Introduction
The θ \theta θ -intermediate dimensions were introduced by Falconer, Fraser and Kempton [35 ] as a continuum of dimensions that interpolates between the Hausdorff dimension and the box-counting dimensions. For θ ∈ [ 0 , 1 ] \theta\in[0,1] θ ∈ [ 0 , 1 ] , the lower and upper θ \theta θ -intermediate dimensions \underline{\mbox{\rm dim}}_{\,\theta} and \overline{\mbox{\rm dim}}_{\,\theta} are defined by restricting the diameters of the covering sets to an interval of the form [ δ 1 / θ , δ ] [\delta^{1/\theta},\delta] [ δ 1/ θ , δ ] (see Definition 2.1 ). Thus θ = 0 \theta=0 θ = 0 corresponds to the Hausdorff dimension and θ = 1 \theta=1 θ = 1 to the lower and upper box dimensions. For every bounded set E ⊂ R d E\subset\mathbb{R}^d E ⊂ R d ,
\dim_{\rm H}E\leq \underline{\mbox{\rm dim}}_{\,\theta}E\leq \overline{\mbox{\rm dim}}_{\,\theta}E\leq \overline{\mbox{\rm dim}}_{\rm B}\,E
\qquad\text{and}\qquad
\underline{\mbox{\rm dim}}_{\,\theta}E\leq \underline{\mbox{\rm dim}}_{\rm B}\,E ,
and both functions of θ \theta θ are non-decreasing and continuous on ( 0 , 1 ] (0,1] ( 0 , 1 ] . These dimensions have been studied in many settings, for instance for fractional Brownian images [28 , 36 ] , orthogonal projections [29 , 23 ] , elliptical polynomial spirals [30 ] . There are also many generalizations, for instance [25 ] for intermediate dimensions in metric spaces and [21 ] for intermediate dimensions of measures.
A natural problem is to determine which functions arise as intermediate dimensions. This problem was solved by Banaji and Rutar [26 ] . For 0 ≤ λ ≤ α ≤ d 0\leq\lambda\leq\alpha\leq d 0 ≤ λ ≤ α ≤ d , they introduced the class H ( λ , α ) \mathcal{H}(\lambda,\alpha) H ( λ , α ) of non-decreasing functions h : [ 0 , 1 ] → [ λ , α ] h:[0,1]\to[\lambda,\alpha] h : [ 0 , 1 ] → [ λ , α ] which are continuous on ( 0 , 1 ] (0,1] ( 0 , 1 ] and whose growth is controlled by the Dini condition
D + h ( θ ) ≤ ( h ( θ ) − λ ) ( α − h ( θ ) ) ( α − λ ) θ , θ ∈ ( 0 , 1 ) . D^+h(\theta)\leq\frac{(h(\theta)-\lambda)(\alpha-h(\theta))}{(\alpha-\lambda)\,\theta},
\qquad \theta\in(0,1). D + h ( θ ) ≤ ( α − λ ) θ ( h ( θ ) − λ ) ( α − h ( θ )) , θ ∈ ( 0 , 1 ) .
They proved the following. As F F F ranges over the bounded subsets of R d \mathbb{R}^d R d with lower Assouad dimension λ \lambda λ and Assouad dimension α \alpha α , the pairs (\underline{\mbox{\rm dim}}_{\,\theta}F,\overline{\mbox{\rm dim}}_{\,\theta}F) are exactly the pairs ( h 1 , h 2 ) (h_1,h_2) ( h 1 , h 2 ) of functions in H ( λ , α ) \mathcal{H}(\lambda,\alpha) H ( λ , α ) with h 1 ≤ h 2 h_1\leq h_2 h 1 ≤ h 2 and h 1 ( 0 ) = h 2 ( 0 ) h_1(0)=h_2(0) h 1 ( 0 ) = h 2 ( 0 ) (see Theorem 4.2 ). In this paper we study the analogous problem within a classical family of subsets of the line: the complementary sets of a sequence.
Let a = { a n } a=\{a_n\} a = { a n } be a positive, non-increasing sequence with ∑ n a n = 1 \sum_n a_n=1 ∑ n a n = 1 . The class C a \mathcal{C}_a C a of complementary sets of a a a consists of all closed sets E ⊂ [ 0 , 1 ] E\subset[0,1] E ⊂ [ 0 , 1 ] whose complement in [ 0 , 1 ] [0,1] [ 0 , 1 ] is a disjoint union of open intervals whose lengths are exactly the terms of a a a . This class contains sets of very different nature. Two of them play a central role. The first is the countable set D a D_a D a , a decreasing sequence of points whose consecutive gaps are a 1 , a 2 , … a_1,a_2,\dots a 1 , a 2 , … . The second is the decreasing Cantor set C a C_a C a , obtained by removing the gaps in decreasing order along a binary Cantor construction (see § 3 ). The question of which values a given dimension takes on C a \mathcal{C}_a C a has a long history. For the Hausdorff dimension it goes back to Besicovitch and Taylor [27 ] , and the set of attainable values is the closed interval [ 0 , dim H C a ] [0,\dim_{\rm H}C_a] [ 0 , dim H C a ] [27 , 31 , 38 ] ; related questions for packing measures were considered in [39 ] . In contrast, the lower and upper box dimensions do not depend on the arrangement of the gaps, so each of them takes a single value on C a \mathcal{C}_a C a [33 Chapter 3.2] . For the Assouad and lower Assouad dimensions, García, Hare and Mendivil [37 ] showed that C a C_a C a and D a D_a D a are extremal. They also showed that, under suitable decay conditions on a a a , the attainable values form the intervals [ dim A C a , 1 ] [\dim_A C_a,1] [ dim A C a , 1 ] and [ 0 , dim L C a ] [0,\dim_L C_a] [ 0 , dim L C a ] , respectively.
The intermediate dimensions of complementary sets were studied in [24 ] for each fixed θ \theta θ . There it was shown that D a D_a D a and C a C_a C a are again extremal: for every E ∈ C a E\in\mathcal{C}_a E ∈ C a ,
\overline{\mbox{\rm dim}}_{\,\theta}D_a\leq\overline{\mbox{\rm dim}}_{\,\theta}E\leq\overline{\mbox{\rm dim}}_{\,\theta}C_a ,
where the upper bound requires a mild assumption on a a a , and the analogous inequalities hold for the lower intermediate dimensions. Moreover, \overline{\mbox{\rm dim}}_{\,\theta}D_a=\theta B/(1-(1-\theta)B) with B=\overline{\mbox{\rm dim}}_{\rm B}\,C_a . This is the smallest value allowed by the bounds of Banaji and Rutar for a set with upper box dimension B B B . Finally, it was proved that every value between these two bounds is attained for each fixed θ \theta θ , both for the upper and for the lower intermediate dimensions, provided that
2 < inf n s n ( a ) s n + 1 ( a ) ≤ sup n s n ( a ) s n + 1 ( a ) < ∞ . 2<\inf_n\frac{s_n(a)}{s_{n+1}(a)}\leq\sup_n\frac{s_n(a)}{s_{n+1}(a)}<\infty . 2 < n inf s n + 1 ( a ) s n ( a ) ≤ n sup s n + 1 ( a ) s n ( a ) < ∞.
Here s n ( a ) = 2 − n ∑ i ≥ 2 n a i s_n(a)=2^{-n}\sum_{i\geq 2^n}a_i s n ( a ) = 2 − n ∑ i ≥ 2 n a i is the average length of the intervals of the n n n -th step in the construction of C a C_a C a . This condition, which we call hypothesis (H) (see § 4 ), combines the conditions under which the Assouad and the lower Assouad dimensions of complementary sets fill an interval [37 ] .
These results describe, for each fixed θ \theta θ , the set of attainable values. They do not describe how the intermediate dimensions of a single complementary set vary with θ \theta θ , which is the information captured by the theorem of Banaji and Rutar. This leads to the question, raised in [24 Question 1] , of describing the families
\mathcal{F}^U(a):=\{\theta\mapsto\overline{\mbox{\rm dim}}_{\,\theta}E : E\in\mathcal{C}_a\}
and
\mathcal{F}^L(a):=\{\theta\mapsto\underline{\mbox{\rm dim}}_{\,\theta}E : E\in\mathcal{C}_a\}.
Three necessary conditions for a function h h h to belong to F U ( a ) \mathcal{F}^U(a) F U ( a ) follow from the results above. By the theorem of Banaji and Rutar, h ∈ H ( 0 , 1 ) h\in\mathcal{H}(0,1) h ∈ H ( 0 , 1 ) . Since all sets in C a \mathcal{C}_a C a have the same upper box dimension, h(1)=\overline{\mbox{\rm dim}}_{\rm B}\,C_a . Finally, by the extremality of C a C_a C a , h\leq\overline{\mbox{\rm dim}}_{\,\theta}C_a .
Our main result shows that, under hypothesis (H) , these conditions are also sufficient. More precisely (Theorem 4.17 and Corollary 4.18 ), if a a a satisfies (H) , then F U ( a ) \mathcal{F}^U(a) F U ( a ) is exactly the set of functions h ∈ H ( 0 , 1 ) h\in\mathcal{H}(0,1) h ∈ H ( 0 , 1 ) such that
h(1)=\overline{\mbox{\rm dim}}_{\rm B}\,C_a \qquad\text{and}\qquad h(\theta)\leq\overline{\mbox{\rm dim}}_{\,\theta}C_a\ \text{ for all }\theta\in[0,1],
with the usual convention dim ‾ 0 = dim H \overline{\dim}_{\,0}=\dim_{\rm H} dim 0 = dim H . In particular, F U ( a ) \mathcal{F}^U(a) F U ( a ) depends on a a a only through the function \theta\mapsto\overline{\mbox{\rm dim}}_{\,\theta}C_a , and it is much larger than the family of functions realized in [24 ] . For instance, suppose that the box dimension of C a C_a C a exists. Since \dim_{\rm H}C_a=\underline{\mbox{\rm dim}}_{\rm B}\,C_a , we then have \overline{\mbox{\rm dim}}_{\,\theta}C_a=\dim_{\rm B} C_a for every θ \theta θ , and the second condition is automatic. In this case F U ( a ) \mathcal{F}^U(a) F U ( a ) consists of all functions h ∈ H ( 0 , 1 ) h\in\mathcal{H}(0,1) h ∈ H ( 0 , 1 ) with h ( 1 ) = dim B C a h(1)=\dim_{\rm B}C_a h ( 1 ) = dim B C a . By the theorem of Banaji and Rutar, these are exactly the upper intermediate dimension functions of bounded subsets of R \mathbb{R} R with this upper box dimension. This applies, for example, to the middle third Cantor set.
We also obtain a sharp lower bound with a rigidity property. Every h ∈ F U ( a ) h\in\mathcal{F}^U(a) h ∈ F U ( a ) satisfies h\geq\overline{\mbox{\rm dim}}_{\,\theta}D_a . If h h h coincides with \overline{\mbox{\rm dim}}_{\,\theta}D_a at a single θ 0 ∈ ( 0 , 1 ] \theta_0\in(0,1] θ 0 ∈ ( 0 , 1 ] , then it coincides with it on the whole interval [ θ 0 , 1 ] [\theta_0,1] [ θ 0 , 1 ] (Proposition 4.8 ). Moreover, equality at some θ 0 ∈ ( 0 , 1 ) \theta_0\in(0,1) θ 0 ∈ ( 0 , 1 ) forces the lower Assouad and Assouad dimensions of the set to be 0 0 0 and 1 1 1 , respectively (Proposition 4.10 ). These facts follow from a simple reformulation of the Dini condition defining H ( 0 , 1 ) \mathcal{H}(0,1) H ( 0 , 1 ) as the monotonicity of the quotient h ( θ ) / ( θ ( 1 − h ( θ ) ) ) h(\theta)/(\theta(1-h(\theta))) h ( θ ) / ( θ ( 1 − h ( θ ))) (Lemma 4.7 ). This reformulation also gives a short proof of the lower bound \overline{\mbox{\rm dim}}_{\,\theta}D_a\leq\overline{\mbox{\rm dim}}_{\,\theta}E .
The proof of Theorem 4.17 follows the strategy of Banaji and Rutar [26 ] . One first constructs, by concatenating mountains and valleys, a function whose infima over windows of length log ( 1 / θ ) \log(1/\theta) log ( 1/ θ ) reproduce h ( θ ) h(\theta) h ( θ ) . This function is then converted into a sequence of scales for a Cantor-type set. In our setting, two new difficulties appear. First, the set must be built using only the terms of a a a as gaps. This forces its scales to be, up to constants, at most those of C a C_a C a , while the hypothesis h\leq\overline{\mbox{\rm dim}}_{\,\theta}C_a only provides information on infima over windows. We overcome this by combining the scales produced by the mountains with those of C a C_a C a , and by placing each mountain on a window where the covering numbers of C a C_a C a are large. Second, the function h h h may saturate the Dini condition defining H ( 0 , 1 ) \mathcal{H}(0,1) H ( 0 , 1 ) . In that case, any set whose upper intermediate dimensions are given by h h h must have lower Assouad dimension 0 0 0 and Assouad dimension 1 1 1 . The ratios between consecutive scales of such a set cannot remain bounded away from 0 0 0 and from 1 / 2 1/2 1/2 . This cannot be achieved with the decreasing Cantor sets used in [24 ] (see Remark 4.20 ). For this reason we work with limit sets of homogeneous binary Cantor schemes. The gaps of these schemes are terms of a a a , but they are not arranged in decreasing order, and their intermediate dimensions depend only on their scales (Lemma 4.12 ). The remaining terms of a a a are used to build a countable set, which does not increase the dimension because of the lower bound above.
The corresponding problem for the lower intermediate dimensions remains open. The main obstruction is that these dimensions are not finitely stable. As a consequence, the gluing of the Cantor part and the countable part used in our construction does not control the lower intermediate dimensions of the union.
The paper is organized as follows. In Section 2 we recall the definitions and basic properties of the dimensions involved. In § 3 we recall the main facts about complementary sets, including the results of [24 ] that we use. § 4 contains our results. In Sub§ 4.1 we recall the results of Banaji and Rutar that we need. In Sub§ 4.2 we prove the lower bound and its rigidity. In Sub§ 4.3 and § 4.4 we develop the homogeneous Cantor schemes and the mountains. Finally, in Sub§ 4.5 we prove the main theorem and the characterization of F U ( a ) \mathcal{F}^U(a) F U ( a ) .
Preliminaries
Definitions
Given a non-empty set A ⊂ R A \subset \mathbb{R} A ⊂ R , we write ∣ A ∣ |A| ∣ A ∣ for the diameter of the set.
We write a n ≍ b n a_n \asymp b_n a n ≍ b n to mean that there exist constants c 1 , c 2 > 0 c_1, c_2 > 0 c 1 , c 2 > 0 , independent of n n n , such that c 1 b n ≤ a n ≤ c 2 b n c_1 b_n \leq a_n \leq c_2 b_n c 1 b n ≤ a n ≤ c 2 b n for all n n n .
The Hausdorff dimension of A A A will be denoted by dim H A \dim_H A dim H A , and \underline{\mbox{\rm dim}}_{\rm B}\,A and \overline{\mbox{\rm dim}}_{\rm B}\,A will denote the lower and upper box-counting dimensions of A A A , respectively. The reader can refer to [34 ] for more information on these dimensions.
This document concerns the θ \theta θ -intermediate dimensions, which were introduced in [35 ] and are defined as follows.
Definition 2.1. Let F ⊆ R F \subseteq \mathbb{R} F ⊆ R be bounded. For 0 ≤ θ ≤ 1 0 \leq \theta \leq 1 0 ≤ θ ≤ 1 , we define the lower θ \theta θ -intermediate dimensions of F F F by \begin{aligned}
\underline{\mbox{\rm dim}}_{\,\theta}F = \inf \big\{& s \geq 0 : \ \forall\ \varepsilon>0, \ \text{and all }\ \delta_0>0, \text{there}\ \exists\ 0<\delta\leq \delta_0, \ \text{and } \\
& \{U_i\}_{i \in I}\ : F \subseteq \cup_{i \in I} U_i \ : \delta^{1/\theta} \leq |U_i| \leq \delta \ \text{and }\ \sum_{i \in I} |U_i|^s \leq \varepsilon\big\}.
\end{aligned}
Similarly, we define the upper θ \theta θ -intermediate dimensions of F F F by \begin{aligned}
\overline{\mbox{\rm dim}}_{\,\theta}F = \inf \big\{& s \geq 0 : \ \forall\ \varepsilon>0, \ \text{there}\ \exists\ \delta_0>0, \ : \ \forall\ 0<\delta\leq \delta_0, \ \text{there}\ \exists \\
& \{U_i\}_{i \in I}\ : F \subseteq \cup_{i \in I} U_i \ : \delta^{1/\theta} \leq |U_i| \leq \delta \ \text{and }\ \sum_{i \in I} |U_i|^s \leq \varepsilon\big\}.
\end{aligned}
For all θ ∈ ( 0 , 1 ] \theta \in (0,1] θ ∈ ( 0 , 1 ] and a compact set A ⊂ R A \subset \mathbb{R} A ⊂ R , we have \dim_H A \leq \overline{\mbox{\rm dim}}_{\,\theta}A \leq \overline{\mbox{\rm dim}}_{\rm B}\,A and similarly with the lower θ \theta θ dimension, replacing \overline{\mbox{\rm dim}}_{\rm B}\, with \underline{\mbox{\rm dim}}_{\rm B}\, . This spectrum of dimensions has the property of being continuous for θ ∈ ( 0 , 1 ] \theta \in (0,1] θ ∈ ( 0 , 1 ] , and sometimes it is also continuous at θ = 0 \theta=0 θ = 0 . The reader can find more information in [35 ] .
We also need to recall the definitions of the Assouad dimension and the lower Assouad dimension. Given a non-empty set E ⊂ R E \subset \mathbb{R} E ⊂ R , N r ( E ) N_r(E) N r ( E ) will denote the minimum number of closed balls of radius r r r needed to cover E E E .
The Assouad dimension of E E E is defined as
dim A E = inf { α : there exist constants c , ρ > 0 such that sup x ∈ E N r ( B ( x , R ) ∩ E ) ≤ c ( R r ) α for all 0 < r < R < ρ } . \begin{align*}
\dim_A E= \inf\{ \alpha : & \text{ there exist constants } c,\rho>0 \text{ such that }\\
& \sup_{x \in E} N_r (B(x,R) \cap E) \leq c\left(\frac{R}{r}\right)^\alpha \text{ for all } 0<r<R<\rho \}.
\end{align*} dim A E = inf { α : there exist constants c , ρ > 0 such that x ∈ E sup N r ( B ( x , R ) ∩ E ) ≤ c ( r R ) α for all 0 < r < R < ρ } .
In the same way, the lower Assouad dimension of E E E is defined as
dim L E = sup { α : there exist constants c , ρ > 0 such that inf x ∈ E N r ( B ( x , R ) ∩ E ) ≥ c ( R r ) α for all 0 < r < R < ρ } . \begin{align*}
\dim_L E= \sup\{ \alpha : & \text{ there exist constants } c,\rho>0 \text{ such that }\\
& \inf_{x \in E} N_r (B(x,R) \cap E) \geq c\left(\frac{R}{r}\right)^\alpha \text{ for all } 0<r<R<\rho \}.
\end{align*} dim L E = sup { α : there exist constants c , ρ > 0 such that x ∈ E inf N r ( B ( x , R ) ∩ E ) ≥ c ( r R ) α for all 0 < r < R < ρ } .
The following inequalities hold for compact sets:
\dim_L E \leq \dim_H E \leq \underline{\mbox{\rm dim}}_{\,\theta}E\leq \overline{\mbox{\rm dim}}_{\,\theta}E\leq \overline{\mbox{\rm dim}}_{\rm B}\,E\leq \dim_A E.
