Gamma Rhythms, PING & the Balanced State
Twenty papers form the literature backbone for the PING-gated-sparsity work in this project. They split into four roles: substrate — the balanced E/I background PING lives on; mechanism — what PING is, how it is generated, and the regime conditions under which the rhythm survives; function — what oscillatory gating actually buys a circuit (spike precision, selective routing); and framing — the wider rate-vs-temporal-coding debate that decides how much we are allowed to claim from a synchronous rhythm. Each entry below summarises the paper in a couple of sentences and then states how it relates to this project.
The papers
Substrate — the balanced background
Wilson & Cowan (1972). Derives coupled nonlinear differential equations for the dynamics of spatially localised E and I populations of model neurons. Phase-plane analysis shows the standard repertoire — fixed points, hysteresis, limit cycles — and crucially that the frequency of the limit-cycle oscillation is a monotonic function of stimulus intensity. The paper is the first principled rate-population mean-field treatment of an E/I circuit.
Relation to this project. This is the ancestor of the mean-field reduction in the methods of ar009. The collection’s PING circuit, averaged over a cycle, behaves like a Wilson–Cowan limit cycle whose frequency depends on input; that is what the rate-vs-input plots in nb036 are reading out at the population level. When the manuscript talks about “the mean-field picture”, Wilson & Cowan is the citation.
van Vreeswijk & Sompolinsky (1996). Shows that large, sparsely connected networks of E and I neurons with strong synapses settle into a balanced state: excitation and inhibition cancel on average, single-neuron membrane fluctuations of order √K drive Poisson-like irregular firing, and the network responds linearly to changes in external input despite the highly nonlinear single-unit dynamics. The balance is self-organising — no fine-tuning required — and produces the irregular, asynchronous baseline that cortical recordings actually show.
Relation to this project. This is the substrate on which PING is imposed. Without a balanced regime as the off-rhythm operating point, the collection’s claim that PING-gated firing is sparser than the alternative would be meaningless — sparser than what? Van Vreeswijk & Sompolinsky defines the “what”. The COBA baselines reported in nb024 and nb025 are balanced-state networks; the PING comparator imposes a gamma envelope on top of that substrate.
Vogels, Sprekeler, Zenke, Clopath & Gerstner (2011). Detailed E/I balance — co-tuning of excitatory and inhibitory inputs cell-by-cell — can be established and maintained homeostatically by Hebbian plasticity at inhibitory synapses. The plasticity rule converges to a state with sparse firing patterns, accommodates embedded memories without destabilising the background, and provides a mechanistic answer to “how does the cortex stay in the van Vreeswijk–Sompolinsky regime as it learns?”.
Relation to this project. This is the warrant for treating wEI and wIE as trainable parameters in the manuscript. The collection’s training recipe adjusts those weights with surrogate-gradient descent rather than with the Vogels rule, but the principle is the same: detailed balance is something a circuit learns, not something you hand-set once. If the manuscript wants to claim that the operating point we land on is biologically plausible (rather than mathematically convenient), this is the paper that says learned detailed balance is what cortex actually does.
Mechanism — what PING is
Börgers & Kopell (2005). The paper that named PING. In a network of E (pyramidal-like) and I (fast-spiking interneuron-like) cells with mutual coupling and noisy external drive, strong PING arises when the drive to I cells is low enough that I cannot fire without being prompted by an E volley — so E gates I, I silences E, and the loop sets the rhythm. The paper identifies the two failure modes when I-drive grows: phase walkthrough (I cells get ahead of E and fire on their own) and E-cell suppression (I activity is high enough that E never escapes). The “G” in PING refers to gamma specifically: the analysis shows the mechanism is only robust to noise in the gamma range when parameters are tuned.
Relation to this project. This is the operational definition of PING that the collection’s architecture is built around. The COBA-PING network in ar009 — fixed E↔I matrices, conductance-based synapses, every E driven on each cycle — is a strong-PING circuit in exactly the sense Börgers & Kopell define. The (wEI, wIE) sweep in nb036 traces the boundary at which the rhythm leaves the strong-PING regime: increase wIE far enough and you walk through the I-drive threshold from this paper.
