Bilevel optimization is a central tool in computational imaging for high-dimensional hyperparameter tuning; prominent applications include learning data-adaptive regularizers and optimizing forward operators in variational regularization. However, due to the nested nature of the problem, exact gradients with respect to the parameters (so-called hypergradients) are not available: the precision of the hypergradient is linearly related to the computational cost, thus requiring algorithms which can solve the problem without unnecessary precision. In learned-regularizer pipelines, the inner reconstruction solver is typically run for a fixed number of iterations, leaving the hypergradient accuracy uncontrolled and existing convergence theory inapplicable. In this work, we analyze the schedules that practitioners actually run: we prove convergence of inexact stochastic bilevel optimization and establish rates under both decaying accuracy and step size schedules, showing that, with optimal configurations, convergence occurs at an $O(k^{-1/4})$ rate in expectation. This is the optimal rate of nonconvex stochastic gradient descent and matches single-loop stochastic bilevel methods; with warm-started inner solves, total costs also agree up to constants. Experiments on image denoising and inpainting with convex ridge regularizers and input-convex neural networks confirm our analysis: decreasing step sizes improve stability, accuracy scheduling is more critical than step size strategy, and adaptive preconditioning (e.g., Adam) further boosts performance, particularly for input-convex neural networks. In a budget-matched comparison with four state-of-the-art single-loop methods, accuracy-scheduled stochastic gradient descent and its Adam variant attain the highest test peak signal-to-noise ratio.
Efficient preconditioning plays an important role in large-scale first-order optimization. In this paper, we propose an adaptive accelerated optimistic dual averaging ($\text{A}^2$ODA) method for unconstrained convex optimization that adopts low-cost preconditioning based on $p$-norm geometry for $p\in[2,\infty]$. The method is line-search-free and employs a new stepsize selection technique based on an aggregated affine lower model. Its update structure is simple and unified across $p\in[2,\infty]$, with the case $p>2$ requiring no more elaborate updates than the Euclidean case $p=2$. Under a generalized local smoothness condition, we derive a convergence bound that separates the effects of constant and gap-dependent curvature, with the latter potentially yielding faster-decaying terms. In the special case of local smoothness, this bound recovers a nearly optimal convergence rate.
Alternating online mirror descent (AltOMD) is a simple learning rule for zero-sum games, which improves upon simultaneous OMD numerically and theoretically. Previous work obtained $O(T^{1/3})$ and $O(T^{1/5})$ regret bounds for AltOMD, by leveraging modified-energy-function-based analyses. However, settling whether AltOMD achieves near-constant regret has remained an open question. In this paper we resolve the question of whether AltOMD achieves near-constant regret in the affirmative. In particular, we show an $O(log(T))$ regret bound on the sum of the players' external regrets for any two-player zero-sum bilinear game with compact convex strategy sets, under mild regularity assumptions on the regularizers. We verify that our assumptions hold for the negative entropy regularizer, thereby obtaining logarithmic regret for the alternating multiplicative weights updates algorithm. The regret bound also yields an $O(log(T)/T)$ duality gap bound for the averaged strategies.
Matrix factorization provides a simple model for studying the training dynamics of neural networks. Recent studies of gradient flow under small initialization have analyzed how trajectories escape from a neighborhood of the origin and approach their first nonzero stationary point. In this paper, we study continuous-time dynamics that applies a finite number of Newton-Schulz iterations to the normalized matrix gradient (Muon). We show that the number of Newton-Schulz steps can change the rank selected during the early stages of training and allow the resulting trajectory to bypass the first nonzero stationary point approached by gradient flow. Under an additional residual-inertia condition, we prove that bypass occurs for all sufficiently large finite Newton-Schulz depths and provide an explicit sufficient threshold. We also construct an example in which the bypass threshold is four.
