad f=P/y=1+ax+x^2(by+d+cy^{-1}), P = b x 2 y 2 + P 1 ( x ) y + c x 2 , P 1 ( x ) = 1 + a x + d x 2 , f = P / y = 1 + a x + x 2 ( b y + d + c y − 1 ) ,
N = y ∂ y P = P 1 ( x ) y + 2 b x 2 y 2 , R = N / f . N=y\partial_yP=P_1(x)y+2bx^2y^2,\qquad R=N/f. N = y ∂ y P = P 1 ( x ) y + 2 b x 2 y 2 , R = N / f .
We follow [19 ] and work in the sheared (central-row) picture. Let A = Z 2 [ y , y − 1 ] [ [ x ] ] A=\mathbb{Z}_2[y,y^{-1}][[x]] A = Z 2 [ y , y − 1 ] [[ x ]] be the ring of power series in x x x whose coefficients are Laurent polynomials in y y y ; f f f is a unit of A A A . Let σ ( x ) = x 2 \sigma(x)=x^2 σ ( x ) = x 2 , σ ( y ) = y 2 \sigma(y)=y^2 σ ( y ) = y 2 , let Λ r , 0 \Lambda_{r,0} Λ r , 0 act by ∑ c i j x i y j ↦ ∑ c 2 i + r , 2 j x i y j \sum c_{ij}x^iy^j\mapsto\sum c_{2i+r,2j}x^iy^j ∑ c ij x i y j ↦ ∑ c 2 i + r , 2 j x i y j , put C = Λ 0 , 0 C=\Lambda_{0,0} C = Λ 0 , 0 , and let CT y \operatorname{CT}_y CT y take the coefficient of y 0 y^0 y 0 . Lemma 2.1(a) holds in A A A with the same proof, and CT y ∘ C = Λ 0 ∘ CT y \operatorname{CT}_y\circ C=\Lambda_0\circ\operatorname{CT}_y CT y ∘ C = Λ 0 ∘ CT y .
Let A ^ + \widehat{A}_+ A + be the set of formal sums ∑ c i j x i y j \sum c_{ij}x^iy^j ∑ c ij x i y j (c i j ∈ Z 2 c_{ij}\in\mathbb{Z}_2 c ij ∈ Z 2 ) supported on { i ≥ 0 , i + j ≥ 0 } \{i\geq0,\ i+j\geq0\} { i ≥ 0 , i + j ≥ 0 } . Each coefficient of a product of two such sums is a finite sum, so A ^ + \widehat{A}_+ A + is a ring. It contains A + : = A ∩ A ^ + A_+:=A\cap\widehat{A}_+ A + := A ∩ A + as a subring, and f , N , R ∈ A + f,N,R\in A_+ f , N , R ∈ A + . The substitution φ ( g ) ( x , y ) = g ( x y , y ) \varphi(g)(x,y)=g(xy,y) φ ( g ) ( x , y ) = g ( x y , y ) , that is x i y j ↦ x i y i + j x^iy^j\mapsto x^iy^{i+j} x i y j ↦ x i y i + j , is a ring isomorphism A ^ + → Z 2 [ [ x , y ] ] \widehat{A}_+\to\mathbb{Z}_2[[x,y]] A + → Z 2 [[ x , y ]] , and
φ ∘ σ = σ ∘ φ , φ ∘ Λ r , 0 = C r ∘ φ , Diag φ = CT y , φ ( f ) = Q F , φ ( N ) = N F . \varphi\circ\sigma=\sigma\circ\varphi,\qquad\varphi\circ\Lambda_{r,0}=C_r\circ\varphi,\qquad\operatorname{Diag}\varphi=\operatorname{CT}_y,\qquad\varphi(f)=Q_F,\qquad\varphi(N)=N_F. φ ∘ σ = σ ∘ φ , φ ∘ Λ r , 0 = C r ∘ φ , Diag φ = CT y , φ ( f ) = Q F , φ ( N ) = N F .
Transporting Sections 2–3 by φ \varphi φ , Lemmas 3.1, 3.2, 3.4 and Proposition 3.6 hold in A ^ + \widehat{A}_+ A + (with its 2-adic topology), with f f f in place of Q Q Q and Λ r , 0 \Lambda_{r,0} Λ r , 0 in place of C r C_r C r . In particular, for a Laurent polynomial M ∈ A + M \in A_{+} M ∈ A + and the all-zero word,
O k ( M / f ) = ∑ j ≥ 0 M j ( k ) σ j ( 1 / f ) , m k = ∑ t = 1 k E t ( m k − t ) , M j ( k ) = ∑ t = 1 k C t ( δ j + t m k − t ) ( j ≥ 1 ) , (21) \mathcal{O}^{k}(M/f)=\sum_{j\geq0}M_{j}^{(k)}\sigma^{j}(1/f),\qquad m_{k}=\sum_{t=1}^{k}E_{t}(m_{k-t}),\qquad M_{j}^{(k)}=\sum_{t=1}^{k}C^{t}(\delta_{j+t}m_{k-t})\quad(j\geq1), \tag*{(21)} O k ( M / f ) = j ≥ 0 ∑ M j ( k ) σ j ( 1/ f ) , m k = t = 1 ∑ k E t ( m k − t ) , M j ( k ) = t = 1 ∑ k C t ( δ j + t m k − t ) ( j ≥ 1 ) , (21)
where E t ( M ) = C t ( δ t M ) E_{t}(M)=C^{t}(\delta_{t}M) E t ( M ) = C t ( δ t M ) , δ t = δ t ( f ) \delta_{t}=\delta_{t}(f) δ t = δ t ( f ) and m 0 = M m_{0}=M m 0 = M . Also CT y R = F \operatorname{CT}_{y}R=F CT y R = F is the Furstenberg series.
Theorem A.1. For every α ≥ 1 \alpha\geq1 α ≥ 1 :
(a) the sequence ( C k R m o d 2 α ) k ≥ 0 (C^{k}R\bmod2^{\alpha})_{k\geq0} ( C k R mod 2 α ) k ≥ 0 has exact preperiod α + 1 \alpha+1 α + 1 and exact period 2 α + 1 2^{\alpha+1} 2 α + 1 ; it takes exactly 2 α + 1 + α + 1 2^{\alpha+1}+\alpha+1 2 α + 1 + α + 1 distinct values;
(b) the sequence ( Λ 0 k F m o d 2 α ) k ≥ 0 (\Lambda_{0}^{k}F\bmod2^{\alpha})_{k\geq0} ( Λ 0 k F mod 2 α ) k ≥ 0 has exact eventual period 2 α + 1 2^{\alpha+1} 2 α + 1 ; in particular K α ( F ) ≥ 2 α + 1 K_{\alpha}(F)\geq2^{\alpha+1} K α ( F ) ≥ 2 α + 1 .
(c) let f ~ ∈ Z 2 [ x , y , y − 1 ] \widetilde{f}\in\mathbb{Z}_{2}[x,y,y^{-1}] f ∈ Z 2 [ x , y , y − 1 ] satisfy f ~ ≡ f ( m o d 2 ) \widetilde{f}\equiv f\pmod2 f ≡ f ( mod 2 ) and have the same monomial support as f f f , let q = 2 α − 1 q=2^{\alpha-1} q = 2 α − 1 , S 0 = N f q − 1 m o d 2 α S_{0}=N^{f^{q-1}}\bmod2^{\alpha} S 0 = N f q − 1 mod 2 α and λ 0 , 0 ( S ) = Λ 0 , 0 ( S f ~ q ) m o d 2 α \lambda_{0,0}(S)=\Lambda_{0,0}(S\widetilde{f}^{q})\bmod2^{\alpha} λ 0 , 0 ( S ) = Λ 0 , 0 ( S f q ) mod 2 α (the initial state and the digit-0 transition of [19 §2] with the lift f ~ \widetilde{f} f ). Then the λ 0 , 0 \lambda_{0,0} λ 0 , 0 -orbit of S 0 S_{0} S 0 has exactly 2 α + 1 + α + 1 2^{\alpha+1}+\alpha+1 2 α + 1 + α + 1 elements.
For P = x 2 y 2 + ( x 2 + x + 1 ) y + x 2 P=x^{2}y^{2}+(x^{2}+x+1)y+x^{2} P = x 2 y 2 + ( x 2 + x + 1 ) y + x 2 , statement (c) is the orbit-size formula conjectured in the appendix of [19 ] , where it was verified for α ≤ 14 \alpha\leq14 α ≤ 14 .
The proof occupies the rest of this appendix. The main step is the annihilation estimate of Proposition A.4.
An ordinary invariant subspace
The Newton polygon of f f f is the triangle Δ \Delta Δ with vertices ( 0 , 0 ) (0,0) ( 0 , 0 ) , ( 2 , − 1 ) (2,-1) ( 2 , − 1 ) and ( 2 , 1 ) (2,1) ( 2 , 1 ) . Its lattice points are ( 0 , 0 ) (0,0) ( 0 , 0 ) , ( 1 , 0 ) (1,0) ( 1 , 0 ) , ( 2 , − 1 ) (2,-1) ( 2 , − 1 ) , ( 2 , 0 ) (2,0) ( 2 , 0 ) and ( 2 , 1 ) (2,1) ( 2 , 1 ) . The linear form ℓ ( i , j ) = i + 2 j \ell(i,j)=i+2j ℓ ( i , j ) = i + 2 j is nonnegative on Δ \Delta Δ and vanishes exactly on the edge E = [ ( 0 , 0 ) , ( 2 , − 1 ) ] E=[(0,0),(2,-1)] E = [( 0 , 0 ) , ( 2 , − 1 )] . Put
J 0 = ( Δ ∩ Z 2 ) ∖ E = { ( 1 , 0 ) , ( 2 , 0 ) , ( 2 , 1 ) } , J_{0}=(\Delta\cap\mathbb{Z}^{2})\setminus E=\{(1,0),(2,0),(2,1)\}, J 0 = ( Δ ∩ Z 2 ) ∖ E = {( 1 , 0 ) , ( 2 , 0 ) , ( 2 , 1 )} ,
so that L J 0 L_{J_{0}} L J 0 is spanned by x x x , x 2 x^{2} x 2 and x 2 y x^{2}y x 2 y . For u ∈ J 0 u\in J_{0} u ∈ J 0 and e ′ ∈ ( 2 s − 1 ) Δ e'\in(2^{s}-1)\Delta e ′ ∈ ( 2 s − 1 ) Δ , a lattice point ( u + e ′ ) / 2 s (u+e')/2^{s} ( u + e ′ ) / 2 s lies in Δ \Delta Δ and satisfies ℓ ≥ 2 − s ℓ ( u ) > 0 \ell\geq2^{-s}\ell(u)>0 ℓ ≥ 2 − s ℓ ( u ) > 0 , hence lies in J 0 J_{0} J 0 . Therefore, by Lemma 3.1 (a), the maps E t E_{t} E t and M ↦ C s ( f 2 s − 1 M ) M\mapsto C^{s}(f^{2^{s-1}}M) M ↦ C s ( f 2 s − 1 M ) preserve L J 0 L_{J_{0}} L J 0 . Let Γ t \Gamma_{t} Γ t and Θ s \Theta_{s} Θ s be their matrices with respect to the ordered basis ( x , x 2 , x 2 y ) (x,x^{2},x^{2}y) ( x , x 2 , x 2 y ) , acting on row vectors: M M M with row vector [ M ] [M] [ M ] is sent to the polynomial with row vector [ M ] Γ t [M]\Gamma_{t} [ M ] Γ t , respectively [ M ] Θ s [M]\Theta_{s} [ M ] Θ s . Then Γ t ≡ 0 ( m o d 2 t − 1 ) \Gamma_{t}\equiv0\pmod{2^{t-1}} Γ t ≡ 0 ( mod 2 t − 1 ) , and ( Θ s ) u , w (\Theta_{s})_{u,w} ( Θ s ) u , w is the coefficient of x 2 s w 1 − u 1 y 2 s w 2 − u 2 x^{2^{s}w_{1}-u_{1}}y^{2^{s}w_{2}-u_{2}} x 2 s w 1 − u 1 y 2 s w 2 − u 2 in f 2 s − 1 f^{2^{s}-1} f 2 s − 1 .
