Introduction
The classical Fourier series on an interval expresses a square-integrable function as a superposition of mutually orthogonal exponentials. The spectral set problem asks which other domains admit such an expansion and how the geometry of a domain is reflected in its frequencies. A measurable set K ⊂ R d K \subset\mathbb{R}^{d} K ⊂ R d with 0 < ∣ K ∣ < ∞ 0 < |K| < \infty 0 < ∣ K ∣ < ∞ is called a spectral set if there exists a countable set Λ ⊂ R d \Lambda\subset\mathbb{R}^{d} Λ ⊂ R d such that
∣ K ∣ − 1 / 2 e 2 π i ⟨ λ , x ⟩ : λ ∈ Λ |K|^{-1/2}e^{2\pi i\langle\lambda,x\rangle}:\lambda\in\Lambda ∣ K ∣ − 1/2 e 2 π i ⟨ λ , x ⟩ : λ ∈ Λ
is an orthonormal basis of L 2 ( K ) L^{2}(K) L 2 ( K ) . It is a translational tile if there is a discrete set A ⊂ R d A \subset\mathbb{R}^{d} A ⊂ R d satisfying
∑ a ∈ A 1 K ( x − a ) = 1 for almost every x ∈ R d . \sum_{a\in A}\mathbf{1}_{K}(x-a)=1 \quad\text{for almost every }x\in\mathbb{R}^{d}. a ∈ A ∑ 1 K ( x − a ) = 1 for almost every x ∈ R d .
Thus the translates cover the ambient space without overlaps of positive measure. Fuglede’s conjecture [15 ] , proposed in 1974, asserts that spectral sets and translational tiles are exactly the same sets. For a fundamental domain of a lattice, the dual lattice provides the familiar exponential basis, illustrating the connection between the two properties.
2010 Mathematics Subject Classification. Primary 28A80; Secondary 42C05.
Key words and phrases. self-similar measure, spectral eigenvalue, Beurling dimension, Beurling density, spectrum.
The research is supported in part by the NNSF of China (Nos. 12501111, 11831007 and 12071125), the Hunan Provincial NSF (No. 2025JJ60065), the Fundamental Research Funds for the Central Universities.
The conjecture stimulated extensive interaction among Fourier analysis, discrete geometry, number theory, and finite-group combinatorics. Tao [35 ] constructed spectral sets that do not tile in dimensions at least five. Kolountzakis and Matolcsi [20 ] constructed tiles without spectra; further counterexamples in dimension four were obtained in [13 , 31 ] . These developments show that neither implication holds for unrestricted sets in general. Geometric restrictions, however, can recover a precise correspondence. For example, Laba [22 ] established the conjecture for unions of two intervals, and Lev and Matolcsi [25 ] proved it for convex bodies in every dimension. These results highlight the role of geometric and arithmetic structure in the relation between spectrality and tiling.
The measure-theoretic formulation extends this question beyond sets of positive Lebesgue measure. A Borel probability measure μ \mu μ on R \mathbb{R} R is called a spectral measure if there is a countable set Λ ⊂ R \Lambda\subset\mathbb{R} Λ ⊂ R such that
E ( Λ ) = { e λ ( x ) = e 2 π i λ x : λ ∈ Λ } E(\Lambda)=\{e_{\lambda}(x)=e^{2\pi i\lambda x}:\lambda\in\Lambda\} E ( Λ ) = { e λ ( x ) = e 2 π iλ x : λ ∈ Λ }
is an orthonormal basis for L 2 ( μ ) L^{2}(\mu) L 2 ( μ ) . Such a set Λ \Lambda Λ is called a spectrum of μ \mu μ . Normalized Lebesgue measure on a spectral set fits this definition. A singular measure, in contrast, is concentrated on a set of Lebesgue measure zero, so the existence of an exponential basis cannot be explained by the usual volume-based tiling picture.
Jorgensen and Pedersen [19 ] showed that this distinction does not prevent Fourier expansions. The self-similar measure with contraction 1 / 4 1/4 1/4 and digits { 0 , 2 } \{0,2\} { 0 , 2 } is spectral, with spectrum
{ ∑ j = 0 n − 1 ε j 4 j : n ≥ 1 , ε j ∈ { 0 , 1 } } , \left\{\sum_{j=0}^{n-1}\varepsilon_j4^j:n\geq1,\ \varepsilon_j\in\{0,1\}\right\}, { j = 0 ∑ n − 1 ε j 4 j : n ≥ 1 , ε j ∈ { 0 , 1 } } ,
whereas the standard middle-third Cantor measure is not spectral. These examples demonstrate that self-similarity alone does not determine spectrality: the arithmetic relation between the contraction and the digit set is essential. They led to systematic studies of self-similar, self-affine, and infinite-convolution measures, including the spectrality results in [1 , 4 , 5 , 8 , 12 , 23 , 28 , 29 ] .
Singular spectral measures are therefore natural objects for studying Fourier analysis on fractal supports. An exponential basis gives coefficients, Parseval identities, and L 2 L^{2} L 2 expansions for functions on such supports, while questions about convergence of the corresponding Fourier series lead to additional analytic problems; see [9 , 33 , 34 ] . At the same time, their spectra connect the small-scale structure of a measure with the large-scale distribution of a discrete frequency set. In the Lebesgue setting, Landau’s density theory [24 ] ties sampling and interpolation to the volume of the underlying domain. For singular measures, Beurling dimension describes the polynomial growth of frequencies; see the study of exponential frames in [10 ] .
For a fixed spectral measure, one may further ask which growth exponents and densities its spectra can attain, and which dilations preserve spectrality. We study these questions simultaneously for singular self-similar measures: we prescribe both the Beurling dimension and the upper density while requiring a family of integer dilates to remain spectra.
We consider self-similar probability measures on the line. Let q ≥ 2 q \ge2 q ≥ 2 be an integer and let D ⊂ Z D \subset\mathbb{Z} D ⊂ Z be finite. The associated measure is the unique probability measure satisfying
μ q , D ( A ) = 1 # D ∑ d ∈ D μ q , D ( q A − d ) ( A ⊂ R Borel ) ; (1) \mu_{q,D}(A)=\frac{1}{\#D}\sum_{d\in D}\mu_{q,D}(qA-d)\qquad(A\subset\mathbb{R}\text{ Borel}); \tag*{(1)} μ q , D ( A ) = # D 1 d ∈ D ∑ μ q , D ( q A − d ) ( A ⊂ R Borel ) ; (1)
see [18 ] . Equivalently, μ q , D = ∗ j = 1 ∞ δ q − j D \mu_{q,D}=\ast_{j=1}^{\infty}\delta_{q^{-j}D} μ q , D = ∗ j = 1 ∞ δ q − j D , where δ F = ( # F ) − 1 ∑ x ∈ F δ x \delta_F=(\#F)^{-1}\sum_{x\in F}\delta_x δ F = ( # F ) − 1 ∑ x ∈ F δ x for a finite nonempty set F F F . A useful algebraic condition for spectrality is the following.
Definition 1.1. Let q ≥ 2 q \ge2 q ≥ 2 be an integer and let D , L ⊂ Z D,L\subset\mathbb{Z} D , L ⊂ Z satisfy # D = # L = m \#D=\#L=m # D = # L = m . The triple ( q , D , L ) (q,D,L) ( q , D , L ) is a Hadamard triple if
1 m ( e 2 π i d ℓ / q ) d ∈ D , ℓ ∈ L \frac{1}{\sqrt{m}}\left(e^{2\pi i d\ell/q}\right)_{d\in D,\ell\in L} m 1 ( e 2 π i d ℓ / q ) d ∈ D , ℓ ∈ L is unitary.
For one-dimensional Hadamard triples, spectrality of the associated self-similar measures was established in [23 ] . The higher-dimensional Hadamard triple theorem of Dutkay, Haussermann, and Lai [12 ] establishes spectrality for the corresponding self-affine measures. Throughout this paper we assume
2 ≤ m : = # D = # L < q , r 0 : = log m log q . 2\le m:=\#D=\#L<q,\qquad r_0:=\frac{\log m}{\log q}. 2 ≤ m := # D = # L < q , r 0 := log q log m .
The existence of a spectrum leaves considerable freedom in the choice of its frequencies. One aspect of this freedom concerns spectral eigenvalues : numbers t t t for which a spectrum and its dilate t Λ t\Lambda t Λ are both spectra of the same measure. This problem has been studied in [6 , 11 , 14 , 17 , 21 ] . For a set A A A of nonzero real numbers containing 1 1 1 , write
V ( μ , A ) = { Λ : a Λ is a spectrum of μ for every a ∈ A } . V(\mu,A)=\{\Lambda:a\Lambda\text{ is a spectrum of }\mu\text{ for every }a\in A\}. V ( μ , A ) = { Λ : a Λ is a spectrum of μ for every a ∈ A } .
We call this collection a spectral eigen-subspace ; the term does not imply a vector-space structure. The distinct spectra constructed below all have minimum zero and are therefore inequivalent under translation.
The multiplier set relevant to a Hadamard triple is
T = { t ∈ Z : ( q , D , t L ) is a Hadamard triple } . (2) \mathcal{T}=\{t\in\mathbb{Z}:(q,D,tL)\text{ is a Hadamard triple}\}. \tag*{(2)} T = { t ∈ Z : ( q , D , t L ) is a Hadamard triple } . (2)
Here 1 ∈ T 1\in\mathcal{T} 1 ∈ T , 0 ∉ T 0\notin\mathcal{T} 0 ∈ / T , and T = − T \mathcal{T}=-\mathcal{T} T = − T . Deng, Fu and Kang [7 ] studied common spectra for the odd integer multipliers of the quarter Cantor measure. The results of [30 ] establish common spectra and Beurling dimension interpolation for the multiplier set (1.2).
To describe the distribution of frequencies more precisely, put B ( x , h ) = ( x − h , x + h ) B(x,h)=(x-h,x+h) B ( x , h ) = ( x − h , x + h ) and define, for s > 0 s>0 s > 0 ,
M Λ ( h ) = sup x ∈ R # ( Λ ∩ B ( x , h ) ) , D s + ( Λ ) = lim sup h → ∞ M Λ ( h ) h s . M_\Lambda(h)=\sup_{x\in\mathbb{R}}\#(\Lambda\cap B(x,h)),\qquad D_s^+(\Lambda)=\limsup_{h\to\infty}\frac{M_\Lambda(h)}{h^s}. M Λ ( h ) = x ∈ R sup # ( Λ ∩ B ( x , h )) , D s + ( Λ ) = h → ∞ lim sup h s M Λ ( h ) .
The denominator is the s s s th power of the radius ; this convention is used consistently throughout. The Beurling dimension is
dim B e Λ = inf { u > 0 : D u + ( Λ ) = 0 } = sup { u > 0 : D u + ( Λ ) > 0 } , \dim_{\mathrm{Be}}\Lambda=\inf\{u>0:D_u^+(\Lambda)=0\}=\sup\{u>0:D_u^+(\Lambda)>0\}, dim Be Λ = inf { u > 0 : D u + ( Λ ) = 0 } = sup { u > 0 : D u + ( Λ ) > 0 } ,
with the usual convention for dimension zero. Beurling dimension quantifies polynomial growth, whereas density at that dimension retains information that the exponent alone does not record. Dimension estimates for spectra of self-similar measures were studied in [10 , 16 ] ; see also the entropy bound in [32 ] . Constructions of sparse spectra and dimension interpolation appear in [3 , 26 , 27 ] .
Theorem A ([30 ] ). Let ( q , D , L ) (q,D,L) ( q , D , L ) be a one-dimensional Hadamard triple with 2 ≤ m < q 2 \le m < q 2 ≤ m < q . For every s ∈ [ 0 , r 0 ] s \in[0,r_0] s ∈ [ 0 , r 0 ] , the collection
V ( s ) ( μ q , D , T ) = { Λ ∈ V ( μ q , D , T ) : dim B e Λ = s } V^{(s)}(\mu_{q,D},\mathcal{T})=\{\Lambda\in V(\mu_{q,D},\mathcal{T}):\dim_{\mathrm{Be}}\Lambda=s\} V ( s ) ( μ q , D , T ) = { Λ ∈ V ( μ q , D , T ) : dim Be Λ = s } has cardinality 2 ℵ 0 2^{\aleph_0} 2 ℵ 0 .
This result prescribes the growth exponent. It does not prescribe the upper density at that exponent: a set of dimension s s s may have zero, finite positive, or infinite upper s s s -Beurling density. We therefore ask the following finer question.
Question 1.1. For 0 < s < r 0 0<s<r_0 0 < s < r 0 and ζ ∈ [ 0 , + ∞ ] \zeta\in[0,+\infty] ζ ∈ [ 0 , + ∞ ] , does
V ζ ( s ) ( μ q , D , T ) = { Λ ∈ V ( μ q , D , T ) : dim B e Λ = s , D s + ( Λ ) = ζ } (3) V^{(s)}_{\zeta}(\mu_{q,D},\mathcal{T})=\{\Lambda\in V(\mu_{q,D},\mathcal{T}):\dim_{\mathrm{Be}}\Lambda=s,\ D^{+}_{s}(\Lambda)=\zeta\} \tag*{(3)} V ζ ( s ) ( μ q , D , T ) = { Λ ∈ V ( μ q , D , T ) : dim Be Λ = s , D s + ( Λ ) = ζ } (3) have cardinality 2 ℵ 0 2^{\aleph_0} 2 ℵ 0 ?
Our main result answers Question 1.1 affirmatively.
Theorem 1.1. Let ( q , D , L ) (q,D,L) ( q , D , L ) be a one-dimensional Hadamard triple with 2 ≤ m : = # D < q 2\le m:=\#D<q 2 ≤ m := # D < q . Let μ q , D \mu_{q,D} μ q , D , T \mathcal{T} T , and V ζ ( s ) ( μ q , D , T ) V^{(s)}_{\zeta}(\mu_{q,D},\mathcal{T}) V ζ ( s ) ( μ q , D , T ) be defined by (1.1), (1.2), and (1.3), respectively. For every 0 < s < log m / log q 0<s<\log m/\log q 0 < s < log m / log q and ζ ∈ [ 0 , + ∞ ] \zeta\in[0,+\infty] ζ ∈ [ 0 , + ∞ ] ,
# V ζ ( s ) ( μ q , D , T ) = 2 ℵ 0 . \#V^{(s)}_{\zeta}(\mu_{q,D},\mathcal{T})=2^{\aleph_0}. # V ζ ( s ) ( μ q , D , T ) = 2 ℵ 0 . The spectra may be chosen in Z ≥ 0 \mathbb{Z}_{\ge0} Z ≥ 0 with minimum zero.
The condition m < q m<q m < q cannot be omitted from the theorem as stated. For ( q , D , L ) = ( 2 , { 0 , 1 } , { 0 , 1 } ) (q,D,L)=(2,\{0,1\},\{0,1\}) ( q , D , L ) = ( 2 , { 0 , 1 } , { 0 , 1 }) , the measure is Lebesgue measure on [ 0 , 1 ] [0,1] [ 0 , 1 ] . After translating a spectrum to contain zero, its orthogonality forces it to be a subset of Z \mathbb{Z} Z . Completeness then forces it to equal Z \mathbb{Z} Z , so every spectrum has Beurling dimension one. Thus the asserted interpolation for 0 < s < 1 0<s<1 0 < s < 1 fails in this full-digit example.
The proof has two components. First, we construct an increasing family of digit words whose numerical images have the prescribed density. Long blocks of zero digits separate consecutive stages; a block of selected digit words is followed by a block of unrestricted digits. For each resulting finite set, its interval counts differ from a fixed proportion of the full digit set by at most twice the number of selected prefixes. This two-sided estimate permits exact density calibration. Second, we preserve these frequencies while adding lacunary frequencies to obtain completeness for every multiplier. The added frequencies contribute only O ( log h ) O(\log h) O ( log h ) points to any interval of radius h h h , so they do not change the upper u u u -Beurling density for any u > 0 u>0 u > 0 . Independent choices among them produce the required cardinality. The endpoint cases ζ = 0 \zeta=0 ζ = 0 and ζ = + ∞ \zeta=+\infty ζ = + ∞ are handled with slowly varying positive target densities.
Section 2 gives the Fourier tools. Section 3 establishes the recursive spectral extension. Section 4 constructs the prescribed-density digit sets and proves Theorem 1.1.
Preliminaries
We recall the finite Fourier identities and completeness criteria used in the construction. We state the cited results without proof and then derive the selection estimates needed for the recursive construction. We use the convention μ ^ ( ξ ) = ∫ e − 2 π i ξ x d μ ( x ) \widehat{\mu}(\xi)=\int e^{-2\pi i\xi x}\,d\mu(x) μ ( ξ ) = ∫ e − 2 π i ξ x d μ ( x ) . Write μ = μ q , D \mu=\mu_{q,D} μ = μ q , D and
μ n = ∗ j = 1 n δ q − j D , m D ( ξ ) = 1 m ∑ d ∈ D e − 2 π i d ξ . \mu_n=\mathop{*}_{j=1}^{n}\delta_{q^{-j}D},\qquad m_D(\xi)=\frac{1}{m}\sum_{d\in D}e^{-2\pi i d\xi}. μ n = ∗ j = 1 n δ q − j D , m D ( ξ ) = m 1 d ∈ D ∑ e − 2 π i d ξ .
Then
μ ^ ( ξ ) = ∏ j = 1 ∞ m D ( q − j ξ ) = μ n ^ ( ξ ) μ ^ ( q − n ξ ) . (4) \widehat{\mu}(\xi)=\prod_{j=1}^{\infty}m_D(q^{-j}\xi)=\widehat{\mu_n}(\xi)\widehat{\mu}(q^{-n}\xi). \tag*{(4)} μ ( ξ ) = j = 1 ∏ ∞ m D ( q − j ξ ) = μ n ( ξ ) μ ( q − n ξ ) . (4)
In particular, μ n ^ \widehat{\mu_n} μ n is q n q^n q n -periodic. A set Λ \Lambda Λ is orthogonal for μ \mu μ if and only if
( Λ − Λ ) ∖ { 0 } ⊂ Z ( μ ^ ) , Z ( f ) = { x : f ( x ) = 0 } . (5) (\Lambda-\Lambda)\setminus\{0\}\subset\mathcal{Z}(\widehat{\mu}),\qquad\mathcal{Z}(f)=\{x:f(x)=0\}. \tag*{(5)} ( Λ − Λ ) ∖ { 0 } ⊂ Z ( μ ) , Z ( f ) = { x : f ( x ) = 0 } . (5)
The following standard consequences of Hadamard unitarity connect the finite Fourier matrix with the spectra of the partial convolutions.
