Estimators from Convex Geometries of the Simplex
P. Navas. Preprint, 2026. arXiv:2610.21551
Several geometries of the probability simplex come from one principle: they are generated by a potential. Convex potentials give the Bregman divergences and the dually flat structures of information geometry, and the exponentially concave potentials of Pal and Wong give logarithmic divergences and projectively flat structures. In each case minimizing the divergence under moment constraints gives an estimator, the U-estimators of Eguchi, Komori and Ohara in the first case and the λ-deformed estimators of Wong and Zhang in the second, and both reduce to maximum likelihood in the undeformed limit. The paper adds a third kind of estimator, taken from the Levi-Civita connection by minimizing geodesic distance; its existence, uniqueness and robustness depend on curvature and diameter instead of flatness, and the curvature is computed explicitly for the square-root generator. The Aitchison geometry turns up as a special case, and a symmetrized version of the Pal–Wong cost gives a robust center of Huber type. An application to compositional difference-in-differences shows that the choice of geometry changes the estimate in practice, and an appendix relates the construction to Föllmer's minimal-entropy pricing.
Characterizing M-estimators using canonical divergences from information geometry
P. Navas. Preprint, 2026. arXiv:2611.18141
On Euclidean space, the losses that give consistent M-estimators of a functional are its strictly consistent scoring functions, and Osband's principle describes them through identification functions, the estimating equations of the method of moments. The paper carries this characterization over to the probability simplex. The residual, a subtraction on Euclidean space, becomes the inverse exponential map of an affine connection. It identifies the barycenter of the connection, which is the arithmetic, the normalized geometric or the Fréchet mean for the mixture, exponential and Levi-Civita connections, and its integral along geodesics is the canonical divergence of Ay and Amari. A geometric form of Osband's principle then decides which losses each connection admits. For the two flat connections every strictly convex generator gives a consistent Bregman loss. For the curved α-connections the main result is a rigidity theorem: curvature forces the second-order term of every consistent loss, its Osband tensor, to be a multiple of the metric. Half the squared Fisher–Rao distance is then, up to scale, the only consistent loss for the Fréchet mean, and for every α other than −1, 0 and 1 no consistent loss exists and only the method of moments remains. The residual is integrable only at α = 0, which sharpens the integrability claim of Ay and Amari. The same characterization gives the estimating equations of regression on the simplex, with instruments in the role of covariates, the calibration decomposition of compositional forecasts, and a way to rank compositional forecasts once a geometry has been chosen.
The information geometry of the Sinkhorn divergence on a Hilbert manifold
P. Navas. Preprint, 2026. arXiv:2610.39850
The Sinkhorn divergence of entropic optimal transport defines a metric tensor on probability measures, introduced by Lavenant et al. The paper builds a real-analytic Hilbert manifold of measures, the smooth stratum of positive Sobolev densities on a closed Riemannian manifold, on which this tensor is a weak Riemannian metric, and carries Eguchi's construction of a geometry from a divergence up to third order. Following Ay, the metric is characterized through gradients of mean values: it is of the same kind as the Fisher–Rao and Otto metrics, and it is bounded by a multiple of the Fisher–Rao metric, but not conversely. At third order the geometry is self-dual: by the transposition symmetry of the transport problem, both connections coincide with the Levi-Civita connection and the Amari–Chentsov tensor vanishes. On smooth densities this is the connection of Ay and Schwachhöfer. The structure is invariant under isometries of the cost, and on the torus the metric is shown not to be a Hessian metric. A second part, in preparation, deals with the geodesics and curvature of the Sinkhorn metric.
