Amir Yehudayoff

Nataly Brukhim, Daniel Carmon, Irit Dinur et al.

A seminal result in learning theory characterizes the PAC learnability of binary classes through the Vapnik-Chervonenkis dimension. Extending this characterization to the general multiclass setting has been open since the pioneering works on multiclass PAC learning in the late 1980s. This work resolves this problem: we characterize multiclass PAC learnability through the DS dimension, a combinatorial dimension defined by Daniely and Shalev-Shwartz (2014). The classical characterization of the binary case boils down to empirical risk minimization. In contrast, our characterization of the multiclass case involves a variety of algorithmic ideas; these include a natural setting we call list PAC learning. In the list learning setting, instead of predicting a single outcome for a given unseen input, the goal is to provide a short menu of predictions. Our second main result concerns the Natarajan dimension, which has been a central candidate for characterizing multiclass learnability. This dimension was introduced by Natarajan (1988) as a barrier for PAC learning. Whether the Natarajan dimension characterizes PAC learnability in general has been posed as an open question in several papers since. This work provides a negative answer: we construct a non-learnable class with Natarajan dimension one. For the construction, we identify a fundamental connection between concept classes and topology (i.e., colorful simplicial complexes). We crucially rely on a deep and involved construction of hyperbolic pseudo-manifolds by Januszkiewicz and Swiatkowski. It is interesting that hyperbolicity is directly related to learning problems that are difficult to solve although no obvious barriers exist. This is another demonstration of the fruitful links machine learning has with different areas in mathematics.

Noga Alon, Shay Moran, Hilla Schefler et al.

We provide a unified framework for characterizing pure and approximate differentially private (DP) learnability. The framework uses the language of graph theory: for a concept class $\mathcal{H}$, we define the contradiction graph $G$ of $\mathcal{H}$. Its vertices are realizable datasets, and two datasets $S,S'$ are connected by an edge if they contradict each other (i.e., there is a point $x$ that is labeled differently in $S$ and $S'$). Our main finding is that the combinatorial structure of $G$ is deeply related to learning $\mathcal{H}$ under DP. Learning $\mathcal{H}$ under pure DP is captured by the fractional clique number of $G$. Learning $\mathcal{H}$ under approximate DP is captured by the clique number of $G$. Consequently, we identify graph-theoretic dimensions that characterize DP learnability: the clique dimension and fractional clique dimension. Along the way, we reveal properties of the contradiction graph which may be of independent interest. We also suggest several open questions and directions for future research.

Zachary Chase, Bogdan Chornomaz, Shay Moran et al.

We use and adapt the Borsuk-Ulam Theorem from topology to derive limitations on list-replicable and globally stable learning algorithms. We further demonstrate the applicability of our methods in combinatorics and topology. We show that, besides trivial cases, both list-replicable and globally stable learning are impossible in the agnostic PAC setting. This is in contrast with the realizable case where it is known that any class with a finite Littlestone dimension can be learned by such algorithms. In the realizable PAC setting, we sharpen previous impossibility results and broaden their scope. Specifically, we establish optimal bounds for list replicability and global stability numbers in finite classes. This provides an exponential improvement over previous works and implies an exponential separation from the Littlestone dimension. We further introduce lower bounds for weak learners, i.e., learners that are only marginally better than random guessing. Lower bounds from previous works apply only to stronger learners. To offer a broader and more comprehensive view of our topological approach, we prove a local variant of the Borsuk-Ulam theorem in topology and a result in combinatorics concerning Kneser colorings. In combinatorics, we prove that if $c$ is a coloring of all non-empty subsets of $[n]$ such that disjoint sets have different colors, then there is a chain of subsets that receives at least $1+ \lfloor n/2\rfloor$ colors (this bound is sharp). In topology, we prove e.g. that for any open antipodal-free cover of the $d$-dimensional sphere, there is a point $x$ that belongs to at least $t=\lceil\frac{d+3}{2}\rceil$ sets.