Lemma 2.1 ([23]). Let q ≥ 2 q\ge2 q ≥ 2 and D , L ⊂ Z D,L\subset\mathbb{Z} D , L ⊂ Z with # D = # L = m \#D=\#L=m # D = # L = m . The following are equivalent:
(i) ( q , D , L ) (q,D,L) ( q , D , L ) is a Hadamard triple;
(ii) the elements of L L L are distinct modulo q q q , and replacing them by their residues gives a Hadamard triple;
(iii) L L L is a spectrum of δ q − 1 D \delta_{q^{-1}D} δ q − 1 D ;
(iv) ∑ ℓ ∈ L ∣ m D ( ( ξ + ℓ ) / q ) ∣ 2 = 1 \sum_{\ell\in L}|m_D((\xi+\ell)/q)|^2=1 ∑ ℓ ∈ L ∣ m D (( ξ + ℓ ) / q ) ∣ 2 = 1 for all ξ ∈ R \xi\in\mathbb{R} ξ ∈ R .
Moreover, if t ∈ T t\in\mathbb{T} t ∈ T , then t ( L + q L + ⋯ + q n − 1 L ) t(L+qL+\cdots+q^{n-1}L) t ( L + q L + ⋯ + q n − 1 L ) is a spectrum of μ n \mu_n μ n .
For the infinite measure, we will need Fourier lower bounds that remain valid under small perturbations. The following continuity property supplies this stability.
Lemma 2.2 ([1, Lemma 2.2]). Let μ \mu μ be a compactly supported Borel probability measure on R \mathbb{R} R . Then μ ^ \widehat{\mu} μ is uniformly continuous on R \mathbb{R} R .
Finite-level orthogonality alone does not guarantee completeness of the limiting system. The next criterion supplements nested finite spectra with a uniform lower bound for the tail Fourier transform. It follows from the argument of [2 Theorem 2.3] , together with Lemma 2.2, and will be applied along a subsequence on which a prescribed multiplier is fixed.
Theorem 2.3. Let μ : = μ q , D \mu:=\mu_{q,D} μ := μ q , D be given as (1) , and let b j → ∞ b_j\to\infty b j → ∞ be strictly increasing positive integers. Suppose 0 ∈ Λ b j ⊂ Λ b j + 1 0\in\Lambda_{b_j}\subset\Lambda_{b_{j+1}} 0 ∈ Λ b j ⊂ Λ b j + 1 , that Λ b j \Lambda_{b_j} Λ b j is a spectrum of μ b j \mu_{b_j} μ b j , and that
inf j inf λ ∈ Λ b j ∣ μ ^ ( q − b j λ ) ∣ 2 > 0. \inf_j\inf_{\lambda\in\Lambda_{b_j}}\left|\widehat{\mu}(q^{-b_j}\lambda)\right|^2>0. j inf λ ∈ Λ b j inf μ ( q − b j λ ) 2 > 0.
Then ⋃ j Λ b j \bigcup_j \Lambda_{b_j} ⋃ j Λ b j is a spectrum of μ \mu μ .
To meet the tail condition in Theorem 2.3 , we use the following selection result from [30 ] . It supplies two integer translations at every point of a compact interval, with a positive lower bound independent of that point.
Lemma 2.4 ([30 ] ). Let μ : = μ q , D \mu:= \mu_{q,D} μ := μ q , D and T \mathcal{T} T be given as (1.1) and (1.2), respectively, where ( q , D , L ) (q,D,L) ( q , D , L ) forms a Hadamard triple. Assume 2 ≤ m < q 2 \le m < q 2 ≤ m < q . For every t ∈ T ∩ N t \in\mathcal{T} \cap\mathbb{N} t ∈ T ∩ N there are ϵ t > 0 \epsilon_t > 0 ϵ t > 0 and an integer N t ≥ 1 N_t \ge1 N t ≥ 1 such that, for each x ∈ [ 0 , t ] x \in[0,t] x ∈ [ 0 , t ] , there are two integers 0 ≤ u 1 ( x , t ) < u 2 ( x , t ) < q N t 0 \le u_1(x,t) < u_2(x,t) < q^{N_t} 0 ≤ u 1 ( x , t ) < u 2 ( x , t ) < q N t satisfying
∣ μ ^ ( x + t u j ( x , t ) ) ∣ ≥ 2 ϵ t , j = 1 , 2. \left|\widehat{\mu}(x+tu_j(x,t))\right| \ge2\epsilon_t,\qquad j=1,2. ∣ μ ( x + t u j ( x , t )) ∣ ≥ 2 ϵ t , j = 1 , 2.
The strict inequality m < q m<q m < q is part of the hypotheses of this selection result. The constants may depend on the multiplier. To use the result for every positive multiplier, we arrange that each one is treated infinitely many times in the construction.
Definition 2.1. A sequence ( a k ) (a_k) ( a k ) in a countable set A A A is called a recurrent enumeration if each element of A A A occurs infinitely often.
We use a recurrent enumeration of T + : = T ∩ N \mathcal{T}_+ := \mathcal{T} \cap\mathbb{N} T + := T ∩ N . Negative multipliers will be recovered by complex conjugation.
We now combine Lemmas 2.2 and 2.4 into a form adapted to the recursive construction. Besides stability under perturbations, we arrange that the first translation is zero near the origin, so previously chosen frequencies can be retained.
Lemma 2.5. Let μ : = μ q , D \mu:= \mu_{q,D} μ := μ q , D and T \mathcal{T} T be given as (1.1) and (1.2), respectively, where ( q , D , L ) (q,D,L) ( q , D , L ) forms a Hadamard triple. For each t ∈ T + : = T ∩ N t \in\mathcal{T}_+ := \mathcal{T} \cap\mathbb{N} t ∈ T + := T ∩ N , there are numbers 0 < ϵ t < 1 / 2 0 < \epsilon_t < 1/2 0 < ϵ t < 1/2 , 0 < δ t < 1 / 4 0 < \delta_t < 1/4 0 < δ t < 1/4 , and an integer N t ≥ 1 N_t \ge1 N t ≥ 1 with t q − N t < δ t / 4 tq^{-N_t} < \delta_t/4 t q − N t < δ t /4 such that there exists an integer-valued map
η t : [ 0 , t ] × { 1 , 2 } ⟶ { 0 , 1 , … , q N t − 1 } \eta_t : [0,t] \times\{1,2\} \longrightarrow\{0,1,\ldots,q^{N_t}-1\} η t : [ 0 , t ] × { 1 , 2 } ⟶ { 0 , 1 , … , q N t − 1 } satisfying
η t ( x , 1 ) < η t ( x , 2 ) ( 0 ≤ x ≤ t ) \eta_t(x,1) < \eta_t(x,2)\qquad(0 \le x \le t) η t ( x , 1 ) < η t ( x , 2 ) ( 0 ≤ x ≤ t ) and
∣ μ ^ ( x + t η t ( x , j ) + y ) ∣ ≥ ϵ t ( ∣ y ∣ < δ t , j = 1 , 2 ) . \left|\widehat{\mu}(x+t\eta_t(x,j)+y)\right| \ge\epsilon_t\qquad(|y| < \delta_t,\ j=1,2). ∣ μ ( x + t η t ( x , j ) + y ) ∣ ≥ ϵ t ( ∣ y ∣ < δ t , j = 1 , 2 ) . The map and constants can be chosen to satisfy η t ( x , 1 ) = 0 \eta_t(x,1)=0 η t ( x , 1 ) = 0 for 0 ≤ x ≤ δ t 0 \le x \le\delta_t 0 ≤ x ≤ δ t and
∣ μ ^ ( v ) ∣ ≥ 1 2 whenever ∣ v ∣ < 2 δ t . \left|\widehat{\mu}(v)\right| \ge\frac{1}{2}\qquad\text{whenever } |v| < 2\delta_t. ∣ μ ( v ) ∣ ≥ 2 1 whenever ∣ v ∣ < 2 δ t . In particular, b t : = η t ( 0 , 2 ) b_t:=\eta_t(0,2) b t := η t ( 0 , 2 ) is a positive integer.
Proof. Fix t t t . Lemma 2.4 supplies a number ϵ t 0 > 0 \epsilon_t^0 > 0 ϵ t 0 > 0 , an integer N t 0 N_t^0 N t 0 , and two integers u 1 ( x ) < u 2 ( x ) u_1(x) < u_2(x) u 1 ( x ) < u 2 ( x ) such that ∣ μ ^ ( x + t u j ( x ) ) ∣ ≥ 2 ϵ t 0 \left|\widehat{\mu}(x+tu_j(x))\right| \ge2\epsilon_t^0 ∣ μ ( x + t u j ( x )) ∣ ≥ 2 ϵ t 0 . Choose 0 < ϵ t < min { ϵ t 0 , 1 / 4 } 0 < \epsilon_t < \min\{\epsilon_t^0,1/4\} 0 < ϵ t < min { ϵ t 0 , 1/4 } . By uniform continuity choose 0 < δ t < 1 / 4 0 < \delta_t < 1/4 0 < δ t < 1/4 sufficiently small that
∣ μ ^ ( z + y ) − μ ^ ( z ) ∣ < ϵ t ( z ∈ R , ∣ y ∣ < δ t ) . \left|\widehat{\mu}(z+y)-\widehat{\mu}(z)\right| < \epsilon_t\qquad(z \in\mathbb{R},\ |y| < \delta_t). ∣ μ ( z + y ) − μ ( z ) ∣ < ϵ t ( z ∈ R , ∣ y ∣ < δ t ) .
Decrease δ t \delta_t δ t again, if needed, to have ∣ μ ^ ( v ) ∣ ≥ 1 / 2 |\hat{\mu}(v)| \ge1/2 ∣ μ ^ ( v ) ∣ ≥ 1/2 for ∣ v ∣ < 2 δ t |v| < 2\delta_t ∣ v ∣ < 2 δ t . For either of the original choices, the triangle inequality gives
∣ μ ^ ( x + t u j ( x ) + y ) ∣ ≥ ∣ μ ^ ( x + t u j ( x ) ) ∣ − ∣ μ ^ ( x + t u j ( x ) + y ) − μ ^ ( x + t u j ( x ) ) ∣ ≥ 2 ϵ t 0 − ϵ t ≥ ϵ t . |\hat{\mu}(x+tu_j(x)+y)| \ge|\hat{\mu}(x+tu_j(x))| - |\hat{\mu}(x+tu_j(x)+y)-\hat{\mu}(x+tu_j(x))|
\ge2\epsilon_t^0-\epsilon_t \ge\epsilon_t. ∣ μ ^ ( x + t u j ( x ) + y ) ∣ ≥ ∣ μ ^ ( x + t u j ( x )) ∣ − ∣ μ ^ ( x + t u j ( x ) + y ) − μ ^ ( x + t u j ( x )) ∣ ≥ 2 ϵ t 0 − ϵ t ≥ ϵ t .
For x > δ t x>\delta_t x > δ t keep these choices. For 0 ≤ x ≤ δ t 0\le x\le\delta_t 0 ≤ x ≤ δ t put η t ( x , 1 ) = 0 \eta_t(x,1)=0 η t ( x , 1 ) = 0 and η t ( x , 2 ) = u 2 ( x ) \eta_t(x,2)=u_2(x) η t ( x , 2 ) = u 2 ( x ) . The second is positive because u 2 ( x ) > u 1 ( x ) ≥ 0 u_2(x)>u_1(x)\ge0 u 2 ( x ) > u 1 ( x ) ≥ 0 . The first satisfies the required bound since ∣ x + y ∣ < 2 δ t |x+y|<2\delta_t ∣ x + y ∣ < 2 δ t and ∣ μ ^ ( x + y ) ∣ ≥ 1 / 2 > ϵ t |\hat{\mu}(x+y)|\ge1/2>\epsilon_t ∣ μ ^ ( x + y ) ∣ ≥ 1/2 > ϵ t . Finally enlarge N t 0 N_t^0 N t 0 to an integer N t N_t N t with t q − N t < δ t / 4 tq^{-N_t}<\delta_t/4 t q − N t < δ t /4 . All selected integers still lie below q N t q^{N_t} q N t , and b t = η t ( 0 , 2 ) > 0 b_t=\eta_t(0,2)>0 b t = η t ( 0 , 2 ) > 0 as required. □
For each t ∈ T + t\in T_+ t ∈ T + , we henceforth fix ϵ t \epsilon_t ϵ t , δ t \delta_t δ t , N t N_t N t , and η t \eta_t η t satisfying all the properties in Lemma 2.5 .
Word spaces and spectral extension
This section explains how to extend a prescribed family of digit frequencies to a spectrum whose dilates by every t ∈ T t\in T t ∈ T are also spectra. We first normalize the digit set and establish the finite-stage orthogonality and Fourier estimates. We then give a recursive construction that retains the prescribed frequencies and adds a q q q -lacunary set of frequencies to ensure completeness. The binary encodings distinguish the resulting spectra; § 4 will show that the added frequencies do not affect the prescribed upper s s s -Beurling density for s > 0 s>0 s > 0 .
Lemma 3.1. A Hadamard triple with m < q m<q m < q admits a frequency digit set L ′ L' L ′ ⊂ \subset ⊂ { 0 , … , q − 2 } \{0,\ldots,q-2\} { 0 , … , q − 2 } containing zero. It has exactly the same multiplier set (1.2) as the original digit set L L L .
Proof. By Lemma 2.1, replace every element of L L L by its residue in { 0 , … , q − 1 } \{0,\ldots,q-1\} { 0 , … , q − 1 } . The resulting residue set is nonempty and proper because 1 ≤ m < q 1\le m<q 1 ≤ m < q .
There is an element c ∈ L c\in L c ∈ L whose predecessor c − 1 c-1 c − 1 modulo q q q does not belong to L L L . Indeed, if every element had its predecessor in L L L , starting from one element and repeatedly taking predecessors would place all q q q residues in L L L , a contradiction. Define
L ′ = ( L − c ) ( m o d q ) L'=(L-c)\pmod q L ′ = ( L − c ) ( mod q ) using residues in { 0 , … , q − 1 } \{0,\ldots,q-1\} { 0 , … , q − 1 } . The element c c c produces zero, whereas q − 1 q-1 q − 1 cannot occur because that would require c − 1 ∈ L c-1\in L c − 1 ∈ L . Thus 0 ∈ L ′ ⊂ { 0 , … , q − 2 } 0\in L'\subset\{0,\ldots,q-2\} 0 ∈ L ′ ⊂ { 0 , … , q − 2 } .
Translation by − t c -tc − t c multiplies row d d d of the matrix for ( q , D , t L ) (q,D,tL) ( q , D , t L ) by e − 2 π i d t c / q e^{-2\pi i dtc/q} e − 2 π i d t c / q , a scalar of modulus one. Replacing a translated digit by a congruent digit adds an integer multiple of 2 π i 2\pi i 2 π i to every exponent, leaving the entries unchanged. Both operations preserve unitarity in both directions. For every integer t t t , ( q , D , t L ) (q,D,tL) ( q , D , t L ) is therefore Hadamard if and only if ( q , D , t L ′ ) (q,D,tL') ( q , D , t L ′ ) is Hadamard. This proves equality of the multiplier sets. □
Henceforth use this normalization and write
L = { 0 = l 0 < l 1 < ⋯ < l m − 1 } ⊂ { 0 , … , q − 2 } , Ω = { 0 , … , m − 1 } . L=\{0=l_0<l_1<\cdots<l_{m-1}\}\subset\{0,\ldots,q-2\},\qquad\Omega=\{0,\ldots,m-1\}. L = { 0 = l 0 < l 1 < ⋯ < l m − 1 } ⊂ { 0 , … , q − 2 } , Ω = { 0 , … , m − 1 } .
For J = j 1 ⋯ j n ∈ Ω n J=j_1\cdots j_n\in\Omega^n J = j 1 ⋯ j n ∈ Ω n , put
T n ( J ) = ∑ r = 1 n l j r q r − 1 . T_n(J)=\sum_{r=1}^{n}l_{j_r}q^{r-1}. T n ( J ) = r = 1 ∑ n l j r q r − 1 .
Words are read from the least significant digit. For example, for q = 4 q=4 q = 4 and L = { 0 , 1 } L=\{0,1\} L = { 0 , 1 } , the word 101 has value T 3 ( 101 ) = 1 + 0 ⋅ 4 + 1 ⋅ 4 2 = 17 T_3(101)=1+0\cdot4+1\cdot4^2=17 T 3 ( 101 ) = 1 + 0 ⋅ 4 + 1 ⋅ 4 2 = 17 , whereas T 3 ( 110 ) = 5 T_3(110)=5 T 3 ( 110 ) = 5 . Thus a prefix specifies the low digits, and a suffix specifies the high digits; deleting a suffix removes the most significant digits. Appending zero digits does not change the numerical value of a word: T n + r ( J 0 r ) = T n ( J ) T_{n+r}(J0^r)=T_n(J) T n + r ( J 0 r ) = T n ( J ) for r ≥ 0 r\geq0 r ≥ 0 . We set T 0 ( ∅ ) = 0 T_0(\emptyset)=0 T 0 ( ∅ ) = 0 .
The next lemma provides the Fourier lower bounds used to prove completeness. We use η t \eta_t η t , b t b_t b t , ϵ t \epsilon_t ϵ t , and δ t \delta_t δ t from Lemma 2.5 .
Lemma 3.2. For t ∈ T + t\in\mathcal{T}_+ t ∈ T + , J ∈ Ω n J\in\Omega^n J ∈ Ω n , and i ∈ { 1 , 2 } i\in\{1,2\} i ∈ { 1 , 2 } , put
ψ n ( t ) ( J , i ) = T n ( J ) + q n η t ( t q − n T n ( J ) , i ) . (6) \psi_n^{(t)}(J,i)=T_n(J)+q^n\eta_t(tq^{-n}T_n(J),i). \tag*{(6)} ψ n ( t ) ( J , i ) = T n ( J ) + q n η t ( t q − n T n ( J ) , i ) . (6) Then ψ n ( t ) ( J , i ) ≡ T n ( J ) ( m o d q n ) \psi_n^{(t)}(J,i)\equiv T_n(J)\pmod{q^n} ψ n ( t ) ( J , i ) ≡ T n ( J ) ( mod q n ) and
∣ μ ^ ( y + t q − n ψ n ( t ) ( J , i ) ) ∣ ≥ ϵ t ( ∣ y ∣ < δ t ) . \left|\widehat{\mu}\left(y+tq^{-n}\psi_n^{(t)}(J,i)\right)\right|\geq\epsilon_t\qquad(|y|<\delta_t). μ ( y + t q − n ψ n ( t ) ( J , i ) ) ≥ ϵ t ( ∣ y ∣ < δ t ) . If t q − n T n ( J ) ≤ δ t tq^{-n}T_n(J)\leq\delta_t t q − n T n ( J ) ≤ δ t , then ψ n ( t ) ( J , 1 ) = T n ( J ) \psi_n^{(t)}(J,1)=T_n(J) ψ n ( t ) ( J , 1 ) = T n ( J ) . Moreover, if x ∈ [ 0 , t ] x\in[0,t] x ∈ [ 0 , t ] , ∣ y ∣ < δ t |y|<\delta_t ∣ y ∣ < δ t , u = η t ( x , j ) u=\eta_t(x,j) u = η t ( x , j ) for some j ∈ { 1 , 2 } j\in\{1,2\} j ∈ { 1 , 2 } , and z = x + y + t u z=x+y+tu z = x + y + t u , then, for every sufficiently large integer r r r ,
∣ μ ^ ( z + t q r b t ) ∣ ≥ ϵ t 2 . \left|\widehat{\mu}(z+tq^rb_t)\right|\geq\epsilon_t^2. ∣ μ ( z + t q r b t ) ∣ ≥ ϵ t 2 . The threshold for r r r may be increased arbitrarily.