Discrimination of Vector Forecasts using Multivariate Center-Outward Ranks
P. Navas. Preprint, 2026. TRR 391 Working Paper Series (forthcoming)
Forecasts of multivariate outcomes raise the question whether a forecaster that models the coordinates jointly does better than one that models each of them on its own. Accuracy, the usual performance measure, combines two attributes: calibration, which statistical post-processing provides by relabeling the values of a forecast, and discrimination, the information about the outcome that the forecast contains, which relabeling does not change. Since forecasts are usually already post-processed when they are delivered, two forecasters differ in the end by their discrimination, and that is what we compare. For scalar forecasts, Gneiting and Walz measure discrimination by the coefficient of predictive ability (CPA), an average of AUCs over all thresholds that equals a rank correlation, so that it depends on the forecast only through its ranks and does not change under strictly increasing recalibration; Pohle, Wermuth and Weiß provide confidence intervals and tests for it through the asymptotics of rank correlations. For vector forecasts of multivariate outcomes, we measure their discrimination by the distance covariance of the center-outward ranks of forecast and outcome, the multivariate probability integral transforms given by optimal transport. The coefficient is zero if and only if forecast and outcome are independent; its values for different forecasters are on a common scale, and it is invariant under a class of relabelings that we identify as the multivariate analogue of the strictly increasing maps. Under independence (the null hypothesis), the exact distribution-free test of Shi, Drton and Han applies; away from it (alternative hypothesis), the estimation of the ranks contributes to the first-order asymptotics, and we carry out inference with entropic ranks (at a fixed level of regularization). The distance covariance of the estimated entropic ranks is asymptotically linear, with an influence function in closed form that has two parts: one for the statistic with known ranks and one for the estimation of the rank maps. We obtain confidence intervals, a paired test that compares two forecasters of the same outcome, and simultaneous intervals over subsets of coordinates. In simulations, the intervals have coverage close to 95% under “clear” dependence, the paired test has size close to 5%, slightly above it only at the smallest sample size, and it is more powerful than a comparison that treats the two coefficients as independent.
Characterizing M-estimators using Canonical Divergences from Information Geometry. Statistische Woche, Essen, September 2026.
The Convex Geometries of the Probability Simplex. Ruhrmetrics, Essen, July 2026.
Bayesian Nonparametrics using the Sinkhorn Divergence. German Probability and Statistics Days, Dresden, March 2025.
Estimating Multivariate Quantiles using Optimal Transport. School and Conference on Metric Measure Spaces, Ricci Curvature, and Optimal Transport, Como, September 2024.
Categorical Approach to Bayesian Geometric Updating. Econometrics Seminar, University of Duisburg-Essen, May 2019.
Reproducing Kernel Hilbert Manifolds coming from the Sinkhorn Divergence. Mathematics of Machine Learning, Hamburg, September 2025.
Bayesian Estimation and Autoencoders with the Sinkhorn Divergence. Mathematical Methods for High-Dimensional Data, Rome, September 2025.
Bayesian Nonparametrics and Hilbert Manifolds using the Sinkhorn Divergence. Uncertainty in Multivariate, Non-Euclidean, and Functional Spaces, Cambridge, May 2025.
Stochastic Geometry and Point Processes, Paderborn University, May 2026.
IMS International Conference on Statistics and Data Science (ICSDS), Sevilla, December 2025.
PEPPER, Probability Event: Point Processes, Exchange and Research, Münster, November 2024.
SPIRES Centre for Topological Data Analysis Conference, Oxford, August 2024.
Synergies between Modern Probability, Geometric Analysis and Stochastic Geometry, Hausdorff Research Institute for Mathematics, Bonn, 2024.
Workshop Mathematics of Data Science, Hausdorff Research Institute for Mathematics, Bonn, April 2022.
Hausdorff Winter School: The Interplay between High-Dimensional Geometry and Probability, 2021.
Hausdorff School on Markov Chain Monte Carlo, 2020.
Economics meets Optimal Transport, Hausdorff Research Institute for Mathematics, Bonn, July 2017.
Triangular quantile maps for multivariate extremes
Multivariate quantiles for the tails, built from a triangular transport map. How large an observation counts as is set by a size gauge: the sum, the maximum, a numéraire coordinate or the Euclidean norm. The work shows how to convert between these gauges and proves a factorization theorem over the triangular class.
Double machine learning for entropic vector quantile regression
Inference for the coefficients of entropic vector quantile regression via orthogonal scores. The orthogonal score at fixed regularization is derived; the theory with high-dimensional dictionaries is the next step.