1: \section{Partition Zeta Functions}\label{pzf}
2:
3: The {\it partition zeta functions} described in this section were
4: first defined in \cite{Rock} and will be discussed further in
5: \cite{LapRock}. These functions are defined for any Borel measure on
6: the unit interval along with a family of partitions and are
7: parameterized by the regularity values attained by such measures.
8: The construction is very similar to that of the geometric zeta
9: function in the sense that the functions are series whose terms come
10: from a sequence of properly defined lengths. For geometric zeta
11: functions the terms are derived from the lengths of the disjoint
12: intervals of a given fractal string. However, for partition zeta
13: functions the terms are derived from a sequence of lengths from a
14: family of partitions which exhibit the same regularity.
15:
16: \ndnt The families of partitions we consider satisfy certain
17: requirements and occur quite naturally in the construction of
18: multinomial measures such as the binomial measure. Consider an
19: ordered family of partitions $\mathfrak{P} =
20: \{\mathcal{P}_n\}_{n=1}^{\infty}$ of $[0,1]$, each of which splits
21: the unit interval into finitely many subintervals. The order is
22: given by the relation $\mathcal{P}_n \succ \mathcal{P}_{n+1}$ taken
23: to mean that each of the intervals $P_{n+1}^k$ which comprise the
24: partition $\mathcal{P}_{n+1}$ is a subinterval of some interval in
25: $\mathcal{P}_n$\footnote{To avoid trivial situations, we further
26: assume that the mesh of the sequence of partitions $\mathcal{P}_n$
27: tends to zero.}.
28:
29: \begin{dfn}\label{def:pzf}
30: For a measure $\mu$ on the interval $[0,1]$ with an ordered family
31: of partitions $\mathfrak{P}$, the \underline{partition zeta function
32: with regularity $\alpha$} is
33: \[
34: \zeta^{\mu}_{\mathfrak{P}}(\alpha,s) =
35: \sum_{n=1}^{\infty}\sum_{A(P^k_n)=\alpha}|P^{k}_{n}|^s,
36: \]
37: where the inner sum is taken over the intervals $P^k_n$ in the
38: partition $\mathcal{P}_n$ which have regularity $A(P^k_n)=\alpha \in
39: [-\infty,\infty]$, and $\textnormal{Re}(s)$ is large enough.
40: \end{dfn}
41:
42:
43: \ndnt Recall the construction of the measure $\mu$ from Section 2.
44: The breakdown of mass and length readily defines a {\it natural
45: family of partitions} $\mathfrak{P}$ for this measure $\mu$ as
46: simply the closed intervals and their complements in the
47: construction of the Cantor set. The mass breakdown allows for the
48: separation of the individual intervals in $\mathfrak{P}$ into
49: collections according to their regularity, found with respect to the
50: measure $\mu$. At each stage, the intervals with the same regularity
51: $\alpha$ have multiplicities given by binomial coefficients. In
52: turn, their lengths and multiplicities constitute the terms in the
53: definition of the partition zeta function with that regularity.
54:
55:
56: \ndnt To determine the intervals which have the same regularity,
57: note that the collection of all intervals from every partition in
58: the family $\mathfrak{P}$ is a countable set. Thus, the regularity
59: values attained on these intervals are a function of the ordered
60: pair of integers $(k_1,k_2)$ (which satisfy the properties mentioned
61: below) as follows:
62: \[
63: \alpha := \alpha(k_1, k_2) =
64: \frac{\log{(2^{nk_1}/3^{nk_2})}}{\log{(1/3^{nk_2})}} = 1 -
65: \frac{k_1}{k_2}\log_3{2},
66: \]
67: for all $n \in \mathbb{N}$ and where $k_1 \in \mathbb{N} \cup
68: \{0\}$, $k_2 \in \mathbb{N}$, $k_1 \leq k_2$ and $k_1$ and $k_2$ are
69: necessarily relatively prime (denoted $(k_1,k_2)=1$), except when
70: $k_1 = 0$ or $1$ and $k_2 = 1$. The integers $k_j$ relate to the
71: measure and regularity of intervals in that, roughly, $k_1$ is the
72: number of times an interval falls to the right and gets $2/3$ of the
73: mass after $k_2$ stages during the construction.