Proof. The digits lie between zero and q − 2 q-2 q − 2 , so
0 ≤ T n ( J ) ≤ ( q − 2 ) q n − 1 q − 1 < q n . 0\leq T_n(J)\leq(q-2)\frac{q^n-1}{q-1}<q^n. 0 ≤ T n ( J ) ≤ ( q − 2 ) q − 1 q n − 1 < q n .
Thus x = t q − n T n ( J ) x=tq^{-n}T_n(J) x = t q − n T n ( J ) belongs to [ 0 , t ] [0,t] [ 0 , t ] , the domain of η t \eta_t η t . Formula (6) immediately gives the congruence modulo q n q^n q n . Moreover,
y + t q − n ψ n ( t ) ( J , i ) = y + x + t η t ( x , i ) . y+tq^{-n}\psi_n^{(t)}(J,i)=y+x+t\eta_t(x,i). y + t q − n ψ n ( t ) ( J , i ) = y + x + t η t ( x , i ) .
Its Fourier modulus is at least ϵ t \epsilon_t ϵ t by Lemma 2.5 . If x ≤ δ t x\leq\delta_t x ≤ δ t , that lemma allows η t ( x , 1 ) = 0 \eta_t(x,1)=0 η t ( x , 1 ) = 0 , proving the zero-correction assertion.
For the last part, Lemma 2.5 gives ∣ μ ^ ( z ) ∣ ≥ ϵ t |\widehat{\mu}(z)|\geq\epsilon_t ∣ μ ( z ) ∣ ≥ ϵ t for either choice j ∈ { 1 , 2 } j\in\{1,2\} j ∈ { 1 , 2 } . Choose r r r so large that ∣ q − r z ∣ < δ t |q^{-r}z|<\delta_t ∣ q − r z ∣ < δ t . Since t b t t b_t t b t is an integer and μ ^ r \widehat{\mu}_r μ r is q r q^r q r -periodic,
∣ μ ^ ( z + t q r b t ) ∣ = ∣ μ ^ r ( z + t q r b t ) ∣ ∣ μ ^ ( q − r z + t b t ) ∣ = ∣ μ ^ r ( z ) ∣ ∣ μ ^ ( q − r z + t b t ) ∣ . \left|\widehat{\mu}(z+tq^rb_t)\right|=\left|\widehat{\mu}_r(z+tq^rb_t)\right|\left|\widehat{\mu}(q^{-r}z+tb_t)\right|=\left|\widehat{\mu}_r(z)\right|\left|\widehat{\mu}(q^{-r}z+tb_t)\right|. ∣ μ ( z + t q r b t ) ∣ = ∣ μ r ( z + t q r b t ) ∣ μ ( q − r z + t b t ) = ∣ μ r ( z ) ∣ μ ( q − r z + t b t ) .
From (4) it follows that ∣ μ ^ r ( z ) ∣ ≥ ∣ μ ^ ( z ) ∣ ≥ ϵ t |\widehat{\mu}_r(z)|\geq|\widehat{\mu}(z)|\geq\epsilon_t ∣ μ r ( z ) ∣ ≥ ∣ μ ( z ) ∣ ≥ ϵ t . The other factor is at least ϵ t \epsilon_t ϵ t , using x = 0 x=0 x = 0 , the second choice b t b_t b t , and perturbation q − r z q^{-r}z q − r z in Lemma 2.5 . Multiplication gives ϵ t 2 \epsilon_t^2 ϵ t 2 . All larger r r r also satisfy ∣ q − r z ∣ < δ t |q^{-r}z|<\delta_t ∣ q − r z ∣ < δ t , so any additional finite lower bound on r r r can be imposed. □
Definition 3.1. A family J n ⊂ Ω n \mathcal{J}_n\subset\Omega^n J n ⊂ Ω n is an increasing word sequence if 0 n ∈ J n 0^n\in\mathcal{J}_n 0 n ∈ J n and
J n 0 ⊂ J n + 1 , ( Ω n ∖ J n ) 0 ⊂ Ω n + 1 ∖ J n + 1 . \mathcal{J}_n0\subset\mathcal{J}_{n+1},\qquad(\Omega^n\setminus\mathcal{J}_n)0\subset\Omega^{n+1}\setminus\mathcal{J}_{n+1}. J n 0 ⊂ J n + 1 , ( Ω n ∖ J n ) 0 ⊂ Ω n + 1 ∖ J n + 1 . It is prefix closed if in addition J n + 1 ⊂ J n Ω \mathcal{J}_{n+1}\subset\mathcal{J}_n\Omega J n + 1 ⊂ J n Ω . Set J 0 = { ∅ } \mathcal{J}_0=\{\emptyset\} J 0 = { ∅ } and Λ ( J ) = ⋃ n T n ( J n ) \Lambda(\mathcal{J})=\bigcup_n T_n(\mathcal{J}_n) Λ ( J ) = ⋃ n T n ( J n ) .
Prefix closure is an additional hypothesis wherever used below. All word sequences constructed in Section 4 have this property.
Lemma 3.3. Let 0 = m 0 < m 1 < m 2 < ⋯ 0=m_{0}<m_{1}<m_{2}<\cdots 0 = m 0 < m 1 < m 2 < ⋯ , n k = m k − m k − 1 n_{k}=m_{k}-m_{k-1} n k = m k − m k − 1 , and λ 0 ( ∅ ) = 0 \lambda_{0}(\varnothing)=0 λ 0 ( ∅ ) = 0 . Define integers recursively by
λ k ( P J ) = λ k − 1 ( P ) + q m k − 1 T n k ( J ) + q m k c k ( P , J ) , P ∈ Ω m k − 1 , J ∈ Ω n k , \lambda_{k}(PJ)=\lambda_{k-1}(P)+q^{m_{k-1}}T_{n_{k}}(J)+q^{m_{k}}c_{k}(P,J),\qquad P\in\Omega^{m_{k-1}},\quad J\in\Omega^{n_{k}}, λ k ( P J ) = λ k − 1 ( P ) + q m k − 1 T n k ( J ) + q m k c k ( P , J ) , P ∈ Ω m k − 1 , J ∈ Ω n k , where c k ( P , J ) ∈ Z c_{k}(P,J)\in\mathbb{Z} c k ( P , J ) ∈ Z and c k ( P , 0 n k ) = 0 c_{k}(P,0^{n_{k}})=0 c k ( P , 0 n k ) = 0 . The other carries may depend on the whole node ( P , J ) (P,J) ( P , J ) and on all earlier choices. Then Λ k = { λ k ( W ) : W ∈ Ω m k } \Lambda_{k}=\{\lambda_{k}(W):W\in\Omega^{m_{k}}\} Λ k = { λ k ( W ) : W ∈ Ω m k } are nested, and t Λ k t\Lambda_{k} t Λ k is a spectrum of μ m k \mu_{m_{k}} μ m k for every t ∈ T t\in T t ∈ T .
Proof. We prove the lemma by induction on k k k . A zero suffix gives
λ k ( P 0 n k ) = λ k − 1 ( P ) + q m k − 1 ⋅ 0 + q m k ⋅ 0 = λ k − 1 ( P ) . \lambda_{k}(P0^{n_{k}})=\lambda_{k-1}(P)+q^{m_{k-1}}\cdot0+q^{m_{k}}\cdot0=\lambda_{k-1}(P). λ k ( P 0 n k ) = λ k − 1 ( P ) + q m k − 1 ⋅ 0 + q m k ⋅ 0 = λ k − 1 ( P ) . Thus every old frequency is retained and zero is retained at every stage. This proves nestedness.
Fix t ∈ T t\in T t ∈ T . Suppose the conclusion has been proved through stage k − 1 k-1 k − 1 , and compare different nodes P J PJ P J and P ′ J ′ P'J' P ′ J ′ at stage k k k . Keep n k = m k − m k − 1 n_{k}=m_{k}-m_{k-1} n k = m k − m k − 1 .
Case 1: P ≠ P ′ P\ne P' P = P ′ . The recursive definition gives an integer z z z with
t ( λ k ( P J ) − λ k ( P ′ J ′ ) ) = t ( λ k − 1 ( P ) − λ k − 1 ( P ′ ) ) + q m k − 1 z . t\bigl(\lambda_{k}(PJ)-\lambda_{k}(P'J')\bigr)=t\bigl(\lambda_{k-1}(P)-\lambda_{k-1}(P')\bigr)+q^{m_{k-1}}z. t ( λ k ( P J ) − λ k ( P ′ J ′ ) ) = t ( λ k − 1 ( P ) − λ k − 1 ( P ′ ) ) + q m k − 1 z . By q m k − 1 q^{m_{k-1}} q m k − 1 -periodicity and the induction hypothesis,
μ ^ m k − 1 ( t ( λ k ( P J ) − λ k ( P ′ J ′ ) ) ) = μ ^ m k − 1 ( t ( λ k − 1 ( P ) − λ k − 1 ( P ′ ) ) ) = 0. \widehat{\mu}_{m_{k-1}}\bigl(t(\lambda_{k}(PJ)-\lambda_{k}(P'J'))\bigr)=\widehat{\mu}_{m_{k-1}}\bigl(t(\lambda_{k-1}(P)-\lambda_{k-1}(P'))\bigr)=0. μ m k − 1 ( t ( λ k ( P J ) − λ k ( P ′ J ′ )) ) = μ m k − 1 ( t ( λ k − 1 ( P ) − λ k − 1 ( P ′ )) ) = 0. This is a factor of the Fourier transform for μ m k \mu_{m_{k}} μ m k , so the latter vanishes as well.
Case 2: P = P ′ P=P' P = P ′ and J ≠ J ′ J\ne J' J = J ′ . Let z = c k ( P , J ) − c k ( P , J ′ ) z=c_{k}(P,J)-c_{k}(P,J') z = c k ( P , J ) − c k ( P , J ′ ) . Now
t ( λ k ( P J ) − λ k ( P J ′ ) ) = t q m k − 1 ( T n k ( J ) − T n k ( J ′ ) + q n k z ) . t\bigl(\lambda_{k}(PJ)-\lambda_{k}(PJ')\bigr)=tq^{m_{k-1}}\bigl(T_{n_{k}}(J)-T_{n_{k}}(J')+q^{n_{k}}z\bigr). t ( λ k ( P J ) − λ k ( P J ′ ) ) = t q m k − 1 ( T n k ( J ) − T n k ( J ′ ) + q n k z ) . The product of Fourier factors from levels m k − 1 + 1 m_{k-1}+1 m k − 1 + 1 through m k − 1 + n k m_{k-1}+n_{k} m k − 1 + n k is
∏ r = 1 n k m D ( q − r t ( T n k ( J ) − T n k ( J ′ ) + q n k z ) ) = μ ^ n k ( t ( T n k ( J ) − T n k ( J ′ ) + q n k z ) ) \prod_{r=1}^{n_{k}}m_{D}\bigl(q^{-r}t(T_{n_{k}}(J)-T_{n_{k}}(J')+q^{n_{k}}z)\bigr)=\widehat{\mu}_{n_{k}}\bigl(t(T_{n_{k}}(J)-T_{n_{k}}(J')+q^{n_{k}}z)\bigr) r = 1 ∏ n k m D ( q − r t ( T n k ( J ) − T n k ( J ′ ) + q n k z ) ) = μ n k ( t ( T n k ( J ) − T n k ( J ′ ) + q n k z ) ) = μ ^ n k ( t ( T n k ( J ) − T n k ( J ′ ) ) ) = 0. =\widehat{\mu}_{n_{k}}\bigl(t(T_{n_{k}}(J)-T_{n_{k}}(J'))\bigr)=0. = μ n k ( t ( T n k ( J ) − T n k ( J ′ )) ) = 0. Periodicity gives the penultimate equality and Lemma 2.1 gives the last. For k = 1 k=1 k = 1 there is only the empty parent, so Case 2 starts the induction. In either case the exponential inner product is zero. Different nodes cannot give the same frequency, since that would give inner product one. There are exactly m m k m^{m_{k}} m m k nodes and, by Lemma 2.1, exactly m m k m^{m_{k}} m m k atoms of μ m k \mu_{m_{k}} μ m k . The orthonormal system therefore has the full dimension and is a basis. □
For the next construction, I = i 1 i 2 ⋯ ∈ { 1 , 2 } N I=i_{1}i_{2}\cdots\in\{1,2\}^{\mathbb{N}} I = i 1 i 2 ⋯ ∈ { 1 , 2 } N is a binary encoding of the choices of frequencies; the two symbols 1 and 2 encode the two alternatives at each stage. Write I ∣ k = i 1 ⋯ i k I|_{k}=i_{1}\cdots i_{k} I ∣ k = i 1 ⋯ i k . More generally, for a finite binary encoding ι = i 1 ⋯ i k ∈ { 1 , 2 } k \iota=i_{1}\cdots i_{k}\in\{1,2\}^{k} ι = i 1 ⋯ i k ∈ { 1 , 2 } k , write ι − = i 1 ⋯ i k − 1 \iota^{-}=i_{1}\cdots i_{k-1} ι − = i 1 ⋯ i k − 1 . At level zero both the digit word and the binary encoding are the empty word ∅ \varnothing ∅ . We call a word J ∈ J n J\in\mathcal{J}_{n} J ∈ J n a core word . At stage k k k , the values λ k ( J , ι ) \lambda_{k}(J,\iota) λ k ( J , ι ) with J ∈ J m k J\in\mathcal{J}_{m_{k}} J ∈ J m k are called core frequency values .
Theorem 3.4 (Recursive spectral extension). Let ( a k ) (a_k) ( a k ) be a recurrent enumeration of T + T_+ T + and let ( J n ) (\mathcal{J}_n) ( J n ) be a prefix-closed increasing word sequence. Put m 0 = 0 m_0=0 m 0 = 0 and λ 0 ( ∅ , ∅ ) = 0 \lambda_0(\varnothing,\varnothing)=0 λ 0 ( ∅ , ∅ ) = 0 . Construct the levels recursively, with
m k − 1 < γ k < m k < γ k + 1 , n k = m k − m k − 1 . m_{k-1}<\gamma_k<m_k<\gamma_{k+1}, \qquad n_k=m_k-m_{k-1}. m k − 1 < γ k < m k < γ k + 1 , n k = m k − m k − 1 . At the start of stage k k k , the finite sets Λ k − 1 ( ι ) \Lambda_{k-1}(\iota) Λ k − 1 ( ι ) from stage k − 1 k-1 k − 1 have been defined, where
Λ k − 1 ( ι ) = { λ k − 1 ( P , ι ) : P ∈ Ω m k − 1 } , ι ∈ { 1 , 2 } k − 1 . \Lambda_{k-1}(\iota)=\{\lambda_{k-1}(P,\iota):P\in\Omega^{m_{k-1}}\}, \qquad\iota\in\{1,2\}^{k-1}. Λ k − 1 ( ι ) = { λ k − 1 ( P , ι ) : P ∈ Ω m k − 1 } , ι ∈ { 1 , 2 } k − 1 . Define the maximum over all these sets by
M k − 1 : = max ι ∈ { 1 , 2 } k − 1 ⋃ Λ k − 1 ( ι ) = max ι ∈ { 1 , 2 } k − 1 P ∈ Ω m k − 1 λ k − 1 ( P , ι ) , M 0 = 0. M_{k-1}:=\max_{\iota\in\{1,2\}^{k-1}}\bigcup\Lambda_{k-1}(\iota)=\max_{\substack{\iota\in\{1,2\}^{k-1}\\P\in\Omega^{m_{k-1}}}}\lambda_{k-1}(P,\iota), \qquad M_0=0. M k − 1 := ι ∈ { 1 , 2 } k − 1 max ⋃ Λ k − 1 ( ι ) = ι ∈ { 1 , 2 } k − 1 P ∈ Ω m k − 1 max λ k − 1 ( P , ι ) , M 0 = 0. Suppose the choices at every stage satisfy
J m k = J γ k 0 m k − γ k , a k q γ k − m k < δ a k / 4 , a k q − m k M k − 1 < δ a k / 4. \mathcal{J}_{m_k}=\mathcal{J}_{\gamma_k}0^{m_k-\gamma_k}, \qquad a_kq^{\gamma_k-m_k}<\delta_{a_k}/4, \qquad a_kq^{-m_k}M_{k-1}<\delta_{a_k}/4. J m k = J γ k 0 m k − γ k , a k q γ k − m k < δ a k /4 , a k q − m k M k − 1 < δ a k /4. Then the integer corrections can be chosen so that, for each binary encoding I ∈ { 1 , 2 } N I\in\{1,2\}^{\mathbb{N}} I ∈ { 1 , 2 } N , the set
Λ ( I ) = ⋃ k ≥ 1 Λ k ( I ∣ k ) \Lambda(I)=\bigcup_{k\geq1}\Lambda_k(I|_k) Λ ( I ) = k ≥ 1 ⋃ Λ k ( I ∣ k ) has the following properties:
(i) t Λ ( I ) t\Lambda(I) t Λ ( I ) is a spectrum of μ \mu μ for every t ∈ T t\in\mathcal{T} t ∈ T ;
(ii) Λ ( I ) = Λ ( J ) ∪ E ( I ) \Lambda(I)=\Lambda(\mathcal{J})\cup E(I) Λ ( I ) = Λ ( J ) ∪ E ( I ) , where explicitly
E ( I ) = ⋃ k ≥ 1 { λ k ( P J , I ∣ k ) : P ∈ Ω m k − 1 , J ∈ Ω n k ∖ { 0 n k } , P J ∉ J m k } . E(I)=\bigcup_{k\geq1}\left\{\lambda_k(PJ,I|_k):P\in\Omega^{m_{k-1}},\ J\in\Omega^{n_k}\setminus\{0^{n_k}\},\ PJ\notin\mathcal{J}_{m_k}\right\}. E ( I ) = k ≥ 1 ⋃ { λ k ( P J , I ∣ k ) : P ∈ Ω m k − 1 , J ∈ Ω n k ∖ { 0 n k } , P J ∈ / J m k } . The set E ( I ) E(I) E ( I ) is disjoint from Λ ( J ) \Lambda(\mathcal{J}) Λ ( J ) , consists of positive integers, and is q q q -lacunary;
(iii) Λ ( I ) ≠ Λ ( I ′ ) \Lambda(I)\ne\Lambda(I') Λ ( I ) = Λ ( I ′ ) whenever I ≠ I ′ I\ne I' I = I ′ . Here a positive set is q q q -lacunary if its increasing enumeration ( e j ) (e_j) ( e j ) satisfies e j + 1 ≥ q e j e_{j+1}\geq qe_j e j + 1 ≥ q e j .
