Leaky Integrate-and-Fire Dynamics

• Leaky integrate-and-fire dynamics are a simplified model that describes the evolution of neuronal membrane potential via leakage, input integration, and threshold-triggered spiking.
• They support diverse applications including spiking neural networks, neuromorphic hardware, and analysis of chimera states, noise-driven variability, and chaotic dynamics.
• Extensions encompass coupled network analyses, stochastic modeling, and analog implementations, providing robust tools for computational neuroscience and efficient hardware design.

Leaky integrate-and-fire (LIF) dynamics describe the time evolution of a state variable—typically the membrane potential of a neuronal unit—under the combined action of passive leakage, active integration of inputs, and threshold-driven firing with reset. The canonical LIF neuron equation reads $\tau_m\,dV/dt = -V(t) + \mu + I(t)$, where $\tau_m$ is the membrane time constant, $V$ is the membrane potential, $\mu$ a bias current, and $I(t)$ the input signal. LIF neurons underpin a wide range of theoretical neuroscience models, spiking neural network architectures, neuromorphic hardware platforms, and mathematical analyses of collective neural dynamics. The model is notable for its analytic tractability, computational efficiency, and versatility in accommodating stochasticity, network interactions, heterogeneous inputs, and diverse implementation substrates.

1. Core Model Equations

The single-neuron LIF model is formulated as a first-order ordinary differential equation: $\frac{du_i}{dt} = -u_i(t) + \mu$ where $u_i(t)$ is the membrane potential of neuron $i$, and $\mu$ is a constant suprathreshold drive. When $u_i(t)$ attains the threshold value $u_{\text{th}}$ (typically $u_{\text{th}} = 0.98$), the neuron emits a spike, and $u_i$ is instantaneously reset to $u_{\text{rest}} = 0$: $$\text{if } u_i(t^-) = u_{\text{th}},\quad u_i(t^+) = 0.$$ In networked contexts, such as non-locally coupled LIF rings, the membrane potential evolves as: $$\frac{du_i}{dt} = -u_i(t) + \mu + \frac{\sigma}{2R} \sum_{j=i-R}^{i+R} (u_j(t) - u_i(t))$$ where $\sigma$ is the coupling strength, $R$ is the coupling range (number of nearest neighbors on either side), and indices are cyclic modulo $N$ (number of neurons) (Tsigkri-DeSmedt et al., 2015).

With an absolute refractory period $p_r$, after each spike neuron $i$ is held at $u_i = 0$ for duration $p_r$, during which its dynamics are suspended.

2. Coupled Network Dynamics and Chimera States

Networked LIF models exhibit intricate spatiotemporal patterns depending on connectivity, coupling parameters, and refractory mechanisms. For a ring of $N=1000$ identical LIF neurons coupled to their $R$ nearest neighbors:

• Chimera states—coexistence of spatially coherent (plateau-like) and incoherent (fluctuating) regions—arise for intermediate $\sigma$ values ($0.50 < \sigma < 0.7$ depending on $R$).
• Increasing $R$ shifts the $\sigma$ band supporting chimeras upward.
• Introducing a refractory period $p_r$ substantially broadens the region in $(\sigma,R)$ space admitting chimera solutions and enhances their stability; further, the chimera multiplicity (number of distinct coherent/incoherent patches) grows quasi-linearly with $p_r$ up to saturation (Tsigkri-DeSmedt et al., 2015).

In multiplex topologies composed of two LIF rings with intra- and inter-ring coupling, varying intra-ring strengths and sign of inter-ring coupling produces:

• Stationary multi-chimera configurations with phase-opposed coherent/incoherent clusters.
• Subthreshold oscillatory regimes in which neurons in one ring remain below firing threshold, entrained by the dynamics of the other ring.
• Distinct synchronization signatures for attractive versus repulsive inter-ring coupling (e.g., inversion of mean-frequency profiles). A systematic parameter scan yields phase diagrams for coherence, chimerism, and subthreshold domains (Anesiadis et al., 2023).

3. Stochasticity, Noise, and First-Passage Analysis

LIF neurons under random input (Gaussian, Poissonian, $1/f$ noise) are modeled by stochastic differential equations, most often of the Ornstein-Uhlenbeck type. The canonical noisy LIF equation is: $$dV(t) = \left[ -\frac{V(t) - V_r}{\tau_m} + \frac{I_0}{C} \right] dt + \frac{\sigma}{C} dW(t)$$ with $V_r$ the resting potential, $I_0$ bias current, $C$ capacitance, and $W(t)$ standard Brownian motion. Spike generation and reset are triggered upon $V(t) \geq \theta$.

The mean firing rate can be explicitly computed by solving the first-passage time problem for the corresponding Fokker-Planck equation, yielding the Siegert formula (Kreutz-Delgado, 2015). Notably, $1/f$ noise leads to non-saturating, super-Poissonian Fano factors and near-optimal fast response to step stimuli (Sobie et al., 2010). In noisy LIF networks, the invariant measure governing spike-train statistics is of Gibbs type, with exponential decay of ISI tails (Cessac, 2010).

4. Mathematical Structure and Mean-Field Theory

The LIF neuron admits rigorous treatment in measure-theoretic and PDE formalisms. For networks, mean-field theory yields nonlinear transport equations for the population density $\rho(t,v)$ over membrane potential $v$: $$\partial_t \rho - \partial_v [ (-v + b + J r(t)) \rho ] = r(t) \delta(v - v_r)$$ where $J$ is the global coupling, $r(t)$ the population firing rate, and $v_r$ the reset potential (Dumont et al., 2017). Strong coupling can induce finite-time blow-up (divergence of firing rate), while moderate coupling ensures global well-posedness, uniqueness, and exponential stability of asynchronous steady states.

In event-driven network simulations, precise storage capacity and attractor dynamics have been established, including phase-coded memory retrieval (Scarpetta et al., 2012).

5. Analog, Discrete, and Neuromorphic Implementations

Discrete-time, analog-output, and hardware-accelerated LIF variants extend classic dynamics. In event-driven systems, the analog fire mechanism transmits continuous values rather than binary spikes, maintaining leaky integration and improving accuracy without additional gating complexity. The LIAF-Net architecture employs leaky integration with analog fire for efficient spatiotemporal processing, achieving competitive or superior performance to LSTM, GRU, and ConvLSTM networks with dramatically lower weight and compute requirements (Wu et al., 2020).

Neuromorphic hardware instantiations include skyrmion-based magnetic tunnel junctions (Lone et al., 2022), domain-wall spintronic devices (Sekh et al., 16 Aug 2024), and photonic RTD circuits (Robertson et al., 28 Jan 2025), all embedding the leaky integration and threshold-reset dynamics of LIF via physical analogues (e.g., VCMA-controlled skyrmion size for integration, magnetic relaxation for leak, domain-wall position and electrical readout for firing threshold). Experimental results demonstrate sub-ns membrane time constants, GHz-range spiking, and energy-efficient leaky integration.

6. Extensions: Periodic, Stochastic, and Generalized LIF

LIF models have been rigorously analyzed under almost-periodic and stochastic inputs. With Stepanov almost periodic forcing, the firing and displacement maps are uniformly continuous and admit unique firing rates independent of initial time (Kasprzak et al., 2016). Discrete-time LIF networks with noise yield exact expressions for firing statistics, correlations, and invariant measures (Cessac, 2010). Generalizations with stochastic firing thresholds (Ornstein-Uhlenbeck or more complex Gauss-Markov processes) recast the problem as first-passage for Brownian motion against a moving boundary, enabling both exact and perturbative computations of interspike interval distributions and revealing non-monotonic mean firing times as a function of threshold noise amplitude (Braun et al., 2015).

7. Criticality, Chaos, and Order Symmetry Breaking

Continuous LIF (c-LIF) models introduce inertial terms in membrane dynamics, enabling analytically solvable non-instantaneous resets and novel dynamical regimes (Stucchi et al., 2021). In fully connected networks with short-term synaptic plasticity, c-LIF units exhibit period-doubling cascades into chaos and support scale-free avalanches. The onset of broad, heavy-tailed event distributions is mechanistically linked to order symmetry breaking in neuron firing sequences. This reveals a minimal route to critical-like dynamics without requiring heterogeneity or fine-tuning of model parameters.

---

Leaky integrate-and-fire dynamics constitute a mathematically transparent and biophysically grounded framework that accommodates a broad spectrum of phenomena—chimera states, noise-driven variability, memory attractors, chaos and criticality, and efficient hardware realization. Their foundational role in network theory, signal processing, and neuromorphic engineering continues to motivate rigorous analysis and experimental innovation across computational neuroscience and beyond.