The compositional nature of the human alpha rhythm

Alpha rhythm, Individual alpha peak frequency, Neural oscillations, Neural fingerprint 12 13 14 15 16 17 18 19 20 21 22 Corresponding author: [email protected] 23 .CC-BY-NC 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted January 25, 2026. ; https://doi.org/10.64898/2026.01.24.701499doi: bioRxiv preprint 2

24 Resting-state brain activity is dominated by oscillations in the alpha band (8 –13 Hz). The dominant 25 frequency within this band, the individual alpha peak frequency (IAPF), has been widely used as a 26 global marker of inter-individual variability in perceptual and cognitive functions, as well as clinical 27 phenotypes. However, the traditional approach implicitly assumes a single oscillatory rhythm varying 28 along a parametric continuum. Here, we analyzed resting -state EEG data from more than 2000 29 participants across multiple independent datasets. We identify three distinct and highly stable alpha 30 components, or archetypes, that jointly and compositionally determine the IAPF. These components 31 exhibit dissociable age-related trajectories as well as distinct scalp topographies and cortical sources. 32 Together, these findings indicate that individual differences in alpha activity do not reflect variation 33 along a single frequency dimension, but rather differences in the relative contribution of discrete alpha 34 generators. This calls for a reevaluation of many reported associations between alpha frequency, brain 35 function, and clinical phenotypes. 36 37 .CC-BY-NC 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted January 25, 2026. ; https://doi.org/10.64898/2026.01.24.701499doi: bioRxiv preprint 3

38 Brain activity is dominated by oscillations in the 8 –13 Hz range, commonly referred to as alpha 39 activity (Figure 1A) [1]. Over the past century, parameters of these oscillations, and in particular the 40 individual alpha peak frequency (IAPF, i.e., the dominant frequency within the alpha band [2]), have 41 emerged as key biomarkers for predicting brain function and dysfunction [3]. 42 In non -invasive scalp electroencephalography (EEG), IAPF is typically estimated as an 43 average alpha peak across all electrodes or a subset (Figure 1B). This measure exhibits a striking 44 degree of inter -individual variability [4] alongside strong test –retest reliability [5,6], making it a 45 potential neurophysiological trait -like marker with both theoretical and clinical value. Indeed, 46 variability in IAPF has been linked to inter-individual differences in a variety of perceptual [7–9] and 47 cognitive tasks [10,11], as well as to general cognitive traits [12] and clinical conditions. In particular, a 48 systematic slowing of IAPF has been reported in conditions such as cognitive decline [13,14], epilepsy 49 [15], autism [16], Alzheimer’s and Lewy body dementia [17,18], Parkinson’s disease [19], schizophrenia 50 [8,11,20] and depression [21]. 51 The heterogeneity of reported associations and the limited understanding of the neural 52 mechanisms underlying scalp IAPF, however, have rendered it a largely transdiagnostic marker, with 53 limited functional and diagnostic specificity and weak causal insight [3]. For example, it remains 54 unclear whether inter -individual variability in IAPF reflects parametric variation within a single 55 dominant generator or instead arises from the superposition of multiple alpha generators with distinct 56 anatomical and functional roles , as suggested by several lines of research [22–30]. If multiple alpha 57 generators are involved, then the single scalar estimate of global IAPF, as traditionally used, provides 58 only a coarse and ambiguous summary, conflating many different combinations of underlying alpha 59 components into the same value. Such ambiguity may contribute to the limited specificity of global 60 IAPF measures and the heterogeneity of reported findings, thus, calling for a revision of how IAPF 61 is estimated, interpreted, and linked to behavior and pathology. 62 Here, in a large-scale sample of task-free EEG recordings drawn from multiple resting -state 63 datasets totaling over 2000 participants, we tested the hypothesis that alpha activity, and therefore the 64 estimated IAFP, is composed of multiple, partially independent components (Figure 1C). Using 65 unsupervised spectral decomposition, we assessed whether distinct alpha components with specific 66 scalp topographies and neural generators coexist within individuals, and evaluated their contribution 67 to inter-individual variability in IAPF and age [31–33], as well as their longitudinal stability, and their 68 cortical sources. Our results reveal highly stable and reproducible alpha “archetypes” , i.e., 69 components that are systematically conflated in conventional global IAPF estimates. Each individual 70 .CC-BY-NC 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted January 25, 2026. ; https://doi.org/10.64898/2026.01.24.701499doi: bioRxiv preprint 4 expresses a characteristic mixture of these alpha components, defining a composite neural trait 71 analogous to a fingerprint. We argue that variation in the relative weighting of these components can 72 substantially alter the interpretation of IAPF, motivating a new framework for alpha “profiling” that 73 enables more mechanistic and reliable links between alpha rhythms, behavior, brain function, and 74 clinical outcomes. 75

76 Resting-state EEG is characterized by a mixture of archetypical alpha components 77 We estimated the average scalp power spectral density (PSD) from resting-state EEG data in a sample 78 of 1244 participants spanning the age range from 5 to 89 years (TDBRAIN dataset [34]). The PSD 79 exhibited the typical pattern observed in resting-state EEG, characterized by a dominant peak in the 80 alpha band and a gradual decrease in power with increasing frequency, consistent with the so-called 81 aperiodic (1/f-like [35]) component (Figure 1A). When IAPF was estimated at the population level 82 using a conventional global scalp approach, thus, implicitly assuming a single dominant generator, 83 the mean IAPF across participants was 9.63 Hz (SD = 1.04), with substantial inter -individual 84 variability (Figure 1B). 85 To test whether global IAPF instead reflects a combination of multiple alpha components 86 (Figure 1C), we applied two complementary unsupervised spectral decomposition approaches. First, 87 we performed a frequency-domain principal component analysis (fPCA) [36–38], retaining components 88 that explained at least 5% of the population-level variance (see Methods). This analysis revealed three 89 distinct components within the alpha range (hereafter referred to as slow, middle and fast alpha 90 components), alongside a low-frequency non-alpha component reflecting power at lower frequencies 91 (Figure 1D). Together, these components accounted for 80.7% of the variance. 92 .CC-BY-NC 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted January 25, 2026. ; https://doi.org/10.64898/2026.01.24.701499doi: bioRxiv preprint 5 93 .CC-BY-NC 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted January 25, 2026. ; https://doi.org/10.64898/2026.01.24.701499doi: bioRxiv preprint 6 Figure 1. Resting-state EEG is characterized by a mixture of archetypical alpha components. A) EEG 94 power spectral density (PSD) typically shows an alpha peak (increased power at ~10 Hz), commonly 95 summarized by the alpha peak frequency. B) Individual alpha peak frequency (IAPF) is often treated as a 96 single, unified index, which varies widely across individuals (each dot represents one participant from the 97 TDBRAIN dataset [34]). C) However, IAPF may be misleading if it reflects a mixture of distinct alpha rhythms, 98 which can be disentangled using frequency -domain principal component analysis (fPCA) or archetypal 99 analysis applied to spectral profiles (see Methods). D) In the TDBRAIN dataset (N = 1244), fPCA identified 100 four main spectral components explaining ≥80% of the variance, including three components within the alpha 101 band. Shaded areas indicate ±1 standard deviation of the estimated loadings across test iterations in a hold-out 102 cross-validation procedure. E) Archetypal analysis identified four components explaining ≥95% of the 103 variance, with spectral profiles closely matching those recovered by fPCA. F) Proportion of participants 104 showing maximal expression of each archetypal spectral profile. G) Purity plot showing the proportion of 105 participants (y-axis) as a function of the dominant archetype weight (x -axis). The distribution peaks around 106 0.5, indicating that most participants express multiple archetypes with moderate contributions. H) Three-107 dimensional fPCA score plot (fPC1 –fPC3), with individual participants projected in fPCA space (scores 108 averaged across the scalp) and color -coded according to the dominant archetype. Archetype centroids are 109 indicated by stars. I) IAPF, estimated from the average scalp PSD, was significantly predicted by both the 110 slowest and fastest alpha spectral components (i.e., Components 2 and 4 in Figure 1D), with effects in opposite 111 directions. Error bars represent 95% confidence intervals for the linear coefficients (β) estimates. J) Age-related 112 variance was explained by changes across all spectral components, suggesting that aging reflects shifts in the 113 relative dominance of alpha components rather than a uniform slowing of IAPF. The solid line and shaded area 114 show predictions from a quadratic model with 95% confidence intervals. 115 116 Second, we applied archetypal analysis [39], an unsupervised method that represents each 117 individual in a dataset as a mixture of “pure type” profiles, to identify a set of prototypical spectral 118 patterns expressed in varying proportions across participants. Despite relying on different 119 mathematical principles than fPCA [39], this analysis also identified four components that explained 120 95.7% of the variance. The resulting archetypal spectral profiles closely resembled those recovered 121 via fPCA (Figure 1E), as indicat ed by strong correlations with the corresponding spectral loadings 122 (slow alpha: r = .87; middle alpha: r = .80; fast alpha: r = .77; low -frequency non-alpha: r = .86). 123 These profiles were expressed broadly across the population (Figures 1F and 1H), with a slightly 124 higher proportion of individuals showing maximal expression of the low -frequency non -alpha 125 component and the middle alpha component (Figure 1F). 126 We then quantified, for each participant, the strength of expression of each archetype using 127 the mixture weights estimated by the model —i.e., a “purity analysis” (see Methods) . Purity was 128 defined as the largest weight in each participant’s mixture vector, representing the extent to which the 129 participant’s PSD was associated with a single archetype or a combination of multiple archetypes. 130 The distribution of purity values peaked around 0.5, indicating that most participants expressed 131 several archetypes with moderate contributions (Figure 1G). This pattern confirmed that individuals 132 generally exhibit mixtures of these spectral profiles, and that no single archetype dominates strongly 133 for most participants. 134 .CC-BY-NC 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted January 25, 2026. ; https://doi.org/10.64898/2026.01.24.701499doi: bioRxiv preprint 7 Individual differences in alpha components drive variability in IAPF and aging 135 The results from both decomposition methods converged, indicating that the global IAPF typically 136 estimated at the scalp reflects a mixture of multiple alpha components rather than a single oscillator. 137 Thus, global IAPF likely captures the relative contributions of these distinct components and their 138 combination in each individual’s scalp EEG, rather than parametric variation of a unitary intrinsic 139 rhythm. 140 To directly test this possibility, we modeled each participant’s IAPF (N = 1244; estimated with 141 standard procedures from the average scalp PSD; see Methods) as a function of the component scores 142 obtained from our fPCA spectral decompositions (see Methods). IAPF was significantly predicted by 143 the slow and fast alpha components, with opposite regression weights ( slow alpha component: 𝛽 = -144 0.85, SE = 0.05, t = -15.15, p < .001; fast alpha component: 𝛽 = 0.61, SE = 0.06, t = 10.12, p < .001), 145 whereas the remaining components made negligible or no contributions (middle alpha component: 𝛽 146 = -0.11, SE = 0.06, t = -1.71, p = .08; low-frequency non-alpha component: 𝛽 = 0.02, SE = 0.06, t = 147 0.32, p = .74) (Figure 1I). On this basis, we constructed a simplified model using the difference 148 between the slow and fast alpha scores to predict IAPF. This contrast significantly predicted IAPF (𝛽 149 = 0.74, SE = 0.02, t = 33.38, p < .001) and accounted for up to 47.3% of its variance. Thus, variability 150 in IAPF is largely due to the relative contribution of two distinct alpha components. 151 To further prove this point, we focused on the known relationship between IAPF and age 152 [40,34,33]. As expected, age was a strong predictor of IAPF , and model comparison showed that a 153 quadratic model provided a substantially better fit than a linear model (linear: BIC = 10732; quadratic: 154 BIC = 3539). In the quadratic model, both the linear age term ( β = 0.04, SE = 0.006, t = 7.15, p < 155 .001) and the quadra tic term ( β = −0.0006, SE = 0.00008, t = −8.42, p < .001 ) were significant, 156 indicating a nonlinear trajectory in which IAPF increases during early adulthood, plateaus, and 157 declines in older age. This result, if taken blindly, would be suggestive of a general slowing of alpha 158 rhythms with age [25,41,42]. 