Critique of the test of multifractality as applied to biological data (Q2177446)

This paper describes the application of multifractal testing in biological data by describing Chhabra and Jensen methods. ``Multifractality'' denotes ``multiplicity of fractal dimensions'' where a range of fractal dimensions is finding instead of one. For the test of multifractality, the fractal dimension becomes a continuous function, $f(\alpha)$, instead of being a constant D, where $\alpha$ is an index of local variability in the fractal dimension. The conditions for multifractality are said to have been met in a given problem when values of $\alpha$ and $f(\alpha)$ can be determined and can be shown to collapse onto a smooth curve with certain characteristics that are expected on theoretical grounds. However, the calculations of $\alpha$ and $f(\alpha)$ are not straightforward in all but a few idealized cases, and discussion of different methods of calculation are therefore at the core of this subject. Because there is a well-defined in mathematics, it is well proved how to obtain $\alpha$ and $f(\alpha)$, for experimental data, biological data, etc., due to experimental errors, especially in biology, because of biological ``variability'' or ``scatter'', the test of multifractal is not easy or directly obtained. In the method Chhabra and Jensen, a discrete set of data points representing measured values of an entity $x$ is ordering from its lowest value $x_{\min}$ to the highest value $x_{\max}$. The continuous range $x_{\min}\leq x\leq x_{\max}$ is then divided into $N$ small subintervals of length ε and centered at $x_i$, $i=1,2,\dots,N$. The number of data points within the subinterval centered at $x_i$ is denoting by the measure $\rho_i(\varepsilon)$ of the subinterval. The measure $\rho_i$ is normalized by writing \[ \mu_i(q,\varepsilon)=\frac{\rho_i^q(\varepsilon)}{\sum^N_{i=1}\rho_i^q(\varepsilon)}, \] where $q$ is a key parameter in the relation between $\alpha$ and $f(\alpha)$. The Hölder exponent $\alpha$ is defined by \[ \alpha(q)=\lim\limits_{\varepsilon\to 0}\frac{\sum^N_{i=1}\mu_i(q,\varepsilon)\log\rho_i(q,\varepsilon)}{\log\varepsilon}, \] and the corresponding Hausdorff dimension is defined by \[ f(q)=\lim\limits_{\varepsilon\to 0}\frac{\sum^N_{i=1}\mu_i(q,\varepsilon)\log\rho_i(q,\varepsilon)}{\log\varepsilon}. \] That is, the value of $\alpha(q)$ and $f(q)$ are obtained from the slope of a plot of $\sum^N_{i=1}\mu_i(q,\varepsilon)\log\rho_i(q,\varepsilon)$ against $\log\varepsilon$ and $\sum^N_{i=1}\mu_i(q,\varepsilon)\log\mu_i(q,\varepsilon)$ against $\log\varepsilon$, respectively. Multifractal is mainly established for discrete data, so it is particularly useful for experimental data or biological data, however, the test of multifractal requires the measure to be tested be binned into $N(\varepsilon)$ boxes or bins of diminishing size $\varepsilon$ to the ultimate limit $\varepsilon\to 0$, the limit cannot be attained in the case of biological or experimental data. At last, some examples are given to illustrate the advantages and disadvantages of the test of multifractality.