Monte Carlo algorithm for least dependent non-negative mixture decomposition.

We propose a simulated annealing algorithm (stochastic non-negative independent component analysis, SNICA) for blind decomposition of linear mixtures of non-negative sources with non-negative coefficients. The demixing is based on a Metropolis-type Monte Carlo search for least dependent components, with the mutual information between recovered components as a cost function and their non-negativity as a hard constraint. Elementary moves are shears in two-dimensional subspaces and rotations in three-dimensional subspaces. The algorithm is geared at decomposing signals whose probability densities peak at zero, the case typical in analytical spectroscopy and multivariate curve resolution. The decomposition performance on large samples of synthetic mixtures and experimental data is much better than that of traditional blind source separation methods based on principal component analysis (MILCA, FastICA, RADICAL) and chemometrics techniques (SIMPLISMA, ALS, BTEM).

Measure profile surrogates: a method to validate the performance of epileptic seizure prediction algorithms.

In a growing number of publications it is claimed that epileptic seizures can be predicted by analyzing the electroencephalogram (EEG) with different characterizing measures. However, many of these studies suffer from a severe lack of statistical validation. Only rarely are results passed to a statistical test and verified against some null hypothesis H0 in order to quantify their significance. In this paper we propose a method to statistically validate the performance of measures used to predict epileptic seizures. From measure profiles rendered by applying a moving-window technique to the electroencephalogram we first generate an ensemble of surrogates by a constrained randomization using simulated annealing. Subsequently the seizure prediction algorithm is applied to the original measure profile and to the surrogates. If detectable changes before seizure onset exist, highest performance values should be obtained for the original measure profiles and the null hypothesis. "The measure is not suited for seizure prediction" can be rejected. We demonstrate our method by applying two measures of synchronization to a quasicontinuous EEG recording and by evaluating their predictive performance using a straightforward seizure prediction statistics. We would like to stress that the proposed method is rather universal and can be applied to many other prediction and detection problems.

Estimating mutual information.

We present two classes of improved estimators for mutual information M(X,Y), from samples of random points distributed according to some joint probability density mu(x,y). In contrast to conventional estimators based on binnings, they are based on entropy estimates from k -nearest neighbor distances. This means that they are data efficient (with k=1 we resolve structures down to the smallest possible scales), adaptive (the resolution is higher where data are more numerous), and have minimal bias. Indeed, the bias of the underlying entropy estimates is mainly due to nonuniformity of the density at the smallest resolved scale, giving typically systematic errors which scale as functions of k/N for N points. Numerically, we find that both families become exact for independent distributions, i.e. the estimator M(X,Y) vanishes (up to statistical fluctuations) if mu(x,y)=mu(x)mu(y). This holds for all tested marginal distributions and for all dimensions of x and y. In addition, we give estimators for redundancies between more than two random variables. We compare our algorithms in detail with existing algorithms. Finally, we demonstrate the usefulness of our estimators for assessing the actual independence of components obtained from independent component analysis (ICA), for improving ICA, and for estimating the reliability of blind source separation.

Least-dependent-component analysis based on mutual information.

We propose to use precise estimators of mutual information (MI) to find the least dependent components in a linearly mixed signal. On the one hand, this seems to lead to better blind source separation than with any other presently available algorithm. On the other hand, it has the advantage, compared to other implementations of "independent" component analysis (ICA), some of which are based on crude approximations for MI, that the numerical values of the MI can be used for (i) estimating residual dependencies between the output components; (ii) estimating the reliability of the output by comparing the pairwise MIs with those of remixed components; and (iii) clustering the output according to the residual interdependencies. For the MI estimator, we use a recently proposed k -nearest-neighbor-based algorithm. For time sequences, we combine this with delay embedding, in order to take into account nontrivial time correlations. After several tests with artificial data, we apply the resulting MILCA (mutual-information-based least dependent component analysis) algorithm to a real-world dataset, the ECG of a pregnant woman.

Bivariate surrogate techniques: necessity, strengths, and caveats.

The concept of surrogates allows testing results from time series analysis against specified null hypotheses. In application to bivariate model dynamics we here compare different types of surrogates, each designed to test against a different null hypothesis, e.g., an underlying bivariate linear stochastic process. Two measures that aim at a characterization of interdependence between nonlinear deterministic dynamics were used as discriminating statistics. We analyze eight different stochastic and deterministic models not only to demonstrate the power of the surrogates, but also to reveal some pitfalls and limitations.