Skip to contents

Bias-corrected bounds are experimental (pre-beta research), incomplete, and not recommended for general-purpose inference. See the experimental-status section below. Selects the research CDF construction or legacy bootstrap centering.

Usage

np_udist_bias_control(construction = c("joint-process", "cdf", "legacy"),
                      calibration = c("default", "score-envelope"))

Arguments

construction

The joint corrected CDF process or legacy correction.

calibration

The default construction or the explicit, side-specific score envelope for supported fixed numeric Gaussian asymptotic bands.

Details

This private research implementation is undergoing full empirical-range qualification. It does not establish a finite-sample coverage guarantee. With center="bias-corrected", omitting bias_control selects "joint-process". The default estimate center remains unchanged. Use errors="asymptotic" for Gaussian calibration or errors="bootstrap" with bootstrap="inid", "fixed" or "geom" for literal independent, circular fixed-length or stationary geometric-block resampling. Explicit construction="legacy" retains previous bootstrap correction.

The default joint-process construction admits fixed and generalized-nearest-neighbor unbounded second-order Gaussian kernels with at most two numeric and two ordered variables. Ordered smoothing retains the canonical categorical representation: nonnumeric ordered labels use ordinal ranks, while numeric labels retain their metric gaps. It uses the normalized Li–Racine kernel and the complete joint corrected contribution, retaining declared levels and integer gaps. It calibrates the maximum absolute corrected CDF displacement, with complete covariance and common observation innovations across target tiles. For fixed bandwidths, Gaussian and literal multinomial inid processes use the same centered observation factor. Critical values are in CDF probability units, not standardized units. Selected bandwidths or neighbor counts and categorical smoothing parameters stay fixed. With fixed bandwidths, fixed/geom processes use literal selected-law block counts and the corresponding finite-sample covariance; their reported scales are not IID scales. The automatic block length remains b.star(xdat, round=TRUE)[1,1], selected once per plot. The first input series may not represent dependence elsewhere in multivariate data. Explicit boot_control=np_boot_control(blocklen=...) overrides are honored. Neither method divides by a replicate studentizer. Full-length fixed blocks with zero process variance are unavailable. No local or replicate studentizer division, variance floor, trimming, endpoint clipping or resampling-law substitution is used. An entirely zero observed process is unavailable; zero marginal scales are reported explicitly.

For generalized nearest-neighbor bandwidths, Gaussian calibration retains the complete influence of the sample-specific query radii. Bootstrap calibration recomputes every physical radius and corrected center using the literal selected law, and uses their absolute displacements. The displayed corrected center retains its original summation; paired evaluations of the same algebraic sums ensure zero displacement for an unchanged sample. Reported bootstrap scales are empirical standard deviations under that selected law, including radius randomness. This construction has no replicate studentizing divisor and does not qualify the preceding spacing-pilot bootstrap-t method. If a bootstrap query radius collapses to zero, the Gaussian integrated kernel uses its exact one-sided limit: one below the query, one half at a tie, zero above it; its correction moment is zero. Positive-radius coordinates retain their original arithmetic. This is an extension of the bootstrap CDF statistic, with no draw discarded, replaced or floored. Observed zero radii and invalid Gaussian spacing pilots remain unavailable. Pure ordered NN inputs retain the fixed-geometry identity. Numeric adaptive NN remains parked at its population bias-remainder gate. These extensions require their own public-call coverage screen and do not inherit coverage from fixed bandwidths.

Gaussian marginal quantiles are analytic. Joint critical values use type-one quantiles bounded by the analytic marginal and Bonferroni maximum bounds. Inid marginal and joint quantiles use type one. Bonferroni intervals envelope the joint band. Singleton bands coincide. Standalone Gaussian Bonferroni uses the default joint draws for its envelope; request band="all" to set gaussian_draws and obtain all bands from the same process. Finite-draw Bonferroni tail resolution and marginal sparse-tail coverage remain limitations. Every positive draw count is supported when the provider arithmetic is defined. Poorly resolved empirical quantiles return all requested bands with one aggregated warning and retained inference[[1]]$resolution diagnostics. No draw count is increased and no requested band is omitted for resolution alone. Numeric GNN bootstrap scales are empirical standard deviations across draws, so fewer than two draws remains a distinct undefined-arithmetic error. Analytic Gaussian marginal quantiles do not require empirical tail resolution. The maximum-process draws are internal computation records, not plot output.

For numeric-only fixed bandwidths with one or two variables, Gaussian and inid pointwise bands additionally envelope a signed-contribution score interval. The contribution has a certified range: writing \(t=\sqrt{1+2/q}\), its lower bound is \(a=\Phi(-t)-(q/2)t\phi(t)\) and upper bound is \(1-a\). A Wilson score interval for the rescaled contribution mean is mapped back to this range. This is approximate inference for the fixed-bandwidth contribution expectation, not an exact confidence interval for the unsmoothed CDF. Data-selected bandwidths and evaluation ranges require separate empirical qualification. The inid interval is a hybrid marginal envelope: literal draws and quantiles are unchanged, but the final endpoints can be wider and asymmetric. The corrected center is unchanged. Simultaneous bands retain their raw process calibration and expand only by the minimum uniform amount required to contain the revised pointwise bands. Bonferroni bands retain nesting; singleton bands coincide. The provider retains the preceding intervals and criticals internally for construction and diagnostic tests; the plot result publishes only selected endpoints. Final critical values report maximum one-sided extents around the preserved center; use bands for the asymmetric endpoints. Settings record the contribution range and nesting expansion. Mixed, NN and block routes do not receive this marginal envelope. An entirely zero observed process remains unavailable.

