COLT 2019 Papers

We establish that algorithmic experiments in zero-sum games “fail miserably” to confirm the unique, sharp prediction of maxmin equilibration. Contradicting nearly a century of economic thought that treats zero-sum games nearly axiomatically as the exemplar symbol of economic stability, we prove that no meaningful prediction can be made about the day-to-day behavior of online learning dynamics in zero-sum games. Concretely, Multiplicative Weights Updates (MWU) with constant step-size is \emph{Lyapunov chaotic} in the dual (payoff) space. Simply put, let’s assume that an observer asks the agents playing Matching-Pennies whether they prefer Heads or Tails (and by how much in terms of aggregate payoff so far). The range of possible answers consistent with any arbitrary small set of initial conditions blows up exponentially with time everywhere in the payoff space. This result is \emph{robust} both \emph{algorithmically} as well as \emph{game theoretically}. \textbf{Algorithmic robustness:} Chaos is robust to agents using any of a general sub-family of Follow-the-Regularized-Leader (FTRL) algorithms, the well known regret-minimizing dynamics, even when agents mix-and-match dynamics, use different or slowly decreasing step-sizes. \textbf{Game theoretic robustness:} Chaos is robust to all affine variants of zero-sum games (strictly competitive games), network variants with arbitrary large number of agents and even to competitive settings beyond these. Our result is in stark contrast with the time-average convergence of online learning to (approximate) Nash equilibrium, a result widely reported as “(weak) convergence to equilibrium”.

We study the question of learning an adversarially robust predictor. We show that any hypothesis class $\mathcal{H}$ with finite VC dimension is robustly PAC learnable with an \emph{improper} learning rule. The requirement of being improper is necessary as we exhibit examples of hypothesis classes $\mathcal{H}$ with finite VC dimension that are \emph{not} robustly PAC learnable with any \emph{proper} learning rule.

In the general submatrix detection problem, the task is to detect the presence of a small $k \times k$ submatrix with entries sampled from a distribution $\mathcal{P}$ in an $n \times n$ matrix of samples from $\mathcal{Q}$. This formulation includes a number of well-studied problems, such as biclustering when $\mathcal{P}$ and $\mathcal{Q}$ are Gaussians and the planted dense subgraph formulation of community detection when the submatrix is a principal minor and $\mathcal{P}$ and $\mathcal{Q}$ are Bernoulli random variables. These problems all seem to exhibit a universal phenomenon: there is a statistical-computational gap depending on $\mathcal{P}$ and $\mathcal{Q}$ between the minimum $k$ at which this task can be solved and the minimum $k$ at which it can be solved in polynomial time. Our main result is to tightly characterize this computational barrier as a tradeoff between $k$ and the KL divergences between $\mathcal{P}$ and $\mathcal{Q}$ through average-case reductions from the planted clique conjecture. These computational lower bounds hold given mild assumptions on $\mathcal{P}$ and $\mathcal{Q}$ arising naturally from classical binary hypothesis testing. Our results recover and generalize the planted clique lower bounds for Gaussian biclustering in Ma and Wu (2015) and Brennan et al. (2018) and for the sparse and general regimes of planted dense subgraph in Hajek et al. (2015) and Brennan et al. (2018). This yields the first universality principle for computational lower bounds obtained through average-case reductions.

The method to derive uniform bounds with Gaussian and Rademacher complexities is extended to the case where the sample average is replaced by a nonlinear statistic. Tight bounds are obtained for U-statistics, smoothened L-statistics and error functionals of l2-regularized algorithms.

In this work, we consider the sample complexity required for testing the monotonicity of distributions over partial orders. A distribution $p$ over a poset is {\em monotone} if, for any pair of domain elements $x$ and $y$ such that $x \preceq y$, $p(x) \leq p(y)$. To understand the sample complexity of this problem, we introduce a new property called \emph{bigness} over a finite domain, where the distribution is $T$-big if the minimum probability for any domain element is at least $T$. We establish a lower bound of $\Omega(n/\log n)$ for testing bigness of distributions on domains of size $n$. We then build on these lower bounds to give $\Omega(n/\log{n})$ lower bounds for testing monotonicity over a matching poset of size $n$ and significantly improved lower bounds over the hypercube poset. We give sublinear sample complexity bounds for testing bigness and for testing monotonicity over the matching poset. We then give a number of tools for analyzing upper bounds on the sample complexity of the monotonicity testing problem.

