Tianhao Wu, Weizhe Yuan, Olga Golovneva, Jing Xu, Yuandong Tian, Jiantao Jiao, Jason E Weston, Sainbayar Sukhbaatar. Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing. 2025.

We introduce a new benchmark for reward models that quantifies their ability to produce strong language models through RLHF (Reinforcement Learning from Human Feedback). The gold-standard approach is to run a full RLHF training pipeline and directly probe downstream LLM performance. However, this process is prohibitively expensive. To address this, we build a predictive model of downstream LLM performance by evaluating the reward model on proxy tasks. These proxy tasks consist of a large-scale human preference and a verifiable correctness preference dataset, in which we measure 12 metrics across 12 domains. To investigate which reward model metrics are most correlated to gold-standard RLHF outcomes, we launch an end-to-end RLHF experiment on a large-scale crowd-sourced human preference platform to view real reward model downstream performance as ground truth. Ultimately, we compile our data and findings into Preference Proxy Evaluations (PPE), the first reward model benchmark explicitly linked to post-RLHF real-world human preference performance, which we opensource for public use and further development at https://github.com/lmarena/PPE.

With hundreds of thousands of language models available on Huggingface today, efficiently evaluating and utilizing these models across various downstream tasks has become increasingly critical. Many existing methods repeatedly learn task-specific representations of Large Language Models (LLMs), which leads to inefficiencies in both time and computational resources. To address this, we propose EmbedLLM, a framework designed to learn compact vector representations of LLMs that facilitate downstream applications involving many models, such as model routing. We introduce an encoder-decoder approach for learning such embedding, along with a systematic framework to evaluate their effectiveness. Empirical results show that EmbedLLM outperforms prior methods in model routing. Additionally, we demonstrate that our method can forecast a model's performance on multiple benchmarks, without incurring additional inference cost. Extensive probing experiments validate that the learned embeddings capture key model characteristics, e.g. whether the model is specialized for coding tasks, even without being explicitly trained on them. We open source our dataset, code and embedder to facilitate further research and application.

LLMs are typically trained to answer user questions or follow instructions similarly to how human experts respond. However, in the standard alignment framework they lack the basic ability of explicit thinking before answering. Thinking is important for complex questions that require reasoning and planning – but can be applied to any task. We propose a training method for equipping existing LLMs with such thinking abilities for general instruction following without use of additional human data. We achieve this by an iterative search and optimization procedure that explores the space of possible thought generations, allowing the model to learn how to think without direct supervision. For each instruction, the thought candidates are scored using a judge model to evaluate their responses only, and then optimized via preference optimization. We show that this procedure leads to superior performance on AlpacaEval and Arena-Hard, and shows gains from thinking on non-reasoning categories such as marketing, health and general knowledge, in addition to more traditional reasoning & problem-solving tasks.

Large Language Models (LLMs) excel at reasoning and planning when trained on chain-of-thought (CoT) data, where the step-by-step thought process is explicitly outlined by text tokens. However, this results in lengthy inputs where many words support textual coherence rather than core reasoning information, and processing these inputs consumes substantial computation resources. In this work, we propose a hybrid representation of the reasoning process, where we partially abstract away the initial reasoning steps using latent discrete tokens generated by VQ-VAE, significantly reducing the length of reasoning traces. We explore the use of latent trace abstractions in two scenarios: 1) training the model from scratch for the Keys-Finding Maze problem, 2) fine-tuning LLMs on this hybrid data with an extended vocabulary including unseen latent tokens, for both logical and mathematical reasoning problems. To facilitate effective learning, we introduce a simple training procedure that randomly mixes latent and text tokens, which enables fast adaptation to new latent tokens. Our approach consistently outperforms the baselines methods in various benchmarks, such as Math (+4.2%, Llama-3.2-1B), GSM8K (+4.1%, Llama-3.2-3B), and Fresh-Gaokao-Math-2023 (+13.3%, Llama-3.1-8B) with an average reduction of 17% in reasoning trace’s length.

Large language models (LLMs) can acquire new knowledge through fine-tuning, but this process exhibits a puzzling duality: models can generalize remarkably from new facts, yet are also prone to hallucinating incorrect information. However, the reasons for this phenomenon remain poorly understood. In this work, we argue that both behaviors stem from a single mechanism known as out-of-context reasoning (OCR): the ability to deduce implications by associating concepts, even those without a causal link. Our experiments across five prominent LLMs confirm that OCR indeed drives both generalization and hallucination, depending on whether the associated concepts are causally related. To build a rigorous theoretical understanding of this phenomenon, we then formalize OCR as a synthetic factual recall task. We empirically show that a one-layer single-head attention-only transformer with factorized output and value matrices can learn to solve this task, while a model with combined weights cannot, highlighting the crucial role of matrix factorization. Our theoretical analysis shows that the OCR capability can be attributed to the implicit bias of gradient descent, which favors solutions that minimize the nuclear norm of the combined output-value matrix. This structure explains why the model learns to associate facts and implications with high sample efficiency, regardless of whether the correlation is causal or merely spurious. Ultimately, our work provides a theoretical foundation for understanding the OCR phenomenon, offering a new lens for analyzing and mitigating undesirable behaviors from knowledge injection.

Imaging systems have traditionally been designed to mimic the human eye and produce visually interpretable measurements. Modern imaging systems, however, process raw measurements computationally before or instead of human viewing. As a result, the information content of raw measurements matters more than their visual interpretability. Despite the importance of measurement information content, current approaches for evaluating imaging system performance do not quantify it: they instead either use alternative metrics that assess specific aspects of measurement quality or assess measurements indirectly with performance on secondary tasks. We developed the theoretical foundations and a practical method to directly quantify mutual information between noisy measurements and unknown objects. By fitting probabilistic models to measurements and their noise characteristics, our method estimates information by upper bounding its true value. By applying gradient-based optimization to these estimates, we also developed a technique for designing imaging systems called Information-Driven Encoder Analysis Learning (IDEAL). Our information estimates accurately captured system performance differences across four imaging domains (color photography, radio astronomy, lensless imaging, and microscopy). Systems designed with IDEAL matched the performance of those designed with end-to-end optimization, the prevailing approach that jointly optimizes hardware and image processing algorithms. These results establish mutual information as a universal performance metric for imaging systems that enables both computationally efficient design optimization and evaluation in real-world conditions. A video summary of this work can be found at: https://waller-lab.github.io/EncodingInformationWebsite/

Large Language Models (LLMs) have demonstrated remarkable performance in many applications, including challenging reasoning problems via chain-of-thought (CoT) techniques that generate ``thinking tokens'' before answering the questions. While existing theoretical works demonstrate that CoT with discrete tokens boosts the capability of LLMs, recent work on continuous CoT lacks a theoretical understanding of why it outperforms discrete counterparts in various reasoning tasks, such as directed graph reachability, a fundamental graph reasoning problem that includes many practical domain applications as special cases. In this paper, we prove that a two-layer transformer with $D$ steps of continuous CoT can solve the directed graph reachability problem, where $D$ is the diameter of the graph, while the best known result of constant-depth transformers with discrete CoT requires $O(n^2)$ decoding steps where $n$ is the number of vertices ($D<n$). In our construction, each continuous thought vector is a superposition state that encodes multiple search frontiers simultaneously (i.e., parallel breadth-first search (BFS)), while discrete CoT must choose a single path sampled from the superposition state, which leads to a sequential search that requires many more steps and may be trapped in local solutions. We also performed extensive experiments to verify that our theoretical construction aligns well with the empirical solution obtained via training dynamics. Notably, encoding of multiple search frontiers as a superposition state automatically emerges in training continuous CoT, without explicit supervision to guide the model to explore multiple paths simultaneously.

Reinforcement Learning from Human Feedback (RLHF) is a pivotal technique that aligns language models closely with human-centric values. The initial phase of RLHF involves learning human values using a reward model from ranking data. It is observed that the performance of the reward model degrades after one epoch of training, and optimizing too much against the learned reward model eventually hinders the true objective. This paper analyzes potential reasons behind the issues, and designs improved reward learning algorithm termed 'Iterative Data Smoothing' (IDS). The core idea is that during each training epoch, we not only update the model with the data, but also update the date using the model, replacing hard labels with soft labels. Our empirical findings highlight the superior performance of this approach over the traditional methods.