Proof. We first define the corrections and the finite sets, including all binary encodings, and then verify the three conclusions.
Step 1: the nodes that create new additional frequencies. Suppose stage k − 1 k-1 k − 1 is complete for all ι − ∈ { 1 , 2 } k − 1 \iota^-\in\{1,2\}^{k-1} ι − ∈ { 1 , 2 } k − 1 . The sets in the definition of M k − 1 M_{k-1} M k − 1 are nonempty and finite: there are 2 k − 1 2^{k-1} 2 k − 1 binary encodings and m m k − 1 m^{m_{k-1}} m m k − 1 digit words for each. Thus M k − 1 M_{k-1} M k − 1 is known and finite before γ k , m k \gamma_k,m_k γ k , m k are chosen. Nestedness, verified below, will also ensure that it bounds every frequency from earlier stages.
For ι ∈ { 1 , 2 } k \iota\in\{1,2\}^k ι ∈ { 1 , 2 } k and a digit word P J ∈ Ω m k PJ\in\Omega^{m_k} P J ∈ Ω m k , where P ∈ Ω m k − 1 P\in\Omega^{m_{k-1}} P ∈ Ω m k − 1 and J ∈ Ω n k J\in\Omega^{n_k} J ∈ Ω n k , define
λ k ( P J , ι ) = λ k − 1 ( P , ι − ) + q m k − 1 T n k ( J ) + q m k c k ( P , J , ι ) . \lambda_k(PJ,\iota)=\lambda_{k-1}(P,\iota^-)+q^{m_{k-1}}T_{n_k}(J)+q^{m_k}c_k(P,J,\iota). λ k ( P J , ι ) = λ k − 1 ( P , ι − ) + q m k − 1 T n k ( J ) + q m k c k ( P , J , ι ) .
For every parent P P P and encoding ι \iota ι , set c k ( P , 0 n k , ι ) = 0 c_k(P,0^{n_k},\iota)=0 c k ( P , 0 n k , ι ) = 0 . Thus a zero suffix will retain its parent’s value. Also set c k ( P , J , ι ) = 0 c_k(P,J,\iota)=0 c k ( P , J , ι ) = 0 whenever P J ∈ J m k PJ\in\mathcal{J}_{m_k} P J ∈ J m k . The only remaining triples are
N k = { ( ℓ , P , J ) : ℓ ∈ { 1 , 2 } k , P ∈ Ω m k − 1 , J ∈ Ω n k ∖ { 0 n k } , P J ∉ J m k } . \mathcal{N}_k=\left\{(\ell,P,J):\ \ell\in\{1,2\}^k,\quad P\in\Omega^{m_{k-1}},\quad J\in\Omega^{n_k}\setminus\{0^{n_k}\},\quad PJ\notin\mathcal{J}_{m_k}\right\}. N k = { ( ℓ , P , J ) : ℓ ∈ { 1 , 2 } k , P ∈ Ω m k − 1 , J ∈ Ω n k ∖ { 0 n k } , P J ∈ / J m k } .
These, and only these, receive a large positive correction at stage k k k . For a triple in N k \mathcal{N}_k N k , put
x k ( J ) = a k q − n k T n k ( J ) , y k ( P , ℓ − ) = a k q − m k λ k − 1 ( P , ℓ − ) . x_k(J)=a_kq^{-n_k}T_{n_k}(J),\qquad y_k(P,\ell^-)=a_kq^{-m_k}\lambda_{k-1}(P,\ell^-). x k ( J ) = a k q − n k T n k ( J ) , y k ( P , ℓ − ) = a k q − m k λ k − 1 ( P , ℓ − ) .
Because 0 ≤ T n k ( J ) < q n k 0\leq T_{n_k}(J)<q^{n_k} 0 ≤ T n k ( J ) < q n k and by the assumed bound on M k − 1 M_{k-1} M k − 1 , these satisfy
0 ≤ x k ( J ) < a k , 0 ≤ y k ( P , ℓ − ) < δ a k / 4. 0\leq x_k(J)<a_k,\qquad0\leq y_k(P,\ell^-)<\delta a_k/4. 0 ≤ x k ( J ) < a k , 0 ≤ y k ( P , ℓ − ) < δ a k /4.
For each such triple, choose an integer r k ( ℓ , P , J ) ≥ 1 r_k(\ell,P,J)\geq1 r k ( ℓ , P , J ) ≥ 1 and put
φ k ( P , J , ℓ ) = q r k ( ℓ , P , J ) b a k , b a k = η a k ( 0 , 2 ) > 0. \varphi_k(P,J,\ell)=q^{r_k(\ell,P,J)}b_{a_k},\qquad b_{a_k}=\eta_{a_k}(0,2)>0. φ k ( P , J , ℓ ) = q r k ( ℓ , P , J ) b a k , b a k = η a k ( 0 , 2 ) > 0.
For ℓ = ℓ 1 ⋯ ℓ k \ell=\ell_1\cdots\ell_k ℓ = ℓ 1 ⋯ ℓ k and ( ℓ , P , J ) ∈ N k (\ell,P,J)\in\mathcal{N}_k ( ℓ , P , J ) ∈ N k , define
σ k ( P , J , ℓ ) = { ℓ k , P = 0 m k − 1 , J = 0 n k − 1 1 , 1 , otherwise . \sigma_k(P,J,\ell)=
\begin{cases}
\ell_k,&P=0^{m_{k-1}},\ J=0^{n_k-1}1,\\
1,&\text{otherwise}.
\end{cases} σ k ( P , J , ℓ ) = { ℓ k , 1 , P = 0 m k − 1 , J = 0 n k − 1 1 , otherwise .
The distinguished word 0 m k − 1 1 0^{m_k-1}1 0 m k − 1 1 lies outside J m k \mathcal{J}_{m_k} J m k , since m k > γ k m_k>\gamma_k m k > γ k and every core word has zero final digit. Its suffix is nonzero, so the corresponding triple belongs to N k \mathcal{N}_k N k for each encoding. Set
c k ( P , J , ℓ ) = η a k ( x k ( J ) , σ k ( P , J , ℓ ) ) + φ k ( P , J , ℓ ) , ( ℓ , P , J ) ∈ N k . (7) c_k(P,J,\ell)=\eta_{a_k}\left(x_k(J),\sigma_k(P,J,\ell)\right)+\varphi_k(P,J,\ell),\qquad(\ell,P,J)\in\mathcal{N}_k. \tag*{(7)} c k ( P , J , ℓ ) = η a k ( x k ( J ) , σ k ( P , J , ℓ ) ) + φ k ( P , J , ℓ ) , ( ℓ , P , J ) ∈ N k . (7)
Thus the distinguished node uses the choice indexed by ℓ k \ell_k ℓ k , while every other triple uses the first choice. Every newly assigned frequency value is a positive integer.
The exponent will depend on the whole triple, including the binary encoding and the parent, so all these values can be separated.
Step 2: an explicit successive choice of the exponents. Give the finite set N k \mathcal{N}_k N k any fixed order and let ρ k : N k → { 1 , … , # N k } \rho_k:\mathcal{N}_k\to\{1,\ldots,\#\mathcal{N}_k\} ρ k : N k → { 1 , … , # N k } be its rank map. Suppose the first j − 1 j-1 j − 1 triples at this stage have been assigned. Let B k , j − 1 B_{k,j-1} B k , j − 1 be the maximum of M k − 1 M_{k-1} M k − 1 and the frequency values already assigned to the first j − 1 j-1 j − 1 triples at stage k k k :
B k , j − 1 : = max ( { M k − 1 } ∪ { λ k ( P J , ℓ ) : ( ℓ , P , J ) ∈ N k , ρ k ( ℓ , P , J ) ≤ j − 1 } ) . B_{k,j-1}:=\max\left(\{M_{k-1}\}\cup\left\{\lambda_k(PJ,\ell):(\ell,P,J)\in\mathcal{N}_k,\ \rho_k(\ell,P,J)\leq j-1\right\}\right). B k , j − 1 := max ( { M k − 1 } ∪ { λ k ( P J , ℓ ) : ( ℓ , P , J ) ∈ N k , ρ k ( ℓ , P , J ) ≤ j − 1 } ) .
All these values have already been determined. For j = 1 j=1 j = 1 the second set is empty, so B k , 0 = M k − 1 B_{k,0}=M_{k-1} B k , 0 = M k − 1 ; in particular B 1 , 0 = 0 B_{1,0}=0 B 1 , 0 = 0 . By nestedness, M k − 1 M_{k-1} M k − 1 bounds all frequencies assigned at earlier stages, for all binary encodings.
For the triple ( ℓ , P , J ) (\ell,P,J) ( ℓ , P , J ) of rank j j j , write
z k ( P , J , ℓ ) = x k ( J ) + y k ( P , ℓ − ) + a k η a k ( x k ( J ) , σ k ( P , J , ℓ ) ) . z_k(P,J,\ell)=x_k(J)+y_k(P,\ell^-)+a_k\eta_{a_k}\left(x_k(J),\sigma_k(P,J,\ell)\right). z k ( P , J , ℓ ) = x k ( J ) + y k ( P , ℓ − ) + a k η a k ( x k ( J ) , σ k ( P , J , ℓ ) ) .
Choose r k ( ℓ , P , J ) r_k(\ell,P,J) r k ( ℓ , P , J ) to be the least positive integer r r r satisfying both
q − r ∣ z k ( P , J , ℓ ) ∣ < δ a k q^{-r}\left|z_k(P,J,\ell)\right|<\delta a_k q − r ∣ z k ( P , J , ℓ ) ∣ < δ a k
and
λ k − 1 ( P , ℓ − ) + q m k − 1 T n k ( J ) + q m k ( η a k ( x k ( J ) , σ k ( P , J , ℓ ) ) + q r b a k ) > max { q B k , j − 1 , M k − 1 , 1 } . \lambda_{k-1}(P,\ell^-)+q^{m_{k-1}}T_{n_k}(J)
+q^{m_k}\left(\eta_{a_k}\left(x_k(J),\sigma_k(P,J,\ell)\right)+q^r b_{a_k}\right)
>\max\{qB_{k,j-1},M_{k-1},1\}. λ k − 1 ( P , ℓ − ) + q m k − 1 T n k ( J ) + q m k ( η a k ( x k ( J ) , σ k ( P , J , ℓ ) ) + q r b a k ) > max { q B k , j − 1 , M k − 1 , 1 } .
Such an integer exists: with all preceding data fixed, the first left-hand side tends to zero and the frequency in the second inequality tends to infinity as r → ∞ r \to\infty r → ∞ . This rule specifies the exponents once the selection maps and the finite ordering have been fixed. It also guarantees that every new additional value is greater than all earlier-stage frequencies and greater than q q q times every previously assigned additional value. In particular, for any two triples in N k \mathcal{N}_k N k ,
ρ k ( ι , P , J ) > ρ k ( ι ′ , P ′ , J ′ ) ⟹ λ k ( P J , ι ) > q λ k ( P ′ J ′ , ι ′ ) . \rho_k(\iota,P,J)>\rho_k(\iota',P',J')\quad\Longrightarrow\quad\lambda_k(PJ,\iota)>q\lambda_k(P'J',\iota'). ρ k ( ι , P , J ) > ρ k ( ι ′ , P ′ , J ′ ) ⟹ λ k ( P J , ι ) > q λ k ( P ′ J ′ , ι ′ ) .
This holds for all encodings, including distinguished nodes, because the size condition uses the actual choice σ k ( P , J , ι ) \sigma_k(P,J,\iota) σ k ( P , J , ι ) . Together with the bound on earlier stages, it gives global strict separation of all newly assigned additional values.
Once all triples in N k \mathcal{N}_k N k have been assigned, set
Λ k ( ι ) = { λ k ( W , ι ) : W ∈ Ω m k } , M k = max ι ∈ { 1 , 2 } k , W ∈ Ω m k λ k ( W , ι ) . \Lambda_k(\iota)=\{\lambda_k(W,\iota):W\in\Omega^{m_k}\},\qquad
M_k=\max_{\iota\in\{1,2\}^k,\ W\in\Omega^{m_k}}\lambda_k(W,\iota). Λ k ( ι ) = { λ k ( W , ι ) : W ∈ Ω m k } , M k = ι ∈ { 1 , 2 } k , W ∈ Ω m k max λ k ( W , ι ) .
These are finite data for the next stage. In particular the construction of stage k k k does not use the unknown future value M k M_k M k in choosing its own word length.
Step 3: finite spectrality and the precise description of E ( I ) E(I) E ( I ) . All corrections are integers and zero suffixes have correction zero. Lemma 3.3 therefore gives, for each binary encoding ι \iota ι ,
0 ∈ Λ k − 1 ( ι − ) ⊂ Λ k ( ι ) , t Λ k ( ι ) is a spectrum of μ m k ( t ∈ T ) . 0\in\Lambda_{k-1}(\iota^-)\subset\Lambda_k(\iota),\qquad t\Lambda_k(\iota)\text{ is a spectrum of }\mu_{m_k}\quad(t\in\mathcal{T}). 0 ∈ Λ k − 1 ( ι − ) ⊂ Λ k ( ι ) , t Λ k ( ι ) is a spectrum of μ m k ( t ∈ T ) .
For a core node, prefix closure implies that its parent is a core node. Induction, starting from zero, gives
λ k ( W , ι ) = T m k ( W ) ( W ∈ J m k , ι ∈ { 1 , 2 } k ) . \lambda_k(W,\iota)=T_{m_k}(W)\qquad(W\in\mathcal{J}_{m_k},\ \iota\in\{1,2\}^k). λ k ( W , ι ) = T m k ( W ) ( W ∈ J m k , ι ∈ { 1 , 2 } k ) .
These values do not depend on the binary encoding.
Define the new additional set at stage k k k for a fixed finite encoding by
E k ( ι ) = { λ k ( P J , ι ) : ( ι , P , J ) ∈ N k } . E_k(\iota)=\{\lambda_k(PJ,\iota):(\iota,P,J)\in\mathcal{N}_k\}. E k ( ι ) = { λ k ( P J , ι ) : ( ι , P , J ) ∈ N k } .
By the three cases in Step 1, the finite sets satisfy the exact recursive identity
Λ k ( ι ) = Λ k − 1 ( ι − ) ∪ T m k ( J m k ) ∪ E k ( ι ) . \Lambda_k(\iota)=\Lambda_{k-1}(\iota^-)\cup T_{m_k}(\mathcal{J}_{m_k})\cup E_k(\iota). Λ k ( ι ) = Λ k − 1 ( ι − ) ∪ T m k ( J m k ) ∪ E k ( ι ) .
Indeed a zero suffix produces the first set, a core node produces the second, and a nonzero suffix outside the core produces the third. Taking unions along an infinite binary encoding gives
Λ ( I ) = Λ ( J ) ∪ E ( I ) , E ( I ) : = ⋃ k ≥ 1 E k ( I ∣ k ) . \Lambda(I)=\Lambda(\mathcal{J})\cup E(I),\qquad E(I):=\bigcup_{k\ge1}E_k(I|_k). Λ ( I ) = Λ ( J ) ∪ E ( I ) , E ( I ) := k ≥ 1 ⋃ E k ( I ∣ k ) .
Here every core word at an intermediate length extends by zeros to some m k m_k m k , which justifies the union of the core terms.
We also verify that no additional value can become a core value later. If λ ∈ E k ( ι ) \lambda\in E_k(\iota) λ ∈ E k ( ι ) , its node is distinct from every core node at stage k k k . They belong to the same finite orthogonal set for the multiplier 1 ∈ T 1\in\mathcal{T} 1 ∈ T . Their frequencies cannot be congruent modulo q m k q^{m_k} q m k , since
μ ^ m k ( q m k z ) = 1 ( z ∈ Z ) . \widehat{\mu}_{m_k}(q^{m_k}z)=1\qquad(z\in\mathbb{Z}). μ m k ( q m k z ) = 1 ( z ∈ Z ) .
Every limiting core value reduces modulo q m k q^{m_k} q m k to an element of T m k ( J m k ) T_{m_k}(\mathcal{J}_{m_k}) T m k ( J m k ) , by prefix closure and zero extension. Thus λ ∉ Λ ( J ) \lambda\notin\Lambda(\mathcal{J}) λ ∈ / Λ ( J ) , and E ( I ) ∩ Λ ( J ) = ∅ E(I) \cap\Lambda(\mathcal{J}) = \varnothing E ( I ) ∩ Λ ( J ) = ∅ .
Step 4: the tail Fourier estimates and completeness. At stage k k k , we estimate the tail Fourier transform for the multiplier a k a_k a k . For a retained frequency, the hypotheses give
0 ≤ a k q − m k λ k − 1 ( P , ι − ) ≤ a k q − m k M k − 1 < δ a k / 4. 0 \le a_kq^{-m_k}\lambda_{k-1}(P,\iota^-) \le a_kq^{-m_k}M_{k-1} < \delta_{a_k}/4. 0 ≤ a k q − m k λ k − 1 ( P , ι − ) ≤ a k q − m k M k − 1 < δ a k /4.
For a core frequency, the zero digits after γ k \gamma_k γ k give
0 ≤ a k q − m k T m k ( W ) < a k q γ k − m k < δ a k / 4. 0 \le a_kq^{-m_k}T_{m_k}(W) < a_kq^{\gamma_k-m_k} < \delta_{a_k}/4. 0 ≤ a k q − m k T m k ( W ) < a k q γ k − m k < δ a k /4.
