Behavior Cloning (BC) of expert controllers for multiscale dynamical systems is challenging: state variables evolving at widely separated timescales induce highly nonlinear couplings that a shared latent space may fail to adequately represent, degrading closed-loop generalization. We propose Structured Representation Learning for Behavior Cloning (SRLBC), an architecture designed to leverage the physical knowledge of multiscale systems in which slow state variables, fast (stiff) state variables, and predictable exogenous inputs evolve at different timescales. On 500 long-horizon roll-outs of a pressurized water reactor (PWR) load-following benchmark -a high-dimensional, stiff industrial system with challenging safety constraints -and across 20 training seeds, separated embeddings systematically outperform shared embedding baselines on cumulative cost, imitation accuracy and constraint satisfaction. The resulting standalone policies approach NMPC closed-loop performance at a computational cost 15 times lower on average. Since standalone policies still leave a few percent of trajectories infeasible, we embed the learned policies in a safety-augmented approximate NMPC framework, in which the optimizer is only called when the predicted input sequence is not certified as safe. This yields controllers still 6.8 times faster than the NMPC controller on average, with a feasibility (i.e., constraint satisfaction) of at least 99.8% for all architectures, SRLBC retaining the lowest closed-loop cost. Under distribution shift, standalone BC policies degrade sharply: SRLBC degrades the least, and once safety-augmented, it retains a feasibility above 98% with the lowest closed-loop cost. Beyond the quantitative results, Jacobian-based sensitivity analysis shows that the resulting controller relies on a sparse and less redundant set of latent features, each attributable to a single group of physical variables, and aligned with the physical timescale decomposition of the system.
Most convergence guarantees for iterative stochastic algorithms are stated at a fixed iteration count and do not justify stopping at a data-dependent time. Recent anytime-valid envelopes do, but deterministic ones are typically too conservative for practical stopping, and existing data-adaptive ones are derived case-by-case, each concerned with a specific algorithm and a specific noise model and boundary construction. This paper provides a general recipe for sharp, data-adaptive anytime-valid envelopes for a large class of iterative algorithms. Our approach has two key steps: an optimization module derives a one-step recursion that performance measures of many such algorithms satisfy, reducing the error analysis to time-uniform concentration for predictably weighted sums of noise. Then, a statistical module ports the sub-$\psi$ condition to vector-valued processes, uniformly over predictable weights, and derives off-the-shelf boundaries via mixtures, stitching, and product betting martingales. Combining the two yields readymade envelopes for many iterative algoithms including but not limited to stochastic gradient descent (SGD), Oja's algorithm, temporal difference learning, Q-learning, and Bures-Wasserstein SGD, many with no prior data-adaptive anytime-valid guarantees. Specifically, for SGD, our framework recovers, often under weaker conditions, or significantly sharpens every existing envelope we are aware of. Our constructions come with explicit implementation details and are substantiated by numerical experiments.
Policy mirror descent (PMD) has a mature finite-time theory in discounted Markov decision processes (MDPs), but less is known in the average-reward setting, a more natural objective for many control applications. We give a finite-time, finite-sample analysis of PMD in ergodic average-reward MDPs built around a single master recursion that governs convergence for any critic, without external regularization. Its specializations yield linear rates for exact, inexact-tabular, and linear function approximation (LFA) updates. For exact PMD, the rate is tight in the worst case, and convergence can become superlinear. We complement these convergence results with critics whose end-to-end sample complexities improve the prior best policy-gradient dependence from $|S||A|\,t_{\mathrm{mix}}^3$ to $|S|\,t_{\mathrm{mix}}^3+|S||A|\,t_{\mathrm{mix}}^2$ in the tabular setting and from $|S||A|\,t_{\mathrm{mix}}^5$ to $d\,t_{\mathrm{mix}}^3$ under LFA. Matching information-theoretic lower bounds establish that both critics are optimal up to logarithmic factors.
