Professional Documents
Culture Documents
Marco F. Duarte
Hamid Dadkhahi
(UMass Amherst)
Karsten Fyhn
(Aalborg University)
linear
measurements
frequency-sparse
signal
Fourier components
=
frequency-sparse
signal
linear
measurements
nonzero
DFT coefficients
Frequency-Sparse Signals
and the DFT Basis
1 h j2f /N j22f /N
e(f ) = p
e
e
. . . ej2(N
N
1)f /N
Frequency-Sparse Signals
and the DFT Basis
1
Integral frequencies
Arbitrary frequencies
0.8
0.6
0.4
0.2
0
0
10
20
30
40
Approximation sparsity K
50
Average SNR, dB
50
40
30
20
10
0
100
200
300
400
Number of measurements M
500
Average SNR, dB
50
40
30
20
10
0
100
200
300
400
Number of measurements M
500
, c = 10
, c = 10
, c = 10
Recovery
1
algorithms
operate similarly
0.5
to matched
filtering:
( )
Dirichlet Kernel
0
0
1
0.02
0.04
0.06
0.08
0.1
0.12
, c = 10
Sparse approximation
algorithms fail
Sparse Approximation of
Frequency-Sparse Signals
3
0
0
10
20
30
40
Approximation sparsity K
50
RN
RN
Structured Restricted
Isometry Property (SRIP)
Preserve the structure only between sparse signals that
follow the structure model
Random (iid Gaussian, Rademacher) matrix has the SRIP
with high probability if
RN
mK K-dim planes
[Blumensath, Davies; Lu, Do]
RM
x1
x2
RN
RN
if
( )
1
0.5
0
0
1
0.02
0.04
0.06
0.08
0.1
0.12
( )
1
0.5
0
0
0.02
0.04
0.06
0.08
0.1
0.12
( ) > 0.1
1
Preserve the structure
only between sparse signals that
follow the structured sparsity model
0.5
RN
Integer Program
Inputs:
Signal vector x
Target sparsity K
Redundancy factor c
Maximum coherence
Output:
Approximation vector
Compute coefficients:
( )
1
0.5
0
0
Solve support:
0.02
0.04
0.06
0.08
0.1
0.12
( ) > 0.1
Mask coefficients:
0.5[0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 1 1 1 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0]
0
0
0.02
0.04
0.06
0.08
0.1
0.12
Return
CS recovery
error
signal K-term
structured sparse approximation error
noise
RN
Inhibition Heuristic
Inputs:
Signal vector x
Target sparsity K
Redundancy factor c
Maximum coherence
Output:
Approximation vector
( )
1
0.5
0
0
0.02
0.04
0.06
0.08
0.1
0.12
Compute coefficients:
Initialize:
While is nonzero and
Find max abs entry
Copy entry
Inhibit coherent entries
( ) > 0.1
1
0.5
Return
of
frequencies
amplitudes
Widely-studied problem:
Line spectral estimation
x
K
MUSIC
Root MUSIC
ESPRIT
PHD
...
Sparse Approximation of
Frequency-Sparse Signals
3
0
0
10
20
30
40
Approximation sparsity K
50
Sparse Approximation of
Frequency-Sparse Signals
3
0
0
10
20
30
40
Approximation sparsity K
50
Average SNR, dB
50
40
30
20
10
0
100
200
300
400
Number of measurements M
500
Average SNR, dB
50
40
30
20
10
0
100
200
300
400
Number of measurements M
500
Structured sparsity
Multitaper
Root MUSIC
Standard Sparsity
400
300
200
100
0
100
200
300
400
Number of measurements M
How to change
signal model
to further
improve
performance?
