Introduction Highly math-anxious (HMA) individuals are characterized by a strong tendency to avoid math, which ultimately undercuts their math competence and forecloses important career paths (Ashcraft, 2002). It is hypothesized that worries and intrusive thoughts associated with math anxiety (MA) reduce working memory resources needed for cognitively demanding math tasks (Chang & Beilock, 2016). However, mental processes that access the memory representations of mathematical knowledge has not been fully uncovered (Ashcraft, 2001). Previous studies indicate that the frontal cortex is dominantly involved in working memory (WM) and more specifically while updating the working memory representations (Smith & Jonides, 1997). Additionally, Klados et. al. 2015 show that higher event-related potential (ERP) measures of HMA subjects are predominantly located at frontocentral sites at cortex, while performing WM tasks. Here, we aim to explore the changes in cortical connectivity profile induced by MA during WM tasks. Methods EEG recordings were measured from 32 adults during performance of WM tasks with two levels of difficulty, 1-back (BT1) and 2-back (BT2) (Klados et al., 2015). According to the Abbreviated Math Anxiety Scale (AMAS), half of the participants were selected among the highly math anxious (HMA) students, whereas the other half had low math anxiety (LMA) (Hopko et al., 2003). ERPs were recorded via a Neurofax EEG-1200 system from 57 electrode sites according to a modified international 10/10 system using an Electrocap. Epoch length was 1200 ms including 200 ms prestimulus baseline. Signals were filtered offline between 0.5 and 45 Hz and submitted to an ICA procedure to identify ocular artifact components (Bell & Sejnowski, 1995). These artifact components were filtered with REGICA (Schlogl et al., 2007, Klados et al., 2009, 2011). Resulting waveforms were inspected visually and epochs containing visible artifacts in the first 500 ms post-stimulus were removed. To overcome the impact of varying electrical conductivity among head compartments on functional connectivity analyses (Nolte et. al., 200), the cortical activity was estimated from 28 EEG signals by adopting a cortical dipole model (He & Wu, 1999, Mattia et. al., 2009). Here, we referenced MNI152 template as an average head model. Scalp, cortex, outer- and inner-skull were extracted by implementing the Boundary Element Method (BEM) with 302 nodes (Uscedu, 2016). Finally, a column-norm normalization was used to prevent the linear inverse problem. This way, we obtained a transition kernel from 57 scalp signals to 302 cortical signals. We constructed connectivity matrices based on 302 cortical signals by using Magnitude Squared Coherence (MSC) in upper alpha band (8-10 Hz) (Lithari et. al., 2012, Klados et. al., 2013). Then, we embed these connectivity networks to the “connectivity components” by employing a nonlinear dimensional reduction technique (Coifman & Lafon, 2016). The first component captures the highest variance, i.e. the similarity/dissimilarity of the connectivity patterns of source regions. We used 2-way ANOVA with factors MA (HMA & LMA) and task difficulty (BT1 & BT2) along the first connectivity component. The p-values were adapted by the FDR correction (Benjamini & Hochberg, 1995). Results Fig.1 demonstrates the mixed effect of group by task interaction on the connectivity profiles. There is a significant interaction between MA and task difficulty, that is prominent in dorsolateral prefrontal cortex (DLPFC), temporal lobe, left ventromedial PFC, right inferior-parietal lobule (IPL), somatosensory and right motor areas. For each significant region (p(G x T)