Repertoire of timescales in uni – and transmodal regions mediate working memory capacity

Working memory (WM) describes the dynamic process of maintenance and manipulation of information over a certain time delay. Neuronally, WM recruits a distributed network of cortical regions like the visual and dorsolateral prefrontal cortex as well as the subcortical hippocampus. How the input dynamics and subsequent neural dynamics impact WM remains unclear though. To answer this question, we combined the analysis of behavioral WM capacity with measuring neural dynamics through task-related power spectrum changes, e


Introduction
Within our daily activities we perform many different cognitive operations on various inputs, for instance manipulating, monitoring and storing specific stimuli in working memory.Working memory (WM) describes the dynamic process of maintenance and manipulation of information over short time delays despite new incoming interfering information (Li et al., 2022;Sreenivasan and D'Esposito, 2019).One key feature is WM capacity or WM span which has been conceptualized as a containment ability referring to the amount of information a system (e.g. short-term memory buffer) can possibly process and gather (Atkinson and Shiffrin, 1968).While in a more dynamic conceptualization, capacity is compared to a system's robustness, that is, how much the system can be loaded before it breaks down; or bandwidth, e.g., how much information can be processed or transmitted in a unit of time (Wenger and Townsend, 2000).Accordingly, a person with a higher capacity can easily manipulate higher and lower cognitive loads in working memory compared to a low-capacity person.The dynamic processes on the neural level that mediate working memory capacity remain yet unclear.Addressing this gap in our knowledge is the goal of our study.
Working memory operates through a neuronally distributed network of brain regions (Berger et al., 2019;Christophel et al., 2017;de Mooij-van Malsen et al., 2023;Li et al., 2022;Owen et al., 2005;Rose, 2020).Higher-order regions like the parietal or dorsolateral prefrontal cortex (DLPFC) as well as lower-order regions like visual cortex (Borders et al., 2022;Christophel et al., 2017;Gayet et al., 2018;Yonelinas, 2013;Zhao et al., 2022) are associated with WM performance.The role of the hippocampus in WM has been a subject of extensive research and debate, with multiple studies providing evidence both for and against its involvement in WM processes (Borders et al., 2022;Leszczynski, 2011;Shavitt et al., 2020;Tambini et al., 2023;Yonelinas, 2013).While visual WM as assessed in the visual match-to-sample task recruits the hippocampus (Von Allmen et al., 2013), the implication of the hippocampus in working memory in other tasks remains highly contentious, primarily attributed to its relatively low activation levels during various WM tasks (Courtney, 2022;Kessels and Bergmann, 2022;Peters and Reithler, 2022;Rose and Chao, 2022;Slotnick, 2023;Stern and Hasselmo, 2022;Wood et al., 2022).
Specific temporal patterns of neural activity have been described as characteristic in WM (Bhattacharya et al., 2022;Liu et al., 2023;Wasmuht et al., 2018).One process, key in mediating actual WM performance, is delay activity.Delay activity is a sustained neural activity over the retention delay (Christophel et al., 2017;Curtis and D'Esposito, 2003;D'Esposito and Postle, 2015;Li et al., 2022;Sreenivasan and D'Esposito, 2019).This delay activity is mainly found in parietal and frontal cortices (Li et al., 2022).Another dynamic process reported in WM literature is cortical short-term plasticity where the information is temporally maintained through synaptic weight changes.This dynamic neural functioning is referred to as 'activity-silent WM' (Beukers et al., 2021;Courtney, 2022) as it does not appear as blood-oxygenation-level-dependent (BOLD) signal increase as commonly measured via general linear modeling (GLM) of hemodynamic signal.To specify the activity-silent neural dynamics in more detail, we suggest the implication of a variety of neural dynamics during WM.Neural dynamics operate on various timescales, ranging from milliseconds over seconds (Cavanagh et al., 2020;Golesorkhi et al., 2021;Huang et al., 2018;Murray et al., 2014;Wengler et al., 2020).Different mental processes have been related to the recruitment of varying ranges of timescales (Wolff et al., 2022;Zilio et al., 2021).For instance, a shift towards slower timescales have been related to the loss of consciousness (Wolff et al., 2022;Zilio et al., 2021Zilio et al., , 2023) ) while the recruitment of multiple timescales has been shown in cognitive processes like attention (Zeraati et al., 2023).How the multiple timescales of the brain's neural activity mediate working memory capacity remains to be explored.
