Mingyuan Xu, Zongqi Xia, Tianxi Cai, Doudou Zhou, Nian Si
Learning a common policy from multiple source environments requires accounting for cross-source heterogeneity, and the offline setting introduces an additional challenge of limited data coverage. We formulate this problem as a distributionally robust Markov decision process under a shared linear MDP representation with source-specific transition and reward components. The uncertainty set is constructed through feature-wise mixtures of these components, yielding a $d$-rectangular structure that captures cross-source heterogeneity and admits a robust Bellman recursion. Building on this formulation, we develop Group-Robust Pessimistic Value Iteration (GR-PEVI), an offline policy-learning algorithm that combines feature-wise worst-case aggregation with a data-dependent pessimism penalty. Under a robust partial coverage condition, we establish finite-sample suboptimality guarantees for GR-PEVI and extend the analysis to clustered sources, characterizing the trade-off between the statistical gain from pooling similar sources and the approximation error induced by within-cluster heterogeneity. Simulation studies demonstrate the effectiveness of GR-PEVI, with empirical behavior consistent with the theoretical analysis. We apply GR-PEVI to a multi-source electronic health record study of multiple sclerosis treatment decisions to illustrate its use in a clinical setting.
Valentin Leplat, Daniil Merkulov, Aleksandr Katrutsa, Daniel Bershatsky, Olga Tsymboi, Ivan Oseledets
Stochastic first-order methods are widely used for large-scale optimization and machine learning, where their performance often depends on the learning rate. We study NAG-GS, a stochastic momentum method based on a semi-implicit Gauss--Seidel discretization of a Nesterov-type flow. For strongly convex quadratics, we determine the exact range of stable step sizes for fixed damping and the step size that minimizes the worst-case spectral radius. When the damping equals the strong-convexity parameter, the method recovers the accelerated asymptotic factor $(\sqrtκ-1)/(\sqrtκ+1)$, where $κ$ is the condition number. We also prove that the quadratic iteration converges linearly with dynamic damping when its limiting fixed-damping iteration is stable. For general strongly convex, possibly nonsmooth objectives, we establish an $O(1/k)$ bound on the expected objective gap of the last iterate, using a conditionally unbiased stochastic first-order oracle with bounded second moment. The result holds with dynamic damping from any $γ_0>0$. We show that NAG-GS can also be written as a Heavy-Ball iteration with explicitly determined step sizes and momentum coefficients. We test the stochastic result on nonsmooth hinge-loss classification problems, and further evaluate NAG-GS on quadratic problems, logistic regression, convolutional neural networks, RoBERTa, and Vision Transformers. The experiments show that NAG-GS can remain effective at learning rates where standard optimizers become unstable or lose accuracy.
James Hazelden, Laura Driscoll, Eli Shlizerman, Eric Shea-Brown
In training a neural network with gradient descent (GD), each iteration induces a linear operator that governs the first-order updates to a model's internal state variables. We call this operator the Global Empirical Neural Tangent Kernel (GeNTK). In finite-width networks, GeNTK is typically intractable to form, leading prior work to focus on restrictive settings, such as tracking only outputs or infinite-width limits. Here, we bridge this gap by studying the structure of GeNTK across many models. Formulating the model state as the solution to a global implicit constraint, we derive GeNTK as a product of two operators: K, accounting for immediate parameter-to-state interactions, and P, describing internal state-to-state dependencies. For a broad class of "weight-based models", such as RNNs and transformers, we prove a universal Kronecker-core theorem showing that K admits an exact, forward-pass-computable form given by the Gram matrix of weight-site variables. This core gives a measurable state-dependent restriction on Euclidean GD, termed "self-referential bias": current activity filters which adjoint directions can drive a first-order update. For recurrent models (GRUs and RNNs), we examine the spectrum of GeNTK and formalize notions of low-rank, biased learning in terms of reduced spatial or temporal ranks of the full operator. Practically, this explains why some RNN initializations yield better learning, while some plateau, by biasing the GeNTK and thus restricting learning to a small domain of learnable tasks. To demonstrate computational viability, we validate the main theorems in a self-attention transformer and a pretrained Pythia transformer language model with 410 million parameters. Overall, our results expose tractable local rank and mode-selective biases in the finite-width GD geometry and open the door to principled study of the GeNTK, a bridge between theory and practice.
