Dual-Blind Deconvolution For Overlaid Radar-Communications Systems

This article has been accepted for publication in IEEE Journal on Selected Areas in Information Theory.
This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/JSAIT.2023.3287823
Dual-Blind Deconvolution for Overlaid

Radar-Communications Systems
Edwin Vargas Student Member, IEEE, Kumar Vijay Mishra Senior Member, IEEE, Roman Jacome Student Member, IEEE,
Brian M. Sadler Life Fellow, IEEE, Henry Arguello Senior Member, IEEE
Abstract—The increasingly crowded spectrum has spurred the design the two received signals. In this paper, we focus on the coexistence
of joint radar-communications systems that share hardware resources problem.
and efficiently use the radio frequency spectrum. We study a general
In general, coexistence systems employ different radar and commu-
spectral coexistence scenario, wherein the channels and transmit signals
of both radar and communications systems are unknown at the receiver. nications waveforms and separate receivers, wherein the management
In this dual-blind deconvolution (DBD) problem, a common receiver of interference from different radio systems is key to retrieving
admits a multi-carrier wireless communications signal that is overlaid useful information [4]. When the received radar signal reflected off
with the radar signal reflected off multiple targets. The communications from the target is overlaid with communications messages occupying
and radar channels are represented by continuous-valued range-time and
Doppler velocities of multiple transmission paths and multiple targets. the same bandwidth, the knowledge of respective waveforms is
We exploit the sparsity of both channels to solve the highly ill-posed DBD useful in designing matched filters to extract the two signals [2].
problem by casting it into a sum of multivariate atomic norms (SoMAN) Usually, the radar signal is known and the goal of the radar receiver
minimization. We devise a semidefinite program to estimate the unknown is to extract unknown target parameters from the received signal.
target and communications parameters using the theories of positive-
hyperoctant trigonometric polynomials (PhTP). Our theoretical analyses
In a typical communications receiver, the channel is known (or
show that the minimum number of samples required for near-perfect estimated) and the unknown transmit message is of interest in the
recovery is dependent on the logarithm of the maximum of number of receive processing. However, these assumptions do not extend to
radar targets and communications paths rather than their sum. We show a general scenario. For example, in a passive [7] or multi-static
that our SoMAN method and PhTP formulations are also applicable to
[8] radar system, the receiver may not have complete information
more general scenarios such as unsynchronized transmission, the presence
of noise, and multiple emitters. Numerical experiments demonstrate great about the radar transmit waveform. Similarly, in communications
performance enhancements during parameter recovery under different over dynamic channels such as millimeter-wave [1] or Terahertz-
scenarios. band [9], the channel coherence times are very short. As a result, the
Index Terms—Atomic norm, dual-blind deconvolution, channel estima- channel state information changes rapidly and cannot be determined
tion, joint radar-communications, passive sensing. a priori. Moreover, different transmitters are employed in the spectral
coexistence scenario leading to a lack of synchronized arrivals of both
signals at a common receiver.
I. I NTRODUCTION In this paper, we consider the aforementioned general spectral
With the advent of new wireless communications systems and coexistence scenario, where both radar and communications channels
novel radar technologies, the electromagnetic spectrum has become and their respective transmit signals are unknown to the common
a contested resource [1]. This has led to the development of various receiver. We model the extraction of all four of these quantities as
system engineering and signal processing approaches toward opti- a dual-blind deconvolution (DBD) problem. This formulation is a
mal, secure, and dynamic sharing of the spectrum [2]. In general, variant of blind deconvolution (BD) — a longstanding problem that
spectrum-sharing technologies follow three major approaches: co- occurs in a variety of engineering and scientific applications, such
design [3], cooperation [4], and coexistence [5]. While co-design as astronomy, image deblurring, system identification, and optics —
requires developing new systems and waveforms to efficiently utilize where two unknown signals are estimated from the observation of
the spectrum [6], cooperation requires an exchange of additional their convolution [10–13]. This problem is ill-posed and, in general,
information between radar and communications to enhance their structural constraints are imposed on signals to derive recovery
respective performances [2]. In the coexistence approach, radar and algorithms with provable performance guarantees. The underlying
communications independently access the spectrum and spectrum- ideas in these techniques are based on compressed sensing and low-
sharing efforts are largely focused on receiver processing. Among all rank matrix recovery, wherein signals lie in the low-dimensional
three approaches, separating the radar and communications signals random subspace and/or in high signal-to-noise ratio (SNR) regime
at a coexistence receiver is highly challenging because the uncoor- [13–17].
dinated sharing of resources does not help in clearly distinguishing
A. Prior Art
E. V., R. J., and H. A. are with Universidad Industrial de Santander,
Bucaramanga, Santander 680002 Colombia, e-mail: {edwin.vargas4@correo., Unlike prior works on spectral coexistence (see, e.g., [1, 25] and
roman2162474@correo., henarfu@}uis.edu.co. references therein), we examine the overlaid radar-communications
K. V. M. and B. M. S. are with the United States CCDC Army signal as the ill-posed DBD problem. Previously, [26] studied dual
Research Laboratory, Adelphi, MD 20783 USA, e-mail: kvm@ieee.org, deconvolution problem which assumed that the radar transmit signal
brian.m.sadler6.civ@mail.mil.
This research was sponsored by the Army Research Office/Laboratory and communications channel were known. Our approach toward a
under Grant Number W911NF-21-1-0099, and the VIE project entitled “Dual more challenging DBD problem is inspired by some recent works
blind deconvolution for joint radar-communications processing”. K. V. M. [19, 20] that have analyzed the basic BD for off-the-grid sparse
acknowledges support from the National Academies of Sciences, Engineering, scenarios. The radar and communications channels are usually sparse
and Medicine via Army Research Laboratory Harry Diamond Distinguished
Fellowship.
[27], more so at higher frequency bands, and their parameters are
The conference precursor of this work was presented at the 2022 IEEE Inter- continuous-valued [28]. Hence, sparse reconstruction in off-the-grid
national Conference on Acoustics, Speech, and Signal Processing (ICASSP). or continuous parameter domain through techniques based on atomic
Authorized licensed use limited to: Universidad Industrial de Santander. Downloaded on June 26,2023 at 20:01:23 UTC from IEEE Xplore. Restrictions apply.
© 2023 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.See https://www.ieee.org/publications/rights/index.html for more information.
This article has been accepted for publication in IEEE Journal on Selected Areas in Information Theory. This is the author's version which has not been fully edited and
TABLE I
C OMPARISON WITH P RIOR A RT
Problem Measurements Unknown parameters Algorithm
L−1 T
Multi-dimensional SRa [18] [α]ℓ e−j2πfℓ n
P
[y]n
e = α, fℓ Multi-dimensional ANM
ℓ=0
1-D BDb [19] PL−1 gn , α, τ ANM

[y]n = ℓ=0 [α]ℓ e−j2πn[τ ]ℓ [g]n
Blind 1-D SRc[20] PL−1 gℓ , α, τ ANM

[y]n = ℓ=0 [α]ℓ e−j2πn[τ ]ℓ [gℓ ]n
SR radard [21] PL−1 rℓ , α ANM

y=G ℓ=0 [α]ℓ a(rℓ )
Blind 2-D SRe[22] PL−1 α, rℓ , gℓ ANM

[y]n
e = ℓ=0 [α]ℓ a(rℓ )H Gn
e gℓ
PI
BdMf [23] y= i=1 diag(Bi hi )Ai xi hℓ , xℓ Nuclear norm minimization
P
PI Li −1 −j2πn[τi ]ℓ
1-D SR-dMg[24] [y]n = i=1 [g]i,n · ℓ=0 [αi ]ℓ e αi , τi SoAN minimization
This paper (DBD)h PL−1 αr , u, τr , νr , SoMAN minimization

[y]n H −j2πnτe e−j2π(n[τr ]ℓ +p[νr ]ℓ ) +
e = ℓ=0 [αr ]ℓ bn ue αc , v, τc , νc
P Q−1 H ve−j2π(n[τc ]q +p[νc ]q )
q=0 [α c ] q dne
a fℓ is a d-dimensional vector of unknown off-the-grid poles, [α]ℓ is complex amplitude, ne is a linear indexing sequence of the multidimensional measurement,
and L is the number of frequencies.
b Vector g is the transmit signal, [τ ] is the unknown off-the-grid pole.
ℓ
c This problem assumes a total of L unknown signals of which the ℓ-th signal and the n-th sample is [g ] .
ℓ n
d G denotes the sensing Gabor matrix and a(r) are the atoms containing 2-D where r = [[τ ] , [ν] ], [τ ] denote time-delay and [ν] Doppler-velocity.
ℓ ℓ ℓ ℓ ℓ
e G is the sensing Gabor matrix.
n
f Matrices B , A model the unknown channels. The unknown signal vectors are h and x .
ℓ ℓ ℓ ℓ
g L is the number of off-the-grid parameters of i-th signal [g]
i i,n .
h See Section II for details. Briefly, b and d are columns of a representation basis; u and v are low-dimensional representations of the radar and
n n
e
communications signal, respectively. Note that, from (19), the index ñ in the communications signal component dn e depends on the pulse index p. The radar
signal component bn only depends on the discrete-time index n.
norm minimization (ANM) [29] is appropriate for our application. In aforementioned works deal with mixed radar-communications signal
BD, sparsity and subspace constraints over the signals have shown structures.
great ability in reducing the search space to obtain unique solutions
[30, 31].
Our work has close connections with a rich heritage of research
in several related signal processing problems; Table I compares A closely related problem is that of blind demixing (BdM) where
our DBD problem with some important prior works. Our approach the sampled signal is a mixture of several BD problems [23, 24],
is based on the semidefinite program (SDP) derived for ANM- all of which have a similar structure. Per Table I, for I = 1
based spectral super-resolution (SR) of a high-dimensional signal in and 2, the BdM reduces to single and double BDs (with uniform
[18]. This formulation has previously been extended to bi-variate structure in each convolution), respectively. However, our DBD
ANM for estimating delay and Doppler frequencies of targets from problem has a different structure in each convolution operation
received radar signal samples [21]. The ANM-based recovery was (e.g., the communications messages change with each transmission).
also employed in the 1-D BD formulation of [19]. In general, the Further, the recovery guarantees previously derived for BdM do not
absence of knowledge about the transmit signal requires one to define consider special signal structures of overlaid radar-communications.
a structural assumption about the signal, e.g., a low-dimensional The recovery algorithms in some BdM formulations [23] are based
subspace representation based on incoherence and isotropy proper- on nuclear norm minimization over a set of rank-one lifted matrices.
ties [32]. This structural assumption was also used in [33], which In [37], the BdM recovery problem had multiple non-convex rank-
considered blind SR with multiple modulating unknown waveforms. one constraints to deal with multiple low-latency communications
A 2-D ANM-based blind SR with unknown transmit signals was systems. These BdM studies do not employ an atomic decomposition
addressed in [34] and extended to 3-D in [35]. The ANM has also over the transfer function and are not well-suited for the estimation
been useful for line spectrum denoising [36]. However, none of the of off-the-grid parameters.
B. Our Contributions C. Organization and Notations
Preliminary results of this work appeared in our conference The rest of the paper is organized as follows. In the next section,
publication [38], where only radar and communications channel we present the signal model of overlaid radar and communications
parameters were estimated using SoMAN minimization. In this paper, receiver. Section III casts the DBD into SoMAN minimization and
we apply our methods also to estimate the communications messages presents the SDP of the dual problem. In Section IV, we derive the
and radar waveforms, provide detailed theoretical guarantees, include conditions for finding the dual certificate of support. We validate
additional numerical validations, address multiple emitter scenarios, our proposed approach through extensive numerical experiments in
and consider unsynchronized and noisy signals. Our main contribu- Section V and conclude in Section VI. In the supplementary material,
tions in this paper are: Appendix A generalizes our method to more complex scenarios.
1) Solving spectral coexistence as a structured continuous-valued Throughout this paper, we reserve boldface lowercase, boldface
DBD. Contrary to the above-mentioned BD, BdM, and SR problems, uppercase, and calligraphic letters for vectors, matrices, and index
our DBD is based on practical radar-communications scenarios that sets, respectively. The notation [x]i indicates the i-th entry of the
lend a special structure to the signal as follows. Both communications vector x and [X]i the i-th row of X. We denote the transpose,
and radar channels are sparse and their parameters are continuous- conjugate, and Hermitian by (·)T , (·)∗ , and (·)H , respectively. The
valued; the radar channel is a function of unknown target parameters identity matrix of size N × N is IN . || · ||p is the ℓp norm. The
such as range-time and Doppler velocities. A similar delay-Doppler notation Tr {·} is the trace of the matrix, | · | is the cardinality of a
channel is considered for communications paths. Whereas the radar set, supp()˙ is the support set of its argument, E [·] is the statistical
transmits a train of pulses, the communications transmitter emits expectation function, and P denotes the probability. The functions
a multi-carrier signal. Our DBD formulation incorporates all these max and min output the maximum and minimum values of their
c
restrictions/structures. arguments, respectively. The sign function is defined as sign(c) = |c| .
2) Novel ANM-based recovery. The continuous-valued nature of The function diag(·) outputs a diagonal matrix with the input vector
radar and communications channel parameters requires using ANM- along its main diagonal. The block diagonal matrix with diagonal
based approaches. In particular, we cast the recovery problem as the matrix elements X1 , . . . , XP is X = blockdiag[X1 , . . . , XP ].
minimization of the sum of multivariate atomic norms (SoMAN).
Previously, the sum of 1-D atomic norms (SoAN) was minimized
to solve a joint (non-blind) demixing (dM) and SR problem in [24]. II. S IGNAL M ODEL
However, this 1-D SoAN formulation is non-blind and hence does not
trivially extend to our multi-variable DBD. Further, the guarantees in
Consider a pulse Doppler radar that transmits a train of Pr
the 1-D SoAN are provided for the same structure of multiple signals
uniformly-spaced pulses s(t) with a pulse repetition interval (PRI)
while, in our formulation, the structures of radar and communications
Tr ; its reciprocal 1/Tr is the pulse repetition frequency (PRF). The
signals are not identical. Moreover, in the sequel, we also generalize
transmit signal of the radar is
SoMAN to include synchronization errors for scenarios in which
radar and communications systems have different clocks. r −1
PX
3) Strong theoretical guarantees. Our method is backed up by strong xr (t) = s(t − pTr ), 0 ≤ t ≤ Pr Tr . (1)
theoretical performance guarantees. We prove that the minimum p=0
number of samples required for perfect recovery scale logarithmically
with the maximum of the radar targets and communications paths. The entire duration of Pr pulses is known as the coherent processing
This is counter-intuitive because the recovery in radar and commu- interval (CPI). The pulse s(t) is a time-limited baseband func-
nications receivers generally depends on the number of unknown Rtion, whose continuous-time Fourier transform (CTFT) is S(f ) =
∞
scatterers and paths. −∞
s(t)e−j2πf t dt. It is common to assume that most of the radar
4) Generalization to practical scenarios. We consider several prac- signal’s energy lies within the frequencies ±B/2,
R B/2 where B denotes
tical issues to generalize our approach (see Supplementary Material). the effective signal bandwidth, i.e., s(t) ≈ −B/2 S(f )ej2πf t df .
We show that, in the presence of noise, our formulation is applicable The radar channel or target scene consists of L non-fluctuating
by adding a regularization term to the dual problem. We address the point-targets, according to the Swerling-I target model [40]. The
unsynchronized transmission of radar and communications signals by transmit signal is reflected back by the targets toward the radar
multiplying the atomic set by a lag-dependent diagonal matrix. We receiver. The unknown target parameter vectors are αr ∈ CL ,
also show that the n-tuple BD problem comprising multiple sources τ r ∈ RL and ν r ∈ RL , where the ℓ-th target is characterized by:
of radar and communications signals is addressed by minimizing the time delay [τ r ]ℓ , which is linearly proportional to the target’s range
addition of multiple SoMANs. Finally, we handle the practical case i.e. [τ r ]ℓ = 2dℓ /c where c is the speed of light and dℓ is the target
of an unequal duration of radar pulse and communications symbol range; Doppler frequency [ν r ]ℓ , proportional to the target’s radial
by formulating the SoMAN problem to compute multiple coefficient velocity i.e. [ν r ]ℓ = 4πfc vℓ /c where fc is the carrier frequency and
vectors for communications signals. Note that our recent follow-up vℓ is the radial velocity; and complex amplitude [αr ]ℓ that models the
work in [39] also considers the multi-antenna DBD problem. path loss and target reflectivity. The target locations are defined with
5) Extensive numerical experiments: We comprehensively validated respect to the polar coordinate system of the radar and their range
our SoMAN approach for several scenarios, including randomly and Doppler are assumed to lie in the unambiguous time-frequency
distributed targets, unequal numbers of radar targets and communica- region, i.e., the time delays (τ ) are no longer than the PRI, and
tions paths, and closely-spaced parameters. To validate the provided Doppler frequencies (ν) are up to the PRF. The radar channel is the
theoretical guarantees numerically, We provide statistical perfor- product of delta functions as
mance by varying system parameters and validate the theoretical
result predicted by Theorem 2. Finally, we refer the reader to the L−1
X
Supplementary Material that compiles several numerical results for hr (ν, τ ) = [αr ]ℓ δ(ν − [ν r ]ℓ )δ(τ − [τ r ]ℓ ). (2)
all the practical scenarios mentioned above. ℓ=0
A. Overlaid Radar-Communications Receiver

In a spectral coexistence scenario, both radar and communications
operate in the same spectrum (Fig. 1). At the receiver, the baseband
signal is a superposition of radar and communications signals as
y(t) = xr (t) ∗ hr (t) + xc (t) ∗ hc (t),
Z ∞
= hr (t, τ )xr (t − τ ) + hc (t, τ )xc (t − τ )dτ, (7)
−∞
where 0 ≤ t ≤ max(Pr Tr , Pc Tc ). Substituting the radar and

communications channel models of, respectively, (3) and (6) in (7)
yields
L−1 Q−1
X X
y(t) = [αr ]ℓ xr (t − [τ r ]ℓ )e−j2π[ν r ]ℓ t + [αc ]q xc (t − [τ c ]q )e−j2π[ν c ]q t .
ℓ=0 q=0
(8)
Substituting expressions of xr (t) and xc (t) into (8) leads to
r −1 L−1
PX X
y(t) = [αr ]ℓ s(t − pTr − [τ r ]ℓ )e−j2π[ν r ]ℓ t
p=0 ℓ=0
c −1 Q−1
PX X
+ [αc ]q xm (t − mTc − [τ c ]q )e−j2π[ν c ]q t (9)
m=0 q=0
r −1 L−1
PX X
Fig. 1. In a spectral coexistence scenario, a joint receiver admits the overlaid ≈ [αr ]ℓ s(t − pTr − [τ r ]ℓ )e−j2π[ν r ]ℓ pTr
radar and communications signals. p=0 ℓ=0
c −1 Q−1
PX X
+ [αc ]q xm (t − mTc − [τ c ]q )e−j2π[ν c ]q mTc , (10)
The delay-Doppler representation of the channel is obtained by m=0 q=0
taking the Fourier transform along the time-axis of the time-delay where the last approximation results from the assumptions [ν r ]ℓ Tr ≪
representation hr (t, τ ) as 1 and [ν c ]q Tc ≪ 1 [42, 43] such that the phase rotation between each
radar pulse or OFDM message in the communications is assumed
L−1
X constant during a CPI. Then, for every P < t < P + 1, the values
hr (t, τ ) = [αr ]ℓ δ(τ − [τ r ]ℓ )e−j2π[ν r ]ℓ t . (3)
[ν r ]ℓ tT and [ν c ]ℓ tT remain constant. The change occurs only during
ℓ=0
the next pulse. Thus, we replace tT with the index p.
For the sake of simplicity, assume that the number of pulses equals
For communications, the transmit signal xc (t) comprises Pc
the number of messages, i.e., Pr = Pc = P and the PRI Tr is the
messages modeled as orthogonal frequency-division multiplexing
same as the message duration Tc i.e. Tr = Tc = T (Fig. 2). In the
(OFDM) signals with K equi-bandwidth sub-carriers which together
sequel (Section A-D), we also show that our formulation is applicable
occupy a bandwidth of B with a symbol duration of Tc and have
to the case when Tr ̸= Tc . Following Fig. 2, the received signal in
a frequencyPseparation of ∆f [41]. Hence, the m-th message is
K−1 j2πk∆f t (10) becomes
xm (t) = k=0 [gm ]k e , where [gm ]k is m-th complex
P −1
symbol modulated onto the k-th sub-carrier frequency fk = k∆f . X
The communications transmitted signal is y(t) = yep (t), (11)
p=0
c −1
PX where
xc (t) = xm (t − mTc ), 0 ≤ t ≤ Pc Tc . (4) L−1
X
m=0 yep (t) = [αr ]ℓ s(t − pT − [τ r ]ℓ )e−j2π[ν r ]ℓ pT
ℓ=0
The communications channel comprises Q propagation paths Q−1
X
characterized by their attenuation coefficients, delay, and frequency + [αc ]q xp (t − pT − [τ c ]q )e−j2π[ν c ]q pT , 0 ≤ t ≤ P T.
shifts, respectively, encapsulated in the parameter vectors αc ∈ CQ , q=0
τ c ∈ RQ , and ν ∈ RQ . The delay-Doppler representation of the (12)
communications channel is Express the measurements in terms of shifted signals yp (t) =
Q−1 yep (t+pT ), where the signals yep (t+pT ) are time-aligned with y0 (t).
Then, the signal y0 (t) and the shifted signals contain the same set
X
hc (ν, τ ) = [αc ]q δ(ν − [ν c ]q )δ(τ − [τ c ]q ), (5)
q=0 of parameters.
The CTFT of yp (t) produces
from which we obtain the time-delay representation as Z pT +T L−1
X
Yp (f ) = e−j2πf t [αr ]ℓ s(t − [τ r ]ℓ )e−j2π[ν r ]ℓ pT
Q−1 pT
X ℓ=0
hc (t, τ ) = [αc ]q δ(τ − [τ c ]q )e−j2π[ν c ]q t . (6) Q−1
X
!
−j2π[ν c ]q pT
q=0 + [αc ]q xp (t − [τ c ]q )e dt, (13)
q=0
and [αc ]q are unit-norm after normalizing the signal by its magnitude,
i.e. |[αr ]ℓ | = |[αc ]q | = 1; and [s]n+N = S(fn ).
We collect all discrete Fourier transform samples in the measure-
ment vector
h iT
y = y0T , · · · , yP
T
−1 . (17)
Then, all samples of P pulses (and communications messages) are

L−1 Q−1
X X
[y]ne = [αr ]ℓ [s]n e−j2π(n[τr ]ℓ +p[νr ]ℓ ) + [αc ]q [g]ne e−j2π(n[τc ]q +p[νc ]q ) ,
ℓ=0 q=0
(18)
where
n
e = n + N + M p, (19)
is a linear indexing sequence of the multidimensional measurement

with n = −N, . . . , N and p = 0, . . . , P −1, and g = [g0T , ..., gP
T
−1 ].
Fig. 2. Sequence of transmission and reception of radar and commu- Note that [y]ne = [yp ]n+N and [g]ne = [gp ]n+N .
nications signals. In synchronized transmission, the radar transmits each
pulse simultaneously as the communications transmitter sends out messages. Remark 1. In general, a cyclic prefix (CP) of length Kcp is
The common receiver of both systems simultaneously admits radar signals added to each modulated OFDM symbol for mitigating inter-symbol
reflected off L targets and communications messages through Q paths. For interference (ISI). The OFDM symbol duration then becomes Tcp +Tc ,
the unsynchronized transmission, see Fig. 1 in the supplementary material.
where Tcp denote the CP duration. Usually, Tcp is chosen such that
the maximum delay [τc ]qmax < Tcp < Tc . While we did not consider
which is a sum of the Fourier transforms of the p-th received symbol CP in the communications model above, our approach could be
of the communications signal and the p-th received pulse of the easily modified to include CP by changing the expression xm (t) =
P K−1 j2πk∆f t PK−1 j2πk∆f t
radar system. In the integral above, computing the radar term and k=0 [gm ] k e to x m (t) = k=K−Kcp +1 [gm ]k e .
expanding xp (t) in the communications part yields The consequence of this change is that the communications messages
L−1
gm is of a larger dimension. This requires a different low-dimensional
Yp (f ) =
X
[αr ]ℓ S(f )e−j2πf [τ r ]ℓ e−j2π[ν r ]ℓ pT subspace representation described in the next subsection. This case
ℓ=0
also resembles the unequal PRI and symbol duration scenario ex-
Q−1 K−1 plained in Section A.D of the Supplementary Material.
X X
+ [αc ]q e−j2πf [τ c ]q e−j2π[ν c ]q pT [gp ]k ×
q=0 k=0
Z pT +T B. Representation in low-dimensional space
e−j2πf t ej2πk∆f t dt. (14)
pT Our goal is to estimate the sets of parameters
Bn [αr ]ℓ , [τr ]ℓ , [νr ]ℓ , [αc ]q , [τc ]q and [νc ]q , without knowing the
Sampling (14) uniformly at the frequencies fn = = n∆f ,M radar pulse signal samples s and the communications symbols
with n = −N, . . . , N , M = 2N +1, and assuming that M = K, i.e., −1
{gp }Pp=0 . This is an ill-posed problem because of a large number
sampling in the frequency domain at the OFDM separation frequency
of unknown variables and fewer measurements. We now represent
∆f [42] gives
the unknown transmit radar and communications signals s and gp
L−1
X −j2πn∆f [τ r ]ℓ −j2π[ν r ]ℓ pT
in a low-dimensional subspace spanned by the columns of known
Yp (fn ) = [αr ]ℓ S(fn )e e random matrices B ∈ CM ×J and Dp ∈ CM ×J , respectively.
ℓ=0
Therefore, s = Bu and gp = Dp vp , where u ∈ CJ and vp ∈ CJ
Q−1 K−1 Z pT +T
X X are the unknown low-dimensional coefficient vector of the radar
+ [αc ]q e−j2πn∆f [τ c ]q e−j2π[ν c ]q pT [gp ]k ej2π∆f (k−n)t dt.
pT
and communications signal respectively. This reduces the number of
q=0 k=0
(15) unknown variables because J ≪ M . The basis matrices are
B = [b−N , . . . , bN ]H , bn ∈ CJ ,
RT
Using the property 0 e−j2π∆f (k−n)t dt = 0 with k ̸= n and (20a)
H
concatenating the values Yp (fn ) in the vector yp , i.e., [yp ]n+N = Dp = [Dp ]−N , · · · , [Dp ]N , [Dp ]n ∈ CJ .

(20b)
Yp (fn ) for n = −N, · · · , N , the discrete values of the Fourier
transform (14) are Rewrite the set of symbols g compactly as
L−1
g = Dv, D = [d0 , · · · , dM P −1 ]H = blockdiag [D0 , · · · , DP −1 ] .
X
[yp ]n+N = [αr ]ℓ [s]n+N e−j2π(n[τr ]ℓ +p[νr ]ℓ )
ℓ=0
(21)
Q−1
X The low-dimensional representation of s and g is inspired by radar
+ [αc ]q [gp ]n+N e−j2π(n[τc ]q +p[νc ]q ) , (16)
q=0
and communications applications. For instance, this structure has
been validated previously for a sensing-based SR of complex ex-
[τ r ]ℓ
where [τr ]ℓ = ∈ [0, 1] and [τc ]ℓ = [τTc ]ℓ ∈ [0, 1] are the
T ponentials that are modulated with unknown waveforms [20]. In the
normalized delays; [νr ]ℓ = [ν∆f
r ]ℓ
∈ [0, 1] and [νc ]ℓ = [ν∆f
c ]ℓ
∈ [0, 1] case of wireless communications, this assumption has been verified
are the normalized Doppler frequencies; the absolute values of [αr ]ℓ for multipath channels with unknown modulation [16] and multi-user
communications [44]. Using the basis decompositions (20a)-(20b) III. PARAMETER R ECOVERY VIA S O MAN MINIMIZATION
and (21), the signal model in (18) becomes We model the radar and communications trails of the received
L−1
X X Q−1 signal in (31) as positive linear combinations of constituent atoms,
−j2π(n[τr ]ℓ +p[νr ]ℓ ) −j2π(n[τc ]q +p[νc ]q )
[y]ne = [αr ]ℓ bH
n ue + [αc ]q dH
e ve
n which .are lifted rank-one matrices ua(r)H and va(c)H belonging
ℓ=0 q=0 to the atomic sets
(22) n o
Ar = ua(r)H : r ∈ [0, 1)2 , ||u||2 = 1 ⊂ CJ×M P , (32)
Denote the channel vectors as
L−1 and
X
hr = [αr ]ℓ a(rℓ ), (23)
n o
Ac = va(c)H : c ∈ [0, 1)2 , ||v||2 = 1 ⊂ CP J×M P , (33)
ℓ=0
and respectively. These atoms are basic units for synthesizing our signal
Q−1
with sparse channels. This definition leads to the following formu-
lation of atomic norms ||Zr ||Ar and ||Zc ||Ac - a sparsity-enforcing
X
hc = [αc ]q a(cq ), (24)
q=0
analog of ℓ1 norm for, respectively, general atomic sets Ar and Ac :
( )
where rℓ = [[τr ]ℓ , [νr ]ℓ ]T and cq = [[τc ]q , [νc ]q ]T such that
X X H
||Zr ||Ar = inf |[αr ]ℓ | Zr = [αr ]ℓ ua(rℓ ) ,
[αr ]ℓ ∈C,rℓ ∈[0,1]2
ℓ ℓ
a(rℓ ) ||u||2 =1
T (34)
= a([[τr ]ℓ , [νr ]ℓ ] )
= ej2π([τr ]ℓ (−N )+[νr ]ℓ (0)) , . . . , ej2π([τr ]ℓ (N )+[νr ]ℓ (P −1)) ∈ CM P ,
and
( )
(25) X X H
||Zc ||Ac = inf |[αc ]q | Zc = [αc ]q va(cq ) .
and [αc ]q ∈C,cq ∈[0,1]2
q q
||v||2 =1
a(cq ) (35)
T
= a([[τc ]q , [νc ]q ] ) To estimate the unknown matrices Zr and Zc , we minimize the
j2π([τc ]q (−N )+[νc ]q ](0))
= e ,...,e j2π([τc ]q (N )+[νc ]q (P −1))
∈C MP
. SoMAN ||Zr ||Ar + ||Zc ||Ac among all possible matrices Zr and
(26) Zc leading to the same observed samples as y. The primal convex
optimization problem for our DBD then becomes
Then, the model in (22) reduces to
minimize ||Zr ||Ar + ||Zc ||Ac subject to y = ℵr (Zr ) + ℵc (Zc ).
Zr ,Zc
[y]ne = hH r en bH
n u + h H
c e n dH
n v, (27)
e e e
(36)
or equivalently A semidefinite characterization of minimization of SoMAN in (36)
does not directly result from the primal problem as is the case with
[y]ne = ⟨Zr , GH H
e ⟩ + ⟨Zc , An
n e ⟩,
the standard atomic norm [18, 28]. Since the primal problem has only
= Tr(Gne Zr ) + Tr(Ane Zc ), (28) equality constraints, Slater’s condition is satisfied, and strong duality
with the waveform-channel matrices holds. This implies, solving the dual problem also yields an exact
solution to the primal problem. Therefore, we now analyze the dual
Zr = uhH
r ∈ C
J×M P
, (29) problem of (36) and derive a new semidefinite program for SoMAN
Zc = vhH P J×M P minimization using theories of positive-hyperoctant trigonometric
c ∈ C , (30)
polynomials (PhTP).
M P ×J M P ×P J
Gne = ene bHn ∈ C and Ane = ene dH
e ∈ C
n , where ene
MP
is the v-th canonical vector of R .
A. Dual problem
Define the linear operators ℵr : CJ×M P → CM P and ℵc :
CP J×M P → CM P such that The dual problem of (36) is obtained via standard Lagrangian
analysis. The Lagrangian function of problem (36) is
[ℵr (Zr )]ne = Tr(Gne Zr ),
L(Zr , Zc , q) = ||Zr ||Ar + ||Zc ||Ac + ⟨q, y − ℵr (Zr ) − ℵc (Zc )⟩,
and (37)
[ℵc (Zc )]ne = Tr(Ane Zc ). where q is the dual variable. Define the dual function g(q) as
The measurement vector now becomes g(q) = inf L(Zr , Zc , q)

Zr ,Zc
y = ℵr (Zr ) + ℵc (Zc ), (31) =⟨q, y⟩ + inf (||Zr ||Ar + ||Zc ||Ac ) − ⟨q, ℵr (Zr )⟩ − ⟨q, ℵc (Zc )⟩)
Zr ,Zc
where the matrices Zr and Zc encapsulate all the unknown variables =⟨q, y⟩ − sup (⟨ℵ∗r (q), Zr ⟩ − ||Zr ||Ar ) − sup (⟨ℵ∗c (q), Zc ⟩ − ||Zc ||Ac ) ,
Zr Zc
to be estimated, i.e., channel vectors hr and hc along with the signal (38)
coefficients u and g. This representation of the unknown variables
in 1-rank matrix is similar to the lifting trick previously employed where ℵ∗r
: C MP
→ C J×M P
and : C ℵ∗c
→ C MP
are P J×M P
∗
in other ill-posed problems such as BD [16, 19] and phase retrieval the adjoint
PP −1 PN operators of ℵr and ℵc , respectively, such that ℵr (q) =
PP −1 PN
∗
[45]. Given that Zr = uhH H
r (Zc = vhc ), one could estimate the p=0 n=−N [q]n e Gn
H
e and ℵc (q) = p=0 n=−N [q]n
H
e Ane . The
vectors u and hr (v and hc ) up to a scaling factor as the left and supremum values in (38) correspond to the convex conjugate function
right singular vectors of Zr (Zc ). of the atomic norms ||Zr ||Ar and ||Zc ||Ac .
Using the expansion of the dual function g(q) in (38), the dual constraints of the dual optimization problem (39) to the following
problem of (36) is LMIs
" # " #
maximize⟨q, y⟩R K bH
Kr K bH
K c
q ⪰ 0, b ⪰ 0, K ⪰ 0,
K
br IJ Kc IJP
subject to ∥ℵ⋆r (q)∥⋆Ar ≤ 1
∥ℵ⋆c (q)∥⋆Ac ≤ 1, (39) where Kb r = PP −1 PN
p=0 n=−N [q]n
e Gn
H
e ∈ CM P ×J and K bc =
PP −1 PN H M P ×P J
where ∥ · ∥∗ represents the dual norm. This dual norm is defined as p=0 n=−N [q]ne Ane ∈ C are the coefficients of the
PhTPs. This leads to the SDP of (39) as
||Z||∗A = sup ⟨U, Z⟩, (40)
||U||A ≤1 maximize ⟨q, y⟩R
q,K
where the indicator function
( subject to K ⪰ 0
∗ 0 if ||Z||∗A ≤ 1, "
bH
#
h (Z) = K K r
∞ otherwise, ⪰0
Kbr IJ
is the conjugate function of the dual norm. "
bH
#
K K c
We identify that the constraints in (39) are positive-hyperoctant ⪰0
trigonometric polynomials (PhTP) in r ∈ [0, 1]2 and c ∈ [0, 1]2 . Kbc IJP
Using the definition of dual-norm (40), rewrite these constraints in Tr(Θn K) = δn . (45)
(39) as
This SDP is solved using an off-the-shelf solver such as CVX [47].
∥ℵ∗r (q)∥∗Ar = sup ⟨u, ℵ∗r (q)H a(r)⟩ = ℵ∗r (q)H a(r) ,
r∈[0,1]2 ,∥u∥=1 2 Remark 2. We highlight some key differences in our SDP with
respect to some related prior works that apply ANM to other (non-
∥ℵ∗c (q)∥∗Ac = sup ⟨v, ℵ∗c (q)H a(c)⟩ = ℵ∗c (q)H a(c) .
c∈[0,1]2 ,∥v∥=1 2 BD) applications. For instance, the SDP in [18], the coefficient
of the polynomial is just the dual variable. In the non-blind SR
In the expressions above, define the bi-variate (delay and Doppler) problem of [21], the coefficient of the polynomial in the LMI is
vector polynomials in r and c as defined as a product of the inverse Fourier matrix basis, the Gabor
P
X −1 N
X sensing matrix, and the dual variable. For blind 2-D SR of [34],
fr (r) = ℵ∗r (q)H a(r) = [q]ne GH J
e a(r) ∈ C ,
n (41) the coefficients are defined in terms of a low-dimensional subspace
p=0 n=−N representation. Furthermore, these ANM approaches have only one
P −1 N
X X LMI while SoMAN has as many LMIs as the number of atomic norms.
fc (c) = ℵ∗c (q)H a(c) = [q]ne AH
e a(c) ∈ C
n
JP
. (42) A multitude of LMIs occur in the SDP for ANM applied to super-
p=0 n=−N
resolution problem in [28] but it employed only one PhTP that was
Evidently, the inequality constraints in (39) are equivalent to the also univariate.
Euclidean norms of the polynomials fr (r) and fc (c) being upper
After the dual variable is obtained by solving the SDP in (45),
bounded by unity. It follows from the Bounded Real Lemma (BRL)
we build the PhTPs fr (r), fc (c). The set of delay and Doppler
[46] that the bound on the magnitude of a multivariate trigonometric
parameters for radar R b = {b rℓ }L−1
ℓ=0 and communications C =
b
polynomial is characterized by a linear matrix inequality (LMI). More Q−1
T {b
rℓ }ℓ=0 are localized in the respective delay-Doppler planes when,
precisely, given a bi-variate PhTP R(λ) = t kt e−j2πλ t , where
P
2 respectively, the PhTPs ∥fr (b
rℓ )∥ and ∥fc (b
rq )∥1 assume a maximum
λ ∈ [0, 1] , t = [t1 , t2 ], 0 ≤ t1 ≤ L1 , 0 ≤ t2 ≤ L2 , with integers
modulus of unity. This procedure yields channel information for both
L1 and L2 . The inequality |R(λ)| < 1 implies that there exists a
systems. Thereafter, to extract the radar and communications transmit
positive semidefinite Gram matrix K such that [46, p. 135]
waveform, we solve a system of linear equations in Section III-D.
K kH

δn = Tr (Θn K) , ⪰ 0, (43)
k IJ
B. Dual Certificate
where n = [n1 , n2 ], 0 < n1 < m1 , −m2 < n2 < m2 , k is the
The primal problem (36) has only equality constraints. Hence,
vector containing all the polynomial coefficients,
( Slater’s condition is satisfied and strong duality holds. This allows
1, n1 = n2 = 0, us to state the following Proposition 1, which presents the dual-
δn = (44) certificate of support for the optimizer of (36).
0, otherwise,
and Θn = Θn2 ⊗Θn1 , where Θn1 , Θn2 are two elementary Toeplitz Proposition 1. Consider the signal y in (31). The atomic sets
matrices with ones in the n1 -th and n2 -th diagonal respectively and Ar and Ac are as defined in (32) and (33), respectively. Denote
Q−1
zero elsewhere. R = {rℓ }L−1 ⋆ ⋆
ℓ=0 and C = {cq }q=0 . The pair {Zr , Zc } is the unique
In the case of uni-variate trigonometric polynomials where the size solution of (36) if there exist two bi-variate PhTPs defined as in (41)-
of the Gram matrix is fixed. However, for the multi-variate case, the (42), one each in {τr , νr } and {τc , νc }, with complex coefficients
size of K depends on the sum-of-squares relaxation degree vector q such that
m = [m1 , m2 ], where m1 ≥ L1 and m2 ≥ L2 . A higher relaxation
fr (rℓ ) = sign([αr ]ℓ )u, ∀rℓ ∈ R, (46a)
degree leads to a better approximation but it also leads to a higher
computational complexity. In practice, this degree may be chosen to fc (cq ) = sign([αc ]q )v, ∀cq ∈ C, (46b)
be the minimum possible value, i.e., m1 = L1 and m2 = L2 and still ∥fr (r)∥22 < 1, ∀r ∈ [0, 1]2 \ R, (46c)
yield the optimal result [18, 21]. Furthermore, it is obvious that when ∥fc (c)∥22 < 1, ∀c ∈ [0, 1]2 \ C. (46d)
such a positive semidefinite matrix K exists, |R(λ)| < 1. Following
the SDP characterization of PhTP shown heretofore, we convert the Proof: The variable q is dual feasible. This gives
⟨q, y⟩R = ⟨q, ℵr (Zr )⟩R + ⟨ℵc (Zc )⟩R D. Estimating the communications messages and radar waveform
= ⟨ℵ∗r (q), Zr ⟩R + ⟨ℵ∗c (q), Zc ⟩R Solving (36) yields matrices Zr and Zc , wherein the information
L−1
X Q−1
X about the channel vectors hr and hc is obtained by following the
= ⟨ℵ∗r (q), [αr ]ℓ ua(rℓ )H ⟩R + ⟨ℵ∗c (q), [αc ]q va(cq )H ⟩R procedure explained at the end of Section III-A. It follows from
ℓ=0 q=0 (29) and (30), that the only unknowns left to estimate are now the
L−1 Q−1 radar and communications waveform vectors u and v. To this end,
X X
= [αr ]∗ℓ ⟨ℵ∗r (q), ua(rℓ )H ⟩R + [αc ]∗q ⟨ℵ∗c (q), va(cq )H ⟩R define the matrices
ℓ=0 q=0
r0 )H A0 rL−1 )H A0
 
L−1 Q−1 a (b ... a (b
=
X
[αr ]∗ℓ ⟨fr (rℓ ), u⟩R +
X
[αc ]∗q ⟨fc (cq ), v⟩R Wr =  .. .. ..
 , (52)
 
. . .
ℓ=0 q=0
a (br0 )H AM P −1 . . . a (b
rL−1 )H AM P −1
L−1 Q−1
X X
= [αr ]∗ℓ sign([αr ]ℓ ) + [αc ]∗q sign([αc ]q ) where Wr ∈ CM P ×LJ and
ℓ=0 q=0
c0 )H G0 cQ−1 )H G0
 
L−1
X Q−1
X a (b ... a (b
= |[αr ]ℓ | + |[αc ]q | Wc =  .. .. ..
 , (53)
 
. . .
ℓ=0 q=0
c0 )H GM P −1
a (b ... a (b H
cQ−1 ) GM P −1
≥ ||Zr ||Ar + ||Zc ||Ac . (47)
where Wc ∈ CM P ×P QJ . Concatenate these matrices as W0 =
On the other hand, based on the Hölder inequality we have [Wr, Wc] ∈ CM P ×J(L+P Q) and the coefficient vectors as
⟨q, y⟩R = ⟨ℵ∗r (q), Zr ⟩R + ⟨ℵ∗c (q), Zc ⟩R z = [[αr ]0 uT , ·, [αr ]L−1 uT , [αc ]0 vT , · · · , [αc ]Q−1 vT ]T ,
≤ ||ℵ∗r (q)||∗Ar ||Zr ||Ar + ||ℵ∗c (q)||∗Ac ||Zc ||Ac
where z ∈ CJ(L+P Q) . We recover this vector z through a least-
≤ ||Zr ||Ar + ||Zc ||Ac , (48)
squares solution of the following problem
where the first inequality follows from the Cauchy-Schwartz in- minimize ∥y − W0 b
z∥2 . (54)
equality and the last follows from (46a)-(46d). From (47) and (48), z
b
⟨q, y⟩R = ||Zr ||Ar + ||Zc ||Ac so that the pair {Zr , Zc } is primal This only requires that the columns of the matrix [Wr Wc ] are
optimal and q is dual optimal based on strong duality. linearly independent because the matrices depend on the values
of steering vectors a(br0 ), . . . , a(b
r0 ), a(b
c0 ), . . . , a(b
c0 ). The radar
waveform s and the communications messages g are then estimated
C. Performance guarantees following the definitions in (20a)-(20b) and (21) as b s = Bb u and
Recall the following useful properties. g
b = Db v. Here, note that the communications messages change at
every pulse. Therefore, g is a concatenation of all messages over P
Definition 1 (Isotropy [32]). The distribution F is isotropic if given pulses.
a random vector x ∈ CN ∼ F satisfies the following condition:
E[xxH ] = IN . (49) E. Computational Complexity

The present work is the first investigation into the spectral coex-
which means that the columns have unit variance and are uncorre- istence DBD problem. Here, our focus is perfectly recovering the
lated. unknown continuous-valued channel and signal parameters. In this
Definition 2 (Incoherence [32]). The distribution F is incoherent context, ANM-based SDP excels and yields strong theoretical guar-
antees. The complexity of the proposed algorithm is O M 3.5 P 3.5 ,

if given a vector x = [x0 , . . . , xN −1 ] ∼ F and given coherence
parameter µ it satisfies that: which corresponds to the complexity of the SDP solver SDPT3
[48, 49]. With non-ANM non-SDP heuristics, the exact recovery
max |xi |2 ≤ µ. (50) of continuous-valued parameters severely degrades. However, other
0<i<N −1
approaches may be employed to reduce the computational complexity
Following these definitions, we now state our main result. of the SDP. These techniques include reweighting the continuous dic-
tionary [50], accelerating proximal gradient [51], and localizing the
Theorem 2. Assume the columns of B and D in, respectively, (20a) solution space using prior knowledge and applying block iterative ℓ1
and (21) are drawn from a distribution that satisfies the isotropy and minimization (BL1M) [52]. A promising alternative is also mentioned
incoherence properties. Further, if ∥v∥22 = ∥u∥22 = 1 and |αi | = 1, in our recent follow-up work [53], where we have formulated a simple
then there exists a numerical constant C such that delay-only DBD problem and solved it via a low-rank Hankel matrix

MP J
recovery approach. However, this method has not been evaluated for
M P ≥ Cµ max(L, Q)J log2 our general DBD problem.
δ

M P QJ M P LJ
× max log , log , (51) IV. P ROOF OF T HEOREM 2
δ δ
To prove Theorem 2, we first construct the radar and commu-
is sufficient to guarantee that, with probability at least 1 − δ, the
nications dual polynomials and then demonstrate that they satisfy
pair {Zr , Zc } is perfectly recovered from M P measurements of the
the required conditions (46a)-(46d) mentioned in the Proposition 1.
observation vector y in (31).
The existence of dual polynomials guarantees the optimal solution
Proof: See Section IV. {Z⋆r , Z⋆c } of the primal problem (36).
A. Construction of the dual polynomials and the vector t = [tTr , tTc ]T ∈ C(3LJ+3QP J)×1 , where
′ ′ ′ ′
∂m ∂n
Using the notation f (m ,n ) (r) = ∂τ ∂ν f (r), the following
m′ n′ tr = [sign([αr ]0 )uT , · · · , sign([αr ]L−1 )uT , 0TJ×1 , · · · , 0TJ×1 ]T ,
conditions ensure that the dual polynomials fr (r) and fc (c) achieve
the maxima at true parameter values tc = [sign([αc ]0 )vT , · · · , sign([αc ]Q−1 )vT , 0TP J×1 , · · · , 0TP J×1 ]T .
(58)
fr (rℓ ) = sign([αr ]ℓ )u, ∀rℓ ∈ R, (55a)
− fr(1,0) (rℓ ) = 0J×1 , ∀rℓ ∈ R, (55b) Then, the optimization problem in (56) becomes
− fr(0,1) (rℓ ) = 0J×1 , ∀rℓ ∈ R, (55c) minimize ∥Wq0 ∥22

q0
fc (cq ) = sign([αc ]q )v, ∀cq ∈ C, (55d)
subject to Fq0 = t. (59)
− fc(1,0) (cq ) = 0P J×1 , ∀cq ∈ C, (55e)
−1
− fc(0,1) (cq ) = 0P J×1 , ∀cq ∈ C. (55f) The least-squares solution of (59) is q0 = WH W FH λ with
In order to build the polynomials fr (r) and fc (c) that satisfy the λ = [β T , γ T , ζ T , η T , θ T , ξT ]T , (60)
conditions in (55), we now introduce random kernels that are obtained
after approximating the dual variable q in (41) and (42). Unlike where β = [β0T , · · · , βL−1 T
]T ∈ CJL×1 , γ = [γ0T , · · · , γL−1
T
]T ∈
JL×1 T T T JL×1
previous related works [19, 21, 29, 54], an additional complication C , ζ = [ζ0 , · · · , ζL−1 ] ∈ C T T
, η = [η0 , · · · , ηQ−1 ]T ∈
arises in the DBD problem because the dual variable q must be the CQP J×1 , θ = [θ0T , · · · , θQ−1 T
]T ∈ CQP J×1 , and ξ =
same for both radar and communications. Our strategy is to make [ξ0T , · · · , ξQ−1
T
]T ∈ CQP J×1 .
a coarse approximation of the polynomials via an initial estimate Using the expressions of F and λ, we rewrite the product b =
q0 ̸= q of the dual variable. To this end, we solve the following FH λ in the expression of q0 as
weighted constrained minimization problem with weights ω0 , · · · ,  a(r )H e
ℓ
H
−N P b−N
  j2π(−N )a(r )b eH 
ℓ −N −N P
L−1
ωM P −1 (to be determined later) X
.. ..
b=   βℓ +   γℓ
. .
minimize ∥Wq0 ∥22 ℓ=0 a(rℓ )H eN P bH
N j2π(N )a(rℓ )bN eH
NP
q0  j2π(−P )a(r H 
ℓ )b−N e−N P
subject to fr (rℓ ) = sign([αr ]ℓ )u ∀rℓ ∈ R ..
+  ζℓ
− fr(1,0) (rℓ ) = 0J×1 ∀rℓ ∈ R .
j2π(P )a(rℓ )bN eH
NP
− fr(0,1) (rℓ ) = 0J×1 ∀rℓ ∈ R H
 
a(cq ) e−N P dH ( q ) −N P eH
 j2π(−N )a c d 
Q−1 −N P −N P
.. ..
X
fc (cq ) = sign([αc ]q )v ∀cq ∈ C +  ηq + 
  θq
 . .
− fc(1,0) (cq ) = 0P J×1 ∀cq ∈ C q=0 H H
a(cq ) eN P dN P j2π(N )a ( q) NP NP
c d e H
fc(0,1) (cq ) ( q ) −N P eH
 j2π(−P )a c d 
− = 0P J×1 ∀cq ∈ C, −N P
+ .
..  ξq .
(56)
j2π(P )a(cq )dN P eH NP
where (61)
W = diag([ω0 , · · · , ωM P −1 ]T ) (57)
Substituting q0 for the approximation of q in (41) yields
is a diagonal matrix. Note that the first and fourth constraints are
the same as (46a) and (46b), while the remaining constraints are fr (r)
conditions necessary to satisfy (46c) and (46d). Using the expressions L−1
X
"
X 1
!
in (20a), (25), and (26), define the matrix F ∈ C(3LJ+3QP J)×M P = bn eH H
e a(r) a(rℓ )en
n
H
e bn βℓ
ωn2e
as ℓ=0 n
e
!
 H H
 X 1
b−N e−N P a(r0 ) ··· bN eN P a(r0 )
+ j2π(n)bn eH H
e a(r) a (rℓ ) en
n
H
e bn γℓ
 .. .. ..  ωn2e

 . . . 
 n
e
b−N eH bN eH
! #
−N P a(rL−1 ) ... N P a(rL−1 ) X 1
 
 −j2π(−N )b−N eH H + j2π(p)bn eH H
e a(r) a (rℓ ) en
H
e bn ζℓ
 
−N P a(r0 ) ... −j2π(N )bN eN P a(r0 ) n
ωn2e

 
 .. .. ..  n
e
 . . .  Q−1
"
X 1
!
 −j2π(−N )b−N eH H

−N P a(rL−1 ) −j2π(N )bN eN P a(rL−1 ) 
 X
... H H H
+ bn ene a(c) a(cq )ene dne ηq
 −j2π(−P )b−N eH −j2π(P )bN eH ωn2e
 
−N P a(r0 ) ... N P a(r0 )  q=0 n
e
.. ..
 
 ..  !
 −j2π(−P )b .eH a r
 . .  X 1
−N −N P ( L−1 ) ... H
−j2π(P )bN eN P a(rL−1 ) 
 + j2π(n)bn eH
n
H
e a(r) a (rℓ ) en
H
e dn θℓ
ωn2e
e
F=

d−N P eH ··· dN P eH n

−N P a(c0 ) a(c 0 ) e
 NP  ! #
 .. .. ..  X 1
j2π(p)bn eH H H
 
. . . + e a(r) a (rℓ ) en
n e dn ξq . (62)
ωn2e
  e
d−N P eH
−N P a(cQ−1 ) ··· dN P eH a (c )
 
 NP Q−1  n
e
 −j2π(−N )d H H
−N P e−N P a(c0 ) ... −j2π(N )dN P eN P a(c0 ) 

Using

 .. .. .. 
.
 
 . . 
 −j2π(−N )d−N P eH a(cQ−1 )
 −N P ... −j2π(N )dN P eH N P ( Q−1 ) 
a c  bn eH
e a(r) = e
n
j2π(nτ +pν)
bn , (63)
 −j2π(−P )d−N P eH a(c0 ) ... −j2π(P )dN P eH N P a(c0 ) H −j2π(nτ +pν)
ene bH bH

−N P
  a(r) n =e n, (64)
 .. .. .. 

. . .

−j2π(−P )d−N P eH −N P a(cQ−1 ) ... H
−j2π(P )dN P eN P a(cQ−1 ) the expression in (62) becomes
10
fr (r) of the random kernel N to interpolate the sign pattern in the dual
L−1 certificate is zero.
XX 1 h
= 2
ej2πn(τ −τℓ ) ej2πp(ν−νℓ ) bn bH
n βℓ
ℓ=0 n
e
ωn
e The dual communications polynomial fc is obtained mutatis mu-
j2πn(τ −τℓ ) j2πp(ν−νℓ ) tandis as
+j2π(n)e e bn bH
n γℓ
i L−1
j2πn(τ −τℓ ) j2πp(ν−νℓ )
bn bH
X
+j2π(p)e e n ζℓ fc (c) = N(c − rℓ )H βℓ + N(1,0) (c − rℓ )H γℓ + N(0,1) (c − rℓ )H ζℓ
Q−1 ℓ=0
XX 1 h j2πn(τ −τq ) j2πp(ν−νq ) Q−1
+ 2
e e bn dH
e ηq
n X
q=0 n
e
ωne + P(0,0) (c − cq )ηq + P(1,0) (c − cq )θq + P(0,1) (c − cq )ξq ,
q=0
+j2π(n)ej2πn(τ −τq ) ej2πp(ν−νq ) bn dH
e θℓ
n (72)
i
+j2π(p)ej2πn(τ −τq ) ej2πp(ν−νq ) bn dH
e ξq .
n (65) where
The expression above involves the random matrices 1 X
P (r) = gN (n)gP (p)ej2πnτ ej2πpν dne dH
e .
n (73)
X 1 j2πnτ j2πpν MP
n
M (r) = e e bn bH
n ∈ C
J×J
, (66)
e
ωn2e Note that the kernel N(r) is common between the expressions of
n
e
and radar and communications polynomials.

X 1 j2πnτ j2πpν Next, we select the vectors β, γ, ζ, η, θ, and ξ to satisfy
N (r) = e e bn dH
e ∈ C
n
J×P J
. (67) the conditions (46a) and (46b) for the radar and communications
ωn2e
n
e polynomials, respectively.
′ ′ ′
,n′ ) ∂m ∂n
Using the notation X(m (r) = ′
∂τ m ∂ν m
′ X(r), we rewrite the
polynomial in (65) as B. To satisfy conditions (46a) and (46b)
L−1 Take the derivatives of kernels
X
fr (r) = M(r − rℓ )βℓ + M(1,0) (r − rℓ )γℓ + M(0,1) (r − rℓ )ζ,ℓ ′ ′ X 1 ′ ′
M(m ,n ) (r) = (j2πn)n (j2πp)m ej2πnτ ej2πnν bn bHn
ℓ=0 ωn2e
n
e
Q−1 X 1
X ′ ′ ′ ′
+ N(r − cq )ηq + N(1,0) (r − cq )θq + N(0,1) (r − cq )ξq . N(m ,n ) (r) = 2
(j2πn)n (j2πp)m ej2πnτ ej2πnν bn dHn
ωne
e
q=0 n
e
(68) ′ ′ X 1 ′ ′
P(m ,n ) (c) = (j2πn)n (j2πp)m ej2πnτ ej2πnν dne dH
e .
n
Recall from (19) that ñ = N + n + M p, n = −N, · · · , N , p = ωn2e
n
e
0, · · · , P − 1, and M = 2N + 1. We choose the weights in (57) as Then, rewrite the derivatives of radar and communications polyno-
mials as
s s
N P
ωñ = , (69) L−1
gN (n) gP (p) ′
,n′ )
X ′
,n′ ) ′
+1,n′ )
fr(m (r) = M(m (r − rℓ )βℓ + M(m (r − rℓ )γℓ
such that M and N are generated from the 2-D squared Fejér kernel ℓ=0
′
,n′ +1)
+ M(m (r − rℓ )ζℓ
Q−1
φ(r) = φN (τ )φP (ν), (70) X ′
,n′ ) ′
+1,n′ )
+ N(m (r − cq )ηq + N(m (r − cq )θq
where q=0
′
,n′ +1)
4
+ N(m

sin(T πτ ) N (r − cq )ξq ,
φN (τ ) = , T = + 1,
T sin(πτ ) 2 (74)
N
X and,
φN (τ ) = gN (n)ej2πτ n , (71)
L−1
n=−N ′
,n′ ) ′
,n′ ) ′
+1,n′ )
X
fc(m (c) = T(m (c − rℓ )βℓ + T(m (c − rℓ )γℓ
with ℓ=0
′
min{n+N,N } ,n′ +1)
+ T(m

1 X |k| |n − k| (c − rℓ )ζℓ
gN (n) = 1− 1− .
N M M Q−1
′
k=max{n−N,−N } ,n′ ) ′
+1,n′ )
X
+ P(m (c − cq )ηq + P(m (c − cq )θq
The Fejér kernel has the maximum modulus of unity and then q=0
it rapidly decays to zero. It has been successfully applied in the ′
,n′ +1)
+ P(m (c − cq )ξq .
construction of polynomials that guarantee optimality for SR and
(75)
BD problems [21, 29, 54].
Next, define the vectors
Remark 3. Unlike prior ANM-based BD approaches [19] that con-
struct the polynomial with one random kernel, the radar polynomial jr = [β T , κγ T , κζ T ]T , jc = [η T , κθ T , κξT ]T , (76)
(68) in our DBD problem is constructed using two distinct random
and matrices H1 ∈ C3LJ×3LJ , H2 ∈ C3LJ×3P QJ , and H3 ∈
kernels and their partial derivatives; as mentioned next, the same is
C3P QJ×3P QJ such that
true for the communications polynomial. Note that the selection of the  
(0,0) 1 (1,0) 1 (0,1)
random kernels ensures that the dual radar polynomial in (68) has Ei κ i
E κ i
E
the same form as (41). However, the basis vectors bn and dne are Hi =  − κ1 E(1,0) (2,0)
− κ12 Ei
(1,1) 
− κ12 Ei  , i = 1, 2, 3

i
statistically independent. Therefore, the contribution in expectation 1 (0,1) 1 (1,1) 1 (0,2)
− κ Ei − κ2 E i − κ2 Ei
11
p ′ ′
with κ = φ′′ (0). The matrix E1 (m ,n ) ∈ CLJ×LJ comprises a The expectation of the matrix H is
2
total of L block matrices, each of size J × J, as " #
H1 H2
. H = E [H] = (81) H
′ ′ ′
,n′ ) H2 H3
M(m ,n ) (r1 − r1 ) M(m
 
(r1 − rL )
...
′ ′ .. ..
E1 (m ,n ) =  ..
 
 . The matrix H in (81) is a blockdiagonal matrix with blocks H1 and
 . . . 
′ ′ ′ ′ H3 . Thus, H is invertible if and only if H1 and H3 are invertible.
M(m ,n ) (rL − r1 ) . . . M(m ,n ) (rL − rL )
Note that H1 and H3 are built upon the 2-D Fejér kernel. This allows
′ ′ ′ ′
Similarly, the matrices E2 (m ,n ) ∈ CLJ×QP J and E3 (m ,n ) ∈ us to apply the result in [54, Lemma C.2] that directly implies that
CQP J×QP J consist of LQ block matrices of size J × P J and Q2 H1 and H3 are invertible. The following Lemma 3 shows that H is
block matrices of size P J × P J respectively very close to E [H] and hence invertible with high probability.
′ ′ ′ ′  Lemma 3. Consider the event
N(m ,n ) (r1 − c1 ) N(m ,n ) (r1 − cQ )

...
′ ′ .. ..
E2 (m ,n ) =  ..
 
 . . .
,
 Eϵ1 = {∥H − E [H] ∥ ≤ ϵ1 },
′ ′
m′ ,n′ )
N(m ,n ) (rL − c1 ) ... N( (rL − cQ ) for every real ϵ1 ∈ (0, 0.8). Then the event Eϵ1 occurs with
probability 1 − 4δ for every δ > 0 provided that
′
,n′ ) ′ ′
P(m P(m ,n ) (c1 − cQ )
 
(c1 − c1 ) ... 90µJ

6LJ 6P QJ

′ ′ .. .. MP > max (L, Q) log max , .
E 3 (m ,n )= .. 

 . . .
.
 ϵ21 δ δ
′ ′ ′ ′
P(m ,n ) (cQ − c1 ) ... P(m ,n ) (cQ − cQ )
Consequently, we rewrite the conditions in (55) as the following Proof: See Appendix B.
system of linear equations From Lemma 3, and the following inequalities [21]

H1 H2 jr tr ∥I − H1 ∥ ≤ 0.1908,
H = , (77)
H2 H3 jc tc
| {z } and
=H∈C3(QP J+LJ)×3(QP J+LJ)
∥I − H3 ∥ ≤ 0.1908,
where tr and tc have been defined in (58). Clearly, if the matrix H
is invertible, the coefficients jr (jc ) for the radar (communications) it follows that H is invertible. Thus, one could obtain βℓ , γℓ , ζℓ , ηq ,
polynomials could be obtained. In order to establish the invertibility θq , and ξq , ℓ = 0, · · · , L − 1, q = 0, · · · , Q − 1 from the solution
of H, we show that E [H] is invertible and then prove that H is close of (77). By the construction above, the conditions (46a) and (46b)
to E [H]. are ipso facto satisfied. Next, we prove that the polynomials fr and
Using the following notation of the vectors fc satisfy the conditions (46c) and (46d).
  e−j2π(n[τr ]0 +p[νr ]0 )
 
1
υne =  −j2πn ⊗ .. 3L
C. To satisfy conditions (46c) and (46d)
∈C ,

κ
−j2πp
 .
−j2π(n[τr ]L−1 +p[νr ]L−1 )
κ e To prove (46c) and (46d) for, respectively, fr (r) and fc (c), we
utilize the strategy initially proposed in [29]. Note again that, in
and
our case, we build each polynomial, i.e. fr (r) and fc (c), with
e−j2π(n[τc ]0 +p[νc ]0 )
 

1
 two different random kernels. Moreover, one of the kernels N(r)
−j2πn ⊗ .. 3Q is common between the two polynomials. This makes the proof
ϱne =  ∈C ,

κ  .
−j2πp in this subsection different from [29]. We show that the random
κ e−j2π(n[τc ]L−1+p[νc ]L−1 ) dual polynomials fr (r) and fc (c) concentrate around, respectively,
the constituent matrices of H become J and P J-dimensional form of the scalar-valued dual polynomials
1 X constructed in [54] on a discrete set Ωgrid ∈ [0, 1)2 . Thereafter, we
H1 = gN (n)gP (p) (υne ⊗ bn ) (υne ⊗ ϱne )H consider an extension to the continuous set [0, 1)2 . Here, we show
MP
n
e this explicitly for fr (r). The proof for fc (c) is similar.
1 X
= gN (n)gP (p) υne υnH ⊗ bn bH n , (78) Consider the polynomial
MP
e
n  ′ ′

M(m ,n ) (r − r0 )H
e
1 X  .. 
gN (n)gP (p) υne ϱH ⊗ bn dH
 
H2 = n n , (79)  . 
MP ′ ′
e e
 M(m ,n ) (r − rL−1 )H
 
n
e 
′ ′
 
 1 M(m +1,n ) (r − r )H 
 κ 0 
1 X
(m′ ,n′ ) 1 .  ∈ C3JL×J
gN (n)gP (p) ϱne ϱH ⊗ dne dH
 
H3 = n n , (80) Ψr (r) = m′ +n′ 
 .
.
MP
e e
κ  1 (m′ +1,n′ )

n H 

 M (r − rL−1 ) 
e
 κ ′ ′
The expected values of these matrices follow from  1 M(m ,n +1) (r − r )H 
 κ 0 
the Pisotropy property in (49) as H1 = E [H1 ] =  .. 
1 H .
 
e gN (n)gP (p) υn e ⊗ IJ ,
e υn H3 = E [H3 ] =
M P Pn
 
1 H 1 (m′ ,n′ +1) H
MP e gN (n)gP (p) ϱn
n e ϱn
e ⊗ IP J and H2 = E [H2 ] = κ
M (r − r L−1 )
03QP J×3LJ . (82)
12
with statistical expectation and

′
,n′ )
E[Ψ(m
 
r (r)] (83) η
m′ n′ jc =  κθ  = L
b1u
e+L
b2v
e,
1 X −j2πn −j2πp
= gM (n)gP (p) κξ
MP κ κ
n
e
× e−j2π(nτ +pν) υne ⊗ IJ e = sign(αr ) ⊗ u and v

where u e = sign(αc ) ⊗ v.

φm
′ ∗ n′ ∗  Using the expressions above, the radar polynomial becomes
M (τ −[τr ]0 ) φP (ν−[νr ]0 )
..

.
 1 ′ ′
fr(m ,n ) (r)
 
′ ′
φm
∗
φn
∗
κm′ +n′
 
 (τ −[τr ]L−1 )
M (ν−[νr ]Q−1 ) 
P
 ′
1 φm +1 (τ −[τ ] )∗ φn′ (ν−[ν] )∗
 h ′ ′
i j
(m′ ,n′ ) r
0 0
= Ψ(m ,n )
(r)H (r)H
 P

1
 κ M  r Ψc jc
=
 ..  ⊗IJ

κm′ +n′ . ′
,n′ ) ′ ′

′ ∗ n′ = Ψ(m r)H L1 u
e + Ψ(m ,n )
L v
 
∗ 
1 φm +1 τ −[τ ]
( L−1 ) φP (ν−[ν]Q−1 ) 
r r
{z 2 }
 e
 κ M | {z } |
′
 1 φm′ (τ −[τ ] )∗ φn +1 (ν−[ν] )∗  (m′ ,n′ ) (m′ ,n′ )
 κ M 0 P 0  =Θr (r) =Θ
er (r)
 ..  ′ ′ ′ ′

. ′
 + Ψ(m
c
,n )
(r)H L e + Ψ(m
b1u c
,n )
(r)H L
b2v
e. (86)
1 φm′ τ −[τ ] ∗ n +1 ∗
M ( L−1 ) φP (ν−[ν]Q−1 )
| {z } | {z }
(m′ ,n′ ) ′ ,n′ )
| κ {z } =Θc (r) e (m
=Θ c (r)
(m′ ,n′ )
ψr (r)
1 ′ ′ (m′ ,n′ )
≜ ψr(m ,n ) (r) ⊗ IJ . Expanding the matrix Θr (r), we obtain
κm′ +n′
′ ′
The analogous polynomial for communications is Θ(m
r
,n )
(r)
′ ′
h ′ ′
i h ′ ′
iH
′ ′
= Ψ(m ,n )
(r) − E Ψ(m ,n )
(r) E Ψ(m ,n )
 
N(m ,n ) (r − c0 )H r r r (r)

 .. 
 × L1 − L′1 ⊗ IJ + L′1 ⊗ IJ u

.
  e
′ ′ iH
(m′ ,n′ )
h
 N(m ,n ) (r − cQ−1 )H 
  ′
= E Ψr (r) L1 ⊗ IJ u e
 1 Nm′ +1,n′ (r − c )H 
 
 κ 0 ′ ′
h ′ ′
iH
+ Ψ(m ,n )
(r) − E Ψ(m ,n )

′ ′ 1 .. (r) L1 u
Ψc(m ,n ) (r) = m′ +n′ 
  r r e

κ  .  | {z }
 1 m′ +1,n′ H 
 (m′ ,n′ )
 N
κ
(r − cQ−1 ) =J1 (r)
′ ′
 
 1 Nm ,n +1 (r − c )H  h ′ ′
iH
0
Ψ(m ,n )
L1 − L′1 ⊗ IJ u

 κ  + E r (r) e, (87)
 .. 
.
  | {z }
(m′ ,n′ )
 
′ ′ H =J2 (r)
1 m ,n +1
κ
N (r − cQ−1 )
m′ n′
1 X −j2πn −j2πp where the first term evaluates to
= gM (n)gP (p)
MP κ κ h ′ ′
iH
n
E Ψ(m ,n )
L′1 ⊗ IJ u
e
r (r)
× e−j2π(nτ +pν) ϱne ⊗ dne bH 3QP J×J e
n ∈ C . ′ ′
H
(84) = ψr(m ,n ) (r) ⊗ IJ L′1 ⊗ IJ (sign(αr ) ⊗ u)

Using E[dne bn ] = 0JP ×J , we have ′ ′

H
= ψr(m ,n ) (r) ⊗ IJ L′1 sign(αr ) ⊗ u

′
,n′ ) 1 ′ ′
E[Ψc(m ψr(m ,n ) ⊗ 0P J×J = 03JP L×J ′ ′ 1 ′ ′

(r)] = (85) = ψr(m ,n ) (r)H L′1 sign(αr ) u ≜ (m′ +n′ ) f (m ,n ) (r)u,
κm′ +n′ κ
In order to simplify the bound computation we follow the pro- T
where L′1 sign(αr ) ≜ ∈ R3L and the scalar-valued

cedure in [29] and we express H−1 in terms of the matrices β κγ κζ
L1 ∈ C3LJ×LJ , R1 ∈ C3LJ×2JL L b 1 ∈ C3QP J×LJ , Rb1 ∈ polynomial
3LJ×2LJ 3LJ×P JQ
C , L2 ∈ C , R2 ∈ C3LJ×2P JQ , L b2 ∈
L−1
C3QP J×P JQ , Rb 2 ∈ C3QP J×2P JQ , L b 2 ∈ C3QP J×P QJ , R
b2 ∈ ′
,n′ )
X ′
,n′ ) ′
+1,n′ )
3LJ×2P QJ f (m = φ(m (r − rℓ )β ℓ + φ(m (r − rℓ )γ ℓ
C as follows
ℓ=0
′
,n′ +1)

L 1 R1 L 2 R2 + φ(m (r − rℓ )ζ ℓ , (88)
H−1 ≜ b b2 .
L1 R b1 L b2 R
where φ is the Fejér kernel defined in (70).
and
Similarly,
L′1 ⊗ IJ R′1 ⊗ IJ

−1 03LJ×P JQ 03LJ×2P JQ
H = b ′2 ⊗ IP J b ′2 ⊗ IP J . ′
,n′ ) ′ ′ ′ ′

03QP J×LJ 03QP J×2LJ L R e (m
Θ (r) = Ψ(m ,n )
(r)H − EΨ(m ,n )
(r)H L2 v
r r r e
Then, the solutions of (77) are
| {z }
(m′ ,n′ )
J3 (r)
 
β h ′ ′
iH
jr =  κγ  = L1 u
e + L2 v
e, +E Ψ(m
r
,n )
(r) (L2 ) v
e (r),
κζ
| {z }
(m′ ,n′ )
J4 (r)
13
Substituting various expressions above, (86) becomes Lemma 6. Assume that v ∈ CJ is sampled on the complex unit
sphere. Let 0 < δ < 1 the failure probability and ϵ4 > 0. There
1 (m′ ,n′ ) 1 ′
,n′ )
′ +n′ ) fr (r) = f (m (r)u exists a numerical constant C such that if
κ (m κm′ +n′
′ ′
(m ,n ) (m′ ,n′ )

+ J1 (r) + J2 (r) 1 |Ωgrid |JL |Ωgrid |J
M P ≥ Cµ max(L, Q)J max 2 log log2 ,
(m′ ,n′ ) (m′ ,n′ ) ϵ4 δ δ
+ J3 (r) + J4 (r)
′ ′ ′
,n′ )
JL P QJ
+ Θc(m ,n ) (r) e (m
+Θ (r). (89) log , log ,
c δ δ
′ ′
1 (m ,n ) 1 (m′ ,n′ ) the following inequality is satisfied
To demonstrate that f
κ r
(r) is close to
′
κm′ +n
′ f (r)u,
′
(m ,n ) (m′ ,n′ )
we prove that the spectral norms of J1 (r), J2 (r), ( )
(m′ ,n′ ) (m′ ,n′ ) (m′ ,n′ ) ′ ′ (m′ ,n′ ) ′ ′
J3 (r), J4 (r), Θc (r), and Θe c(m ,n ) (r) are small P sup J3 (e
r) ≥ ϵ4 , m , n = 0, 1, 2, 3 ≤ 48δ.
rd ∈Ωgrid 2
with high probability on the set of grid points Ωgrid . Compared e
with the single-BD approaches such as [19, 20] where bounding

(m′ ,n′ ) (m′ ,n′ ) Proof: The proof follows mutatis mutandis from the proof of
the dual polynomial requires bounding J1 (r) and J2 (r), (m′ ,n′ )
Lemma 4 using the fact that J1 is the sum of zero-mean random
the DBD problem requires the same on the additional terms
′ ′
(m ,n ) ′ ′
(m ,n ) ′ ′
(m ,n ) ′ ′ matrices.
J3 (r), J4 (r), Θc (r), and Θ e (m
c
,n )
(r). Then, we
1 (m′ ,n′ ) Lemma 7. Assume that v ∈ CJ is sampled on the complex unit
show, with high probability, that κm′ +n′ fr (r) is close to
1 (m′ ,n′ ) sphere. Let 0 < δ < 1 be the failure probability and ϵ5 > 0 a
′ ′ f (r)u on Ωgrid . Finally, we show that the polynomial
κm +n constant. There exists a numerical constant C such that if
constructed as above satisfies the expected condition ∥fr (r)∥2 < 1,
r\R. With a proper grid size, we extend this result for the continuous
|Ωgrid |J

JL

P QJ

domain [0, 1]2 . M P ≥Cµ max(L, Q)J max log2 , log , log ,
(m′ ,n′ )
δ δ δ
1) Bounds for ∥Ji (r)∥2 , where i = 1, 2, 3, 4,
(m′ ,n′ ) e (m
′ ′
,n ) the following inequality is satisfied
∥Θc (r)∥2 , and ∥Θ c (r)∥ 2 : To bound these matrices,
we use the fact that they can be decomposed as a sum of random ( )
zero-mean matrices and, consequently, matrix Bernstein inequality (m′ ,n′ ) ′ ′
P sup J4 (e
r) ≥ ϵ5 , m , n = 0, 1, 2, 3 ≤ 32δ.
is applicable. See similar argument in [20, 29]. The following rd ∈Ωgrid
e 2
(m′ ,n′ ) (m′ ,n′ )
Lemmata 4 and 5 show that ∥J1 (r)∥2 and ∥J2 (r)∥2 are
bounded on Ωgrid . Proof: The proof follows mutatis mutandis from the proof of
(m′ ,n′ )
Lemma 5 using the fact that J2 is the sum of zero-mean random
Lemma 4. Assume that u ∈ CJ is sampled on the complex unit matrices.
sphere. Let 0 < δ < 1 the failure probability, and ϵ2 > 0. There (m′ ,n′ )
Finally, Lemmata 8 and 9 show that ∥Θc (r)∥2 , and
′ ′
exists a numerical constant C such that if e (m
∥Θ c
,n )
(r)∥2 are bounded on Ωgrid .

1 |Ωgrid |JL |Ωgrid |J PJ
M P ≥ Cµ max(L, Q)J max 2 log log2 , Lemma 8. Assume that v ∈ C is sampled on the complex unit
ϵ2 δ δ sphere. Let 0 < δ < 1 be the failure probability and ϵ6 > 0 a

LJ P QJ constant. There exists a numerical constant C such that if
log , log ,
δ δ
1 |Ωgrid |P QJ |Ωgrid |J
M P ≥Cµ max(L, Q)J max 2 log log2 ,
the following inequality is satisfied ϵ6 δ δ

( ) LJ P QJ
(m′ ,n′ ) ′ ′
log , log ,
P sup J1 r) ≥ ϵ2 , m , n = 0, 1, 2, 3 ≤ 48δ.
(e δ δ
rd ∈Ωgrid
e 2
then, the following inequality is satisfied
Proof: See Appendix C. ( )
′ ′
′ ′
J
Lemma 5. Assume that u ∈ C is sampled on the complex unit P sup Θ(m
c
,n )
(e
r) ≥ ϵ6 , m , n = 0, 1, 2, 3 ≤ 48δ.
rd ∈Ωgrid 2
sphere. Let 0 < δ < 1 the failure probability, and ϵ3 > 0. There
e
exists a numerical constant C such that if Proof: See Appendix E.

CµJ max(L, Q) 2 |Ωgrid |J Lemma 9. Assume that v ∈ CP J is sampled on the complex unit
MP ≥ log
ϵ2 δ sphere. Let 0 < δ < 1 be the failure probability and ϵ6 > 0 a
3
JL P QJ constant. There exists a numerical constant C such that if
× max log , log ,
δ δ
2 |Ωgrid |J LJ P QJ
the following inequality is satisfied M P ≥Cµ max(L, Q)J max log , log , log ,
δ δ δ
( )
(m′ ,n′ ) ′ ′ the following inequality is satisfied
P sup J2 (e
r) ≥ ϵ3 , m , n = 0, 1, 2, 3 ≤ 32δ.
rd ∈Ωgrid
e 2 ( )
′ ′
e (m ,n ) ′ ′
P sup Θ r (e
r ) ≥ ϵ 7 , m , n = 0, 1, 2, 3 ≤ 32δ.
Proof: See Appendix D. rd ∈Ωgrid
e 2
Similarly, the following Lemmata 6 and 7 show that

(m′ ,n′ ) (m′ ,n′ ) Proof: The proof is obtained, mutatis mutandis, from that of
∥J3 (r)∥2 , and ∥J4 (r)∥2 are bounded on Ωgrid .
Lemma 8.
14
(m′ ,n′ ) ′ ′
2) κ1 fr is close to κm′1+n′ f (m ,n ) u on Ωgrid : We obtained problem, the dual polynomials are also multi-dimensional bi-variate
this as a consequence of the following Lemma 10. but contain cross-terms (see (89)). Following the procedure in [34,
Lemma 14] but instead using Lemma 11 for our specific polynomial
Lemma 10. Assume that Ωgrid ∈ [0, 1)2 is a finite set of points. Let
with ϵ = 0.0005 concludes the proof.
0 < δ < 1 be the failure probability and ϵ > 0 a constant. Then,
there exists a numerical constant C such that if

1

|Ωgrid |P QJ
D. Proof of the theorem
M P ≥ Cµ max(L, Q)J max log ,
ϵ2 δ Using the aforementioned results, a dual polynomial can be con-

1 |Ωgrid |LJ 2 |Ωgrid |J structed to satisfy (46a), (46b), (46c), and (46d) as long as the
log , log ,
ϵ2 δ δ condition on M P as in (92) holds. In particular, Lemma 3 proves

LJ P QJ (46a) and (46b) by showing that the linear system of equations in
log , log ,
δ δ (77) is invertible with high probability. Conditioned on this event,
then Lemmata 4-9 lead to Lemma 10 showing a measure concentration
′ ′
of fr (r) over f (m ,n ) u on Ωgrid . Finally, this result is extended in
P[Eϵ ] ≥ 1 − δ,
Lemma 12 for the continuous domain [0, 1]2 thereby proving that
where the polynomial satisfies (46c). The proof for satisfying the condition
( (46d) for the communications polynomial fc (c) follows a similar
1 ′ ′ 1 ′ ′ ϵ
Eϵ = sup fr(m ,n ) (e
r) − m′ +n′ f (m ,n ) (e
r)u ≤ , procedure as above for the radar polynomial. This completes the proof
rd ∈Ωgrid κm′ +n′ κ 2 3
e
of Theorem 2.
m′ , n′ = 0, 1, 2, 3 ,
Proof: Using Lemmata 4-9, it directly follows that the event Eϵ V. N UMERICAL E XPERIMENTS
holds with a high probability. This completes the proof. We evaluated our model and methods through numerical exper-
3) ∥fr (r)∥2 < 1 for r ∈ [0, 1]2 \ R: Define the minimum separa- iments using the CVX SDP3 [47] solver for the problem in (45).
tion condition for the off-the-grid delay and Doppler parameters as, Throughout all experiments, we generated the columns of the trans-
respectively, formation matrices B and Dp as bn = [1, ej2πσn , . . . , ej2π(J−1)σn ],
max (|[τr ]ℓ − [τ ]r ]ℓ′ | , |τc ]q − [τ ]c ]q′ |) ≥ ∆τ , (90) where σn ∼ N (0, 1). The parameters αr and αc were drawn from
a normal distribution with |[αr ]ℓ | = |[αc ]q | = 1, ∀q and ℓ. The
and real and imaginary components of the coefficients vectors u and v
max (|[νr ]ℓ − [ν]r ]ℓ′ | , |[νc ]q − [νc ]q′ |) ≥ ∆ν , (91) followed a uniform distribution over the interval [0, 1].
′ ′
where ℓ ̸= ℓ , q ̸= q , and |a − b| is the wrap-around metric on [0,1),
e.g. |0.2 − 0.1| = 0.1 but |0.8 − 0.1| = 0.3. The following lemma A. Channel parameter estimation
(m′ ,n′ ) ′ ′
shows that κ1 fr is close to κm′1+n′ f (m ,n ) u everywhere on
We first evaluated our approach for specific realizations of various
[0, 1)2
channel conditions.
Lemma 11. Assume that ∆τ ≤ 1/M and ∆ν ≤ 1/P . Let 0 < δ < Randomly spaced targets: Fig. 3 illustrates a simple recovery
1 the failure probability and ϵ > 0. Then, for all r = [τ, ν] and scenario with number of samples M = 13, pulses/messages P = 9,
m′ , n′ = 0, 1, 2, 3, there exists a numerical constant C such that if radar targets L = 3, communications paths Q = 3, and size

1

M P QJ
of the low-dimensional subspace J = 3. The target delays and
M P ≥ Cµ max(L, Q)J max 2 log , Dopplers were drawn from the interval [0, 1] uniformly at random
ϵ δ
(without replacement). In particular, we used τr = [0.23, 0.68, 0.87],
1 M P LJ 2 MP J LJ P QJ νr = [0.45, 0.42, 0.71], τc = [0.12, 0.21, 0.95], and νc =
log , log , log , log ,
ϵ2 δ δ δ δ [0.09, 0.25, 0.87]. Fig. 3 shows the resulting dual polynomials fr (r)
the following inequality is satisfied and fc (c) after solving (45). The estimates of the channel parameters
of radar and communications are located at ∥fr (r)∥ = 1 and
1 (m′ ,n′ ) 1 ′
,n′ ) ϵ
′ +n′ fr (r) − f (m (r)u ≤ ∥fc (c)∥ = 1. The SDP in (45) perfectly recovers the delay-Doppler
κ m κm′ +n′ 2 3
pairs of both channels.
Proof: See Appendix F. Unequal number of targets and paths: Fig. 4 shows the perfect
The following lemma states that ∥fr (r)∥2 < 1 with a high recovery of all parameters for a scenario with L ̸= Q. In particular,
probability. we set L = 8, Q = 2, M = 17 and P = 11. The delay-
Doppler parameters were drawn uniformly at random as before
Lemma 12. There exists a constant C such that if
so that τr = [0.54, 0.11, 0.49, 0.89, 0.05, 0.23, 0.41, 0.06], νr =
MP J [0.01, 0.51, 0.95, 0.61, 0.45, 0.11, 0.78, 0.22], τc = [0.15, 0.24], and
M P ≥ Cµ max(L, Q)J log2
δ νc = [0.92, 0.05].

M P QJ M P LJ Closely spaced targets: Fig. 5 depicts perfect recovery in a scenario
max log , log , (92)
δ δ when the delay-Doppler pairs of both channels are closely located
in the [0, 1]2 domain with M = 17, P = 11, and the subspace
then with probability 1 − δ, where 0 < δ < 1, ∥fr (r)∥2 < 1, for all
size J = 5. Here, the minimum separation of the time-delays,
r ∈ [0, 1]2 \ R.
and Doppler parameter were set to ∆τ = 0.1 M
and ∆ν = 0.1 P
Proof: The proof requires showing ∥fr (r)∥2 < 1 in the set respectively, which is much lower than theoretical results (see [21,
of points near and far from rj ∈ R. We refer the reader to [34, 29]). For this experiment, the maximum separation was set to 0.3 M
Lemma 14] for details, which proved an analogous result for single- and 0.3
P
τr = [0.520, 0.512, 0.531], νr = [0.210, 0.225, 0.230] and
BD with multi-dimensional bivariate dual polynomials. In our DBD τc = [0.779, 0.796, 0.773], and νc = [0.233, 0.228, 0.25].
15
a b
Fig. 3. Noiseless channel parameter localization for (a) radar with L = 3 targets and (b) communications with Q = 3 paths. Here, M = 13, P = 9, and
J = 3. The parameter estimates are localized in the delay-Doppler plane when the norm of the polynomials assumes a value of unity.
a b
Fig. 4. As in Fig. 3 but for (a) L = 8 targets and (b) Q = 2 paths.
True parameters
Estimated parameters (0.23, 0.531) (0.233, 0.779)

(0.25,0.773)
(0.21, 0.52)
(0.225, 0.512) (0.228, 0.796)
a b
Fig. 5. As in Fig. 3 but for closely spaced targets. The inset plots show the cross-section of PhTPs in the delay-Doppler plane.
B. Statistical performance Figs. 6a and b plot the probability of success over 100 trials with
As we remarked earlier, radar transmit signals and communications varying L = Q and subspace size J to recover radar channel
channels are usually known to the traditional receivers. In Figs. 6, we and communications messages, respectively. We note an inverse
compare the statistical performance of our algorithm in recovering relationship between J and L = Q, as also predicted by Theorem 2,
the conventional parameters-of-interest, i.e., radar channel r and in which the theoretical curve is superposed on the figure. Figs. 6c
communications transmit messages g. The comparison is over several and d show the recovery of the same radar and communications
random realizations of delay and Doppler pairs with the minimum parameters for fixed P = 9 but varying L and time samples M . Here,
separation ∆τ = M 1
and ∆ν = P1 . We declare a successful as suggested by Theorem 2, a logarithmic relationship is observed
estimation when ∥r − b r∥2 < 10−3 and ∥g − g b∥2 < 10−3 . Note that between M and L = Q. This is validated by plotting the curve
the error in communications messages is obtained by concatenating obtained by the right-hand side expression of Eq. (51) in Theorem 2
all P messages in a single vector. The number of targets L and on Figs 6a-d. Finally, Figs. 6e and f show the probability of successful
propagation paths Q were the same. For fixed M = 13 and P = 9, recovery for fixed M = 13, P = 11, and J = 3 but varying Q and
16
a b c d e f 1
0.5
2 5 8 11 2 5 8 11 3 8 13 18 3 8 13 18 0
J J M M
Fig. 6. Probability of successfully recovering (a) radar channel parameters and (b) communications messages for varying L (= Q) and J but fixed M = 13
and P = 9 over 100 trials. (c)-(d) As in (a)-(b), respectively, but for fixed P = 11 and J = 3. (e)-(f) As in (a)-(b), respectively, but for fixed M = 13,
P = 11, and J = 3. In Figs. (a)-(d), the blue curve shows the theoretical result as predicted by Theorem 2.
L. [5] L. Wu, K. V. Mishra, M. R. Bhavani Shankar, and B. Otter-

sten, “Resource allocation in heterogeneously-distributed joint radar-
communications under asynchronous Bayesian tracking framework,”
VI. S UMMARY
IEEE Journal on Selected Areas in Communications, vol. 40, no. 7,
We investigated the joint radar-communications processing from pp. 2026–2042, 2022.
a DBD perspective. The proposed approach leverages the sparse [6] G. Duggal, S. Vishwakarma, K. V. Mishra, and S. S. Ram, “Doppler-
structure of the channels and recovers the parameters of interest in resilient 802.11ad-based ultrashort range automotive joint radar-
communications system,” IEEE Transactions on Aerospace and Elec-
this highly ill-posed problem by minimizing a SoMAN objective. We tronic Systems, vol. 56, no. 5, pp. 4035–4048, 2020.
employed PhTP theories to formulate an SDP problem. Numerical [7] S. Sedighi, K. V. Mishra, M. B. Shankar, and B. Ottersten, “Localiza-
experiments showed perfect recovery of radar and communications tion with one-bit passive radars in narrowband internet-of-things using
parameters in several different scenarios. multivariate polynomial optimization,” IEEE Transactions on Signal
Processing, vol. 69, pp. 2525–2540, 2021.
A lot more system variables may also be considered unknown
[8] S. H. Dokhanchi, B. S. Mysore, K. V. Mishra, and B. Ottersten,
in further extensions of this problem. From a practical standpoint, “A mmWave automotive joint radar-communications system,” IEEE
some parameters such as the PRT T may be accurately estimated by Transactions on Aerospace and Electronic Systems, vol. 55, no. 3, pp.
the common receiver over a few observations through, for instance, 1241–1260, 2019.
periodicity estimation algorithms [55]. Knowing the PRT also yields [9] A. M. Elbir, K. V. Mishra, and S. Chatzinotas, “Terahertz-band joint
ultra-massive MIMO radar-communications: Model-based and model-
an estimate of the number of pulses. In this work, we did not consider
free hybrid beamforming,” IEEE Journal of Special Topics in Signal
the continuous-wave transmission model for radar or even a common Processing, vol. 15, no. 6, pp. 1468–1483, 2021.
waveform for co-design radar-communications systems. These are [10] S. M. Jefferies and J. C. Christou, “Restoration of astronomical images
interesting problems for future work. Similarly, we excluded a by iterative blind deconvolution,” The Astrophysical Journal, vol. 415,
multiple-antenna receiver in this paper. But some of our initial results p. 862, 1993.
[11] G. Ayers and J. C. Dainty, “Iterative blind deconvolution method and its
on multi-antenna DBD have recently appeared in [39].
applications,” Optics Letters, vol. 13, no. 7, pp. 547–549, 1988.
Solving the linear system in (54) provides estimates of the [12] K. Abed-Meraim, W. Qiu, and Y. Hua, “Blind system identification,”
transmitted encoded OFDM symbols, which remain to be decoded. Proceedings of the IEEE, vol. 85, no. 8, pp. 1310–1322, 1997.
This may be achieved by the conventional blind decoding (BdC) [13] A. Ahmed and L. Demanet, “Leveraging diversity and sparsity in blind
and blind equalization (BE) algorithms, on which a large body of deconvolution,” IEEE Transactions on Information Theory, vol. 64,
no. 6, pp. 3975–4000, 2018.
literature exists [56–58]. For instance, in the case of BdC, blind
[14] K. Lee, Y. Li, M. Junge, and Y. Bresler, “Blind recovery of sparse sig-
selection mapping methods [59, 60] have been shown to recover nals from subsampled convolution,” IEEE Transactions on Information
the original symbol sequence using a maximum likelihood decoder Theory, vol. 63, no. 2, pp. 802–821, 2016.
without requiring additional side-information transmission. The BE [15] X. Li, S. Ling, T. Strohmer, and K. Wei, “Rapid, robust, and reliable
techniques are used for joint channel estimation and symbol detection. blind deconvolution via nonconvex optimization,” Applied and Compu-
tational Harmonic Analysis, vol. 47, no. 3, pp. 893–934, 2019.
For instance, [61] leverages null symbols in a linear equalizer to
[16] A. Ahmed, B. Recht, and J. Romberg, “Blind deconvolution using con-
minimize a quadratic criterion. When the channel changes on a vex programming,” IEEE Transactions on Information Theory, vol. 60,
symbol-by-symbol basis, [62] has proposed a BE algorithm to recover no. 3, pp. 1711–1732, 2013.
the transmitted symbols without requiring any statistical prior on the [17] H.-W. Kuo, Y. Lau, Y. Zhang, and J. Wright, “Geometry and symme-
symbols. try in short-and-sparse deconvolution,” in International Conference on
Machine Learning, 2019, pp. 3570–3580.
[18] W. Xu, J.-F. Cai, K. V. Mishra, M. Cho, and A. Kruger, “Precise
R EFERENCES semidefinite programming formulation of atomic norm minimization for
[1] K. V. Mishra, M. R. Bhavani Shankar, V. Koivunen, B. Ottersten, and recovering d-dimensional (d ≥ 2) off-the-grid frequencies,” in IEEE
S. A. Vorobyov, “Toward millimeter wave joint radar-communications: Information Theory and Applications Workshop, 2014, pp. 1–4.
A signal processing perspective,” IEEE Signal Processing Magazine, [19] Y. Chi, “Guaranteed blind sparse spikes deconvolution via lifting and
vol. 36, no. 5, pp. 100–114, 2019. convex optimization,” IEEE Journal of Selected Topics in Signal Pro-
[2] B. Paul, A. R. Chiriyath, and D. W. Bliss, “Survey of RF communi- cessing, vol. 10, no. 4, pp. 782–794, 2016.
cations and sensing convergence research,” IEEE Access, vol. 5, pp. [20] D. Yang, G. Tang, and M. B. Wakin, “Super-resolution of complex
252–270, 2016. exponentials from modulations with unknown waveforms,” IEEE Trans-
[3] J. Liu, K. V. Mishra, and M. Saquib, “Co-designing statistical MIMO actions on Information Theory, vol. 62, no. 10, pp. 5809–5830, 2016.
radar and in-band full-duplex multi-user MIMO communications,” IEEE [21] R. Heckel, V. I. Morgenshtern, and M. Soltanolkotabi, “Super-resolution
Transactions on Aerospace and Electronic Systems, 2022, in press. radar,” Information and Inference: A Journal of the IMA, vol. 5, no. 1,
[4] M. Bică and V. Koivunen, “Radar waveform optimization for target pp. 22–75, 2016.
parameter estimation in cooperative radar-communications systems,” [22] M. A. Suliman and W. Dai, “Blind two-dimensional super-resolution and
IEEE Transactions on Aerospace and Electronic Systems, vol. 55, no. 5, its performance guarantee,” IEEE Transactions on Signal Processing,
pp. 2314–2326, 2018. vol. 70, pp. 2844–2858, 2022.
17
[23] S. Ling and T. Strohmer, “Blind deconvolution meets blind demixing: [46] B. Dumitrescu, Positive trigonometric polynomials and signal processing
Algorithms and performance bounds,” IEEE Transactions on Informa- applications. Springer, 2007.
tion Theory, vol. 63, no. 7, pp. 4497–4520, 2017. [47] M. Grant, S. Boyd, and Y. Ye, “CVX: MATLAB software for disciplined
[24] Y. Li and Y. Chi, “Stable separation and super-resolution of mixture convex programming,” 2009.
models,” Applied and Computational Harmonic Analysis, vol. 46, no. 1, [48] Z. Zhang, Y. Wang, and Z. Tian, “Efficient two-dimensional line spec-
pp. 1–39, 2019. trum estimation based on decoupled atomic norm minimization,” Signal
[25] A. R. Chiriyath, B. Paul, and D. W. Bliss, “Radar-communications con- Processing, vol. 163, pp. 95–106, 2019.
vergence: Coexistence, cooperation, and co-design,” IEEE Transactions [49] K. Krishnan and T. Terlaky, “Interior point and semidefinite approaches
on Cognitive Communications and Networking, vol. 3, no. 1, pp. 1–12, in combinatorial optimization,” Graph theory and combinatorial opti-
2017. mization, pp. 101–157, 2005.
[26] M. Farshchian and I. Selesnick, “A dual-deconvolution algorithm for [50] L. Mingjiu, W. Chen, J. Ma, J. Yang, Q. Cheng, and X. Ma, “Gridless
radar and communication black-space spectrum sharing,” in IEEE Inter- sparse isar imaging via 2-d fast reweighted atomic norm minimization,”
national Workshop on Compressed Sensing Theory and its Applications IEEE Geoscience and Remote Sensing Letters, vol. 19, pp. 1–5, 2022.
to Radar, Sonar and Remote Sensing, 2016, pp. 6–10. [51] Y. Wang and Z. Tian, “Ivdst: A fast algorithm for atomic norm min-
[27] K. V. Mishra and Y. C. Eldar, “Sub-Nyquist radar: Principles and imization in line spectral estimation,” IEEE Signal Processing Letters,
prototypes,” in Compressed Sensing in Radar Signal Processing, A. D. vol. 25, no. 11, pp. 1715–1719, 2018.
Maio, Y. C. Eldar, and A. Haimovich, Eds. Cambridge University Press, [52] M. Cho, K. V. Mishra, J.-F. Cai, and W. Xu, “Block iterative reweighted
2019, pp. 1–48. algorithms for super-resolution of spectrally sparse signals,” IEEE Signal
[28] K. V. Mishra, M. Cho, A. Kruger, and W. Xu, “Spectral super-resolution Processing Letters, vol. 22, no. 12, pp. 2319–2313, 2015.
with prior knowledge,” IEEE Transactions on Signal Processing, vol. 63, [53] J. Monsalve, E. Vargas, K. V. Mishra, B. M. Sadler, and H. Arguello,
no. 20, pp. 5342–5357, 2015. “Beurling-Selberg extremization for dual-blind deconvolution recovery
[29] G. Tang, B. N. Bhaskar, P. Shah, and B. Recht, “Compressed sensing in joint radar-communications,” arXiv preprint arXiv:2211.09253, 2022.
off the grid,” IEEE Transactions on Information Theory, vol. 59, no. 11, [54] E. J. Candès and C. Fernandez-Granda, “Towards a mathematical theory
pp. 7465–7490, 2013. of super-resolution,” Communications on Pure and Applied Mathematics,
[30] Y. Li, K. Lee, and Y. Bresler, “Identifiability and stability in blind decon- vol. 67, no. 6, pp. 906–956, 2014.
volution under minimal assumptions,” IEEE Transactions on Information [55] A. P. Klapuri, “Multiple fundamental frequency estimation based on
Theory, vol. 63, no. 7, pp. 4619–4633, 2017. harmonicity and spectral smoothness,” IEEE Transactions on Speech
[31] Y. Zhang, H.-W. Kuo, and J. Wright, “Structured local optima in sparse and Audio Processing, vol. 11, no. 6, pp. 804–816, 2003.
blind deconvolution,” IEEE Transactions on Information Theory, vol. 66, [56] M. Ghosh, “Blind decision feedback equalization for terrestrial television
no. 1, pp. 419–452, 2019. receivers,” Proceedings of the IEEE, vol. 86, no. 10, pp. 2070–2081,
1998.
[32] E. J. Candès and Y. Plan, “A probabilistic and RIPless theory of
[57] J. Vı́a and I. Santamarı́a, “On the blind identifiability of orthogonal
compressed sensing,” IEEE Transactions on Information Theory, vol. 57,
space-time block codes from second-order statistics,” IEEE Transactions
no. 11, pp. 7235–7254, 2011.
on Information theory, vol. 54, no. 2, pp. 709–722, 2008.
[33] D. Yang, G. Tang, and M. B. Wakin, “Super-resolution of complex
[58] T. R. Dean, M. Wootters, and A. J. Goldsmith, “Blind joint MIMO
exponentials from modulations with unknown waveforms,” IEEE Trans-
channel estimation and decoding,” IEEE Transactions on Information
actions on Information Theory, vol. 62, no. 10, pp. 5809–5830, 2016.
Theory, vol. 65, no. 4, pp. 2507–2524, 2018.
[34] M. A. Suliman and W. Dai, “Mathematical theory of atomic norm
[59] H.-S. Joo, S.-J. Heo, H.-B. Jeon, J.-S. No, and D.-J. Shin, “A new blind
denoising in blind two-dimensional super-resolution,” IEEE Transactions
SLM scheme with low decoding complexity for OFDM systems,” IEEE
on Signal Processing, vol. 69, pp. 1681–1696, 2021.
Transactions on Broadcasting, vol. 58, no. 4, pp. 669–676, 2012.
[35] ——, “Exact three-dimensional estimation in blind super-resolution via [60] A. D. S. Jayalath and C. Tellambura, “SLM and PTS peak-power reduc-
convex optimization,” in Annual Conference on Information Sciences tion of OFDM signals without side information,” IEEE Transactions on
and Systems, 2019, pp. 1–9. Wireless Communications, vol. 4, no. 5, pp. 2006–2013, 2005.
[36] S. Li, M. B. Wakin, and G. Tang, “Atomic norm denoising for complex [61] M. de Courville, P. Duhamel, P. Madec, and J. Palicot, “Blind equal-
exponentials with unknown waveform modulations,” IEEE Transactions ization of OFDM systems based on the minimization of a quadratic
on Information Theory, vol. 66, no. 6, pp. 3893–3913, 2019. criterion,” in IEEE International Conference on Communications, vol. 3,
[37] J. Dong, K. Yang, and Y. Shi, “Blind demixing for low-latency com- 1996, pp. 1318–1322.
munication,” IEEE Transactions on Wireless Communications, vol. 18, [62] T. Y. Al-Naffouri, A. A. Dahman, M. S. Sohail, W. Xu, and B. Hassibi,
no. 2, pp. 897–911, 2018. “Low-complexity blind equalization for OFDM systems with general
[38] E. Vargas, K. V. Mishra, R. Jacome, B. M. Sadler, and H. Arguello, constellations,” IEEE Transactions on Signal Processing, vol. 60, no. 12,
“Joint radar-communications processing from a dual-blind deconvolution pp. 6395–6407, 2012.
perspective,” in IEEE International Conference on Acoustics, Speech and [63] A. K. Karthik and R. S. Blum, “Recent advances in clock synchroniza-
Signal Processing, 2022, pp. 5622–5626. tion for packet-switched networks,” Foundations and Trends® in Signal
[39] R. Jacome, K. V. Mishra, E. Vargas, B. M. Sadler, and H. Ar- Processing, vol. 13, no. 4, 2020.
guello, “Multi-dimensional dual-blind deconvolution approach toward [64] P. Chen, Z. Chen, Z. Cao, and X. Wang, “A new atomic norm for
joint radar-communications,” in IEEE International Workshop on Signal DOA estimation with gain-phase errors,” IEEE Transactions on Signal
Processing Advances in Wireless Communication, 2022, pp. 1–5. Processing, vol. 68, pp. 4293–4306, 2020.
[40] M. I. Skolnik, Radar handbook, 3rd ed. McGraw-Hill, 2008. [65] J. A. Tropp, “User-friendly tail bounds for sums of random matrices,”
[41] L. Cimini, “Analysis and simulation of a digital mobile channel using Foundations of Computational Mathematics, vol. 12, no. 4, pp. 389–434,
orthogonal frequency division multiplexing,” IEEE Transactions on 2012.
Communications, vol. 33, no. 7, pp. 665–675, 1985. [66] L. A. Harris, “Bernstein’s polynomial inequalities and functional analy-
[42] L. Zheng and X. Wang, “Super-resolution delay-Doppler estimation for sis,” Irish Mathematical Society Bulletin, vol. 36, pp. 19–33, 1996.
OFDM passive radar,” IEEE Transactions on Signal Processing, vol. 65,
no. 9, pp. 2197–2210, 2017.
[43] G. R. Muns, K. V. Mishra, C. B. Guerra, Y. C. Eldar, and K. R.
Chowdhury, “Beam alignment and tracking for autonomous vehicular
communication using IEEE 802.11 ad-based radar,” in IEEE Conference
on Computer Communications Workshops - Hot Topics in Social and
Mobile Connected Smart Objects, 2019, pp. 535–540.
[44] X. Luo and G. B. Giannakis, “Low-complexity blind synchronization
and demodulation for (ultra-) wideband multi-user ad hoc access,” IEEE
Transactions on Wireless communications, vol. 5, no. 7, pp. 1930–1941,
2006.
[45] E. J. Candes, T. Strohmer, and V. Voroninski, “Phaselift: Exact and
stable signal recovery from magnitude measurements via convex pro-
gramming,” Communications on Pure and Applied Mathematics, vol. 66,
no. 8, pp. 1241–1274, 2013.

Dual-Blind Deconvolution For Overlaid Radar-Communications Systems

Uploaded by

Document Information

Original Description:

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Dual-Blind Deconvolution For Overlaid Radar-Communications Systems

Uploaded by

Copyright:

Available Formats

This article has been accepted for publication in IEEE Journal on Selected Areas in Information Theory.

Dual-Blind Deconvolution for Overlaid

Problem Measurements Unknown parameters Algorithm

1-D BDb [19] PL−1 gn , α, τ ANM

Blind 1-D SRc[20] PL−1 gℓ , α, τ ANM

SR radard [21] PL−1 rℓ , α ANM

Blind 2-D SRe[22] PL−1 α, rℓ , gℓ ANM

This paper (DBD)h PL−1 αr , u, τr , νr , SoMAN minimization

B. Our Contributions C. Organization and Notations

A. Overlaid Radar-Communications Receiver

where 0 ≤ t ≤ max(Pr Tr , Pc Tc ). Substituting the radar and

Then, all samples of P pulses (and communications messages) are

is a linear indexing sequence of the multidimensional measurement

The measurement vector now becomes g(q) = inf L(Zr , Zc , q)

E[xxH ] = IN . (49) E. Computational Complexity

− fr(0,1) (rℓ ) = 0J×1 , ∀rℓ ∈ R, (55c) minimize ∥Wq0 ∥22

and radar and communications polynomials.

with statistical expectation and

× e−j2π(nτ +pν) υne ⊗ IJ e = sign(αr ) ⊗ u and v

Using E[dne bn ] = 0JP ×J , we have  ′ ′

with the single-BD approaches such as [19, 20] where bounding

exists a numerical constant C such that if Proof: See Appendix E.

Similarly, the following Lemmata 6 and 7 show that

Estimated parameters (0.23, 0.531) (0.233, 0.779)

(0.225, 0.512) (0.228, 0.796)

L. [5] L. Wu, K. V. Mishra, M. R. Bhavani Shankar, and B. Otter-

You might also like

Using E[dne bn ] = 0JP ×J , we have ′ ′