Model-based estimation and imputation with torus missing values

arXiv:2609.21549 · stat.ME, stat.CO · Submitted 2026-09-18 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Statistics. The day in statistics: methods, theory and the inference under them.

Tom: We're looking at the paper: "Model-based estimation and imputation with torus missing values".

Jane: Here's the gist of the paper. Model-based estimation and imputation with torus missing values addresses parameter estimation and model-based imputation for multivariate circular data lying on a p-dimensional torus in the presence of missing values.

Tom: The Core Problem The paper tackles the challenge of parameter estimation and model-based imputation for multivariate circular data residing on a p-dimensional torus when those observations are incomplete.

Jane: The periodic nature of the sample space invalidates conventional imputation techniques designed for Euclidean data.

Tom: The proposed framework focuses on maximum likelihood estimation under the wrapped elliptically symmetric family of distributions, particularly the multivariate wrapped normal distribution when missing data is ignorable.

Title and authors: Tom: So, focusing specifically on the title and the authors of "Model-based estimation and imputation with torus missing values," what's the simple breakdown of what that actually means for us?

Jane: Well, it’s about taking data that exists on a circle, or a torus in higher dimensions, and figuring out two things: first, how to estimate the true parameters of that circular data—like its mean or covariance structure—and second, how to fill in the missing spots when we don't have all those measurements >

Lu: The paper points out that the periodic nature of this sample space is what invalidates standard imputation techniques usually designed for straight-line Euclidean data >

Meng: So, the authors are proposing a new way to do it, moving away from those old Euclidean assumptions and into something that respects the wrapping topology >

Tom: They suggest using a general framework based on the wrapped elliptically symmetric family of distributions, with special focus on the multivariate wrapped normal distribution when we assume the missing data is ignorable >

Jane: That means they’re not just doing some simple filling in; they are building a model that understands how these circular observations depend on each other across that periodic boundary >

Lu: The core idea is leveraging the conditional properties of these distributions on the unwrapped space to embed the imputation right into an expectation-maximization algorithm, treating both the wrapping coefficients and those missing entries as latent variables >

Meng: So, instead of trying to guess every single missing angle independently, they use this two-part structure—the wrapping coefficient and the missing entry—to solve it all at once during training >

Tom: The EM algorithm structure involves calculating a complete data log-likelihood that depends on these wrapping coefficients and a scale mixture representation, with the E step computing conditional expected values involving weights for each possible wrapping coefficient vector >

Jane: And in the M step, they update the parameters by maximizing those complete data score equations derived from those expectations, which is how they get their final estimates >

Lu: The imputation strategy then involves averaging over all plausible wrapping configurations while respecting that mixture nature of the conditional distribution, and they achieve principled imputations by reducing that average value at convergence modulo two pi >

Meng: That means the imputed values aren't just some arbitrary guess; they are coherent with the underlying torus geometry and their estimated covariance structure because they come from averaging over all possibilities >

Tom: And multiple imputations can be generated too, drawn from the conditional distribution of x infinity given xi observed, using those final parameter estimates and then averaging over the posterior distribution of wrapping coefficients >

Jane: So, it’s a comprehensive approach. It handles the circular constraints directly in both the estimation and imputation stages of their model >

Lu: This paper is reaching for a way to apply likelihood-based inference to these non-Euclidean manifolds, which is particularly useful when dealing with complex data structures like those in structural bioinformatics or movement ecology >

Meng: It’s ambitious, but it's the kind of representation learning idea that could fundamentally change how we model things that naturally wrap around, like molecular conformations >

The paper's summary: Tom: So, summarizing the paper "Model-based estimation and imputation with torus missing values," they’re solving parameter estimation and imputation for multivariate circular data living on a p-dimensional torus when it has missing values >

Jane: The key takeaway is that the periodic nature of the sample space breaks conventional imputation techniques meant for normal Euclidean data, so they propose a general framework using the wrapped elliptically symmetric family of distributions >

Lu: They focus particularly on the multivariate wrapped normal distribution because it inherits desirable properties from its Normal distribution, and this is where they put their main emphasis >

Meng: They use the conditional properties of these distributions on the unwrapped space to set up an EM algorithm that handles both unknown wrapping coefficients and missing angular measurements as latent variables >

Tom: The complete data log-likelihood is formulated as a function involving the wrapping coefficients and a scale mixture representation, which sets up how they handle the missingness >

Jane: The E step requires computing conditional expected values of this likelihood, which means calculating weights that represent the conditional probability of an observation having a specific wrapping coefficient vector >

Lu: Then in the M step, they update the parameters by maximizing those complete data score equations derived from those expectations to get their final solution >

