Many neural models of event data, such as hospital visits, trades, or pixel flips on an event camera, describe the instantaneous rate of events as the exponential of a learned function. Training them requires the integral of that rate over every quiet stretch between events. This integral is called the compensator. It has no closed form, so practitioners approximate it with a handful of grid points or random samples. The cost of those approximations dominates training, and their error is silently folded into reported likelihoods. We propose a different split of the work. A Gaussian process is fit to the log-rate, which is smooth and cheap to describe, using a few expensive network evaluations. The integral of the exponentiated surrogate is then computed from moments that are exact in one dimension. In multi-dimensional evidence integration these moments forced Taylor approximations. For an integral over time they reduce to one- and two-dimensional integrals of cheap analytic functions, which can be computed to machine precision.
The resulting estimator gives a posterior mean for the compensator and a calibrated error bar for it. The error bar decides where to spend additional network evaluations. The proposal also makes one structural link. When the log-rate is built from continuous convolution kernels, the kernel's own bandwidth fixes the smoothness of the Gaussian process prior. Here a continuous kernel means a filter defined as a smooth function of time rather than as a fixed array of weights. Fixing smoothness from the kernel side avoids the hyperparameter-fitting pathologies documented for warped quadrature. The claim is specific and testable. At matched likelihood accuracy, this estimator needs substantially fewer network evaluations than adaptive Gauss–Kronrod quadrature. Its reported error bars should cover the true integration error at their nominal rate.
Core Claim
Consider a temporal point process with rate $\lambda(t) = \exp(h(t))$, where $h(t) = \mu + \sum_{t_j < t} \psi(t - t_j)$ and $\psi$ is a continuous kernel. The kernel could be a sinusoidal coordinate network as in CKConv, a Gabor network with a learnable Gaussian mask as in FlexConv, or a Jacobi-polynomial expansion as in PLEIADES. On each inter-event interval we evaluate $h$ at a few nodes and condition a Gaussian process on those values, which gives a posterior mean $m(t)$, variance $s^2(t)$ and covariance $c(t,t')$. The compensator $\Lambda = \int_a^b \exp(h(t))\,dt$ then has log-normal-process moments.
$$ \mathbb{E}[\Lambda] = \int_a^b e^{m(t) + s^2(t)/2}\,dt, \qquad \mathrm{Var}[\Lambda] = \int_a^b\!\!\int_a^b e^{m(t)+m(t') + (s^2(t)+s^2(t'))/2}\left(e^{c(t,t')}-1\right)dt\,dt'. $$
Both integrands involve only the surrogate and never the network, so a dense deterministic grid evaluates them to negligible error at negligible cost. This removes the specific approximation that Improving Quadrature for Constrained Integrands (MMLT, a log-transformed Bayesian quadrature with moment matching) needed. In its multi-dimensional evidence setting, the moments of the exponentiated process were intractable and were approximated by Taylor expansion. That truncation error never entered the uncertainty envelope. The novelty is therefore not "Bayesian quadrature for point processes" in general. It is the observation that in the one-dimensional temporal regime the log-warp is free: it costs nothing in tractability. Combined with a kernel-informed prior, this makes calibrated log-space quadrature cheap enough to run inside every training step.
Relation to Prior Work
Three works motivate and supply the method. Each sits in the setting of static, low-dimensional evidence integration. Sampling for Inference in Probabilistic Models with Fast Bayesian Quadrature (WSABI) uses a square-root warp because its moments stay tractable. Fast Bayesian Inference with Batch Bayesian Quadrature via Kernel Recombination (BASQ) keeps that warp and adds fast batch selection. Square-root warps only halve the dynamic range, while log-rates in self-exciting processes can swing by many nats within one interval. Chai and Garnett's MMLT accepts the log warp at the price of Taylor moments. We reuse their moment-matching view. We also reuse their warning that fitting hyperparameters in the warped space can miscalibrate uncertainty in the original space. One difference matters for that warning. Here we observe $h$ directly and exactly, whereas they observe $f$ and infer $g$. Their trap may therefore not transfer unchanged. We test this rather than assume it.
The continuous-kernel papers supply the model class and one constraint. CKConv and FlexConv evaluate kernels at arbitrary coordinates. This is what makes $h$ queryable at any node. FlexConv's analytic bound on the maximum frequency a kernel can represent gives a principled lengthscale for the Gaussian process on $h$. PLEIADES integrates polynomial kernels exactly over time bins, but only for unconstrained, linear-in-kernel outputs. Exact antiderivatives do not survive the exponential that guarantees a positive rate, and that gap is the one this proposal fills.
Log-intensity Gaussian process models such as Scalable Nonparametric Bayesian Inference on Point Processes with Gaussian Processes treat the log-rate itself as the random object. We instead treat a deterministic neural log-rate as unknown only because it has been evaluated at finitely many points. In other words, our uncertainty is numerical and not statistical, in the sense of Probabilistic numerics and uncertainty in computations. Finally, the file notes that Monte Carlo compensators inject gradient variance that destabilizes training (Zhu et al.). This motivates a deterministic-mean estimator.
Research Program
The first component is the estimator. For each interval we start from a small fixed node set. We build the Gaussian process on $h$ with a constant or linear prior mean, following the polynomial-mean logic of A Bayes-Sard Cubature Method so that the estimator is exact when $h$ is affine. We compute the two moments on a dense surrogate grid. Training uses $\mathbb{E}[\Lambda]$ in the log-likelihood. That loss is differentiable through $m$ and $s^2$, so gradients reach the kernel parameters. The variance plays two roles. It reports integration error on held-out likelihoods. It also drives adaptive refinement: new nodes are added at the maximizer of the integrand's posterior variance until the relative standard deviation of $\Lambda$ falls below a tolerance. Because intervals are independent and each Gram matrix is only a few nodes wide, the cubic cost of Gaussian process inference is irrelevant here. Its usual price in Bayesian quadrature is paid in the number of nodes, which stays small.
The second component is the prior. We compare three ways of setting the lengthscale. The first derives it from the kernel's analytic bandwidth, using FlexConv's maximum-frequency computation or the polynomial degree in PLEIADES. The second maximizes the marginal likelihood of $h$ at the nodes. The third is the $f$-space fit of Chai and Garnett. For each we measure calibration of the compensator error bar. The third component is analysis. In one dimension, deterministic interpolation error bounds for $h$ translate into bounds on $\Lambda$ that scale with $e^{\max h}$. This explains why polynomial rules applied directly to $\lambda$ degrade when the log-rate has a large range within an interval. It is also the regime in which we expect the method to help, and outside it we expect no gain.
Baselines
The naive reference point is current practice: a uniform trapezoidal or Riemann rule on $\lambda$ with a fixed node count per interval, applied directly to the network output. Every method is compared at equal numbers of network evaluations of $h$. That count is the real cost, and we also report wall-clock time.
The strongest classical competitor is adaptive Gauss–Kronrod quadrature on $\lambda$. It has exponential convergence for smooth integrands, comes with a built-in error estimate, and runs in linear time. We give it the same tolerance-driven adaptivity and the same evaluation budget. Beating it means reaching a held-out log-likelihood within a fixed tolerance of a high-precision reference with at least two times fewer evaluations, in the regime where $h$ varies over several nats per interval. It also means that our error bars cover the true error at near-nominal rates while Gauss–Kronrod's error estimate under-covers. If Gauss–Kronrod matches us on both counts, the claim fails.
A Monte Carlo compensator with the same number of evaluations is included because it is unbiased and common in neural point processes. We judge it on gradient variance and training stability as well as accuracy. Three ablations isolate the ingredients. Linear-space Bayesian quadrature on $\lambda$ removes the log warp. A version that uses MMLT's Taylor moments instead of exact one-dimensional moments isolates the exact-moment step. Marginal-likelihood lengthscales instead of kernel-derived ones isolate the prior.
The most important ablation is deterministic: interpolate $h$ with a Chebyshev polynomial at the same nodes and integrate $\exp$ of the interpolant densely. This isolates whether the probabilistic layer adds anything beyond the log-space interpolation itself.