Zachary Chase, Shay Moran, Amir Yehudayoff

Replicability is essential in science as it allows us to validate and verify research findings. Impagliazzo, Lei, Pitassi and Sorrell (`22) recently initiated the study of replicability in machine learning. A learning algorithm is replicable if it typically produces the same output when applied on two i.i.d. inputs using the same internal randomness. We study a variant of replicability that does not involve fixing the randomness. An algorithm satisfies this form of replicability if it typically produces the same output when applied on two i.i.d. inputs (without fixing the internal randomness). This variant is called global stability and was introduced by Bun, Livni and Moran ('20) in the context of differential privacy. Impagliazzo et al. showed how to boost any replicable algorithm so that it produces the same output with probability arbitrarily close to 1. In contrast, we demonstrate that for numerous learning tasks, global stability can only be accomplished weakly, where the same output is produced only with probability bounded away from 1. To overcome this limitation, we introduce the concept of list replicability, which is equivalent to global stability. Moreover, we prove that list replicability can be boosted so that it is achieved with probability arbitrarily close to 1. We also describe basic relations between standard learning-theoretic complexity measures and list replicable numbers. Our results, in addition, imply that besides trivial cases, replicable algorithms (in the sense of Impagliazzo et al.) must be randomized. The proof of the impossibility result is based on a topological fixed-point theorem. For every algorithm, we are able to locate a "hard input distribution" by applying the Poincaré-Miranda theorem in a related topological setting. The equivalence between global stability and list replicability is algorithmic.

Daniel Carmon, Roi Livni, Amir Yehudayoff

Stochastic convex optimization is one of the most well-studied models for learning in modern machine learning. Nevertheless, a central fundamental question in this setup remained unresolved: "How many data points must be observed so that any empirical risk minimizer (ERM) shows good performance on the true population?" This question was proposed by Feldman (2016), who proved that $Ω(\frac{d}ε+\frac{1}{ε^2})$ data points are necessary (where $d$ is the dimension and $ε>0$ is the accuracy parameter). Proving an $ω(\frac{d}ε+\frac{1}{ε^2})$ lower bound was left as an open problem. In this work we show that in fact $\tilde{O}(\frac{d}ε+\frac{1}{ε^2})$ data points are also sufficient. This settles the question and yields a new separation between ERMs and uniform convergence. This sample complexity holds for the classical setup of learning bounded convex Lipschitz functions over the Euclidean unit ball. We further generalize the result and show that a similar upper bound holds for all symmetric convex bodies. The general bound is composed of two terms: (i) a term of the form $\tilde{O}(\frac{d}ε)$ with an inverse-linear dependence on the accuracy parameter, and (ii) a term that depends on the statistical complexity of the class of $\textit{linear}$ functions (captured by the Rademacher complexity). The proof builds a mechanism for controlling the behavior of stochastic convex optimization problems.

Egor Bakaev, Florestan Brunck, Christoph Hertrich et al.

This work studies the expressivity of ReLU neural networks with a focus on their depth. A sequence of previous works showed that $\lceil \log_2(n+1) \rceil$ hidden layers are sufficient to compute all continuous piecewise linear (CPWL) functions on $\mathbb{R}^n$. Hertrich, Basu, Di Summa, and Skutella (NeurIPS'21 / SIDMA'23) conjectured that this result is optimal in the sense that there are CPWL functions on $\mathbb{R}^n$, like the maximum function, that require this depth. We disprove the conjecture and show that $\lceil\log_3(n-1)\rceil+1$ hidden layers are sufficient to compute all CPWL functions on $\mathbb{R}^n$. A key step in the proof is that ReLU neural networks with two hidden layers can exactly represent the maximum function of five inputs. More generally, we show that $\lceil\log_3(n-2)\rceil+1$ hidden layers are sufficient to compute the maximum of $n\geq 4$ numbers. Our constructions almost match the $\lceil\log_3(n)\rceil$ lower bound of Averkov, Hojny, and Merkert (ICLR'25) in the special case of ReLU networks with weights that are decimal fractions. The constructions have a geometric interpretation via polyhedral subdivisions of the simplex into ``easier'' polytopes.

Egor Bakaev, Florestan Brunck, Christoph Hertrich et al.

We study two models of ReLU neural networks: monotone networks (ReLU$^+$) and input convex neural networks (ICNN). Our focus is on expressivity, mostly in terms of depth, and we prove the following lower bounds. For the maximum function MAX$_n$ computing the maximum of $n$ real numbers, we show that ReLU$^+$ networks cannot compute MAX$_n$, or even approximate it. We prove a sharp $n$ lower bound on the ICNN depth complexity of MAX$_n$. We also prove depth separations between ReLU networks and ICNNs; for every $k$, there is a depth-2 ReLU network of size $O(k^2)$ that cannot be simulated by a depth-$k$ ICNN. The proofs are based on deep connections between neural networks and polyhedral geometry, and also use isoperimetric properties of triangulations.

