energy across all signals. Then the projection coefficients are computed by the orthogonal
projection of Y onto Φ
~
F
k . The backward step shrinks
~
F
k
to F
k
by removing the β indices with
the smallest magnitude projection coefficients. Finally, the projection coefficients W in Φ
F
k
are computed via the orthogonal projection of Y onto Φ
F
k
, and the residue is updated as
res
k
¼ Y−Φ
F
k W ¼ Y−Φ
F
k Φ
†
F
k
Y. The pursuit algorithm iterates this procedure until all of
the coordinates in the correct support set are included in the estimated support set. Then we can
obtain an estimation
~
X of the sources X through solving a least-square problem.
The pseudo code of the DCSFBP algorithm is described as follows:
Algorithm: DCSFBP (Distributed compressed sensing forword-backword pursuit)
Algorithm DCSFBP (Distributed compressed sensing forword-backword pursuit)
Input: random sampling matrix
Φ
, measurement matrix
Y
Define:
max
K
Initialization:
0
F
(Estimated support set)
0
res Y
(Initial residue)
1k
(Iteration index)
Repeat
Forward step:
T1
:
arg max
k
S
row
SS
T Φres
(Preliminary forward test)
1kk
FF T
(Make candidate list)
†
2
argmin
kk
FF
W
WYΦWΦY
Backward step:
:
arg min
S
row
SS
T W
(Backward test)
kk
FFT
(Update the final list)
Projection:
†
2
argmin
kk
FF
W
WYΦWΦY
(Estimate the signals)
†
kkk
k
FFF
res Y Φ W Y Φ Φ Y
(Update the residue)
1kk
(Update the iteration index)
Termination rule:
max
2
or
kk
FKres
Output:
The estimate
K
sparse signals
X
that each column
1, 2, ,
k
j
NFx0
and
†
k
k
jj
F
FxΦy
.
, , ,
Multimed Tools Appl (2017) 76:20587–20608 20591