⚡ electrical engineering

Filtered Voltage Feedback in a Piecewise-Linear Electronic Neuron: First-Switch Threshold and a Saddle-Node of Spiking Cycles

This paper analyzes a planar piecewise-linear electronic neuron with filtered voltage feedback, demonstrating that its onset of persistent spiking occurs via a saddle-node bifurcation of limit cycles at a current significantly lower than the equilibrium loss threshold, resulting in a hysteretic coexistence of rest and spiking states and a finite burst of spikes below the bifurcation point.

Original authors: Seyedeh Rahemeh Amiri

Published 2026-10-30

📖 7 min read🧠 Deep dive

CC BY 4.0

Original authors: Seyedeh Rahemeh Amiri

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

In the study of how the brain thinks and reacts, scientists often build simplified models of neurons, the tiny cells that carry electrical signals. These models try to capture the essential behavior of a real neuron: sitting quietly until a signal pushes it over a limit, then firing a sharp spike of electricity, and finally resetting to wait for the next signal. For decades, researchers have known that the timing of that first spike matters. If a neuron is pushed just barely hard enough to fire, it might take a long time to react, or it might fire immediately. The shape of the electrical circuit inside the model determines whether the reaction is slow and gradual or sudden and sharp. Understanding these timing rules helps scientists predict how networks of neurons might behave, which is crucial for everything from understanding brain diseases to designing better artificial intelligence.

A researcher at the University at Albany recently took a specific, simplified electronic model of a neuron and added a single new feature to see how it changed the system's behavior. The original model was a flat, two-dimensional design that could produce spike-like electrical waves without needing to manually reset itself after firing. However, it lacked a way to "remember" its recent history. The researcher introduced a new variable that acts like a memory filter, smoothing out the voltage changes over time. This small addition transformed the system's behavior in surprising ways, creating a scenario where the neuron could sit in a quiet state or fire repeatedly, depending on how it was started, even when the electrical current pushing it was exactly the same.

The researcher began by asking how much electrical current was needed to make the neuron fire its very first spike when starting from a completely quiet state. In the original model without the memory filter, the time it took to fire would stretch out infinitely if the current was just barely enough to trigger it. But with the new filter added, the researcher found a precise, mathematical limit to this waiting time. Even when the current was set to the absolute minimum required to cross the firing threshold, the neuron fired in a finite amount of time, roughly 50 milliseconds. This means the system does not get stuck in an endless wait; it has a built-in speed limit for its first reaction. The researcher calculated this exact threshold and the time it takes to reach it, finding that the delay follows a specific curve that flattens out rather than stretching forever.

The most significant discovery, however, concerned what happens when the current is increased just a little bit more. The researcher found that the system developed a kind of double personality. In a specific range of electrical currents, the neuron could exist in two different states at the same time. If the neuron started in a quiet, resting position, it would stay quiet forever. But if it was given a tiny nudge to push it past a certain point, it would switch into a state of continuous, rhythmic firing. This phenomenon, known as bistability, means the system has a memory of its starting condition. The researcher mapped out exactly where this switch happens. They found that the transition from a quiet state to a firing state is not a smooth slide but a sudden jump, similar to a light switch that clicks on.

To understand how this firing starts and stops, the researcher looked at the cycle of the spikes. They discovered that the onset of this continuous firing is caused by a specific mathematical event where two invisible paths in the system's behavior collide and vanish. One path represents a stable rhythm of firing, and the other represents an unstable rhythm that the system cannot hold onto. When the electrical current is lowered to a critical point, these two paths merge and disappear. Below this point, the continuous firing stops completely. This explains why the system can be quiet or firing, but not something in between, once the current drops below that threshold. The researcher measured this critical point with high precision, finding it occurs at a current level that is about 71% of the maximum current the resting state can handle.

The study also revealed what happens when the current is set just below this critical point, where the continuous firing cannot sustain itself. Instead of staying quiet, the neuron fires a burst of spikes and then settles back down. The researcher found a clear rule for how long this burst lasts. The closer the current is to the critical point, the longer the burst becomes. If the current is just a tiny fraction below the limit, the neuron might fire hundreds of times before stopping. If it is further away, it might fire only a few times. The researcher confirmed that the number of spikes in these bursts follows a predictable pattern: as the distance to the critical point gets smaller, the number of spikes grows larger in a specific, inverse relationship. This behavior was tested across thousands of simulations, and the predictions matched the computer results perfectly, from bursts of just two spikes to bursts of over eight hundred.

The researcher also tested how changing the speed and strength of the new memory filter affected these results. They varied the filter's settings across a wide grid of possibilities. They found that making the filter stronger or slower widened the range of currents where the quiet and firing states could coexist. This means that by tuning the memory component, one could control how easily the neuron gets stuck in a firing loop or how hard it is to get it to start. The study confirmed that these behaviors are robust and depend directly on the specific values chosen for the filter's speed and strength.

Throughout the investigation, the researcher used two different computer methods to solve the equations governing the neuron's behavior. One method adapted its calculation speed to follow the sharp turns in the voltage, while the other used a fixed, exact mathematical formula for every step. Both methods produced identical results, giving the researcher high confidence in their findings. They also checked their work by comparing the filtered system to a simpler version without the filter, confirming that the filter was indeed the cause of the new behaviors, such as the ability to fire from a quiet start and the existence of the double states.

The work provides a complete picture of how a simple electronic neuron with a memory filter behaves. It establishes exact rules for when the first spike occurs, how long it takes to happen, and how the system transitions between silence and continuous firing. The findings show that adding a single memory component to a simple circuit can create complex behaviors like hysteresis, where the system's history matters, and precise control over burst lengths. While the results are based on computer simulations of a theoretical circuit, they offer a clear, quantitative understanding of how discontinuous switching and filtered feedback interact. This detailed map of the system's behavior helps clarify the fundamental mechanics of how neurons can switch between states, providing a solid foundation for future studies on more complex neural networks.

Technical Summary: Filtered Voltage Feedback in a Piecewise-Linear Electronic Neuron

Problem Statement
The paper investigates a planar piecewise-linear electronic neuron model characterized by six switching regions. While this original model produces spike-like excursions without an imposed reset, it lacks an internal memory variable, leading to specific dynamical behaviors such as logarithmic first-spike latency. The study addresses the extension of this model by introducing a single exponentially filtered voltage state, $w$ , which feeds back negatively onto the voltage derivative $\dot{v}$ . This modification creates a three-dimensional hybrid system. The analysis is guided by three primary questions: (1) How does the filter alter the current threshold and latency for the first switching event? (2) Can the system support the coexistence of a stable resting state and a repetitive spiking orbit at the same input current? (3) What is the quantitative law governing the termination of the spiking branch as current is lowered?

Methodology
The author employs a combination of analytical derivation and rigorous numerical verification:

Analytical Derivation: For the first-switch event (starting from rest), the system is reduced to a linear subsystem. The author derives a closed-form expression for the first-switch threshold current ( $I_{\text{first}}$ ) and the time to reach the threshold ( $t_{\text{first}}$ ). They also analyze the latency behavior near this threshold.
Numerical Integration: The hybrid system is integrated using two independent methods to ensure robustness: (a) an adaptive Runge-Kutta solver (DOP853) with event detection, and (b) an exact matrix-exponential solver that propagates flows through affine regions and locates crossings via bracketing.
Return Map Analysis: To study persistent spiking, the author constructs a Poincaré return map on the section $v = v_{\text{th}}$ . Due to the strong contraction of the recovery variable $k$ , the dynamics effectively reduce to a one-dimensional map $w_{n+1} = \Phi(w_n)$ .
Bifurcation Analysis: The onset of spiking is identified by locating the saddle-node bifurcation of limit cycles (a "fold" of cycles) where the stable and unstable fixed points of the return map merge. The normal form coefficients are measured via finite differences to characterize the scaling laws near this bifurcation.

Key Contributions and Results

First-Switch Threshold and Latency:
- The paper provides an exact closed-form formula for the current required to reach the switching surface from rest: $I_{\text{first}} = v_{\text{th}}/s_{\text{max}}$ , where $s_{\text{max}}$ is the peak of the linear step response.
- This threshold lies significantly below the equilibrium current $I_{\text{eq}}$ (where the resting state becomes inadmissible).
- Unlike the original planar model, where latency diverges logarithmically as current approaches the threshold, the filtered model exhibits a square-root cusp in latency. The first-spike time approaches a finite limit ( $t_p$ ) as $I \downarrow I_{\text{first}}$ .
Rest-Spiking Bistability and Hysteresis:
- The addition of the filter introduces a region of bistability where a stable resting state and a stable spiking limit cycle coexist.
- This coexistence occurs in the current range $I_{\text{fold}} < I < I_{\text{eq}}$ . For the main example parameters ( $g=0.05, \tau=50$ ), this window spans approximately 71% to 100% of $I_{\text{eq}}$ .
- The onset of spiking is a saddle-node bifurcation of cycles (a fold). The spiking branch exhibits hysteresis: persistent spiking can be maintained at currents lower than the fold point $I_{\text{fold}}$ if the system is already in the spiking state, but the system will return to rest if started from rest at those same currents.
Quantitative Scaling Laws:
- Onset Rate: The firing rate at the onset of spiking (at $I_{\text{fold}}$ ) is finite (approximately 12.6 Hz for the main example), contrasting with models where the period diverges.
- Burst Length: Below the fold current ( $I < I_{\text{fold}}$ ), the system emits a finite burst of spikes before returning to rest. The number of spikes $N$ follows an inverse-square-root law: $N \simeq \pi / \sqrt{|a|b(I_{\text{fold}} - I)}$ . This "ghost" bottleneck behavior was verified numerically for burst lengths ranging from $N=2$ to $N=837$ .
- Parameter Dependence: The ratio $I_{\text{fold}}/I_{\text{eq}}$ was mapped over a $5 \times 5$ grid of feedback gain ( $g$ ) and time constant ( $\tau$ ). The ratio decreases monotonically with increasing $g$ and $\tau$ , indicating that stronger and slower feedback widens the bistability window.
Model Validation:
- The results were cross-verified using the independent matrix-exponential solver and a matched instantaneous-feedback control system, confirming that the observed overshoot and bistability are due to the filtered memory, not the instantaneous feedback.

Significance and Claims
The author explicitly states that they do not claim the phenomena of bistability, resonance, or saddle-node scaling laws are new; these are well-documented in conductance-based and other piecewise-linear models.

The specific contribution of this work is the quantitative realization of these phenomena in a discontinuous, six-region electronic circuit with a filtered feedback state. The paper provides:

An exact first-switch threshold and latency law for this specific discontinuous system.
The identification of the spiking onset as a fold of cycles within the full discontinuous system.
A measured normal form that quantitatively predicts burst length based on return-map iterations.
A detailed characterization of how the fold current depends on the filter parameters ( $g, \tau$ ).

The study remains modest regarding its scope, noting that the fold location is numerical (no closed form for $I_{\text{fold}}$ exists due to transcendental crossing times), the reduction to one dimension is numerically verified but not analytically proven, and the results are specific to this theoretical circuit model without hardware validation.

More like this