State-dependent neural correlations can be understood from a neural coding framework. Noise correlations—trial-to-trial or moment-to-moment covariability—can be interpreted only if the underlying signal correlation—similarity of task selectivity between pairs of neural units—is known. Despite many investigations in local spiking circuits, it remains unclear how this coding framework applies to large-scale brain networks. Here, we investigated relationships between large-scale noise correlations and signal correlations in a multitask human fMRI dataset. We found that task-state noise correlation changes (e.g., functional connectivity) did not typically change in the same direction as their underlying signal correlation (e.g., tuning similarity of two regions). Crucially, noise correlations that changed in the opposite direction as their signal correlation (i.e., anti-aligned correlations) improved information coding of these brain regions. In contrast, noise correlations that changed in the same direction (aligned noise correlations) as their signal correlation did not. Interestingly, these aligned noise correlations were primarily correlation increases, suggesting that most functional correlation increases across fMRI networks actually degrade information coding. These findings illustrate that state-dependent noise correlations shape information coding of functional brain networks, with interpretation of correlation changes requiring knowledge of underlying signal correlations. Author Summary We leveraged insights from the neural coding theory to interpret large-scale patterns of correlated activity across many cognitive tasks in human fMRI. We characterized the signal correlation (i.e., task tuning similarity between regions), the noise correlation (i.e., functional connectivity), and the change in noise correlation by task engagement. Following neural coding theory, we parsed noise correlation changes into components that are aligned with the signal correlation, which degrade information coding across regions, versus those that are anti-aligned, which enhance information coding. Anti-alignment of noise correlation changes with the correlation minimizes the amount of interference between the two brain regions. Together, these findings provide a task information coding perspective to interpret task-state correlation changes in human functional brain networks.

Functional connectivity (FC) of blood oxygen level-dependent (BOLD) fMRI time series can be estimated using methods that differ in sensitivity to the temporal order of time points (static vs. dynamic) and the number of regions considered in estimating a single edge (bivariate vs. multivariate). Previous research suggests that dynamic FC explains variability in FC fluctuations and behavior beyond static FC. Our aim was to systematically compare methods on both dimensions. We compared five FC methods: Pearson’s/full correlation (static, bivariate), lagged correlation (dynamic, bivariate), partial correlation (static, multivariate), and multivariate AR model with and without self-connections (dynamic, multivariate). We compared these methods by (i) assessing similarities between FC matrices, (ii) by comparing node centrality measures, and (iii) by comparing the patterns of brain-behavior associations. Although FC estimates did not differ as a function of sensitivity to temporal order, we observed differences between the multivariate and bivariate FC methods. The dynamic FC estimates were highly correlated with the static FC estimates, especially when comparing group-level FC matrices. Similarly, there were high correlations between the patterns of brain-behavior associations obtained using the dynamic and static FC methods. We conclude that the dynamic FC estimates represent information largely similar to that of the static FC. Author Summary Functional connectivity (FC) methods differ in their sensitivity to temporal order (static vs. dynamic) and the number of regions used to estimate a single edge (bivariate vs. multivariate). Dynamic connectivity measures are sensitive to temporal order, whereas static ones are not. We compared the FC methods with respect to both dimensions and showed that the dynamic and static FC estimates represent largely similar information. Using simulations, we have shown that the similarity between dynamic and static FC estimates may be due to the convolution of the neural signal with the hemodynamic response function (HRF) alone. On the other hand, bivariate FC estimates showed better generalizability in cross-validation compared with multivariate FC estimates in the study of brain-behavior associations.