A key inequality involving these dimensions, which we will use later, is given in the following result.
Proposition 2.3 (Corollary 2.8 [26 ] ).
If dim L E < dim A E \dim_L E<\dim_A E dim L E < dim A E , then for all θ ∈ ( 0 , 1 ] \theta \in (0,1] θ ∈ ( 0 , 1 ] , \overline{\mbox{\rm dim}}_{\,\theta}E \geq \frac{\theta\dim_A E (\overline{\mbox{\rm dim}}_{\rm B}\,E - \dim_L E) + \dim_L E (\dim_A E - \overline{\mbox{\rm dim}}_{\rm B}\,E)}{\theta(\overline{\mbox{\rm dim}}_{\rm B}\,E - \dim_L E) + (\dim_A E-\overline{\mbox{\rm dim}}_{\rm B}\,E)}. The same holds replacing \overline{\mbox{\rm dim}}_{\,\theta} and \overline{\mbox{\rm dim}}_{\rm B}\, with \underline{\mbox{\rm dim}}_{\,\theta} and \underline{\mbox{\rm dim}}_{\rm B}\, everywhere.
Complementary sets and their fractal dimensions
Let a = { a n } a=\{a_n\} a = { a n } be a positive, non-increasing, summable sequence with sum 1 1 1 . Let I a = [ 0 , 1 ] I_a=[0,1] I a = [ 0 , 1 ] . C a \mathcal{C}_a C a will denote the family of all closed sets E E E contained in I a I_a I a that are of the form E = I a ∖ ∪ j ≥ 1 U j E=I_a \setminus \displaystyle\cup_{j \geq 1} U_j E = I a ∖ ∪ j ≥ 1 U j , where { U j } \{U_j\} { U j } is a disjoint family of open intervals contained in I a I_a I a such that ∣ U j ∣ = a j |U_j|=a_j ∣ U j ∣ = a j for all j j j . The sets in C a \mathcal{C}_a C a are called the complementary sets of a a a . Note that if a = { a n } \mathbf{a}=\{\mathbf{a}_n\} a = { a n } is any rearrangement of the sequence { a n } \{a_n\} { a n } , then C a = C a \mathcal{C}_a=\mathcal{C}_{\mathbf{a}} C a = C a .
An important set in C a \mathcal{C}_a C a is the countable set, defined as follows:
D a = { ∑ i = k ∞ a i } k ∈ N . D_a=\left\{\sum_{i=k}^{\infty} a_i\right\}_{k\in\mathbb{N}}. D a = { i = k ∑ ∞ a i } k ∈ N .
As the reader notes, D a D_a D a is a countable, decreasing set of points with decreasing gaps given by the elements of the sequence a a a .
Another set of great importance in this family is the non-increasing Cantor set associated with the sequence a a a , which is constructed inductively in the following way. In the first step, remove from I a I_a I a an open interval of length a 1 a_1 a 1 , leaving two closed intervals I 1 , 1 I_{1,1} I 1 , 1 and I 1 , 2 I_{1,2} I 1 , 2 . Having performed the k k k -th step, we have 2 k 2^k 2 k closed intervals I k , 1 , . . . , I k , 2 k I_{k,1},...,I_{k,2^{k}} I k , 1 , ... , I k , 2 k contained in I a I_a I a . Then, from each interval I k , j I_{k,j} I k , j , remove an open interval of length a 2 k + j − 1 a_{2^k +j-1} a 2 k + j − 1 to obtain two closed intervals I k + 1 , 2 j − 1 I_{k+1,2j -1} I k + 1 , 2 j − 1 and I k + 1 , 2 j I_{k+1,2j} I k + 1 , 2 j . Finally, define the Cantor set as
C a : = ⋂ k = 1 ∞ ⋃ j = 1 2 k I k , j . C_a := \displaystyle\bigcap_{k=1}^{\infty}\bigcup_{j=1}^{2^{k}} I_{k,j}. C a := k = 1 ⋂ ∞ j = 1 ⋃ 2 k I k , j .
Given a non-increasing, summable sequence a = { a n } a=\{a_n\} a = { a n } , define
x n ( a ) = ∑ i = n ∞ a i and s n ( a ) = 2 − n x 2 n ( a ) . \displaystyle x_n(a)=\sum_{i=n}^{\infty}a_i
\quad \text{ and }\quad s_n(a)=2^{-n} x_{2^n}(a). x n ( a ) = i = n ∑ ∞ a i and s n ( a ) = 2 − n x 2 n ( a ) .
When the sequence a a a is fixed and clear from context, we will simply write s n s_n s n instead of s n ( a ) s_n (a) s n ( a ) .
Note that since a a a is non increasing we have the estimate
s n + 1 < ∣ I n , j ∣ < s n − 1 for all 1 ≤ j ≤ 2 n , s_{n+1} < |I_{n,j}| < s_{n-1}
\quad \text{for all $1 \leq j \leq 2^n$,} s n + 1 < ∣ I n , j ∣ < s n − 1 for all 1 ≤ j ≤ 2 n ,
where I n , j I_{n,j} I n , j are the closed intervals of step n n n in the construction of the Cantor set C a C_{a} C a .
The Hausdorff dimension of sets in C a \mathcal{C}_a C a was studied in [27 ] , [31 ] , and [38 ] . The following proposition holds:
Proposition 3.1. Let a = { a n } a = \{a_n\} a = { a n } be a positive summable sequence, and let C a C_a C a be the decreasing Cantor set associated with a a a . Then dim H C a = lim inf n → ∞ log n − log ( x n / n ) . \dim_H C_{a} = \liminf_{n \to \infty} \frac{\log n}{-\log (x_n/n)}. dim H C a = lim inf n → ∞ − l o g ( x n / n ) l o g n . Furthermore, any set E ∈ C a E \in \mathcal{C}_{a} E ∈ C a satisfies dim H E ≤ dim H C a \dim_H E \leq \dim_H C_{a} dim H E ≤ dim H C a . Conversely, given any s ∈ [ 0 , dim H C a ] s \in [0, \dim_H C_{a}] s ∈ [ 0 , dim H C a ] , there exists a set E ∈ C a E \in \mathcal{C}_a E ∈ C a such that dim H E = s \dim_H E = s dim H E = s .
Moreover, it was shown in [38 Proposition 3.3] that \dim_H C_a = \underline{\mbox{\rm dim}}_{\rm B}\,C_a .
The box dimensions of sets in C a \mathcal{C}_a C a were carefully studied in [33 Chapter 3.2] . There, Falconer shows that the upper and lower box dimensions of sets in C a \mathcal{C}_a C a depend only on the sequence a a a , and not on the rearrangement. That is, given any complementary set, its lower box dimension equals that of every other complementary set, and the same holds for the upper box dimension. Furthermore, for any E ∈ C a E \in \mathcal{C}_a E ∈ C a ,
\liminf_{n \to \infty} \frac{\log n}{-\log a_n} \leq \underline{\mbox{\rm dim}}_{\rm B}\,E \leq \overline{\mbox{\rm dim}}_{\rm B}\,E \leq \limsup_{n \to \infty} \frac{\log n}{-\log a_n},
and the box dimension exists if and only if these limits coincide.
The upper box dimension of C a C_a C a (and therefore of any other complementary set in C a \mathcal{C}_a C a ) is given by
lim sup n → ∞ log n − log ( x n / n ) , \limsup_{n \to \infty} \frac{\log n}{-\log (x_n/n)}, n → ∞ lim sup − log ( x n / n ) log n ,
which, as shown in [40 ] , is equal to
lim sup n → ∞ log n − log a n . \limsup_{n \to \infty} \frac{\log n}{-\log a_n}. n → ∞ lim sup − log a n log n .
This equivalence does not hold for the lower box dimension. In fact, the inequality
lim inf n → ∞ log n − log a n ≤ lim inf n → ∞ log n − log ( x n / n ) \liminf_{n \to \infty} \frac{\log n}{-\log a_n} \leq \liminf_{n \to \infty} \frac{\log n}{-\log (x_n/n)} n → ∞ lim inf − log a n log n ≤ n → ∞ lim inf − log ( x n / n ) log n
can be strict; see, for instance, [32 Proposition 5] . The Assouad dimension and the lower Assouad dimension of sets in C a \mathcal{C}_a C a were studied in [37 ] . There, the authors prove that, given a non-increasing summable sequence a = { a n } a = \{a_n\} a = { a n } ,
dim A C a = lim sup n → ∞ ( sup k n log 2 log ( s k / s k + n ) ) , \dim_A C_a = \limsup_{n \to \infty} \left( \sup_k \frac{n \log 2}{\log(s_k / s_{k+n})} \right), dim A C a = n → ∞ lim sup ( k sup log ( s k / s k + n ) n log 2 ) ,
and
dim L C a = lim inf n → ∞ ( inf k n log 2 log ( s k / s k + n ) ) . \dim_L C_a = \liminf_{n \to \infty} \left( \inf_k \frac{n \log 2}{\log(s_k / s_{k+n})} \right). dim L C a = n → ∞ lim inf ( k inf log ( s k / s k + n ) n log 2 ) .
Moreover, it was shown that dim A D a = 0 \dim_A D_a = 0 dim A D a = 0 or 1 1 1 , and if E E E is any complementary set in C a \mathcal{C}_a C a , then
dim A C a ≤ dim A E ≤ dim A D a , \dim_A C_a \leq \dim_A E \leq \dim_A D_a, dim A C a ≤ dim A E ≤ dim A D a ,
and
0 = dim L D a ≤ dim L E ≤ dim L C a . 0 = \dim_L D_a \leq \dim_L E \leq \dim_L C_a. 0 = dim L D a ≤ dim L E ≤ dim L C a .
Furthermore, if the sequence a a a is doubling, i.e., there exists a constant c > 0 c > 0 c > 0 such that a n ≤ c a 2 n a_n \leq c \, a_{2n} a n ≤ c a 2 n for all n n n , then
{ dim A E : E ∈ C a } = [ dim A C a , 1 ] , \{\dim_A E: E \in \mathcal{C}_a\} = [\dim_A C_a, 1], { dim A E : E ∈ C a } = [ dim A C a , 1 ] ,
and if instead of doubling, we assume that inf s n s n + 1 > 2 \inf \frac{s_n}{s_{n+1}} > 2 inf s n + 1 s n > 2 , then
{ dim L E : E ∈ C a } = [ 0 , dim L C a ] . \{\dim_L E: E \in \mathcal{C}_a\} = [0, \dim_L C_a]. { dim L E : E ∈ C a } = [ 0 , dim L C a ] .
For θ \theta θ -intermediate dimensions, the problem was studied recently by Angelini and Molter [24 ] . They computed the intermediate dimensions of D a D_a D a and, under mild assumptions on a a a , those of C a C_a C a [24 Theorems 4.2 and 4.3] . They also proved the following results for complementary sets. We state them only for the upper intermediate dimensions; the analogous statements for the lower intermediate dimensions also hold.
Lemma 3.2. [24 Lemma 5.2 and Lemma 5.3] Let a = { a n } a=\{a_n\} a = { a n } be a summable non-increasing sequence and let E ∈ C a E\in \mathcal{C}_a E ∈ C a be any complementary set. Then \overline{\mbox{\rm dim}}_{\,\theta}D_a \leq \overline{\mbox{\rm dim}}_{\,\theta}E. Moreover, if the sequence satisfies sup n s n ( a ) / s n + 1 ( a ) < ∞ \sup_n s_n(a)/s_{n+1}(a)<\infty sup n s n ( a ) / s n + 1 ( a ) < ∞ then \overline{\mbox{\rm dim}}_{\,\theta}E \leq \overline{\mbox{\rm dim}}_{\,\theta}C_a.
Theorem 3.3. [24 Theorem 6.1] Let θ ∈ ( 0 , 1 ] \theta\in(0,1] θ ∈ ( 0 , 1 ] and let a = { a n } a=\{a_n\} a = { a n } be a positive, non-increasing, summable sequence. If 2 < inf n s n ( a ) s n + 1 ( a ) ≤ sup n s n ( a ) s n + 1 ( a ) < ∞ , 2<\inf_n \frac{s_n(a)}{s_{n+1}(a)}\leq \sup_n \frac{s_n(a)}{s_{n+1}(a)}<\infty, 2 < inf n s n + 1 ( a ) s n ( a ) ≤ sup n s n + 1 ( a ) s n ( a ) < ∞ , then for every t\in[\overline{\mbox{\rm dim}}_{\,\theta}D_a , \overline{\mbox{\rm dim}}_{\,\theta}C_a], there exists a set E = E ( t , θ ) ∈ C a E=E(t,\theta)\in\mathcal{C}_a E = E ( t , θ ) ∈ C a such that \overline{\mbox{\rm dim}}_{\,\theta}E=t .
The lower bound in Lemma 3.2 is given explicitly by the following result.
Theorem 3.4. [24 Theorem 4.3] Let a = { a n } a=\{a_n\} a = { a n } be a non-increasing summable sequence. Then, for all θ ∈ [ 0 , 1 ] \theta\in[0,1] θ ∈ [ 0 , 1 ] , \overline{\mbox{\rm dim}}_{\,\theta}D_a=\frac{\theta\,\overline{\dim}_{\mathrm{B}} D_a}
{1-(1-\theta)\,\overline{\dim}_{\mathrm{B}} D_a}.
Since dim L D a = 0 \dim_{\mathrm{L}} D_a=0 dim L D a = 0 and dim A D a = 1 \dim_{\mathrm{A}} D_a=1 dim A D a = 1 , this is the smallest value allowed by Proposition 2.3 . Moreover, since the upper (and respectively lower) box dimension is the same for every set in C a \mathcal{C}_a C a , the lower bound depends only on dim ‾ B C a \overline{\dim}_{\mathrm{B}} C_a dim B C a .
The proof of Theorem 3.3 is constructive, and it gives more than the statement. Indeed, the set constructed there depends only on a parameter s > 1 s>1 s > 1 , and not on θ \theta θ .
Thus the functions h s h_s h s , s > 1 s>1 s > 1 , are realized as upper intermediate dimension functions of complementary sets. By Lemma 3.2 and Theorem 3.4 , every such function h h h satisfies
\frac{\theta B}{1-(1-\theta)B}\leq h(\theta)\leq \overline{\mbox{\rm dim}}_{\,\theta}C_a .
This led to the question, raised in [24 Question 1] , of describing the families
\mathcal{F}^U(a):=\{\theta\mapsto \overline{\mbox{\rm dim}}_{\,\theta}E: E\in\mathcal{C}_a\}
\quad\text{and}\quad
\mathcal{F}^L(a):=\{\theta\mapsto \underline{\dim}_{\theta} E:
E\in\mathcal{C}_a\}.
The main result of this paper answers this question for F U ( a ) \mathcal{F}^U(a) F U ( a ) under the hypothesis of Theorem 3.3 .
For a recent survey of complementary sets and their dimensions, including developments beyond the scope of the present section, we refer the reader to [22 ] .
Attainable dimension functions
In this section we prove the characterization of F U ( a ) \mathcal{F}^U(a) F U ( a ) announced above. In Sub§ 4.1 we recall the class of functions H ( λ , α ) \mathcal{H}(\lambda,\alpha) H ( λ , α ) and the results from [26 ] that we need. In Sub§ 4.2 we reformulate the defining condition of H ( 0 , 1 ) \mathcal{H}(0,1) H ( 0 , 1 ) as the monotonicity of a quotient. This yields a sharp lower bound for the functions in F U ( a ) \mathcal{F}^U(a) F U ( a ) and a rigidity property. In Sub§ 4.3 and § 4.4 we develop the two tools of the construction: the homogeneous Cantor schemes, whose intermediate dimensions depend only on their scales, and the mountains. The main result, Theorem 4.17 is proved in Sub§ 4.5 .
Throughout this section we write
B:=\overline{\mbox{\rm dim}}_{\rm B}\,C_a \qquad\text{and}\qquad \varphi_B(\theta):=\frac{\theta B}{1-(1-\theta)B}\quad \text{for } \theta\in[0,1].
Recall that \overline{\mbox{\rm dim}}_{\rm B}\,E=B for every E ∈ C a E\in\mathcal{C}_a E ∈ C a , moreover \overline{\mbox{\rm dim}}_{\,\theta}D_a=\varphi_B(\theta) by Theorem [24 Theorem 4.3] . As usual, we adopt the convention δ 1 / 0 = 0 \delta^{1/0}=0 δ 1/0 = 0 in Definition 2.1 , so that dim ‾ 0 E = dim H E \overline{\dim}_{\,0}E=\dim_{\rm H}E dim 0 E = dim H E for every bounded set E E E . We will say that the sequence a a a satisfies hypothesis (H) if
2 < κ 1 : = inf n s n ( a ) s n + 1 ( a ) ≤ κ 2 : = sup n s n ( a ) s n + 1 ( a ) < ∞ . \begin{equation}
\tag{H}
2<\kappa_1:=\inf_n \frac{s_n(a)}{s_{n+1}(a)}\leq \kappa_2:=\sup_n \frac{s_n(a)}{s_{n+1}(a)}<\infty .
\end{equation} 2 < κ 1 := n inf s n + 1 ( a ) s n ( a ) ≤ κ 2 := n sup s n + 1 ( a ) s n ( a ) < ∞. ( H )
Classes of functions and Banaji–Rutar results
Given a function f f f defined on an interval, we write D + f D^+f D + f , D + f D_+f D + f , D − f D^-f D − f and D − f D_-f D − f for its upper right, lower right, upper left and lower left Dini derivatives, respectively. For instance,
D + f ( x ) = lim sup ε → 0 + f ( x + ε ) − f ( x ) ε D^+f(x)=\limsup_{\varepsilon\to 0^+}\frac{f(x+\varepsilon)-f(x)}{\varepsilon} D + f ( x ) = ε → 0 + lim sup ε f ( x + ε ) − f ( x )
and
D − f ( x ) = lim inf ε → 0 + f ( x ) − f ( x − ε ) ε , D_-f(x)=\liminf_{\varepsilon\to 0^+}\frac{f(x)-f(x-\varepsilon)}{\varepsilon}, D − f ( x ) = ε → 0 + lim inf ε f ( x ) − f ( x − ε ) ,
whenever the corresponding one sided increments are defined.
Definition 4.1 ([26 Definition 1.3] ). Let 0 ≤ λ ≤ α ≤ 1 0\leq\lambda\leq\alpha\leq 1 0 ≤ λ ≤ α ≤ 1 . If λ < α \lambda<\alpha λ < α , we denote by H ( λ , α ) \mathcal{H}(\lambda,\alpha) H ( λ , α ) the set of all functions h : [ 0 , 1 ] → [ λ , α ] h:[0,1]\to[\lambda,\alpha] h : [ 0 , 1 ] → [ λ , α ] which are non-decreasing, continuous on ( 0 , 1 ] (0,1] ( 0 , 1 ] , and satisfy D + h ( θ ) ≤ ( h ( θ ) − λ ) ( α − h ( θ ) ) ( α − λ ) θ for all θ ∈ ( 0 , 1 ) .
D^+h(\theta)\leq \frac{(h(\theta)-\lambda)(\alpha-h(\theta))}{(\alpha-\lambda)\theta}\quad\text{for all } \theta\in(0,1).
D + h ( θ ) ≤ ( α − λ ) θ ( h ( θ ) − λ ) ( α − h ( θ )) for all θ ∈ ( 0 , 1 ) . If λ = α \lambda=\alpha λ = α , we let H ( λ , α ) \mathcal{H}(\lambda,\alpha) H ( λ , α ) consist only of the constant function h ≡ λ h\equiv\lambda h ≡ λ .
The next theorem due to Banaji and Rutar provide a necessary and sufficient condition for a given function to be realized as the intermediate dimensions of a bounded subset F ⊂ R d F\subset\mathbb{R}^d F ⊂ R d . We write the case d = 1 d=1 d = 1 .
Theorem 4.2 ([26 Theorem B] ). Let F ⊂ R F\subset \mathbb{R} F ⊂ R be a bounded set with dim L F = λ \dim_L F=\lambda dim L F = λ and dim A F = α \dim_A F=\alpha dim A F = α . Then with h_1(\theta)=\underline{\mbox{\rm dim}}_{\,\theta}F and h_2(\theta)=\overline{\mbox{\rm dim}}_{\,\theta}F we have h 1 , h 2 ∈ H ( λ , α ) h_1,h_2\in \mathcal{H}(\lambda,\alpha) h 1 , h 2 ∈ H ( λ , α ) , h 1 ≤ h 2 h_1\leq h_2 h 1 ≤ h 2 and h 1 ( 0 ) = h 2 ( 0 ) h_1(0)=h_2(0) h 1 ( 0 ) = h 2 ( 0 ) .
Conversely, if 0 ≤ λ ≤ α ≤ 1 0\leq \lambda\leq \alpha\leq 1 0 ≤ λ ≤ α ≤ 1 and h 1 , h 2 ∈ H ( λ , α ) h_1,h_2\in \mathcal{H}(\lambda,\alpha) h 1 , h 2 ∈ H ( λ , α ) satisfy h 1 ≤ h 2 h_1\leq h_2 h 1 ≤ h 2 and h 1 ( 0 ) = h 2 ( 0 ) h_1(0)=h_2(0) h 1 ( 0 ) = h 2 ( 0 ) , then there exists a compact perfect set F ⊂ R F\subset \mathbb{R} F ⊂ R such that dim L F = λ \dim_L F=\lambda dim L F = λ , dim A F = α \dim_A F=\alpha dim A F = α , \underline{\mbox{\rm dim}}_{\,\theta}F=h_1(\theta) and \overline{\mbox{\rm dim}}_{\,\theta}F=h_2(\theta) for all θ ∈ [ 0 , 1 ] \theta\in[0,1] θ ∈ [ 0 , 1 ] .
We also need some tools from [26 Section 3] , which we state only for d = 1 d=1 d = 1 , λ = 0 \lambda=0 λ = 0 and α = 1 \alpha=1 α = 1 . We denote by G ( 0 , 1 ) \mathcal{G}(0,1) G ( 0 , 1 ) the set of continuous functions g : R → [ 0 , 1 ] g:\mathbb{R}\to[0,1] g : R → [ 0 , 1 ] such that
D + g ( x ) ∈ [ − g ( x ) , 1 − g ( x ) ] for all x ∈ R , D^+g(x)\in[-g(x),\,1-g(x)]\quad\text{for all } x\in\mathbb{R}, D + g ( x ) ∈ [ − g ( x ) , 1 − g ( x )] for all x ∈ R ,
see [26 Definition 3.1] . Given g ∈ G ( 0 , 1 ) g\in\mathcal{G}(0,1) g ∈ G ( 0 , 1 ) and w ∈ R w\in\mathbb{R} w ∈ R , the offset κ w ( g ) ∈ G ( 0 , 1 ) \kappa_w(g)\in\mathcal{G}(0,1) κ w ( g ) ∈ G ( 0 , 1 ) is defined by κ w ( g ) ( x ) = g ( x − w ) \kappa_w(g)(x)=g(x-w) κ w ( g ) ( x ) = g ( x − w ) if x ≥ w x\geq w x ≥ w and κ w ( g ) ( x ) = g ( 0 ) \kappa_w(g)(x)=g(0) κ w ( g ) ( x ) = g ( 0 ) if x ≤ w x\leq w x ≤ w . A function g ∈ G ( 0 , 1 ) g\in\mathcal{G}(0,1) g ∈ G ( 0 , 1 ) is rapidly decreasing if there exist y ∈ R y\in\mathbb{R} y ∈ R and C > 0 C>0 C > 0 such that g ( x ) ≤ g ( y ) exp ( y − x ) + C exp ( − x ) g(x)\leq g(y)\exp(y-x)+C\exp(-x) g ( x ) ≤ g ( y ) exp ( y − x ) + C exp ( − x ) for all x ≥ y x\geq y x ≥ y . Note that a rapidly decreasing function satisfies lim x → ∞ g ( x ) = 0 \lim_{x\to\infty}g(x)=0 lim x → ∞ g ( x ) = 0 .
Let ρ = ( ρ k ) k ≥ 0 \rho=(\rho_k)_{k\geq0} ρ = ( ρ k ) k ≥ 0 be a strictly decreasing sequence of positive numbers converging to 0 0 0 . For ϕ ∈ ( 0 , ρ 0 ) \phi\in(0,\rho_0) ϕ ∈ ( 0 , ρ 0 ) let k = k ρ ( ϕ ) ≥ 1 k=k_\rho(\phi)\geq1 k = k ρ ( ϕ ) ≥ 1 be the unique integer such that ρ k ≤ ϕ < ρ k − 1 \rho_k\leq \phi<\rho_{k-1} ρ k ≤ ϕ < ρ k − 1 , and set
s ρ ( ϕ ) : = k ρ ( ϕ ) log 2 − log ϕ . \begin{equation}
s_\rho(\phi):=\frac{k_\rho(\phi)\log 2}{-\log\phi}. \tag*{(2)}
\end{equation} s ρ ( ϕ ) := − log ϕ k ρ ( ϕ ) log 2 . (2)
Observe that k ρ ( ϕ ) = min { k ≥ 0 : ρ k ≤ ϕ } k_\rho(\phi)=\min\{k\geq0 : \rho_k\leq\phi\} k ρ ( ϕ ) = min { k ≥ 0 : ρ k ≤ ϕ } . If r = ( r j ) j ≥ 1 ⊂ ( 0 , 1 / 2 ] \mathbf{r}=(r_j)_{j\geq1}\subset(0,1/2] r = ( r j ) j ≥ 1 ⊂ ( 0 , 1/2 ] , we write s r : = s ρ s_{\mathbf r}:=s_\rho s r := s ρ where ρ 0 = 1 \rho_0=1 ρ 0 = 1 and ρ k = r 1 ⋯ r k \rho_k=r_1\cdots r_k ρ k = r 1 ⋯ r k . This is exactly the function s r ( δ ) s_{\mathbf r}(\delta) s r ( δ ) of [26 Section 2.3] .
Lemma 4.4 ([26 Lemma 3.4] ). Let g ~ ∈ G ( 0 , 1 ) \tilde g\in\mathcal{G}(0,1) g ~ ∈ G ( 0 , 1 ) be a function which is not rapidly decreasing and such that 0 < g ~ ( 0 ) < 1 0<\tilde g(0)<1 0 < g ~ ( 0 ) < 1 . Then there is a constant w 0 ∈ R w_0\in\mathbb{R} w 0 ∈ R , depending only on g ~ ( 0 ) \tilde g(0) g ~ ( 0 ) , such that for every w ≥ w 0 w\geq w_0 w ≥ w 0 there exists a sequence r = ( r j ) j ≥ 1 ⊂ ( 0 , 1 / 2 ] \mathbf r=(r_j)_{j\geq1}\subset(0,1/2] r = ( r j ) j ≥ 1 ⊂ ( 0 , 1/2 ] for which g : = κ w ( g ~ ) g:=\kappa_w(\tilde g) g := κ w ( g ~ ) satisfies ∣ s r ( exp ( − exp ( x ) ) ) − g ( x ) ∣ ≤ log 2 exp ( − x ) for all x ≥ w 0 . \bigl|s_{\mathbf r}(\exp(-\exp(x)))-g(x)\bigr|\leq \log 2\,\exp(-x)\quad\text{for all } x\geq w_0 . s r ( exp ( − exp ( x ))) − g ( x ) ≤ log 2 exp ( − x ) for all x ≥ w 0 .
Finally, we recall two elementary facts on Dini derivatives.
Lemma 4.5 ([26 Lemma 2.2] ). Let f f f and g g g be continuous functions on [ a , b ] [a,b] [ a , b ] with D + g ≤ D + f D_+g\leq D^+f D + g ≤ D + f on [ a , b ) [a,b) [ a , b ) and g ( a ) = f ( a ) g(a)=f(a) g ( a ) = f ( a ) . Then g ≤ f g\leq f g ≤ f on [ a , b ] [a,b] [ a , b ] .
Lemma 4.6 ([26 Lemma 2.4] ). Let g : R → ( 0 , 1 ) g:\mathbb{R}\to(0,1) g : R → ( 0 , 1 ) be continuous and let V ⊂ R V\subset\mathbb{R} V ⊂ R be open. Then D + g ( x ) ∈ [ − g ( x ) , 1 − g ( x ) ] D^+g(x)\in[-g(x),1-g(x)] D + g ( x ) ∈ [ − g ( x ) , 1 − g ( x )] for all x ∈ V x\in V x ∈ V if and only if D − g ( x ) ∈ [ − g ( x ) , 1 − g ( x ) ] D_-g(x)\in[-g(x),1-g(x)] D − g ( x ) ∈ [ − g ( x ) , 1 − g ( x )] for all x ∈ V x\in V x ∈ V .
An integrated form of the Dini condition
The following elementary observation is the key tool of this section. Condition (1) is a pointwise inequality for a Dini derivative, which is awkward to manipulate directly. The following lemma replaces it, for λ = 0 \lambda=0 λ = 0 and α = 1 \alpha=1 α = 1 , by an equivalent global statement: the quotient Q h ( θ ) = h ( θ ) / ( θ ( 1 − h ( θ ) ) ) Q_h(\theta)=h(\theta)/(\theta(1-h(\theta))) Q h ( θ ) = h ( θ ) / ( θ ( 1 − h ( θ ))) is non-increasing. This simple reformulation is used repeatedly in the rest of the section. First, comparing Q h Q_h Q h with the constant quotient of φ B \varphi_B φ B yields the lower bound h ≥ φ B h\ge\varphi_B h ≥ φ B and its rigidity (Proposition 4.8 ). In the proof of the main Theorem 4.17 , this bound is what allows the countable part of the construction to be added without increasing the dimension. Second, the monotonicity of Q h Q_h Q h is exactly what makes the mountains of Lemma 4.15 well defined, and it ensures that their slopes are admissible. These mountains are the building blocks of the set realizing a prescribed function in Theorem 4.17 .
Lemma 4.7. Let h ∈ H ( 0 , 1 ) h\in\mathcal{H}(0,1) h ∈ H ( 0 , 1 ) with 0 < h ( 1 ) < 1 0<h(1)<1 0 < h ( 1 ) < 1 . Then 0 < h ( θ ) < 1 0<h(\theta)<1 0 < h ( θ ) < 1 for every θ ∈ ( 0 , 1 ] \theta\in(0,1] θ ∈ ( 0 , 1 ] , and the function Q h ( θ ) : = h ( θ ) θ ( 1 − h ( θ ) ) Q_h(\theta):=\frac{h(\theta)}{\theta\,(1-h(\theta))} Q h ( θ ) := θ ( 1 − h ( θ )) h ( θ ) is non-increasing on ( 0 , 1 ] (0,1] ( 0 , 1 ] .
Proof. Since h h h is non-decreasing, h ( θ ) ≤ h ( 1 ) < 1 h(\theta)\leq h(1)<1 h ( θ ) ≤ h ( 1 ) < 1 for all θ \theta θ . Let 0 < c < 1 0<c<1 0 < c < 1 be such that h > 0 h>0 h > 0 on [ c , 1 ] [c,1] [ c , 1 ] , and define on [ c , 1 ] [c,1] [ c , 1 ] the continuous function
P ( θ ) : = log h ( θ ) 1 − h ( θ ) − log θ . P(\theta):=\log\frac{h(\theta)}{1-h(\theta)}-\log\theta . P ( θ ) := log 1 − h ( θ ) h ( θ ) − log θ . The map v ↦ log ( v / ( 1 − v ) ) v\mapsto \log (v/(1-v)) v ↦ log ( v / ( 1 − v )) is continuously differentiable on ( 0 , 1 ) (0,1) ( 0 , 1 ) with derivative 1 / ( v ( 1 − v ) ) > 0 1/(v(1-v))>0 1/ ( v ( 1 − v )) > 0 . Since h h h is continuous at each θ ∈ [ c , 1 ) \theta\in[c,1) θ ∈ [ c , 1 ) and 0 ≤ D + h ( θ ) ≤ h ( θ ) ( 1 − h ( θ ) ) / θ < ∞ 0\leq D^+h(\theta)\leq h(\theta)(1-h(\theta))/\theta<\infty 0 ≤ D + h ( θ ) ≤ h ( θ ) ( 1 − h ( θ )) / θ < ∞ , the mean value theorem gives
D + P ( θ ) = D + h ( θ ) h ( θ ) ( 1 − h ( θ ) ) − 1 θ ≤ 0 for all θ ∈ [ c , 1 ) . D^+P(\theta)=\frac{D^+h(\theta)}{h(\theta)(1-h(\theta))}-\frac{1}{\theta}\leq 0\qquad\text{for all }\theta\in[c,1). D + P ( θ ) = h ( θ ) ( 1 − h ( θ )) D + h ( θ ) − θ 1 ≤ 0 for all θ ∈ [ c , 1 ) . Applying Lemma 4.5 to g = P g=P g = P and to the constant function f ≡ P ( c ′ ) f\equiv P(c') f ≡ P ( c ′ ) on [ c ′ , 1 ] [c',1] [ c ′ , 1 ] , for any c ′ ∈ [ c , 1 ) c'\in[c,1) c ′ ∈ [ c , 1 ) , we obtain P ≤ P ( c ′ ) P\leq P(c') P ≤ P ( c ′ ) on [ c ′ , 1 ] [c',1] [ c ′ , 1 ] . Hence P P P is non-increasing on [ c , 1 ] [c,1] [ c , 1 ] .
Suppose now that h ( θ 0 ) = 0 h(\theta_0)=0 h ( θ 0 ) = 0 for some θ 0 ∈ ( 0 , 1 ] \theta_0\in(0,1] θ 0 ∈ ( 0 , 1 ] . Then θ 0 < 1 \theta_0<1 θ 0 < 1 , and since h h h is continuous on ( 0 , 1 ] (0,1] ( 0 , 1 ] , the set { θ ∈ [ θ 0 , 1 ] : h ( θ ) = 0 } \{\theta\in[\theta_0,1] : h(\theta)=0\} { θ ∈ [ θ 0 , 1 ] : h ( θ ) = 0 } is compact and does not contain 1 1 1 . Let θ 1 < 1 \theta_1<1 θ 1 < 1 be its maximum. Then h > 0 h>0 h > 0 on ( θ 1 , 1 ] (\theta_1,1] ( θ 1 , 1 ] , so by the previous paragraph P ( θ ) ≥ P ( 1 ) > − ∞ P(\theta)\geq P(1)>-\infty P ( θ ) ≥ P ( 1 ) > − ∞ for every θ ∈ ( θ 1 , 1 ] \theta\in(\theta_1,1] θ ∈ ( θ 1 , 1 ] . On the other hand h ( θ ) → h ( θ 1 ) = 0 h(\theta)\to h(\theta_1)=0 h ( θ ) → h ( θ 1 ) = 0 as θ → θ 1 + \theta\to\theta_1^+ θ → θ 1 + , which gives P ( θ ) → − ∞ P(\theta)\to-\infty P ( θ ) → − ∞ , a contradiction. Therefore h > 0 h>0 h > 0 on ( 0 , 1 ] (0,1] ( 0 , 1 ] , the function P P P is non-increasing on ( 0 , 1 ] (0,1] ( 0 , 1 ] , and so is Q h = exp ( P ) Q_h=\exp(P) Q h = exp ( P ) . ◻
Proposition 4.8. Let h ∈ H ( 0 , 1 ) h\in\mathcal{H}(0,1) h ∈ H ( 0 , 1 ) with h ( 1 ) = B ∈ ( 0 , 1 ) h(1)=B\in(0,1) h ( 1 ) = B ∈ ( 0 , 1 ) . Then the following hold.
h ( θ ) ≥ φ B ( θ ) h(\theta)\geq\varphi_B(\theta) h ( θ ) ≥ φ B ( θ ) for all θ ∈ [ 0 , 1 ] \theta\in[0,1] θ ∈ [ 0 , 1 ] .
If h ( θ 0 ) = φ B ( θ 0 ) h(\theta_0)=\varphi_B(\theta_0) h ( θ 0 ) = φ B ( θ 0 ) for some θ 0 ∈ ( 0 , 1 ] \theta_0\in(0,1] θ 0 ∈ ( 0 , 1 ] , then h ( θ ) = φ B ( θ ) h(\theta)=\varphi_B(\theta) h ( θ ) = φ B ( θ ) for all θ ∈ [ θ 0 , 1 ] \theta\in[\theta_0,1] θ ∈ [ θ 0 , 1 ] .
Proof. A direct computation shows that φ B ( θ ) / ( 1 − φ B ( θ ) ) = θ B / ( 1 − B ) \varphi_B(\theta)/(1-\varphi_B(\theta))=\theta B/(1-B) φ B ( θ ) / ( 1 − φ B ( θ )) = θ B / ( 1 − B ) , that is, Q φ B ( θ ) = B / ( 1 − B ) Q_{\varphi_B}(\theta)=B/(1-B) Q φ B ( θ ) = B / ( 1 − B ) for all θ ∈ ( 0 , 1 ] \theta\in(0,1] θ ∈ ( 0 , 1 ] . By Lemma 4.7 ,
Q h ( θ ) ≥ Q h ( 1 ) = B 1 − B = Q φ B ( θ ) for all θ ∈ ( 0 , 1 ] . Q_h(\theta)\geq Q_h(1)=\frac{B}{1-B}=Q_{\varphi_B}(\theta)\qquad\text{for all }\theta\in(0,1]. Q h ( θ ) ≥ Q h ( 1 ) = 1 − B B = Q φ B ( θ ) for all θ ∈ ( 0 , 1 ] . Since v ↦ v / ( 1 − v ) v\mapsto v/(1-v) v ↦ v / ( 1 − v ) is strictly increasing on [ 0 , 1 ) [0,1) [ 0 , 1 ) , the inequality Q h ( θ ) ≥ Q φ B ( θ ) Q_h(\theta)\geq Q_{\varphi_B}(\theta) Q h ( θ ) ≥ Q φ B ( θ ) is equivalent to h ( θ ) ≥ φ B ( θ ) h(\theta)\geq\varphi_B(\theta) h ( θ ) ≥ φ B ( θ ) . For θ = 0 \theta=0 θ = 0 we simply have h ( 0 ) ≥ 0 = φ B ( 0 ) h(0)\geq0=\varphi_B(0) h ( 0 ) ≥ 0 = φ B ( 0 ) . This proves (i). If h ( θ 0 ) = φ B ( θ 0 ) h(\theta_0)=\varphi_B(\theta_0) h ( θ 0 ) = φ B ( θ 0 ) , then Q h ( θ 0 ) = Q h ( 1 ) Q_h(\theta_0)=Q_h(1) Q h ( θ 0 ) = Q h ( 1 ) , and since Q h Q_h Q h is non-increasing, Q h Q_h Q h is constant on [ θ 0 , 1 ] [\theta_0,1] [ θ 0 , 1 ] . Thus Q h = Q φ B Q_h=Q_{\varphi_B} Q h = Q φ B on [ θ 0 , 1 ] [\theta_0,1] [ θ 0 , 1 ] , which means h = φ B h=\varphi_B h = φ B on [ θ 0 , 1 ] [\theta_0,1] [ θ 0 , 1 ] , proving (ii). ◻
The equality case in Proposition 4.8 also forces the Assouad and lower dimensions to be extremal.
