--- title: Luce-Distributed Permutations url: https://www.emergentmind.com/topics/luce-distributed-permutations type: topic --- # Luce-Distributed Permutations Luce-distributed permutations are probability distributions on permutations generated by sequential Luce choices: at each step, one samples without replacement from the remaining items with probabilities proportional to prescribed positive weights. In the statistical literature, the standard Plackett–Luce model, the Extended Plackett–Luce model, and inverse-Luce variants are all instances of this general idea, while broader Gibbs-posterior formulations recover Luce-family noisy-comparison models such as Bradley–Terry–Luce. Recent work studies these distributions from several directions: parametric ranking models, Bayesian inference, message-passing approximations, asymptotic limit theory, stochastic-gradient methods for latent permutations, and secretary-type online decision problems [2002.05953] [2110.00513] [2509.07729] [2605.10452] [1911.10036]. ## 1. Sequential-choice definition and equivalent continuous representations In the standard Plackett–Luce formulation, a complete ranking of \(K\) entities is a permutation \(x=(x_1,\dots,x_K)\in \mathcal S_K\), where \(x_j\) is the entity assigned to rank \(j\). Each entity \(k\) has a positive worth parameter \(\lambda_k>0\), and the model assigns probability \[ \Pr(X=x\mid \lambda)=\prod_{j=1}^{K} \frac{\lambda_{x_j}}{\sum_{m=j}^{K}\lambda_{x_m}}. \] This is a discrete distribution on all \(K!\) permutations. In the logit parameterization used for black-box optimization, for a permutation \(b=(b_1,\dots,b_k)\in S_k\) and score vector \(\theta=(\theta_1,\dots,\theta_k)\), \[ p(b \mid \theta)=\prod_{j=1}^k \frac{\exp \theta_{b_j}}{\sum_{u=j}^k \exp \theta_{b_u}}. \] Both expressions encode the same stagewise mechanism: the first item is chosen from all items, the second from the remaining items, and each selection obeys Luce’s choice axiom [2002.05953] [1911.10036]. A weight-based formulation emphasizes sampling without replacement. For positive weights \(\theta_1,\dots,\theta_n\) with total mass \[ w_n=\theta_1+\dots+\theta_n, \] define \[ p(\theta_1,\dots,\theta_n):=\frac{\theta_1}{w_n}\times\frac{\theta_2}{w_n-\theta_1}\times\frac{\theta_3}{w_n-\theta_1-\theta_2}\times\cdots\times\frac{\theta_n}{\theta_n}. \] The law on \(S_n\) is then \[ P(\sigma_n)=p(\theta_{\sigma_n^{-1}(1)},\dots,\theta_{\sigma_n^{-1}(n)}). \] Equivalently, if \(\tau=\tau_1\cdots\tau_n\) is the draw order, then \(\sigma(\tau_j)=j\), so \(\sigma(i)\) is the draw time of label \(i\) [2509.07729]. Two continuous representations are central. In one, if \(E_i\) are independent exponentials with mean \(1/\theta_i\), then comparisons among the \(E_i\)’s encode the Luce order. In the other, if \[ z_i \sim G(\theta_i,1), \qquad i=1,\dots,k, \] then the event \(z_{b_1}\ge \cdots \ge z_{b_k}\) has probability exactly equal to the Plackett–Luce probability. These representations are the basis of both asymptotic analysis and low-variance gradient estimators [2509.07729] [1911.10036]. The mode has a simple form in the standard model. Under standard Plackett–Luce, the modal ordering is \[ \hat x=\operatorname{order}_{\downarrow}(\lambda), \] and, in the score parameterization, the mode is the permutation that sorts scores in descending order [2002.05953] [1911.10036]. ## 2. Extended, reverse, and inverse Luce models The Extended Plackett–Luce (EPL) model generalizes standard Plackett–Luce by introducing a choice order \[ \sigma=(\sigma_1,\dots,\sigma_K)\in\mathcal S_K, \] which specifies the order in which ranks are filled during the ranking process. Its probability mass function is \[ \Pr(X_i=x_i\mid \lambda,\sigma) = \prod_{j=1}^{K} \frac{\lambda_{x_{i\sigma_j}}}{\sum_{m=j}^{K}\lambda_{x_{i\sigma_m}}}. \] Equivalently, it