A path-dependent entropic Lagrangian calculus that extends static state selection to probability-path evolution through restricted generators, upper-limit history terms, and explicit balance--entropy port routing is developed.
Abstract
Probability distributions are central to information theory, statistical inference, and modern probabilistic learning. Maximum entropy selects a probability state under prescribed constraints, but it does not specify how that state is reached, how probability is transported, or how dissipation and external information exchange are accounted for along the path. We develop a path-dependent entropic Lagrangian calculus that extends static state selection to probability-path evolution through restricted generators, upper-limit history terms, and explicit balance--entropy port routing. The construction yields the thermal state relation, conservative probability balance, and nonnegative production under standard mobility closure. Its KL/Shannon sector recovers maximum-entropy and Bayesian laws as stationary no-flux states, while time-dependent information potentials separate internal dissipation from supplied information power. Composable information and structural potentials control tails, sparsity, robustness, regularization, and nonlocal multimodality without changing the accounting architecture. Two numerical examples verify mass conservation, energy decomposition, and the total free-energy ledger.
We develop a least-action framework for describing how a probability distribution can evolve from an equilibrium state to a prescribed nonequilibrium state under constrained incremental changes. Taking a Gibbs distribution as the equilibrium reference, the framework gives a direct physical meaning to the geometry of the probability simplex: distance from equilibrium corresponds to nonequilibrium free energy, while changes between successive distributions carry an informational kinetic cost. The Pythagorean structure of relative entropy then provides the central insight of the work. It shows that intermediate distributions chosen via sequential information projections can reduce the kinetic cost of large transitions and establishes an energy-conservation-like relation between the kinetic expenditure along a path and the free energy accumulated in reaching the target distribution. Motivated by this geometry, we construct a greedy least-action path through successive information projections, obtain a closed-form characterization of each projection through the Lambert W function, and establish a finite-step performance guarantee. We further show that state-dependent costs can be incorporated naturally by reshaping the underlying Gibbs reference, providing a thermodynamic interpretation of path penalties as modifications of the effective energy landscape. Together, these results provide a unified view of distributional evolution through least action, information geometry, and nonequilibrium thermodynamics.
We study infinite-horizon time-inconsistent Markov decision processes with a countably infinite state space and unbounded reward functions. The reward is allowed to depend explicitly on the initial time and initial state, thereby accommodating general sources of time inconsistency. We seek relaxed feedback equilibria, and our approach is based on entropy regularization and weighted functional analytic methods. With entropy regularization, we characterize a regular relaxed equilibrium through a fixed-point operator. By introducing two weight functions with distinct roles, one controlling the growth of rewards and values and the other defining the ambient weighted space, we construct a compact invariant set under a product topology and apply the Schauder-Tychonoff fixed-point theorem to establish existence of regularized equilibria. Importantly, the invariant set can be chosen uniformly for small entropy weight $\lambda\in(0,1]$. We then let $\lambda\to0+$ and show, through compactness, concentration of Gibbs policies, and uniform-integrability arguments, that a subsequential limit is a relaxed equilibrium of the original unregularized problem. We further study a policy iteration algorithm (PIA) for the entropy-regularized equilibrium problem. Under a weighted-discounting structure and sufficiently strong discounting, we establish exponential convergence and uniqueness of the regularized equilibrium in a suitable weighted Banach space. Combining the policy-iteration error with a quantitative soft-max approximation bound, we show that the iterated policies constitute weighted $\varepsilon$-equilibria for the original unregularized problem and derive an explicit regret estimate. A numerical example illustrating the convergence of PIA under strong discounting and a counterexample demonstrating its failure under weak discounting are also provided.
Bayesian posterior sampling is a ubiquitous paradigm for problems where a point estimate of parameters is not sufficient, such as risk analysis and uncertainty quantification. However, likelihoods may be misspecified, intractable, computationally expensive, or not representative of the discrepancy of interest. Generalized Bayes extends likelihood-based posterior updates by using other losses. Sinkhorn divergences have appealing geometric properties: they compare empirical measures directly and yield smooth gradients thanks to entropic regularization. In this work, we introduce Sinkhorn divergences as Generalized Bayes losses for Hamiltonian Monte Carlo (HMC) and No-U-Turn Sampler (NUTS). We also propose heuristics to set hyperparameters that affect the stability and calibration quality, such as the number of Sinkhorn iterations, the entropic regularization strength, and the marginal relaxation penalty. In regimes where the forward model relies on a stochastic simulator, we combine HMC/NUTS with a common-random-numbers strategy to obtain a deterministic surrogate objective that preserves gradients and Hamiltonian dynamics. We study both mass-preserving balanced and relaxed unbalanced settings. We evaluate our method empirically on (1) a simple Gaussian model as a sanity check; (2) a distribution supported on a noisy spiral manifold where a likelihood-based approach is a poor fit; (3) a Gaussian pulse model with misalignment due to errors-in-variables, emphasizing robustness to misspecification; and (4) CIFAR-10 image patch alignment under perturbations, highlighting differences between balanced and unbalanced regimes.