Proposition 4.10. Let F ⊂ R F\subset\mathbb{R} F ⊂ R be a bounded set with \overline{\mbox{\rm dim}}_{\rm B}\,F=B\in(0,1) . If dim ‾ θ 0 F = φ B ( θ 0 ) \overline{\dim}_{\,\theta_0}F=\varphi_B(\theta_0) dim θ 0 F = φ B ( θ 0 ) for some θ 0 ∈ ( 0 , 1 ) \theta_0\in(0,1) θ 0 ∈ ( 0 , 1 ) , then dim L F = 0 \dim_L F=0 dim L F = 0 and dim A F = 1 \dim_A F=1 dim A F = 1 .
Proof. Write λ = dim L F \lambda=\dim_L F λ = dim L F and α = dim A F \alpha=\dim_A F α = dim A F , so that λ ≤ B ≤ α \lambda\leq B\leq\alpha λ ≤ B ≤ α . If λ = α \lambda=\alpha λ = α , then \theta\mapsto\overline{\mbox{\rm dim}}_{\,\theta}F is constant by Theorem 4.2 , hence equal to B B B , and B > φ B ( θ 0 ) B>\varphi_B(\theta_0) B > φ B ( θ 0 ) because 0 < B < 1 0<B<1 0 < B < 1 and θ 0 < 1 \theta_0<1 θ 0 < 1 . Thus λ < α \lambda<\alpha λ < α , and Proposition 2.3 gives φ B ( θ 0 ) = dim ‾ θ 0 F ≥ ψ ( θ 0 ) \varphi_B(\theta_0)=\overline{\dim}_{\,\theta_0}F\geq\psi(\theta_0) φ B ( θ 0 ) = dim θ 0 F ≥ ψ ( θ 0 ) , where
ψ ( θ ) : = θ α ( B − λ ) + λ ( α − B ) θ ( B − λ ) + ( α − B ) . \psi(\theta):=\frac{\theta\alpha(B-\lambda)+\lambda(\alpha-B)}{\theta(B-\lambda)+(\alpha-B)}. ψ ( θ ) := θ ( B − λ ) + ( α − B ) θ α ( B − λ ) + λ ( α − B ) . Set u : = B − λ ∈ [ 0 , B ] u:=B-\lambda\in[0,B] u := B − λ ∈ [ 0 , B ] and v : = α − B ∈ [ 0 , 1 − B ] v:=\alpha-B\in[0,1-B] v := α − B ∈ [ 0 , 1 − B ] . A direct computation gives
ψ ( θ ) = B − ( 1 − θ ) u v v + θ u and φ B ( θ ) = B − ( 1 − θ ) B ( 1 − B ) ( 1 − B ) + θ B . \psi(\theta)=B-\frac{(1-\theta)\,uv}{v+\theta u}
\qquad\text{and}\qquad
\varphi_B(\theta)=B-\frac{(1-\theta)\,B(1-B)}{(1-B)+\theta B}. ψ ( θ ) = B − v + θ u ( 1 − θ ) uv and φ B ( θ ) = B − ( 1 − B ) + θ B ( 1 − θ ) B ( 1 − B ) . If u v = 0 uv=0 uv = 0 , then ψ ( θ 0 ) = B > φ B ( θ 0 ) \psi(\theta_0)=B>\varphi_B(\theta_0) ψ ( θ 0 ) = B > φ B ( θ 0 ) , which is impossible. If u v > 0 uv>0 uv > 0 , then u v / ( v + θ 0 u ) = ( u − 1 + θ 0 v − 1 ) − 1 uv/(v+\theta_0 u)=(u^{-1}+\theta_0v^{-1})^{-1} uv / ( v + θ 0 u ) = ( u − 1 + θ 0 v − 1 ) − 1 is strictly increasing in u u u and in v v v , so ψ ( θ 0 ) ≥ φ B ( θ 0 ) \psi(\theta_0)\geq\varphi_B(\theta_0) ψ ( θ 0 ) ≥ φ B ( θ 0 ) with equality if and only if u = B u=B u = B and v = 1 − B v=1-B v = 1 − B . Hence λ = 0 \lambda=0 λ = 0 and α = 1 \alpha=1 α = 1 . ◻
Homogeneous Cantor schemes
The set realizing a prescribed function in Theorem 4.17 cannot, in general, be a decreasing Cantor set (see Remark 4.20 ). We therefore work with a more flexible class of sets. These are limit sets of binary Cantor schemes in which the intervals of a given level have comparable lengths, which need not be equal. Their gaps also need not be arranged in decreasing order. The main result of this subsection, Lemma 4.12 , shows that for such sets the upper intermediate dimensions depend only on the sequence of scales ρ \rho ρ . They are given by window infima of the function S ρ S_\rho S ρ defined below. This reduces the construction of Theorem 4.17 to two tasks: choosing a suitable sequence of scales, and placing terms of a a a as gaps so that the resulting scheme is homogeneous with those scales.
Let Σ ∗ = ⋃ k ≥ 0 { 1 , 2 } k \Sigma^*=\bigcup_{k\geq0}\{1,2\}^k Σ ∗ = ⋃ k ≥ 0 { 1 , 2 } k be the set of finite words over the alphabet { 1 , 2 } \{1,2\} { 1 , 2 } , including the empty word ∅ \varnothing ∅ , and write ∣ σ ∣ |\sigma| ∣ σ ∣ for the length of σ \sigma σ . We write τ ⪰ σ \tau\succeq\sigma τ ⪰ σ if σ \sigma σ is a prefix of τ \tau τ (including τ = σ \tau=\sigma τ = σ ).
Definition 4.11. A family { I σ } σ ∈ Σ ∗ \{I_\sigma\}_{\sigma\in\Sigma^*} { I σ } σ ∈ Σ ∗ of non-degenerate closed intervals is a binary Cantor scheme if for every σ ∈ Σ ∗ \sigma\in\Sigma^* σ ∈ Σ ∗ the intervals I σ 1 I_{\sigma1} I σ 1 and I σ 2 I_{\sigma2} I σ 2 are disjoint and contained in I σ I_\sigma I σ , and max ∣ σ ∣ = k ∣ I σ ∣ → 0 \max_{|\sigma|=k}|I_\sigma|\to0 max ∣ σ ∣ = k ∣ I σ ∣ → 0 as k → ∞ k\to\infty k → ∞ . Its limit set is F = ⋂ k ≥ 0 ⋃ ∣ σ ∣ = k I σ . F=\bigcap_{k\geq0}\ \bigcup_{|\sigma|=k}I_\sigma . F = ⋂ k ≥ 0 ⋃ ∣ σ ∣ = k I σ . Let ρ = ( ρ k ) k ≥ 0 \rho=(\rho_k)_{k\geq0} ρ = ( ρ k ) k ≥ 0 be a strictly decreasing sequence converging to 0 0 0 with ρ 0 ≤ 1 \rho_0\leq1 ρ 0 ≤ 1 , and let K ≥ 1 K\geq1 K ≥ 1 . The scheme is ( ρ , K ) (\rho,K) ( ρ , K ) -homogeneous if K − 1 ρ k ≤ ∣ I σ ∣ ≤ K ρ k K^{-1}\rho_k\leq|I_\sigma|\leq K\rho_k K − 1 ρ k ≤ ∣ I σ ∣ ≤ K ρ k for all σ ∈ Σ ∗ \sigma\in\Sigma^* σ ∈ Σ ∗ with ∣ σ ∣ = k |\sigma|=k ∣ σ ∣ = k and all k ≥ 0 k\geq0 k ≥ 0 .
For instance, the closed intervals I k , j I_{k,j} I k , j used in § 3 to construct C a C_a C a form, after the obvious relabelling by words, a binary Cantor scheme with limit set C a C_a C a , and s k + 1 ( a ) ≤ ∣ I k , j ∣ ≤ s k − 1 ( a ) s_{k+1}(a)\leq|I_{k,j}|\leq s_{k-1}(a) s k + 1 ( a ) ≤ ∣ I k , j ∣ ≤ s k − 1 ( a ) . Therefore, if a a a satisfies (H) and C ≥ 1 C\geq1 C ≥ 1 , this scheme is ( ρ , K ) (\rho,K) ( ρ , K ) -homogeneous with ρ k = C − 1 s k ( a ) \rho_k=C^{-1}s_k(a) ρ k = C − 1 s k ( a ) and K = C κ 2 K=C\kappa_2 K = C κ 2 .
For a strictly decreasing sequence ρ \rho ρ as in Definition 4.11 we will also use the function
S ρ ( x ) : = s ρ ( exp ( − exp ( x ) ) ) , x > log log ( 1 / ρ 0 ) . S_\rho(x):=s_\rho\bigl(\exp(-\exp(x))\bigr),\qquad x>\log\log(1/\rho_0). S ρ ( x ) := s ρ ( exp ( − exp ( x )) ) , x > log log ( 1/ ρ 0 ) .
Since ϕ = exp ( − exp ( y ) ) \phi=\exp(-\exp(y)) ϕ = exp ( − exp ( y )) is a decreasing bijection, ϕ ∈ [ δ 1 / θ , δ ] \phi\in[\delta^{1/\theta},\delta] ϕ ∈ [ δ 1/ θ , δ ] if and only if y ∈ [ x , x + log ( 1 / θ ) ] y\in[x,x+\log(1/\theta)] y ∈ [ x , x + log ( 1/ θ )] , where x = log log ( 1 / δ ) x=\log\log(1/\delta) x = log log ( 1/ δ ) . The next lemma is essentially [26 Lemma 2.10 and Proposition 2.11] , where it is proved for homogeneous Moran sets. We include the proof since in our setting the intervals of a fixed level need not have the same length.
Lemma 4.12. Let F F F be the limit set of a ( ρ , K ) (\rho,K) ( ρ , K ) -homogeneous binary Cantor scheme. Then for all θ ∈ ( 0 , 1 ] \theta\in(0,1] θ ∈ ( 0 , 1 ] ,
\overline{\mbox{\rm dim}}_{\,\theta}F=\limsup_{\delta\to0}\ \inf_{\phi\in[\delta^{1/\theta},\delta]}s_\rho(\phi)
=\limsup_{x\to\infty}\ \inf_{y\in[x,x+\log(1/\theta)]}S_\rho(y),
and
\dim_{\rm H}F=\underline{\mbox{\rm dim}}_{\rm B}\,F=\liminf_{\phi\to0}s_\rho(\phi)=\liminf_{x\to\infty}S_\rho(x).
Proof. The second equalities in (3) and (4) follow from the change of variables described above, so we only prove the first ones. Let μ \mu μ be the image of the uniform Bernoulli measure on { 1 , 2 } N \{1,2\}^{\mathbb{N}} { 1 , 2 } N under the coding map which sends ω \omega ω to the unique point of ⋂ k I ω ∣ k \bigcap_k I_{\omega|k} ⋂ k I ω ∣ k . Since the intervals of each level are pairwise disjoint, μ ( I σ ) = 2 − ∣ σ ∣ \mu(I_\sigma)=2^{-|\sigma|} μ ( I σ ) = 2 − ∣ σ ∣ for every σ \sigma σ , and μ ( F ) = 1 \mu(F)=1 μ ( F ) = 1 .
We first prove two estimates. Let U ⊂ R U\subset\mathbb{R} U ⊂ R with 0 < ∣ U ∣ < ρ 0 0<|U|<\rho_0 0 < ∣ U ∣ < ρ 0 and let k = k ρ ( ∣ U ∣ ) ≥ 1 k=k_\rho(|U|)\geq1 k = k ρ ( ∣ U ∣ ) ≥ 1 , so that ρ k ≤ ∣ U ∣ < ρ k − 1 \rho_k\leq|U|<\rho_{k-1} ρ k ≤ ∣ U ∣ < ρ k − 1 . The intervals I σ I_\sigma I σ with ∣ σ ∣ = k − 1 |\sigma|=k-1 ∣ σ ∣ = k − 1 are pairwise disjoint and have length at least K − 1 ρ k − 1 > K − 1 ∣ U ∣ K^{-1}\rho_{k-1}>K^{-1}|U| K − 1 ρ k − 1 > K − 1 ∣ U ∣ . If U U U intersects N N N of them, at least N − 2 N-2 N − 2 of them are contained in the convex hull of U U U , so ( N − 2 ) K − 1 ρ k − 1 ≤ ∣ U ∣ < ρ k − 1 (N-2)K^{-1}\rho_{k-1}\leq|U|<\rho_{k-1} ( N − 2 ) K − 1 ρ k − 1 ≤ ∣ U ∣ < ρ k − 1 , that is, N < K + 2 N<K+2 N < K + 2 . Since F ⊂ ⋃ ∣ σ ∣ = k − 1 I σ F\subset\bigcup_{|\sigma|=k-1}I_\sigma F ⊂ ⋃ ∣ σ ∣ = k − 1 I σ , we obtain
μ ( U ) ≤ ( K + 2 ) 2 − ( k − 1 ) = 2 ( K + 2 ) ∣ U ∣ s ρ ( ∣ U ∣ ) , \begin{equation}
\mu(U)\leq (K+2)\,2^{-(k-1)}=2(K+2)\,|U|^{s_\rho(|U|)}, \tag*{(5)}
\end{equation} μ ( U ) ≤ ( K + 2 ) 2 − ( k − 1 ) = 2 ( K + 2 ) ∣ U ∣ s ρ ( ∣ U ∣ ) , (5) where we used that ∣ U ∣ s ρ ( ∣ U ∣ ) = exp ( − k log 2 ) = 2 − k |U|^{s_\rho(|U|)}=\exp(-k\log2)=2^{-k} ∣ U ∣ s ρ ( ∣ U ∣ ) = exp ( − k log 2 ) = 2 − k by (2) . On the other hand, if ϕ ∈ ( 0 , ρ 0 ) \phi\in(0,\rho_0) ϕ ∈ ( 0 , ρ 0 ) and k = k ρ ( ϕ ) k=k_\rho(\phi) k = k ρ ( ϕ ) , the 2 k 2^k 2 k intervals I σ I_\sigma I σ with ∣ σ ∣ = k |\sigma|=k ∣ σ ∣ = k cover F F F and have length at most K ρ k ≤ K ϕ K\rho_k\leq K\phi K ρ k ≤ K ϕ . Covering each of them by ⌈ K ⌉ \lceil K\rceil ⌈ K ⌉ intervals of length ϕ \phi ϕ we obtain a cover V ϕ \mathcal{V}_\phi V ϕ of F F F by ⌈ K ⌉ 2 k \lceil K\rceil 2^k ⌈ K ⌉ 2 k sets of diameter ϕ \phi ϕ , and for every t ≥ 0 t\geq0 t ≥ 0 ,
∑ V ∈ V ϕ ∣ V ∣ t = ⌈ K ⌉ 2 k ϕ t = ⌈ K ⌉ ϕ t − s ρ ( ϕ ) . \begin{equation}
\sum_{V\in\mathcal{V}_\phi}|V|^t=\lceil K\rceil\,2^k\phi^t=\lceil K\rceil\,\phi^{\,t-s_\rho(\phi)} . \tag*{(6)}
\end{equation} V ∈ V ϕ ∑ ∣ V ∣ t = ⌈ K ⌉ 2 k ϕ t = ⌈ K ⌉ ϕ t − s ρ ( ϕ ) . (6) Fix θ ∈ ( 0 , 1 ] \theta\in(0,1] θ ∈ ( 0 , 1 ] and let d d d denote the lim sup \limsup lim sup in (3) . Let t > t ′ > d t>t'>d t > t ′ > d . There is δ 0 ∈ ( 0 , ρ 0 ) \delta_0\in(0,\rho_0) δ 0 ∈ ( 0 , ρ 0 ) such that for all δ ∈ ( 0 , δ 0 ) \delta\in(0,\delta_0) δ ∈ ( 0 , δ 0 ) there exists ϕ ∈ [ δ 1 / θ , δ ] \phi\in[\delta^{1/\theta},\delta] ϕ ∈ [ δ 1/ θ , δ ] with s ρ ( ϕ ) < t ′ s_\rho(\phi)<t' s ρ ( ϕ ) < t ′ . By (6) , V ϕ \mathcal{V}_\phi V ϕ is a cover of F F F by sets with diameters in [ δ 1 / θ , δ ] [\delta^{1/\theta},\delta] [ δ 1/ θ , δ ] and
∑ V ∈ V ϕ ∣ V ∣ t = ⌈ K ⌉ ϕ t − s ρ ( ϕ ) ≤ ⌈ K ⌉ δ t − t ′ → δ → 0 0. \sum_{V\in\mathcal{V}_\phi}|V|^t=\lceil K\rceil\,\phi^{\,t-s_\rho(\phi)}\leq\lceil K\rceil\,\delta^{\,t-t'}\xrightarrow[\delta\to0]{}0 . V ∈ V ϕ ∑ ∣ V ∣ t = ⌈ K ⌉ ϕ t − s ρ ( ϕ ) ≤ ⌈ K ⌉ δ t − t ′ δ → 0 0. Hence \overline{\mbox{\rm dim}}_{\,\theta}F\leq t , and letting t → d t\to d t → d gives \overline{\mbox{\rm dim}}_{\,\theta}F\leq d .
Conversely, suppose d > 0 d>0 d > 0 and let 0 < t < d 0<t<d 0 < t < d . There exist δ n → 0 \delta_n\to0 δ n → 0 with δ n < ρ 0 \delta_n<\rho_0 δ n < ρ 0 such that s ρ ( ϕ ) > t s_\rho(\phi)>t s ρ ( ϕ ) > t for all ϕ ∈ [ δ n 1 / θ , δ n ] \phi\in[\delta_n^{1/\theta},\delta_n] ϕ ∈ [ δ n 1/ θ , δ n ] . Let { U i } \{U_i\} { U i } be any cover of F F F with δ n 1 / θ ≤ ∣ U i ∣ ≤ δ n \delta_n^{1/\theta}\leq|U_i|\leq\delta_n δ n 1/ θ ≤ ∣ U i ∣ ≤ δ n , where we may assume that each U i U_i U i intersects F F F . By (5) , and since ∣ U i ∣ < 1 |U_i|<1 ∣ U i ∣ < 1 ,
1 = μ ( F ) ≤ ∑ i μ ( U i ) ≤ 2 ( K + 2 ) ∑ i ∣ U i ∣ s ρ ( ∣ U i ∣ ) ≤ 2 ( K + 2 ) ∑ i ∣ U i ∣ t . 1=\mu(F)\leq\sum_i\mu(U_i)\leq2(K+2)\sum_i|U_i|^{s_\rho(|U_i|)}\leq2(K+2)\sum_i|U_i|^{t}. 1 = μ ( F ) ≤ i ∑ μ ( U i ) ≤ 2 ( K + 2 ) i ∑ ∣ U i ∣ s ρ ( ∣ U i ∣ ) ≤ 2 ( K + 2 ) i ∑ ∣ U i ∣ t . Thus every such cover has t t t -cost at least 1 / ( 2 ( K + 2 ) ) 1/(2(K+2)) 1/ ( 2 ( K + 2 )) , for arbitrarily small δ n \delta_n δ n , and by Definition 2.1 we get \overline{\mbox{\rm dim}}_{\,\theta}F\geq t . Letting t → d t\to d t → d proves (3) .
