Semiparametric Bandwidth Selectors

• Semiparametric bandwidth selectors are methods that bridge fully nonparametric and parametric approaches by estimating unknown curvature functionals using structured pilot models.
• They are implemented in diverse settings, including plug-in selectors for circular densities, pooled kernel estimators in multivariate models, and Bayesian bandwidth inference in mixture models.
• These techniques improve smoothing accuracy by integrating pilot-based adjustments and semiparametric likelihoods, even as optimality proofs and full consistency remain challenging.

Semiparametric bandwidth selectors are procedures for choosing smoothing parameters that occupy an intermediate position between “very nonparametric” selectors and “very parametric” reference rules. Taken together, the literature indicates at least three recurring usages of the term: selectors that estimate unknown risk functionals by means of a structured pilot model; selectors embedded in semiparametric models in which the bandwidth interacts with finite- and infinite-dimensional components; and Bayesian treatments in which a bandwidth-like scale parameter is itself the object of semiparametric inference rather than the output of a classical smoothing rule (Hjort, 13 Feb 2026, Oliveira et al., 2012, Shang, 2020, Kleijn, 2013).

1. Conceptual scope

The most explicit recent attempt to define the field treats semiparametric bandwidth selection as the “middle ground” between cross-validation or other “very nonparametric” procedures and normal-reference rules or other “very parametric” procedures. In that formulation, the selector is built by approximating an auxiliary density through short Hermite expansions around the normal distribution and then minimizing an exact MISE-based criterion (Hjort, 13 Feb 2026). In classical smoothing language, this is a selector for kernel density estimation; in methodological language, it is a low-dimensional correction to a parametric reference rule.

A second usage appears in plug-in bandwidth selection for circular density estimation. There the final estimator is fully nonparametric, but the unknown curvature term in the AMISE is estimated from a fitted finite mixture of von Mises distributions. The procedure is therefore semiparametric in spirit: the target density is not constrained to lie in the parametric family, yet the bandwidth is chosen by plugging in quantities extracted from that family (Oliveira et al., 2012).

A third usage arises when the bandwidth is not merely a smoothing hyperparameter but part of a semiparametric model. In the multivariate density-ratio setting, the bandwidth enters a pooled kernel estimator derived from constrained empirical likelihood, so the AMISE constant depends on semiparametric pooling across samples rather than on a single-sample kernel variance alone (Voulgaraki et al., 2010). In functional partial linear regression, the regression bandwidth $h$, the residual-density bandwidth $b$, and even localized residual-density bandwidths $(\tau,\tau_\varepsilon)$ are selected inside a model that combines a functional linear component with a functional nonparametric component (Shang, 2020).

The term also has an important negative boundary. In “Semiparametric posterior limits,” the bandwidth example concerns the Gaussian kernel scale parameter $\sigma$ in a normal location mixture, but the paper is explicit that it is not constructing a selector “like cross-validation or plug-in AMISE methods.” Its contribution is asymptotic Bayesian inference for a bandwidth parameter, not a classical bandwidth-selection algorithm (Kleijn, 2013).

2. Risk expansions and the nuisance-functionals problem

The common mathematical core of semiparametric bandwidth selectors is the replacement of an unknown curvature or roughness functional by an estimate that uses some structured information. In circular kernel density estimation with von Mises kernels, the relevant criterion is the asymptotic mean integrated squared error

$$AMISE(\nu)= \frac{1}{16}\left[1-\frac{I_2(\nu)}{I_0(\nu)}\right]^2 \int_{0}^{2\pi} \left[f^{\prime\prime}(\theta)\right]^2 d\theta + \frac{I_0(2\nu)}{2n\pi \left(I_0(\nu)\right)^2},$$

so the unknown nuisance quantity is $\int_0^{2\pi}[f''(\theta)]^2\,d\theta$ (Oliveira et al., 2012).

For length-biased data, the corrected kernel estimator has the AMISE

$$\operatorname{AMISE}(h) = \frac{1}{4}h^4\mu_2^2(K)R(f'') + \frac{R(K)\mu c}{nh},$$

with $c=\int y^{-1}f(y)\,dy$. The rate remains the familiar $n^{-1/5}$, but the variance constant is modified by the length-bias mechanism through $\mu c$ (Borrajo et al., 2016).

In the multiple-sample semiparametric density-ratio model, the MISE expansion retains the standard order structure,

$b$0

hence $b$1, but the variance constant is changed to

$b$2

which reflects the efficiency gain from pooling information across samples (Voulgaraki et al., 2010).

The 2026 Hermite-expansion framework reformulates the problem at a different level. It writes

$b$3

for the difference density and shows that bandwidth selection can be recast through an exact MISE identity involving $b$4. In that framework, the decisive asymptotic quantity is $b$5, so estimating bandwidth well becomes equivalent, at leading order, to estimating the curvature functional through a structured approximation to the difference density (Hjort, 13 Feb 2026).

