How does an adaptive fractional state link cortical circuits to visual dynamics?

A regime that standard mean-field theory cannot describe

Classical balanced-network models explain irregular, low-rate spiking by cancelling excitation and inhibition on average, leaving Gaussian, short-memory fluctuations to drive neurons. That reduction cannot represent heavy-tailed excursions or correlations that persist across long timescales. The paper's central claim is that mouse visual cortex needs both departures at once, and that they are jointly measurable.

The authors name this working regime the adaptive fractional (AF) state. It combines three features that earlier models treated separately: heavy-tailed, superdiffusive fluctuations; long-range temporal dependence; and oscillations. The formal description is an effective mean-field theory in which a spatial fractional derivative generates the heavy-tailed excursions and a temporal fractional derivative generates the long-range memory, while oscillatory structure is preserved.

Three exponents that measure different things

The AF state is summarized by exponents that must not be conflated. The diffusion exponent a describes how far activity wanders over a time lag; values above 0.5 indicate superdiffusion, meaning excursions grow faster than ordinary diffusion. The spectral exponent b describes the slope of the power spectrum and indexes long-range temporal correlations. The variability exponent c is fitted from the rising portion of the Fano factor, the spike-count variance divided by its mean, across windows from 1 ms to 10 s.

These come from different signals and different aggregation. The diffusion and spectral exponents are derived from local field potentials, while the variability exponent is derived from spikes. Fano factor fits use median curves across units within each session, area, and layer, requiring at least five qualifying units, and exponents are reported as medians across sessions with bootstrap intervals. A change in one exponent therefore cannot be read directly as a change in another.

Recordings span six areas and four layers

The empirical base is the Allen Institute Visual Behavior Neuropixels dataset: local field potentials sampled at 1250 Hz with channels spaced 40 µm apart, plus spike data, from probes placed in primary visual cortex (VISp) and five higher areas (VISl, VISrl, VISal, VISpm, VISam). The analysis keeps 69 sessions from 41 mice, requiring each session to intersect all six areas with usable local field potentials and annotated target structures. Spontaneous activity is characterized during gray-screen epochs of a passive-viewing task with 250 ms stimuli.

Hierarchy is quantified with an anatomical score based on the feedforward or feedback character of interareal projections. The authors correlate each exponent with that score using Kendall's tau, a rank correlation chosen because it does not assume a functional form and tolerates heavy-tailed exponent distributions. Significance comes from a permutation test that shuffles hierarchy scores within each session, with correction across the twelve exponent-by-layer cells.

Higher areas trade superdiffusive excursions for temporal memory

Along the hierarchy, the two local-field-potential exponents move in opposite directions. From VISp to VISam the diffusion exponent falls from 0.62 to 0.51, meaning weaker superdiffusion, while the spectral exponent rises from -1.75 to -1.57, meaning stronger temporal persistence. The authors describe this as a coordinated shift along a single axis in the (a, b) plane, ordered by anatomical hierarchy score, rather than as a change in one characteristic timescale.

The variability exponent, measured from spikes rather than local field potentials, increases with hierarchy score at every layer and significantly at all but layer 4. Because it comes from a different signal in the same recordings, the authors treat it as corroborating the strengthening of long-range temporal dependence toward higher areas.

The shift is strongest in superficial layers

The hierarchical organization is not a uniform property of an area. Both gradients are strongest in layer 2/3. By layer 6, the correlation of the diffusion exponent with hierarchy score has reversed sign and the correlation of the spectral exponent has vanished. The authors conclude that the joint axis is a property of superficial circuits rather than a global feature of an area, consistent with the role they attribute to mouse superficial layers in recurrent and interareal processing.

This layer dependence also constrains how the result should be used. A session contributes a tau at a given layer only if at least four of the six areas carry an exponent there, a requirement that binds mainly for the variability exponent because its per-area fits need at least five units. That leaves between 51 and 69 sessions per layer for the variability exponent, against 68 for the diffusion and spectral exponents.

Weaker effective inhibition moves both exponents together

To test a mechanism, the authors built a spatially embedded recurrent spiking network with excitatory and inhibitory populations, conductance-based synapses, and spike-frequency adaptation in excitatory neurons, simulated with a 0.1 ms time step. A single control parameter, the synaptic inhibition-to-excitation ratio delta, displaces both exponents along a direction closely aligned with the experimentally measured hierarchical axis. Lowering delta weakens large-excursion dynamics while increasing temporal persistence.

The authors attribute this to the model's proximity to a transition between localized and delocalized activity. Under strong relative inhibition, firing is sparse and spatially disorganized, so a neuron's input is dominated by rare large synaptic events and its statistics are strongly superdiffusive. As inhibition weakens, activity packets coalesce and propagate coherently, so the input a neuron receives depends increasingly on the packet's recent trajectory, converting rapid excursions into temporal persistence. The diffusion exponent saturates near 0.64 at high delta and falls below 0.5 during the transition to the excitation-dominated regime.

The mechanism is consistent, not uniquely identified

The authors explicitly caution against reading the exponents as a direct measurement of excitation-inhibition balance. Synaptic kinetics, connectivity, adaptation, and other circuit properties can produce related displacements, and anatomical measures of inhibitory cell density do not specify the effective synaptic balance an individual neuron experiences. What the agreement among cortical dynamics, circuit simulations, and anatomical gradients supports is effective circuit balance as one axis along which cortical working regimes can be organized.

The anatomical correspondence is offered as consistency rather than proof: parvalbumin-expressing interneurons, which supply fast inhibition, are relatively more abundant in lower visual areas, and the inhibitory contribution to interareal connectivity decreases toward higher levels, with both gradients strongest in layer 2/3. The recordings are also spontaneous activity during gray-screen epochs in mice, so the exponents describe that condition rather than evoked visual responses.

Why two coordinates matter more than one timescale

The paper's most consequential reframing is that a single intrinsic timescale is an incomplete summary of the cortical hierarchy. The classical view orders areas by progressively longer integration windows toward higher-order cortex. The AF account keeps the idea of increasing temporal integration but adds a second coordinate: higher visual areas exchange superdiffusive excursions for temporal persistence, so the two exponents move in opposite directions along one axis.

The authors suggest this joint regime lets cortical areas express rapid responses and long-term temporal integration at the same time, which is the tension the opening question poses. The practical priority for readers is therefore to compare areas on both the diffusion and spectral exponents, layer by layer, rather than on either exponent alone or on a single fitted timescale. The variability exponent adds a spike-based check on temporal memory, but it is not interchangeable with the local-field-potential measures.

Sources used in this answer