1

Analog to Digital Cognitive Radio:

Sampling, Detection and Hardware
Deborah Cohen, Shahar Tsiper, Yonina C. Eldar
Technion – Israel Institute of Technology, Haifa, Israel
{debby@tx, tsiper@tx, yonina@ee}.technion.ac.il

I. I NTRODUCTION
arXiv:1708.05201v1 [cs.CE] 17 Aug 2017

The radio spectrum is the radio frequency (RF) portion of

the electromagnetic spectrum. These spectral resources are tra-
ditionally allocated to licensed or primary users (PUs) by gov-
ernmental organizations. As can be observed in “RF Spectral
Resources”, most of the frequency bands are already allocated
to one or more PUs. Consequently, new users can hardly find
free frequency bands. Following the ever-increasing demand
from new wireless communication applications, this issue has
become critical over the past few years.
Various studies [1]–[3] have shown that this over-crowded
spectrum is usually significantly underutilized in frequency,
time and space. A 2002 FCC report shows large temporal
and geographic variations in terms of the spectrum usage,
from 15% to 85% utilization. Figure 1 presents the spectrum
occupancy measurements conducted by the Shared Spectrum
Company (SSC) in two major metropolitans, New York and
Chicago. We observe that the occupancy of most frequency
bands does not exceed 50%, and half of them are even below
10% [4]. This shows that spectrum shortage is often due to
the inflexibility of the frequency allocation, resulting in low Fig. 1. Spectrum occupancy measurements in New York City, New York
utilization efficiency. from Aug. 30th.-Sept. 3rd 2004 and Chicago, Illinois from Nov. 16th-18th
2005 (figure courtesy of SSC [5]).
In order to bridge between spectrum scarcity and sparsity
and make better use of spectral resources, the cognitive
radio (CR) [6]–[8] paradigm has been proposed. The idea of
adaptive learning for spectrum sensing can be traced back to PUs, the spectrum sensing task performed by a CR should be
Shannon [9]. This technology, which is under development, reliable and fast [11]–[13] and obey the requirements of the
allows the spectrum to be used more efficiently by granting IEEE 802.22 protocol, described in “IEEE 802.22 Standard
secondary users opportunistic access to licensed spectral bands for WRAN”.
when the corresponding PU is not active. A CR transceiver For the past ten years, CR and its challenges have been
scans for unused bands and changes its transmission and thoroughly reviewed in the literature. Several works [14]–
reception parameters to different frequencies depending on the [19] focus on spectrum sensing and survey sensing tech-
spectrum utilization. niques, along with their performance and limitations. These
CR faces many issues at various levels, and challenges techniques are essentially energy detection, matched filter and
traditional analog, digital and network processing techniques cyclostationary detection. Collaborative CR networks, where
to meet its specific radio sensitivity requirements and wide- different users share their sensing results and cooperatively
band frequency agility [10]. The CR cycle includes two decide on the licensed spectrum occupancy, have also been
major functionalities: spectrum sensing and spectrum access. proposed to overcome practical propagation issues such as
Through spectrum sensing, CRs collect information about the path loss, channel fading and shadowing. Other works [17],
status of the surrounding spectrum’s occupancy, that is the [20] deal with spectrum access, which uses the information
PUs’ activity, allowing for adapted exploitation of the vacant gathered from spectrum sensing to plan the spectrum exploita-
spectral bands. In order to minimize the interference caused to tion by unlicensed users. This functionality includes spectrum
analysis, spectrum access, and spectrum mobility [21]. It
This work was funded by the European Union’s Horizon 2020 research
and innovation program under grant no. 646804-ERC-COG-BNYQ, and by ensures coexistence with PUs and other CRs, by minimizing
the Israel Science Foundation under grant no. 335/14. interference to the former and sharing spectrum with the latter.
 2

In this article, we focus on the issue of spectrum sensing Additional issues of security against malicious users and
from the analog to digital interface point of view, which has various attacks to the network are discussed in [16], [19], [20].
received little attention so far. Our motivation stems from one We now focus on spectrum sensing, which is the funda-
of the main challenges of spectrum sensing in the context mental enabler to spectrum access. In order to increase the
of CR, which is the sampling rate bottleneck. This issue chance of finding an unoccupied spectral band, CRs have to
arises since CRs typically deal with wideband signals with sense a wide band of spectrum. Nyquist rates of wideband
prohibitively high Nyquist rates. Sampling at this rate requires signals are high and can even exceed today’s best ADCs front-
very sophisticated and expensive analog to digital converters end bandwidths. Besides, such high sampling rates generate
(ADCs), leading to a torrent of samples. Therefore, the classic a large number of samples to process, affecting speed and
spectrum sensing methods described above, which are tradi- power consumption. To overcome the rate bottleneck, several
tionally performed on Nyquist rate samples, are difficult to sampling methods have been proposed that leverage the a
implement in practice on wideband signals. Our main goal priori known received signal’s structure, enabling sampling
is to provide an analog to digital CR framework including an reduction. These include the random demodulator [22], [23],
analog preprocessing and sub-Nyquist sampling front-end, and multi-rate sampling [24], multicoset sampling and the modu-
subsequent low rate digital processing. lated wideband converter (MWC) [13], [25]–[27].
We first survey recent methods for spectrum sensing at The CR then performs spectrum sensing on the acquired
sub-Nyquist rates, paving the way to efficient sensing with samples to detect the present of PUs’ transmissions. The
low computational and power requirements. We then review simplest and most common spectrum sensing approach is
spectrum sensing strategies that aim at overcoming other energy detection [28], which does not require any a priori
diverse challenges in the context of CR, such as coping with knowledge on the input signal. Unfortunately, energy detection
low signal to noise ratio (SNR) regimes, channel fading and is very sensitive to noise and performs poorly in low SNRs.
shadowing effects. Throughout the paper, we consider both This becomes even more critical in sub-Nyquist regimes since
theoretical and practical aspects and present hardware im- the sensitivity of energy detection is amplified due to aliasing
plementation of theoretical concepts, demonstrating real-time of the noise [29]. Therefore, this scheme fails to meet CR
wideband spectrum sensing for CR from low rate samples. performance requirements in low SNRs. In contrast, matched
filter (MF) detection [30], [31], which correlates a known
waveform with the input signal to detect the presence of a
II. CR CHALLENGES
transmission, is the optimal linear filter for maximizing SNR
In this article, we focus on CR spectrum sensing. In practice, in the presence of additive stochastic noise. However, this
the information gathered from spectrum sensing is used to plan technique requires perfect knowledge of the potential received
spectrum access by the unlicensed users. For completeness, transmission. When no a priori knowledge can be assumed
we first quickly review the main components and challenges on the received signals’ waveform, MF is difficult to imple-
of spectrum access. ment. A compromise between both methods is cyclostationary
Spectrum analysis or management, which directly follows detection [32], [33]. This strategy is more robust to noise
spectrum sensing, ensures coexistence with PUs and other than energy detection but at the same time only assumes that
CRs. The ambient RF environment is analyzed in order to the signal of interest exhibits cyclostationarity, which is a
characterize the behavior of PUs and the properties of the typical characteristic of communication signals. Consequently,
detected spectrum holes in terms of interference, duration of cyclostationary detection is a natural candidate for spectrum
availability and more. Then, spectrum access can be optimized sensing from sub-Nyquist samples in low SNRs.
to maximize the CR throughput while maintaining interference Besides noise, the task of spectrum sensing for CRs is
caused to the licensed users below a target threshold [20]. further complicated as a result of path loss, fading and
Several techniques have been proposed to minimize inter- shadowing [34]–[36]. These phenomena are due to the signal’s
ference to PUs as well as ensure proper reception of sec- propagation, that can be affected by obstacles and multipath,
ondary signals, such as waveform design and multi-carrier and result in the attenuation of the signal’s power. To overcome
approaches [17], [20]. These are regrouped under the term these practical issues, collaborative CR networks have been
of spectrum sculpting [17]. Besides minimizing interference considered, where different users share their sensing results
to the PUs, spectrum sharing needs to be coordinated within and cooperatively decide on the licensed spectrum occupancy.
the CR network. Various power control and resource allocation Cooperative spectrum sensing may be classified into three
schemes that deal with this issue are reviewed in [16], [20]. catagories based on the way the data is shared by the CRs in
Spectrum access further requires synchronization between the the network: centralized, distributed and relay-assisted. In each
CR transmitter and receiver [20]. of these settings, two options of data fusion arise: decision
The function of spectrum mobility ensures adaptation to fusion, or hard decision, where the CRs only report their binary
changes in the spectrum occupancy. When a licensed user local decisions, and measurement fusion, or soft decision,
starts accessing the channel currently being used by a CR, the where they share their samples [34]. Cooperation has been
latter has to vacate the band and switch to another free spectral shown to improve detection performance and relax sensitivity
band. This operation is referred to as hand-off [16], [20]. The requirements by exploiting spatial diversity [36], [37]. At the
multi-carrier transmission approach evoked above allows to medium access control (MAC) level, cooperation introduces
maintain uninterrupted communication in such scenarios [17]. the need for a tailored communication protocol and a control
 TIONAL TELEC
NA
U.S
OM
M .

DE

UN
PA

IC
RTM

ATIO
Primary
SERVICE

Secondary

ENT OF CO

NS & INF O
MM

RM
E

A
TIO RC
N E

FIXED
AD
MINISTRATIO N

MOBILE

AMATEUR

SATELLITE
SATELLITE

FIXED
Mobile
AERONAUTICAL
AERONAUTICAL
AERONAUTICAL

BROADCASTING
BROADCASTING

FIXED SATELLITE

ACTIVITY CODE
RADIONAVIGATION

EXAMPLE
MOBILE SATELLITE

AMATEUR SATELLITE

EARTH EXPLORATION

GOVERNMENT EXCLUSIVE

October 2003
NON-GOVERNMENT EXCLUSIVE

Table to determine the current status of U.S. allocations.

AIDS

DESCRIPTION
MOBILE

Capital Letters
MARITIME

SATELLITE
SATELLITE
SATELLITE

Office of Spectrum Management

LAND MOBILE
LAND MOBILE
RADIO SERVICES COLOR LEGEND

ALLOCATION USAGE DESIGNATION

INTER-SATELLITE

MARITIME MOBILE
MARITIME MOBILE

RADIONAVIGATION

MOBILE SATELLITE
METEOROLOGICAL
METEOROLOGICAL

1st Capital with lower case letters

U.S. DEPARTMENT OF COMMERCE

STATES
UNITED

FCC and NTIA. As such, it does not completely reflect all aspects, i.e., footnotes and recent changes
This chart is a graphic single-point-in-time portrayal of the Table of Frequency Allocations used by the

the scope of this paper.

made to the Table of Frequency Allocations. Therefore, for complete information, users should consult the
SATELLITE
SATELLITE

RADIOLOCATION

GOVERNMENT/ NON-GOVERNMENT SHARED

AND TIME SIGNAL
SPACE RESEARCH
RADIONAVIGATION
RADIONAVIGATION

SPACE OPERATION

National Telecommunications and Information Administration

FREQUENCY

RADIO ASTRONOMY

RADIODETERMINATION

STANDARD FREQUENCY
STANDARD FREQUENCY
RADIOLOCATION SATELLITE

AND TIME SIGNAL SATELLITE

ALLOCATIONS
THE RADIO SPECTRUM
30.0 3.0 300.0 30.0 3.0 300 3
Standard AERONAUTICAL MOBILE (R)
Frequency and FIXED MOBILE MARITIME FIXED MOBILE 3.025
Time Signal SATELLITE SATELLITE Radiolocation 30.56
(E-S) (E-S) RADIONAVIGATION AERONAUTICAL
Satellite (S-E) LAND
3.1

3 kHz
31.0 MOBILE (OR)
Stand. Frequency

3 GHz
3 MHz

FIXED
FIXED

MOBILE

MOBILE
and Time Signal FIXED MOBILE MOBILE

30 GHz
30 MHz

SATELLITE
300 kHz
Satellite (S-E) 3.155

MARITIME
31.3

Aeronautical

300 MHz
SPACE EARTH

Radionavigation
32.0

(Radio Beacons)
RADIO RESEARCH MOBILE* FIXED
EXPLORATION 322.0

(RADIO BEACONS)
RADIONAVIGATION

RADIO-
ASTRONOMY (Passive) SAT. (Passive) 3.230

LOCATION
31.8 FIXED MOBILE FIXED MOBILE 325

Radiolocation
SPACE Maritime

** EXCEPT AERO MOBILE

RADIONAVIGATION 328.6 Aeronautical AERONAUTICAL
RESEARCH (deep space) 3.3 33.0 Radionavigation RADIONAVIGATION

* EXCEPT AERO MOBILE (R)

32.0 AERONAUTICAL RADIONAVIGATION Mobile
SPACE RES. INTER- SAT LAND (Radio Beacons) (RADIO BEACONS)
RADIONAVIGATION 32.3 335.4 335

FIXED

Radio-
location
FIXED

MOBILE**
RADIONAVIGATION INTER-SATELLITE MOBILE
3.4

RADIO-
34.0

Amateur
33.0
AERONAUTICAL

LOCATION
RADIONAVIGATION

Radiolocation
33.4 MOBILE (R)

FIXED
MOBILE

MOBILE
3.5 FIXED MOBILE

SATELLITE
AERONAUTICAL Radio-
3.5
RADIO-
RADIOLOCATION Radiolocation RADIONAVIGATION 35.0
(Ground) LOCATION location
399.9
3.6 RADIONAVIGATION SATELLITE MOBILE SATELLITE (E-S) LAND
36.0 AERO. RADIO- RADIO- FIXED SAT. Radio-
SPACE RE. EARTH EXPL. NAV.(Ground) LOCATION (S-E) location 400.05 FIXED MOBILE
FIXED MOBILE STD. FREQ. & TIME SIGNAL SAT. (400.1 MHz)
.(Passive) SAT. (Passive) 3.65 400.15
37.0 FIXED SAT. MET. AIDS MOBILE. SPACE RES. Space Opn. MET. SAT.
MOBILE** FIXED (Radiosonde) (S-E)

Mobile
SPACE RESEARCH (S-E) SAT. (S-E) (S-E) (S-E) 36.0
FIXED MOBILE 401.0
3.7

Aeronautical
(space-to-Earth) MET. AIDS SPACE OPN. MET-SAT. EARTH EXPL Earth
Met-Satellite Expl.
Earth Expl Sat
(Radio- (E-S) SAT. (E-S)

AERONAUTICAL
37.6 sonde) (S-E) (E-S)
(E-S)Satellite(E-S) FIXED MOBILE

(RADIO BEACONS)
RADIONAVIGATION
SPACE FIXED

AMATEUR

Office of Spectrum Management - January 2016).

F I X E D MOBILE MET-SAT. EARTH EXPL Met-Satellite Earth Expl Sat 402.0 37.0
RES. SATELLITE (S-E) MET. AIDS
(E-S) SAT. (E-S) (E-S) (E-S) LAND MOBILE
(NTIA), responsible for the Federal use.

38.0 (Radiosonde) 37.5

FIXED FIXED 403.0 Radio Astronomy LAND MOBILE
MOBILE METEOROLOGICAL AIDS (RADIOSONDE) 38.0
SAT. (S-E) 406.0 RADIO ASTRONOMY FIXED MOBILE
38.6 38.25

(S-E)
MOBILE SATELLITE (E-S)

FIXED

FIXED
FIXED-SATELLITE FIXED MOBILE 406.1 FIXED MOBILE

SATELLITE
39.5 RADIO 39.0 4.0
F I X E D MOBILE FIXED MOBILE FIXED MARITIME MOBILE 405
MOBILE FIXED ASTRONOMY 410.0 Aeronautical Mobile RADIONAVIGATION
SATELLITE SAT. SPACE RESEARCH
LAND MOBILE 4.063
NOT ALLOCATED

40.0 FIXED MOBILE (S-S) 40.0 415

F I X E D MOBILE SPACE Earth EARTH 420.0
Expl. EXPL
SAT SAT. RES. (E-S) Sat (s - e) SAT (E-S) 4.2 AERONAUTICAL MARITIME
40.5 RADIOLOCATION Amateur
BCST BROAD- FX-SAT Fixed Mobile RADIONAVIGATION MOBILE

FIXED
SAT. CASTING (S-E)

MOBILE
41.0 AERONAUTICAL 450.0 435
RF S PECTRAL R ESOURCES

MOBILE

ISM – 40.68 ± .02 MHz

BROAD- BCST LAND MOBILE 42.0

ACTIVITIES
DESIGNATIONS
BAND
WAVELENGTH
MARITIME
FIXED MOBILE RADIONAVIGATION 454.0
CASTING SAT.

FREQUENCY 0
FIXED LAND MOBILE
42.5 455.0
RADIO 4.4
LAND MOBILE
456.0 FIXED LAND
FIXED 4.438
ASTRONOMY
F I X E D M O B I L E * * SATELLITE (E-S) FIXED LAND MOBILE MOBILE
43.5 FIXED MOBILE
4.5 LAND MOBILE
460.0 FIXED
FIXED MOBILE FIXED 43.69 MOBILE*
SATELLITE (E-S) 462.5375

Infra-sonics
SATELLITE (E-S) LAND MOBILE
45.5 462.7375 4.65

MOBILE
RADIONAV. MOBILE LAND MOBILE

10 Hz
LAND AERONAUTICAL MOBILE (R)

MARITIME
FIXED

3 x 107m
Aeronautical

SATELLITE MOBILE 467.5375 4.7

SAT (E-S).
Radionavigation

46.9 LAND MOBILE 467.7375 MOBILE

FIXED
AERONAUTICAL MOBILE (OR)

Satellite (S-E)
RADIONAV.SAT.

Meteorological

FIXED
MOB. SAT(E-S) MOBILE FIXED LAND MOBILE FIXED

MOBILE
47.0 470.0 4.75

SATELLITE (S-E)
46.6
AMATEUR AMATEUR SATELLITE FIXED MOBILE
47.0 MOBILE* FIXED
47.2 BROADCASTING
FX 4.8 4.85
FIXED MOBILE LAND MOBILE FIXED 495

100 Hz
SAT(E-S) (TV CHANNELS 14 - 20)

3 x 106m
48.2 FIXED MOBILE FIXED MOBILE MOBILE (DISTRESS AND CALLING)
FX 4.94 505

LAND
FIXED MOBILE 512.0 4.995

MOBILE
SAT(E-S) 50.2 FIXED MOBILE** STANDARD FREQ. AND TIME SIGNAL (5000 KHZ) MARITIME MOBILE 510
EARTH 4.99 5.003
49.6

Sonics
SPACE RESEARCH EXPLORATION RADIO ASTRONOMY Space Research (Passive) STANDARD FREQ. Space Research
FIXED MOBILE 5.005 MARITIME AERONAUTICAL 9
SATELLITE 5.0 FIXED

Audible Range
50.4

1 kHz
FI XED MOBILE AERONAUTICAL 50.0 5.060 MOBILE RADIONAVIGATION

3 x 105m
FIXED MOBILE SATELLITE (E-S) SATELLITE (E-S) (SHIPS ONLY) (RADIO BEACONS)
RADIONAVIGATION

3 kHz
51.4
FIXED MOBILE 5.15 525
AERO. RADIONAV. FIXED SAT (S-E) AERONAUTICAL

VERY LOW FREQUENCY (VLF)

5.25 MOBILE RADIONAVIGATION
52.6

FIXED
RADIOLOCATION Radiolocation (RADIO BEACONS)
EARTH SPACE 535

MOBILE**
EXPLORATION 5.35

AMATEUR
RESEARCH

10 kHz
SATELLITE (Passive) AERONAUTICAL RADIO- Radio- 5.45

30,000 m
(Passive) RADIONAV. LOCATION location
54.25 5.46 54.0
SPACE RES. INTER- SAT EARTH EXPL-SAT (Passive) RADIONAVIGATION Radiolocation

BROADCASTING
55.78 5.47 AERONAUTICAL MOBILE (R)
FIXED MOBILE INTER- SAT SPACE RES. EARTH-ES MARITIME

(TV CHANNELS 21-36)

56.9 Radiolocation

LF
FIXED MOBILE SPACE RES. EARTH-ES INTER- SAT RADIONAVIGATION
57.0 5.6 5.68
EARTH MARITIME METEOROLOGICAL AERONAUTICAL MOBILE (OR)
SPACE Radiolocation 5.73

100 kHz
INTER RADIONAVIGATION

3,000 m
MOBILE FIXED AIDS

Ultra-sonics
EXPLORATION 5.65

AM Broadcast
RES. - SAT MOBILE* FIXED 5.90
SAT. (Passive) RADIOLOCATION Amateur
58.2 5.83 FIXED
SPACE EARTH MOBILE* BROADCASTING
MOBILE FIXED RESEARCH EXPLORATION RADIO- Amateur- sat (s-e) Amateur 608.0
LOCATION LAND MOBILE RADIO ASTRONOMY
RADIONAVIGATION

