Low dimensional dynamics of a sparse balanced synaptic network of quadratic integrate-and-fire neurons

arXiv:2508.06253 · q-bio.NC, cond-mat.dis-nn, cond-mat.stat-mech · Submitted 2025-08-08 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Genomics Radio. Generated commentary on the latest computational biology and genomics papers.

Ines: We're looking at the paper: "Low dimensional dynamics of a sparse balanced synaptic network of quadratic integrate-and-fire neurons".

Marcus: Here's the gist of the paper. Low dimensional dynamics of a sparse balanced synaptic network of quadratic integrate-and-fire neurons is well representable by the stochastic dynamics of a generic neuron subject to an effective shot noise.

Ines: Importantly, the most sophisticated (but robust) collective regimes of the network do not allow for the diffusion approximation, which is routinely adopted for a shot noise in mathematical neuroscience.

Marcus: This paper derives low dimensional model reductions on the basis of circular cumulants specifically for a population of quadratic integrate-and-fire neurons (QIFs) and a shot noise.

Ines: The accuracy of a low dimensional model indicates and explains a low embedding dimensionality of the macroscopic collective dynamics of the network.

Title and authors: Ines: So, we're looking at this paper, "Low dimensional dynamics of a sparse balanced synaptic network of quadratic integrate-and-fire neurons," and the authors are proposing a way to look at these complex networks using something called circular cumulants.

Marcus: Yeah, it’s about taking the full mean field theory for these inhibitory networks and reducing it down to a much simpler model, specifically based on two circular cumulants. The core idea is that this reduction is more accurate than the standard diffusion approximation we usually use for shot noise in neuroscience.

Yuki: From a population genetics standpoint, what I want to know is how this relates to the actual biological systems we see in nature. Are these dynamics something that would actually emerge when you have both excitation and inhibition interacting?

Ines: Exactly, Yuki. The paper says that for the most robust collective regimes in this network, the diffusion approximation just doesn't work anymore, which is a huge piece of information because those regimes can be biologically relevant. They are showing how to model things that standard methods miss >

Marcus: Right. And what they do is derive equations for these two circular cumulants from an infinite chain of equations and then truncate that chain down to just two terms, which we call the two circular cumulant reduction. This is basically a shortcut to getting a low dimensional model for the macroscopic dynamics.

Ines: It’s interesting because it shows that you can get a low embedding dimensionality for the macroscopic collective dynamics of the network, and they validate this by showing it works well for time-independent regimes and even dynamic responses to periodic modulation deep into parameter spaces where we usually give up on approximations >

Yuki: So, if this model is accurate at very low excitatory currents, like i0 equals zero point zero one times g squared zero, what does that tell us about the kinds of biological conditions we can realistically study with these models?

Marcus: It tells us that for those low current levels, this two circular cumulant reduction gives reasonable accuracy for the time independent states. However, if you look at their comparison with other methods like the Ott-Antonsen Ansatz, they find that the Ott-Antonsen Ansatz only gets accurate when i0 is much higher >

Title and authors: Ines: That’s a key distinction. The Ott-Antonsen Ansatz fundamentally cannot reproduce those noise induced oscillations we see in neural circuits when you use the diffusion approximation >

Marcus: The two circular cumulant model, though, it captures the dynamic response for i0 equal point zero two times g squared zero and its error drops rapidly as i0 over g squared zero increases when there is strong modulation >

Yuki: That sounds promising because it suggests a way to study those noise induced oscillations more accurately than what we could get from the simpler approximations we usually rely on >

Ines: It does, but they also lay out some limitations. They state that the diffusion approximation is valid as long as the ratio of a over sigma V is small, and they give an estimate for this accuracy measure a over sigma V to be approximately two minus g times a-g to the power of four root K nu much less than one >

Marcus: And this estimate gives us an optimistic boundary for when that diffusion approximation might hold up, suggesting it's accurate for K one to two orders of magnitude larger than the KD value, which is calculated based on i0 over g squared zero squared >

Yuki: That helps frame what we can expect from these models in terms of sample size and network density. It points toward regimes where we might need to use these more sophisticated tools instead of just assuming the standard noise behavior applies >

Ines: And for time dependent modulation, they modify the continuity equation to include a time dependent external current I(t) equal I zero times one plus eta of t, and this leads to a modified dynamical system that depends on i0 through alpha equal g zero over K three over four root i0 >