Taken together, these constructions suggest that semiparametric bandwidth selection is less about altering the bias–variance tradeoff itself than about changing how the unknown constants in that tradeoff are estimated.

3. Reference-model and pilot-based constructions

The circular plug-in selector is a canonical semiparametric construction. The estimator of interest is

$b$6

and the selector $b$7 is defined by fitting a finite mixture of von Mises distributions

$b$8

using AIC to choose $b$9, maximum likelihood via the EM algorithm to estimate $(\tau,\tau_\varepsilon)$0, numerical quadrature to compute $(\tau,\tau_\varepsilon)$1, and numerical minimization of $(\tau,\tau_\varepsilon)$2 (Oliveira et al., 2012). The procedure is especially motivated by multimodality, antipodal modes, skewness, and peakedness, precisely the settings in which a single-von-Mises reference can be misleading. In the reported real-data examples, computation times were about $(\tau,\tau_\varepsilon)$3 seconds for likelihood cross-validation and $(\tau,\tau_\varepsilon)$4 seconds for the plug-in selector (Oliveira et al., 2012).

The Hermite-expansion program generalizes the same plug-in logic beyond a single reference family. It models the difference density by

$(\tau,\tau_\varepsilon)$5

using only even Hermite terms because $(\tau,\tau_\varepsilon)$6 is symmetric. When all correction coefficients vanish except the Gaussian base term, the construction collapses to the normal-reference rule; with nonzero coefficients, it becomes an “extended rule of thumb” that corrects normal reference by a small number of estimated parameters (Hjort, 13 Feb 2026). The same paper derives an explicit correction for the roughness functional $(\tau,\tau_\varepsilon)$7, which makes the selector interpretable as a low-dimensional modification of the normal-reference value (Hjort, 13 Feb 2026).

Length-biased kernel density estimation offers another spectrum of pilot-based selectors. The rule-of-thumb replaces $(\tau,\tau_\varepsilon)$8 by its Gaussian-reference value, whereas the bootstrap selectors $(\tau,\tau_\varepsilon)$9 and $\sigma$0 replace $\sigma$1 by $\sigma$2 from a pilot density estimator, with optimal pilot order $\sigma$3 in the higher-order theory (Borrajo et al., 2016). This is a textbook semiparametric pattern: the final estimator remains nonparametric, but the bandwidth is chosen through a structured or pilot-based approximation to an otherwise intractable curvature functional.

4. Bandwidths inside semiparametric models

In the multivariate multiple-sample density-ratio model, the semiparametric structure is

$\sigma$4

with pooled estimation via constrained empirical likelihood. The corresponding kernel estimator

$\sigma$5

uses the whole pooled dataset, and its optimal bandwidth has the standard rate $\sigma$6 but a semiparametrically improved variance constant (Voulgaraki et al., 2010). Practical bandwidth choice is handled by a plug-in rule and by two cross-validation criteria, $\sigma$7 and $\sigma$8, and the paper explicitly allows coordinatewise bandwidths $\sigma$9 in implementation (Voulgaraki et al., 2010).

In the functional partial linear model

$$AMISE(\nu)= \frac{1}{16}\left[1-\frac{I_2(\nu)}{I_0(\nu)}\right]^2 \int_{0}^{2\pi} \left[f^{\prime\prime}(\theta)\right]^2 d\theta + \frac{I_0(2\nu)}{2n\pi \left(I_0(\nu)\right)^2},$$0

the bandwidth $$AMISE(\nu)= \frac{1}{16}\left[1-\frac{I_2(\nu)}{I_0(\nu)}\right]^2 \int_{0}^{2\pi} \left[f^{\prime\prime}(\theta)\right]^2 d\theta + \frac{I_0(2\nu)}{2n\pi \left(I_0(\nu)\right)^2},$$1 enters the functional Nadaraya–Watson estimator for the nonparametric component, but it also influences the estimate of the linear component through the partialing-out matrix $$AMISE(\nu)= \frac{1}{16}\left[1-\frac{I_2(\nu)}{I_0(\nu)}\right]^2 \int_{0}^{2\pi} \left[f^{\prime\prime}(\theta)\right]^2 d\theta + \frac{I_0(2\nu)}{2n\pi \left(I_0(\nu)\right)^2},$$2. The paper then estimates $$AMISE(\nu)= \frac{1}{16}\left[1-\frac{I_2(\nu)}{I_0(\nu)}\right]^2 \int_{0}^{2\pi} \left[f^{\prime\prime}(\theta)\right]^2 d\theta + \frac{I_0(2\nu)}{2n\pi \left(I_0(\nu)\right)^2},$$3 jointly with the residual-density bandwidth $$AMISE(\nu)= \frac{1}{16}\left[1-\frac{I_2(\nu)}{I_0(\nu)}\right]^2 \int_{0}^{2\pi} \left[f^{\prime\prime}(\theta)\right]^2 d\theta + \frac{I_0(2\nu)}{2n\pi \left(I_0(\nu)\right)^2},$$4 by means of the leave-one-out kernel pseudo-likelihood