is the standard Plackett–Luce probability applied to permuted data \[ x_i^*=(x_{i\sigma_1},\dots,x_{i\sigma_K}). \] The model therefore induces a flexible discrete distribution on \(\mathcal S_K\) in which the sequence of Luce choices is itself permutation-valued [2002.05953]. Two special cases are explicit. If \(\sigma=\mathcal I=(1,2,\dots,K)\), EPL reduces to the standard forward Plackett–Luce model. If \(\sigma=(K,K-1,\dots,1)\), it becomes the reverse Plackett–Luce model. The worth parameters \(\lambda_k\) still govern selection probabilities, but for general \(\sigma\) their direct interpretation as “preference for rank \(1\), rank \(2\), \dots” is no longer immediate because ranking stages are traversed in the order \(\sigma\), not necessarily from best to worst. The modal ranking under EPL is \[ \hat x^{(\sigma,\lambda)} = \hat x \circ \sigma^{-1}, \] so the choice order re-labels the ranking stages [2002.05953]. The parameterization is mixed discrete/continuous: - discrete component: \(\sigma\in\mathcal S_K\), - continuous component: \(\lambda=(\lambda_1,\dots,\lambda_K)\in\mathbb R_{>0}^K\). The likelihood is invariant under scalar multiplication, \[ \lambda \mapsto c\lambda,\qquad c>0, \] so the model is identifiable only up to scale. The cited work notes that this is not a major substantive issue, but it can affect MCMC mixing [2002.05953]. A distinct but closely related variant is the inverse-Luce model. For positive weights \(\{\theta_i\}_{i=1}^n\), the direct Luce distribution on \(S_n\) is \[ P_n^{\mathcal{L};\{\theta_i\}_{i=1}^n}(\sigma_1\cdots\sigma_n) = \theta_{\sigma_1} \frac{\theta_{\sigma_2}}{\omega_n-\theta_{\sigma_1}} \frac{\theta_{\sigma_3}}{\omega_n-\theta_{\sigma_1}-\theta_{\sigma_2}} \cdots \frac{\theta_{\sigma_n}}{\theta_{\sigma_n}}, \qquad \omega_n=\sum_{i=1}^n \theta_i. \] Its inverse-Luce version is the pushforward under inversion \(\sigma\mapsto \sigma^{-1}\): \[ P_n^{\mathcal{L}_{\mathrm{inv};\{\theta_i\}_{i=1}^n}}(\sigma_1\cdots\sigma_n) = \theta_{\sigma^{-1}_1} \frac{\theta_{\sigma^{-1}_2}}{\omega_n-\theta_{\sigma^{-1}_1}} \frac{\theta_{\sigma^{-1}_3}}{\omega_n-\theta_{\sigma^{-1}_1}-\theta_{\sigma^{-1}_2}} \cdots \frac{\theta_{\sigma^{-1}_n}}{\theta_{\sigma^{-1}_n}}. \] The inverse-Luce model is the one that becomes tractable in the secretary problem when the smallest number is the top rank [2605.10452]. ## 3. Statistical inference and computational methods Bayesian inference for EPL starts from data \(\mathcal D=\{x_1,\dots,x_n\}\) and likelihood \[ \pi(\mathcal D\mid \sigma,\lambda) = \prod_{i=1}^{n}\prod_{j=1}^{K} \frac{\lambda_{x_{i\sigma_j}}}{\sum_{m=j}^{K}\lambda_{x_{i\sigma_m}}}. \] The prior is factored as \[ \pi(\sigma,\lambda)=\pi(\lambda\mid \sigma)\pi(\sigma). \] For the choice order, the cited work uses a Plackett–Luce prior \[ \sigma\mid q \sim \mathrm{PL}(q), \qquad \Pr(\sigma\mid q)=\prod_{j=1}^{K} \frac{q_{\sigma_j}}{\sum_{m=j}^{K}q_{\sigma_m}}, \] which is uniform over \(\mathcal S_K\) if \(q_k=q\) for all \(k\). Conditional on \(\sigma\), the worth parameters have independent Gamma priors, \[ \lambda_k\mid \sigma \stackrel{ind}{\sim} \mathrm{Ga}(a_k^{(\sigma)},1), \qquad k=1,\dots,K. \] A central feature is a mode-preserving prior predictive framework: the hyperparameters \(a_k^{(\sigma)}\) are chosen so that the modal ordering of the prior predictive distribution is preserved across different choice orders [2002.05953]. Posterior sampling is carried out with Metropolis-coupled MCMC (MC\(^3\)) / parallel tempering because the posterior over \(\sigma\) can have local modes separated by large distances in permutation space. The algorithm alternates between updating \(\lambda\), updating \(\sigma\), rescaling \(\lambda\), and swapping