Marcus: In Fourier space, when they extend the extension of equation six, they get equation twenty-three. The relation between the firing rate and the mean membrane potential is given by pi nu minus i v equals p over I zero W psi >

Yuki: So what does this mean for predicting population firing rates under different external driving frequencies? Can we use this framework to predict how a network will respond when you change the input signal over time >

Title and authors: Ines: The two circular cumulant model captures the dynamic response for i0 equal point zero two times g squared zero and its error drops as i0 over g squared zero increases for strong modulation, which means it should give better predictions for resonant and off-resonance modulation frequencies >

Marcus: And they also show how the diffusion approximation version of the continuity equation yields a modified version of equation four, leading to a modified version of equation twenty-three with a truncated matrix I DA >

Yuki: So this is about how noise affects the dynamics when you consider these external drives. It’s not just about steady states anymore >

Ines: Yeah, it shows that when you move into dynamic regimes, like those driven by time dependent currents, the two circular cumulant reduction still gives us a useful low dimensional model for understanding what happens >

Marcus: The paper concludes that the two circular cumulant reduction model might be applicable for theoretical studies of self organized global oscillations in sparse networks where both excitatory and inhibitory synaptic links are present and collective oscillations emerge due to the interplay of excitation and inhibition >

Yuki: So, if we can use this model to study those complex, balanced systems with both excitation and inhibition, what does that mean for understanding how these things organize themselves across different scales of the system >

Ines: It means we have a low dimensional neural mass model that explains the low embedding dimensionality of attracting macroscopic regimes under constant and periodically modulated conditions, which is a big piece of information >

Marcus: So to wrap up on this paper, it’s about using circular cumulants to derive a low dimensional neural mass model for sparse balanced synaptic networks of quadratic integrate-and-fire neurons >

Yuki: The two circular cumulant reduction model is useful for studying self organized global oscillations in sparse networks where both excitation and inhibition are present and collective oscillations emerge due to the interplay of excitation and inhibition >

Ines: That's right. We’ve got a low dimensional neural mass model that explains the low embedding dimensionality of attracting macroscopic regimes both under constant and periodically modulated conditions, which is what this paper is about >

The paper's summary: Ines: So we've been looking at this paper about quadratic integrate-and-fire neurons in sparse networks, and here’s the summary of what they found on how to simplify that huge math problem.

Marcus: Basically, they're taking this really complicated set of equations for the whole network and reducing it down to a very simple one based only on two circular cumulants.

Yuki: So, if I understand that right, it’s like boiling down a massive chemical reaction to just two key components that still tell you how the whole thing behaves?

Ines: Exactly. They say this reduction is more accurate than the standard diffusion approximation we usually use for noise in these systems.

Marcus: That’s the crucial part. The diffusion approximation breaks down in certain regimes, but this two circular cumulant model still holds up well when you look at time-independent states and even when things are changing periodically.

Yuki: So, if we think about a real biological system, like a neuron network where excitation and inhibition are balanced—which is what these QIF neurons represent—this model seems to capture the collective behavior better than the standard simplified noise models.

Ines: It does. They show that this reduction gives a low dimensionality for the macroscopic dynamics, which means we can understand complex network states with much less math than before.

Marcus: And they give us some concrete numbers on where it works, like for very low excitatory currents, around zero point zero one times g squared zero. But there are limits.

Yuki: What do those limits mean for the actual biology? Are we only talking about very weak stimuli in these models?

Ines: It means the model is good for those specific low current levels, but if you push it into a regime where the noise-induced oscillations kick in, like when excitation becomes stronger, this model starts underestimating how much those oscillations suppress collective modes.

Marcus: They compare it to another common method called the Ott–Antonsen Ansatz. The OA works fine for higher excitatory currents, but it just can't reproduce those noise-induced dynamics that we see in real neural circuits when you use the diffusion approximation.

Yuki: So, if we want to study those self-organized global oscillations in sparse networks with both excitation and inhibition present, this two circular cumulant model might actually be a better tool than what we currently rely on.

Ines: That’s the point. It offers a way to model those complex, balanced systems where everything is interacting, and it explains why the macroscopic states of these networks have such low embedding dimensionality under both constant and time-varying conditions.

Marcus: But they also flag when the diffusion approximation fails, giving us an estimate involving a ratio of parameters, which tells you exactly when you need to stop using that simpler method.

Yuki: So we’ve got a way to build a more detailed picture of how these networks organize themselves across different scales, even when the standard assumptions break down.

The paper's improvements: Tom: So we’re looking at how this paper suggests ways to make their model even better than what they first presented on low dimensional dynamics for these sparse networks of quadratic integrate-and-fire neurons.

Ines: The main improvement is moving from just a two circular cumulant reduction to using more complex ones, specifically introducing four circular cumulants.

Marcus: So instead of just two terms, they use four terms to describe the dynamics, and this makes the model more detailed in capturing the behavior of those collective states.

Yuki: What does adding those extra terms actually let us see that we couldn't before? Does it help us connect this to how these networks evolve over time?

Ines: It helps capture more complex behaviors, like noise-induced oscillations, which we talked about earlier. The authors suggest that the higher order cumulants are necessary to accurately model those specific dynamic responses.

Marcus: They also show that by using this expanded set of equations, they can get a better handle on how the network reacts when you introduce time dependent modulation into the external current.

Yuki: So if we look at population dynamics over time, does this four circular cumulant approach give us a clearer picture of those resonant and off-resonant frequencies?

Ines: It should. They show that by using these higher order terms, the model gets better accuracy for i0 equal point zero two times g squared zero under strong modulation compared to the simpler two term reduction.

Marcus: That means when you’re trying to predict how a network responds to an external drive that changes over time, this four circular cumulant version gives a more reliable prediction.

Yuki: It seems like it provides a better tool for population genetics because it lets us better support the kind of complex, dynamic organization we see in real species.

Ines: Exactly. The biological implication is that if you want to study self organized global oscillations in these balanced networks, this improved model gives you a more accurate way to map out those attracting macroscopic regimes.

Marcus: They also address the limitations we discussed earlier by providing a clearer framework for assessing when the diffusion approximation actually fails, giving us a better estimate for that accuracy measure.

Yuki: So what does this all mean for how we study these biological systems across different scales? Does it change our way of thinking about these sparse networks?

Conclusion: Ines: So to wrap up this look at "Low dimensional dynamics of a sparse balanced synaptic network of quadratic integrate-and-fire neurons," the main thing is that they’ve built a low dimensional neural mass model for these networks using circular cumulants.

Marcus: They found that this two circular cumulant reduction accurately captures the macroscopic collective dynamics, even when the standard diffusion approximation breaks down.

Yuki: It seems like they’ve given us a better way to see how both excitation and inhibition organize themselves in these sparse systems across different conditions.

Ines: That’s right. For time independent states, this model works down to very low excitatory currents, around zero point zero one times g squared zero, which is a specific number for the dynamics.

Marcus: And they show that even under periodic modulation—when the external current changes over time—this reduction still gives reasonable predictions for i0 equal point zero two times g squared zero.

Yuki: So what does this mean for the wider history of species? Can we use these kinds of models to support claims about how these complex interactions shape evolution?

Ines: It means we have a low dimensional neural mass model that explains the low embedding dimensionality of attracting macroscopic regimes both under constant and periodically modulated conditions.

Marcus: And they also give us a clearer estimate for when the diffusion approximation fails, which is important context for any big claim you make about these networks.

Yuki: So it’s about getting a better statistical handle on what the sample can actually support when we look at these large, sparse systems.

Ines: Exactly. The work on "Low dimensional dynamics of a sparse balanced synaptic network of quadratic integrate-and-fire neurons" gives us a framework for theoretical studies in this area.

Marcus: It's a solid step forward for understanding how these networks organize themselves when the noise is more complex than simple shot noise.

Yuki: It opens up possibilities for modeling those self organized global oscillations we’ve been interested in seeing emerge from excitation and inhibition interplay.

Institute of Continuous Media Mechanics, UB RAS · Institute of Physics and Mathematics, Perm State University

q-bio.NC, cond-mat.dis-nn, cond-mat.stat-mech

Submitted: 2025-08-08

Updated: 2025-10-30

Comments: 12 pages, 5 figures

Journal ref: Eur. Phys. J. Spec. Top., Vol. 235, no. 10, 2511-2523 (2026)

DOI: 10.1140/epjs/s11734-025-02096-8

License: http://creativecommons.org/licenses/by/4.0/

Importance score: 79/100

The gist: Low dimensional dynamics of a sparse balanced synaptic network of quadratic integrate-and-fire neurons is well representable by the stochastic dynamics of a generic neuron subject to an effective…

Key concepts

Circular Cumulants (CC)
These are mathematical tools used to reduce an infinite chain of equations describing network dynamics into a much smaller set. By focusing on specific moments (like the first and second), researchers can capture the essential, low-dimensional behavior of the system without needing all the complex information.
Quadratic Integrate-and-Fire Neurons (QIFs)
These are a type of artificial neuron model used to simulate biological neurons. They have a specific mathematical rule for how their membrane potential changes when they receive input, making them suitable for modeling synaptic networks where spikes occur based on reaching a certain voltage threshold.
Diffusion Approximation
This is a common simplification in neuroscience where complex stochastic noise is treated as simple Brownian motion. The paper shows that this approximation fails in certain collective regimes of these neural networks, meaning it cannot accurately describe the system's behavior when it exhibits strong, coordinated activity.

Terminology

Summary

Low dimensional dynamics of a sparse balanced synaptic network of quadratic integrate-and-fire neurons is well representable by the stochastic dynamics of a generic neuron subject to an effective shot noise. p. 0

Where the paper says it

Goldobin1,2,b 1 Institute of Continuous Media Mechanics, UB RAS, Academician Korolev Street 1, 614013 Perm, Russia 2 Institute of Physics and Mathematics, Perm State University, Bukirev Street 15, 614990 Perm, Russia October 31, 2025 Abstract. Kinetics of a balanced network of neurons with a sparse grid of synaptic links is well representable by the stochastic dynamics of a generic neuron subject to an effective shot noise.

Importantly, the most sophisticated (but robust) collective regimes of the network do not allow for the diffusion approximation, which is routinely adopted for a shot noise in mathematical neuroscience. p. 0
Where the paper says it

The rate of delta-pulses of the noise is determined self-consistently from the probability density of the neuron states. Importantly, the most sophisticated (but robust) collective regimes of the network do not allow for the diffusion approximation, which is routinely adopted for a shot noise in mathematical neuroscience.

This paper derives low dimensional model reductions on the basis of circular cumulants specifically for a population of quadratic integrate-and-fire neurons (QIFs) and a shot noise. p. 0
Where the paper says it

Neither “next-generation neural mass models” nor their upgraded versions for Gaussian noise can be employed for this case. In this paper we derive low dimensional model reductions on the basis of circular cumulants specifically for a population of quadratic integrate-and-fire neurons (QIFs) and a shot noise.

The accuracy of a low dimensional model indicates and explains a low embedding dimensionality of the macroscopic collective dynamics of the network. p. 0
Where the paper says it

The low dimensional model is examined with numerical simulations and found to be accurate for time-independent states and dynamic response to a periodic modulation deep into the parameter domain where the diffusion approximation is not applicable. The accuracy of a low dimensional model indicates and explains a low embedding dimensionality of the macroscopic collective dynamics of the network.

Model Formulation for Network Dynamics

The kinetics equations for the complete mean field theory of a homogeneous inhibitory network of quadratic integrate-and-fire neurons are considered. The membrane potentials Vj of QIFs evolve according to equation (1). p. 0

Where the paper says it

We consider a dynamically balanced network of N pulse-coupled QIFs with a sparse grid of inhibitory synaptic links. Membrane potentials Vj of QIFs evolve according to the following equations [38,39]: V˙ j = V 2 j + I − a X N k=1 X n ǫjkδ(t − t (n) k ) , (1) a e-mail: ageeva mv00@mail.

For a sufficiently sparse network with in-degree K ≪ N, the spike train can be assumed to be uncorrelated and Poissonian. p. 1
Where the paper says it

Here we explicitly indicate the scaling with K required for the dynamical balance. For a sufficiently sparse network with in- degree K ≪ N the spike train can be assumed to be uncorrelated and Poissonian.

In this case the mean field dynamics of a generic QIF can be represented in terms of the Langevin equation (2). p. 1
Where the paper says it

For a sufficiently sparse network with in- degree K ≪ N the spike train can be assumed to be uncorrelated and Poissonian. In this case the mean field dynamics of a generic QIF can be represented in the terms of the following Langevin equation: V˙ = V 2 + I − aS(t), (2) where S(t) is a Poissonian train of δ spikes with rate R(t) = Kν(t) and ν(t) is the population firing rate.

The population dynamics are described by the continuity equation (3). p. 1
Where the paper says it

In this case the mean field dynamics of a generic QIF can be represented in the terms of the following Langevin equation: V˙ = V 2 + I − aS(t), (2) where S(t) is a Poissonian train of δ spikes with rate R(t) = Kν(t) and ν(t) is the population firing rate. For a homogeneous population within the mean-field framework the population dynamics can be described in terms of the membrane potential probability density function P(V, t), whose time evolution is given by the continuity equation: ∂P(V, t) ∂t = − ∂ ∂V hi0 √ K + V 2 P(V, t) i + Kνh P(V + a, t) − P(V, t) i , a = g0 √ K .

Derivation of Low Dimensional Reductions

The paper introduces circular cumulants (CC) to reduce the infinite chain of equations for Kuramoto–Daido order parameters. A two circular cumulant reduction is derived from the infinite equation chain. p. 1

Where the paper says it

In Section 2, we present a detailed derivation of the complete mean field model— an infinite chain of equations for the dynamics of the Kuramoto–Daido order parameters of the genuine phase. In Section 3, a rigorous two circular cumulant truncation of the infinite equation chain is derived.

The characteristic function is approximated by (1 + z1k + z2k 2/2! p. 3
Where the paper says it

; in particular, two first cumulants are κ1 = z1 , κ2 = z2 − z 2 1 , where the first CC is identical to the Kuramoto order parameter and the second CC measures the deviation of the second circular moment z2 from the Ott–Antonsen manifold zn = (z1) n [27,28]. With a two circular cumulant reduction, the characteristic function 1 + z1k + z2 k 2 2!

+ z3k 3/3! + z4k 4/4! + · · · ≈ eκ1k+κ2 / 2 ≈ (1 + κ2k 2/2)eκ1k. p. 3
Where the paper says it

− nX−1 m=1 κmzn−m (n − m)! ; in particular, two first cumulants are κ1 = z1 , κ2 = z2 − z 2 1 , where the first CC is identical to the Kuramoto order parameter and the second CC measures the deviation of the second circular moment z2 from the Ott–Antonsen manifold zn = (z1) n [27,28].

The resulting equations for the first and second circular cumulants are given by (13) and (14). p. 3
Where the paper says it

− nX−1 m=1 κmzn−m (n − m)! ; in particular, two first cumulants are κ1 = z1 , κ2 = z2 − z 2 1 , where the first CC is identical to the Kuramoto order parameter and the second CC measures the deviation of the second circular moment z2 from the Ott–Antonsen manifold zn = (z1) n [27,28].

Validation and Applicability of Models

The accuracy of the 2CC model reduction is validated with numerical simulations for time-independent regimes. The model is accurate for time-independent states and dynamic response to a periodic modulation deep into the parameter domain where the diffusion approximation is not applicable. p. 0

Where the paper says it

For the kinetics equations of the complete mean field theory of a homogeneous inhibitory network of quadratic integrate-and-fire neurons, we introduce circular cumulants of the genuine phase variable and derive a rigorous two cumulant reduction for both timeindependent conditions and modulation of the excitatory current. The low dimensional model is examined with numerical simulations and found to be accurate for time-independent states and dynamic response to a periodic modulation deep into the parameter domain where the diffusion approximation is not applicable.

The 2CC reduction provides reasonable accuracy for the time-independent states (AS) for as low excitatory current as i0 = 0. p. 5
Where the paper says it

03 and larger currents, one can expect a decent accuracy of the 2CC reduction. In Figure 2, one can see that 2CC reduction (19)–(21) provides reasonable accuracy for the time-independent states (AS) for as low excitatory current as i0 = 0.

01g 2 0.

Comparison with Approximation Methods

The Ott–Antonsen Ansatz (OA) reduction, which sets κ2 to zero, becomes accurate only for much higher values of the excitatory current i0. The OA model reduction is fundamentally unable to reproduce the noise-induced oscillations in neural circuits within the diffusion approximation. p. 5

Where the paper says it

One can see that the OA model reduction becomes accurate only for much higher values of the excitatory current i0. Moreover, it is known to be fundamentally unable to reproduce the noise-induced oscillations in neural circuits within the diffusion approximation [3–5,33,34].

The 2CC model captures the dynamic response for i0 = 0. p. 7
Where the paper says it

(25) The relation between the firing rate and {zn} remains unchanged. In Figures 3 and 4, one can see that the 2CC model reduction decently captures the “exact” dynamic response to strong modulation I(τ) = I0[1 + 0.

p. 7
Where the paper says it

(25) The relation between the firing rate and {zn} remains unchanged. In Figures 3 and 4, one can see that the 2CC model reduction decently captures the “exact” dynamic response to strong modulation I(τ) = I0[1 + 0.

2g 2 0 and its error rapidly decreases as i0/g 2 0 increases for strong modulation. p. 9
Where the paper says it

Even for a strong modulation of synaptic current with a resonant frequency, where the system is most sensitive to the problem of the overestimation of the noise-induced self-excitation of oscillations, the 2CC model captures the dynamic response for i0 = 0. 2g 2 0 and its error rapidly decreases as i0/g2 0 increases.

p. 9
Where the paper says it

Even for a strong modulation of synaptic current with a resonant frequency, where the system is most sensitive to the problem of the overestimation of the noise-induced self-excitation of oscillations, the 2CC model captures the dynamic response for i0 = 0. 2g 2 0 and its error rapidly decreases as i0/g2 0 increases.

Theoretical Limits of Diffusion Approximation

The diffusion approximation is valid as long as the ratio a/σV is small, where σV can be estimated for time-independent macroscopic states. The quantitative measure for the approximation accuracy a/σV should be small: a/σV ≈ 2(-Ag)1/4√Kν≪ 1. p. 9

Where the paper says it

The width σV of the distribution in the local minimum of this potential can be estimated σ 2 V = D/[2p −Ag] = g 2 0ν/[4p −Ag]. The quantitative measure for the approximation accuracy a/σV should be small: a σV ≈ 2(−Ag) 1/4 √ Kν ≪ 1 .

This estimate gives an optimistic boundary of the applicability of the diffusive approximation. p. 9
Where the paper says it

5 ln(g 2 0 /i0)]1/6 (i0/g2 0 ) 1/3 √ K = r KD K . (32) This estimate gives an optimistic boundary of the applicability of the diffusive approximation, as the applicability can be also violated for time-dependent macroscopic regimes.

The approximation is accurate for K 1–2 orders of magnitude larger than KD = 4[0. p. 6
Where the paper says it

The boundary between the GO and AS regimes is the Hopf bifurcation curve plotted with the black solid line for the original complete mean field system (6) and (15) and with the red dashed line for the diffusion approximation (Section 7). The cyan line: KD(i0/g2 0), the diffusion approximation is accurate for in- degree K 1–2 orders of magnitude larger than KD (Section 7.

5 ln(g 2 0/i0)]1/3/(i0/g 2 0)2.

Conclusion on Model Utility

The two circular cumulant reduction model might be expected to be applicable for theoretical studies of self-organized global oscillations in sparse networks where both excitatory and inhibitory synaptic links are present and collective oscillations emerge due to the interplay of excitation and inhibition. An accurate agreement between the dynamics of the 2CC reduction and the complete mean-field model indicates and explains the low embedding dimensionality of attracting macroscopic regimes of the system both under constant and periodically modulated conditions. p. 9

Where the paper says it

Summarizing, the 2CC model might be expected to be applicable for theoretical studies of self-organized global oscillations in sparse networks where both excitatory and inhibitory synaptic links are present and collective oscillations emerge due to the interplay of excitation and inhibition [1,2,47–51]. An accurate agreement between the dynamics of the 2CC reduction and the complete mean-field model indicates and explains the low embedding dimensionality of attracting macroscopic regimes of the system both under constant and periodically modulated conditions.

The calculations with the derived 2CC reduction also do not require computation of coefficients Inm. p. 9
Where the paper says it

An accurate agreement between the dynamics of the 2CC reduction and the complete mean-field model indicates and explains the low embedding dimensionality of attracting macroscopic regimes of the system both under constant and periodically modulated conditions. Calculations with the derived 2CC reduction also do not require computation of coefficients Inm.

Time-Dependent Modulation Dynamics

The continuity equation (3) is modified to incorporate time-dependent modulation of the external current I(t) = I0[1 + η(t)]. The resulting dynamical system (19)–(21) is not independent of i0 as α = g0/(K3/4√i0). p. 5

Where the paper says it

, (19) with rescaled time τ = K1/4√ i0 t, one can rewrite (13)–(14) as dκ1 dτ = i2κ1 + Kνe ζ1(1 + κ1) 2 1 − ζ1κ1 + ζ1(1 + ζ1) 2 (1 − ζ1κ1) 3 κ2 , (20) dκ2 dτ = i4κ2 + Kνe ζ 2 1 (1 + κ1) 4 (1 − ζ1κ1) 2 − κ2 1 − 2(1 + ζ1) 2 (1 − ζ1κ1) 2 + 4(1 + ζ1) 3 (1 − ζ1κ1) 3 − 3(1 + ζ1) 4 (1 − ζ1κ1) 4 . (21) Noteworthy, the dynamical system (19)–(21) is not independent of i0 as α = g0/(K3/4√ i0).

In the Fourier space, the extension of equation (6) yields equation (23). p. 6
Where the paper says it

The values of i0 and K for these plots are shown with magenta lines in Figure 1a. In Fourier space, the latter equation yields an extension of (6): z˙n = inp I0 2 + η(t) zn + η(t) zn−1 + zn+1 2 + Kν(t) "X +∞ m=0 Inm zm − zn # ; (23)

The relation between the firing rate and mean membrane potential is given by πν − iv = p/I0Wψ. p. 4
Where the paper says it

(12) One can calculate ζ1 − ζ2 = 4/[α(α − 2i)], ζ2 = −2 − 1/ζ1 and finally write: κ˙ 1 = i2K 1 4 √ i0κ1 + Kν ζ1(1 + κ1) 2 1 − ζ1κ1 + ζ1(1 + ζ1) 2 (1 − ζ1κ1) 3 κ2 , (13) κ˙ 2 = i4K 1 4 √ i0κ2 + Kν ζ 2 1 (1 + κ1) 4 (1 − ζ1κ1) 2 − κ2 1 − 2(1 + ζ1) 2 (1 − ζ1κ1) 2 + 4(1 + ζ1) 3 (1 − ζ1κ1) 3 − 3(1 + ζ1) 4 (1 − ζ1κ1) 4 . (14) 4 Relation between firing rate, mean membrane potential and probability density of the genuine phase In Eq.

The 2CC model captures the dynamic response for i0 = 0.2g 2 0 and its error rapidly decreases as i0/g 2 0 increases for strong modulation.

Diffusion Approximation in Dynamic Regimes

The diffusion approximation version of the continuity equation (3) yields a modified version of (4). This leads to a modified version of equation system (23) with the truncated matrix I[DA]. p. 8

Where the paper says it

Hence, for the Kuramoto–Daido order parameters zn, one finds a modified version of equation system (23): z˙n = inp I0 2 + η(t) zn + η(t) zn−1 + zn+1 2 + Kν(t) X +∞ m=0 I [DA] nm zm , n = 1, 2, 3, . , (28) where z0 = 1; the “truncated” matrix I [DA] nm ≡ αQnm + α 2 2 (Q2 )nm , Q =   0 0 0 0 0 .

The relation between the firing rate and mean membrane potential is given by equation (29). p. 1
Where the paper says it

In Section 3, a rigorous two circular cumulant truncation of the infinite equation chain is derived. In Section 4, we explicitly show that, for the shot noise, the relation between the firing rate, the mean membrane voltage and the probability density is still given by the same conventional Montbri´o–Paz´o–Roxin order parameter [17].

The 2CC reduction for this modified system yields equations (30) and (31). p. 8
Where the paper says it

(26) Here one approximately represents the shot-noise term with the mean drift and continuous diffusion parts; technically, its Taylor expansion with respect to a is truncated after the two leading terms. For the probability density w(ψ, t) of the genuine phase ψ = 2 arctan(V /√ I0), continuity equation (26) yields a modified version of (4) : ∂w(ψ, t) ∂t = − ∂ ∂ψ np I0 2 + η(t)(1 + cosψ) w(ψ, t) o + Kν(t) a √ I0 Qw(ψ, t) + a 2 2I0 Q 2w(ψ, t) , (27) where the operator Q(.

Final Summary of Findings

The authors have derived a low dimensional neural mass model for the case where the effect on intrinsic fluctuations is represented by an effective shot noise but essentially cannot be reduced to the diffusive approximation. The model reduction is based on the circular cumulant formalism and adopts a two circular cumulant truncation (24)–(25) for an infinite chain of circular moment equations (23). p. 9

Where the paper says it

The model reduction is based on the circular cumulant formalism and, as a first step, requires the rewriting of the continuity equation in terms of the genuine phase. Further, we adopt a two circular cumulant truncation (24)–(25) for an infinite chain of circular moment equations (23).

For time-independent solutions the 2CC reduction is accurate for as low excitatory currents as i0 = 0. p. 9
Where the paper says it

For completeness, we examined both the 2CC reduction and its further downgrade to the Ott–Antonsen Ansatz [27,28] by setting κ2 = 0. For time-independent solutions the 2CC reduction is accurate for as low excitatory currents as i0 = 0.

01g 2 0. For self-excited noise-induced oscillations the 2CC reduction is less accurate since it underestimates the diffusive suppression of collective modes and gives a lowered value of the Hopf bifurcation threshold as compared to the “exact” solution. p. 9
Where the paper says it

01g 2 0 , which extends by two orders of magnitude farther than the applicability domain of the diffusion approximation (Figure 1a). For self-excited noise-induced oscillations the 2CC reduction is less accurate since it underestimates the diffusive suppression of collective modes and gives a lowered value of the Hopf bifurcation threshold as compared to the “exact” solution; for heterogeneous populations the threshold is more accurate [33,5], but this is beyond the scope of our paper.

The 2CC model might be expected to be applicable for theoretical studies of self-organized global oscillations in sparse networks where both excitatory and inhibitory synaptic links are present and collective oscillations emerge due to the interplay of excitation and inhibition. p. 9
Where the paper says it

For off-resonance frequencies, the dynamic response error is much smaller. Summarizing, the 2CC model might be expected to be applicable for theoretical studies of self-organized global oscillations in sparse networks where both excitatory and inhibitory synaptic links are present and collective oscillations emerge due to the interplay of excitation and inhibition [1,2,47–51].

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Improve dynamics modeling by incorporating a low dimensional model based on two circular cumulants to accurately describe macroscopic collective dynamics where diffusion approximation fails. This allows for theoretical studies of inhibitory-excitatory balanced neural networks that are biologically relevant in regimes not covered by conventional models.

  2. Enhance the ability of AI systems to predict population firing rates under time-dependent synaptic modulation using the derived 2CC reduction equations. The improved system can accurately capture dynamic responses for resonant and off-resonance modulation frequencies, leading to better predictions of network behavior during external driving.

  3. Develop a more robust framework for assessing the applicability limits of the diffusion approximation in neural networks by utilizing the theoretical boundary estimate a/sigmaV approximately equals rKD K where rKD is calculated based on i0/g20. This provides a quantitative measure for when conventional approximations fail, guiding researchers to use more sophisticated models.

  4. Implement an AI capable of distinguishing between macroscopic regimes like global oscillations (GOs) and asynchronous dynamics (AS) by analyzing the behavior of circular cumulants in Fourier space. This allows the system to classify network states based on the hierarchy of smallness observed in the exact solutions versus approximations like Ott-Antonsen Ansatz.

  5. Create a system that can accurately model noise-induced self-excitation of collective oscillations by comparing 2CC reductions with other models like the Ott-Antonsen Ansatz. This enables researchers to quantify the error in predicting Hopf bifurcation thresholds when dealing with shot noise, which is a known weakness of simpler approximations.

Abstract

Kinetics of a balanced network of neurons with a sparse grid of synaptic links is well representable by the stochastic dynamics of a generic neuron subject to an effective shot noise. The rate of delta-pulses of the noise is determined self-consistently from the probability density of the neuron states. Importantly, the most sophisticated (but robust) collective regimes of the network do not allow for the diffusion approximation, which is routinely adopted for a shot noise in mathematical neuroscience. These regimes can be expected to be biologically relevant. For the kinetics equations of the complete mean field theory of a homogeneous inhibitory network of quadratic integrate-and-fire neurons, we introduce circular cumulants of the genuine phase variable and derive a rigorous two cumulant reduction for both time-independent conditions and modulation of the excitatory current. The low dimensional model is examined with numerical simulations and found to be accurate for time-independent states and dynamic response to a periodic modulation deep into the parameter domain where the diffusion approximation is not applicable. The accuracy of a low dimensional model indicates and explains a low embedding dimensionality of the macroscopic collective dynamics of the network. The reduced model can be instrumental for theoretical studies of inhibitory-excitatory balanced neural networks.

Related papers