In addition to the brain's neural timescales, working memory capacity may also be mediated by the timescales of the inputs themselves, e.g., input dynamics.One key feature shaping neural activity is the task design (Mumford et al., 2014).External stimuli are typically presented with a certain temporal structure (e.g., longer or shorter, regular or irregular trial durations or intertrial intervals (Fiveash et al., 2020;Rhodes and Di Luca, 2016;Thunell and Thorpe, 2019;Zeithamova et al., 2017).The emerging temporal structure over the duration of a task with a series of trials is here referred to as 'task-periodicity'.It's regularity and frequency may be processed and ultimately encoded by the brain's neural activity in a corresponding frequency range, leading to periodic activations (Fiveash et al., 2020;Huettel et al., 2014;Klar et al., 2023bKlar et al., , 2023a)).Specifically, two recent studies demonstrated that inter-trial or inter-stimulus-intervals in the infra-slow frequency range (periods of 15.5 to 60 s) yield power increases in corresponding frequency ranges of the brain's power spectrum density (PSD) (Klar et al., 2023b(Klar et al., , 2023a)).They show that the loss of consciousness in anesthesia goes along with the loss of task-periodic power peaks in the brain's PSD (Klar et al., 2023b(Klar et al., , 2023a)).This leaves open whether such task-periodicity in specific single frequencies also modulates cognitive functions like working memory on both neural and behavioral levels.
To capture the implication of different timescales of neural dynamics, we will here focus on the frequency domain of the fMRI signal.The frequency domain is displayed using a power spectral distribution (PSD).The PSD is the distribution of power of different frequencies obtained via Fourier transformation of the signal.Hence it indicates if there is one predominant or multiple equally strong frequencies implicated.An important measure of summarizing this information is the median frequency (MF; for studies applying MF in fMRI, see (Golesorkhi et al., 2022a;Huang et al., 2018;Klar et al., 2023bKlar et al., , 2023a))).MF divides the power spectrum into two halves with an equal area under the curve and indicates the frequency of that dividing lineit thus provides an estimate of whether the PSD is tilted more towards the slower or faster frequency ranges (Golesorkhi et al., 2022a).Reflecting the power balance within the whole or global PSD across both slower and faster frequencies (measured in broadband), MF indexes the global range of the power spectrum (Huang et al., 2018).Is such range of the brain's PSD related to the range of WM capacity?
Considering the temporal nature of working memory (see above), we think that dynamic neural measures should yield important insights.Therefore, the goal of our fMRI study was to investigate how WM capacity is related to the brain's neural dynamics, that is its single and multiple timescales.For this, participants performed a working memory task with regularly presented trials.We will explore the task induced frequency power within the single more local PSD (narrowband), as well as the multiple timescales of the global power distribution across the whole frequency range as measured by MF in the PSD (broadband).We performed these analyses in typical working memory regions like DLPFC and hippocampus, as well as sensory regions like the primary visual area and the motor cortex (as reaction time was required inducing a regular motor response).
Our first specific aim consisted in investigating single (task-periodicity induced, narrowband) and multiple frequency power indices based on the PSD.Task-periodicity is estimated via the power at the frequency corresponding to the trial duration while the multiple frequencies are indexed by the MF which tests for the balance between slow and fast frequency power.Given the previous observations of task-periodicity (Fiveash et al., 2020;Klar et al., 2023bKlar et al., , 2023a)), we hypothesized a single power peak in specifically the visual cortex changes in a frequency range that corresponds to the trial durations of the VSTM task but not in the control recording.This would illustrate that the task-periodicity impacts neural activity in the same frequency in which the task was presented.The second specific aim was to link the neuronal dynamics to behavior.We hypothesized that the task-induced single power PSD peak and the global PSD characteristics are related to WM capacity.Specifically, we hypothesized that subjects with higher WM capacity show a better neural alignment to the task-periodicity as manifest in higher PSD peak in visual cortex.Secondly, these same subjects are expected to show a particular balance of slow and fast frequency power on the global PSD in especially those regions typically associated with WM, that is, hippocampus and DLPFC.

