Skip to contents

Changelog

rbbnp 1.2.0

This release introduces a smaller, conventional public API and the reproducibility infrastructure for the JSS software paper.

Interface

  • Added bias_bound_density() and bias_bound_regression() as the two primary estimators. Both use lowercase x/y, a single bw argument for fixed or automatically selected bandwidths, conf_level, a validated kernel argument, and a named control list for advanced settings.
  • Retained biasBound_density() and biasBound_condExpectation() as deprecation shims. Formerly exported implementation helpers are now internal.
  • Made the Fourier-envelope grid resolution explicit as control$envelope_points_per_log instead of an undocumented literal.
  • confint() now rejects a confidence level different from the fitted level instead of silently ignoring it.
  • Vector-valued eval inputs now return one estimate, standard error, and interval per requested evaluation point for both primary estimators.

Data and diagnostics

  • Replaced duplicate sample_data files with the descriptively named two_fold_uniform_example dataset and matching Stata file.
  • Removed the unused DATA_PATH and EXT_DATA_PATH namespace objects.
  • Made the internal mixed simulation design a genuine random mixture.

Validation

  • Added a full benchmark against h/4 undersmoothing, nprobust, np, and sm over four designs and three sample sizes. The results document the intended tradeoff: average pointwise coverage for the bias-bound intervals ranges from 0.991 to 1.000, with a smallest point-specific coverage of 0.945. They are generally wider, especially in regression, with one near-equal density comparison at the largest sample size.

  • Corrected the FFT density cross-validation objective and added direct-versus- FFT agreement tests, including the Old Faithful density sample. The previous fast path could select a bandwidth far from the exact least-squares-CV minimizer; separate tests verify that regression CV responds to y.

  • Regression cross-validation now minimizes leave-one-out Nadaraya–Watson prediction error and therefore uses the response. Previously the automatic regression selector incorrectly reused density CV on the predictor alone.

  • smoothness_floor = TRUE now enforces the full theoretical restriction r in {2, 3, ...} with a mixed-integer envelope solve. The earlier continuous-slope behavior is available only as an explicit diagnostic extension and is not covered by Schennach’s 2020 theorem.

  • Regression marginal and numerator Fourier windows are selected separately, matching the two applications of the theory. The exact Theorem 2 window now errors with guidance when its feasible set is empty instead of substituting an unselected upper cap.

  • Corrected regression standard errors to the conventional Nadaraya–Watson ratio scale and fixed stored conf_level metadata to contain the fitted confidence level rather than the tail probability.

  • Restored Lemma 3’s deliberately unbounded regression intervals when a bias-adjusted denominator reaches zero. This supersedes the 1.1.0 release note’s overbroad statement that regression bands were always finite.

  • Empty control = list() values now correctly request all documented defaults.

  • Restored the standard tests/testthat.R runner so source-package checks execute the complete test suite.

  • Stabilized the Epanechnikov Fourier transform at and near zero using its Taylor expansion; the former closed-form evaluation returned NaN at zero and suffered catastrophic cancellation nearby.

  • Exposed the envelope solver’s former fixed upper exponent 50 as control$smoothness_max and made the LP and fallback grid use the same nonnegative range.

  • The signal-to-noise window threshold now uses the smooth form tau_n = sqrt(2 * log(exp(2) + n)) in place of max(2, sqrt(2 * log(n))). The two agree to within 0.05 percent for n at or above 1000 and select identical windows on realistic data, so coverage and width are unchanged; the new form only regularises the small-n behaviour smoothly (tau -> 2 as n -> 0) and removes the hard max() floor. Following referee feedback, the window rule is now documented as a calibrated heuristic rather than a strict bound.

rbbnp 1.1.0

CRAN release: 2026-06-09

This release improves how the frequency window for smoothness estimation is chosen. By default the confidence bands are now narrower, and regression bands are always finite. The previous behaviour is still available (see the note below).

Highlights

  • A new default window rule (methods_get_xi = "snr"). The upper cutoff frequency is now chosen from the signal-to-noise ratio of the empirical Fourier transform. The original rule of Schennach (2020) relies on a worst-case bound that selects no frequency at realistic sample sizes, so the package used to fall back to a wide window that gave a flat smoothness envelope and very wide bands. The new rule always returns a usable window and produces confidence intervals roughly 0.4 times as wide.

  • More accurate regression bias bounds. Two corrections are now on by default. noise_floor = "auto" uses the noise floor appropriate for a general response, and envelope_use_Y = TRUE fits the smoothness envelope to the joint spectrum of Y and X rather than the marginal spectrum of X, which had under-estimated the bias.

  • Finite bands for difficult cases (integer_r = TRUE). When the data do not show a clear power-law decay, the fitted slope can fall below the minimum the method assumes. It is now raised to that minimum and the amplitude refit, which keeps the bias bound finite. This avoids the very wide bands that could otherwise appear for very smooth densities or polynomial conditional means. A warning is shown when this adjustment is made.

  • Refreshed plots. Density, regression, and Fourier-transform plots now use a cleaner theme with a colorblind-friendly palette and a legend. The Fourier-transform plot shows a wider frequency range with the selected window shaded, so you can see where the signal gives way to noise; the new xi_range and expand arguments control the displayed range. Custom colors through fill_ci and fill_bias still work.

Notes

  • To reproduce the previous behaviour, set methods_get_xi = "Schennach_loose", noise_floor = "compact", and envelope_use_Y = FALSE.

rbbnp 1.0.0

First stable release with a modern S3 interface, performance improvements, and expanded documentation.

Highlights

  • Stable S3 API for density and regression objects (bbnp_density, bbnp_regression).
  • Bias-bound inference with MSE-optimal bandwidths (no undersmoothing), following Schennach (2020).
  • Faster bandwidth selection (FFT-based CV) and improved A/r estimation (LP-based envelope).
  • Expanded vignettes and pkgdown website.

rbbnp 0.3.0 (development)

CRAN release: 2025-04-30

This release focuses on modernizing the package interface and improving performance.

Highlights

  • Modern S3 interface: biasBound_density() and biasBound_condExpectation() now return S3 objects (bbnp_density, bbnp_regression) with standard methods: print(), summary(), plot(), coef(), confint(), and fitted() (regression).

  • More faithful implementation of Schennach (2020):

    • Improved estimation of smoothness parameters (A, r) using a linear-programming formulation.

  • Performance improvements:

    • Faster bandwidth cross-validation via an FFT-based approach for larger samples.
    • Vectorized conditional variance computation to reduce bottlenecks.

  • Docs & usability:

    • Expanded vignettes and pkgdown website organization.

Notes

  • Some internal functions were refactored; most user-facing APIs remain backward compatible.