Finally, if t < lim inf ϕ → 0 s ρ ( ϕ ) t<\liminf_{\phi\to0}s_\rho(\phi) t < lim inf ϕ → 0 s ρ ( ϕ ) , then (5) gives μ ( U ) ≤ 2 ( K + 2 ) ∣ U ∣ t \mu(U)\leq2(K+2)|U|^t μ ( U ) ≤ 2 ( K + 2 ) ∣ U ∣ t for every set U U U of sufficiently small diameter, and the mass distribution principle (see [34 ] ) yields dim H F ≥ t \dim_{\rm H}F\geq t dim H F ≥ t . On the other hand, (6) shows that F F F can be covered by ⌈ K ⌉ 2 k ρ ( ϕ ) \lceil K\rceil2^{k_\rho(\phi)} ⌈ K ⌉ 2 k ρ ( ϕ ) sets of diameter ϕ \phi ϕ , so
\underline{\mbox{\rm dim}}_{\rm B}\,F\leq\liminf_{\phi\to0}\frac{\log\bigl(\lceil K\rceil2^{k_\rho(\phi)}\bigr)}{-\log\phi}=\liminf_{\phi\to0}s_\rho(\phi).
Since \dim_{\rm H}F\leq\underline{\mbox{\rm dim}}_{\rm B}\,F , (4) follows. ◻
We also record some elementary consequences of (H) .
Lemma 4.14. Let a a a be a positive, non-increasing sequence with ∑ a n = 1 \sum a_n=1 ∑ a n = 1 satisfying (H) . Then the following hold.
( 1 − 2 / κ 1 ) s k ( a ) ≤ a 2 k ≤ s k − 1 ( a ) ≤ κ 2 s k ( a ) (1-2/\kappa_1)\,s_k(a)\leq a_{2^k}\leq s_{k-1}(a)\leq\kappa_2\,s_k(a) ( 1 − 2/ κ 1 ) s k ( a ) ≤ a 2 k ≤ s k − 1 ( a ) ≤ κ 2 s k ( a ) for all k ≥ 1 k\geq1 k ≥ 1 .
a 2 k ≤ c a a 2 k + 1 a_{2^k}\leq c_a\,a_{2^{k+1}} a 2 k ≤ c a a 2 k + 1 for all k ≥ 1 k\geq1 k ≥ 1 , where c a : = κ 1 κ 2 2 / ( κ 1 − 2 ) c_a:=\kappa_1\kappa_2^2/(\kappa_1-2) c a := κ 1 κ 2 2 / ( κ 1 − 2 ) .
The Cantor set C a C_a C a satisfies 0 < log 2 log κ 2 ≤ dim L C a ≤ dim H C a ≤ B ≤ dim A C a ≤ log 2 log κ 1 < 1. 0<\frac{\log2}{\log\kappa_2}\leq\dim_L C_a\leq\dim_{\rm H}C_a\leq B\leq\dim_A C_a\leq\frac{\log2}{\log\kappa_1}<1 . 0 < l o g κ 2 l o g 2 ≤ dim L C a ≤ dim H C a ≤ B ≤ dim A C a ≤ l o g κ 1 l o g 2 < 1.
Proof. Write s k = s k ( a ) s_k=s_k(a) s k = s k ( a ) . Since a a a is non-increasing and s k = 2 − k ∑ i = 2 k 2 k + 1 − 1 a i + 2 s k + 1 s_k=2^{-k}\sum_{i=2^k}^{2^{k+1}-1}a_i+2s_{k+1} s k = 2 − k ∑ i = 2 k 2 k + 1 − 1 a i + 2 s k + 1 , we have
a 2 k ≥ 2 − k ∑ i = 2 k 2 k + 1 − 1 a i = s k − 2 s k + 1 ≥ ( 1 − 2 κ 1 ) s k . a_{2^k}\geq2^{-k}\sum_{i=2^k}^{2^{k+1}-1}a_i=s_k-2s_{k+1}\geq\Bigl(1-\frac{2}{\kappa_1}\Bigr)s_k . a 2 k ≥ 2 − k i = 2 k ∑ 2 k + 1 − 1 a i = s k − 2 s k + 1 ≥ ( 1 − κ 1 2 ) s k . Moreover a 2 k ≤ a i a_{2^k}\leq a_i a 2 k ≤ a i for 2 k − 1 ≤ i < 2 k 2^{k-1}\leq i<2^k 2 k − 1 ≤ i < 2 k , so a 2 k ≤ 2 − ( k − 1 ) ∑ i = 2 k − 1 2 k − 1 a i ≤ s k − 1 ≤ κ 2 s k a_{2^k}\leq2^{-(k-1)}\sum_{i=2^{k-1}}^{2^k-1}a_i\leq s_{k-1}\leq\kappa_2s_k a 2 k ≤ 2 − ( k − 1 ) ∑ i = 2 k − 1 2 k − 1 a i ≤ s k − 1 ≤ κ 2 s k . This proves (a). By (a),
a 2 k ≤ κ 2 s k ≤ κ 2 2 s k + 1 ≤ κ 2 2 1 − 2 / κ 1 a 2 k + 1 = c a a 2 k + 1 , a_{2^k}\leq\kappa_2s_k\leq\kappa_2^2s_{k+1}\leq\frac{\kappa_2^2}{1-2/\kappa_1}\,a_{2^{k+1}}=c_a\,a_{2^{k+1}}, a 2 k ≤ κ 2 s k ≤ κ 2 2 s k + 1 ≤ 1 − 2/ κ 1 κ 2 2 a 2 k + 1 = c a a 2 k + 1 , which is (b). Finally, (H) gives κ 1 n ≤ s k / s k + n ≤ κ 2 n \kappa_1^n\leq s_k/s_{k+n}\leq\kappa_2^n κ 1 n ≤ s k / s k + n ≤ κ 2 n for all k k k and n n n , so every quotient n log 2 / log ( s k / s k + n ) n\log2/\log(s_k/s_{k+n}) n log 2/ log ( s k / s k + n ) lies in [ log 2 / log κ 2 , log 2 / log κ 1 ] [\log2/\log\kappa_2,\log2/\log\kappa_1] [ log 2/ log κ 2 , log 2/ log κ 1 ] . Part (c) now follows from the formulas for dim A C a \dim_A C_a dim A C a and dim L C a \dim_L C_a dim L C a recalled in § 3 and from the inequalities \dim_L\leq\dim_{\rm H}\leq\underline{\mbox{\rm dim}}_{\rm B}\,\leq\overline{\mbox{\rm dim}}_{\rm B}\,\leq\dim_A . ◻
Mountains
By Lemma 4.12 , prescribing the upper intermediate dimensions of a set amounts to constructing a function G G G , playing the role of S ρ S_\rho S ρ , whose window infima over intervals of length log ( 1 / θ ) \log(1/\theta) log ( 1/ θ ) reproduce h ( θ ) h(\theta) h ( θ ) . In addition, the function must belong to the class G ( 0 , 1 ) \mathcal{G}(0,1) G ( 0 , 1 ) ; that is, its Dini derivatives must lie between − G -G − G and 1 − G 1-G 1 − G . These are the slopes that a function of the form S ρ S_\rho S ρ can have. This condition allows Lemma 4.4 to produce a sequence of scales ρ \rho ρ with S ρ S_\rho S ρ close to the constructed function. The basic building block of this function is a mountain associated with h h h and a parameter ε > 0 \varepsilon>0 ε > 0 . This is a continuous function with admissible slopes which starts and ends at h ( ε ) h(\varepsilon) h ( ε ) . For every θ ∈ [ ε , 1 ] \theta\in[\varepsilon,1] θ ∈ [ ε , 1 ] , it contains a window of length log ( 1 / θ ) \log(1/\theta) log ( 1/ θ ) on which it stays above h ( θ ) h(\theta) h ( θ ) , while every such window meets a point where it is at most h ( θ ) h(\theta) h ( θ ) . In the proof of Theorem 4.17 , infinitely many mountains, with ε → 0 \varepsilon\to 0 ε → 0 , are joined by valleys to produce the required function.
The next lemma is a version of the mountains of [26 Theorem 3.9] in the case λ = 0 \lambda=0 λ = 0 and α = 1 \alpha=1 α = 1 . We present it in a slightly different way, based on Lemma 4.7 , which makes all the required properties transparent.
Lemma 4.15. Let h ∈ H ( 0 , 1 ) h\in\mathcal{H}(0,1) h ∈ H ( 0 , 1 ) with h ( 1 ) = B ∈ ( 0 , 1 ) h(1)=B\in(0,1) h ( 1 ) = B ∈ ( 0 , 1 ) , let ε ∈ ( 0 , 1 ) \varepsilon\in(0,1) ε ∈ ( 0 , 1 ) and write ℓ ε : = log ( 1 / ε ) \ell_\varepsilon:=\log(1/\varepsilon) ℓ ε := log ( 1/ ε ) . There exists a continuous function f = f h , ε : [ 0 , ℓ ε ] → [ h ( ε ) , B ] f=f_{h,\varepsilon}:[0,\ell_\varepsilon]\to[h(\varepsilon),B] f = f h , ε : [ 0 , ℓ ε ] → [ h ( ε ) , B ] with the following properties.
f ( 0 ) = f ( ℓ ε ) = h ( ε ) f(0)=f(\ell_\varepsilon)=h(\varepsilon) f ( 0 ) = f ( ℓ ε ) = h ( ε ) and max [ 0 , ℓ ε ] f = B \max_{[0,\ell_\varepsilon]}f=B max [ 0 , ℓ ε ] f = B .
D − f ( y ) ∈ [ − f ( y ) , 1 − f ( y ) ] D_-f(y)\in[-f(y),\,1-f(y)] D − f ( y ) ∈ [ − f ( y ) , 1 − f ( y )] for every y ∈ ( 0 , ℓ ε ] y\in(0,\ell_\varepsilon] y ∈ ( 0 , ℓ ε ] .
For every θ ∈ [ ε , 1 ] \theta\in[\varepsilon,1] θ ∈ [ ε , 1 ] there exists x θ ∈ [ 0 , ℓ ε − log ( 1 / θ ) ] x_\theta\in[0,\ell_\varepsilon-\log(1/\theta)] x θ ∈ [ 0 , ℓ ε − log ( 1/ θ )] such that f ( y ) ≥ h ( θ ) f(y)\geq h(\theta) f ( y ) ≥ h ( θ ) for all y ∈ [ x θ , x θ + log ( 1 / θ ) ] y\in[x_\theta,x_\theta+\log(1/\theta)] y ∈ [ x θ , x θ + log ( 1/ θ )] .
For every θ ∈ [ ε , 1 ] \theta\in[\varepsilon,1] θ ∈ [ ε , 1 ] and every y ∈ [ 0 , ℓ ε − log ( 1 / θ ) ] y\in[0,\ell_\varepsilon-\log(1/\theta)] y ∈ [ 0 , ℓ ε − log ( 1/ θ )] there exists z ∈ [ y , y + log ( 1 / θ ) ] z\in[y,y+\log(1/\theta)] z ∈ [ y , y + log ( 1/ θ )] with f ( z ) ≤ h ( θ ) f(z)\leq h(\theta) f ( z ) ≤ h ( θ ) .
Proof. By Lemma 4.7 , 0 < h < 1 0<h<1 0 < h < 1 on ( 0 , 1 ] (0,1] ( 0 , 1 ] and Q h Q_h Q h is non-increasing. For θ ∈ [ ε , 1 ] \theta\in[\varepsilon,1] θ ∈ [ ε , 1 ] define
x L ( θ ) : = log 1 − h ( ε ) 1 − h ( θ ) and x R ( θ ) : = x L ( θ ) + log 1 θ . x_L(\theta):=\log\frac{1-h(\varepsilon)}{1-h(\theta)}\qquad\text{and}\qquad x_R(\theta):=x_L(\theta)+\log\frac1\theta . x L ( θ ) := log 1 − h ( θ ) 1 − h ( ε ) and x R ( θ ) := x L ( θ ) + log θ 1 . The function x L x_L x L is continuous and non-decreasing, with x L ( ε ) = 0 x_L(\varepsilon)=0 x L ( ε ) = 0 and x L ( 1 ) = x ∗ : = log ( ( 1 − h ( ε ) ) / ( 1 − B ) ) x_L(1)=x^*:=\log\bigl((1-h(\varepsilon))/(1-B)\bigr) x L ( 1 ) = x ∗ := log ( ( 1 − h ( ε )) / ( 1 − B ) ) . Since 1 / ( 1 − h ( θ ) ) = 1 + θ Q h ( θ ) 1/(1-h(\theta))=1+\theta Q_h(\theta) 1/ ( 1 − h ( θ )) = 1 + θ Q h ( θ ) , we can write
x R ( θ ) = log ( 1 − h ( ε ) ) + log ( 1 θ + Q h ( θ ) ) , x_R(\theta)=\log\bigl(1-h(\varepsilon)\bigr)+\log\Bigl(\frac1\theta+Q_h(\theta)\Bigr), x R ( θ ) = log ( 1 − h ( ε ) ) + log ( θ 1 + Q h ( θ ) ) , and since Q h Q_h Q h is non-increasing while 1 / θ 1/\theta 1/ θ is strictly decreasing, x R x_R x R is continuous and strictly decreasing on [ ε , 1 ] [\varepsilon,1] [ ε , 1 ] , with x R ( ε ) = ℓ ε x_R(\varepsilon)=\ell_\varepsilon x R ( ε ) = ℓ ε and x R ( 1 ) = x ∗ x_R(1)=x^* x R ( 1 ) = x ∗ . In particular 0 ≤ x ∗ ≤ ℓ ε 0\leq x^*\leq\ell_\varepsilon 0 ≤ x ∗ ≤ ℓ ε . Let τ : [ x ∗ , ℓ ε ] → [ ε , 1 ] \tau:[x^*,\ell_\varepsilon]\to[\varepsilon,1] τ : [ x ∗ , ℓ ε ] → [ ε , 1 ] be the inverse of x R x_R x R , which is continuous and strictly decreasing, and define
f ( y ) : = { 1 − ( 1 − h ( ε ) ) e − y if y ∈ [ 0 , x ∗ ] , h ( τ ( y ) ) if y ∈ [ x ∗ , ℓ ε ] . f(y):=
\begin{cases}
1-(1-h(\varepsilon))\,e^{-y} & \text{if } y\in[0,x^*],\\[2pt]
h(\tau(y)) & \text{if } y\in[x^*,\ell_\varepsilon].
\end{cases} f ( y ) := { 1 − ( 1 − h ( ε )) e − y h ( τ ( y )) if y ∈ [ 0 , x ∗ ] , if y ∈ [ x ∗ , ℓ ε ] . Both expressions are equal to B B B at y = x ∗ y=x^* y = x ∗ , so f f f is continuous. On [ 0 , x ∗ ] [0,x^*] [ 0 , x ∗ ] the function f f f increases from h ( ε ) h(\varepsilon) h ( ε ) to B B B , and on [ x ∗ , ℓ ε ] [x^*,\ell_\varepsilon] [ x ∗ , ℓ ε ] it is non-increasing, from B B B to h ( τ ( ℓ ε ) ) = h ( ε ) h(\tau(\ell_\varepsilon))=h(\varepsilon) h ( τ ( ℓ ε )) = h ( ε ) . This gives (M1) and shows that f f f takes values in [ h ( ε ) , B ] [h(\varepsilon),B] [ h ( ε ) , B ] . Moreover, for every θ ∈ [ ε , 1 ] \theta\in[\varepsilon,1] θ ∈ [ ε , 1 ] ,
f ( x L ( θ ) ) = h ( θ ) = f ( x R ( θ ) ) . \begin{equation}
f(x_L(\theta))=h(\theta)=f(x_R(\theta)). \tag*{(7)}
\end{equation} f ( x L ( θ )) = h ( θ ) = f ( x R ( θ )) . (7) To prove (M3), let θ ∈ [ ε , 1 ] \theta\in[\varepsilon,1] θ ∈ [ ε , 1 ] and put x θ : = x L ( θ ) x_\theta:=x_L(\theta) x θ := x L ( θ ) , so that x θ + log ( 1 / θ ) = x R ( θ ) ≤ ℓ ε x_\theta+\log(1/\theta)=x_R(\theta)\leq\ell_\varepsilon x θ + log ( 1/ θ ) = x R ( θ ) ≤ ℓ ε . If y ∈ [ x L ( θ ) , x ∗ ] y\in[x_L(\theta),x^*] y ∈ [ x L ( θ ) , x ∗ ] , then f ( y ) ≥ f ( x L ( θ ) ) = h ( θ ) f(y)\geq f(x_L(\theta))=h(\theta) f ( y ) ≥ f ( x L ( θ )) = h ( θ ) because f f f is increasing on [ 0 , x ∗ ] [0,x^*] [ 0 , x ∗ ] . If y ∈ [ x ∗ , x R ( θ ) ] y\in[x^*,x_R(\theta)] y ∈ [ x ∗ , x R ( θ )] , then τ ( y ) ≥ θ \tau(y)\geq\theta τ ( y ) ≥ θ and so f ( y ) = h ( τ ( y ) ) ≥ h ( θ ) f(y)=h(\tau(y))\geq h(\theta) f ( y ) = h ( τ ( y )) ≥ h ( θ ) .
To prove (M4), let θ ∈ [ ε , 1 ] \theta\in[\varepsilon,1] θ ∈ [ ε , 1 ] and y ∈ [ 0 , ℓ ε − log ( 1 / θ ) ] y\in[0,\ell_\varepsilon-\log(1/\theta)] y ∈ [ 0 , ℓ ε − log ( 1/ θ )] . If y ≤ x L ( θ ) y\leq x_L(\theta) y ≤ x L ( θ ) , then f ( y ) ≤ f ( x L ( θ ) ) = h ( θ ) f(y)\leq f(x_L(\theta))=h(\theta) f ( y ) ≤ f ( x L ( θ )) = h ( θ ) . Otherwise z : = y + log ( 1 / θ ) z:=y+\log(1/\theta) z := y + log ( 1/ θ ) satisfies x ∗ ≤ x R ( θ ) < z ≤ ℓ ε x^*\leq x_R(\theta)<z\leq\ell_\varepsilon x ∗ ≤ x R ( θ ) < z ≤ ℓ ε , so τ ( z ) < θ \tau(z)<\theta τ ( z ) < θ and f ( z ) = h ( τ ( z ) ) ≤ h ( θ ) f(z)=h(\tau(z))\leq h(\theta) f ( z ) = h ( τ ( z )) ≤ h ( θ ) . In both cases we have found a point of [ y , y + log ( 1 / θ ) ] [y,y+\log(1/\theta)] [ y , y + log ( 1/ θ )] at which f f f is at most h ( θ ) h(\theta) h ( θ ) .