Large language model (LLM) inference, the process by which a pre-trained model receives an input prompt and generates output tokens sequentially, has become increasingly important with the widespread deployment of LLM-based applications. As inference workloads continue to scale, GPU memory has emerged as a critical bottleneck: the long input prompts and generation sequences common in agentic and reasoning workloads lead to substantial key-value (KV) cache usage. In this paper, we study the resulting online scheduling problem from a queueing perspective. Requests arrive stochastically with heterogeneous prompt and output lengths, undergo prefill and autoregressive decode, and compete for a finite KV-cache capacity. Unlike classical queueing models, the duration of a processing step depends on both prefill computation and KV-cache I/O, so service capacity is jointly determined by computation, memory access, and memory occupancy. We develop a fluid model that reduces the system to a multiclass single-server queue with effective service rates capturing computation, memory access, and memory capacity. This fluid analysis yields a $c\mu$ priority rule, where the effective service rate $\mu$ is jointly determined by computational speed, memory bandwidth, and memory capacity. We prove that this rule minimizes cumulative holding cost in the fluid model and, for a finite workload, also minimizes completion time and maximizes full-run output-token throughput. We then construct a feasible scheduling policy for the original finite system and prove that its scaled dynamics converge to the optimal fluid policy. Finally, experiments on a real LLM serving engine using an NVIDIA H800 show that practical implementations of the proposed priority rule substantially reduce average latency while maintaining high throughput relative to existing scheduling policies.
The classical model-based optimal design criteria (D-, T-, L-, C-, and A-optimality) are usually motivated via the Cram\'er--Rao bound as optimizing scalar functions of the Fisher information matrix (FIM). This paper offers a complementary reading in which each criterion appears asymptotically as $N \to \infty$ as the leading-order term of an expected dissimilarity between prior and posterior, with the FIM entering through the Bernstein--von Mises limit. The Kullback--Leibler (KL) divergence of the posterior relative to the prior recovers D-optimality (classically known); a reverse KL of the prior relative to the posterior recovers T-optimality (here defined as maximization of the trace of the FIM, following one common convention) and its weighted-trace generalization; and a first-order approximation of the squared 2-Wasserstein distance recovers L-, C- and A-optimality. The five criteria emerge as instances of a common Bayesian-geometric template, indexed by three choices: the dissimilarity, the prior parametrization, and the approximation scheme used to evaluate it (e.g.\ Taylor expansions of the matrix square root). For the weighted criteria, the Gaussian prior covariance encodes the criterion's weight matrix, so any choice of weights corresponds to a specific prior in the family. The T-, L-, C-, and A-optimality derivations appear novel to the best of my knowledge. A sixth criterion, E-optimality, is recovered not as an independent expected-dissimilarity limit but as a min-max robustification of C-optimality against uncertainty in the modeler's prior direction. The framework is extended to distinguish between a designer's prior over unknown parameters at design time and a modeler's prior used for posterior inference, recovering the classical expected information gain objective as a special case.
Data-driven algorithms are ubiquitous across industries, and there is an extensive literature on analyzing their \emph{average} performance. The risk profile of such algorithms, while critical, is much less understood. This paper quantifies the risk of data-driven policies and isolates its sources. We evaluate a policy through the conditional value-at-risk (CVaR) of its loss over the training sample, any policy randomization, and the future outcome, at a fixed unknown distribution. Regret against an oracle that knows the distribution decomposes into expected post-sample regret (PS), the average outcome-only CVaR gap of the selected action, and a sample-risk premium (SR), the additional tail risk omitted by averaging those evaluations. The analogous premium vanishes under expected loss. We characterize when the premium vanishes and how large it is relative to post-sample regret. For losses with a continuous density near the tail cutoff, we identify conditions under which both components decay at the same rate, yet the limiting ratio of the premium to post-sample regret can reach the thousands. When a probability mass at the cutoff moves with the action, the premium can decay more slowly and dominate post-sample regret. We establish conditions under which sample average approximation learns the oracle action at rate $n^{-1/2}$ in both settings, but its combined-risk regret is of order $n^{-1}$ in the first and $n^{-1/2}$ in the second. In finite-sample numerical examples of SAA in both settings, the premium accounts for approximately $74\%$--$97\%$ of regret.
