intensify 0.3.2

Point process library with Hawkes specialization. Rust-accelerated MLE, EM, online, and Bayesian inference.

intensify

A Python library for point process modeling — Poisson, Cox, and Hawkes — with a Rust-backed implementation of the likelihood, gradient, and simulation hot paths. Provides closed-form gradients, Ogata thinning, cluster simulation, and time-rescaling goodness-of-fit diagnostics.

pip install intensify

Binary wheels are published for Linux (x86_64, aarch64), macOS (Intel and Apple Silicon), and Windows (x86_64) on Python 3.10–3.12. Source builds require a Rust toolchain; install via pip install 'intensify[fast]'.

Quickstart

import numpy as np
import intensify as its

# Simulate event times from a self-exciting process.
model = its.Hawkes(mu=0.6, kernel=its.ExponentialKernel(alpha=0.55, beta=1.4))
events = model.simulate(T=80.0, seed=1)

# Fit mu, alpha, and beta jointly from the observed events.
result = model.fit(events, T=80.0)

print(f"Branching ratio: {result.branching_ratio_:.3f}")
print(f"Log-likelihood:  {result.log_likelihood:.3f}")
print(f"Fitted params:   {result.flat_params()}")

# Visualize the fitted intensity.
fig = its.plot_intensity(result)
fig.savefig("quickstart_intensity.png", dpi=160)

Representative output:

Branching ratio: 0.547
Log-likelihood:  -66.671
Fitted params:   {'mu': 0.6271952244498643, 'alpha': 0.5470266035451343, 'beta': 1.0562059205150198}

Additional workflows:

# Multivariate connectivity: estimate directed excitation strengths.
kernels = [
    [its.ExponentialKernel(0.20, 1.0), its.ExponentialKernel(0.05, 1.0)],
    [its.ExponentialKernel(0.10, 1.0), its.ExponentialKernel(0.25, 1.0)],
]
mh = its.MultivariateHawkes(n_dims=2, mu=[0.5, 0.6], kernel=kernels)
mv_events = mh.simulate(T=100.0, seed=4)
mv_result = mh.fit(mv_events, T=100.0, fit_decay=False)
print(mv_result.connectivity_matrix())

# Goodness of fit: time-rescaling theorem residuals.
from intensify.core.diagnostics.goodness_of_fit import time_rescaling_test

ks_stat, p_value = time_rescaling_test(result)
print(f"KS stat={ks_stat:.3f}, p={p_value:.3f}")

# Cox process: latent, time-varying intensity for event-stream data.
lgcp = its.LogGaussianCoxProcess(n_bins=80, mu_prior=-0.2, sigma_prior=0.6)
events = lgcp.simulate(T=10.0, seed=11)
print(len(events), "events from an LGCP prior sample")

Features

• Process families:
  • Poisson: HomogeneousPoisson, InhomogeneousPoisson (callable intensity or piecewise-constant rates).
  • Cox: LogGaussianCoxProcess (LGCP), ShotNoiseCoxProcess.
  • Hawkes: univariate, multivariate, marked, nonlinear (softplus, sigmoid, relu, identity links), multivariate-nonlinear, and signed (inhibitory).
• Kernels: exponential, sum-of-exponentials, power-law, approximate power-law (Bacry–Muzy), and nonparametric (piecewise-constant). Each kernel is supported across all MLE code paths.
• Inference: MLE with hand-derived closed-form gradients (no autodiff in the hot path); recursive O(N) likelihood for exponential-family kernels; EM and online (streaming) updates routed through the same Rust core; optional Bayesian MCMC via NumPyro ([bayesian] extra).
• Diagnostics: time-rescaling theorem (KS and QQ on inter-compensator increments), AIC/BIC, and residual intensity.
• Simulation: Ogata thinning (general) and cluster/branching (Galton–Watson), both implemented in Rust.
• Stationarity enforcement: projected gradient for multivariate Hawkes; the spectral radius of the kernel-norm matrix is stored as branching_ratio_ on every multivariate FitResult.
• Architecture: Rust core (intensify._libintensify) for kernel, likelihood, gradient, and simulation hot paths. Pure-Python user-facing API. The package raises ImportError at import time if the compiled extension is unavailable.

Comparison with tick

tick is the established Python library for Hawkes-process inference. intensify covers a broader set of process families and kernels, and its exponential-Hawkes path is competitive with tick on wall-clock time while requiring less manual configuration. The summary below is intended as a quick orientation for users choosing between the two; both libraries remain valuable, and tick is a sensible choice when its feature set is sufficient.

Feature coverage

Capability	intensify	tick
Inhomogeneous Poisson (arbitrary rate function or piecewise-constant)	yes	simulation only
Log-Gaussian Cox Process (LGCP)	yes	—
Shot-Noise Cox Process	yes	—
Joint MLE of (μ, α, β); Hawkes decay fit from data	yes	decay must be supplied
MLE for power-law, approximate-power-law, and nonparametric kernels	yes	—
Marked Hawkes fit with any kernel	yes	—
Nonlinear (softplus/sigmoid/relu) Hawkes; signed kernels	yes	—
Multivariate stationarity enforcement (projected gradient)	yes	—
Time-rescaling test on inter-compensator increments	yes	yes
Python 3.10–3.12 support, prebuilt wheels	yes	3.8 only, C++ build

Performance

The benchmark suite builds seeded synthetic Hawkes datasets, fits the same model repeatedly, and reports median wall time over three runs. The comparison with tick locks the exponential decay β, because tick requires the user to supply it; intensify also runs in joint-decay mode, where β is estimated from data. Full methodology and reproduction commands are in docs/benchmarks.md.

Multivariate exponential, decay-given (mv_exp_5d), median wall time:

N	tick (ms)	intensify 0.3.1 (ms)
501	1.0	0.5
2,249	2.0	0.8
9,271	6.0	2.4
27,519	15.0	6.9
91,249	48.0	22.2

Parameter-recovery RMSE on the same problems is within 0.01 of tick at every N, and slightly lower at the larger sizes (full table in docs/benchmarks.md).

In joint-decay mode — where the kernel decay is fit alongside μ and α rather than supplied by the user — mv_exp_5d at N=1099 runs in about 14 ms. This is the path most lab users want, since avoiding a separate cross-validation loop over β is one of the practical motivations for the library. tick does not provide a comparable mode.

For kernels outside the exponential family — power-law, approximate-power-law, nonparametric, signed, marked, and nonlinear — intensify provides MLE paths that tick does not. The nonparametric path is viable at modest N: a 500-event fit completes in under one second using the binary-search bin lookup introduced in 0.3.0.

Documentation

Full documentation: https://hillmatt7.github.io/intensify

• Getting started
• User guide: inference, kernels, simulation, diagnostics
• API reference
• Tutorials (Jupyter notebooks)

Citation

If you use intensify in academic work, please cite it:

@software{intensify,
  author = {Hill, Matthew},
  title  = {intensify: A Python point process library with Rust-backed inference},
  year   = {2026},
  url    = {https://github.com/hillmatt7/intensify}
}

See CITATION.cff for the machine-readable form.

Contributing and changes

• Contributing guide
• Changelog
• Open issues

License

MIT — see LICENSE.