It remains to prove (M2). For y ∈ ( 0 , x ∗ ] y\in(0,x^*] y ∈ ( 0 , x ∗ ] we have D − f ( y ) = ( 1 − h ( ε ) ) e − y = 1 − f ( y ) ∈ [ − f ( y ) , 1 − f ( y ) ] D_-f(y)=(1-h(\varepsilon))e^{-y}=1-f(y)\in[-f(y),1-f(y)] D − f ( y ) = ( 1 − h ( ε )) e − y = 1 − f ( y ) ∈ [ − f ( y ) , 1 − f ( y )] . On ( x ∗ , ℓ ε ] (x^*,\ell_\varepsilon] ( x ∗ , ℓ ε ] the function f f f is non-increasing, so D − f ( y ) ≤ 0 ≤ 1 − f ( y ) D_-f(y)\leq0\leq1-f(y) D − f ( y ) ≤ 0 ≤ 1 − f ( y ) . For the lower bound we claim that
f ( y ′ ) ≤ f ( y ) e y − y ′ whenever x ∗ ≤ y ′ ≤ y ≤ ℓ ε . \begin{equation}
f(y')\leq f(y)\,e^{\,y-y'}\qquad\text{whenever } x^*\leq y'\leq y\leq\ell_\varepsilon. \tag*{(8)}
\end{equation} f ( y ′ ) ≤ f ( y ) e y − y ′ whenever x ∗ ≤ y ′ ≤ y ≤ ℓ ε . (8) Indeed, let θ = τ ( y ) ≤ θ ′ = τ ( y ′ ) \theta=\tau(y)\leq\theta'=\tau(y') θ = τ ( y ) ≤ θ ′ = τ ( y ′ ) . From the definition of Q h Q_h Q h we get h = Q h / ( θ − 1 + Q h ) h=Q_h/(\theta^{-1}+Q_h) h = Q h / ( θ − 1 + Q h ) , hence
f ( y ) = Q h ( θ ) θ − 1 + Q h ( θ ) , f ( y ′ ) = Q h ( θ ′ ) θ ′ − 1 + Q h ( θ ′ ) , f(y)=\frac{Q_h(\theta)}{\theta^{-1}+Q_h(\theta)},\qquad f(y')=\frac{Q_h(\theta')}{\theta'^{-1}+Q_h(\theta')}, f ( y ) = θ − 1 + Q h ( θ ) Q h ( θ ) , f ( y ′ ) = θ ′ − 1 + Q h ( θ ′ ) Q h ( θ ′ ) , and
e y − y ′ = e x R ( θ ) − x R ( θ ′ ) = θ − 1 + Q h ( θ ) θ ′ − 1 + Q h ( θ ′ ) . e^{\,y-y'}=e^{\,x_R(\theta)-x_R(\theta')}=\frac{\theta^{-1}+Q_h(\theta)}{\theta'^{-1}+Q_h(\theta')}. e y − y ′ = e x R ( θ ) − x R ( θ ′ ) = θ ′ − 1 + Q h ( θ ′ ) θ − 1 + Q h ( θ ) . Therefore f ( y ) e y − y ′ = Q h ( θ ) / ( θ ′ − 1 + Q h ( θ ′ ) ) ≥ Q h ( θ ′ ) / ( θ ′ − 1 + Q h ( θ ′ ) ) = f ( y ′ ) f(y)e^{\,y-y'}=Q_h(\theta)/(\theta'^{-1}+Q_h(\theta'))\geq Q_h(\theta')/(\theta'^{-1}+Q_h(\theta'))=f(y') f ( y ) e y − y ′ = Q h ( θ ) / ( θ ′ − 1 + Q h ( θ ′ )) ≥ Q h ( θ ′ ) / ( θ ′ − 1 + Q h ( θ ′ )) = f ( y ′ ) , since Q h ( θ ) ≥ Q h ( θ ′ ) Q_h(\theta)\geq Q_h(\theta') Q h ( θ ) ≥ Q h ( θ ′ ) . Now if y ∈ ( x ∗ , ℓ ε ] y\in(x^*,\ell_\varepsilon] y ∈ ( x ∗ , ℓ ε ] and 0 < t < y − x ∗ 0<t<y-x^* 0 < t < y − x ∗ , then (8) gives
f ( y ) − f ( y − t ) t ≥ f ( y ) 1 − e t t → t → 0 + − f ( y ) , \frac{f(y)-f(y-t)}{t}\geq f(y)\,\frac{1-e^{t}}{t}\xrightarrow[t\to0^+]{}-f(y), t f ( y ) − f ( y − t ) ≥ f ( y ) t 1 − e t t → 0 + − f ( y ) , so D − f ( y ) ≥ − f ( y ) D_-f(y)\geq-f(y) D − f ( y ) ≥ − f ( y ) . This completes the proof. ◻
Prescribing the upper intermediate dimensions of complementary sets
We can now state and prove the main result of this section.
Theorem 4.17. Let a = { a n } a=\{a_n\} a = { a n } be a positive, non-increasing sequence with ∑ n a n = 1 \sum_n a_n=1 ∑ n a n = 1 such that 2 < inf n s n ( a ) s n + 1 ( a ) ≤ sup n s n ( a ) s n + 1 ( a ) < ∞ . 2<\inf_n\frac{s_n(a)}{s_{n+1}(a)}\leq\sup_n\frac{s_n(a)}{s_{n+1}(a)}<\infty . 2 < inf n s n + 1 ( a ) s n ( a ) ≤ sup n s n + 1 ( a ) s n ( a ) < ∞. Let 0 ≤ λ < α ≤ 1 0\leq\lambda<\alpha\leq1 0 ≤ λ < α ≤ 1 and let h ∈ H ( λ , α ) h\in\mathcal{H}(\lambda,\alpha) h ∈ H ( λ , α ) be such that h(\theta)\leq\overline{\mbox{\rm dim}}_{\,\theta}C_a\ \ \text{for all }\theta\in[0,1]\qquad\text{and}\qquad h(1)=\overline{\mbox{\rm dim}}_{\rm B}\,C_a . Then there exists E ∈ C a E\in\mathcal{C}_a E ∈ C a such that \overline{\mbox{\rm dim}}_{\,\theta}E=h(\theta) for all θ ∈ [ 0 , 1 ] \theta\in[0,1] θ ∈ [ 0 , 1 ] .
Before giving the proof, let us describe the strategy. As in the proof of Theorem 3.3 , the set E E E will be the union of a Cantor type set M M M , built with some of the terms of a a a , and a countable set D ′ D' D ′ , built with the remaining terms. The set M M M will satisfy \overline{\mbox{\rm dim}}_{\,\theta}M=h(\theta) , while \overline{\mbox{\rm dim}}_{\,\theta}D'\leq\varphi_B(\theta)\leq h(\theta) by Proposition 4.8 . To construct M M M we follow the approach of [26 ] . We concatenate mountains associated with h h h and valleys to obtain a function G ∈ G ( 0 , 1 ) G\in\mathcal{G}(0,1) G ∈ G ( 0 , 1 ) whose window infima reproduce h h h , and Lemma 4.4 provides a sequence of scales ρ ( 2 ) \rho^{(2)} ρ ( 2 ) adapted to G G G . Two new difficulties appear. First, since M M M must be built with terms of a a a , the lengths of its intervals of level k k k should not exceed, up to a constant factor, those of C a C_a C a (see Remark 4.19 ), while the hypothesis h\leq\overline{\mbox{\rm dim}}_{\,\theta}C_a only gives information about infima over windows. We overcome this by taking the minimum of ρ ( 2 ) \rho^{(2)} ρ ( 2 ) and the scales of C a C_a C a , and by placing each mountain on a window where the covering numbers of C a C_a C a are large. Second, the function h h h may saturate (1) for λ = 0 \lambda=0 λ = 0 and α = 1 \alpha=1 α = 1 , so the ratios of consecutive scales may approach 0 0 0 and 1 / 2 1/2 1/2 . For this reason M M M is not a decreasing Cantor set C b C_b C b as in Theorem 3.3 , but the limit set of a homogeneous Cantor scheme in which the gaps are not arranged in decreasing order (see Remark 4.20 ).
Proof. Step 0. Preliminary reductions. By Remark 4.3 , h ∈ H ( 0 , 1 ) h\in\mathcal{H}(0,1) h ∈ H ( 0 , 1 ) . By Lemma 4.14 (c), B=\overline{\mbox{\rm dim}}_{\rm B}\,C_a\in(0,1) , and then Lemma 4.7 and Proposition 4.8 give
0 < h ( θ ) ≤ B < 1 for θ ∈ ( 0 , 1 ] and h ≥ φ B on [ 0 , 1 ] . \begin{equation}
0<h(\theta)\leq B<1\quad\text{for }\theta\in(0,1]\qquad\text{and}\qquad h\geq\varphi_B\ \text{ on }[0,1]. \tag*{(9)}
\end{equation} 0 < h ( θ ) ≤ B < 1 for θ ∈ ( 0 , 1 ] and h ≥ φ B on [ 0 , 1 ] . (9) We write U(\theta):=\overline{\mbox{\rm dim}}_{\,\theta}C_a for θ ∈ [ 0 , 1 ] \theta\in[0,1] θ ∈ [ 0 , 1 ] , so that U ( 0 ) = dim H C a U(0)=\dim_{\rm H}C_a U ( 0 ) = dim H C a .
Step 1. The scales of C a C_a C a . Let κ 1 , κ 2 \kappa_1,\kappa_2 κ 1 , κ 2 be as in (H) , let c a c_a c a be as in Lemma 4.14 , and set
C : = κ 1 κ 2 κ 1 − 2 > 1 and ρ k ( 1 ) : = C − 1 s k ( a ) ( k ≥ 0 ) , C:=\frac{\kappa_1\kappa_2}{\kappa_1-2}>1\qquad\text{and}\qquad \rho^{(1)}_k:=C^{-1}s_k(a)\quad(k\geq0), C := κ 1 − 2 κ 1 κ 2 > 1 and ρ k ( 1 ) := C − 1 s k ( a ) ( k ≥ 0 ) , and write S 1 : = S ρ ( 1 ) S_1:=S_{\rho^{(1)}} S 1 := S ρ ( 1 ) . As observed after Definition 4.11 , C a C_a C a is the limit set of a ( ρ ( 1 ) , C κ 2 ) (\rho^{(1)},C\kappa_2) ( ρ ( 1 ) , C κ 2 ) -homogeneous binary Cantor scheme. Hence Lemma 4.12 gives
U ( θ ) = lim sup x → ∞ inf y ∈ [ x , x + log ( 1 / θ ) ] S 1 ( y ) for θ ∈ ( 0 , 1 ] , U ( 0 ) = lim inf x → ∞ S 1 ( x ) . \begin{equation}
\begin{split}
U(\theta)&=\limsup_{x\to\infty}\ \inf_{y\in[x,x+\log(1/\theta)]}S_1(y)\qquad\text{for }\theta\in(0,1],\\
U(0)&=\liminf_{x\to\infty}S_1(x).
\end{split} \tag*{(10)}
\end{equation} U ( θ ) U ( 0 ) = x → ∞ lim sup y ∈ [ x , x + l o g ( 1/ θ )] inf S 1 ( y ) for θ ∈ ( 0 , 1 ] , = x → ∞ lim inf S 1 ( x ) . (10) Step 2. Parameters. Let { q 1 , q 2 , … } \{q_1,q_2,\dots\} { q 1 , q 2 , … } be an enumeration of Q ∩ ( 0 , 1 ) \mathbb{Q}\cap(0,1) Q ∩ ( 0 , 1 ) and let ( θ j ) j ≥ 1 (\theta_j)_{j\geq1} ( θ j ) j ≥ 1 be the sequence
q 1 , q 1 , q 2 , q 1 , q 2 , q 3 , q 1 , q 2 , q 3 , q 4 , … , q_1,\ q_1,q_2,\ q_1,q_2,q_3,\ q_1,q_2,q_3,q_4,\ \dots, q 1 , q 1 , q 2 , q 1 , q 2 , q 3 , q 1 , q 2 , q 3 , q 4 , … , in which every q i q_i q i appears infinitely many times. Put
ε j : = min { θ 1 , … , θ j , 1 j + 1 } . \varepsilon_j:=\min\Bigl\{\theta_1,\dots,\theta_j,\frac{1}{j+1}\Bigr\}. ε j := min { θ 1 , … , θ j , j + 1 1 } . Then the sequence ( ε j ) (\varepsilon_j) ( ε j ) is non-increasing and converges to 0 0 0 , and 0 < ε j ≤ θ j < 1 0<\varepsilon_j\leq\theta_j<1 0 < ε j ≤ θ j < 1 for every j j j . By (9) , h ( ε j ) > 0 h(\varepsilon_j)>0 h ( ε j ) > 0 , and we define
β j : = max { h ( 0 ) , h ( ε j + 1 ) j + 1 } . \beta_j:=\max\Bigl\{h(0),\ \frac{h(\varepsilon_{j+1})}{j+1}\Bigr\}. β j := max { h ( 0 ) , j + 1 h ( ε j + 1 ) } . Then 0 < β j ≤ h ( ε j + 1 ) ≤ h ( ε j ) 0<\beta_j\leq h(\varepsilon_{j+1})\leq h(\varepsilon_j) 0 < β j ≤ h ( ε j + 1 ) ≤ h ( ε j ) and β j → h ( 0 ) \beta_j\to h(0) β j → h ( 0 ) as j → ∞ j\to\infty j → ∞ . For each j j j let f j : = f h , ε j f_j:=f_{h,\varepsilon_j} f j := f h , ε j be the mountain given by Lemma 4.15 , defined on [ 0 , ℓ j ] [0,\ell_j] [ 0 , ℓ j ] with ℓ j : = log ( 1 / ε j ) \ell_j:=\log(1/\varepsilon_j) ℓ j := log ( 1/ ε j ) , and let x j ∈ [ 0 , ℓ j − log ( 1 / θ j ) ] x_j\in[0,\ell_j-\log(1/\theta_j)] x j ∈ [ 0 , ℓ j − log ( 1/ θ j )] be the point given by (M3) for θ = θ j ∈ [ ε j , 1 ] \theta=\theta_j\in[\varepsilon_j,1] θ = θ j ∈ [ ε j , 1 ] .
For j ≥ 1 j\geq1 j ≥ 1 and F ≥ 0 F\geq0 F ≥ 0 we also define a valley v j , F v_{j,F} v j , F on [ 0 , T j + F ] [0,T_j+F] [ 0 , T j + F ] , where T j : = Δ j + Δ j ′ T_j:=\Delta_j+\Delta'_j T j := Δ j + Δ j ′ with
Δ j : = log h ( ε j ) β j ≥ 0 and Δ j ′ : = log 1 − β j 1 − h ( ε j + 1 ) ≥ 0 , \Delta_j:=\log\frac{h(\varepsilon_j)}{\beta_j}\geq0\qquad\text{and}\qquad \Delta'_j:=\log\frac{1-\beta_j}{1-h(\varepsilon_{j+1})}\geq0, Δ j := log β j h ( ε j ) ≥ 0 and Δ j ′ := log 1 − h ( ε j + 1 ) 1 − β j ≥ 0 , by
v j , F ( t ) : = { h ( ε j ) e − t if t ∈ [ 0 , Δ j ] , β j if t ∈ [ Δ j , Δ j + F ] , 1 − ( 1 − β j ) e − ( t − Δ j − F ) if t ∈ [ Δ j + F , T j + F ] . v_{j,F}(t):=
\begin{cases}
h(\varepsilon_j)\,e^{-t} & \text{if } t\in[0,\Delta_j],\\
\beta_j & \text{if } t\in[\Delta_j,\Delta_j+F],\\
1-(1-\beta_j)\,e^{-(t-\Delta_j-F)} & \text{if } t\in[\Delta_j+F,\,T_j+F].
\end{cases} v j , F ( t ) := ⎩ ⎨ ⎧ h ( ε j ) e − t β j 1 − ( 1 − β j ) e − ( t − Δ j − F ) if t ∈ [ 0 , Δ j ] , if t ∈ [ Δ j , Δ j + F ] , if t ∈ [ Δ j + F , T j + F ] . The function v j , F v_{j,F} v j , F is continuous, v j , F ( 0 ) = h ( ε j ) v_{j,F}(0)=h(\varepsilon_j) v j , F ( 0 ) = h ( ε j ) , v j , F ( T j + F ) = h ( ε j + 1 ) v_{j,F}(T_j+F)=h(\varepsilon_{j+1}) v j , F ( T j + F ) = h ( ε j + 1 ) , β j ≤ v j , F ≤ h ( ε j ) \beta_j\leq v_{j,F}\leq h(\varepsilon_j) β j ≤ v j , F ≤ h ( ε j ) , and D − v j , F ( t ) ∈ [ − v j , F ( t ) , 1 − v j , F ( t ) ] D_-v_{j,F}(t)\in[-v_{j,F}(t),1-v_{j,F}(t)] D − v j , F ( t ) ∈ [ − v j , F ( t ) , 1 − v j , F ( t )] for every t ∈ ( 0 , T j + F ] t\in(0,T_j+F] t ∈ ( 0 , T j + F ] .
Step 3. The function G G G . Let w 0 w_0 w 0 be the constant given by Lemma 4.4 for functions g ~ \tilde g g ~ with g ~ ( 0 ) = h ( ε 1 ) ∈ ( 0 , 1 ) \tilde g(0)=h(\varepsilon_1)\in(0,1) g ~ ( 0 ) = h ( ε 1 ) ∈ ( 0 , 1 ) . By (10) we can choose inductively numbers X 1 < X 2 < ⋯ X_1<X_2<\cdots X 1 < X 2 < ⋯ as follows. Let X 1 ≥ w 0 + x 1 X_1\geq w_0+x_1 X 1 ≥ w 0 + x 1 be such that
inf y ∈ [ X 1 , X 1 + log ( 1 / θ 1 ) ] S 1 ( y ) ≥ U ( θ 1 ) − 1 , \inf_{y\in[X_1,X_1+\log(1/\theta_1)]}S_1(y)\geq U(\theta_1)-1, y ∈ [ X 1 , X 1 + l o g ( 1/ θ 1 )] inf S 1 ( y ) ≥ U ( θ 1 ) − 1 , and put P 1 : = X 1 − x 1 ≥ w 0 P_1:=X_1-x_1\geq w_0 P 1 := X 1 − x 1 ≥ w 0 and R 1 : = P 1 + ℓ 1 R_1:=P_1+\ell_1 R 1 := P 1 + ℓ 1 . Once P j P_j P j and R j = P j + ℓ j R_j=P_j+\ell_j R j = P j + ℓ j have been defined, choose X j + 1 ≥ R j + T j + x j + 1 + 1 X_{j+1}\geq R_j+T_j+x_{j+1}+1 X j + 1 ≥ R j + T j + x j + 1 + 1 such that
inf y ∈ [ X j + 1 , X j + 1 + log ( 1 / θ j + 1 ) ] S 1 ( y ) ≥ U ( θ j + 1 ) − 1 j + 1 , \begin{equation}
\inf_{y\in[X_{j+1},X_{j+1}+\log(1/\theta_{j+1})]}S_1(y)\geq U(\theta_{j+1})-\frac{1}{j+1}, \tag*{(11)}
\end{equation} y ∈ [ X j + 1 , X j + 1 + l o g ( 1/ θ j + 1 )] inf S 1 ( y ) ≥ U ( θ j + 1 ) − j + 1 1 , (11) and set P j + 1 : = X j + 1 − x j + 1 P_{j+1}:=X_{j+1}-x_{j+1} P j + 1 := X j + 1 − x j + 1 , F j : = P j + 1 − R j − T j ≥ 1 F_j:=P_{j+1}-R_j-T_j\geq1 F j := P j + 1 − R j − T j ≥ 1 and R j + 1 : = P j + 1 + ℓ j + 1 R_{j+1}:=P_{j+1}+\ell_{j+1} R j + 1 := P j + 1 + ℓ j + 1 . Since x j ≤ ℓ j x_j\leq\ell_j x j ≤ ℓ j we have R j ≥ X j R_j\geq X_j R j ≥ X j , and therefore X j + 1 ≥ X j + 1 X_{j+1}\geq X_j+1 X j + 1 ≥ X j + 1 . In particular P 1 < R 1 < P 2 < R 2 < ⋯ P_1<R_1<P_2<R_2<\cdots P 1 < R 1 < P 2 < R 2 < ⋯ and P j → ∞ P_j\to\infty P j → ∞ . Define G : R → R G:\mathbb{R}\to\mathbb{R} G : R → R by
G ( x ) : = { h ( ε 1 ) if x ≤ P 1 , f j ( x − P j ) if x ∈ [ P j , R j ] , j ≥ 1 , v j , F j ( x − R j ) if x ∈ [ R j , P j + 1 ] , j ≥ 1. G(x):=
\begin{cases}
h(\varepsilon_1) & \text{if } x\leq P_1,\\
f_j(x-P_j) & \text{if } x\in[P_j,R_j],\ j\geq1,\\
v_{j,F_j}(x-R_j) & \text{if } x\in[R_j,P_{j+1}],\ j\geq1.
