PU Probability Prediction based Bayesian
Compressive Spectrum Sensing
Nuoya Zhang, Xuekang Sun, Caili Guo, Li Gao
Institute of Educational Technology
Beijing University of Posts and Telecommunications, Beijing, 100876
Email: noahicy@bupt.edu.cn
Abstract—Bayesian Compressive Sensing (BCS) can effectively
relax the requirement of hardware operational bandwidth and
perfectly recover sparse wideband signal at sub-Nyquist rate in
wideband spectrum sensing. However, one of the problem of
BCS is the long recovery time caused by the high computational
complexity. In this paper, a PU Probability Prediction based
Bayesian Compressive Sensing algorithm (PBCS) is proposed, in
which PU (Primary User) probability prediction results are utilized
as an index to select basis functions in the fast RVM (Relevance
Vector Machine) algorithm to decrease the iterative times and
thus reduce the recovery time. In addition, the simulation results
are illustrated that it needs less measurements and has enhanced
robustness against noise.
Index Terms—Bayesian Compressive Sensing, sparse Bayesian
learning, fast RVM, PU probability prediction, spectrum sensing
I. INTRODUCTION
The traditional static spectrum allocation strategy is becoming
inadequate to meet the constantly increasing demands of the
wireless spectrum, as there is no longer any bandwidth left to
allocate. Cognitive Radio (CR) is an efficient communications
paradigm to solve this problem [1]. In Cognitive Radio, spec-
trum sensing is an important process for secondary users to
opportunistically access to a licensed band when the primary
user (PU) is absent. But due to the limitation of hardware op-
erational bandwidth, the sampling becomes a more challenging
problem in wideband spectrum sensing.
Compressive Sensing (CS) is adopted as an important method
for wideband spectrum sensing [2],[3]. Different from Nyquist
sampling in narrow-band signal sensing, compressive sensing
can recover wideband sparse signals at sub-Nyquist rate which
can reduce the sampling rate greatly by implementing the prior
knowledge of signal sparsity. Due to the low percentage of spec-
trum occupancy by PU, wideband signal is sparse in frequency
domain, which allows us to use CS. The three core theories
of CS are: (1) Signal sparse representation, (2) Measurement
matrix compressive sampling, (3) Signal recovery. In this paper,
we concentrate on the signal recovery theory.
Existing methods of signal recovery for compressive sensing
can generally be classified into the following categories [4],
i.e. the greedy pursuit approach, the convex relaxation-type
approach and the nonconvex optimization method.
Bayesian Compressive Sensing (BCS) technique that has
received increasing attention is proposed to recover sparse
This work is supported by Chinese National Nature Science Foundation
(61372116)
signals from the point of Bayesian view [5],[6]. Sparse Bayesian
learning (SBL) and relevance vector machines (RVM) have been
widely used to achieve compressive representations based on
the prior information. Fast Marginal Likelihood Maximisation
(FMLM) proposed by M.E.Tipping [7],[8] extended to fast RVM
algorithm is an excellent algorithm in BCS, which provides
a tighter approximation to the l
0
-norm sparsity measure than
the l
1
-norm. Thus it can get a better performance than tradi-
tional recovery algorithms. However, maximizing the marginal
likelihood function, the key process in fast RVM, is quite a
complex iterative process. And it’s a main shortcoming of the
BCS algorithms.
In this paper, we improve the iterative algorithm of marginal
likelihood maximization in fast RVM by incorporating PU
probability prediction to reduce the complexity of BCS signal
recovery in wideband spectrum sensing. Unlike [7], the PU
probability prediction results are used as an index to choose
basis vector at the initialization of the iterative algorithm, which
can reduce the process of choosing all the needed basis function
and the effect of noise. Thus, the total times of iterations can
decrease, meaning the reduction of complexity.
The rest of this paper is organized as follow. In Section II, we
will introduce the signal model of BCS. In Section III, we will
introduce our algorithm called PU probability prediction based
Bayesian compressive sensing (PBCS) and demonstrate how the
PBCS reduce complexity. Simulation results will be presented
in Section IV. Some conclusions are discussed in Section V.
II. SIGNAL MODEL AND PROBLEM STATEMENT
According to the CS theorem, the received signal sampled at
sub-Nyquist rate can be represented as
y = Φx + n (1)
where y ∈ R
M
is the measurement vector, R is the set of real
numbers, Φ is the (M × N) Gaussian random measurement
matrix (M < N ) with Φ = [ϕ
1
, ϕ
2
, . . . , ϕ
N
], wherein ϕ
i
∈
R
M
is the basis vector, x ∈ R
N
denotes the wideband spectrum
signal vector, in which the number of non-zero elements is k,
and n is zero mean Gaussian noise with variance σ
2
.
In this section, the signal model is introduced in three
part. Firstly, the theorem of reconstructing signal by Bayesian
Compressive Sensing is presented. Then, the second part details
the recovery procedure of fast RVM which we will improve. At
last, the PU probability prediction is introduced.