Egor Bakaev, Florestan Brunck, Amir Yehudayoff

We study approximations of polytopes in the standard model for computing polytopes using Minkowski sums and (convex hulls of) unions. Specifically, we study the ability to approximate a target polytope by polytopes of a given depth. Our main results imply that simplices can only be ``trivially approximated''. On the way, we obtain a characterization of simplices as the only ``outer additive'' convex bodies.

Steve Hanneke, Shay Moran, Alexander Shlimovich et al.

Learning theory has traditionally followed a model-centric approach, focusing on designing optimal algorithms for a fixed natural learning task (e.g., linear classification or regression). In this paper, we adopt a complementary data-centric perspective, whereby we fix a natural learning rule and focus on optimizing the training data. Specifically, we study the following question: given a learning rule $\mathcal{A}$ and a data selection budget $n$, how well can $\mathcal{A}$ perform when trained on at most $n$ data points selected from a population of $N$ points? We investigate when it is possible to select $n \ll N$ points and achieve performance comparable to training on the entire population. We address this question across a variety of empirical risk minimizers. Our results include optimal data-selection bounds for mean estimation, linear classification, and linear regression. Additionally, we establish two general results: a taxonomy of error rates in binary classification and in stochastic convex optimization. Finally, we propose several open questions and directions for future research.

Elisabetta Cornacchia, Jan Hązła, Ido Nachum et al.

We study the implicit bias of ReLU neural networks trained by a variant of SGD where at each step, the label is changed with probability $p$ to a random label (label smoothing being a close variant of this procedure). Our experiments demonstrate that label noise propels the network to a sparse solution in the following sense: for a typical input, a small fraction of neurons are active, and the firing pattern of the hidden layers is sparser. In fact, for some instances, an appropriate amount of label noise does not only sparsify the network but further reduces the test error. We then turn to the theoretical analysis of such sparsification mechanisms, focusing on the extremal case of $p=1$. We show that in this case, the network withers as anticipated from experiments, but surprisingly, in different ways that depend on the learning rate and the presence of bias, with either weights vanishing or neurons ceasing to fire.

Iftach Haitner, Noam Mazor, Rotem Oshman et al.

Key-agreement protocols whose security is proven in the random oracle model are an important alternative to protocols based on public-key cryptography. In the random oracle model, the parties and the eavesdropper have access to a shared random function (an "oracle"), but the parties are limited in the number of queries they can make to the oracle. The random oracle serves as an abstraction for black-box access to a symmetric cryptographic primitive, such as a collision resistant hash. Unfortunately, as shown by Impagliazzo and Rudich [STOC '89] and Barak and Mahmoody [Crypto '09], such protocols can only guarantee limited secrecy: the key of any $\ell$-query protocol can be revealed by an $O(\ell^2)$-query adversary. This quadratic gap between the query complexity of the honest parties and the eavesdropper matches the gap obtained by the Merkle's Puzzles protocol of Merkle [CACM '78]. In this work we tackle a new aspect of key-agreement protocols in the random oracle model: their communication complexity. In Merkle's Puzzles, to obtain secrecy against an eavesdropper that makes roughly $\ell^2$ queries, the honest parties need to exchange $Ω(\ell)$ bits. We show that for protocols with certain natural properties, ones that Merkle's Puzzle has, such high communication is unavoidable. Specifically, this is the case if the honest parties' queries are uniformly random, or alternatively if the protocol uses non-adaptive queries and has only two rounds. Our proof for the first setting uses a novel reduction from the set-disjointness problem in two-party communication complexity. For the second setting we prove the lower bound directly, using information-theoretic arguments.

Gal Yehuda, Amir Yehudayoff

We prove that at least $Ω(n^{0.51})$ hyperplanes are needed to slice all edges of the $n$-dimensional hypercube. We provide a couple of applications: lower bounds on the computational complexity of parity, and a lower bound on the cover number of the hypercube by skew hyperplanes.

Olivier Bousquet, Steve Hanneke, Shay Moran et al.