By (3) and Lemma 2.1(a),
C s ( f 2 s − 1 M ) = ∑ t = 1 s C s − t ( f 2 s − t − 1 E t ( M ) ) , that is, Θ s = ∑ t = 1 s Γ t Θ s − t ( Θ 0 = I ) . C^{s}(f^{2^{s-1}}M)=\sum_{t=1}^{s}C^{s-t}\left(f^{2^{s-t-1}}E_{t}(M)\right),\qquad\text{that is,}\qquad\Theta_{s}=\sum_{t=1}^{s}\Gamma_{t}\Theta_{s-t}\quad(\Theta_{0}=I). C s ( f 2 s − 1 M ) = t = 1 ∑ s C s − t ( f 2 s − t − 1 E t ( M ) ) , that is, Θ s = t = 1 ∑ s Γ t Θ s − t ( Θ 0 = I ) .
In terms of generating series in a formal variable Z Z Z , ( I − ∑ t Γ t Z t ) ( ∑ s Θ s Z s ) = I (I-\sum_{t}\Gamma_{t}Z^{t})(\sum_{s}\Theta_{s}Z^{s})=I ( I − ∑ t Γ t Z t ) ( ∑ s Θ s Z s ) = I . Writing H ( Z ) = ∑ t Γ t Z t H(Z)=\sum_{t}\Gamma_{t}Z^{t} H ( Z ) = ∑ t Γ t Z t , the series ∑ n ≥ 0 H ( Z ) n \sum_{n\geq0}H(Z)^{n} ∑ n ≥ 0 H ( Z ) n is defined coefficientwise, since H ( 0 ) = 0 H(0)=0 H ( 0 ) = 0 , and is a two-sided inverse of I − H ( Z ) I-H(Z) I − H ( Z ) . Thus also
Θ s = ∑ t = 1 s Θ s − t Γ t ( s ≥ 1 ) . (22) \Theta_{s}=\sum_{t=1}^{s}\Theta_{s-t}\Gamma_{t}\qquad(s\geq1). \tag*{(22)} Θ s = t = 1 ∑ s Θ s − t Γ t ( s ≥ 1 ) . (22)
With (A.1) and induction, this gives [ m k ] = [ M ] Θ k [m_{k}]=[M]\Theta_{k} [ m k ] = [ M ] Θ k for M ∈ L J 0 M\in L_{J_{0}} M ∈ L J 0 . The matrix
Θ 1 = Γ 1 = ( a 0 0 1 d 0 0 c b ) \Theta_{1}=\Gamma_{1}=
\begin{pmatrix}
a & 0 & 0\\
1 & d & 0\\
0 & c & b
\end{pmatrix} Θ 1 = Γ 1 = a 1 0 0 d c 0 0 b
is invertible modulo 2, and Θ s ≡ Θ 1 s ( m o d 2 ) \Theta_s \equiv\Theta_1^s \pmod{2} Θ s ≡ Θ 1 s ( mod 2 ) because Γ t ≡ 0 ( m o d 2 ) \Gamma_t \equiv0 \pmod{2} Γ t ≡ 0 ( mod 2 ) for t ≥ 2 t \ge2 t ≥ 2 . Hence Θ s ∈ G L 3 ( Z 2 ) \Theta_s \in GL_3(\mathbb{Z}_2) Θ s ∈ G L 3 ( Z 2 ) . Put Φ s = Θ s + 1 Θ s − 1 \Phi_s = \Theta_{s+1}\Theta_s^{-1} Φ s = Θ s + 1 Θ s − 1 .
Lemma A.2. Φ s − Φ s − 1 ∈ 2 s M 3 ( Z 2 ) \Phi_s-\Phi_{s-1} \in2^sM_3(\mathbb{Z}_2) Φ s − Φ s − 1 ∈ 2 s M 3 ( Z 2 ) for s ≥ 1 s \ge1 s ≥ 1 . Hence Φ = lim Φ s \Phi= \lim\Phi_s Φ = lim Φ s exists and Φ − Φ s ∈ 2 s + 1 M 3 ( Z 2 ) \Phi-\Phi_s \in2^{s+1}M_3(\mathbb{Z}_2) Φ − Φ s ∈ 2 s + 1 M 3 ( Z 2 ) . If χ ( t ) = det ( t I − Φ ) = ∑ r = 0 3 χ r t r \chi(t) = \det(tI-\Phi) = \sum_{r=0}^3 \chi_rt^r χ ( t ) = det ( t I − Φ ) = ∑ r = 0 3 χ r t r , then ∑ r = 0 3 χ r Θ n + r ∈ 2 n + 1 M 3 ( Z 2 ) \sum_{r=0}^3 \chi_r\Theta_{n+r} \in2^{n+1}M_3(\mathbb{Z}_2) ∑ r = 0 3 χ r Θ n + r ∈ 2 n + 1 M 3 ( Z 2 ) for every n ≥ 0 n \ge0 n ≥ 0 .
Proof. By (A.2), Θ s + 1 − Φ s − 1 Θ s = ∑ t = 1 s ( Θ s + 1 − t − Φ s − 1 Θ s − t ) Γ t + Γ s + 1 \Theta_{s+1}-\Phi_{s-1}\Theta_s = \sum_{t=1}^s(\Theta_{s+1-t}-\Phi_{s-1}\Theta_{s-t})\Gamma_t+\Gamma_{s+1} Θ s + 1 − Φ s − 1 Θ s = ∑ t = 1 s ( Θ s + 1 − t − Φ s − 1 Θ s − t ) Γ t + Γ s + 1 . Using Θ s + 1 − t = Φ s − t Θ s − t \Theta_{s+1-t}=\Phi_{s-t}\Theta_{s-t} Θ s + 1 − t = Φ s − t Θ s − t , this equals ∑ t = 2 s ( Φ s − t − Φ s − 1 ) Θ s − t Γ t + Γ s + 1 \sum_{t=2}^s(\Phi_{s-t}-\Phi_{s-1})\Theta_{s-t}\Gamma_t+\Gamma_{s+1} ∑ t = 2 s ( Φ s − t − Φ s − 1 ) Θ s − t Γ t + Γ s + 1 . By induction on s s s , Φ s − t − Φ s − 1 = − ∑ i = s − t + 1 s − 1 ( Φ i − Φ i − 1 ) ∈ 2 s − t + 1 \Phi_{s-t}-\Phi_{s-1}=-\sum_{i=s-t+1}^{s-1}(\Phi_i-\Phi_{i-1})\in2^{s-t+1} Φ s − t − Φ s − 1 = − ∑ i = s − t + 1 s − 1 ( Φ i − Φ i − 1 ) ∈ 2 s − t + 1 , while Γ t ∈ 2 t − 1 \Gamma_t\in2^{t-1} Γ t ∈ 2 t − 1 . So the right-hand side lies in 2 s 2^s 2 s , and multiplying by Θ s − 1 \Theta_s^{-1} Θ s − 1 gives the first claim. Then Θ n + r = Φ n + r − 1 ⋯ Φ n Θ n ≡ Φ r Θ n ( m o d 2 n + 1 ) \Theta_{n+r}=\Phi_{n+r-1}\cdots\Phi_n\Theta_n\equiv\Phi^r\Theta_n\pmod{2^{n+1}} Θ n + r = Φ n + r − 1 ⋯ Φ n Θ n ≡ Φ r Θ n ( mod 2 n + 1 ) , and χ ( Φ ) = 0 \chi(\Phi)=0 χ ( Φ ) = 0 gives ∑ r χ r Θ n + r ≡ χ ( Φ ) Θ n = 0 ( m o d 2 n + 1 ) \sum_r\chi_r\Theta_{n+r}\equiv\chi(\Phi)\Theta_n=0\pmod{2^{n+1}} ∑ r χ r Θ n + r ≡ χ ( Φ ) Θ n = 0 ( mod 2 n + 1 ) . To see the matrix identity directly, write adj ( t I − Φ ) = t 2 I + t B 1 + B 0 \operatorname{adj}(tI-\Phi)=t^2I+tB_1+B_0 adj ( t I − Φ ) = t 2 I + t B 1 + B 0 in the adjugate identity. Coefficient comparison gives B 1 = Φ + χ 2 I B_1=\Phi+\chi_2I B 1 = Φ + χ 2 I , B 0 = Φ 2 + χ 2 Φ + χ 1 I B_0=\Phi^2+\chi_2\Phi+\chi_1I B 0 = Φ 2 + χ 2 Φ + χ 1 I , and − Φ B 0 = χ 0 I -\Phi B_0=\chi_0I − Φ B 0 = χ 0 I , hence χ ( Φ ) = 0 \chi(\Phi)=0 χ ( Φ ) = 0 . □
Lemma A.2 is the analogue for J 0 J_0 J 0 of the congruences of [22 Theorem 1(ii)] , with the identity as Frobenius on Z 2 \mathbb{Z}_2 Z 2 ; there the index set consists of the interior or of all lattice points of the Newton polytope. The subspace L J 0 L_{J_0} L J 0 corresponds to an open set of the Newton polygon in the sense of [6 §3] , where p > 2 p>2 p > 2 is assumed.
The eigenvalues
For u = ( 1 , 0 ) u=(1,0) u = ( 1 , 0 ) and w ∈ { ( 2 , 0 ) , ( 2 , 1 ) } w\in\{(2,0),(2,1)\} w ∈ {( 2 , 0 ) , ( 2 , 1 )} , the exponent 2 s w − u 2^sw-u 2 s w − u has x x x -degree 2 s + 1 − 1 > deg x f 2 2 s − 1 2^{s+1}-1>\deg_x f_2^{2^s-1} 2 s + 1 − 1 > deg x f 2 2 s − 1 . For u = ( 2 , 0 ) u=(2,0) u = ( 2 , 0 ) and w = ( 2 , 1 ) w=(2,1) w = ( 2 , 1 ) , it is ( 2 s + 1 − 2 , 2 s ) (2^{s+1}-2,2^s) ( 2 s + 1 − 2 , 2 s ) , whose y y y -degree exceeds that of the top x x x -degree part ( x 2 ( b y + d + c y − 1 ) ) 2 s − 1 (x^2(by+d+cy^{-1}))^{2^s-1} ( x 2 ( b y + d + c y − 1 ) ) 2 s − 1 . Hence Θ s \Theta_s Θ s is lower triangular, with diagonal entries
β s = [ x 2 s − 1 y 0 ] f 2 2 s − 1 , ϑ s = CT y ( b y + d + c y − 1 ) 2 s − 1 , b 2 s − 1 . \beta_s=[x^{2^s-1}y^0]f_2^{2^s-1},\qquad
\vartheta_s=\operatorname{CT}_y(by+d+cy^{-1})^{2^s-1},\qquad
b^{2^s-1}. β s = [ x 2 s − 1 y 0 ] f 2 2 s − 1 , ϑ s = CT y ( b y + d + c y − 1 ) 2 s − 1 , b 2 s − 1 .
So Φ \Phi Φ is lower triangular, with diagonal entries μ 1 = lim β s + 1 / β s \mu_1=\lim\beta_{s+1}/\beta_s μ 1 = lim β s + 1 / β s , μ 2 = lim ϑ s + 1 / ϑ s \mu_2=\lim\vartheta_{s+1}/\vartheta_s μ 2 = lim ϑ s + 1 / ϑ s and μ 3 = lim b 2 s = 1 \mu_3=\lim b^{2^s}=1 μ 3 = lim b 2 s = 1 . They are odd, since Φ ≡ Θ 1 ( m o d 2 ) \Phi\equiv\Theta_1\pmod{2} Φ ≡ Θ 1 ( mod 2 ) .
Lemma A.3. μ 2 = − 1 \mu_2=-1 μ 2 = − 1 .
Proof. Let κ = 1 + ( d / c ) y + ( b / c ) y 2 ∈ Z 2 [ y ] \kappa=1+(d/c)y+(b/c)y^2\in\mathbb{Z}_2[y] κ = 1 + ( d / c ) y + ( b / c ) y 2 ∈ Z 2 [ y ] , so that 1 / ( b y + d + c y − 1 ) = ( y / c ) / κ = ∑ n ≥ 1 u n y n 1/(by+d+cy^{-1})=(y/c)/\kappa=\sum_{n\ge1}u_ny^n 1/ ( b y + d + c y − 1 ) = ( y / c ) / κ = ∑ n ≥ 1 u n y n .
We apply the one-variable case (m = 1 m=1 m = 1 ) of the constructions above to κ \kappa κ , with the digit-0 operators E t ( M ) = C y t ( δ t ( κ ) M ) E_t(M)=C_y^t(\delta_t(\kappa)M) E t ( M ) = C y t ( δ t ( κ ) M ) . The Newton polytope of κ \kappa κ is [ 0 , 2 ] [0,2] [ 0 , 2 ] , and { 1 } \{1\} { 1 } is stable under these operators, since [ 1 + e ′ / 2 t ] ∈ [ 2 − t , 2 − 2 − t ] [1+e'/2^t]\in[2^{-t},2-2^{-t}] [ 1 + e ′ / 2 t ] ∈ [ 2 − t , 2 − 2 − t ] for e ′ ∈ [ 0 , 2 ( 2 t − 1 ) ] e'\in[0,2(2^t-1)] e ′ ∈ [ 0 , 2 ( 2 t − 1 )] . The corresponding 1 × 1 1\times1 1 × 1 matrices are θ s = [ y 2 s − 1 ] κ 2 s − 1 = c 1 − 2 s ϑ s \theta_s=[y^{2^s-1}]\kappa^{2^s-1}=c^{1-2^s}\vartheta_s θ s = [ y 2 s − 1 ] κ 2 s − 1 = c 1 − 2 s ϑ s , so their ratio limit is μ 2 ⋅ lim c 2 s − 2 s + 1 = μ 2 \mu_2\cdot\lim c^{2^s-2^{s+1}}=\mu_2 μ 2 ⋅ lim c 2 s − 2 s + 1 = μ 2 . Lemma A.2 and the proof of Proposition A.4 (i) below, which uses only Lemma A.2 and (A.1), apply verbatim with M = y / c M=y/c M = y / c . They give ( C y − μ 2 ) C y k ( y / ( c κ ) ) ∈ 2 k + 1 Z 2 [ [ y ] ] (C_y-\mu_2)C_y^k(y/(c\kappa))\in2^{k+1}\mathbb{Z}_2[[y]] ( C y − μ 2 ) C y k ( y / ( c κ )) ∈ 2 k + 1 Z 2 [[ y ]] . The coefficient of y y y in this series is u 2 k + 1 − μ 2 u 2 k u_{2^{k+1}}-\mu_2u_{2^k} u 2 k + 1 − μ 2 u 2 k .
Modulo 2, ∑ u n y n ≡ y / ( 1 + y + y 2 ) = ( y + y 2 ) / ( 1 + y 3 ) \sum u_ny^n\equiv y/(1+y+y^2)=(y+y^2)/(1+y^3) ∑ u n y n ≡ y / ( 1 + y + y 2 ) = ( y + y 2 ) / ( 1 + y 3 ) , so u 2 k u_{2^k} u 2 k is odd. Hence μ 2 = lim u 2 k + 1 / u 2 k \mu_2=\lim u_{2^{k+1}}/u_{2^k} μ 2 = lim u 2 k + 1 / u 2 k .
Work in the complete ring O = Z 2 [ T ] / ( T 2 + T + 1 ) \mathcal{O}=\mathbb{Z}_2[T]/(T^2+T+1) O = Z 2 [ T ] / ( T 2 + T + 1 ) , with ζ \zeta ζ the class of T T T and residue field F 4 \mathbb{F}_4 F 4 . The polynomial b Y 2 + d Y + c bY^2+dY+c b Y 2 + d Y + c has the two simple residue roots ζ , ζ 2 \zeta,\zeta^2 ζ , ζ 2 . Each lifts uniquely: a root r n r_n r n modulo 2 n 2^n 2 n is corrected to r n + 2 n ϵ r_n+2^n\epsilon r n + 2 n ϵ modulo 2 n + 1 2^{n+1} 2 n + 1 by choosing ϵ \epsilon ϵ in the residue field, since the derivative 2 b r n + d 2br_n+d 2 b r n + d is a unit. Completeness supplies roots ρ 1 = ζ ξ 1 , ρ 2 = ζ 2 ξ 2 \rho_1=\zeta\xi_1,\rho_2=\zeta^2\xi_2 ρ 1 = ζ ξ 1 , ρ 2 = ζ 2 ξ 2 with ξ 1 , ξ 2 ∈ 1 + 2 O \xi_1,\xi_2\in1+2\mathcal{O} ξ 1 , ξ 2 ∈ 1 + 2 O . By (F2), ξ i 2 k ≡ 1 ( m o d 2 k + 1 ) \xi_i^{2^k}\equiv1\pmod{2^{k+1}} ξ i 2 k ≡ 1 ( mod 2 k + 1 ) . Partial fractions give u n = ( ρ 2 − n − ρ 1 − n ) / ( b ( ρ 1 − ρ 2 ) ) u_n=(\rho_2^{-n}-\rho_1^{-n})/(b(\rho_1-\rho_2)) u n = ( ρ 2 − n − ρ 1 − n ) / ( b ( ρ 1 − ρ 2 )) . As k → ∞ k\to\infty k → ∞ we have ξ 1 2 k , ξ 2 2 k → 1 \xi_1^{2^k},\xi_2^{2^k}\to1 ξ 1 2 k , ξ 2 2 k → 1 , so ρ 2 − 2 k − ρ 1 − 2 k \rho_2^{-2^k}-\rho_1^{-2^k} ρ 2 − 2 k − ρ 1 − 2 k is asymptotic to ζ − 2 ⋅ 2 k − ζ − 2 k \zeta^{-2\cdot2^k}-\zeta^{-2^k} ζ − 2 ⋅ 2 k − ζ − 2 k . This equals ζ − ζ 2 \zeta-\zeta^2 ζ − ζ 2 if 2 k ≡ 1 ( m o d 3 ) 2^k\equiv1\pmod{3} 2 k ≡ 1 ( mod 3 ) and ( ζ − ζ 2 ) (\zeta-\zeta^2) ( ζ − ζ 2 ) if 2 k ≡ 2 ( m o d 3 ) 2^k\equiv2\pmod{3} 2 k ≡ 2 ( mod 3 ) . Since ( ζ − ζ 2 ) 2 = − 3 (\zeta-\zeta^2)^2=-3 ( ζ − ζ 2 ) 2 = − 3 is a unit and 2 2 k ≡ − 2 k ( m o d 3 ) 2^{2^k}\equiv-2^k\pmod{3} 2 2 k ≡ − 2 k ( mod 3 ) , the ratio u 2 k + 1 / u 2 k u_{2^{k+1}}/u_{2^k} u 2 k + 1 / u 2 k tends to − 1 -1 − 1 . □
Thus χ ( t ) = ( t 2 − 1 ) ( t − μ ) \chi(t)=(t^2-1)(t-\mu) χ ( t ) = ( t 2 − 1 ) ( t − μ ) with μ = μ 1 ∈ Z 2 \mu=\mu_1\in\mathbb{Z}_2 μ = μ 1 ∈ Z 2 odd. We do not need the value of μ \mu μ . For the five odd integer parameter sets in Table 1 and Table 2, exact computations give μ ≡ 91 ( m o d 128 ) \mu\equiv91\pmod{128} μ ≡ 91 ( mod 128 ) and μ 2 − μ + 2 ≡ 0 ( m o d 128 ) \mu^2-\mu+2\equiv0\pmod{128} μ 2 − μ + 2 ≡ 0 ( mod 128 ) . This is a finite observation; no exact algebraic identity for μ \mu μ is proved or used.
Annihilation
Proposition A.4. (i) For M ∈ L J 0 M \in L_{J_0} M ∈ L J 0 and k ≥ 0 k \ge0 k ≥ 0 : χ ( C ) C k ( M / f ) ∈ 2 k + 1 A ^ + \chi(C)C^k(M/f) \in2^{k+1}\widehat{A}_+ χ ( C ) C k ( M / f ) ∈ 2 k + 1 A + .
(ii) For k ≥ 1 k \ge1 k ≥ 1 : χ ( C ) C k R ∈ 2 k − 1 A ^ + \chi(C)C^kR \in2^{k-1}\widehat{A}_+ χ ( C ) C k R ∈ 2 k − 1 A + .
Proof. By (A.1) and linearity, the level-j j j numerator of χ ( C ) C k ( M / f ) \chi(C)C^k(M/f) χ ( C ) C k ( M / f ) is ∑ r χ r M j ( k + r ) \sum_r \chi_r M_j^{(k+r)} ∑ r χ r M j ( k + r ) .
(i) At level 0 it equals [ M ] ∑ r χ r Θ k + r ∈ 2 k + 1 [M]\sum_r \chi_r\Theta_{k+r} \in2^{k+1} [ M ] ∑ r χ r Θ k + r ∈ 2 k + 1 by Lemma A.2. At level j ≥ 1 j \ge1 j ≥ 1 it is
∑ t ≥ 1 C t ( δ j + t + ∑ r : k + r ≥ t χ r m k + r − t ) . \sum_{t\ge1} C^t\left(\delta_{j+t}+\sum_{r:\,k+r\ge t}\chi_r m_{k+r-t}\right). t ≥ 1 ∑ C t ( δ j + t + r : k + r ≥ t ∑ χ r m k + r − t ) .
For t ≤ k t \le k t ≤ k the inner sum is the level-0 residual at k − t k-t k − t , which lies in 2 k − t + 1 2^{k-t+1} 2 k − t + 1 ; multiplied by δ j + t ∈ 2 j + t − 1 \delta_{j+t} \in2^{j+t-1} δ j + t ∈ 2 j + t − 1 it lies in 2 k + j ⊂ 2 k + 1 2^{k+j} \subset2^{k+1} 2 k + j ⊂ 2 k + 1 . For t > k t>k t > k we have δ j + t ∈ 2 j + k ⊂ 2 k + 1 \delta_{j+t} \in2^{j+k} \subset2^{k+1} δ j + t ∈ 2 j + k ⊂ 2 k + 1 .
