Super-Resolution of Complex Exponentials from

Modulations with Known Waveforms

Zhihui Zhu1 , Manuel Lopez-Santillana2 , Michael B. Wakin1

1
Department of Electrical Engineering, Colorado School of Mines, Golden, CO, 80401, USA
{zzhu, mwakin}@mines.edu
2
Department of Applied Mathematics and Statistics, Colorado School of Mines, Golden, CO, 80401, USA
[email protected]

Abstract—In this paper, we investigate parametric estima- be parametrized by the front and interior walls [1, 3]. The
tion of complex exponentials from modulations with known attenuator ξl (t) is determined by the material of the wall and
waveforms. This problem arises naturally in radar systems other factors and is usually modeled as unknown [1].
and wireless communications, especially in applications which
suffer from multipath effects. Viewing the problem as a group Spike detection for neural recording: Neuron spikes (action
sparse recovery, we recast it naturally into an atomic norm potentials) are often captured with a microelectrode tip which
minimization, which has an equivalent semidefinite program is surrounded by many neurons, and therefore, receives a mix-
(SDP) characterization and thus can be solved efficiently. We ture of neurons’ electrical activities [7]. The neural recording—
experimentally demonstrate the advantage of our approach when obtained through wireless neural recording systems—can be
compared with a super-resolution method that does not consider
modeled as a superposition of returns due to radio signal
multipath effects.
reflection from surrounding materials [8], as in (1).
Keywords—Atomic norm, super-resolution, group sparsity, mul-
tipath exploitation, semidefinite programming When there is only a single component with L = 1, (1)
reduces to
I. I NTRODUCTION x(t) = (h(t)ξ1 (t)) ? gl (t).

We consider an acquired signal consisting of a superposi- Without loss of generality, suppose ξ1 (t) = 1, g1 (t) = δ(t).
tion of the same point source convolved with different point Otherwise we can deconvolve x(t) with g1 (t) and ξ1 (t) can
spread functions. Mathematically, we consider the following be absorbed into h(t). Identifying the parameters in (1) from
parametric superpositon model for the acquired signal low-frequency measurements reduces to the super-resolution
problem or line spectrum estimation (if we exchange the time
L
X and frequency domains), where one can apply conventional
x(t) = (h(t)ξl (t)) ? gl (t) (1) approaches for parameter estimation such as matrix pencil [9]
l=1 and MUSIC [10]. Chandrasekaran et al. [11] propose to use
where ? denotesP the circular convolution operator, the signal of the atomic norm, induced by the convex hull of a set of
K
interest h(t) = k=1 σ ek δ(t − tk ) is a weighted superposition atoms, as the general convex penalty function for linear inverse
of spikes, ξl (t) is an unknown (nonlinear) function which acts problems. The atomic norm generates the l1 norm for sparse
as an amplifier or attenuator, and gl (t) is a known point spread recovery problems and the nuclear norm for low rank matrix
function. Our goal is to identify the spike locations {tk } and recovery. Importantly, it provides a powerful framework to
coefficients {e σk } from low-frequency measurements of x(t) handle a dictionary with an infinite number of atoms. For
(which are described formally in Section II). This model arises super-resolution or line spectrum estimation, the atomic norm
in applications such as radar imaging, DNA microarrays, and minimization approach exactly inverts the parameters when
spike detection. A particular area that motivates this work there is no noise [12–15] and a certain minimum separation
is multipath exploitation in radar systems [1] and wireless condition is met. Recently, Zhu et al. [16] investigated SAR
communication [2]. We list two stylized applications below. imaging via sparse atomic norm reconstruction. Atomic norm
minimization has also been utilized for modal analysis [17].
Multipath exploitation for urban radar imaging: The in- Chi [18] and Yang et al. [19] utilized atomic norm minimiza-
direct multipath reflections of electromagnetic waves off of tion to exactly recover parameters of complex exponentials
targets (in conjunction with the walls, floors, etc.) present a from their modulations with unknown waveforms via a lifting
challenge in synthetic aperture radar (SAR) imaging [1, 3–6]. trick. Li and Chi [20] applied atomic norm minimization to
They result in ghost artifacts that can clutter the reconstructed simultaneously identify multiple sets of spikes from a superpo-
image. Let σ ek and tk respectively represent the complex sition of their modulations with known point spread functions.
reflectivity and locations of the k-th target. In this case, l = 1 In [21–23], the authors solved blind deconvolution problems
corresponds to the direct path and ξ1 (t) = 1, g1 (t) = δ(t), with nuclear norm minimization by assuming the signals live
i.e, there is no attenuator and convolution. The point spread in known subspaces or are observed using random masks.
function gl (t) corresponding to the l-th (l ≥ 2) path is δ(t−τl ) Our problem differs from [21–23] in that the modulations are
where τl is the additive two-way travel time of the l-th path known but the target signal is parameterized (by t).
compared to the direct path. With prior information about the
room geometry, the point spread function gl (t) (or τl ) can The approach for estimating the spikes using sparse model-

567
 qP
ing in [3] (where the spikes correspond to the target locations) 2 L
where ck =: l |σlk | and αk ∈ C with elements αk [l] :=
is to first divide the range of t uniformly to reduce the σkl
ck for all k = 1, . . . , K and l = 1, . . . , L. By definition, we
continuous parameter space into a finite set of grid points; have kαk k2 = 1. So x can be viewed as a sparse combination
then construct a dictionary for each component based on these of elements from the atomic set
grid points; and finally recover the spikes via sparse recovery ( )
algorithms. This approach suffers from the basis mismatch X L

problem and has no theoretical guarantees. In this paper, A := a(t, α) = α[l]B l et , α ∈ CL , kαk2 = 1 ,
we apply atomic norm techniques to invert the parameters l=1
in (1) from low-frequency measurements. The atomic norm is which can be viewed as an infinite dictionary governed by the
utilized to promote group sparsity and has an equivalent SDP parameters t and α. Note that the diagonal matrices B l are
characterization. Thus the problem can be solved efficiently fixed and known. The atomic norm of x is then defined as
using an off-the-shelf solver [24]. The outline of this paper is ( )
as follows. In Section II, the main problem is illustrated. Our X X
approach is discussed in Section III. Section IV presents some kxkA = inf ck x = ck a(tk , αk ) ,
ck ≥0, (4)
simulations to support our proposed methods. kαk k2 =1 k k
tk ∈[0,1]

II. P ROBLEM S ETUP which can be viewed as a penalty for promoting group sparsity
Suppose tk ∈ [0, 1] and x(t) and gl (t) are supported on of x, i.e., representing x by picking as few items as possible
[0, 1] for all k = 1, 2, . . . , K and l = 1, 2, . . . , L. We rewrite from the group set
(1) as {{B l et , l = 1, . . . , L} , t ∈ [0, 1]} .
L L K
!
X X X
x(t) = (h(t)ξl (t)) ? gl (t) = σlk gl (t tk ) , B. Semidefinite Program Characterization
l=1 l=1 k=1
The following result indicates that the atomic norm kxkA
where σlk = σ ek ξl (tk ) and denotes the subtraction operator admits an equivalent SDP characterization.
on the unit circle [0, 1]. Taking the Fourier series of x(t), we
obtain the measurements Theorem 1. The atomic norm kxkA can be written equiva-
L K
! lently as
X X
−j2πtk n
x[n] = σlk g l [n]e (2)

1 1
l=1 k=1
kxkA = inf trace(Toep(u)) + c
u∈CN ,c 2N 2
for n = −2M, −2M + 1, . . . , 2M . Thus x only contains low- Toep(u)  0, (5)
frequency information about x since we only observe the 4M +  P H
 
1 lowest Fourier series coefficients. Denote N = 4M +1. Here l B l Toep(u)B l x
0 .
g l ∈ CN contains the lowest N Fourier series coefficients of xH c
gl (t) with elements
Z 1 Proof of Theorem 1: Denote P the value of the right hand
g l [n] = gl (t)e−j2πtn dt side as SDP(x).P Suppose x = Pk ck a(fk , αk ) with ck > 0.
0 Define u = k c k e tk
and c = k ck . We have Toep(u) =
H
P
for n = −2M, −2M + 1, . . . , 2M . k ck etk etk  0. Note that
H
 P 
We use Ω = {t1 , . . . , tK } to denote the unknown set of l B l Toep(u)B l x
=
spike locations. Our goal is to estimate the spike locations Ω xH t
and {σlk , l = 1, . . . , L, k = 1, . . . , K} from x.   H H 
e B