Motivated by the expressive power of neural ordinary differential equations in flow-based generative modeling, we study the approximation of probability measures by pushforwards generated by parameterized families of Lipschitz vector fields. For Wasserstein distances, we establish a quantitative relationship between finite-sample point interpolation and distributional approximation, and derive universality criteria based on $L^p$-approximation of transport maps. In particular, affine-invariant families containing a non-affine vector field are universal from compactly supported absolutely continuous base measures. For total variation, we show that coordinate-wise dynamics are insufficient in dimensions $d\geq 2$ and establish sufficient conditions based on one-dimensional Sobolev approximation and shear-type mixing. We further examine the role of point interpolation, establishing when different regimes relate to one another and when approximate interpolation can be made exact.
We introduce the Axiomatic Learning Structures (ALS) and Axiomatic Probabilistic Structure (APS) in order to study the interactions of Machine Learning, Optimization, with research frameworks- which, we take the view, can be characterized at the core by the corresponding axioms. The ALS explores the interplays between structures that one is interested in Learning Theory such as Loss functions and the axioms of the objects and transformations involved in these functions. The APS studies the interaction between structures of interest in Probability Theory such as Gibbs distributions, equations with noise, etc. and the axioms of the corresponding objects and transformations. We also introduce in Section 6 Transformations-Axiomatic Based Learning (TABLe) on Mathematical data-sets. One feature of this approach is the ranking of transformations from a table for a certain task based on scores which reflect which transformations work better than others, similarly to when human investigators try various methods on a task. We also study structure recovery from noisy data. In particular, to exemplify the ideas, we consider loss functions which contain derivations and derivations-like transformations. As a first application we study recovery in the context of polynomials with integer coefficients using standard derivatives along with other types of derivations which appear in Arithmetic: the p-derivations which satisfy \textit{different axioms} from the usual derivatives. As a first application of the APS we study a class of noisy (partial/mixed) derivations equations. In the last sections, to provide further potential applications of this framework, we give an overview of an initial simplified Opinion Dynamics model and its (simultaneous) study of rules of updates.
We study efficient algorithms for realizing the first-order oracle complexity of optimization of $G$-Lipschitz convex functions with respect to the $\ell_q$-norm over an $\ell_p$-ball of radius $R$, where $1\leq p,q\leq\infty$. For $p<q$, we obtain error $\widetilde{O}_{p,q}(GR/T^{1/p-(1/q-1/2)_{+}})$ after $T$ oracle queries, attaining the lower bounds in a COLT 2015 open problem (Guz15b) with efficient algorithms. In particular, the rate is $\widetilde{O}(GR/T)$ for Euclidean Lipschitzness with an $\ell_1$-ball domain of radius $R$ ($p=1,q=2$). Our solution consists of reducing convex Lipschitz optimization to the chasing nested convex sets problem (BBE+20) in sublevel sets of an evolving maximum of affine lower bounds, that is, the bundle (LNN95): at each query we either find a point with low function value or we produce a deep cut in the current sublevel of the bundle that we chase. The dichotomy between stability of selectors and forced movement by deep cuts bounds the number of iterations of the algorithm nearly optimally. For nested convex subsets of the $\ell_p$ ball $R B_p^d$, we introduce novel stable selectors whose movement is nearly optimal in high dimensions. A Monte Carlo average of the proposed selector achieves near-optimal rates with high probability and can be implemented for our optimization algorithm in polynomial time in the real-arithmetic model.
We study first-order black-box convex optimization over an $\ell_p$-ball for objectives Lipschitz in the $\ell_q$-norm, solving in the affirmative the nonsmooth version of the COLT open question (Guz15b) on whether the geometry of a smaller feasible set ($p < q$) can improve convergence rates in convex optimization, and matching prior lower bounds up to logarithmic factors. Our rates include $\widetilde{O}(1/T)$ for convex Euclidean-Lipschitz optimization over the $\ell_1$-ball, improving on the $O(1/\sqrt{T})$ classical rate under general assumptions. The key technical device is a new online learning game, where the comparator is evaluated using the maximum of affine losses observed so far. We bound the value of this game above and below in terms of a combinatorial online learning quantity: the sequential fat-shattering dimension, which we characterize for the $\ell_p / \ell_q$ case. Our results generally apply when the feasible set $\mathcal{X}$ and the set of possible subgradients $\mathcal{H}$ admit a type of minimax theorem, advancing on a fundamental question by Sridharan (Sri12, Section 10.1.2, Q3). As a geometric consequence of our analysis, of independent interest, we obtain estimates for the expected distance of a convex hull of samples to their mean in several Banach geometries, a version of the celebrated Wendel's theorem (Wen62), but quantitative and for bounded general distributions as opposed to centrally symmetric ones.
