Skip to contents

Changelog

Source: NEWS.md

spikeyglass 0.1.0

PIP collapse (pi handling in the EM)

A diagnostic of the PIP collapse seen in the DiNA simulation is in data-raw/pip-collapse-diagnostic.md. It was caused by ladders ending at a small lambda1/v0 (e.g. 16): an edge needs roughly sum_g |theta_g| > (G log r - logit(pi)) / (r - 1), r = lambda1/v0, to reach PIP 0.5. Below that bar the pi update shrinks pi, which raises the bar again. The following changes make the EM robust to this and make it visible:

  • pi_delta and pi_xi are kept inside [1/N, 1 - 1/N], N = p(p-1)/2. pi_xi could previously reach exactly 0 and stay there (log(0) in the E-step), so problist2 was identically 0.
  • EM convergence now also requires the change in log(pi) to be below the new tol.pi (default 1e-3). Checking theta alone could stop a ladder step while pi was still falling by orders of magnitude per iteration.
  • ssjgl() warns when the last ladder step has no edge with PIP > 0.5 or pi_delta is at its floor.
  • New carry_pi = FALSE. If TRUE, pi starts each ladder step from the previous step’s value instead of the prior mean.
  • ?ssjgl gains a section on what the PIPs are (EM conditional probabilities that rank edges, not calibrated posterior probabilities) and on the collapse condition.

For fits that did not collapse, the final estimates are unchanged (to 1e-14 on the reference fits); some first ladder steps take one iteration more or fewer.

CRAN preparation

  • ssjgl() gains verbose = FALSE. Progress (each EM iteration on the R path, each ladder step, re-initialisation notices) is reported with message() only when verbose = TRUE. It was printed unconditionally with cat(). SSJGL_select_v0_cv(), SSJGL_final_with_pcor_CI() and SSJGL_CV_final_pcorCI() now default to verbose = FALSE and use message().
  • simulate_ssjgl_data(), SSJGL_select_v0_cv() and SSJGL_final_with_pcor_CI() no longer change the caller’s random-number stream: the RNG state is restored on exit after using seed.
  • plot_path() restores par() on exit, has dead code removed, and its example now runs. Its legend now says what is drawn (the old “Edge” / “Non-edge” legend referred to a true graph the function never had).
  • simdat is regenerated reproducibly with simulate_ssjgl_data() (data-raw/simdat.R) and now also contains the true precision matrices (Omega). The previous version came from a script that could not be run.
  • The unused fit1.Rdata files are removed, so the package no longer depends on R >= 3.5.0 through serialized data. It now depends on R >= 4.1.0.
  • New inst/CITATION (method paper and package), CITATION.cff and cran-comments.md. The DESCRIPTION has a fuller description citing Li et al. (2019), BugReports, and the GitHub URL. The Title is in title case.
  • GitHub Actions: R CMD check on macOS, Windows and Ubuntu (devel, release, oldrel-1), plus test coverage.
  • The profiling script moved from inst/profiling/ to scripts/ (not installed).

spikeyglass 0.0.0.9000

  • Added a NEWS.md file to track changes to the package.