X  [B 1 et · · · B L et ]    tk . 1 
ck k k
..  αk   0

III. O UR A PPROACH αH k
 
k H
eH tk B L
A. Atoms and Atomic Norm
and
Let B l := diag(g l ) denote an N ×N diagonal matrix with 1 1 X
diagonal g l . Also let trace(Toep(u)) + t = ck .
2N 2
 −j2πt(−2M )  k
e
.. N Therefore, SDP(x) ≤ kxkA .
et :=   ∈ C , t ∈ [0, 1]
 
.
e−j2πt(2M ) On the other hand, suppose for some u and x,
H
 P 
denote a length-N vector of samples from a discrete-time com- l B l Toep(u)B l x
 0, Toep(u)  0.
plex exponential signal with digital frequency t. We rewrite the xH c
measurements x in (2) with matrix notation
! By the Vandermonde decomposition lemma [14], we have
XK X L K
X XL X
x= σlk B l etk = ck αk [l]B l etk , (3) Toep(u) = V DV H = dk etk eH
tk
k=1 l=1 k=1 l=1 k

568
with D = diag(dK ),Pdk > 0. It follows that x is in the range C. Recovery Guarantee
c
of B l V , i.e., x = l B l V wl . Let γ = P kw 2 . Now we
l k2
l We can certify the optimality of minimizing the atomic
have norm defined in (4) using the following proposition.
H
 P 
l B l Toep(u)B l x
= AΣAH  0, Proposition 2. Suppose x# = k ck a(tk , αk ) with ck > 0,
P
xH c k = 1, 2, . . . , K and {B l etk , l = 1, . . . L, k = 1, . . . , K} are
where linearly independent. If there is a vector p ∈ CN such that
  the corresponding vector-valued dual polynomial q(t)[l] =
[ B1V ··· BLV ]  0 H
A=  , eHt B l p satisfies
0 wH
1 ··· wH
L
  q(tk )[l] = αk [l], tk ∈ Ω,

D

(6)
  .. 
I  kq(t)k2 < 1, t ∈ / Ω,
Σ =  .  .
D then x# = k ck a(fk , αP
  P
k ) is the unique atomic decomposi-
I γI tion satisfying kx# kA = k ck .
This implies Σ  0. It follows from the Schur complement
lemma that The above optimality conditions are derived from the facts
−1
that the atomic norm minimization is convex with strong

D
.. duality holding and that both primal and dual optimal values
γI ≥  .
 
.  are attained. We omit the proof due to space limitations.
D We note that the construction of such a dual polynomial
heavily depends on g l . Inspired by [12, 14], where the dual
Thus, we have polynomial is constructed with the square of the Fejér kernel,
!
X X P |wl [k]|2 we can construct a dual polynomial q(t) that satisfies (6) as
c= kwl k22 γ≥ l
. long as g l is populated from certain distributions.
dk
l k
Theorem 2. [19] Suppose x# = k ck a(tk , αk ) with ck >
P
1
It follows from the fact trace(Toep(u)) = trace(D) that
N 0, k = 1, 2, . . . , K. Suppose
2
g m := [ g 1 [m] g 2 [m] · · ·
P
1 1 1X 1 X l |wl [k]| g L [m] ]
trace(Toep(u)) + c ≥ dk +
2N 2 2 2 dk for m = −2M, . . . , 2M are i.i.d. samples from a distribution
k k
F that satisfies the following two conditions
v ! !
X P |wl [k]|2
u
u X
l
≥t dk ≥ kxkA , 2
dk E f H f = IL , max |f [l]| ≤ µ(F), f ∼ F. (7)
k k 1≤l≤L