159 However, when modeling age as a function of the four spectral components derived above, a 160 more nuanced scenario emerged (Figure 1J). For both the fast and middle alpha components, 161 quadratic models provided a better fit than linear models (fast alpha: BIC = −26.9 vs. 0.79; middle 162 alpha: BIC = −24.27 vs. −2.65). In both cases, age showed significant positive linear effects and 163 significant negative quadratic effects, consistent with inverted U -shaped trajectories (fast alpha, 164 linear: β = 0.02, SE = 0.004, t = 5.56, p < .001; quadratic: β = −0.0002, SE = 0.00004, t = −5.86, p < 165 .001; middle alpha, linear β = 0.02, SE = 0.004, t = 4.97, p < .001; quadratic: β = −0.0002, SE = 166 0.00004, t = −5.77, p < .001), indicating increases from early adulthood to midlife followed by 167 declines at older ages. In contrast, the slow alpha component was best described by a nonlinear model 168 .CC-BY-NC 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted January 25, 2026. ; https://doi.org/10.64898/2026.01.24.701499doi: bioRxiv preprint 8 showing an increasingly positive association with age (BIC = -17.9 vs. -3.2 for the linear model). 169 While the linear term alone was not significant ( β = −0.007, SE = 0.004, t = −1.74, p = .08), the 170 quadratic term was positive and significant (β = 0.0002, SE = 0.00005, t = 4.56, p < .001), consistent 171 with a progressive increase later in life. Importantly, this component accounted for the largest 172 proportion of age -related variance (adjusted R² = 0.68; Figure 1J). The low -frequency non -alpha 173 component also showed a nonlinear association with age (BIC = −25.3 vs 3.8 for the linear model ; 174 linear term: β = −0.03, SE = 0.004, t = −7.48, p < .001; quadratic term: β = 0.0003, SE = 0.00004, t 175 = 6.34, p < .001). 176 Collectively, these results show that aging is not associated with a uniform slowing of global 177 alpha activity, but rather with complex, component -specific changes in spectral contributions, 178 including a growing dominance of the slow alpha component. 179 Alpha components are stable individual traits, replicable across datasets 180 We ran the same analyses on an independent resting -state EEG dataset, the DVS dataset (N = 586; 181 age range 20–70) from a longitudinal study [43], which includes repeated recordings on the same day 182 and again after five years for a subset of participants. Using the same criteria and procedures applied 183 to the TDBRAIN dataset, we recovered the four main spectral components, which together explained 184 80.5% of the variance. The frequency profiles closely overlapped with those reported in the 185 TDBRAIN dataset (Figure 2A), with comparable archetypal expression patterns (Figure 2B; 186 Supplementary Figures 1A, 1B and 1C) and similar relationships with both IAPF and age 187 (Supplementary Figures 2A and 2B). 188 These results demonstrate that the spectral components are robust and replicable across 189 independent datasets, recorded in a different laboratory, with different participants, hardware, and 190 preprocessing pipelines. 191 .CC-BY-NC 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted January 25, 2026. ; https://doi.org/10.64898/2026.01.24.701499doi: bioRxiv preprint 9 192 Figure 2. Alpha components are stable individual traits, replicable across datasets. A) In the DVS dataset 193 (N = 586), fPCA identified four main spectral components explaining ≥80% of the variance, including three 194 components within the alpha band. As in Figure 1, shaded areas indicate ±1 standard deviation of the estimated 195 loadings across test iterations in a hold-out cross-validation procedure. B) Three-dimensional fPCA score plot 196 (fPC1–fPC3) as in Figure 1H. C) Short -term test –retest stability, quantified as the correlation between 197 component loadings estimated from EEG recordings separated by approximately 2 hours. The solid line and 198 shaded area show predictions from a linear model with 95% confidence intervals. D) Long -term test–retest 199 stability assessed over a 5-year interval (N = 198). E) Fingerprint analysis. Identifiability was quantified using 200 subject-by-subject correlations of spectral component factor scores across sessions. Higher within -subject 201 similarity (ISelf) relative to between -subject similarity (IOthers) indicates strong subject -specific reliability, 202 demonstrating that alpha spectral components constitute stable and distinct individual fingerprints. 203 204 We then exploited the longitudinal structure of the dataset to assess the test –retest reliability 205 of the spectral components. Applying the same spectral decomposition to each session, we quantified 206 the stability of individual component scores, that is, each component’s contribution to a participant’s 207 scalp PSD (see Methods). All components showed strong reliability, both within the same day (two 208 sessions separated by approximately two hours; Figure 2C; low-frequency non-alpha component: β 209 .CC-BY-NC 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted January 25, 2026. ; https://doi.org/10.64898/2026.01.24.701499doi: bioRxiv preprint 10 = 0.94, SE = 0.02, t = 35.98, p < .001; slow alpha: β = 0.89, SE = 0.02, t = 40.54, p < .001; mid alpha: 210 β = 0.81, SE = 0.02, t = 33.46, p < .001, fast alpha: β = 0.84, SE = 0.02, t = 36.31, p < .001) and across 211 five years (for 198 participants; Figure 2D; low-frequency non-alpha component: β = 0.87, SE = 0.03, 212 t = 25.92, p < .001; slow alpha: β = 0.88, SE = 0.03, t = 22.32, p < .001; mid alpha: β = 0.81, SE = 213 0.04, t = 18.29, p < .001, fast alpha: β = 0.85, SE = 0.04, t = 20.15, p < .001). The associations between 214 IAPF, age, and component scores, as well as the strong test -retest reliability were also reproduced 215 independently in each session (Supplementary Figures 2A, 2B, 2C and 2D; Supplementary Figures 216 3A and 3B). 217 At the individual level, the spatial pattern of component scores across electrodes was highly 218 specific. A fingerprinting analysis (see Methods) comparing within -subject correlations across 219 sessions with between -subject correlations revealed near -perfect identifiability (success rate = 220 99.88%; Figure 2D). In other words, spectral component topographies are trait-like and highly stable 221 for each participant over time. 222 Alpha spectral components map onto distinct spatial and cortical patterns 223 We replicated the results in another independent dataset, the LEMON dataset (N = 198; young: 20–224 35; older: 59–77 [44]), identifying the same four spectral components (Figures 3A and 3B ; 225 Supplementary 2D, 2E and 2F), which together explained 83.1% of the variance. These components 226 also exhibited similar associations with age and IAPF (Supplementary Figures 2E and 2F). 227 The scalp topographies of the four components differed (see Methods and Figure 3C). In 228 particular, the slow alpha component was expressed more strongly over occipito-temporal electrodes, 229 whereas the fast alpha component showed maximal expression over occipito-parietal electrodes 230 (Figure 3C). Comparable spatial patterns were also observed in all the other datasets (Supplementary 231 Figure 4; see also Supplementary Figure 5 for an additional replication in a smaller (N = 19) task-232 free EEG dataset). 233 .CC-BY-NC 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted January 25, 2026. ; https://doi.org/10.64898/2026.01.24.701499doi: bioRxiv preprint 11 234 .CC-BY-NC 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted January 25, 2026. ; https://doi.org/10.64898/2026.01.24.701499doi: bioRxiv preprint 12 Figure 3. Alpha spectral components map onto distinct spatial and cortical patterns . A) In the LEMON 235 dataset (N = 198), fPCA again identified four main spectral components explaining ≥80% of the variance, 236 including three components within the alpha band (results are displayed and color -coded as in Figures 1 and 237 2). B) Three-dimensional fPCA score plot (fPC1–fPC3), as in Figure 1H and 2B. C) Scalp EEG topographies 238 of each spectral component, estimated from fPCA factor scores and averaged across participants. D) Pairwise 239 spatial comparisons between component topographies demonstrate that the components are spatially distinct, 240 consistent with partially separable large -scale neural generators or networks. E) Source -level fPCA results 241 projected onto 68 cortical regions (see Methods) recovered the three posterior alpha components and revealed 242 distinct cortical expression patterns. F) Difference map between the slow and fast alpha components at the 243 source level, highlighting relative differences in cortical expression. The fast alpha component showed stronger 244 expression in dorsal occipito-parietal regions, whereas the slow alpha component was more expressed in lateral 245 occipital and inferior temporal areas, suggesting a spatial dissociation broadly consistent with dorsal –ventral 246 functional organization. 247 248 The LEMON dataset is multimodal and includes T1-weighted MRI for each participant. Using 249 forward models derived from individual anatomy, we reconstructed EEG signals at the cortical level, 250 estimated PSDs, and applied the same fPCA -based spectral decomposition used for the scalp data 251 (see Methods). Using the same criterion of at least 5% explained variance, this analysis identified 252 three alpha components in the source PSDs, together accounting for 87% of the variance, but did not 253 recover the low -frequency non -alpha component. The spectral profiles of these source -level 254 components closely matched those obtained at the scalp (Figure 3E). 255 To characterize their spatial organization, we averaged component scores across cortical 256 sources within the 68 regions defined by the Lausanne parcellation (scale 1; see Methods). All three 257 components showed strong expression in occipito-parieto-temporal regions (Figure 3E). Importantly, 258 the slow and fast alpha components exhibited distinct cortical distributions: the fast component 259 showed higher expression in dorsal occipito -parietal areas, whereas the slow component showed 260 higher expression in lateral occipital and inferior temporal regions (Figure 3F). 261 This dissociation confirms that the multiple alpha components identified at the scalp reflect 262 partially distinct cortical generators. 263 264

265 For more than a century, alpha activity has been considered among the key biomarkers of brain 266 function and dysfunction [45,3]. One of the most extensively investigated features, the IAPF, has been 267 used as a global measure to quantify the speed of alpha activity, implicitly assuming a continuum in 268 which parametric variations relate to inter-individual variability in a range of perceptual and cognitive 269 functions [46,47,10,9]. Here, across multiple independent datasets, we show that the conventional 270 estimate of IAPF does not reflect a unitary global rhythm. Instead, we show that IAPF emerges from 271 the superposition of multiple alpha components with distinct spectral and spatial characteristics. 272 .CC-BY-NC 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted January 25, 2026. ; https://doi.org/10.64898/2026.01.24.701499doi: bioRxiv preprint 13 These components are differentially expressed across individuals, and their relative contributions 273 form stable individual profiles. Inter -individual variability in alpha activity is therefore better 274 described as a compositional property rather than as parametric variation along a single -frequency 275 continuum. 276 Our main finding is that highly similar spectral profiles and archetypal expression patterns 277 can be consistently recovered across datasets that differ substantially in sample size, recording setup, 278 laboratory of origin, and preprocessing pipelines. Across all datasets, the same set of components 279 explained a large proportion of resting -state EEG variance, revealing that resting -state EEG is 280 composed of a mixture of spectral components, including three robust alpha -band components. To 281 our knowledge, this provides the first large-scale evidence that scalp EEG alpha activity is inherently 282 compositional, and that its component structure is stable and reproducible at the population level. 283 These findings provide strong empirical support for a body of previous work suggesting the 284 coexistence of multiple alpha rhythms rather than a single global generator. Evidence for multiple 285 alpha generators has been reported both at rest [23–25] and during active tasks [26–30], and was already 286 hypothesized decades ago [22]. Related ideas are also implicit in the long -standing practice of 287 subdividing alpha activity into “lower” and “upper” bands [48,2], as well as in reports of individuals 288 exhibiting multiple clearly identifiable alpha peaks in their power spectra [49]. However, such 289 distinctions have typically remained secondary and, at least in the context of estimating IAPF, have 290 largely been overshadowed by the dominant practice of relying on a single global measure. Our results 291 go beyond these earlier approaches by demonstrating that alpha activity is not merely divisible along 292 an arbitrary frequency boundary, but instead reflects a structured mixture of reproducible components 293 with distinct spectral properties, spatial organization, and stable individual expression patterns. 294 If IAPF reflects the weighted summation of multiple underlying alpha components, then 295 inferences drawn from the global IAPF alone are inherently ambiguous and, in some cases, potentially 296 misleading. In particular, a shift in IAPF cannot be straightforwardly interpreted as a uniform 297 speeding or slowing of alpha activity. Instead, it may reflect changes in the relative dominance of 298 distinct generators, such as slower versus faster alpha components, without any change in their 299 intrinsic frequencies. This fundamentally alters the interpretation of IAPF as a biomarker. 