Lacunary function

Pick an integer ${\displaystyle a\geq 2}$. Consider the following function defined by a simple power series:

${\displaystyle f(z)=\sum _{n=0}^{\infty }z^{a^{n}}=z+z^{a}+z^{a^{2}}+z^{a^{3}}+z^{a^{4}}+\cdots \,}$

The power series converges locally uniform on any open domain |z| < 1. This can be proved by comparing f with the geometric series, which is absolutely convergent when |z| < 1. So f is analytic on the open unit disk. Nevertheless, f has a singularity at every point on the unit circle, and cannot be analytically continued outside of the open unit disk, as the following argument demonstrates.

Clearly f has a singularity at z = 1, because

${\displaystyle f(1)=1+1+1+\cdots \,}$

is a divergent series. But if z is allowed to be non-real, problems arise, since

${\displaystyle f\left(z^{a}\right)=f(z)-z\qquad f\left(z^{a^{2}}\right)=f(z^{a})-z^{a}\qquad f\left(z^{a^{3}}\right)=f\left(z^{a^{2}}\right)-z^{a^{2}}\qquad \cdots \qquad f\left(z^{a^{n+1}}\right)=f\left(z^{a^{n}}\right)-z^{a^{n}}}$

we can see that f has a singularity at a point z when z^a = 1, and also when z^a2 = 1. By the induction suggested by the above equations, f must have a singularity at each of the aⁿ-th roots of unity for all natural numbers n. The set of all such points is dense on the unit circle, hence by continuous extension every point on the unit circle must be a singularity of f.[1]

Evidently the argument advanced in the simple example shows that certain series can be constructed to define lacunary functions. What is not so evident is that the gaps between the powers of z can expand much more slowly, and the resulting series will still define a lacunary function. To make this notion more precise some additional notation is needed.

We write

${\displaystyle f(z)=\sum _{k=1}^{\infty }a_{k}z^{\lambda _{k}}=\sum _{n=1}^{\infty }b_{n}z^{n}\,}$

where b_n = a_k when n = λ_k, and b_n = 0 otherwise. The stretches where the coefficients b_n in the second series are all zero are the lacunae in the coefficients. The monotonically increasing sequence of positive natural numbers {λ_k} specifies the powers of z which are in the power series for f(z).

Now a theorem of Hadamard can be stated.[2] If

${\displaystyle {\frac {\lambda _{k}}{\lambda _{k-1}}}>1+\delta \,}$

for all k, where δ > 0 is an arbitrary positive constant, then f(z) is a lacunary function that cannot be continued outside its circle of convergence. In other words, the sequence {λ_k} doesn't have to grow as fast as 2^k for f(z) to be a lacunary function – it just has to grow as fast as some geometric progression (1 + δ)^k. A series for which λ_k grows this quickly is said to contain Hadamard gaps. See Ostrowski–Hadamard gap theorem.

Mathematicians have also investigated the properties of lacunary trigonometric series

${\displaystyle S((\lambda _{k})_{k},\theta )=\sum _{k=1}^{\infty }a_{k}\cos(\lambda _{k}\theta )\qquad S((\lambda _{k})_{k},\theta ,\omega )=\sum _{k=1}^{\infty }a_{k}\cos(\lambda _{k}\theta +\omega )\,}$

for which the λ_k are far apart. Here the coefficients a_k are real numbers. In this context, attention has been focused on criteria sufficient to guarantee convergence of the trigonometric series almost everywhere (that is, for almost every value of the angle θ and of the distortion factor ω).

• Kolmogorov showed that if the sequence {λ_k} contains Hadamard gaps, then the series S(λ_k, θ, ω) converges (diverges) almost everywhere when

${\displaystyle \sum _{k=1}^{\infty }a_{k}^{2}}$

converges (diverges).

• Zygmund showed under the same condition that S(λ_k, θ, ω) is not a Fourier series representing an integrable function when this sum of squares of the a_k is a divergent series.[3]

Greater insight into the underlying question that motivates the investigation of lacunary power series and lacunary trigonometric series can be gained by re-examining the simple example above. In that example we used the geometric series

${\displaystyle g(z)=\sum _{n=1}^{\infty }z^{n}\,}$

and the Weierstrass M-test to demonstrate that the simple example defines an analytic function on the open unit disk.

The geometric series itself defines an analytic function that converges everywhere on the closed unit disk except when z = 1, where g(z) has a simple pole.[4] And, since z = e^iθ for points on the unit circle, the geometric series becomes

${\displaystyle g(z)=\sum _{n=1}^{\infty }e^{in\theta }=\sum _{n=1}^{\infty }\left(\cos n\theta +i\sin n\theta \right)\,}$

at a particular z, |z| = 1. From this perspective, then, mathematicians who investigate lacunary series are asking the question: How much does the geometric series have to be distorted – by chopping big sections out, and by introducing coefficients a_k ≠ 1 – before the resulting mathematical object is transformed from a nice smooth meromorphic function into something that exhibits a primitive form of chaotic behavior?

• Analytic continuation
• Szolem Mandelbrojt
• Benoit Mandelbrot
• Mandelbrot set
• Fabry gap theorem
• Ostrowski–Hadamard gap theorem

• Katusi Fukuyama and Shigeru Takahashi, Proceedings of the American Mathematical Society, vol. 127 #2 pp. 599–608 (1999), "The Central Limit Theorem for Lacunary Series".
• Szolem Mandelbrojt and Edward Roy Cecil Miles, The Rice Institute Pamphlet, vol. 14 #4 pp. 261–284 (1927), "Lacunary Functions".
• E. T. Whittaker and G. N. Watson, A Course in Modern Analysis, fourth edition, Cambridge University Press, 1927.

• Fukuyama and Takahashi, 1999 A paper (PDF) entitled The Central Limit Theorem for Lacunary Series, from the AMS.
• Mandelbrojt and Miles, 1927 A paper (PDF) entitled Lacunary Functions, from Rice University.
• MathWorld article on Lacunary Functions