74:
75:
76: \ndnt The regularity value $\alpha(k_1,k_2)$ with fixed $(k_1,k_2)$
77: as above only occurs in the partitions $\mathcal{P}_{nk_2}$ for each
78: $n \in \mathbb{N}$, with multiplicity $\binom{nk_2}{nk_1}$. In
79: summation,
80: \[
81: \zeta^{\mu}_{\mathfrak{P}}(\alpha(k_1,k_2),s) =
82: \zeta^{\mu}_{\mathfrak{P}}(\alpha(k_2-k_1,k_2),s) =
83: \sum_{n=1}^{\infty} \binom{nk_2}{nk_1} 3^{-k_2ns}.
84: \]
85: See Figure 4 for the first several intervals with regularity
86: $\alpha{(1,2)}$.
87:
88:
89: \begin{figure}
90: \epsfysize=3.7cm\epsfbox{csmassdistalpha12toward.eps}
91: \caption{\textit{Construction of the multifractal binomial
92: measure $\mu$, with emphasis on the intervals with regularity
93: $\alpha(1,2)$.}}
94: \end{figure}
95:
96:
97: \ndnt There is a notion of multifractal spectrum which stems
98: immediately from this set up and is reminiscent of similar results
99: on multifractals found in Chapter 17 of \cite{Falc}, especially the
100: graphs of the spectra in Figures 17.2 on page 259 (reproduced in
101: Section 2 of this work) and 17.3 on page 261. In our context, the
102: spectrum $\sigma(\alpha)$ is defined as the function which yields
103: the abscissa of convergence of the partition zeta function with
104: regularity $\alpha$, $\zeta^{\mu}_{\mathfrak{P}}(\alpha,s)$, for all
105: $\alpha$. See Figure 5 for an approximation of the graph for this
106: function and compare to Figure 2 which contains the graph of the
107: multifractal spectrum $f(\alpha)$ from \cite{Falc}.
108:
109:
110:
111: \ndnt This graph exhibits some interesting properties, such as its
112: maximum coincides with the Minkowski dimension of the support of
113: $\mu$. Further, this structure holds in greater generality. If, in
114: the construction of the Cantor set, the initial length is replaced
115: by some $h^{-1}$ and the smaller weight by a $w^{-1}$, we have the
116: following theorem. The proof is omitted, but can be found in
117: \cite{LapRock,Rock}.
118:
119: \begin{thm}\label{thm:max} As a function of the regularity values,
120: the abscissa of convergence function $\sigma$ associated with the
121: partition zeta function of the measure $\mu(h,w)$ with $h\geq2$ and
122: $w>2$ has the form
123: \begin{eqnarray}
124: \sigma(\alpha) &=& \frac{(\alpha - \log_h{w})}{\log_h{(w-1)}} \cdot
125: \log_h\left( \frac{-(\alpha - \log_h{w})}{\log_h{(w-1)} }\right)\\
126: & &- \left(1+ \frac{(\alpha - \log_h{w})}{\log_h{(w-1)}}\right)
127: \cdot \log_h{\left(1+ \frac{(\alpha - \log_h{w}) }{
128: \log_h{(w-1)}}\right) }.\notag
129: \end{eqnarray}
130: \ndnt As the abscissa of convergence function, $\sigma$ is defined
131: on a dense subset of the interval $[\log_h{w}-\log_h{(w-1)},
132: \log_h{w}]$, and it attains its maximum at
133: \begin{eqnarray}
134: \alpha =
135: \alpha(1,2) = \log_h{w}-(1/2)\log_h{(w-1)}.