300 The association with age provides a clear example. Numerous studies have reported that IAPF 301 follows a non-monotonic trajectory across the lifespan and that the slowing observed in older adults 302 is associated with a general decline in cognitive abilities [42], leading to the interpretation of IAPF as 303 a neurophysiological marker of global brain functioning [10]. Our results suggest a different 304 interpretation: age -related changes in IAPF primarily reflect shifts in the relative expression of 305 specific alpha components rather than a homogeneous slowing of a single oscillatory process (see 306 .CC-BY-NC 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted January 25, 2026. ; https://doi.org/10.64898/2026.01.24.701499doi: bioRxiv preprint 14 also [50]). A similar argument applies to schizophrenia. Previous studies have reported a slower IAPF 307 in individuals with schizophrenia and have linked this effect to disease -related perceptual and 308 cognitive deficits [51–54]. However, prior work using the same spectral decomposition approach 309 employed here has shown that schizophrenia is instead associated with the emergence or enhancement 310 of a distinct lower-alpha/theta component with a specific spatial topography [38]. Taken together, these 311 findings clearly show how many results attributed to global changes in IAPF may instead arise from 312 alterations in the compositional structure of alpha activity, reflecting not simply “slower” alpha 313 activity, but a quantitatively different mixture of underlying generators. 314 As mentioned above, previous studies have already reported multiple alpha components with 315 patterns largely consistent with those observed here [24,38,50,55–57]. Building on this work, our large -316 scale investigation provides a general and decisive framework supporting these earlier findings, 317 demonstrating that the compositional structure of alpha activity constitutes a highly stable and reliable 318 neural trait. Importantly, we show that these components , and their relative contributions at the 319 individual level, can be robustly estimated even from relatively short resting -state recordings and 320 modest sample sizes, comparable to those commonly used in event -related EEG studies 321 (Supplementary Fig. 5). This provides a practical tool for moving beyond brain-behavior associations 322 based on a single, ambiguous global IAPF. This framework is also critical for causal intervention 323 approaches, such as magnetic or electric brain stimulation, which aim to modulate ongoing alpha 324 rhythms [58–62]. Our results indicate that both stimulation target sites and stimulation frequency should 325 be informed by the underlying compositional structure of alpha activity rather than by a unitary IAPF. 326 A key question naturally raised by our findings is about the functional roles, if any, to which distinct 327 alpha components can be mapped. 328 In source analysis, we found that the two components most strongly involved in explaining 329 variability in both IAPF and age were associated with sources located in occipito-parietal (fast alpha) 330 and occipito-temporal (slow alpha) regions. These source locations are consistent with distinct alpha 331 generators reported in previous studies using smaller samples [24], and have recently been linked to 332 different functional roles [63]. Their cortical distribution is also loosely reminiscent of the classic 333 distinction between dorsal and ventral visual pathways , suggesting that different alpha components 334 may support partially dissociable computational or functional roles, potentially related to distinct 335 streams of visuo-spatial processing, a hypothesis with important implications for future research. It 336 should be noted that while scalp -level fPCA systematically revealed an additional low -frequency, 337 non-alpha component, this component did not meet the variance -based selection criterion in source 338 space (5% explained variance), accounting instead for 4.6% of the variance. One possibility is that 339 this component is attenuated in source space because it reflects more spatially diffuse, low signal-to-340 .CC-BY-NC 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted January 25, 2026. ; https://doi.org/10.64898/2026.01.24.701499doi: bioRxiv preprint 15 noise activity. In contrast, alpha components remain spatially coherent and continue to dominate the 341 low-dimensional structure of source-level spectra. 342 We hope that focusing on the compositional nature of alpha activity, rather than on the global 343 IAPF, can provide a more fine-grained tool to help resolve inconsistent findings in the literature and 344 replication failures [64–68]. Essentially, we might have been looking at the wrong measure: true 345 associations between behavioral metrics, cognitive scores, or clinical phenotypes and alpha 346 components may have been conflated or weakened because these components were combined into a 347 single global IAPF estimate. This is also highly relevant for clinical applications, where IAPF has 348 been used, for example, to predict treatment [69,70], pain sensitivity [71], and for patient stratification 349 [72]. Similarly, from a methodological point of view, there are currently widespread methods to 350 parametrize spectral components and estimate IAPF or other properties of oscillatory and “aperiodic” 351 components in neural activity, which rely, however, on the assumption that global peaks are 352 representative of a single oscillatory process that combines with an aperiodic one. These approaches 353 require several analytical choices [73–76]. Our approach and findings show instead that distinct 354 components can be separated in a reliable and unsupervised data-driven way, allowing quantification 355 of their contribution at the scalp and individual -participant levels, variables that can be readily used 356 for brain–behavior predictions and as biomarkers in clinical research. 357 In sum, human alpha activity is composed of multiple rhythms with distinct generators, 358 resulting in a mixture of alpha archetypal components that combine in each individual as a highly 359 stable trait. Despite over a century of research, the functional roles of these components and their 360 generating mechanisms remain largely unknown, obscured by the traditional reliance on a single 361 IAPF estimate. We propose that compositional metrics of alpha activity should be employed to revisit 362 the extensive literature, opening new avenues for more precise, individual-specific profiling of alpha 363 rhythms. 364 365

366 Sample description 367 We used three publicly available large -scale datasets of resting -state EEG (RS -EEG) recordings, 368 comprising a total of 2085 participants (see Supplementary Material and Supplementary Figure 5 for 369 validation in an additional, smaller (N = 19) in-house dataset [77]). 370 The first dataset consists of RS -EEG recordings from 1274 participants, representing a 371 heterogeneous sample of healthy individuals and patients with various clinical diagnoses, with ages 372 ranging from 5 to 89 years (620 females, 654 males). These recordings were collected in Nijmegen, 373 .CC-BY-NC 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted January 25, 2026. ; https://doi.org/10.64898/2026.01.24.701499doi: bioRxiv preprint 16 the Netherlands, over a 20-year period as part of the Two Decades–Brainclinics Research Archive for 374 Insights in Neurophysiology (TDBRAIN; for details, see [34]). Each participant completed a 4-minute 375 RS-EEG recording, beginning with 2 minutes in an eyes -open (EO) condition and followed by 2 376 minutes in an eyes-closed (EC) condition. 377 The second dataset involves RS-EEG recordings from 608 healthy participants aged 20 to 70 378 years (376 females, 232 males), collected in Dortmund, Germany, as part of the Dortmund Vital Study 379 (DVS; for details, see [43]. For each p articipant, a 6 -minute RS-EEG was initially recorded, with 3 380 minutes in an EC condition followed by 3 minutes in a n EO condition (referred to as "pre -test" 381 recordings). The participants then performed a series of five cognitive tasks lasting approximately 382 two hours in total before completing a second 6 -minute RS-EEG recording, again consisting of 3 383 minutes in both EC and EO conditions (referred to as "post -test" recordings) . A subset of 208 384 participants (130 females, 78 males) returned for a second session, approximately five years later, 385 following the same protocol. 386 The third dataset contains RS-EEG recordings from 203 healthy participants, collected in 387 Leipzig, Germany, as part of the Leipzig Study for Mind -Brain-Body Interactions (LEMON; for 388 details, see [44]. Participants were divided into two age groups: 138 young adults aged 20 to 35 years 389 (42 females, 96 males) and 65 older adults aged 59 to 77 years ( 32 females, 33 males). Each 390 participant completed a 16-minute resting-state EEG recording with 16 alternating 60-second blocks 391 of EC and EO conditions, always starting with an EC block. In addition, individual structural 392 magnetic resonance imaging (MRI) data were acquired for each of these participants (see [44] for 393 details). 394 Data collection for the TDBRAIN, DVS, and LEMON datasets was conducted in accordance 395 with local ethics committee guidelines and the Declaration of Helsinki. All participants provided 396 written informed consent before data collection. For additional details, please refer to the original 397 studies [44,34,43]. 398 RS-EEG acquisition and preprocessing 399 RS-EEG recordings from the TDBRAIN dataset were obtained using either a Quickcap 400 (Compumedics, NC, USA) or an ANT -Neuro Waveguard Cap (Brain Products GmbH, Gilching, 401 Germany) equipped with sintered Ag/AgCl electrodes. The EEG setup comprised 26 channels 402 arranged according to the international 10 –10 system. Signals were recorded at a sampling rate of 403 500 Hz. A virtual ground was used during acquisition, with offline re -referencing to the averaged 404 mastoids (A1 and A2) and a ground placed at AFz. Skin–electrode impedance was maintained below 405 10 kΩ throughout the recordings. 406 .CC-BY-NC 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted January 25, 2026. ; https://doi.org/10.64898/2026.01.24.701499doi: bioRxiv preprint 17 For our study, we preprocessed the 2-minute EC segments from the TDBRAIN dataset using 407 an automated pipeline recommended by Delorme (2023) in EEGLAB (version v2024.2 [79]). First, 408 EEG recordings were downsampled from 500 Hz to 250 Hz and a 0.5 Hz high -pass Finite Impulse 409 Response (FIR) filter ( pop_eegfiltnew()) was applied. Next, we identified bad channels and noisy 410 data segments using the clean_rawdata plugin: Channels with correlation values below 0.85 and data 411 segments exceeding the Artifact Subspace Reconstruction (ASR [80]) threshold of 20 were removed. 412 Independent component analysis (ICA) was then performed using Infomax ( Picard plugin, 413 pop_runica()) and components associated with muscle and eye artifacts, identified with >90% 414 probability by the ICLabel algorithm, were removed. After preprocessing, an average of 1.9 415 electrodes, 9.4% of the data, and 0.7 independent components were discarded across the 2 minutes of 416 EC recordings. Finally, removed electrodes were interpolated using spherical spline interpolation 417 (pop_interp() [81]), all electrodes were re -referenced to the average reference ( pop_reref()), and the 418 clean data were epoched into 2-second windows. To ensure a minimal amount of data per participant, 419 we excluded 16 participants for whom the automated pipeline retained less than 1 minute of 420 recording. In addition, one participant was removed due to missing demographic and clinical 421 information, resulting in 1257 retained participants from the initial 1274. 422 RS-EEG data from the DVS dataset were recorded with a BrainVision BrainAmp DC 423 amplifier using a 64 -channel elastic cap (Brain Products GmbH, Gilching, Germany) , configured 424 according to the 10 –20 system, with the FCz electrode serving as the online reference. The EEG 425 signal was sampled at 1000 Hz and all electrode impedances were maintained below 10 kΩ. 426 From the DVS dataset, we considered and preprocessed the 3-minute EC segments, recorded 427 before and after the cognitive tasks, during the first session and, when available, the 3 -minute EC 428 segments from the second session 5 years later . We followed the same automated pipeline , as 429 described above for the TDBRAIN dataset, except that EEG recordings were downsampled here from 430 1000 Hz to 250 Hz. An average of 3.5 electrodes, 15.5% of the data, and 4.7 independent components 431 were discarded across the 12 minutes of EC recordings. Finally, we excluded participants if the 432 automated pipeline retained less than one minute of usable data in either or both EC segments of the 433 pre- or post -test recordings for any session . This resulted in 601 retained participants in the first 434 session (7 removed from the initial 608) and 206 participants in the second session (2 removed from 435 the initial 208). 436 RS-EEG data from the LEMON dataset were recorded with a BrainVision BrainAmp MR plus 437 amplifier using 61 active ActiCAP electrodes (Brain Products GmbH, Gilching, Germany), arranged 438 according to the 10-10 international system, with the FCz electrode serving as the online reference. 439 The ground electrode was positioned on the sternum, and all electrode impedances were kept below 440 .CC-BY-NC 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted January 25, 2026. ; https://doi.org/10.64898/2026.01.24.701499doi: bioRxiv preprint 18 5 kΩ. Data were continuously recorded with an online bandpass filter ranging from 0.015 Hz to 1 441 kHz and digitized at a sampling rate of 2500 Hz. 442 We used the EEG data preprocessed by Babayan and colleagues [44]. Specifically, after data 443 acquisition, EEG recordings were downsampled from 2500 Hz to 250 Hz, bandpass filtered between 444 1 and 45 Hz using an eighth-order Butterworth filter, and separated into EO and EC conditions. Next, 445 electrodes exhibiting frequent voltage shifts or poor signal quality were identified via visual 446 inspection and excluded. Similarly, data segments containing extreme peak -to-peak deflections or 447 high-frequency bursts were manually removed. Dimensionality reduction was then performed using 448 principal component analysis (PCA), retaining a sufficient number of principal components (N ≥ 30) 449 to explain 95% of the total variance and independent component analysis using Infomax ( Picard 450 plugin, pop_runica()) was applied to visually identify and remove components associated with eye 451 movements, blinks, or cardiac activity. Further information about EEG data acquisition and 452 preprocessing can be found in Babayan et al. [44]. In the present work, we analyzed only the EC 453 recordings, applying additional preprocessing steps that included interpolation of removed electrodes, 454 re-referencing to the average reference, and epoching into 4-second windows. Two participants were 455 excluded due to differing sampling rates in their original EEG recordings, resulting in 201 retained 456 participants (138 young adults and 63 older adults). 