Generalized moments of additive functions. II (Q1347385)

[Part I, cf. J. Number Theory 32, 281-288 (1989; Zbl 0678.10038).] The first moment inequality valid for every additive function, the celebrated Turán-Kubilius inequality, can be stated as ``the variance of an additive function on the first \(x\) integers is less than a constant times the variance of a sum of naturally defined random variables''. This was extended to moments of order \(\beta >0\) by Elliott, and to general moments of the type \[ \sum _{n\leq x} \varphi (|f(n)-A) |) \tag{1} \] defined via an arbitrary increasing nonnegative function \(\varphi \) by the reviewer [J. Number Theory 18, 27-33 (1984; Zbl 0524.10042)]. This estimate is a constant times the variance of the sum of variables whenever the moment function satisfies \(\varphi (2x)\ll \varphi (x)\), and weaker for fast growing functions. For fast growing (say, exponential) functions the moment of an additive function can exceed the moment of the corresponding sum. In another direction, the reviewer proved that by first substracting a proper logarithmic function one can find an expression which yields not just an upper estimate, but the correct order of magnitude for the variance of an additive function, the variance of a logarithmic function being much smaller than that of the corresponding sum of independent variables. \textit{A. Hildebrand} improved this in two directions: first, for the variance he found an actual asymptotic formula, next, he proved that the order of magnitude can be found for moments of order \(\beta >0\) in this way [Ann. Inst. Fourier 33, No. 3. 1-22 (1983; Zbl 0486.10043)]. The present paper extends Hildebrand's theorem to general moments of type (1), under the assumptions that the moment function \(\varphi \) satisfies \(\varphi (0)=0\), \(\varphi (x)\rightarrow \infty \) as \(x\rightarrow \infty \) and grows at most polynomially, \(\varphi (2x)\ll \varphi (x)\). This assumption is natural for two reasons: first, the \(\varphi \)-moment of a sum of independent variables can be estimated via a sum of moments of the terms for this class, next, because of the real limitations of this phenomenon as mentioned after (1). This inequality is then applied to describe classes of additive functions that satisfy \[ {\limsup x}^{-1} \sum _{n\leq x} \varphi (|f(n)-\alpha (x) |/\beta (x)) < \infty \] for suitable norming and centering, or which satisfy \[ x^{-1} \sum _{n\leq x} \varphi ( |f(n)-\alpha (x) |) \leq F(x) \] with suitable centering and a given function \(F\).