The curse of dimensionality associated with solving the Hamilton-Jacobi-Bellman equation for the value function remains a key obstacle for real-time optimal feedback control of high-dimensional nonlinear dynamical systems. This is especially severe in aerospace applications such as relaminarizing unsteady turbulent flows and recovering from aerodynamic stall where nonlinear dynamics play a significant role. We introduce Balanced Active Subspaces for Optimal Control (BASOC), a model reduction framework which seeks to find directions in state space along which the value function is most sensitive, and the distribution of states exhibits maximum variance along optimal trajectories. BASOC leverages trajectory optimization and the Pontryagin minimum principle to obtain state and co-state (value function gradient) data along locally optimal trajectories for constructing state and gradient covariance matrices. Balancing these matrices provides an oblique projection for approximating the value function, which can also be used to construct reduced-order models and surrogate feedback control laws. We show that for linear-quadratic optimal control problems, BASOC becomes equivalent to a snapshot method for initial condition balancing of the optimally controlled closed-loop realization. The framework is demonstrated on a nonlinear three-dimensional system where POD and balanced truncation are known to perform poorly, and on the nonlinear complex Ginzburg-Landau equation to demonstrate feasibility for larger-scale systems.
Large-scale linear programs with millions of variables and constraints strain time and memory budgets. First-order methods such as the primal-dual hybrid gradient scale to this regime because they avoid matrix factorizations, but they require many iterations. Learned accelerators can reduce these iterations, yet fewer iterations do not guarantee a shorter solve. Learning inference also takes time, and running the accelerator longer trades solver time for inference time. Existing unrolled accelerators handle this trade-off poorly: each layer processes every variable and constraint, the predicted iterates still need solver refinement, and every instance runs the same layers. We therefore ask: What neural accelerator design achieves end-to-end speedups on large-scale linear programs across sizes and target accuracies? We introduce AHEAD, an accelerator that predicts step sizes for first-order linear programming solvers while keeping their update rules. Its weight-shared predictor reads compact solver statistics and rescales the primal and dual step sizes, so each warm-start iteration stays inexpensive and a short warm-start phase hands the solver a warm start. A learned handoff policy sets the number of warm-start iterations for each instance and target accuracy, after which the solver finishes under its own termination criterion. On packing linear programs with up to $10^7$ variables, AHEAD achieves $6.5$--$11.6\times$ end-to-end speedups over Gurobi's first-order solver from a cold start, including on instances of unseen sizes and distributions. The gains extend to PageRank linear programs and hold across difficulty levels and target accuracies.
We propose an optimization-informed deep neural network approach, named iUzawa-Net, aiming for the first solver that enables real-time solutions for a class of nonsmooth optimal control problems of linear partial differential equations (PDEs). The iUzawa-Net unrolls an inexact Uzawa method for saddle point problems, replacing classical preconditioners and PDE solvers with specifically designed learnable neural networks. We prove universal approximation properties and establish the asymptotic $\varepsilon$-optimality for the iUzawa-Net, and validate its promising numerical efficiency through nonsmooth elliptic and parabolic optimal control problems. Our techniques offer a versatile framework for designing and analyzing various optimization-informed deep learning approaches to optimal control and other PDE-constrained optimization problems. The proposed learning-to-control approach synergizes model-based optimization algorithms and data-driven deep learning techniques, inheriting the merits of both methodologies.
