Functional MRI has been slow to benefit from data acceleration techniques based on non-linear image reconstruction for a number of reasons. This is in part due to the uncertain impact of non-linear filtering effects on image and noise statistics, as well as the difficulty in implementing robust trajectories that satisfy the sampling requirements of the acceleration scheme (e.g. incoherence) while providing optimal BOLD contrast and whole-brain coverage. More importantly, no consensus has been reached on the validity of different constraints used to regularise under-sampled FMRI data reconstruction. Although early work in sparsity-constrained FMRI acceleration used the k-t FOCUSS framework,^1,2 recently arguments have emerged for rank-constrained reconstruction.³ In this abstract, we compare the reconstruction performance of two techniques, k-t FOCUSS (temporal sparsity) and k-t FASTER (low-rank), on synthetic FMRI data based on realistic signal characteristics. We demonstrate that the strict rank-constraint method outperforms spectral- and Karhunen-Loeve Transform (KLT)-sparsity across different metrics.The original k-t FOCUSS algorithm uses an asymptotic L1 minimization program to solve for a sparse x-f reconstruction and can be extended with the KLT as its sparsifying temporal transform (KLT-FOCUSS).¹ In contrast, k-t FASTER solves for a spatio-temporally low-rank reconstruction using an iterative hard thresholding and matrix shrinkage algorithm, without requiring a pre-specified basis. We compared k-t FASTER and both FOCUSS methods on 2D pseudo-random Cartesian and golden-ratio radial undersampling of a synthetic data set. To generate the data, five ROIs (spelling FMRIB) were superimposed on fully-sampled FMRI data acquired at TR=0.8 s. The ROI time-courses ranged from completely periodic (single frequency component) to aperiodic (broadband), and background voxels contained a 1/f power spectral density.⁴ RMS error, difference images, and statistical maps from general linear model regression (FSL FEAT) were used to evaluate reconstruction integrity.k-t FASTER produced the lowest total RMS error for both Cartesian and radial undersampling, as well as the lowest RMS errors in individual ROIs. Reconstructions from 8x undersampled radial data had the following RMS errors: k-t FASTER, 2.88%; k-t FOCUSS, 4.12%; and KLT-FOCUSS, 4.18%. Figs. 1 and 2 illustrate the impact of FOCUSS spectral penalties, in both time-series fidelity and spectral content. In Fig. 3, difference images show increased error in the “R” and “I” ROIs, which contain the most broadband spectral information. Fig. 4 shows the k-t FASTER z-statistic map for the “M” ROI as virtually identical to the ground truth, with almost no false positive signal localization. The k-t FOCUSS map, however, shows identification of the “B” ROI. k-t FOCUSS captured the shared, high power low-frequency component, but it failed to characterise the high-frequency components that distinguish the (orthogonal) time-courses. While the KLT-FOCUSS reconstruction retains more broadband spectral content, some false positive localization in the “B” is still evident.These results demonstrate the limitation of sparsity regularization on signals that, aside from simple block design task FMRI, are not spectrally sparse. In fact, it has long been known that FMRI data exhibit 1/f power spectra,⁴ in which resting state FMRI signals reside up to frequencies of about 0.1 Hz. While hemodynamic impulse response filters limit the frequency content of FMRI data, ultra-high temporal resolution studies have found resting-state network information in frequency bands up to 0.8 Hz,⁵ recovery of which would be impacted by sparse regularisation over typical sampling bandwidths. In contrast, rank-based regularisation constrains the reconstruction to a low number of significant modes or components, without requiring that signals conform to any specific structure or representation in an a priori basis. The KLT-FOCUSS method nominally enforces sparsity in an adaptive KLT domain, which permits signals of arbitrary bandwidth. The adaptive KLT domain is similar in principle to low-rank constraints. However, because the KLT transform uses an initial k-t FOCUSS estimate that enforces x-f sparsity, the estimated KLT basis is biased towards spectrally sparse components. KLT-FOCUSS is therefore not as robust as the low-rank basis estimation procedure used by rank-constrained matrix reconstruction. Finally, while the temporal constraints imposed by k-t or KLT-FOCUSS examined here may not be appropriate for FMRI, careful investigation into the use of alternative sparsity constraints (e.g. spatial transform sparsity) or prior information in conjunction with the powerful rank constraints is needed to fully evaluate heavily constrained, non-linear reconstructions in FMRI.