457 Although the automated preprocessing pipelines for the TDBRAIN and DVS datasets are 458 similar, substantial differences in sample size (1257 vs. 601 participants retained after preprocessing) 459 and the number of electrodes (26 vs. 64) provide additional control, ensuring that our results are not 460 driven solely by sample size or EEG coverage. The same applies to the LEMON dataset (201 461 participants retained; 61 electrodes) and an additional in -house dataset (19 participants; 128 462 electrodes, see Supplementary Material and Supplementary Figure 5), which were additionally 463 preprocessed using different pipelines and segmented into longer epoch durations (4 seconds vs. 2 464 seconds in TDBRAIN and DVS) . These differences further reduce the likelihood that our findings 465 are specific to, or artifacts of, a particular preprocessing pipeline. 466 Power spectral density 467 For each participant in all datasets, we estimated the power spectral density (PSD) at each electrode 468 and epoch within the 1 –40 Hz frequency range using a 1024 -point fast Fourier transform (fft(), 469 MATLAB R2024b). Prior to PSD estimation, resting -state EEG (RS -EEG) data were globally z -470 scored across all electrodes and epochs, and single epochs were demeaned. 471 To estimate the individual alpha peak frequency (IAPF), we employed the FOOOF algorithm, 472 which separates the 1/f background component from oscillatory peaks in the PSD [75]. FOOOF was 473 .CC-BY-NC 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted January 25, 2026. ; https://doi.org/10.64898/2026.01.24.701499doi: bioRxiv preprint 19 applied to the PSD averaged across electrodes and epochs, with the most prominent peak in the 7–14 474 Hz range identified as the IAPF. Participants for whom IAPF detection via FOOOF was unsuccessful 475 were excluded from further analysis. This resulted in final sample sizes of 1244 participants from the 476 TDBRAIN dataset (13 excluded); 586 and 593 participants for the pre- and post-test recordings from 477 the first session in the DVS dataset (15 and 8 excluded, respectively); 203 and 205 participants for 478 the pre- and post-test recordings from the second session in the DVS dataset (3 and 1 excluded, 479 respectively); 198 participants from the LEMON dataset (3 excluded); and 19 participants from the 480 in-house dataset (0 excluded). 481 Spectral decomposition via frequency-PCA 482 To estimate the spectral components characteristics of RS -EEG we applied the frequency -PCA 483 (frequency principal component analysis, fPCA) method proposed by Nakhnikian and colleagues [38] 484 (see also [36,37]). First, the PSD estimated at each electrode and for each participant, averaged over 485 epochs, was concatenated in a matrix: 486 487 𝑋 ∈ ℝ(𝑠×𝑒)×𝑓 488 [1] 489 where s is the number of subjects, e is the number of electrodes and f is the number of frequency bins 490 obtained from the PSD estimation (159 bins, restricted to the 1 -40 Hz range). Each row of 𝑋 491 corresponds to the PSD of a specific subject-electrode pair across frequencies. We then applied PCA: 492 493 𝑋 = 𝑈Σ𝑉𝑇 494 [2] 495 where 𝑈 ∈ ℝ(𝑠×𝑒)×(𝑠×𝑒) contains the subject-electrode loadings, Σ ∈ ℝ(𝑠×𝑒)×𝑓 is the diagonal matrix 496 of singular values, and V ∈ ℝ𝑓×𝑓contains the principal components, representing spectral patterns 497 across frequencies. The first few principal components in V describe dominant spectral modes, while 498 the corresponding subject -electrode loadings in 𝑈 indicate how strongly each subject -electrode 499 combination contributes to each spectral component. A PCA was initially conducted to determine the 500 number of components to retain, using the criterion of retaining components that explained at least 501 5% of the variance in the PSD. This resulted in the retention of four components in all analyses 502 performed (except for source reconstruction), which combined always retained more than 80% of the 503 variance. 504 After performing a subsequent PCA with the retained components, we applied Varimax 505 rotation following the approach in Nakhnikian and colleagues [38], i.e., using the PCA toolbox [82] 506 .CC-BY-NC 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted January 25, 2026. ; https://doi.org/10.64898/2026.01.24.701499doi: bioRxiv preprint 20 (doPCA(), configured for covariance-based PCA with unscaled Varimax rotation and without Kaiser 507 normalization). This method maximally separates the factors while maintaining orthogonality, 508 resulting in PCs that are strongly correlated with the variables they explain and weakly correlated 509 with others. This facilitates the interpretation of the results in the original PSD space (e.g., Figures 510 1D). 511 Spatial expressions of each spectral component were derived from the factor scores obtained 512 during fPCA. Factor scores were reshaped into s × e × k arrays (where 𝑘 indexes component), 513 allowing scalp loadings to be estimated by averaging factor scores across participants for each 514 electrode and component (e.g., Figures 3C). These averaged factor scores were interpreted as spatial 515 weights reflecting the relative contribution of each electrode to a given spectral component. In 516 addition, scalp-averaged factor scores were computed at the single -subject level to provide a global 517 measure of the contribution of each component to an individual’s scalp PSD , which was then used 518 for predictive models (e.g., Figure 1J). 519 Unlike classic global IAPF approaches, which typically collapse spectral information across 520 electrodes or rely on a predefined subset, the present method explicitly models the joint spectral –521 spatial structure of RS-EEG. That is, by decomposing PSDs at the scalp level, fPCA allows multiple 522 alpha-related components with distinct spectral profiles and spatial distributions to be identified 523 simultaneously. 524 The estimation of spectral components was embedded within a hold -out cross -validation 525 framework. On each iteration, participants were randomly split into independent training and test sets 526 (50% each), and this procedure was repeated 1000 times. For each iteration, fPCA was performed 527 separately on the training and test sets, retaining the same number of components as determined in 528 the full dataset. This procedure was designed to assess the generalizability of the extracted 529 components and to ensure that the identified structure was not driven by sample-specific noise. 530 Component generalizability was quantified by computing Pearson correlations between the 531 spectral loadings obtained from the training and test sets. To assess whether these correlations 532 reflected genuine spectral structure rather than methodological artifacts, we compared them against a 533 null distribution derived from surrogate data. Surrogate datasets were generated by independently 534 shuffling PSD values across frequencies for each participant and electrode (1000 repetitions), 535 preserving overall power distributions while disrupting frequency-specific structure. fPCA was then 536 applied to the surrogate data using the same cross-validation procedure. 537 Across all datasets, the observed train –test correlations fell within a .64–.99 range and were 538 consistently higher than those obtained from surrogate data (-.14–.10), with all comparisons reaching 539 statistical significance (paired t -test with Fisher z -transformation; all ps < .001). The standard 540 .CC-BY-NC 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted January 25, 2026. ; https://doi.org/10.64898/2026.01.24.701499doi: bioRxiv preprint 21 deviation of each estimated factor across the iterations in the test set was then used as a disper sion 541 measure for the plots in Figures 1D, 2A and 3A, and Supplementary Figure 5A. 542 Archetypal analysis 543 To further characterize dominant spectral profiles at the group level, we applied archetypal analysis, 544 implemented via Principal Convex Hull Analysis (PCHA() [39]), to RS-EEG PSD. Unlike PCA, which 545 identifies orthogonal components that maximize variance, archetypal analysis represents data as 546 convex combinations of a small set of extreme patterns (“archetypes”), which lie on the boundary of 547 the data cloud and can be interpreted as prototypical spectral profiles. 548 For this analysis, PSD estimates were first averaged across electrodes for each participant, 549 yielding a subject-by-frequency matrix 𝑋 ∈ ℝ𝑠×𝑓, where 𝑠 denotes the number of subjects and 𝑓the 550 number of frequency bins. To ensure non -negativity and comparability across participants, PSD 551 values were rescaled to the [0,1] interval prior to analysis. Archetypal analysis was then performed 552 on the transposed matrix 𝑋⊤, such that each archetype corresponds to a spectral profile across 553 frequencies. 554 Formally, this analysis approximates the data matrix as: 555 556 𝑋 ≈ 𝑋𝐶𝑆 558 [3] 557 where the columns of 𝑋𝐶 define the archetypes, 𝐶 is a non-negative matrix whose columns sum to 559 one and select convex combinations of observed spectra to form archetypes, and 𝑆 is a non-negative 560 coefficient matrix whose columns sum to one and describe how strongly each archetype contributes 561 to each participant’s spectrum. This formulation ensures that both archetypes and subject -level 562 reconstructions remain within the convex hull of the observed data. 563 The number of archetypes 𝑘 was selected in a data-driven manner by evaluating models with 564 𝑘 = 2 to 10 archetypes and computing the fraction of variance explained by each solution. The 565 smallest 𝑘 accounting for more than 95% of the total variance was retained for subsequent analyses. 566 As in the fPCA approach, this resulted in 𝑘 = 4—i.e., four main archetypes. Final model estimation 567 was performed using this value of 𝑘. Note that, in contrast to fPCA, which captures dominant modes 568 of variance, archetypal analysis emphasizes extreme spectral patterns and represents individual 569 participants as mixtures of these prototypical profiles. Nevertheless, both approaches converged to 570 the identification of four main components/archetypes. For consistency and graphical purposes, we 571 reordered the archetypes based on their close correspondence with the spectral components identified 572 via fPCA. 573 .CC-BY-NC 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted January 25, 2026. ; https://doi.org/10.64898/2026.01.24.701499doi: bioRxiv preprint 22 To characterize how strongly individual participants expressed specific archetypal spectral 574 profiles, we quantified the dominance and purity of archetypal representations based on the subject-575 level weight matrix 𝑆. Each column of 𝑆 contains non-negative weights summing to one and describes 576 the convex combination of archetypes used to reconstruct a given participant’s scalp -averaged PSD. 577 We identified the dominant archetypes by taking the index of the maximum value, per participant, in 578 the related row of 𝑆. Purity was then defined as the corresponding maximum weight. 579 We assessed the distribution of dominant archetypes across participants by computing the 580 proportion of participants for whom each archetype had the highest weight (e.g., Figure 1F). This 581 analysis provides a view of how frequently each archetypal spectral profile serves as the primary 582 contributor to individual PSDs. The distribution of purity values across participants, visualized using 583 a probability-normalized histogram (e.g., Figure 1G), allows instead to assess whether the population 584 is dominated by individuals with highly archetype -specific spectral profiles or, instead, by more 585 mixed expressions. For example, a distribution skewed toward high purity values (e.g., purity = 1) 586 would suggest that the identified archetypes capture relatively discrete and internally coherent modes 587 of spectral patterns that tend to dominate individual PSDs, with a high clustering of the data around 588 the main archetypes. In contrast, a distribution centered at lower purity values would indicate that 589 individual PSDs are more consistently composed of combinations of multiple archetypes, consistent 590 with a more continuous or overlapping organization of spectral components across participants. 591 Together, these two analyses address complementary questions: dominance reflects how often 592 each archetype is the primary contributor at the individual level, whereas purity reflects how 593 exclusively an archetype characterizes individual spectral profiles. 594 To visualize how participants are distributed in the reduced spectral space of 595 components/archetypes, we projected individual archetypal expressions into the first fPCs obtained 596 from the fPCA and represented each participant as a point in a three -dimensional fPC space ( e.g., 597 Figure 1E). Each point therefore corresponds to a participant’s position along the dominant spectral 598 component axes, capturing the major sources of inter -individual spectral variability as derived from 599 fPCA. Participants were color-coded according to their dominant archetype, defined as the archetype 600 with the highest weight for that individual. This visualization allows a qualitative assessment of how 601 archetypal dominance relates to the geometry of the spectral space, that is, whether participants 602 associated with different archetypes occupy distinct regions, form clusters, or instead exhibit smooth 603 transitions across the fPC dimensions. 604 .CC-BY-NC 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted January 25, 2026. ; https://doi.org/10.64898/2026.01.24.701499doi: bioRxiv preprint 23 Relationships between IAPF, age and the spectral components 605 IAPF is known to vary systematically across the lifespan, increasing during early development and 606 declining in later adulthood [34,33]. As an initial validation step, we confirmed this well -established 607 relationship in the TDBRAIN dataset. Linear and quadratic regression models were fitted using the 608 fitlm() function in MATLAB (R2024b), with IAPF as the dependent variable and age as the predictor. 609 Model comparison was performed using the Bayesian Information Criterion (BIC), and the model 610 with the lower BIC was retained. For the selected model, regression coefficients (β), standard errors, 611 test statistics, and associated p values for the age terms were reported. 612 To further determine whether distinct age -related changes were present in each spectral 613 component identified by the fPCA analysis in the TDBRAIN dataset, linear and quadratic regression 614 models were also fitted, with the score of each spectral component as the dependent variable and age 615 as the predictor. Model selection was again based on BIC, with the model exhibiting the lower BIC 616 retained. For each selected model, regression coefficients ( β), standard errors, test statistics, and p 617 values for the age terms were reported. Adjusted R² values were also reported to characterize the 618 proportion of age-related variance explained by each component. 619 In addition, we assessed the relationship between IAPF and the spectral components in the 620 TDBRAIN dataset. A multiple linear regression model was fitted with IAPF as the dependent variable 621 and the four spectral components identified by the fPCA as simultaneous predictors. For this model, 622 regression coefficients ( β), standard errors, test statistics, and p values were reported for each 623 predictor. Given the strong and opposing contributions of the two dominant alpha-related components 624 (corresponding to slow and fast alpha; see Figure 1I), an additional analysis was conducted to quantify 625 their shared relationship with IAPF. A contrast term was computed as the difference between the slow- 626 and fast-alpha component scores, and a separate linear regression model was fitted with this contrast 627 as the sole predictor of IAPF. 628 Similar analyses examining the relationships between the spectral components, age, and IAPF 629 were also conducted using the DVS and LEMON datasets (see Supplementary Figure 2). 630 Stability and fingerprinting analysis 631 As mentioned, the DVS dataset included multiple RS -EEG recordings from the same individuals, 632 with two recordings on the same day (i.e., pre - and post-test recordings from the first session) and, 633 for a subset of participants, two additional recordings approximately five years later (i.e., pre - and 634 post-test recordings from the second session). This longitudinal design allowed us to evaluate the 635 stability of spectral components across both short - and long-term intervals. Factor scores derived 636 from the fPCA decomposition were averaged across electrodes for each participant and session. 637 .CC-BY-NC 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted January 25, 2026. ; https://doi.org/10.64898/2026.01.24.701499doi: bioRxiv preprint 24 Within-session stability was quantified by correlating factor scores between pre - and post -test 638 recordings on the same day. Specifically, we used linear regression to model the factor scores in one 639 session as a function of those in the other, and adjusted R² values provided a measure of the reliability 640 of each spectral component. Longitudinal stability across sessions separated by several years was 641 assessed in the same manner . Only participants with data available for both recordings in a given 642 comparison were included, resulting in 582 participants for pre- vs post-test recordings from the first 643 session and 198 participants for pre-test recordings from the first session vs pre-test recordings from 644 the second session (see Supplementary Figure 3 for replicated results using other recording 645 combinations). 646 To assess the uniqueness of individual spectral profiles, we performed a fingerprint analysis 647 [83,84]. We computed a subject -by-subject correlation matrix comparing factor score topographies 648 across two sessions (N = 198). The diagonal elements of this matrix (ISelf) quantify, for each 649 participant, the correlation between their own spectral profile across sessions. In contrast, the mean 650 correlation between a participant’s profile and those of all other participants (IOthers) captures 651 between-subject similarity. The difference between the population means of ISelf and IOthers (IDiff) 652 provides an estimate of subject-specific reliability, while the proportion of participants for whom ISelf 653 exceeds IOthers defines the identifiability success rate. High IDiff values and high success rates 654 indicate that individual spectral profiles are both stable over time and distinguishable across 655 participants, supporting the use of spectral component factor scores as reliable individual 656 “fingerprint”. 657 EEG inverse modeling 658 The LEMON dataset includes individual structural magnetic resonance imaging (MRI) data acquired 659 from the same participants who underwent RS-EEG recordings (see [44] for detailed acquisition 660 parameters). Anatomical T1-weighted MRI data were preprocessed using the Connectome Mapper 3 661 (v3.1.0; [85]) open -source pipeline with FreeSurfer (v7.1.1) software package . Because digitized 662 electrode positions were not available for all participants , we employed a standard 64 -channel 663 BioSemi EEG cap template. This template montage was co-registered to a template anatomical MRI 664 in MNI space, and then warped to each individual's native space using the nonlinear transformation 665 matrix from MNI to individual anatomy. 666 T1-weighted MRI data were segmented to extract scalp, skull, and brain surfaces. V olume 667 conduction models were constructed using the boundary element method (BEM) implemented in 668 OpenMEEG [86]. Source models were defined using a regular 7 mm grid of equivalent current dipoles 669 constrained to the cortical gray matter of the MNI template brain, yielding 2501 solution points in 670 .CC-BY-NC 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted January 25, 2026. ; https://doi.org/10.64898/2026.01.24.701499doi: bioRxiv preprint 25 cortical tissue. Each participant’s anatomical image was then registered to the MNI template, and the 671 inverse deformation field was applied to project the MNI grid into individual anatomical space [87,88]. 672 Forward models were computed for each participant using unconstrained dipole orientations, 673 allowing three orthogonal dipole components per location. EEG source activity was reconstructed 674 using standardized low -resolution brain electromagnetic tomography (sLORETA) [89] with a 675 regularization parameter λ = 10% . For each dipole and direction, the PSD was computed per epoch, 676 then averaged across the three dipole orientations at each solution point using the same method as in 677 the scalp-level analysis. 678 Single-subject PSD data from all solution points were concatenated and analyzed using the 679 fPCA approach described above, replacing electrode-level input with source-level data. Components 680 were retained using the same criterion of explaining at least 5% of variance, which in the source 681 analysis yielded three components corresponding to the three alpha -related sources, excluding the 682 low-frequency non-alpha component identified at the scalp. Seven participants were excluded from 683 the source-level analysis due to failed anatomical segmentation or forward model computation, likely 684 caused by the facial anonymization procedure in the LEMON dataset that affected tissue 685 segmentation and head model accuracy. For visualization, component scores were averaged within 686 68 cortical regions of interest defined by the Lausanne parcellation (scale 1) [90], which is derived 687 from the Desikan–Killiany anatomical atlas [91]. Only cortical gray-matter regions were included. 688 689 .CC-BY-NC 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted January 25, 2026. ; https://doi.org/10.64898/2026.01.24.701499doi: bioRxiv preprint 26

690 1. Berger, H. (1929). Über das elektroenkephalogramm des menschen. Archiv für psychiatrie und 691 nervenkrankheiten, 87(1), 527–570. 692 2. Klimesch, W. (1999). EEG alpha and theta oscillations reflect cognitive and memory performance: 693 A review and analysis. Brain Research Reviews , 29(2–3), 169 –195. 694 https://doi.org/10.1016/S0165-0173(98)00056-3 695 3. Pascucci, D., Menétrey, M. Q., Passarotto, E., Luo, J., Paramento, M., & Rubega, M. (2025). EEG 696 brain waves and alpha rhythms: Past, current and future direction. Neuroscience & 697 Biobehavioral Reviews, 176, 106288. https://doi.org/10.1016/j.neubiorev.2025.106288 698 4. Haegens, S., Cousijn, H., Wallis, G., Harrison, P. J., & Nobre, A. C. (2014). Inter - and intra-699 individual variability in alpha peak frequency. NeuroImage, 92, 46 –55. 700 https://doi.org/10.1016/j.neuroimage.2014.01.049 701 5. Grandy, T. H., Werkle-Bergner, M., Chicherio, C., Schmiedek, F., Lövdén, M., & Lindenberger, U. 702 (2013). Peak individual alpha frequency qualifies as a stable neurophysiological trait marker 703 in healthy younger and older adults: Alpha stability. Psychophysiology, 50(6), 570 –582. 704 https://doi.org/10.1111/psyp.12043 705 6. Popov, T., Tröndle, M., Baranczuk‐Turska, Z., Pfeiffer, C., Haufe, S., & Langer, N. (2023). Test–706 retest reliability of resting‐state EEG in young and older adults. Psychophysiology, 60(7), 707 e14268. https://doi.org/10.1111/psyp.14268 708 7. Arioli, M., Mattersberger, M., Hoehl, S., & Brzozowska, A. (2024). Peak alpha frequency is linked 709 to visual temporal attention in 6 -month-olds. Scientific Reports , 14(1), 28173. 710 https://doi.org/10.1038/s41598-024-79129-0 711 8. Menétrey, M. Q., Roinishvili, M., Chkonia, E., Herzog, M. H., & Pascucci, D. (2024). Alpha peak 712 frequency affects visual performance beyond temporal resolution. Imaging Neuroscience, 2, 713 1–12. https://doi.org/10.1162/imag_a_00107 714 .CC-BY-NC 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted January 25, 2026. ; https://doi.org/10.64898/2026.01.24.701499doi: bioRxiv preprint 27 9. Samaha, J., & Romei, V . (2024). Alpha-Band Frequency and Temporal Windows in Perception: A 715 Review and Living Meta -analysis of 27 Experiments (and Counting). Journal of Cognitive 716 Neuroscience, 36(4), 640–654. https://doi.org/10.1162/jocn_a_02069 717 10. Grandy, T. H., Werkle-Bergner, M., Chicherio, C., Lövdén, M., Schmiedek, F., & Lindenberger, 718 U. (2013). Individual alpha peak frequency is related to latent factors of general cognitive 719 abilities. NeuroImage, 79, 10–18. https://doi.org/10.1016/j.neuroimage.2013.04.059 720 11. Ramsay, I. S., Lynn, P. A., Schermitzler, B., & Sponheim, S. R. (2021). Individual alpha peak 721 frequency is slower in schizophrenia and related to deficits in visual perception and cognition. 722 Scientific Reports, 11(1), 17852. https://doi.org/10.1038/s41598-021-97303-6 723 12. Grandy, T. H., Werkle-Bergner, M., Chicherio, C., Lövdén, M., Schmiedek, F., & Lindenberger, 724 U. (2013). Individual alpha peak frequency is related to latent factors of general cognitive 725 abilities. NeuroImage, 79, 10–18. https://doi.org/10.1016/j.neuroimage.2013.04.059 726 13. Finley, A. J., Angus, D. J., Knight, E. L., Van Reekum, C. M., Lachman, M. E., Davidson, R. J., 727 & Schaefer, S. M. (2024). Resting EEG Periodic and Aperiodic Components Predict 728 Cognitive Decline Over 10 Years. The Journal of Neuroscience , 44(13), e1332232024. 729 https://doi.org/10.1523/JNEUROSCI.1332-23.2024 730 14. Lejko, N., Larabi, D. I., Herrmann, C. S., Aleman, A., & Ćurčić -Blake, B. (2020). Alpha Power 731 and Functional Connectivity in Cognitive Decline: A Systematic Review and Meta-Analysis. 732 Journal of Alzheimer’ s Disease, 78(3), 1047–1088. https://doi.org/10.3233/JAD-200962 733 15. Janiukstyte, V ., Kozma, C., Owen, T. W., Chaudhary, U. J., Diehl, B., Lemieux, L., Duncan, J. S., 734 Rugg-Gunn, F., De Tisi, J., Wang, Y ., & Taylor, P. N. (2024). Alpha rhythm slowing in 735 temporal lobe epilepsy across scalp EEG and MEG. Brain Communications, 6(6), fcae439. 736 https://doi.org/10.1093/braincomms/fcae439 737 16. Dickinson, A., DiStefano, C., Senturk, D., & Jeste, S. S. (2017). Peak alpha frequency is a neural 738 marker of cognitive function across the autism spectrum. European Journal of Neuroscience, 739 9. 740 .CC-BY-NC 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted January 25, 2026. ; https://doi.org/10.64898/2026.01.24.701499doi: bioRxiv preprint 28 17. Babiloni, C., Jakhar, D., Tucci, F., Del Percio, C., Lopez, S., Soricelli, A., Salvatore, M., Ferri, 741 R., Catania, V ., Massa, F., Arnaldi, D., Famà, F., Güntekin, B., Yener, G., Stocchi, F., Vacca, 742 L., Marizzoni, M., Giubilei, F., Yıldırım, E., … Noce, G. (2024). Resting state 743 electroencephalographic alpha rhythms are sensitive to Alzheimer’s disease mild cognitive 744 impairment progression at a 6 -month follow -up. Neurobiology of Aging , 137, 19 –37. 745 https://doi.org/10.1016/j.neurobiolaging.2024.01.013 746 18. Cecchetti, G., Agosta, F., Canu, E., Basaia, S., Rugarli, G., Curti, D. G., Coraglia, F., Cursi, M., 747 Spinelli, E. G., Santangelo, R., Caso, F., Fanelli, G. F., Magnani, G., & Filippi, M. (2024). 748 Analysis of individual alpha frequency in a large cohort from a tertiary memory center. 749 European Journal of Neurology, 31(10). https://doi.org/10.1111/ene.16424 750 19. Zhao, Y ., Cai, J., Song, J., Shi, H., Kong, W., Li, X., Wei, W., & Xue, X. (2025). Peak alpha 751 frequency and alpha power spectral density as vulnerability markers of cognitive impairment 752 in Parkinson’s disease: An exploratory EEG study. Frontiers in Neuroscience, 19, 1575815. 753 https://doi.org/10.3389/fnins.2025.1575815 754 20. Murphy, M., & Öngür, D. (2019). Decreased peak alpha frequency and impaired visual evoked 755 potentials in first episode psychosis. NeuroImage: Clinical , 22, 101693. 756 https://doi.org/10.1016/j.nicl.2019.101693 757 21. Zhou, P., Wu, Q., Zhan, L., Guo, Z., Wang, C., Wang, S., Yang, Q., Lin, J., Zhang, F., Liu, L., Lin, 758 D., Fu, W., & Wu, X. (2023). Alpha peak activity in resting -state EEG is associated with 759 depressive score. Frontiers in Neuroscience , 17, 1057908. 760 https://doi.org/10.3389/fnins.2023.1057908 761 22. Başar, E., Schürmann, M., Başar -Eroglu, C., & Karakaş, S. (1997). Alpha oscillations in brain 762 functioning: An integrative theory. International Journal of Psychophysiology , 26(1–3), 5–763 29. https://doi.org/10.1016/S0167-8760(97)00753-8 764 .CC-BY-NC 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted January 25, 2026. ; https://doi.org/10.64898/2026.01.24.701499doi: bioRxiv preprint 29 23. Ben-Simon, E., Podlipsky, I., Arieli, A., Zhdanov, A., & Hendler, T. (2008). Never Resting Brain: 765 Simultaneous Representation of Two Alpha Related Processes in Humans. PLoS ONE, 3(12), 766 e3984. https://doi.org/10.1371/journal.pone.0003984 767 24. Barzegaran, E., Vildavski, V . Y ., & Knyazeva, M. G. (2017). Fine Structure of Posterior Alpha 768 Rhythm in Human EEG: Frequency Components, Their Cortical Sources, and Temporal 769 Behavior. Scientific Reports, 7(1), 8249. https://doi.org/10.1038/s41598-017-08421-z 770 25. Park, J., Ho, R. L. M., Wang, W., Nguyen, V . Q., & Coombes, S. A. (2024). The effect of age on 771 alpha rhythms in the human brain derived from source localized resting -state 772 electroencephalography. NeuroImage, 292, 120614. 773 https://doi.org/10.1016/j.neuroimage.2024.120614 774 26. Capilla, A., Schoffelen, J. -M., Paterson, G., Thut, G., & Gross, J. (2014). Dissociated α-Band 775 Modulations in the Dorsal and Ventral Visual Pathways in Visuospatial Attention and 776 Perception. Cerebral Cortex, 24(2), 550–561. https://doi.org/10.1093/cercor/bhs343 777 27. Sokoliuk, R., Mayhew, S. D., Aquino, K. M., Wilson, R., Brookes, M. J., Francis, S. T., 778 Hanslmayr, S., & Mullinger, K. J. (2019). Two Spatially Distinct Posterior Alpha Sources 779 Fulfill Different Functional Roles in Attention. The Journal of Neuroscience , 39(36), 7183–780 7194. https://doi.org/10.1523/JNEUROSCI.1993-18.2019 781 28. Rodriguez-Larios, J., ElShafei, A., Wiehe, M., & Haegens, S. (2022). Visual Working Memory 782 Recruits Two Functionally Distinct Alpha Rhythms in Posterior Cortex. Eneuro, 9(5), 783 ENEURO.0159-22.2022. https://doi.org/10.1523/ENEURO.0159-22.2022 784 29. Cruz, G., Melcón, M., Sutandi, L., Matias Palva, J., Palva, S., & Thut, G. (2025). Oscillatory 785 Brain Activity in the Canonical Alpha-Band Conceals Distinct Mechanisms in Attention. The 786 Journal of Neuroscience , 45(1), e0918242024. https://doi.org/10.1523/JNEUROSCI.0918 -787 24.2024 788 .CC-BY-NC 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted January 25, 2026. ; https://doi.org/10.64898/2026.01.24.701499doi: bioRxiv preprint 30 30. Zhou, Y . J., Van Es, M. W. J., & Haegens, S. (2025). Distinct alpha networks modulate different 789 aspects of perceptual decision -making. PLOS Biology , 23(10), e3003461. 790 https://doi.org/10.1371/journal.pbio.3003461 791 31. Tröndle, M., Popov, T., Pedroni, A., Pfeiffer, C., Barańczuk -Turska, Z., & Langer, N. (2021). 792 Decomposing age effects in EEG alpha power [Preprint]. Neuroscience. 793 https://doi.org/10.1101/2021.05.26.445765 794 32. Finley, A. J., Angus, D. J., Van Reekum, C. M., Davidson, R. J., & Schaefer, S. M. (2022). Periodic 795 and aperiodic contributions to theta‐beta ratios across adulthood. Psychophysiology, 59(11), 796 e14113. https://doi.org/10.1111/psyp.14113 797 33. Turner, C., Baylan, S., Bracco, M., Cruz, G., Hanzal, S., Keime, M., Kuye, I., McNeill, D., Ng, 798 Z., Van Der Plas, M., Ruzzoli, M., Thut, G., Trajkovic, J., Veniero, D., Wale, S. P., Whear, S., 799 & Learmonth, G. (2023). Developmental changes in individual alpha frequency: Recording 800 EEG data during public engagement events. Imaging Neuroscience , 1, 1 –14. 801 https://doi.org/10.1162/imag_a_00001 802 34. Van Dijk, H., Van Wingen, G., Denys, D., Olbrich, S., Van Ruth, R., & Arns, M. (2022). The two 803 decades brainclinics research archive for insights in neurophysiology (TDBRAIN) database. 804 Scientific Data, 9(1), 333. https://doi.org/10.1038/s41597-022-01409-z 805 35. Pritchard, W. S. (1992). The Brain in Fractal Time: 1/F -Like Power Spectrum Scaling of the 806 Human Electroencephalogram. International Journal of Neuroscience , 66(1–2), 119 –129. 807 https://doi.org/10.3109/00207459208999796 808 36. Tenke, C., & Kayser, J. (2005). Reference-free quantification of EEG spectra: Combining current 809 source density (CSD) and frequency principal components analysis (fPCA). Clinical 810 Neurophysiology, 116(12), 2826–2846. https://doi.org/10.1016/j.clinph.2005.08.007 811 37. Barry, R. J., & De Blasio, F. M. (2017). EEG frequency PCA in EEG‐ERP dynamics. 812 Psychophysiology, 55(5), e13042. https://doi.org/10.1111/psyp.13042 813 .CC-BY-NC 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted January 25, 2026. ; https://doi.org/10.64898/2026.01.24.701499doi: bioRxiv preprint 31 38. Nakhnikian, A., Oribe, N., Hirano, S., Fujishima, Y ., Hirano, Y ., Nestor, P. G., Francis, G. A., 814 Levin, M., & Spencer, K. M. (2024). Spectral decomposition of resting state 815 electroencephalogram reveals unique theta/alpha activity in schizophrenia. European Journal 816 of Neuroscience, 59(8), 1946–1960. https://doi.org/10.1111/ejn.16244 817 39. Mørup, M., & Hansen, L. K. (2012). Archetypal analysis for machine learning and data mining. 818 Neurocomputing, 80, 54–63. https://doi.org/10.1016/j.neucom.2011.06.033 819 40. Walter, W. G. (1938). CRITICAL REVIEW: THE TECHNIQUE AND APPLICATION OF 820 ELECTRO-ENCEPHALOGRAPHY . Journal of Neurology, Neurosurgery & Psychiatry , 821 1(4), 359–385. https://doi.org/10.1136/jnnp.1.4.359 822 41. Scally, B., Burke, M. R., Bunce, D., & Delvenne, J. -F. (2018). Resting -state EEG power and 823 connectivity are associated with alpha peak frequency slowing in healthy aging. Neurobiology 824 of Aging, 71, 149–155. https://doi.org/10.1016/j.neurobiolaging.2018.07.004 825 42. Cesnaite, E., Steinfath, P., Jamshidi Idaji, M., Stephani, T., Kumral, D., Haufe, S., Sander, C., 826 Hensch, T., Hegerl, U., Riedel-Heller, S., Röhr, S., Schroeter, M. L., Witte, A., Villringer, A., 827 & Nikulin, V . V . (2023). Alterations in rhythmic and non‐rhythmic resting‐state EEG activity 828 and their link to cognition in older age. NeuroImage, 268, 119810. 829 https://doi.org/10.1016/j.neuroimage.2022.119810 830 43. Getzmann, S., Gajewski, P. D., Schneider, D., & Wascher, E. (2024). Resting -state EEG data 831 before and after cognitive activity across the adult lifespan and a 5-year follow-up. Scientific 832 Data, 11(1), 988. https://doi.org/10.1038/s41597-024-03797-w 833 44. Babayan, A., Erbey, M., Kumral, D., Reinelt, J. D., Reiter, A. M. F., Röbbig, J., Schaare, H. L., 834 Uhlig, M., Anwander, A., Bazin, P. -L., Horstmann, A., Lampe, L., Nikulin, V . V ., Okon -835 Singer, H., Preusser, S., Pampel, A., Rohr, C. S., Sacher, J., Thöne-Otto, A., … Villringer, A. 836 (2019). A mind -brain-body dataset of MRI, EEG, cognition, emotion, and peripheral 837 physiology in young and old adults. Scientific Data , 6(1), 180308. 838 https://doi.org/10.1038/sdata.2018.308 839 .CC-BY-NC 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted January 25, 2026. ; https://doi.org/10.64898/2026.01.24.701499doi: bioRxiv preprint 32 45. Clayton, M. S., Yeung, N., & Cohen Kadosh, R. (2018). The many characters of visual alpha 840 oscillations. European Journal of Neuroscience , 48(7), 2498 –2508. 841 https://doi.org/10.1111/ejn.13747 842 46. Klimesch, W., Schimke, H., & Pfurtscheller, G. (1993). Alpha frequency, cognitive load and 843 memory performance. Brain Topography , 5(3), 241 –251. 844 https://doi.org/10.1007/BF01128991 845 47. Clark, C. R., Veltmeyer, M. D., Hamilton, R. J., Simms, E., Paul, R., Hermens, D., & Gordon, E. 846 (2004). Spontaneous alpha peak frequency predicts working memory performance across the 847 age span. International Journal of Psychophysiology, 9. 848 48. Klimesch, W., Schimke, H., & Schwaiger, J. (1994). Episodic and semantic memory: An analysis 849 in the EEG theta and alpha band. Electroencephalography and Clinical Neurophysiology , 850 91(6), 428–441. https://doi.org/10.1016/0013-4694(94)90164-3 851 49. Corcoran, A. W., Alday, P. M., Schlesewsky, M., & Bornkessel-Schlesewsky, I. (2018). Toward a 852 reliable, automated method of individual alpha frequency (IAF) quantification. 853 Psychophysiology, 55(7), e13064. https://doi.org/10.1111/psyp.13064 854 50. Knyazeva, M. G., Barzegaran, E., Vildavski, V . Y ., & Demonet, J. -F. (2018). Aging of human 855 alpha rhythm. Neurobiology of Aging , 69, 261 –273. 856 https://doi.org/10.1016/j.neurobiolaging.2018.05.018 857 51. Menétrey, M. Q., Roinishvili, M., Chkonia, E., Herzog, M. H., & Pascucci, D. (2024). Alpha peak 858 frequency affects visual performance beyond temporal resolution. Imaging Neuroscience, 2, 859 1–12. https://doi.org/10.1162/imag_a_00107 860 52. Ramsay, I. S., Lynn, P. A., Schermitzler, B., & Sponheim, S. R. (2021). Individual alpha peak 861 frequency is slower in schizophrenia and related to deficits in visual perception and cognition. 862 Scientific Reports, 11(1), 17852. https://doi.org/10.1038/s41598-021-97303-6 863 53. Catalano, L. T., Reavis, E. A., Wynn, J. K., & Green, M. F. (2024). Peak alpha frequency in 864 schizophrenia, bipolar disorder, and healthy volunteers: Associations with visual information 865 .CC-BY-NC 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted January 25, 2026. ; https://doi.org/10.64898/2026.01.24.701499doi: bioRxiv preprint 33 processing and cognition. Biological Psychiatry: Cognitive Neuroscience and Neuroimaging, 866 9(11), 1132–1140. 867 54. Murphy, M., & Öngür, D. (2019). Decreased peak alpha frequency and impaired visual evoked 868 potentials in first episode psychosis. NeuroImage: Clinical, 22, 101693. 869 55. Tenke, C. E., Kayser, J., Alvarenga, J. E., Abraham, K. S., Warner, V ., Talati, A., Weissman, M. 870 M., & Bruder, G. E. (2018). Temporal stability of posterior EEG alpha over twelve years. 871 Clinical Neurophysiology, 129(7), 1410–1417. https://doi.org/10.1016/j.clinph.2018.03.037 872 56. Arutiunian, V ., Opdahl, M., Sullivan, C. A. W., Santhosh, M., Neuhaus, E., Borland, H., Bernier, 873 R. A., Bookheimer, S. Y ., Dapretto, M., Jack, A., Jeste, S., McPartland, J. C., Naples, A., Van 874 Horn, J. D., Pelphrey, K. A., Webb, S. J., & Gupta, A. R. (2025). Number of Alpha Peaks in 875 the Electroencephalogram Is Associated With Clinical Phenotype and Copy Number Variants 876 in Youths With Autism. Biological Psychiatry: Cognitive Neuroscience and Neuroimaging , 877 S2451902225003027. https://doi.org/10.1016/j.bpsc.2025.10.001 878 57. Chiang, A. K. I., Rennie, C. J., Robinson, P. A., Roberts, J. A., Rigozzi, M. K., Whitehouse, R. 879 W., Hamilton, R. J., & Gordon, E. (2008). Automated characterization of multiple alpha peaks 880 in multi -site electroencephalograms. Journal of Neuroscience Methods , 168(2), 396 –411. 881 https://doi.org/10.1016/j.jneumeth.2007.11.001 882 58. Ronconi, L., Busch, N. A., & Melcher, D. (2018). Alpha -band sensory entrainment alters the 883 duration of temporal windows in visual perception. Scientific Reports , 8(1), 11810. 884 https://doi.org/10.1038/s41598-018-29671-5 885 59. Lin, Y .-J., Shukla, L., Dugué, L., Valero-Cabré, A., & Carrasco, M. (2021). Transcranial magnetic 886 stimulation entrains alpha oscillatory activity in occipital cortex. Scientific Reports , 11(1), 887 18562. https://doi.org/10.1038/s41598-021-96849-9 888 60. Trajkovic, J., Di Gregorio, F., Marcantoni, E., Thut, G., & Romei, V . (2022). A TMS/EEG protocol 889 for the causal assessment of the functions of the oscillatory brain rhythms in perceptual and 890 .CC-BY-NC 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted January 25, 2026. ; https://doi.org/10.64898/2026.01.24.701499doi: bioRxiv preprint 34 cognitive processes. STAR Protocols , 3(2), 101435. 891 https://doi.org/10.1016/j.xpro.2022.101435 892 61. Millard, S. K., Speis, D. B., Skippen, P., Chiang, A. K. I., Chang, W., Lin, A. J., Furman, A. J., 893 Mazaheri, A., Seminowicz, D. A., & Schabrun, S. M. (2024). Can non‐invasive brain 894 stimulation modulate peak alpha frequency in the human brain? A systematic review and 895 meta‐analysis. European Journal of Neuroscience , 60(3), 4182 –4200. 896 https://doi.org/10.1111/ejn.16424 897 62. D’Angelo, M., Lanfranco, R. C., Chancel, M., & Ehrsson, H. H. (2026). Parietal alpha frequency 898 shapes own-body perception by modulating the temporal integration of bodily signals. Nature 899 Communications, 17(1), 53. https://doi.org/10.1038/s41467-025-67657-w 900 63. Capilla, A., Schoffelen, J. -M., Paterson, G., Thut, G., & Gross, J. (2014). Dissociated α-Band 901 Modulations in the Dorsal and Ventral Visual Pathways in Visuospatial Attention and 902 Perception. Cerebral Cortex, 24(2), 550–561. https://doi.org/10.1093/cercor/bhs343 903 64. Mahjoory, K., Cesnaite, E., Hohlefeld, F. U., Villringer, A., & Nikulin, V . V . (2019). Power and 904 temporal dynamics of alpha oscillations at rest differentiate cognitive performance involving 905 sustained and phasic cognitive control. NeuroImage, 188, 135 –144. 906 https://doi.org/10.1016/j.neuroimage.2018.12.001 907 65. Hülsdünker, T., & Mierau, A. (2021). Visual Perception and Visuomotor Reaction Speed Are 908 Independent of the Individual Alpha Frequency. Frontiers in Neuroscience , 15, 620266. 909 https://doi.org/10.3389/fnins.2021.620266 910 66. Buergers, S., & Noppeney, U. (2022). The role of alpha oscillations in temporal binding within 911 and across the senses. Nature Human Behaviour , 6(5), 732 –742. 912 https://doi.org/10.1038/s41562-022-01294-x 913 67. Busch, N., Geyer, T., & Zinchenko, A. (2024). Individual peak alpha frequency does not index 914 individual differences in inhibitory cognitive control. Psychophysiology, 61(8), e14586. 915 https://doi.org/10.1111/psyp.14586 916 .CC-BY-NC 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted January 25, 2026. ; https://doi.org/10.64898/2026.01.24.701499doi: bioRxiv preprint 35 68. Cottier, T., Turner, W., Chae, V . J., Holcombe, A. O., & Hogendoorn, H. (2025). No Evidence 917 That Resting‐State Individual Alpha Frequency Represents a Mechanism Underlying Motion‐918 Position Illusions. European Journal of Neuroscience , 62(9), e70250. 919 https://doi.org/10.1111/ejn.70250 920 69. Arns, M. (2012). EEG-Based Personalized Medicine in ADHD: Individual Alpha Peak Frequency 921 as an Endophenotype Associated with Nonresponse. Journal of Neurotherapy , 16(2), 123–922 141. https://doi.org/10.1080/10874208.2012.677664 923 70. Arns, M., V ollebregt, M. A., Palmer, D., Spooner, C., Gordon, E., Kohn, M., Clarke, S., Elliott, 924 G. R., & Buitelaar, J. K. (2018). Electroencephalographic biomarkers as predictors of 925 methylphenidate response in attention -deficit/hyperactivity disorder. European 926 Neuropsychopharmacology, 28(8), 881–891. 927 https://doi.org/10.1016/j.euroneuro.2018.06.002 928 71. Furman, A. J., Meeker, T. J., Rietschel, J. C., Yoo, S., Muthulingam, J., Prokhorenko, M., Keaser, 929 M. L., Goodman, R. N., Mazaheri, A., & Seminowicz, D. A. (2018). Cerebral peak alpha 930 frequency predicts individual differences in pain sensitivity. NeuroImage, 167, 203 –210. 931 https://doi.org/10.1016/j.neuroimage.2017.11.042 932 72. V oetterl, H. T. S., Sack, A. T., Olbrich, S., Stuiver, S., Rouwhorst, R., Prentice, A., Pizzagalli, D. 933 A., Van Der Vinne, N., Van Waarde, J. A., Brunovsky, M., Van Oostrom, I., Reitsma, B., 934 Fekkes, J., Van Dijk, H., & Arns, M. (2023). Alpha peak frequency-based Brainmarker-I as a 935