Participants
The inclusion criteria for participants are described in (Steffener et al., 2022).40 young and 40 old participants have been recorded in functional magnetic resonance imaging (fMRI).The final sample includes 36 subjects (18 female; Age: m = 36.1,SD = 23, min = 19.2,max = 80.8)..The high exclusion was necessary due to motion-censoring as the time dependent measures like MF highly sensitive to atypical fluctuations.(See Section 2.8.pre-processing).The study received ethical approval from the Research Ethics Board (REB) of the University of Ottawa, and all participants signed informed consent forms.

Task administration in MRI
The VSTM task, a delayed matching task as in (Nagel et al., 2009), consists of regularly presented trials and was chosen to study the relationship between the task's periodicity effect and memory capacity.One or several dots are presented for 2.5 s at random position on the screen.The number of dots presented varies in each block.Next, a mask is shown for 0.3 s before the participant is to keep the locations of the dots in memory for 3.2s.Then, a single dot was shown at one position of the screen for 2.5s and the participant was asked to indicate with a button press if there has been a dot at this location during the encoding period.The task is presented in blocks of 6 trials (each 9.5 s).The inter-block-interval is 23 s (20 s break, 3 s countdown before the next block).The whole task takes 329.5 s (with a TR of 1.11, this makes 362TR + 2TR at the end to capture the rather slow BOLD activity).Hence, working memory is solicitated periodically every 9.5 s (0.105Hz).
The load level is adapted to each subject's capacity.The capacity was assessed in an earlier, purely behavioral staircase procedure (see Steffener et al., 2022).The task was implemented in PsychoPy2 (Peirce et al., 2019).All software to deliver this task is publicly available at: https://github.com/NCMlab/CognitiveTasks .
To assure that the periodicity peak is related to the task structure and not to scanner noise or other nuisances, we chose another recording of the same subject and same recording session.This recording of similar duration and does not follow a periodic trial presentation of 9.5 s.Hence, we do not expect to find a task-periodicity peak at this frequency in the control task.
Two identical runs for each task were administered within the MRI.

Assessment of working memory capacity
Capacity was determined with a staircase method via the VSTM task before participants entered the scanner (Fig. 1A).Capacity was defined as the level with 80 % correct trials using a 3-up/1-down staircase.For details on the procedure, see (Steffener et al., 2022).Subsequently, the difficulty of the VSTM task in the scanner was adapted to each subject's capacity.Hence, capacity here is considered as cognitive repertoire (McIntosh and Jirsa, 2019).Hence, when faced with a specific trial, each participant realizes the task based on their own range or repertoire of behaviors (here WM capacity).For instance, a load level of 3 dots will be perceived as much more challenging for a participant with a low capacity while the same stimuli is perceived as simple for a participant with a capacity of 12. Within the same load level, one's ease to give the correct response is fluctuating due to dynamic nature of brain fluctuations.If not apparent in accuracy, it commonly emerges in reaction time Fig. 1.Behavioral procedure.A Participants WM capacity was assessed with the VSTM task before entering the scanner in a staircase procedure.B The difficulty of the VSTM task in the scanner was adapted to each participant based on their capacity.Participants were asked to remember the position of one or several dots before they were to indicate if the dot probe is at a position where a dot appeared during the encoding period.One trial takes 9.5 s.The VSTM is presented in blocks of 6 trials (each 9.5 s).The inter-block-interval is 23 s (20 s break, 3 s countdown before the next block).Hence, memory is solicitated every 9.5 s (0.105Hz) Hence, the VSTM task has a very rhythmic structure with a response expected every 9.5 s.
(RT).Longer RT is evidence of a current higher difficulty compared to shorter RT.
For some analysis we categorized the variable capacity.For this, we split the capacity in high, medium and low values.Groups were defined by to have an equal size to avoid groups with a very low sample size.