(Passive) SAT. (Passive) 5.85 614.0 5.95

MF
EARTH 59.0 MOBILE FIXED SAT(E-S) Amateur
MOBILE SPACE RADIO- INTER- 5.925

1 MHz
EXPLORATION F I X E D

300 m
SAT. (Passive) RES.. LOC. SAT FIXED
59.3 FIXED SATELLITE (E-S) BROADCASTING

ISM – 5.8 ± .075 GHz

missions, arrays of sensors have been considered and DOA

jointly the carrier frequencies and DOAs of the received trans-
receivers do not sense the PU’s signal. In order to estimate
the latter for transmission to the opposite direction, where
a certain direction with respect to the CR, may be used by
For example, a spectral band occupied by a PU situated in
of vacant bands in space in addition to the frequency domain.
recovery enhances CR performance by allowing exploitation
trum sensing and direction of arrival (DOA) estimation. DOA
Finally, CRs may require, or at least benefit from joint spec-
channel [10]. These CR communication challenges are outside
frequency channel or as an underlay ultra-wideband (UWB)
channel [14], [15], which can be implemented as a dedicated
FIXED MOBILE RADIO- INTER- 6.425
LOCATION SATELLITE FIXED
SATELLITE (E-S) MOBILE 6.2 14

HF
64.0 6.525
INTER-

UNLICENSED DEVICES
FIXED

30 m
MOBILE** FIXED

ISM – 61.25 ± .250 GHz

SATELLITE FIXED
SATELLITE (E-S)

10 MHz
59-64 GHz IS DESIGNATED FOR
65.0 FIXED SATELLITE (S-E)(E-S) FIXED 6.70 MARITIME MOBILE
EARTH SPACE INTER-

BROADCASTING
6.875
National Telecommunications and Information Administration
which administers spectrum for non-Federal use, and the
between the Federal Communications Commission (FCC),
regulatory responsibility for the radio spectrum is divided
tions that allocate them to fixed users. In the United States,
Spectral resources are managed by governmental organiza-

EXPLORATION RESEARCH F I X E D M O B I L E * * SATELLITE

(TV CHANNELS 2-4)

FM Broadcast
SATELLITE MOBILE FIXED FIXED

TV BROADCASTING
66.0 SATELLITE (E-S) 6.525
RADIO- RADIO FIXED 7.025
INTER- MOBILE FIXED SAT (E-S) AERONAUTICAL MOBILE (R)

VHF
MOBILE MOBILE 7.075

MAGNIFIED ABOVE
SATELLITE

3m
NAVIGATION SATELLITE NAVIGATION

100 MHz
SATELLITE MOBILE FIXED 698 6.685

P
7.125 AERONAUTICAL MOBILE (OR)
71.0 6.765
FIXED MOBILE BROADCAST
Fixed

FIXED

FIXED

L
Mobile FIXED

THE RADIO SPECTRUM

746

UHF
7.19 FIXED MOBILE BROADCAST 7.0
FIXED SPACE RESEARCH (E-S)

1 GHz
30 cm
7.235 AMATEUR AMATEUR SATELLITE
MARITIME MOBILE

FIXED 764 7.1

7.25
ISM – 6.78 ± .015 MHz

(E-S)

(E-S)
FIXED
MOBILE

FIXED

S C
MOBILE
MOBILE
FIXED SATELLITE (S-E) Fixed FIXED MOBILE 72.0 AMATEUR
SATELLITE (S-E)

SATELLITE

SATELLITE
7.30 776 FIXED MOBILE
FIXED SATELLITE (S-E) FIXED Mobile Satellite (S-E) 7.3
73.0 FIXED
FIXED BROADCASTING
BROADCASTING
Mobile

X
7.45 Mobile 7.35 19.95

SHF
FIXED MET. Mobile FIXED MOBILE BROADCAST RADIO ASTRONOMY

3 cm
SATELLITE (S-E) SATELLITE (S-E) FIXED Satellite (S-E)

10 GHz
Microwaves
74.0 7.55 74.6 STANDARD FREQ. AND TIME SIGNAL (20 kHz)
FIXED MOBILE FIXED Mobile FIXED MOBILE
SATELLITE (E-S) FIXED FIXED Satellite (S-E) 794 74.8 20.05
75.5 SATELLITE (S-E) AERONAUTICAL RADIONAVIGATION
7.75 FIXED MOBILE 75.2
AMATEUR AMATEUR SATELLITE FIXED FIXED MOBILE
Mobile

76.0 806 75.4

7.90
FIXED

RADIOLOC. Amateur FIXED MOBILE

EHF
77.0 FIXED MOBILE LAND MOBILE FIXED 76.0
Fixed

Radar
Bands
Amateur SATELLITE (E-S) SATELLITE (E-S)

0.3 cm

Radar
RADIOLOC.

100 GHz
Amateur Sat. 77.5 8.025 821
RADIOLOC. FIXED EARTH EXPL. Mobile LAND MOBILE
AMATEUR AMATEUR SAT 78.0 FIXED 824

300 GHz
RADIO- Amateur SATELLITE (E-S) SATELLITE(S-E) Satellite (E-S) LAND MOBILE FIXED 8.1
8.175 849 FIXED
FIXED MARITIME
MARITIMEMOBILE
MOBILE
LOCATION Amateur Satellite EARTH EXPL. FIXED MET. Mobile AERONAUTICAL MOBILE 8.195
81.0 SAT. (S-E) SATELLITE FIXED SATELLITE Satellite (E-S) 851
(E-S) (E-S) (no airborne) LAND MOBILE FIXED
FIXED 8.215 866
MOBILE EARTH EXPL. FIXED Mobile Satellite LAND MOBILE

1 THz
FIXED MOBILE SATELLITE SATELLITE SATELLITE FIXED 869

0.03 cm
SATELLITE (S-E) (E-S)(no airborne)
FIXED

(S-E) (S-E) (E-S) 8.4 LAND MOBILE FIXED

SPACE RESEARCH (S-E) 894
84.0 FIXED AERONAUTICAL MOBILE
(deep space only) 896
BROAD- 8.45 LAND MOBILE FIXED
BROADCASTING
MOBILE

BROAD- SPACE RESEARCH (S-E) FIXED 901901

Sub-Millimeter
MARITIME

(TV CHANNELS 5-6)

FIXED MOBILE CASTING CASTING

MARITIME MOBILE

8.5 MOBILE FIXED 902

SATELLITE RADIOLOCATION Radiolocation
86.0

1013Hz
9.0

INFRARED
3 x 105Å
AERONAUTICAL RADIOLOCATION Amateur 8.815

Infrared
RADIONAVIGATION Radiolocation 88.0
9.2 928 AERONAUTICAL MOBILE (R)
MARITIME Radiolocation FIXED 8.965
RADIONAVIGATION 929 AERONAUTICAL MOBILE (OR) 30
9.3 9.040

RADIO
SPACE
EARTH

(Passive)
(Passive)
Meteorological LAND MOBILE FIXED 930

SATELLITE

RESEARCH
ISM – 915.0 ± 13 MHz
RADIONAVIGATION Radiolocation

ASTRONOMY
EXPLORATION
Aids MOBILE FIXED 931 FIXED
9.5
92.0

1014Hz
LAND MOBILE FIXED

3 x 104Å
FIXED 932 9.4
RADIO- SATELLITE
FIXED 935 FIXED BROADCASTING
FIXED MOBILE LOCATION LAND MOBILE FIXED 9.5

Visible
(E-S) RADIOLOCATION Radiolocation 940
95.0 LAND MOBILE FIXED

VISIBLE
941
(AM RADIO)

FIXED 944 BROADCASTING

10.0
BROADCASTING

RADIO- FIXED

1015Hz
Radiolocation Amateur 960

3 x 103Å
LOCATION 9.9

Radio-
location
30

RADIO-

RADIO-

MOBILE

MOBILE
10.45 FIXED

SATELLITE
SATELLITE
NAVIGATION
Amateur

NAVIGATION
9.995
(FM RADIO)

Radiolocation Amateur STANDARD FREQ. AND TIME SIGNAL (10,000 kHz)

100.0 Satellite 10.5 10.003
BROADCASTING

EARTH EXPL. SPACE RESEARCH RADIOLOCATION STANDARD FREQ. Space Research

10.55 10.005
SATELLITE (Passive) (Passive) FIXED AERONAUTICAL MOBILE (R)
102.0 10.6 10.1
FIXED SPACE RESEARCH EARTH EXPL. RADIO AMATEUR

1016Hz
3 x 102Å
FIXED 10.15
FIXED SATELLITE (Passive) SAT. (Passive) ASTRONOMY
(S-E) 10.68

Ultraviolet
105.0 RADIO SPACE EARTH EXPL.
ASTRONOMY RESEARCH (Passive) SATELLITE (Passive)
AERONAUTICAL

10.7
RADIONAVIGATION

108.0

ULTRAVIOLET
FIXED

Mobile*

1017Hz
3 x 10Å
1215

(S-E)

FIXED
FIXED
RADIONAVIGATION 11.175

RADIO
SPACE
EARTH
FIXED

(Passive)
(Passive)
SATELLITE
RADIOLOCATION AERONAUTICAL MOBILE (OR)

SATELLITE

RESEARCH
SATELLITE (S-E) 11.275

ASTRONOMY
1240

EXPLORATION
AERONAUTICAL MOBILE (R)
11.7 11.4
FIXED
11.6

3Å
RADIOLOCATION Amateur FIXED BROADCASTING

1018Hz
11.65
MARITIME MOBILE

AERONAUTICAL

116.0
RADIONAVIGATION

1300

(S-E)
SPACE EARTH
FIXED
INTER- 117.975

Mobile **
BROADCASTING
SATELLITE
RESEARCH EXPL SAT.

X-RAY
F I X E D MOBILE SATELLITE
(Passive)
AERONAUTICAL AERONAUTICAL
(Passive) Radiolocation 12.05
12.2 RADIONAVIGATION

X-ray
119.98 MOBILE (R) FIXED BROADCASTING
MO- INTER-
SPACE EARTH Amatuer 1350
FIXED SAT. 121.9375 12.10
BILE

1019Hz
RES. EXPL .SAT FIXED MOBILE RADIOLOCATION AERONAUTICAL MOBILE FIXED
120.02

3 x 10 -1Å
BROADCASTING FIXED 1390 123.0875
INTER- SPACE EARTH FIXED-SAT (E-S) AERONAUTICAL MOBILE 12.23
SATELLITE FIXED MOBILE ** 1392 123.5875
MOBILE SATELLITE RESEARCH EXPL SAT.

ISM – 122.5 ± .500 GHz

FIXED (Passive) MOBILE ** FIXED MARITIME
(Passive) 1395
126.0 12.7 LAND MOBILE AERONAUTICAL MOBILE
1400

Gamma-ray
FIXED RADIO ASTRONOMY EARTH EXPL SAT (Passive) SPA CE RESEARCH ( Passive) MOBILE (R)
MOBILE FIXED 1427 13.2
SATELLITE (E-S) AERONAUTICAL MOBILE (OR)

1020Hz
LAND MOBILE Fixed (TLM) 128.8125 13.26

3 x 10 -2Å
12.75 1429.5
SPACE FIXED AERONAUTICAL AERONAUTICAL MOBILE (R)
LAND MOBILE (TLM) FIXED (TLM) 13.36

FIXED
INTER-

RADIO-
RESEARCH (S-E) SATELLITE MOBILE

MOBILE
FIXED 1430 MOBILE (R) RADIOASTRONOMY

LOCATION
SATELLITE
(Deep Space) (E-S) FIXED-SAT (S-E) FIXED (TLM) LAND MOBILE (TLM)
132.0125 13.41
13.25 FIXED

GAMMA-RAY
AERONAUTICAL RADIONAV. Space Research (E-S) 1432 AERONAUTICAL MOBILE (R) Mobile* 13.57
134.0 13.4 FIXED** MOBILE 136.0 FIXED BROADCASTING
Standard RADIO- Radio- 1435 AERONAUTICAL MOBILE (R) 13.6 59
Freq. and LOCATION location MOBILE (AERONAUTICAL TELEMETERING) 1525 137.0 BROADCASTING STANDARD FREQ. AND TIME SIGNAL (60 kHz)

1021Hz
13.8

3 x 10 -3Å
Time Signal 13.75 MOBILE SAT. MOB. SAT. (S-E) SPACE RES. (S-E) SPACE OPN. (S-E) MET. SAT. (S-E)
Space

RADIO- FIXED Radio- 137.025 FIXED BROADCASTING 61

Research

Satellite (E-S) Mobile ** (Space to Earth) Mob. Sat. (S-E) SPACE RES. (S-E) SPACE OPN. (S-E) MET. SAT. (S-E) Mobile* 13.87
LOCATION SAT.(E-S) location
14.0 1530 137.175

Radio-
ISM – 13.560 ± .007 MHz

FIXED

RADIO-
location

RADIO-
MARITIME MOBILE SAT. MOBILE SAT. Mobile

MOBILE
MOBILE
RADIO MOB. SAT. (S-E) SPACE RES. (S-E) SPACE OPN. (S-E) MET. SAT. (S-E) 14.0

SATELLITE
FIXED Land Mobile

SATELLITE
Space (Space to Earth) (Space to Earth) (Aero. TLM) 137.825

NAVIGATION
AMATEUR AMATEUR SATELLITE

NAVIGATION
Research NAVIGATION SAT. (E-S) Satellite (E-S) 1535 Mob. Sat. (S-E) SPACE RES. (S-E) SPACE OPN. (S-E) MET. SAT. (S-E) 14.25
MARITIME MOBILE SATELLITE 138.0
14.2 (space to Earth) MOBILE SATELLITE (S-E) AMATEUR
142.0 14.35
AMATEUR AMATEUR SATELLITE FIXED Land Mobile 1544 FIXED MOBILE
144.0 Mobile** MOBILE SATELLITE (S-E)

1022Hz
SATELLITE (E-S) Satellite (E-S) 144.0
FIXED

3 x 10 -4Å
RADIO- 1545 AMATEUR Mobile*
MOBILE

LOCATION Amateur Amateur Satellite AERONAUTICAL MOBILE SATELLITE (R) AMATEUR SATELLITE 146.0
FIXED
Mobile Satellite (S- E)
MARITIME

149.0 FIXED Land Mobile 14.4 (space to Earth) AMATEUR

FIXED Fixed Mobile 1549.5 148.0 14.990
FIXED MOBILE SAT. (E-S) Satellite (E-S) 14.47 AERONAUTICAL MOBILE SATELLITE (R) MOBILE SATELLITE MOBILE SATELLITE (E-S) FIXED MOBILE 149.9 STANDARD FREQ. AND TIME SIGNAL (15,000 kHz)
SATELLITE (S-E) (space to Earth) (Space to Earth) RADIONAV-SATELLITE 15.005 70
Fixed Mobile FX SAT.(E-S) L M Sat(E-S) 1558.5 MOBILE SATELLITE (E-S) 150.05
150.0 14.5 AERONAUTICAL MOBILE SATELLITE (R) (space to Earth) STANDARD FREQ. Space Research
FIXED EARTH FIXED Mobile Space Research FIXED MOBILE 15.010
FIXED SATELLITE MOBILE EXPL. SAT. SPACE RES. 14.7145 1559 150.8 AERONAUTICAL MOBILE (OR)
(S-E) (Passive) (Passive) Space Research AERONAUTICAL RADIONAVIGATION RADIONAV. SATELLITE (Space to Earth) 15.10
MOBILE FIXED

1023Hz
Fixed 1610 LAND MOBILE

3 x 10 -5Å
151.0 AERO. RADIONAVIGATION RADIO DET. SAT. (E-S) M O B I L E S A T ( E - S ) 152.855
15.1365
1610.6 BROADCASTING
FIXED Mobile Space Research AERO. RADIONAV. RADIO DET. SAT. (E-S) MOBILE SAT. (E-S) RADIO ASTRONOMY LAND MOBILE
15.35 AERO. RADIONAV. RADIO DET. SAT. (E-S) MOBILE SAT. (E-S) Mobile Sat. (S-E)
1613.8
SPACE RESEARCH EARTH EXPL. SAT. 1626.5 154.0 15.6
RADIO ASTRONOMY (Passive) (Passive) LAND MOBILE FIXED BROADCASTING
15.4 FIXED 15.8
(S-E)

Cosmic-ray
FIXED
FIXED

AERONAUTICAL RADIONAVIGATION 156.2475

FIXED

MARITIME MOBILE
SATELLITE

15.43 MOBILE SATELLITE (E-S) 157.0375

COSMIC-RAY

1024Hz
MARITIME MOBILE

3 x 10 -6Å
AERO RADIONAV FIXED SAT (E-S) 157.1875
FIXED
Radiolocation

15.63 MARITIME MOBILE LAND MOBILE 16.36

164.0 AERONAUTICAL RADIONAVIGATION 157.45 1605
EARTH RADIO 15.7 1660 FIXED LAND MOBILE MOBILE BROADCASTING
SPACE RES.
MARITIME MOBILE

EXPLORATION RADIOLOCATION Radiolocation RADIO ASTRONOMY MOBILE SAT. (E-S) 161.575 MARITIME 1615
SATELLITE (Passive) ASTRONOMY (Passive) 16.6 1660.5 MARITIME MOBILE
168.0 RADIOLOCATION Space Res.(act.) Radiolocation RADIO ASTRONOMY SPACE RESEARCH (Passive) 161.625 MOBILE 90
MOBILE 17.1 1668.4 LAND MOBILE
FIXED METEOROLOGICAL 161.775
RADIOLOCATION Radiolocation MARITIME MOBILE LAND MOBILE 17.41

1025Hz
170.0 RADIO ASTRONOMY

3 x 10 -7Å
INTER- AIDS (RADIOSONDE) 162.0125 BROADCASTING
FIXED MOBILE 17.2 1670 FIXED
SATELLITE Earth Expl Sat Space Res. RADIOLOC. Radioloc. 17.48
174.5 Radiolocation
17.3 MOBILE** FIXED FIXED BROADCASTING
SPACE EARTH BCST SAT. FX SAT (E-S) 1675 17.55
RESEARCH INTER- 17.7 METEOROLOGICAL METEOROLOGICAL
FIXED MOBILE SATELLITE EXPLORATION FIXED SATELLITE (E-S) FIXED BROADCASTING 1705
(Passive) SAT. (Passive) SATELLITE (s-E) AIDS (Radiosonde)
FIXED

17.8 1700 17.9

MOBILE

176.5 FIXED SATELLITE (S-E) FIXED AERONAUTICAL MOBILE (R)

INTER- 18.3 17.97
FIXED MOBILE FIXED SATELLITE (S-E) 173.2 AERONAUTICAL MOBILE (OR)
SATELLITE 18.6 FIXED Land Mobile 18.03
(s-E)
Fixed

SPACE RES. FX SAT (S-E) EARTH EXPL. SAT.

FIXED

182.0 173.4 FIXED

FIXED

18.8
RADIO-

18.068
MET. SAT.
MOBILE

EARTH FIXED MOBILE

TRAVELERS INFORMATION STATIONS (G) AT 1610 kHz

RADIO SPACE RESEARCH 1710 AMATEUR

LOCATION

EXPLORATION FIXED SATELLITE (S-E) 174.0 AMATEUR SATELLITE

ASTRONOMY (Passive) 18.168 110
SATELLITE (Passive) 19.3 FIXED MOBILE Mobile FIXED 1800
185.0 FIXED SATELLITE (S-E) FIXED 1755 18.78
INTER- 19.7 MARITIME MOBILE
FIXED MOBILE FIXED MOBILE 18.9
SATELLITE FIXED SATELLITE (S-E) MOBILE SAT. (S-E) 1850 FIXED BROADCASTING
190.0 20.1 19.02 AMATEUR
FX SAT (S-E) MOBILE SATELLITE (S-E) FIXED
20.2 FIXED MOBILE 19.68
STD FREQ. & TIME SAT (S-E) MOBILE SAT (S-E)
FXFIXED MARITIME MOBILE
21.2 19.80
FIXED

FIXED 1900
MOBILE

SPACE RES. F I X E D MOBILE EARTH EXPL. SAT. 2000 19.990

MARITIME

21.4 MOBILE SATELLITE (E-S) STAND. FREQ. & TIME SIG. Space Research
RADIO-

RADIO-

MOBILE

19.995
MOBILE
RADIONAVIGATION Radiolocation

FIXED MOBILE
SATELLITE

SATELLITE

22.0 2020 STANDARD FREQUENCY & TIME SIGNAL (20,000 KHZ)

NAVIGATION

NAVIGATION

MOBILE** Space Research

20.005 RADIOLOCATION 130
FIXED FIXED MOBILE STANDARD FREQ.
200.0 S P A C E R A D . A S T MOBILE** F I X E D EARTH EXPL. SAT.
22.21 2025 20.010
SPACE RES. EARTH RES. 22.5 SPACE RES. EARTH EXPL. SPACE OP.
FIXED MOBILE EXPLORATION SAT. (E-S)(s-s) SAT. (E-S)(s-s) (E-S)(s-s) MOB. FX. Mobile
(Passive) (Passive) 2110 FIXED 2000
MOBILE
BROADCASTING

202.0 FIXED FIXED MOBILE MARITIME

(TV CHANNELS 7-13)

21.0 MOBILE FIXED

22.55 2155 MOBILE
S) AMATEUR AMATEUR SATELLITE
FIXED 2065
FIXED MOBILE INTER-SATELLITE 2160 21.45 MARITIME MOBILE (TELEPHONY)
FIXED

(E-S)

BROADCASTING 2107
MOBILE

FIXED

FIXED

FIXED MOBILE
MOBILE

2180 21.85
MARITIME

23.55
SATELLITE

FIXED LAND MARITIME

FIXED MOBILE MOBILE SATELLITE (S-E) 21.924 FIXED MOBILE MOBILE
2200 216.0 AERONAUTICAL MOBILE (R) MOBILE
Fixed Mobile Radio- FIXED MOBILE 22.0
217.0 SPACE SPACE EARTH location Amateur 2170
23.6 FIXED MOBILE MARITIME MOBILE (TELEPHONY) 160
(LOS) RESEARCH OPERATION EXPLORATION 220.0 MARITIME MOBILE 2173.5
SPACE RES. EARTH EXPL. (LOS) (s-E)(s-s) (s-E)(s-s) SAT. (s-E)(s-s) LAND MOBILE FIXED Radiolocation
RADIO ASTRONOMY 222.0 MOBILE (DISTRESS AND CALLING)
(Passive) SAT. (Passive) 2290 AMATEUR Radiolocation 22.855 2190.5
24.0 FIXED MARITIME MOBILE (TELEPHONY)
OF SPECTRUM OCCUPIED.

SPACE RES..(S-E) FIXED MOBILE** 225.0 23.0 2194

RADIO
SPACE

2300
EARTH

(Passive)
(Passive)

FIXED Mobile*
SATELLITE

RESEARCH

Amateur 23.2
ASTRONOMY

AMATEUR 2305
EXPLORATION

AMATEUR SATELLITE
require occasional usage of this spectral range.

Amateur RADIOLOCATION MOBILE** FIXED AERONAUTICAL MOBILE (OR)

231.0 2310 23.35
FIXED Radio-
FIXED

Radiolocation Mobile Fixed MOB FX R- LOC. B-SAT 2320

FIXED MOBILE SATELLITE (S-E) location 24.05
MOBILE

FIXED MOBILE**
FIXED
MARITIME

MOBILE

FIXED
235.0 Earth Expl. RADIO- Radio- Mobile Radio- Fixed BCST-SATELLITE
SPACE RES. EARTH EXPL. Satellite Amateur
FIXED MOBILE SATELLITE(S-E) (Passive) SAT. (Passive) LOCATION location location 2345 24.89
(Active) AMATEUR SATELLITE AMATEUR
FIXED

238.0 190
MOBILE
MOBILE

FIXED Radio- 24.25 235.0 24.99

MARITIME

FIXED MOBILE SATELLITE (S-E)

Radiolocation Mobile Fixed
MOB FX R- LOC. B-SAT
location FIXED STANDARD FREQ. AND TIME SIGNAL (25,000 kHz) AERONAUTICAL
LAND MOBILE

241.0 24.45 2360 25.005

RADIO- STANDARD FREQ. Space Research RADIONAVIGATION
Amateur Satellite Amateur RADIONAVIGATION INTER-SATELLITE MOBILE RADIOLOCATION Fixed 25.01 200
LOCATION LAND MOBILE
248.0 25.07 2495
AMATEUR SATELLITE AMATEUR 24.65 2385 MARITIME MOBILE
25.21 STANDARD FREQ. AND TIME SIGNAL (2500kHz)
ISM – 245.0 ± 1GHz

250.0 Earth Expl.

INTER-SATELLITE RADIOLOCATION SATELLITE (E-S) MOBILE FIXED 2390 LAND MOBILE 2501
EARTH EXPLORATION Satellite (Active) 25.33
SPACE RES. (Passive) AMATEUR FIXED MOBILE** STANDARD FREQ. Space Research
SATELLITE (Passive) 24.75 2400 25.55 2502
252.0 INTER-SATELLITE F I XRADIOLOCATION
SATELLITE
E D (E-S) RADIO ASTRONOMY STANDARD FREQ. AND TIME SIGNAL
AMATEUR 25.67
ISM – 2450.0 ± 50 MHz

RADIONAVIGATION SATELLITE BROADCASTING 2505

Earth Standard 2417 26.1
(E-S)
ISM – 24.125 ± 0.125 GHz

Exploration Frequency and Radiolocation Amateur MARITIME MOBILE

FIXED MOBILE 2450 26.175
RADIO-

RADIO-

MOBILE

MOBILE

Satellite Time Signal 25.05 LAND MOBILE

SATELLITE
SATELLITE

(E-S) FIXED MOBILE Radiolocation 26.48

265.0 (S-S) F I X ESatellite
D SATELLITE FIXED FIXED MOBILE**
2483.5 26.95
Mobile

(E-S) FIXED
TRUM SEGMENTS SHOWN IS NOT PROPORTIONAL TO THE ACTUAL AMOUNT
PLEASE NOTE: THE SPACING ALLOTTED THE SERVICES IN THE SPEC-

RADIODETERMINATION SAT. (S-E)

FIXED

Earth 25.25 MOBILE SATELLITE (S-E) 2500

MOBILE

MOBILE** 26.96
std Exploration
freq e-e-sat FIXED INTER-SAT. MOBILE FX-SAT (S - E)
Aeronautical

BCST - SAT. MOBILE** FIXED

FIXED

25.5 2655
(E-S)

27.23
RADIO-

FIXED
MOBILE

FIXED
MOBILE

MOBILE

&Satellite
time e-e-sat
(S-S) (s-s) INTER-SAT. MOBILE E-Expl Sat Radio Ast Space res. FIXED MOBILE**
MARITIME

FIXED MOB** B- SAT. FX FX-SAT

NAVIGATION ASTRONOMY

NAVIGATION SATELLITE
AERONAUTICAL

MOBILE 27.0 2690 27.41

LAND MOBILE

MOBILE SATELLITE

275.0 Earth Exploration RADIO ASTRON. SPACE RESEARCH EARTH EXPL SAT FIXED LAND MOBILE
RADIONAVIGATION

FIXED INTER- MOBILE 2700 27.54

Satellite (S-S) SATELLITE FIXED MOBILE
27.5 28.0
FIXED MOBILE FIXED AMATEUR AMATEUR SATELLITE 2850

same low rate samples obtained from an array of sensors.

SAT (E-S) AERONAUTICAL METEOROLOGICAL Radiolocation
Maritime
275
29.5 AIDS 29.7 AERONAUTICAL Aeronautical
RADIONAVIGATION LAND MOBILE Radionavigation
FIXED

29.8 RADIONAVIGATION Mobile

MOBILE

FIXED SATELLITE (E-S) MOBILE SATELLITE (E-S) AERONAUTICAL (Radio Beacons) 285
2900 FIXED
ISM – 27.12 ± .163 MHz

29.9 29.89 MOBILE (R) MARITIME Aeronautical

MARITIME FIXED MOBILE RADIONAVIGATION Radionavigation
FIXED SATELLITE (E-S) MOBILE SATELLITE (E-S) Radiolocation 29.91 (RADIO BEACONS) (Radio Beacons)
300 GHz
3 MHz

RADIONAVIGATION FIXED
300 kHz

3 GHz
300 MHz

3000 300
30 MHz

30 GHz

300.0 30.0 300 30.0 3000

tailored recovery that can be implemented in real CR settings.

emphasis is on practical low rate acquisition schemes and
sensing and joint carrier frequency and DOA estimation. Our
including cyclostationary detection, collaborative spectrum
aspects of spectrum sensing performed on low rate samples,
methods for multiband signals and then consider different
CR outlined above. We first review sub-Nyquist sampling
This article focuses on the spectrum sensing challenges for
frequencies and DOAs of the received transmissions, from the
We review several algorithms that jointly estimate the carrier
be adapted to the joint carrier and DOA estimation problem.
or compressed sensing (CS) [41] based approaches, may then
recovery techniques such as MUSIC [38], [39], ESPRIT [40]
3

Fig. 2. United States frequency allocation chart (courtesy of the U.S. Department of Commerce, National Telecommunications and Information Administration,
poses inherent difficulties in introducing new technologies that
are pre-allocated to one or even several licensed PUs. This
the VHF, UHF and SHF frequency ranges (0.003–30GHz)
allocation chart. As can be seen, all of the bands spanning
ment for radio spectrum. Figure 2 shows the US frequency
(ECC) cooperate on aspects related to the regulatory environ-
stitute (ETSI) and the Electronic Communications Committee
In Europe, the European Telecommunications Standards In-
 4

In particular, we examine the analog to digital interface of sub-Nyquist methods, that theoretically achieve the minimal
CRs, as opposed to the review in [42], which is concerned sampling rate. Two main limitations of sub-Nyquist sampling
with the digital implications of compressive spectrum sensing. are noise enhancement [29], which is discussed later on and
This allows to demonstrate the realization of the theoretical may be partially overcome by techniques reviewed below
concepts on hardware prototypes. We focus in particular on such as cyclostationary detection and collaborative spectrum
the implementation of one sampling scheme reviewed here, sensing, and the inherent assumption that the channel is sparse
the MWC, and show how the same low rate samples can be and contains spectral holes.
used in the different extensions of spectrum sensing described
above. A. Multitone Model and the Random Demodulator
We note that CRs are faced with additional fundamental
difficulties, such as the hidden-node problem and potential Tropp et al. [22], [23] consider a discrete multitone model
interference from secondary users to existing primary com- and suggest sampling using the random demodulator, depicted
munication links. These are beyond the scope of this review in Fig. 5. Multitone functions are composed of K active tones
but need to be addressed as well for CRs to become practical. spread over a bandwidth W , such that
f (t) = ∑ bω e−2πiωt , t ∈ [0, 1). (6)
ω∈Ω
III. S UB -N YQUIST S AMPLING FOR CR
Here, Ω is composed of K normalized frequencies, or tones,
CR receivers sense signals composed of several transmis-
that are a subset of the integers in the interval [−W /2, W /2],
sions with unknown support, spread over a wide spectrum.
and bω , for ω ∈ Ω, are a set of complex-valued amplitudes.
Such sparse wideband signals belong to the so-called multi-
The number of active tones K is assumed to be much smaller
band model [25], [26], [43]. An example of a multiband signal
than the bandwidth W . The goal is to recover both the tones
x(t) with K bands is illustrated in Fig. 4, with individual
ω and the corresponding amplitudes bω .
bandwidths not greater than B, centered around unknown
carrier frequencies ∣fi ∣ ≤ fNyq /2, where fNyq denotes the
signals’ Nyquist rate and i indexes the transmissions. Note
that, for real-valued signals, K is an even integer due to
spectral conjugate symmetry and the number of transmissions
is Nsig = K/2.

𝐵
~

−𝑓Nyq -𝑓3 -𝑓2 −𝑓1 0 𝑓1 𝑓2 𝑓3 𝑓Nyq

2 2
Fig. 5. Block diagram for the random demodulator, including a random
number generator, a mixer, an accumulator and a sampler [23].
Fig. 4. Multiband model with K = 6 bands. Each band does not exceed
the bandwidth B and is modulated by an unknown carrier frequency ∣fi ∣ ≤
fNyq /2, for i = 1, 2, 3.
To sample the signal f (t), it is first modulated by a high-rate
sequence pc (t) created by a pseudo-random number generator.
When the frequency support of x(t) is known, classic sam- It is then integrated and sampled at a low rate, as shown
pling methods such as demodulation, undersampling ADCs in Fig. 5. The random sequence used for demodulation is a
and interleaved ADCs (see [43], [46] and references therein) square wave, which alternates between the levels ±1 with equal
reduce the sampling rate below Nyquist. Here, since the probability. The K tones present in f (t) are thus aliased by
frequency location of the transmissions are unknown, classic the pseudorandom sequence. The resulting demodulated signal
processing first samples x(t) at its Nyquist rate fNyq , which y(t) = f (t)pc (t) is then integrated over a period 1/R and
may be prohibitively high. Sensing such wideband signals may sampled at the low rate R. This integrate-and-dump approach
be performed by splitting them into several frequency bands results in the following samples
and processing each band separately. This can be done either (m+1)/R
in series or in parallel. However, this increases the latency in ym = R ∫ y(t)dt, m = 0, 1, . . . , R − 1. (7)
m/R
the first case and the hardware complexity in the second.
To overcome the sampling rate bottleneck, several blind sub- The samples ym acquired by the random demodulator can
Nyquist sampling and recovery schemes have been proposed be written as a linear combination of the W × 1 sparse
that exploit the signal’s structure and in particular its sparsity amplitude vector b that contains the coefficients bω at the
in the frequency domain, but do not require knowledge of the corresponding locations ω [22], [23]. In matrix form, we write
carrier frequencies. It has been shown in [25] that the minimal
y = Φb, (8)
sampling rate for perfect blind recovery in multiband settings
is twice the Landau rate [47], that is twice the occupied where y is the vector of size R that contains the samples
bandwidth. This rate can be orders of magnitude lower than ym and Φ is the known sampling matrix that describes the
Nyquist. In the remainder of this section, we survey several overall action of the system on the vector of amplitudes b,
 5

IEEE 802.22 S TANDARD FOR WRAN to the operating band during designated quiet times, in order
During the past decade, the use of analog TV radio bands has to make sure the incumbent user does not wake up and require
been slowly decaying, mainly due to the introduction of digital its use. The media access control (MAC), situated above the
TV broadcasts based on new transmitting media such as cable PHY layer in the network hierarchy, is responsible for CR
and satellite. The once crowded VHF and UHF frequency communication management. Its specifications include fast
range (between 54 and 860 MHz) is less occupied today, and dynamic adaptation to changes in the environment by
particularly in remote rural areas. Concurrently, a new need constantly sensing the spectrum. The requirements for each of
has emerged in these areas for fast wireless data networks with the different network layers are particularly demanding when
long range capabilities, up to 30 km. For these operating range compared to other RF standards from the 802.xx family.
requirements, the VHF/UHF radio spectrum, able to traverse To conclude, the 802.22 WRAN standard holds great promise
these distances with relatively low transmitting power, is a in delivering fast broadband connections to remote rural areas,
natural choice. Combining the need for available frequencies with broadcast ranges of up to 30 km, by smartly sharing the
in this range with the fact that it is mostly underutilized today, already allocated VHF and UHF spectrum. Spectrum coexis-
has laid the foundation for designing efficient spectrum sharing tence is made possible by CR tools, which require advanced
techniques, including CR. technologies. These introduce tremendous challenges in terms
IEEE 802.22 is the first international standard especially of the necessary hardware to meet such harsh demands, while
designed to achieve this goal. It describes a wireless regional preserving small form factors and low power levels. Sub-
area network (WRAN) relying on new CR technologies, that Nyquist methods for spectral sensing and reconstruction offer
exploits the vacant white space in the VHF/UHF range, while a way to alleviate part of this burden by exploiting the signals’
still reserved to licensed TV bands. Managing the spectrum spectral structure, effectively making CR technology more
and transmitting broadband communications only in vacant accessible.
frequency slots enables fast internet access in hard-to-reach
Backbone
areas with low population density, and therefore has high Network
potential for extensive worldwide usage ~ 30 km

To exploit analog TV bands’ white spaces, an opportunistic

approach is selected. Both base stations (BS) and clients,
referred to as customer-premises equipment (CPE), are spread ~ 100 km
CPE
apart geographically, as seen in Fig. 3. In order to enable a
true CR network, the standard defines necessary capabilities BS CPE

from both the BSs and CPEs to enable cognitive sensing and
CPE
adapt to possibly rapid changes incurred by incumbent users. CPE BS
The physical (PHY) layer is the lowest network layer. It CPE

contains the circuitry used to transmit and receive analog

CPE
communication signals. For implementing CR technologies,
the PHY has to rapidly adapt to spectral changes, and be both
agile and flexible enough to jump between carrier frequencies Fig. 3. CR network illustration, including BSs and CPEs.
without losing information. It is required to constantly listen

namely demodulation and filtering (see [23] for more details). prohibitive and very sensitive to the grid choice (see [46] for
Capitalizing on the sparsity of the vector b, the amplitudes a detailed analysis). Furthermore, the time-domain approach
bω and their respective locations ω can be recovered from the precludes processing at a low rate, even for multitone inputs
low rate samples y using CS [41] techniques such as those since interpolation to the Nyquist rate is an essential ingredient
discussed in “Compressed Sensing Recovery”, in turn allowing of signal reconstruction. In terms of hardware and practical
for the recovery of f (t). The minimal required number of implementation, the random demodulator requires accurate
samples R for perfect recovery of f (t) in noiseless settings modulation by a periodic square mixing sequence and accurate
is 2K [41]. integration, which may be challenging when using analog
The random demodulator is one of the pioneer and innova- signal generators, mixers and filters.
tive attempts to extend the inherently discrete and finite CS In contrast to the random demodulator, which adopts a
theory to analog signals. However, truly analog signals, as discrete multitone model, the rest of the approaches we focus
those we consider here, require a prohibitively large number of on treat the analog multiband model, illustrated in Fig. 4,
harmonics to approximate them well within the discrete model. which is of interest in the context of CR.
When attempting to approximate signals such as those from
the multiband model, the number of tones W is of the order of B. Multi-Rate Sampling
the Nyquist rate and the number of samples R is a multiple of An alternative sampling approach is based on the syn-
KB. This in turn renders the reconstruction computationally chronous multi-rate sampling (SMRS) [24] scheme, which
 6

C OMPRESSED S ENSING R ECOVERY where the absolute value is computed element-wise. The
Compressed sensing (CS) [41], [43] is a framework for simul- residual is obtained by subtracting the contribution of a partial
taneous sensing and compression of finite-dimensional vectors, estimate x̂` of the signal at the `th iteration, from z, as follows
which relies on linear dimensionality reduction. In particular,
r = z − Ax̂` . (4)
the field of CS focuses on the following recovery problem
z = Ax, (1) Once the support set is updated by adding the index i, the
coefficients of x̂` over the support set are updated, so as to
where x is a N × 1 sparse vector, namely with few non zero minimize the residual error.
entries, and z is a vector of measurements of size M < N . CS Other greedy techniques include thresholding algorithms. We
provides recovery conditions and algorithms to reconstruct x focus here on the iterative hard thresholding (IHT) method
from the low-dimensional vector z. proposed in [45]. Starting from an initial estimate x̂0 = 0,
The different spectrum sensing applications described in this the algorithm iterates a gradient descent step with step size µ
paper deal primarily with analog signals and sampling tech- followed by hard thresholding, i.e.
niques, while CS inherently defines a digital framework. We
will discuss how the analog approaches of low rate sampling x̂` = T (x̂`−1 + µAH (z − Ax̂`−1 ), k) (5)
use CS as a tool for recovery and adapt it to the analog
setting. To that end, we describe here two CS greedy recovery until a convergence criterion is met. Here T (x, k) denotes a
algorithms, that solve the optimization problem thresholding operator on x that sets all but the k entries of x
with the largest magnitudes to zero, and k is the sparsity level
x̂ = arg min ∣∣x∣∣0 s.t. z = Ax, (2) of x, assumed to be known.
x
These two greedy algorithms and other CS recovery techniques
where ∣∣ ⋅ ∣∣0 denotes the `0 -norm. The first algorithm we
can be adapted to further settings, such as multiple measure-
consider belongs to the family of matching pursuit (MP)
ment vectors (MMV), where the measurements z and sparse
methods [44]. The orthogonal matching pursuit (OMP) algo-
objective x become matrices, infinite measurement vectors
rithm iteratively proceeds by finding the column of A most
(IMV), block sparsity and more, as we will partially discuss
correlated to the signal residual r,
in the paper. Further details on CS recovery conditions and
i = arg max ∣AH r∣, (3) techniques can be found in [41], [43].

has been proposed in the context of electro-optical systems to account for the absence of signals of interest in all the fre-
undersample multiband signals. The SMRS samples the input quencies that are down-converted to that baseband frequency,
signal at P different sampling rates Fi , each of which is an which allows to reduce the number of sampling channels.
integer multiple of a basic sampling rate ∆f . This procedure This assumption does not hold in the case where two or
aliases the signal with different aliasing intervals, as illustrated more frequency components cancel each other due to aliasing,
in Fig. 6. The Fourier transform of the undersampled signals is which happens with probability zero. Once the corresponding
then related to the original signal through an underdetermined components are eliminated from (9), the reduced system is
system of linear equations, inverted using the Moore-Penrose pseudo-inverse to recover
x(f ).
z(f ) = Qx(f ). (9)

