On the singularity of the Fisher Information matrix in the sine-skewed family on the d d -dimensional torus

Emily Schutte, Sophia Loizidou, Vincent Laheurte · 2026

Read On the singularity of the Fisher Information matrix in the sine-skewed family on the d d -dimensional torus by Emily Schutte, Sophia Loizidou, Vincent Laheurte (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2603-04369.txt.

Introduction

Schutte and colleagues pin down when Fisher information turns singular for sine-skewed torus models. Sine skewing, the only skewing mechanism on the hyper-torus, adds asymmetry to torus distributions. Some models suffer singular Fisher matrices near symmetry, harming inference. It was open which models show the flaw. The paper gives a general characterization of the singular class in full d-dimensional settings.

Catalog record

CREATOREmily Schutte, Sophia Loizidou, Vincent Laheurte
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE26340
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2603.04369

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

On the singularity of the Fisher Information matrix in the sine-skewed family on the d d -dimensional torus Emily Schutte, Sophia Loizidou, Vincent Laheurte Abstract Skewed distributions are fundamental in modelling asymmetric data on the d d -dimensional torus. In this context, asymmetry is introduced through the sine-skewing mechanism, which is the only skewing mechanism that has been proposed on the hyper-torus in the literature. Some sine-skewed models are known to suffer from a singular Fisher information matrix in the vicinity of symmetry, which poses a significant issue for inferential purposes. It is an open question to determine for which sine-skewed models Fisher information singularity occurs. In this paper, a general characterization of the class of models that exhibit this singularity is given in the general d d -dimensional setting. Keywords Directional statistics \cdot Asymmetry \cdot Fisher information singularity On the singularity of the Fisher Information matrix in the sine-skewed family on the d -dimensional torus 1 Introduction Several real-world datasets can naturally be represented as data on the d -dimensional torus. Notable examples arise from the area of bioinformatics, particularly in the context of the protein folding prediction problem, where angles play a central role [ 13 , 19 , 20 ] . Other examples are for instance RNA data [ 24 ] and the circadian clock of two different tissues of a mouse [ 18 ] . Beyond the bioinformatics world, toroidal data arise in various domains: the direction of animal movement [ 22 ] , morphological data from human neurons [ 29 ] , wind direction data [ 15 ] and directions of steepest descent before and after an earthquake [ 28 ] , to cite but these. Many distributions have been proposed in the literature to model d -dimensional angular data. On the circle ( d=1 ), key distributions are the von Mises, the cardioid, the wrapped Cauchy and the wrapped Normal distributions, to name a few. On the two-dimensional torus, commonly used models are the bivariate wrapped Cauchy distribution [ 14 , 15 ] , the bivariate von Mises distribution [ 21 ] and its submodels, the Sine [ 31 ] and Cosine [ 20 ] distributions. Considering higher dimensions, the trivariate wrapped Cauchy copula [ 13 ] can model three angles, while the multivariate wrapped normal [ 6 ] and the multivariate extensions of the Sine and Cosine distributions [ 19 ] can be used for any dimension. For a general overview, we refer the reader to [ 16 ] and [ 26 ] . A strong limitation of the above-mentioned models arises from their symmetric nature, which is not able to capture the asymmetric phenomena that can be observed in circular and toroidal data [ 2 , 27 , 10 ] . Recognizing this need on the circle, [ 32 ] and [ 1 ] took inspiration from the popular skew-normal distribution of [ 5 ] and proposed the sine-skewed distributions on the circle. [ 3 ] extended this proposal to the d -dimensional torus, leading to a density of the form \displaystyle\boldsymbol{\theta}\mapsto f_{\boldsymbol{\mu},\boldsymbol{\lambda}}(\boldsymbol{\theta};\boldsymbol{\vartheta}):=f_{0}(\boldsymbol{\theta}-\boldsymbol{\mu};\boldsymbol{\vartheta})\left(1+\sum_{j=1}^{d}\lambda_{j}\sin(\theta_{j}-\mu_{j})\right) (1) subject to \sum_{j=1}^{d}|\lambda_{j}|\leq 1, where f_{0} is a symmetric base density on the d -dimensional torus (like all those mentioned above), \boldsymbol{\mu}=(\mu_{1},\ldots,\mu_{d})^{\intercal}\in[-\pi,\pi)^{d} is the location parameter, \boldsymbol{\vartheta} stands for any non-location parameters and \boldsymbol{\lambda}=(\lambda_{1},\ldots,\lambda_{d})^{\intercal}\in(-1,1)^{d} plays the role of the skewness parameter. For the rest of the paper, we omit \boldsymbol{\vartheta} from the notation for the sake of readability. These skewing models respond to the need of asymmetric distributions and have increased the flexibility of distributions on the d -dimensional torus, yet this comes at a cost. Indeed, a major problem arises from the Fisher information matrix (FIM) of these models and the ensuing inferential challenges. On the real line, it has been observed by [ 5 ] that the FIM for the skew-normal distribution becomes singular in the vicinity of symmetry, that is, when \lambda=0 . A lot of literature has been dedicated to this issue, see in particular [ 8 ] for a thorough account. Unfortunately, this singularity also arises in the sine-skewed family of distributions as defined in ( 1 ), which may come as no surprise since their construction follows the same principle as the skew-normal distribution of [ 5 ] . [ 17 ] showed that, in the circular case, the FIM is singular in the vicinity of symmetry if and only if the sine-skewing transformation is applied to the von Mises distribution, hence no other symmetric density yields such a result. This fact is due to an unfortunate collinearity between location and skewness scores. [ 4 ] note that in the two-dimensional case singularity is observed for the sine-skewed Cosine distribution. This observation is to be expected, as the Cosine distribution is an extension of the von Mises model on the torus. However, interestingly, the Sine distribution, which is a different two-dimensional extension of the same model, does not exhibit the same issue. This contrast highlights that extending the singularity results beyond the circle is a non-trivial task. Such singularities are problematic because singularity of the FIM indicates lack of information for the parameters to be uniquely identifiable from the data. A major inferential consequence is that the asymptotic normality of the maximum likelihood estimator fails to hold. Moreover, when the FIM is non-singular the rate of convergence to the true parameter is O(n^{-\frac{1}{2}}) . However, when the matrix is singular this rate can be significantly slower and the likelihood can also exhibit irregularities, such as bimodality [ 30 ] . This has severe implications on many statistical procedures that rely on the asymptotic normality property, such as hypothesis testing and construction of confidence intervals. Singularity of the FIM is also closely related to the smeariness of (nonparametric) sample Fréchet means [ 12 ] . Theorem 3 of [ 11 ] shows that, under the assumption that the underlying density differs from the uniform distribution at -\pi for the first time in its k^{\text{th}} derivative there, the convergence rate of the intrinsic sample mean is n^{-\frac{1}{2(k+1)}} , which is slower than the previously mentioned n^{-\frac{1}{2}} rate. Given the lack of knowledge as to which sine-skewed distributions on the d -dimensional torus suffer from a FIM singularity, we fully answer this open question in this paper. In Section 2 we provide a general form of symmetric density functions on the hyper-torus that is satisfied if and only if their sine-skewed version admits a singular FIM in the vicinity of symmetry. In Section 3 we consider some well-known distributions and study whether they satisfy the proposed characterization or not, hence suffer from FIM singularity. Lastly, we conclude the paper with a discussion in Section 4 . 2 Fisher information singularity for the d -dimensional torus In this section we establish the main result of our paper. Considering model ( 1 ), we require f_{0} to be a component-wise 2\pi -periodic, symmetric density around \boldsymbol{\mu}\in[-\pi,\pi)^{d} . More formally, we require f_{0}\in\mathcal{F} , where \mathcal{F} is defined as \displaystyle\mathcal{F}:=\left\{\begin{array}[]{ll}&f_{0}(\boldsymbol{\theta})>0\,\mbox{a.e.}\,\ \forall\boldsymbol{\theta}\in[-\pi,\pi)^{d},\\ f_{0}(\boldsymbol{\theta}):&f_{0}(\boldsymbol{\theta}+2\pi\boldsymbol{k})=f_{0}(\boldsymbol{\theta})\,\ \forall\boldsymbol{k}\in\mathbb{Z}^{d},\\ &f_{0}(\boldsymbol{\mu}+\boldsymbol{\theta})=f_{0}(\boldsymbol{\mu}-\boldsymbol{\theta})\,\ \forall\boldsymbol{\theta}\in[-\pi,\pi)^{d},\\ &\int_{[-\pi,\pi)^{d}}f_{0}(\boldsymbol{\theta}){\mathrm{d}}\boldsymbol{\theta}=1.\end{array}\right\}. We also require the following mild assumption. Assumption 1 . The mapping \boldsymbol{\theta}\mapsto f_{0}(\boldsymbol{\theta}) is differentiable on [-\pi,\pi)^{d} . In Theorem 1 we derive a general expression for symmetric densities which is

Read the complete text