Thus, for every λ ∈ Λ k − 1 ( ι − ) ∪ T m k ( J m k ) \lambda\in\Lambda_{k-1}(\iota^-) \cup T_{m_k}(\mathcal{J}_{m_k}) λ ∈ Λ k − 1 ( ι − ) ∪ T m k ( J m k ) , we have ∣ μ ^ ( a k q − m k λ ) ∣ ≥ 1 / 2 |\widehat{\mu}(a_kq^{-m_k}\lambda)| \ge1/2 ∣ μ ( a k q − m k λ ) ∣ ≥ 1/2 by the choice of δ a k \delta_{a_k} δ a k in Lemma 2.5. For a new additional frequency in E k ( ι ) E_k(\iota) E k ( ι ) , (7) gives
a k q − m k λ k ( P , J , ι ) = z k ( P , J , ι ) + a k q r k ( ι , P , J ) b a k . a_kq^{-m_k}\lambda_k(P,J,\iota) = z_k(P,J,\iota) + a_kq^{r_k(\iota,P,J)}b_{a_k}. a k q − m k λ k ( P , J , ι ) = z k ( P , J , ι ) + a k q r k ( ι , P , J ) b a k .
Since σ k ( P , J , ι ) ∈ { 1 , 2 } \sigma_k(P,J,\iota) \in\{1,2\} σ k ( P , J , ι ) ∈ { 1 , 2 } and ∣ y k ( P , ι − ) ∣ < δ a k / 4 |y_k(P,\iota^-)| < \delta_{a_k}/4 ∣ y k ( P , ι − ) ∣ < δ a k /4 , Lemma 2.5 yields ∣ μ ^ ( z k ( P , J , ι ) ) ∣ ≥ ϵ a k |\widehat{\mu}(z_k(P,J,\iota))| \ge\epsilon_{a_k} ∣ μ ( z k ( P , J , ι )) ∣ ≥ ϵ a k . Step 2 ensures q − r k ∣ z k ∣ < δ a k q^{-r_k}|z_k| < \delta_{a_k} q − r k ∣ z k ∣ < δ a k for this same z k z_k z k . The factorization in the proof of Lemma 3.2 , valid for either selection value, therefore gives
∣ μ ^ ( a k q − m k λ k ( P , J , ι ) ) ∣ ≥ ∣ μ ^ ( z k ( P , J , ι ) ) ∣ ∣ μ ^ ( q − r k ( ι , P , J ) z k ( P , J , ι ) + a k b a k ) ∣ ≥ ϵ a k 2 . |\widehat{\mu}(a_kq^{-m_k}\lambda_k(P,J,\iota))| \ge|\widehat{\mu}(z_k(P,J,\iota))||\widehat{\mu}(q^{-r_k(\iota,P,J)}z_k(P,J,\iota)+a_kb_{a_k})| \ge\epsilon_{a_k}^2. ∣ μ ( a k q − m k λ k ( P , J , ι )) ∣ ≥ ∣ μ ( z k ( P , J , ι )) ∣∣ μ ( q − r k ( ι , P , J ) z k ( P , J , ι ) + a k b a k ) ∣ ≥ ϵ a k 2 .
Since 0 < ϵ a k < 1 / 2 0 < \epsilon_{a_k} < 1/2 0 < ϵ a k < 1/2 , the three cases imply
inf λ ∈ Λ k ( ι ) ∣ μ ^ ( a k q − m k λ ) ∣ 2 ≥ ϵ a k 4 ( ι ∈ { 1 , 2 } k ) . \inf_{\lambda\in\Lambda_k(\iota)}|\widehat{\mu}(a_kq^{-m_k}\lambda)|^2 \ge\epsilon_{a_k}^4 \qquad(\iota\in\{1,2\}^k). λ ∈ Λ k ( ι ) inf ∣ μ ( a k q − m k λ ) ∣ 2 ≥ ϵ a k 4 ( ι ∈ { 1 , 2 } k ) .
Fix t ∈ T + t \in\mathcal{T}_+ t ∈ T + . By recurrence, the indices satisfying a k = t a_k=t a k = t form an infinite sequence k 1 < k 2 < ⋯ k_1<k_2<\cdots k 1 < k 2 < ⋯ . Apply Theorem 2.3 with b j = m k j b_j=m_{k_j} b j = m k j , finite spectra t Λ k j ( I ∣ k j ) t\Lambda_{k_j}(I|_{k_j}) t Λ k j ( I ∣ k j ) , and the uniform constant ϵ t 4 > 0 \epsilon_t^4>0 ϵ t 4 > 0 . Nestedness and k j → ∞ k_j\to\infty k j → ∞ imply that their union is t Λ ( I ) t\Lambda(I) t Λ ( I ) , so it is a spectrum of μ \mu μ . Complex conjugation takes its exponential basis to the basis with frequencies − t Λ ( I ) -t\Lambda(I) − t Λ ( I ) . Since T = − T \mathcal{T}=-\mathcal{T} T = − T , this proves (i) for every multiplier.
Step 5: sparsity and distinctness. List all additional values in their order of assignment, first by stage and then by the rank ρ k \rho_k ρ k . Step 2 makes each one strictly greater than q q q times the preceding one. Thus the global set of additional values is q q q -lacunary, and so is its subset E ( I ) E(I) E ( I ) . Together with Step 3 this proves (ii).
For (iii), suppose I ≠ I ′ I\ne I' I = I ′ and let k k k be their first differing position. At the node 0 m k − 1 1 0^{m_k-1}1 0 m k − 1 1 , choose the additional frequency λ ∗ \lambda_* λ ∗ belonging to I ∣ k I|_k I ∣ k . It is retained in Λ ( I ) \Lambda(I) Λ ( I ) at every later stage. The value λ ∗ \lambda_* λ ∗ is assigned to a triple whose binary encoding is I ∣ k I|_k I ∣ k . Since I ∣ k ≠ I ′ ∣ k I|_k\ne I'|_k I ∣ k = I ′ ∣ k , this triple does not occur along I ′ I' I ′ . Step 2 assigns distinct frequency values to all triples, both within each stage and across different stages. Hence λ ∗ ∉ E ( I ′ ) \lambda_* \notin E(I') λ ∗ ∈ / E ( I ′ ) . Step 3 gives λ ∗ ∉ Λ ( J ) \lambda_* \notin\Lambda(\mathcal{J}) λ ∗ ∈ / Λ ( J ) as well. Consequently λ ∗ ∉ Λ ( I ′ ) \lambda_* \notin\Lambda(I') λ ∗ ∈ / Λ ( I ′ ) and the two spectra differ. □
The size conditions in Theorem 3.4 are imposed recursively. In § 4 , we choose each new core stage after M k − 1 M_{k-1} M k − 1 is known and verify these conditions while preserving the prescribed density.
Lemma 3.5. If E ⊂ ( 0 , ∞ ) E\subset(0,\infty) E ⊂ ( 0 , ∞ ) is b b b -lacunary for some b > 1 b>1 b > 1 , then M E ( h ) = O ( 1 + log h ) M_E(h)=O(1+\log h) M E ( h ) = O ( 1 + log h ) as h → ∞ h\to\infty h → ∞ . Consequently D + u ( E ) = 0 D_+^u(E)=0 D + u ( E ) = 0 for every u > 0 u>0 u > 0 .
Proof. If E E E is empty or finite the lemma is immediate. Otherwise write its increasing enumeration as 0 < e 1 < e 2 < ⋯ 0<e_1<e_2<\cdots 0 < e 1 < e 2 < ⋯ . Suppose B ( x , h ) B(x,h) B ( x , h ) contains r ≥ 1 r\ge1 r ≥ 1 elements and let e j e_j e j be the first of them. Their last element is at least b r − 1 e j b^{r-1}e_j b r − 1 e j . Both endpoints lie in an interval of length 2 h 2h 2 h , so
( b r − 1 − 1 ) e 1 ≤ ( b r − 1 − 1 ) e j < 2 h . (b^{r-1}-1)e_1\le(b^{r-1}-1)e_j<2h. ( b r − 1 − 1 ) e 1 ≤ ( b r − 1 − 1 ) e j < 2 h .
Solving this inequality for r r r gives
# ( E ∩ B ( x , h ) ) ≤ 1 + log b ( 1 + 2 h / e 1 ) . \#(E\cap B(x,h))\le1+\log_b(1+2h/e_1). # ( E ∩ B ( x , h )) ≤ 1 + log b ( 1 + 2 h / e 1 ) .
The right-hand side is independent of x x x , hence also bounds M E ( h ) M_E(h) M E ( h ) . For h ≥ 1 h\ge1 h ≥ 1 it is at most C ( 1 + log h ) C(1+\log h) C ( 1 + log h ) for a constant depending on b , e 1 b,e_1 b , e 1 . For each u > 0 u>0 u > 0 ,
0 ≤ M E ( h ) h u ≤ C 1 + log h h u ⟶ 0. 0\le\frac{M_E(h)}{h^u}\le C\frac{1+\log h}{h^u}\longrightarrow0. 0 ≤ h u M E ( h ) ≤ C h u 1 + log h ⟶ 0.
Taking limsup proves D u + ( E ) = 0 D_u^+(E)=0 D u + ( E ) = 0 . Adjoining zero changes each interval count by at most one. □ \square □
Prescribed Beurling density
In this section, only the digit assumptions 0 ∈ L ⊂ { 0 , … , q − 2 } 0\in L\subset\{0,\ldots,q-2\} 0 ∈ L ⊂ { 0 , … , q − 2 } and m = # L ≥ 2 m=\#L\ge2 m = # L ≥ 2 are needed. Write
Γ 0 = { 0 } , Γ n = T n ( Ω n ) , Γ = ⋃ n ≥ 1 Γ n , r 0 = log m / log q . \Gamma_0=\{0\},\qquad\Gamma_n=T_n(\Omega^n),\qquad\Gamma=\bigcup_{n\ge1}\Gamma_n,\qquad r_0=\log m/\log q. Γ 0 = { 0 } , Γ n = T n ( Ω n ) , Γ = n ≥ 1 ⋃ Γ n , r 0 = log m / log q .
Lemma 4.1. Let ( J n ) (\mathcal J_n) ( J n ) be prefix closed and Λ ( J ) \Lambda(\mathcal J) Λ ( J ) be given in Definition 3.1. If h ≥ 1 h\ge1 h ≥ 1 and q n h ≤ 2 h < q n h + 1 q^{n_h}\le2h<q^{n_h+1} q n h ≤ 2 h < q n h + 1 for some integer n h ∈ N n_h\in\mathbb N n h ∈ N , then
M Λ ( J ) ( h ) = M T n h + 2 ( J n h + 2 ) ( h ) . (8) M_{\Lambda(\mathcal J)}(h)=M_{T_{n_h+2}(\mathcal J_{n_h+2})}(h). \tag*{(8)} M Λ ( J ) ( h ) = M T n h + 2 ( J n h + 2 ) ( h ) . (8)
The supremum on the right is attained by a center in [ 0 , q n h + 2 ) [0,q^{n_h+2}) [ 0 , q n h + 2 ) .
Proof. Let λ , λ ′ ∈ Λ ( J ) \lambda,\lambda'\in\Lambda(\mathcal J) λ , λ ′ ∈ Λ ( J ) be distinct. Write λ = ∑ r ≥ 1 w r q r − 1 \lambda=\sum_{r\ge1}w_rq^{r-1} λ = ∑ r ≥ 1 w r q r − 1 and λ ′ = ∑ r ≥ 1 w r ′ q r − 1 \lambda'=\sum_{r\ge1}w'_rq^{r-1} λ ′ = ∑ r ≥ 1 w r ′ q r − 1 , where w r , w r ′ ∈ L w_r,w'_r\in L w r , w r ′ ∈ L and both digit sequences are eventually zero. Pad with zeros so they have equal length, and let j j j be their highest differing position. The coefficient at that position has absolute value at least one, while all lower differences have absolute value at most q − 2 q-2 q − 2 . Consequently
∣ λ − λ ′ ∣ ≥ q j − 1 − ( q − 2 ) ∑ r = 0 j − 2 q r = q j − 1 − ( q − 2 ) q j − 1 − 1 q − 1 = q j − 1 + q − 2 q − 1 . |\lambda-\lambda'|\ge q^{j-1}-(q-2)\sum_{r=0}^{j-2}q^r=q^{j-1}-(q-2)\frac{q^{j-1}-1}{q-1}=\frac{q^{j-1}+q-2}{q-1}. ∣ λ − λ ′ ∣ ≥ q j − 1 − ( q − 2 ) r = 0 ∑ j − 2 q r = q j − 1 − ( q − 2 ) q − 1 q j − 1 − 1 = q − 1 q j − 1 + q − 2 .
If j ≥ n h + 3 j\ge n_h+3 j ≥ n h + 3 , this is at least q n h + 2 / ( q − 1 ) > q n h + 1 > 2 h q^{n_h+2}/(q-1)>q^{n_h+1}>2h q n h + 2 / ( q − 1 ) > q n h + 1 > 2 h . Two points of B ( x , h ) B(x,h) B ( x , h ) cannot have such a distance. Therefore all points in B ( x , h ) ∩ Λ ( J ) B(x,h)\cap\Lambda(\mathcal J) B ( x , h ) ∩ Λ ( J ) have identical digits above position $n_h+2.
For a nonempty intersection write this common tail as v = ∑ r ≥ n h + 3 w r q r − 1 v=\sum_{r\ge n_h+3}w_rq^{r-1} v = ∑ r ≥ n h + 3 w r q r − 1 . For every selected λ ∈ Λ ( J ) ∩ B ( x , h ) \lambda\in\Lambda(\mathcal J)\cap B(x,h) λ ∈ Λ ( J ) ∩ B ( x , h ) ,
λ − v ∈ T n h + 2 ( J n h + 2 ) , λ − v ∈ B ( x − v , h ) . \lambda-v\in T_{n_h+2}(\mathcal J_{n_h+2}),\qquad\lambda-v\in B(x-v,h). λ − v ∈ T n h + 2 ( J n h + 2 ) , λ − v ∈ B ( x − v , h ) .
So
# ( B ( x , h ) ∩ Λ ( J ) ) ≤ # ( B ( x − v , h ) ∩ T n h + 2 ( J n h + 2 ) ) ≤ M T n h + 2 ( J n h + 2 ) ( h ) . \#(B(x,h)\cap\Lambda(\mathcal J))\le\#(B(x-v,h)\cap T_{n_h+2}(\mathcal J_{n_h+2}))\le M_{T_{n_h+2}(\mathcal J_{n_h+2})}(h). # ( B ( x , h ) ∩ Λ ( J )) ≤ # ( B ( x − v , h ) ∩ T n h + 2 ( J n h + 2 )) ≤ M T n h + 2 ( J n h + 2 ) ( h ) .
Taking the supremum over x x x gives one direction of (4.1). The other direction follows from T n h + 2 ( J n h + 2 ) ⊂ Λ ( J ) T_{n_h+2}(\mathcal{J}_{n_h+2}) \subset\Lambda(\mathcal{J}) T n h + 2 ( J n h + 2 ) ⊂ Λ ( J ) .
For a nonempty finite set the possible interval counts belong to a finite set of integers, so their supremum is attained. If a maximizing center lies outside the convex hull of the selected points, move it to the nearest endpoint of that hull. This only reduces its distance from the selected points. The new center remains maximizing and lies in [ 0 , q n h + 2 ) [0,q^{n_h+2}) [ 0 , q n h + 2 ) , since all the finite digit values lie in that interval. □
Remark 4.1. For J n = Ω n \mathcal{J}_n=\Omega^n J n = Ω n , Lemma 4.1 gives M Γ ( h ) = M Γ n h + 2 ( h ) M_{\Gamma}(h)=M_{\Gamma_{n_h+2}}(h) M Γ ( h ) = M Γ n h + 2 ( h ) .
Lemma 4.2. There is a constant C = C ( q , m ) C=C(q,m) C = C ( q , m ) such that M Γ ( h ) ≤ C h r 0 M_{\Gamma}(h)\le Ch^{r_0} M Γ ( h ) ≤ C h r 0 for all h ≥ 1 h\ge1 h ≥ 1 . In particular,
1 ≤ D r 0 + ( Γ ) < ∞ . 1\le D_{r_0}^{+}(\Gamma)<\infty. 1 ≤ D r 0 + ( Γ ) < ∞. For every ε > 0 \varepsilon>0 ε > 0 , there is H ε H_\varepsilon H ε such that M Γ ( h ) / h r 0 ≤ D r 0 + ( Γ ) + ε M_{\Gamma}(h)/h^{r_0}\le D_{r_0}^{+}(\Gamma)+\varepsilon M Γ ( h ) / h r 0 ≤ D r 0 + ( Γ ) + ε for h ≥ H ε h\ge H_\varepsilon h ≥ H ε .
Proof. For h ≥ 1 h\ge1 h ≥ 1 choose the integer n h ≥ 0 n_h\ge0 n h ≥ 0 with q n h ≤ 2 h < q n h + 1 q^{n_h}\le2h<q^{n_h+1} q n h ≤ 2 h < q n h + 1 . Lemma 4.1 for the full word set gives M Γ ( h ) = M Γ n h + 2 ( h ) ≤ m n h + 2 M_{\Gamma}(h)=M_{\Gamma_{n_h+2}}(h)\le m^{n_h+2} M Γ ( h ) = M Γ n h + 2 ( h ) ≤ m n h + 2 . Since q r 0 = m q^{r_0}=m q r 0 = m , we obtain the pointwise estimate
M Γ ( h ) ≤ m 2 q n h r 0 ≤ 2 r 0 m 2 h r 0 . M_{\Gamma}(h)\le m^2q^{n_hr_0}\le2^{r_0}m^2h^{r_0}. M Γ ( h ) ≤ m 2 q n h r 0 ≤ 2 r 0 m 2 h r 0 .
This proves the upper bound with C = 2 r 0 m 2 C=2^{r_0}m^2 C = 2 r 0 m 2 . For the lower bound note that all m n m^n m n elements of Γ n \Gamma_n Γ n belong to B ( 0 , q n ) B(0,q^n) B ( 0 , q n ) , because their values are nonnegative and strictly less than q n q^n q n . Hence
M Γ ( q n ) ( q n ) r 0 ≥ # Γ n q n r 0 = m n m n = 1. \frac{M_{\Gamma}(q^n)}{(q^n)^{r_0}}\ge\frac{\#\Gamma_n}{q^{nr_0}}=\frac{m^n}{m^n}=1. ( q n ) r 0 M Γ ( q n ) ≥ q n r 0 # Γ n = m n m n = 1.
Taking limsup along h = q n → ∞ h=q^n\to\infty h = q n → ∞ proves D r 0 + ( Γ ) ≥ 1 D_{r_0}^{+}(\Gamma)\ge1 D r 0 + ( Γ ) ≥ 1 .
