Since the discovery of networks of brain regions whose members have correlated fluctuations in BOLD timeseries signals during periods when the subject is not performing an intentional cognitive task (Biswal, Yetkin et al. 1995), a large number of such “resting-state” analyses have been performed both to extend basic underpinnings of brain function as well as to study neurological disorders. Until recently, these analyses have tacitly assumed that the networks are static, and a single map of Functional Connectivity (FC) is typically calculated from the full scan using either region-of-interest (ROI) correlation methods as in Biswal, or a data-driven method such as Independent Components Analysis (ICA) (Beckmann and Smith 2004). However, more recently FC has been observed to vary across typical scan time scales of 5-10 minutes (Chang and Glover 2010), and considerable interest in the reasons for non-stationarity and methods for elucidating such changes has been demonstrated (see (Hutchison, Womelsdorf et al. 2013) for a review). Here we discuss the methods of observing dynamic FC (dFC), the pitfalls and issues associated with drawing inferences from dynamic changes in the FC maps, and attempts to correlate dFC with measures of behavior or psychometric dynamics. The simplest method of observing dFC is to temporally segment the scan using a sliding window of time frames (Chang and Glover 2010), calculating FC with either an ROI or ICA-based method that produces multiple maps depicting apparently varying brain states such as CAPS (Liu and Duyn 2013) and/or quantitative measures of the dynamics (Chen, Chang et al. 2015). Both intra-network and inter-network dynamics have been observed. An alternative method uses temporally restricted basis functions such as wavelets to demonstrate continuously evolving network patterns within a series of temporal scales commensurate with the temporal extent of the basis functions (Chang and Glover 2010). There are several intrinsic problems with extracting FC in the resting brain experiment. The first is the lack of “ground truth”, i.e. there is no hypothesis for how to model the timeseries in any given voxel, unless independently collected measures of some psychometric or behavioral construct have been collected that could be reasonably expected to reflect changes in brain state. The second significant problem with FC and especially dFC is the low signal-to-noise ratio of the BOLD response in the resting state. The “noise” in FC is any variance in the timeseries that is non-neural and not related to the specific network under study. Since this noise is often not distinguishable from signal of interest, false connectivity can be generated by correlations that are caused by non-neural signals common to many voxels, such as BOLD modulations caused by respiration (Birn, Diamond et al. 2006)and the significance of any observed correlations can be reduced by the spurious variance. The third problem in using sliding window correlations is the tradeoff between short windows for good temporal resolution and poor statistics. It is important to verify the null hypothesis to avoid false connectivity.(Chang and Glover 2010, Hindriks, Adhikari et al. 2016). Therefore, it is important to remove any known sources of spurious fluctuations in the timeseries data before attempting to identify networks with any form of FC. The usual motion co-registration must be employed and motion regressors projected from the timeseries. However, additional care must be taken to remove other sources of noise. Typically the frame repetition time (TR) employed in resting state scanning with conventional acquisitions is ~2s. While respiration may be correctly sampled at this TR, cardiovascular pulsation is undersampled and may be temporally aliased to low frequencies where it can pose as true neural signal. One approach is to use a retrospective sorting algorithm like RETROICOR(Glover, Li et al. 2000) to correctly identify and remove such aliased pseudo-periodic signals. Another source of spurious signal is due to modulation of CO2 content from variations in heart-rate and respiration leading to slow, non-cyclic BOLD modulations. It has been found that a respiration response function (Birn, Smith et al. 2008) and cardiac response function (Chang, Cunningham et al. 2009) can be used to form nuisance regressors to remove these sources of noise. Alternatively, rapid scanning methods using Simultaneous Multi-Slice acquisitions (Feinberg and Setsompop 2013) have been recently introduced that can reduce the TR to a few hundred ms, allowing cardiac-related signals to be faithfully sampled and removed by low-pass or notch filtering. When combined with other independently collected metrics such as heart-rate variability (Chang, Metzger et al. 2013), galvanic skin response (Buchel, Morris et al. 1998), or EEG (Goldman, Stern et al. 2000), dFC can provide valuable insights into brain networks by hypothesizing connections between these measures, brain state variations and fluctuations in dFC within and between different networks. When thus confirmed, it has also been observed that dFC can demonstrated differences between healthy and diseased populations, potentially allowing resting state fMRI to be used as a biomarker.