(ii) Now M = N M=N M = N and m k = m k ( N ) m_k=m_k(N) m k = m k ( N ) . The induction in the proof of Lemma 4.2, with v = 0 v=0 v = 0 and η = 0 \eta=0 η = 0 , shows that supp m k ⊂ ( 1 − 2 − k ) Δ + 2 − k K N \operatorname{supp}m_k \subset(1-2^{-k})\Delta+2^{-k}K_N supp m k ⊂ ( 1 − 2 − k ) Δ + 2 − k K N , where K N = conv supp N ⊂ conv { ( 0 , 1 ) , ( 1 , 1 ) , ( 2 , 1 ) , ( 2 , 2 ) } K_N=\operatorname{conv}\operatorname{supp}N \subset\operatorname{conv}\{(0,1),(1,1),(2,1),(2,2)\} K N = conv supp N ⊂ conv {( 0 , 1 ) , ( 1 , 1 ) , ( 2 , 1 ) , ( 2 , 2 )} . The facets of Δ \Delta Δ are i ≤ 2 i\le2 i ≤ 2 , 2 j ≤ i 2j\le i 2 j ≤ i and − 2 j ≤ i -2j\le i − 2 j ≤ i . On K N K_N K N the first and third inequalities hold, and 2 j − i ≤ 2 2j-i\le2 2 j − i ≤ 2 . So a lattice point of ( 1 − 2 − k ) Δ + 2 − k K N (1-2^{-k})\Delta+2^{-k}K_N ( 1 − 2 − k ) Δ + 2 − k K N satisfies 2 j − i ≤ 2 1 − k 2j-i\le2^{1-k} 2 j − i ≤ 2 1 − k , hence 2 j − i ≤ 0 2j-i\le0 2 j − i ≤ 0 for k ≥ 2 k\ge2 k ≥ 2 ; it also satisfies ℓ ≥ 2 − k min K N ℓ > 0 \ell\ge2^{-k}\min_{K_N}\ell>0 ℓ ≥ 2 − k min K N ℓ > 0 . Hence m k ∈ L J 0 m_k\in L_{J_0} m k ∈ L J 0 for k ≥ 2 k\ge2 k ≥ 2 .
For k ≥ 2 k\ge2 k ≥ 2 write m k = ∑ t = 1 k − 2 E t ( m k − t ) + s k m_k=\sum_{t=1}^{k-2}E_t(m_{k-t})+s_k m k = ∑ t = 1 k − 2 E t ( m k − t ) + s k with s k = E k − 1 ( m 1 ) + E k ( m 0 ) s_k=E_{k-1}(m_1)+E_k(m_0) s k = E k − 1 ( m 1 ) + E k ( m 0 ) . Then s k ∈ 2 k − 2 s_k\in2^{k-2} s k ∈ 2 k − 2 , and s k ∈ L J 0 s_k\in L_{J_0} s k ∈ L J 0 because all other terms are. Induction with (A.2) gives [ m k ] = ∑ l = 2 k [ s l ] Θ k − l [m_k]=\sum_{l=2}^k[s_l]\Theta_{k-l} [ m k ] = ∑ l = 2 k [ s l ] Θ k − l for k ≥ 2 k\ge2 k ≥ 2 . For k ≥ 2 k\ge2 k ≥ 2 , the level-0 residual is therefore
∑ r χ r [ m k + r ] = ∑ l = 2 k [ s l ] ∑ r χ r Θ k − l + r + ∑ r χ r ∑ l = k + 1 k + r [ s l ] Θ k + r − l . \sum_r\chi_r[m_{k+r}]
=\sum_{l=2}^k[s_l]\sum_r\chi_r\Theta_{k-l+r}
+\sum_r\chi_r\sum_{l=k+1}^{k+r}[s_l]\Theta_{k+r-l}. r ∑ χ r [ m k + r ] = l = 2 ∑ k [ s l ] r ∑ χ r Θ k − l + r + r ∑ χ r l = k + 1 ∑ k + r [ s l ] Θ k + r − l .
In the first sum each term lies in 2 l − 2 ⋅ 2 k − l + 1 = 2 k − 1 2^{l-2}\cdot2^{k-l+1}=2^{k-1} 2 l − 2 ⋅ 2 k − l + 1 = 2 k − 1 , and in the second each term lies in 2 l − 2 ⊂ 2 k − 1 2^{l-2}\subset2^{k-1} 2 l − 2 ⊂ 2 k − 1 . At a level j ≥ 1 j\ge1 j ≥ 1 we argue as in (i). For t ≤ k − 2 t\le k-2 t ≤ k − 2 the inner sum lies in 2 k − t − 1 2^{k-t-1} 2 k − t − 1 , and the product with δ j + t \delta_{j+t} δ j + t lies in 2 k + j − 2 ⊂ 2 k − 1 2^{k+j-2}\subset2^{k-1} 2 k + j − 2 ⊂ 2 k − 1 . For t ≥ k − 1 t\ge k-1 t ≥ k − 1 we have δ j + t ∈ 2 j + k − 2 ⊂ 2 k − 1 \delta_{j+t}\in2^{j+k-2}\subset2^{k-1} δ j + t ∈ 2 j + k − 2 ⊂ 2 k − 1 . For k = 1 k=1 k = 1 the claim is trivial. □
Reduction modulo 2
Modulo 2 all parameters may be replaced by 1, and f 2 ≡ σ ( f ) f^2\equiv\sigma(f) f 2 ≡ σ ( f ) . Hence C ( M / f ) ≡ C ( M f ) / f ( m o d 2 ) C(M/f)\equiv C(Mf)/f\pmod2 C ( M / f ) ≡ C ( M f ) / f ( mod 2 ) , and C k R ≡ M k / f C^kR\equiv M_k/f C k R ≡ M k / f with M 0 = N M_0=N M 0 = N and M k + 1 = C ( f M k ) M_{k+1}=C(fM_k) M k + 1 = C ( f M k ) . A direct computation gives, modulo 2,
M 0 = ( 1 + x + x 2 ) y , M 1 = x + x y + x 2 + x 2 y , M 2 = x 2 y , M_0=(1+x+x^2)y,\qquad M_1=x+xy+x^2+x^2y,\qquad M_2=x^2y, M 0 = ( 1 + x + x 2 ) y , M 1 = x + x y + x 2 + x 2 y , M 2 = x 2 y ,
M 3 = x 2 ( 1 + y ) , M 4 = x + x 2 y , M 5 = x + x 2 ( 1 + y ) , M_3=x^2(1+y),\qquad M_4=x+x^2y,\qquad M_5=x+x^2(1+y), M 3 = x 2 ( 1 + y ) , M 4 = x + x 2 y , M 5 = x + x 2 ( 1 + y ) ,
and M 6 = M 2 M_6=M_2 M 6 = M 2 . Hence ( C 2 − 1 ) C k R ≡ x / f ( m o d 2 ) (C^2-1)C^kR\equiv x/f\pmod2 ( C 2 − 1 ) C k R ≡ x / f ( mod 2 ) for k ≥ 2 k\ge2 k ≥ 2 , and the sequence ( C k R m o d 2 ) k ≥ 2 (C^kR\bmod2)_{k\ge2} ( C k R mod 2 ) k ≥ 2 has exact period 4. The coefficient of x 1 y 0 x^1y^0 x 1 y 0 in x / f x/f x / f is 1. Since this monomial lies in the central row, ( Λ 0 2 − 1 ) Λ 0 k F (\Lambda_0^2-1)\Lambda_0^kF ( Λ 0 2 − 1 ) Λ 0 k F also has odd coefficient of x x x for k ≥ 2 k\ge2 k ≥ 2 , and ( Λ 0 k F m o d 2 ) k ≥ 2 (\Lambda_0^kF\bmod2)_{k\ge2} ( Λ 0 k F mod 2 ) k ≥ 2 has exact period 4.
The preperiod witness
Let F ^ = c / ( b F ) \widehat{F}=c/(bF) F = c / ( b F ) be the second root of P P P in y y y , so that P = b x 2 ( y − F ) ( y − F ^ ) P=bx^2(y-F)(y-\widehat{F}) P = b x 2 ( y − F ) ( y − F ) in the ring Z 2 [ y , y − 1 ] ( ( x ) ) \mathbb{Z}_2[y,y^{-1}]((x)) Z 2 [ y , y − 1 ] (( x )) , where x x x is invertible. In this ring
R = y 2 ∂ y P P = y 2 ( 1 y − F + 1 y − F ^ ) = ∑ n ≥ 0 F n y 1 − n − ∑ n ≥ 0 y n + 2 F ^ − n − 1 . R=\frac{y^2\partial_yP}{P}
=y^2\left(\frac{1}{y-F}+\frac{1}{y-\widehat{F}}\right)
=\sum_{n\ge0}F^n y^{1-n}-\sum_{n\ge0}y^{n+2}\widehat{F}^{-n-1}. R = P y 2 ∂ y P = y 2 ( y − F 1 + y − F 1 ) = n ≥ 0 ∑ F n y 1 − n − n ≥ 0 ∑ y n + 2 F − n − 1 .
both series converging x x x -adically, since F F F and 1 / F ^ = b F / c 1/\widehat{F}=bF/c 1/ F = b F / c have positive x x x -order; both lie in A A A . Hence [ y j ] R = − ( b F / c ) j − 1 [y^j]R=-(bF/c)^{j-1} [ y j ] R = − ( b F / c ) j − 1 for j ≥ 2 j\ge2 j ≥ 2 . Write F = − c x 2 V F=-cx^2V F = − c x 2 V . Then P 1 V = 1 + b c x 4 V 2 P_1V=1+bcx^4V^2 P 1 V = 1 + b c x 4 V 2 , so V = 1 − a x + ( a 2 − d ) x 2 + O ( x 3 ) V=1-ax+(a^2-d)x^2+O(x^3) V = 1 − a x + ( a 2 − d ) x 2 + O ( x 3 ) . For k ≥ 1 k\ge1 k ≥ 1 and n = 2 k n=2^k n = 2 k ,
B k : = [ x 2 y ] C k R = [ x 2 n y n ] R = b n − 1 [ x 2 ] V n − 1 = b n − 1 ( n − 1 ) ( n 2 a 2 − d ) . (23) B_k:=[x^2y]C^kR=[x^{2n}y^n]R=b^{n-1}[x^2]V^{n-1}=b^{n-1}(n-1)\left(\frac{n}{2}a^2-d\right). \tag*{(23)} B k := [ x 2 y ] C k R = [ x 2 n y n ] R = b n − 1 [ x 2 ] V n − 1 = b n − 1 ( n − 1 ) ( 2 n a 2 − d ) . (23)
For ℓ ≥ 1 \ell\ge1 ℓ ≥ 1 and n ′ = 2 ℓ n n' = 2^{\ell}n n ′ = 2 ℓ n , put U ( n ) = ( n − 1 ) ( n 2 a 2 − d ) U(n)=(n-1)\left(\frac{n}{2}a^2-d\right) U ( n ) = ( n − 1 ) ( 2 n a 2 − d ) . Then
U ( n ′ ) − U ( n ) = 2 k − 1 ( 2 ℓ − 1 ) ( a 2 2 k ( 2 ℓ + 1 ) − ( 2 d + a 2 ) ) U(n')-U(n)=2^{k-1}(2^{\ell}-1)\left(a^2 2^k(2^{\ell}+1)-(2d+a^2)\right) U ( n ′ ) − U ( n ) = 2 k − 1 ( 2 ℓ − 1 ) ( a 2 2 k ( 2 ℓ + 1 ) − ( 2 d + a 2 ) )
has 2-adic valuation exactly k − 1 k-1 k − 1 , while b n ′ − b n ∈ 2 k + 2 b^{n'}-b^n \in2^{k+2} b n ′ − b n ∈ 2 k + 2 because b 2 k ≡ 1 ( m o d 2 k + 2 ) b^{2^k}\equiv1\pmod{2^{k+2}} b 2 k ≡ 1 ( mod 2 k + 2 ) : an odd b b b has b 2 ≡ 1 ( m o d 8 ) b^2\equiv1\pmod8 b 2 ≡ 1 ( mod 8 ) , and each further squaring raises this precision by at least one, by (F2). Since B k + ℓ − B k = b − 1 ( b n ′ ( U ( n ′ ) − U ( n ) ) + ( b n ′ − b n ) U ( n ) ) B_{k+\ell}-B_k=b^{-1}\left(b^{n'}(U(n')-U(n))+(b^{n'}-b^n)U(n)\right) B k + ℓ − B k = b − 1 ( b n ′ ( U ( n ′ ) − U ( n )) + ( b n ′ − b n ) U ( n ) ) , we get
v 2 ( B k + ℓ − B k ) = k − 1 ( k , ℓ ≥ 1 ) . v_2(B_{k+\ell}-B_k)=k-1 \qquad(k,\ell\ge1). v 2 ( B k + ℓ − B k ) = k − 1 ( k , ℓ ≥ 1 ) .