Buzsáki & Wang (2012). Review of gamma mechanisms across regions. Their main claims: gamma-band rhythmogenesis is “inextricably tied to perisomatic inhibition”; gamma is typically short-lived and emerges from coordinated E/I rather than steady oscillation; gamma frequency varies extensively depending on which mechanism (PING, ING, etc.) is engaged; and gamma rhythm coexists with irregular single-neuron firing — power in the gamma band is not the same as periodic spiking. They also stress that gamma is modulated by slower rhythms (theta, alpha), and that this cross-frequency coupling is part of what gamma is for.
Relation to this project. The “gamma power ≠ periodic single-neuron firing” point is the disclaimer the collection relies on. The notebooks measure population rhythmicity and per-spike economy, not per-neuron periodicity, and Buzsáki & Wang is the citation that authorises that distinction. The review also serves as the umbrella reference for any time the manuscript (ar009, ar010) needs to invoke “the gamma mechanism” without picking apart which sub-variant — Buzsáki & Wang is where that umbrella points.
Börgers, Talei Franzesi, LeBeau, Boyden & Kopell (2012). Follow-up to Börgers & Kopell 2005. Studies how gamma rhythms break down as the driven E ensemble becomes small or as synaptic strengths become heterogeneous. They distinguish strong PING (tonic drive to all E cells, every E participates every cycle) from weak PING (stochastic drive, fractional participation per cycle). The headline result is a soft lower bound on assembly size: heterogeneous networks must lock to synchrony within a few gamma cycles or not at all, and that requires E↔I synapses strong enough that a single E volley reliably triggers the I population.
Relation to this project. This is the regime taxonomy the collection lives inside. The (wEI, wIE) coupling sweeps in nb036 are explicitly probing the Börgers et al. boundary between rhythmic and arhythmic operation; the accuracy-vs-energy frontier we report sits at the strong-PING corner of that map. When a sweep cell loses rhythmicity it is failing the “E volley triggers I” criterion stated here.
Atallah & Scanziani (2009). In vivo and in vitro recordings from rat CA3 show that gamma amplitude and frequency vary cycle-by-cycle: one cycle’s amplitude predicts the latency to the next. The mechanism is that synaptic excitation varies over an order of magnitude across cycles, and inhibition tracks it immediately and proportionally — so the E/I ratio is preserved while the absolute conductances scale, and the resulting decay time of the I-tracked envelope sets the next inter-cycle interval. Tight detailed E/I balance is therefore the knob that sets instantaneous γ frequency.
Relation to this project. This paper is the experimental warrant for the project’s central design choice: making wEI and wIE the trainable scientific knobs. nb036’s (wEI, wIE) heatmap is reading out exactly the Atallah-Scanziani relationship — re-scaling I tracking of E moves the rhythm’s mean rate without destroying it, and the diagonal where E and I scale together is the “preserved-ratio” axis the paper predicts. If we found a useful operating point off the diagonal we would be claiming something the Atallah-Scanziani data argues against; instead we don’t, and the frontier hugs the diagonal.
Börgers (2017). Chapter 30 of An Introduction to Modeling Neuronal Dynamics — the textbook treatment of PING. Builds the E-cell / I-cell loop from first principles: an E volley recruits the I population, I silences E, the GABA decay releases E, and the cycle repeats at gamma frequency. The cleanest pedagogical statement of the mechanism, with the cycle period explicitly tied to the inhibitory decay time constant.
Relation to this project. The newcomer-facing reference for “what is PING?” — the loop diagram in this chapter is the architecture that ar003 implements as COBANet and nb023 characterises in isolation. Where Börgers & Kopell (2005) is the research-paper definition, this is the one to send a reader who has never met PING before.
Kopell, Börgers, Pervouchine, Malerba & Tort (2010). A chapter surveying reduced biophysical models of hippocampal gamma (30–90 Hz) and theta (4–12 Hz) rhythms and their interaction, deliberately simple enough to expose why the models behave as they do rather than just reproducing data. Classifies rhythms by mechanism rather than frequency, and treats the dynamical role of fast-spiking and O-LM interneurons separately before combining them.
Relation to this project. Background for the modelling philosophy the collection adopts: reduced, mechanism-first networks chosen to make the dynamics legible. The collection uses only the gamma half of this picture (no theta, no O-LM cells), but the “classify by mechanism, keep the model minimal” stance behind ar009’s COBA-PING reduction is the one this chapter argues for.
Viriyopase, Memmesheimer & Gielen (2016). Asks what happens when both gamma-generating mechanisms — interneuronal gamma (ING, mutual I↔I inhibition) and pyramidal-interneuronal gamma (PING, the E↔I loop) — are present in the same network. The mechanisms compete: the one generating the higher frequency wins and suppresses the other. For type-I interneurons the network frequency is at or slightly above the higher of the two component frequencies; gap-junction and chemical-synapse routes can cooperate or compete depending on interneuron type.
Relation to this project. This is the paper behind the architecture’s – init ratio — strong I→E relative to E→I keeps the network in the PING-dominated rather than ING-dominated corner. It is the reference for why adding (the inhibitory self-coupling exposed in nb050) matters: turn it up and you move toward the ING regime this paper maps, which competes with the PING loop the collection relies on.
Lee & Jones (2013). A laminar-cortex spiking model built to interpret human MEG/EEG gamma. Two distinct generators are compared: spiking PING, whose period is set by the GABA decay time constant, and subthreshold gamma driven by gamma-periodic exogenous input. The paper shows the two leave distinguishable current-dipole signatures, and — the load-bearing result for us — that high-frequency gamma (100–150 Hz) is not a separate rhythm but emerges from a temporal feature of single PING cycles, i.e. harmonics of the fundamental.
Relation to this project. Two direct connections. First, “GABA decay sets the period” is exactly the relationship the sweep in nb041 measures — the affine law rides on this paper’s mechanism. Second, the high-gamma-as-harmonics result explains the comb of peaks at integer multiples of in the PSDs of nb023 and nb050: those are not extra rhythms, they are the non-sinusoidal shape of one PING cycle.
Nguyen & Rubchinsky (2021). Studies the temporal patterning of synchrony in a conductance-based PING network — not just how strongly cells synchronise but the moment-to-moment structure of synchronous and desynchronous intervals. The key finding: changes in synaptic strength and membrane kinetics can alter the temporal pattern of synchrony independently of the average synchrony strength, with a tendency toward short desynchronisation events of the kind seen experimentally.
Relation to this project. Relevant to the (wEI, wIE) sweeps in nb036: it says the same weight knobs the collection trains do more than set a mean rate — they reshape the fine temporal structure of synchrony. A caution, really: a single average-rhythmicity number (a PSD peak height) can hide changes in the temporal patterning this paper resolves.
Segneri, Bi, Olmi & Torcini (2020). Uses a next-generation neural mass model (an exact mean-field reduction of spiking quadratic-integrate-and-fire networks) to study theta-nested gamma. Both PING and ING set-ups produce gamma via a Hopf bifurcation under slow theta-frequency forcing, with gamma amplitude growing with forcing intensity and phase-amplitude coupling to the theta phase.
Relation to this project. The Hopf-bifurcation framing is the mean-field counterpart of nb033, where the collection derives the PING recruitment threshold and oscillation frequency from a Wilson–Cowan reduction. Segneri et al. is the exact-reduction version of the same move (QIF → neural mass rather than spiking → Wilson–Cowan), and the reference for treating gamma onset as a Hopf bifurcation of an E/I mean field.
Nandi, Valla & di Volo (2024). Shows that an exact neural mass model of E/I quadratic-integrate-and-fire networks predicts intrinsic bursting gamma — gamma that appears in short bursts rather than as a steady cycle — arising from deterministic collective chaos, with no external noise required. The bursting regime coexists with highly irregular single-neuron firing.
Relation to this project. Directly relevant to nb050’s finding that the same architecture supports a chaotic balanced (Vreeswijk–Sompolinsky) state alongside the gamma-clocked PING one. Nandi et al. is the mean-field statement that collective chaos and gamma are not mutually exclusive — gamma can be a chaotic, bursty signal — which is the bridge between the rhythmic and asynchronous-irregular regimes the collection now spans.
Function — what oscillatory gating buys you
Schaefer, Angelo, Spors & Margrie (2006). Membrane potential oscillations (MPOs) in olfactory mitral cells, both synaptically driven and intrinsic, sharpen action potential timing by removing the slow-drift component of Vm that otherwise lets jitter accumulate across a spike train. Empirically this enables discrimination of up to ≈1000 stimuli at oscillation frequencies from 3–65 Hz, and stimuli arriving on the trough / early rising phase are discriminated most robustly.
Relation to this project. This is the function-level justification for using PING at all. The collection’s claim is that gamma-locked spikes are sparser and more informative per spike than free-running spikes; Schaefer et al. is the cellular-level explanation of why per-spike information goes up under an MPO. The phase-of-firing readouts in nb023 and the rate-vs-accuracy gap reported in nb024/nb025 (PING beats COBA at matched accuracy) are the network-scale reflection of this single-cell precision effect.
Fries (2015). The canonical statement of the communication-through-coherence (CTC) hypothesis: gamma-band synchronisation gates communication between neuronal groups by carving out windows of excitation flanked by inhibition, and the postsynaptic group only “listens” during those windows. Attention sharpens the gamma signature of the attended representation, which selectively entrains the postsynaptic group and shuts out competing inputs. Fries also embeds gamma in a multi-rhythm hierarchy — theta (≈7 Hz) samples attention, alpha-beta (8–20 Hz) carries top-down control, gamma (30–90 Hz) carries bottom-up content.
Relation to this project. CTC is the function-level claim the collection is closest to making. The PING substrate in the manuscript provides exactly the excitation-flanked-by-inhibition window Fries describes, and the per-spike economy result is what you would expect if the postsynaptic side is only listening during the gamma trough. The collection does not currently test CTC directly — there is no second cortical area downstream in our networks — but the manuscript’s gating-and-routing framing rests on Fries being right at least directionally.
Akam & Kullmann (2012). A rate-coded model of selective communication through coherent gain modulation, and the formal critique of CTC. Two converging input networks each carry a population code; only one is “target”. The receiving network can recover the target by oscillatory gain that is coherent with the target’s modulation — but only if the distractor’s oscillation differs from the target’s in amplitude, phase, or frequency. Distractors that oscillate incoherently in the same band severely degrade transfer because the overlap with the gain modulation fluctuates trial-to-trial. The paper also requires that target modulation be strong relative to single-neuron noise.
Relation to this project. Where Fries asserts that gamma gates communication, Akam & Kullmann states the conditions under which it can. This is the paper that disciplines the manuscript’s claims: when our notebooks show PING improving accuracy, we should resist the temptation to call that “communication-through-coherence” unless the experiment actually has competing inputs to separate. The PING-as-gate framing in ar009 cites Fries for the proposition and Akam & Kullmann for the small print.
Framing — how much can we claim?
Ainsworth, Lee, Cunningham, Traub, Kopell & Whittington (2012). Review that rejects the rate-vs-temporal-coding dichotomy and argues for a synergistic view: in sensory cortex, multiple types of gamma rhythm coexist, and rate and timing are co-expressed on the same circuit elements as stimulus strength varies. Different gamma sub-types arise from different interneuron mechanisms and serve different population-coding roles.
Relation to this project. This is the paper that authorises the project’s readout structure. We train on rate-style readouts but report under a rhythmic substrate, and treat that as ordinary cortex-like behaviour rather than a contradiction. The collection’s headline plots are rate vs energy — a rate-coded outcome — but the underlying activity is gamma-locked; Ainsworth et al. is why those two facts coexist without requiring us to pick a side.
Shadlen & Movshon (1999). A skeptical review of the temporal-binding hypothesis. The argument: synchrony among neurons may exist and may be modulated, but it is neither necessary nor sufficient as a representational code for binding; spike timing relative to a population code suffers from severe encoding/decoding problems, and most evidence cited in favour of binding-by-synchrony is consistent with synchrony being an epiphenomenon of shared input.
Relation to this project. The deliberate counterweight on the shelf. It is easy, when working inside a PING-gated network, to slide from “the rhythm is useful” to “the rhythm is the code”. Shadlen & Movshon is the reason the collection’s claims are scoped narrowly — rate-accuracy-per-energy at matched accuracy, not binding or content-by-synchrony. When the manuscript (ar009, ar010) talks about what gamma buys, it buys spike-economy and precision, not feature binding. Citing Shadlen & Movshon is the project’s pre-emptive disclaimer.
Xing, Shen, Burns, Yeh, Shapley & Li (2012). Records gamma-band LFP in monkey V1 in both awake and anaesthetised states and finds the temporal statistics of gamma identical across states: large variability in peak frequency, brief oscillatory epochs (< 100 ms on average), and stochastic incidence and duration of events. The conclusion: gamma-band activity is temporally unstructured — better described as filtered neural noise from a damped oscillator than as a clock — and is “too random to serve as a clock signal” for synchronising responses between areas.
Relation to this project. The strongest in-band skeptic. Where Shadlen & Movshon challenges synchrony-as-code in general, Xing et al. challenges the gamma clock specifically, on data. It is the experimental reason the collection never claims its gamma is a precise metronome: the rhythm provides a statistical envelope and a per-spike economy, not cycle-accurate timing. Notably, the “damped-oscillator filtered-noise” description is exactly the regime nb050 lands in when it drives the same network into the asynchronous-irregular state — Xing et al. is the data this project would point to if asked whether real cortical gamma looks more like the clean PING limit cycle or the noisy balanced state. (The honest answer: more like the latter.)
How the twenty fit together
Read in this order. Substrate: Wilson & Cowan gives the mean-field language; van Vreeswijk & Sompolinsky shows that the natural off-rhythm state is balanced and irregular; Vogels et al. shows that the balance can be learned, which justifies training wEI and wIE. Mechanism: Börgers & Kopell 2005 defines PING and Börgers (2017) is the textbook version of that definition; Buzsáki & Wang and Kopell et al. 2010 frame it inside the broader gamma (and theta) literature; Börgers et al. 2012 fixes the regime conditions; Viriyopase et al. sets PING against its ING rival (and motivates the wIE/wEI ratio and wII knobs); Atallah & Scanziani and Lee & Jones give the cycle-by-cycle and GABA-decay readings of what sets gamma frequency; Nguyen & Rubchinsky warns that the same weights reshape fine temporal patterning; Segneri et al. and Nandi et al. supply the exact-mean-field counterparts — gamma onset as a Hopf bifurcation, and gamma that can itself be chaotic and bursty. Function: Schaefer et al. and Fries argue what the oscillation is good for (precision, selective routing); Akam & Kullmann supplies the constraints under which the routing actually works. Framing: Ainsworth et al. lets rate-style readouts coexist with a rhythmic substrate; Shadlen & Movshon scopes the claims down from the rhythm is the code to the rhythm is useful; Xing et al. supplies the data that real gamma is noisy and clock-like only statistically. The PING-gated-sparsity manuscript (ar009, ar010) sits in the intersection of all four: it accepts the substrate, accepts the mechanism, claims the precision-and-economy function, frames results in rate-and-rhythm synergy, and does not claim binding.