Safe Reinforcement Learning (RL) aims to find a policy that achieves high rewards while satisfying cost constraints. When learning from scratch, safe RL agents tend to be overly conservative, which impedes exploration and restrains the overall performance. In many realistic tasks, e.g. autonomous driving, large-scale expert demonstration data are available. We argue that extracting expert policy from offline data to guide online exploration is a promising solution to mitigate the conserveness issue. Large-capacity models, e.g. decision transformers (DT), have been proven to be competent in offline policy learning. However, data collected in realworld scenarios rarely contain dangerous cases (e.g., collisions), which makes it prohibitive for the policies to learn safety concepts. Besides, these bulk policy networks cannot meet the computation speed requirements at inference time on real-world tasks such as autonomous driving. To this end, we propose Guided Online Distillation (GOLD), an offline-to-online safe RL framework. GOLD distills an offline DT policy into a lightweight policy network through guided online safe RL training, which outperforms both the offline DT policy and online safe RL algorithms. Experiments in both benchmark safe RL tasks and real-world driving tasks based on the Waymo Open Motion Dataset (WOMD) [1] demonstrate that GOLD can successfully distill lightweight policies and solve decision-making problems in challenging safety-critical scenarios.

Current LLMs are generally aligned to follow safety requirements and tend to refuse toxic prompts. However, LLMs can fail to refuse toxic prompts or be overcautious and refuse benign examples. In addition, state-of-the-art toxicity detectors have low TPRs at low FPR, incurring high costs in real-world applications where toxic examples are rare. In this paper, we introduce Moderation Using LLM Introspection (MULI), which detects toxic prompts using the information extracted directly from LLMs themselves. We found we can distinguish between benign and toxic prompts from the distribution of the first response token's logits. Using this idea, we build a robust detector of toxic prompts using a sparse logistic regression model on the first response token logits. Our scheme outperforms SOTA detectors under multiple metrics.

While there has been a large body of research attempting to circumvent tokenization for language modeling (Clark et al. 2022, Xue et al. 2022), the current consensus is that it is a necessary initial step for designing state-of-the-art performant language models. In this paper, we investigate tokenization from a theoretical point of view by studying the behavior of transformers on simple data generating processes. When trained on data drawn from certain simple $k^{\text{th}}$-order Markov processes for $k > 1$, transformers exhibit a surprising phenomenon - in the absence of tokenization, they empirically are incredibly slow or fail to learn the right distribution and predict characters according to a unigram model (Makkuva et al. 2024). With the addition of tokenization, however, we empirically observe that transformers break through this barrier and are able to model the probabilities of sequences drawn from the source near-optimally, achieving small cross-entropy loss. With this observation as starting point, we study the end-to-end cross-entropy loss achieved by transformers with and without tokenization. With the appropriate tokenization, we show that even the simplest unigram models (over tokens) learnt by transformers are able to model the probability of sequences drawn from $k^{\text{th}}$-order Markov sources near optimally. Our analysis provides a justification for the use of tokenization in practice through studying the behavior of transformers on Markovian data.

Auto-regressive large language models (LLMs) show impressive capacities to solve many complex reasoning tasks while struggling with some simple logical reasoning tasks such as inverse search: when trained on ''$A \to B$'' (e.g., *Tom is the parent of John*), LLM fails to directly conclude ''$B \gets A$'' (e.g., *John is the child of Tom*) during inference even if the two sentences are semantically identical, which is known as the ''reversal curse''. In this paper, we theoretically analyze the reversal curse via the training dynamics of (stochastic) gradient descent for two auto-regressive models: (1) a bilinear model that can be viewed as a simplification of a one-layer transformer; (2) one-layer transformers under certain assumptions. Our analysis reveals that for both models, the reversal curse is a consequence of the (effective) model weights *asymmetry*, i.e., the increase of weights from a token $A$ to token $B$ during training does not necessarily cause the increase of the weights from $B$ to $A$, which is caused by the training dynamics under certain choice of loss function and the optimization space of model parameters. Moreover, our analysis can be naturally applied to other logical reasoning tasks such as chain-of-thought (COT), which provides a new perspective different from previous work that focuses on expressivity. Finally, we conduct experiments to validate our theory on multi-layer transformers under different settings. Our code is available at [https://github.com/marlo-z/reversal_curse_analysis/](https://github.com/marlo-z/reversal_curse_analysis/).

Secure aggregation is a critical component in federated learning (FL), which enables the server to learn the aggregate model of the users without observing their local models. Conventionally, secure aggregation algorithms focus only on ensuring the privacy of individual users in a single training round. We contend that such designs can lead to significant privacy leakages over multiple training rounds, due to partial user selection/participation at each round of FL. In fact, we show that the conventional random user selection strategies in FL lead to leaking users' individual models within number of rounds that is linear in the number of users. To address this challenge, we introduce a secure aggregation framework, Multi-RoundSecAgg, with multi-round privacy guarantees. In particular, we introduce a new metric to quantify the privacy guarantees of FL over multiple training rounds, and develop a structured user selection strategy that guarantees the long-term privacy of each user (over any number of training rounds). Our framework also carefully accounts for the fairness and the average number of participating users at each round. Our experiments on MNIST, CIFAR-10 and CIFAR-100 datasets in the IID and the non-IID settings demonstrate the performance improvement over the baselines, both in terms of privacy protection and test accuracy.

We propose Byzantine-robust federated learning protocols with nearly optimal statistical rates based on recent progress in high dimensional robust statistics. In contrast to prior work, our proposed protocols improve the dimension dependence and achieve a near-optimal statistical rate for strongly convex losses. We also provide statistical lower bound for the problem. For experiments, we benchmark against competing protocols and show the empirical superiority of the proposed protocols.

Reinforcement learning (RL) provides a theoretical framework for continuously improving an agent’s behavior via trial and error. However, efficiently learning policies from scratch can be very difficult, particularly for tasks that present exploration challenges. In such settings, it might be desirable to initialize RL with an existing policy, offline data, or demonstrations. However, naively performing such initialization in RL often works poorly, especially for value-based methods. In this paper, we present a meta algorithm that can use offline data, demonstrations, or a pre-existing policy to initialize an RL policy, and is compatible with any RL approach. In particular, we propose Jump-Start Reinforcement Learning (JSRL), an algorithm that employs two policies to solve tasks: a guide-policy, and an exploration-policy. By using the guide-policy to form a curriculum of starting states for the exploration-policy, we are able to efficiently improve performance on a set of simulated robotic tasks. We show via experiments that it is able to significantly outperform existing imitation and reinforcement learning algorithms, particularly in the small-data regime. In addition, we provide an upper bound on the sample complexity of JSRL and show that with the help of a guide-policy, one can improve the sample complexity for non-optimism exploration methods from exponential in horizon to polynomial.

We study the problem of online learning in a two-player decentralized cooperative Stackelberg game. In each round, the leader first takes an action, followed by the follower who takes their action after observing the leader’s move. The goal of the leader is to learn to minimize the cumulative regret based on the history of interactions. Differing from the traditional formulation of repeated Stackelberg games, we assume the follower is omniscient, with full knowledge of the true reward, and that they always best-respond to the leader’s actions. We analyze the sample complexity of regret minimization in this repeated Stackelberg game. We show that depending on the reward structure, the existence of the omniscient follower may change the sample complexity drastically, from constant to exponential, even for linear cooperative Stackelberg games. This poses unique challenges for the learning process of the leader and the subsequent regret analysis.

We provide a theoretical framework for Reinforcement Learning with Human Feedback (RLHF). We show that when the underlying true reward is linear, under both Bradley-Terry-Luce (BTL) model (pairwise comparison) and Plackett-Luce (PL) model ($K$-wise comparison), MLE converges under certain semi-norm for the family of linear reward. On the other hand, when training a policy based on the learned reward model, we show that MLE fails while a pessimistic MLE provides policies with good performance under certain coverage assumption. We also show that under the PL model, both the true MLE and a different MLE which splits the $K$-wise comparison into pairwise comparisons converge, while the true MLE is asymptotically more efficient. Our results validate the empirical success of the existing RLHF algorithms, and provide new insights for algorithm design. Our analysis can also be applied for the problem of online RLHF and inverse reinforcement learning.