We establish complexity lower bounds for stochastic first-order algorithms in nonconvex--concave minimax optimization, allowing algorithms to use variance reduction. Our main contribution is a lower bound for a zero-respecting algorithm class that permits variance reduction, extending beyond the algorithmic restrictions imposed by some existing lower bounds. We consider objectives with an $L$-Lipschitz continuous joint gradient, a compact convex dual domain of Euclidean radius at most $D_Y$, and a primal value function, defined by maximizing the objective over the dual variable, with initial suboptimality at most $\Delta$. The target accuracy $\varepsilon$ is measured by the gradient norm of the Moreau envelope of the constrained primal value function with parameter $1/(2L)$. Under an unbiased stochastic first-order oracle with variance at most $\sigma^2$ and mean-squared smoothness, we prove the lower bound $\Omega\!\left(L^2D_Y\Delta\varepsilon^{-3}+L^3D_Y^2\Delta\sigma^2\varepsilon^{-6}\right)$. This result quantifies the dependence on accuracy, dual-domain radius, and oracle noise even when variance reduction is allowed. We also establish complementary lower bounds for nonconvex--strongly-concave minimax optimization. With dual strong-concavity parameter $\mu>0$ and condition number $\kappa:=L/\mu$, we obtain $\Omega\!\left(L\Delta\sqrt{\kappa}\,\varepsilon^{-2}+L\Delta\kappa\sigma^2\varepsilon^{-4}\right)$ under the bounded-variance oracle model. Under the additional mean-squared smoothness condition with constant $\bar L$, we obtain $\Omega\!\left(L\Delta\sqrt{\kappa}\,\varepsilon^{-2}+\Delta\bar L\sigma\kappa^{3/2}\varepsilon^{-3}\right)$. Together, these results identify complexity barriers across the concave and strongly concave regimes, with the main nonconvex--concave bound remaining valid for algorithms that use variance reduction.
A shortest-path decision is often made from edge costs that are only observed through noisy samples, such as the average of $m$ simulation replications. When the mean cost vectors of different instances lie near a $k$-dimensional subspace, projecting the averaged sample onto that subspace should remove noise. We make this precise. The mean squared error of any projection estimator splits exactly into a variance term and a bias term. The best projector is the positive eigenspace of $\Sigma_c - D/m$, where $\Sigma_c$ is the covariance of the true costs and $D$ is the noise covariance. For a rank-$k$ signal, the projection has the error of a sample mean with a factor $d/(\mu k)$ more replications, where $d$ is the number of edges and $\mu$ is an incoherence constant. For a signal that is only approximately low-rank, every rank-$k$ projection has an error floor, and the plain average wins beyond an explicit replication count. A margin bound shows that the chance of choosing a suboptimal path falls at the same improved rate. A projector learned from $n$ noisy instances loses at most an $O(d/n)$ term.
This monograph provides an introduction to mean field reinforcement learning through the lens of Markov decision processes arising from large-population stochastic control with mean field interactions and common noise. Starting from the connection between multi-agent reinforcement learning and mean field control, it develops the probabilistic, mathematical, and control-theoretic framework needed to formulate representative-agent learning problems, analyze their relationship with finite-population systems, and study both general and linear-quadratic models. The presentation includes dynamic programming principles, propagation-of-chaos limits, and theoretical analyses of tabular Q-learning and policy-gradient methods. It also discusses numerical implementations, including tabular schemes and deep reinforcement learning methods such as deep deterministic policy gradient. The goal is to give readers a coherent bridge between mean field control theory and reinforcement learning methodology, emphasizing the mathematical structure of the problems and the design of tractable learning approaches for large stochastic populations.
We study stochastic density control between Gaussian-mixture endpoint distributions under Brownian prior dynamics. Since the direct Schr\"odinger bridge between Gaussian mixtures is generally not available in closed form, we introduce a lifted path-space construction in which each trajectory is augmented with a source--target component label. Consequently, the problem decomposes into Gaussian component-to-component Schr\"odinger bridges with explicit marginal, drift, and cost formulas, while the mixture-level assignment reduces to a finite-dimensional entropic coupling problem with a Sinkhorn scaling form. We then analyze the projection obtained by discarding or forgetting the label. By construction, the projected law satisfies the original Gaussian-mixture endpoint constraints, but its relative entropy generally differs from the lifted relative entropy by a nonnegative conditional label-information gap. This gap reveals a path-space obstruction: the lifted optimizer cannot, in general, be identified with the direct unlabeled Schr\"odinger bridge after projection. We also derive the posterior-averaged Markov drift associated with the projected marginal flow, prove a kinetic-energy upper bound, and identify a common path-potential condition under which the projection gap vanishes. Several numerical illustrations showing density and shape control are recorded for a self-contained exposition.
We develop a quantitative statistical theory of transformers in the large-context regime by adopting the abstraction of contextual flow maps (CFMs): dynamical systems that evolve a distinguished token in the presence of a contextual measure across a stack of attention blocks. Within this framework, the finite-context model approximates an idealized infinite-context system in which the contextual measure is replaced by its underlying population, so that the context length $n$ becomes a statistical resource. Exploiting the McKean--Vlasov structure of the dynamics and the classical machinery of propagation of chaos, we establish a forward bound controlling the deviation between the finite- and infinite-context CFMs uniformly along depth, and a backward bound controlling the deviation between the corresponding training trajectories uniformly across iterations of online gradient descent. Both bounds achieve the optimal Wasserstein rate $n^{-1/d}$ for general CFMs and parametric rate $n^{-1/2}$ for a restricted class of CFMs that includes transformers as a special case. The analysis rests on a new Eulerian adjoint formulation of the loss gradient and stability estimates for the resulting forward--adjoint system, both of which may be of independent interest.