Finally put D ∗ = D r 0 + ( Γ ) < ∞ D_* = D_{r_0}^{+}(\Gamma)<\infty D ∗ = D r 0 + ( Γ ) < ∞ . If the last assertion were false for some ε > 0 \varepsilon>0 ε > 0 , one could choose h j ≥ j h_j\ge j h j ≥ j with M Γ ( h j ) / h j r 0 > D ∗ + ε M_{\Gamma}(h_j)/h_j^{r_0}>D_*+\varepsilon M Γ ( h j ) / h j r 0 > D ∗ + ε . The limsup along this sequence, and hence the full limsup, would be at least D ∗ + ε D_*+\varepsilon D ∗ + ε , a contradiction. □
Lemma 4.3. Let β , γ \beta,\gamma β , γ be integers with 0 ≤ β < γ 0\le\beta<\gamma 0 ≤ β < γ , and let S ⊂ Γ β S\subset\Gamma_\beta S ⊂ Γ β with P = # S P=\#S P = # S . Put
F = S + q β Γ γ − β , θ = P / m β . F=S+q^\beta\Gamma_{\gamma-\beta},\qquad\theta=P/m^\beta. F = S + q β Γ γ − β , θ = P / m β . Then every interval U U U satisfies
∣ # ( F ∩ U ) − θ # ( Γ γ ∩ U ) ∣ ≤ 2 P . (9) \left|\#(F\cap U)-\theta\#(\Gamma_\gamma\cap U)\right|\le2P. \tag*{(9)} ∣ # ( F ∩ U ) − θ # ( Γ γ ∩ U ) ∣ ≤ 2 P . (9) Consequently ∣ M F ( h ) − θ M Γ γ ( h ) ∣ ≤ 2 P |M_F(h)-\theta M_{\Gamma_\gamma}(h)|\le2P ∣ M F ( h ) − θ M Γ γ ( h ) ∣ ≤ 2 P for every h > 0 h>0 h > 0 .
Proof. We group the digit words by their high-digit suffixes. Every word J J J of length γ \gamma γ has a unique decomposition J = I V J=IV J = I V , with I ∈ Ω β I\in\Omega^\beta I ∈ Ω β and V ∈ Ω γ − β V\in\Omega^{\gamma-\beta} V ∈ Ω γ − β , and
T γ ( I V ) = T β ( I ) + q β T γ − β ( V ) . T_\gamma(IV)=T_\beta(I)+q^\beta T_{\gamma-\beta}(V). T γ ( I V ) = T β ( I ) + q β T γ − β ( V ) .
Thus each fixed suffix corresponds to a block B v = Γ β + q β v B_v=\Gamma_\beta+q^\beta v B v = Γ β + q β v , v ∈ Γ γ − β v\in\Gamma_{\gamma-\beta} v ∈ Γ γ − β . The selected points of that block form S + q β v S+q^\beta v S + q β v and number P P P . Because Γ β ⊂ [ 0 , q β ) \Gamma_\beta\subset[0,q^\beta) Γ β ⊂ [ 0 , q β ) and distinct integer v v v differ by at least one, the convex hulls of these blocks are disjoint and ordered in the same way as v v v . Fix an interval U U U . Let V \mathcal{V} V be the set of suffix values whose full blocks meet U U U , and write K = # V K=\#\mathcal{V} K = # V . If K ≥ 2 K\ge2 K ≥ 2 , only the first and last of these blocks can be partly included: every point of any intermediate block lies between two points of U U U , and hence belongs to U U U . If K = 1 K=1 K = 1 , there is only one possibly partial block. For K = 0 K=0 K = 0 both counts are zero.
We now compare the two counts block by block. On a full block their difference is P − θ m β = 0 P-\theta m^{\beta}=0 P − θ m β = 0 . On a partial block, both
# ( ( S + q β v ) ∩ U ) and θ # ( B v ∩ U ) \#((S+q^{\beta}v)\cap U)\quad\text{and}\quad\theta\#(B_v\cap U) # (( S + q β v ) ∩ U ) and θ # ( B v ∩ U )
lie between zero and P P P , so the absolute value of their difference is at most P P P . Summing over at most two partial blocks proves (4.2). Equivalently,
θ # ( Γ γ ∩ U ) − 2 P ≤ # ( F ∩ U ) ≤ θ # ( Γ γ ∩ U ) + 2 P . \theta\#(\Gamma_{\gamma}\cap U)-2P\le\#(F\cap U)\le\theta\#(\Gamma_{\gamma}\cap U)+2P. θ # ( Γ γ ∩ U ) − 2 P ≤ # ( F ∩ U ) ≤ θ # ( Γ γ ∩ U ) + 2 P .
For every x ∈ R x\in\mathbb{R} x ∈ R , (4.2) gives
# ( F ∩ B ( x , h ) ) ≤ θ M Γ γ ( h ) + 2 P . \#(F\cap B(x,h))\le\theta M_{\Gamma_{\gamma}}(h)+2P. # ( F ∩ B ( x , h )) ≤ θ M Γ γ ( h ) + 2 P .
Taking the supremum over x x x , we obtain
M F ( h ) ≤ θ M Γ γ ( h ) + 2 P . M_F(h)\le\theta M_{\Gamma_{\gamma}}(h)+2P. M F ( h ) ≤ θ M Γ γ ( h ) + 2 P .
Conversely, (4.2) also gives, for every x ∈ R x\in\mathbb{R} x ∈ R ,
θ # ( Γ γ ∩ B ( x , h ) ) ≤ M F ( h ) + 2 P . \theta\#(\Gamma_{\gamma}\cap B(x,h))\le M_F(h)+2P. θ # ( Γ γ ∩ B ( x , h )) ≤ M F ( h ) + 2 P .
Since θ ≥ 0 \theta\ge0 θ ≥ 0 , taking the supremum over x x x yields
θ M Γ γ ( h ) ≤ M F ( h ) + 2 P . \theta M_{\Gamma_{\gamma}}(h)\le M_F(h)+2P. θ M Γ γ ( h ) ≤ M F ( h ) + 2 P .
Combining these inequalities proves the lemma. □
A finite subset of Γ g \Gamma_g Γ g is called closed under digit truncation if deleting any number of its most significant digits keeps the resulting number in the set.
Lemma 4.4 (Finite-stage calibration). Fix 0 < s < r 0 0<s<r_0 0 < s < r 0 , an integer g ≥ 2 g\ge2 g ≥ 2 , and a finite set F − ⊂ Γ g F_-\subset\Gamma_g F − ⊂ Γ g containing zero and closed under digit truncation. Put N = # F − N=\#F_- N = # F − and H = q g H=q^g H = q g . Given τ > 0 \tau>0 τ > 0 and e > 0 e>0 e > 0 , arbitrarily large integers γ \gamma γ admit integers α < β < γ \alpha<\beta<\gamma α < β < γ , an integer A ≥ 2 A\ge2 A ≥ 2 , and a set S ⊂ Γ β S\subset\Gamma_\beta S ⊂ Γ β with # S = P = A N \#S=P=AN # S = P = A N such that, for F = S + q β Γ γ − β F=S+q^{\beta}\Gamma_{\gamma-\beta} F = S + q β Γ γ − β ,
(i) α > g + 3 \alpha>g+3 α > g + 3 , F − ⊂ F F_-\subset F F − ⊂ F , F F F is closed under digit truncation, and retaining only the first α \alpha α digits of any element of F F F gives an element of F − F_- F − ;
(ii) with z = sup h ≥ H M Γ γ ( h ) / h s z=\sup_{h\ge H}M_{\Gamma_\gamma}(h)/h^s z = sup h ≥ H M Γ γ ( h ) / h s and θ = P / m β \theta=P/m^\beta θ = P / m β , τ − e / 2 ≤ θ z ≤ τ \tau-e/2\le\theta z\le\tau τ − e /2 ≤ θ z ≤ τ ;
(iii) # F / q s γ ≤ τ \#F/q^{s\gamma}\le\tau # F / q s γ ≤ τ , and
τ − e ≤ sup H ≤ h ≤ q γ M F ( h ) h s ≤ max { N / H s , τ } + e . \tau-e\le\sup_{H\le h\le q^\gamma}\frac{M_F(h)}{h^s}\le\max\{N/H^s,\tau\}+e. τ − e ≤ H ≤ h ≤ q γ sup h s M F ( h ) ≤ max { N / H s , τ } + e . Moreover, α \alpha α can exceed any prescribed finite bound, and β ≥ c γ \beta\ge c\gamma β ≥ c γ and β − α ≍ log β \beta-\alpha\asymp\log\beta β − α ≍ log β for sufficiently large γ \gamma γ , where one may take c = ( r 0 − s ) / ( 2 r 0 ) > 0 c=(r_0-s)/(2r_0)>0 c = ( r 0 − s ) / ( 2 r 0 ) > 0 .
Proof. We first choose a real parameter and then round it to obtain an integer number of selected prefixes. The parameter β \beta β determines the proportion of selected prefixes, and α \alpha α determines the length of the intervening zero block. The digits in positions β + 1 , … , γ \beta+1,\ldots,\gamma β + 1 , … , γ are unrestricted. Throughout this proof g g g , N N N , τ \tau τ , e e e are fixed and only γ \gamma γ is allowed to tend to infinity. Constants in this proof may depend on these fixed data.
Step 1: estimate the full-block counting quotient. Since Γ γ ⊂ B ( 0 , q γ ) \Gamma_\gamma\subset B(0,q^\gamma) Γ γ ⊂ B ( 0 , q γ ) and # Γ γ = m γ \#\Gamma_\gamma=m^\gamma # Γ γ = m γ , we have
z ≥ M Γ γ ( q γ ) q s γ = q ( r 0 − s ) γ . z \ge\frac{M_{\Gamma_\gamma}(q^\gamma)}{q^{s\gamma}}=q^{(r_0-s)\gamma}. z ≥ q s γ M Γ γ ( q γ ) = q ( r 0 − s ) γ .
For H ≤ h ≤ q γ H\le h\le q^\gamma H ≤ h ≤ q γ , Lemma 4.2 gives
M Γ γ ( h ) h s ≤ M Γ ( h ) h s ≤ C h r 0 − s ≤ C q ( r 0 − s ) γ . \frac{M_{\Gamma_\gamma}(h)}{h^s}\le\frac{M_\Gamma(h)}{h^s}\le Ch^{r_0-s}\le Cq^{(r_0-s)\gamma}. h s M Γ γ ( h ) ≤ h s M Γ ( h ) ≤ C h r 0 − s ≤ C q ( r 0 − s ) γ .
For h ≥ q γ h\ge q^\gamma h ≥ q γ all points fit inside an interval of radius h h h , so the quotient is m γ / h s m^\gamma/h^s m γ / h s and is decreasing. Therefore
q ( r 0 − s ) γ ≤ z ≤ C q ( r 0 − s ) γ , z = sup H ≤ h ≤ q γ M Γ γ ( h ) h s . q^{(r_0-s)\gamma}\le z\le Cq^{(r_0-s)\gamma},\qquad z=\sup_{H\le h\le q^\gamma}\frac{M_{\Gamma_\gamma}(h)}{h^s}. q ( r 0 − s ) γ ≤ z ≤ C q ( r 0 − s ) γ , z = H ≤ h ≤ q γ sup h s M Γ γ ( h ) .
In particular lim γ → ∞ z = ∞ \lim_{\gamma\to\infty}z=\infty lim γ → ∞ z = ∞ , since r 0 − s > 0 r_0-s>0 r 0 − s > 0 .
Step 2: choose a real parameter, then round it. For this fixed γ \gamma γ consider the continuous function
G γ ( x ) = ( x + 1 ) z N m x . G_\gamma(x)=\frac{(x+1)zN}{m^x}. G γ ( x ) = m x ( x + 1 ) z N .
At x = g + 2 x=g+2 x = g + 2 its value tends to infinity as γ → ∞ \gamma\to\infty γ → ∞ . At the other endpoint,
G γ ( γ ) ≤ C N ( γ + 1 ) q ( r 0 − s ) γ − r 0 γ = C N ( γ + 1 ) q − s γ ⟶ 0. G_\gamma(\gamma)\le CN(\gamma+1)q^{(r_0-s)\gamma-r_0\gamma}=CN(\gamma+1)q^{-s\gamma}\longrightarrow0. G γ ( γ ) ≤ C N ( γ + 1 ) q ( r 0 − s ) γ − r 0 γ = C N ( γ + 1 ) q − s γ ⟶ 0.
For all sufficiently large γ \gamma γ , these values lie respectively above and below τ \tau τ . The intermediate value theorem gives b ∈ ( g + 2 , γ ) b\in(g+2,\gamma) b ∈ ( g + 2 , γ ) with
( b + 1 ) z N m b = τ . \frac{(b+1)zN}{m^b}=\tau. m b ( b + 1 ) z N = τ .
Taking logarithms to base q q q yields the exact identity
r 0 b = log q z + log q ( b + 1 ) + log q ( N / τ ) . r_0b=\log_q z+\log_q(b+1)+\log_q(N/\tau). r 0 b = log q z + log q ( b + 1 ) + log q ( N / τ ) .
Step 1 gives log q z = ( r 0 − s ) γ + O ( 1 ) \log_q z=(r_0-s)\gamma+O(1) log q z = ( r 0 − s ) γ + O ( 1 ) . As g + 2 < b < γ g+2<b<\gamma g + 2 < b < γ , the other variable term is O ( log γ ) O(\log\gamma) O ( log γ ) . Hence, with a = ( r 0 − s ) / r 0 ∈ ( 0 , 1 ) a=(r_0-s)/r_0\in(0,1) a = ( r 0 − s ) / r 0 ∈ ( 0 , 1 ) ,
b = a γ + O ( log γ ) . b=a\gamma+O(\log\gamma). b = aγ + O ( log γ ) .
Here O ( log γ ) O(\log\gamma) O ( log γ ) means that the absolute value of the error is at most a fixed constant times log γ \log\gamma log γ for large γ \gamma γ .
Put
β = ⌊ b ⌋ , κ = τ m β z N , A = ⌊ κ ⌋ . \beta=\lfloor b\rfloor,\qquad\kappa=\frac{\tau m^\beta}{zN},\qquad A=\lfloor\kappa\rfloor. β = ⌊ b ⌋ , κ = z N τ m β , A = ⌊ κ ⌋ .
The symbols ⌊ x ⌋ \lfloor x\rfloor ⌊ x ⌋ and ⌈ x ⌉ \lceil x\rceil ⌈ x ⌉ denote the greatest integer at most x x x and the least integer at least x x x , respectively. The defining equation for b b b gives κ = ( b + 1 ) m β − b \kappa=(b+1)m^{\beta-b} κ = ( b + 1 ) m β − b . Since b − 1 < β ≤ b b-1<\beta\le b b − 1 < β ≤ b ,
b + 1 m ≤ κ ≤ b + 1 , b + 1 m − 1 ≤ A ≤ b + 1. \frac{b+1}{m}\le\kappa\le b+1,\qquad\frac{b+1}{m}-1\le A\le b+1. m b + 1 ≤ κ ≤ b + 1 , m b + 1 − 1 ≤ A ≤ b + 1.
In particular A ≥ 2 A\ge2 A ≥ 2 for large γ \gamma γ . Using 0 ≤ κ − A < 1 0\le\kappa-A<1 0 ≤ κ − A < 1 , we obtain
0 ≤ τ − A N z m β = ( κ − A ) N z m β < N z m β = τ m β − b b + 1 ≤ m τ b + 1 ⟶ 0. 0\le\tau-\frac{ANz}{m^\beta}=(\kappa-A)\frac{Nz}{m^\beta}<\frac{Nz}{m^\beta}=\frac{\tau m^{\beta-b}}{b+1}\le\frac{m\tau}{b+1}\longrightarrow0. 0 ≤ τ − m β A N z = ( κ − A ) m β N z < m β N z = b + 1 τ m β − b ≤ b + 1 m τ ⟶ 0.
Increase γ \gamma γ until this error is below ϵ / 2 \epsilon/2 ϵ /2 . This proves (ii) once the set of A N AN A N prefixes is constructed. The threshold for γ \gamma γ may depend on τ \tau τ and ϵ \epsilon ϵ , which are fixed throughout this construction.
Step 3: construct the words with the exact count. Set d = ⌈ log m A ⌉ d=\lceil\log_m A\rceil d = ⌈ log m A ⌉ and α = β − d \alpha=\beta-d α = β − d . Then m d ≥ A m^d\ge A m d ≥ A , so Γ d \Gamma_d Γ d contains at least A A A elements. Let Q A Q_A Q A consist of its A A A smallest elements and define
S = F − + q α Q A , F = S + q β Γ γ − β . S=F_-+q^\alpha Q_A,\qquad F=S+q^\beta\Gamma_{\gamma-\beta}. S = F − + q α Q A , F = S + q β Γ γ − β .
The estimates in Step 2 imply
β = a γ + O ( log γ ) , d = log m γ + O ( 1 ) , α = a γ + O ( log γ ) . \beta=a\gamma+O(\log\gamma),\qquad d=\log_m\gamma+O(1),\qquad\alpha=a\gamma+O(\log\gamma). β = aγ + O ( log γ ) , d = log m γ + O ( 1 ) , α = aγ + O ( log γ ) .
Thus α > g + 3 \alpha>g+3 α > g + 3 and α < β < γ \alpha<\beta<\gamma α < β < γ for large γ \gamma γ ; α \alpha α can exceed any fixed lower bound. Also β ≥ ( a / 2 ) γ = c γ \beta\ge(a/2)\gamma=c\gamma β ≥ ( a /2 ) γ = c γ and d ≍ log β d\asymp\log\beta d ≍ log β . The latter notation means that d / log β d/\log\beta d / log β stays between two positive constants for all sufficiently large γ \gamma γ .
The digit blocks of an element of F F F occupy disjoint positions: the first g g g positions contain a word from F − F_- F − ; positions g + 1 , … , α g+1,\ldots,\alpha g + 1 , … , α are zero; the next d d d positions give a word from Q A Q_A Q A ; and positions β + 1 , … , γ \beta+1,\ldots,\gamma β + 1 , … , γ are arbitrary. Base-q q q expansion is unique, so these sums have no repeated elements. In particular
# S = A N = P , # F = A N m γ − β . \#S=AN=P,\qquad\#F=ANm^{\gamma-\beta}. # S = A N = P , # F = A N m γ − β .
Both Q A Q_A Q A and the free block contain zero, so F − ⊂ F F_-\subset F F − ⊂ F . Deleting high digits from an element of Q A Q_A Q A decreases its value and keeps it in Γ d \Gamma_d Γ d , hence in the first A A A elements Q A Q_A Q A . Together with the corresponding property of F − F_- F − , this proves closure under digit truncation for F F F : a cut in the free block leaves a free word, a cut in the selected block leaves a word in Q A Q_A Q A , and a cut in the lowest g g g positions leaves an element of F − F_- F − . A cut in the intervening zero block simply leaves the old word. All assertions in (i) follow.