MRI data collection parameters
All neuroimaging used the 3T Siemens Biograph mMR MR-PET scanner at the Brain Imaging Centre (BIC) at the Royal Ottawa Mental Health Centre (ROMHC).Participants wore protective earplugs during the scans and held a squeeze ball they could activate if they felt uncomfortable and wished to terminate the scan.A multi-band accelerated EPI sequence (Moeller et al., 2010) using an acceleration factor of 6, TR = 1110ms, TE = 16.6ms,52-degree flip angle, phase partial Fourier 6/8, 56 slices collected in an alternating increasing slice order, 2.5 × 2.5 mm in-plane resolution, slice thickness = 2.75 mm, field of view: 220 × 200 mm (Steffener et al., 2022).A total of 365 volumes were recorded per run.

Pre-processing
All image preprocessing has been done in AFNI using afni_proc.py(https://afni.nimh.nih.gov)(Cox, 1996).For each task and subject, we applied slice timing correction.We applied motion censoring of volumes with Enorm (Euclidean norm of first differences of motion parameters) > 0.25 mm.Anatomical scans were normalized into MNI152 2009c stereotactic space and regression of linear and non-linear drift, and regression of local white matter signals to reduce non-neuronal noise (Jo et al., 2013).Finally, we applied a spatial smoothing using an 8 mm fullwidth at half-maximum isotropic Gaussian kernel and a bandpass between 0.01 and the Nyquist frequency of 0.45 Hz.The censored volumes were interpolated with a polynomial of degree 3.
If more than 10 % of data was removed during the preprocessing due to motion, the run was discarded.This strict removal is important because our measures are very timeseries sensitive.If several timepoints are missing in a row, the interpolation creates a smoothed section which impacts strongly measures like MF.

Definition of ROIs
ROIs were extracted based on Glasser et al. (Glasser et al., 2016).The primary visual area (V1) and the primary motor area (M1) were chosen in regard to the visual nature of the task requiring a motor response.The hippocampus and dorsolateral prefrontal cortex (DLPFC) were chosen due to their implication in WM (Borders et al., 2022;Christophel et al., 2017;Gayet et al., 2018;Yonelinas, 2013;Zhao et al., 2022;Courtney, 2022;Kessels and Bergmann, 2022;Peters and Reithler, 2022;Rose and Chao, 2022;Slotnick, 2023;Stern and Hasselmo, 2022;Wood et al., 2022).Voxels of these regions were extracted and calculations were computed on these voxels before averaging across voxels per ROI.

Periodogram
First, we computed the power spectrum for each voxel in the desired ROIs before it was averaged into one PSD per ROI and per subject.The PSD was computed via the signal.periodogramfunction with a hamming window in python 3.10.

Task-periodicity peak: a measure of a single timescale
The peak indicates the area under the curve of the narrowband power spectrum between 0.1 to 0.11Hz (9 to 10 s period, 4 frequency sampling points).This surrounds the frequency of the task periodicity of 9.5s.It was estimated with and integral approximation of the spectrum using Simpson's rule with the scipy.integrate.simpsfunction in python.
This measure considers only a part of the power spectrum.Hence, we call it a local variable.

MF-post
To capture the global MF independent of the power peak, we computed MF on the section of the PSD following the periodicity-peak, hence in the Hz range from 0.11 to 0.45Hz (136 frequency data points).The section on the PSD before the peak was too small for MF calculation.

Functional connectivity
To compute functional connectivity (FC) between the selected ROIs, mean time-series of each ROI were extracted from the pre-processed data.Next, we extracted the correlation coefficient between each ROI with the hippocampus, subject-based.

Statistical analyses
Mixed effect models were used to account for the repeated measures nature of ROI, hemisphere and run.Hence, these factors were always included as fixed effects, while subjects were considered as random factor.Linear mixed models with repeated measures were computed with the r package 'lme4' and the lmer() function (model <-lmer(DV ~ ROI*side*Run + (1|Subject), df).Further, Wilcoxon tests were chosen to compare the groups 'high' vs. 'low' due to their unequal variances.