Meng: For imputation, they achieve it by averaging over all plausible wrapping configurations while respecting that mixture nature of the conditional distribution, and principled imputations are obtained by reducing that average imputed value at convergence modulo two pi >

Tom: They also show how multiple imputations can be generated by drawing from the conditional distribution of x infinity given xi observed using their final parameter estimates and averaging over the posterior distribution of wrapping coefficients >

Jane: So, it’s a complete process: they estimate parameters, they handle missing data via EM, and they provide a way to get consistent imputations that respect the circular geometry >

Lu: The paper shows the potential of this distributional family for unsupervised learning on non-Euclidean manifolds and how it can be used for model-based clustering of heterogeneous torus data >

Meng: It’s interesting because it moves beyond simple point estimates by using a richer distributional family to guide the imputation, which is what you need when things are inherently messy >

The paper's improvements: Tom: Now let's talk about the suggested improvements they offer for this work and what that actually means for us in terms of making it more robust >

Jane: They suggest a few key things. First, they want to make the parameter estimation more robust by treating both wrapping coefficients and missing entries as latent variables within that principled expectation-maximization algorithm >

Lu: That’s the first big improvement: you get a way to estimate parameters from partially observed torus data by treating those two things as hidden variables in an EM process >

Meng: So instead of just guessing the wrap, it integrates the uncertainty about the wrap directly into the estimation process itself, which should lead to more stable results if done right >

Tom: Second improvement is about imputation: they want to imputations to have geometric consistency, defined as that "average imputed value at convergence modulo two pi" >

Jane: That means we aren't just getting some random number; the imputed value is coherent with the estimated covariance structure because it respects the torus geometry >

Lu: And they also address handling complex missingness mechanisms, allowing inference under ignorable missing data mechanisms, which lets them get a likelihood-based estimation even when missingness depends on observed data >

Meng: That’s useful for real-world scenarios where you know something is MAR, and the framework doesn't break because it treats that dependence as a nuisance parameter >

Tom: And finally, they address computational efficiency by using a pattern-based optimization strategy to partition the dataset into unique missingness patterns to reduce redundant high-dimensional matrix operations >

Jane: They refine their complexity to be O of N patterns times p cubed plus n times d times p squared, and they also use Cholesky decomposition for numerical stability when performing those calculations >

Lu: So, if we can get the computational efficiency down, this method becomes much more feasible for large-scale applications where you need to scale up >

Conclusion: Tom: Wrapping up this discussion on "Model-based estimation and imputation with torus missing values," the big implication is that we have a principled way to estimate parameters and impute missing data in circular spaces >

Jane: The paper shows that for multivariate circular data, standard techniques fail because of periodicity, but their EM approach handles both wrapping coefficients and missing entries together >

Lu: This gives us a principled way to get consistent imputations that respect the geometry of the torus when dealing with incomplete measurements >

Meng: For deployment, it means we have a method that can be compared against existing baselines, and they actually showed it performs better across different settings >

Tom: So for listeners who are driving or cooking, this paper gives you a concrete idea: if your data is circular and you have missing values, use this EM approach to get the most geometrically sound estimates possible >

Jane: We've seen that even though the authors noted computational demands in high dimensions, they still proved it works well on real datasets like TIM8 protein data >

Lu: It’s a solid result because they showed that their method consistently improves over other methods in every setting tested >

Meng: So, we have a model for this problem now, and the next step is figuring out how to make that model run fast enough for real-time use >

University Giustino Fortunato · University of Trento

stat.ME, stat.CO

Submitted: 2026-09-18

Updated: 2026-11-04

Comments: 23 pages (without Appendix)

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Importance score: 83/100

The gist: Model-based estimation and imputation with torus missing values addresses parameter estimation and model-based imputation for multivariate circular data lying on a p-dimensional torus in the presence…

Key concepts

Wrapped Elliptically Symmetric Distributions (WES)
These are probability distributions specifically designed for data on a torus. They describe the probability density of data points that wrap around a circle. The paper focuses on the multivariate wrapped normal distribution as a specific case to model the circular nature of the observations.
Expectation-Maximization (EM) Algorithm
This iterative statistical method is used to solve problems where some data is missing. In this context, it's used to simultaneously estimate two things: the unknown wrapping coefficients and the actual missing angular measurements. The E-step calculates expected values based on current estimates, and the M-step updates those estimates.
Wrapped Normal Distribution
This is a specific type of distribution within WES used in the estimation framework when data is ignorable (missingness doesn't depend on observed data). It provides a mathematical model for circular data where observations are periodic, allowing for likelihood-based parameter estimation.
Conditional Imputation Modulo 2π
This strategy generates imputations by averaging all plausible wrapping configurations and then reducing the result modulo 2π. This ensures the imputed values are geometrically consistent with the torus structure and provides a principled way to handle missing angular measurements.

