ABSTRACT. We consider a privacy problem where a user shares their location information in order to use some service while limiting the amount to information leaked concerning visits to certain sensitive locations. At some time steps, the user will interact with the service and must share their true location to receive any benefit. At other times, the user would like to falsify their location to preserve as much privacy as possible, but their sequence of locations must be statistically plausible to avoid detection. We investigate trade-offs between utility in privacy via explicit algorithms for location trace generation and converse bounds.

ABSTRACT. In several modern applications involving online platforms and digital marketplaces, the data-driven allocation of supply to demand involves new difficulties such as accommodating combinatorial feasibility environments or reusable resources (or both). Modeling these settings as instances of a Bayesian online allocation problem raises a natural question: Is there a universal technical framework for designing algorithmic solutions across all these settings? Online contention resolution schemes (OCRSs) aim to provide such solutions by converting ex-ante relaxations into approximately optimal online feasible policies. Despite their importance---and relative universality---existing constructions of (near-)optimal OCRSs are often technically challenging. Many constructions rely on indirect reductions to prophet inequalities and LP duality, resulting in algorithms that are ad hoc and difficult to interpret or implement. Moreover, it is unclear whether such vanilla OCRSs can accommodate new paradigms such as resource reusability.

In this paper, we introduce stationary online contention resolution schemes (S-OCRSs), a permutation-invariant class of OCRSs in which the distribution of the selected feasible set is independent of the arrival order. We establish a distributional characterization of S-OCRSs and provide a universal online implementation, thereby reducing the design of online policies to the construction of suitable distributions over feasible sets. This reduction applies to arbitrary, potentially complex feasibility environments. We further show that S-OCRSs are precisely the tool needed to address reusable resource allocation problems. We then develop a general `maximum-entropy-based' approach for constructing and analyzing S-OCRSs across a variety of settings. Together, these results provide a universal and principled technical framework for designing simple and potentially improved OCRSs, yielding approximately optimal online allocation algorithms for a wide range of modern Bayesian online allocation problems.

We demonstrate the power of our new framework across several canonical feasibility environments. For bipartite matchings, we obtain an improved (3-\sqrt{5})/2-selectable OCRS, attaining the independence benchmark conjectured to be optimal for OCRSs. For knapsack constraints, we obtain a 1/4-selectable S-OCRS. We also obtain a 1-\sqrt{2/(\pi k)}+O(1/k)-selectable S-OCRS for k-uniform matroids; a simple, explicit 1/2-selectable S-OCRS for weakly Rayleigh matroids, including all \mathbb{C}-representable matroids such as graphic and laminar matroids; a 1-\sqrt{2\log L/k}-O(1/\sqrt{k})-selectable S-OCRS for rank-L hypergraph matchings with capacity k (also known as the network revenue management setting); and a \Theta(1/\sqrt{N})-selectable S-OCRS for general downward-closed constraints on sets of size N. These latter guarantees match the best-known bounds for OCRSs via simple and principled algorithms, in contrast to existing algorithms that are either ad-hoc, complex, or non-explicit. All of the aforementioned S-OCRS guarantees for different feasibility environments extend to approximation factors for the Bayesian online allocation problems of reusable products in the same feasibility environments.

ABSTRACT. Natural policy gradient (NPG) is a central tool in modern reinforcement learning, with variants such as TRPO and PPO serving as major workhorses in today’s applications. Despite extensive study, existing convergence analyses either do not provide sharp guarantees, such as distribution-free global geometric convergence, or obtain such guarantees only through additional regularization or adaptive stepsizes. A key difficulty is that, unlike classical dynamic programming methods such as value iteration and policy iteration, gradient-based methods are often viewed as departing from Bellman-operator-based algorithm design and analysis, making it difficult to directly exploit defining features of the Bellman operator, such as contraction and monotonicity. In this talk, we present a new perspective showing that NPG belongs to a unified framework, which we call doubly smoothed policy iteration, where the policy is updated by taking a smoothed greedy step with respect to a weighted average of historical Q-functions. This framework includes NPG, policy iteration, dual-averaged policy iteration, and general policy dual averaging as special cases. By leveraging the contraction and monotonicity properties of the Bellman operator, this viewpoint yields strong convergence guarantees for a broad class of algorithms through a simpler analysis. The same algorithmic and analytical framework also extends to stochastic shortest path and average-reward problems.

