Nuisance Tangent Space in Semiparametric Models

• Nuisance tangent space is a closed Hilbert subspace defined by the span of scores for nuisance parameters in semiparametric models.
• It enables the orthogonal projection of full score functions to extract efficient scores and derive the Fisher information bound.
• Its geometric structure unifies the treatment of finite- and infinite-dimensional nuisances, offering robust solutions to model deviations.

A nuisance tangent space is a closed subspace of a suitable Hilbert space associated with a semiparametric or robust statistical model, spanned by the scores for nuisance parameters—parameters not of direct inferential interest. The geometry of the nuisance tangent space underpins the construction of semiparametric efficient estimators and provides a unified language for handling both finite- and infinite-dimensional nuisance effects, including neighborhoods of model deviations and structural parameters. Key to semiparametric theory is the orthogonal projection of score functions onto this space, yielding the efficient score and Fisher information in the presence of nuisance.

1. Formal Definition and Theoretical Setting

Consider a semiparametric model

$$P = \{ p(x | \theta, \eta) : \theta \in \Theta \subset \mathbb{R}^q, \;\; \eta \in \mathcal{H}_\text{nuisance} \},$$

where $\theta$ are the parameters of interest and $\eta$ are nuisance parameters, possibly infinite-dimensional. All construction occurs in the Hilbert space $$H = \{ h : \mathcal{X} \to \mathbb{R} \mid E_0[h]=0, E_0[h^2]<\infty \}$$ with $\langle h_1, h_2\rangle = E_0[h_1 h_2]$ (Fortunati et al., 29 Jun 2025).

The full model tangent space at the true parameter value is the closure of the linear span of all score functions for both $\theta$ and $\eta$:

$$T_{\text{full}} = \overline{\mathrm{span}}\{ \text{score}_{\theta_1}, \dots, \text{score}_{\theta_q} ;\, \text{score}_\eta(f)\ \forall f \}.$$

The nuisance tangent space $T_{\text{nuisance}}$ is the closed linear span of all directions corresponding to $\eta$ alone.

For robust models, parameter deviations via neighborhoods (e.g., Hellinger, total variation, contamination) are also treated as infinite-dimensional nuisances; the associated tangent sets $\theta$0 (balls in $\theta$1) define the nuisance tangent space as $\theta$2 (Rieder, 2014).

2. Geometric Decomposition and Projection

Assuming mild regularity, $\theta$3 as a direct sum; that is, the space of interest-parameter perturbations and the nuisance tangent space are orthogonal and their intersection is trivial. This geometric split enables the “removal” of nuisance influence via orthogonal projection.

Given the total score vector $\theta$4 for $\theta$5, the efficient (“residual”) score is

$\theta$6

where $\theta$7 denotes the orthogonal projection in $\theta$8. This extraction yields the efficient information and underpins the Cramér–Rao lower bound in semiparametric settings (Fortunati et al., 29 Jun 2025, Rieder, 2014).

Orthogonal projections in robust statistics obey analogous rules; for any $\theta$9, the projection $\eta$0 onto $\eta$1 satisfies

$\eta$2

3. Explicit Structure for Elliptically Symmetric and Robust Models

Real Elliptically Symmetric (RES) Families

Let $\eta$3, with $\eta$4. Parametrize $\eta$5 as $\eta$6 with scale $\eta$7 and shape $\eta$8. The nuisance tangent space comprises:

• A finite-dimensional part from scale and shape: $\eta$9,
• An infinite-dimensional part from the density generator $$H = \{ h : \mathcal{X} \to \mathbb{R} \mid E_0[h]=0, E_0[h^2]<\infty \}$$0: $$H = \{ h : \mathcal{X} \to \mathbb{R} \mid E_0[h]=0, E_0[h^2]<\infty \}$$1, with $$H = \{ h : \mathcal{X} \to \mathbb{R} \mid E_0[h]=0, E_0[h^2]<\infty \}$$2.

