Low frequency oscillations – neural correlates of stability and flexibility in cognition

The study investigates how low-frequency, large-scale brain synchronization supports the balance between cognitive flexibility (encoding/updating information) and stability (maintenance) during visuospatial working memory (vsWM). The central hypothesis is that distinct functional brain states, characterized by synchronized theta (4–8 Hz) and alpha (8–14 Hz) networks, govern information flow via phase-amplitude coupling and that optimal transitions between these states relate to better cognitive performance. Building on evidence of low-dimensional spatiotemporal activation patterns and cross-frequency interactions, the authors seek to identify and characterize these states in human MEG, relate them to task phases (encoding vs maintenance), examine load and distractor effects, and test mechanistic routing via an in-silico model.

Prior work indicates that large-scale low-frequency oscillations in theta and alpha are pervasive across cognitive processes and can modulate higher-frequency activity via phase-amplitude coupling. M/EEG and intracranial recordings show phase coupling between theta/alpha phases and gamma amplitude, and gamma bursts linked to low-frequency oscillations, suggesting a mechanism for routing information. Large-scale synchronization predicts perception and supports working memory and attention. Alpha synchronization in dorsal networks has been linked to vsWM capacity and attention, and occipital theta/alpha have been associated with visual processing and event-related responses. Frontobasal ganglia-thalamic loops have been implicated in controlling access to working memory and stabilizing representations, and pulvinar/thalamic oscillations may regulate flexibility and stability. The literature also describes an inverted-U relationship between dopamine and cognitive control, analogous to the hypothesized optimal state-switching rate.

Datasets: Two primary MEG datasets and one additional distractor dataset were analyzed. HCP dataset: 83 adults (ages 22–35; 45 men) performed a 2-back/0-back WM task during MEG (MAGNES 3600). Stimuli: 2000 ms presentation and 500 ms fixation, blocks by category (tools/faces). Preprocessing (provided by HCP): bad channels/segments removed via correlation/variance and ICA; non-brain components discarded; downsampled to 508.63 Hz; epoched −150 to 2650 ms. Structural MRI: Siemens 3T. Four-subject dataset: 2 men and 2 women (ages 21–26) scanned on seven occasions over eight weeks. Tasks: WM-Grid (4×4 grid; 5 or 6 items; 300 ms stimuli, 1000 ms delays; 40 trials each load), Odd One Out (five 300 ms stimuli with 1000 ms delays; positions of odd shape; 40 trials), and a verbal recognition control (letters with 1300 ms spacing; detect Q). MEG: Elekta Neuromag TRIUX, 306 channels, 1000 Hz; tSSS (MaxFilter); ICA to remove heart/eye artifacts; notch at 50 Hz and harmonics; epoched with 300 ms padding. MRI: GE 3T. Distractor dataset: 17 participants (ages 21–41; 13 retained after excluding those without identifiable theta networks). Task: four sequential 500 ms bars with 500 ms delays; distractor trials had bars 2 and 3 as distractors; fixation and recall cues followed; 400 trials total. MEG preprocessing matched 4-subject dataset; downsampled to 250 Hz; trials −2000 ms to 6000 ms. Common MEG processing: Anatomical reconstructions via FreeSurfer; parcellation into 200 regions (Schaefer atlas, 7 networks). Source reconstruction using MNE/dSPM with 5 mm dipoles (fixed orientations) and noise covariance estimated per dataset; collapse to parcels optimized as in prior work. Wavelet filtering: Morlet wavelets at 6 Hz (theta) and 10 Hz (alpha), 5 cycles; downsampled to 5× center frequency; edge windows removed. Synchronization metric: imaginary part of phase-locking value. Network identification: Within each subject and frequency band, ICA (scikit-learn, 2–5 components) on absolute wavelet amplitudes, normalized within frequency/session/subject/task; ICA time series averaged across trials per session/subject/task. ICA spatial operators (200-weight vectors) defined network topographies. K-means clustering classified time points into four states by component strengths; states corresponded to dominant networks: posterior theta (state 1), posterior alpha (state 2), dorsal alpha (state 3), dorsal theta (state 4). A frontal theta component emerging over time in the 4-subject dataset yielded a fifth cluster but was excluded from main results. Reliability of networks assessed via cosine similarity across sessions, tasks, and subjects. Behavioral analysis (HCP): Reaction time in 2-back (accuracy ceiling) and performance on three NIH Toolbox tasks (Flanker, Dimensional Change Card Sort, Pattern Comparison Processing Speed); PCA across tasks, using PC1 as general performance. State-switching rate defined as number of transitions per trial; quadratic relationships tested using OLS with standardized coefficients. Load/distractor analyses (4-subject and distractor datasets): Time spent in states 1 and 3 during stimulus and delay periods regressed on WM load; distractor vs no-distractor differences in state 1 durations for stimuli 2–4 tested with one-sided paired t-tests. In-silico whole-brain model: 202 nodes (200 cortical + 2 basal ganglia–thalamus, one per hemisphere) with connectivity and inter-node distances from MICA-MICS diffusion MRI; cortical nodes retained; subcortical nodes (excluding amygdala, hippocampus) combined into a BG-thalamus unit with cortical inputs to BG and thalamic outputs. Two layers per node: oscillatory Kuramoto layer (delayed coupling; ω=0.04·2π; k=0.22; delays τ=2·D; white noise variance 0.0025; Euler dt=0.1 ms) and spiking rate layer (du/dt = Σ C u(t−τ) − a u + I; a=0.25; dt=0.1 ms). Phase-amplitude coupling: spiking modulated by oscillatory phase, enhancing at troughs and suppressing at peaks. Generation of synchronized networks by increasing thalamic outputs to targeted cortical regions (posterior vs dorsal networks) in the Kuramoto layer; modulation of synchronization frequency (theta/alpha) via subcortical oscillator frequency. Information flow quantified by transfer entropy (TE): time series normalized, binned 0–20, TE summed over delays 0.1–40 ms. Simulation conditions: Stimulate V1 or IPS under posterior or dorsal synchronization at 10 Hz; frequency comparisons at theta (6 Hz) vs alpha (10 Hz). Statistics: t-tests (paired/independent, one- or two-sided per hypothesis), standardized coefficients, Cohen’s d, FDR corrections for node-wise TE differences.