to stratify to pharmacotherapy and brain stimulation treatments in depression. Nature 936 Mental Health, 1(12), 1023–1032. https://doi.org/10.1038/s44220-023-00160-7 937 73. Wen, H., & Liu, Z. (2016). Separating Fractal and Oscillatory Components in the Power Spectrum 938 of Neurophysiological Signal. Brain Topography , 29(1), 13 –26. 939 https://doi.org/10.1007/s10548-015-0448-0 940 .CC-BY-NC 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted January 25, 2026. ; https://doi.org/10.64898/2026.01.24.701499doi: bioRxiv preprint 36 74. Gao, R., Peterson, E. J., & V oytek, B. (2017). Inferring synaptic excitation/inhibition balance 941 from field potentials. NeuroImage, 158, 70 –78. 942 https://doi.org/10.1016/j.neuroimage.2017.06.078 943 75. Donoghue, T., Haller, M., Peterson, E. J., Varma, P., Sebastian, P., Gao, R., Noto, T., Lara, A. H., 944 Wallis, J. D., Knight, R. T., Shestyuk, A., & V oytek, B. (2020). Parameterizing neural power 945 spectra into periodic and aperiodic components. Nature Neuroscience, 23(12), 1655–1665. 946 https://doi.org/10.1038/s41593-020-00744-x 947 76. Gerster, M., Waterstraat, G., Litvak, V ., Lehnertz, K., Schnitzler, A., Florin, E., Curio, G., & 948 Nikulin, V . (2022). Separating Neural Oscillations from Aperiodic 1/f Activity: Challenges 949 and Recommendations. Neuroinformatics, 20(4), 991–1012. https://doi.org/10.1007/s12021-950 022-09581-8 951 77. Bai, S., Menétrey, M. Q., & Pascucci, D. (2026). What is next? Predictable visual sequences are 952 encoded with anticipatory biases and reduced neural responses. iScience, 114697. 953 https://doi.org/10.1016/j.isci.2026.114697 954 78. Delorme, A. (2023). EEG is better left alone. Scientific Reports , 13(1), 2372. 955 https://doi.org/10.1038/s41598-023-27528-0 956 79. Delorme, A., & Makeig, S. (2004). EEGLAB: An open source toolbox for analysis of single-trial 957 EEG dynamics including independent component analysis. Journal of Neuroscience Methods, 958 134(1), 9–21. https://doi.org/10.1016/j.jneumeth.2003.10.009 959 80. Mullen, T. R., Kothe, C. A. E., Chi, Y . M., Ojeda, A., Kerth, T., Makeig, S., Jung, T. -P., & 960 Cauwenberghs, G. (2015). Real-time neuroimaging and cognitive monitoring using wearable 961 dry EEG. IEEE Transactions on Biomedical Engineering , 62(11), 2553 –2567. 962 https://doi.org/10.1109/TBME.2015.2481482 963 81. Perrin, F., Pernier, J., Bertrand, O., & Echallier, J. F. (1989). Spherical splines for scalp potential 964 and current density mapping. Electroencephalography and Clinical Neurophysiology, 72(2), 965 184–187. https://doi.org/10.1016/0013-4694(89)90180-6 966 .CC-BY-NC 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted January 25, 2026. ; https://doi.org/10.64898/2026.01.24.701499doi: bioRxiv preprint 37 82. Dien, J. (2010). The ERP PCA Toolkit: An open source program for advanced statistical analysis 967 of event -related potential data. Journal of Neuroscience Methods , 187(1), 138 –145. 968 https://doi.org/10.1016/j.jneumeth.2009.12.009 969 83. Van De Ville, D., Farouj, Y ., Preti, M. G., Liégeois, R., & Amico, E. (2021). When makes you 970 unique: Temporality of the human brain fingerprint. Science Advances , 7(42), eabj0751. 971 https://doi.org/10.1126/sciadv.abj0751 972 84. Finn, E. S., Shen, X., Scheinost, D., Rosenberg, M. D., Huang, J., Chun, M. M., Papademetris, 973 X., & Constable, R. T. (2015). Functional connectome fingerprinting: Identifying individuals 974 using patterns of brain connectivity. Nature Neuroscience , 18(11), Articolo 11. 975 https://doi.org/10.1038/nn.4135 976 85. Tourbier, S., Rue Queralt, J., Glomb, K., Aleman -Gomez, Y ., Mullier, E., Griffa, A., Schöttner, 977 M., Wirsich, J., Tuncel, A., Jancovic, J., Bach Cuadra, M., & Hagmann, P. (2022). 978 connectomicslab/connectomemapper3: Connectome Mapper v3.1.0 (Versione v3.1.0) 979 [Software]. Zenodo. https://doi.org/10.5281/ZENODO.7249263 980 86. Gramfort, A., Papadopoulo, T., Olivi, E., & Clerc, M. (2010). OpenMEEG: Opensource software 981 for quasistatic bioelectromagnetics. BioMedical Engineering OnLine , 9(1), 45. 982 https://doi.org/10.1186/1475-925X-9-45 983 87. Pagnotta, M. F., Pascucci, D., & Plomp, G. (2022). Selective attention involves a feature-specific 984 sequential release from inhibitory gating. Neuroimage, 246, 118782. 985 88. Pascucci, D., Hervais-Adelman, A., & Plomp, G. (2018). Gating by induced Α-Γ asynchrony in 986 selective attention. Human Brain Mapping. https://doi.org/10.1002/hbm.24216 987 89. Grech, R., Cassar, T., Muscat, J., Camilleri, K. P., Fabri, S. G., Zervakis, M., Xanthopoulos, P., 988 Sakkalis, V ., & Vanrumste, B. (2008). Review on solving the inverse problem in EEG source 989 analysis. Journal of NeuroEngineering and Rehabilitation , 5(1), 25. 990 https://doi.org/10.1186/1743-0003-5-25 991 .CC-BY-NC 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted January 25, 2026. ; https://doi.org/10.64898/2026.01.24.701499doi: bioRxiv preprint 38 90. Daducci, A., Gerhard, S., Griffa, A., Lemkaddem, A., Cammoun, L., Gigandet, X., Meuli, R., 992 Hagmann, P., & Thiran, J. -P. (2012). The connectome mapper: An open -source processing 993 pipeline to map connectomes with MRI. PloS one, 7(12), e48121. 994 91. Desikan, R. S., Ségonne, F., Fischl, B., Quinn, B. T., Dickerson, B. C., Blacker, D., Buckner, R. 995 L., Dale, A. M., Maguire, R. P., & Hyman, B. T. (2006). An automated labeling system for 996 subdividing the human cerebral cortex on MRI scans into gyral based regions of interest. 997 Neuroimage, 31(3), 968–980. 998 999 1000 .CC-BY-NC 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted January 25, 2026. ; https://doi.org/10.64898/2026.01.24.701499doi: bioRxiv preprint 39 Acknowledgments 1001 This work was supported by the Swiss National Science Foundation (Grant number : 1002 TMSGI1_218247). 1003 Author contribution 1004 Conceptualization: MQM & DP 1005 Methodology: MQM & DP 1006 Investigation: MQM & DP 1007 Visualization: MQM & DP 1008 Supervision: DP 1009 Writing—original draft: MQM & DP 1010 Writing—review & editing: MQM & DP 1011 Competing interests 1012 Authors declare that they have no competing interests. 1013 Data and materials availability 1014 All data and material available in the main text or the supplementary material will be made available 1015 in an online repository upon publication. 1016 1017 .CC-BY-NC 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted January 25, 2026. ; https://doi.org/10.64898/2026.01.24.701499doi: bioRxiv preprint 40 Supplementary Material 1018 1019 Cross-dataset replication of four archetypical components 1020 The four components identified using the fPCA approach (e.g., Figure 1D) were also replicated in the 1021 TDBRAIN dataset using a complementary decomposition method based on archetypal analysis (see 1022 Methods), which relies on mathematical principles distinct from those of fPCA. This approach 1023 yielded highly similar spectral profiles, revealing three components peaking within the alpha band at 1024 distinct frequencies and one component reflecting lower-frequency activity (Figure 1E). 1025 To further confirm these findings, archetypal analysis was also applied to the DVS and 1026 LEMON datasets (Supplementary Figures 1A and 1B). In both datasets, four archetypal components 1027 with spectral profiles closely resembling those obtained in TDBRAIN were identified, explaining 1028 95.9% and 95.6% of the variance, respectively. All correlations between the spectral loadings and the 1029 corresponding archetypical profiles fell within a .66–.86 range. 1030 Consistent with the TDBRAIN results (Figure 1F), these archetypal profiles were broadly 1031 expressed across participants, with higher proportion of individuals showing maximal expression of 1032 the low-frequency non-alpha component and the middle alpha component (Supplementary Figures 1033 1B and 1E). The LEMON dataset also showed specifically high maximal expression of the fast alpha 1034 component, potentially explained by the larger proportion of younger individuals in this cohort. 1035 Moreover, a purity analysis, which quantified, for each participant, the strength of expression of each 1036 archetype based on the mixture weights estimated by the model , revealed distributions peaking 1037 slightly above 0.5. This pattern, consistent with the TDBRAIN dataset (Figure 1G), indicates that 1038 most participants expressed multiple archetypes with moderate contributions rather than being 1039 strongly dominated by a single archetype (Supplementary Figures 1C and 1F). 1040 Together, these results confirm that the four distinct components can be replicated with an 1041 alternative method, showing that they are mixed even at individual levels. 1042 .CC-BY-NC 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted January 25, 2026. ; https://doi.org/10.64898/2026.01.24.701499doi: bioRxiv preprint 41 1043 Supplementary Figure 1. Archetypal analysis in the A) DVS and D) LEMON datasets. Four components 1044 were consistently identified, with spectral profiles closely matching those recovered by the same analysis in 1045 the TDBRAIN dataset (Figure 1E) and by fPCA in both datasets (Figures 2A and 3A). B) and E) The proportion 1046 of participants showing maximal expression of each archetypal profile was similar to that observed in 1047 TDBRAIN (Figure 1F). C) and F) The distribution of participants (y-axis) as a function of dominant archetype 1048 weight (x-axis) is peaking around 0.5 in both datasets, consistent with TDBRAIN (Figure 1G). 1049 1050 .CC-BY-NC 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted January 25, 2026. ; https://doi.org/10.64898/2026.01.24.701499doi: bioRxiv preprint 42 Cross-dataset replication of distinct alpha component contributions to IAPF and age 1051 Using the TDBRAIN dataset, we further showed that IAPF estimated at the scalp (see Methods) was 1052 significantly predicted by both the slow and fast alpha components, but with opposite regression 1053 weights (Figure 1I). 1054 These results were replicated in the DVS dataset across both available sessions 1055 (Supplementary Figures 2A and 2C). In the first session (N = 586), IAPF was significantly predicted 1056 with opposite effects by the slow and fast alpha components (slow alpha: 𝛽 = -0.86, SE = 0.08, t = -1057 10.83, p < .001; fast alpha: 𝛽 = 0.47, SE = 0.08, t = 5.48, p < .001). In addition, the middle alpha 1058 component showed a modest but significant effect ( 𝛽 = -0.20, SE = 0.09, t = -2.17, p = .03), which 1059 was not observed in TDBRAIN, whereas the low -frequency non -alpha component again did not 1060 significantly predict IAPF (𝛽 = 0.10, SE = 0.09, t = 1.63, p = .10). Similar patterns were observed in 1061 the second session, conducted five years later (N = 203). IAPF was significantly predicted by all alpha 1062 components except the low-frequency non-alpha component (slow alpha: 𝛽 = -1.09, SE = 0.12, t = -1063 8.61, p < .001; fast alpha: 𝛽 = 0.32, SE = 0.14, t = 2.32, p = .02; middle alpha: 𝛽 = -0.52, SE = 0.15, 1064 t = -3.39, p < .001; low-frequency non-alpha: 𝛽 = -0.03, SE = 0.13, t = -0.02, p = .97). The contrast 1065 between slow and fast components remained a strong predictor of IAPF in both sessions (first session: 1066 𝛽 = 0.69, SE = 0.03, t = 20.24, p < .001; second session: 𝛽 = 0.74, SE = 0.05, t = 12.69, p < .001), 1067 accounting for almost half of the variance (first session: 41.2%; second session: 44.2%). 1068 These findings were also replicated in the LEMON dataset, which showed results very similar 1069 to TDBRAIN (Supplementary Figure 2E). In this dataset, IAPF was significantly predicted by both 1070 the slow and fast alpha components (slow alpha: 𝛽 = -0.94, SE = 0.17, t = -5.32, p < .001; fast alpha: 1071 𝛽 = 0.63, SE = 0.21, t = 2.94, p = .003), but not by the middle or low-frequency non-alpha components 1072 (middle alpha: 𝛽 = -0.32, SE = 0.20, t = -1.61, p = .10; low -frequency non-alpha: 𝛽 = -0.10, SE = 1073 0.17, t = -0.60, p = .54). The contrast between slow and fast components was again a strong predictor 1074 of IAPF (𝛽 = 0.81, SE = 0.05, t = 14.48, p < .001), accounting for 51.7% of the variance. 1075 In parallel with the finding that alpha components exert distinct effects on IAPF, we also 1076 observed in the TDBRAIN dataset that aging was not associated with a uniform slowing of all alpha 1077 components, as would be suggested by previous studies reporting a general age -related decrease in 1078 IAPF (e.g., Scally et al., 2018; Park et al., 2024). Instead, aging was linked to complex, component-1079 specific changes in spectral contributions, specifically showing a concomitant increase in the 1080 dominance of the slow alpha component, while the fast and middle alpha components decreased with 1081 age (Figure 1J). 1082 .CC-BY-NC 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted January 25, 2026. ; https://doi.org/10.64898/2026.01.24.701499doi: bioRxiv preprint 43 Despite a narrower age range, these results were replicated in both sessions of the DVS dataset 1083 when modeling age as a function of the four spectral components (Supplementary Figures 2B and 1084 2D). Comparisons between linear and quadratic models showed that the better -fitting model varied 1085 depending on the component and dataset, but the direction of effects remained consistent across 1086