\end{cases} G ( x ) := ⎩ ⎨ ⎧ h ( ε 1 ) f j ( x − P j ) v j , F j ( x − R j ) if x ≤ P 1 , if x ∈ [ P j , R j ] , j ≥ 1 , if x ∈ [ R j , P j + 1 ] , j ≥ 1. By (M1) and the properties of the valleys, G G G is well defined and continuous, and G G G takes values in ( 0 , 1 ) (0,1) ( 0 , 1 ) , since 0 < β j ≤ G ≤ B < 1 0<\beta_j\leq G\leq B<1 0 < β j ≤ G ≤ B < 1 on [ R j , P j + 1 ] [R_j,P_{j+1}] [ R j , P j + 1 ] and h ( ε j ) ≤ G ≤ B h(\varepsilon_j)\leq G\leq B h ( ε j ) ≤ G ≤ B on [ P j , R j ] [P_j,R_j] [ P j , R j ] . Moreover, by (M2) and the corresponding property of the valleys, D − G ( x ) ∈ [ − G ( x ) , 1 − G ( x ) ] D_-G(x)\in[-G(x),1-G(x)] D − G ( x ) ∈ [ − G ( x ) , 1 − G ( x )] for every x ∈ R x\in\mathbb{R} x ∈ R , because the left Dini derivative at each point only depends on the piece of G G G lying to its left. By Lemma 4.6 with V = R V=\mathbb{R} V = R , we obtain G ∈ G ( 0 , 1 ) G\in\mathcal{G}(0,1) G ∈ G ( 0 , 1 ) .
We now collect three properties of G G G .
For every j ≥ 1 j\geq1 j ≥ 1 , G ( y ) ≥ h ( θ j ) G(y)\geq h(\theta_j) G ( y ) ≥ h ( θ j ) for all y ∈ [ X j , X j + log ( 1 / θ j ) ] y\in[X_j,X_j+\log(1/\theta_j)] y ∈ [ X j , X j + log ( 1/ θ j )] .
For every θ ∈ ( 0 , 1 ] \theta\in(0,1] θ ∈ ( 0 , 1 ] , lim sup x → ∞ inf y ∈ [ x , x + log ( 1 / θ ) ] G ( y ) ≤ h ( θ ) \displaystyle\limsup_{x\to\infty}\ \inf_{y\in[x,x+\log(1/\theta)]}G(y)\leq h(\theta) x → ∞ lim sup y ∈ [ x , x + l o g ( 1/ θ )] inf G ( y ) ≤ h ( θ ) .
lim inf x → ∞ G ( x ) = h ( 0 ) \displaystyle\liminf_{x\to\infty}G(x)=h(0) x → ∞ lim inf G ( x ) = h ( 0 ) .
Property (G1) follows from (M3), since [ X j , X j + log ( 1 / θ j ) ] = P j + [ x j , x j + log ( 1 / θ j ) ] ⊂ [ P j , R j ] [X_j,X_j+\log(1/\theta_j)]=P_j+[x_j,x_j+\log(1/\theta_j)]\subset[P_j,R_j] [ X j , X j + log ( 1/ θ j )] = P j + [ x j , x j + log ( 1/ θ j )] ⊂ [ P j , R j ] . For (G2), let θ ∈ ( 0 , 1 ] \theta\in(0,1] θ ∈ ( 0 , 1 ] , let x ≥ P 1 x\geq P_1 x ≥ P 1 and let W = [ x , x + log ( 1 / θ ) ] W=[x,x+\log(1/\theta)] W = [ x , x + log ( 1/ θ )] . If W ⊂ [ P j , R j ] W\subset[P_j,R_j] W ⊂ [ P j , R j ] for some j j j , then necessarily θ ≥ ε j \theta\geq\varepsilon_j θ ≥ ε j , and (M4) gives inf W G ≤ h ( θ ) \inf_W G\leq h(\theta) inf W G ≤ h ( θ ) . Otherwise, since W W W is an interval contained in [ P 1 , ∞ ) = ⋃ j ( [ P j , R j ] ∪ [ R j , P j + 1 ] ) [P_1,\infty)=\bigcup_{j}\bigl([P_j,R_j]\cup[R_j,P_{j+1}]\bigr) [ P 1 , ∞ ) = ⋃ j ( [ P j , R j ] ∪ [ R j , P j + 1 ] ) , it intersects some valley interval [ R j , P j + 1 ] [R_j,P_{j+1}] [ R j , P j + 1 ] with P j + 1 ≥ x P_{j+1}\geq x P j + 1 ≥ x , on which G ≤ h ( ε j ) G\leq h(\varepsilon_j) G ≤ h ( ε j ) . Let j ( x ) : = min { j : P j + 1 ≥ x } j(x):=\min\{j : P_{j+1}\geq x\} j ( x ) := min { j : P j + 1 ≥ x } , so that j ( x ) → ∞ j(x)\to\infty j ( x ) → ∞ as x → ∞ x\to\infty x → ∞ . Since h ( ε j ) h(\varepsilon_j) h ( ε j ) is non-increasing in j j j , in all cases
inf y ∈ W G ( y ) ≤ max { h ( θ ) , h ( ε j ( x ) ) } . \inf_{y\in W}G(y)\leq\max\{h(\theta),\,h(\varepsilon_{j(x)})\}. y ∈ W inf G ( y ) ≤ max { h ( θ ) , h ( ε j ( x ) )} . As j → ∞ j\to\infty j → ∞ , h ( ε j ) h(\varepsilon_j) h ( ε j ) decreases to inf t ∈ ( 0 , 1 ] h ( t ) ≤ h ( θ ) \inf_{t\in(0,1]}h(t)\leq h(\theta) inf t ∈ ( 0 , 1 ] h ( t ) ≤ h ( θ ) , and (G2) follows. For (G3), note that on [ P 1 , ∞ ) [P_1,\infty) [ P 1 , ∞ ) we have G ≥ min { h ( ε j ) , β j : j ≥ 1 } ≥ h ( 0 ) G\geq\min\{h(\varepsilon_j),\beta_j : j\geq1\}\geq h(0) G ≥ min { h ( ε j ) , β j : j ≥ 1 } ≥ h ( 0 ) , while G ( R j + Δ j ) = β j → h ( 0 ) G(R_j+\Delta_j)=\beta_j\to h(0) G ( R j + Δ j ) = β j → h ( 0 ) and R j → ∞ R_j\to\infty R j → ∞ .
Let g ~ ( y ) : = G ( y + w 0 ) \tilde g(y):=G(y+w_0) g ~ ( y ) := G ( y + w 0 ) . Then g ~ ∈ G ( 0 , 1 ) \tilde g\in\mathcal{G}(0,1) g ~ ∈ G ( 0 , 1 ) , g ~ ( 0 ) = G ( w 0 ) = h ( ε 1 ) \tilde g(0)=G(w_0)=h(\varepsilon_1) g ~ ( 0 ) = G ( w 0 ) = h ( ε 1 ) because w 0 ≤ P 1 w_0\leq P_1 w 0 ≤ P 1 , and g ~ \tilde g g ~ is not rapidly decreasing, since by (M1) we have max [ P j , R j ] G = B \max_{[P_j,R_j]}G=B max [ P j , R j ] G = B for every j j j , so that lim sup y → ∞ g ~ ( y ) = B > 0 \limsup_{y\to\infty}\tilde g(y)=B>0 lim sup y → ∞ g ~ ( y ) = B > 0 . Moreover κ w 0 ( g ~ ) ( x ) = G ( x ) \kappa_{w_0}(\tilde g)(x)=G(x) κ w 0 ( g ~ ) ( x ) = G ( x ) for all x ≥ w 0 x\geq w_0 x ≥ w 0 . By Lemma 4.4 (with w = w 0 w=w_0 w = w 0 ) there is a sequence r = ( r j ) ⊂ ( 0 , 1 / 2 ] \mathbf{r}=(r_j)\subset(0,1/2] r = ( r j ) ⊂ ( 0 , 1/2 ] such that, writing ρ 0 ( 2 ) : = 1 \rho^{(2)}_0:=1 ρ 0 ( 2 ) := 1 , ρ k ( 2 ) : = r 1 ⋯ r k \rho^{(2)}_k:=r_1\cdots r_k ρ k ( 2 ) := r 1 ⋯ r k and S 2 : = S ρ ( 2 ) S_2:=S_{\rho^{(2)}} S 2 := S ρ ( 2 ) , we have
∣ S 2 ( x ) − G ( x ) ∣ ≤ log 2 exp ( − x ) for all x ≥ w 0 . \begin{equation}
|S_2(x)-G(x)|\leq\log2\,\exp(-x)\qquad\text{for all }x\geq w_0 . \tag*{(12)}
\end{equation} ∣ S 2 ( x ) − G ( x ) ∣ ≤ log 2 exp ( − x ) for all x ≥ w 0 . (12) Step 4. The combined scales. Define
ρ k : = min { ρ k ( 1 ) , ρ k ( 2 ) } ( k ≥ 0 ) , \rho_k:=\min\{\rho^{(1)}_k,\rho^{(2)}_k\}\qquad(k\geq0), ρ k := min { ρ k ( 1 ) , ρ k ( 2 ) } ( k ≥ 0 ) , so that ρ 0 = C − 1 < 1 \rho_0=C^{-1}<1 ρ 0 = C − 1 < 1 . If ρ k − 1 = ρ k − 1 ( i ) \rho_{k-1}=\rho^{(i)}_{k-1} ρ k − 1 = ρ k − 1 ( i ) with i ∈ { 1 , 2 } i\in\{1,2\} i ∈ { 1 , 2 } , then ρ k ≤ ρ k ( i ) ≤ ρ k − 1 ( i ) / 2 = ρ k − 1 / 2 \rho_k\leq\rho^{(i)}_k\leq\rho^{(i)}_{k-1}/2=\rho_{k-1}/2 ρ k ≤ ρ k ( i ) ≤ ρ k − 1 ( i ) /2 = ρ k − 1 /2 , because s k ( a ) / s k − 1 ( a ) ≤ 1 / κ 1 < 1 / 2 s_k(a)/s_{k-1}(a)\leq1/\kappa_1<1/2 s k ( a ) / s k − 1 ( a ) ≤ 1/ κ 1 < 1/2 and r k ≤ 1 / 2 r_k\leq1/2 r k ≤ 1/2 . Hence
ρ k ≤ ρ k − 1 / 2 for all k ≥ 1 , \begin{equation}
\rho_k\leq\rho_{k-1}/2\qquad\text{for all }k\geq1, \tag*{(13)}
\end{equation} ρ k ≤ ρ k − 1 /2 for all k ≥ 1 , (13) and in particular ρ \rho ρ is strictly decreasing and converges to 0 0 0 . Since both ρ ( 1 ) \rho^{(1)} ρ ( 1 ) and ρ ( 2 ) \rho^{(2)} ρ ( 2 ) are decreasing, for ϕ ∈ ( 0 , ρ 0 ) \phi\in(0,\rho_0) ϕ ∈ ( 0 , ρ 0 ) we have ρ k ≤ ϕ \rho_k\leq\phi ρ k ≤ ϕ if and only if ρ k ( 1 ) ≤ ϕ \rho^{(1)}_k\leq\phi ρ k ( 1 ) ≤ ϕ or ρ k ( 2 ) ≤ ϕ \rho^{(2)}_k\leq\phi ρ k ( 2 ) ≤ ϕ , and therefore k ρ ( ϕ ) = min { k ρ ( 1 ) ( ϕ ) , k ρ ( 2 ) ( ϕ ) } k_\rho(\phi)=\min\{k_{\rho^{(1)}}(\phi),k_{\rho^{(2)}}(\phi)\} k ρ ( ϕ ) = min { k ρ ( 1 ) ( ϕ ) , k ρ ( 2 ) ( ϕ )} . By (2) ,
s ρ = min { s ρ ( 1 ) , s ρ ( 2 ) } on ( 0 , ρ 0 ) , \begin{equation}
s_\rho=\min\{s_{\rho^{(1)}},s_{\rho^{(2)}}\}\ \text{ on }(0,\rho_0), \tag*{(14)}
\end{equation} s ρ = min { s ρ ( 1 ) , s ρ ( 2 ) } on ( 0 , ρ 0 ) , (14) that is, S ρ = min { S 1 , S 2 } S_\rho=\min\{S_1,S_2\} S ρ = min { S 1 , S 2 } for all sufficiently large x x x .
We claim that
Φ ( θ ) : = lim sup x → ∞ inf y ∈ [ x , x + log ( 1 / θ ) ] S ρ ( y ) = h ( θ ) for all θ ∈ ( 0 , 1 ] , lim inf x → ∞ S ρ ( x ) = h ( 0 ) . \begin{equation}
\begin{split}
\Phi(\theta):=\limsup_{x\to\infty}\ \inf_{y\in[x,x+\log(1/\theta)]}S_\rho(y)&=h(\theta)\qquad\text{for all }\theta\in(0,1],\\
\liminf_{x\to\infty}S_\rho(x)&=h(0).
\end{split} \tag*{(15)}
\end{equation} Φ ( θ ) := x → ∞ lim sup y ∈ [ x , x + l o g ( 1/ θ )] inf S ρ ( y ) x → ∞ lim inf S ρ ( x ) = h ( θ ) for all θ ∈ ( 0 , 1 ] , = h ( 0 ) . (15) Since S ρ ≤ S 2 ≤ G + log 2 e − x S_\rho\leq S_2\leq G+\log2\,e^{-x} S ρ ≤ S 2 ≤ G + log 2 e − x by (14) and (12) , property (G2) gives Φ ( θ ) ≤ h ( θ ) \Phi(\theta)\leq h(\theta) Φ ( θ ) ≤ h ( θ ) . For the reverse inequality, fix q ∈ Q ∩ ( 0 , 1 ) q\in\mathbb{Q}\cap(0,1) q ∈ Q ∩ ( 0 , 1 ) and let j j j be any of the infinitely many indices with θ j = q \theta_j=q θ j = q . On the window [ X j , X j + log ( 1 / q ) ] [X_j,X_j+\log(1/q)] [ X j , X j + log ( 1/ q )] we have S 1 ≥ U ( q ) − 1 / j ≥ h ( q ) − 1 / j S_1\geq U(q)-1/j\geq h(q)-1/j S 1 ≥ U ( q ) − 1/ j ≥ h ( q ) − 1/ j by (11) and the hypothesis h ≤ U h\leq U h ≤ U , and S 2 ≥ G − log 2 e − X j ≥ h ( q ) − log 2 e − X j S_2\geq G-\log2\,e^{-X_j}\geq h(q)-\log2\,e^{-X_j} S 2 ≥ G − log 2 e − X j ≥ h ( q ) − log 2 e − X j by (G1) and (12) . Hence
inf y ∈ [ X j , X j + log ( 1 / q ) ] S ρ ( y ) ≥ h ( q ) − 1 j − log 2 e − X j , \inf_{y\in[X_j,X_j+\log(1/q)]}S_\rho(y)\geq h(q)-\frac1j-\log2\,e^{-X_j}, y ∈ [ X j , X j + l o g ( 1/ q )] inf S ρ ( y ) ≥ h ( q ) − j 1 − log 2 e − X j , and since X j → ∞ X_j\to\infty X j → ∞ along these indices, Φ ( q ) ≥ h ( q ) \Phi(q)\geq h(q) Φ ( q ) ≥ h ( q ) . Now Φ \Phi Φ is non-decreasing in θ \theta θ , because the windows become shorter as θ \theta θ increases. Thus, if θ ∈ ( 0 , 1 ] \theta\in(0,1] θ ∈ ( 0 , 1 ] and q n ∈ Q ∩ ( 0 , 1 ) q_n\in\mathbb{Q}\cap(0,1) q n ∈ Q ∩ ( 0 , 1 ) increase to θ \theta θ , then Φ ( θ ) ≥ Φ ( q n ) ≥ h ( q n ) → h ( θ ) \Phi(\theta)\geq\Phi(q_n)\geq h(q_n)\to h(\theta) Φ ( θ ) ≥ Φ ( q n ) ≥ h ( q n ) → h ( θ ) by the continuity of h h h on ( 0 , 1 ] (0,1] ( 0 , 1 ] . This proves the first part of (15) . For the second part, (14) , (10) , (G3) and (12) give
lim inf x → ∞ S ρ ( x ) = min { lim inf x → ∞ S 1 ( x ) , lim inf x → ∞ S 2 ( x ) } = min { U ( 0 ) , h ( 0 ) } = h ( 0 ) , \liminf_{x\to\infty}S_\rho(x)=\min\Bigl\{\liminf_{x\to\infty}S_1(x),\ \liminf_{x\to\infty}S_2(x)\Bigr\}=\min\{U(0),h(0)\}=h(0), x → ∞ lim inf S ρ ( x ) = min { x → ∞ lim inf S 1 ( x ) , x → ∞ lim inf S 2 ( x ) } = min { U ( 0 ) , h ( 0 )} = h ( 0 ) , where we used the hypothesis h ( 0 ) ≤ U ( 0 ) h(0)\leq U(0) h ( 0 ) ≤ U ( 0 ) .
Step 5. A complementary Cantor set with scales ρ \rho ρ . For m ≥ 0 m\geq0 m ≥ 0 let
π m : = ∏ i = 1 m ( 1 − 1 ( i + 1 ) 2 ) = m + 2 2 ( m + 1 ) ∈ [ 1 2 , 1 ] and ρ m ′ : = π m ρ m . \pi_m:=\prod_{i=1}^{m}\Bigl(1-\frac{1}{(i+1)^2}\Bigr)=\frac{m+2}{2(m+1)}\in\Bigl[\frac12,1\Bigr]\qquad\text{and}\qquad\rho'_m:=\pi_m\rho_m . π m := i = 1 ∏ m ( 1 − ( i + 1 ) 2 1 ) = 2 ( m + 1 ) m + 2 ∈ [ 2 1 , 1 ] and ρ m ′ := π m ρ m . By (13) , ρ m ′ / ρ m − 1 ′ ≤ ( 1 − ( m + 1 ) − 2 ) / 2 < 1 / 2 \rho'_m/\rho'_{m-1}\leq\bigl(1-(m+1)^{-2}\bigr)/2<1/2 ρ m ′ / ρ m − 1 ′ ≤ ( 1 − ( m + 1 ) − 2 ) /2 < 1/2 , so the numbers
η m : = ρ m − 1 ′ − 2 ρ m ′ ( m ≥ 1 ) \eta_m:=\rho'_{m-1}-2\rho'_m\qquad(m\geq1) η m := ρ m − 1 ′ − 2 ρ m ′ ( m ≥ 1 ) satisfy 0 < η m < ρ m − 1 ′ 0<\eta_m<\rho'_{m-1} 0 < η m < ρ m − 1 ′ . Moreover 2 m ρ m ′ ≤ 2 m ρ m ( 1 ) = C − 1 x 2 m ( a ) → 0 2^m\rho'_m\leq2^m\rho^{(1)}_m=C^{-1}x_{2^m}(a)\to0 2 m ρ m ′ ≤ 2 m ρ m ( 1 ) = C − 1 x 2 m ( a ) → 0 , so a telescoping argument gives
∑ m > k 2 m − k − 1 η m = ρ k ′ for all k ≥ 0. \begin{equation}
\sum_{m>k}2^{m-k-1}\eta_m=\rho'_k\qquad\text{for all }k\geq0 . \tag*{(16)}
\end{equation} m > k ∑ 2 m − k − 1 η m = ρ k ′ for all k ≥ 0. (16) We now select the gaps. For p ≥ 0 p\geq0 p ≥ 0 let Λ p : = { n ∈ N : 2 p ≤ n < 2 p + 1 } \Lambda_p:=\{n\in\mathbb{N} : 2^p\leq n<2^{p+1}\} Λ p := { n ∈ N : 2 p ≤ n < 2 p + 1 } , and for m ≥ 1 m\geq1 m ≥ 1 let p ( m ) : = min { p ≥ 0 : a 2 p ≤ η m } p(m):=\min\{p\geq0 : a_{2^p}\leq\eta_m\} p ( m ) := min { p ≥ 0 : a 2 p ≤ η m } . By Lemma 4.14 (a) and the choice of C C C ,
η m < ρ m − 1 ′ ≤ ρ m − 1 ( 1 ) = C − 1 s m − 1 ( a ) ≤ C − 1 κ 2 s m ( a ) ≤ C − 1 κ 2 κ 1 κ 1 − 2 a 2 m = a 2 m . \begin{split}
\eta_m<\rho'_{m-1}\leq\rho^{(1)}_{m-1}&=C^{-1}s_{m-1}(a)\leq C^{-1}\kappa_2\,s_m(a)\\
&\leq C^{-1}\kappa_2\frac{\kappa_1}{\kappa_1-2}\,a_{2^m}=a_{2^m}.