ABSTRACT. This talk will present some recent work on stochastic iterative algorithms or dynamics that depend on the entire past, motivated by some applications in machine learning and elephant random walks.

ABSTRACT. This work focuses on learning from a large number of devices with each device holding only a single sample of privacy-sensitive data. Several real-world applications exist to this one sample per client setup up including learning from fitness trackers, data/app usage aggregators, body-worn sensing devices, and daily event monitors to name a few. When a client has only one sample, the standard federated learning paradigm breaks down as a local update based on that single point is far from being useful, especially in the earlier rounds for estimation of the model coefficients. This utility is further weakened by the privacy-inducing noise applied at every round. This work caters to this problem to enable such clients to collaboratively contribute to effectively learn a global model without leaking the privacy of their data. The proposed approach injects a single, carefully calibrated noisy perturbation to transform the sample at each client, followed by a post-processed representation which is shared with the server. These representations aggregated at the server are processed to obtain an unbiased gradient update that in expectation matches the non-private centralized gradient while preserving data privacy. This approach is different than traditional private federated learning, where the communication payloads involve model coefficients as opposed to privately transformed data samples. This method enables devices with extremely limited data to collaborate and learn accurate, privacy-preserving models without requiring large local datasets or sacrificing individual privacy.

ABSTRACT. Spectral optimizers such as Muon, and the more general Freon family D = (GGT) −cG, have shown strong empirical performance in large-scale learning. Recent evidence suggests their success may not rely on an exact target geometry: randomizing the singular values (the Kaon variant) performs comparably to carefully whitened updates. We turn this observation into a communication question for distributed learning: if exact spectral geometry is unnecessary for optimization, must we communicate the exact spectrum? We propose Sketch-Kaon and SketchFreon, communication-efficient federated optimizers that transmit only a two-round randomized low-rank subspace sketch of the aggregated gradient while replacing the exact singular values with randomized or coarsely quantized ones. To stabilize training under client heterogeneity and a noisy uplink, we introduce a client-split alignment estimator γˆ that adapts the global step size using agreement between two independently aggregated client groups— computed by reusing already-received uploads, hence at no additional communication. We give convergence, sketchapproximation, and safe-step analyses, and evaluate on synthetic distributed regression, a controlled spectrum ablation, federated edge modulation classification, and federated wireless channel prediction. Across these settings the exact singular values can be randomized with negligible effect at identical communication, and the sketched methods achieve markedly better bits-to-accuracy than full-gradient spectral descent while being more robust to a noisy uplink.

ABSTRACT. We investigate decentralized gradient descent (DGD) with emphasis on efficient communication and computation resource utilization under budget constraints. Specifically, we propose and analyze a multi-consensus decentralized gradient descent (mcDGD) scheme, where the number of consensus rounds and the stepsize are allowed to vary across iterations. Building on a unified analytical framework for DGD, we derive finite-time convergence bounds that explicitly characterize the interaction between consensus quality and optimization dynamics. Our analysis requires only convexity of the local objective functions while assuming smoothness and strong convexity of the global objective. The resulting bounds enable a principled communication allocation strategy under resource constraints, including structural properties of optimal consensus scheduling. Numerical experiments corroborate the theoretical findings and demonstrate improved communication-computation efficiency compared to fixed-consensus decentralized optimization methods.

ABSTRACT. This article develops a large-signal stability analysis for a sampled-data optimization-based secondary controller for distributed energy resources (DERs) in power systems. The induced closed loop combines nonlinear inverter power-flow dynamics, filtered active and reactive power measurements, constrained optimization updates, and interpolation-based actuation between sampling instants. We study this optimization-in-the-loop nonlinear sampled-data system beyond local linearization. The analysis provides computable bounds on the voltage, filtered reactive power, secondary control input, and input derivative. It further characterizes steady-state operating points and establishes how the optimizer conditions connect voltage regulation with equal per-unitized reactive power sharing. Finally, input-to-state stability of the phase-frequency dynamics is established with respect to active-power mismatch. These results provide a rigorous mathematical foundation for sampled-data optimization-based secondary control of DER networks.