Estimation error in machine learning prognostic models can propagate into poor maintenance decisions. We study a data-driven predictive maintenance problem coupling health-condition estimation with a joint ordering-replacement model. We show that the conventional Estimate-Then-Optimize (ETO) approach yields inconsistent, suboptimal decisions for a broad class of such problems. To bridge the gap, we propose an Integrated Estimate-Optimize (IEO) framework that directly targets maintenance cost as the prognostic learning objective, bypassing surrogate statistical accuracy metrics such as Kullback--Leibler divergence or mean squared error. Theoretically, we prove out-of-sample performance guarantees of the IEO framework under certain conditions, including i.i.d. and sliding-window data-generating processes. Computationally, we develop a training procedure combining Monte Carlo dropout, central-difference gradient approximation, and an ETO warm-start to handle discrete maintenance decisions and nonlinear costs. Experiments on the four CMAPSS subsets demonstrate that, under the nominal setting, IEO reduces mean maintenance cost by 14.8--22.7% relative to ETO and by 7.2--11.8% relative to a decision-oriented baseline, while also reducing failure rates and cost variation. These improvements are obtained despite higher prediction errors. Increasing the ratio of unplanned-downtime costs to planned-downtime costs makes this tradeoff more pronounced: IEO shifts predictions further toward RUL underestimation, accompanied by greater cost savings and fewer failures. These findings support aligning prognostic learning with maintenance objectives when imperfect predictions must guide decisions under costly unplanned downtime.
Recently, rectified flow has emerged as a fundamental framework for large-scale image generation, powering state-of-the-art systems such as FLUX.1 and Stable Diffusion 3. Despite its remarkable empirical success, the computational and statistical guarantees of iterative rectified flow have remained largely unexplored. We address this problem by studying c-rectified flow, a cost-aware class of rectified flow that projects velocity fields onto a gradient class while preserving endpoint marginals. The ordinary rectified flow can fail to recover the optimal transport coupling: in a Gaussian case study, the iteration converges to the optimal coupling if and only if the source and target covariance matrices commute. More generally, for a Gaussian to log-concave coupling, the ordinary rectified flow can reach a fixed point after one step from independent initialization. In contrast, under suitable regularity and uniform-integrability assumptions, iterative c-rectified flow always converges to the optimal transport coupling. We further establish quantitative one-step contraction and exponential convergence guarantees under projection-stability assumptions for both quadratic and strongly convex displacement costs. Finally, under a Hölder ball assumption, we develop new minimax-optimal score estimation rates and show that, when combined with iterative c-rectified flow, they yield a rate-optimal estimator of the optimal transport for the dimension \(d \ge 3\) and a nearly parametric rate for \(d=1,2\).
We study full-information online linear optimisation on the unit $\ell_p$ ball in $\mathbb{R}^d$, for fixed $2\le p<\infty$ and losses of dual norm at most one. Let $Δ_T$ denote the cumulative loss of the best fixed action above each round's minimum. A follow-the-regularised-leader (FTRL) algorithm attains, on every loss sequence and without knowing $T$ or $Δ_T$, regret $O_p(1+\min\{Δ_T^{1-1/p},\, d^{1/2-1/p}\sqrt{Δ_T}\})$. Matching lower bounds show that this rate is minimax for each fixed $p$, and its constant $(2+o(1))p/\log p$ is optimal as $p\to\infty$ among guarantees uniform in $d$ and $T$. The algorithm uses a bounded radial regulariser and switches at most once to its quadratic augmentation. On some paths where follow-the-leader (FTL) has regret one, its regret grows with $T$. A switching extension keeps the sharp constant and also guarantees $\mathrm{Reg}_T\le 2\,\mathrm{Reg}_T^{\mathrm{FTL}}$. For i.i.d. losses whose mean has no zero coordinates, we compute the exact $\log T$ coefficient of the expected regret, which separates FTL's variance term from the bias of the regulariser.
Nicolás García Trillos, Adam Quinn Jaffe, Bodhisattva Sen
The quantity of interest in the classical Cramér-Rao theory of unbiased estimation (e.g., the Cramér-Rao lower bound, its exact attainment for exponential families, and asymptotic efficiency of maximum likelihood estimation) is the variance, which represents the instability of an estimator when its value is compared to the value for an independently-sampled data set from the same distribution. In this paper we are interested in a quantity which represents the instability of an estimator when its value is compared to the value for an infinitesimal additive perturbation of the original data set; we refer to this as the "sensitivity" of an estimator. The resulting theory of sensitivity is based on the Wasserstein geometry in the same way that the classical theory of variance is based on the Fisher-Rao (equivalently, Hellinger) geometry, and this insight allows us to determine a collection of results which are analogous to the classical case: a Wasserstein-Cramér-Rao lower bound for the sensitivity of any unbiased estimator, a characterization of models in which there exist unbiased estimators achieving the lower bound exactly, and various results on the asymptotic achievement of the lower bound. More precisely, we show that Wasserstein projection estimators (and one-step approximations thereof) are asymptotically sensitivity-efficient under suitable assumptions, and that there exist Hodges-type estimators which exhibit super-sensitivity-efficiency. We use these results to treat many statistical examples, sometimes revealing new optimality properties for existing estimators and other times revealing entirely new estimators.
Zhirayr Tovmasyan, Slavomír Hanzely, Egor Shulgin, Peter Richtárik, Sebastian U. Stich
One of the main challenges of federated learning is heterogeneous data split among the clients, leading to biased local updates called client drift. The SCAFFOLD algorithm (Karimireddy et al., 2020) corrects client drift with control variates, but its relation to the similarity of client data is not yet fully understood: the original similarity-based guarantees cover only quadratics with a stronger second-order similarity assumption, and a follow-up weakly convex analysis (Luo et al., 2025) obtains a $\sqrt{Lδ}$ dependence for $L$-smooth losses with averaged Hessian similarity $δ$. We analyze SCAFFOLD with minibatch-based control variates and establish linear dependence on Hessian similarity beyond quadratics, for both convex and nonconvex objectives. In particular, for convex objectives with exact gradients, $R$ SCAFFOLD rounds with $K$ local steps lead to an objective gap of order $(L/K+δ)D_0^2/R$, where $D_0$ is the initial distance to a minimizer. When $K=Θ(L/δ)$, it improves the minibatch SGD rate by a factor $Θ(L/δ)$. For nonconvex objectives, we prove an analogous stationarity guarantee under uniform Hessian similarity for a client-selection variant. In both settings, we extend these guarantees to stochastic gradients and prove matching lower bounds for the analyzed methods with constant local stepsizes. We also show that, under partial participation, even identical clients need not obtain the full communication benefit of additional local steps.
We formulate flow matching on probability measures over probability measures. Directly from the Wasserstein-on-Wasserstein (WoW) continuity equation, we show that admissible nested outer and inner couplings induce a valid metameasure flow and non-local velocity field. Validity requires neither optimal plans nor a smooth manifold. Practically, this opens up the design space of multisample flow matching: exact Wasserstein plans are distinguished by energy and straightness, while sliced and lazy linear plans yield scalable alternatives. Point-cloud experiments, including controlled Bures-Wasserstein and Euclidean minibatch baselines, demonstrate the resulting geometry-cost trade-off.
We study Multi-Level Aggregation with Maximum Delay (MLA-MD) on finite weighted rooted trees. Each service pays for a rooted subtree and for the maximum waiting time among the requests it clears. For the offline problem, we give an exact polynomial-time algorithm that computes both the optimum and an optimal service schedule. For the online problem, we give a deterministic $2$-competitive algorithm and, against an oblivious adversary, a randomized $e/(e-1)$-competitive algorithm. Both competitive ratios are optimal, even on rooted paths. We also extend the offline result and both online guarantees to service systems with submodular costs.
By virtue of the sheer compute demands of large-scale artificial intelligence (AI) and high-performance computing (HPC), modern multi-core microprocessors suffer from severe power density limits and thermal throttling, entailing an urgent need for principled methods for predictive, physics-informed dynamical thermal management (DTM). Here, we introduce a hardware/software co-design physics-informed PDE-constrained DTM throttling optimal control algorithm, ADACT (Adjoint-Driven Adaptive Co-Design Thermal Control), that minimizes thermal risk and cooling energy overhead while maximizing core execution throughput. The underlying physical temperature field is governed by a two-dimensional nonhomogeneous parabolic reaction-diffusion equation with spatially localized, time-varying power source terms and variable convective cooling subject to nonhomogeneous Dirichlet boundary conditions. We employ a discretization of the continuous state PDE via a Crank-Nicolson finite difference scheme with a 5-point discrete Laplacian operator on a uniform spatial mesh, which preserves unconditional numerical stability and $O(Δx^2 + Δy^2 + Δt^2)$ second-order temporal and spatial accuracy. To efficiently compute exact gradients, we construct and implement a continuous backward-in-time adjoint state method with L-BFGS as our optimization algorithm of choice. The accuracy and second-order convergence rates of both forward state and backward adjoint solvers are quantitatively verified via the Method of Manufactured Solutions (MMS), and gradient accuracy is validated using directional derivative Taylor tests. Numerical experiments demonstrate that ADACT dynamically throttles software power workloads and modulates hardware cooling efforts to strictly uphold physical thermal thresholds while maintaining operational throughput.