On the question of reported numbers, there is no reported result on Lacuna for compensator accuracy or calibration in neural temporal point processes. The closest quantitative claims in the file are qualitative. MMLT reports lower median absolute evidence error than WSABI, Bayesian Monte Carlo and sequential Monte Carlo on Damped Lyman-$\alpha$ detection, but the file gives no number. Baselines therefore have to be run by us under identical conditions.
Why not just…?
Why not just use Gauss–Kronrod or Gauss–Legendre on the rate? These rules are cheap and accurate for smooth functions, and adaptive Gauss–Kronrod even reports an error estimate. Their accuracy constant, however, grows with the range of $\exp(h)$ over the interval. A burst of excitation makes $\lambda$ sharply peaked even when $h$ is a gentle curve. The log-space surrogate interpolates the gentle curve instead. This is a real threat on intervals where $h$ is nearly flat, and there we expect no advantage. The experiments stratify intervals by within-interval log-rate range, so that any gain can be located precisely.
Why not just interpolate the log-rate deterministically and integrate its exponential? This is the most serious alternative. It would plausibly match our posterior mean, because most of the accuracy comes from the log-space interpolation. We therefore claim the mean improvement for log-space interpolation in general and not for Bayesian quadrature. The probabilistic contribution has to come from the error bar. Chebyshev tail coefficients give a heuristic error estimate. Our variance gives a calibrated one that propagates exactly through the exponential. If coverage experiments show the tail-coefficient estimate is equally well calibrated and drives refinement equally well, we will report that the probabilistic layer is unnecessary and narrow the contribution.
Why not just use Monte Carlo compensators? They are unbiased and trivially parallel. However, Zhu et al. is cited in the file as showing that sampling-based compensators add gradient variance, and Monte Carlo provides no smoothness-exploiting accuracy at small budgets. We measure gradient variance and convergence curves at equal budgets to check this directly.
Why not just choose kernels with closed-form compensators? Exponential Hawkes kernels, or linear rates with polynomial kernels as in PLEIADES, avoid quadrature entirely. They do so by giving up either kernel flexibility or the guarantee of a positive rate. We include an exact-compensator exponential Hawkes model as a reference. Its likelihood gap to the continuous-kernel models measures what flexibility buys, and that gap is the price the alternative pays.
Proposed Experiments
The first experiment is synthetic. We use Hawkes and self-correcting processes with exponential and sum-of-Gaussian kernels, where compensators are known exactly, and sweep the excitation strength to control within-interval log-rate range. On these we measure absolute compensator error against network evaluations. We also measure the empirical coverage of 50%, 90% and 99% intervals from our variance and from each competitor's error estimate.
The second experiment trains neural point processes whose log-rates are built from SIREN, MAGNet-with-mask and Jacobi kernels. The Jacobi implementation comes from the PeaBrane/Pleiades codebase. Training data are irregularly sampled PhysioNet records, the setting CKConv evaluated on. As an exploratory event source we also use per-pixel event streams from DVS128, the PLEIADES benchmark. For both datasets we report held-out log-likelihood with integration error bars, the deviation from a high-precision reference of a few hundred nodes, evaluations per interval, and training stability.
The decisive checks are three. The first is whether, at matched reference-relative likelihood error, the log-space estimator uses at least two times fewer evaluations than adaptive Gauss–Kronrod in the high-log-range stratum. The second is whether its intervals achieve near-nominal coverage where competitor estimates do not. The third is whether either advantage disappears in the deterministic-interpolation ablation. A positive answer to the first two together with a negative answer to the third would establish that calibrated numerical uncertainty, not just log-space interpolation, earns its place in training continuous-kernel event models.
References
- CKConv: Continuous Kernel Convolution For Sequential Data
- FlexConv: Continuous Kernel Convolutions With Differentiable Kernel Sizes
- PLEIADES: Building Temporal Kernels with Orthogonal Polynomials
- Improving Quadrature for Constrained Integrands
- Sampling for Inference in Probabilistic Models with Fast Bayesian Quadrature
- Fast Bayesian Inference with Batch Bayesian Quadrature via Kernel Recombination
- Scalable Nonparametric Bayesian Inference on Point Processes with Gaussian Processes
- Probabilistic Numerics and Uncertainty in Computations
- Neural Spectral Marked Point Processes
- A Bayes-Sard Cubature Method