states between tempered chains. The paper recommends geometric temperature spacing and swapping adjacent chains only. Predictive inference is explicit: \[ \Pr(\tilde x\mid \mathcal D) = \mathbb E_{\sigma,\lambda\mid \mathcal D}\left[\Pr(\tilde x\mid \sigma,\lambda)\right]. \] The posterior predictive mode is interpreted as the aggregate ranking, while discrepancies \[ d_{jk} = \left| \Pr(\tilde x_j=k\mid \mathcal D)-\Pr(x_j=k) \right| \] are used to detect lack of fit [2002.05953]. A different inference perspective is given by Gibbs posteriors on permutations from pairwise comparison data. With a directed comparison graph \(G\), a uniform prior over permutations, and edge factor \(f\), the posterior is \[ P(\bm{\pi} \mid G) = \frac{\prod_{(i,j)\in G} f(\pi_i,\pi_j)} {\sum_{\bm{\pi}'} \prod_{(i,j)\in G} f(\pi_i',\pi_j')}. \] This is interpreted as \[ P(\bm{\pi}\mid G)=\frac{e^{-\beta H(\bm{\pi})}}{Z}, \qquad H(\bm{\pi})=\sum_{(i,j)\in G} h(\pi_i,\pi_j). \] The framework includes a hard step-function Hamiltonian for partial orders and a Bradley–Terry–Luce noisy-comparison model with \[ f(x_i,x_j)=\frac{e^{\beta(x_i-x_j)}}{e^{\beta(x_i-x_j)}+e^{\beta(x_j-x_i)}}=\frac{r_i}{r_i+r_j}, \qquad r_i=e^{2\beta x_i}. \] Belief propagation computes approximate marginal position distributions, while the Bethe free energy is used both to approximate the number of linear extensions and to perform model selection between competing ranking models such as the step-function model, Bradley–Terry–Luce, and related alternatives [2110.00513]. For optimization over latent permutations, low-variance score-function estimators have been developed specifically for the Plackett–Luce distribution. The task is \[ \min_{\theta}\ \mathbb{E}_{p(b\mid \theta)}[f(b)], \] where \(b\) is a permutation and \(f\) may be non-differentiable or only available as a black box. The cited work extends REBAR and RELAX to permutations by using the Gumbel representation \(b=H(z)\), deriving a conditional distribution \(p(z\mid b,\theta)\) as a sequence of truncated Gumbels, and constructing PL-REBAR and PL-RELAX control variates. The support size is \(k!\), but Plackett–Luce uses \(O(k)\) parameters and can be sampled in \(O(k\log k)\), making it usable in causal structure learning and other black-box permutation problems [1911.10036]. ## 4. Global and local asymptotic theory A recent asymptotic theory studies permutations \(\sigma_n\sim \Luce(\theta_1,\dots,\theta_n)\) through their permuton and local limits. Set \[ f_n(y):=\theta_{\lfloor yn \rfloor}, \qquad y\in[0,1], \] and assume \[ f_n \to f \quad \text{a.e. on }[0,1], \] for some positive, finite, measurable function \(f\). Define \[ F(x):=1-\int_0^1\mathrm{e}^{-x f(y)}\,d y,\quad \text{for all }x\geq 0. \] The limiting permuton \(\mu\) is the law of \[ (X,Y):=\Big(U,F\big(E/f(U)\big)\Big), \] where \(U\sim \mathrm{Unif}[0,1]\), \(E\sim \mathrm{Exp}(1)\), independent. Under these assumptions, \((\sigma_n)_n\) converges in probability in the permuton sense to the deterministic permuton \(\mu\) [2509.07729]. The limit is explicit. The permuton is absolutely continuous with density \[ \rho(x,y):= \frac{f(x)\mathrm{e}^{-f(x)F^{-1}(y)}}{\int_0^1f(t)\mathrm{e}^{-f(t)F^{-1}(y)}\,d t}, \quad\text{for all }(x,y)\in(0,1)^2. \] For Sukhatme weights \(\theta_i=n-i+1\), one has \(f(x)=1-x\), and the density is asymmetric and singular near \((1,1)\). Pattern densities are also identified: for \(\pi\in S_k\), \[ (\pi,\mu)= E\Big[ P(\Luce(f(U^1),\dots,f(U^k))=\pi \mid U^1,\dots,U^k) \Big], \] where \((U^1,\dots,U^k)\) are the order statistics of \(k\) i.i.d. uniform variables. Thus the limiting pattern law is an average of exact Luce laws with random weights. For Sukhatme weights, \((12,\mu)=\log 2\) and \((21,\mu)=1-\log 2\), showing a marked global