Global synchronization peaked in theta and alpha bands across datasets. ICA revealed four consistent large-scale networks in both bands: posterior (occipital/posterior-parietal) and dorsal (parietal/posterior-frontal). Temporal clustering showed subjects cycling through four states dominated by these networks; up to ~90% of subjects concurrently occupied the same state at certain times (mean ~65%). Network reliability was significant across task (t=37.1, df=154, p<10^-78), session (t=13.2, df=77, p<10^-21), and subjects (t=23.7, df=58, p<10^-31). Cognitive performance related to state-switching rate: a significant quadratic relationship between number of state switches and 2-back reaction time (t=2.88, df=80, p=0.0051) and between PC1 of the cognitive battery and switching (t=−3.18, df=80, p=0.0021), with optimal performance at ~nine switches per trial. Encoding vs maintenance: In vsWM tasks, state 1 (posterior theta) was most active during stimulus presentation (encoding), while state 3 (dorsal alpha) dominated during delay (maintenance). A control verbal recognition task did not show systematic switching between states 1 and 3. Load effects: Time in state 1 during stimulus presentations decreased linearly with load in WM-Grid (β_SD=−0.79, 95% CI[−1.01, −0.57], t=−7.33, df=33, p<10^-7); effect not significant in Odd One Out. During delays, state 3 showed a significant quadratic relation with peak at load 3 (WM-Grid: β_SD=−1.00, 95% CI[−1.19, −0.80], t=−10.33, df=39, p<10^-11; Odd One Out: β_SD=−0.86, 95% CI[−1.14, −0.58], t=−6.19, df=33, p<10^-6), coinciding with a ceiling in alpha synchronization strength at load ≥3. Distractor dataset: Internally regulated posterior theta gating was supported by shorter state 1 durations for distractor stimuli 2–3 and longer for stimulus 4 compared to no-distractor trials; the mean difference across periods was 128.2 ms (95% CI[43.8, ∞], t=2.73, df=12, p=0.009, d=0.79). In-silico model: Increased thalamic input selectively synchronized targeted posterior or dorsal networks; synchronization inside vs outside network increased by ~63% (posterior) and ~62% (dorsal); posterior: mean synchronization difference=0.08 (t=10.20, df=9, p<10^-5, d=3.40), dorsal: 0.09 (t=14.81, df=9, p<10^-7, d=4.94). Information routing: TE from V1 was 201% (SD=80%) larger under posterior vs dorsal synchronization; TE from IPS was 207% (SD=28%) larger under dorsal vs posterior synchronization (posterior: mean ΔTE=0.018, t=7.83, df=9, p<10^-4, d=2.61; dorsal: mean ΔTE=0.021, t=26.88, df=9, p<10^-9, d=8.88), with strongest increases within the synchronized networks (posterior network vs rest: Δ=0.018, t=8.60, p<10^-4, d=2.87; dorsal network vs rest: Δ=0.019, t=15.95, p<10^-7, d=5.32). Frequency effects: Posterior synchronization frequency (theta vs alpha) did not affect TE (p=0.85), whereas dorsal theta increased TE relative to dorsal alpha (mean difference=0.011, 95% CI[0.008, 0.013], t=11.64, df=9, p<10^-6, d=3.88); dorsal alpha selectively increased TE to specific regions (e.g., frontal eye field, ipsilateral posterior cortex) after FDR correction.

