**3. Results**

## *3.1. Minimum Required Fixed-Point Iterations*

The fixed point iteration algorithm described in Equations (15a)–(16b) is used to estimate the Kronecker–Toeplitz-structured covariance for the STBF-STRUCT classifier. Fixedpoint iteration is an iterative procedure starting from (in our case) non-informed initial guesses for the spatial and temporal covariance matrices. As a stopping criterion, one could impose a threshold on the difference in outcome of successive steps, e.g., based on the covariance norm or the classifier accuracy. However, few iterations or even just one [59] suffice to achieve satisfactory performance in practice.

Figure 2 confirms these results for the STBF-STRUCT classifier. Using more than one fixed-point iteration does not significantly improve the accuracy across the amounts of training data and the number of trials used for evaluation. Hence, only one iteration is used for the STBF-STRUCT classifier, leading to a drastic speed-up of the training process.

**Figure 2.** Average cross-validated STBF-STRUCT accuracy using one trial per block for validation over all 21 subjects relative to the number of iterations used to estimate the Kronecker–Toeplitz-structured shrunk covariance. Error bars represent the first and third quartiles. The accuracy does not improve when using more than one iteration. (**A**) Results for 1, 2, and 5 trials using only the first block in each training fold for training. (**B**) Results for 1, 2, and 5 trials using all nine training blocks in the training folds.