bias [2509.07729]. The same work emphasizes that exact Luce permutations and permutations sampled from the limiting permuton can differ significantly on fine-scale statistics, even though they share the same global permuton limit. A key example is the distribution of the first or top values: a naive approximation replacing random weights \(f(U^i)\) by deterministic weights is false in general. This suggests a separation between global convergence in the permuton sense and finer local or top-rank behavior [2509.07729]. Local convergence is formulated through consecutive patterns. For fixed \(k\) and \(\pi\in S_k\), if \[ \Lambda(\pi):=\lim_{n\to +\infty}\frac{1}{n}\sum_{i=0}^{n-k}p(\theta_{i+\pi^{-1}(1)},\dots, \theta_{i+\pi^{-1}(k)}) \] exists, then \[ \pcocc(\pi,\sigma_n)\to\Lambda(\pi) \] in probability. For Sukhatme weights \(\theta_i=n-i+1\), the limiting local frequencies are uniform, \[ \Lambda(\pi)=\frac{1}{k!}, \] for every fixed \(\pi\in S_k\). Hence the model is locally uniform even though it is globally non-uniform. The paper further proves a Berry–Esseen-type CLT for consecutive pattern occurrences and a CLT for inversions. For any positive weights, \[ E(\occ(21,\sigma_n)) = \sum_{1\le i0\), \[ M_n^* \sim n\left(\frac1c \log\!\left(1+\frac{e^c-1}{e}\right)\right), \qquad \lim_{n\to\infty} P_n(\mathcal{S}^\downarrow(n,M_n^*))=\frac1e. \] For \(q_n=1\pm c/n^\alpha\), \(0<\alpha<1\), \[ M_n^* \sim \frac{1-\log(e-1)}{c}\, n^\alpha \approx 0.459\, \frac{n^\alpha}{c}, \] again with limiting success probability \(1/e\). For fixed \(q>1\), \[ M_n^* = n-L, \qquad \frac1{q-1}\le L < \frac{q}{q-1}, \] and the limiting success probability is \[ \frac{q-1}{q^L}L. \] For Sukhatme weights \(\theta_j^{(n)}=n+1-j\), \[ M_n^* \sim \left(1-\sqrt{1-\frac1e}\right)n \approx 0.2049\,n, \] while for reverse Sukhatme weights \(\theta_j=j\), \[ M_n^* \sim e^{-1/2}n \approx 0.6065\,n, \] and in both cases the limiting success probability is \(1/e\) [2605.10452]. ## 6. Scope of the term and distinction from unrelated permutation coordinatizations The phrase “Luce-distributed permutations” refers in these sources to probabilistic models on \(S_n\): sequential weighted sampling without replacement, parametric ranking distributions such as Plackett–Luce and EPL, inverse-Luce laws, and Gibbs posteriors that include Bradley–Terry–Luce as a special case [2002.05953] [2110.00513] [2509.07729] [2605.10452] [1911.10036]. It should be distinguished from deterministic combinatorial uses of permutations that do not define probability distributions. A clear example is the theory of join-distributive lattices of length \(n\) and join-width at most \(k\). In that setting, there are two descriptions by \(k-1\) permutations acting on an \(n\)-element set: the Edelman–Jamison construction inside a powerset lattice and a lattice-theoretic coordinatization by eligible tuples. The key result is \[ \mathsf{EJ}(\boldsymbol{\sigma})\cong \mathsf{Cz}(\boldsymbol{\sigma}^{-1}), \] and the paper also characterizes join-distributive lattices by trajectories. This is a theory of finite join-distributive lattices, semimodularity, meet-semidistributivity, and trajectory classes of prime intervals, rather than a Luce-type stochastic law on permutations [1210.3376]. The distinction matters because both areas use permutation data, but for different purposes. In Luce-type models, permutations are random outcomes whose probabilities are determined by worth parameters, choice orders, or pairwise interaction factors. In the join-distributive lattice setting, permutations serve as coordinates for a deterministic representation theorem. The two subjects therefore share notation but not mathematical content [1210.3376]. Source: https://www.emergentmind.com/topics/luce-distributed-permutations