Behavioral
We first show large inter-individual differences in both capacity and RT (Fig. 2A+B).Next, we correlated the capacity with RT across all runs and load levels.In both runs, RT relates similarly to the capacity (Run1: r = -0.52,p = 0.0027; Run2: r = -0.53,p = 0.0029; Fig. 2B).We conclude that both recorded runs are equivalent on the behavioral side.As in the literature, this sample also presents a faster RT with a higher capacity (Unsworth and Engle, 2005).
To specify the link between RT and capacity in relation to the load level, we defined equally sized groups into high, medium, and low based on their capacity (Fig. 1A right).Descriptively, Fig. 2C depicts how participants with high vs. low capacity respond to different degrees of task-difficulty (Load).Each participant was presented with five different load levels.We observe how participants with low capacity take a long RT on simple load levels while high-capacity participants show a short RT for these same load levels.To confirm this observation statistically, we found significant Wilcoxon signed-rank tests (p < 0.05) between high -and low-capacity participants for load levels 1 and 3 (Fig. 2D).Load levels 1 and 3 were presented to all participants, independent of their capacity level, hence they offer a possibility for direct comparison of all subjects.It is clearly visible that high-capacity participants respond faster to the same load levels than low-capacity participants.This indicates that the extend of the working memory capacity also affects the We observe significant differences in RT between participants with high compared to low capacity in terms of their mean RT to load levels 1 and 3.This indicates that participants respond differently to the same stimulus based on their own repertoire of capacity.
processing of low load items.Finally, it shall be mentioned that, as the VSTM task was administered twice in the scanner, we also investigated if both trials were equivalent in terms of response time (RT).No significant differences were found (p > 0.05, Fig. 2A).

Single Timescale (task-periodicity peak)
To assess the impact of the temporal structure of the task, e.g., taskperiodicity on neural activity, we compare the power spectrum density (PSD) from the VSTM task with the one of the control recording (Fig. 3A + B, non-log transformed PSD are in the supplements) at log(0.105Hz) (peak).The interaction effect of task and ROI on the peak of the mixed model was significant (F(3,766.93)= 63.244,p <.000).Post-hoc pairwise comparisons indicate a significantly stronger peak in V1 during the VSTM task compared to the control (p <.0001; Fig. 3C).Next, we shuffled the timeseries of the BOLD activity of both tasks and all ROIs and computed the peak.No significant difference between VSTM and control recording was found for any of the ROIs (p >.05; see supplements).This indicates a clear effect of the task-periodicity on neural activations in the sensory region V1.

Multiple Timescales (MF):
To capture properties of a wider range of neural fluctuations, we computed the PSD and extracted MF across all frequencies (broadband; Fig. 4A).MF is an indicator of the distribution of power between faster and slower frequencies.A mixed model indicates a significant main effect of ROI on MF (F(3,432.47)= 161.99,p <.0000).Bonferroni-corrected pairwise comparisons show significant differences between all ROIs except for M1 and DLPFC (Fig. 4B).V1 To test if this difference remains without the task-induced periodicity in the lower frequency of 0.105Hz, we computed the MF on the section of the PSD after the task-induced power peak (0.11-0.45Hz).Results remain consistent when excluding the slower frequencies (see supplements).Together, this indicates that regions clearly differ in their balance of higher and lower frequencies, with V1 demonstrating the strongest and the hippocampus weakest tendency towards higher power in slow frequencies.

Neuronal-behavioral relationshipslinking single and multiple timescales to capacity and reaction time
In the previous section we showed a clear impact of the taskperiodicity on neural BOLD-fluctuations in V1.The impact is centered to a single, local frequency (0.105Hz) of the PSD.Secondly, we investigated the global power distribution over the full PSD via the MF and observed clear differences between regions in their fast-slow frequency balance.Here, V1 and the hippocampus position themselves at the extremes, with the V1 demonstrating the highest expression of slow frequencies.How are these dynamic neuronal features related to behavior?