We revisit the problem of computing with noisy information considered in Feige et al. [1], which includes computing the OR function from noisy queries, and computing the MAX, SEARCH, and SORT functions from noisy pairwise comparisons. For K given elements, the goal is to correctly recover the desired function with probability at least 1 – δ when the outcome of each query is flipped with probability p. We consider both the adaptive sampling setting where each query can be adaptively designed based on past outcomes, and the non-adaptive sampling setting where the query cannot depend on past outcomes. The prior work provides tight bounds on the worst-case query complexity in terms of the dependence on K. However, the upper and lower bounds do not match in terms of the dependence on δ and p. We improve the lower bounds for all the four functions under both adaptive and non-adaptive query models. Most of our lower bounds match the upper bounds up to constant factors when either p or δ is bounded away from 0, while the ratio between the best prior upper and lower bounds go to infinity when p → 0 or p → 1/2. On the other hand, we also provide matching upper and lower bounds for the number of queries in expectation, improving both the upper and lower bounds for variable-length query model.

Large Language Models (LLMs) and other large foundation models have achieved impressive results, but their size exacerbates existing resource consumption and latency challenges. In particular, the large-scale deployment of these models is hindered by the significant resource requirements during inference. In this paper, we study two approaches for mitigating these challenges: employing a cache to store previous queries and learning a model selector to choose from an ensemble of models for query processing. Theoretically, we provide an optimal algorithm for jointly optimizing both approaches to reduce the inference cost in both offline and online tabular settings. By combining a caching algorithm, namely Greedy Dual Size with Frequency (GDSF) or Least Expected Cost (LEC), with a model selector, we achieve optimal rates in both offline and online settings. Empirically, simulations show that our caching and model selection algorithm greatly improves over the baselines, with up to $50\times$ improvement over the baseline when the ratio between the maximum cost and minimum cost is $100$. Experiments on real datasets show a $4.3\times$ improvement in FLOPs over the baseline when the ratio for FLOPs is $10$, and a $1.8\times$ improvement in latency when the ratio for average latency is $1.85$.

Self-training is a well-established technique in semi-supervised learning, which leverages unlabeled data by generating pseudo-labels and incorporating them with a limited labeled dataset for training. The effectiveness of self-training heavily relies on the accuracy of these pseudo-labels. In this paper, we introduce doubly-robust self-training, an innovative semi-supervised algorithm that provably balances between two extremes. When pseudo-labels are entirely incorrect, our method reduces to a training process solely using labeled data. Conversely, when pseudo-labels are completely accurate, our method transforms into a training process utilizing all pseudo-labeled data and labeled data, thus increasing the effective sample size. Through empirical evaluations on both the ImageNet dataset for image classification and the nuScenes autonomous driving dataset for 3D object detection, we demonstrate the superiority of the doubly-robust loss over the self-training baseline.

We propose A-Crab (Actor-Critic Regularized by Average Bellman error), a new practical algorithm for offline reinforcement learning (RL) in complex environments with insufficient data coverage. Our algorithm combines the marginalized importance sampling framework with the actor-critic paradigm, where the critic returns evaluations of the actor (policy) that are pessimistic relative to the offline data and have a small average (importance-weighted) Bellman error. Compared to existing methods, our algorithm simultaneously offers a number of advantages: (1) It achieves the optimal statistical rate of $1/\sqrt{N}$---where $N$ is the size of offline dataset---in converging to the best policy covered in the offline dataset, even when combined with general function approximators. (2) It relies on a weaker \textit{average} notion of policy coverage (compared to the $\ell_\infty$ single-policy concentrability) that exploits the structure of policy visitations. (3) It outperforms the data-collection behavior policy over a wide range of specific hyperparameters. We provide both theoretical analysis and experimental results to validate the effectiveness of our proposed algorithm. The code is available at https://github.com/zhuhl98/ACrab.

Contract theory studies the interactions between a principal and an agent when the two parties transact in the presence of private information [Bolton and Dewatripont, 2004, Faure-Grimaud et al., 2001, Salanié, 2005]. The principal would like to achieve her desired outcomes by hiring agents to work for her. The agent wishes to make money by working for the principal. They develop agreements in the form of a contract, which specifies how much the principal would pay under the different possible outcomes of the agent's work.

Policy optimization methods are one of the most widely used classes of Reinforcement Learning (RL) algorithms. However, theoretical understanding of these methods remains insufficient. Even in the episodic (time-inhomogeneous) tabular setting, the state-of-the-art theoretical result of policy-based method in Shani et al. (2020) is only $\tilde{O}(\sqrt{S^2AH^4K})$ where $S$ is the number of states, $A$ is the number of actions, $H$ is the horizon, and $K$ is the number of episodes, and there is a $\sqrt{SH}$ gap compared with the information theoretic lower bound $\tilde{\Omega}(\sqrt{SAH^3K})$ (Jin et al., 2018). To bridge such a gap, we propose a novel algorithm Reference-based Policy Optimization with Stable at Any Time guarantee (RPO-SAT), which features the property “Stable at Any Time”. We prove that our algorithm achieves $\tilde{O}(\sqrt{SAH^3K} + \sqrt{AH^4K})$ regret. When $S > H$, our algorithm is minimax optimal when ignoring logarithmic factors. To our best knowledge, RPO-SAT is the first computationally efficient, nearly minimax optimal policy-based algorithm for tabular RL.

We provide a general framework for designing Generative Adversarial Networks (GANs) to solve high-dimensional robust statistics problems, which aim at estimating unknown parameter of the true distribution given adversarially corrupted samples. Prior work [1], [2] focus on the problem of robust mean and covariance estimation when the true distribution lies in the family of Gaussian distributions or elliptical distributions, and analyze depth or scoring rule based GAN losses for the problem. Our work extend these to robust mean estimation, second-moment estimation, and robust linear regression when the true distribution only has bounded Orlicz norms, which includes the broad family of sub-Gaussian, sub-exponential and bounded moment distributions. We also provide a different set of sufficient conditions for the GAN loss to work: we only require its induced distance function to be a cumulative density function of some light-tailed distribution, which is easily satisfied by neural networks with sigmoid activation. In terms of techniques, our proposed GAN losses can be viewed as a smoothed and generalized Kolmogorov-Smirnov distance, which overcomes the computational intractability of the original Kolmogorov-Smirnov distance used in the weaken the distance approach in [3].

Online imitation learning is the problem of how best to mimic expert demonstrations, given access to the environment or an accurate simulator. Prior work has shown that in the \textit{infinite} sample regime, exact moment matching achieves value equivalence to the expert policy. However, in the \textit{finite} sample regime, even if one has no optimization error, empirical variance can lead to a performance gap that scales with $H^2 / N_{\text{exp}}$ for behavioral cloning and $H / N_{\text{exp}}$ for online moment matching, where $H$ is the horizon and $N_{\text{exp}}$ is the size of the expert dataset. We introduce the technique of ``replay estimation'' to reduce this empirical variance: by repeatedly executing cached expert actions in a stochastic simulator, we compute a smoother expert visitation distribution estimate to match. In the presence of general function approximation, we prove a meta theorem reducing the performance gap of our approach to the \textit{parameter estimation error} for offline classification (i.e. learning the expert policy). In the tabular setting or with linear function approximation, our meta theorem shows that the performance gap incurred by our approach achieves the optimal $\widetilde{O} \left( \min( H^{3/2} / N_{\text{exp}}, H / \sqrt{N_{\text{exp}}} \right)$ dependency, under significantly weaker assumptions compared to prior work. We implement multiple instantiations of our approach on several continuous control tasks and find that we are able to significantly improve policy performance across a variety of dataset sizes.

In the infinite-armed bandit problem, each arm's average reward is sampled from an unknown distribution, and each arm can be sampled further to obtain noisy estimates of the average reward of that arm. Prior work focuses on the best arm, i.e. estimating the maximum of the average reward distribution. We consider a general class of distribution functionals beyond the maximum and obtain optimal sample complexities in both offline and online settings. We show that online estimation, where the learner can sequentially choose whether to sample a new or existing arm, offers no advantage over the offline setting for estimating the mean functional, but significantly reduces the sample complexity for other functionals such as the median, maximum, and trimmed mean. We propose unified meta algorithms for the online and offline settings and derive matching lower bounds using different Wasserstein distances. For the special case of median estimation, we identify a curious thresholding phenomenon on the indistinguishability between Gaussian convolutions with respect to the noise level, which may be of independent interest.