ABSTRACT. Private Information Retrieval (PIR) and Private Set Intersection (PSI) are fundamental privacy-enhancing primitives with growing relevance in cloud-based applications. In the multi-server setting, information-theoretically secure PSI can be obtained from information-theoretically secure symmetric PIR. We extend this connection to the single-server setting by relaxing the security requirement from information-theoretic to computational guarantees. Our main result shows that any computationally secure symmetric PIR protocol can be transformed into a computationally secure PSI scheme, enabling a generic way to define new PSI schemes under widely used cryptographic assumptions.

ABSTRACT. With the recent boom in data science, the demand for privacy protection in data analytics has increased. In particular, private selection tasks for extracting significant information from data are essential, and the development and application of trustworthy methods to address them while satisfying differential privacy have been actively pursued. However, within existing mechanisms, there remains considerable room for further refinement of the concepts of sensitivity that determine the noise scale. Theoretically, enhancing these concepts allows stricter and smaller perturbations, thereby improving the reliability of the analysis results. Therefore, this study proposes novel differentially private selection mechanisms using enhanced sensitivity concepts. First, the concepts of global and smooth sensitivity are enhanced to simultaneously handle changes in the values of both the target and the other candidates. In addition, direction-aware concepts are proposed that consider the direction of changes as well. Furthermore, enhanced concepts are proposed that can be employed when using a one-sided noise distribution along with rigorous theoretical guarantees. Experiments using genomic statistical analysis as an example demonstrated that the proposed mechanism using enhanced and direction-aware smooth sensitivity with a one-sided distribution can provide higher accuracy than existing methods. Overall, this study can serve as a critical foundation for the development of optimal and reliable methods for differentially private selection. The Python implementation of the experiments and the supplemental results are available at https://github.com/ay0408/EnhancedSPS.

ABSTRACT. We study private Boolean function evaluation in 2.5D/3D heterogeneous systems, where a small trusted chiplet coordinates with isolated but untrusted chiplets. The function is known, but the private input and all input-dependent intermediate values must remain hidden. Since the input is private, a natural approach is to use k-server Private Information Retrieval (PIR) over the function truth table: the trusted chiplet acts as the PIR client, and the isolated untrusted chiplets act as replicated non-colluding PIR servers. However, direct PIR is infeasible because an n-input Boolean function has a truth table with 2^n entries. We propose Private Information Retrieval for Directed Acyclic Graphs (PIR-DAG), which maps the Boolean function onto a directed acyclic graph (DAG) of q-input lookup tables (LUTs), where q is typically around 4, and evaluates each LUT truth table through a PIR-protected evaluation. This changes the PIR database used in each evaluation from the full 2^n-entry truth table to a 2^q -entry LUT truth table, giving an exponential reduction in per-evaluation database size, and a large reduction in total trusted-to-untrusted communication, relative to direct full truth table PIR. PIR-DAG therefore replaces one infeasible full truth table PIR execution with a sequence of small LUT-level PIR executions. We prove that this sequential evaluation preserves information-theoretic privacy against any individual honest-but-curious untrusted chiplet, even when later LUT indices depend on previously reconstructed intermediate values. We evaluate PIR-DAG against conventional full truth table PIR under finite trusted-to-untrusted upload bandwidth on ISCAS ’85 benchmarks.

ABSTRACT. Cancer treatment is at the core a sequential decision-making problem with partial observability, latent patient heterogeneity, and explicit constraints on the budget for medical measurements. Unlike standard Reinforcement Learning (RL) approaches that control state trajectories, cancer treatments permanently modify patients' transition dynamics, changing how states evolve over time. We model cancer treatment as a belief-space planning problem using active inference, deriving an expected free-energy objective that unifies goal-directed control and information acquisition under measurement budgets without. We implement this framework using real clinical cancer data from the AACR Project GENIE Biopharma Collaborative dataset. Results on clinical data demonstrate a simultaneous patient categorization and high treatment efficacy, under real measurement and treatment constraints.