Single timescale and behavior
First, we will investigate the relationship between the taskperiodicity effect and behavior.For this, we computed several mixed models (see methods).Capacity and RT are related to the peak via an interaction effect with ROI while controlling for the effect of hemisphere and run (Capacity: F(3,418.67)= 4.96, p = .002;RT: F(3, 418.51) = 6.828, p = .000).A higher peak in the visual cortex is related to a better capacity and a faster RT (In Fig. 5A).Hence, the more the power spectrum in V1 adapts to the task periodicity, the better the subjects' performance.Or in other terms, a more adaptive V1 to external temporal information is related to a better memory capacity.

Multiple timescales and behavior
Next, we associated behavioral variables with the global PSD measure MF.A significant interaction effect between ROI and MF has been found for capacity (F(3,417.34)= 4.034, p < .007)and RT (F(3,417.20)= 4.603, p < .003).In Fig. 5B we observe the strongest link between MF and capacity as well as between MF and RT in the hippocampus.We confirm this by investigating MF on the PSD after the periodicity-related power peak (post-MF, 0.11-0.45Hz) to verify that this effect is not due to the peak itself (Fig. 5B.b).And indeed, the results with post-MF confirm those on the global MF.A higher MF in this study means a median frequency at about half of the total spectrum.This indicates that a more equilibrated relationship between the power in low and high frequencies goes jointly with better WM capacity.
Together, we observe differential regional involvement of single and multiple neuronal timescales in WM capacity.The sensory region V1 is related to WM capacity and RT via the encoding of the task-periodicity while the hippocampus relates to the WM variables via its balance of fast and slow frequencies.

Functional connectivity
WM has a cue or input component, a delay or maintenance component and an response or output component (Curtis and D'Esposito, 2003).These components need to coordinate among each other.We hypothesize that a better motor-hippocampus coordination/synchrony leads to faster RT.
To further established and support that the Hippocampus is really involved in WM, we conducted FC, indicating the degree of synchronization, between the hippocampus and the other three ROIs (Fig. 6, see supplements for correlation table).Next, we correlated the individual association strength between ROIs with the behavioral RT and capacity.And indeed, a highly significant correlation was found between the left hippocampus-left motor cortex synchronization with RT (r = -0.546,p < .001)and marginally with capacity (r = 0.327, p = 0.072).This indicates that a higher motor-hippocampus synchrony is associated with faster RT and better capacity.We conclude that the hippocampus is implicated in WM.

Hierarchy of timescales in WM
So far, we could show that capacity and RT are linked to the taskperiodicity (b95rel) via V1 and to the fast-slow frequency power balance (MF) in the hippocampus.Do both regions with their respective timescale impact working memory capacity equally?To answer this question, we computed a stepwise regression with RT as dependent, MF in hippocampus (MF_hippo) and the task-periodicity peak in V1 (peak_V1) as independent variable.Both independent variables have a significant effect on RT (MF_hippo: p < .001;peak_V1: p = .008)with an R 2 of 0.203 for MF_hippo and of 0.046 for peak_V1 (the full results of this regression are in the supplements).We computed the same analysis with capacity as dependent variable and found similar results (MF_hippo: p < .001,R 2 = 0.149; peak_V1: p = 0.013, R 2 = 0.44; full results are in the supplements).In this sense, the global distribution of power in faster and slower frequencies in the hippocampus is the main factor predicting RT and WM capacity.The stronger task-periodicity effect in V1 may rather play a role as a facilitator.

Mediation model
WM capacity was assessed before the main recording.RT indicates the actual performance on each trial, capturing trial-to-trial variations reflecting the dynamic repertoire of brain dynamics.Hence, we placed capacity as independent and RT as outcome variable (Fig. 7).As mediator we included the MF_of the hippocampus and the power spectrum peak in V1.Run and hemisphere were added as background confounders.The indirect path of the model with peak_V1 as mediator was not significant (p = 0.127).Instead, the mediative character of MF_hippo between the global capacity and the actual RT is significant (p = .005).Neither of the confounders influenced the model.(Full results of the model and the non-significant results of the indirect path RT → MF_hippo → capacity can be found in supplements).
A. Wolman et al.