Guilhem Nespoulous, F. Bertrand, Myriam Maumy et al.· 0 citations
Entropy functionals and their associated divergences underlie many statistical methods, including maximum entropy inference, minimum divergence estimation, and goodness-of-fit testing, yet choosing among Shannon, R\'enyi, Tsallis, and more general entropies is often a matter of convention rather than structural principle. We introduce a measure theoretic framework in which admissibility requires the entropy of an input measure to be bounded above by that of its reference measure whenever the former is absolutely continuous with respect to the latter. Under generalized mean-value composition, we characterize all such entropy functionals and obtain a four-level hierarchy determined successively by the mean generator, entropy scale, and additivity assumptions. A continuous strictly monotone generator $g$ is admissible exactly when $t\mapsto g(1/t)$ is strictly convex for increasing $g$, or strictly concave for decreasing $g$. This resolves a question posed by R\'enyi (Proc. 4th Berkeley Sympos. Math. Statist. Prob., 1961) concerning which generalized means may replace the arithmetic mean in his entropy axiomatization. The same criterion is equivalent to strict convexity of an associated Csisz\'ar $f$-divergence generator and therefore yields data processing under Markov kernels with an exact equality condition. Within this hierarchy, product additivity singles out the R\'enyi family, while internal additivity, or product additivity together with arithmetic mean-value composition, singles out Shannon entropy. The characterization is constructive and yields new admissible entropy and divergence families, including integral-transform examples.
We consider coupled stochastic systems decomposed into exterior, boundary, and interior variables, with the boundary variables sometimes carrying the directed structure of a sensor and actuator. The central question is when the conditional law of histories factorises, and how this path space statement is detected by log likelihoods, by Girsanov changes of measure, and by information theoretic quantities used in nonequilibrium statistical physics. The basic object is a regular conditional probability on a path space. Under domination by clamped reference laws, the boundary property becomes multiplicative separation of a Radon--Nikodym derivative, or equivalently additive separation of a path log likelihood. For It\=o diffusions this log likelihood is computed by Girsanov's theorem; its expectation is the quadratic control energy appearing in the F\"ollmer entropy identity and in the stochastic control formulation of Schr\"odinger bridge problems. When exact factorisation fails, the remaining coupling is measured by conditional mutual information, namely the relative entropy between the true boundary-conditioned path law and the product of its conditional marginals. This gives a common language for boundary screening, path likelihood inference, controlled changes of path law, and the thermodynamic value of mutual information.
Information engines exploit feedback to extract work from thermal fluctuations, extending the second law of thermodynamics through information-theoretic bounds. While several such bounds have been proposed, their relative performance under realistic conditions---where measurements are noisy and feedback is temporally correlated---remains largely unclear. Here, we experimentally and theoretically investigate this problem in an underdamped feedback-controlled system with Markovian measurements but non-Markovian control sequences. We compare three representative bounds derived from transfer entropy, unavailable information, and Markovian mutual information, and find that none is universally optimal. Instead, measurement noise preferentially affects information measures that rely on detailed trajectory statistics, while leaving quantities based on instantaneous correlations comparatively robust. As a consequence, trajectory-dependent bounds deteriorate rapidly, giving rise to a crossover in which the Markovian mutual-information bound becomes tighter than the unavailable-information bound over a broad range of measurement noise. Our results reveal a general limitation of information-theoretic descriptions that rely on detailed trajectory statistics in realistic settings and provide a unified perspective on information thermodynamics beyond idealised feedback protocols.
Natalia Ruiz-Pino, Ludovic Bellon, Antonio Prados· 0 citations
We use cookies to run the site and, with your consent, for analytics and to show ads.
See our Cookie Policy.