Most first-order optimizers treat matrix-valued parameters as vectors, ignoring the intrinsic geometry of hidden-layer weights in neural networks. Muon addresses this mismatch by updating along the polar factor of a momentum matrix, but its theoretical understanding has lagged behind practice. In particular, practical implementations incorporate Nesterov momentum, compute the polar factor only approximately, and operate with stochastic gradients that may be heavy-tailed. We close this gap by developing a convergence theory for Muon with Nesterov momentum and inexact polar decomposition in non-convex matrix optimization under heavy-tailed noise. Our analysis builds on a unified framework for inexact polar decomposition that captures practical iterative approximations such as Newton-Schulz and quantifies how their errors propagate through the optimization dynamics. Under uniform bounds on polar errors, we establish an optimal iteration and sample complexity of $O \left(\varepsilon^{\frac{-(3\alpha-2)}{(\alpha-1)}} \right)$ for finding an $\varepsilon$-stationary point, where $\alpha\in(1,2]$~denotes the heavy-tail index. For the inexact-polar setting with $\sigma_1=0$, we also provide guarantees that do not require prior knowledge of $\alpha$. We analyze a randomized low-rank polar decomposition that is substantially more efficient than full-space methods while remaining compatible with our theory. Numerical experiments further demonstrate the effectiveness of the proposed inexact and randomized~variants.
For a risk-averse finite-horizon Markov Decision Process, we introduce a special class of Markov coherent risk measures, called mini-batch measures. We also define the class of multipattern risk-averse problems that reduce to linear MDPs in the risk-neutral case. We use both concepts in a feature-based $Q$-learning method with multipattern $Q$-factor approximation and we prove a high-probability regret bound of $\widetilde{\mathcal{O}}\big(d^{3/2}HN^{H-1}\sqrt{K}\big)$, where $d$ is the number of features, $H$ is the horizon, $N\ge2$ is the mini-batch size, $K$ is the number of episodes, and $\widetilde{\mathcal{O}}$ hides logarithmic factors. We also propose an economical version of the $Q$-learning method that streamlines the policy evaluation (backward) step. The theoretical results are illustrated on a stochastic assignment problem and a short-horizon multi-armed bandit problem.
In this paper we study continuum-marginal optimal transport. Given a time-continuous family of probability marginals, the problem is to recover the minimum-energy velocity field whose flow reproduces every marginal. This problem is the continuum limit of the classical two-marginal Benamou-Brenier formulation, and also the deterministic limit of the Nelson problem of stochastic optimal transport. We propose a practical mesh-free solver for this problem. The weak continuity equation is embedded in a reproducing kernel Hilbert space, yielding a sample-only objective that requires no spatial discretization. The velocity is parametrized by any linear-in-parameters dictionary or neural network, and is optimized by mini-batch stochastic methods. Synthetic experiments confirm that the method achieves accurate drift recovery and marginal consistency. The same computational framework also applies to the Nelson problem.
LLM-based multi-agent planning delegates operational decisions to specialists that communicate through finite reports. In this research, we study its reliability limits when a coordinator can revisit source records before completing a resource-constrained plan. A useful report can support an immediate decision or direct later reading, so its value depends on how these stages communicate. We derive exact communication and access tradeoffs and prove that two report forms suffice for a broad class of allocation requirements, one favoring immediate resource identification and the other distinguishing absence from unnamed availability. For shared capacity, this reduction permits joint optimization of reporting roles and adaptive reads for a fixed number of source types with unequal workloads and reading costs. Identical specialists can optimally report different facts, and maximizing immediate completion can impose an unbounded recovery cost penalty even with full-record access. We also conduct numerical experiments to quantify the benefits of the proposed joint optimization and show how source access improves language-agent decisions that already use their reports optimally. In particular, an exhaustive evaluation of 207,632 report designs on small allocation instances recovers the predicted joint optima. Our analysis suggests that reliable agent communication should be evaluated by the complete decisions it supports and the cost of finishing its unresolved cases, instead of solely by report length, local prediction accuracy, or immediate task completion, as is often the case in current practice.