Many goal-reaching reinforcement learning (RL) tasks have empirically verified that rewarding the agent on subgoals improves convergence speed and practical performance. We attempt to provide a theoretical framework to quantify the computational benefits of rewarding the completion of subgoals, in terms of the number of synchronous value iterations. In particular, we consider subgoals as one-way intermediate states, which can only be visited once per episode and propose two settings that consider these one-way intermediate states: the one-way single-path (OWSP) and the one-way multi-path (OWMP) settings. In both OWSP and OWMP settings, we demonstrate that adding intermediate rewards to subgoals is more computationally efficient than only rewarding the agent once it completes the goal of reaching a terminal state. We also reveal a trade-off between computational complexity and the pursuit of the shortest path in the OWMP setting: adding intermediate rewards significantly reduces the computational complexity of reaching the goal but the agent may not find the shortest path, whereas with sparse terminal rewards, the agent finds the shortest path at a significantly higher computational cost. We also corroborate our theoretical results with extensive experiments on the MiniGrid environments using Q-learning and some popular deep RL algorithms.

We study the problem of off-policy evaluation in the multi-armed bandit model with bounded rewards, and develop minimax rate-optimal procedures under three settings. First, when the behavior policy is known, we show that the Switch estimator, a method that alternates between the plug-in and importance sampling estimators, is minimax rate-optimal for all sample sizes. Second, when the behavior policy is unknown, we analyze performance in terms of the competitive ratio, thereby revealing a fundamental gap between the settings of known and unknown behavior policies. When the behavior policy is unknown, any estimator must have mean-squared error larger—relative to the oracle estimator equipped with the knowledge of the behavior policy— by a multiplicative factor proportional to the support size of the target policy. Moreover, we demonstrate that the plug-in approach achieves this worst-case competitive ratio up to a logarithmic factor. Third, we initiate the study of the partial knowledge setting in which it is assumed that the minimum probability taken by the behavior policy is known. We show that the plug-in estimator is optimal for relatively large values of the minimum probability, but is sub-optimal when the minimum probability is low. In order to remedy this gap, we propose a new estimator based on approximation by Chebyshev polynomials that provably achieves the optimal estimation error. Numerical experiments on both simulated and real data corroborate our theoretical findings.

Offline reinforcement learning (RL) algorithms seek to learn an optimal policy from a fixed dataset without active data collection. Based on the composition of the offline dataset, two main methods are used: imitation learning which is suitable for expert datasets, and vanilla offline RL which often requires uniform coverage datasets. From a practical standpoint, datasets often deviate from these two extremes and the exact data composition is usually unknown. To bridge this gap, we present a new offline RL framework, called single-policy concentrability, that smoothly interpolates between the two extremes of data composition, hence unifying imitation learning and vanilla offline RL. Under this new framework, we ask: can one develop an algorithm that achieves a minimax optimal rate adaptive to unknown data composition? To address this question, we consider a lower confidence bound (LCB) algorithm developed based on pessimism in the face of uncertainty in offline RL. We study finite-sample properties of LCB as well as information-theoretic limits in multi-armed bandits, contextual bandits, and Markov decision processes (MDPs). Our analysis reveals surprising facts about optimality rates. In particular, in both contextual bandits and RL, LCB achieves a fast convergence rate for nearly-expert datasets, analogous to the one achieved by imitation learning, contrary to the slow rate achieved in offline RL. In contextual bandits, we prove that LCB is adaptively optimal for the entire data composition range, achieving a smooth transition from imitation learning to offline RL. We further show that LCB is almost adaptively optimal in MDPs.

We study the statistical guarantees for the Imitation Learning (IL) problem in episodic MDPs.Rajaraman et al. (2020) show an information theoretic lower bound that in the worst case, a learner which can even actively query the expert policy suffers from a suboptimality growing quadratically in the length of the horizon, $H$. We study imitation learning under the $\mu$-recoverability assumption of Ross et al. (2011) which assumes that the difference in the $Q$-value under the expert policy across different actions in a state do not deviate beyond $\mu$ from the maximum. We show that the reduction proposed by Ross et al. (2010) is statistically optimal: the resulting algorithm upon interacting with the MDP for $N$ episodes results in a suboptimality bound of $\widetilde{\mathcal{O}} \left( \mu |\mathcal{S}| H / N \right)$ which we show is optimal up to log-factors. In contrast, we show that any algorithm which does not interact with the MDP and uses an offline dataset of $N$ expert trajectories must incur suboptimality growing as $\gtrsim |\mathcal{S}| H^2/N$ even under the $\mu$-recoverability assumption. This establishes a clear and provable separation of the minimax rates between the active setting and the no-interaction setting. We also study IL with linear function approximation. When the expert plays actions according to a linear classifier of known state-action features, we use the reduction to multi-class classification to show that with high probability, the suboptimality of behavior cloning is $\widetilde{O}(dH^2/N)$ given $N$ rollouts from the optimal policy. This is optimal up to log-factors but can be improved to $\widetilde{O}(dH/N)$ if we have a linear expert with parameter-sharing across time steps. In contrast, when the MDP transition structure is known to the learner such as in the case of simulators, we demonstrate fundamental differences compared to the tabular setting in terms of the performance of an optimal algorithm, Mimic-MD (Rajaraman et al. (2020)) when extended to the function approximation setting. Here, we introduce a new problem called confidence set linear classification, that can be used to construct sample-efficient IL algorithms.

Offline (or batch) reinforcement learning (RL) algorithms seek to learn an optimal policy from a fixed dataset without active data collection. Based on the composition of the offline dataset, two main methods are used: imitation learning which is suitable for expert datasets, and vanilla offline RL which often requires uniform coverage datasets. From a practical standpoint, datasets often deviate from these two extremes and the exact data composition is usually unknown. To bridge this gap, we present a new offline RL framework that smoothly interpolates between the two extremes of data composition, hence unifying imitation learning and vanilla offline RL. The new framework is centered around a weak version of the concentrability coefficient that measures the deviation of the behavior policy from the expert policy alone. Under this new framework, we ask: can one develop an algorithm that achieves a minimax optimal rate adaptive to unknown data composition? To address this question, we consider a lower confidence bound (LCB) algorithm developed based on pessimism in the face of uncertainty in offline RL. We study finite-sample properties of LCB as well as information-theoretic limits in multi-armed bandits, contextual bandits, and Markov decision processes (MDPs). Our analysis reveals surprising facts about optimality rates. In particular, in both contextual bandits and RL, LCB achieves a faster rate of $1/N$ for nearly-expert datasets compared to the usual rate of $1/\sqrt{N}$ in offline RL, where $N$ is the batch dataset sample size. In contextual bandits with at least two contexts, we prove that LCB is adaptively optimal for the entire data composition range, achieving a smooth transition from imitation learning to offline RL. We further show that LCB is almost adaptively optimal in MDPs.

In online reinforcement learning (RL), efficient exploration remains particularly challenging in high-dimensional environments with sparse rewards. In low-dimensional environments, where tabular parameterization is possible, count-based upper confidence bound (UCB) exploration methods achieve minimax near-optimal rates. However, it remains unclear how to efficiently implement UCB in realistic RL tasks that involve non-linear function approximation. To address this, we propose a new exploration approach via maximizing the deviation of the occupancy of the next policy from the explored regions. We add this term as an adaptive regularizer to the standard RL objective to balance exploration vs. exploitation. We pair the new objective with a provably convergent algorithm, giving rise to a new intrinsic reward that adjusts existing bonuses. The proposed intrinsic reward is easy to implement and combine with other existing RL algorithms to conduct exploration. As a proof of concept, we evaluate the new intrinsic reward on tabular examples across a variety of model-based and model-free algorithms, showing improvements over count-only exploration strategies. When tested on navigation and locomotion tasks from MiniGrid and DeepMind Control Suite benchmarks, our approach significantly improves sample efficiency over state-of-the-art methods.