For supported corrected empirical bootstrap bands, explicit boot_control=np_boot_control(calibration="separate-tails") uses signed lower and upper empirical tails in CDF units. The simultaneous roots are separate quantiles of the family minimum and maximum; endpoints invert those signed estimate displacements. calibration="tail-envelope" additionally takes the interval hull with the unchanged absolute construction. The default is unchanged, and calibration="absolute" requests it explicitly. These options preserve the selected inid, fixed or geom law, the complete contribution process, block choice, corrected center and displayed variance. They currently admit fixed and purely ordered NN geometry; numeric NN asymmetric calibration requires separate qualification. Density's "percentile-tails" choice is not a CDF method.

For fixed numeric Gaussian2 inputs with one or two variables, the inid signed bands retain the existing marginal score protection. Its two sides are nested outward separately, preserving asymmetry instead of imposing the larger uniform displacement on both sides. An explicit asymptotic option, bias_control=np_udist_bias_control(calibration="score-envelope"), uses this same side-specific composition with the unchanged Gaussian process. Gaussian draws remain reflection-symmetric; this option's possible asymmetry comes from the deterministic score envelope, and some bands can remain symmetric. Default pointwise numeric bands already include score protection. The new Gaussian option changes how that protection extends to the simultaneous and Bonferroni bands; it does not simulate a skewed Gaussian law.

Explicit asymmetric options require band="pointwise", "simultaneous", "bonferroni" or "all"; they do not silently alter normal-scale "pmzsd" intervals. Singleton bands coincide, and the declared family remains the full returned grid within each panel. Results retain absolute.intervals, raw.process.intervals, raw.intervals, signed tail orientation where applicable and side-specific nesting diagnostics. No [0,1] clipping or monotonicity projection is applied. These are private candidates requiring their own coverage evidence; containing a baseline interval is not a proof of nominal coverage.

Explicit construction="cdf" retains the preceding private studentized construction, including normalized Li–Racine ordered smoothing and bounded small-grid covariance calibration. Its bootstrap uses replicate scales and retains the extended degenerate-pivot contract. This explicit construction does not inherit qualification from the absolute process. Explicit "legacy" retains the historical bootstrap center. All constructions remain research candidates; full empirical-range screening is distinct from confirmation.

Omitting band, or selecting band="pmzsd", uses the corrected center plus or minus \(z_{1-\alpha/2}\) times the returned std_error. The existing pointwise provider supplies the selected-law scale, retaining block dependence and exact GNN resampling where applicable. Its computation and RNG contract are retained. Zero marginal scales give zero normal width. These normal intervals do not apply the score or joint envelopes and do not inherit their coverage results. Endpoints are in the interval rows with band="pmzsd". Explicit "pointwise", "simultaneous", "bonferroni" and "all" retain their calibrated constructions. Numeric adaptive nearest-neighbor bandwidths, unordered factors, bounded or higher-order kernels, large-bandwidth limits and factor boxplots remain unqualified for this private corrected route and fail clearly. Their legacy routes remain available. No explicit law is substituted.

The finite family is the returned grid within each panel. Above two input variables, ordinary plots return slices; their bands do not establish simultaneous coverage over the union of panels or a full multidimensional rectangle. Ordered and block extensions have separate qualification gates and do not inherit coverage evidence from the numeric IID construction.

Value

A family-specific control object consumed by plot methods. Data-returning plots use np_plot_data: original and corrected values, available standard errors, absolute endpoints and compact method and resolution diagnostics. Select a panel, scalar target and band with as.data.frame(result, panel=1, target=1, band="simultaneous"). Full fitted state, covariance workspaces, bootstrap draws and RNG states are internal computation records, not public plot fields. The fitted object is unchanged.

Experimental status of bias-corrected inference

All bias-corrected confidence intervals and bands are experimental (pre-beta research), incomplete, and not recommended for general-purpose inference. This status remains in effect unless the documentation explicitly states otherwise for a particular method and scope. It applies wherever bias-corrected bounds are available: asymptotic and bootstrap constructions, including legacy methods, pointwise intervals, Bonferroni intervals and simultaneous bands, for levels and supported derivatives.

These methods form an ongoing research project. Implementation support, numerical checks and favorable simulation results in selected settings do not establish generally valid coverage. Coverage can fail outside the studied settings, and available options do not constitute a general solution to bias-corrected inference. Method-specific assumptions, restrictions and known failures must be considered when experimenting with these bounds.

See also