Inference-time alignment for diffusion models aims to adapt a pre-trained reference diffusion model toward a target distribution without retraining the reference score network, thereby preserving the generative capacity of the reference model while enforcing desired properties at inference time. A central mechanism for achieving such alignment is guidance, which modifies the sampling dynamics through an additional drift term. In this work, we introduce variationally stable Doob's matching, a novel framework for provable guidance estimation grounded in Doob's $h$-transform. Our approach formulates guidance as the gradient of the logarithm of an underlying Doob's $h$-function and employs gradient-regularized regression to simultaneously estimate both the $h$-function and its gradient, resulting in a consistent estimator of the guidance. Theoretically, we establish non-asymptotic convergence rates for the estimated guidance. Moreover, we analyze the resulting controllable diffusion processes and prove non-asymptotic convergence guarantees for the generated distributions in the Wasserstein-2 distance. Finally, we show that variationally stable guidance estimators are adaptive to unknown low dimensionality, effectively mitigating the curse of dimensionality under low-dimensional linear subspace assumptions. Numerical experiments demonstrate the effectiveness of our proposed method and show the advantages compared to the baselines.
Pipeline-parallel distributed optimization is essential for large-scale machine learning but is challenged by significant communication overhead from transmitting high-dimensional activations and gradients between workers. Existing approaches often depend on impractical unbiased gradient assumptions or incur sample-size memory overhead. This paper introduces Clapping, a Communication compression algorithm with LAzy samPling for Pipeline-parallel learnING. Clapping adopts a lazy sampling strategy that reuses data samples across steps, breaking sample-wise memory barrier and supporting convergence in few-epoch or online regimes. Clapping comprises two variants including Clapping-FC and Clapping-FU, both of which achieve convergence without unbiased assumption for compressed gradient, effectively addressing compression error propagation in multi-worker settings. Numerical experiments validate the performance of Clapping across different tasks.
Robust subspace estimation is fundamental to many machine learning and data analysis tasks. Iteratively Reweighted Least Squares (IRLS) is an elegant and empirically effective approach to this problem, yet its theoretical properties remain poorly understood. This paper establishes that, under deterministic conditions, a variant of IRLS with dynamic smoothing regularization converges linearly to the underlying subspace from any initialization. We extend these guarantees to affine subspace estimation, a setting that lacks prior recovery theory. Additionally, we illustrate the practical benefits of IRLS through an application to low-dimensional neural network training. Our results provide the first global convergence guarantees for IRLS in robust subspace recovery and, more broadly, for nonconvex IRLS on a Riemannian manifold.
Group distributionally robust optimization (GDRO) aims to learn models that perform well across $m$ distributions simultaneously. Existing methods either (i) query $m$ samples per round, which imposes strict implementation requirements, or (ii) query only a single sample per round, which requires many rounds to converge and therefore leads to long runtime. To avoid these limitations, we introduce a novel setting called \emph{flexible sample queries} (FSQ), which allows the number of samples that can be queried to vary across rounds. Under the FSQ setting, we cast GDRO as a two-player game: one performs non-oblivious online convex optimization, while the other tackles a non-oblivious prediction with limited advice (PLA) problem. In this game, the first player updates its decision using follow-the-regularized-leader with stochastic gradients. For the second player, we develop a new PLA method equipped with an adaptive exploration strategy that can leverage an arbitrary per-round sample size. We then establish the \emph{first} high-probability regret bound for non-oblivious PLA. By integrating these components, we develop an anytime GDRO algorithm that supports FSQ, and derive a high-probability optimization error bound of $O\left(\frac{1}{t}\sqrt{\sum_{j=1}^t \frac{m}{r_j}\log m}\right)$, where $r_j$ is the sample size at round $j$. This result demonstrates that the optimization error decreases as the per-round sample sizes increase, and implies the same near-optimal sample complexity of $O(m\log (m)/\epsilon^2)$ for any fixed sample size $r\in[m]$.
Many real-world problems, such as those with fairness constraints, involve complex expectation constraints and large datasets, necessitating the design of efficient stochastic methods to solve them. Most existing research focuses on cases with no {constraint} or easy-to-project constraints or deterministic constraints. In this paper, we consider nonconvex nonsmooth stochastic optimization problems with expectation constraints, for which we build a novel exact penalty model. We first show the relationship between the penalty model and the original problem. Then on solving the penalty problem, we present a single-loop SPIDER-type stochastic subgradient method, which utilizes the subgradients of both the objective and constraint functions, as well as the constraint function value at each iteration. Under certain regularity conditions (weaker than Slater-type constraint qualification or strong feasibility assumed in existing works), we establish an iteration complexity result of $O(\epsilon^{-4})$ to reach a near-$\epsilon$ stationary point of the penalized problem in expectation, matching the lower bound for such tasks. Building on the exact penalization, an $(\epsilon,\epsilon)$-KKT point of the original problem is obtained. For a few scenarios, our complexity of either the {objective} sample subgradient or the constraint sample function values can be lower than the state-of-the-art results by a factor of $\epsilon^{-2}$. Moreover, on solving two fairness-constrained problems and a multi-class Neyman-Pearson classification problem, our method is significantly (up to 466 times) faster than the state-of-the-art algorithms, including switching subgradient method and inexact proximal point methods.