We analyze the performance of the Tukey median estimator under total variation (TV) distance corruptions. Previous results show that under Huber’s additive corruption model, the breakdown point is 1/3 for high-dimensional halfspace- symmetric distributions. We show that under TV corruptions, the breakdown point reduces to 1/4 for the same set of distributions. We also show that a certain projection algorithm can attain the optimal breakdown point of 1/2. Both the Tukey median estimator and the projection algorithm achieve sample complexity linear in dimension.

Imitation learning (IL) aims to mimic the behavior of an expert policy in a sequential decision-making problem given only demonstrations. In this paper, we focus on understanding the minimax statistical limits of IL in episodic Markov Decision Processes (MDPs). We first consider the setting where the learner is provided a dataset of $N$ expert trajectories ahead of time, and cannot interact with the MDP. Here, we show that the policy which mimics the expert whenever possible is in expectation $\lesssim \frac{|\mathcal{S}| H^2 \log (N)}{N}$ suboptimal compared to the value of the expert, even when the expert plays a stochastic policy. Here $\mathcal{S}$ is the state space and $H$ is the length of the episode. Furthermore, we establish a suboptimality lower bound of $\gtrsim |\mathcal{S}| H^2 / N$ which applies even if the expert is constrained to be deterministic, or if the learner is allowed to actively query the expert at visited states while interacting with the MDP for $N$ episodes. To our knowledge, this is the first algorithm with suboptimality having no dependence on the number of actions, under no additional assumptions. We then propose a novel algorithm based on minimum-distance functionals in the setting where the transition model is given and the expert is deterministic. The algorithm is suboptimal by $\lesssim |\mathcal{S}| H^{3/2} / N$, matching our lower bound up to a $\sqrt{H}$ factor, and breaks the $\mathcal{O}(H^2)$ error compounding barrier of IL.

We present an efficient and practical (polynomial time) algorithm for online prediction in unknown and partially observed linear dynamical systems (LDS) under stochastic noise. When the system parameters are known, the optimal linear predictor is the Kalman filter. However, in unknown systems, the performance of existing predictive models is poor in important classes of LDS that are only marginally stable and exhibit long-term forecast memory. We tackle this problem by bounding the generalized Kolmogorov width of the Kalman filter coefficient set. This motivates the design of an algorithm, which we call spectral LDS improper predictor (SLIP), based on conducting a tight convex relaxation of the Kalman predictive model via spectral methods. We provide a finite-sample analysis, showing that our algorithm competes with the Kalman filter in hindsight with only logarithmic regret. Our regret analysis relies on Mendelson’s small-ball method, providing sharp error bounds without concentration, boundedness, or exponential forgetting assumptions. Empirical evaluations demonstrate that SLIP outperforms state-of-the-art methods in LDS prediction. Our theoretical and experimental results shed light on the conditions required for efficient probably approximately correct (PAC) learning of the Kalman filter from partially observed data.

We analyze bias correction methods using jackknife, bootstrap, and Taylor series. We focus on the binomial model, and consider the problem of bias correction for estimating f(p), where f ∈ C[0, 1] is arbitrary. We characterize the supremum norm of the bias of general jackknife and bootstrap estimators for any continuous functions, and demonstrate the in delete-d jackknife, different values of d may lead to drastically different behaviors in jackknife. We show that in the binomial model, iterating the bootstrap bias correction infinitely many times may lead to divergence of bias and variance, and demonstrate that the bias properties of the bootstrap bias corrected estimator after r - 1 rounds are of the same order as that of the r-jackknife estimator if a bounded coefficients condition is satisfied.

Generative Adversarial Networks (GANs) are a thriving unsupervised machine learning technique that has led to significant advances in various fields such as computer vision, natural language processing, among others. However, GANs are known to be difficult to train and usually suffer from mode collapse and the discriminator winning problem. To interpret the empirical observations of GANs and design better ones, we deconstruct the study of GANs into three components and make the following contributions. Formulation: we propose a perturbation view of the population target of GANs. Building on this interpretation, we show that GANs can be connected to the robust statistics framework, and propose a novel GAN architecture, termed as Cascade GANs, to provably recover meaningful low-dimensional generator approximations when the real distribution is high-dimensional and corrupted by outliers. Generalization: given a population target of GANs, we design a systematic principle, projection under admissible distance, to design GANs to meet the population requirement using only finite samples. We implement our principle in three cases to achieve polynomial and sometimes near-optimal sample complexities: (1) learning an arbitrary generator under an arbitrary pseudonorm; (2) learning a Gaussian location family under total variation distance, where we utilize our principle to provide a new proof for the near-optimality of the Tukey median viewed as GANs; (3) learning a low-dimensional Gaussian approximation of a high-dimensional arbitrary distribution under Wasserstein distance. We demonstrate a fundamental trade-off in the approximation error and statistical error in GANs, and demonstrate how to apply our principle in practice with only empirical samples to predict how many samples would be sufficient for GANs in order not to suffer from the discriminator winning problem. Optimization: we demonstrate alternating gradient descent is provably not locally asymptotically stable in optimizing the GAN formulation of PCA. We found that the minimax duality gap being non-zero might be one of the causes, and propose a new GAN architecture whose duality gap is zero, where the value of the game is equal to the previous minimax value (not the maximin value). We prove the new GAN architecture is globally asymptotically stable in solving PCA under alternating gradient descent.

We identify a trade-off between robustness and accuracy that serves as a guiding principle in the design of defenses against adversarial examples. Although this problem has been widely studied empirically, much remains unknown concerning the theory underlying this trade-off. In this work, we decompose the prediction error for adversarial examples (robust error) as the sum of the natural (classification) error and boundary error, and provide a differentiable upper bound using the theory of classification-calibrated loss, which is shown to be the tightest possible upper bound uniform over all probability distributions and measurable predictors. Inspired by our theoretical analysis, we also design a new defense method, TRADES, to trade adversarial robustness off against accuracy. Our proposed algorithm performs well experimentally in real-world datasets. The methodology is the foundation of our entry to the NeurIPS 2018 Adversarial Vision Challenge in which we won the 1st place out of 2,000 submissions, surpassing the runner-up approach by 11.41% in terms of mean L_2 perturbation distance.

A blockchain is a database of sequential events that is maintained by a distributed group of nodes. A key consensus problem in blockchains is that of determining the next block (data element) in the sequence. Many blockchains address this by electing a new node to propose each new block. The new block is (typically) appended to the tip of the proposer's local blockchain, and subsequently broadcast to the rest of the network. Without network delay (or adversarial behavior), this procedure would give a perfect chain, since each proposer would have the same view of the blockchain. A major challenge in practice is forking. Due to network delays, a proposer may not yet have the most recent block, and may therefore create a side chain that branches from the middle of the main chain. Forking reduces throughput, since only one a single main chain can survive, and all other blocks are discarded. We propose a new P2P protocol for blockchains called Barracuda, in which each proposer, prior to proposing a block, polls ℓ other nodes for their local blocktree information. Under a stochastic network model, we prove that this lightweight primitive improves throughput as if the entire network were a factor of ℓ faster. We provide guidelines on how to implement Barracuda in practice, guaranteeing robustness against several real-world factors.

We propose an efficient algorithm for approximate computation of the profile maximum likelihood (PML), a variant of maximum likelihood maximizing the probability of observing a sufficient statistic rather than the empirical sample. The PML has appealing theoretical properties, but is difficult to compute exactly. Inspired by observations gleaned from exactly solvable cases, we look for an approximate PML solution, which, intuitively, clumps comparably frequent symbols into one symbol. This amounts to lower-bounding a certain matrix permanent by summing over a subgroup of the symmetric group rather than the whole group during the computation. We extensively experiment with the approximate solution, and the empirical performance of our approach is competitive and sometimes significantly better than state-of-the-art performances for various estimation problems.

We show through case studies that it is easier to estimate the fundamental limits of data processing than to construct the explicit algorithms to achieve those limits. Focusing on binary classification, data compression, and prediction under logarithmic loss, we show that in the finite space setting, when it is possible to construct an estimator of the limits with vanishing error with n samples, it may require at least n ln n samples to construct an explicit algorithm to achieve the limits.

We present \emph{Local Moment Matching (LMM)}, a unified methodology for symmetric functional estimation and distribution estimation under Wasserstein distance. We construct an efficiently computable estimator that achieves the minimax rates in estimating the distribution up to permutation, and show that the plug-in approach of our unlabeled distribution estimator is “universal" in estimating symmetric functionals of discrete distributions. Instead of doing best polynomial approximation explicitly as in existing literature of functional estimation, the plug-in approach conducts polynomial approximation implicitly and attains the optimal sample complexity for the entropy, power sum and support size functionals.