ABSTRACT. We study the problem of system identification in heterogeneous settings, where different systems may follow distinct underlying dynamics. Existing clustered system identification approaches often rely on iterative training-based cluster assignment, which can be sensitive to learning uncertainty and model initialization. In contrast, we propose a one-shot, training-free clustering method that identifies similar systems using the structure of their locally observed data. Specifically, each system estimates a local state covariance matrix, and cluster identities are inferred by measuring the alignment between the leading covariance eigenspaces of different systems. We provide a mathematical interpretation of the proposed similarity score and develop a finite-sample analysis that characterizes how covariance estimation error induces eigenspace perturbations in terms of the underlying system dynamics. We then derive a probability bound for pairwise false merges and a global clustering success guarantee. Numerical experiments demonstrate that the proposed eigenspace-based clustering method effectively identifies systems with shared dynamics, leading to lower personalized model-estimation error compared with training-based clustering and non-clustered baselines.

ABSTRACT. Despite its popularity, mean Age of Information (AoI) or peak AoI is often used in wireless network design as indicators of closed-loop performance without an explicit physical interpretation. This paper examines whether minimizing mean AoI is in fact optimal for the linear quadratic regulator (LQR) tracking objective in networked control systems. For scalar linear time-invariant systems with delayed intermittent updates, we show that, under state-independent scheduling policies, the infinite-horizon LQR problem reduces to an optimization over the distribution of AoI. The resulting objective depends on higher-order statistics of the freshness of samples, not only on its mean. Consequently, policies with the same mean AoI can induce substantially different tracking costs. More strikingly, we show that there exist schedulers for which reducing the mean AoI increases the tracking cost. Finally, we demonstrate the theory in a data-driven car-following simulation using empirical disturbances extracted from NGSIM US-101 vehicle trajectories. The empirical results match the predicted performance trends and show that lower mean AoI can lead to both higher LQR cost and higher crash probability, confirming that mean AoI alone is insufficient for control-oriented network design.

ABSTRACT. A robot is fundamentally a device that takes converts information about the world into useful physical work. Inside the robot are a series of information processing computations going by names such as observation, representation, prediction, and planning that all slowly prune away information. We build on this information-centric view to inform the design of minimalist resource-efficient robots, asking questions like: What does a robot learner need to sense & attend to at each stage of computation to learn and perform a task? How does this vary as a function of task characteristics, agent characteristics, and learning phase? I will present a series of empirical results that offer clues towards some of the answers, while staying grounded in the modern practice of robot learning.

ABSTRACT. Robots are starting to integrate into manufacturing, healthcare, and our daily lives. However, their development remains largely an iterative and intuition-driven process. Robot structures are typically designed, fabricated, and tested manually, followed by the development of control policies to achieve desired behaviors. This separation between design and control often leads to suboptimal performance and repeated manufacturing iterations. I aim to overcome these challenges with a unified, data-driven co-design framework for robots. Such a framework integrates simulation, optimization, and machine learning to automatically generate and evaluate robot designs, along with their corresponding control policies. With an end-to-end approach, we can not only optimize for single tasks, but also uncover physically intelligent designs that simplify the control problem, enabling robust performance across varying use cases. I will show how this computational approach yields robots that are not only customizable but fundamentally more generalizable than human-designed baselines, paving the way to unlocking capabilities previously thought to be infeasible.

ABSTRACT. Simulation offers a pathway for collecting large amounts of data cheaply, unlocking methods like reinforcement learning for data generation and controller synthesis. However simulation in itself is always an approximate model of the world, with the so-called sim-to-real gap leading to policy errors and imprecisions. In this talk, I take the position that just because the simulator is wrong doesn't mean it is not useful. We present a framework for sim-to-real finetuning that goes beyond transferring policies alone, showing that transferring models, value functions and exploration strategies can enable fast, stable real-world adaptation on deployment with reinforcement learning in the real world.