Work

Sparse graph codes for compression, sensing, and secrecy

Year: 2010
Type: dissertation
Abstract: Sparse graph codes were first introduced by Gallager over 40 years ago.Over the last two decades, such codes have been the subject of intense research, and capacityapproaching sparse graph codes with low complexity encoding and decoding algorithms have been designed for many channels.Motivated by the success of sparse graph codes for channel coding, we explore the use of sparse graph codes for four other problems related to compression, sensing, and security.First, we construct locally encodable and decodable source codes for a simple class of sources.Local encodability refers to the property that when the original source data changes slightly, the compression produced by the source code can be updated easily.Local decodability refers to the property that a single source symbol can be recovered without having to decode the entire source block.Second, we analyze a simple message-passing algorithm for compressed sensing recovery, and show that our algorithm provides a nontrivial f1/f1 guarantee.We also show that very sparse matrices and matrices whose entries must be either 0 or 1 have poor performance with respect to the restricted isometry property for the f2 norm.Third, we analyze the performance of a special class of sparse graph codes, LDPC codes, for the problem of quantizing a uniformly random bit string under Hamming distortion.We show that LDPC codes can come arbitrarily close to the rate-distortion bound using an optimal quantizer.This is a special case of a general result showing a duality between lossy source coding and channel coding-if we ignore computational complexity, then good channel codes are automatically good lossy source codes.We also prove a lower bound on the average degree of vertices in an LDPC code as a function of the gap to the rate-distortion bound.Finally, we construct efficient, capacity-achieving codes for the wiretap channel, a model of communication that allows one to provide information-theoretic, rather than computational, security guarantees.Our main results include the introduction of a new security critertion which is an information-theoretic analog of semantic security, First, I would like to thank my advisors, Greg and Devavrat.Greg has been a wonderful mentor since my days as an M. Eng.student, and I have learned so much from both Greg and Devavrat over the years.It goes without saying that they have taught me many valuable research skills, but they also taught me that if a conference deadline is at 7:00, don't try to upload your paper at 6:59! On a more serious note, they showed me the importance of taking the time to present the results of my research clearly.Indeed, this thesis would probably be incomprehensible but for their efforts to help me improve the clarity of the presentation.As I neared graduation, their career advice and help with the job search have been very useful, and I am thankful that I got to work with two advisors who, in addition to being great research mentors, take so much interest in their students' professional development.I would also like to thank Piotr, not only for serving on my thesis committee, but for teaching me about compressed sensing in his streaming algorithms class.Chapter 4 started out as my class project for his class, and he originally suggested that I extend my results for binary matrices to general sparse matrices.Thanks to all of the friends I have made over the last eight years.
Language: English
Open Access status: green
OpenAlex ID: W2168428256
Cites: 126
Cited by: 30
Related to: 9