1: \section{Regular variation}
2: \label{sec:rv}
3: Several formal calculations by physicists working on Smoluchowski's
4: equations take the following form : (1) assume that the number density
5: $n(x) \sim x^{\alpha}$ for some scaling exponent $\alpha$, and (2)
6: conclude based on physical arguments that $\alpha$ takes a particular value.
7: The theory of regular variation helps us makes these formal
8: calculations precise, and lays bare the mechanism controlling
9: the approach to scaling form. Our primary source is Feller's book, and
10: we restate below useful results from~\cite[VIII.8]{Feller}. The theory
11: of regular variation has many applications in analysis and probability, and an
12: authoritative text, rich in examples, is~\cite{Bingham}.
13:
14: \subsection{Rigidity of scaling limits}
15: Loosely speaking, a function is {\em slowly varying\/} if it is
16: asymptotically flat under changes of scale. Precisely, we say that
17: a positive function $L(x)$ is {\em slowly varying at infinity\/} if
18: \begin{equation}
19: \label{eq:rv1}
20: \lim_{x \to \infty} \frac{L(tx)}{L(x)} = 1, \quad\mbox{for all $t >0$}.
21: \end{equation}
22: For example, all powers and iterates of $\log x$ are slowly varying at
23: infinity. If we consider the limit $x \to 0$ instead, we obtain functions
24: that are slowly varying at $0$.
25:
26: A function $N(x)$ is {\em regularly varying at infinity\/} with index
27: $\rho \in \mathbb{R}$ if there is a slowly varying function $L(x)$ such that
28: \begin{equation}
29: \label{eq:rv2}
30: N(x) \sim x^\rho
31: L(x)\;\;\mathrm{as}\;\;{x \to \infty} .
32: \end{equation}
33: The notation $\sim$ means $\lim_{x \to \infty} N(x)/x^\rho L(x) =1$.
34:
35: The notion of regular variation is intimately related to necessary
36: and sufficient condtions for the existence of scaling limits.
37: This is reflected in the following classical ``rigidity''
38: lemma~\cite[Lemma VIII.8.3]{Feller}, which will be one of our principal tools.
39:
40: \begin{lemma}
41: \label{le:rv}
42: Suppose that
43: ${a_{n+1}}/{a_n} \to 1$ {and} $\lambda_n \to \infty$ as $n\to\infty$.
44: %\[ \frac{a_{n+1}}{a_n} \to 1 \quad\mbox{and}\quad \lambda_n \to \infty
45: %\qquad\mbox{as $n\to\infty$.}\]
46: If $\varphi$ is a positive, monotone function such that
47: \[ \lim_{n \to \infty} a_n \varphi\left(\frac{s}{\lambda_n}\right) = g(s) \leq
48: \infty \]
49: exists for $s$ in a dense subset of $(0,\infty)$, and $g$ is finite
50: and positive on some
51: interval, then $\varphi$ varies regularly at $0$ and $g(s) = cs^\rho$
52: for $-\infty < \rho < \infty$ and some $c >0$.
53: \end{lemma}
54:
55: \subsection{Tauberian theorems}
56: We will rigorously deduce the
57: asymptotics of $\num$ by the beautiful
58: Hardy-Littlewood-Karamata Tauberian theorem~\cite[XIII.5]{Feller}.
59: \begin{thm}
60: \label{thm:tauber}
61: If $L$ is slowly varying at infinity and $0 \leq \alpha < \infty$, then the
62: following are equivalent:
63: %\[ \num(x) \sim \frac{x^\alpha L(x)}{\Gamma(1+\alpha)} \quad \mbox{as
64: % $x \to \infty$} ,\]
65: \[ \num(x) \sim {x^\alpha L(x)} \quad \mbox{as $x \to \infty$} ,\]
66: and
67: \[ \eta(s) \sim s^{-\alpha} L\left(\frac{1}{s}\right)\Gamma(1+\alpha) \quad
68: \mbox{as $s \to 0$}.\]
69: Moreover, this equivalence remains true when we interchange the roles of the
70: origin and infinity, namely when $s \to \infty$ and $x \to 0$.
71: \end{thm}
72: We will use the following lemma to show that there is no loss of
73: generality in working with $\varphi$ instead of $\eta$.
74: \begin{lemma}
75: \label{le:diff_phi}
76: Suppose $\partial_s \psi$ is the Laplace transform of a positive
77: measure. Let $\alpha <1$ and $L$ be a function slowly varying at $0$.
78: The following are equivalent.
79: \begin{enumerate}
80: \item $\psi(s)-\psi (0) \sim s^{1-\alpha} L(s)$ as $s \to 0$.
81: \item $\partial_s \psi(s) \sim (1-\alpha) s^{-\alpha} L(s)$ as $s \to 0$.
82: \end{enumerate}
83: \end{lemma}
84: \begin{proof}
85: Suppose $(1)$. Since $\psi(s)-\psi(0) = s^\alpha L(s) h(s)$ with $\lim_{s
86: \to 0} h(s) =1$, without loss of generality we may write $\psi(s)-\psi(0)
87: = s^\alpha L(s)$. Fix $a >1$. Then by the mean value theorem and
88: the complete monotonicity of $\partial_s \psi$ we have
89: \[s(a-1) \partial_s \psi (s) \geq \psi(as) - \psi(s) = s^{1-\alpha}
90: L(s) \left( a^{1-\alpha} \frac{L(as)}{L(s)} -1 \right). \]
91: Thus, letting $s \to 0$, and using (\ref{eq:rv1}) we have
92: \[ \liminf_{s \to 0} \frac{s^\alpha \partial_s \psi (s)}{L(s)} \geq
93: \frac{a^{1-\alpha}-1}{a-1}. \]
94: Since $a >1$ is arbitrary, we may maximize the right hand side to
95: obtain
96: \[ \liminf_{s \to 0} \frac{s^\alpha \partial_s \psi (s)}{L(s)} \geq
97: 1-\alpha.\]
98: Choosing $a <1$ and using a similar argument yields,
99: \[ \limsup_{s \to 0} \frac{s^\alpha \partial_s \psi (s)}{L(s)} \leq
100: 1-\alpha.\]
101: Thus, $\partial_s \psi (s) \sim (1-\alpha) s^{-\alpha} L(s)$.
102:
103: Conversely, assume (2). Then we have
104: \[ \psi(s)-\psi(0) = (1-\alpha) \int_0^s t^{-\alpha}L(t) dt
105: =(1-\alpha) s^{1-\alpha} L(s) \int_0^1 t^{-\alpha}\frac{L(st)}{L(s)}\,dt.
106: \]
107: Since $L$ is slowly varying at zero, then for any constants
108: $A>1$ and $\delta>0$ there exists $s_0$ such that for $0<s\le s_0$,
109: $0<t\le1$ we have $L(st)/L(s) \le A t^{-\delta}$
110: (this is not hard to show, but see Theorem 1.5.6 in \cite{Bingham}).
111: Then (1) follows by dominated convergence.
112: \end{proof}
113: