Functional Differentiation losses in the brain with aging/autism [and measuring it with fMRI/fNIRS/etc]

#1

functional differentiation

2026 PNAS paper tied to DOI 10.1073/pnas.2527522123 is not mainly about “reduced functional differentiation” as an abstract theme; more specifically, it reports that age-related decline in large-scale network organization is shared across humans and mice. Researchers at The University of Texas at Dallas and Columbia University found that mice, like humans, show reduced brain-system segregation with age, although human brains are more integrated in youth and decline faster across the lifespan

Seems important to measure with fMRI and mb kernel
https://www.linkedin.com/in/matteo-vinao-carl-phd-181011123/recent-activity/all/.
https://www.youtube.com/watch?v=g2RMbDCKhVM [connectome.health is trying to track some of this!!]

just by intuition, functional differentiation helps with emotional intelligence/gracefulness/not being overly controlled by one narrative, being able to “agile switch” states when one’s focus is disrupted, not letting bad things completely destroy one’s day, etc…

Aging desegregation is a form of excess global coherence — everything bleeding into everything else, specialized modules losing their independence. This maps onto strong/excess central coherence in Frith’s sense: global integration at the expense of local specialization. The tethering hypothesis is consistent with this: the decreased divergence between association and primary cortex may impede the promotion of abstract information integration in the human brain — the association cortex becomes susceptible to external stimuli interference

" reduced dynamic control over when networks integrate versus segregate"

===

Chain A (plausible, ~0.45): Noise → blurred networks. If local populations within each area are less internally coherent (flatter slope = more async firing), the aggregate signal from that area is noisier. When you then compute correlations between areas, noise leaks through shared physiological sources and contaminates the between-network correlations disproportionately (between-network baseline is weak, so noise floor relatively matters more than for strong within-network correlations). Net effect: within-network correlation drops, between-network correlation rises → fMRI segregation score falls. The mechanism for “how increased noise causes dedifferentiation” runs through SNR asymmetry.

Chain B (plausible, ~0.40): E/I shift → loss of oscillatory fingerprints → networks lose identity. Different networks have distinct oscillatory signatures (DMN - alpha/low-frequency; FPN - beta; sensorimotor - mu; DAN - beta/gamma). These oscillatory fingerprints are how networks maintain segregation in the frequency domain, not just the spatial domain. Critically, with increasing participant age, there is increased 1/f noise [and] age-related changes in frontal and auditory PAC are specific to theta/high gamma PAC PubMed — the cross-frequency coupling that enables hierarchical network organization specifically collapses. PV interneurons are most vulnerable to aging/tau; losing them specifically kills gamma; losing gamma kills high-frequency network signatures; networks can no longer maintain their distinct “carrier frequencies.” Dedifferentiation via loss of frequency-specific organization.

I find Chain B more mechanistically compelling than A but A is probably more what fMRI segregation measures are capturing.

===

#2

longitudinal study in healthy older adults found progressive loss of intra-network functional specialization in both the DMN and executive control network, while internetwork DMN–ECN connectivity followed a nonlinear trajectory rather than a simple monotonic takeover by one system. Other work shows that older adults often have weaker ability to suppress irrelevant network communication and lower flexibility in reconfiguring networks for task demands.

A better formulation is that aging often involves functional blurring: network boundaries get less sharp, especially for associative systems. That blurring is linked to worse domain-general cognition and is partly explained by declining white-matter integrity. One influential 2021 study found that associative-system segregation declines especially from the late fifties onward, that this decline tracks global cognitive ability, and that white-matter integrity partly accounts for the change.

At the task level, reduced resting-state segregation also corresponds to less selective task recruitment. The Wig group’s 2017 work linked older adults’ less segregated resting-state topology to less differentiated task-evoked activity. So the picture is not merely “DMN too loud.” It is closer to: the system becomes less modular, less selective, and less able to recruit the right subnetworks at the right moments while keeping the others out of the way

==

Prior studies have suggested that at this stage of disease progression, cortical activity becomes less flexible, with slower temporal dynamics at low frequencies 86–89 and reduced entropy and complexity 87,90,91 This points to a potential mechanistic shift: structural pathology may constrain the flexibility with which cortical activity is rerouted. To

==

The principles by which diverse flow patterns are orchestrated across scales therefore remain 86 unresolved—specifically, how the brain alternates between global integration (long-range coordination of distributed 87 regions) and local segregation (fine-scale routing supporting specialised computations). Addressing this gap requires 88 methods that recover the higher-order, directed (causal) architecture of information flow and quantify how routing 89 motifs are conserved, coupled, and reconfigured across scales

==

Pang et al. derived geometric eigenmodes of the brain by applying the Laplace–Beltrami operator to cortical surfaces reconstructed from structural MRI and demonstrated that these geometric eigenmodes can approximate the spatial distribution of functional connectivity during rest and average task-evoked activity in the HCP dataset99. Overall, canonical structural eigenmodes have been shown to successfully capture spatial maps of brain activity, but they do not represent its causal spatiotemporal flow. In contrast, the canonical routing eigenmodes introduced here are distinct, as they are derived directly from directional EEG time series and thus intrinsically capture the directed spatiotemporal dynamics of brain activity.

#3

circuit-level causes of stuck dynamics / reduced state switching

Listed in rough order of what I think matters most for the “bad brain states” family (AD, unconsciousness, depression, aging). Confidence varies.

1. Neuromodulatory collapse (high confidence this matters)

The ascending arousal systems — locus coeruleus (NE), nucleus basalis of Meynert (ACh), VTA/SN (DA), raphe nuclei (5HT) — are what enable flexible state switching. They gate cortical gain, precision-weighting, and E-I balance dynamically.

  • LC-NE: earliest AD site; tau pathology starts in LC before cortex. these measures may probe effects of AD on ascending activating systems and reciprocal thalamocortical circuits in which oscillatory (de)synchronizing signals dynamically underpin cortical arousal in the regulation of quiet vigilance PubMed Central. Lose LC → lose the dynamic gain modulation that permits rapid state shifts.
  • NBM-ACh: critical for cortical desynchronization / alpha suppression, attention switching. cholinergic-monoaminergic interactions contribute to EEG slowing and dementia Jcimcr.
  • DA: controls prefrontal flexibility; D1/D2 both decline ~10%/decade in aging. Very likely contributes to reduced FPN flexibility.
  • 5-HT2A: the psychedelic target; enhances cortical entropy and flexibility. Antiparallel to the AD / stuck-brain picture — why psychedelics can in principle unstick things, at least acutely.

2. E/I imbalance (high confidence)

rsEEG revealed global neural slowing and disrupted synchrony… TMS-EEG identified network-specific local hyperexcitability in the parietal DMN and disrupted connectivity with frontal DMN regions Science. In AD, PV interneurons are specifically vulnerable → loss of fast inhibition → gamma breakdown → beta/gamma traveling waves disrupted. The system loses the ability to sustain the fast dynamics needed for flexible mode transitions.

This connects to your mitochondrial question — PV interneurons are the most energetically demanding cells in cortex (highest mitochondrial density, highest firing rates). Complex II/III deficits would hit PV interneurons disproportionately, biasing toward exactly this E-I imbalance picture. Relevant to your personal metabolomics thread, though I’d hold this connection weakly.

3. Thalamocortical gating (medium-high confidence)

The thalamic reticular nucleus (TRN) is the gate-keeper for cortical state switching. Mediodorsal thalamus gates prefrontal flexibility. In AD and aging, thalamic volume loss + altered TRN function → reduced capacity to reconfigure cortical states. in the basal ganglia the largest psilocybin-induced FC disruptions were seen in mediodorsal (MD) thalamus and anteromedial caudate PubMed Central — which is suggestive that thalamic gating is one of the key switches for getting “unstuck.”