For any Markov source, there exist universal codes whose normalized codelength approaches the Shannon limit asymptotically as the number of samples goes to infinity. This paper investigates how fast the gap between the normalized codelength of the “best” universal compressor and the Shannon limit (i.e. the compression redundancy) vanishes non-asymptotically in terms of the alphabet size and mixing time of the Markov source. We show that, for Markov sources whose relaxation time is at least 1+ [((2+c))/(√k)], where k is the state space size (and c > 0 is a constant), the phase transition for the number of samples required to achieve vanishing compression redundancy is precisely Θ(k2).

Entropy estimation is one of the prototypical problems in distribution property testing. To consistently estimate the Shannon entropy of a distribution on $S$ elements with independent samples, the optimal sample complexity scales sublinearly with $S$ as $\Theta(\frac{S}{\log S})$ as shown by Valiant and Valiant \cite{Valiant--Valiant2011}. Extending the theory and algorithms for entropy estimation to dependent data, this paper considers the problem of estimating the entropy rate of a stationary reversible Markov chain with $S$ states from a sample path of $n$ observations. We show that \begin{itemize} \item Provided the Markov chain mixes not too slowly, \textit{i.e.}, the relaxation time is at most $O(\frac{S}{\ln^3 S})$, consistent estimation is achievable when $n \gg \frac{S^2}{\log S}$. \item Provided the Markov chain has some slight dependency, \textit{i.e.}, the relaxation time is at least $1+\Omega(\frac{\ln^2 S}{\sqrt{S}})$, consistent estimation is impossible when $n \lesssim \frac{S^2}{\log S}$. \end{itemize} Under both assumptions, the optimal estimation accuracy is shown to be $\Theta(\frac{S^2}{n \log S})$. In comparison, the empirical entropy rate requires at least $\Omega(S^2)$ samples to be consistent, even when the Markov chain is memoryless. In addition to synthetic experiments, we also apply the estimators that achieve the optimal sample complexity to estimate the entropy rate of the English language in the Penn Treebank and the Google One Billion Words corpora, which provides a natural benchmark for language modeling and relates it directly to the widely used perplexity measure.

We analyze the Kozachenko–Leonenko (KL) fixed k-nearest neighbor estimator for the differential entropy. We obtain the first uniform upper bound on its performance for any fixed k over H\"{o}lder balls on a torus without assuming any conditions on how close the density could be from zero. Accompanying a recent minimax lower bound over the H\"{o}lder ball, we show that the KL estimator for any fixed k is achieving the minimax rates up to logarithmic factors without cognizance of the smoothness parameter s of the H\"{o}lder ball for $s \in (0,2]$ and arbitrary dimension d, rendering it the first estimator that provably satisfies this property.

We consider the problem of estimating the L1distance between two discrete probability measures P and Q from empirical data in a nonasymptotic and large alphabet setting. When Q is known and one obtains n samples from P, we show that for every Q, the minimax rate-optimal estimator with n samples achieves performance comparable to that of the maximum likelihood estimator with n ln n samples. When both P and Q are unknown, we construct minimax rate-optimal estimators, whose worst case performance is essentially that of the known Q case with Q being uniform, implying that Q being uniform is essentially the most difficult case. The effective sample size enlargement phenomenon, identified by Jiao et al., holds both in the known Q case for every Q and the Q unknown case. However, the construction of optimal estimators for ∥P - Q∥1requires new techniques and insights beyond the approximation-based method of functional estimation by Jiao et al.

Fundamental relations between information and estimation have been established in the literature for the continuous time Gaussian and Poisson channels. In this paper, we demonstrate that such relations hold for a much larger family of continuous-time channels. We introduce the family of semi-martingale channels where the channel output is a semi-martingale stochastic process, and the channel input modulates the characteristics of the semi-martingale. For these channels, which includes as a special case the continuous time Gaussian and Poisson models, we establish new representations relating the mutual information between the channel input and output to an optimal causal filtering loss, thereby unifying and considerably extending results from the Gaussian and Poisson settings. Extensions to the setting of mismatched estimation are also presented where the relative entropy between the laws governing the output of the channel under two different input distributions is equal to the cumulative difference between the estimation loss incurred by using the mismatched and optimal causal filters, respectively. The main tool underlying these results is the Doob-Meyer decomposition of a class of sub-martingales. The results in this paper can be viewed as the continuous-time analogues of recent generalizations for relations between information and estimation for discrete-time Lévy channels.

We propose a framework to analyze and quantify the bias in adaptive data analysis. It generalizes that proposed by Russo and Zou'15, applying to measurements whose moment generating function exists, measurements with a finite p-norm, and measurements in general Orlicz spaces. We introduce a new class of dependence measures which retain key properties of mutual information while more effectively quantifying the exploration bias for heavy tailed distributions. We provide examples of cases where our bounds are nearly tight in situations where the original framework of Russo and Zou'15 does not apply.

We consider the problem of estimating functionals of discrete distributions, and focus on a tight (up to universal multiplicative constants for each specific functional) nonasymptotic analysis of the worst case squared error risk of widely used estimators. We apply concentration inequalities to analyze the random fluctuation of these estimators around their expectations and the theory of approximation using positive linear operators to analyze the deviation of their expectations from the true functional, namely their bias. We explicitly characterize the worst case squared error risk incurred by the maximum likelihood estimator (MLE) in estimating the Shannon entropy H(P) = Σi=1S-piln pi, and the power sum Fα(P) = Σi=1Spiα, α > 0, up to universal multiplicative constants for each fixed functional, for any alphabet size S ≤ ∞ and sample size n for which the risk may vanish. As a corollary, for Shannon entropy estimation, we show that it is necessary and sufficient to have n ≫ S observations for the MLE to be consistent. In addition, we establish that it is necessary and sufficient to consider n ≫ S1/αsamples for the MLE to consistently estimate Fα(P), 01/α/ ln S samples, which implies that the MLE has a strictly sub-optimal sample complexity. When 1-2(α-1)for infinite alphabet size, while the minimax squared error rate is (n ln n)-2(α-1). When α ≥ 3/2, the MLE achieves the minimax optimal rate n-1regardless of the alphabet size. As an application of the general theory, we analyze the Dirichlet prior smoothing techniques for Shannon entropy estimation. In this context, one approach is to plug-in the Dirichlet prior smoothed distribution into the entropy functional, while the other one is to calculate the Bayes estimator for entropy under the Dirichlet prior for squared error, which is the conditional expectation. We show that in general such estimators do not improve over the maximum likelihood estimator. No matter how we tune the parameters in the Dirichlet prior, this approach cannot achieve the minimax rates in entropy estimation. The performance of the minimax rate-optimal estimator with n samples is essentially at least as good as that of Dirichlet smoothed entropy estimators with n ln n samples.

Fundamental relations between information and estimation have been established in the literature for the discrete-time Gaussian and Poisson channels. In this paper, we demonstrate that such relations hold for a much larger class of observation models. We introduce the natural family of discrete-time Lévy channels where the distribution of the output conditioned on the input is infinitely divisible. For Lévy channels, we establish new representations relating the mutual information between the channel input and output to an optimal expected estimation loss, thereby unifying and considerably extending results from the Gaussian and Poisson settings. We demonstrate the richness of our results by working out two examples of Lévy channels, namely the gamma channel and the negative binomial channel, with corresponding relations between information and estimation. Extensions to the setting of mismatched estimation are also presented.

We consider the problem of estimating the L1distance between two discrete probability measures P and Q from empirical data in a nonasymptotic and large alphabet setting. We construct minimax rate-optimal estimators for L1(P,Q) when Q is either known or unknown, and show that the performance of the optimal estimators with n samples is essentially that of the Maximum Likelihood Estimators (MLE) with n ln n samples. Hence, we demonstrate that the effective sample size enlargement phenomenon, discovered and discussed in Jiao et al. (2015), holds for this problem as well. However, the construction of optimal estimators for L1(P,Q) requires new techniques and insights outside the scope of the Approximation methodology of functional estimation in Jiao et al. (2015).

Fundamental relations between information and estimation have been established in the literature for the Gaussian and Poisson channels. In this work, we demonstrate that such relations hold for a much larger family of continuous-time channels. We introduce the family of semi-martingale channels where the channel output is a continuous time semi-martingale, and the channel input modulates the characteristics of the semi-martingale. For these channels, which includes as a special case the Gaussian and Poisson models, we establish new representations relating the mutual information between the channel input and output to an optimal causal filtering loss, thereby unifying and considerably extending results from the Gaussian and Poisson settings. Extensions to the setting of mismatched estimation are also presented where the relative entropy between the laws governing the output of the channel under two different input distributions is equal to the cumulative difference between the estimation loss incurred by using the mismatched and optimal causal filters respectively. The results in this work can be viewed as the continuous-time analogues of recent generalizations for relations between information and estimation for scalar transformations via Lévy channels.

The Dirichlet prior is widely used in estimating discrete distributions and functionals of discrete distributions. In terms of Shannon entropy estimation, one approach is to plug-in the Dirichlet prior smoothed distribution into the entropy functional, while the other one is to calculate the Bayes estimator for entropy under the Dirichlet prior for squared error, which is the conditional expectation. We show that in general they do not improve over the maximum likelihood estimator, which plugs-in the empirical distribution into the entropy functional. No matter how we tune the parameters in the Dirichlet prior, this approach cannot achieve the minimax rates in entropy estimation, as recently characterized by Jiao, Venkat, Han, and Weissman [1], and Wu and Yang [2]. The performance of the minimax rate-optimal estimator with n samples is essentially at least as good as that of the Dirichlet smoothed entropy estimators with n ln n samples. We harness the theory of approximation using positive linear operators for analyzing the bias of plug-in estimators for general functionals under arbitrary statistical models, thereby further consolidating the interplay between these two fields, which was thoroughly exploited by Jiao, Venkat, Han, and Weissman [3] in estimating various functionals of discrete distributions. We establish new results in approximation theory, and apply them to analyze the bias of the Dirichlet prior smoothed plug-in entropy estimator. This interplay between bias analysis and approximation theory is of relevance and consequence far beyond the specific problem setting in this paper.

We consider estimating the Shannon entropy of a discrete distribution P from n i.i.d. samples. Recently, Jiao, Venkat, Han, and Weissman (JVHW), and Wu and Yang constructed approximation theoretic estimators that achieve the minimax L2rates in estimating entropy. Their estimators are consistent given n ≫ S/lnS samples, where S is the support size, and it is the best possible sample complexity. In contrast, the Maximum Likelihood Estimator (MLE), which is the empirical entropy, requires n ≫ S samples. In the present paper we significantly refine the minimax results of existing work. To alleviate the pessimism of minimaxity, we adopt the adaptive estimation framework, and show that the JVHW estimator is an adaptive estimator, i.e., it achieves the minimax rates simultaneously over a nested sequence of subsets of distributions P, without knowing the support size S or which subset P lies in. We also characterize the maximum risk of the MLE over this nested sequence, and show, for every subset in the sequence, that the performance of the minimax rate-optimal estimator with n samples is essentially that of the MLE with n ln n samples, thereby further substantiating the generality of “effective sample size enlargement” phenomenon discovered by Jiao, Venkat, Han, and Weissman. We provide a “pointwise” explanation of the sample size enlargement phenomenon, which states that for sufficiently small probabilities, the bias function of the JVHW estimator with n samples is nearly that of the MLE with n ln n samples.

We analyze the problem of discrete distribution estimation under ℓ1loss. We provide non-asymptotic upper and lower bounds on the maximum risk of the empirical distribution (the maximum likelihood estimator), and the minimax risk in regimes where the alphabet size S may grow with the number of observations n. We show that among distributions with bounded entropy H, the asymptotic maximum risk for the empirical distribution is 2H / ln n, while the asymptotic minimax risk is H / ln n. Moreover, a hard-thresholding estimator, whose threshold does not depend on the unknown upper bound H, is asymptotically minimax. We draw connections between our work and the literature on density estimation, entropy estimation, total variation distance (ℓ1divergence) estimation, joint distribution estimation in stochastic processes, normal mean estimation, and adaptive estimation.

The Maximum Likelihood Estimator (MLE) is widely used in estimating information measures, and involves “plugging-in” the empirical distribution of the data to estimate a given functional of the unknown distribution. In this work we propose a general framework and procedure to analyze the nonasymptotic performance of the MLE in estimating functionals of discrete distributions, under the worst-case mean squared error criterion. We show that existing theory is insufficient for analyzing the bias of the MLE, and propose to apply the theory of approximation using positive linear operators to study this bias. The variance is controlled using the well-known tools from the literature on concentration inequalities. Our techniques completely characterize the maximum L2risk incurred by the MLE in estimating the Shannon entropy H(P) = Σi=1S-piln pi, and Fα(P) = Σi=1Spiαup to a multiplicative constant. As a corollary, for Shannon entropy estimation, we show that it is necessary and sufficient to have n ≪ S observations for the MLE to be consistent, where S represents the support size. In addition, we obtain that it is necessary and sufficient to consider n ≪ S1/αsamples for the MLE to consistently estimate Fα(P); 01/α/ ln S samples, which implies that the MLE is strictly sub-optimal. When 12rate of convergence for the MLE is n-2(α-1)for infinite support size, while the minimax L2rate is (n ln n)-2(α-1). When α ≥ 3/2, the MLE achieves the minimax optimal L2convergence rate n-1regardless of the support size.

We propose a general methodology for the construction and analysis of minimax estimators for functionals of discrete distributions, where the support size S is unknown and may be comparable to the number of observations n. We illustrate the merit of our approach by thoroughly analyzing non-asymptotically the performance of the resulting schemes for estimating two important information measures: the entropy H(P) = Σi=1S-piln piand Fα(P) = Σi=1Spiα, α > 0. We obtain the minimax L2risks for estimating these functionals up to a universal constant. In particular, we demonstrate that our estimator achieves the optimal sample complexity n ≫ S / ln S for entropy estimation. We also demonstrate that the sample complexity for estimating Fα(P), 01/a/ln S, which can be achieved by our estimator and not by the popular plug-in Maximum Likelihood Estimator (MLE). For 12rate for estimating Fα(P) is (n ln n)-2(α-1)regardless of the support size, while the exact L2rate for the MLE is n-2(α-1). For all the above cases, the behavior of the minimax rate-optimal estimators with n samples is essentially that of the MLE with n ln n samples. Finally, we highlight the practical advantages of our schemes for the estimation of entropy and mutual information.

We consider the problem of discrete distribution estimation under l1loss. We provide tight upper and lower bounds on the maximum risk of the empirical distribution (the maximum likelihood estimator), and the minimax risk in regimes where the support size S may grow with the number of observations n. We show that among distributions with bounded entropy H, the asymptotic maximum risk for the empirical distribution is 2H/ln n, while the asymptotic minimax risk is H/ ln n. Moreover, we show that a hard-thresholding estimator oblivious to the unknown upper bound H, is essentially minimax. However, if we constrain the estimates to lie in the simplex of probability distributions, then the asymptotic minimax risk is again 2H/ ln n. We draw connections between our work and the literature on density estimation, entropy estimation, total variation distance (I1divergence) estimation, joint distribution estimation in stochastic processes, normal mean estimation, and adaptive estimation.

We consider a natural measure of relevance: the reduction in optimal prediction risk in the presence of side information. For any given loss function, this relevance measure captures the benefit of side information for performing inference on a random variable under this loss function. When such a measure satisfies a natural data processing property, and the random variable of interest has alphabet size greater than two, we show that it is uniquely characterized by the mutual information, and the corresponding loss function coincides with logarithmic loss. In doing so, our work provides a new characterization of mutual information, and justifies its use as a measure of relevance. When the alphabet is binary, we characterize the only admissible forms the measure of relevance can assume while obeying the specified data processing property. Our results naturally extend to measuring the causal influence between stochastic processes, where we unify different causality measures in the literature as instantiations of directed information.