Fig. 3. AþB
Power spectra of each ROI (in log scale for better visibility of the peak.See supplements for PSDs in linear scale).Each line depicts one subject, the black line indicates the average of all subjects.The upper row depicts the right hemisphere, the lower row the left hemisphere.The blue vertical dotted line indicates 0.105Hz (9.5 s), the red one 0.416Hz (2.5 s), indicative of the expected trial periodicity for the VSTM and Control recording, respectively.By visual inspection we remark a power peak at the expected task-periodicity for the VSTM task but not for the Control recording.C Boxplot of the peak measuring the relative power between 0.1-0.11Hzthat corresponds to the frequency of the task-periodicity in the VSTM task.Significant differences (Bonferroni corrected) indicate a higher power peak in the VSTM than the Control recording.This indicates that we find the expected effect of the task-periodicity in our dataset.

Discussion
Different mental processes have been related to the recruitment of varying ranges of timescales (Wolff et al., 2022;Zilio et al., 2021Zilio et al., , 2023)).For instance, the recruitment of multiple timescales has been shown in cognitive processes like attention (Zeraati et al., 2023) while EEG studies indicate a shift towards slower timescales during the loss of consciousness (Wolff et al., 2022;Zilio et al., 2021Zilio et al., , 2023)).In this study, we aimed to elucidate neural timescales implicated in WM.First, we showed that a higher memory capacity is related to a faster response time averaged across all load levels.This is commonly observed in memory tasks and hence validates our task administration (Bo et al., 2019).
Next, we tested for the implication of the hippocampus in WM (Schneider et al., 2017) via functional connectivity and found the synchrony between the left motor cortex and the left hippocampus to be significantly related to behavior.This validates the further inclusion of the hippocampus in the analysis.
Our first specific aim was the investigation of dynamic BOLD fluctuations implicated in WM capacity and the influence of the tasks' periodic structure on these dynamic fluctuations.Both neural measures were considered to provide information about the probabilistic functional modes underlying brain dynamics and behavior (McIntosh and Jirsa, 2019).Here we provide first evidence that the temporal characteristics of the task design plays a role in WM performance during input processing.Results showed that the regular task structure of 9.5 s (0.105Hz) clearly elicited repeatedly single BOLD fluctuations at 9.5 s as indexed by the local power peak.This predominantly affected the primary visual area.This is not surprising (Huettel et al., 2014) given that we have a visual task.Indeed, music studies and neural oscillators provide evidence of the ability of the brain to synchronize to a rhythmic external stimulation like tones (Doelling and Assaneo, 2021;Fiveash et al., 2020;Rhodes and Di Luca, 2016).Strikingly, the observed task-induced power peak was not independent of task performance.Here, a stronger neural response in the primary visual area to the task periodicity was observed in subjects with a higher memory capacity, suggesting its role in input processing.
The second aim of this study was to investigate if WM capacity is reflected in the dynamic neuronal fluctuations of multiple timescales as measured by MF.Different neural fluctuations have been linked to WM in different frequency bands (e.g., theta, gamma), inter-peak-intervals (Noguchi and Kakigi, 2020), delay activity (Christophel et al., 2017;Curtis and D'Esposito, 2003;D'Esposito and Postle, 2015;Li et al., 2022;Sreenivasan and D'Esposito, 2019) or activity-silent processes (Beukers et al., 2021;Courtney, 2022;Rose, 2020).These studies provide evidence of the importance of neural fluctuations in WM.Our results extend these findings by showing a close relationship between WM capacity and the balance of power between higher and lower frequencies in the hippocampus.Higher capacity and faster RT were found to be associated with a rather balanced power distribution across all frequencies, e.g., MF, in the hippocampus.Higher power in slower frequencies was linked to a decreased capacity and longer RT.This is in line with findings related to activity-silent WM (Beukers et al., 2021;Rose, 2020;Rose and Chao, 2022) where processes operate which are invisible to BOLD amplitude measures.We can conclude on a close relationship between the recruitment of multiple neural timescales in the hippocampus and fluctuations in the WM repertoire.
This work focused on the timescale of 9.5 seconds per trial.Future directions which alter the stimulus duration and retention duration will help tease apart what aspect of the task periodicity is the "most" important.The other line of evidence implies the global power spectrum's balance of slow and fast frequency power, e.g., MF in the hippocampus, suggesting its role in information storage or maintenance in shaping WM capacity via multiple timescales.