#4

what an eigenvalue spectrum is

Start with the basic object. When you decompose signals into modes (like the divergence/vorticity eigenmodes in this paper), each mode has an associated eigenvalue. The eigenvalue tells you how much variance that mode captures in your data. Modes are conventionally ordered by eigenvalue magnitude: mode 1 has the largest eigenvalue, mode 2 the next largest, and so on.

Plot eigenvalue on the y-axis against mode order on the x-axis and you get the eigenvalue spectrum. In panel i of the figures you shared, this is what “VE” (variance explained) plots against “mode order” — same idea, eigenvalue normalized to fraction of total variance.

what “curvature” of that spectrum means

The shape of that decay curve tells you how variance is distributed across modes.

Steep drop-off (high curvature, sharp elbow): mode 1 captures a huge fraction of variance, mode 2 somewhat less, and by mode 10 there’s almost nothing left. The system’s activity is well-described by a small number of dominant modes. Low-dimensional, concentrated.

Gradual flat decay (low curvature, no clear elbow): variance is spread more evenly across many modes. Mode 1 isn’t that much bigger than mode 30. You need many modes to describe the system. High-dimensional, distributed.

Visually, a steep spectrum looks like the curve in Fig 3i — initial high values, rapid drop, long tail near zero. A flat spectrum would look like a near-horizontal line with a gentle slope.

related formalizations

People quantify this curvature a few different ways depending on field:

Effective dimensionality / participation ratio: PR = (Σλᵢ)² / Σλᵢ². This gives a single number: roughly, “how many modes does the system effectively use.” A steep spectrum gives low PR (system uses few modes); a flat spectrum gives high PR (system uses many modes).

Entropy of eigenvalue distribution: treat the normalized eigenvalues as a probability distribution and compute Shannon entropy. Low entropy = concentrated (steep); high entropy = spread (flat).

Exponent of power-law fit: if eigenvalues follow λₙ ∝ n^(-α), the exponent α characterizes the decay rate. Steeper decay = larger α.

All three are different ways of summarizing the same thing: how concentrated vs. distributed the variance is across modes.

why it matters for the aging/dedifferentiation discussion

This is where it connects to everything we’ve been talking about.

If aging/AD causes mode dominance shift (the Axis 2 compression from my last response) — more variance concentrated in low-order modes — the spectrum would get steeper with age. Fewer modes effectively used.

If aging causes loss of structured mode recruitment (system just becomes noisier, less organized) — variance distributed randomly rather than concentrated in the canonical modes — the spectrum would get flatter. Many modes used but none dominantly.

These are opposite predictions and they’d distinguish two fundamentally different dedifferentiation mechanisms:

  • Steeper spectrum with age → compression to global/integrative regime, loss of high-order recruitment. “Stuck in default mode” kind of picture.
  • Flatter spectrum with age → loss of organization, approach to random dynamics. “Noisy collapse” picture.

The 1/f flattening story I told earlier maps onto the second. The low-order mode dominance / stuck-dynamics story maps onto the first. In principle, these could both be happening (to different mode subsets or at different timescales), or one could dominate.

This is why I wanted the eigenvalue spectrum stratified by age — it would adjudicate between these two mechanisms in a way that single summary statistics can’t.

quick reality check on what’s in the figures you’ve shown me

Panel i in both the young and old figures shows the eigenvalue spectrum. Both look qualitatively similar: initial high values around 0.25 for vorticity, 0.1 for divergence, dropping rapidly by mode ~10, long tail to mode 128. I can’t reliably eyeball curvature differences between young and old from what’s visible — they look similar, which is consistent with the paper’s “preserved across age” thesis.

But as I keep noting: similar-looking at gross visual inspection doesn’t mean statistically indistinguishable. A 10% change in effective dimensionality would be biologically meaningful and probably not visible to eye. The question is whether the authors computed and reported effective dimensionality (or equivalent scalar curvature metric) age-stratified. If they did and it was null, that’s genuine evidence for preservation. If they didn’t compute it, that’s a gap.