Consider the problem of minimizing functions that are Lipschitz and strongly convex, but not necessarily differentiable. We prove that after $T$ steps of stochastic gradient descent, the error of the final iterate is $O(\log(T)/T)$ \emph{with high probability}. We also construct a function from this class for which the error of the final iterate of \emph{deterministic} gradient descent is $\Omega(\log(T)/T)$. This shows that the upper bound is tight and that, in this setting, the last iterate of stochastic gradient descent has the same general error rate (with high probability) as deterministic gradient descent. This resolves both open questions posed by Shamir (2012). An intermediate step of our analysis proves that the suffix averaging method achieves error $O(1/T)$ \emph{with high probability}, which is optimal (for any first-order optimization method). This improves results of Rakhlin et al. (2012) and Hazan and Kale (2014), both of which achieved error $O(1/T)$, but only in expectation, and achieved a high probability error bound of $O(\log \log(T)/T)$, which is suboptimal.

We introduce and study a class of probabilistic generative models, where the latent object is a finite-dimensional diffusion process on a finite time interval and the observed variable is drawn conditionally on the terminal point of the diffusion. We make the following contributions: We provide a unified viewpoint on both sampling and variational inference in such generative models through the lens of stochastic control. We quantify the expressiveness of diffusion-based generative models. Specifically, we show that one can efficiently sample from a wide class of terminal target distributions by choosing the drift of the latent diffusion from the class of multilayer feedforward neural nets, with the accuracy of sampling measured by the Kullback–Leibler divergence to the target distribution. Finally, we present and analyze a scheme for unbiased simulation of generative models with latent diffusions and provide bounds on the variance of the resulting estimators. This scheme can be implemented as a deep generative model with a random number of layers.

Gradient descent, when applied to the task of logistic regression, outputs iterates which are biased to follow a unique ray defined by the data. The direction of this ray is the maximum margin predictor of a maximal linearly separable subset of the data; the gradient descent iterates converge to this ray in direction at the rate $\cO(\nicefrac{\ln\ln t }{\ln t})$. The ray does not pass through the origin in general, and its offset is the bounded global optimum of the risk over the remaining data; gradient descent recovers this offset at a rate $\cO(\nicefrac{(\ln t)^2}{\sqrt{t}})$.

We prove that, for a broad range of problems, maximum-a-posteriori (MAP) estimation and approximate sampling of the posterior are at least as computationally difficult as maximum-likelihood (ML) estimation. By way of illustration, we show how hardness results for ML estimation of mixtures of Gaussians and topic models carry over to MAP estimation and approximate sampling under commonly used priors.

Consider the following problem: given two arbitrary densities $q_1,q_2$ and a sample-access to an unknown target density $p$, find which of the $q_i$’s is closer to $p$ in total variation. A remarkable result due to Yatracos shows that this problem is tractable in the following sense: there exists an algorithm that uses $O(\epsilon^{-2})$ samples from $p$ and outputs $q_i$ such that with high probability, $TV(q_i,p) \leq 3\cdot OPT + \epsilon$, where $OPT= \min\{TV(q_1,p),TV(q_2,p)\}$. Moreover, this result extends to any finite class of densities $\mathcal{Q}$: there exists an algorithm that outputs the best density in $\mathcal{Q}$ up to a multiplicative approximation factor of 3. We complement and extend this result by showing that: (i) the factor 3 can not be improved if one restricts the algorithm to output a density from $\mathcal{Q}$, and (ii) if one allows the algorithm to output arbitrary densities (e.g. a mixture of densities from $\mathcal{Q}$), then the approximation factor can be reduced to 2, which is optimal. In particular this demonstrates an advantage of improper learning over proper in this setup. We develop two approaches to achieve the optimal approximation factor of $2$: an adaptive one and a static one. Both approaches are based on a geometric point of view of the problem and rely on estimating surrogate metrics to the total variation. Our sample complexity bounds exploit techniques from {\it Adaptive Data Analysis}.

The effectiveness of model-based versus model-free methods is a long-standing question in reinforcement learning (RL). Motivated by recent empirical success of RL on continuous control tasks, we study the sample complexity of popular model-based and model-free algorithms on the Linear Quadratic Regulator (LQR). We show that for policy evaluation, a simple model-based plugin method requires asymptotically less samples than the classical least-squares temporal difference (LSTD) estimator to reach the same quality of solution; the sample complexity gap between the two methods can be at least a factor of state dimension. For policy optimization, we study a simple family of problem instances and show that nominal (certainty equivalence principle) control also requires several factors of state and input dimension fewer samples than the policy gradient method to reach the same level of control performance on these instances. Furthermore, the gap persists even when employing baselines commonly used in practice. To the best of our knowledge, this is the first theoretical result which demonstrates a separation in the sample complexity between model-based and model-free methods on a continuous control task.