How quickly can a given class of concepts be learned from examples? It is common to measure the performance of a supervised machine learning algorithm by plotting its "learning curve", that is, the decay of the error rate as a function of the number of training examples. However, the classical theoretical framework for understanding learnability, the PAC model of Vapnik-Chervonenkis and Valiant, does not explain the behavior of learning curves: the distribution-free PAC model of learning can only bound the upper envelope of the learning curves over all possible data distributions. This does not match the practice of machine learning, where the data source is typically fixed in any given scenario, while the learner may choose the number of training examples on the basis of factors such as computational resources and desired accuracy. In this paper, we study an alternative learning model that better captures such practical aspects of machine learning, but still gives rise to a complete theory of the learnable in the spirit of the PAC model. More precisely, we consider the problem of universal learning, which aims to understand the performance of learning algorithms on every data distribution, but without requiring uniformity over the distribution. The main result of this paper is a remarkable trichotomy: there are only three possible rates of universal learning. More precisely, we show that the learning curves of any given concept class decay either at an exponential, linear, or arbitrarily slow rates. Moreover, each of these cases is completely characterized by appropriate combinatorial parameters, and we exhibit optimal learning algorithms that achieve the best possible rate in each case. For concreteness, we consider in this paper only the realizable case, though analogous results are expected to extend to more general learning scenarios.

Ido Nachum, Amir Yehudayoff

This work provides an additional step in the theoretical understanding of neural networks. We consider neural networks with one hidden layer and show that when learning symmetric functions, one can choose initial conditions so that standard SGD training efficiently produces generalization guarantees. We empirically verify this and show that this does not hold when the initial conditions are chosen at random. The proof of convergence investigates the interaction between the two layers of the network. Our results highlight the importance of using symmetry in the design of neural networks.

Ido Nachum, Amir Yehudayoff

How many bits of information are revealed by a learning algorithm for a concept class of VC-dimension $d$? Previous works have shown that even for $d=1$ the amount of information may be unbounded (tend to $\infty$ with the universe size). Can it be that all concepts in the class require leaking a large amount of information? We show that typically concepts do not require leakage. There exists a proper learning algorithm that reveals $O(d)$ bits of information for most concepts in the class. This result is a special case of a more general phenomenon we explore. If there is a low information learner when the algorithm {\em knows} the underlying distribution on inputs, then there is a learner that reveals little information on an average concept {\em without knowing} the distribution on inputs.

Shay Moran, Ido Nachum, Itai Panasoff et al.

We study and provide exposition to several phenomena that are related to the perceptron's compression. One theme concerns modifications of the perceptron algorithm that yield better guarantees on the margin of the hyperplane it outputs. These modifications can be useful in training neural networks as well, and we demonstrate them with some experimental data. In a second theme, we deduce conclusions from the perceptron's compression in various contexts.

Ofer M. Shir, Amir Yehudayoff

We consider Evolution Strategies operating only with isotropic Gaussian mutations on positive quadratic objective functions, and investigate the covariance matrix when constructed out of selected individuals by truncation. We prove that the covariance matrix over $(1,λ)$-selected decision vectors becomes proportional to the inverse of the landscape Hessian as the population-size $λ$ increases. This generalizes a previous result that proved an equivalent phenomenon when sampling was assumed to take place in the vicinity of the optimum. It further confirms the classical hypothesis that statistical learning of the landscape is an inherent characteristic of standard Evolution Strategies, and that this distinguishing capability stems only from the usage of isotropic Gaussian mutations and rank-based selection. We provide broad numerical validation for the proven results, and present empirical evidence for its generalization to $(μ,λ)$-selection.

Ido Nachum, Jonathan Shafer, Amir Yehudayoff

How many bits of information are required to PAC learn a class of hypotheses of VC dimension $d$? The mathematical setting we follow is that of Bassily et al. (2018), where the value of interest is the mutual information $\mathrm{I}(S;A(S))$ between the input sample $S$ and the hypothesis outputted by the learning algorithm $A$. We introduce a class of functions of VC dimension $d$ over the domain $\mathcal{X}$ with information complexity at least $Ω\left(d\log \log \frac{|\mathcal{X}|}{d}\right)$ bits for any consistent and proper algorithm (deterministic or random). Bassily et al. proved a similar (but quantitatively weaker) result for the case $d=1$. The above result is in fact a special case of a more general phenomenon we explore. We define the notion of information complexity of a given class of functions $\mathcal{H}$. Intuitively, it is the minimum amount of information that an algorithm for $\mathcal{H}$ must retain about its input to ensure consistency and properness. We prove a direct sum result for information complexity in this context; roughly speaking, the information complexity sums when combining several classes.

Daniel M. Kane, Roi Livni, Shay Moran et al.

This work studies distributed learning in the spirit of Yao's model of communication complexity: consider a two-party setting, where each of the players gets a list of labelled examples and they communicate in order to jointly perform some learning task. To naturally fit into the framework of learning theory, the players can send each other examples (as well as bits) where each example/bit costs one unit of communication. This enables a uniform treatment of infinite classes such as half-spaces in $\mathbb{R}^d$, which are ubiquitous in machine learning. We study several fundamental questions in this model. For example, we provide combinatorial characterizations of the classes that can be learned with efficient communication in the proper-case as well as in the improper-case. These findings imply unconditional separations between various learning contexts, e.g.\ realizable versus agnostic learning, proper versus improper learning, etc. The derivation of these results hinges on a type of decision problems we term "{\it realizability problems}" where the goal is deciding whether a distributed input sample is consistent with an hypothesis from a pre-specified class. From a technical perspective, the protocols we use are based on ideas from machine learning theory and the impossibility results are based on ideas from communication complexity theory.

Shai Ben-David, Pavel Hrubes, Shay Moran et al.

We consider the following statistical estimation problem: given a family F of real valued functions over some domain X and an i.i.d. sample drawn from an unknown distribution P over X, find h in F such that the expectation of h w.r.t. P is probably approximately equal to the supremum over expectations on members of F. This Expectation Maximization (EMX) problem captures many well studied learning problems; in fact, it is equivalent to Vapnik's general setting of learning. Surprisingly, we show that the EMX learnability, as well as the learning rates of some basic class F, depend on the cardinality of the continuum and is therefore independent of the set theory ZFC axioms (that are widely accepted as a formalization of the notion of a mathematical proof). We focus on the case where the functions in F are Boolean, which generalizes classification problems. We study the interaction between the statistical sample complexity of F and its combinatorial structure. We introduce a new version of sample compression schemes and show that it characterizes EMX learnability for a wide family of classes. However, we show that for the class of finite subsets of the real line, the existence of such compression schemes is independent of set theory. We conclude that the learnability of that class with respect to the family of probability distributions of countable support is independent of the set theory ZFC axioms. We also explore the existence of a "VC-dimension-like" parameter that captures learnability in this setting. Our results imply that that there exist no "finitary" combinatorial parameter that characterizes EMX learnability in a way similar to the VC-dimension based characterization of binary valued classification problems.

Raef Bassily, Shay Moran, Ido Nachum et al.

We study learning algorithms that are restricted to using a small amount of information from their input sample. We introduce a category of learning algorithms we term $d$-bit information learners, which are algorithms whose output conveys at most $d$ bits of information of their input. A central theme in this work is that such algorithms generalize. We focus on the learning capacity of these algorithms, and prove sample complexity bounds with tight dependencies on the confidence and error parameters. We also observe connections with well studied notions such as sample compression schemes, Occam's razor, PAC-Bayes and differential privacy. We discuss an approach that allows us to prove upper bounds on the amount of information that algorithms reveal about their inputs, and also provide a lower bound by showing a simple concept class for which every (possibly randomized) empirical risk minimizer must reveal a lot of information. On the other hand, we show that in the distribution-dependent setting every VC class has empirical risk minimizers that do not reveal a lot of information.

Noga Alon, Moshe Babaioff, Yannai A. Gonczarowski et al.

In this work we derive a variant of the classic Glivenko-Cantelli Theorem, which asserts uniform convergence of the empirical Cumulative Distribution Function (CDF) to the CDF of the underlying distribution. Our variant allows for tighter convergence bounds for extreme values of the CDF. We apply our bound in the context of revenue learning, which is a well-studied problem in economics and algorithmic game theory. We derive sample-complexity bounds on the uniform convergence rate of the empirical revenues to the true revenues, assuming a bound on the $k$th moment of the valuations, for any (possibly fractional) $k>1$. For uniform convergence in the limit, we give a complete characterization and a zero-one law: if the first moment of the valuations is finite, then uniform convergence almost surely occurs; conversely, if the first moment is infinite, then uniform convergence almost never occurs.

Ofir David, Shay Moran, Amir Yehudayoff

This work continues the study of the relationship between sample compression schemes and statistical learning, which has been mostly investigated within the framework of binary classification. The central theme of this work is establishing equivalences between learnability and compressibility, and utilizing these equivalences in the study of statistical learning theory. We begin with the setting of multiclass categorization (zero/one loss). We prove that in this case learnability is equivalent to compression of logarithmic sample size, and that uniform convergence implies compression of constant size. We then consider Vapnik's general learning setting: we show that in order to extend the compressibility-learnability equivalence to this case, it is necessary to consider an approximate variant of compression. Finally, we provide some applications of the compressibility-learnability equivalences: (i) Agnostic-case learnability and realizable-case learnability are equivalent in multiclass categorization problems (in terms of sample complexity). (ii) This equivalence between agnostic-case learnability and realizable-case learnability does not hold for general learning problems: There exists a learning problem whose loss function takes just three values, under which agnostic-case and realizable-case learnability are not equivalent. (iii) Uniform convergence implies compression of constant size in multiclass categorization problems. Part of the argument includes an analysis of the uniform convergence rate in terms of the graph dimension, in which we improve upon previous bounds. (iv) A dichotomy for sample compression in multiclass categorization problems: If a non-trivial compression exists then a compression of logarithmic size exists. (v) A compactness theorem for multiclass categorization problems.

Ofer M. Shir, Jonathan Roslund, Amir Yehudayoff

We study the theoretical capacity to statistically learn local landscape information by Evolution Strategies (ESs). Specifically, we investigate the covariance matrix when constructed by ESs operating with the selection operator alone. We model continuous generation of candidate solutions about quadratic basins of attraction, with deterministic selection of the decision vectors that minimize the objective function values. Our goal is to rigorously show that accumulation of winning individuals carries the potential to reveal valuable information about the search landscape, e.g., as already practically utilized by derandomized ES variants. We first show that the statistically-constructed covariance matrix over such winning decision vectors shares the same eigenvectors with the Hessian matrix about the optimum. We then provide an analytic approximation of this covariance matrix for a non-elitist multi-child $(1,λ)$-strategy, which holds for a large population size $λ$. Finally, we also numerically corroborate our results.

Shay Moran, Amir Yehudayoff

Sample compression schemes were defined by Littlestone and Warmuth (1986) as an abstraction of the structure underlying many learning algorithms. Roughly speaking, a sample compression scheme of size $k$ means that given an arbitrary list of labeled examples, one can retain only $k$ of them in a way that allows to recover the labels of all other examples in the list. They showed that compression implies PAC learnability for binary-labeled classes, and asked whether the other direction holds. We answer their question and show that every concept class $C$ with VC dimension $d$ has a sample compression scheme of size exponential in $d$. The proof uses an approximate minimax phenomenon for binary matrices of low VC dimension, which may be of interest in the context of game theory.

Shay Moran, Amir Shpilka, Avi Wigderson et al.

In this work we study the quantitative relation between VC-dimension and two other basic parameters related to learning and teaching. Namely, the quality of sample compression schemes and of teaching sets for classes of low VC-dimension. Let $C$ be a binary concept class of size $m$ and VC-dimension $d$. Prior to this work, the best known upper bounds for both parameters were $\log(m)$, while the best lower bounds are linear in $d$. We present significantly better upper bounds on both as follows. Set $k = O(d 2^d \log \log |C|)$. We show that there always exists a concept $c$ in $C$ with a teaching set (i.e. a list of $c$-labeled examples uniquely identifying $c$ in $C$) of size $k$. This problem was studied by Kuhlmann (1999). Our construction implies that the recursive teaching (RT) dimension of $C$ is at most $k$ as well. The RT-dimension was suggested by Zilles et al. and Doliwa et al. (2010). The same notion (under the name partial-ID width) was independently studied by Wigderson and Yehudayoff (2013). An upper bound on this parameter that depends only on $d$ is known just for the very simple case $d=1$, and is open even for $d=2$. We also make small progress towards this seemingly modest goal. We further construct sample compression schemes of size $k$ for $C$, with additional information of $k \log(k)$ bits. Roughly speaking, given any list of $C$-labelled examples of arbitrary length, we can retain only $k$ labeled examples in a way that allows to recover the labels of all others examples in the list, using additional $k\log (k)$ information bits. This problem was first suggested by Littlestone and Warmuth (1986).