Changelog

0.1.6 (2026-11-04)

Changed (breaking)

  • DataStream.compute_statistics() now always computes the standard error of the mean as sd(block_means) / sqrt(ess_blocks), where ess_blocks is the Geyer effective sample size of the block means. In 0.1.5 it divided by the raw number of blocks whenever the window came from the Ljung-Box search, whether or not the test passed. That test has low power for typical block counts and often passes while the block means are still correlated, so the standard error could be too small. mean_uncertainty and the confidence intervals are therefore larger for almost every call. The means do not change.

  • The default block window is now w = round(tau_int) (the “tau-window” rule) instead of the Ljung-Box search. This changes window_size, n_short_averages and ess_blocks. independence_status is now "tau_window". The Ljung-Box results are still reported, but no longer affect the window or the standard error.

  • se_method is now always "ess_blocks". "iid_blocks", "iid_blocks_best_p" and "ess_blocks_fallback" are no longer returned.

  • Ensemble statistics follow the same approach:

    • Pooled block means (T1) always use the ESS-corrected standard error (se_method="sem_ess", previously "sem_n" when all members passed Ljung-Box).

    • The ensemble average (T0) and inverse-variance weighted member means (T2) by default use a block window based on a correlation time pooled over all members (tau_mode="pooled").

  • MeanVariationTrimStrategy now uses the same defaults as RobustWorkflow: decor_multiplier=6.0 (was 2.0), fudge_fac=0.1 (was 1e-6) and final_smoothing_window=10 (was 5). This changes the detected SSS start when the strategy is created without these arguments, including build_trim_strategy("mean_variation") and DataStream.trim(method="mean_variation"). RobustWorkflow results do not change.

  • The regression baselines in tests/*/expected/ were regenerated.

Added

  • DataStream.compute_statistics(window_rule=...): "tau" (default) or "autotune" for the legacy Ljung-Box window search. Both rules use the ESS-based standard error. An explicit window_size overrides the rule.

  • Ensemble statistics: tau_mode="pooled"|"per_member" for T0 and T2 (use "per_member" when members really differ, e.g. parameter scans), and pooled_ess_method="concat"|"sum" for T1. T0 metadata also reports tau_mode and tau_star.

  • quends.base.utils.tau_window_blocks() and quends.base.utils.pooled_tau_int() (pooled correlation time for homogeneous ensembles).

  • quends.MEAN_VARIATION_DEFAULTS: read-only mapping of the start-of-SSS detection defaults, shared by MeanVariationTrimStrategy and RobustWorkflow.

Fixed

  • Tau-window results no longer get the misleading warning “Too few blocks for independence test; SE via Geyer ESS.”.

  • _process_column() now rejects an unknown window_rule even when an explicit window is given.

Documentation

  • New README section “Detecting the Start of Statistical Steady State”: when to use RobustWorkflow, how the mean_variation heuristic works, the verbosity=2 diagnostic plots, and how to tune its hyperparameters.

  • The DataStream.compute_statistics(), Ensemble, ensemble statistics and ensemble statistics workflow docstrings now describe the tau-window blocking and the ESS-based standard error.

Security

  • Added pillow>=13.0.0 as a dependency.

  • Updated locked dependencies to patched releases: Black 26.10.1, certifi 2026.10.23, filelock 4.2.7, Pillow 13.0.0, Tornado 6.6, and virtualenv 21.15.2.

0.1.5 (2026-10-15)

Changed (breaking)

  • The minimum supported Python version is now 3.12 (previously 3.8). Python 3.8 through 3.11 are no longer supported.

Development

  • Updated the dependency lockfile for Python 3.12 and later.

  • Updated the test and documentation CI workflows to run on Python 3.12.

0.1.4 (2026-10-08)

Removed (breaking)

  • The max_lag_frac and autocorr_sig_level arguments were removed from RobustWorkflow and MeanVariationTrimStrategy. They no longer had any effect on the decorrelation time, which is computed by DataStream.compute_decorrelation_time(). Passing either argument now raises a TypeError. Remove them from existing calls.

Changed

  • In MeanVariationTrimStrategy, the smoothing window used to detect the start of statistical steady state is now decor_multiplier * tau_int, capped at the signal length (and at least 3 points). Before, it was also capped at max_lag_frac * n_pts. With the default decor_multiplier this can give a wider smoothing window, and so a different (often later) SSS start time and different statistics than in 0.1.3.

  • With verbosity > 1, the autocorrelation plot now covers 1.5 * tau_int lags (capped at the signal length), with a dashed vertical line at the decorrelation time and a legend.

Documentation

  • Added a full parameter description to the MeanVariationTrimStrategy docstring, and fixed the trim() parameter docs.

  • Fixed the documented default of smoothing_window_correction in RobustWorkflow (0.5, not 0.8) and clarified decor_multiplier.

  • Removed the obsolete arguments from the tutorial notebooks and the robustworkflow_advanced_guide.py example script.

0.1.3 (2026-10-07)

Changed

  • The integrated autocorrelation time (tau_int) estimate now extends the autocorrelation lag window when needed. The ACF is first computed up to min(n // 4, 2000) lags. If the Geyer positive-pair stopping criterion is not reached, the number of lags is doubled repeatedly, up to the full series length (n - 1).

  • As a result, tau_int can be larger, and the effective sample size smaller, than in 0.1.2 for strongly correlated series whose initial lag window was too short.

Fixed

  • RobustWorkflow.process_data_stream() no longer raises IndexError: single positional indexer is out-of-bounds when start_time is later than the last valid data point. In that case it now returns NaN statistics if operate_safe=True. If operate_safe=False, it returns an ad-hoc estimate over the tail of the data stream, set by no_sss_tail_fraction. Both cases report the new status "StartTimeBeyondData".

Warnings

  • A UserWarning saying that tau_int is potentially under-estimated is now raised only when the Geyer criterion is not reached even with the full-length ACF. The same warning appears in the compute_statistics() result metadata.

  • This replaces the previous “decorrelation time is large compared to the max lag” warning, which was based on a tau_int >= 0.5 * nlags rule.

Documentation

  • Updated the README section on publishing to PyPI to describe the automated GitHub Release workflow.

  • Synchronized __version__ in src/quends/__init__.py with the version in pyproject.toml.

0.1.2

0.1.1

0.1.0 (2025-02-28)

  • First release on PyPI.