where the last line follows from the Cauchy-Schwartz inequal- Also assume ∆(Ω) := mink6=k0 |tk − tk0 |, the smallest wrap-
ity. around distance between any pair of the spikes, is greater
1
Rewrite the measurements x in (3) as than M and M ≥ 64. Additionally, assume that αk are
i.i.d. randomly generated from the uniform distribution on the
complex unit sphere CSL−1 . Then, there exists a numerical
L K
! L
X X X
x= Bl ck αk [l]etk = Blx
el constant C such that
l=1 k=1 l=1    
M KL ML
PK M ≥ Cµ(F)KL log log
with xe = k=1 ck αk [l]etk . Denote X = [e x1 · · · x
eL] = δ δ
PK l
f
T
k=1 ck et k
α k , where T represents the nonconjugate trans-
is sufficient to guarantee that x# = k ck a(fk ,P
P
pose operator. This is the lifting scheme utilized in [18, 19], αk ) is the
where a different atomic norm induced by a different atomic unique atomic decomposition satisfying kx# kA = k ck with
set is applied for the augmented matrix X. f Note that the probability at least 1 − δ.
atomic norm utilized in [18, 19] for X is equivalent to kxkA ,
f
the atomic norm utilized for x. Thus the SDP characteriza- The main idea is to construct a dual polynomial q(t) (which
tion for solving the corresponding atomic norm minimization is similar to what is used in [19]) that satisfies (6).
in [18, 19] is also equivalent to kxkA . Remark 1. The two conditions in (7) are referred as the
isotropy and incoherence properties of F [18, 19]. We note
Proposition 1. The SDP characterization in (5) has the that not all components need to be random. Without loss of
following equivalent form generality, we suppose f [1] = 1, f ∼ F. Such a distribution
F also satisfies (7) as long as the distribution of the other

1 1
kxkA = inf trace(Toep(u)) + trace(C) components satisfies certain conditions. In this case, g 1 [m] = 1
u∈CN 2N 2
C,X for all m which corresponds to the direct path in many

Toep(u) X
  applications that involve multipath. Theorem 2 guarantees that
x = X (X), 0 . the spikes can be recoverd from x (a superpostion of signals
XH C
P from all paths) as long as the modulations g l corresponding
Here X (X) = l B l xl with xl being the l-th column of X. to the l-th path for all l ≥ 2 jointly satisfy (7).

569
D. Localizing the spikes 1 1

The SDP formulation (5) can be used to recover the spikes. 0.8 0.8

Suppose u is an optimal solution to (5). Then the Vandermonde 0.6

dual polynomial
locations of spikes 0.6
dual polynomial
locations of spikes

q(t)||2

q (t)|
decomposition of Toep(u) characterizes the spikes.

|e
||b
0.4 0.4
The dual norm of kxkA can be defined as
0.2 0.2
s
X 2
∗ H
kpkA = sup hp, xiR = sup eH
t Bl p .
0
0 0.2 0.4
t
0.6 0.8 1
0
0 0.2 0.4
t
0.6 0.8 1
kxkA t∈[0,1] l
Figure 1. Illustration of dual polynomials. Left: kbq (t)k2 ; Right: |e
q (t)|. The
The dual problem of minimizing the atomic norm (4) can be dashed red lines represent the locations of spikes.
written as

maximize hx, piR , subject to kpk∗A ≤ 1 (8) IV. N UMERICAL S IMULATIONS

We present a synthetic experiment arising in [3] to support
which also has an equivalent SDP formulation. the proposed approach. Let N = 64. Without loss of generality,
The spike locations can alternatively be identified from p
b, we set the measurement index n ∈ {0, . . . , N − 1} instead
the optimal solution to (8). To be precise, consider the vector of n ∈ {−2M, . . . , 2M }. We generate K = 5 spikes with
valued dual polynomial spike locations uniformly at random satisfying the minimum
separation ∆(Ω) > 1/N and coefficients generated with
H dynamic range of 10 and uniform phase. We generate L = 2
b(t)[l] = eH
q f Bl p
b.
modulations with g 1 [n] = 1 and g 2 [n] = ej2πnτ , where
The set of frequencies can be obtained by finding the peaks τ = 0.013 ≈ N1 , i.e., the corresponding g1 (t) = δ(t) and
of g2 (t) = δ(t − τ ). We choose ξ1 (t) = 1 and ξ2 (tk ) with i.i.d.
kb b = {t : kb
q (t)k2 : Ω q (t)k2 = 1}. standard Gaussian entries for k = 1, . . . , K. Note that in this
case
K
X
E. Revisiting the case L = 1 (h(t)ξ1 (t)) ? g 1 (t) = σ1k δ(t − tk )
k=1
When L = 1, (1) reduces to x(t) P = h(t). Then the
measurements in (3) simplify to x = k σ1k etk . Recovering and
K
the spikes from x is the super-resolution [12] or line spectrum X
estimation problem [14]. One can define an atomic set A0 = (h(t)ξ2 (t)) ? g 2 (t) = σ2k δ(t − tk − τ ),
{et , t ∈ [0, 1]} and obtain an equivalent SDP characterization k=1

(which is similar to (5), see [12, 14]) for the corresponding which implies we can alternatively apply a super-resolution
atomic norm. The dual norm is given by algorithm [12] (also see Section III-E) to recover all the spikes
{tk , tk + τ }K
k=1 . Figure 1 illustrates the dual polynomials of
kpk∗A0 = sup eH
t p . kb
q (t)k2 and |eq (t)| (see Sections III-D and III-E, respectively).
t∈[0,1]
We observe that the spikes can be localized correctly from the
Let p
e denote the optimal solution to the dual problem peaks of the dual polynomial kb q (t)k2 , while |e
q (t)| provides
many spurious spikes.
maximize hx, piR , subject to kpk∗A0 ≤ 1.
ACKNOWLEDGMENT
The spikes can be localized by finding the peaks of qe(t) =
eH
t p
e [12, 14]. This work was supported by NSF grant CCF-1409261.

F. Atomic Norm Soft Thresholding (AST) R EFERENCES

[1] P. Setlur, T. Negishi, N. Devroye, and D. Erricolo, “Multipath
We conclude this section with a discussion about the exploitation in non-los urban synthetic aperture radar,” IEEE
presence of noise in the measurements: Jour. Selected Top. Sign. Proc., vol. 8, pp. 137–152, Feb 2014.
[2] W. U. Bajwa, J. Haupt, A. M. Sayeed, and R. Nowak, “Com-
y[n] = x[n] + w[n] pressed channel sensing: A new approach to estimating sparse
multipath channels,” Proc. IEEE, vol. 98, no. 6, pp. 1058–1076,
for n = −2M, −2M +1, . . . , 2M . Here w[n] is additive noise. 2010.
In this case, we can obtain an estimate x
b that solves the atomic [3] M. Leigsnering, F. Ahmad, M. Amin, and A. Zoubir, “Multipath
norm soft thresholding (AST): exploitation in through-the-wall radar imaging using sparse
reconstruction,” IEEE Trans. Aerosp. Electron. Syst., vol. 50,
1 no. 2, pp. 920–939, 2014.
x − yk2
minimize λkxkA + ke
e ∈C
x N 2 [4] Z. Zhu and M. B. Wakin, “Wall clutter mitigation and target de-
tection using Discrete Prolate Spheroidal Sequences,” in 3rd Int.
where λ is an appropriately chosen regularization parameter Workshop on Compressed Sensing Theory and its Applications
that depends on the noise level [13, 16]. to Radar, Sonar and Remote Sensing (CoSeRa), June 2015.

570
 [5] F. Ahmad, J. Qian, and M. Amin, “Wall clutter mitigation using
Discrete Prolate Spheroidal Sequences for sparse reconstruction
of indoor stationary scenes,” IEEE Trans. Geosci. Remote Sens.,
vol. 53, pp. 1549–1557, March 2015.
[6] Z. Zhu and M. B. Wakin, “On the dimensionality of wall
and target return subspaces in through-the-wall radar imaging,”
in 4th Int. Workshop on Compressed Sensing Theory and its
Applications to Radar, Sonar and Remote Sensing (CoSeRa),
September 2016.
[7] S. Shahid, J. Walker, and L. S. Smith, “A new spike detection
algorithm for extracellular neural recordings,” IEEE Trans.
Biomed. Eng., vol. 57, no. 4, pp. 853–866, 2010.
[8] I. Obeid, M. A. Nicolelis, and P. D. Wolf, “A multichannel
telemetry system for single unit neural recordings,” J. Neurosci.
Methods., vol. 133, no. 1?, pp. 33 – 38, 2004.
[9] Y. Hua and T. K. Sarkar, “Matrix pencil method for estimating
parameters of exponentially damped/undamped sinusoids in
noise,” IEEE Trans. Acoust., Speech, Signal Processing, vol. 38,
no. 5, pp. 814–824, 1990.
[10] R. Schmidt, “Multiple emitter location and signal parameter
estimation,” IEEE Trans. Antennas Propag., vol. 34, no. 3,
pp. 276–280, 1986.
[11] V. Chandrasekaran, B. Recht, P. A. Parrilo, and A. S. Willsky,
“The convex geometry of linear inverse problems,” Foundations
Comput. Math., vol. 12, no. 6, pp. 805–849, 2012.
[12] E. J. Candès and C. Fernandez-Granda, “Towards a mathe-
matical theory of super-resolution,” Comm. Pure Appl. Math.,
vol. 67, no. 6, pp. 906–956, 2014.
[13] B. N. Bhaskar, G. Tang, and B. Recht, “Atomic norm denoising
with applications to line spectral estimation,” IEEE. Trans.
Signal Process., vol. 61, no. 23, pp. 5987–5999, 2013.
[14] G. Tang, B. N. Bhaskar, P. Shah, and B. Recht, “Compressed
sensing off the grid,” IEEE Trans. Inf. Theory, vol. 59, no. 11,
pp. 7465–7490, 2013.
[15] Q. Li and G. Tang, “Approximate support recovery of atomic
line spectral estimation: A tale of resolution and precision,”
arXiv preprint arXiv:1612.01459, 2016.
[16] Z. Zhu, G. Tang, P. Setlur, S. Gogineni, M. Wakin, and M. Ran-
gaswamy, “Super-resolution in SAR imaging: Analysis with the
atomic norm,” in Proc. IEEE Sensor Array and Multichannel
Signal Processing Workshop (SAM), July 2016.
[17] S. Li, D. Yang, G. Tang, and M. B. Wakin, “Atomic norm
minimization for modal analysis from random and compressed
samples,” arXiv preprint arXiv:1703.00938, 2017.
[18] Y. Chi, “Guaranteed blind sparse spikes deconvolution via lifting
and convex optimization,” IEEE J. Selected Top. Sign. Proc.,
vol. 10, no. 4, pp. 782–794, 2016.
[19] D. Yang, G. Tang, and M. B. Wakin, “Super-resolution of com-
plex exponentials from modulations with unknown waveforms,”
to appear in IEEE Trans. Inf. Theory, 2016.
[20] Y. Li and Y. Chi, “Stable separation and super-resolution of
mixture models,” arXiv preprint arXiv:1506.07347, 2015.
[21] A. Ahmed, B. Recht, and J. Romberg, “Blind deconvolution
using convex programming,” IEEE Trans. Inf. Theory, vol. 60,
no. 3, pp. 1711–1732, 2014.
[22] G. Tang and B. Recht, “Convex blind deconvolution with
random masks,” in Computational Optical Sensing and Imaging,
pp. CW4C–1, Optical Society of America, 2014.
[23] S. Bahmani and J. Romberg, “Lifting for blind deconvolution
in random mask imaging: Identifiability and convex relaxation,”
SIAM J. Imaging Sci., vol. 8, no. 4, pp. 2203–2238, 2015.
[24] K.-C. Toh, M. J. Todd, and R. H. Tütüncü, “SDPT3—a MAT-
LAB software package for semidefinite programming, version
1.3,” Optimiz. Methods Softw., vol. 11, no. 1-4, pp. 545–581,
1999.

571