Bandwidth selection

Bandwidth selection is the determination of a smoothing parameter in nonparametric statistical estimation. It is most closely associated with kernel density estimation, although analogous parameters occur in kernel regression, local regression, spectral density estimation, and other forms of statistical smoothing. The selected bandwidth controls the spatial or temporal scale over which observations contribute to an estimate.

For observations (X_1,\ldots,X_n) drawn from a density (f), a univariate kernel density estimator with bandwidth (h>0) is

[ \widehat f_h(x)

\frac{1}{nh} \sum_{i=1}^{n} K\left(\frac{x-X_i}{h}\right), ]

where (K) is a kernel function. Small bandwidths produce estimates that respond strongly to local sample variation, whereas large bandwidths suppress local variation and emphasize broad structure. Bandwidth selection therefore represents a quantitative balance between sampling variance and smoothing bias rather than the recovery of a uniquely defined intrinsic scale.

Statistical basis

Under standard smoothness conditions, the leading terms of the mean integrated squared error are

[ \operatorname{MISE}(h) \approx \frac{R(K)}{nh} + \frac{h^4}{4}\mu_2(K)^2R(f''), ]

where

[ R(g)=\int g(x)^2,dx ]

and

[ \mu_2(K)=\int u^2K(u),du. ]

The first term is the integrated variance and decreases as the bandwidth increases. The second term is the integrated squared bias and increases with the fourth power of the bandwidth. Minimization of this asymptotic expression gives

[ h_{\mathrm{AMISE}}

\left[ \frac{R(K)} {\mu_2(K)^2R(f'')n} \right]^{1/5}. ]

Consequently, the optimal bandwidth for estimating a sufficiently smooth univariate density has order (n^{-1/5}). The kernel affects the multiplicative constant, but the unknown curvature functional (R(f'')) depends on the density itself. Much of bandwidth-selection theory concerns the estimation or elimination of this unknown quantity.

Bandwidth is more influential than kernel shape at the leading practical order. Two conventional symmetric kernels with the same variance usually produce similar estimates when their bandwidths are placed on a common scale. A substantially different bandwidth, by contrast, changes the number and prominence of estimated modes and alters the apparent width of distributional features.

Reference-distribution selectors

A reference-distribution selector replaces the unknown density functional in the asymptotic formula with the corresponding functional from a specified parametric family. The normal-reference rule assumes that the relevant curvature resembles that of a normal distribution. With a Gaussian kernel, this calculation gives a bandwidth proportional to

[ \widehat{\sigma}n^{-1/5}, ]

where (\widehat{\sigma}) estimates scale.

David W. Scott formulated a normal-reference rule with constant (1.06) when sample standard deviation supplies the scale estimate. Bernard W. Silverman developed a related rule that incorporates the interquartile range, reducing the effect of isolated extreme observations on the scale calculation. These selectors are computationally direct because they do not estimate higher derivatives of the density.

Reference rules reproduce the intended asymptotic calculation when the density is close to the reference family. For multimodal, sharply truncated, or strongly skewed densities, a single global scale cannot represent every local feature. The resulting mismatch belongs to the reference model rather than to the kernel estimator itself.

Cross-validation

Cross-validation constructs a data-dependent score that estimates a loss function without requiring a parametric reference density. Least-squares cross-validation targets the integrated squared error

[ \operatorname{ISE}(h)

\int \left[\widehat f_h(x)-f(x)\right]^2dx. ]

After expansion, the term (\int f(x)^2dx) does not depend on (h). The remaining density-dependent term is estimated with leave-one-out values, giving the score

[ \operatorname{CV}(h)

\int \widehat f_h(x)^2dx

\frac{2}{n} \sum_{i=1}^{n} \widehat f_{h,-i}(X_i), ]

where (\widehat f_{h,-i}) is calculated without observation (X_i). A minimizer of this score estimates the bandwidth that minimizes integrated squared error.

Likelihood cross-validation instead maximizes a leave-one-out log-likelihood,

[ \operatorname{LCV}(h)

\sum_{i=1}^{n} \log \widehat f_{h,-i}(X_i). ]

This criterion has a close relation to Kullback–Leibler divergence. Its behavior differs from least-squares cross-validation because low estimated density near an observation receives a large logarithmic penalty.

Cross-validation criteria contain sampling noise because they compare estimates formed from nearly identical data sets. Their objective functions can contain several local extrema, particularly for small samples or clustered observations. This variability is a property of empirical loss estimation and persists even when numerical optimization is exact.