We give nearly matching upper and lower bounds on the oracle complexity of finding $\epsilon$-stationary points $(\|\nabla F(x)\|\leq\epsilon$ in stochastic convex optimization. We jointly analyze the oracle complexity in both the local stochastic oracle model and the global oracle (or, statistical learning) model. This allows us to decompose the complexity of finding near-stationary points into optimization complexity and sample complexity, and reveals some surprising differences between the complexity of stochastic optimization versus learning. Notably, we show that in the global oracle/statistical learning model, only logarithmic dependence on smoothness is required to find a near-stationary point, whereas polynomial dependence on smoothness is necessary in the local stochastic oracle model. In other words, the separation in complexity between the two models can be exponential, and the folklore understanding that smoothness is required to find stationary points is only weakly true for statistical learning. Our upper bounds are based on extensions of a recent “recursive regularization” technique proposed by Allen-Zhu (2018). We show how to extend the technique to achieve near-optimal rates, and in particular show how to leverage the extra information available in the global oracle model. Our algorithm for the global model can be implemented efficiently through finite sum methods, and suggests an interesting new computational-statistical tradeoff.

We study the problem of recovering a hidden binary $k$-sparse $p$-dimensional vector $\beta$ from $n$ noisy linear observations $Y=X\beta+W$ where $X_{ij}$ are i.i.d. $\mathcal{N}(0,1)$ and $W_i$ are i.i.d. $\mathcal{N}(0,\sigma^2)$. A closely related hypothesis testing problem is to distinguish the pair $(X,Y)$ generated from this structured model from a corresponding null model where $(X,Y)$ consist of purely independent Gaussian entries. In the low sparsity $k=o(p)$ and high signal to noise ratio $k/\sigma^2=\Omega\left(1\right)$ regime, we establish an “All-or-Nothing” information-theoretic phase transition at a critical sample size $n^*=2 k\log \left(p/k\right) /\log \left(1+k/\sigma^2\right)$, resolving a conjecture of [GamarnikZadik17]. Specifically, we show that if $\liminf_{p\rightarrow \infty} n/n^*>1$, then the maximum likelihood estimator almost perfectly recovers the hidden vector with high probability and moreover the true hypothesis can be detected with a vanishing error probability. Conversely, if $\limsup_{p\rightarrow \infty} n/n^*<1$, then it becomes information-theoretically impossible even to recover an arbitrarily small but fixed fraction of the hidden vector support, or to test hypotheses strictly better than random guess. Our proof of the impossibility result builds upon two key techniques, which could be of independent interest. First, we use a conditional second moment method to upper bound the Kullback-Leibler (KL) divergence between the structured and the null model. Second, inspired by the celebrated area theorem, we establish a lower bound to the minimum mean squared estimation error of the hidden vector in terms of the KL divergence between the two models.

We study the problem of identity testing of symmetric markov chains. In this setting, we are given access to a single trajectory from a markov chain with unknown transition matrix $\bm{Q}$ and the goal is to determine whether $\bm{Q} = \bm{P}$ for some known matrix $\bm{P}$ or $\text{Dist}(\bm{P}, \bm{Q}) \geq \epsilon$ where $\text{Dist}$ is suitably defined. In recent work by Daskalakis et al, 2018, it was shown that it is possible to distinguish between the two cases provided the length of the observed trajectory is at least super-linear in the hitting time of $\bm{P}$ which may be arbitrarily large. In this paper, we propose an algorithm that avoids this dependence on hitting time thus enabling efficient testing of markov chains even in cases where it is infeasible to observe every state in the chain. Our algorithm is based on combining classical ideas from approximation algorithms with techniques for the spectral analysis of markov chains.

There has been significant study on the sample complexity of testing properties of distributions over large domains. For many properties, it is known that the sample complexity can be substantially smaller than the domain size. For example, over a domain of size $n$, distinguishing the uniform distribution from distributions that are far from uniform in $\ell_1$-distance uses only $O(\sqrt{n})$ samples. However, the picture is very different in the presence of arbitrary noise, even when the amount of noise is quite small. In this case, one must distinguish if samples are coming from a distribution that is $\epsilon$-close to uniform from the case where the distribution is $(1-\epsilon)$-far from uniform. The latter task requires nearly linear in $n$ samples (Valiant, 2008; Valiant and Valiant, 2017a). In this work, we present a noise model that on one hand is more tractable for the testing problem, and on the other hand represents a rich class of noise families. In our model, the noisy distribution is a mixture of the original distribution and noise, where the latter is known to the tester either explicitly or via sample access; the form of the noise is also known \emph{a priori}. Focusing on the identity and closeness testing problems leads to the following mixture testing question: Given samples of distributions $p, q_1,q_2$, can we test if $p$ is a mixture of $q_1$ and $q_2$? We consider this general question in various scenarios that differ in terms of how the tester can access the distributions, and show that indeed this problem is more tractable. Our results show that the sample complexity of our testers are exactly the same as for the classical non-mixture case.