Step 4: control the endpoint error. Set
R = q α 4 q , R ′ = q β 4 q . R=\frac{q^\alpha}{4q},\qquad R'=\frac{q^\beta}{4q}. R = 4 q q α , R ′ = 4 q q β .
For large γ \gamma γ , H < R < R ′ < q γ H<R<R'<q^\gamma H < R < R ′ < q γ . Since A ≤ b + 1 ≤ γ + 1 A\le b+1\le\gamma+1 A ≤ b + 1 ≤ γ + 1 , we have P ≤ N ( γ + 1 ) P\le N(\gamma+1) P ≤ N ( γ + 1 ) . Also α / γ → a > 0 \alpha/\gamma\to a>0 α / γ → a > 0 , so eventually α ≥ a γ / 2 \alpha\ge a\gamma/2 α ≥ aγ /2 . Consequently
2 P R s ≤ 2 N ( γ + 1 ) ( 4 q ) s q − s a γ / 2 ⟶ 0. \frac{2P}{R^s}\le2N(\gamma+1)(4q)^s q^{-sa\gamma/2}\longrightarrow0. R s 2 P ≤ 2 N ( γ + 1 ) ( 4 q ) s q − s aγ /2 ⟶ 0.
Choose γ \gamma γ large enough that $2P/R^s<\epsilon/4.
Step 5: upper bounds in three ranges of radii. If H ≤ h < R H \le h < R H ≤ h < R , two points coming from different translated copies of F − F_{-} F − must differ in some digit above α \alpha α . The highest such digit and the separation estimate in Lemma 4.1 give distance at least q α / ( q − 1 ) > 2 R > 2 h q^{\alpha}/(q-1)>2R>2h q α / ( q − 1 ) > 2 R > 2 h . Thus an interval of radius h h h meets only one copy of F − F_{-} F − , and
M F ( h ) h s ≤ N H s . \frac{M_F(h)}{h^s}\le\frac{N}{H^s}. h s M F ( h ) ≤ H s N .
If R ≤ h < R ′ R\le h<R' R ≤ h < R ′ , the same argument above position β \beta β shows that an interval meets at most one translated copy of S S S . Therefore
M F ( h ) h s ≤ P R s < ϵ / 8. \frac{M_F(h)}{h^s}\le\frac{P}{R^s}<\epsilon/8. h s M F ( h ) ≤ R s P < ϵ /8.
If R ′ ≤ h ≤ q γ R'\le h\le q^\gamma R ′ ≤ h ≤ q γ , Lemma 4.3 and Steps 1–2 give
M F ( h ) h s ≤ θ M F γ ( h ) h s + 2 P h s ≤ θ z + 2 P R s ≤ τ + ϵ / 4. \frac{M_F(h)}{h^s}\le\theta\frac{M_{F_\gamma}(h)}{h^s}+\frac{2P}{h^s}\le\theta z+\frac{2P}{R^s}\le\tau+\epsilon/4. h s M F ( h ) ≤ θ h s M F γ ( h ) + h s 2 P ≤ θ z + R s 2 P ≤ τ + ϵ /4.
Combining the three estimates proves the upper bound in (iii).
Step 6: find a radius giving the lower bound. We must first show that the supremum defining z z z can be taken over [ R , q γ ] [R,q^\gamma] [ R , q γ ] . Lemma 4.2 and Step 1 imply
1 z sup H ≤ h < R M F γ ( h ) h s ≤ C ( 4 q ) − ( r 0 − s ) q − ( r 0 − s ) ( γ − α ) ⟶ 0. \frac{1}{z}\sup_{H\le h<R}\frac{M_{F_\gamma}(h)}{h^s}\le C(4q)^{-(r_0-s)}q^{-(r_0-s)(\gamma-\alpha)}\longrightarrow0. z 1 H ≤ h < R sup h s M F γ ( h ) ≤ C ( 4 q ) − ( r 0 − s ) q − ( r 0 − s ) ( γ − α ) ⟶ 0.
Indeed, Step 3 gives α / γ → a = ( r 0 − s ) / r 0 < 1 \alpha/\gamma\to a=(r_0-s)/r_0<1 α / γ → a = ( r 0 − s ) / r 0 < 1 , so γ − α → ∞ \gamma-\alpha\to\infty γ − α → ∞ . For sufficiently large γ \gamma γ , the supremum over [ H , R ) [H,R) [ H , R ) is therefore strictly smaller than z z z . Since Step 1 identifies z z z with the supremum over [ H , q γ ] [H,q^\gamma] [ H , q γ ] , we conclude that
z = sup R ≤ h ≤ q γ M F γ ( h ) h s . z=\sup_{R\le h\le q^\gamma}\frac{M_{F_\gamma}(h)}{h^s}. z = R ≤ h ≤ q γ sup h s M F γ ( h ) .
Now the definition of supremum and θ = P / m β > 0 \theta=P/m^\beta>0 θ = P / m β > 0 yield h ∗ ∈ [ R , q γ ] h_*\in[R,q^\gamma] h ∗ ∈ [ R , q γ ] such that
θ M F γ ( h ∗ ) h ∗ s > θ z − ϵ / 4. \theta\frac{M_{F_\gamma}(h_*)}{h_*^s}>\theta z-\epsilon/4. θ h ∗ s M F γ ( h ∗ ) > θ z − ϵ /4.
This choice uses only the definition of supremum and does not require that it be attained. The lower inequality from Lemma 4.3 now gives
M F ( h ∗ ) h ∗ s ≥ θ M F γ ( h ∗ ) h ∗ s − 2 P h ∗ s > θ z − ϵ / 4 − ϵ / 4 ≥ τ − ϵ . \frac{M_F(h_*)}{h_*^s}\ge\theta\frac{M_{F_\gamma}(h_*)}{h_*^s}-\frac{2P}{h_*^s}>\theta z-\epsilon/4-\epsilon/4\ge\tau-\epsilon. h ∗ s M F ( h ∗ ) ≥ θ h ∗ s M F γ ( h ∗ ) − h ∗ s 2 P > θ z − ϵ /4 − ϵ /4 ≥ τ − ϵ .
This proves the required lower bound.
Finally the exact count in Step 3 gives
# F q s γ = P m γ − β q s γ = θ q ( r 0 − s ) γ ≤ θ z ≤ τ . \frac{\#F}{q^{s\gamma}}=\frac{Pm^{\gamma-\beta}}{q^{s\gamma}}=\theta q^{(r_0-s)\gamma}\le\theta z\le\tau. q s γ # F = q s γ P m γ − β = θ q ( r 0 − s ) γ ≤ θ z ≤ τ .
All the requirements hold simultaneously for every sufficiently large γ \gamma γ , completing the proof. □ \square □
Lemma 4.5. Let s > 0 s>0 s > 0 . Suppose, for all sufficiently large h h h , M Λ ( h ) / h s ≤ C log h M_{\Lambda}(h)/h^{s}\le C\log h M Λ ( h ) / h s ≤ C log h , and suppose
lim sup h → ∞ ( log h ) M Λ ( h ) h s > 0. \limsup_{h\to\infty}(\log h)\frac{M_{\Lambda}(h)}{h^{s}}>0. h → ∞ lim sup ( log h ) h s M Λ ( h ) > 0. Then dim B e Λ = s \dim_{\mathrm{Be}}\Lambda=s dim Be Λ = s .
Proof. We prove the upper and lower dimension bounds separately. For any u > s u>s u > s , the assumed upper estimate gives, for all large h h h ,
0 ≤ M Λ ( h ) h u = M Λ ( h ) h s h s − u ≤ C log h h u − s ⟶ 0. 0\le\frac{M_{\Lambda}(h)}{h^{u}}=\frac{M_{\Lambda}(h)}{h^{s}}h^{s-u}\le C\frac{\log h}{h^{u-s}}\longrightarrow0. 0 ≤ h u M Λ ( h ) = h s M Λ ( h ) h s − u ≤ C h u − s log h ⟶ 0.
Thus D u + ( Λ ) = 0 D_{u}^{+}(\Lambda)=0 D u + ( Λ ) = 0 for every u > s u>s u > s and dim B e Λ ≤ s \dim_{\mathrm{Be}}\Lambda\le s dim Be Λ ≤ s . For the reverse inequality, a positive limsup means that there are c > 0 c>0 c > 0 and a sequence h j → ∞ h_j\to\infty h j → ∞ satisfying
M Λ ( h j ) h j s ≥ c log h j . \frac{M_{\Lambda}(h_j)}{h_j^{s}}\ge\frac{c}{\log h_j}. h j s M Λ ( h j ) ≥ log h j c .
If the limsup is finite, take c c c smaller than it; if infinite, any fixed positive c c c suffices. For 0 < u < s 0<u<s 0 < u < s it follows that
M Λ ( h j ) h j u = M Λ ( h j ) h j s h j s − u ≥ c h j s − u log h j ⟶ ∞ . \frac{M_{\Lambda}(h_j)}{h_j^{u}}=\frac{M_{\Lambda}(h_j)}{h_j^{s}}h_j^{s-u}\ge\frac{ch_j^{s-u}}{\log h_j}\longrightarrow\infty. h j u M Λ ( h j ) = h j s M Λ ( h j ) h j s − u ≥ log h j c h j s − u ⟶ ∞.
The limsup over all radii is at least the limsup over this sequence, so D u + ( Λ ) = ∞ D_{u}^{+}(\Lambda)=\infty D u + ( Λ ) = ∞ for every 0 < u < s 0<u<s 0 < u < s . Hence dim B e Λ ≥ s \dim_{\mathrm{Be}}\Lambda\ge s dim Be Λ ≥ s , completing the proof. □
Theorem 4.6. For every 0 < s < r 0 0<s<r_0 0 < s < r 0 and ζ ∈ [ 0 , + ∞ ] \zeta\in[0,+\infty] ζ ∈ [ 0 , + ∞ ] , there is a prefix-closed increasing word sequence ( J n ) (\mathcal{J}_n) ( J n ) such that
dim B e Λ ( J ) = s , D s + ( Λ ( J ) ) = ζ . \dim_{\mathrm{Be}}\Lambda(\mathcal{J})=s,\qquad D_s^{+}(\Lambda(\mathcal{J}))=\zeta. dim Be Λ ( J ) = s , D s + ( Λ ( J )) = ζ . It may be constructed with integers γ k − 1 < α k < β k < γ k \gamma_{k-1}<\alpha_k<\beta_k<\gamma_k γ k − 1 < α k < β k < γ k satisfying γ k − 1 = o ( γ k ) \gamma_{k-1}=o(\gamma_k) γ k − 1 = o ( γ k ) , β k ≥ c γ k \beta_k\ge c\gamma_k β k ≥ c γ k , β k − α k ≍ log β k \beta_k-\alpha_k\asymp\log\beta_k β k − α k ≍ log β k , and
J α k = J γ k − 1 0 α k − γ k − 1 , J γ k = J β k Ω γ k − β k . \mathcal{J}_{\alpha_k}=\mathcal{J}_{\gamma_{k-1}}0^{\alpha_k-\gamma_{k-1}},\qquad\mathcal{J}_{\gamma_k}=\mathcal{J}_{\beta_k}\Omega^{\gamma_k-\beta_k}. J α k = J γ k − 1 0 α k − γ k − 1 , J γ k = J β k Ω γ k − β k . At stage k k k , α k \alpha_k α k can exceed any finite lower bound specified after the preceding stage. More precisely, put
τ k = { ζ , 0 < ζ < ∞ , 1 / log m ( k + 1 ) , ζ = 0 , log m ( k + 1 ) , ζ = ∞ , e k = min { 1 , τ k } ( k + 1 ) 2 , H k = q γ k . \tau_k=
\begin{cases}
\zeta, & 0<\zeta<\infty,\\
1/\log_m(k+1), & \zeta=0,\\
\log_m(k+1), & \zeta=\infty,
\end{cases}
\qquad
e_k=\frac{\min\{1,\tau_k\}}{(k+1)^2},
\qquad
H_k=q^{\gamma_k}. τ k = ⎩ ⎨ ⎧ ζ , 1/ log m ( k + 1 ) , log m ( k + 1 ) , 0 < ζ < ∞ , ζ = 0 , ζ = ∞ , e k = ( k + 1 ) 2 min { 1 , τ k } , H k = q γ k . For k ≥ 2 k\ge2 k ≥ 2 , the sequence can be chosen so that
τ k − e k ≤ sup H k − 1 ≤ h ≤ H k M Λ ( J ) ( h ) h s ≤ max { τ k − 1 , τ k } + e k . (10) \tau_k-e_k\le\sup_{H_{k-1}\le h\le H_k}\frac{M_{\Lambda(\mathcal{J})}(h)}{h^s}\le\max\{\tau_{k-1},\tau_k\}+e_k. \tag*{(10)} τ k − e k ≤ H k − 1 ≤ h ≤ H k sup h s M Λ ( J ) ( h ) ≤ max { τ k − 1 , τ k } + e k . (10)
Proof. We construct the word sets stage by stage, then estimate their interval counts at all large radii.
Step 1: initialization and induction. Choose γ 0 ≥ 2 \gamma_0\ge2 γ 0 ≥ 2 , put F 0 = { 0 } F_0=\{0\} F 0 = { 0 } , and take J n = { 0 n } \mathcal{J}_n=\{0^n\} J n = { 0 n } for 0 ≤ n ≤ γ 0 0\le n\le\gamma_0 0 ≤ n ≤ γ 0 ; 0 0 0^0 0 0 denotes the empty word. Suppose the finite set F k − 1 ⊂ Γ γ k − 1 F_{k-1} \subset\Gamma_{\gamma_{k-1}} F k − 1 ⊂ Γ γ k − 1 and all word sets through length γ k − 1 \gamma_{k-1} γ k − 1 have been fixed. Apply Lemma 4.4 with
g = γ k − 1 , F − = F k − 1 , N k = # F k − 1 , τ = τ k , e = e k . g=\gamma_{k-1},\quad F_{-}=F_{k-1},\quad N_k=\#F_{k-1},\quad\tau=\tau_k,\quad e=e_k. g = γ k − 1 , F − = F k − 1 , N k = # F k − 1 , τ = τ k , e = e k .
Its hypotheses hold initially and are preserved by part (i) of that lemma. Choose γ k \gamma_k γ k sufficiently large that in addition
γ k ≥ k 2 , γ k ≥ ( k + 1 ) γ k − 1 , α k > γ k − 1 + 3 , \gamma_k \ge k^2,\qquad\gamma_k \ge(k+1)\gamma_{k-1},\qquad\alpha_k>\gamma_{k-1}+3, γ k ≥ k 2 , γ k ≥ ( k + 1 ) γ k − 1 , α k > γ k − 1 + 3 ,
and any specified finite lower bound on α k \alpha_k α k is satisfied. Since the asymptotic estimates in the lemma are statements for each fixed stage, enlarge γ k \gamma_k γ k also until β k ≥ c γ k \beta_k\ge c\gamma_k β k ≥ c γ k and 1 2 log m β k ≤ β k − α k ≤ 2 log m β k \frac{1}{2}\log_m\beta_k\le\beta_k-\alpha_k\le2\log_m\beta_k 2 1 log m β k ≤ β k − α k ≤ 2 log m β k . Indeed, for each fixed stage, the proof of Lemma 4.4 gives
β = a γ + O ( log γ ) , β − α = log m γ + O ( 1 ) , a = r 0 − s r 0 > 0. \beta=a\gamma+O(\log\gamma),\qquad\beta-\alpha=\log_m\gamma+O(1),\qquad a=\frac{r_0-s}{r_0}>0. β = aγ + O ( log γ ) , β − α = log m γ + O ( 1 ) , a = r 0 r 0 − s > 0.
Consequently, ( β − α ) / log m β → 1 (\beta-\alpha)/\log_m\beta\to1 ( β − α ) / log m β → 1 as γ → ∞ \gamma\to\infty γ → ∞ , which justifies these bounds. This gives constants independent of k k k in the asserted comparison. Let d k = β k − α k d_k=\beta_k-\alpha_k d k = β k − α k , and obtain
S k = F k − 1 + q α k Q A k , F k = S k + q β k Γ γ k − β k . S_k=F_{k-1}+q^{\alpha_k}Q_{A_k},\qquad F_k=S_k+q^{\beta_k}\Gamma_{\gamma_k-\beta_k}. S k = F k − 1 + q α k Q A k , F k = S k + q β k Γ γ k − β k .
Their sizes are # S k = A k N k \#S_k=A_kN_k # S k = A k N k and # F k = A k N k m γ k − β k \#F_k=A_kN_km^{\gamma_k-\beta_k} # F k = A k N k m γ k − β k . In particular, for H k = q γ k H_k=q^{\gamma_k} H k = q γ k , the last assertion of Lemma 4.4 gives
# F k H k s ≤ τ k . (11) \frac{\#F_k}{H_k^{s}} \le\tau_k. \tag*{(11)} H k s # F k ≤ τ k . (11)
The inequality γ k ≥ ( k + 1 ) γ k − 1 \gamma_k\ge(k+1)\gamma_{k-1} γ k ≥ ( k + 1 ) γ k − 1 implies γ k − 1 / γ k ≤ 1 / ( k + 1 ) → 0 \gamma_{k-1}/\gamma_k\le1/(k+1)\to0 γ k − 1 / γ k ≤ 1/ ( k + 1 ) → 0 , as required.
Step 2: define all intermediate word sets. Let Q k ⊂ Ω d k Q_k\subset\Omega^{d_k} Q k ⊂ Ω d k be the words corresponding to Q A k Q_{A_k} Q A k . At the endpoint set explicitly
J γ k = J γ k − 1 0 α k − γ k − 1 Q k Ω γ k − β k . \mathcal{J}_{\gamma_k}=\mathcal{J}_{\gamma_{k-1}}0^{\alpha_k-\gamma_{k-1}}Q_k\Omega^{\gamma_k-\beta_k}. J γ k = J γ k − 1 0 α k − γ k − 1 Q k Ω γ k − β k .
Concatenation here means taking every choice from each of the listed sets of words. Its numerical image is exactly F k F_k F k . For γ k − 1 ≤ n ≤ γ k \gamma_{k-1}\le n\le\gamma_k γ k − 1 ≤ n ≤ γ k , define J n \mathcal{J}_n J n as the set of first n n n digits of words in J γ k \mathcal{J}_{\gamma_k} J γ k . Part (i) of Lemma 4.4 ensures that at the lower endpoint this gives the already specified J γ k − 1 \mathcal{J}_{\gamma_{k-1}} J γ k − 1 . Thus consecutive stages agree and no earlier definition is changed.
