OK, I'm going to hijack this thread even though there's an answer as I haven't found any quality, localized information about multifractals.
As mentioned in the comments, I first heard about multifractals from a Google Tech Talk by Rogene M. Eichler West, which can be found, without sound, on YouTube, called "Multifractals: Theory, Algorithms, & Applications" . Unfortunately Google Video got discontinued after they bought out YouTube and I can't find the original video that had the sound included.
I still do not understand on a deep level what, how and why multifractals are doing, are better than another method or how they do it, but from what I understand the idea is to generalize the concept of spectrum to include functions that have a scale symmetry, where the scale symmetry can be on many different scales (thus multi-fractal, instead of just being fractal). Just as the Fourier spectra is constructing a profile of the translation invariances of a function, the multifractal spectra gives information about the scale invariances of a function.
The general methodology seems to be, for a given function $f(t)$:
Where $D(\alpha) \stackrel{def}{=} D_F\{x, h(x) = \alpha\}$, and $D_F\{\cdot\}$ is the (Hausdorff?) dimension of a point-set.
I believe the idea is that for chaotic/fractal/discontinuous functions, at any point they can be characterized, locally, by the largest term of their Taylor expansion and the Hölder exponent is a way to characterize this. Once you have the function, $h(t)$, characterizing the Hölder exponent, you use that to construct the singularity spectrum. I believe the singularity spectrum is a synonym for the multi-fractal spectrum.
From what I can tell, the specifics of how to calculate $h(t)$ and $D(\alpha)$ in practice vary from approximating them outright by their definition or by using wavelets to approximate the Hölder exponent and then using a Legendre transform to approximate the multifractal spectrum.
From what I understand, $D(\alpha)$ tends to be (or is always?) concave. I have only the vaguest notion of why this is so. How one relates wavelet transforms to finding the Hölder exponent, how one uses the Legendre transform to find the multi-fractal spectrum, why the multi-fractal spectrum should be concave, what kind of intuitive feeling one should get about a function from viewing the spectrum, amongst many others, I still have no idea about.
The multiplicative cascade seems to be a canonical example of a multifractal process.
Online, "A Brief Overview of Multifractal Time Series" gives a terse run through of multifractals. They claim to be able to tell a healthy heart from one that is suffering from congestive heart failure (see here).
Here are some slides giving a brief overview of multifractals. Near the end of the slides, they give a wavelet transform of the Devil's staircase function and talk a bit about using Wavelet Transform Modulus Maxima Method (WTMM), which appears to be a standard tool when doing this type of analysis (anyone have any good links for this?).
Looking around, I found Wavelets in physics by J. C. van den Berg that had this section web accessible for a definition of the singularity spectrum.
Rudolf H. Riedi seems to have a few papers out there that describe multifractal processes. Here are a few:
While focused on finance, Laurent Calvet and Adlai Fisher have a lot of introduction to terminology in "Multifractility in asset returns: Theory and evidence".
And of course Mandelbrot, along with other authors, has many papers, some of which are:
Fractional Brownian Motion is also mentioned frequently, but I have no real idea of how they relate. Large Deviation Theory also seems to be mentioned, but I don't know how this relates to multifractals either. I believe I've also seen entropy, phase transitions and statistical mechanics mentioned here and there. I would be curious if and what the relation to these subjects and multifractals is.
I feel like I'm stumbling around trying to understand this subject and I have yet to find a cohesive text that brings together enough intuition, math and implementation details so that I feel like I have a firm grasp of what's going on. I would welcome any additional resources or corrections to this answer.
You could start with the first chapter of this book, and then with this three-volumes book. The former is a very nice mathematical introduction to finance, from the viewpoint of someone on the mathematical (or physical) side. The latter may seem, and is, a book on interest rates, but it allows you to cover all mathematical techniques used in finance nowadays, and its first volume is the best introduction I have ever seen on mathematical finance ; it has btw a very nice bibliography that will redirect you to central papers in the discipline etc. I am not that fan of this book, even if I started in the field with him, but it could be ok nevertheless for what you are looking for. Finally, there is a book that is not very good on mathematical finance at all, but it is central on FX implied volatility quoting conventions, and is a must have for this.
