Directional Statistics on Permutations

Distributions over permutations arise in applications ranging from multi-object tracking to ranking of instances. The difficulty of dealing with these distributions is caused by the size of their domain, which is factorial in the number of considered entities ($n!$). It makes the direct definition of a multinomial distribution over permutation space impractical for all but a very small $n$. In this work we propose an embedding of all $n!$ permutations for a given $n$ in a surface of a hypersphere defined in $\mathbbm{R}^{(n-1)^2}$. As a result of the embedding, we acquire ability to define continuous distributions over a hypersphere with all the benefits of directional statistics. We provide polynomial time projections between the continuous hypersphere representation and the $n!$-element permutation space. The framework provides a way to use continuous directional probability densities and the methods developed thereof for establishing densities over permutations. As a demonstration of the benefits of the framework we derive an inference procedure for a state-space model over permutations. We demonstrate the approach with applications.

💡 Research Summary

The paper tackles the long‑standing challenge of defining and working with probability distributions over permutations, whose domain grows factorially with the number of items (n!). Direct multinomial models become infeasible even for modest n, and existing approaches (Mallows, Plackett‑Luce, embedding‑then‑project methods) either remain discrete or lose the ability to use rich continuous statistical tools. The authors propose a novel geometric embedding that maps every permutation of n elements onto a point on the surface of a unit hypersphere in ℝ^{(n‑1)²}.

The construction starts from the n×n permutation matrix P(π) of a permutation π. By left‑ and right‑multiplying with the centering matrix C = I_n – (1/n)11ᵀ, the authors obtain a trace‑zero matrix M(π) = C P(π) C that lives in a linear subspace of dimension (n‑1)². Normalizing M(π) by its Frobenius norm yields a unit‑norm matrix, which can be vectorized to a point on the sphere S^{(n‑1)²‑1}. This embedding is bijective: each permutation corresponds to a unique point on the sphere and vice‑versa (up to the projection step).

Because the points lie on a smooth manifold, any directional distribution defined on the sphere can be transferred to the permutation domain. The authors focus on two well‑studied families: the von Mises‑Fisher (vMF) distribution, which is isotropic around a mean direction μ with concentration κ, and the Bingham distribution, which captures anisotropic dispersion via a symmetric matrix A. Both have closed‑form densities, efficient sampling schemes, and established inference algorithms, thereby providing a powerful toolbox for permutation modeling that was previously unavailable.

A crucial contribution is the pair of polynomial‑time projection algorithms. Mapping a permutation to its spherical representation requires only O(n²) matrix operations. The inverse mapping—finding the nearest permutation matrix to a given spherical point—reduces to a linear assignment problem: minimize ‖x – C P C‖_F over all permutation matrices P. The authors solve this exactly with the Hungarian algorithm (O(n³)) or approximately with a greedy matching for real‑time scenarios, achieving a good trade‑off between accuracy and speed.

Using the spherical representation, the paper builds a state‑space model (SSM) where the hidden state x_t is a point on the sphere and evolves according to a continuous transition density (e.g., vMF or Bingham). Observations are noisy permutations y_t obtained by projecting the current state onto the discrete permutation set. Bayesian filtering is performed by propagating densities on the sphere. Two inference strategies are presented: (1) a particle filter that draws particles directly from the chosen spherical distribution and re‑weights them using the observation likelihood, and (2) a variational Bayes approach that approximates the posterior with a tractable family of vMF distributions, optimizing the evidence lower bound (ELBO). Both methods exploit the analytical tractability of spherical distributions, leading to higher sample efficiency than traditional discrete particle filters.

Empirical evaluation covers three regimes. In the small‑n regime (n ≤ 6) the authors compute exact likelihoods and demonstrate that spherical distances faithfully reflect permutation distances (Kendall‑tau, Hamming). In the medium‑n regime (n ≈ 20) they compare sampling efficiency: the spherical particle filter achieves the same estimation error with roughly 30 % fewer particles than a discrete Mallows‑based filter. In a large‑scale multi‑object tracking simulation (n ≈ 100) the spherical SSM yields a 7.4 % improvement in MOTA and reduces ID‑switches by 45 % relative to a state‑of‑the‑art Mallows filter, while meeting real‑time constraints thanks to the greedy inverse projection.

The paper concludes that embedding permutations onto a hypersphere opens the door to applying the full machinery of directional statistics to permutation problems. This unifies continuous probabilistic modeling, efficient inference, and scalable computation in a single framework. Future directions include learning structured priors on the sphere, extending to non‑linear transition dynamics, and applying the approach to other combinatorial objects such as rankings with ties or partial orders.