In this paper, we derive an explicit solution of the stationary prediction with expert advice PDE for five experts. The formula is given in three regions. In the first two regions, it is the four-expert solution plus a single integral with an elementary positive density. In the third region, it is a finite sum of hyperbolic products whose coefficients are determined by one scalar quadrature. Our formula establishes that the rank ordered adversary $(1,0,1,0,0)$ achieves the Hamiltonian maximum in the PDE everywhere in the ordered sector, while the COMB strategy $(1,0,1,0,1)$ is optimal only on a lower dimensional subset of the sector (where $x_1=x_2$ and $x_3=x_4$). The verification of the Hamiltonian inequalities is a tedious task, part of which is completed with a computer assisted proof. The verification reduces to 21 scalar inequalities, which we prove using 147 exact rational Bernstein polynomial certificates. The exact certificates and their independent arithmetic checks are included in a supplement to this paper, and a Lean 4 formalization machine-checks the verification and both main theorems, apart from the viscosity characterization.
Stochastic nonconvex optimization is central to training deep networks and LLMs in modern machine learning. We give a black-box reduction from stochastic nonconvex optimization to ordinary static regret minimization in online convex optimization (OCO), addressing the open problem posed by Chen and Hazan (2024). Our reduction maintains a predictable gradient tracker, while a black-box online learner $\mathcal A$ selects a preconditioner that transforms this tracker into the update direction. Given a $\beta$-smooth function with initial gap at most $M$ and an unbiased gradient oracle with variance bounded by $\sigma^2$, we bound the expected average squared gradient norm by $O(\sigma\sqrt{M\beta/T}+\sqrt{M\beta}\mathrm{Reg}_T(\mathcal A)/T+\frac{M\beta}{T})$, where $\mathrm{Reg}_T(\mathcal A)$ is the static regret of $\mathcal A$ against the identity preconditioner. Thus, any OCO oracle with $O(\sqrt T)$ regret recovers the classical $O(T^{-1/2})$ convergence rate. We also prove that the additive variance term cannot be removed: every conversion making a constant number of stochastic-oracle queries per round, whose guarantee is stated through static regret against a comparator that the online learner can play, must contain an additive term of order $\sigma\sqrt{M\beta/T}$ for sufficiently large $T$. Moreover, for every online learner, our conversion bound is tight up to universal constants. We also extend the framework to nonsmooth nonconvex optimization, still relying only on ordinary static regret, and attain the optimal convergence rate for Goldstein-type stationarity. Finally, we conduct numerical experiments on nonconvex objectives to illustrate how the reduction exploits online-selected preconditioners while using the same stochastic-oracle budget as stochastic gradient descent.
The tail trimming strategy is widely adopted for robustness against outlier contamination in statistical learning and operations research. However, it generally lacks the thorough understanding that to what extent the tail-trimming technique is effective for nonlinear regression. In this paper, we address the robust nonlinear regression based on the so-called interval conditional value-at-risk (In-CVaR), which is introduced to guard against outliers by trimming extreme losses. We first analyze the trimming effect of the In-CVaR based estimator under distributional perturbation. Our first finding is that the In-CVaR based estimator is qualitatively robust under distributional perturbation in terms of the Prokhorov metric if and only if the largest portion of losses is trimmed. Furthermore, under sample contamination, we derive a probably approximately correct (PAC) bound for the In-CVaR based estimator, showing that the preliminary covariate filtering that is customary in the literature is not needed for In-CVaR based regression. The nonconvexity of the In-CVaR objective, together with unboundedness of covariates, responses and parameter space create the technical challenges for our developed theory, and thus necessitate the new treatment of the qualitative robustness and PAC analysis. Overall, this study rigorously analyzes robustness properties of In-CVaR based nonlinear regression models, which illustrates to what extent the right tail-trimming of the loss by In-CVaR risk measure is essential and effective for robust nonlinear regression. These theoretical findings are further supported by numerical experiments.