Every allowed prefix is still allowed at a shorter length, so the family is prefix closed. Closure of F k F_k F k under digit truncation ensures that, if an allowed word is extended by a zero digit, that extension also occurs. Explicitly, truncate a word of F k F_k F k to the given prefix and pad the resulting element of F k F_k F k by zeros. This proves J n 0 ⊂ J n + 1 \mathcal{J}_n0\subset\mathcal{J}_{n+1} J n 0 ⊂ J n + 1 . The second inclusion in Definition 3.1 follows from prefix closure: if J 0 ⊂ J n + 1 \mathcal{J}_0\subset\mathcal{J}_{n+1} J 0 ⊂ J n + 1 , then J ∈ J n \mathcal{J}\in\mathcal{J}_n J ∈ J n . Zero belongs at every length, so this is an increasing word sequence in the stated sense.
All digits between γ k − 1 + 1 \gamma_{k-1}+1 γ k − 1 + 1 and α k \alpha_k α k are zero, and the last block is free. Therefore
J α k = J γ k − 1 0 α k − γ k − 1 , J γ k = J β k Ω γ k − β k . \mathcal{J}_{\alpha_k}=\mathcal{J}_{\gamma_{k-1}}0^{\alpha_k-\gamma_{k-1}},\qquad\mathcal{J}_{\gamma_k}=\mathcal{J}_{\beta_k}\Omega^{\gamma_k-\beta_k}. J α k = J γ k − 1 0 α k − γ k − 1 , J γ k = J β k Ω γ k − β k .
Also F k − 1 ⊂ F k F_{k-1}\subset F_k F k − 1 ⊂ F k and Λ ( J ) = ⋃ k F k \Lambda(\mathcal{J})=\bigcup_k F_k Λ ( J ) = ⋃ k F k .
Step 3: later stages do not alter counts at earlier radii. Fix k k k and 0 < h ≤ H k 0<h\le H_k 0 < h ≤ H k . Suppose two points of B ( x , h ) ∩ Λ ( J ) B(x,h)\cap\Lambda(\mathcal{J}) B ( x , h ) ∩ Λ ( J ) have different tails above position γ k \gamma_k γ k . All their digits from γ k + 1 \gamma_k+1 γ k + 1 through α k + 1 \alpha_{k+1} α k + 1 are zero, so their highest differing position is at least α k + 1 + 1 \alpha_{k+1}+1 α k + 1 + 1 . Lemma 4.1’s separation estimate would give distance at least
q α k + 1 q − 1 > 2 q γ k = 2 H k ≥ 2 h , \frac{q^{\alpha_{k+1}}}{q-1}>2q^{\gamma_k}=2H_k\ge2h, q − 1 q α k + 1 > 2 q γ k = 2 H k ≥ 2 h ,
contradicting membership in the same interval. Thus all points share the same tail. Subtracting it puts their prefixes in F k F_k F k by prefix closure and keeps them in an interval of radius h h h . It follows that M Λ ( J ) ( h ) ≤ M F k ( h ) M_{\Lambda(\mathcal{J})}(h)\le M_{F_k}(h) M Λ ( J ) ( h ) ≤ M F k ( h ) . Since F k ⊂ Λ ( J ) F_k\subset\Lambda(\mathcal{J}) F k ⊂ Λ ( J ) , the reverse inequality holds:
M Λ ( J ) ( h ) = M F k ( h ) ( 0 < h ≤ H k ) . M_{\Lambda(\mathcal{J})}(h)=M_{F_k}(h)\qquad(0<h\le H_k). M Λ ( J ) ( h ) = M F k ( h ) ( 0 < h ≤ H k ) .
This equality is why estimates for one finite stage remain valid for the final infinite set.
Step 4: upper and lower bounds on one stage of radii. For k ≥ 2 k\ge2 k ≥ 2 and H k − 1 ≤ h ≤ H k H_{k-1}\le h\le H_k H k − 1 ≤ h ≤ H k , Lemma 4.4(iii) and Step 3 give
τ k − ϵ k ≤ sup H k − 1 ≤ h ≤ H k M Λ ( J ) ( h ) h s ≤ max { # F k − 1 H k − 1 s , τ k } + ϵ k . \tau_k-\epsilon_k\le\sup_{H_{k-1}\le h\le H_k}\frac{M_{\Lambda(\mathcal{J})}(h)}{h^s}\le\max\left\{\frac{\#F_{k-1}}{H_{k-1}^s},\tau_k\right\}+\epsilon_k. τ k − ϵ k ≤ H k − 1 ≤ h ≤ H k sup h s M Λ ( J ) ( h ) ≤ max { H k − 1 s # F k − 1 , τ k } + ϵ k .
Use (4.4) from stage k − 1 k-1 k − 1 in the right side to obtain (4.3). The definition of supremum also supplies, for every k ≥ 2 k\ge2 k ≥ 2 , a radius h k ∈ [ H k − 1 , H k ] h_k\in[H_{k-1},H_k] h k ∈ [ H k − 1 , H k ] with
M Λ ( J ) ( h k ) h k s > τ k − 2 ϵ k . \frac{M_{\Lambda(\mathcal{J})}(h_k)}{h_k^s}>\tau_k-2\epsilon_k. h k s M Λ ( J ) ( h k ) > τ k − 2 ϵ k .
As h k ≥ H k − 1 → ∞ h_k\ge H_{k-1}\to\infty h k ≥ H k − 1 → ∞ , these radii can be used to bound the limsup from below.
Step 5: compute the density in each of the three cases. If 0 < ζ < ∞ 0<\zeta<\infty 0 < ζ < ∞ , all targets equal ζ \zeta ζ and ϵ k → 0 \epsilon_k\to0 ϵ k → 0 . The upper estimate valid on every stage gives D s + ≤ ζ D_s^+\le\zeta D s + ≤ ζ ; the radii h k h_k h k give D s + ≥ ζ D_s^+\ge\zeta D s + ≥ ζ .
If ζ = 0 \zeta=0 ζ = 0 , then τ k ↓ 0 \tau_k\downarrow0 τ k ↓ 0 . On the k k k th stage the upper bound is τ k − 1 + ϵ k → 0 \tau_{k-1}+\epsilon_k\to0 τ k − 1 + ϵ k → 0 . Since the stages cover all sufficiently large radii and counting quotients are nonnegative, D s + = 0 D_s^+=0 D s + = 0 . At the same time the lower estimates remain positive:
τ k − 2 ϵ k ≥ τ k ( 1 − 2 ( k + 1 ) 2 ) ≥ 1 2 log m ( k + 1 ) ( k ≥ 2 ) . \tau_k-2\epsilon_k\ge\tau_k\left(1-\frac{2}{(k+1)^2}\right)\ge\frac{1}{2\log_m(k+1)}
\qquad(k\ge2). τ k − 2 ϵ k ≥ τ k ( 1 − ( k + 1 ) 2 2 ) ≥ 2 log m ( k + 1 ) 1 ( k ≥ 2 ) .
This quantitative lower bound will preserve the dimension.
If ζ = ∞ \zeta=\infty ζ = ∞ , then τ k = log m ( k + 1 ) → ∞ \tau_k=\log_m(k+1)\to\infty τ k = log m ( k + 1 ) → ∞ and ϵ k → 0 \epsilon_k\to0 ϵ k → 0 . The quotients at h k h_k h k tend to infinity, so D s + = ∞ D_s^+=\infty D s + = ∞ . We have proved D s + ( Λ ( J ) ) = ζ D_s^+(\Lambda(\mathcal{J}))=\zeta D s + ( Λ ( J )) = ζ in every case.
Step 6: verify that the dimension is always s s s . When ζ = 0 \zeta=0 ζ = 0 or ζ = ∞ \zeta=\infty ζ = ∞ , the density alone does not determine the dimension. We verify the hypotheses of Lemma 4.5 for all three cases. For all large k k k , all three choices of target satisfy max { τ k − 1 , τ k } + ϵ k ≤ C 1 ( 1 + log ( k + 1 ) ) \max\{\tau_{k-1},\tau_k\}+\epsilon_k\le C_1(1+\log(k+1)) max { τ k − 1 , τ k } + ϵ k ≤ C 1 ( 1 + log ( k + 1 )) , where C 1 C_1 C 1 may depend on the fixed ζ \zeta ζ when it is finite positive. For h ∈ [ H k − 1 , H k ] h \in[H_{k-1},H_k] h ∈ [ H k − 1 , H k ] we have log h ≥ γ k − 1 log q ≥ ( k − 1 ) 2 log q \log h \ge\gamma_{k-1}\log q \ge(k-1)^2\log q log h ≥ γ k − 1 log q ≥ ( k − 1 ) 2 log q . Increasing a constant to cover the initial stages, (4.3) therefore gives
M Λ ( J ) ( h ) h s ≤ C log h (12) \frac{M_{\Lambda(\mathcal{J})}(h)}{h^s} \le C\log h \tag*{(12)} h s M Λ ( J ) ( h ) ≤ C log h (12)
for all sufficiently large h h h . For ζ = 0 \zeta=0 ζ = 0 , Step 5 gives M Λ ( J ) ( h k ) / h k s ≥ 1 / ( 2 log m ( k + 1 ) ) M_{\Lambda(\mathcal{J})}(h_k)/h_k^s \ge1/(2\log_m(k+1)) M Λ ( J ) ( h k ) / h k s ≥ 1/ ( 2 log m ( k + 1 )) . For 0 < ζ < ∞ 0<\zeta<\infty 0 < ζ < ∞ , this quotient is eventually at least ζ / 2 \zeta/2 ζ /2 , and for ζ = ∞ \zeta=\infty ζ = ∞ it is eventually at least 1 1 1 . Since 1 / ( 2 log m ( k + 1 ) ) → 0 1/(2\log_m(k+1))\to0 1/ ( 2 log m ( k + 1 )) → 0 , the same lower bound thus holds in all three cases for sufficiently large k k k . Consequently,
( log h k ) M Λ ( J ) ( h k ) h k s ≥ ( k − 1 ) 2 log q 2 log m ( k + 1 ) ⟶ ∞ . (\log h_k)\frac{M_{\Lambda(\mathcal{J})}(h_k)}{h_k^s}\ge\frac{(k-1)^2\log q}{2\log_m(k+1)}\longrightarrow\infty. ( log h k ) h k s M Λ ( J ) ( h k ) ≥ 2 log m ( k + 1 ) ( k − 1 ) 2 log q ⟶ ∞.
Thus the positive-limsup condition of Lemma 4.5 holds in all cases. That lemma proves dim B e Λ ( J ) = s \dim_{\mathrm{Be}}\Lambda(\mathcal{J})=s dim Be Λ ( J ) = s . □
Proof of Theorem 1.1. We combine the prescribed word construction with the spectral extension, keeping the order of the choices explicit.
Step 1: normalize the digits and fix the targets. Lemma 3.1 allows us to assume 0 ∈ L ⊂ { 0 , … , q − 2 } 0\in L\subset\{0,\ldots,q-2\} 0 ∈ L ⊂ { 0 , … , q − 2 } , without changing μ q , D \mu_{q,D} μ q , D or T \mathcal{T} T . Fix 0 < s < r 0 0<s<r_0 0 < s < r 0 and ζ ∈ [ 0 , + ∞ ] \zeta\in[0,+\infty] ζ ∈ [ 0 , + ∞ ] . Use the targets τ k \tau_k τ k and errors ϵ k \epsilon_k ϵ k of Theorem 4.6 . Also choose a recurrent enumeration ( a k ) (a_k) ( a k ) of T + \mathcal{T}_+ T + . For example, if T + = { t 1 , t 2 , … } \mathcal{T}_+=\{t_1,t_2,\ldots\} T + = { t 1 , t 2 , … } , concatenate ( t 1 ) (t_1) ( t 1 ) , ( t 1 , t 2 ) (t_1,t_2) ( t 1 , t 2 ) , ( t 1 , t 2 , t 3 ) (t_1,t_2,t_3) ( t 1 , t 2 , t 3 ) , and so on; for a finite set repeat its entire list. Every multiplier then occurs infinitely often.
Step 2: perform the two constructions in alternating order. Initially take m 0 = 0 m_0=0 m 0 = 0 and M 0 = 0 M_0=0 M 0 = 0 , and initialize the finite density core as in Theorem 4.6 . Suppose the choices through stage k − 1 k-1 k − 1 have been made. At this point the integers m k − 1 m_{k-1} m k − 1 and all frequencies for the finitely many binary encodings through stage k − 1 k-1 k − 1 are known. In particular the maximum M k − 1 M_{k-1} M k − 1 defined in Theorem 3.4 is known. First choose an integer v k ≥ 1 v_k\ge1 v k ≥ 1 with
a k q − v k < δ a k / 4. a_kq^{-v_k}<\delta_{a_k}/4. a k q − v k < δ a k /4.
This choice depends only on a k a_k a k and δ a k \delta_{a_k} δ a k .
Apply the next stage of Lemma 4.4 with the prescribed τ k , ϵ k \tau_k,\epsilon_k τ k , ϵ k , requiring additionally α k > m k − 1 + 3 \alpha_k>m_{k-1}+3 α k > m k − 1 + 3 and
a k q − ( γ k + v k ) M k − 1 < δ a k / 4. a_kq^{-(\gamma_k+v_k)}M_{k-1}<\delta_{a_k}/4. a k q − ( γ k + v k ) M k − 1 < δ a k /4.
The first requirement is possible because α k \alpha_k α k can exceed any specified finite bound; the second holds for sufficiently large γ k \gamma_k γ k because M k − 1 M_{k-1} M k − 1 is already fixed. All other density requirements are those in Theorem 4.6 . Put m k = γ k + v k m_k=\gamma_k+v_k m k = γ k + v k . After the density core is fixed through γ k \gamma_k γ k , extend it by zero digits through m k m_k m k . In the next density stage impose α k + 1 > m k + 3 \alpha_{k+1}>m_k+3 α k + 1 > m k + 3 , ensuring that these zero extensions remain exactly the allowed words there. Thus
J m k = J γ k 0 m k − γ k , a k q γ k − m k < δ a k / 4 , a k q − m k M k − 1 < δ a k / 4. \mathcal{J}_{m_k}=\mathcal{J}_{\gamma_k}0^{m_k-\gamma_k},\qquad a_kq^{\gamma_k-m_k}<\delta_{a_k}/4,\qquad a_kq^{-m_k}M_{k-1}<\delta_{a_k}/4. J m k = J γ k 0 m k − γ k , a k q γ k − m k < δ a k /4 , a k q − m k M k − 1 < δ a k /4.
These are precisely the size conditions for spectral stage k k k . Carry out that stage for all its finite binary encodings using Theorem 3.4 . Its successive choices are next density stage. This proves inductively that all stages can be completed; no parameter is chosen using a future unknown maximum.
Step 3: obtain common spectra with the prescribed core. Theorem 4.6 gives
dim B e Λ ( J ) = s , D s + ( Λ ( J ) ) = ζ . \dim_{\mathrm{Be}}\Lambda(\mathcal{J})=s,\qquad D_s^+(\Lambda(\mathcal{J}))=\zeta. dim Be Λ ( J ) = s , D s + ( Λ ( J )) = ζ .
Theorem 3.4 supplies, for every I ∈ { 1 , 2 } N I\in\{1,2\}^{\mathbb{N}} I ∈ { 1 , 2 } N ,
Λ ( I ) = Λ ( J ) ∪ E ( I ) , t Λ ( I ) is a spectrum of μ ( t ∈ T ) , \Lambda(I)=\Lambda(\mathcal{J})\cup E(I),\qquad t\Lambda(I)\text{ is a spectrum of }\mu\quad(t\in\mathcal{T}), Λ ( I ) = Λ ( J ) ∪ E ( I ) , t Λ ( I ) is a spectrum of μ ( t ∈ T ) ,
where E ( I ) E(I) E ( I ) is positive and q q q -lacunary. In the recursion, a core word has carry zero and a core parent; induction gives λ k ( J , I ∣ k ) = T m k ( J ) \lambda_k(J,I|_k)=T_{m_k}(J) λ k ( J , I ∣ k ) = T m k ( J ) for every J ∈ J m k J\in\mathcal{J}_{m_k} J ∈ J m k . Thus the same numerical core occurs for every binary encoding, as required.
Step 4: preserve the upper Beurling densities. For any h > 0 h>0 h > 0 , inclusion and the union bound give
M Λ ( J ) ( h ) ≤ M Λ ( I ) ( h ) ≤ M Λ ( J ) ( h ) + M E ( I ) ( h ) . M_{\Lambda(\mathcal{J})}(h)\leq M_{\Lambda(I)}(h)\leq M_{\Lambda(\mathcal{J})}(h)+M_{E(I)}(h). M Λ ( J ) ( h ) ≤ M Λ ( I ) ( h ) ≤ M Λ ( J ) ( h ) + M E ( I ) ( h ) .
Fix u > 0 u>0 u > 0 and divide by h u h^u h u . Lemma 3.5 implies M E ( I ) ( h ) / h u → 0 M_{E(I)}(h)/h^u\to0 M E ( I ) ( h ) / h u → 0 , so the two other quotients have the same limsup. If the core limsup is infinite, the lower inequality alone already gives infinity for the union. Hence in the extended range [ 0 , + ∞ ] [0,+\infty] [ 0 , + ∞ ] we have
D u + ( Λ ( I ) ) = D u + ( Λ ( J ) ) ( u > 0 ) . D_u^+(\Lambda(I))=D_u^+(\Lambda(\mathcal{J}))\quad(u>0). D u + ( Λ ( I )) = D u + ( Λ ( J )) ( u > 0 ) .
Taking u = s u=s u = s proves the required density, including zero and infinity. Taking the infimum of the exponents for which this density is zero proves the required Beurling dimension s s s .
Step 5: count the distinct spectra. Theorem 3.4 (iii) proves that different infinite binary sequences give different sets Λ ( I ) \Lambda(I) Λ ( I ) . The set of such sequences has cardinality 2 ℵ 0 2^{\aleph_0} 2 ℵ 0 , so the required family has at least that cardinality. For the reverse bound, every countable subset of R \mathbb{R} R can be listed in a sequence, and
# R N = ( 2 ℵ 0 ) ℵ 0 = 2 ℵ 0 . \#\mathbb{R}^{\mathbb{N}}=(2^{\aleph_0})^{\aleph_0}=2^{\aleph_0}. # R N = ( 2 ℵ 0 ) ℵ 0 = 2 ℵ 0 .
There are therefore at most continuum many countable frequency sets, and equality follows. All constructed frequencies are nonnegative integers and zero is retained at every stage. This completes the proof. □
Remark 4.2. The family has continuum cardinality, although every individual spectrum is countable.
School of Mathematics, Hunan University, Changsha Hunan 410082, China
Email address: zyluhnsd@163.com, zyluzky@amss.ac.cn