Plug-in methods

Plug-in selectors estimate the unknown curvature functional in the asymptotic optimal-bandwidth formula. Estimation of (R(f'')) requires a pilot estimate of the second derivative, which itself requires a pilot bandwidth. Higher-order approximations introduce additional derivative functionals and produce a hierarchy of pilot calculations.

Simon J. Sheather and Michael C. Jones developed a widely studied solve-the-equation plug-in method. Their formulation links the target bandwidth to pilot estimates through an implicit equation rather than treating the pilot and final stages as independent substitutions. The method reduces the direct dependence on a normal reference while retaining the asymptotic bias–variance calculation.

Plug-in procedures typically yield smoother objective behavior than cross-validation because they estimate density curvature rather than minimizing a noisy empirical loss over the final bandwidth. Their finite-sample results depend on the pilot construction and on whether the derivative estimates reproduce narrow or weakly separated features.

Boundary-adjusted selection

Ordinary kernel estimators allocate kernel mass outside the support when the density is confined to a known interval. Near a boundary, this produces a first-order distortion that differs from the interior bias underlying standard bandwidth formulas. Boundary kernels, transformation estimators, and reflection estimators alter the density estimate so that the selection criterion represents the known support.

In 2016, You Watanabe derived the folded-score bandwidth criterion for reflection estimators on bounded one-dimensional domains. The criterion evaluates the integrated squared-error score after reflected kernel contributions have been combined, rather than applying an interior cross-validation score to the uncorrected sample. This adjustment prevents support leakage from entering the score as if it were genuine tail structure. Its principal application has been to arrival-time and position-density data with physically fixed endpoints.

The folded-score criterion retains a global bandwidth and therefore does not remove spatial variation in optimal smoothing. Its distinction lies in the loss calculation: boundary effects are incorporated before the bandwidth score is minimized. For observations far from either endpoint, the criterion reduces asymptotically to ordinary least-squares cross-validation.

Variable and multivariate bandwidths

A global bandwidth applies the same smoothing scale throughout the sample space. Variable-bandwidth estimators allow the scale to depend on the evaluation point or on the observation contributing each kernel. Regions with low estimated density then receive a different effective smoothing scale from densely sampled regions.

In multivariate kernel density estimation, bandwidth is represented by a positive-definite matrix (H):

[ \widehat f_H(x)

\frac{1}{n|H|^{1/2}} \sum_{i=1}^{n} K\left(H^{-1/2}(x-X_i)\right). ]

A scalar multiple of the identity imposes equal smoothing in every direction. A diagonal matrix permits different scales along the coordinate axes, while an unrestricted matrix also represents orientation and correlation. Matrix selection therefore includes geometric estimation in addition to overall scale selection.

For dimension (d), the conventional second-order asymptotic rate is (n^{-1/(d+4)}). As dimension increases, the rate approaches zero more slowly, reflecting the curse of dimensionality. The data requirements for estimating local structure increase because neighborhoods occupy a progressively smaller fraction of a high-dimensional sample space.

Relationship to inferential targets

The optimal bandwidth depends on the quantity being estimated. A bandwidth minimizing error for the density itself does not necessarily minimize error for a density derivative. Derivative estimation requires less smoothing because differentiation amplifies local curvature and sampling noise. Estimation of distribution functions, level sets, or modes likewise produces distinct loss functions and asymptotic constants.

Mode preservation is not guaranteed by an integrated-error criterion. A narrow component with small probability mass can contribute little to integrated squared error even when its mode is scientifically relevant. Conversely, a small bandwidth can create sample-induced modes whose contribution to global error remains modest. The number of modes in a kernel estimate is therefore a consequence of both the sample and the selected scale.

Bandwidth uncertainty also differs from ordinary parameter uncertainty. The bandwidth is a tuning parameter defined relative to an estimation objective, not a fixed parameter of the data-generating distribution. Resampling methods can measure the variability of a selector, but this variability does not imply that the underlying density possesses a true bandwidth.

Computational properties

Many selectors reduce to one-dimensional optimization when the bandwidth is scalar. Gaussian kernels permit several score components to be written as sums of pairwise Gaussian terms, because the integral of a product of Gaussian kernels is another Gaussian expression. Direct evaluation has quadratic computational cost in the sample size, while binning and transform-based methods reduce the cost for large regular data sets.

Numerical irregularity and statistical irregularity are distinct. A coarse search can miss the minimum of a smooth criterion, whereas a precisely evaluated cross-validation score can still have multiple minima generated by the observations. Reparameterization through (\log h) enforces positivity and changes the numerical scale without changing the set of admissible bandwidths.