High-quality matches between patients and primary care providers improve downstream health outcomes, yet provider turnover breaks these matches and forces patients to be reassigned. Health systems facing this problem must decide not only which providers to recommend, but how much choice to offer. Direct assignment avoids competition for scarce provider capacity but relies on noisy estimates of patient preferences; broader choice allows patients to correct these estimates using private information but can create costly coordination failures. We study this tradeoff through the design of fixed, personalized menus from which patients choose providers. We show that constructing the optimal menu is NP-hard and establish an information-theoretic lower bound. We find that offering each patient a single provider matches this bound up to a constant factor, while unrestricted choice can perform arbitrarily poorly. However, we show that additional choice improves on single-provider assignment when the system cannot reliably distinguish among a patient's alternatives. We formalize this tradeoff through two forms of loss: menu loss, which arises when menus offer too few options, and coordination loss, which arises when scarce capacity is exposed to too much choice. Motivated by this decomposition, we develop the Scenario-Averaged Marginals (SAM) algorithm, which tailors menus based on provider shadow prices so choice is offered only when utility exceeds the capacity cost. We identify sufficient conditions which SAM's regret matches the lower bound, namely when patients have valuable alternatives and offered menus don't create competition for scarce capacity. In a semi-synthetic environment calibrated to a provider network using Medicare data, SAM improves total utility by up to 19% over single-provider assignment as preference uncertainty grows, while improving utility for the least-served ZIP codes.
Convexity has three equivalent descriptions: support by affine functions, monotonicity of the subdifferential, and increasing secant slopes. We weaken all three by one distortion with parameter $\gamma\in(0,1]$, which multiplies positive changes by $\gamma$ and divides negative ones by $\gamma$. The resulting classes are the convex functions at $\gamma=1$ and, for $\gamma<1$, are distinct and nested, with the parameter dropping by at most a square along the chain. For proper lower semicontinuous functions in these classes, the effective domain is convex, local minimizers are global, and the function is locally Lipschitz on the relative interior of its domain. The classes separate offline from online optimization over compact convex sets with efficient projection, with access to function values and bounded support vectors. A smooth member of any class can be minimized to additive accuracy $\varepsilon$ with $\mathrm{poly}(1/\varepsilon)$ first-order queries and projections. Online, with nonpositive losses and $\alpha$-regret measured against $\alpha$ times the best fixed point, a polynomial-time algorithm achieves $\gamma$-regret $O(\sqrt T)$ on the largest class. For every fixed, effectively specified $\gamma\in(0,1)$ and every $\alpha\in(\gamma,1]$, no randomized polynomial-time algorithm has sublinear expected $\alpha$-regret unless $\mathbf{NP}=\mathbf{RP}$, already for smooth concave polynomial losses in the smallest class, each minimized by one linear optimization. Thus, unless $\mathbf{NP}=\mathbf{RP}$, $\gamma$ is the exact online coefficient of each class, while the offline coefficient is one.
