IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 63, NO. 11, JUNE 1, 2015 2915
On the Characterization of -Compressible
Ergodic Sequences
Jorge F. Silva, Member, IEEE, and Milan S. Derpich, Member, IEEE
Abstract—This work offers a necessary and sufficient condition
for a stationary and ergodic process to be -compressible in the
sense proposed by Amini, Unser and Marvasti [“Compressibility of
deterministic and random infinity sequences,” IEEE Trans. Signal
Process., vol. 59, no. 11, pp. 5193–5201, 2011, Def. 6]. The condition
reduces to check that the -moment of the invariant distribution
of the process is well defined, which contextualizes and extends the
result presented by Gribonval, Cevher and Davies in [“Compress-
ible distributions for high-dimensional statistics,” IEEE Trans. Inf.
Theory, vol. 58, no. 8, pp. 5016–5034, 2012, Prop. 1]. Furthermore,
for the scenario of non- -compressible ergodic sequences, we pro-
vide a closed-form expression for the best -term relative approx-
imation error (in the -norm sense) when only a fraction (rate) of
the most significant sequence coefficients are kept as the sequence-
length tends to infinity. We analyze basic properties of this rate-ap-
proximation error curve, which is again a function of the invariant
measure of the process. Revisiting the case of i.i.d. sequences, we
completely identify the family of -compressible processes, which
reduces to look at a polynomial order decay (heavy-tail) property
of the distribution.
Index Terms—Asymptotic analysis, best -term approximation
error analysis, compressed sensing, compressibility of infinite se-
quences, compressible priors, ergodic processes, heavy-tail distri-
butions.
I. INTRODUCTION
D
EFINING notions of compressibility for a stochastic
process, meaning that with high probability realizations
of the process can be well-approximated in some sense by
its best -term sparse version [3], has been a recent topic of
active research [1], [2], [4]–[6]. Quantifying compressibility
for random sequences and the identification of compressible
and sparse distributions (priors) are relevant problems consid-
ering the recent development of the compressed sensing theory
[7]–[9] and its applications. These results can play an important
role in regression [10], signal reconstruction (for instance in
Manuscript received July 28, 2014; revised January 11, 2015; accepted March
13, 2015. Date of publication April 02, 2015; date of current version May 05,
2015. The associate editor coordinating the review of this manuscript and ap-
proving it for publication was Prof. Fancesco Verde. This material is based
on work supported by grants of CONICYT-Chile, Fondecyt Grant 1140840
and the Advanced Center for Electrical and Electronic Engineering (AC3E),
Basal Project FB0008. In addition, the work of M. S. Derpich was partially sup-
ported by CONICYT Fondecyt Grant 1140384. (Corresponding Author: Jorge
F. Silva.)
J. F. Silva is with the Department of Electrical Engineering, Information and
Decision Systems Group, University of Chile, 412-3 Santiago, Chile (e-mail:
josilva@ing.uchile.cl).
M. S. Derpich is with the Department of Electronic Engineering, Universidad
Técnica Federico Santa María, Valparaíso 2390123, Chile (e-mail: milan.der-
pich@usm.cl).
Color versions of one or more of the figures in this paper are available online
at http://ieeexplore.ieee.org.
Digital Object Identifier 10.1109/TSP.2015.2419183
the classical compressed sensing setting [2, Th. 2]), inference,
and decision-making problems [11], [12]. One important case
is defining such a compressibility notion for i.i.d. processes
where the probability measure is equipped with a density
function
1
[1], [2]. In this context, realizations of the process are
non-sparse (almost surely), and conventional ways of defining
compressibility for finite dimensional signals, based on the
power-law decay of the best -term approximation error (or
sequences that belong to the weak- ball), are not applicable
either, as shown in [1], [2].
Motivated by this problem, Amini et al. [1] and Gribonval et
al. [2] have introduced new definitions for compressible random
sequences. These notions are not based on the typical absolute
approximation error decay pattern of the signals, but on a rela-
tive -best -term approximation error behavior. In particular,
Amini et al. [1] formally define the concept of -compressible
process (details in Section II below). This new definition pro-
vides a meaningful way of categorizing i.i.d. random sequences
(and their distributions), in terms of the probability that almost
all the -relative energy of the process is concentrated in an
arbitrarily small sub-dimension of the coordinate domain, as
the block-length tends to infinity. Under this context, they pro-
vide two important results using the theory of order statistics
[1]. First of all, [1, Theorem 3] shows that a concrete family of
i.i.d. heavy-tail distributions is -compressible (including the
generalized Pareto, Students’s and log-logistic), while on the
other side, [1, Theorem 1] demonstrates that families with expo-
nentially decaying tails (such as Gaussian, Laplace, generalized
Gaussian) are not -compressible. Therefore, it is interesting to
ask about the compressibility of i.i.d processes not considered
in that analysis. In this direction, we highlight the work of Gri-
bonval et al. [2], which under an alternative notion of relative
-compressibility (involving almost sure convergences instead
of convergence in measure, which was the criterion adopted in
[1]) and a different analysis setting (fixed-rate instead of the
variable rate used in [1]), elaborates an exact dichotomy be-
tween compressible and non-compressible i.i.d. sequences. This
raises the question of whether it is possible to connect Amini
et al. [1] -compressibility with the more refined almost sure
(a.s.) convergence analysis of the best -term relative approx-
imation error in [2, Prop. 1], with the idea of completing the
analysis of [1, Ths. 1 and 3].
To address this question, we extend the analysis from i.i.d.
sequences to stationary and ergodic processes. In this broader
setting, the main result (Theorem 1) provides a necessary and
sufficient condition for a stationary and ergodic process to be
-compressible (in the sense of Amini et al. [1, Def. 6]), for
1
The probability is absolutely continuous with respect to the Lebesgue mea-
sure [13].
1053-587X © 2015 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.