Constraint learning is an emerging paradigm of modeling optimization problems with the help of machine learning. In this paradigm, trained machine learning models are embedded into the optimization model. For example, we can learn constraints empirically from available data, or we can avoid using a nonlinear objective function by learning a simpler approximation for it. Neural networks become a compelling option if the machine learning models must capture more elaborate and nuanced relationships. However, optimization models embedding neural networks may become too large or too slow to solve. In this tutorial, we present practical considerations for an effective and efficient use of constraint learning with neural networks and validate it with a case study highlighting the benefit of smaller and sparser networks.
Alberto De Marchi, Yura Malitsky, Adrien B. Taylor
We derive convergence guarantees for mirror descent and proximal mirror descent with a logarithmic barrier as the distance-generating function. When the solution lies on the boundary, the Bregman divergence blows up and classical guarantees become vacuous. While existing remedies recover $O(\log k/k)$ rates by adapting classical bounds, we propose a more direct route in which the Bregman distance to the solution never appears, and which achieves the same rate with constant stepsize. We also construct an example showing that this rate cannot be improved in general. Finally, we compare the proposed approach with interior-point methods.
High-fidelity digital twins in continuous biomanufacturing are often bottlenecked by incomplete kinetic models, while standard data-driven algorithms (e.g., SINDy, PINNs) lack interpretability or impose rigid physical assumptions. This paper introduces a hybrid closed-loop architecture for autonomous kinetic state-space discovery driven by a Large Language Model (LLM) agent. The framework employs a sequential "isolation-peeling" methodology across orthogonal operational regimes. This algorithmic peeling mathematically decouples entangled multiplicative (inhibitory) and additive (catalytic) biological rate law. Grounded in a Two-Level Retrieval-Augmented Generation (RAG) library, the framework utilizes Moving Horizon Estimation (MHE) and logarithmic decomposition to decouple complex cell growth kinetics into orthogonal multiplicative (inhibitory) and additive (regulatory) residual channels. When statistical degeneracy occurs an Active Design of Experiments (DoE) controller dynamically triggers targeted perturbations to decorrelate competing model hypotheses. Evaluated on a continuous stirred-tank reactor exhibiting unmodeled Haldane inhibition, Luong product toxicity, and secondary metabolite autocatalysis, the framework autonomously expanded the state space and recovered ground-truth parameters with under 6% error without a priori kinetic knowledge.
We develop a unified framework for assortment optimization and matching under multinomial logit (MNL) choice, yielding simple algorithms and analyses across a broad class of problems. The framework centers on a single-customer assortment problem in which the firm commits to eligible products before their independent availabilities are revealed. A simple threshold policy achieves a tight $1/2$ guarantee against a benchmark that selects its assortment after observing availability and may control dependence subject to the same marginal probability bounds. This guarantee accommodates cardinality constraints and adversarial removal of excess available products. Combined with problem-specific linear programming relaxations, the framework yields $1/2$ guarantees for standard two-sided assortment optimization, customization under mixtures of MNL models, ride-hailing notifications, multi-round broadcast matching, and MNL prophet inequalities, and a $1/4$ guarantee for fully static two-sided assortment optimization. Across these applications, we improve approximation factors, extend guarantees to more general revenues and cardinality constraints, and retain existing factors against stronger benchmarks.