\end{split} η m < ρ m − 1 ′ ≤ ρ m − 1 ( 1 ) = C − 1 s m − 1 ( a ) ≤ C − 1 κ 2 s m ( a ) ≤ C − 1 κ 2 κ 1 − 2 κ 1 a 2 m = a 2 m . Since a a a is non-increasing, a 2 p ≥ a 2 m > η m a_{2^p}\geq a_{2^m}>\eta_m a 2 p ≥ a 2 m > η m for p ≤ m p\leq m p ≤ m , hence p ( m ) ≥ m + 1 p(m)\geq m+1 p ( m ) ≥ m + 1 . If n ∈ Λ p ( m ) n\in\Lambda_{p(m)} n ∈ Λ p ( m ) , then a n ≤ a 2 p ( m ) ≤ η m a_n\leq a_{2^{p(m)}}\leq\eta_m a n ≤ a 2 p ( m ) ≤ η m , and by Lemma 4.14 (b) and the minimality of p ( m ) p(m) p ( m ) ,
a n ≥ a 2 p ( m ) + 1 ≥ c a − 2 a 2 p ( m ) − 1 > c a − 2 η m . a_n\geq a_{2^{p(m)+1}}\geq c_a^{-2}\,a_{2^{p(m)-1}}>c_a^{-2}\,\eta_m . a n ≥ a 2 p ( m ) + 1 ≥ c a − 2 a 2 p ( m ) − 1 > c a − 2 η m . Thus
c a − 2 η m < a n ≤ η m for all n ∈ Λ p ( m ) . \begin{equation}
c_a^{-2}\,\eta_m<a_n\leq\eta_m\qquad\text{for all }n\in\Lambda_{p(m)} . \tag*{(17)}
\end{equation} c a − 2 η m < a n ≤ η m for all n ∈ Λ p ( m ) . (17) For each p ≥ 1 p\geq1 p ≥ 1 , the levels m m m with p ( m ) = p p(m)=p p ( m ) = p satisfy m ≤ p − 1 m\leq p-1 m ≤ p − 1 , so
∑ m : p ( m ) = p 2 m − 1 ≤ ∑ m = 1 p − 1 2 m − 1 < 2 p − 1 = # Λ p 2 . \sum_{m : p(m)=p}2^{m-1}\leq\sum_{m=1}^{p-1}2^{m-1}<2^{p-1}=\frac{\#\Lambda_p}{2}. m : p ( m ) = p ∑ 2 m − 1 ≤ m = 1 ∑ p − 1 2 m − 1 < 2 p − 1 = 2 # Λ p . Therefore we can choose pairwise disjoint sets A m ⊂ Λ p ( m ) A_m\subset\Lambda_{p(m)} A m ⊂ Λ p ( m ) , m ≥ 1 m\geq1 m ≥ 1 , with # A m = 2 m − 1 \#A_m=2^{m-1} # A m = 2 m − 1 , in such a way that at least 2 p − 1 2^{p-1} 2 p − 1 indices of each Λ p \Lambda_p Λ p , p ≥ 1 p\geq1 p ≥ 1 , do not belong to A : = ⋃ m A m A:=\bigcup_mA_m A := ⋃ m A m . For each m ≥ 1 m\geq1 m ≥ 1 fix a bijection between { 1 , 2 } m − 1 \{1,2\}^{m-1} { 1 , 2 } m − 1 and A m A_m A m , and for σ ∈ { 1 , 2 } m − 1 \sigma\in\{1,2\}^{m-1} σ ∈ { 1 , 2 } m − 1 let b σ : = a n b_\sigma:=a_n b σ := a n , where n ∈ A m n\in A_m n ∈ A m corresponds to σ \sigma σ .
For σ ∈ Σ ∗ \sigma\in\Sigma^* σ ∈ Σ ∗ let ℓ ( σ ) : = ∑ τ ⪰ σ b τ \ell(\sigma):=\sum_{\tau\succeq\sigma}b_\tau ℓ ( σ ) := ∑ τ ⪰ σ b τ . If ∣ σ ∣ = k |\sigma|=k ∣ σ ∣ = k , then for each m > k m>k m > k there are exactly 2 m − 1 − k 2^{m-1-k} 2 m − 1 − k words τ ⪰ σ \tau\succeq\sigma τ ⪰ σ with ∣ τ ∣ = m − 1 |\tau|=m-1 ∣ τ ∣ = m − 1 , and by (17) , (16) and π k ≥ 1 / 2 \pi_k\geq1/2 π k ≥ 1/2 ,
ρ k 2 c a 2 ≤ ρ k ′ c a 2 ≤ ℓ ( σ ) ≤ ρ k ′ ≤ ρ k . \begin{equation}
\frac{\rho_k}{2c_a^2}\leq\frac{\rho'_k}{c_a^{2}}\leq\ell(\sigma)\leq\rho'_k\leq\rho_k . \tag*{(18)}
\end{equation} 2 c a 2 ρ k ≤ c a 2 ρ k ′ ≤ ℓ ( σ ) ≤ ρ k ′ ≤ ρ k . (18) Note also that ℓ ( σ ) = ℓ ( σ 1 ) + b σ + ℓ ( σ 2 ) \ell(\sigma)=\ell(\sigma1)+b_\sigma+\ell(\sigma2) ℓ ( σ ) = ℓ ( σ 1 ) + b σ + ℓ ( σ 2 ) . We define intervals recursively by I ∅ : = [ 0 , ℓ ( ∅ ) ] I_\varnothing:=[0,\ell(\varnothing)] I ∅ := [ 0 , ℓ ( ∅ )] and, if I σ = [ u , u + ℓ ( σ ) ] I_\sigma=[u,u+\ell(\sigma)] I σ = [ u , u + ℓ ( σ )] ,
I σ 1 : = [ u , u + ℓ ( σ 1 ) ] and I σ 2 : = [ u + ℓ ( σ 1 ) + b σ , u + ℓ ( σ ) ] , I_{\sigma1}:=[u,\,u+\ell(\sigma1)]\qquad\text{and}\qquad I_{\sigma2}:=[u+\ell(\sigma1)+b_\sigma,\,u+\ell(\sigma)], I σ 1 := [ u , u + ℓ ( σ 1 )] and I σ 2 := [ u + ℓ ( σ 1 ) + b σ , u + ℓ ( σ )] , and we let J σ : = ( u + ℓ ( σ 1 ) , u + ℓ ( σ 1 ) + b σ ) J_\sigma:=\bigl(u+\ell(\sigma1),\,u+\ell(\sigma1)+b_\sigma\bigr) J σ := ( u + ℓ ( σ 1 ) , u + ℓ ( σ 1 ) + b σ ) be the open interval between them. By (18) , { I σ } \{I_\sigma\} { I σ } is a ( ρ , 2 c a 2 ) (\rho,2c_a^2) ( ρ , 2 c a 2 ) -homogeneous binary Cantor scheme. Let M M M be its limit set. A point of I ∅ I_\varnothing I ∅ which does not belong to M M M lies in some I σ I_\sigma I σ but in none of its two children, so it lies in J σ J_\sigma J σ . Conversely J σ ∩ M = ∅ J_\sigma\cap M=\varnothing J σ ∩ M = ∅ . Hence
I ∅ ∖ M = ⋃ σ ∈ Σ ∗ J σ , I_\varnothing\setminus M=\bigcup_{\sigma\in\Sigma^*}J_\sigma, I ∅ ∖ M = σ ∈ Σ ∗ ⋃ J σ , a disjoint union of open intervals whose lengths are the numbers a n a_n a n , n ∈ A n\in A n ∈ A , each of them used exactly once. Furthermore ℓ ( ∅ ) = ∑ n ∈ A a n < 1 \ell(\varnothing)=\sum_{n\in A}a_n<1 ℓ ( ∅ ) = ∑ n ∈ A a n < 1 . By Lemma 4.12 and (15) ,
\begin{equation}
\overline{\mbox{\rm dim}}_{\,\theta}M=h(\theta)\ \text{ for all }\theta\in(0,1]\qquad\text{and}\qquad\dim_{\rm H}M=h(0). \tag*{(19)}
\end{equation}
Step 6. The countable part and conclusion. Let a ′ = { a i ′ } a'=\{a'_i\} a ′ = { a i ′ } be the non-increasing rearrangement of { a n : n ∉ A } \{a_n : n\notin A\} { a n : n ∈ / A } , which is an infinite sequence with ∑ i a i ′ = 1 − ℓ ( ∅ ) \sum_ia'_i=1-\ell(\varnothing) ∑ i a i ′ = 1 − ℓ ( ∅ ) . Define
y k : = ℓ ( ∅ ) + ∑ i ≥ k a i ′ ( k ≥ 1 ) , D ′ : = { y k : k ≥ 1 } ∪ { ℓ ( ∅ ) } , y_k:=\ell(\varnothing)+\sum_{i\geq k}a'_i\quad(k\geq1),\qquad D':=\{y_k : k\geq1\}\cup\{\ell(\varnothing)\}, y k := ℓ ( ∅ ) + i ≥ k ∑ a i ′ ( k ≥ 1 ) , D ′ := { y k : k ≥ 1 } ∪ { ℓ ( ∅ )} , and E : = M ∪ D ′ E:=M\cup D' E := M ∪ D ′ . Then y 1 = 1 y_1=1 y 1 = 1 , y k y_k y k decreases to ℓ ( ∅ ) \ell(\varnothing) ℓ ( ∅ ) , and E E E is a closed subset of [ 0 , 1 ] [0,1] [ 0 , 1 ] such that
[ 0 , 1 ] ∖ E = ⋃ σ ∈ Σ ∗ J σ ∪ ⋃ k ≥ 1 ( y k + 1 , y k ) , [0,1]\setminus E=\bigcup_{\sigma\in\Sigma^*}J_\sigma\ \cup\ \bigcup_{k\geq1}(y_{k+1},y_k), [ 0 , 1 ] ∖ E = σ ∈ Σ ∗ ⋃ J σ ∪ k ≥ 1 ⋃ ( y k + 1 , y k ) , which is a disjoint union of open intervals whose lengths are exactly the terms of a a a . Thus E ∈ C a E\in\mathcal{C}_a E ∈ C a . The set D ′ D' D ′ is a translate of D a ′ D_{a'} D a ′ , so by Lemma 3.2 , \overline{\mbox{\rm dim}}_{\,\theta}D'=\varphi_{B'}(\theta) with B':=\overline{\mbox{\rm dim}}_{\rm B}\,D'\leq\overline{\mbox{\rm dim}}_{\rm B}\,E=B . Since φ B ( θ ) \varphi_B(\theta) φ B ( θ ) is non-decreasing in B B B , (9) gives \overline{\mbox{\rm dim}}_{\,\theta}D'\leq\varphi_B(\theta)\leq h(\theta) . By the finite stability of the upper intermediate dimensions (see [35 ] ) and (19) , for every θ ∈ ( 0 , 1 ] \theta\in(0,1] θ ∈ ( 0 , 1 ] ,
\overline{\mbox{\rm dim}}_{\,\theta}E=\max\{\overline{\mbox{\rm dim}}_{\,\theta}M,\ \overline{\mbox{\rm dim}}_{\,\theta}D'\}=\max\{h(\theta),\varphi_{B'}(\theta)\}=h(\theta).
Finally, for θ = 0 \theta=0 θ = 0 , dim H E = max { dim H M , dim H D ′ } = max { h ( 0 ) , 0 } = h ( 0 ) \dim_{\rm H}E=\max\{\dim_{\rm H}M,\dim_{\rm H}D'\}=\max\{h(0),0\}=h(0) dim H E = max { dim H M , dim H D ′ } = max { h ( 0 ) , 0 } = h ( 0 ) . This completes the proof. ◻
As a consequence, we obtain a complete description of the family \mathcal{F}^U(a)=\{\theta\mapsto\overline{\mbox{\rm dim}}_{\,\theta}E : E\in\mathcal{C}_a\} .
Corollary 4.18. Let a = { a n } a=\{a_n\} a = { a n } be a positive, non-increasing sequence with ∑ n a n = 1 \sum_na_n=1 ∑ n a n = 1 satisfying (H) . Then F U ( a ) \mathcal{F}^U(a) F U ( a ) is exactly the set of functions h ∈ H ( 0 , 1 ) h\in\mathcal{H}(0,1) h ∈ H ( 0 , 1 ) such that h(1)=\overline{\mbox{\rm dim}}_{\rm B}\,C_a\qquad\text{and}\qquad h(\theta)\leq\overline{\mbox{\rm dim}}_{\,\theta}C_a\quad\text{for all }\theta\in[0,1]. Moreover, every h ∈ F U ( a ) h\in\mathcal{F}^U(a) h ∈ F U ( a ) satisfies \overline{\mbox{\rm dim}}_{\,\theta}D_a\leq h(\theta)\leq\overline{\mbox{\rm dim}}_{\,\theta}C_a for all θ ∈ [ 0 , 1 ] \theta\in[0,1] θ ∈ [ 0 , 1 ] , and if h ( θ 0 ) = dim ‾ θ 0 D a h(\theta_0)=\overline{\dim}_{\,\theta_0}D_a h ( θ 0 ) = dim θ 0 D a for some θ 0 ∈ ( 0 , 1 ] \theta_0\in(0,1] θ 0 ∈ ( 0 , 1 ] , then h(\theta)=\overline{\mbox{\rm dim}}_{\,\theta}D_a for all θ ∈ [ θ 0 , 1 ] \theta\in[\theta_0,1] θ ∈ [ θ 0 , 1 ] .
Proof. The inclusion ⊇ \supseteq ⊇ is Theorem 4.17 with λ = 0 \lambda=0 λ = 0 and α = 1 \alpha=1 α = 1 . Conversely, let E ∈ C a E\in\mathcal{C}_a E ∈ C a and h(\theta)=\overline{\mbox{\rm dim}}_{\,\theta}E . By Theorem 4.2 and Remark 4.3 , h ∈ H ( dim L E , dim A E ) ⊆ H ( 0 , 1 ) h\in\mathcal{H}(\dim_L E,\dim_A E)\subseteq\mathcal{H}(0,1) h ∈ H ( dim L E , dim A E ) ⊆ H ( 0 , 1 ) . Moreover h(1)=\overline{\mbox{\rm dim}}_{\rm B}\,E=\overline{\mbox{\rm dim}}_{\rm B}\,C_a . Since (H) implies lim n → ∞ ( s n / s n + 1 ) 1 / n = 1 \lim_{n\to\infty}(s_n/s_{n+1})^{1/n}=1 lim n → ∞ ( s n / s n + 1 ) 1/ n = 1 , Lemma 3.2 gives h(\theta)\leq\overline{\mbox{\rm dim}}_{\,\theta}C_a for θ ∈ ( 0 , 1 ] \theta\in(0,1] θ ∈ ( 0 , 1 ] , and for θ = 0 \theta=0 θ = 0 this inequality is dim H E ≤ dim H C a \dim_{\rm H}E\leq\dim_{\rm H}C_a dim H E ≤ dim H C a , which holds by Proposition 3.1 . The last statement follows from Lemma 4.14 (c) and Proposition 4.8 , since \overline{\mbox{\rm dim}}_{\,\theta}D_a=\varphi_B(\theta) with B=\overline{\mbox{\rm dim}}_{\rm B}\,C_a\in(0,1) . ◻
In particular, the functions h s h_s h s , s > 1 s>1 s > 1 , of Remark 4.9 form only a small part of F U ( a ) \mathcal{F}^U(a) F U ( a ) . We conclude with two remarks on the construction.
Further Work
In [26 ] , the authors characterized the functions which arise as the intermediate dimensions of bounded subsets of R n \mathbb{R}^n R n (see Theorem 4.2 ). Theorem 4.17 and Corollary 4.18 provide the analogous characterization for the upper intermediate dimensions of complementary sets, under hypothesis (H) . The corresponding problem for the lower intermediate dimensions remains open.
Let us briefly describe what is known in that case. Since \underline{\mbox{\rm dim}}_{\,\theta}C_a=\underline{\mbox{\rm dim}}_{\rm B}\,C_a for all θ \theta θ , every function h(\theta)=\underline{\mbox{\rm dim}}_{\,\theta}E with E ∈ C a E\in\mathcal{C}_a E ∈ C a satisfies h(\theta)\leq h(1)=\underline{\mbox{\rm dim}}_{\rm B}\,C_a . By Theorem 4.2 and Remark 4.3 , h ∈ H ( 0 , 1 ) h\in\mathcal{H}(0,1) h ∈ H ( 0 , 1 ) , and if 0<\underline{\mbox{\rm dim}}_{\rm B}\,C_a<1 , Proposition 4.8 applies to h h h with B B B replaced by \underline{\mbox{\rm dim}}_{\rm B}\,C_a , so that \underline{\mbox{\rm dim}}_{\,\theta}D_a\leq h(\theta) with the same rigidity property. The main difficulty is that the lower intermediate dimensions are not finitely stable, so the gluing argument of Step 6 in the proof of Theorem 4.17 does not control the lower intermediate dimensions of the union. The following questions arise naturally.
Question 1. Let a = { a n } a=\{a_n\} a = { a n } be a non-increasing summable sequence satisfying 2 < inf n s n ( a ) s n + 1 ( a ) ≤ sup n s n ( a ) s n + 1 ( a ) < ∞ . 2<\inf_n\frac{s_n(a)}{s_{n+1}(a)}\leq \sup_n\frac{s_n(a)}{s_{n+1}(a)}<\infty . 2 < inf n s n + 1 ( a ) s n ( a ) ≤ sup n s n + 1 ( a ) s n ( a ) < ∞. Is it true that \mathcal{F}^L(a):=\bigl\{\,\theta\mapsto \underline{\mbox{\rm dim}}_{\,\theta}E : E\in\mathcal{C}_a\,\bigr\}=\bigl\{h\in\mathcal{H}(0,1) : h(1)=\underline{\mbox{\rm dim}}_{\rm B}\,C_a\bigr\}? More generally, which pairs of functions ( h 1 , h 2 ) (h_1,h_2) ( h 1 , h 2 ) arise as (\underline{\mbox{\rm dim}}_{\,\theta}E,\overline{\mbox{\rm dim}}_{\,\theta}E) for some E ∈ C a E\in\mathcal{C}_a E ∈ C a ?
Funding
This work was supported by FIS-2023-02725, singular structures in the geometry of measures: decompositions, rigidity and rectifiability, CUP E53C25001800001. The author was additionally partially supported by grants PICT 2022-4875 (ANPCyT), PIP 202287/22 (CONICET), and PROICO 3-0720 “Análisis Real y Funcional. Ec. Diferenciales”.