We consider the infinite-horizon average-reward restless bandit problem. We propose a novel two-set policy that maintains two dynamic subsets of arms: one subset of arms has a nearly optimal state distribution and takes actions according to an LP-Priority routine; the other subset of arms is driven towards the optimal state distribution and gradually merged into the first subset. We show that our two-set policy is asymptotically optimal with an $O(\exp(-C N))$ optimality gap for an $N$-armed problem, under the assumptions of aperiodic-unichain, non-degeneracy, and local stability. Our policy is the first to achieve exponential asymptotic optimality under the above set of easy-to-verify assumptions, whereas prior work either requires a global attractor assumption or only achieves an $O(1/\sqrt{N})$ optimality gap. We further discuss obstacles in weakening the assumptions by demonstrating examples where exponential asymptotic optimality is not achievable when any of the three assumptions is violated. Notably, we prove a lower bound for a large class of locally unstable restless bandits, showing that local stability is particularly fundamental for exponential asymptotic optimality. Finally, we use simulations to demonstrate that the two-set policy outperforms previous policies on certain restless bandit problems and performs competitively overall.
We establish mathematical foundations for optimal spatiotemporal vector fields governing continuous score-guided probability flows over bounded submanifolds $M \subset SE(3)$ on the Special Euclidean group $SE(3) \cong \mathbb{R}^3 \times SO(3)$. Modeling dynamic density evolution via a score-perturbed Lie-algebraic continuity equation under zero-flux boundary conditions, we resolve the non-local score drift singularity $s(g,t) = \nabla_{SE(3)} \log \rho(g,t)$ as $\rho \rightarrow 0^+$ via weighted parabolic flux dissipation bounds. Using left-invariant exponential group mollifiers, we derive boundary-preserving DiPerna--Lions commutator estimates tracking structure constants $\{c_{ij}^k\}$. Second, interpreting the optimality system as a 7D space-time elliptic operator $L_{7D} = -\Delta_{SE(3)} - \partial_t^2$ with zero-order corner compatibility on $\partial M \times \{0,T\}$, we establish Agmon--Douglis--Nirenberg $W_p^2(Q_T)$ shift estimates tracking Lie algebra constants, boundary geometry $\|\partial M\|_{C^2}$, and Ricci curvature $\|Ric_{SE(3)}\|_{L^\infty}$, yielding a full $C^\infty(\overline{Q}_T)$ Schauder bootstrap expansion. Third, we extend this system to non-cooperative Markov Mean-Field Games (MFGs) on $SE(3)$, establishing score-guided Nash equilibria via Schauder--Tychonoff fixed points. Crucially, the constant diagonal Ricci curvature $Ric_{SE(3)} = \mathrm{diag}(0,0,0,1/2,1/2,1/2)$ from $SO(3)$ induces a sharp Bochner--Weitzenb\"ock spectral gap boost $\lambda_1(-\Delta_{SE(3)}) \ge \lambda_1(-\Delta_{\mathbb{R}^3}) + 1/2 > 0$. This positive rotational curvature enhances the Lie-algebraic Poincar\'e coercivity constant, demonstrating that intrinsic group curvature stabilizes score-guided flows against dispersion and expands the critical Nash uniqueness horizon $T_{SE(3)}^*$ beyond flat Euclidean limits.
Best subset selection is a central problem in statistical regression, drawing increasing attention from the broader optimization and machine learning communities. Solving the hard L0 sparsity constraint directly has been computationally challenging; as a result, most existing work proposes formulations which do not enforce the constraint, but instead use a mix of L0/L1/L2 penalties to encourage sparsity. Modern methods for solving such penalized formulations are tractable for p = 100,000, but give up direct control over sparsity. In this paper, we propose a two-stage proximal gradient descent-based approach which both enforces the sparsity constraint and incorporates L1/L2 penalties, leveraging GPUs for computational acceleration to find approximate solutions. Our method, ParaDeAl, offers the user the ability to select the problem hyperparameters, solves all configurations in parallel, and obtains solutions up to 15x faster than the best state-of-the-art CPU competitors while achieving almost identical test accuracy for the same support size for large (p > 100,000) synthetic and real-world datasets.