Juan Ramirez, Meraj Hashemizadeh, Simon Lacoste-Julien
Deep learning models are increasingly subject to explicit requirements, or constraints, such as safety levels, fairness limits, and sparsity budgets. In practice, such requirements are often handled through fixed weighted-sum penalization: the constraints are added to the task loss with fixed coefficients, and the resulting scalarized objective is minimized. This position paper argues that fixed penalization is often ill-suited to deep learning problems with non-negotiable requirements, for three main reasons. First, in non-convex settings, the penalized and constrained problems are not, in general, equivalent, so a minimizer of the penalized objective need not solve the constrained problem. Second, a penalty coefficient sets an objective-constraint trade-off rather than the constraint level, so finding an acceptable coefficient often takes trial and error, each trial a full training run. Third, fixed penalization weakens hard requirements into soft penalties to be traded off against task performance. As a result, compliance is no longer built into the problem, and requirements can be quietly relaxed. We therefore argue that, when a deep learning problem specifies non-negotiable requirements, one should solve the constrained problem itself instead of a surrogate defined by fixed penalization.
Modern machine learning relies on practical optimization methods that often differ from the algorithms studied in classical analyses: they train submodels, clip or normalize gradients, use layer-wise non-Euclidean directions, or approximate costly inner computations. This thesis analyzes the objective, assumptions, update, and computation defined by these procedures, rather than those of simplified proxies. For submodel training, we show that Independent Subnetwork Training can converge to a shifted fixed point because sketching before differentiation biases the update for the original objective. Model-Agnostic Sparsified Training instead incorporates the sketch distribution into its objective, which makes sketched gradients unbiased and admits convergence guarantees. For private optimization, we show that adaptive quantile clipping retains bias at any fixed quantile, while suitable quantile schedules recover near-optimal nonconvex rates under heavy-tailed noise. We then combine smoothed normalization with error compensation and obtain distributed private convergence guarantees under standard assumptions, without assuming bounded gradients. For optimizers based on linear minimization oracles, we develop and validate a layer-wise generalized smoothness model and build on it the Gluon framework for Muon- and Scion-style methods. We also show that the approximate orthogonalization in Muon matters: less precise orthogonalization calls for smaller stepsizes, so its precision should be tuned together with the learning rate. Overall, practical methods are better understood when theory retains the features that determine their optimization behavior.
We develop and analyze multilevel gradient-regularized Newton methods for unconstrained convex optimization with Lipschitz-continuous Hessians. By computing Newton corrections in lower-dimensional subspaces, the proposed methods aim to reduce computational costs while preserving the global $\mathcal{O}\left(k^{-2}\right)$ convergence rate in function values of the full-space method. To establish this rate for multilevel methods, we introduce a class of row-nested restriction operators that ensures monotonicity of the reduced-gradient norms across levels. Numerical experiments with row-nested operators demonstrate convergence behavior comparable to full-space gradient-regularized Newton, while offering substantial computational savings on large-scale instances.
L. Mainini, A. Serani, H. Pehlivan, F. Di Fiore, M. P. Rumpfkeil, E. Minisci, D. Quagliarella, S. Yildiz 等 15 位作者
This paper introduces a comprehensive benchmarking framework that includes: (i) a suite of analytical benchmark optimization problems designed to stress-test and validate multifidelity methods; (ii) a set of assessment metrics for quantifying and comparing performance over measurable objectives; and (iii) the classification, evaluation, and comparison of several families of multifidelity optimization methods and frameworks using the proposed benchmarks to identify their respective strengths and weaknesses in real-world scenarios. The proposed benchmark problems are analytically defined functions carefully selected to capture mathematical challenges commonly encountered in real-world applications, including high dimensionality, multimodality, discontinuities, and noise. Their closed-form nature ensures computational efficiency, high reproducibility, and a clear separation of algorithmic behavior from numerical artifacts. The accompanying performance metrics support the systematic evaluation of multifidelity methods, measuring both optimization effectiveness and global approximation accuracy. By providing a rigorous, reproducible, and accessible benchmarking framework, this work aims to enable the broader community to understand, compare, and advance multifidelity optimization methods for complex problems in science and engineering.