$$AMISE(\nu)= \frac{1}{16}\left[1-\frac{I_2(\nu)}{I_0(\nu)}\right]^2 \int_{0}^{2\pi} \left[f^{\prime\prime}(\theta)\right]^2 d\theta + \frac{I_0(2\nu)}{2n\pi \left(I_0(\nu)\right)^2},$$5

and extends the construction to localized residual-density bandwidths $$AMISE(\nu)= \frac{1}{16}\left[1-\frac{I_2(\nu)}{I_0(\nu)}\right]^2 \int_{0}^{2\pi} \left[f^{\prime\prime}(\theta)\right]^2 d\theta + \frac{I_0(2\nu)}{2n\pi \left(I_0(\nu)\right)^2},$$6 (Shang, 2020). The same framework uses marginal likelihood to select the semi-metric that defines neighborhoods among curves; for smooth functional data the second-derivative semi-metric performed best, while for rough functional data FPCA-based semi-metrics were often preferable (Shang, 2020).

A contemporary applied extension appears in semiparametric Bayesian networks, where some conditional probability distributions are linear-Gaussian and others are conditional KDEs. There the practical question is which multivariate bandwidth-matrix selector to use for the nonparametric CKDE nodes. The paper studies the normal rule, unbiased cross-validation, smooth cross-validation, and plug-in selection inside that semiparametric architecture, reporting that unbiased cross-validation generally outperforms the normal rule in high-sample-size scenarios, whereas the normal rule remains robust but tends to oversmooth and stagnate with more data (Alejandre et al., 20 Jun 2025).

5. Bayesian and posterior interpretations

The Bayesian literature introduces a different meaning of semiparametric bandwidth selection. In the normal location mixture

$$AMISE(\nu)= \frac{1}{16}\left[1-\frac{I_2(\nu)}{I_0(\nu)}\right]^2 \int_{0}^{2\pi} \left[f^{\prime\prime}(\theta)\right]^2 d\theta + \frac{I_0(2\nu)}{2n\pi \left(I_0(\nu)\right)^2},$$7

the parameter of interest is the Gaussian kernel scale $$AMISE(\nu)= \frac{1}{16}\left[1-\frac{I_2(\nu)}{I_0(\nu)}\right]^2 \int_{0}^{2\pi} \left[f^{\prime\prime}(\theta)\right]^2 d\theta + \frac{I_0(2\nu)}{2n\pi \left(I_0(\nu)\right)^2},$$8, restricted to a compact interval $$AMISE(\nu)= \frac{1}{16}\left[1-\frac{I_2(\nu)}{I_0(\nu)}\right]^2 \int_{0}^{2\pi} \left[f^{\prime\prime}(\theta)\right]^2 d\theta + \frac{I_0(2\nu)}{2n\pi \left(I_0(\nu)\right)^2},$$9, and the nuisance parameter is the unknown mixing distribution $\int_0^{2\pi}[f''(\theta)]^2\,d\theta$0 (Kleijn, 2013). The paper treats $\int_0^{2\pi}[f''(\theta)]^2\,d\theta$1 explicitly as the kernel variance or bandwidth parameter and formulates the conjectured semiparametric Bernstein–von Mises limit

$\int_0^{2\pi}[f''(\theta)]^2\,d\theta$2

Its practical message is inferential rather than algorithmic: if the conjecture holds, posterior credible intervals for $\int_0^{2\pi}[f''(\theta)]^2\,d\theta$3 behave asymptotically like efficient confidence intervals, even though the efficient Fisher information is not available in closed form (Kleijn, 2013).

The functional partial linear model also adopts a Bayesian bandwidth view, but with a pseudo-likelihood rather than a posterior limit theorem. The priors $\int_0^{2\pi}[f''(\theta)]^2\,d\theta$4 and an adaptive random-walk Metropolis algorithm deliver posterior means for $\int_0^{2\pi}[f''(\theta)]^2\,d\theta$5, $\int_0^{2\pi}[f''(\theta)]^2\,d\theta$6, $\int_0^{2\pi}[f''(\theta)]^2\,d\theta$7, and $\int_0^{2\pi}[f''(\theta)]^2\,d\theta$8, yet the paper is explicit that the likelihood is a pseudo-likelihood and that credible sets may fail to be asymptotically valid (Shang, 2020).