Here, x(f ) contains frequency slices of size ∆f of the There are several drawbacks to the SMRS that limit its per-
original signal x(t) and z(f ) is composed of the Fourier formance and potential implementation. First, the discretiza-
transform of the sampled signal. Each channel contributes tion process affects the SNR since some of the samples are
Mi = Fi /∆f equations to the system (9), which concatenates thrown out. Furthermore, spectral components down-converted
the observation vector of all the channels. The measurement to off the grid frequencies are missed. In addition, the first
matrix Q has exactly P non-zero elements in each column, recovery approach requires a large number of sampling chan-
that correspond to the locations of the spectral replica in each nels, proportional to the number of active bands K, whereas
channel baseband [0, Fi ]. the reduction procedure does not ensure a unique solution
This approach assumes that either the signal or the sampling and the inversion problem is ill-posed in many cases. Another
time window are finite. The continuous variable f is then difficulty is that, in practice, synchronization between channels
discretized to a frequency resolution of ∆f . Since x(t) is sampling at different rates is challenging. Finally, this scheme
sparse in the frequency domain, the vector x(f ) is sparse samples wideband signals using low rate samplers. Practical
and can be recovered from (9) using CS techniques, for each ADCs introduce an inherent bandwidth limitation, modeled by
discrete frequency f . An alternative recovery method, referred an anti-aliasing low pass filter (LPF) with cut-off frequency
to as the reduction procedure and illustrating in Fig. 6, consists determined by the sampling rate, which distorts the samples.
of detecting baseband frequencies in which there is no signal, Thus, the SMRS implementation requires low rate samplers
by observing the samples. These frequencies are assumed to with large analog bandwidth.
 7

Original signals where the cosets ci are ordered integers so that 0 ≤ c1 < c2 <
⋯ < cM < N . A possible implementation of the sampling
(a) sequences (10) is depicted in Fig. 8. The building blocks are
M uniform samplers at rate 1/N TNyq , where the ith sampler
is shifted by ci TNyq from the origin. When sampling at the
Nyquist rate, M = N and ci = (i − 1).
Signals sampled at rate 𝑓1
(b) Time Shifts 𝑡 = 𝑛𝑁𝑇Nyq

Δ𝑡 = 𝑐1 𝑇Nyq 𝑧1 𝑛
𝑓1 2𝑓1 3𝑓1

Signals sampled at rate 𝑓2

⋅⋅⋅
𝑥 𝑡

(c) 𝑡 = 𝑛𝑁𝑇Nyq

𝑓2 2𝑓2 Δ𝑡 = 𝑐𝑀 𝑇Nyq 𝑧𝑀 𝑛

Possible occupied support

Fig. 8. Schematic implementation of multicoset sampling. The input signal
x(t) is inserted into the multicoset sampler that splits the signal into M
(d) branches, and delays each one by a fixed coefficient ci TNyq . Every branch
is then sampled at the low rate 1/(N TNyq ), and then digitally processed to
perform spectrum sensing and signal reconstruction.
Fig. 6. Action of the SMRS on a multiband signal: (a) The input signal with
K = 2 bands, (b) signals sampled at rate F1 in channel 1, (c) signals sampled
at rate F2 in channel 2, and (d) possible support which is the intersection of
The samples in the Fourier domain can then be written as
the supports in channel 1 and 2 [24]. linear combinations of spectrum slices of x(t), such that [25]
z(f ) = Ax(f ), f ∈ Fs . (11)

C. Multicoset sampling Here, x(f ) denotes the spectrum slices of x(t), fs = ≥ 1

N TNyq
B is the sampling rate of each channel and Fs = [−fs /2, fs /2].
A popular sampling scheme for sampling wideband signals This relation is illustrated in Fig. 9. The M × N sampling
at the Nyquist rate is multicoset or interleaved ADCs [25], matrix A is determined by the selected cosets ci . The recovery
[46], [48] in which several channels are used, each operating processing described below is performed in the time domain,
at a lower rate. We now discuss how such systems can be used where we can write
in the sub-Nyquist regime.
Multicoset sampling may be described as the selection of z[n] = Ax[n], n ∈ Z, (12)
certain samples from the uniform Nyquist grid, as shown
The vector z[n] collects the measurements at t = n/fs and
in Fig. 7, where TNyq = 1/fNyq denotes the Nyquist period.
x[n] contains the samples sequences corresponding to the
More precisely, the uniform grid is divided into blocks of
spectrum slices of x(t).
The goal can then be stated as the recovery of x[n] from
𝑓 𝑡 𝑁𝑇Nyq the samples z[n]. The system (12) is underdetermined due
to the sub-Nyquist setup and known as infinite measurement
𝑇Nyq 𝑐1 vectors (IMV) in the CS literature [41], [43]. With respect
to these two properties, the digital reconstruction algorithm
𝑐3
𝑐𝑀 encompasses the following three stages [25] that we explain
𝑐2 in more detail below:
1) The continuous-to-finite (CTF) block constructs a finite
frame (or basis) from the samples.
𝑡 2) The support recovery formulates an optimization prob-
𝑀 out of 𝑁
active cosets lem whose solution’s support is identical to the support
S of x[n], that is the active slices.
Fig. 7. Illustration of multicoset sampling. 3) The signal is digitally recovered by reducing (12) to the
support of x[n].
The recovery of x[n] for every n independently is inef-
N consecutive samples, from which only M < N are kept.
ficient and not robust to noise. Instead, the CTF method,
Mathematically, the ith sampling sequence is defined as
developed in [25], exploits the fact that the bands occupy
x(nTNyq ), n = mN + ci , m ∈ Z continuous spectral intervals so that x[n] are jointly sparse,
xci [n] = { (10) that is they have the same spectral support S over time. The
0, otherwise,
 8

Although this sampling scheme seems relatively simple

𝑋(𝑓) – Spectrum of 𝑥(𝑡)
and straightforward, it suffers from several practical draw-
backs [46]. First, as in the SMRS approach, multicoset sam-
pling requires low rate ADCs with large analog bandwidth.
~

~
𝑙1 𝑓𝑝 𝑙2 𝑓𝑝 𝑙3 𝑓𝑝 𝑓Nyq Another issue arises from the time shift elements, since main-
0
2 taining accurate time delays between the ADCs on the order
of the Nyquist interval TNyq is difficult. Finally, the number
×𝑎 ×𝑎 ×𝑎 ×𝑎 ×𝑎 ×𝑎
𝑖 1 𝑙1 𝑖 1 𝑙2 𝑖 1 𝑙3 𝑖 2 𝑙1 𝑖 2 𝑙2 𝑖 2 𝑙3 of channels M required for recovery of the active bands
can be prohibitively high. The MWC, presented in the next
section, uses similar recovery techniques while overcoming
+ + these practical sampling issues.
𝑖1 Channel 𝑖2 Channel

D. MWC sampling
𝑓𝑝 𝑓𝑝 𝑓𝑝 𝑓𝑝
−
2 2
−
2 2
The MWC [26] exploits the blind recovery ideas developed
in [25] and combines them with the advantages of analog
Fig. 9. The spectrum slices of the input signal x(f ) are shown here to be RF demodulation. To circumvent the analog bandwidth issue
multiplied by the coefficients ail of the sensing matrix A, resulting in the in the ADCs, an RF front-end mixes the input signal x(t)
measurements zi for the ith channel. Note that in multicoset sampling, only with periodic waveforms. This operation imitates the effect
the slices’ complex phase is modified by the coefficients ail . In the MWC
sampling described below, both the phases and amplitudes are affected. of delayed undersampling used in the multicoset scheme and
results in folding the spectrum to baseband with different
weights for each frequency interval. These characteristics of
the MWC enable practical hardware implementation, which
CTF produces a finite system of equations, called multiple
will be described later on.
measurement vectors (MMV) [41], [43] from the infinite
More specifically, the MWC is composed of M parallel
number of linear systems described by (12). The samples are
channels. In each channel, x(t) is multiplied by a periodic
first summed as
mixing function pi (t) with period Tp = 1/fp and Fourier
Q = ∑ z[n]zH [n], (13) expansion
n ∞ 2π
jT lt
H pi (t) = ∑ ail e p . (17)
and then decomposed to a frame V such that Q = VV . l=−∞
Clearly, there are many possible ways to select V. One option
The mixing process aliases the spectrum, such that each band
is to construct it by performing an eigendecomposition of Q
appears in baseband. The signal then goes through a LPF with
and choosing V as the matrix of eigenvectors corresponding to
cut-off frequency fs /2 and is sampled at rate fs ≥ fp . The
the non zero eigenvalues. The finite dimensional MMV system
analog mixture boils down to the same mathematical relation
V = AU, (14) between the samples and the N = fNyq /fs frequency slices of
x(t) as in multicoset sampling, namely (11) in frequency and
is then solved for the sparsest matrix U with minimal number (12) in time, as shown in Fig. 9. Here, the M × N sampling
of non-identically zero rows using CS techniques [41], [43]. matrix A contains the Fourier coefficients ail of the periodic
The key observation of this recovery strategy is that the indices mixing functions. The recovery conditions and algorithm are
of the non zero rows of U coincide with the active spectrum identical to those described for multicoset sampling.
slices of z[n] [25]. These indices are referred to as the support Choosing the channels’ sampling rate fs to be equal to
of z[n] and are denoted by S. the mixing rate fp results in a similar configuration as the
Once the support S is known, x[n] is recovered by reducing multicoset scheme in terms of the number of channels. In
the system of equations (12) to S. The resulting matrix AS , this case, the minimal number of channels required for the
that contains the columns of A corresponding to S, is then recovery of K bands is 2K. The number of branches dictates
inverted as follows the total number of hardware devices and thus governs the
implementation complexity. Reducing the number of channels
xS [n] = A†S z[n]. (15)
is a crucial challenge for practical implementation of a CR
Here, xS [n] denotes the vector x[n] reduced to its support. receiver. The MWC architecture presents an interesting flexi-
The remaining entries of x[n] are equal to zero. bility that permits trading channels for sampling rate, allowing
The overall sampling rate of the multicoset system is to drastically reduce the number of channels, even down to a
single channel.
M Consider a configuration where fs = qfp , with odd q. In
ftot = M fs =
fNyq . (16)
N this case, the ith physical channel provides q equations over
The minimal number of channels is dictated by CS results [41] Fp = [−fp /2, fp /2], as illustrated in Fig. 10. Conceptually,
such that M ≥ 2K with fs ≥ B per channel. The sampling M physical channels sampling at rate fs = qfp are then
rate can thus be as low as 2KB, which is twice the Landau equivalent to M q channels sampling at fs = fp . The number
rate [47]. of channels is thus reduced at the expense of higher sampling
 9

This configuration has three main advantages over the stan-

dard MWC. First, it allows for a simpler design of the mixing
functions which can be identical in all sensors. The only
requirement on p(t), besides being periodic with period Tp , is
5𝑓𝑝 3𝑓𝑝 𝑓𝑝 0 𝑓𝑝 3𝑓𝑝 5𝑓𝑝
−
2
−
2
−
2 2 2 2 that none of its Fourier series coefficients within the signal’s
(a) Spectrum of 𝑧ǁ𝑖 𝑛 for 𝑓𝑠 = 5𝑓𝑝 .
Nyquist bandwidth is zero. Second, the ULA based system
outperforms the MWC in terms of recovery performance in
low SNR regimes. Since all the MWC channels belong to the
same sensor, they are all affected by the same additive sensor
noise. In the ULA architecture, each channel belongs to a dif-
𝑓𝑝 𝑓𝑝 𝑓𝑝 𝑓𝑝 𝑓𝑝 𝑓𝑝 𝑓𝑝 𝑓𝑝 𝑓𝑝 𝑓𝑝
−
2
0
2
−
2
0
2
−
2
0
2
−
2
0
2
−
2
0
2
ferent sensor with uncorrelated sensor noise between channels.
This alternative architecture benefits from the same flexibility
(b) Spectrum of 𝑧𝑖,𝑗 𝑛 after expansion.
as the standard MWC in terms of collapsing the channels,
which translates here into reducing the antennas. This leads
Fig. 10. Illustration of the expander configuration for q = 5. (a) Spectrum
of the output z̃i [n] of the physical ith channel, (b) spectrum of the samples to a trade-off between hardware complexity, governed by the
zi,j [n] of the q = 5 equivalent virtual channels, for j = 1, . . . , 5, after digital number of antennas, and SNR. Finally, as will be shown later
expansion. on, the modified system can be easily extended to enable joint
spectrum sensing and DOA estimation.
rate fs in each channel and additional digital processing. Similarly to the previous sampling schemes, the samples
The output of each of the M physical channels is digitally z(f ) can be expressed as a linear transformation of the
demodulated and filtered to produce samples that would result unknown vector of slices x(f ), such that
from M q equivalent virtual branches. This happens in the so- z(f ) = Ax(f ), f ∈ Fs . (19)
called expander module, directly after the sampling stage and
before the digital processing described above, in the context Here, x(f ) is a non sparse vector that contains cyclic shifted,
of multicoset sampling. At its brink, this strategy allows to scaled and sampled versions of the active bands, as shown
collapse a system with M channels to a single branch with in Fig. 13. In contrast to the previous schemes, in this
sampling rate fs = M fp (further details can be found in [26], configuration, the matrix A, defined by
[49], [50]).
⎛e
j2πf1 τ1
⋯ ej2πfN τ1 ⎞
The MWC sampling and recovery processes are illustrated
⎜ ⋮ ⋮ ⎟
in Fig. 11. This approach results in a hardware-efficient sub- A=⎜ ⎟, (20)
⎜ ⎟
Nyquist sampling method that does not suffer from the practi- ⎝ej2πf1 τM j2πfN τM ⎠
⋯ e
cal limitations described in previous sections, in particular, the
analog bandwidth limitation of low rate ADCs. In addition, the depends on the unknown carrier frequencies.
number of MWC channels can be drastically reduced below Two approaches are presented in [53] to recover the carrier
2K to as few as one, using higher sampling rate fs in each frequencies of the transmissions composing the input signal.
channel and additional digital processing. This tremendously The first is based on CS algorithms and assumes that the
reduces the burden on hardware implementation. However, the carriers lie on a predefined grid. In this case, the resulting
choice of appropriate periodic functions pi (t) to ensure correct sensing matrix, which extends A with respect to the grid, is
recovery is challenging. Some guidelines are provided in [43], known and the expanded vector x(f ) is sparse. This leads to
[51], [52]. a similar system as (11) or (12) which can be solved using
the recovery paradigm from [25], described in the context
E. Uniform Linear Array based MWC of multicoset sampling. In the second technique, the grid
An alternative sensing configuration, composed of a uniform assumption is dropped and ESPRIT [40] is used to estimate the
linear array (ULA) and relying on the sampling paradigm of carrier frequencies. This approach first computes the sample
the MWC, is presented in [53]. The sensing system consists covariance of the measurements z[n] as
of a ULA composed of M sensors, with two adjacent sensors R = ∑ z[n]zH [n], (21)
separated by a distance d, such that d < c/(∣ cos(θ)∣fNyq ), n
where c is the speed of light and θ is the angle of arrival
and performs a singular value decomposition (SVD). The non
(AOA) of the signal x(t). This system, illustrated in Fig. 12,
zero singular values correspond to the signal’s subspace and
capitalizes on the different accumulated phases of the input
the carrier frequencies are then estimated from these. Once
signal between sensors, given by ej2πfi τm , where
the carriers are recovered, the signal itself is reconstructed by
dm inverting the sampling matrix A in (19).
τm = cos(θ) (18)
c The minimal number of sensors required by both reconstruc-
is the delay at the mth sensor with respect to the first one. Each tion methods in noiseless settings is M = 2K, with each sensor
sensor implements one channel of the MWC, that is the input sampling at the minimal rate of fs = B to allow for perfect
signal is mixed with a periodic function, low-pass filtered and signal recovery [53]. The proposed system thus achieves the
then sampled at a low rate. minimal sampling rate 2KB derived in [25]. We note that the
 10

𝑋(𝑓)
𝑓Nyq /2 𝑓Nyq /2

~
~
𝐱 (𝑓)
𝑝1 (𝑡) 𝐻(𝑓) 𝑧1 (𝑓)
𝑛𝑇𝑠
−𝑓𝑝 𝑓𝑝 −𝑓𝑝 𝑓𝑝
−𝑓𝑠 𝑓𝑠

Expander 𝑀 Support Signal

𝑥(𝑡) LNA + 𝑀/𝑞 Recovery Reconstruction
Carrier 𝐳 𝑛 = 𝐀𝐱 𝑛 −𝑓𝑝 𝑓𝑝
Filters Detection
𝑝𝑀/𝑞 (𝑡)
𝐻(𝑓) 𝑞 𝐱 𝑛 = 𝐀†𝑠 𝐳[𝑛]
𝑛𝑇𝑠
−𝑓𝑝 𝑓𝑝 −𝑓𝑝 𝑓𝑝
−𝑓𝑠 𝑓𝑠 𝑧𝑚 (𝑓)

Analog Front End - MWC Low Rate A/D Digital

Fig. 11. Schematic implementation of the MWC analog sampling front-end and digital signal recovery from low rate samples.

expander strategy proposed in the context of the MWC can be

applied in this configuration as well.
𝑥 𝑡 source

𝜃 𝑥(𝑡 − 𝜏1 )
𝐻(𝑓)
𝑛𝑇𝑠
IV. MWC H ARDWARE
−𝑓𝑠 𝑓𝑠

𝑀 channels
𝑥(𝑡 − 𝜏𝑀 ) A. MWC Prototype
𝐻(𝑓)
𝑛𝑇𝑠

−𝑓𝑠 𝑓𝑠 One of the main aspects that distinguish the sub-Nyquist

MWC from other sampling schemes is its practical implemen-
𝑝(𝑡)
tation [49], proving the feasibility of sub-Nyquist sampling
even under distorting effects of analog components and phys-
Fig. 12. ULA configuration with M sensors, with distance d between two ical phenomena. A hardware prototype, shown in Fig.14, was
adjacent sensors. Each sensor includes an analog front-end composed of a
mixer with the same periodic function p(t), a LPF and a sampler, at rate fs .
developed and built according to the block diagram in Fig. 11.
In particular, the system receives an input signal with Nyquist
rate of 6 GHz and spectral occupancy of up to 200 MHz, and
samples at an effective rate of 480 MHz, that is only 8% of
the Nyquist rate and 2.4 times the Landau rate. This rate
constitutes a relatively small oversampling factor of 20% with
respect to the theoretical lower sampling bound. This section
describes the different elements of the hardware prototype,
shown in Fig. 14, explaining the various considerations that
were taken into account when implementing the theoretical
0 0
concepts on actual analog components.
At the heart of the system lies the proprietary MWC
board [49] that implements the sub-Nyquist analog front-
0 0 end. The card uses a high speed 1-to-4 analog splitter that
duplicates the wideband signal to M = 4 channels, with an
expansion factor of q = 5, yielding M q = 20 virtual channels
after digital expansion. Then, an analog preprocessing step,
0 0 composed of preliminary equalization, impedance corrections
(a) (b) and gain adjustments, aims at maintaining the dynamic range
and fidelity of the input in each channel. Indeed, the signal and
mixing sequences must be amplified to specific levels before
Fig. 13. The left pane shows the original source signals at baseband (before
modulation). The right pane presents the output signals at baseband x(f ) entering the analog mixers to ensure proper behavior emulating
after modulation, mixing, filtering and sampling. mathematical multiplication with the mixing sequences. The
entire analog path of the multiband input signal can be seen
in Fig. 15.
 11

Fig. 14. MWC CR system prototype: (a) vector signal Generators, (b) FPGA mixing sequences generator, (c) MWC analog front-end board, (d) RF combiner,
(e) spectrum analyzer, (f) ADC and DSP.

Value Notes
The mixing sequences that modulate the signal play an
fs 120 MHz(q + 1) fp - Sampling rate
fp 20 MHz 1/Tp
essential part in signal recovery. They must have low cross-
q 5 Expansion factor correlations with each other, while spanning a large bandwidth
M /q 4 # Hardware Channels determined by the Nyquist rate of the input signal, and yet be
fmax 3 GHz fmax = fNyq /2 easy enough to generate with relatively cheap, off-the-shelf
B Bandwidth on each carrier
18.5 MHz
Number of ±1 intervals in each
hardware. The sequences pi (t), for i = 1, . . . , 4, are chosen as
Mp 305 truncated versions of Gold Codes [54], which are commonly
period of pi (t)
TABLE I used in telecommunication (CDMA) and satellite navigation
T HE MWC PARAMETERS USED IN THE SETUP OF F IG . 14.
(GPS). Mixing sequences based on Gold codes were found to
give good results in the MWC system [55], primarily due to
their small cross-correlations.
Non-Uniform Noise & LPF – Noise Quantization Noise Since Gold codes are binary, the mixing sequences are
Freq. Response Harmonics & Phase Shift + Phase Noise
restricted to alternating ±1 values. This fact allows to digitally
Splitter +
Pre-Processing Mixer LPF
generate the sequences on a dedicated FPGA. Alternatively,
AMP 𝑛𝑇𝑠 𝑧𝑖 [𝑛]
𝑥(𝑡) AMP ATT
Digital they can be implemented on a small chip with very low power
Expander
and complexity. The added benefit of producing the mixing se-
𝑝𝑖 (𝑡) Sync Trigger + CLK quences on such a platform is that the entire sampling scheme
Unknown Splitter
Phase Noise
𝑛𝑇𝑠 may be synchronized and triggered using the same FPGA
Clock Delta
with minimally added phase noise and jitter, keeping a closed
synchronization loop with the samplers and mixers. A XiLinX
Fig. 15. Hardware RF chain detailed schematics, including amplifiers, VC707 FPGA acts as the central timing unit of the entire sub-
attenuators, filters, mixers, samplers and synchronization signals required for Nyquist CR setup by generating the mixing sequences and
precise and accurate operation. The distortions induced by each component
are indicated as well.
the synchronization signals required for successful operation.
It is crucial that both the mixing period Tp and the low rate
samplers operating at (q+1)fp (due to intended oversampling)
The modulated signal next passes through an analog anti- are fully synchronized, in order to ensure correct modeling of
aliasing LPF. The anti-aliasing filter must be characterized by the entire system, and consequently guarantee accurate support
both an almost linear phase response in the pass band between detection and signal reconstruction.
0 to 50 MHz and an attenuation of more than 20 dB at fs /2 = The digital back-end is implemented using a National In-
60 MHz. A Chebyshev LPF of 7th order with cut-off frequency struments PXIe-1065 computer with DC coupled ADC. Since
(−3 dB) of 50 MHz was chosen for the implementation. After the digital processing is performed at the low rate fs , very low
impedance and gain corrections, the signal now has a spectral computational load is required in order to achieve real time
content limited to 50 MHz, that contains a linear combination recovery. MATLAB® and LabVIEW® environments are used
of the occupied bands with different amplitudes and phases, as for implementing the various digital operations and provide
seen in Fig. 9. Finally the low rate analog signal is sampled by an easy and flexible research platform for further experimen-
a National Instruments© ADC operating at 120 MHz, leading tations, as discussed in the next sections. The sampling matrix
to a total sampling rate of 480 MHz. A is computed once off-line, using the calibration process
 12

outlined in [50] and described in “Hardware Calibration of

the MWC”.

B. Support Recovery
The prototype is fed with RF signals composed of up to 5
carrier transmissions with an unknown total bandwidth occu-
pancy of up to 200 MHz, and Nyquist rate of 6 GHz. An RF
input x(t) is generated using vector signal generators (VSG),
each producing one modulated data channel with individual
bandwidth of up to 20 MHz. The input transmissions then go
through an RF combiner, resulting in a dynamic multiband
input signal. This allows to test the system’s ability to rapidly
sense the input spectrum and adapt to changes, as required by
modern CR standards (see “IEEE 802.22 Standard for WRAN”
for details). In addition, the described setup can simulate
more complex scenarios, including collaborative spectrum
Fig. 16. Screen shot from the MWC recovery software: low rate samples
sensing [56], [57], joint DOA estimation [53], cyclostationary acquired from one MWC channel at rate 120 MHz (top), digital reconstruction
based detection [58] and various modulation schemes such of the entire spectrum, performed from sub-Nyquist samples (middle), true
as PSK, OFDM and more, for verifying sub-Nyquist data input signal x(t) showed using a fast spectrum analyzer (bottom).
reconstruction capabilities.
Support recovery is digitally performed on the low rate
samples, as presented above in the context of multicoset
sampling. The prototype successfully recovers the support of
the transmitted bands, when SNR levels are above 15 dB,
as demonstrated in Fig. 16. More sophisticated detection
schemes, such as cyclostationary detection, allow to achieve
perfect support recovery from the same sub-Nyquist samples
in lower SNR regimes of 0 − 10 dB, as seen in Fig. 27, and
will be further discussed below.
The main advantage of the MWC is that sensing is per-
formed in real-time for the entire spectral range, even though
the operation is performed solely on sub-Nyquist samples,
which results in substantial savings in both computational
and memory complexity. In additional tests, it is shown that
the bandwidth occupied in each band can also be very small
without impeding the performance, as shown in Fig. 17, where
the support of signals with very low bandwidth (just 10%
occupancy within the 20 MHz band) is correctly detected. Fig. 17. The setup is identical to Fig. 16. In this case, the individual
transmissions have low bandwidth, highlighting the structure of the signal
when folding to baseband.
C. Signal Reconstruction
Once the support is recovered, the data is reconstructed
from the sub-Nyquist samples, by applying (15). This step is
performed in real-time, reconstructing the signal bands z[n] the user broadcasts text strings, that are then deciphered and
one sample at a time, with low complexity due to the small displayed on screen. There are no restrictions regarding the
dimensions of the matrix-vector multiplication. We note that modulation type, bandwidth or other parameters, since the
reconstruction does not require interpolation to the Nyquist baseband information is exactly reconstructed regardless of its
grid. The active transmissions are recovered at the low rate of respective content. Therefore, any digital modulation method,
20 MHz, corresponding to the bandwidth of the slices z(f ). as well as analog broadcasts, can be transmitted and deci-
The prototype’s digital recovery stage is further expanded phered without loss of information, by applying any desirable
to support decoding of common communication modulations, decoding scheme directly on the sub-Nyquist samples.
including BPSK, QPSK, QAM and OFDM. An example for By combining both spectrum sensing and signal reconstruc-
the decoding of three QPSK modulated bands is presented in tion, the MWC prototype serves as two separate communica-
Fig. 18, where the I/Q constellations are shown after recon- tion devices. The first is a state-of-the-art CR that performs
structing the original transmitted signals xS using (15), from real time spectrum sensing at sub-Nyquist rates, and the
their low-rate and aliased sampled signals zn . The I/Q constel- second is a receiver able to decode multiple data transmissions
lations of the baseband signals is displayed, each individually simultaneously, regardless of their carrier frequencies, while
decoded using a general QPSK decoder. In this example, adapting to spectral changes in real time. In cases where the
 13

twice the Landau rate may exceed the Nyquist rate. Second,
robustness to noise is increased due to the averaging performed
to estimate statistics. This is drastically improved in the case
of cyclostationary signals in the presence of stationary noise.
Indeed, exploiting cyclostationarity properties exhibited by
communication signals allows to separate them from stationary
noise, leading to better detection in low SNR regimes [72].
In this section, we first review power spectrum detection
techniques in stationary settings and then extend these to cyclic
spectrum detection of cyclostationary signals.

A. Power Spectrum Based Detection

In the statistical setting, the signal x(t) is modeled as
the sum of uncorrelated wide-sense stationary transmissions.
The stationarity assumption is key to further reducing the
Fig. 18. Demodulation, reconstruction and detection of Nsig = 3 inputs from
sub-Nyquist samples using the MWC CR prototype: I/Q phase diagrams, sampling rate. In frequency, stationarity is expressed by the
showing the modulation pattern of the transmitted bands after reconstruction absence of correlation between distinct frequency components.
from the low rate samples (top left), sub-Nyquist samples from an MWC Specifically, as shown in [73], the Fourier transform of a wide-
channel zi [n] in the Fourier domain (middle), signal sampled by an external
spectrum analyzer showing the entire bandwidth of 3 GHz (bottom), informa- sense stationary signal is a nonstationary white process, such
tion sent on each carrier, proving successful reconstruction (top right). that
E[X(f1 )X ∗ (f2 )] = Sx (f1 )δ(f1 − f2 ). (24)
Here, the power spectrum Sx (f ) of x(t) is the Fourier
support of the potential active transmissions is a priori known transform of its autocorrelation rx (τ ). Thus, obviously, the
(e.g. potential cellular carriers), the MWC may be used as support of Sx (f ) is identical to that of X(f ). In this case,
an RF demodulator that efficiently acquires several frequency we will see that the autocorrelation matrix of the N spectrum
bands simultaneously. frequency slices of x(t) contained in x(f ) is diagonal and
Other schemes would require a dedicated demodulation presents only N degrees of freedom, which is key to sampling
channel for each potentially active band. In this case, the rate reduction.
mixing sequences should be designed so that their Fourier co- Another intuitive interpretation can be given in the time
efficients are non zero only in the bands of interest, increasing domain. The autocorrelation of stationary signals rx (τ ) =
SNR, and the support recovery stage is not needed [59]. E [x(t)x(t − τ )] is a function only of the time lags τ and
not of the sampling times t. Therefore, the autocorrelation
V. S TATISTICS D ETECTION recovery capabilities of a sampling set is determined by the
In the previous section, we reviewed recent sub-Nyquist density of the associated difference set, namely the set that
sampling methods that reconstruct a multiband signal, such contains the time lags. The cardinality of the difference set,
as a CR signal, from low rate samples. However, the final which depends on the choice of sampling times, is, in the worst
goal of CRs often only requires detection and not necessarily case, equal to twice that of the original set. This corresponds
perfect reconstruction of the PUs’ transmissions. In this regard, for example to a uniform sampling set. As a consequence,
several works have proposed performing detection on second- when the sampling scheme is not tailored to power spectrum
order signal statistics, which share the same frequency support recovery, the minimal required sampling rate in the blind
as the original signal. In particular, power and cyclic spectra setting is the Landau rate [66]. However, for an appropriate
have been considered for stationary and cyclostationary [32] choice of sampling times [62], [74], the cardinality of the
(see “Cyclostationarity”) signals, respectively. Instead of re- difference set can be on the order of the square of that of the
covering the signal from the low rate samples, its statistics original sampling set. In this case, under certain conditions,
are reconstructed and its support is estimated [58], [60]–[71]. the density of the different set can be arbitrarily high even if
Recovering second-order statistics rather than the signal the density of the sampling set goes to zero. With appropriate
itself benefits from two main advantages. First, it allows to design, the autocorrelation or power spectrum may thus be
further reduce the sampling rate, as we will discuss below. estimated from samples with arbitrarily low average sampling
Intuitively, statistics have less degrees of freedom than the rate [60]–[62], [74]–[76] at the expense of increased latency.
signal itself, requiring less samples for their reconstruction. We first review power spectrum recovery techniques that
This follows from the assumption that the signal of interest is do not exploit any specific design. We then present methods
either stationary or cyclostationary. Going one step further, that further reduce the sampling rate by adapting the sampling
the sparsity constraint can even be removed in this case scheme to the purpose of autocorrelation or power spectrum
and the power/cyclic spectrum of non sparse signals can be estimation.
recovered from samples obtained below the Nyquist rate [58], 1) Power Spectrum Recovery: In this section, we first focus
[60], [61], [65], [66], [70]. This is useful for CRs operating on sampling with generic MWC or multicoset schemes without
in less sparse environments, in which the lower bound of specific design of the mixing sequences or cosets, respectively.
 14

H ARDWARE C ALIBRATION OF THE MWC prior knowledge on the mixing series pi (t) except for their
In the sampling system described above, the system matrix period length Tp .
A is theoretically known and contains the Fourier series Since our system is not time invariant (e.g. samplers), nor
coefficients of the mixing sequences, such that linear (e.g. mixers), one cannot find the entries of the system
Tp
matrix by simply measuring its response to an impulse. To
1 −j T
2π
lt circumvent this difficulty, the system’s response is investigated
Ail = cil = ∫ pi (t) e p dt . (22) for every frequency band of the spectrum by injecting known
Tp
0 sinusoidal inputs sequentially. In each iteration, the following
When calculating the matrix coefficients using (22), perfect input
support recovery and signal reconstruction is guaranteed both
theoretically and by numerical verification performed in soft- xl (t) = αl sin [2π (lfp + f0 ) t] , l ∈ [0, 1, . . . , L0 ] , (23)
ware simulations. However, in practice, analog and physical
distortions and imperfections affect the mixing and sampling is fed to the system. Here, f0 = 0.1fp was heuristically chosen
process, and some modeling assumptions, that describe the and the amplitudes αl increase with l to compensate for the
system matrix in theory, no longer hold. The main effects that attenuation of the Fourier coefficients of the mixing sequences
distort the transfer function are: at high frequencies. Every sine wave corresponds to a specific
● The mixing procedure introduces nonlinearities. In gen- spectral band and translates to a relevant column of the matrix
eral, mixers are intended to modulate narrowband signals A. The sub-Nyquist samples are then digitally processed to
with one sine carrier, as opposed to our mixing sequences estimate the system matrix coefficients, column by column.
that effectively contain over a hundred different harmon- Performance of the calibrated system is illustrated in Fig. 19.
ics. 1
● The analog filters have a non-flat frequency response, Support Recovery Rate
both in amplitude and phase. 0.8

● The actual design uses amplifiers and attenuators. These

components exhibit non-linear frequency response. 0.6
Calibrated Matrix
● The phase noise and jitter, due to variations in com- Theoretical Matrix
0.4
ponents, cables and clock deltas lead to unknown and
varying time shifts between the mixing and sampling 0.2
stages.
An accurate method for estimating the effective A is crucial 0
-10 0 10 20 30 40 50
for successful support recovery and signal reconstruction. SNR [dB]
An adaptive calibration scheme is proposed in [50]. The
calibration procedure estimates the elements of A with no Fig. 19. Hardware reconstruction success rate of the calibrated matrix versus
the theoretical matrix.

To recover Sx (f ) from the low rates samples z(f ) obtained Generic choices of the sampling parameters, either mixing
via multicoset sampling or the MWC [66], consider the sequences or cosets, which are only required to ensure that A
correlation matrix of the latter Rz (f ) = E[z(f )zH (f )]. From is full spark, are investigated in [66]. The Khatri-Rao product
(11), Rz (f ) can be related to correlations between the slices (Ā ⊙ A) is full spark as well if M > N /2, that is the number
x(f ), that is Rx (f ) = E[x(f )xH (f )], as follows of rows of A is at least half the number of slices N . The
minimal sampling rate to recover rx (f ), and consequently
Rz (f ) = ARx (f )AH , f ∈ Fs . (25)
Sx (f ), from rz (f ) in (27) is thus equal to the Landau rate
From (24), the correlation matrix Rx (f ) is diagonal and KB, namely half the rate required for signal recovery [66].
contains the power spectrum Sx (f ) at the corresponding The recovery of rx (f ) is performed using the procedure
frequencies, as presented in the context of signal recovery on (27), that is CTF,
fNyq support recovery and power spectrum reconstruction (rather
Rx(i,i) (f ) = Sx (f + ifs − ), f ∈ Fs . (26) than signal reconstruction).
2
The same result for the minimal sampling rate is valid for
Recovering the power spectrum Sx (f ) is thus equivalent to non sparse signals, for which KB is on the order of fNyq [66].
recovering the matrix Rx (f ). Exploiting the fact that Rx (f ) The power spectrum of such signals can be recovered at half
is diagonal and denoting by rx (f ) its diagonal, (25) can be their Nyquist rate. This means that even without any sparsity
reduced to constraints on the signal in crowded environments, a CR can
rz (f ) = (Ā ⊙ A)rx (f ), (27) retrieve its power spectrum by exploiting stationarity. In this
where rz (f ) = vec(Rz (f )) concatenates the columns of case, the system (27) is overdetermined and rx (f ) is obtained
Rz (f ). The matrix Ā is the conjugate of A and ⊙ denotes by a simple pseudo-inverse operation.
the Khatri-Rao product [77]. Obviously, in practice, we do not have access to Rz (f ),
 15

which thus needs to be estimated. The overall sensing time

is divided into Nf frames or windows of length Ns samples.
In [66], different choices of Nf and Ns are examined for
a fixed sensing time. In order to estimate the autocorrela-
tion matrix Rz (f ) in the frequency domain, estimates of Fig. 20. Minimal sparse ruler of order M = 6 and length N = 10.
zi (f ), 1 ≤ i ≤ M , denoted by ẑi (f ), are first computed
using the fast Fourier transform (FFT) on the samples zi [n]
over a finite time window. The elements of Rz (f ) are then sparse ruler; however, the following bounds hold
obtained by averaging the correlations between ẑi (f ) over
√ √
the frames. In practice, the number of samples dictates the τ (N − 1) M 3(N − 1)
number of discrete Fourier transform (DFT) coefficients in ≤ ≤ , (28)
N N N
the frequency domain and therefore the resolution of the
reconstructed power spectrum. Once r̂x (f ) is reconstructed, where τ ≈ 2.4345 [79]. A circular or modular sparse ruler
several detection statistics can be computed such as power or extends this idea to include periodicity. Such designs that seek
eigenvalue based test statistics [78]. minimal sparse rulers, that is rulers with minimal number of
2) Sampling Scheme Design: Sampling approaches specif- marks M√ , allow to achieve compression ratios M /N on the
ically designed for estimating the autocorrelation of stationary order of N . As N increases, the compression ratio may be
signals at much finer lags than the sample spacings have been arbitrarily low.
studied recently in detail [62], [74], [75], [79]. These rely on Two additional sampling techniques specifically designed
the observation that the autocorrelation is a function of the lags for autocorrelation recovery are nested arrays [74] and co-
only, namely the differences between pairs of sampling times. prime sampling [75], presented in the context of autocorrela-
Thus, the correlation may be estimated at time lags contained tion estimation as well as beamforming and DOA estimation
in the difference set, also referred to as the difference co-array, applications. In nested and co-prime structures, similarly to
which is the set composed of all differences between pairs of multicoset, the corresponding co-arrays have more degrees of
sampling times. Since the size of the difference set may be freedom than those of the original arrays, leading to a finer grid
greater than that of the original sampling set, depending on for the time lags with respect to the sampling times. We now
the choice of sampling times, we may need less sampling quickly review both sampling structures and their correspond-
times for autocorrelation recovery than for signal recovery. ing difference co-arrays and show how the autocorrelation of
Therefore, the sampling times should be carefully chosen so an arbitrary stationary signal can be recovered on the Nyquist
as to maximize the cardinality of the difference set. grid from low rate samples.
The first approach we present, for autocorrelation recovery In its simplest form, the nested array [74] structure has
at sub-Nyquist rates, adopts multicoset sampling and designs two levels of sampling density. The first level samples are at
the cosets in (10) to obtain a maximal number of differences. the N1 locations {`TNyq }1≤`≤N1 and the second level samples
In the previous section, the results were derived for any coset are at the N2 locations {(N1 + 1)kTNyq }1≤k≤N2 . This nonuni-
selection. Here, we show that the sampling rate may be lower form sampling is then repeated with period (N1 + 1)N2 TNyq .
if the cosets are carefully chosen. When using multicoset Since there are N1 + N2 samples in intervals of length
sampling, the sampling matrix A in (25) or (27), is a partial (N1 + 1)N2 TNyq , the average sampling rate of a nested array
2π
Fourier matrix with (i, k)th element ej N ci k . A typical element sampling set is given by
j 2π (c −c )k
of (Ā ⊙ A) is then e N i j
. If all cosets are distinct, then
the size of the difference set over one period is greater than N1 + N2 1 1
fs = ≡ + . (29)
or equal to 2M − 1. This bound corresponds to a worst case (N1 + 1)N2 TNyq N1 TNyq N2 TNyq
scenario, as discussed in the previous section and leads to a
sampling rate of at least half Nyquist in the non sparse setting This rate can be arbitrarily low since N1 and N2 may be as
and at least Landau for a sparse signal with unknown support. large as we choose, at the expense of latency.
This happens for example is we select the first or last M cosets Now, consider the difference co-array which has contribu-
or if we keep only the even or odd cosets. tion from the cross-differences and the self-differences. The
To maximize the size of the difference set and increase the non negative cross-differences, normalized by TNyq for clarity,
rank of (Ā ⊙ A), the cosets can be chosen [62], [79] using are given by
minimal linear and circular sparse rulers [80]. A linear sparse
ruler is a set of integers from the interval [0, N ], such that n = (N1 + 1)k − `, 1 ≤ k ≤ N2 , 1 ≤ ` ≤ N1 . (30)
the associated difference sets contains all integers in [0, N ].
Intuitively, it may be viewed as a ruler with some marks erased All differences in the range 1 ≤ n ≤ (N1 + 1)N2 − 1 are
but still able to measure all integer distances between 0 and covered, except for multiples of N1 +1. These are are precisely
its length. For example, consider the minimal sparse ruler of the self differences among the second array. As a result, the
length N = 10. This ruler requires M = 6 marks, as shown in difference co-array is a filled array composed of all integers
Fig. 20. Obviously, all the lags 0 ≤ τ ≤ 10 on the integer grid −[(N1 + 1)N2 − 1] ≤ n ≤ [(N1 + 1)N2 − 1]. Going back to our
are identifiable. There is no closed form expression for the autocorrelation estimation problem, this result shows that by
maximum compression ratio M /N that is achievable using a proper averaging, we can estimate R(τ ) at any lag τ on the
 16

Nyquist grid from nested array samples with arbitrarily low general background on cyclostationarity and then review sub-
sampling rate, as Nyquist cyclostationary detection approaches.
⎧
⎪ Q−1 1) Cyclic Spectrum Recovery: In the previous section, we
⎪
⎪
⎪
1
⎪ q=0∑ x(Ñ (k + q))x∗ (` + Ñ q), n = Ñ k − `, showed how the power spectrum Sx (f ) can be reconstructed
⎪ Q
R̂[n] = ⎨ Q−1 from correlations Rz (f ) between the samples obtained using
⎪
⎪
⎪
⎪
⎪
1
Q ∑
x(Ñ (k + q))x∗ (Ñ (` + q)), n = Ñ (k − `) , the MWC or multicoset sampling. To that end, we first
⎪
⎩ q=0 related Sx (f ) to the slices’ correlation matrix Rx (f ) and
(31)
then recovered the latter from Rz (f ). Here, this approach
where Q is the number of snapshots used for averaging, and
is extended to the cyclic spectrum Sxα (f ). We first show
Ñ = N1 + 1. Here k and ` are such that (30) holds.
how it is related to shifted correlations between the slices,
Another sampling technique designed for autocorrelation re-
namely Rax (f˜) = E [x(f˜)xH (f˜ + a)], for a ∈ [0, fs ] and
covery is co-prime sampling. It involves two uniform sampling
f˜ ∈ [0, fs − a]. Next, similarly to power spectrum recovery,
sets with spacing N1 TNyq and N2 TNyq respectively, where N1
Rax (f˜) is reconstructed from shifted correlations of the sam-
and N2 are co-prime integers. The average sampling rate of
ples Raz (f˜) = E [z(f˜)zH (f˜ + a)]. Once the cyclic spectrum
such a sampling set, given by
Sxα (f ) is recovered, the transmissions’ carriers and bandwidth
1 1 may be estimated by locating its peaks. Since the cyclic spec-
fs = + , (32)
N1 TNyq N2 TNyq trum of stationary noise n(t) is zero for α ≠ 0, cyclostationary
can again be made arbitrarily small compared to the Nyquist detection is more robust to noise than stationary detection.
rate 1/TNyq by choosing arbitrarily large co-prime numbers N1 The alternative definition of the cyclic spectrum (38), pre-
and N2 . The associated difference set, normalized by TNyq , is sented in the cyclostationary box, implies that the elements in
composed of elements of the form n = N1 k − N2 `. Since N1 the matrix Rax (f˜) are equal to Sxα (f ) at the corresponding α
and N2 are co-prime, there exist integers k and ` such that and f . Indeed, it can easily be shown [58] that
the above difference achieves any integer value n. Therefore, Rax (f˜)(i,j) = Sxα (f ), (39)
the autocorrelation may be estimated by proper averaging, as
for
1 Q−1 ∗
R̂[n] = ∑ x(N1 (k + N2 q))x (N2 (` + N1 q)), (33) α = (j − i)fs + a
Q q=0
fNyq ˜ fs (j + i)fs a
where k and ` are such that n = N1 k − N2 `. f = − +f − + + . (40)
2 2 2 2
The main drawback of both techniques, besides the practical
Here Rax (f˜)(i,j) denotes the (i, j)th element of Rax (f˜). This
issue of analog bandwidth similarly to multicoset sampling,
means that each entry of the cyclic spectrum Sxα (f ) can be
is the added latency required for averaging. Furthermore, in
mapped to an element from one of the correlation matrices
practice, synchronizing ADCs with different sampling rates
Rax (f˜), and vice versa. From (11) and similarly to (25) in
can be challenging. Finally, nested array sampling still requires
the context of power spectrum recovery, we relate the shifted
one sampler operating at the Nyquist rate. Thus, there is
correlation matrices of x(f ) and z(f ) as
no saving in terms of hardware, but only in memory and
computation. Raz (f˜) = ARax (f˜)AH , f˜ ∈ [0, fs − a] , (41)
for all a ∈ [0, fs ].
B. Cyclostationary Detection Recall that, in the context of stationary signals, Rx (f )
Communication signals often exhibit statistical periodicity, is diagonal. Here, understanding the structure of Rax (f˜) is
due to modulation schemes such as carrier modulation or more involved. In [58], it is shown that Rax (f˜) contains
periodic keying [81]. Therefore, such signals are cyclosta- non zero elements on its 0, 1 and −1 diagonals and anti-
tionary processes [72]. A characteristic function of these diagonals. Besides the non zero entries being contained only
processes, the cyclic spectrum Sxα (f ), exhibits spectral peaks in the three main and anti-diagonals, additional structure is
at the cyclic frequencies, which are determined by the signal’s exhibited, limiting to two the number of non zero elements
parameters of periodicity, such as the carrier frequency and per row and column of the matrix Rax (f˜). The above pattern
symbol rate [72]. The formal definition of the cyclic spectrum follows from two main considerations. First, each frequency
is presented in “Cyclostationarity”. Cyclostationary based de- component, namely each entry of x(f ), is correlated to at most
tection distinguishes between the signals of interest, assumed two frequencies from the shifted vector of slices x(f˜+a), one
to be cyclostationary, and stationary noise and interference from the same frequency band and one from the symmetric
by measuring spectral correlation [72]. Since stationary noise band. Second, the correlated component can be either in the
and interference exhibit no spectral correlation, as shown same/symmetric slice or in one of the adjacent slices.
in (24), such detectors are highly robust to noise. With Figures 24 and 25 illustrate these correlations for a = 0
noise enhancement being one of the main limitations of sub- and a = fs /2, respectively. First, in Fig. 23, a sketch of the
Nyquist sampling based techniques, cyclostationary detection spectrum of x(t), namely X(f ), is presented for the case
performed on the reconstructed cyclic spectrum from com- of a sparse signal buried in stationary noise. It can be seen
pressive measurements is a natural candidate for improving that frequency bands of X(f ) may either appear in one fp -
sub-Nyquist detection. In this section, we first provide some slice or split between two slices at most since fp ≥ B. The
 17

C YCLOSTATIONARITY Then, from [73], it holds that

Cyclostationary processes have statistical characteristics that
vary periodically with time. Examples of periodic phenomena α α
that give rise to random data abound in engineering and Sxα (f ) = Suv (f ) = E [X (f + ) X ∗ (f − )] . (38)
2 2
science. In particular, in communications, periodicity arises
from the underlying data modulation mechanisms, such as
carrier modulation, periodic keying or pulse modulation [81]. As expressed in (38), the cyclic spectrum Sxα (f ) measures
A characteristic function of such processes, referred to as the correlations between different spectral components of x(t).
cyclic spectrum, extends the traditional power spectrum to a Stationary signals, which do not exhibit spectral correlation
two dimensional map, with respect to two frequency variables, between distinct frequency components, appear only at α = 0.
angular and cyclic. The cyclic spectrum exhibits spectral peaks This property is the key to robust detection of cyclostationary
at certain frequency locations, the cyclic frequencies, which signals in the presence of stationary noise, in low SNR
are determined by the signal’s parameters, particularly the regimes.
carrier frequency and symbol rate [72]. The support region in the (f, α) plane of the cyclic spectrum
A process x(t) is said to be wide-sense cyclostationary with of a bandpass cyclostationary signal is composed of four dia-
period T0 if its mean µx (t) = E[x(t)] and autocorrelation monds, as shown in Fig. 21. Thus, the cyclic spectrum Sxα (f )
Rs (t, τ ) = E[x(t)x(t + τ )] are both periodic with period of a multiband signal with K uncorrelated transmissions is
T0 [32], that is supported over 4K diamond-shaped areas. Figure 22 illustrates
the cyclic spectrum of two modulation types, AM and BPSK.
µx (t + T0 ) = µx (t), Rx (t + T0 , τ ) = Rx (t, τ ), (34)
𝛼
for all t ∈ R. Given a wide-sense cyclostationary random
process, its autocorrelation Rx (t, τ ) can be expanded in a
2𝑓𝑐
Fourier series
Rx (t, τ ) = ∑ Rxα (τ )ej2παt , (35)
α −𝑓𝑐 𝑓𝑐
where the sum is over integer multiples of the fundamental
frequency 1/T0 and the Fourier coefficients Rxα (τ ) are referred 𝑓
to as cyclic autocorrelation functions. The cyclic spectrum
is obtained by taking the Fourier transform of the cyclic
autocorrelation functions with respect to τ , namely 2𝐵 −2𝑓𝑐
∞
Sxα (f ) = ∫ Rxα (τ )e−j2πf τ dτ, (36) 𝐵
−∞
where α is referred to as the cyclic frequency and f is the
Fig. 21. Support region of the cyclic spectrum of a bandpass cyclostationary
angular frequency [32]. If there is more than one fundamental signal with carrier frequency fc and bandwidth B.
frequency 1/T0 , then the process x(t) is said to be poly-
cyclostationary in the wide sense. In this case, the cyclic
spectrum contains harmonics (integer multiples) of each of the
fundamental cyclic frequencies [72]. These cyclic frequencies
are governed by the transmissions’ carrier frequencies and
symbol rates as well as modulation types.
An alternative and more intuitive interpretation of the cyclic
spectrum expresses it as the cross-spectral density Sxα (f ) =
Suv (f ) of two frequency-shifted versions of x(t), u(t) and
v(t), such that
u(t) ≜ x(t)e−jπαt , v(t) ≜ x(t)e+jπαt . (37) Fig. 22. Cyclic spectrum magnitude of bandpass signals with (a) AM
modulation and (b) BPSK modulation.

resulting vector of spectrum slices x(f ) and the correlations those of the blue trapezes are contained in the −1 and +1
between these slices without any shift, namely R0x (f˜), are anti diagonals. The red rectangles do not contribute any cross-
shown in Fig. 24(a) and (b), respectively. In Fig. 24(b), we correlations for a = 0.
observe that self-correlations appear only on the main diagonal
Figures 25(a) and (b) show the vector x(f˜) and its shifted
since every frequency component is correlated with itself.
version x(f˜ + a) for a = fs /2, respectively. The resulting
In particular, the main diagonal contains the noise’s power
correlation matrix Rax (f˜) appears in Figure 25(c). Here, the
spectrum (in green). Cross-correlations between the yellow
self correlations of the yellow triangle appear in the main
symmetric triangles appear in the 0-anti diagonal, whereas
diagonal and that of the blue trapeze in the −1 diagonal. The
 18

Fig. 23. Original spectrum X(f ). The cyclostationary transmissions are

shown in yellow, blue and red, buried in stationary noise in green.

Fig. 26. Carrier frequency and bandwidth estimation from the cyclic spec-
trum: preprocessing (left), thresholding and clustering (middle), parameter
estimation (right).

2) Carrier frequency and bandwidth estimation: Many de-

tection and classification algorithms based on cyclostationarity
have been proposed (see reviews [32], [33]). We first survey
several detection and classification approaches and then ex-
plain why they do not adhere to CR requirements. To assess
the presence or absence of a signal, a first group of techniques
requires a priori knowledge of its parameters and particularly
Fig. 24. (a) Spectrum slices vector x(f˜), (b) correlated slices of x(f˜) in of the carrier frequency [82]–[87], which is the information
the matrix R0x (f˜). that CRs needs to determine in the first place. A second strat-
egy focuses on a single transmission [84], [88], which does not
fit the multiband model. Alternative approaches apply machine
cross-correlations all appear in the anti-diagonal, for the shift learning tools to the modulation classification of a single signal
a = fs /2. Note that since the noise is assumed to be wide- with unknown carrier frequency and symbol rate [89]–[92].
sense stationary, from (24), a noise frequency component is Besides being only suitable for a single transmission, these
correlated only with itself. Thus, n(t) only contributes non- methods require a training phase. Thus, these techniques can
zero elements on the diagonal of R0x (f˜). only cope with PUs whose modulation type and parameters
To recover Rax (f˜) from Raz (f˜), structured CS techniques were part of the training set.
are used in [58] that aim at reconstructing a sparse matrix For CR purposes, we seek a detector designed to comply
while taking into account its specific structure, as described with the following requirements: (1) carrier frequency and
above. Once the cyclic spectrum is reconstructed, the number bandwidth estimation rather than simple detection of the
of transmissions and their respective carrier frequencies and presence or absence of a signal; (2) blind detection, namely
bandwidths are estimated, as discussed in the next section. The without knowledge of the carrier frequencies, bandwidths and
detection performed on the cyclic spectrum is more robust to symbol rates of the transmissions; (3) simultaneous detection
stationary noise than power spectrum based detection, at the of several transmissions; (4) preferably a non-learning ap-
expense of a slightly higher sampling rate, as shown in [58]. proach, i.e. with no training phase. The parameter estimation
More precisely, in the presence of stationary noise, the cyclic algorithm, presented in [93], is a simple parameter extraction
spectrum may be reconstructed from samples obtained at 4/5 method from the cyclic spectrum of multiband signals, that
of the Nyquist rate, without any sparsity assumption on the satisfies these requirements. It allows the estimation of several
signal. If the signal of interest is sparse, then the minimal carriers and bandwidths simultaneously, as well as the number
sampling rate is further reduced to 8/5 of the Landau rate [58]. of transmissions, namely half the number of occupied bands
K/2 for real-valued signals. The proposed parameter estima-
tion algorithm can be decomposed into four main steps: pre-
processing, thresholding, clustering and parameter estimation,
as illustrated in Fig. 26.
The preprocessing simply aims at compensating for the
presence of stationary noise in the cyclic spectrum at the
cyclic frequency α = 0, by attenuating the energy of the cyclic
spectrum at this frequency. Thresholding is then applied to
the resulting cyclic spectrum in order to find its peaks. The
locations and values of the selected peaks are then clustered
using k-means to find the corresponding cyclic feature, after
estimating the number of clusters by applying the elbow
method [94]. It follows that, apart from the cluster present in
Fig. 25. (a) Spectrum slices vector x(f˜), (b) spectrum slices shifted vector DC, the number of real signals, namely Nsig = K/2, is equal
x(f˜ + a), for a = fs /2, (c) correlated slices of x(f˜) and x(f˜ + a) in the
matrix Ra ˜ to half the number of clusters. Next, the carrier frequency fi ,
x (f ), with a = fs /2.
which corresponds to the highest peak [72], is estimated for
 19

each transmission. The bandwidth Bi is found by locating We consider the following collaborative model. A network
the edge of the support of the angular frequencies at the of Nrec CRs receives the Nsig transmissions, such that the
corresponding cyclic frequency αi = 2fi . received signal at the jth CR is given by
Results presented in [58] demonstrate that cyclostationary Nsig Nsig
based detection as described in this section outperforms energy x(j) (t) = ∑ rij (t) = ∑ si (t) ∗ hij (t). (46)
detection carried on the signal’s spectrum or power spectrum, i=1 i=1
at the expense of increased complexity. We now show similar
results obtained from hardware simulations, performed using The channel response hij (t) is determined by fading and
the prototype from Fig. 14. shadowing effects. Typical models are Rayleigh fading, or
small-scale fading, and log-normal shadowing, or large-scale
fading [35], [36], [95], as described in “Channel Fading and
C. Hardware Simulations: Robustness to Noise Shadowing”. In the frequency domain, the Fourier transform
Cyclostationary detection has been implemented in the of the jth received signal is given by
MWC CR prototype. The analog front-end is identical to that Nsig
of the original prototype and only the digital recovery part is X (j) (f ) = ∑ Si (f )Hij (f ). (47)
modified since the cyclic spectrum is recovered directly from i=1
the MWC low rate samples. Preliminary testing suggests that
Therefore, the support of x(j) (t) is included in the support of
sensing success is still achievable at SNRs lower by 10 dB than
the original signal x(t). Since the transmissions are affected
those allowed by energy detection performed on the recovered
differently by fading and shadowing from each transmitter to
spectrum or power spectrum. Representative results shown in
each CR, we can assume that the union of their respective
Fig. 27 demonstrate the advantage of cyclostationary detection
supports is equivalent to the frequency support of x(t). The
over energy detection in the presence of noise. The figure
goal here is to assess the support of the transmitted signal x(t)
presents the reconstructed cyclic spectrum from samples of the
from sub-Nyquist samples of the received x(j) (t), 1 ≤ j ≤ Nrec ,
MWC prototype, as well as cross-sections at f = 0 and α = 0,
by exploiting their joint frequency sparsity.
which corresponds to the power spectrum. This increased
A simple and naive approach is to perform support recovery
robustness to noise comes at the expense of more involved
at each CR from its low rate samples and combine the local
digital processing on the low rate samples. The additional
binary decisions, either in a fusion center for centralized
complexity stems from the higher dimensionality involved,
collaboration or in a distributed manner. In this hard decision
since we reconstruct the 2-dimensional cyclic spectrum rather
strategy, the combination can be performed using several
than the 1-dimensional (power) spectrum.
fusion rules such as AND, OR or majority rule. Although this
method is attractive due to its simplicity and low communi-
VI. C OLLABORATIVE S PECTRUM S ENSING cation overhead, it typically achieves lower performance than
A. Collaborative Model its soft decision counterpart. To mitigate the communication
overhead, soft decision based methods may rely on sharing
Until now, we assumed direct observation of the spectrum.
observations based on the low rate samples, rather than the
In practice, the task of spectrum sensing for CR is further
samples themselves. In the next section, we review such
complicated due to physical channel effects such as path loss,
techniques both in centralized and distributed contexts.
fading and shadowing [34]–[36], as described in the channel
box. To overcome these practical issues, collaborative CR
networks have been considered, where different users share B. Centralized Collaborative Support Recovery
their sensing results and cooperatively decide on the licensed
One approach [97], [98] to centralized spectrum sensing
spectrum occupancy.
considers a digital model based upon a linear relation between
the M sub-Nyquist samples z(j) at CR j and N Nyquist
The different collaborative approaches can be distinguished
samples x(j) obtained for a given sensing time frame, namely
according to several criteria [34]. First, cooperation can be
either centralized or distributed. In centralized settings, the
data is sent to a fusion center which combines the shared data z(j) = Ax(j) , (48)
to jointly estimate the spectrum or determine its occupancy.
In the distributed approach, the CRs communicate among where A is the sampling matrix. This model has been exten-
themselves and iteratively converge to a common estimate or sively studied by Leus et al. [97]–[99]. In all these works,
decision. While centralized cooperation does not require iter- the channel state information (CSI) is assumed to be known
ations and can reach the optimal estimate based on the shared and the joint power spectrum of x is reconstructed, where
data, convergence to this estimate is not always guaranteed x denotes the N × 1 vector of Nyquist samples. CSI is
in its distributed counterpart. On the other hand, the latter is traditionally unknown by the CRs and should be estimated
less power hungry and more robust to node and link failure, prior to detection to enable the use of this method. The
increasing the network survivability. An additional criterion autocorrelation of the Nyquist samples is first related to that of
concerns the shared data type; the CRs may share local binary the sub-Nyquist samples. Then, in [97], the common sparsity
decisions on the spectrum occupation (hard decision) or a of rx (j) is exploited in the frequency domain across all CRs to
portion of their samples (soft decision). jointly reconstruct them at the fusion center, using a modified
 20

Fig. 27. Screen shot from the MWC with cyclostationary detection. The input signal is composed of Nsig = 3 transmissions (or K = 6 bands) with carriers
f1 = 320MHz, f2 = 600MHz and f3 = 860MHz. Recovered cyclic spectrum from low rate samples (left), cyclic spectrum profile at the angular frequency
f = 0, where the cyclic peaks are clearly visible at twice the carrier frequencies (middle), power spectrum recovery failing in the presence of noise (right).

C HANNEL FADING AND S HADOWING Log-normal shadowing: Large-scale fading represents the
Analog signals transmitted over physical channels are affected average signal power attenuation or path loss due to motion
by two main phenomena: Rayleigh fading, or small-scale over large areas. The resulting channel frequency response is
fading, and log-normal shadowing, or large-scale fading [35], therefore a constant. This phenomenon is affected by promi-
[36], [95]. The received signal is generally described in terms nent terrain contours between the transmitter and receiver.
of the transmitted signal si (t) convolved with the impulse Empirical measurements suggest that this type of fading, or
response of the channel hij (t), namely shadowing, follows a normal distribution in dB units [96],
or alternatively, the linear channel gain may be modeled as
rij (t) = si (t) ∗ hij (t), (42) a log-normal random variable [36]. Therefore, the path loss
(PL) measured in dB is expressed as
where rij (t) is the received signal corresponding to the ith
transmission at the jth CR and ∗ denotes convolution. Fading d
P L = P L0 + 10γ log + Xσ . (44)
and shadowing affect the channel response hij (t). d0
Rayleigh fading: For most practical channels, the free-space Here, the reference distance d0 corresponds to a point located
propagation model, which only accounts for path loss, is in the far field of the antenna (typically 1 km for large cells).
inadequate to describe the channel. A signal typically travels The path loss to the reference point P L0 is usually found
from transmitter to receiver over multiple reflective paths, through field measurements or calculated using free-space
which is traditionally modeled as Rayleigh fading. This im- path loss. The value of the path loss exponent γ depends on
plies that the amplitude and phase of the channel response the frequency, antenna heights, and propagation environment.
hij (t) = R(t)ejφ(t) are stochastically independent and iden- Finally, Xσ denotes a Gaussian random variable (in dB)
tically distributed processes. The amplitude R(t), for t ∈ R, with variance σ 2 determined heuristically as well [95]. The
follows the Rayleigh distribution, given by shadowed received signal is thus given by
r −r 2 /2σ 2
pR (r) = { σ2
e r≥0
(43) rij (t) = 10−P Lij /20 ⋅ si (t), (45)
0 otherwise,
where P Lij denotes the path loss between the ith transmitter
where 2σ 2 is the mean power [95]. The phase φ(t), for t ∈ R, and the jth receiver and the channel response hij (t) =
is uniformly distributed over the interval [0, 2π). 10−P Lij /20 δ(t).

simultaneous orthogonal matching pursuit (SOMP) [41] algo- increasing the number of receivers.
rithm. In [98], [99], besides exploiting joint sparsity, cross-
correlations between measurements from different CRs are An alternative approach [57] relies on the analog model
used, namely Rzj zk , where j and k are the indices of two CRs. from (11) and does not assume any a priori knowledge on the
These cross-correlations are related to the common power CSI. This method considers collaborative spectrum sensing
spectrum sx = Frx . It is shown that if the total number of from samples acquired via multicoset sampling or the MWC
samples Nrec M is greater than N and these are suitably chosen at each CR. In this approach, the jth CR shares its observation
to account for enough measurement diversity, then the power matrix V(j) , as defined in (14), rather than the sub-Nyquist
spectrum sx of a non sparse signal can be recovered from samples themselves, and its measurement matrix A(j) , with
compressed samples from a sufficient number of CRs. This a fusion center. The sampling matrices are considered to be
shows that the number of receivers may be traded for the different from one another in order to allow for more mea-
number of samples per CR. However, increasing the number surement diversity. However, the same known matrix can be
of samples per CR does not increase spatial diversity, as does used to reduce the communication overhead. The underlying
matrices U(j) are jointly sparse since fading (42), (43) and
 21

shadowing (45) do not affect the original signal’s support. chosen with some probability Pij , according to the Metropolis-
Capitalizing on the joint support of U(j) , the support of the Hastings scheme for random transition probabilities,
transmitted signal x(t) is recovered at the fusion center by
solving ⎧
⎪ min{ d1i , d1j } (i, j) ∈ E,
⎪
⎪
⎪
Pij = ⎨ ∑(i,k)∈E max{0, d1 − 1
} i = j,
∣∣U(j) ∣∣0
N (50)
arg min ⋃i=1rec (49) ⎪
⎪
⎪
i dk
U(j) ⎪
⎩ 0 otherwise.
s.t. V(j) = A(j) U(j) , for all 1 ≤ j ≤ Nrec .
Here di denotes the cardinality of the neighbor set of the ith
To recover the joint support of U(j) from the observation CR and E is the set of communication links between CRs in
matrices V(j) , both OMP and IHT, described in “Compressed the network.
Sensing Recovery”, are extended to the collaborative set- A single vector, computed from the low rate samples (and
ting [57]. Previously we considered support recovery from that will be defined below for each recovery algorithm), is
an individual CR, which boils down to an MMV system of passed between the CR nodes in the network, rather than
equations (14). CS algorithms have been extended to this the samples themselves, effectively reducing communication
case, such as SOMP from [100] and simultaneous IHT (SIHT) overhead. When a CR receives this vector, it performs local
presented in [101]. These now need to account for the joint computation to update both the shared vector and its own
sparsity across the CRs. estimate of the signal support accordingly. Finally, the vector
The distributed CS-SOMP (DCS-SOMP) algorithm [102], is sent to a neighbor CR, chosen according to the random
which extends the original SOMP to allow for different walk with probability (50). Two distributed algorithms are
sampling matrices A(j) for each receiver, is adapted to the CR presented in [56]. The first, distributed one-step greedy algo-
collaborative setting in [57]. The main modification appears in rithm (DOSGA), extends the OSGA from [102] to distributed
the computation of the index that accounts for the greatest settings. The second method, referred to as randomized dis-
amount of residual energy. Here, the selected index is the tributed IHT (RDSIHT), adapts the centralized BSIHT [57] to
one that maximizes the sum of residual projections over the distributed case.
all the receivers. Once the shared support is updated, the
To describe the DOSGA algorithm, we first present its
residual matrices can be computed for each CR separately.
centralized counterpart OSGA. Each CR computes the `2 -
norm of the projections of the observation matrix V(j) onto
The resulting modified algorithm is referred to as block sparse
the columns of the measurement matrix A(j) , stored in the
OMP (BSOMP) [57]. Next, the block sparse IHT (BSIHT)
vector w(j) of size N . The fusion center then averages over
algorithm [57] extends SIHT by selecting the indices of the
common support though averaging over all the estimated
U(j) in each iteration. Once the support is selected, the
all receivers’ vectors, such that
update calculations are performed separately for each receiver.
1 Nrec (j)
Both methods are suitable for centralized cooperation, in the ŵ = ∑w , (51)
presence of a fusion center. As in the previous approach, if the Nrec j=1
CSI is known, then the sampling rate per CR can be reduced
by a factor of Nrec with respect to the rate required from an and retains the highest values of ŵ whose indices constitute
individual CR. the support of interest. In the absence of a fusion center,
finding this average is a standard distributed average consensus
problem, also referred to as distributed averaging or distributed
C. Distributed Collaborative Support Recovery consensus. DOSGA [56] then uses a randomized gossip al-
In the distributed approach, there is no fusion center and gorithm [105] for this purpose, with the Metropolis-Hastings
the CRs can only communicate with their neighbors. Both the transition probabilities.
digital and analog centralized approaches have been extended Next, we turn to the RDSIHT algorithm, which adapts the
to the distributed settings. In [103], [104], a digital model centralized BSIHT algorithm [57] to the distributed scenario.
(48) is used where the spectrum is divided into N known The distributed approach from [56] was inspired by the ran-
slots. Both unknown and known CSI cases are considered. In domized incremental subgradient method proposed in [106],
the first case, each CR computes its local binary decision for and recent work on a stochastic version of IHT [107]. A vector
each spectral band by recovering the sparse spectrum using CS w of size N , that sums the `2 -norms of the rows of the
techniques. Then, an average consensus approach is adopted, estimates of U(j) before thresholding, is shared in the network
with respect to the shared hard decision. If the CSI is known, through random walk. The indices of its k largest values
similarly to [98], [99], then the spectrum is recovered in a correspond to the current estimated support. When the ith CR
distributed fashion. In [103], the proposed algorithm iterates receives w, it locally updates it by performing a gradient step
through the following steps: local spectrum reconstruction using its own objective function that is then added to w. Next,
given the support and consensus averaging on the support. it selects a neighbor j to send the vector to with probability
In [104], a distributed augmented Lagrangian is adopted. Pij (50). The joint sparsity across the CRs is exploited by
The centralized approach, based on the analog model (11), sharing one common vector w by the network. It is shown
presented in [57], is modified in [56] to comply with dis- numerically in [56] that both DOSGA and RDSIHT converge
tributed settings. The ith CR contacts a random neighbor j, to their centralized counterparts.
 22

D. Hardware Simulations: Collaborative vs. Individual Spec- B. Multicoset Approach

trum Sensing A few works have recently considered joint DOA and spec-
We now confirm that the collaborative algorithms for spec- trum sensing of multiband signals from sub-Nyquist samples.
trum sensing perform better than their individual counterparts In [108] and [109], low rate samples are obtained using
by demonstrating a collaborative setting implemented using the multicoset sampling scheme. In [108], which considers
the MWC CR prototype, as can be seen in Fig. 28. During the digital model (48), both time and spatial compression
the conducted experiments, Nrec = 5 CR receivers, spread are applied by selecting samples from the Nyquist grid and
across different locations are emulated, denoted by white receivers from a uniform linear array (ULA), such that
circles on the transmitter/receiver map. The transmitters are
also positioned in various locations depicted by green x-marks. Z[n] = Cs X[n]Ct . (52)
The transmitter positions and broadcasts mimic the effects of
Here, X[n] is the matrix of Nyquist samples from all receivers
physical channel phenomena, i.e. fading and shadowing. The
in the ULA and the selection matrices Cs and Ct operate
frequency support recovered by each of the CRs is false, since
on the spatial and time domain, respectively, to form the
they individually receive only a partial spectral image of their
matrix of compressed samples Z[n]. The 2D power spectrum
surroundings, as expected in a real-world scenario.
matrix of the underlying signal is then reconstructed from the
In all simulated scenarios, collaborative spectrum sensing samples, where every row describes the power spectrum in
outperforms detection realized by individual CRs. This result the frequency domain for a given AOA and every column
is expected, since the soft collaborative methods take advan- corresponds to the power spectrum information in the angular
tage of the spatial deployment of the receivers to reproduce domain for a given frequency.
the exact spectral map of the environment. Moreover, the
In [109], an L-shaped array with two interleaved, or multi-
centralized and distributed algorithms BSOMP, BSIHT and
coset channels, with a fixed delay between the two, samples
RDSIHT, based on soft decisions, showed superior results in
the signal below the Nyquist rate. Then, exploiting correlations
comparison with a hard decision method. The same result can
between samples from the direct path and its delayed version,
be seen in Fig. 28, where the hard decision support algorithm
the frequencies and their corresponding AOAs are estimated
(Hard Co-Op) fails to recover the entire active frequency
using MUSIC [38], [39]. However, the pairing issue between
support (depicted by red bins).
the two, that is matching each frequency with its corresponding
angle, is not discussed.
VII. J OINT C ARRIER F REQUENCY AND D IRECTION In the next section, we describe the compressed carrier and
E STIMATION DOA estimation (CaSCADE) system, presented in [53], that
utilizes the sampling principles of the MWC. This technique
Many signal processing applications, including CR, may addresses the pairing issue and avoids the hardware issues
require or at least benefit from joint spectrum sensing and involved in multicoset sampling.
DOA estimation. In order for CRs to map vacant bands more
efficiently, spatial information about the PUs’ locations can
be of great interest. Consider the network of CRs presented C. The CaSCADE System
in Fig. 29 and focus on CR1. Now, picture a scenario where The CaSCADE system implements the modified, or ULA
PU2, with angle of arrival (AOA) θ2 with respect to CR1, based, MWC over an L-shaped array with 2M − 1 sensors
is transmitting in a certain frequency band with carrier f2 . (M sensors along the y axis and M sensors along the z
Assuming that CR2 does not receive PU2’s transmission, CR1 axis including a common sensor at the origin) in the yz
could transmit in the same frequency band in the opposite plane. Each transmission si (t) impinges on the array with
direction of PU2 towards CR2. DOA estimation can thus its corresponding AOA θi , as shown in Fig. 30. The array
enhance CR performance by allowing exploitation of vacant sensors have the same sampling pattern as the alternative
bands in space in addition to the frequency domain. MWC. Each sensor is composed of an analog mixing front-
end, implementing one physical branch of the MWC, that
includes a mixer, a LPF and a sampler.
A. Model and System Description
By treating the L-shaped array as two orthogonal ULAs, one
To formulate our problem mathematically, assume that the along the y axis and the other along the z axis, two systems of
input signal x(t) is composed of Nsig source signals si (t) with equations similar to (19) can be derived. For the ULA along
both unknown and different carrier frequencies fi and AOAs the y axis, we obtain
θi . The main difference between this scenario and the one that
has been discussed in the previous sections is the additional y(f ) = Ay x(f ), f ∈ Fs , (53)
unknown AOAs θi . Figure 29 illustrates this signal model. To
recover the unknown AOAs, an array of sensors is required. where
A similar problem thoroughly treated in the literature, is the ⎡ ej2πf1 τ1y (θ1 ) ⋯ ej2πfK τ1 (θK ) ⎤⎥
y
⎢
⎢ ⋮ ⋮ ⎥
2D-DOA recovery problem, where two angles are traditionally ⎢ ⎥
Ay = ⎢ ⎥, (54)
recovered and paired. In our case, the second variable is the ⎢ ⎥
⎢ j2πf τ y (θ ) ⎥
j2πfK τM (θK ) ⎥
signal’s carrier frequency instead of an additional angle. ⎢e ⋯ e
y
⎣ 1 M 1
⎦
 23

Fig. 28. Screen shot from the MWC CR collaborative hardware prototype. On the upper left side, we see the spatial map of the receivers in white, and
transmitters in green. On the bottom left, the occupied band indices of the real spectral support are shown, while to the right of the transmitter/receiver
map, the estimated indices by each CR individually are presented. On the right, we see the spectrum sensing results of 4 different algorithms: Hard Co-Op
(hard decision collaboration that selects the most popular frequency band indices), BSIHT, BSOMP and RDSIHT. These results show both the superiority of
collaborative spectrum sensing methods over individual detection and that of soft decision methods compared to the plain union of all CR results.

f z
f2 0

PU2
f zM
f1 0

PU1 θ2
CR1
θ1 zM −1
z
θ3
CR2
y (fj , θj ) (fi, θi)

PU3
z3
f
0 f3
z2
Fig. 29. Illustration of Nsig = 3 source signals in the yz plane. Each
θj θi
transmission is associated with a carrier frequency fi and AOA θi . d
y
z1
y1 y2 y3 yM −1 yM
sub-array y1

and τmy
(θ) = dm
c
cos (θ) denote the delays at the mth sensors sub-array y2

with respect to the first sensor. Along the z axis, the samples
Fig. 30. CaSCADE system: L-shaped array with M sensors along the y axis
z(f ), sampling matrix Az and delays τm z
are similarly defined. and M sensors along the z axis including a common sensor at the origin. The
sub-arrays y1 and y2 (and similarly z1 and z2 ) are defined in the derivation
Two joint recovery approaches for the carrier frequencies of the 2D-ESPRIT algorithm.
and DOAs of the transmissions are proposed in [53]. Note
that once the carriers and AOAs are estimated, the signals can
be reconstructed, as shown for the alternative MWC. For the T
v[n] = [yT [n] zT [n]] , whose correlation matrix,
sake of simplicity, a statistical model where x(t) is wide-sense
stationary is considered. The first recovery approach is based R = E [v[n]vH [n]] = ARx AH , (55)
on CS techniques and allows recovery of both parameters T
assuming the electronic angles fi cos θi and fi sin θi lie on a is computed. Here, A = [ATy ATz ] and the autocorrelation
predefined grid. The CS problem is formulated in such a way matrix Rx = E [x[k]xH [k]] is sparse and diagonal, from the
that no pairing issue arises between the carrier frequencies stationarity of x(t). From the grid assumption, (55) can be
and their corresponding DOAs. To that end, the time-domain discretized with respect to the possible values taken by the
samples from both ULAs are concatenated into one vector electronic angles. The resulting sparse matrix derived from Rx
 24

is diagonal as well, and its sparse diagonal can be recovered scratch. Additional preliminary assumptions on the structure
using traditional CS techniques, similarly to (27). or statistical behavior of the potentially active signals, such as
The second recovery approach, inspired by [110], [111], statistics on channel occupancy, can also be exploited.
extends the ESPRIT algorithm to the joint estimation of We then considered detection challenges in the presence of
carriers and DOAs, while overcoming the pairing issue. The noise, where second-order statistics recovery, and in particular
2D-ESPRIT algorithm presented in [53] is directly applied cyclostationary detection, are shown to perform better than
to the sub-Nyquist samples, by considering cross-correlation simple energy detection. Next, fading and shadowing channel
matrices between the sub-arrays of both axis. Dropping the effects were overcome by collaborative CR networks. Finally,
time variable n for clarity, the samples from the sub-arrays we addressed the joint spectrum sensing and DOA estimation
can be written as problem, allowing for better exploitation of frequency vacant
bands by exploiting spatial sparsity as well. All these methods
y1 = Ay1 x, y 2 = A y2 x
should be combined in order to map the occupied spectrum,
z1 = Az1 x, z2 = Az2 x, (56) in frequency, time and space, thus maximizing the CR net-
work’s throughput. This requires an adequate spectrum access
where y1 , y2 , z1 , z2 are samples from the sub-arrays shown
protocol as well, that translates the data acquired by spectrum
in Fig. 30. The matrices Ay1 and Ay2 are the first and last
sensing into transmission opportunities for the CRs. Spectrum
M −1 rows of Ay , respectively and Az1 and Az2 are similarly
access challenges and algorithms were outside the scope of
defined. Each couple of sub-array matrices along the same axis
this review.
are related as
An essential part of the approach adopted in this sur-
Ay2 = Ay1 Φ vey is the relation between the theoretical algorithms and
Az2 = Az1 Ψ, (57) practical implementation, demonstrating real-time spectrum
sensing from low rate samples using off-the-shelf hardware
where components. We believe that prototype development is key to
Φ ≜ diag [ej2πf1 τ1 (θ1 ) ej2πfK τ1 (θK ) ]
y y enabling sub-Nyquist sampling as a solution to the task of
⋯
spectrum sensing in CR platforms. A natural next step in that
≜ diag [ej2πf1 τ1 (θ1 ) ej2πfK τ1 (θK ) ] .
z z
Ψ ⋯ (58) direction is the implementation of a complete CR network,
We see from (58) that the carrier frequencies fi and AOAs θi collaboratively performing joint spectrum sensing and DOA
are embedded in the diagonal matrices Φ and Ψ. Applying the estimation followed by spectrum access. This prototype should
ESPRIT framework on cross-correlations matrices between the then be tested on real data in order to assess its true capabili-
subarrays of both axis, allows to jointly recover Φ and Ψ [53]. ties. We believe that a sub-Nyquist digital to analog interface
This leads to proper pairing of the corresponding elements can alleviate many of the bottlenecks currently hindering the
fi τ1y (θi ) and fi τ1z (θi ). The AOAs θi and carrier frequencies development of CR systems, allowing fast deployment of low
fi are then given by rate, simple and efficient CR devices, using currently available
hardware.
∠Ψii ∠Φii
θi = tan−1 ( ), fi = . (59) In order for CRs to become a viable solution to spec-
∠Φii 2π dc cos (θi ) trum shortage, other main challenges need to be addressed,
It is proven in [53] that the minimal number of sensors as discussed earlier. The issue of coexistence with existing
required for perfect recovery is 2K +1. This leads to a minimal communication links from PUs is crucial and is very particular
sampling rate of (2K + 1)B, which is slightly higher than to the CR scenario. Here, coexistence is not a symmetric
the minimal rate 2KB required for spectrum sensing in the requirement as CRs are prohibited from interfering with PUs.
absence of DOA recovery. These ideas can be extended to Another challenge is the establishment of a communication
jointly recovery the transmissions’ carrier frequencies, azimuth channel for CRs to be able to exchange the locations of
and elevation angles in a 3D framework. their current transmission bands. Finally, the issue of secu-
rity against attacks to the CR networks still has numerous
unresolved questions.
VIII. C ONCLUSION AND F UTURE C HALLENGES
In this paper, we reviewed several challenges imposed on
R EFERENCES
the traditional task of spectrum sensing by the new application
of CR. We first investigated sub-Nyquist sampling schemes, [1] FCC, “Spectrum policy task force report: Federal com-
enabling sampling and processing of wideband signals at low munications commission, tech. rep. 02-135. [online],”
http://www.gov/edocs public/attachmatch/DOC228542A1.pdf, Nov.
rate, by exploiting their a priori known structure. A possible 2002.
extension of these works is to include adaptive updating of [2] M. McHenry, “NSF spectrum occupancy measurements
the detected support, triggered by a change in a PU’s activity, project summary. shared spectrum co., tech. rep. [online],”
http://www.sharedspectrum.com, Aug. 2005.
either starting a transmission in a previously vacant band or [3] R. I. C. Chiang, G. B. Rowe, and K. W. Sowerby, “A quantitative
withdrawing from an active band. To increase efficiency, this analysis of spectral occupancy measurements for cognitive radio,” Proc.
should be performed by taking the current detected support as IEEE Vehicular Technology Conf., pp. 3016–3020, Apr. 2007.
[4] Y. Liu, “Beyond white space: Robust spectrum sensing and channel
a prior and updating it with respect to the new acquired sam- statistics based spectrum accessing strategies for cognitive radio net-
ples, without going through the support recovery process from work,” Ph.D. dissertation, University of Texas at Austin, Aug. 2013.
 25

[5] T. Erpek, M. Lofquist, and K. Patton, “Spectrum occupancy measure- [32] W. A. Gardner, A. Napolitano, and L. Paura, “Cyclostationarity: Half
ments Loring Commerce Centre Limestone, Maine September 18-20, a century of research,” Signal Process., vol. 86, pp. 639–697, 2006.
2007,” The Shared Spectrum Company, 2007. [33] A. Napolitano, “Cyclostationarity: New trends and applications,” Signal
[6] J. Mitola, “Software radios: Survey, critical evaluation and future Process., vol. 120, pp. 385–408, Mar. 2016.
directions,” IEEE Aerospace Electron. Syst. Mag., vol. 8, pp. 25–36, [34] I. F. Akyildiz, B. F. Lo, and R. Balakrishnan, “Cooperative spectrum
Apr. 1993. sensing in cognitive radio networks: A survey,” Phys. Commun., vol. 4,
[7] S. Haykin, “Cognitive radio: Brain-empowered wireless communica- pp. 40–62, Mar. 2011.
tions,” IEEE J. Select. Areas Commun., vol. 23, pp. 201–220, Feb. [35] S. M. Mishra, A. Sahai, and R. W. Brodersen, “Cooperative sensing
2005. among cognitive radios,” IEEE Int. Conf. Commun., pp. 1658–1663,
[8] J. Mitola and C. Q. Maguire Jr., “Cognitive radio: Making software 2006.
radios more personal,” IEEE Personal Commun., vol. 6, pp. 13–18, [36] A. Ghasemi and E. S. Sousa, “Collaborative spectrum sensing for
Aug. 1999. opportunistic access in fading environments,” IEEE Int. Symp. New
[9] C. E. Shannon, “Computers and automata,” Proc. IRE, vol. 41, pp. Frontiers in Dynamic Spectrum Access Networks, pp. 131–136, Nov.
1234–1241, Oct. 1953. 2005.
[10] D. Cabric, S. Mishra, and R. Brodersen, “Implementation issues in [37] Z. Quan, S. Cui, A. H. Sayed, and H. V. Poor, “Optimal multiband
spectrum sensing for cognitive radios,” Asilomar Conf. on Signals, Syst. joint detection for spectrum sensing in cognitive radio networks,” IEEE
and Computers, vol. 1, pp. 772–776, Nov. 2004. Trans. Signal Process., vol. 57, pp. 1128–1140, 2009.
[11] E. G. Larsson and M. Skoglund, “Cognitive radio in a frequency- [38] V. F. Pisarenko, “The retrieval of harmonics from a covariance func-
planned environment: Some basic limits,” IEEE Trans. Wireless Com- tion,” Geophysics, J. Roy. Astron. Soc., vol. 33, pp. 347–366, 1973.
mun., vol. 7, pp. 4800–4806, Dec. 2008. [39] R. O. Schmidt, “Multiple emitter location and signal parameter estima-
[12] A. Sahai, N. Hoven, and R. Tandra, “Some fundamental limits on tion,” IEEE Trans. Antennas Propagation, vol. 34, pp. 276–280, Mar.
cognitive radio,” Proc. Annu. Allerton Conf. Commun., Control, and 1986.
Computing, pp. 1662–1671, Oct. 2004. [40] R. Roy and T. Kailath, “ESPRIT-estimation of signal parameters via
[13] M. Mishali and Y. C. Eldar, “Wideband spectrum sensing at sub- rotational invariance techniques,” IEEE Trans. Signal Process., vol. 37,
Nyquist rates,” IEEE Signal Process. Mag., vol. 28, pp. 102–135, Jul. pp. 984–995, Jul. 1989.
2011. [41] Y. C. Eldar and G. Kutyniok, Compressed Sensing: Theory and
[14] A. Ghasemi and E. S. Sousa, “Spectrum sensing in cognitive radio net- Applications. Cambridge University Press, 2012.
works: requirements, challenges and design trade-offs,” IEEE Commun. [42] S. K. Sharma, E. Lagunas, S. Chatzinotas, and B. Ottersten, “Appli-
Mag., vol. 46, pp. 32–39, Apr. 2008. cation of compressive sensing in cognitive radio communications: A
[15] E. Axell, G. Leus, E. G. Larsson, and H. V. Poor, “Spectrum sensing survey,” IEEE Commun. Surveys and Tutorials, 2016.
for cognitive radios: State-of-the-art and recent advances,” IEEE Signal [43] Y. C. Eldar, Sampling Theory: Beyond Bandlimited Systems. Cam-
Process. Mag., vol. 29, pp. 101–116, May 2012. bridge University Press, 2015.
[16] B. Wang and K. Liu, “Advances in cognitive radio networks: A survey,” [44] S. G. Mallat and Z. Zhang, “Matching pursuits with time-frequency
IEEE J. Select. Topics in Signal Process., vol. 5, pp. 5–23, 2011. dictionaries,” IEEE Trans. Signal Process., vol. 41, pp. 3397–3415,
[17] J. Ma, G. Y. Li, and B. H. F. Juang, “Signal processing in cognitive Dec. 1993.
radio,” Proc. IEEE, vol. 97, pp. 805–823, 2009. [45] T. Blumensath and M. Davies, “Gradient pursuits,” IEEE Trans. Signal
[18] Y. Zeng, Y.-C. Liang, A. T. Hoang, and R. Zhang, “A review on spec- Process., vol. 56, pp. 2370–2382, Jun. 2008.
trum sensing for cognitive radio: challenges and solutions,” EURASIP [46] M. Mishali and Y. C. Eldar, “Sub-Nyquist sampling: Bridging theory
J. Advances Signal Process., vol. 2010, pp. 1–15, 2010. and practice,” IEEE Signal Process. Mag., vol. 28, no. 6, pp. 98–124,
[19] T. Yücek and H. Arslan, “A survey of spectrum sensing algorithms for Nov. 2011.
cognitive radio applications,” IEEE Commun. Surveys and Tutorials, [47] H. Landau, “Necessary density conditions for sampling and interpola-
vol. 11, pp. 116–130, 2009. tion of certain entire functions,” Acta Math, vol. 117, pp. 37–52, Jul.
[20] E. Hossain, D. Niyato, and D. I. Kim, “Evolution and future trends of 1967.
research in cognitive radio: a contemporary survey,” Wireless Commun. [48] R. Venkataramani and Y. Bresler, “Perfect reconstruction formulas
and Mobile Computing, vol. 15, pp. 1530–1564, 2015. and bounds on aliasing error in sub-Nyquist nonuniform sampling of
[21] E. Biglieri, A. J. Goldsmith, L. J. Greenstein, N. B. Mandayam, and multiband signals,” IEEE Trans. Inf. Theory, vol. 46, pp. 2173–2183,
H. V. Poor, Principles of cognitive radio. Cambridge University Press, Sept. 2000.
2012. [49] M. Mishali, Y. C. Eldar, O. Dounaevsky, and E. Shoshan, “Xampling:
[22] J. N. Laska, S. Kirolos, M. F. Duarte, T. S. Ragheb, R. G. Baraniuk, and Analog to digital at sub-Nyquist rates,” IET Circuits, Devices and Syst.,
Y. Massoud, “Theory and implementation of an analog-to-information vol. 5, pp. 8–20, Jan. 2011.
converter using random demodulation,” IEEE Int. Symp. Circuits and [50] E. Israeli, S. Tsiper, D. Cohen, A. Reysenson, R. Hilgendorf,
Syst., pp. 1959–1962, May 2007. E. Shoshan, and Y. C. Eldar, “Hardware calibration of the modulated
[23] J. A. Tropp, J. N. Laska, M. F. Duarte, J. K. Romberg, and R. G. wideband converter,” IEEE Global Commun. Conf., pp. 948–953, Dec.
Baraniuk, “Beyond Nyquist: Efficient sampling of sparse bandlimited 2014.
signals,” IEEE Trans. Inf. Theory, vol. 56, pp. 520–544, Jan. 2010. [51] M. Mishali and Y. C. Eldar, “Expected RIP: Conditioning of the
[24] M. Fleyer, A. Linden, M. Horowitz, and A. Rosenthal, “Multirate modulated wideband converter,” IEEE Inf. Theory Workshop, pp. 343–
synchronous sampling of sparse multiband signals,” IEEE Trans. Signal 347, 2009.
Process., vol. 58, pp. 1144–1156, Mar. 2010. [52] H. W. L. Gan and H. Wang, “Deterministic binary sequences for
[25] M. Mishali and Y. C. Eldar, “Blind multi-band signal reconstruction: modulated wideband converter,” in Int. Conf. Sampling Theory and
Compressed sensing for analog signals,” IEEE Trans. Signal Process., Applications, 2013.
vol. 57, no. 3, pp. 993–1009, Mar. 2009. [53] S. Stein, O. Yair, D. Cohen, and Y. C. Eldar, “CaSCADE: Compressed
[26] ——, “From theory to practice: Sub-Nyquist sampling of sparse carrier and DOA estimation,” arXiv:1604.02723 [cs.IT], 2016.
wideband analog signals,” IEEE J. Select. Topics Signal Process., [54] R. Gold, “Optimal binary sequences for spread spectrum multiplexing
vol. 4, no. 2, pp. 375–391, Apr. 2010. (corresp.),” IEEE Trans. Inf. Theory, vol. 13, no. 4, pp. 619–621, 1967.
[27] M. Mishali, Y. C. Eldar, and A. J. Elron, “Xampling: Signal acquisition [55] M. Mishali and Y. C. Eldar, “Expected RIP: conditioning of the
and processing in union of subspaces,” IEEE Trans. Signal Process., modulated wideband converter,” IEEE Inf. Theory Workshop, pp. 343–
vol. 59, pp. 4719–4734, Oct. 2011. 347, Oct. 2009.
[28] H. Urkowitz, “Energy detection of unknown deterministic signals,” [56] D. Cohen, A. Akiva, B. Avraham, and Y. C. Eldar, “Distributed
Proc. IEEE, vol. 55, pp. 523–531, Apr. 1967. cooperative spectrum sensing from sub-Nyquist samples for cognitive
[29] E. Arias-Castro and Y. C. Eldar, “Noise folding in compressed sensing,” radios,” IEEE Workshop Signal Proc. Advances Wireless Commun., pp.
IEEE Signal Process. Lett., vol. 18, no. 8, pp. 478–481, Aug. 2011. 336–340, Jun. 2015.
[30] D. O. North, “An analysis of the factors which determine signal/noise [57] ——, “Centralized cooperative spectrum sensing from sub-Nyquist
discrimination in pulsed carrier systems,” Proc. IEEE, vol. 51, pp. samples for cognitive radios,” IEEE Int. Conf. Commun., pp. 7487–
1016–1027, Jul. 1963. 7491, Jun. 2015.
[31] G. L. Turin, “An introduction to matched filters,” IRE Trans. Inf. [58] D. Cohen and Y. C. Eldar, “Sub-Nyquist cyclostationary detection for
Theory, vol. 6, pp. 311–329, Jun. 1960. cognitive radio,” arXiv:1604.02659 [cs.IT], 2016.
 26

[59] D. Adams, Y. C. Eldar, and B. Murmann, “A mixer frontend for a four- [86] Y. T. Chan, J. W. Plews, and K. C. Ho, “Symbol rate estimation by the
channel modulated wideband converter with 62 db blocker rejection,” wavelet transform,” IEEE Int. Symp. Circuits and Syst., pp. 177–180,
in 2016 IEEE Radio Frequency Integrated Circuits Symposium (RFIC), Jun. 1997.
May 2016, pp. 286–289. [87] E. Rebeiz and D. Cabric, “Blind modulation classification based on
[60] M. A. Lexa, M. E. Davies, J. S. Thompson, and J. Nikolic, “Com- spectral correlation and its robustness to timing mismatch,” Military
pressive power spectral density estimation,” IEEE Int. Conf. Acoustics, Commun. Conf., pp. 277–282, 2011.
Speech and Signal Process., vol. 57, pp. 22–27, May 2011. [88] R. J. Mammone, R. J. Rothaker, C. I. Podilchuk, S. Davidovici, and
[61] M. A. Lexa, M. E. Davies, and J. S. Thompson, “Compressive D. L. Schilling, “Estimation of carrier frequency, modulation type and
and noncompressive power spectral density estimation from periodic bit rate of an unknown modulated signal,” IEEE Int. Conf. Commun.,
nonuniform samples,” CoRR, vol. abs/1110.2722, 2011. pp. 1006–1012, 1987.
[62] D. D. Ariananda and G. Leus, “Compressive wideband power spectrum [89] A. Badawy and T. Khattab, “Robust, block based cyclostationary
estimation,” IEEE Trans. Signal Process., vol. 60, pp. 4775–4789, Sept. detection,” IEEE Wireless Commun. and Networking Conf., pp. 253–
2012. 258, Apr. 2014.
[63] D. D. Ariananda, G. Leus, and Z. Tian, “Muti-coset sampling for power [90] A. Fehske, J. Gaeddert, and J. H. Reed, “A new approach to signal
spectrum blind sensing,” Int. Conf. Digital Signal Process., pp. 1–8, classification using spectral correlation and neural networks,” IEEE
Jul. 2011. Int. Symp. on New Frontiers in Dynamic Spectrum Access Networks,
[64] D. Romero and G. Leus, “Compressive covariance sampling,” Proc. pp. 144–150, Nov. 2005.
Inf. Theory and Applications Workshop, pp. 1–8, Feb. 2013. [91] Aparna P.S and M. Jayasheela, “Cyclostationary feature detection in
[65] C. P. Yen, Y. Tsai, and X. Wang, “Wideband spectrum sensing based cognitive radio using different modulation schemes,” Int. J. Computer
on sub-Nyquist sampling,” IEEE Trans. Signal Process., vol. 61, pp. Applications, vol. 47, pp. 12–16, June 2012.
3028–3040, Jun. 2013. [92] X. He, Z. Zeng, and C. Guo, “Signal classification based on cyclosta-
[66] D. Cohen and Y. C. Eldar, “Sub-Nyquist sampling for power spectrum tionary spectral analysis and HMM/SVM in cognitive radio,” Int. Conf.
sensing in cognitive radios: A unified approach,” IEEE Trans. Signal on Measuring Technology and Mechatronics Automation, pp. 309–312,
Process., vol. 62, pp. 3897–3910, Aug. 2014. 2009.
[67] Z. Tian, “Cyclic feature based wideband spectrum sensing using [93] D. Cohen, L. Pollak, and Y. C. Eldar, “Carrier frequency and bandwidth
compressive sampling,” IEEE Int. Conf. Commun., pp. 1–5, Jun. 2011. estimation of cyclostationary multiband signals,” IEEE ICASSP, Mar.
[68] Z. Tian, Y. Tafesse, and B. M. Sadler, “Cyclic feature detection with 2016.
sub-nyquist sampling for wideband spectrum sensing,” IEEE J. Select. [94] R. L. Thorndike, “Who belong in the family?” Psychometrika, vol. 18,
Topics Signal Process., vol. 6, no. 1, pp. 58–69, Feb. 2012. pp. 267–276, Dec. 1953.
[69] S. A. Razavi, M. Valkama, and D. Cabric, “High-resolution cyclic [95] B. Sklar, “Rayleigh fading channels in mobile digital communication
spectrum reconstruction from sub-Nyquist samples,” IEEE Workshop systems part I: Characterization,” IEEE Commun. Mag., vol. 35, pp.
Signal Process. Advances Wireless Commun., pp. 250–254, Jun. 2013. 90–100, Jul. 1997.
[70] G. Leus and Z. Tian, “Recovering second-order statistics from compres- [96] V. Erceg, L. J. Greenstein, S. Y. Tjandra, S. R. Parkoff, A. Gupta,
sive measurements,” IEEE Int. Workshop on Computational Advances B. Kulic, A. A. Julius, and R. Bianchi, “An empirically based path
in Multi-Sensor Adaptive Processing, pp. 337–340, Dec. 2011. loss model for wireless channels in suburban environments,” IEEE J.
Select. Areas Commun., vol. 17, pp. 1205–1211, Sep. 2006.
[71] D. D. Ariananda and G. Leus, “Non-uniform sampling for compressive
[97] Y. Wang, A. Pandharipande, Y. L. Polo, , and G. Leus, “Distributed
cyclic spectrum reconstruction,” IEEE Int. Conf. Acoustics, Speech and
compressive wide-band spectrum sensing,” IEEE Inf. Theory and
Signal Process., pp. 41–45, 2014.
Applications Workshop, pp. 178–183, Feb. 2009.
[72] W. Gardner, Statistical spectral analysis: a non probabilistic theory.
[98] D. D. Ariananda and G. Leus, “A study on cooperative compressive
Prentice Hall, 1988.
wideband power spectrum sensing,” Joint WIC/IEEE Symp. Inf. Theory
[73] A. Papoulis, Probability, Random Variables, and Stochastic Processes.
and Signal Process., pp. 102–109, 2012.
McGraw Hill, 1991.
[99] ——, “Cooperative compressive wideband power spectrum sensing,”
[74] P. Pal and P. P. Vaidyanathan, “Nested array: A novel approach to array IEEE Asilomar Conf. on Signals, Syst., and Computers, pp. 303–307,
processing with enhanced degrees of freedom,” IEEE Trans. Signal Nov. 2012.
Process., vol. 58, pp. 4167–4181, Aug. 2010.
[100] J. Tropp, A. C. Gilbert, and M. J. Strauss, “Simultaneous sparse
[75] P. P. Vaidyanathan and P. Pal, “Sparse sensing with co-prime samplers approximation via greedy pursuit,” IEEE Int. Conf. Acoustics, Speech
and arrays,” IEEE Trans. Signal Process., vol. 59, pp. 573–586, 2011. and Signal Process., vol. 5, pp. 721–724, Mar. 2005.
[76] D. Qu and A. Tarczynski, “A novel spectral estimation method by [101] A. Makhzani and S. Valaee, “Reconstruction of jointly sparse signals
using periodic nonuniform sampling,” Asilomar Conf. Signals, Syst. using iterative hard thresholding,” IEEE Int. Conf. Commun., pp. 3564–
and Computers, pp. 1134–1138, 2007. 3568, 2012.
[77] C. G. Khatri and C. R. Rao, “Solutions to some functional equations [102] M. F. Duarte, S. Sarvotham, D. Baron, M. B. Wakin, and R. G.
and their applications to characterization of probability distributions,” Baraniuk, “Distributed compressed sensing of jointly sparse signals,”
Sankhyā: The Indian J. Statistics, Series A, pp. 167–180, 1968. IEEE Asilomar Conf. Signals, Syst. and Computers, pp. 1537–1541,
[78] M. Arts, A. Bollig, and R. Mathar, “Analytical test statistic distributions 2005.
of the mmme eigenvalue-based detector for spectrum sensing,” in [103] Z. Tian, “Compressed wideband sensing in cooperative cognitive radio
2015 International Symposium on Wireless Communication Systems networks,” IEEE Global Commun. Conf., pp. 1–5, Dec. 2008.
(ISWCS). IEEE, 2015, pp. 496–500. [104] F. Zeng, C. Li, and Z. Tian, “Distributed compressive spectrum sensing
[79] D. Romero, R. López-Valcarce, and G. Leus, “Compression limits for in cooperative multihop cognitive networks,” J. Select. Topics Signal
random vectors with linearly parameterized second-order statistics,” Process., vol. 5, pp. 37–48, Feb. 2011.
IEEE Trans. Inf. Theory, vol. 61, no. 3, pp. 1410–1425, 2015. [105] S. Boyd, A. Ghosh, B. Prabhakar, and D. Shah, “Randomized gossip
[80] J. Leech, “On the representation of 1, 2,, n by differences,” J. London algorithms,” IEEE Trans. Inf. Theory, vol. 52, pp. 2508–2530, 2006.
Mathematical Soc., vol. 1, no. 2, pp. 160–169, 1956. [106] B. J. M. Rabi and M. Johansson, “A randomized incremental subgradi-
[81] W. A. Gardner, “The spectral correlation theory of cyclostationary time- ent method for distributed optimization in networked systems,” SIAM
series,” Signal Process., vol. 11, pp. 13–36, Jul. 1986. J. Optimization, vol. 20, no. 3, pp. 1157–1170, 2009.
[82] P. S. Aparna and M. Jayasheela, “Cyclostationary feature detection in [107] N. Nguyen, D. Needell, and T. Woolf, “Linear convergence of
cognitive radio using different modulation schemes,” Int. J. Computer stochastic iterative greedy algorithms with sparse constraints,” CoRR,
Applications, vol. 47, pp. 12–16, Jun. 2012. vol. abs/1407.0088, 2014. [Online]. Available: http://arxiv.org/abs/
[83] S. Y. Kumar and M. Z. A. Khan, “Robust, block based cyclostationary 1407.0088
detection,” Int. Conf. Telecommunications, pp. 1–5, May 2013. [108] D. D. Ariananda and G. Leus, “Compressive joint angular-frequency
[84] A. V. Dandawate and G. B. Giannakis, “Statistical tests for presence of power spectrum estimation,” Proc. European Signal Process. Conf., pp.
cyclostationarity,” IEEE Trans. Signal Process., pp. 2355–2369, 1994. 1–5, Sep. 2013.
[85] M. Ghozzi, F. Marx, M. Dohler, and J. Palicot, “Cyclostatilonarity- [109] A. A. Kumar, S. G. Razul, and C. S. See, “An efficient sub-Nyquist
based test for detection of vacant frequency bands,” Int. ICST Conf. receiver architecture for spectrum blind reconstruction and direction
Cognitive Radio Oriented Wireless Networks and Commun., pp. 1–5, of arrival estimation,” IEEE Int. Conf. Acoustics, Speech and Signal
Jun. 2006. Process., pp. 6781–6785, May 2014.
 27

[110] J.-F. Gu and P. Wei, “Joint SVD of two cross-correlation matrices to

achieve automatic pairing in 2-D angle estimation problems,” IEEE
Antennas and Wireless Propagation Lett., vol. 6, 2007.
[111] J.-F. Gu, W.-P. Zhu, and M. N. S. Swamy, “Joint 2-D DOA estimation
via sparse L-shaped array,” IEEE Trans. Signal Process., vol. 63, Mar.
2015.