Terminology

Summary

Model-based estimation and imputation with torus missing values addresses parameter estimation and model-based imputation for multivariate circular data lying on a p-dimensional torus in the presence of missing values. p. 0

The Core Problem

The paper tackles the challenge of parameter estimation and model-based imputation for multivariate circular data residing on a p-dimensional torus when those observations are incomplete. The periodic nature of the sample space invalidates conventional imputation techniques designed for Euclidean data. p. 0 The proposed framework focuses on maximum likelihood estimation under the wrapped elliptically symmetric family of distributions, particularly the multivariate wrapped normal distribution when missing data is ignorable. p. 0

The Proposed Methodology

The methodology employs a general framework for maximum likelihood estimation within the wrapped elliptically symmetric distributions (WES) family. This approach leverages the conditional properties of these distributions on the unwrapped space to embed imputation into an Expectation-Maximization algorithm that treats both wrapping coefficients and missing entries as latent variables. p. 0 The EM algorithm is used to simultaneously handle two sources of incompleteness: unknown wrapping coefficients and missing angular measurements. p. 4

EM Algorithm Structure

The complete data log-likelihood is formulated as a function involving the wrapping coefficients and the scale mixture representation. The E-step involves computing conditional expected values of this likelihood, which requires calculating weights that represent the conditional probability of an observation having a specific wrapping coefficient vector. p. 5 The M-step then updates the parameters by maximizing the complete-data score equations derived from these expectations. p. 6

Imputation Strategy

Model-based imputations are achieved by averaging over all plausible wrapping configurations, respecting the true mixture nature of the conditional distribution. Principled conditional imputations are obtained by reducing the average imputed value at convergence modulo 2π. p. 1 Multiple imputations can be generated by drawing from the conditional distribution of x(∞) given xi,obs using final parameter estimates and averaging over the posterior distribution of wrapping coefficients. p. 5

Performance Assessment

Monte Carlo simulations are used to assess the finite-sample performance of the maximum likelihood estimator under wrapped normal specification with various missingness patterns. Numerical studies compare the proposed EM algorithm against baselines like MLE over complete data, MLE after discarding cases with missing entries, and MLE based on imputed data using marginal means. p. 4 The results generally show that the proposed EM algorithm exhibits better performance than benchmark procedures in all considered settings. p. 4

Empirical Validation

The methodology is illustrated on real datasets, such as TIM8 protein data, where missing values are artificially introduced to benchmark imputation accuracy. Empirical applications compare the MLE obtained from the original complete data with those derived from the proposed EM algorithm and other methods. p. 5 The results indicate that missing values largely affect variance-covariance estimation, and deleting cases or imputing them with marginal means can be misleading. p. 8 The proposed EM algorithm always performs satisfactorily and improves over other methods in all considered settings. p. 0

Limitations

A limitation noted is that the algorithm becomes computationally demanding in high-dimensional settings, suggesting future research may explore different directions and the use of parallelization. The paper also notes that for a generator h not closed under marginalization, hpobs in (12) must be evaluated explicitly for each distinct missingness pattern occurring in the data. p. 9 The computational complexity per iteration is refined to O(Npatterns·p 3+n·d·p 2) where d denotes the cardinality of CJ. p. 4 The paper also incorporates a spectral regularization step to ensure numerical stability by thresholding eigenvalues. p. 4

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Robust Parameter Estimation for Circular Data: The system can estimate parameters of wrapped elliptically symmetric distributions from partially observed torus data by treating wrapping coefficients and missing entries as latent variables within a principled Expectation-Maximization algorithm.

  2. Imputation of Missing Values with Geometric Consistency: The AI system can impute missing multivariate circular data using conditional imputations, defined as the average imputed value at convergence modulo 2pi, which ensures the imputed values are coherent with the underlying torus geometry and with the estimated covariance structure.

  3. Handling Complex Missingness Mechanisms: The framework supports inference under ignorable missing data mechanisms, allowing for likelihood-based estimation even when missingness depends on observed data (MAR), treating the missingness process as a nuisance parameter that can be disregarded for estimation purposes.

  4. Performance Benchmarking Against Baselines: The AI system can quantitatively compare its proposed EM algorithm against baseline methods like MLE over complete data, MLE after discarding cases with missing entries, and MLE based on imputed data using marginal means, showing satisfactory and improves over the other methods in all considered settings.

  5. High-Dimensional Computational Efficiency: By leveraging a pattern-based optimization strategy that partitions the dataset into unique missingness patterns, the system can reduce redundant high-dimensional matrix operations, refining complexity to O(Npatterns·p 3+n·d·p 2) and using Cholesky decomposition for numerical stability.

Related papers