NEWS

---

cxreg 1.1.0

New features

• Spectral inference pipeline — complete end-to-end workflow for high-dimensional sparse spectral precision matrices (Deb, Kim, and Basu 2026):
  • select_m(): data-driven bandwidth selection via generalised cross-validation (GCV) on the diagonal periodogram (Ombao et al. 2001).
  • decglasso(): one-step debiased (desparsified) spectral precision estimator. Returns an object of class "decglasso".
  • var.cov(): asymptotic variance and pseudovariance estimation with real/imaginary decomposition. Supports plug-in and HAC estimators. Returns an object of class "varcov".
  • spec.test(): entry-wise Z-statistics, Mahalanobis chi-squared statistics, CI half-widths, and joint confidence ellipse areas. Returns an object of class "spectest".
  • spec.fdr(): FDR-controlled multiple testing for support recovery of the spectral precision matrix. Returns an object of class "specfdr".
• fhat_at(): smoothed periodogram matrix at a single Fourier frequency, added to spectral_functions.R.
• Parallel cross-validation — cv.classo(parallel = TRUE) is now fully operational via foreach / %dopar%. Requires a registered backend (e.g. doParallel::registerDoParallel()). A graceful fallback warning is issued when no backend is registered.

Bug fixes

• classo.control() was ignoring all arguments and always returning hardcoded factory defaults, making cv.classo(trace.it = 1) have no effect. Replaced with an environment-backed settings store so that changes (e.g. classo.control(itrace = 1)) persist for the R session.
• print.classo(): %Dev column was reading from the non-existent $dev field; corrected to $dev.ratio.
• cv.classo(): stop message incorrectly said cv.glasso; corrected to cv.classo. CV-only arguments (alignment, parallel) were not stripped before forwarding the call to classo().
• cv.classo.raw(): standardized = TRUE (wrong argument name) changed to standardize = TRUE in all classo() fold calls.
• cv_classofit(): removed a glmnet-inherited family$initialize block that was incompatible with complex regression. MSE was computed via abs() (MAE) rather than Mod()^2; corrected.
• buildPredmat(): S3 generic was defined after its method; reordered. Prediction matrix initialised as real NA instead of NA_complex_, causing silent type-coercion. The alignment switch() was commented out and bypassed; reactivated.
• dev_comp() and get_start() in classoFlex.R: residuals were split into real and imaginary parts incorrectly. Fixed to Mod(y - x %*% beta)^2. weighted.mean() on complex y (unsupported in base R) replaced with sum(w * y) / sum(w). t(rv) %*% x for lambda_max lacked conjugation; corrected to Conj(t(rv)) %*% x.
• plot.cglasso(): x[[index]] accessed a named field of the fit object by position rather than x$Theta_list[[index]]. is.integer(index) rejected all user-supplied plain integers (e.g. index = 1); replaced with is.numeric + round check. Matrix orientation was wrong (S[, nrow(S):1] reverses columns, not rows); corrected to t(S)[, p:1]. The imaginary panel in "both" mode used zlim = z_re instead of zlim = z_im. par() was restored twice (explicitly and via on.exit).
• predict.classo(): cbind2() (a sparse-matrix S4 generic) replaced with plain matrix multiplication. "link" type documented but missing from match.arg choices. nonzeroCoef() always dropped row 1 assuming an intercept; now conditional on object$a0.
• family.classo() / family.classofit(): were mapping glmnet-style class names (elnet, lognet, etc.) that do not exist in cxreg, returning NA. Both now return "gaussian" directly.
• coef.cv.classo(): missing names(lambda) <- s when s is a character string, making the output unnamed; added to match predict.cv.classo().
• lambda.interp(): top-level = assignment; two index assignments using =; comment formula sfrac*left+(1-sfrac*right) missing parentheses. All fixed.
• plotCoef(): switch(length(which) + 1, "0" = ...) — the "0" case was unreachable because length(NULL) + 1 = 1. Fixed to switch(length(active), "0" = ...). The which variable was reused for original indices and then overwritten inside each panel, causing labels to show local rather than original variable indices.
• cvtype(): subclass.ch = c(1, 2, 5) produced an NA entry (only 2 elements in type.measures); changed to c(1, 2). Both entries labelled "Mean-Absolute Error"; corrected to "Mean-Squared Error".
• error.bars(): returned range(upper, lower) visibly instead of invisible(NULL).

Internal changes

• Distinct S3 classes assigned to all return objects: "decglasso", "varcov", "selectm", "spectest", "specfdr" (previously some returned plain lists or had the wrong class).
• classo.path(): duplicate intercept/cbind block removed (would have appended the intercept column twice when intercept = TRUE). Orphaned classofit list object removed. classo.control() called only once.
• cxreg-package.R: removed unused @import foreach, @importFrom fields image.plot, and @importFrom Rcpp sourceCpp. Added @importFrom stats qnorm qchisq.

---

cxreg 1.0.5

• A minor fix was made during the revision for the Journal of Open Statistical Software.

cxreg 1.0.0 (2025-09-17)

• First stable release: CLASSO and CGLASSO estimation, cross-validation, coefficient and prediction methods, regularization path plots, and precision matrix heatmaps.

cxreg 0.8.0

• Pathwise estimation function for complex-valued graphical lasso updated.

cxreg 0.1.1

• plot.classo(): changed to display real and imaginary parts in separate panels.

cxreg 0.1.0

• Initial development version.