We investigate the problem of identity testing for multidimensional histogram distributions. A distribution $p: D \to \mathbb{R}_+$, where $D \subseteq \mathbb{R}^d$, is called a $k$-histogram if there exists a partition of the domain into $k$ axis-aligned rectangles such that $p$ is constant within each such rectangle. Histograms are one of the most fundamental nonparametric families of distributions and have been extensively studied in computer science and statistics. We give the first identity tester for this problem with {\em sub-learning} sample complexity in any fixed dimension and a nearly-matching sample complexity lower bound. In more detail, let $q$ be an unknown $d$-dimensional $k$-histogram distribution in fixed dimension $d$, and $p$ be an explicitly given $d$-dimensional $k$-histogram. We want to correctly distinguish, with probability at least $2/3$, between the case that $p = q$ versus $\|p-q\|_1 \geq \epsilon$. We design an algorithm for this hypothesis testing problem with sample complexity $O((\sqrt{k}/\epsilon^2) 2^{d/2} \log^{2.5 d}(k/\epsilon))$ that runs in sample-polynomial time. Our algorithm is robust to model misspecification, i.e., succeeds even if $q$ is only promised to be {\em close} to a $k$-histogram. Moreover, for $k = 2^{\Omega(d)}$, we show a sample complexity lower bound of $(\sqrt{k}/\epsilon^2) \cdot \Omega(\log(k)/d)^{d-1}$ when $d\geq 2$. That is, for any fixed dimension $d$, our upper and lower bounds are nearly matching. Prior to our work, the sample complexity of the $d=1$ case was well-understood, but no algorithm with sub-learning sample complexity was known, even for $d=2$. Our new upper and lower bounds have interesting conceptual implications regarding the relation between learning and testing in this setting.

We study tensor completion in the agnostic setting. In the classical tensor completion problem, we receive $n$ entries of an unknown rank-$r$ tensor and wish to exactly complete the remaining entries. In agnostic tensor completion, we make \emph{no assumption} on the rank of the unknown tensor, but attempt to predict unknown entries as well as the best rank-$r$ tensor. For agnostic learning of third-order tensors with the square loss, we give the first polynomial time algorithm that obtains a “fast” (i.e., $O(1/n)$-type) rate improving over the rate obtained by reduction to matrix completion. Our prediction error rate to compete with the best $d\times{}d\times{}d$ tensor of rank-$r$ is $\tilde{O}(r^{2}d^{3/2}/n)$. We also obtain an exact oracle inequality that trades off estimation and approximation error. Our algorithm is based on the degree-six sum-of-squares relaxation of the tensor nuclear norm. The key feature of our analysis is to show that a certain characterization for the subgradient of the tensor nuclear norm can be encoded in the sum-of-squares proof system. This unlocks the standard toolbox for localization of empirical processes under the square loss, and allows us to establish restricted eigenvalue-type guarantees for various tensor regression models, with tensor completion as a special case. The new analysis of the relaxation complements Barak and Moitra (2016), who gave slow rates for agnostic tensor completion, and Potechin and Steurer (2017), who gave exact recovery guarantees for the noiseless setting. Our techniques are user-friendly, and we anticipate that they will find use elsewhere.

We provide a novel computer-assisted technique for systematically analyzing first-order methods for optimization. In contrast with previous works, the approach is particularly suited for handling sublinear convergence rates and stochastic oracles. The technique relies on semidefinite programming and potential functions. It allows simultaneously obtaining worst-case guarantees on the behavior of those algorithms, and assisting in choosing appropriate parameters for tuning their worst-case performances. The technique also benefits from comfortable tightness guarantees, meaning that unsatisfactory results can be improved only by changing the setting. We use the approach for analyzing deterministic and stochastic first-order methods under different assumptions on the nature of the stochastic noise. Among others, we treat unstructured noise with bounded variance, different noise models arising in over-parametrized expectation minimization problems, and randomized block-coordinate descent schemes.

We study discrete time dynamical systems governed by the state equation $h_{t+1}=\phi(Ah_t+Bu_t)$. Here $A,B$ are weight matrices, $\phi$ is an activation function, and $u_t$ is the input data. This relation is the backbone of recurrent neural networks (e.g. LSTMs) which have broad applications in sequential learning tasks. We utilize stochastic gradient descent to learn the weight matrices from a finite input/state trajectory $\{u_t,h_t\}_{t=0}^N$. We prove that SGD estimate linearly converges to the ground truth weights while using near-optimal sample size. Our results apply to increasing activations whose derivatives are bounded away from zero. The analysis is based on i) a novel SGD convergence result with nonlinear activations and ii) careful statistical characterization of the state vector. Numerical experiments verify the fast convergence of SGD on ReLU and leaky ReLU in consistence with our theory.