for both slow and fast alpha rhythms. In the first session, the fast alpha component exhibited 1087 a linear negative association with age (BIC = −19.3 vs. -15.4 for the quadratic model; β = -0.004, SE 1088 = 0.001, t = -2.6, p = .01), while the middle alpha component showed a nonlinear association (BIC = 1089 −46.5 vs. -45.7 for the linear model; linear term: β = 0.01, SE = 0.009, t = 1.47, p = .10; quadratic 1090 term: β = −0.0002, SE = 0.0001, t = −2.17, p = .03), both consistent with an overall decline with age. 1091 In contrast, the slow alpha component showed a nonlinear increase ( BIC = −18.1 vs. -14.1 for the 1092 linear model; linear: β = −0.02, SE = 0.01, t = −2.14, p = .03; quadratic: β = 0.0003, SE = 0.0001, t = 1093 2.84, p = .006), indicating an increasingly positive association with age. The low-frequency non-alpha 1094 component exhibited no significant linear relationship (BIC = −38.5 vs. -35.5 for the quadratic model; 1095 β = 0.002, SE = 0.001, t = 1.69, p = .09). In the second session, despite fewer participants being 1096 recorded, the results still show opposite associations for the fast and slow alpha components, with the 1097 fast component decreasing and the slow component increasing linearly with age (fast alpha: BIC = 1098 48.1 vs. 51.7 for the quadratic model; β = -0.007, SE = 0.002, t = -2.62, p = .01; slow alpha: BIC = 1099 38.6 vs. 41.6 for the quadratic model; β = 0.01, SE = 0.003, t = 3.76, p < .001). The middle alpha and 1100 low-frequency non-alpha components were not significantly associated with age (middle alpha: BIC 1101 = 40.2 vs. 42.4 for the quadratic model; β = 0.0004, SE = 0.003, t = 0.11, p = .91; low-frequency non-1102 alpha: BIC = 22.8 vs. 24.7 for the quadratic model; β = -0.002, SE = 0.003, t = -0.68, p = .49). In both 1103 sessions, the slow alpha component accounted for the largest proportion of age -related variance 1104 (adjusted R² = 0.36 and 0.22, respectively). 1105 Finally, we applied a similar analysis in the LEMON dataset . Since these data included two 1106 groups of individuals, young (20–35 years) and older (59 –77 years), we ran independent-samples t-1107 tests to compare scores between groups for each component (Supplementary Figure 2F). Significant 1108 differences were observed in the slow alpha (t(196) = -3.48, p < .001) and fast alpha components 1109 (t(196) = 1.98, p = .04), whereas no significant differences were found in the middle alpha (t(196) = 1110 0.70, p = .48) or low -frequency non-alpha components (t(196) = 0.85, p = .39). These results again 1111 confirm that age is not associated with a general slowing of alpha oscillations, but rather with two 1112 distinct, component-specific effects: the slow alpha component increases with age, while the fast 1113 alpha component decreases with age. 1114 .CC-BY-NC 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted January 25, 2026. ; https://doi.org/10.64898/2026.01.24.701499doi: bioRxiv preprint 44 1115 Supplementary Figure 2. A) In the DVS datasets (pre-test recordings from the first session), IAPF, estimated 1116 from the average scalp PSD, was significantly predicted by the three different alpha components (i.e., 1117 Components 2, 3 and 4). Error bars represent 95% confidence intervals for the linear coefficients (β) estimates. 1118 B) In the DVS dataset (pre -test recordings from the first session), age -related variance was explained by 1119 changes across all alpha components, suggesting that aging reflects shifts in the relative dominance of alpha 1120 components rather than a uniform slowing of IAPF. The solid line and shaded area show predictions from the 1121 best-fitting linear or quadratic model, with 95% confidence intervals. C-D) In the DVS dataset (pre -test 1122 recordings from the second session), similar results were found, except for the lack of a significant relationship 1123 between the middle alpha component and age. E) In the LEMON dataset, only the slow and fast alpha 1124 components significantly predicted IAPF, consistent with the TDBRAIN dataset (Figure 1I). F) Only the slow 1125 and fast alpha components showed significant differences between the two groups (young and old individuals). 1126 Error bars represent 95% confidence intervals of the group means. 1127 1128 .CC-BY-NC 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted January 25, 2026. ; https://doi.org/10.64898/2026.01.24.701499doi: bioRxiv preprint 45 Replication of test-retest reliability 1129 In the TDBRAIN dataset, taking advantage of the two repeated recordings obtained during the first 1130 session, as well as the two additional recordings from a subset of participants who returned five years 1131 later for a second session, we showed that the four components obtained with fPCA were highly stable 1132 both across the 2 -hour interval (pre - vs. post-test recordings from the first session; Figure 2C) and 1133 over the five-year period (pre-test recordings from the first session vs. pre -test recordings from the 1134 second session; Figure 2D). 1135 To further confirm th ese results, similar analyses were performed using additional 1136 combinations of the available recordings : the pre- vs. post-test recordings from the second session, 1137 and the post-test recordings from the first session vs. the post-test recordings from the second session. 1138 Only participants with data available for both recordings in a given comparison were included, 1139 resulting in 203 participants for the first analysis (Supplementary Figure 3A) and 201 participants for 1140 the second (Supplementary Figure 3B). 1141 All components exhibited strong test –retest reliability, both within the same day (low -1142 frequency non-alpha: β = 0.94, SE = 0.04, t = 20.77, p < .001; slow alpha: β = 0.82, SE = 0.04, t = 1143 20.59, p < .001; middle alpha: β = 0.84, SE = 0.04, t = 19.59, p < .001; fast alpha: β = 0.83, SE = 1144 0.04, t = 20.63, p < .001) and across the five-year interval (low-frequency non-alpha: β = 0.83, SE = 1145 0.03, t = 22.46, p < .001; slow alpha: β = 0.83, SE = 0.03, t = 23.38, p < .001; middle alpha: β = 0.72, 1146 SE = 0.05, t = 14.98, p < .001; fast alpha: β = 0.84, SE = 0.04, t = 20.48, p < .001), confirming that 1147 these components reflect stable individual traits. 1148 .CC-BY-NC 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted January 25, 2026. ; https://doi.org/10.64898/2026.01.24.701499doi: bioRxiv preprint 46 1149 Supplementary Figure 3. A) Short-term test–retest stability, quantified as the correlation between component 1150 loadings estimated from EEG recordings separated by approximately 2 hours (pre - and post-test recordings 1151 from the second session of the DVS dataset; N = 203). The solid line and shaded area show predictions from 1152 a linear model with 95% confidence intervals. B) Long-term test–retest stability assessed over a 5-year interval 1153 (post-test recordings from the first session vs. the post-test recordings from the second session N = 201). 1154 1155 1156 Cross-dataset replication of distinct spectral component maps 1157 The fPCA-based spectral decomposition of the LEMON dataset identified four spectral components 1158 (Figure 3A). For each component, distinct scalp EEG topographies were obtained by averaging fPCA 1159 factor scores across participants at each electrode (Figure 3C). 1160 Applying the same procedure with the four spectral components identified in the TDBRAIN 1161 (Figure 1C) and DVS (Figure 2A) datasets , we replicated these findings and yielded highly 1162 comparable topographical maps (Supplementary Figures 4A and 4B). Specifically, the three alpha -1163 related components exhibited relatively similar posteriorly distributed topographies, whereas the low-1164 frequency non-alpha component showed a clearly distinct spatial pattern. 1165 In addition, as shown with the LEMON dataset (Figures 3D, 3E and 3F), further comparisons 1166 across topographies in both TDBRAIN and DVS datasets confirmed that the slow alpha component 1167 was more strongly expressed over occipito -temporal electrodes, whereas the two others alpha 1168 components showed greater expression over occipito-parietal electrodes (Supplementary Figures 4C 1169 and 4D). 1170 .CC-BY-NC 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted January 25, 2026. ; https://doi.org/10.64898/2026.01.24.701499doi: bioRxiv preprint 47 1171 Supplementary Figure 4. A) Using the TDBRAIN dataset and B) the DVS dataset, scalp EEG topographies 1172 of the four obtained spectral components (see Figures 1D and 2A, respectively) were similar to those observed 1173 in the LEMON dataset (see Figure 3C). Pairwise spatial comparisons between component C) in the TDBRAIN 1174 dataset and D) the DVS dataset confirm that the components are spatially distinct as shown with the LEMON 1175 dataset (see Figures 3D, 3E and 3F). 1176 1177 .CC-BY-NC 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted January 25, 2026. ; https://doi.org/10.64898/2026.01.24.701499doi: bioRxiv preprint 48 Replication of distinct spectral component in a small-sample dataset 1178 The four distinct components consistently obtained with the two complementary unsupervised 1179 spectral decomposition approaches—fPCA and archetypal analysis—were, however, all identified in 1180 large datasets (>200 participants; e.g., Figures 1D and 1E). 1181 To test whether these components can also be recovered in smaller, more typical EEG samples, 1182 we used an in -house dataset comprising resting -state (RS) EEG recordings from 19 healthy 1183 participants aged 18 to 25 (8 females, 11 males), collected prior to participation in an independent 1184 study (Bai et al., 2026). 1185 These participants were recruited from École Polytechnique Fédérale de Lausanne (EPFL) 1186 and the University of Lausanne (UNIL) and received 30 CHF per hour for their participation. Two 1187 2.5-minute resting-state EEG sessions were recorded —one with eyes open (EO) and one with eyes 1188 closed (EC) —using a 128 -channel BioSemi ActiveTwo system (BioSemi, Amsterdam, the 1189 Netherlands). The cap was positioned to ensure that the A1 electrode was equidistant from the inion 1190 and nasion, as well as from both ears. EEG signals were sampled at 2048 Hz and referenced online 1191 to the common mode sense (CMS) and driven right leg (DRL) electrodes, with electrode voltages 1192 maintained within -20 µV to 20 µV . 1193 Offline preprocessing of the EC recordings included downsampling to 250 Hz, band -pass 1194 filtering between 0.5 and 40 Hz, segmentation into 4-second epochs, and manual cleaning to remove 1195 bad channels, epochs, and independent components identified via ICA. The final preprocessing steps 1196 were similar to those used in the TDBRAIN and DVS datasets, including electrode interpolation and 1197 average re-referencing. Across the 2.5 minutes of EC recordings, an average of 1.6 electrodes, 0.6 1198 trials, and 0.7 independent components were discarded. 1199 Despite the much smaller number of participants, four spectral components were recovered 1200 using both fPCA (Supplementary Figure 5A) and archetypal analysis (Supplementary Figure 5B). In 1201 both cases, the spectral profiles closely resembled those obtained in the larger datasets (e.g., Figures 1202 1D and 1E), albeit with slightly more noise. The variance explained remained high, with fPCA 1203 accounting for 88.4% and archetypal analysis for 95.4% of the variance. These results confirm that 1204 the four components—and notably the three distinct alpha components—are broadly expressed across 1205 participants and can be reliably detected even in smaller samples. 1206 Additionally, for each fPCA component, topographical maps —obtained by averaging factor 1207 scores across participants at each electrode (Supplementary Figure 5C)—closely matched those from 1208 the larger datasets (e.g., Figure 3C). This further supports that the different alpha components exhibit 1209 consistent and reliable spatial patterns at the scalp level. 1210 .CC-BY-NC 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted January 25, 2026. ; https://doi.org/10.64898/2026.01.24.701499doi: bioRxiv preprint 49 1211 Supplementary Figure 5. A) In the in-house dataset (N = 19), fPCA identified four main spectral components 1212 explaining ≥80% of the variance, including three components within the alpha band. B) Archetypal analysis 1213 identified four components explaining ≥95% of the variance, with spectral profiles closely matching those 1214 recovered by fPCA. C) Scalp EEG topographies of each spectral component, estimated from fPCA factor 1215 scores and averaged across participants. 1216 1217 .CC-BY-NC 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted January 25, 2026. ; https://doi.org/10.64898/2026.01.24.701499doi: bioRxiv preprint 50

1218 Bai, S., Menétrey, M. Q., & Pascucci, D. (2026). What is next? Predictable visual sequences are 1219 encoded with anticipatory biases and reduced neural responses. iScience, 114697. 1220 https://doi.org/10.1016/j.isci.2026.114697 1221 Park, J., Ho, R. L. M., Wang, W., Nguyen, V . Q., & Coombes, S. A. (2024). The effect of age on alpha 1222 rhythms in the human brain derived from source localized resting -state 1223 electroencephalography. NeuroImage, 292, 120614. 1224 https://doi.org/10.1016/j.neuroimage.2024.120614 1225 Scally, B., Burke, M. R., Bunce, D., & Delvenne, J. -F. (2018). Resting -state EEG power and 1226 connectivity are associated with alpha peak frequency slowing in healthy aging. Neurobiology 1227 of Aging, 71, 149–155. https://doi.org/10.1016/j.neurobiolaging.2018.07.004 1228 1229 .CC-BY-NC 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted January 25, 2026. ; https://doi.org/10.64898/2026.01.24.701499doi: bioRxiv preprint