Heteroscedastic Concomitant Lasso for sparse multimodal electromagnetic brain imaging

In high dimension, it is customary to consider $\ell_1$-penalty regularized estimators to enforce sparsity, the Lasso being a canonical example. For statistical efficiency, they rely on tuning a parameter trading data fitting versus sparsity. For the Lasso theory to hold, this tuning parameter should be proportional to the noise level, yet the latter is often unknown in practice. A possible remedy is to consider estimators, such as the Scaled Lasso or the Concomitant Lasso, that jointly optimize over the regression parameter as well as over the noise level, making the choice of the regularization independent from the noise level. However, when data from different sources with varying noise levels are observed, as it is common with multimodal datasets, new dedicated estimators are necessary. In this work we provide statistical and computational solutions to deal with such heteroscedastic regression model, with a special emphasis on functional brain imaging with combined MEG and EEG signals. Adopting the formulation of Concomitant/Scaled Lasso-type estimators, we propose a jointly convex formulation leading to an efficient algorithm whose computational cost is no more expensive than the one for the Lasso. Empirical results on simulations and real M/EEG datasets demonstrate that our model successfully estimates the different noise levels and provides a stable estimation of the support as the noise levels vary.