For any $$H = \{ h : \mathcal{X} \to \mathbb{R} \mid E_0[h]=0, E_0[h^2]<\infty \}$$3,

$$H = \{ h : \mathcal{X} \to \mathbb{R} \mid E_0[h]=0, E_0[h^2]<\infty \}$$4

where the projections are given by

$$H = \{ h : \mathcal{X} \to \mathbb{R} \mid E_0[h]=0, E_0[h^2]<\infty \}$$5

Robust Neighborhoods

When the nuisance is model deviation captured by a neighborhood (e.g. Hellinger ball), the tangent set $$H = \{ h : \mathcal{X} \to \mathbb{R} \mid E_0[h]=0, E_0[h^2]<\infty \}$$6 is a convex subset of $$H = \{ h : \mathcal{X} \to \mathbb{R} \mid E_0[h]=0, E_0[h^2]<\infty \}$$7:

• Hellinger: $$H = \{ h : \mathcal{X} \to \mathbb{R} \mid E_0[h]=0, E_0[h^2]<\infty \}$$8,
• total variation: $$H = \{ h : \mathcal{X} \to \mathbb{R} \mid E_0[h]=0, E_0[h^2]<\infty \}$$9,
• contamination: $\langle h_1, h_2\rangle = E_0[h_1 h_2]$0 (Rieder, 2014).

The closed linear span of $\langle h_1, h_2\rangle = E_0[h_1 h_2]$1 forms the nuisance tangent space, and projections are characterized as best $\langle h_1, h_2\rangle = E_0[h_1 h_2]$2-approximations onto these closed linear sets.

4. Efficient Score, Influence Curves, and Information Bound

The semiparametric efficient score, and the influence function, result from subtracting the nuisance-space projection:

$\langle h_1, h_2\rangle = E_0[h_1 h_2]$3

Its covariance matrix $\langle h_1, h_2\rangle = E_0[h_1 h_2]$4 is the semiparametric Fisher information, yielding the sharpest achievable lower bound for regular estimators:

$\langle h_1, h_2\rangle = E_0[h_1 h_2]$5

For robust neighborhoods, the semiparametric influence curve is constructed as

$\langle h_1, h_2\rangle = E_0[h_1 h_2]$6

where $\langle h_1, h_2\rangle = E_0[h_1 h_2]$7 is the efficient score and $\langle h_1, h_2\rangle = E_0[h_1 h_2]$8 its variance. In the Hellinger-case, the semiparametric and minimax robust approaches coincide exactly; for total variation and contamination, the “clipping” of the score is recovered only in the sense of influence function “truncation” (Rieder, 2014).

5. Illustrative Examples and Applications

In RES models with low-rank structure (e.g., covariance $\langle h_1, h_2\rangle = E_0[h_1 h_2]$9, $\theta$0 the parameter of interest, $\theta$1 as nuisance), projecting the score for $\theta$2 onto the nuisance tangent space (including both $\theta$3, $\theta$4, and the generator $\theta$5) demonstrates that the efficient score coincides with that of the parametric submodel. Semiparametric and parametric information bounds thus collapse for $\theta$6 in this structure, demonstrating that infinite-dimensional nuisances (such as the density generator) do not reduce efficiency in directions orthogonal to their influence (Fortunati et al., 29 Jun 2025).

6. Extensions: Complex-Valued Models and Testing

Complex Elliptically Symmetric Distributions

In Circular and Noncircular Complex Elliptically Symmetric (C-CES, NC-CES) models, the nuisance tangent space construction generalizes with either real embeddings ($\theta$7) or by directly working in complex Hilbert spaces using Wirtinger calculus. The nuisance spaces and projection formulas remain structurally unaltered; the efficient score is defined analogously with the hermitian inner product $\theta$8 (Fortunati et al., 29 Jun 2025).

Testing Problems

In robust hypothesis testing, projecting the score function onto the difference of convex tangent cones underpins maxmin test construction. In settings with one-sided alternatives or robustified composite hypotheses, the Neyman–Pearson critical direction is determined geometrically as the minimum-norm direction in $\theta$9, the difference of tangent sets. Classical asymptotic robust $\eta$0-tests for contaminated hypotheses are thus recovered as a geometric projection problem in the nuisance tangent space framework (Rieder, 2014).

7. Unified Perspective and Significance

The concept of the nuisance tangent space unifies handling of both finite and infinite-dimensional nuisance parameters (including model deviation and functional/law parameters) as closed Hilbert subspaces. All projection, decomposition, and efficiency procedures operate at the level of orthogonal geometry in $\eta$1 or similar Hilbert spaces. This approach provides a general recipe for efficient estimation and optimal testing in semiparametric, robust, and misspecified models: efficient procedures are characterized by residuals after optimal $\eta$2-projection onto the nuisance tangent space. In Hellinger neighborhoods, the semiparametric and robust approaches coincide; in total variation and contamination cases, the geometric argument captures the structure of optimally robust procedures, including the appearance of clipping in the influence functions (Fortunati et al., 29 Jun 2025, Rieder, 2014).