Limitations
Even though opting for a frequency domain approach instead of investigating task activations in the time domain is still quite rare in the field of neuroscience, it offers some interesting advantages.Power spectra analyses allow the investigation of recurrent fluctuations at specific frequencies and the relationship among them.This additional level of abstraction has great potential for new measures of neural dynamics and a new understanding of how brain and behavior are related.

Conclusion
We demonstrate how the input dynamics with its single timescales in the task's periodicity is processed in a unimodal region like the visual cortex and, importantly, relates to working memory capacity.While the more transmodal hippocampus relates to working memory capacity through multiple timescales with their balance of shorter and longer ones.This underscores the key relevance of the brain's different timescales for working memory capacity which, as our data show, interact in especially unimodal sensory regions with the timescales of the input itself, e.g., task periodicity.Timescales may thus provide the bridge or  In A.a, we confirm that participants with a higher peak in V1 activity also have a higher capacity.This is not the case for the peak in the hippocampus.B Relationship between multiple timescales (MF) and behavioral measures.Here, these are the multiple timescales in the hippocampus which are related to WM. B.b Relationship between capacity and post-MF.connection from the input dynamics over the neural dynamics to the output's cognitive function, e.g., working memory capacity.Accordingly, the dynamics of both input and brain shape the output, e.g., working memory which is seemingly shared by all three as their "common currency" as suggested in Spatiotemporal Neuroscience (Northoff et al., 2020a;Northoff et al., 2020b).
Finally, one important methodological implication of our results is that the study of cognitive processes and neural activities should adopt a broader view of the dynamics of both task and brain.This, for instance, means to incorporate multifaceted temporal and thus dynamic features of the task structure itself, e.g., its periodicity.While on the neural level both single more local (narrowband) timescales as well as more global timescales need to be distinguished and thus measured separately.This is further reiterated by our results showing a complex cognitive process like working memory requires both single and multiple timescales at different processing states.

Declaration of competing interest
The authors declare no conflicts of interest.

A
.Wolman et al.

Fig. 2 .
Fig. 2. Behavioral results.A Boxplots of RT in the VSTM task.RT does not differ between runs.B Similar correlations between RT and capacity can be observed in both runs.C Relationship between RT and load level, colored by high and low capacity for simplicity.Participants with a lower capacity range have a longer RT on low load trials compared to subjects with a high capacity.D Load levels 1 and 3 are shared across all subjects.We observe significant differences in RT between participants with high compared to low capacity in terms of their mean RT to load levels 1 and 3.This indicates that participants respond differently to the same stimulus based on their own repertoire of capacity.

Fig. 4 .
Fig. 4. A Illustration of MF calculated on the PSD in linear space.Lower values of MF indicate higher power in slow frequencies, hence a particular importance of slow frequencies components in the measured BOLD activity.Depicted are extreme (highest/lowest) power distributions observed in fMRI.B Descriptive plot for the multiple timescale measure Median Frequency in the VSTM task.We observe a low MF for V1 and a medium MF with a high variability in the hippocampus.

Fig. 5 .
Fig. 5. Single and multiple timescale measures in relation to behavior.A Capacity and RT relate to the single timescale task-periodicity peak in V1.P-values indicate a significant difference in the slope of V1 compared to M1.In A.a, we confirm that participants with a higher peak in V1 activity also have a higher capacity.This is not the case for the peak in the hippocampus.B Relationship between multiple timescales (MF) and behavioral measures.Here, these are the multiple timescales in the hippocampus which are related to WM. B.b Relationship between capacity and post-MF.

Fig. 6 .
Fig. 6.Heatmap of the synchronization strength between the mean timeseries of the hippocampus with the other ROIs.B Correlation between M1-Hippocampus synchronization and behavioral measures RT and capacity.

Fig. 7 .
Fig. 7. Mediation model suggesting that the global factor capacity influences RT via an equilibrated balance of faster and slower fluctuations in the hippocampus.