In a method to select an undersampling scheme of k-space and an associated set of reconstruction kernels to acquire reduced magnetic resonance (MR) data sets with multiple coils, a calibration data set is acquired for each of the respective coils, a noise covariance is determined from autocorrelations and correlations of the noise of the various coils. At least one set of reconstruction kernels is calculated for each of the multiple undersampling schemes from the calibration data sets of the various coils. For each set of reconstruction kernels, a characteristic value is calculated from the noise covariance and the respective reconstruction kernels of the coils, with the characteristic value being proportional to a spatial mean value of a signal noise of an MR image. A selected undersampling scheme and a selected set of reconstruction kernels are selected based on the calculated characteristic values.