500
d sinusoids frequency response (Dirichlet kernel) in the main lobe by a quadratic function. The Dirichlet kernels maxi75 radians/second. The maximum/peak of the quadratic least-squares fit to three samples of a redundant DFT of the signal
! = 1.9775, i.e., it is an accurate estimate of the sinusoids frequency with precision to four
und its peak is located at
1 h j2f /N j22f /N
e(f ) = p
e
e
. . . ej2(N
N
1)f /N
N
f
Parameter space
e(3) ...
e(2)
e(1)
e(0)
N
f
Parameter space
e(3) ...
e(2)
e(1)
e(0)
in manifold by interpolating
from closest samples with
polar coordinates:
N
f
frequency estimates
(parameter space)
2 B0 (S n ), i = i+1
1022
stisti-
10 11
10
nthesis, `1 -analysis,
Average
Average
costcost
in frequency
in frequency
estimation
estimation
Average
Averagecost
costininfrequency
frequencyestimation
estimation
-analysis
``11-synthesis
`1 -synthesis
SIHT
SIHT
SDP
SDP
BOMP
BOMP
CBP
CBP
BISP
BISP
102
102
101
101
100
100
10 1
10 1
10 2
0
10 2
0
nthesis,
gorithm`[15]
on the
1 -analysis,
gorithm
the
In[15]
theon
BISP
+encies.
1
4
10
+encies.
1
4
the10
BISP
ms withIncoherence
ms with coherence
5
10
15
20
a range of subsam5
10[dB]
15
20
SNR
7
a range
of the
subsamts
to verify
10level
7
SNR [dB]
20
40
60
80
100
10
tsestimation.
to verify the
level
We set
20Fig. 2. Frequency
40
60 performance
80
100case.
estimation
in
noisy
estimation.
We nuset
he
result of the
Number of measurements (M)
Fig. 2.
Frequency
estimation performance
noisy case.
Number
of measurements
(M)Liao in
he result
the nu1.
In theofnoiseless
BOMP
[Fannjiang
and
2012]
N = 100, K = 4,
+1
Noiseless
Noisy
1.
In
the
noiseless
lar interpolation alSDP: Atomic
Norm Minimization
+1
Fig.
1.
Frequency
estimation
performance
in noise-less
case.
c
=
5,
=
0.2
Hz
`1 -analysis
9.5245
8.8222
Noiseless
Noisy
interpolation
alolara given
estimation
[Tang,
Rhaskar,
Shah,
Recht
2012]
Fig. 1. Frequency estimation performance in noise-less case.
-synthesis 9.5245
2.9082
2.7340
`1 -analysis
8.8222
estimation
elo aofgiven
approximation
`SIHT
1 -synthesis
0.2628
2.9082
0.1499
2.7340
sis,
sis,
the
the
ISP
ISP
nce
nce
amamvel
vel
set
set
nu-
Average
Average
costcost
in frequency
in frequency
estimation
estimation
+1
+1
103
103
102
102
101
101
100
100
10 1
10 1
10 2
0
10 2
0
`1 -analysis
-analysis
``11-synthesis
`1 -synthesis
SIHT
SIHT
SDP
SDP
BOMP
BOMP
CBP
CBP
BISP
BISP
5
5
10
SNR10[dB]
N = 100, K = 4, M =SNR
50, [dB]
c = 5,
15
15
20
20
= 0.2 Hz
15
20
SNR10[dB]
SNR [dB] Estimation:
Compressive Line Spectral
Fig. 2. Frequency
estimation performance
in noisy case.
Computational
Expense
Fig. 2. Frequency estimation performance in noisy case.
Noiseless
`1 -analysis
9.5245
Time
(seconds) Noiseless
-synthesis 9.5245
2.9082
`1 -analysis
0.2628
`SIHT
2.9082
1 -synthesis
SDP
8.2355
SIHT
0.2628
BOMP
0.0141
SDP
8.2355
CBP
46.9645
BOMP
0.0141
BISP
5.4265
CBP
46.9645
BISP
5.4265
Noisy
8.8222
Noisy
2.7340
8.8222
0.1499
2.7340
9.9796
0.1499
0.0101
9.9796
40.3477
0.0101
1.4060
40.3477
1.4060
Conclusions
Spectral CS provides significant improvements on
frequency-sparse signal recovery
address coherent dictionaries via structured sparsity
simple-to-implement modifications to recovery algs
can leverage decades of work on spectral estimation
robust to model mismatch, presence of noise
mduarte@ecs.umass.edu