Younes Naceur, Jie Wang, Victor Magron, Antonio Acín
Semidefinite relaxations of polynomial optimization have become a central tool for addressing the non-convex optimization problems over non-commutative operators that are ubiquitous in quantum information theory and, more in general, quantum physics. Yet, as these global relaxation methods are solved numerically, the bounds issued by the semidefinite solver can - and often do - exceed the global optimum, undermining their certifiability. To counter this issue, we introduce a framework for extracting rigorous certified bounds on non-commutative optimization problems from inexact numerical data, and apply it to several paradigmatic problems in quantum information theory. An extension to sparsity and symmetry-adapted semidefinite relaxations is also provided and compared to the general dense scheme. Our results establish rational post-processing as a practical route to reliable certification, pushing semidefinite optimization toward a certifiable standard for quantum information science.
Ioannis Anagnostides, Ioannis Panageas, Nikolas Patris, Tuomas Sandholm
Follow the regularized leader (FTRL) is a foundational algorithm in online learning whose regret properties have been extensively studied for decades. However, in stark contrast to mirror descent---and gradient descent in particular---the convergence of FTRL to first-order stationary points in constrained nonconvex optimization was not well understood. In this paper, we show that continuous-time FTRL can fail to converge even asymptotically in infinitely smooth constrained optimization problems: its Karush-Kuhn-Tucker (KKT) gap can remain bounded away from zero indefinitely. This is especially surprising in light of the fact that continuous-time FTRL guarantees a monotonic improvement of the objective value. As a result, we establish that non-convergent behavior of no-regret dynamics---which besets general variational inequality problems---can occur even in potential systems. From a technical standpoint, our construction relies on an infinitely differentiable function where the gradient flow dynamics exhibit decelerating oscillations without ever converging pointwise, which we in turn embed into higher-dimensional FTRL dynamics that sustain a perpetually large KKT gap. We complement our negative result by showing that FTRL with Euclidean regularization does converge when the objective function is analytic. This shows that the analytic property plays a more fundamental role for FTRL than for gradient flow: it governs not only pointwise convergence, but convergence to first-order stationarity itself.
We consider the well-studied setting of minimizing a convex Lipschitz function using either gradient descent (GD) or its stochastic variant (SGD), and examine the last iterate convergence. We prove that $Θ(\sqrt{\log(T)/T})$ is the optimal anytime rate attainable with a fixed stepsize schedule (that does not depend on the total number of steps $T$). The upper bound holds already for SGD, and improves on the $\log(T)/\sqrt{T}$ rate of Shamir and Zhang [2013], as well as the $\sqrt{\log(T)/T}$ rate of Zamani and Glineur [2023] which was proven only for deterministic GD. The matching lower bound holds already for GD, and confirms the conjecture of Jain et al. [2019], namely that the optimal $1/\sqrt{T}$ rate for the last iterate cannot be obtained without prior knowledge of $T$.
We study continuous-time stochastic linear--quadratic control with a scalar state and nonnegative control constraints. With quadratic running costs, randomized exploration can change the optimal mean feedback. We introduce an auxiliary problem without entropy in which the conditional policy mean is nonnegative and the covariance is prescribed in proportion to the squared state. Moment reduction and state separation yield two regional Riccati equations and characterize the finite-horizon value function and optimal conditional mean. We give a sufficient control--cost compatibility condition under which this mean coincides with the optimal feedback of the original constrained problem. For time-invariant coefficients, we characterize the undiscounted infinite-horizon optimum and establish mean square asymptotic stability under the stated assumptions. For unknown dynamics, we develop model-free actor--critic schemes for finite and infinite horizons. Numerical experiments compare the learned policies with their analytical targets and examine the effect of exploration on the optimal mean.
Zhen-Fan Wang, Kai Zhang, Xianping Du, Zhaolong Han, Dai Zhou, Onur Bilgen
Wind farm layout optimization becomes increasingly difficult as turbine count grows, especially when high-fidelity aerodynamic models are used. Conventional Kriging-based Bayesian optimization often suffers from low sample efficiency in high-dimensional layout spaces because sparse training data provide limited domain coverage. We develop a Graph-informed Kriging framework that embeds reusable aerodynamic interaction information into the surrogate mean function. The wind farm is represented as a directed aerodynamic graph, where pairwise turbine interactions defined in a low-dimensional relative-coordinate space are assembled into a farm-scale mean. This structured mean captures the dominant variation in annual energy production (AEP), allowing the Gaussian process to model only the residual structure and uncertainty. Numerical experiments show that Graph-informed Kriging consistently improves AEP prediction accuracy over ordinary and generic-trend Kriging under sparse-data conditions, with increasing benefits as layout dimension and wake-interaction strength grow. Embedded in Bayesian optimization, it substantially improves sample efficiency and achieves competitive solutions for the IEA Wind Task 37 benchmark with about 280 AEP evaluations. In a controlled 16-turbine case, it also outperforms ordinary and multi-fidelity Kriging under the same high-fidelity evaluation budget. The framework is further extended to a 48-turbine wind farm, demonstrating applicability to large-scale CFD-based layout optimization. These results show that Graph-informed surrogate modeling can enable more efficient use of high-fidelity aerodynamic models for reliable large-scale wind farm layout design.
Product reviews significantly influence purchasing decisions on e-commerce platforms. However, the sheer volume of reviews can overwhelm users, obscuring information most relevant to their individual needs. Existing e-commerce summarization systems typically produce generic, static summaries that fail to account for the fact that (i) different users prioritize different product characteristics, and (ii) these preferences may evolve with interactions. To address these challenges, we propose an online learning framework that generates personalized summaries for each user by representing user interests in a latent aspect space and selecting preference-aligned review evidence. Our system then iteratively refines its understanding of user preferences by incorporating feedback on the generated summaries directly over time. We evaluate our framework on Amazon Reviews'23 using controlled feedback simulations and robustness checks with held-out-history, independent-embedding, and external LLM-based evaluators. These experiments show improvements in preference-aligned evidence selection under the evaluated feedback settings. A separate human-based study finds that scenario-conditioned summaries are preferred over generic (uniformly-aggregated) alternatives in 69.9% of paired comparisons.
We present a parallel solver for finite-horizon linear-quadratic regulator (LQR) problems with arbitrary stagewise linear equality constraints, using linear work and logarithmic depth in the horizon for fixed local dimensions. The solver uses a two-part reduction that separates feasibility from optimization. First, an associative suffix scan computes the states from which each remaining horizon is feasible. Second, we use these feasible-continuation sets to eliminate the constraints independently at each stage, producing an unconstrained LQR problem with smaller or equal, possibly stage-varying dimensions. A parallel unconstrained-LQR solver completes the primal solution. Finally, we recover valid multipliers from the original KKT stationarity equations using a logarithmic-depth contraction-expansion procedure, even when the multipliers are not unique. Sequential and parallel implementations are validated against dense KKT solutions and original residuals. We release a C++/CUDA implementation with JAX bindings under the MIT license.
Alexandra Suvorikova, Darina Dvinskikh, Yuriy Dorn, Alexander Gasnikov
Polyak--Ruppert averaging gives SGD and momentum methods the same asymptotic covariance. We characterize how fast finite runs reach it and how far they can be parallelized. For strongly convex quadratics with additive noise, we prove an online--offline bridge: the empirical risk minimizer on the samples of a tail average is one Newton step from the tail-averaged query point of any online procedure. For SGD, heavy ball, and Nesterov's method with constant momentum, this step is a boundary term at the two ends of the tail and splits the gap to the offline estimator into center, covariance, and shape errors. Necessary and sufficient conditions show that, on the least curved direction, the center error sets the pace: SGD needs $m\ggκ^2$ rounds but momentum only $m\ggκ$, where $κ$ is the condition number, so at sample size $N$ the critical batch size grows from $Θ(\sqrt{N}/κ)$ to $Θ(\sqrt{N/κ})$. The covariance error separates the momentum methods: heavy ball has relative variance excess $Θ(κ^{3/2}/m)$ along the stiffest direction, whereas Nesterov has relative error $O(\sqrtκ/m)$ in every direction. The shape error is method-independent by a Berry--Esseen bound uniform in $κ$ and batch size, so confidence intervals are valid when the other two errors vanish. We further examine how the conclusions change under heavy-tailed or state-dependent noise and clipping, and establish a first-order comparison with ERM for nonlinear Nesterov averaging under explicit conditions.