Findings support the existence of large-scale functional brain states, defined by synchronized theta and alpha networks, that balance flexibility and stability in vsWM. The association of optimal state-switching (~nine transitions) with superior performance suggests that cognitive control involves regulating transitions among these states. State 1 (posterior theta) aligns with encoding and visual processing, while state 3 (dorsal alpha) maps onto maintenance and stable visuospatial representations, modulated by WM load and task demands. Distractor analyses indicate internal gating of posterior theta independent of stimulus identity. The computational model provides a mechanistic account: low-frequency synchronization, driven by thalamic inputs, shapes information routing via phase-amplitude coupling, with routing specificity determined by spatial network (posterior vs dorsal) and, in dorsal regions, by frequency (theta vs alpha). Dorsal alpha prioritizes intra-hemispheric frontoparietal routing consistent with maintenance and attention, whereas dorsal theta enhances interhemispheric transmission, potentially supporting manipulation. Posterior theta/alpha distinctions remain complex; both are implicated in visual processing, although alpha has documented roles in distractor suppression and eyes-closed states. The observed U-shaped relationship between state switching and performance resembles dopaminergic inverted-U effects, suggesting neuromodulatory involvement and fronto-basal ganglia/pulvinar mechanisms in state regulation.

The study identifies and characterizes four synchronized low-frequency networks forming functional brain states in vsWM, with posterior theta supporting encoding (flexibility) and dorsal alpha supporting maintenance (stability). Optimal control over transitions among these states correlates with improved cognitive performance. An in-silico whole-brain model demonstrates that thalamically driven synchronization can selectively route information and that spatial location and oscillatory frequency modulate transfer entropy, particularly in dorsal networks. These results advance understanding of how large-scale oscillatory states can optimize information flow to enable stable and flexible visuospatial representations. Future work should clarify functional differences between posterior theta and alpha, investigate frontal theta contributions under improved SNR or training, explore neuromodulatory and thalamocortical mechanisms controlling state transitions, and refine computational models to bridge simulated and empirical dynamics.

Frontal theta was not robustly observed in MEG vsWM data, possibly due to low signal-to-noise at frontal sensors or training-related emergence; a mid-frontal theta component appeared only in later sessions of the 4-subject dataset and was excluded. Not all networks were identifiable in every subject/session/task subset (63% identifiable after partitioning), which may limit generalizability. The 4-subject dataset is small for intra-individual reliability and task comparisons. Posterior theta vs alpha functional distinctions remain unresolved. Transfer entropy analysis required long continuous data and was therefore conducted in simulations rather than empirical MEG, limiting direct validation of high-frequency spiking dynamics. Some effects (e.g., load dependency in Odd One Out) were non-significant, suggesting task-dependent variability.