136: \end{eqnarray}
137: This maximum value coincides with the Minkowski dimension of the
138: support of the measure $\mu(h,w)$. That is,
139: \begin{eqnarray}
140: \dim_M(supp(\mu(h,w))&=& \max\{\sigma(\alpha) \hs | \hs \alpha =
141: \alpha(k_1,k_2), (k_1,k_2)=1\}\\
142: &=& \log_h{2}.\notag
143: \end{eqnarray}
144: \end{thm}
145:
146:
147: \begin{figure}
148: \epsfysize=10cm\epsfbox{svsalphaoncst.eps}
149: \caption{\textit{$\sigma$ as a function of $\alpha$ for the measure $\mu$.}}
150: \end{figure}
151:
152:
153: \ndnt Theorem \ref{thm:max} almost contains the following
154: multifractal spectrum $f(\alpha)$ discussed on page 934 of Appendix
155: B in \cite{PeitJS} as a specific case:
156:
157: \begin{eqnarray*}
158: \sigma(\alpha) = f(\alpha) =
159: & &-\frac{\alpha_{max}-\alpha}{\alpha_{max}-\alpha_{min}}\log_2\left(\frac{\alpha_{max}-\alpha}{\alpha_{max}-\alpha_{min}}\right)\\
160: & &-\frac{\alpha-\alpha_{min}}{\alpha_{max}-\alpha_{min}}\log_2\left(\frac{\alpha-\alpha_{min}}{\alpha_{max}-\alpha_{min}}\right),
161: \end{eqnarray*}
162: where $\alpha_{max}$ and $\alpha_{min}$ denote the maximum and
163: minimum regularity values attained by the measure in question. To
164: fit the measure from page 934 of Appendix B in \cite{PeitJS} to our
165: setting, take $h=2$ and $w=3$. The pertinent regularity values
166: become
167: \[
168: \alpha := \alpha(k_1, k_2) =
169: \frac{\log{(2^{nk_1}/3^{nk_2})}}{\log{(1/2^{nk_2})}} = \log_2{3} -
170: \frac{k_1}{k_2},
171: \]
172: where $\alpha_{max}$ and $\alpha_{min}$ are attained when $k_1 = 0$
173: with $k_2=1$ and $k_1 =k_2 =1$, respectively. We say that $f$ is
174: ``almost'' a specific case because the equation $\sigma(\alpha) =
175: f(\alpha)$ only holds on a dense and discrete subset of
176: $[\alpha_{min},\alpha_{max}]$, as opposed to the full interval on
177: which $f$ is defined.
178:
179: \ndnt In addition, the graph of the spectrum $f(\alpha)$ depicted in
180: Section 2 of this work and Chapter 17 of \cite{Falc} also stems from
181: an equation such as (5) from Theorem \ref{thm:max} (see page 261 of
182: \cite{Falc}).
183:
184: \ndnt Among the common features of these graphs are the coincidence
185: of the maximum height of the curve with the Minkowksi dimension of
186: the support of the underlying measure and the symmetry about the
187: vertical line that passes through this maximum height. The symmetry
188: of $\sigma(\alpha)$ is evident from the equality of the binomial
189: coefficients $\binom{nk_2}{nk_1}$ and $\binom{nk_2}{n(k_2-k_1)}$ in
190: the respective partition zeta functions.
191:
192: \ndnt The distinctions between the abscissa of convergence function
193: $\sigma$ of this text and the multifractal spectrum $f$ of
194: \cite{Falc} and \cite{PeitJS} lie in their developments. The
195: function $\sigma$ follows directly from the partition zeta functions
196: defined by the weighted partitions which define the measure $\mu$,
197: whereas $f$ follows from the same heuristic development but takes
198: its values from an appropriate Legendre transform.
199:
200: \ndnt Further generalizations to measures with multiplicative
201: structure similar to that of $\mu$ have been made, but for brevity
202: we shall merely mention their existence. Also, in \cite{LVM},
203: another type of zeta function which describes multifractal measures
204: in a manner very similar to the partition zeta functions has been
205: defined and investigated. The results contained in that paper
206: provide even more connections between the analysis of fractal
207: strings via zeta functions and this new approach to multifractal
208: analysis.
209:
210: \ndnt It is important to note that the partition zeta functions do
211: not yield the geometric zeta function as some kind of special case.
212: Indeed, there is no underlying fractal string or closed set that
213: results from a construction like that of the Cantor set for the
214: intervals of $\mathfrak{P}$ and a fixed regularity. So, although
215: they do not recover the geometric zeta function for fractal strings,
216: the partition zeta functions provide some interesting information
217: for multifractal measures, further solidifying the existing results
218: described, for example, in \cite{Falc} and \cite{PeitJS}.
219:
220: \ndnt The next section describes a precursor to the partition zeta
221: function and the similar zeta function in \cite{LVM}. Despite its
222: name, the multifractal zeta function does not connect to
223: multifractal analysis as thoroughly as these other zeta functions,
224: but it is a generalization of the geometric zeta functions of
225: fractal strings and provides topological information which can not
226: be obtained from their complex dimensions.
227: