Modeling Inhibitory Interactions in Temporal Point Processes

54 concepts40 papers1 proposal

In many real-world systems, events do not just trigger future activity; they can also actively suppress it. A neuron firing often forces its immediate neighbors into silence. Making a major vehicle purchase drastically reduces the likelihood that you will buy another one next week. A counter-campaign or censorship effort can actively dampen the spread of viral information.

Yet, for decades, the mathematical frameworks used to model event sequences have suffered from a unidirectional excitation bias. While classical models excel at capturing cascading, self-exciting behavior, extending them to handle inhibitory dynamics introduces a fundamental mathematical roadblock.

To understand why this problem occurs—and how modern research resolves it—we must look at the mechanics of how event probabilities are calculated over time.

The Mathematical Trap of Additive Intensity

The gold standard for modeling irregular, continuous-time event sequences is the temporal point process. At its core, a point process relies on a conditional intensity function, denoted as $\lambda(t)$. This function represents the instantaneous arrival rate of an event given the history of everything that has happened up to time $t$.

In a classical linear Hawkes process, this intensity is modeled additively:

$$\lambda(t) = \mu + \sum_{t_i < t} \alpha e^{-\beta(t - t_i)}$$

Here, $\mu$ is a baseline rate, and each past event $t_i$ contributes a positive bump ($\alpha > 0$) to the intensity, which decays over time. Because everything is added together, $\lambda(t)$ is guaranteed to remain non-negative. This non-negativity is a strict mathematical requirement: an intensity function cannot drop below zero, just as a probability cannot be negative.

The trap arises when we try to model inhibition within this additive framework. As noted by Wingate et al. in The Infinite Latent Events Model, restricting interactions to purely excitatory triggers leaves models incapable of representing suppressive causality. If we simply allow the impact parameter $\alpha$ to be negative to represent an inhibitory event, the math breaks down. If a cluster of inhibitory events occurs, or if the system enters a quiet period where the baseline $\mu$ is low, the summation can easily drag $\lambda(t)$ below zero.

As highlighted by Apostolopoulou et al. in Mutually Regressive Point Processes, this constraint prevents standard formulations from safely capturing regulatory mechanisms like neural refractory periods. If a model cannot guarantee non-negativity under negative influences, it cannot be reliably trained.

To allow history to exert a negative pull without breaking the probability space, researchers introduced nonlinear link functions. Instead of calculating the intensity directly as a sum of historical influences, the history is aggregated into an unconstrained latent state—which can freely swing positive or negative—and then passed through a "squashing" function that maps $(-\infty, \infty)$ to $[0, \infty)$.

One prominent approach is the use of a sigmoid link function. In Efficient Inference of Flexible Interaction in Spiking-neuron Networks, Zhou et al. propose the Sigmoid Nonlinear Multivariate Hawkes Process (SNMHP). By passing the historical event terms through a sigmoid function, the final intensity is bounded and strictly positive, allowing the model to seamlessly integrate both excitatory and inhibitory biological mechanisms. Similarly, Zhang et al. demonstrate in Conjugate Bayesian Two-step Change Point Detection for Hawkes Process that a sigmoid link function provides the exact flexibility needed to discover change points in systems where events suppress rather than trigger subsequent arrivals.

Excitatory History (+) \
                       --> Latent Sum (Unconstrained) --> [ Sigmoid / Exp ] --> Valid Intensity λ(t) ≥ 0
Inhibitory History (-) /

However, this elegant solution introduces a new trade-off: it destroys the conjugacy of the model. In a standard linear Hawkes process, the likelihood function has a clean mathematical form that makes parameter optimization straightforward. Introducing a nonlinear link function makes the exact calculation of the network's history mathematically intractable to integrate directly.

To overcome this, the authors typically rely on data augmentation tricks. For example, by introducing auxiliary latent variables (such as Polya-Gamma random variables), the non-linear inference problem can be transformed back into a series of quasi-linear steps, allowing fast Expectation-Maximization (EM) or variational updates without sacrificing the ability to model negative influences.

Neurally Modulated and Continuous Latent Dynamics

Rather than relying on fixed, parametric link functions to squash historical sums, another paradigm uses continuous-time latent states to track the system's evolution.

In The Neural Hawkes Process: A Neurally Self-Modulating Multivariate Point Process, Mei et al. bridge point processes with Recurrent Neural Networks (RNNs). Instead of maintaining a discrete sum of past event kernels, the model maintains a continuous hidden state vector that updates abruptly whenever an event occurs, and decays continuously between events. Because the hidden state interacts with the output through non-linear transfer functions, the network can naturally learn complex real-world dynamics where certain occurrences actively prevent or delay future events. This concept is mirrored in the Self-Attentive Hawkes Process, which uses attention layers to look back at history and dynamically modulate parameters above or below a baseline.

Alternatively, non-parametric frameworks leverage Gaussian Processes (GPs) to achieve a similar goal without deep neural architectures. Liu et al. demonstrate in Nonparametric Regressive Point Processes Based on Conditional Gaussian Processes that modeling intensities as non-linear functions of relative history via conditional GPs allows the model to capture a fluid mix of inhibition and excitation. By placing a GP prior over the history-dependent intensity, Wang et al. show in Self-Modulating Nonparametric Event-Tensor Factorization that models can successfully capture long-horizon human activity patterns—such as cyclical resource consumption—where an event inherently dampens the immediate probability of its next occurrence.

Alternative Paradigms for Repulsive Distributions

What if we abandon the conditional intensity framework altogether? When a system is dominated by strict "winner-take-all" dynamics or spatial-temporal exclusion, trying to force inhibition into an intensity function can feel like an indirect workaround.

Snoek et al. tackle this problem from a completely different geometric perspective in A Determinantal Point Process Latent Variable Model for Inhibition in Neural Spiking Data. Instead of modeling event rates over time, the authors leverage Determinantal Point Processes (DPPs).

A DPP defines the probability of a subset of events occurring based on the determinant of a similarity kernel matrix, $L$:

$$P(Y) \propto \det(L_Y)$$

In linear algebra, the determinant of a matrix measures the volume of the parallelepiped spanned by its feature vectors. If two vectors are highly similar (meaning two events are close in time or space), they are nearly parallel, causing the volume—and thus the determinant—to shrink toward zero.

Excitatory Model (Hawkes): Events cluster together because occurrence increases local rate.
[● ● ●● ●]       [      ]       [● ●●  ●]

Repulsive Model (DPP): Events push each other away because similarity drops probability to zero.
[  ●   ●   ●   ●   ●   ●   ●   ●  ]

By utilizing this geometric property, DPPs inherently assign a low probability to event configurations where similar events occur close together. This bypasses the non-negativity problem entirely: inhibition is not modeled by subtracting from a rate, but by a mathematical formulation that naturally penalizes co-occurrence, providing an elegant and tractable framework for capturing pure repulsion.

Key Foundational Work

The Infinite Latent Events Model (2009) — contributed 1 concept elements to this direction.
Neuronal connectivity inference from spike trains using an empirical probabilistic causality measure (2009) — contributed 1 concept elements to this direction.
Spatial Organization of Neural Networks: A Probabilistic Modeling Approach (1987) — contributed 1 concept elements to this direction.
History & citationVersion 2 · 2026-10-04

Current · Permanent link

  1. Version 2 (latest) · 2026-10-04 · 39 papers
  2. Version 1 · 2026-10-04 · 60 papers
Cite this direction

Latest version (a living page):

@misc{lacuna_inhibitory_event_modeling_c24791cb,
  title = {inhibitory event modeling},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/inhibitory-event-modeling/rdr_c24791cb357e485897ec13dff454aecf},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 2, frozen text):

@misc{lacuna_inhibitory_event_modeling_c24791cb_v2,
  title = {inhibitory event modeling},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/inhibitory-event-modeling/rdr_c24791cb357e485897ec13dff454aecf/v/2},
  note = {Version 2, frozen 2026-10-04. Accessed 2026-11-20},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake