arXiv subjects

Cheuk Ting Li

Publications and source records attributed to Cheuk Ting Li.

53 records · Page 3Linked to original sources

arXiv2021

Efficient Approximate Minimum Entropy Coupling of Multiple Probability Distributions

Given a collection of probability distributions $p_{1},\ldots,p_{m}$, the minimum entropy coupling is the coupling $X_{1},\ldots,X_{m}$ ($X_{i}\sim p_{i}$) with the smallest entropy $H(X_{1},\ldots,X_{m})$. While this problem is known to be NP-hard, we present an efficient algorithm for computing a coupling with entropy within 2 bits from the optimal value. More precisely, we construct a coupling with entropy within 2 bits from the entropy of the greatest lower bound of $p_{1},\ldots,p_{m}$ with respect to majorization. This construction is also valid when the collection of distributions is infinite, and when the supports of the distributions are infinite. Potential applications of our results include random number generation, entropic causal inference, and functional representation of random variables.

cs.IT↗

arXiv2020

Infinite Divisibility of Information

We study an information analogue of infinitely divisible probability distributions, where the i.i.d. sum is replaced by the joint distribution of an i.i.d. sequence. A random variable $X$ is called informationally infinitely divisible if, for any $n\ge1$, there exists an i.i.d. sequence of random variables $Z_{1},\ldots,Z_{n}$ that contains the same information as $X$, i.e., there exists an injective function $f$ such that $X=f(Z_{1},\ldots,Z_{n})$. While there does not exist informationally infinitely divisible discrete random variable, we show that any discrete random variable $X$ has a bounded multiplicative gap to infinite divisibility, that is, if we remove the injectivity requirement on $f$, then there exists i.i.d. $Z_{1},\ldots,Z_{n}$ and $f$ satisfying $X=f(Z_{1},\ldots,Z_{n})$, and the entropy satisfies $H(X)/n\le H(Z_{1})\le1.59H(X)/n+2.43$. We also study a new class of discrete probability distributions, called spectral infinitely divisible distributions, where we can remove the multiplicative gap $1.59$. Furthermore, we study the case where $X=(Y_{1},\ldots,Y_{m})$ is itself an i.i.d. sequence, $m\ge2$, for which the multiplicative gap $1.59$ can be replaced by $1+5\sqrt{(\log m)/m}$. This means that as $m$ increases, $(Y_{1},\ldots,Y_{m})$ becomes closer to being spectral infinitely divisible in a uniform manner. This can be regarded as an information analogue of Kolmogorov's uniform theorem. Applications of our result include independent component analysis, distributed storage with a secrecy constraint, and distributed random number generation.

cs.IT↗

arXiv2020

Asymptotically Scale-invariant Multi-resolution Quantization

A multi-resolution quantizer is a sequence of quantizers where the output of a coarser quantizer can be deduced from the output of a finer quantizer. In this paper, we propose an asymptotically scale-invariant multi-resolution quantizer, which performs uniformly across any choice of average quantization step, when the length of the range of input numbers is large. Scale invariance is especially useful in worst case or adversarial settings, ensuring that the performance of the quantizer would not be affected greatly by small changes of storage or error requirements. We also show that the proposed quantizer achieves a tradeoff between rate and error that is arbitrarily close to the optimum.

cs.IT↗

arXiv2019

Pairwise Near-maximal Grand Coupling of Brownian Motions

The well-known reflection coupling gives a maximal coupling of two one-dimensional Brownian motions with different starting points. Nevertheless, the reflection coupling does not generalize to more than two Brownian motions. In this paper, we construct a coupling of all Brownian motions with all possible starting points (i.e., a grand coupling), such that the coupling for any pair of the coupled processes is close to being maximal, that is, the distribution of the coupling time of the pair approaches that of the maximal coupling as the time tends to $0$ or $\infty$, and the coupling time of the pair is always within a multiplicative factor $2e^{2}$ from the maximal one. We also show that a grand coupling that is pairwise exactly maximal does not exist.

math.PR↗

arXiv2019

Pairwise Multi-marginal Optimal Transport and Embedding for Earth Mover's Distance

We investigate the problem of pairwise multi-marginal optimal transport, that is, given a collection of probability distributions $\{P_α\}$ on a Polish space $\mathcal{X}$, to find a coupling $\{X_α\}$, $X_α\sim P_α$, such that $\mathbf{E}[c(X_α,X_β)]\le r\inf_{X\sim P_α,Y\sim P_β}\mathbf{E}[c(X,Y)]$ for all $α,β$, where $c$ is a cost function and $r\ge1$. In other words, every pair $(X_α,X_β)$ has an expected cost at most a factor of $r$ from its lowest possible value. This can be regarded as a locality sensitive hash function for probability distributions, and has applications such as robust and distributed computation of transport plans. It can also be considered as a bi-Lipschitz embedding of the collection of probability distributions into the space of random variables taking values on $\mathcal{X}$. For $c(x,y)=\Vert x-y\Vert_2^q$ on $\mathbb{R}^n$, where $q>0$, we show that a finite $r$ is attainable if and only if either $n=1$ or $0<q<1$. As $n\to\infty$, the growth rate of the smallest possible $r$ is exactly $Θ(n^{q/2})$ if $0<q<1$. Hence, the metric space of probability distributions on $\mathbb{R}^n$ with finite $q$-th absolute moments, $0<q<1$, with the earth mover's distance (or 1-Wasserstein distance) with respect to the snowflake metric $c(x,y)=\Vert x-y\Vert_2^q$, is bi-Lipschitz embeddable into $L_1$ with distortion $O(n^{q/2})$. If we consider $c(x,y)=\Vert x-y\Vert_2$ (i.e., $q=1$) on the grid $[0..s]^n$ instead of $\mathbb{R}^n$, then $r=O(\sqrt{n}\log s)$ is attainable, which implies the embeddability of the space of probability distributions on $[0..s]^n$ into $L_1$ with distortion $O(\sqrt{n}\log s)$, and improves upon the $O(n\log s)$ result by Indyk and Thaper. The case of the discrete metric cost $c(x,y)=\mathbf{1}\{x\neq y\}$ and more general metric and ultrametric costs are also investigated.

math.PR↗

arXiv2019

A Unified Framework for One-shot Achievability via the Poisson Matching Lemma

We introduce a fundamental lemma called the Poisson matching lemma, and apply it to prove one-shot achievability results for various settings, namely channels with state information at the encoder, lossy source coding with side information at the decoder, joint source-channel coding, broadcast channels, distributed lossy source coding, multiple access channels, channel resolvability and wiretap channels. Our one-shot bounds improve upon the best known one-shot bounds in most of the aforementioned settings (except multiple access channels, channel resolvability and wiretap channels, where we recover bounds comparable to the best known bounds), with shorter proofs in some settings even when compared to the conventional asymptotic approach using typicality. The Poisson matching lemma replaces both the packing and covering lemmas, greatly simplifying the error analysis. This paper extends the work of Li and El Gamal on Poisson functional representation, which mainly considered variable-length source coding settings, whereas this paper studies fixed-length settings, and is not limited to source coding, showing that the Poisson functional representation is a viable alternative to typicality for most problems in network information theory.

cs.IT↗

arXiv2018

Minimax Learning for Remote Prediction

The classical problem of supervised learning is to infer an accurate predictor of a target variable $Y$ from a measured variable $X$ by using a finite number of labeled training samples. Motivated by the increasingly distributed nature of data and decision making, in this paper we consider a variation of this classical problem in which the prediction is performed remotely based on a rate-constrained description $M$ of $X$. Upon receiving $M$, the remote node computes an estimate $\hat Y$ of $Y$. We follow the recent minimax approach to study this learning problem and show that it corresponds to a one-shot minimax noisy source coding problem. We then establish information theoretic bounds on the risk-rate Lagrangian cost and a general method to design a near-optimal descriptor-estimator pair, which can be viewed as a rate-constrained analog to the maximum conditional entropy principle used in the classical minimax learning problem. Our results show that a naive estimate-compress scheme for rate-constrained prediction is not in general optimal.

cs.IT↗

arXiv2018

One-Shot Variable-Length Secret Key Agreement Approaching Mutual Information

This paper studies an information-theoretic one-shot variable-length secret key agreement problem with public discussion. Let $X$ and $Y$ be jointly distributed random variables, each taking values in some measurable space. Alice and Bob observe $X$ and $Y$ respectively, can communicate interactively through a public noiseless channel, and want to agree on a key length and a key that is approximately uniformly distributed over all bit sequences with the agreed key length. The public discussion is observed by an eavesdropper, Eve. The key should be approximately independent of the public discussion, conditional on the key length. We show that the optimal expected key length is close to the mutual information $I(X;Y)$ within a logarithmic gap. Moreover, an upper bound and a lower bound on the optimal expected key length can be written down in terms of $I(X;Y)$ only. This means that the optimal one-shot performance is always within a small gap of the optimal asymptotic performance regardless of the distribution of the pair $(X,Y)$. This one-shot result may find applications in situations where the components of an i.i.d. pair source $(X^{n},Y^{n})$ are observed sequentially and the key is output bit by bit with small delay, or in situations where the random source is not an i.i.d. or ergodic process.

cs.IT↗

arXiv2018

Strong Functional Representation Lemma and Applications to Coding Theorems

This paper shows that for any random variables $X$ and $Y$, it is possible to represent $Y$ as a function of $(X,Z)$ such that $Z$ is independent of $X$ and $I(X;Z|Y)\le\log(I(X;Y)+1)+4$ bits. We use this strong functional representation lemma (SFRL) to establish a bound on the rate needed for one-shot exact channel simulation for general (discrete or continuous) random variables, strengthening the results by Harsha et al. and Braverman and Garg, and to establish new and simple achievability results for one-shot variable-length lossy source coding, multiple description coding and Gray-Wyner system. We also show that the SFRL can be used to reduce the channel with state noncausally known at the encoder to a point-to-point channel, which provides a simple achievability proof of the Gelfand-Pinsker theorem.

cs.IT↗

arXiv2017

Extended Gray-Wyner System with Complementary Causal Side Information

We establish the rate region of an extended Gray-Wyner system for 2-DMS $(X,Y)$ with two additional decoders having complementary causal side information. This extension is interesting because in addition to the operationally significant extreme points of the Gray-Wyner rate region, which include Wyner's common information, G{á}cs-K{ö}rner common information and information bottleneck, the rate region for the extended system also includes the K{ö}rner graph entropy, the privacy funnel and excess functional information, as well as three new quantities of potential interest, as extreme points. To simplify the investigation of the 5-dimensional rate region of the extended Gray-Wyner system, we establish an equivalence of this region to a 3-dimensional mutual information region that consists of the set of all triples of the form $(I(X;U),\,I(Y;U),\,I(X,Y;U))$ for some $p_{U|X,Y}$. We further show that projections of this mutual information region yield the rate regions for many settings involving a 2-DMS, including lossless source coding with causal side information, distributed channel synthesis, and lossless source coding with a helper.

cs.IT↗

arXiv2016

Distributed Simulation of Continuous Random Variables

We establish the first known upper bound on the exact and Wyner's common information of $n$ continuous random variables in terms of the dual total correlation between them (which is a generalization of mutual information). In particular, we show that when the pdf of the random variables is log-concave, there is a constant gap of $n^{2}\log e+9n\log n$ between this upper bound and the dual total correlation lower bound that does not depend on the distribution. The upper bound is obtained using a computationally efficient dyadic decomposition scheme for constructing a discrete common randomness variable $W$ from which the $n$ random variables can be simulated in a distributed manner. We then bound the entropy of $W$ using a new measure, which we refer to as the erosion entropy.

cs.IT↗

arXiv2016

Maximal Correlation Secrecy

This paper shows that the Hirschfeld-Gebelein-Rényi maximal correlation between the message and the ciphertext provides good secrecy guarantees for cryptosystems that use short keys. We first establish a bound on the eavesdropper's advantage in guessing functions of the message in terms of maximal correlation and the Rényi entropy of the message. This result implies that maximal correlation is stronger than the notion of entropic security introduced by Russell and Wang. We then show that a small maximal correlation $ρ$ can be achieved via a randomly generated cipher with key length $\approx2\log(1/ρ)$, independent of the message length, and by a stream cipher with key length $2\log(1/ρ)+\log n+2$ for a message of length $n$. We establish a converse showing that these ciphers are close to optimal. This is in contrast to entropic security for which there is a gap between the lower and upper bounds. Finally, we show that a small maximal correlation implies secrecy with respect to several mutual information based criteria but is not necessarily implied by them. Hence, maximal correlation is a stronger and more practically relevant measure of secrecy than mutual information.

cs.IT↗

arXiv2016

An Efficient Feedback Coding Scheme with Low Error Probability for Discrete Memoryless Channels

Existing fixed-length feedback communication schemes are either specialized to particular channels (Schalkwijk--Kailath, Horstein), or apply to general channels but either have high coding complexity (block feedback schemes) or are difficult to analyze (posterior matching). This paper introduces a new fixed-length feedback coding scheme which achieves the capacity for all discrete memoryless channels, has an error exponent that approaches the sphere packing bound as the rate approaches the capacity, and has $O(n\log n)$ coding complexity. These benefits are achieved by judiciously combining features from previous schemes with new randomization technique and encoding/decoding rule. These new features make the analysis of the error probability for the new scheme easier than for posterior matching.

cs.IT↗

arXiv2016

Channel Diversity needed for Vector Space Interference Alignment

We consider vector space interference alignment strategies over the $K$-user interference channel and derive an upper bound on the achievable degrees of freedom as a function of the channel diversity $L$, where the channel diversity is modeled by $L$ real-valued parallel channels with coefficients drawn from a non-degenerate joint distribution. The seminal work of Cadambe and Jafar shows that when $L$ is unbounded, vector space interference alignment can achieve $1/2$ degrees of freedom per user independent of the number of users $K$. However wireless channels have limited diversity in practice, dictated by their coherence time and bandwidth, and an important question is the number of degrees of freedom achievable at finite $L$. When $K=3$ and if $L$ is finite, Bresler et al show that the number of degrees of freedom achievable with vector space interference alignment is bounded away from $1/2$, and the gap decreases inversely proportional to $L$. In this paper, we show that when $K\geq4$, the gap is significantly larger. In particular, the gap to the optimal $1/2$ degrees of freedom per user can decrease at most like $1/\sqrt{L}$, and when $L$ is smaller than the order of $2^{(K-2)(K-3)}$, it decays at most like $1/\sqrt[4]{L}$.

cs.IT↗

arXiv2016

A Universal Coding Scheme for Remote Generation of Continuous Random Variables

We consider a setup in which Alice selects a pdf $f$ from a set of prescribed pdfs $\mathscr{P}$ and sends a prefix-free codeword $W$ to Bob in order to allow him to generate a single instance of the random variable $X\sim f$. We describe a universal coding scheme for this setup and establish an upper bound on the expected codeword length when the pdf $f$ is bounded, orthogonally concave (which includes quasiconcave pdf), and has a finite first absolute moment. A dyadic decomposition scheme is used to express the pdf as a mixture of uniform pdfs over hypercubes. Alice randomly selects a hypercube according to its weight, encodes its position and size into $W$, and sends it to Bob who generates $X$ uniformly over the hypercube. Compared to previous results on channel simulation, our coding scheme applies to any continuous distribution and does not require two-way communication or shared randomness. We apply our coding scheme to classical simulation of quantum entanglement and obtain a better bound on the average codeword length than previously known.

cs.IT↗

arXiv2014

Locally Weighted Learning for Naive Bayes Classifier

As a consequence of the strong and usually violated conditional independence assumption (CIA) of naive Bayes (NB) classifier, the performance of NB becomes less and less favorable compared to sophisticated classifiers when the sample size increases. We learn from this phenomenon that when the size of the training data is large, we should either relax the assumption or apply NB to a "reduced" data set, say for example use NB as a local model. The latter approach trades the ignored information for the robustness to the model assumption. In this paper, we consider using NB as a model for locally weighted data. A special weighting function is designed so that if CIA holds for the unweighted data, it also holds for the weighted data. The new method is intuitive and capable of handling class imbalance. It is theoretically more sound than the locally weighted learners of naive Bayes that base classification only on the $k$ nearest neighbors. Empirical study shows that the new method with appropriate choice of parameter outperforms seven existing classifiers of similar nature.

stat.ML↗

arXiv2014

Exact Common Information

This paper introduces the notion of exact common information, which is the minimum description length of the common randomness needed for the exact distributed generation of two correlated random variables $(X,Y)$. We introduce the quantity $G(X;Y)=\min_{X\to W \to Y} H(W)$ as a natural bound on the exact common information and study its properties and computation. We then introduce the exact common information rate, which is the minimum description rate of the common randomness for the exact generation of a 2-DMS $(X,Y)$. We give a multiletter characterization for it as the limit $\bar{G}(X;Y)=\lim_{n\to \infty}(1/n)G(X^n;Y^n)$. While in general $\bar{G}(X;Y)$ is greater than or equal to the Wyner common information, we show that they are equal for the Symmetric Binary Erasure Source. We do not know, however, if the exact common information rate has a single letter characterization in general.

cs.IT↗