betaFeedback

WorksEngineering

Excitability of High-Frequency Positive Feedback

Abstract

What is the threshold of an excitable system, i.e. the criterion that distinguishes rest from excitation? Previous research has shown that this question is far from trivial to answer. This paper studies the threshold of a simple excitable primitive in which excitability arises from the positive feedback interconnection of a first-order high-pass filter and a static, strictly monotone and strictly convex activation function. This threshold is a simple inequality on the difference between the instantaneous input and the motif's internal memory. While this result may help neuro-engineers characterize spiking in biological and artificial neurons, we argue that excitability is not always a desirable property, and that our threshold may also be relevant to safety considerations in fragile systems with rate-induced tipping points. We illustrate our framework with three examples: the FitzHugh-Nagumo model of neuronal dynamics, a model of stop-loss cascades in the stock market, and a social contagion model.

Introduction

The robust fragility of excitable systems

Excitability is a property by which a system responds in an all-or-none fashion to the transient of an input signal. In neurophysiology, it is the well-studied spiking mechanism of neurons, where a large, stereotyped action potential is produced in response to a sufficiently large and/or sharp input current [1] (Figure 1, Left; Figure 2).

First intracellular recording of an action potential and flash crash of the Dow Jones

Figure 1. Left: First intracellular recording of an action potential captured in the squid giant axon [20]. Right: Flash crash of the Dow Jones on May 6th, 2010 [21].

Figure 2. Inspired by [1]. Simulation of a FitzHugh-Nagumo neuron model. The spiking threshold is neither static, neither easily captured by simple energetic arguments.

From both the neurophysiological and neuromorphic engineering standpoint, excitability is a desirable property. Excitability makes the spike timing ultra-sensitive to the transient of the input signal rather than to the neuron’s parametric uncertainties, so the production of action potentials is reliable under neuronal heterogeneity and/or neuromorphic hardware mismatch [2], [3], [4] (Figure 3). As such, excitability enables the robust encoding of information in the spike timing, rather than merely in the spiking frequency [5]. This mechanism bridges the continuous, analog world with the discrete, digital, event-based one: smooth, continuous-time analog signals are robustly reshaped into a discrete set of sparse timed events [6].

Responses to constant and fluctuating inputs

Figure 3. Reproduced from [2]. Left: A constant input yields unreliable spike timing across trials. Right: A fluctuating input yields reliable spike timing across trials.

Although excitability is desirable in neural systems, that may be the exception rather than the rule. Indeed, the excitability property is also a defining feature of fragile systems, that may lead to catastrophic failures. A blackout propagates across a continent in minutes; a systemic financial collapse breaks the world economy through cascading bank failures; the weakening of the Atlantic Meridional Overturning Circulation (AMOC) is self-amplified through global warming. All

This work was supported by the Belgian Government through the Federal Public Service Policy and Support.

of these examples involve a fast self-reinforcing loop whose activation depends on the rate of change of a signal [7], [8]. The same property that makes neurons spike timing reliable could be making other systems fail reliably. In the tipping-point literature, a system’s ultra-sensitivity to a rate of change is also known as rate-induced tipping [9]. Excitable systems are a natural class in which rate-induced tipping occurs, and in this paper, we call such systems excitably fragile.

A surprisingly large range of systems may qualify as excitably fragile. In the Earth system, we find evidence of rate-induced tipping in the carbon emissions from permafrost soils [10]. In mechanical systems, stick-slip friction, which is observed in machine-tool chatter or in earthquakes, produces slip events with a spike-like signature [11]. In power systems, cascading blackouts have been attributed to rate-induced tipping [12], [7]. Related phenomena may arise in electronics, where CMOS latch-up and thermal runaway in power devices are already plausible candidates [13], [14]. Financial markets provide other speculative examples, with flash crashes and liquidity spirals exhibiting similar behaviors [8], [15], [16] (Figure 1, Right).

Whether viewed from a neurophysiological, neuroengineering, tipping-point, or bifurcation-control perspective, the capacity of an excitable systems to blow up in response to a peculiar transient would benefit from a well-defined notion of threshold. [7], [9], [17], [18], [19] (Figure 2). This work addresses this question for a newly introduced motif, where excitability is understood as a high-frequency positive feedback loop.

A dynamical threshold for high-pass excitability

An elementary motif of neural excitability has recently been proposed in neuroengineering [22]. Succinctly, it consists of the positive feedback interconnection of a high-pass filter and a high-gain sigmoid (Figure 4, Top). The motif can reproduce both Hodgkin’s Type I and Type II excitability [23], depending on the shape of the sigmoid. Furthermore, it offers a frequency-based point of view on the mechanism: the motif produces a spike through a positive feedback amplification of the high-frequencies components of its input signal.

Figure 4. Top: Block diagram of the excitable motif... Bottom: ...and its stable inverse.

The analysis carried out in this paper is strictly focused on obtaining an exact spike-initiation threshold. For this reason, this work further simplifies this recently introduced motif. We consider a strictly monotone and convex activation function rather than a sigmoid, so the local spike-producing bifurcation remains, but no inflection point produces a limit cycle terminating the spike. This simplification is akin to that used in hybrid neuron models, where the spike termination is replaced by an artificial reset rule. Examples include the quadratic/exponential integrate-and-fire models and their adaptive extensions, or the Izhikevich model [23]. Here, however, the simplification serves a theoretical rather than a computational purpose. It allows us to focus the analysis on the excitability threshold, which we can then define as the point at which the inverse motif ceases to be well-defined (Figure 4, Bottom).

Characterizing the threshold of an excitable system is generally considered to be a difficult task [23]. In neurophysiology, spike generation is associated with a bifurcation in the state space of the neuron model at a static input level: the rheobase. Although the rheobase is the static threshold of spike generation, it is well documented that spikes can be triggered by inputs that remain at all times below this static threshold, and a dynamical threshold must be characterized in order to anticipate the production of a spike [23], [17] (Figure 2). While this dynamical threshold is not easily captured by simple energetic arguments, a recent work in neuromorphic control successfully introduced an history-dependent energy-based threshold, by computing the minimal external supply that is required to trigger a spike [18]. This work is more general than ours, and it requires solving an optimal control problem. By contrast, thanks the simplicity of the motif considered here, our threshold is much simpler. It consists of a simple inequality on the difference between the instantaneous input and the motif’s internal memory. This threshold naturally accounts for both excitatory and post-inhibitory spikes.

Beyond neuroscience and neuroengineering, if the motif of excitably fragile systems is akin to the one introduced in this paper, then this bound may constitute a generic tipping-point result. Because the blocks involved are extremely elementary, we argue that our proposed motif is very likely to be found in rate-induced tipping point systems. Positive feedback high-pass filtering emerges naturally in systems that are excited by their own rate of change. We describe two examples of such excitably fragile systems at the end of the paper: stop-loss cascades in stock market dynamics, which have notably been implicated in amplifying the 2010 flash crash [21] (Figure 1, Right), and social contagion, when a behaviour spreads in response to a recent increase in participation.

The contributions of this paper are the following.

  1. In Section II, we introduce the minimal high-frequency positive feedback excitability motif, as well as its stable inverse.

  2. In Section III, we derive the explicit dynamical threshold above which the motif blows up, and below which the motif remains stable.

  3. In Section IV, we provide three examples fitting the proposed excitable motif: the FitzHugh-Nagumo model of a neuron, a stop-loss cascade model in the stock market, and a social contagion model.

Finally, Section V concludes the paper and offers some perspectives.

The excitable motif and its stable inverse

The high-pass excitable motif

The excitable motif studied in this paper can be described as the positive feedback interconnection between a static activation function α\alpha and a high pass filter H(s)H(s) (Figure 4, Top). For slow-varying inputs, the system is predominantly feedforward, as the high-pass filter in the feedback loop suppresses constant signals. In contrast, for fast-varying inputs, the positive feedback loop becomes active and may lead to a fast blow-up.

Assumption 1 (Activation function). The activation function α∈C2(R,R>0)\alpha\in C^2(\mathbb{R},\mathbb{R}_{>0}) is taken strictly increasing and strictly

convex. Moreover, it is assumed that:

lim⁡x→−∞α(x)=0,lim⁡x→+∞α(x)=+∞,(1)\lim_{x\to-\infty}\alpha(x)=0,\qquad\lim_{x\to+\infty}\alpha(x)=+\infty, \tag*{(1)}
lim⁡x→−∞α′(x)=0,lim⁡x→+∞α′(x)>1.(2)\lim_{x\to-\infty}\alpha'(x)=0,\qquad\lim_{x\to+\infty}\alpha'(x)>1. \tag*{(2)}

A direct consequence of Assumption 1 is that the inverse maps α−1∈C1(R>0,R)\alpha^{-1}\in C^{1}(\mathbb{R}_{>0},\mathbb{R}) and (α′)−1∈C0((0,1],R)(\alpha')^{-1}\in C^{0}((0,1],\mathbb{R}) are well-defined and increasing.

Assumption 2 (High-pass filter). The high-pass filter is first order:

H(s)≔τsτs+1=1−1τs+1.(3)H(s)\coloneqq\frac{\tau s}{\tau s+1}=1-\frac{1}{\tau s+1}. \tag*{(3)}

Its internal state is denoted xx, with x(0)=α(u(0))x(0)=\alpha(u(0)).

We denote uu the input signal, yy the output signal, and ww the feedback from the high-pass filter. The full excitable motif reads:

y=α(u+w),w=H(s)y,(4)y=\alpha(u+w),\qquad w=H(s)y, \tag*{(4)}

with state-space representation:

y=α(u+y−x),(5)y=\alpha(u+y-x), \tag*{(5)}
τx˙=−x+y,(6)\tau\dot{x}=-x+y, \tag*{(6)}

where x≔y−w=1τs+1yx\coloneqq y-w=\frac{1}{\tau s+1}y is the internal state of the high-pass filter. This system is implicitly defined and possibly multivalued, hence it is acausal and admits no admissible physical realization in general. However, it may be seen as the singular limit of a causal system, where (4a) is replaced by

ϵy˙=−y+α(u+y−x),(7)\epsilon\dot{y}=-y+\alpha(u+y-x), \tag*{(7)}

with y(0)=x(0)=α(u(0))y(0)=x(0)=\alpha(u(0)) and ϵ→0+\epsilon\to0^{+}. Whenever α\alpha is a sigmoid, this is a rewriting of the FitzHugh-Nagumo neuron model (see Section IV-A). If we introduce the variable z≔u+w=α−1(y)z\coloneqq u+w=\alpha^{-1}(y), (3) can also be re-written as

τ(1−α′(z))z˙+z=τu˙+u,(8)\tau(1-\alpha'(z))\dot{z}+z=\tau\dot{u}+u, \tag*{(8)}

where it becomes intuitive that the system remains stable as long as the left-hand side is Hurwitz, i.e. as long as α′(z)<1\alpha'(z)<1.

The inverse motif

In their seminal paper introducing the electrical dynamics of neurons, Hodgkin and Huxley identified their state-space model through a series of voltage-clamp experiments, effectively computing the inverse model of a squid giant axon membrane [20]. Indeed, while the direct model of a neuron (current-to-voltage) is excitable, and is therefore unstable, by contrast, its inverse model (voltage-to-current) is typically feedforward and stable, which greatly facilitates its identification and analysis [24, 25]. The inverse model of a neuron generally contains both a fast positive branch and a slow negative branch. This observation has motivated the mixed-feedback principle, which suggests that excitability always results from the combination of a fast positive feedback and a slow negative feedback [1].

Inspired by this perspective, we notice that our motif also possesses a stable feedforward inverse (Figure 4, Bottom). Its state-space model reads:

u=β(y)+x,(9)u=\beta(y)+x, \tag*{(9)}
τx˙=−x+y,(10)\tau\dot{x}=-x+y, \tag*{(10)}

where β≔α−1−Id⁡\beta\coloneqq\alpha^{-1}-\operatorname{Id}.

While our proposed motif of excitability follows the mixed-feedback principle, since the high-pass filter may be understood as a direct (hence, fast) feedthrough minus a slow low-pass filter, we instead decompose our inverse model into a static path and a dynamical one. The dynamical path (7b) accounts for the internal memory of the system, and the static path consists of a concave activation function β\beta. This β\beta bloc is the inverse of the instantaneous positive feedback loop of high-pass motif (Figure 6). The unique maximum of β\beta is obtained at α′(z)=1\alpha'(z)=1, as proved hereafter.

Figure 6. Collapsing the positive feedback loop of the high-pass motif yields the hysteresis-and-adaptation motif of the FitzHugh-Nagumo neuron model [26]. After a causal regularization, the hysteretic loop reads εy˙=−β(y)+u−x\varepsilon\dot{y} = -\beta(y) + u - x.

Lemma 1 (Maximum of β\beta). The map β∈C1(R>0,R)\beta\in C^{1}(\mathbb{R}_{>0},\mathbb{R}) is strictly concave and attains a maximum at the unique point where α′(α−1(y))=1\alpha'(\alpha^{-1}(y))=1.

arg⁡max⁡β=α((α′)−1(1)),(11)\arg\max\beta=\alpha((\alpha')^{-1}(1)), \tag*{(11)}
max⁡β=(α′)−1(1)−α((α′)−1(1)).(12)\max\beta=(\alpha')^{-1}(1)-\alpha((\alpha')^{-1}(1)). \tag*{(12)}

Proof. Since α\alpha is a strictly convex and strictly increasing bijection from R\mathbb{R} onto R>0\mathbb{R}_{>0}, by the inverse function theorem, β\beta is C1C^{1} on R>0\mathbb{R}_{>0} with β′(y)=1/α′(α−1(y))−1\beta'(y)=1/\alpha'(\alpha^{-1}(y))-1. As α′\alpha' is strictly increasing and α−1\alpha^{-1} is increasing, β′\beta' is strictly decreasing, so β\beta is strictly concave. Finally, since α′\alpha' increases continuously from 00 to a limit greater than 11, β′\beta' has a unique zero at α′(α−1(y))=1\alpha'(\alpha^{-1}(y))=1, which maximizes β\beta.

This inverse model provides a new lens through which to investigate excitability. In the singular limit under consideration, the system explodes whenever its inverse becomes ill-defined. If the measured values of uu and xx at a given time are such that no value of yy can satisfy (7a), i.e. if u−x>max⁡βu-x>\max\beta, then the system blows up. This idea is made rigorous in the next section.

Thresholding the motif

Building on the intuition gained from the inverse motif, we now demonstrate the exact dynamical threshold of our primitive: u−x=max⁡βu - x = \max\beta. Above it, the system blows up; below it, the system remains stable.

Theorem 1 (Internal state threshold). Assume u(0)≤(α′)−1(1)u(0) \le(\alpha')^{-1}(1) and let T≔inf⁡{t≥0:u(t)−x(t)>max⁡β}T \coloneqq\inf\{t \ge0 : u(t) - x(t) > \max\beta\}. If T=+∞T = +\infty, then the excitable motif never blows up. Otherwise it blows up at time TT.

Proof. Let f(y)≔−y+α(u+y−x)f(y) \coloneqq-y + \alpha(u + y - x) for y>0y > 0, so εy˙=f(y)\varepsilon\dot{y} = f(y). Since f′(y)=α′(u+y−x)−1f'(y) = \alpha'(u + y - x) - 1 is strictly increasing and changes sign, we easily find:

arg min⁡f=(α′)−1(1)−(u−x),(13)\operatorname*{arg\,min} f = (\alpha')^{-1}(1) - (u - x), \tag*{(13)}
min⁡f=(u−x)−max⁡β.(14)\min f = (u - x) - \max\beta. \tag*{(14)}

If T<+∞T < +\infty, then at time TT, we have min⁡f>0\min f > 0, so εy˙≥min⁡f>0\varepsilon\dot{y} \ge\min f > 0 and lim⁡ε→0+y˙(T)=+∞\lim_{\varepsilon\to0^+} \dot{y}(T) = +\infty, so the motif blows up.

If T=+∞T = +\infty, we have u−x≤max⁡βu - x \le\max\beta, which is equivalent to f(α((α′)−1(1)))≤0f(\alpha((\alpha')^{-1}(1))) \le0. Since 0≤y(0)≤α((α′)−1(1))0 \le y(0) \le\alpha((\alpha')^{-1}(1)) and yy is continuous, {y≤α((α′)−1(1))}\{y \le\alpha((\alpha')^{-1}(1))\} is positively invariant, so yy remains in [0,α((α′)−1(1))][0,\alpha((\alpha')^{-1}(1))]. Then (4b) gives x˙≥0\dot{x} \ge0 at x=0x = 0 and x˙≤0\dot{x} \le0 at x=α((α′)−1(1))x = \alpha((\alpha')^{-1}(1)), so xx remains in [0,α((α′)−1(1))][0,\alpha((\alpha')^{-1}(1))]. Both xx and yy stay bounded, so the motif never blows up. □

Remark 1. In order to achieve the condition u−x>max⁡βu - x > \max\beta, the variable z≔u+wz \coloneqq u + w must reach the threshold α′(z)=1\alpha'(z) = 1 from below. Since z≔u+wz \coloneqq u + w, the threshold α′(z)=1\alpha'(z) = 1 is expressed in terms of the input uu and of the feedback ww, that is, the output of the high-pass filter rather than its internal state xx.

At rest, the state variables sit at x≡y≡α(u)x \equiv y \equiv\alpha(u) for a constant input uu. The dynamic threshold then reduces to a static one, which is the rheobase of the motif. In the proposed motif, the rheobase exactly corresponds to the point at which the linearized loop gain becomes greater than one, where the loop becomes unstable.

Examples of the high-pass motif

FitzHugh-Nagumo neuron model

In this section, α\alpha is a sigmoid function with sufficiently high gain, such that its maximal slope exceeds one, and the motif (3) is excitable. The static positive feedback loop around α\alpha, which we have interpreted so far as the direct feedthrough of the high-pass filter, can then be interpreted as an hysteretic loop, as it collapses with the sigmoid into a bistable element (Figure 6). This bistable element may be interpreted as the inverse of the static path β\beta from the stable inverse motif introduced earlier (Figure 4, Bottom). The remaining low-pass part of the high-pass filter then takes the role of a slow adaptation variable. Overall, the motif may also be interpreted as a hysteresis-and-adaptation

Corollary 1. Rheobase

Corollary 1 (Rheobase). Assume the input uu is constant, then the motif is unstable, if and only if u>(α′)−1(1)u > (\alpha')^{-1}(1). It will blow up if infinitesimally disturbed.

Proof. Linearizing (5), (4b), and taking ε→0+\varepsilon\to0^+, we easily find that the equilibrium x≡y≡α(u)x \equiv y \equiv\alpha(u) is unstable if and only if α′(u)>1\alpha'(u) > 1, i.e. u>(α′)−1(1)u > (\alpha')^{-1}(1). A disturbance that pushes yy upwards blows up immediately. A disturbance that pushes yy downwards sends yy to the stable branch. There, y<arg⁡max⁡βy < \arg\max\beta, so as long as u−x≤max⁡βu - x \le\max\beta, i.e. x≥arg⁡max⁡β+u−(α′)−1(1)x \ge\arg\max\beta+ u - (\alpha')^{-1}(1), equation (4b) gives τx˙=−x+y<−x+arg⁡max⁡β≤−(u−(α′)−1(1))\tau\dot{x} = -x + y < -x + \arg\max\beta\le-(u - (\alpha')^{-1}(1)), so xx decreases at a rate of at least (u−(α′)−1(1))/τ(u - (\alpha')^{-1}(1))/\tau until u−x>max⁡βu - x > \max\beta, and the motif blows up by Theorem 1. □

Figure 5 illustrates both the dynamic threshold of Theorem 1 and the rheobase of Corollary 1. It can be interpreted as follows. For a slow-varying input uu, the state xx remains close to the equilibrium curve x=α(u)x = \alpha(u), which stays below the dynamic threshold. Consequently, for a sufficiently slow input uu, the motif does not explode until the input reaches the rheobase. In contrast, for a fast-varying input uu, the state xx lags behind α(u)\alpha(u), so a rapid increase in uu allows the dynamic threshold to be reached, leading to a blow-up of the motif below its rheobase. Whenever uu is a step signal, it changes instantaneously, before the internal state xx has time to respond, and it is straightforward to determine graphically whether the step will cross the threshold or not. interconnection, which is the structure of a FitzHugh-Nagumo neuron model [26].

Figure 5. Dynamic threshold udyn∗=max⁡β+xu^*_{\mathrm{dyn}} = \max\beta+ x (Theorem 1) and static threshold ustat∗=(α′)−1(1)u^*_{\mathrm{stat}} = (\alpha')^{-1}(1) (Corollary 1) as functions of xx. For udyn∗<u≤ustat∗u^*_{\mathrm{dyn}} < u \le u^*_{\mathrm{stat}}, the spike produced is isolated: this is the excitable regime. For u>ustat∗u > u^*_{\mathrm{stat}}, the system is past the bifurcation: this is the tonic spiking regime. The curve x=α(u)x = \alpha(u) represents the equilibria in the stable regime.

Up to a positive time reparameterization of the fast dynamics, the causally regularized motif with its hysteretic loop collapsed reads as the following FitzHugh-Nagumo model:

εy˙=−β(y)−x+u,(15)\varepsilon\dot{y}=-\beta(y)-x+u, \tag*{(15)}
τx˙=−x+y.(16)\tau\dot{x}=-x+y. \tag*{(16)}

Given the usual FitzHugh-Nagumo nonlinearity β(y)=y3/3−y\beta(y)=y^3/3-y, our motif yields α−1(y)=β(y)+y=y3/3\alpha^{-1}(y)=\beta(y)+y=y^3/3, so α(z)=(3z)1/3\alpha(z)=(3z)^{1/3}. The two bifurcations of this FitzHugh-Nagumo model are located at y=±1y=\pm1. The rheobase is obtained by Corollary 1 at u=α−1(1)=−1/3u=\alpha^{-1}(1)=-1/3. By Theorem 1, the spike is triggered from the lower branch exactly when u−x>max⁡y≤0β=β(−1)=2/3u-x>\max_{y\leq0}\beta=\beta(-1)=2/3. This threshold exactness is verified numerically in Figure 7. The input uu stays below the rheobase at all times, yet the dynamic threshold is crossed twice: once by an excitatory pulse, once on release of an inhibitory pulse. This results in two output spikes.

Figure 7. Simulation of the FitzHugh-Nagumo neuron (10) with τ=0.1\tau=0.1 and ε=0.001\varepsilon=0.001. From top to bottom: time evolution of the input uu, the internal state xx, the thresholded quantity u−xu-x, and the output yy.

A stop-loss cascade model

As mentioned in the introduction, the excitability property may not be specific to neurons, and it also may not always be desirable.

In the stock market, a stop-loss order refers to an automatic sell order of an asset triggered when the price of this asset falls below a predetermined level, typically defined as a percentage of a reference price. A stop-loss cascade is a phenomenon in which a sudden decline in the price of an asset triggers many stop-loss orders, initiating a self-reinforcing selling loop [27]. We show in this section that stop-loss cascades may be interpreted as an instance of our motif.

Since stop-loss orders are typically specified as percentages of a reference price and the price impact of sell orders is multiplicative, we work with logarithmic prices rather than prices. Let pp be the log-price of an asset and pˉ\bar{p} its reference level. For simplicity, pˉ\bar{p} is modeled as an exponentially weighted moving average of memory τ\tau.

τpˉ˙=−pˉ+p.(17)\tau\dot{\bar{p}}=-\bar{p}+p. \tag*{(17)}

In trading, the drawdown d≜pˉ−pd\triangleq\bar{p}-p of an asset stands for the loss of price of this asset with respect to a reference price pˉ\bar{p}. Given the dynamics of the reference price pˉ\bar{p}, we easily see that

d=−H(s)p,(18)d=-H(s)p, \tag*{(18)}

so the drawdown of an asset can be roughly interpreted as its high-pass-filtered log-price. A slow decline in pp is tracked by the moving average reference pˉ\bar{p} and produces no drawdown, whereas a fast decline in pp will be registered by dd.

The drawdown signal dd is tracked by stop-loss orders, which liquidate a position once its loss exceeds a preset tolerance. We denote by Φ(d)\Phi(d) the volume of selling orders that are triggered for a drawdown signal of value dd or below. Since Φ\Phi is a cumulative distribution of tolerances across all stop-loss orders in the market, it is increasing. A selling order depresses the log-price pp of an asset by λΦ(d)\lambda\Phi(d) below an exogenous log-price p0p_0, where p0p_0 represents the log-price the asset would follow in the absence of forced selling:

p0−p=λΦ(d).(19)p_0-p=\lambda\Phi(d). \tag*{(19)}

Combining (17), (18) and (19), we exactly recover the high-pass motif of excitability (3). The output y≜p0−p=λΦ(d)y\triangleq p_0-p=\lambda\Phi(d) is the log-price decline caused by the cascade, the input u≜−H(s)p0u\triangleq-H(s)p_0 is the part of the exogenous decline that is too abrupt to be absorbed by the moving-average reference, the feedback w≜H(s)(p0−p)w\triangleq H(s)(p_0-p) is the observed drawdown minus the part attributable to the exogenous decline, the variable z≜dz\triangleq d is simply the drawdown signal, and the activation function is given by α=λΦ\alpha=\lambda\Phi. The regularization time constant ε\varepsilon corresponds to the execution time of sell orders, which is typically negligible compared to the moving-average time constant τ\tau.

The system becomes excitably fragile whenever α′(z)=λΦ′(z)>1\alpha'(z)=\lambda\Phi'(z)>1, which depends on both the distribution of volume of selling orders Φ\Phi, and the multiplicative price impact of a sell order λ\lambda. Under conditions of excitable fragility, a stop-loss cascade may be triggered by a fast shock, for example by the rapid spread of false information. The sudden price collapse is then driven primarily by the excitability property rather than by any fundamental change in the value of the asset [15].

A social contagion model

The same excitability structure could be governing the spread of a behavior in a population, for example: a rumor, a protest, a fashion trend, or panic buying. Let pp denote the fraction of the population currently engaged in a behaviour, and let pˉ\bar{p} denote the fraction of the population that people are accustomed to seeing participate in it. We model pˉ\bar{p} as a moving average of memory τ\tau, so that (17) holds. Because of social habituation, individuals may respond more strongly to a recent increase in participation in a behavior than to its absolute level. The relevant signal for social contagion could therefore be the perceived salience of a behaviour, which is, once again, captured by a high-pass filtering mechanism: w≔p−pˉ=H(s)pw \coloneqq p-\bar{p}=H(s)p. The perceived salience ww is large under a sudden increase of pp, but vanishes under steady participation.

Following [28], we consider that each individual starts to participate in the behavior once the perceived salience exceeds a personal tolerance, so that p=Φ(z)p=\Phi(z), where Φ\Phi is the cumulative distribution of tolerances and z=u+wz=u+w is the total salience, uu being its exogenous part, e.g. a surge in media coverage. Setting y≔py\coloneqq p and α≔Φ\alpha\coloneqq\Phi recovers once again the high-pass motif (4). Here, the regularization time constant ε\varepsilon corresponds to the time taken to join the behavior, which we again consider negligible against τ\tau. The excitability condition α′(z)=Φ′(z)>1\alpha'(z)=\Phi'(z)>1 states that one additional unit of perceived salience recruits enough participants to generate more than one unit in return.

Conclusions and perspective

We introduced a minimal motif for excitability based on a high-frequency positive feedback loop. Using the stable inverse of this motif, we derived an explicit dynamical threshold separating signals that trigger excitation and cause the motif to blow up from those that maintain stability. This threshold is a simple inequality on the difference between the instantaneous input and the motif’s internal memory. We illustrated the relevance of this framework with three examples: the FitzHugh-Nagumo model of neuronal dynamics, where the threshold provides a direct characterization of excitability, and a stop-loss cascade model in the stock market, as well as a social contagion model, showing that similar excitability can arise in rate-induced tipping systems.

On the conceptual side, the current work remains to be generalized to a setting in which the high-frequency positive feedback is implemented by a linear, nnth-order, high-/bandpass filter. Under this setting, a key objective would be to develop a unified framework for excitability that connects its classical geometric interpretation in terms of bifurcations [23], its recently introduced energy-based interpretation [18], and a yet-to-be-established frequency-domain perspective.

On the practical side, our proposed threshold remains to be applied to real-world systems spanning neurobiology and neuromorphic engineering, as well as the critical systems studied in the tipping-point literature. For the latter specifically, our motif remains to be identified within existing models to establish its relevance beyond the stop-loss cascade model and the social contagion model considered in this paper.

References

  1. [1]R. Sepulchre, G. Drion, and A. Franci, “Excitable behaviors,” in Emerging Applications of Control and Systems Theory, p. 269–280, Cham, Switzerland: Springer International Publishing, 2018.arxiv.org/abs/1704.04989
  2. [2]Z. F. Mainen and T. J. Sejnowski, “Reliability of spike timing in neocortical neurons,” Science, vol. 268, no. 5216, p. 1503–1506, 1995.
  3. [3]R. Sepulchre, A. Cecconi, M. Bin, and L. Marconi, “Regulation without calibration: From trajectory to event regulation,” IEEE Control Systems, vol. 46, p. 55–69, Feb. 2026.
  4. [4]L. Mendolia, C. Wen, E. Chicca, G. Indiveri, R. Sepulchre, J.-M. Redouté, and A. Franci, “A neuromodulable current-mode silicon neuron for robust and adaptive neuromorphic systems,” 2025.
  5. [5]S. Dènève and C. K. Machens, “Efficient codes and balanced networks,” Nature Neuroscience, vol. 19, p. 375–382, Feb. 2016.DOI
  6. [6]R. Sepulchre, “Spiking control systems,” Proceedings of the IEEE, vol. 110, p. 577–589, May 2022.
  7. [7]P. D. L. Ritchie, H. Alkhayoun, P. M. Cox, and S. Wieczorek, “Rate-induced tipping in natural and human systems,” Earth System Dynamics, vol. 14, no. 3, p. 669–683, 2023.
  8. [8]M. K. Brunnermeier and L. H. Pedersen, “Market liquidity and funding liquidity,” Review of Financial Studies, vol. 22, p. 2201–2238, Nov. 2008.
  9. [9]P. Ashwin, S. Wieczorek, R. Vitolo, and P. Cox, “Tipping points in open systems: bifurcation, noise-induced and rate-dependent examples in the climate system,” Philosophical Transactions of the Royal Society A: Mathematical, Physical and Engineering Sciences, vol. 370, p. 1166–1184, Mar. 2012.DOI
  10. [10]E. O’Sullivan, K. Mulchrone, and S. Wieczorek, “Rate-induced tipping to metastable zombie fires,” Proceedings of the Royal Society A: Mathematical, Physical and Engineering Sciences, vol. 479, no. 2275, 2023.DOI
  11. [11]A. D. Berman, W. A. Ducker, and J. N. Israelachvili, “Origin and characterization of different stick-slip friction mechanisms,” Langmuir, vol. 12, p. 4559–4563, Jan. 1996.DOI
  12. [12]D. P. Nedic, I. Dobson, D. S. Kirschen, B. A. Carreras, and V. E. Lynch, “Criticality in a cascading failure blackout model,” International Journal of Electrical Power & Energy Systems, vol. 28, p. 627–633, Nov. 2006.DOI
  13. [13]R. Troutman and H. Zappe, “A transient analysis of latchup in bulk CMOS,” IEEE Transactions on Electron Devices, vol. 30, p. 170–179, Feb. 1983.DOI
  14. [14]P. Spirito, G. Breglio, V. d’Alessandro, and N. Rinaldi, “Analytical model for thermal instability of low voltage power MOS and SOA in pulse operation,” in Proceedings of the 14th International Symposium on Power Semiconductor Devices and ICs, ISPSD-02, p. 269–272, IEEE, 2002.
  15. [15]A. Kirilenko, A. S. Kyle, M. Samadi, and T. Tuzun, “The flash crash: High-frequency trading in an electronic market,” The Journal of Finance, vol. 72, no. 3, p. 967–998, 2017.
  16. [16]A. Fosset, J.-P. Bouchaud, and M. Benzanquen, “Endogenous liquidity crises,” J. Stat. Mech., vol. 2020, p. 063401, June 2020.
  17. [17]J. Mitry, M. McCarthy, N. Kopell, and M. Wechselberger, “Excitable neurons, firing threshold manifolds and canards,” The Journal of Mathematical Neuroscience, vol. 3, no. 1, p. 12, 2013.DOI
  18. [18]R. Sepulchre and G. Tong, “On the threshold of excitable systems: An energy-based perspective,” in 2025 IEEE 64th Conference on Decision and Control (CDC), p. 980–985, IEEE, Dec. 2025.arxiv.org/abs/2504.02171
  19. [19]Bifurcation Control. Springer Berlin Heidelberg, 2003.DOI
  20. [20]A. L. Hodgkin and A. F. Huxley, “A quantitative description of membrane current and its application to conduction and excitation in nerve,” The Journal of Physiology, vol. 117, p. 500–544, Aug. 1952.
  21. [21]U.S. Securities and Exchange Commission and U.S. Commodity Futures Trading Commission, “Preliminary findings regarding the market events of may 6, 2010,” joint report, U.S. Securities and Exchange Commission and U.S. Commodity Futures Trading Commission, May 2010.
  22. [22]G. Bainier, C. Fernandez Lorden, A. Fyon, C. Adriaans, L. Mendolia, G. Drion, P. Sacré, and A. Franci, “A two-block model of neuronal excitability.” Poster presented at Neuromorphic Computing Netherlands 2026, Maastricht, The Netherlands, Oct. 2026.DOI
  23. [23]E. M. Izhikevich, Dynamical Systems in Neuroscience: The Geometry of Excitability and Bursting. The MIT Press, July 2006.DOI
  24. [24]T. Burghi, Feedback for neuronal system identification. PhD thesis, 2019.DOI
  25. [25]T. B. Burghi, M. Schoukens, and R. Sepulchre, “Feedback identification of conductance-based models,” Automatica, vol. 123, p. 109297, Jan. 2021.DOI
  26. [26]P. Sacré, Systems Analysis of Oscillator Models in the Space of Phase Response Curves. PhD thesis, Universite de Liege (Belgium), 2013.
  27. [27]C. L. Osler, “Stop-loss orders and price cascades in currency markets,” Journal of International Money and Finance, vol. 24, p. 219–241, Mar. 2005.
  28. [28]M. Granovetter, “Threshold models of collective behavior,” American Journal of Sociology, vol. 83, p. 1420–1443, May 1978.DOI

Paper details

vlnc:W_zm63zuxr
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
  • Engineering

Not recorded for this paper

View statement
Looking for related papers…
Contents