Last point, previous books are not books on stochastic processes or PDE's or other mathematical subjects that are used in mathematical finance, they are books on mathematical finance roughly covering these subjects, and using and applying them to fianance - essentially pricing and hedging, curve building etc. This means that sometimes you will need to put your nose in a book or another on stochastic processes or even probabilities (note that this book on probabilities and discrete time martingales is a must-have), or PDE's etc. Theses are my complementary advises. I know that this wasn't you primary question, but I don't see myself giving a piece of advise on mathematical finance without mentioning this.
Best Answer
There is a slight ambiguity in your question since you do not explain what kind of formula you would like to see: however, since it seems, from the references you cited, that you mean a power series formula, you can build it by using Newton's generalization of the binomial theorem $$ t\mapsto (1+t)^\alpha = \sum_{k=0}^\infty \binom{\alpha}{k} t^k.\label{1}\tag{1} $$ By using it, you have that $$ \left(\sum_{k=0}^\infty a_kx^k\right)^{\alpha}= \begin{cases} \displaystyle \sum_{n=0}^\infty \binom{\alpha}{n} \left[\sum_{k=1}^\infty a_kx^k -1\right]^n & a_0=0\\ \\ \displaystyle a_0^\alpha\sum_{n=0}^\infty \binom{\alpha}{n} \left[\sum_{k=1}^\infty a_kx^k\right]^n & a_0\neq0\\ \end{cases}\label{2}\tag{2} $$ and you could further expand this formula by using the generalization of Cauchy Product formula. However, the computational work involved in the whole procedure seems daunting.
The convergence properties of \eqref{2} can be inferred from the convergence properties of \eqref{1}, listed in the related wikipedia entry, and from the properties of the power series $s(x) = \sum_{k=1}^\infty a_k x^k$: call $D(R)$ the convergence disk (or interval, if $x, a_k\in\Bbb R$ for all $n\in\Bbb N$) of $s(x)$, and $C\big(s^\alpha(x)\big)$ the convergence region of \eqref{2}, i.e the largest open set in $D$ where \eqref{2} converges. Since \eqref{1} converges absolutely for any real $\alpha$ if $|t|<1$ then $$ C\big(s^\alpha(x)\big) = \begin{cases} \big\{x\in D(R) : |s(x)-1|<1\big\} & a_0=0 \\ \\ \big\{x\in D(R) : |s(x)-a_0|<{a_0}\big\} & a_0\neq0 \\ \end{cases} $$ In a similar fashion, you can analyze the boundary of $C\big(s^\alpha(x)\big)$ by the convergence properties of \eqref{1} on the boundary $|t|=1$ of its convergence disk.
Edit: comments and further notes
Following a comment of the Asker, let's prove the following formula $$ s^n(x)=\left(\sum_{k=1}^{\infty} a_k x^k\right)^n = \sum_{k=1}^{\infty} \sum_{\substack{0< k_1, \ldots, k_n\le k \\ k_1+\ldots+k_n=k}}a_{k_1} \cdots a_{k_n} x^k \quad\forall n\in\Bbb N,\, n\ge 2\label{3}\tag{3} $$ Let's proceed by induction
Thus formula \eqref{3} is proven true for every $n\in\Bbb N$ by induction.
A final note. There are several ways of representing the composition of functions $\left(\sum_{k=0}^\infty a_kx^k\right)^{\alpha}$ for a real $\alpha$: for example it is possible to express it by using Faa di Bruno's formula in its various flavors. As stated in the comments, If you know the derivative of the composition of two analytic functions you know the Taylor series of their composition since this is likewise an analytic function. In our specific case we have $$ s^\alpha(x)=\sum_{k=0}^\infty \frac{1}{k!}\frac{\mathrm{d}^k}{\mathrm{d}x^k}s^\alpha(x)\big|_{x=0} x^k $$ and, as remarked also in the Wikipedia entry on the said formula, you can calculate the Taylor expansion coefficients directly.
What is the "most convenient" way of proceeding? It depends heavily on the structure of the power series $s(x)$ and on on your purpose: in a word it depends on the context of your application.