Proof of Theorem A.1
(a) Let S ∗ = C α + 1 R ( m o d 2 α ) S^*=C^{\alpha+1}R\pmod{2^\alpha} S ∗ = C α + 1 R ( mod 2 α ) . By Proposition A.4(ii), χ ( C ) S ∗ = 0 \chi(C)S^*=0 χ ( C ) S ∗ = 0 . For m ≥ 1 m\ge1 m ≥ 1 let g m ( t ) = ∑ j = 0 m − 1 t 2 j g_m(t)=\sum_{j=0}^{m-1}t^{2j} g m ( t ) = ∑ j = 0 m − 1 t 2 j . Then
t 2 m − 1 − g m ( μ ) ( t 2 − 1 ) = ( t 2 − 1 ) ( g m ( t ) − g m ( μ ) ) t^{2m}-1-g_m(\mu)(t^2-1)=(t^2-1)(g_m(t)-g_m(\mu)) t 2 m − 1 − g m ( μ ) ( t 2 − 1 ) = ( t 2 − 1 ) ( g m ( t ) − g m ( μ ))
is divisible by χ ( t ) \chi(t) χ ( t ) , so ( C 2 m − 1 ) S ∗ = g m ( μ ) ( C 2 − 1 ) S ∗ (C^{2m}-1)S^*=g_m(\mu)(C^2-1)S^* ( C 2 m − 1 ) S ∗ = g m ( μ ) ( C 2 − 1 ) S ∗ . Write m = 2 r u ′ m=2^r u' m = 2 r u ′ with u ′ u' u ′ odd. Then g m ( μ ) = g u ′ ( μ ) ∏ i < r ( 1 + μ 2 i ) g_m(\mu)=g_{u'}(\mu)\prod_{i<r}(1+\mu^{2^i}) g m ( μ ) = g u ′ ( μ ) ∏ i < r ( 1 + μ 2 i ) , where g u ′ ( μ ) g_{u'}(\mu) g u ′ ( μ ) is a sum of u ′ u' u ′ odd terms and each factor is ≡ 2 ( m o d 4 ) \equiv2\pmod4 ≡ 2 ( mod 4 ) . Hence v 2 ( g m ( μ ) ) = v 2 ( m ) v_2(g_m(\mu))=v_2(m) v 2 ( g m ( μ )) = v 2 ( m ) .
Taking m = 2 α m=2^\alpha m = 2 α shows that S ∗ S^* S ∗ is periodic with period dividing 2 α + 1 2^{\alpha+1} 2 α + 1 , so the preperiod is at most α + 1 \alpha+1 α + 1 . Let ϖ \varpi ϖ be the exact period. If ϖ \varpi ϖ were odd, then reduction modulo 2 would give an odd period of ( C k R ( m o d 2 ) ) k ≥ 2 (C^kR\pmod2)_{k\ge2} ( C k R ( mod 2 ) ) k ≥ 2 , contradicting the exact period 4. So ϖ = 2 m \varpi=2m ϖ = 2 m , and 0 = ( C 2 m − 1 ) S ∗ = g m ( μ ) ( C 2 − 1 ) S ∗ 0=(C^{2m}-1)S^*=g_m(\mu)(C^2-1)S^* 0 = ( C 2 m − 1 ) S ∗ = g m ( μ ) ( C 2 − 1 ) S ∗ . The coefficient of x x x in ( C 2 − 1 ) S ∗ (C^2-1)S^* ( C 2 − 1 ) S ∗ is a unit, so 2 α ∣ g m ( μ ) 2^\alpha\mid g_m(\mu) 2 α ∣ g m ( μ ) , hence 2 α ∣ m 2^\alpha\mid m 2 α ∣ m and ϖ = 2 α + 1 \varpi=2^{\alpha+1} ϖ = 2 α + 1 .
The witness shows that C α R ≢ C α + ℓ R ( m o d 2 α ) C^\alpha R\not\equiv C^{\alpha+\ell}R\pmod{2^\alpha} C α R ≡ C α + ℓ R ( mod 2 α ) for every ℓ ≥ 1 \ell\ge1 ℓ ≥ 1 , so C α R C^\alpha R C α R is not periodic and the preperiod is exactly α + 1 \alpha+1 α + 1 . The orbit of R R R under C C C consists of the non-periodic elements C k R C^kR C k R with k ≤ α k\le\alpha k ≤ α , followed by a cycle of length 2 α + 1 2^{\alpha+1} 2 α + 1 . By minimality of the preperiod and of the period, these 2 α + 1 + α + 1 2^{\alpha+1}+\alpha+1 2 α + 1 + α + 1 elements are pairwise distinct.
(b) Put e k = Λ 0 k F ( m o d 2 α ) = C T y ( C k R ) ( m o d 2 α ) e_k=\Lambda_0^kF\pmod{2^\alpha}=C T_y(C^kR)\pmod{2^\alpha} e k = Λ 0 k F ( mod 2 α ) = C T y ( C k R ) ( mod 2 α ) . Applying C T y C T_y C T y , the argument of (a) uses three facts:
χ ( Λ 0 ) e α + 1 = 0 \chi(\Lambda_0)e_{\alpha+1}=0 χ ( Λ 0 ) e α + 1 = 0 , by Proposition A.4(ii) and C T y ∘ C = Λ 0 ∘ C T y C T_y\circ C=\Lambda_0\circ C T_y C T y ∘ C = Λ 0 ∘ C T y ;
for k ≥ 2 k\ge2 k ≥ 2 the coefficient of x x x in ( Λ 0 2 − 1 ) e k (\Lambda_0^2-1)e_k ( Λ 0 2 − 1 ) e k is odd;
( e k ( m o d 2 ) ) k ≥ 2 (e_k\pmod2)_{k\ge2} ( e k ( mod 2 ) ) k ≥ 2 has exact period 4.
Facts (ii) and (iii) are proved in the subsection on reduction modulo 2. With them, the argument of (a) shows that the eventual period of ( e k ) (e_k) ( e k ) is exactly 2 α + 1 2^{\alpha+1} 2 α + 1 . Hence the kernel of F ( m o d 2 α ) F\pmod{2^\alpha} F ( mod 2 α ) contains at least 2 α + 1 2^{\alpha+1} 2 α + 1 distinct elements e k e_k e k . The preperiod of ( e k ) (e_k) ( e k ) can be smaller than α + 1 \alpha+1 α + 1 , because the witness monomial x 2 y x^2y x 2 y does not lie in the central row.
(c) Since f ~ ≡ f ( m o d 2 ) \widetilde f\equiv f\pmod2 f ≡ f ( mod 2 ) , facts (F1) and (F2) give f ~ q ≡ f q ( m o d 2 α ) \widetilde f^q\equiv f^q\pmod{2^\alpha} f q ≡ f q ( mod 2 α ) and f ~ 2 q ≡ σ ( f ~ ) q ( m o d 2 α ) \widetilde f^{2q}\equiv\sigma(\widetilde f)^q\pmod{2^\alpha} f 2 q ≡ σ ( f ) q ( mod 2 α ) . Hence, by Lemma 2.1(a) in A,
C ( S / f ~ q ) = C ( S f ~ q / f ~ 2 q ) ≡ C ( S f ~ q ) / f ~ q ( m o d 2 α ) . C(S/\widetilde f^q)=C(S\widetilde f^q/\widetilde f^{2q})\equiv C(S\widetilde f^q)/\widetilde f^q\pmod{2^\alpha}. C ( S / f q ) = C ( S f q / f 2 q ) ≡ C ( S f q ) / f q ( mod 2 α ) .
Division by the unit f ~ q \widetilde f^q f q is injective. It maps S 0 S_0 S 0 to N / f = R ( m o d 2 α ) N/f=R\pmod{2^\alpha} N / f = R ( mod 2 α ) and intertwines λ 0 , 0 \lambda_{0,0} λ 0 , 0 with C C C . So the λ 0 , 0 \lambda_{0,0} λ 0 , 0 -orbit of S 0 S_0 S 0 is in bijection with the orbit of R R R , and (a) applies. □
This appendix proves the contact estimate used in § 7 . Throughout the appendix the parameters are a = b = c = d = 1 a=b=c=d=1 a = b = c = d = 1 . Put
q = 1 + x + x 2 , D = q 2 − 4 x 4 , B = D − 1 / 2 , U = ( 1 , x ) . q=1+x+x^2,\qquad D=q^2-4x^4,\qquad B=D^{-1/2},\qquad U=(1,x). q = 1 + x + x 2 , D = q 2 − 4 x 4 , B = D − 1/2 , U = ( 1 , x ) .
The square root has constant coefficient 1. For a scalar series h h h , let L ( h ) = ( [ x 0 ] h , [ x 1 ] h ) t L(h)=([x^0]h,[x^1]h)^{\mathsf t} L ( h ) = ([ x 0 ] h , [ x 1 ] h ) t , and apply L L L columnwise to a two-column frame. Write
π U h = h − U L ( h ) , N ( P ) = P ( L P ) − 1 (24) \pi_Uh=h-UL(h),\qquad\mathcal{N}(P)=P(LP)^{-1} \tag*{(24)} π U h = h − U L ( h ) , N ( P ) = P ( L P ) − 1 (24)
when L P LP L P is invertible. All valuations below are the minimum of the 2-adic valuations of the coefficients, including both columns for a frame. A binary word is written in order of application of its sections. Define
A = B ( x ) / B ( x 2 ) , C r ( h ) = Λ r ( A h ) , P ∅ = U , P w r = N ( C r P w ) . A = B(x)/B(x^2), \qquad C_r(h) = \Lambda_r(Ah), \qquad P_{\varnothing} = U, \qquad P_{wr} = \mathcal{N}(C_rP_w). A = B ( x ) / B ( x 2 ) , C r ( h ) = Λ r ( A h ) , P ∅ = U , P w r = N ( C r P w ) .
These frames are the rational normalizations of the physical frames E w E_w E w of § 7 . More explicitly, with
H = L ( B U ) = ( 1 0 − 1 1 ) , E w = B P w H − 1 , H = L(BU) = \begin{pmatrix} 1 & 0 \\ -1 & 1 \end{pmatrix}, \qquad E_w = BP_wH^{-1}, H = L ( B U ) = ( 1 − 1 0 1 ) , E w = B P w H − 1 ,
the first two rows of E w E_w E w are the identity. Indeed Λ r ( B h ) = B C r ( h ) \Lambda_r(Bh) = BC_r(h) Λ r ( B h ) = B C r ( h ) , and this identity intertwines the two normalizations. Multiplication by B B B and B − 1 B^{-1} B − 1 , and right multiplication by a constant integral unit matrix, preserve valuation. Consequently v 2 ( E w − E v ) = v 2 ( P w − P v ) v_2(E_w-E_v) = v_2(P_w-P_v) v 2 ( E w − E v ) = v 2 ( P w − P v ) .
An all-precision rational multiplier
Lemma B.1. In the 2-adic completion of Z 2 [ x , q − 1 ] \mathbb{Z}_2[x,q^{-1}] Z 2 [ x , q − 1 ] one has
A = q + 2 T q + ∑ n ≥ 2 2 n A n q 4 n − 5 , T = − x − x 2 − x 3 + x 4 , (25) A = q + \frac{2T}{q} + \sum_{n\ge2}\frac{2^nA_n}{q^{4n-5}}, \qquad T = -x-x^2-x^3+x^4, \tag*{(25)} A = q + q 2 T + n ≥ 2 ∑ q 4 n − 5 2 n A n , T = − x − x 2 − x 3 + x 4 , (25) where A n ∈ Z 2 [ x ] A_n \in\mathbb{Z}_2[x] A n ∈ Z 2 [ x ] , deg A n ≤ 8 n − 8 \deg A_n \le8n-8 deg A n ≤ 8 n − 8 . In particular every summand has growth at infinity at most 2 2 2 . The identity Λ r ( B h ) = B C r ( h ) \Lambda_r(Bh) = BC_r(h) Λ r ( B h ) = B C r ( h ) holds for every integral formal series h h h .
Proof. Set
C = 1 + x 2 + 3 x 4 , H ∗ = x 5 + x 6 + 2 x 7 + 4 x 9 + 2 x 10 + 3 x 11 − 3 x 12 . C = 1+x^2+3x^4, \qquad H_* = x^5+x^6+2x^7+4x^9+2x^{10}+3x^{11}-3x^{12}. C = 1 + x 2 + 3 x 4 , H ∗ = x 5 + x 6 + 2 x 7 + 4 x 9 + 2 x 10 + 3 x 11 − 3 x 12 .
Polynomial multiplication gives C 2 D − q 2 D ( x 2 ) = 8 H ∗ C^2D-q^2D(x^2)=8H_* C 2 D − q 2 D ( x 2 ) = 8 H ∗ . Therefore, on the branches with constant coefficient 1 1 1 ,
A = C q ( 1 − 8 H ∗ D C 2 ) 1 / 2 . A = \frac{C}{q}\left(1-\frac{8H_*}{DC^2}\right)^{1/2}. A = q C ( 1 − D C 2 8 H ∗ ) 1/2 .
For k ≥ 1 k\ge1 k ≥ 1 , v 2 ( 8 k ( 1 / 2 k ) ) = k + s 2 ( k ) v_2\left(8^k\binom{1/2}{k}\right)=k+s_2(k) v 2 ( 8 k ( k 1/2 ) ) = k + s 2 ( k ) , where s 2 ( k ) s_2(k) s 2 ( k ) is the number of nonzero binary digits of k k k . To check this, use ( 1 / 2 k ) = ( − 1 ) k − 1 ( 2 k k ) / ( 4 k ( 2 k − 1 ) ) \binom{1/2}{k}=(-1)^{k-1}\binom{2k}{k}/(4^k(2k-1)) ( k 1/2 ) = ( − 1 ) k − 1 ( k 2 k ) / ( 4 k ( 2 k − 1 )) and v 2 ( n ! ) = n − s 2 ( n ) v_2(n!)=n-s_2(n) v 2 ( n !) = n − s 2 ( n ) . This also proves convergence in the indicated completion. It agrees coefficientwise with the formal-series expansion, since H ∗ H_* H ∗ has positive order.
Put h = x − 3 ( 1 + x 3 ) 2 h=x^{-3}(1+x^3)^2 h = x − 3 ( 1 + x 3 ) 2 and J ∗ = ( H ∗ − q 2 h ) / 2 J_*=(H_*-q^2h)/2 J ∗ = ( H ∗ − q 2 h ) /2 . The latter is integral, since H ∗ ≡ q 2 h ( m o d 2 ) H_*\equiv q^2h\pmod{2} H ∗ ≡ q 2 h ( mod 2 ) . As C = q 2 + 2 T C=q^2+2T C = q 2 + 2 T ,
H ∗ D C 2 = ( h q 4 + 2 J ∗ q 6 ) ( 1 − 4 x 4 / q 2 ) − 1 ( 1 + 2 T / q 2 ) − 2 . \frac{H_*}{DC^2}=\left(\frac{h}{q^4}+\frac{2J_*}{q^6}\right)(1-4x^4/q^2)^{-1}(1+2T/q^2)^{-2}. D C 2 H ∗ = ( q 4 h + q 6 2 J ∗ ) ( 1 − 4 x 4 / q 2 ) − 1 ( 1 + 2 T / q 2 ) − 2 .
In the k k k th square-root term choose j j j factors 2 J ∗ / q 6 2J_*/q^6 2 J ∗ / q 6 and indices l , m l,m l , m in the last two geometric expansions. Before multiplication by C / q C/q C / q the pole exponent and coefficient valuation satisfy
e = 4 k + 2 j + 2 l + 2 m , n ≥ k + s 2 ( k ) + j + 2 l + m , e=4k+2j+2l+2m, \qquad n\ge k+s_2(k)+j+2l+m, e = 4 k + 2 j + 2 l + 2 m , n ≥ k + s 2 ( k ) + j + 2 l + m ,
hence e ≤ 4 n − 4 e\le4n-4 e ≤ 4 n − 4 . Multiplication by the q q q part of C / q C/q C / q leaves pole at most 4 n − 5 4n-5 4 n − 5 ; its 2 T / q 2T/q 2 T / q part, with one more unit of valuation, leaves pole at most 4 n − 7 4n-7 4 n − 7 at the new weight. The k = 0 k=0 k = 0 term is q + 2 T / q q+2T/q q + 2 T / q . Every displayed factor has growth at infinity at most zero, except C / q C/q C / q , which has growth at most two. Collecting at each valuation therefore gives (B.1), including its degree bound. At any fixed valuation only finitely many k , j , l , m k,j,l,m k , j , l , m contribute, so this collection controls the entire tail. Finally B h = B ( x 2 ) A h Bh=B(x^2)Ah B h = B ( x 2 ) A h and Λ r ( S ( x 2 ) T ( x ) ) = S ( x ) Λ r ( T ( x ) ) \Lambda_r(S(x^2)T(x))=S(x)\Lambda_r(T(x)) Λ r ( S ( x 2 ) T ( x )) = S ( x ) Λ r ( T ( x )) prove the operator identity. □
The normal lattice and the difference cocycle
For e ≥ 0 e \ge0 e ≥ 0 put
S e = { N / q e : N ∈ Z 2 [ x ] , deg N ≤ 2 e + 1 } . S_e=\{N/q^e:N\in\mathbb{Z}_2[x],\ \deg N\le2e+1\}. S e = { N / q e : N ∈ Z 2 [ x ] , deg N ≤ 2 e + 1 } .
Thus S 0 = ⟨ 1 , x ⟩ S_0=\langle1,x\rangle S 0 = ⟨ 1 , x ⟩ . Define closed normal modules
L = ker L ∩ { ∑ j ≥ 0 2 j h j : h j ∈ S 1 + 2 j } , \mathcal{L}=\ker L\cap\left\{\sum_{j\ge0}2^j h_j:h_j\in S_{1+2j}\right\}, L = ker L ∩ { j ≥ 0 ∑ 2 j h j : h j ∈ S 1 + 2 j } ,
H 3 = ker L ∩ { ∑ j ≥ 3 2 j h j : h j ∈ S 1 + 2 j } , Q = L / ( 2 L + H 3 ) . \mathcal{H}_3=\ker L\cap\left\{\sum_{j\ge3}2^j h_j:h_j\in S_{1+2j}\right\},\qquad Q=\mathcal{L}/(2\mathcal{L}+\mathcal{H}_3). H 3 = ker L ∩ { j ≥ 3 ∑ 2 j h j : h j ∈ S 1 + 2 j } , Q = L / ( 2 L + H 3 ) .
All sums converge 2 2 2 -adically coefficientwise. Each summand in a normal element may be made normal by applying π U \pi_U π U : the removed polynomial pieces sum to U L ( h ) = 0 UL(h)=0 U L ( h ) = 0 .
Lemma B.2. The divided operators T r = π U C r / 2 T_r=\pi_U C_r/2 T r = π U C r /2 preserve L \mathcal{L} L and send H 3 \mathcal{H}_3 H 3 into 2 L 2\mathcal{L} 2 L . The frames satisfy P w = U + 2 Q w P_w=U+2Q_w P w = U + 2 Q w with Q w ∈ L Q_w\in\mathcal{L} Q w ∈ L . If Q r = ( N ( C r U ) − U ) / 2 Q_r=(N(C_rU)-U)/2 Q r = ( N ( C r U ) − U ) /2 , then
Q w r ≡ Q r ( m o d 2 L ) . (26) Q_{wr}\equiv Q_r\pmod{2\mathcal{L}}. \tag*{(26)} Q w r ≡ Q r ( mod 2 L ) . (26)
The quotient Q Q Q is ten-dimensional over F 2 \mathbb{F}_2 F 2 , with ordered basis
x 2 q , x 3 q ; 2 x i q 3 ( 4 ≤ i ≤ 7 ) ; 4 x i q 5 ( 8 ≤ i ≤ 11 ) . (27) \frac{x^2}{q},\ \frac{x^3}{q};\qquad\frac{2x^i}{q^3}\ (4\le i\le7);\qquad\frac{4x^i}{q^5}\ (8\le i\le11). \tag*{(27)} q x 2 , q x 3 ; q 3 2 x i ( 4 ≤ i ≤ 7 ) ; q 5 4 x i ( 8 ≤ i ≤ 11 ) . (27)
If two normalized frames differ by 2 s Z 2^sZ 2 s Z , where s ≥ 1 s\ge1 s ≥ 1 and Z Z Z has columns in L \mathcal{L} L , then after a common digit r r r the difference is 2 s + 1 Z ′ 2^{s+1}Z' 2 s + 1 Z ′ , with
[ Z ′ ] = [ T r Z − Q r L C r Z ] M r − 1 , M 0 = ( 1 0 1 1 ) , M 1 = ( 1 1 0 1 ) over F 2 . (28) [Z']=[T_rZ-Q_rLC_rZ]M_r^{-1},\qquad M_0=\begin{pmatrix}1&0\\1&1\end{pmatrix},\qquad M_1=\begin{pmatrix}1&1\\0&1\end{pmatrix}\quad\text{over }\mathbb{F}_2. \tag*{(28)} [ Z ′ ] = [ T r Z − Q r L C r Z ] M r − 1 , M 0 = ( 1 1 0 1 ) , M 1 = ( 1 0 1 1 ) over F 2 . (28)
The expression in brackets depends only on [ Z ] ∈ Q 2 [Z]\in Q^2 [ Z ] ∈ Q 2 .
Proof. Here is an exact rational-section formula that controls every pole. For e > 1 e>1 e > 1 , in the term indexed by i i i write e − i = 2 j + ε e-i=2j+\varepsilon e − i = 2 j + ε , with ε ∈ { 0 , 1 } \varepsilon\in\{0,1\} ε ∈ { 0 , 1 } . The identities q ( x ) q ( − x ) = q ( x 2 ) q(x)q(-x)=q(x^2) q ( x ) q ( − x ) = q ( x 2 ) , q ( − x ) = q − 2 x q(-x)=q-2x q ( − x ) = q − 2 x and q 2 = q ( x 2 ) + 2 x q q^2=q(x^2)+2xq q 2 = q ( x 2 ) + 2 x q give
Λ r ( N q e ) = ∑ i = 0 e ∑ l = 0 j ( e i ) ( − 2 ) i ( j l ) 2 l Λ r ( N x i + l q ε + l ) q e − j + l . (29) \Lambda_r\left(\frac{N}{q^e}\right)=\sum_{i=0}^{e}\sum_{l=0}^{j}\binom{e}{i}(-2)^i\binom{j}{l}2^l\frac{\Lambda_r(Nx^{i+l}q^{\varepsilon+l})}{q^{e-j+l}}. \tag*{(29)} Λ r ( q e N ) = i = 0 ∑ e l = 0 ∑ j ( i e ) ( − 2 ) i ( l j ) 2 l q e − j + l Λ r ( N x i + l q ε + l ) . (29)
If deg N ≤ 2 e + 3 \deg N\le2e+3 deg N ≤ 2 e + 3 , the degree of each numerator on the right is at most 2 ( e − j + l ) + 1 2(e-j+l)+1 2 ( e − j + l ) + 1 . Indeed the upper bound before section is 2 e + 3 + i + 3 l + 2 ε 2e+3+i+3l+2\varepsilon 2 e + 3 + i + 3 l + 2 ε ; after section its degree is at most the floor of half that number minus r / 2 r/2 r /2 . The required inequality follows from i + l + r ≥ 1 i+l+r\ge1 i + l + r ≥ 1 . In the remaining case i = l = r = 0 i=l=r=0 i = l = r = 0 the upper degree before section is odd, so taking the floor proves the same inequality. Since e − j + l = ⌈ ( e + i ) / 2 ⌉ + l ≤ ⌈ e / 2 ⌉ + i + l e-j+l=\lceil(e+i)/2\rceil+l\le\lceil e/2\rceil+i+l e − j + l = ⌈( e + i ) /2 ⌉ + l ≤ ⌈ e /2 ⌉ + i + l , the section belongs to ∑ t ≥ 0 2 t S ⌈ e / 2 ⌉ + t \sum_{t\ge0}2^tS_{\lceil e/2\rceil+t} ∑ t ≥ 0 2 t S ⌈ e /2 ⌉ + t . For e = 0 e=0 e = 0 a polynomial of degree at most three sections to degree at most one.
Multiply an input 2 j S 1 + 2 j 2^jS_{1+2j} 2 j S 1 + 2 j by the three parts of (B.1), section, and divide by two. The resulting weight g g g and pole bound p p p are
Table 4.
Every product before section has growth at infinity at most three, so the degree estimate just proved applies. In each row p ≤ 1 + 2 g p\le1+2g p ≤ 1 + 2 g ; the sole exception j = t = 0 j=t=0 j = t = 0 in the first row is a polynomial of degree at most one, annihilated by π U \pi_U π U . This proves T r L ⊆ L T_r\mathcal{L}\subseteq\mathcal{L} T r L ⊆ L without dropping a tail. For $j\ge 3 the stronger inequality p ≤ 2 g − 1 p \le2g - 1 p ≤ 2 g − 1 holds in all three rows, proving T r H 3 ⊆ 2 L T_r\mathcal{H}_3 \subseteq2\mathcal{L} T r H 3 ⊆ 2 L . Projection by π U \pi_U π U preserves each pole and growth bound.
The same calculation starting with the polynomial frame U U U gives N ( C r U ) ∈ U + 2 L \mathcal{N}(C_rU) \in U + 2\mathcal{L} N ( C r U ) ∈ U + 2 L . Its normalization is a unit: modulo two its first rows are precisely M r M_r M r above. If P = U + 2 Q P = U + 2Q P = U + 2 Q and P ′ = N ( C r P ) P' = \mathcal{N}(C_rP) P ′ = N ( C r P ) , comparison with N ( C r U ) = U + 2 Q r \mathcal{N}(C_rU) = U + 2Q_r N ( C r U ) = U + 2 Q r gives a difference in 4 L 4\mathcal{L} 4 L . For, the normal part of C r Q C_rQ C r Q is in 2 L 2\mathcal{L} 2 L , and replacing π U \pi_U π U by π U + 2 Q r \pi_{U+2Q_r} π U + 2 Q r subtracts 2 Q r L C r Q 2Q_rLC_rQ 2 Q r L C r Q , also in 2 L 2\mathcal{L} 2 L . Right multiplication by normalization units preserves the module. This proves the frame assertion and (B.2).
For the quotient, normal S e S_e S e has rank 2 e 2e 2 e , with numerator basis x 2 , … , x 2 e + 1 x^2,\ldots,x^{2e+1} x 2 , … , x 2 e + 1 . At weights zero, one and two, the new ranks after quotienting the previous pole space are 2 2 2 , 6 − 2 = 4 6-2=4 6 − 2 = 4 and 10 − 6 = 4 10-6=4 10 − 6 = 4 . All higher weights are killed by H 3 \mathcal{H}_3 H 3 , and even coefficients in these ten generators are killed by 2 L 2\mathcal{L} 2 L . This proves both spanning and independence in (B.3). Equivalently, multiply by q 5 q^5 q 5 modulo eight and eliminate successively the pivots x 2 , x 3 x^2,x^3 x 2 , x 3 with coefficient one, x 4 , … , x 7 x^4,\ldots,x^7 x 4 , … , x 7 with coefficient two, and x 8 , … , x 11 x^8,\ldots,x^{11} x 8 , … , x 11 with coefficient four.
For the cocycle put P ′ = N ( C r P ) P'=\mathcal{N}(C_rP) P ′ = N ( C r P ) . Exact matrix algebra gives
N ( C r ( P + 2 s Z ) ) − P ′ = 2 s ( C r Z − P ′ L C r Z ) [ L C r ( P + 2 s Z ) ] − 1 . \mathcal{N}\bigl(C_r(P+2^sZ)\bigr)-P'=2^s(C_rZ-P'LC_rZ)[LC_r(P+2^sZ)]^{-1}. N ( C r ( P + 2 s Z ) ) − P ′ = 2 s ( C r Z − P ′ L C r Z ) [ L C r ( P + 2 s Z ) ] − 1 .
Substitute P ′ = U + 2 Q ′ P'=U+2Q' P ′ = U + 2 Q ′ and divide by 2 s + 1 2^{s+1} 2 s + 1 . The result has columns in L \mathcal{L} L by the closure just proved. In the quotient replace Q ′ Q' Q ′ by Q r Q_r Q r using (B.2), and the right inverse by its reduction M r − 1 M_r^{-1} M r − 1 . This yields (B.4). The scalar map kills 2 L 2\mathcal{L} 2 L in the quotient. It also kills H 3 \mathcal{H}_3 H 3 : its T r T_r T r part lies in 2 L 2\mathcal{L} 2 L , and L C r Z LC_rZ L C r Z is divisible by eight for such an input. This proves that the finite map is well-defined. □
Lemma B.3. One has Q 0 − Q 1 ∈ 2 L Q_0-Q_1 \in2\mathcal{L} Q 0 − Q 1 ∈ 2 L . Consequently ( P w − P v ) / 4 (P_w-P_v)/4 ( P w − P v ) /4 has columns in L \mathcal{L} L for every pair of nonempty words. Also, for words h , v h,v h , v ,
P h v − P v ∈ 2 ∣ v ∣ + 1 L 2 . (30) P_{hv}-P_v \in2^{|v|+1}\mathcal{L}^2. \tag*{(30)} P h v − P v ∈ 2 ∣ v ∣ + 1 L 2 . (30)
Proof. The forcing frames have the sharper pole bounds
Q r ∈ S 1 + ∑ k ≥ 1 2 k S 2 k columnwise . Q_r \in S_1+\sum_{k\ge1}2^kS_{2k}\quad\text{columnwise}. Q r ∈ S 1 + k ≥ 1 ∑ 2 k S 2 k columnwise .
Indeed the 2 T / q 2T/q 2 T / q term stays in S 1 S_1 S 1 after section and division by two. A term n ≥ 2 n\ge2 n ≥ 2 of (B.1), with additional section weight t t t , has weight k = n + t − 1 k=n+t-1 k = n + t − 1 and pole at most 2 n − 2 + t = 2 k − t 2n-2+t=2k-t 2 n − 2 + t = 2 k − t . More explicitly, Q r = ( π U C r U / 2 ) ( L C r U ) − 1 Q_r=(\pi_UC_rU/2)(LC_rU)^{-1} Q r = ( π U C r U /2 ) ( L C r U ) − 1 , and the positive-weight terms of the constant inverse multiply only lower-weight rational terms. They preserve the displayed bounds. Modulo 2 L 2\mathcal{L} 2 L , terms with t ≥ 1 t\ge1 t ≥ 1 already have the one-pole margin required at positive weight. Thus at weight k ≥ 1 k\ge1 k ≥ 1 only the term n = k + 1 n=k+1 n = k + 1 , t = 0 t=0 t = 0 may retain pole 2 k 2k 2 k . Its incoming pole is 4 k − 1 4k-1 4 k − 1 , which is odd. At weight zero the only border is likewise an odd incoming pole, namely one.
To compare these borders, take an arbitrary incoming numerator N N N at pole 2 m − 1 2m-1 2 m − 1 and work over F 2 \mathbb{F}_2 F 2 . Write N q = a ( x 2 ) + x b ( x 2 ) Nq=a(x^2)+xb(x^2) N q = a ( x 2 ) + x b ( x 2 ) . The two raw sections at denominator q m q^m q m have numerator rows ( a , x b ) (a,xb) ( a , x b ) and ( b , a ) (b,a) ( b , a ) for the input row ( N , x N ) (N,xN) ( N , x N ) . Multiplication by M 0 − 1 M_0^{-1} M 0 − 1 and M 1 − 1 M_1^{-1} M 1 − 1 gives
( a + x b , x b ) , ( b , a + b ) . (a+xb,xb),\qquad(b,a+b). ( a + x b , x b ) , ( b , a + b ) .
At a root ζ \zeta ζ of q q q in F 4 \mathbb{F}_4 F 4 , evaluation of the splitting identity at ζ 2 \zeta^2 ζ 2 gives a ( ζ ) + ζ 2 b ( ζ ) = 0 a(\zeta)+\zeta^2b(\zeta)=0 a ( ζ ) + ζ 2 b ( ζ ) = 0 . Since ζ 2 + ζ = 1 \zeta^2+\zeta=1 ζ 2 + ζ = 1 , the two displayed rows agree at ζ \zeta ζ , so their difference is divisible by q q q . The top pole of the forcing difference therefore drops from 2 k 2k 2 k to at most 2 k − 1 2k-1 2 k − 1 . Scalar normalization corrections from weight k − 1 k-1 k − 1 have pole at most
max { 1 , 2 k − 2 } ≤ 2 k − 1 ( k ≥ 1 ) , \max\{1,2k-2\}\le2k-1\qquad(k\ge1), max { 1 , 2 k − 2 } ≤ 2 k − 1 ( k ≥ 1 ) ,
including the boundary k = 1 k=1 k = 1 . At weight zero the pole-free normal remainder is a polynomial of degree at most one with its first two coefficients zero, and hence vanishes. This cancellation lifts integrally: divide a borderline numerator by the monic polynomial q q q , writing it as q M + 2 N 1 qM+2N_1 q M + 2 N 1 . The first term has the required reduced pole at the current weight; the second carries the old pole 2 k 2k 2 k to weight k + 1 k+1 k + 1 , where the allowed bound in 2 L 2\mathcal{L} 2 L is 2 k + 1 2k+1 2 k + 1 . At weight zero its carry is in 2 S 1 2S_1 2 S 1 , also allowed. Iterating at every weight proves Q 0 − Q 1 ∈ 2 L Q_0-Q_1\in2\mathcal{L} Q 0 − Q 1 ∈ 2 L , not merely a congruence at a fixed precision. Formula (B.2) now shows that all nonempty P w P_w P w are congruent modulo 4 L 2 4\mathcal{L}^2 4 L 2 . Finally P h − U ∈ 2 L 2 P_h-U\in2\mathcal{L}^2 P h − U ∈ 2 L 2 ; each common appended digit increases this factor by two by the exact difference calculation in Lemma B.2. Induction proves (B.6). □
The finite certificate
Encode a vector in the ordered basis (B.3) by the integer whose ten binary bits are its coordinates, with the first basis vector in bit zero. This encoding is used only for finite linear algebra; addition is bitwise exclusive-or. Let S r S_r S r be the scalar map [ z ] ↦ [ T r z − Q r L C r z ] [z]\mapsto[T_rz-Q_rLC_rz] [ z ] ↦ [ T r z − Q r L C r z ] in (B.4).
Lemma B.4. The columns of the two 10 × 10 10\times10 10 × 10 matrices S r S_r S r , in the stated encoding, are
S 0 = ( 1 , 3 , 29 , 0 , 58 , 0 , 28 , 0 , 56 , 0 ) , S_0=(1,3,29,0,58,0,28,0,56,0), S 0 = ( 1 , 3 , 29 , 0 , 58 , 0 , 28 , 0 , 56 , 0 ) , S 1 = ( 2 , 1 , 0 , 29 , 0 , 58 , 0 , 28 , 0 , 56 ) . S_1=(2,1,0,29,0,58,0,28,0,56). S 1 = ( 2 , 1 , 0 , 29 , 0 , 58 , 0 , 28 , 0 , 56 ) . Both forcing frames Q r Q_r Q r have column classes ( 477 , 954 ) (477,954) ( 477 , 954 ) . For every 0 ≤ u < 4 0\le u<4 0 ≤ u < 4 and 4 ≤ v < 8 4\le v<8 4 ≤ v < 8 , interpreting u , v u,v u , v as three-digit words with least significant digit applied first, both columns of ( P u − P v ) / 4 (P_u-P_v)/4 ( P u − P v ) /4 have class 39.
Proof. We give bounded exact arithmetic for all entries. With h h h , J ∗ J_* J ∗ as in Lemma B.1, the expansion there gives
A ≡ q 8 + 2 T q 6 − 4 h q 4 − 8 ( J ∗ + T h ) q 2 − 8 h 2 q 7 ( m o d 16 ) . (31) A\equiv\frac{q^8+2Tq^6-4hq^4-8(J_*+Th)q^2-8h^2}{q^7}\pmod{16}. \tag*{(31)} A ≡ q 7 q 8 + 2 T q 6 − 4 h q 4 − 8 ( J ∗ + T h ) q 2 − 8 h 2 ( mod 16 ) . (31)
Here it suffices to retain the terms of the square-root expansion whose valuation is less than four; expanding its two geometric factors gives the displayed expression. Its numerator, by increasing degree from zero to sixteen, is
( 1 , 6 , 6 , 8 , 2 , 14 , 2 , 6 , 1 , 6 , 10 , 14 , 0 , 0 , 12 , 14 , 15 ) ( m o d 16 ) . (1,6,6,8,2,14,2,6,1,6,10,14,0,0,12,14,15)\pmod{16}. ( 1 , 6 , 6 , 8 , 2 , 14 , 2 , 6 , 1 , 6 , 10 , 14 , 0 , 0 , 12 , 14 , 15 ) ( mod 16 ) .
For any numerator f f f at pole e e e , an alternative exact form of (B.5) is
Λ r ( f / q e ) = Λ r ( f q ( − x ) e ) q e . \Lambda_r(f/q^e)=\frac{\Lambda_r(fq(-x)^e)}{q^e}. Λ r ( f / q e ) = q e Λ r ( f q ( − x ) e ) .
Apply this to (B.7) times each of the ten basis vectors and to its two polynomial columns U U U . The first-row matrices of C r U C_rU C r U modulo sixteen are
L C 0 U = ( 1 0 1 15 ) , L C 1 U = ( 15 1 0 1 ) . LC_0U=\begin{pmatrix}1&0\\1&15\end{pmatrix},\qquad
LC_1U=\begin{pmatrix}15&1\\0&1\end{pmatrix}. L C 0 U = ( 1 1 0 15 ) , L C 1 U = ( 15 0 1 1 ) .
Subtract the first two coefficients, divide by two, and reduce to (B.3). This yields the following complete column calculation. The columns headed T r T_r T r are encoded in the same basis; the only nonzero columns of L C r LC_r L C r modulo two are L C 0 ( e 1 ) = ( 0 , 1 ) t LC_0(e_1)=(0,1)^t L C 0 ( e 1 ) = ( 0 , 1 ) t and L C 1 ( e 2 ) = ( 0 , 1 ) t LC_1(e_2)=(0,1)^t L C 1 ( e 2 ) = ( 0 , 1 ) t .
Table 5.
The last row lists the two forcing columns rather than ten map columns. Thus subtracting the forcing contribution changes 955 to 955 xor 954 = 1 955\mathbin{\operatorname{xor}}954=1 955 xor 954 = 1 , and leaves all other entries unchanged. This proves both matrices and the forcing classes.
For completeness the seed calculation can be performed without rational square roots. Write B = ∑ b n x n B=\sum b_nx^n B = ∑ b n x n , set b 0 = 1 b_0=1 b 0 = 1 and b n = 0 b_n=0 b n = 0 for n < 0 n<0 n < 0 , and use 2 D B ′ + D ′ B = 0 2DB'+D'B=0 2 D B ′ + D ′ B = 0 to compute
2 n b n = − ∑ j = 1 4 d j ( 2 n − j ) b n − j , ( d 1 , d 2 , d 3 , d 4 ) = ( 2 , 3 , 2 , − 3 ) . (32) 2nb_n=-\sum_{j=1}^{4}d_j(2n-j)b_{n-j},\qquad(d_1,d_2,d_3,d_4)=(2,3,2,-3). \tag*{(32)} 2 n b n = − j = 1 ∑ 4 d j ( 2 n − j ) b n − j , ( d 1 , d 2 , d 3 , d 4 ) = ( 2 , 3 , 2 , − 3 ) . (32)
This is an exact integer recurrence; integrality also follows from B = q − 1 ∑ k ≥ 0 ( 2 k k ) ( x 4 / q 2 ) k B=q^{-1}\sum_{k\geq0}\binom{2k}{k}(x^4/q^2)^k B = q − 1 ∑ k ≥ 0 ( k 2 k ) ( x 4 / q 2 ) k . For each 0 ≤ v < 8 0\leq v<8 0 ≤ v < 8 form the first twelve coefficients of
( ∑ n ≥ 0 b 8 n + v x n , ∑ n ≥ 0 b 8 n + v − 1 x n ) . \left(\sum_{n\geq0}b_{8n+v}x^n,\sum_{n\geq0}b_{8n+v-1}x^n\right). ( n ≥ 0 ∑ b 8 n + v x n , n ≥ 0 ∑ b 8 n + v − 1 x n ) .
Multiply by B − 1 B^{-1} B − 1 and normalize its first two rows modulo thirty-two. Subtract any column for u < 4 u<4 u < 4 from that for v ≥ 4 v\geq4 v ≥ 4 , divide by four, and eliminate the pivots in (B.3). The sixteen resulting pairs of classes are
Table 1 (PDF p. 36).
This is a finite certificate: (B.8) requires only b 0 , … , b 95 b_0,\ldots,b_{95} b 0 , … , b 95 , and inversion and normalization use only twelve coefficients. Its sufficiency follows from the previous analytic lemmas. Every normal lattice element modulo eight has denominator q 5 q^5 q 5 and numerator degree at most eleven. Since q 5 ( 0 ) = 1 q^5(0)=1 q 5 ( 0 ) = 1 , its first twelve coefficients determine that numerator uniquely modulo eight. Thus no possible further coefficient can alter the quotient calculation. Division by two for the maps needs precision sixteen; division by four for seeds needs precision thirty-two, exactly as used above. These formulas, all matrix entries and the sixteen seeds are reproduced by the self-contained script in anc/exponent_three/.
Lemma B.5. Every product of the scalar maps S 0 , S 1 S_0,S_1 S 0 , S 1 sends the seed 39 to a nonzero vector. Every actual lattice element whose class is one of these images has physical valuation at most one.
Proof. Let W = ⟨ 1 , 2 , 28 , 36 ⟩ W=\langle1,2,28,36\rangle W = ⟨ 1 , 2 , 28 , 36 ⟩ and N = ⟨ 1 , 2 ⟩ N=\langle1,2\rangle N = ⟨ 1 , 2 ⟩ in Q \mathbb{Q} Q . The four generators of W W W are independent, and direct addition of the certified columns gives
Table 2 (PDF p. 36).
Thus both spaces are invariant. In the ordered basis [ 28 ] , [ 36 ] [28],[36] [ 28 ] , [ 36 ] of W / N W/N W / N the induced matrices are
S ‾ 0 = ( 0 1 1 0 ) , S ‾ 1 = ( 1 1 0 1 ) . \overline{S}_0=\begin{pmatrix}0&1\\1&0\end{pmatrix},\qquad\overline{S}_1=\begin{pmatrix}1&1\\0&1\end{pmatrix}. S 0 = ( 0 1 1 0 ) , S 1 = ( 1 0 1 1 ) .
They are invertible. Since 39 = 36 + 1 + 2 39=36+1+2 39 = 36 + 1 + 2 projects to the nonzero second basis vector, no product annihilates it.
Every vector of W W W has zero coordinates in the final four basis positions. If at least one of its first two coordinates is nonzero, any actual representative has a coefficient odd modulo two. If these coordinates are zero and its class is nonzero, its next four coordinates represent a nonzero element of 2 S 3 / ( 2 S 1 + 4 S 3 ) 2S_3/(2S_1+4S_3) 2 S 3 / ( 2 S 1 + 4 S 3 ) . Terms from 2 L 2L 2 L contribute only 2 S 1 2S_1 2 S 1 modulo four; terms from H 3 H_3 H 3 are divisible by eight. They cannot cancel this nonzero middle-layer class. The representative therefore has a coefficient nonzero modulo four. This proves the physical valuation bound. For a two-column frame the unit matrix on the right of (B.4) preserves the column span, so the same nonannihilation conclusion applies to its common one-dimensional seed span. □
Proposition B.6. For distinct full windows w , v w,v w , v of length α − 1 \alpha-1 α − 1 with longest common terminal suffix of length ℓ \ell ℓ , their clipped physical contact satisfies
min { α , ν 2 ( E w − E v ) } ≤ min { α , ℓ + 3 } . \min\{\alpha,\nu_2(E_w-E_v)\}\le\min\{\alpha,\ell+3\}. min { α , ν 2 ( E w − E v )} ≤ min { α , ℓ + 3 } . Proof. Only ℓ + 3 < α \ell+3<\alpha ℓ + 3 < α needs proof. Retain the last ℓ + 3 \ell+3 ℓ + 3 digits in each word. Their three-digit prefixes have different last digits and hence belong, in either order, to one of the sixteen seed pairs in Lemma B.4 . By Lemma B.3 their difference is 4 Z 0 4Z_0 4 Z 0 , with columns in L \mathcal{L} L and common seed class 39. Append the common ℓ \ell ℓ -digit suffix. Repeated use of (B.4) makes their difference 2 ℓ + 2 Z ℓ 2^{\ell+2}Z_\ell 2 ℓ + 2 Z ℓ . Lemma B.5 supplies a nonzero column class with physical valuation at most one, so the retained windows have contact at most ℓ + 3 \ell+3 ℓ + 3 . History erasure (B.6), applied to the retained word of length ℓ + 3 \ell+3 ℓ + 3 , changes either frame by an element divisible by 2 ℓ + 4 2^{\ell+4} 2 ℓ + 4 . It therefore preserves this witness of smaller valuation. The same contact bound holds for the original windows, and the fixed physical normalization preserves it. The cases with ℓ + 3 ≥ α \ell+3\ge\alpha ℓ + 3 ≥ α , including the shorter windows, are immediate from clipping. □
Data and code
The ancillary directory anc/ contains the independent rational-Cartier construction, the exact history and polynomial constructions, the support certificates, scripts checking the finite identities used in the proofs, the rational-lattice certificate for Appendix B, and exact results and run logs. Its README.md gives reproduction commands and explains the graph formats; SHA256SUMS identifies every included file. Large complete graphs are deterministically regenerated by the supplied code, and their SHA-256 values are recorded with the computation results. No probabilistic state identification or approximate polynomial arithmetic is used in these computations.
Independent researcher, Haifa, Israel
Email address: alex@shvets.io