Changelog • ngme2

ngme2 0.9.5 (2026-03-17)

Add var1() — a Vector Autoregressive order-1 latent field for bivariate time series modelling.
Stationarity is guaranteed by construction via the Cayley reparameterization: four unconstrained parameters $(p_1, p_2, p_3, p_4)$ are mapped to a $2\times2$ VAR coefficient matrix $A$ with spectral radius $\rho(A) < 1$ at every iteration.
The C++ backend implements the Cayley transform natively (class RCallback in src/latents/rcallback.cpp).
Supports all ngme2 noise distributions (NIG, GAL, normal) as a single shared innovation noise.
print() method displays the recovered $A$ matrix, its spectral radius, and the raw $(p_1, p_2, p_3, p_4)$ values.
Add vignette “VAR(1) Bivariate Model in ngme2” (vignettes/var1-model.Rmd) covering: model specification, Cayley reparameterization, simulation study with parameter recovery, convergence trace plots, and NIG vs Gaussian model comparison.

Add posterior distribution plotting for SGLD samples via posterior_plot().
Add plot() support for ngme_sgld_ci objects, reusing stored SGLD samples to visualize marginal posterior distributions.

Refine fixed-effect standardization: SVD now applies only to non-intercept columns; intercept columns remain on their original parameterization.
Preserve no-intercept model semantics (~ 0 + ...): skip fixed-effect centering when no intercept is present.
Improve fe() centering with structural zeros: grouped fe() columns are centered using in-group rows only, so out-of-group structural zeros remain zero.
Fix fixed-effect scale restoration for multi-replicate models: when needed, reconstruction now uses replicate row indices (data_idx) instead of always taking the first n rows.
Improve warm-start robustness (start = previous_fit) across standardization settings by remapping fixed effects through the current model parameterization.
Update default fixed-effect priors: intercept-like columns ("(Intercept)"*) now default to prior_none(), while non-intercept columns keep the default N(0,10) prior.
For standardize_fixed = TRUE, add prior compatibility handling: isotropic normal priors on standardized columns are transformed to the SVD basis; incompatible custom prior_beta specifications now automatically disable fixed-effect standardization with a warning.

Add prior_inv_exponential(lambda, lower) for nu, implementing kappa = 1 / nu ~ Exp(lambda) as a first-class prior option.
Add shorthand alias prior_inv_exp(...) for the same prior.
Add calibration helpers calibrate_inv_exp_lambda_driven_nig() and calibrate_inv_exp_lambda() for choosing lambda from a driven-noise tail-inflation target.
Improve calibration robustness for non-monotone R_c(nu) curves: the helper now scans for crossings and reports observed R_c range when the requested target is unattainable.
Update default nu prior in f() for NIG-driven noise: when nu prior is not explicitly set and nu is stationary, use prior_inv_exp(lambda = log(2)/median(h), lower = nu_lower_bound). For non-stationary nu, keep the legacy N(0,10) default prior.

Harden error handling in ngme() estimation/sampling path: C++ exceptions are now propagated as R errors (including OpenMP parallel regions) instead of potentially terminating the R session.
Fix nu initialization in noise helper constructors to respect nu_lower_bound, using theta_nu = log(nu - nu_lower_bound) and validating nu > nu_lower_bound.
Align normal_nig conversion, printing, and plotting with effective parameterization nu = nu_lower_bound + exp(theta_nu).

Refactor prior API (breaking change): use prior_normal(), prior_pc_sd(), prior_half_cauchy(), prior_none(), and priors(...).
Update f() and ngme_noise() to accept unified prior = ... inputs (remove prior_theta_K and prior_mu/prior_sigma/prior_nu arguments).
Remove legacy ngme_prior() interface and its documentation entry.
Add prior target support (coef/field) for noise parameter priors and per-parameter operator prior compilation.
Add fixed-effect prior support via ngme(..., prior_beta = ...), using the same prior_*()/priors(...) API.
Add user-facing vignette: Prior Templates for Stationary and Non-Stationary Models.

Add grad.norm plateau-based step decay via control_opt(stepsize_decay = "grad_norm_plateau") (epoch-level, synchronized across chains)
Add stepsize_decay() helper for configuring decay options
Update verbose output to include stepsize decay scale and effective stepsize
Improve cross_validation(data = ...) model rebuild for refit-on-new-data workflows: it now resolves external formula symbols (for example mesh, B, n_basis) from the fitted object when needed, and falls back to rebuild-without-start plus hyperparameter transplant if start state dimensions differ.
Add chain-aware prediction/CV aggregation via chain_combine = "predictive_average" in predict() and cross_validation(), which averages predictions across optimization chains instead of averaging parameters first.

Update nu parameterization to be relative to nu_lower_bound (nu = nu_lower_bound + exp(B_nu * theta_nu))
Allow NULL specification for noise parameter in non-stationary case
Update fixed effects initialization
Update start=fit logic

Simplify solver structure
Update to use model() instead of string for model definition
Big update on the R interface, use model() instead of string for model definition
Refactor on the codebase
Add more vignettes