We classify real and complex state-space models (SSMs) with linear transitions, distinguishing fixed-basis diagonal equivalence from input-output equivalence and tracking coefficient-conditioning windows under coordinate changes. We characterize diagonal equivalence through commuting transition families whose members are diagonalizable over the chosen field. With time-shared token-local matrix maps, we identify this class's state-mediated kernels with unnormalized scalar-mode linear attention with multiplicative masks. Mamba 3's multiple-input, multiple-output (MIMO) construction lifts one value per local interface; SSD-MIMO routes independent values for local retrieval rank above one. HybridMamba combines Mamba 1's per-coordinate decay, Mamba 2-style heads and Mamba 3's complex rotations and two-endpoint writes, mixing real and rotating state blocks with different decay laws. They recover Mamba 1-3 single-input, single-output (SISO) configurations as exact special cases. On five large H100 bfloat16 workloads, our fused GPU kernels improve geometric-mean training time over official Mamba 1/3 in matched SISO recoveries, with Mamba 2 near parity at SSM-core scope. Memory tradeoffs depend on the baseline and metric: saved-for-backward versus peak allocation. In small-model multi-token recall experiments, both designs outperform Mamba 1-3 on the hardest tier; higher routing ranks perform best in gated or phase-diverse configurations. In single-seed language models with about 500M parameters, both improve on Mamba 1/2 in test loss at 2B tokens, but the routing gains do not translate into lower language-model loss. Our best configurations approach Mamba 3 SISO through 10B tokens while remaining behind it and Mamba 3 MIMO in language-model loss.
Multiproduct pipelines transport refined products to delivery stations in sequential batches, and their detailed scheduling must meet market demand while respecting nonlinear hydraulic constraints. Retaining the full hydraulic nonlinearity yields a mixed-integer nonlinear program that is computationally intractable, while the linearization commonly adopted for solver-based optimization degrades hydraulic accuracy without resolving the heavy computational burden. This paper develops an action-decoupled proximal policy optimization (PPO) framework that efficiently produces detailed schedules while preserving hydraulic fidelity. The scheduling process is cast as a semi-Markov decision process (SMDP) with flexible action durations. To handle the discrete-continuous hybrid action space, a cascaded dual-actor architecture decouples the discrete operating duration from the continuous flow decisions, eliminating the gradient interference that destabilizes conventional coupled-action policies and thereby yielding high-quality schedules far more efficiently. Validated on a real-world multiproduct pipeline in China and benchmarked against a single-actor baseline and a linearized MILP solved by Gurobi, the proposed framework achieves delivery accuracy close to the exact optimum while substantially improving flow rate stability and substantially reducing solution time.
Dynamic regret measures an online learner against a moving comparator, and standard analyses assume the learner may move freely at every round. We study what happens when it may not: the learner is given a hard budget $S$ on its total movement $\sum_t \|x_t - x_{t-1}\|$. Our main result is that a movement budget creates a feasibility barrier. On a domain of diameter $D$ with $G$-Lipschitz convex losses, for every comparator path length $P_T \in [D, DT]$ we exhibit a deterministic adversary that forces $R_d = \Omega(GDT)$ on every budget-feasible algorithm whenever $S \le c P_T$ for a universal constant $c>0$. This barrier is caused by the budget rather than by a volatile environment: for fixed $D$, it applies when $P_T = o(T)$, where the unconstrained minimax rate $\Theta(G\sqrt{D(D+P_T)T})$ is sublinear, so the same instances are learnable with an unbounded budget and unlearnable without one. Conversely, projected online gradient descent with a budget-aware step size is budget-feasible and attains $o(GDT)$ once $S = \omega(D + P_T)$. Thus the critical movement scale is $D+P_T$: linear regret is unavoidable below a sufficiently small constant multiple of $P_T$, while sublinear regret is achievable when $S/(D+P_T)\to\infty$ in the learnable regime. The construction rests on a distinction we make explicit: what makes a comparator expensive to track under a movement budget is the distance it forces the learner to travel, not its path length. Above the threshold we give a budget-dependent upper bound and identify the tight minimax rate in this region as an open problem rather than claiming a product law. Experiments support the predicted transition location and the $1/S$ decay.
This paper introduces a reinforcement learning-based approximate dynamic programming approach to optimize the deployment of Walker-type satellite constellations under uncertainties, including launch failures and satellite malfunctions. We formulate the problem as a Markov Decision Process that incorporates low-injection orbit launches, continuous low-thrust orbit raising, and variable satellite loads. We adopt the offline approximate value iteration solution method, using linear function approximation with basis functions and action pruning strategies to substantially reduce the computational and memory requirements for finding optimal decision policies. Preliminary results demonstrate convergence to policies with low expected costs and deployment times, both of which are lower than those of a rolling-horizon baseline, showcasing a novel application of approximate dynamic programming in space logistics.