We propose a general methodology for the construction and analysis of essentially minimax estimators for a wide class of functionals of finite dimensional parameters, and elaborate on the case of discrete distributions, where the support size S is unknown and may be comparable with or even much larger than the number of observations n. We treat the respective regions where the functional is nonsmooth and smooth separately. In the nonsmooth regime, we apply an unbiased estimator for the best polynomial approximation of the functional whereas, in the smooth regime, we apply a bias-corrected version of the maximum likelihood estimator (MLE). We illustrate the merit of this approach by thoroughly analyzing the performance of the resulting schemes for estimating two important information measures: 1) the entropy H(P) = ΣSi=1-piln piand 2) Fα(P) = ΣSi=1pαi, α > 0. We obtain the minimax L2rates for estimating these functionals. In particular, we demonstrate that our estimator achieves the optimal sample complexity n × S/ln S for entropy estimation. We also demonstrate that the sample complexity for estimating Fα(P), 01/α/ln S, which can be achieved by our estimator but not the MLE. For 12rate for estimating Fα(P) is (n ln n)-2(α-1)for infinite support size, while the maximum L2rate for the MLE is n-2(α-1). For all the above cases, the behavior of the minimax rate-optimal estimators with n samples is essentially that of the MLE (plug-in rule) with n ln n samples, which we term “effective sample size enlargement.” We highlight the practical advantages of our schemes for the estimation of entropy and mutual information. We compare our performance with various existing approaches, and demonstrate that our approach reduces running time and boosts the accuracy. Moreover, we show that the minimax rate-optimal mutual information estimator yielded by our framework leads to significant performance boosts over the Chow-Liu algorithm in learning graphical models. The wide use of information measure estimation suggests that the insights and estimators obtained in this paper could be broadly applicable.

Four problems related to information divergence measures defined on finite alphabets are considered. In three of the cases we consider, we illustrate a contrast which arises between the binary-alphabet and larger-alphabet settings. This is surprising in some instances, since characterizations for the larger-alphabet settings do not generalize their binary-alphabet counterparts. For example, we show that f-divergences are not the unique decomposable divergences on binary alphabets that satisfy the data processing inequality, despite contrary claims in the literature.

We consider a natural measure of the benefit of side information: the reduction in optimal estimation risk when side information is available to the estimator. When such a measure satisfies a natural data processing property, and the source alphabet has cardinality greater than two, we show that it is uniquely characterized by the optimal estimation risk under logarithmic loss, and the corresponding measure is equal to mutual information. Further, when the source alphabet is binary, we characterize the only admissible forms the measure of predictive benefit can assume. These results unify many causality measures in the literature as instantiations of directed information, and present a natural axiomatic characterization of mutual information without requiring the sum or recursivity property.

Fundamental relations between information and estimation have been established in the literature for the scalar Gaussian and Poisson channels. In this work, we demonstrate that such relations hold for a much larger class of observation models. We introduce the natural family of scalar Lévy channels where the distribution of the output conditioned on the input is infinitely divisible. For Lévy channels, we establish new representations relating the mutual information between the channel input and output to an optimal estimation loss, thereby unifying and considerably extending results from the Gaussian and Poissonian settings. We demonstrate the richness of our results by working out two examples of Lévy channels, namely the Gamma channel and the Negative Binomial channel, with corresponding relations between information and estimation. Extensions to the setting of mismatched estimation are also presented.

Four problems related to information divergence measures defined on finite alphabets are considered. In three of the cases we consider, we illustrate a contrast that arises between the binary-alphabet and larger alphabet settings. This is surprising in some instances, since characterizations for the larger alphabet settings do not generalize their binary-alphabet counterparts. In particular, we show that f-divergences are not the unique decomposable divergences on binary alphabets that satisfy the data processing inequality, thereby clarifying claims that have previously appeared in the literature. We also show that Kullback-Leibler (KL) divergence is the unique Bregman divergence, which is also an f-divergence for any alphabet size. We show that KL divergence is the unique Bregman divergence, which is invariant to statistically sufficient transformations of the data, even when nondecomposable divergences are considered. Like some of the problems we consider, this result holds only when the alphabet size is at least three.

Identities yielding optimal estimation interpretations for mutual information and relative entropy - paralleling those known for minimum mean squared estimation under additive Gaussian noise - were recently discovered for the Poisson channel by Atar and Weissman. We express these identities as equalities between expectations of the associated estimation and information theoretic random variables such as the actual estimation loss and the information density. By explicitly characterizing the relations between these random variables we show that they are related in much stronger pointwise senses that directly imply the known expectation identities while deepening our understanding of them. As an example for the nature of our results, consider the equality between the mutual information and the mean cumulative filtering loss of the optimal filter in continuous-time estimation. We show that the difference between the information density and the cumulative filtering loss is a martingale expressible as a stochastic integral. This explicit characterization not only directly recovers the previously known expectation relation, but allows to characterize other distributional properties of the random variables involved where some of the original objects of interest emerge in new and surprising roles. For example, we find that the increasing predictable part of the Doob-Meyer decomposition of the information density (which is a sub-martinagle) is nothing but the cumulative loss of the optimal filter.

Four estimators of the directed information rate between a pair of jointly stationary ergodic finite-alphabet processes are proposed, based on universal probability assignments. The first one is a Shannon–McMillan–Breiman-type estimator, similar to those used by Verdú in 2005 and Caiin 2006 for estimation of other information measures. We show the almost sure and$L_{1}$convergence properties of the estimator for any underlying universal probability assignment. The other three estimators map universal probability assignments to different functionals, each exhibiting relative merits such as smoothness, nonnegativity, and boundedness. We establish the consistency of these estimators in almost sure and$L_{1}$senses, and derive near-optimal rates of convergence in the minimax sense under mild conditions. These estimators carry over directly to estimating other information measures of stationary ergodic finite-alphabet processes, such as entropy rate and mutual information rate, with near-optimal performance and provide alternatives to classical approaches in the existing literature. Guided by these theoretical results, the proposed estimators are implemented using the context-tree weighting algorithm as the universal probability assignment. Experiments on synthetic and real data are presented, demonstrating the potential of the proposed schemes in practice and the utility of directed information estimation in detecting and measuring causal influence and delay.

We propose four approaches to estimating the directed information rate between a pair of jointly stationary ergodic processes with the help of universal probability assignments. The four approaches yield estimators with different merits such as nonnegativity and boundedness. We establish consistency of these estimators in various senses and derive near-optimal rates of convergence in the minimax sense under mild conditions. The estimators carry over directly to estimating other information measures of stationary ergodic processes, such as entropy rate and mutual information rate, and provide alternatives to classical approaches in the existing literature. Guided by the theoretical results, we use context tree weighting as the vehicle for the implementations of the proposed estimators. Experiments on synthetic and real data are presented, demonstrating the potential of the proposed schemes in practice and the efficacy of directed information estimation as a tool for detecting and measuring causality and delay.

In many signal detection and classification problems, we have knowledge of the distribution under each hypothesis, but not the prior probabilities. This paper is aimed at providing theory to quantify the performance of detection via estimating prior probabilities from either labeled or unlabeled training data. The error or risk is considered as a function of the prior probabilities. We show that the risk function is locally Lipschitz in the vicinity of the true prior probabilities, and the error of detectors based on estimated prior probabilities depends on the behavior of the risk function in this locality. In general, we show that the error of detectors based on the maximum likelihood estimate (MLE) of the prior probabilities converges to the Bayes error at a rate of$n^{-1/2}$, where$n$is the number of training data. If the behavior of the risk function is more favorable, then detectors based on the MLE have errors converging to the corresponding Bayes errors at optimal rates of the form$n^{-(1+\alpha)/2}$, where$\alpha > 0$is a parameter governing the behavior of the risk function with a typical value$\alpha = 1$. The limit$\alpha \rightarrow{} \infty$corresponds to a situation where the risk function is flat near the true probabilities, and thus insensitive to small errors in the MLE; in this case, the error of the detector based on the MLE converges to the Bayes error exponentially fast with$n$. We show that the bounds are achievable no matter given labeled or unlabeled training data and are minimax-optimal in the labeled case.

With the advancement of the sensor network technology and cyber physical systems [2], the merging between the virtual cyber space and the real physical world is bound to happen, which will impact the lifestyle of the human being. The metropolitan area sensor networks, composed of millions of heterogeneous sensors and penetrating every aspect of the city life, provide gigantic amount of real-time information of the city, which makes it possible to reconstruct the real world in an on-line cyber space.