NEWS

---

emhawkes 1.0.2

• Match native simulation and likelihood accumulation to the precision used by the running R build, including R configured with --disable-long-double. Preserve event/RNG order and callback state consistency without changing existing numerical test thresholds. Retain extended precision for derivative calculations that do not reproduce an R accumulation.
• Materialize marked likelihood jumps before addition on every R build, matching simulation rounding when fused multiply-add optimization is enabled. Add an optimized-arithmetic regression to native-safety CI.
• Add a no-long-double CI check alongside the standard platform checks.

emhawkes 1.0.1 (2026-10-18)

• Correct the stats-only subprocess test when maxLik is installed in a standard or site library. Always check stats fitting and saved-result restoration without loading maxLik; check the missing-provider error only when maxLik is genuinely unavailable. Verify the package path used by fresh-session tests.
• Give callbacks with ... consistent count paths and intensity histories with and without inference output, correcting likelihoods that depended on infer.
• Support stats::logLik() on hspec objects. Lazily load maxLik's methods when saved maxLik fits are first printed, summarized or queried; keep their classes and preserve the stats optimizer's independence from maxLik.
• Integrate density-based waiting-time moments in log time, including finite truncation limits, to avoid missing narrow densities or accepting large relative errors through the integrator's default absolute tolerance. Handle singular residual densities at times or cumulative hazards that round to zero during quadrature; diagnose underflow when its error bound exceeds tolerance.
• Return empty legacy residual vectors for types with fewer than two events. Report unavailable stationary mean-intensity systems in hvol() explicitly.
• Stabilize the trapezoid-exponential coefficients and log-tail probabilities near its small-a, ell = 1 limit.
• Avoid quantized or zero arrival times from subnormal decay intermediates in the R and C++ equal-decay simulators. Keep the RNG draw order unchanged.
• Add default R CMD check jobs with _R_CHECK_SUGGESTS_ONLY_=false and maxLik exposed through a site library, alongside the existing --as-cran jobs. Retain a separate hard-dependencies-only CI job that requires maxLik's absence.

emhawkes 1.0.0 (2026-10-17)

Initialization and estimation

• Default simulation, likelihood and fitting to lambda_component0 = "auto" (NULL is an alias). Report whether mean or zero intensity components were selected. Mean initialization applies to supported stable constant linear models, including mapped kernels; other models use zero and recommend an explicit initial state. Components exclude the baseline intensity.
• Correct the univariate half-mean initialization. Preserve explicit initial states, hold the selected fitting rule fixed and differentiate mean components. Store the rule and actual components in fit diagnostics; recover the actual initial states when replaying older saved fits. Automatic initialization can change simulated paths and fitted estimates relative to previous releases.
• Add hfit(optimizer = "stats"), using R's optim() or constrOptim() without maxLik. Support fixed/shared parameters, derivatives, linear inequalities and box bounds, with separate result methods and fitted-model reconstruction. Make maxLik optional while retaining it as the default provider; install it explicitly when using the default. Keep numerical score steps within linear constraints and reject unsupported controls and equality constraints.
• Accept both constraint and constraints, preserving positional arguments. Reject simultaneous non-NULL values and normalize either spelling for maxLik.
• Add fitted_hspec(), fit_initial_state() and fit_diagnostics() to restore fitted specifications and inspect optimizer, engine and curvature choices.
• Add final_hessian = "stable", with scaled differences, analytic-score differentiation where available and step-stability checks. Withhold Wald covariance when curvature or constraint checks fail. Stable curvature is optional for maxLik and the default for stats fits; it does not establish finite-sample confidence-interval coverage.

Computation

• Add registered Rcpp likelihood and intensity/residual inference engines for supported numeric, parameter-only and observed-state specifications. Preserve callback defaults and event timing; unsupported models retain R evaluation. emhawkes.engine selects automatic, reference R or required C++ likelihoods. Source installation now requires Rcpp and a C++ compiler.
• Add opt-in analytic scores for numeric models and supported component functions through gradient = "analytic" and jacobian. Preserve shared/fixed parameters and initial-state derivatives; reduce observed-baseline Jacobian storage. Numerical optimizer differentiation remains the default.
• Add a complete native simulation event loop for supported numeric and callback specifications, including multivariate and mapped multi-kernel models. Reduce history copying and reuse equal decay factors within each row. Preserve R fallback, explicit initial states, callback behavior and random-draw order when comparing the engines. The likelihood engine option does not select the simulation path.
• Preserve native RNG state around callback and warning handlers, including errors, nested simulation and handlers that draw random numbers. Add checks for older R/Rcpp versions, garbage collection, interruption and memory access.

Corrections and diagnostics

• Register S4 initialization and display methods through the package namespace. Preserve construction and saved-fit restoration in sessions where methods is installed but not attached, including isolated stats-optimizer sessions.
• Validate observations, initial components, decay rates, mappings, counting processes and callback parameters consistently. Correct multi-kernel event types, univariate multi-kernel likelihoods and constant-model stability checks. Preserve admissible signed components and stop when no further event occurs.
• Correct event-side baseline recording and return empty residuals for event types with fewer than two events. Use stable exponential integrals and log-scale residual distributions. Correct trapezoid-exponential generation and add log/tail distribution options. Paths can change where event selection or random generation was corrected.
• Correct expected_tau() for widely separated time scales and a positive probability of no further event. Correct mark handling in hvol() and reject models outside the bivariate moment formula's scope.
• Preserve display options when printing paths, honor zero displayed rows and retain single-row matrix formatting. Make initialization, callback, dimension, early-stopping and unavailable-curvature messages actionable.

emhawkes 0.9.8 (2025-08-26)

Fixed some bugs

Breaking changes

• Flexible residual point process with self/mutually exciting and exponential decay model is available.

emhawkes 0.9.7 (2023-02-02)

Fixed some bugs, improved efficiency in this version and several features have been added.

Breaking changes

• The lambda0 argument name used in previous versions has been changed to lambda_component0 in this version. This is to clearly indicate the meaning of the argument and to avoid confusion.

• The name of method vol is changed to hvol and this feature is currently experimental.

• The Vignette file contains more examples and explanations.

emhawkes 0.9.6

Fixed some bugs, improved efficiency in this version and several features have been added.

Breaking changes

• Slot eta is introduced which represents the constant part of impact.

• The concepts of rambda and rambda_component are introduced. They are closely related to the right-continuous version of the intensity process.

• For inference of intensity and goodness of fit, infer_lambda and residual_process functions are implemented.

• The method vol to measure the volatility is introduced and this feature is currently experimental.