For nonnegative-orthant data, the semiparametric estimator

$\int_0^{2\pi}[f''(\theta)]^2\,d\theta$9

uses associated kernels to smooth the unknown weight function $$\operatorname{AMISE}(h) = \frac{1}{4}h^4\mu_2^2(K)R(f'') + \frac{R(K)\mu c}{nh},$$0. Bandwidth matrices are then selected by adaptive Bayesian rules for semicontinuous supports and local Bayesian rules for count supports, with the semiparametric leave-one-out fit replacing the purely nonparametric criterion inside the posterior update (Kokonendji et al., 2021). The same paper introduces the diagnostic

$$\operatorname{AMISE}(h) = \frac{1}{4}h^4\mu_2^2(K)R(f'') + \frac{R(K)\mu c}{nh},$$1

which is used to decide empirically between parametric, semiparametric, and nonparametric modelling (Kokonendji et al., 2021).

6. Boundaries of the concept and recurring issues

Several neighboring literatures clarify what semiparametric bandwidth selectors are not. The multivariate level-set and highest-density-region selector is a target-specific, asymptotic plug-in rule minimizing symmetric-difference risk for sets, but the paper is explicit that it is “not semiparametric in the standard model-theoretic sense” (Doss et al., 2018). The same caution applies to the posterior-limit treatment of $$\operatorname{AMISE}(h) = \frac{1}{4}h^4\mu_2^2(K)R(f'') + \frac{R(K)\mu c}{nh},$$2 in normal location mixtures: it is asymptotic semiparametric inference for a bandwidth parameter, not a proposal of $$\operatorname{AMISE}(h) = \frac{1}{4}h^4\mu_2^2(K)R(f'') + \frac{R(K)\mu c}{nh},$$3, $$\operatorname{AMISE}(h) = \frac{1}{4}h^4\mu_2^2(K)R(f'') + \frac{R(K)\mu c}{nh},$$4, or $$\operatorname{AMISE}(h) = \frac{1}{4}h^4\mu_2^2(K)R(f'') + \frac{R(K)\mu c}{nh},$$5 (Kleijn, 2013).

A second boundary concerns low-dimensional targets estimated by nonparametric means. In sharp and fuzzy regression discontinuity, the proposed selectors jointly choose left and right bandwidths for local linear estimators of a discontinuity or a local Wald ratio, and higher-order bias terms become essential when first-order bias can be eliminated by appropriate bandwidth relations (Arai et al., 2014, Arai et al., 2015). These methods are semiparametric in the sense of targeting a low-dimensional parameter through nonparametric nuisance estimation, but their logic is local-polynomial bias cancellation rather than reference-model-based curvature estimation.

A third boundary is methodological. Sequential cross-validation for locally stationary processes and for dependent monitoring problems develops asymptotically optimal bandwidth selection under dependence, with one-sided or global CV criteria and functional limit theorems, yet these papers are better read as bandwidth theory for dependent-data local M-estimation and change detection than as part of the classical semiparametric plug-in tradition (Richter et al., 2017, Steland, 2012).

The literature also shares several unresolved issues. Formal asymptotic optimality is not universal: the circular plug-in selector is methodologically strong but does not prove consistency of $$\operatorname{AMISE}(h) = \frac{1}{4}h^4\mu_2^2(K)R(f'') + \frac{R(K)\mu c}{nh},$$6 (Oliveira et al., 2012); the functional partial linear model emphasizes empirical performance rather than selector optimality theorems (Shang, 2020); the Hermite-expansion program deliberately stops short of recommending a single final procedure among many possible constructions (Hjort, 13 Feb 2026); and the semiparametric posterior-limit treatment of the mixture bandwidth is explicitly a conjecture because of missing proofs for approximately least-favourable submodels and marginal parametric-rate concentration (Kleijn, 2013). A further practical caution is that good componentwise bandwidths need not optimize the final target: in the two-bandwidth Nadaraya–Watson decomposition, the ratio of two individually good estimators can be worse than a direct single-bandwidth cross-validated estimator in small-noise settings, even though the two approaches become similar in large-noise models (Comte et al., 2020).

Taken together, these works suggest that semiparametric bandwidth selectors are best understood as a heterogeneous family united by one principle: bandwidth choice is informed by more structure than a purely nonparametric criterion, but less structure than a fully parametric model. The structure may enter through a parametric pilot, a semiparametric likelihood, a pooled density-ratio model, a Bayesian prior on a bandwidth-like parameter, or a structured predictive criterion. The resulting selectors differ sharply in target, theory, and interpretation, but they share the aim of improving smoothing by exploiting partial information that standard fully nonparametric selectors leave unused.