IEEE SIGNAL PROCESSING MAGAZINE SPECIAL ISSUE: “SIGNAL PROCESSING FOR THE ISAC REVOLUTION” 1

From OTFS to AFDM: A Comparative Study

of Next-Generation Waveforms for ISAC
in Doubly-Dispersive Channels

Hyeon Seok Rou, Graduate Student Member, IEEE,

arXiv:2401.07700v2 [eess.SP] 12 Feb 2024

Giuseppe Thadeu Freitas de Abreu, Senior Member, IEEE, Junil Choi, Senior Member, IEEE,
David González G., Senior Member, IEEE, Marios Kountouris, Fellow, IEEE,
Yong Liang Guan, Senior Member, IEEE, and Osvaldo Gonsa.

Abstract

Next-generation wireless systems will offer integrated sensing and communications (ISAC) function-
alities not only in order to enable new applications, but also as a means to mitigate challenges such as
doubly-dispersive channels, which arise in high mobility scenarios and/or at millimeter-wave (mmWave)
and Terahertz (THz) bands. An emerging approach to accomplish these goals is the design of new
waveforms, which draw from the inherent relationship between the doubly-dispersive nature of time-
variant (TV) channels and the environmental features of scatterers manifested in the form of multipath
delays and Doppler shifts. Examples of such waveforms are the delay-Doppler domain orthogonal time
frequency space (OTFS) and the recently proposed chirp domain affine frequency division multiplexing
(AFDM), both of which seek to simultaneously combat the detrimental effects of double selectivity and
exploit them for the estimation (or sensing) of environmental information. This article aims to provide a
consolidated and comprehensive overview of the signal processing techniques required to support reliable
ISAC over doubly-dispersive channels in beyond fifth generation (B5G)/sixth generation (6G) systems,
with an emphasis on OTFS and AFDM waveforms, as those, together with the traditional orthogonal
frequency division multiplexing (OFDM) waveform, suffice to elaborate on the most relevant properties
of the trend. The analysis shows that OTFS and AFDM indeed enable significantly improved robustness
against inter-carrier interference (ICI) arising from Doppler shifts compared to OFDM. In addition, the
inherent delay-Doppler domain orthogonality of the OTFS and AFDM effective channels is found to
provide significant advantages for the design and the performance of integrated sensing functionalities.
IEEE SIGNAL PROCESSING MAGAZINE SPECIAL ISSUE: “SIGNAL PROCESSING FOR THE ISAC REVOLUTION” 2

I. I NTRODUCTION
It is expected that beyond fifth generation (B5G) and sixth generation (6G) wireless systems will employ
extremely high-frequency (EHF) technologies, operating in the millimeter-wave (mmWave) and Terahertz
(THz) bands [1] as a means to support applications [2], such as Internet-of-Things (IoT), edge computing
and smart cities; and scenarios such as vehicle-to-everything (V2X), high-speed rail, and non-terrestrial
networks (NTNs), which are often subjected to heterogeneous and high-mobility conditions [3]. High-
mobility scenarios are known to pose a significant challenge to wireless communications systems due to
the resulting doubly-dispersive wireless channel, also referred to as time-variant (TV) multipath, or time-
frequency selectivity [4]. Such heterogeneous scattering environments deteriorate the received signal in
the form of path delays and Doppler shifts, resulting in inter-symbol interference (ISI) and inter-carrier
interference (ICI) which can drastically decrease communication performance under conventional and
highly effective modulation schemes, such as orthogonal frequency division multiplexing (OFDM) [5].
Concomitant with this challenge, there is a growing expectation that B5G and 6G systems will
offer integrated sensing and communications (ISAC) capabilities, possibly with unified hardware and
signal processing techniques [6]. In addition to providing environment perception and accurate/reliable
localization information to serve the aforementioned applications, the enhancements introduced by ISAC
are fundamental to improve spectrum and energy efficiency, and to lower hardware costs of systems
operating in high-mobility scenarios [7].
While it is difficult to foresee which of the upcoming generations/standards – i.e., B5G or 6G – will
see ISAC adopted and implemented into commercial systems, the topic is one of the most intensively
discussed among pre-standardization fora on wireless systems in recent years1 , with notable examples
being the 6G Smart Networks and Services Industry Association (6G-IA), where ISAC has been identified
as a priority technology for its members, and the 5G Automotive Association (5GAA), where ISAC is
considered an enabling technology for Cellular-V2X (C-V2X) services. Although it is reasonable to
anticipate that any form of practically deployed fifth generation (5G)-based ISAC will likely leverage
OFDM, more specifically cyclic prefix (CP)-OFDM and DFT-spread OFDM (DFT-s-OFDM) for down-
link/sidelink, and uplink, respectively, but for 6G, new waveforms, such as orthogonal time frequency
space (OTFS) and/or affine frequency division multiplexing (AFDM), should be considered to fully exploit
the benefits of ISAC. In fact, important standardization bodies such as European Telecommunications
Standards Institute (ETSI) and the 3rd generation partnership project (3GPP), which produce technical
specifications for mobile broadband systems worldwide, have recently added ISAC to their work plans
and roadmaps, with ETSI launching a new group dedicated to ISAC in November 2023.
1
See https://5gaa.org/ and https://6g-ia.eu/ for additional information on 5GAA and 6G-IA, respectively.
IEEE SIGNAL PROCESSING MAGAZINE SPECIAL ISSUE: “SIGNAL PROCESSING FOR THE ISAC REVOLUTION” 3

In line with this trend, novel waveforms have been recently proposed which, thanks to their ability
to retain symbol orthogonality under doubly-dispersive conditions, are both robust to high-mobility and
advantageous for ISAC, as they inherently enable the estimation of environmental parameters, such as
distance and velocity of scatterer objects (i.e., delay and Doppler shifts). One of the most popular methods
is OTFS signaling [8], which leverages the inverse symplectic finite Fourier transform (ISFFT) in order
to modulate a two-dimensional (2D) grid of information symbols directly in the delay-Doppler domain,
gaining great attention for high-mobility B5G systems thanks to its superior performance compared to
currently used waveforms such as OFDM [9].
It is easy to show, indeed, that the full delay-Doppler representation of the channel in OTFS inherently
conveys the velocity and range information of the scatterers in the form of the respective multipath delays
and Doppler shifts, thus implying significant benefits in terms of ISAC. As a consequence, a plethora
of OTFS-based ISAC techniques have been proposed to extract the delay and Doppler parameters of the
resolvable paths directly from the channel state information (CSI), which have been shown to compete
with the sensing performances of OFDM and frequency modulated continuous wave (FMCW) radars,
with higher robustness to mobility and achievable capacity [10].
An alternative strategy to design ISAC-friendly and mobility-robust waveforms is to employ chirp-based
multicarrier approaches [11]. While the chirp-domain design is attractive due to the inherent spread-
spectrum property and potential for full-duplex operations, an important and common drawback of these
earlier approaches is the lack of adaptability to the channel delay and Doppler spreads, which is a
consequence of the non-parametrizable transforms in their design.
A more recent take on the idea, which seeks to mitigate the latter drawback, is the AFDM waveform
[12], which leverages the inverse discrete affine Fourier transform (IDAFT) [13] in order to modulate
information symbols into a twisted time-frequency domain, yielding the desired delay-Doppler orthogo-
nality while maintaining the necessary flexibility. The optimizable parametrization of AFDM is further
accompanied by other desirable properties, such as full diversity guarantee and increased throughput [12],
making AFDM a strong candidate of ISAC-enabling waveform for B5G and 6G systems.
This article aims to offer a thorough analysis of the fundamentals and the future of ISAC technology
in heterogeneous high-mobility scenarios, in the form of a comprehensive comparison of prominent
candidate waveforms, focusing on OTFS and AFDM. The analysis reveals that the novel delay-Doppler
orthogonal designs of OTFS and AFDM benefit the signal processing for both communication and sensing
functionalities, advocating the integration of the two. These insights may hold significant interest and
value not only for academia, but also for standardization engineers across various industry verticals who
are increasingly participating in the development of future generations of mobile broadband systems.
IEEE SIGNAL PROCESSING MAGAZINE SPECIAL ISSUE: “SIGNAL PROCESSING FOR THE ISAC REVOLUTION” 4

The remainder of the article is organized as follows: the fundamental system models and the required
ISAC signal processing techniques for the doubly-dispersive wireless channel2 , emphasizing the inherent
transformations between time, frequency, delay, and Doppler dimensions, are described in Section II.
In Section III, the signal models of the identified candidate waveforms for B5G/6G ISAC in doubly-
dispersive environments are consolidated, highlighting their interrelationships in terms of the multiplexing
domain, transmitter structure, and the core linear canonical transform (LCT). In Section IV, we discuss
the radar sensing techniques leveraging the identified candidate waveforms in terms of the radar target
detection problem (DP) and radar parameter estimation problem (EP), elaborating on signal processing
techniques and solutions categorized into correlation-based methods, and direct/indirect CSI-based ap-
proaches. In Section V, the candidate waveforms are compared with basis on different key performance
indicators (KPIs) for both communications and radar sensing performances, in addition to implications
onto hardware implementation, requirements, and potential challenges. Finally, the key insights provided
by the article are summarized, and some future directions of the research are identified.

II. S IGNAL P ROCESSING F UNDAMENTALS OF D OUBLY-D ISPERSIVE C HANNELS

A long history of research on wireless communications has resulted in the identification and character-
ization of two fundamental and distinct types of small-scale fading effects, namely frequency- and time-
selectivity, also known as time and frequency dispersion, respectively. In particular, an electromagnetic
(EM) signal propagated through a given path is subject to a specific path delay proportional to the total
propagation distance between the transmitter and the receiver, and a Doppler shift3 proportional to the
relative velocities among transmitter, receiver, and scatterer, and the carrier frequency.
In a channel with multiple distinguishable propagation paths, the different copies of the originally
transmitted signal with varying time delays and Doppler shifts are superposed at the receiver, resulting
in interference that impacts on the reliability and performance of the wireless communication link,
unless appropriate signal processing techniques are employed. In this section, we first consolidate the
fundamental doubly-dispersive channel model with all of its representations in the time, frequency, delay,
and Doppler domains, along with the associated transformation methods, followed by the corresponding
signal processing mechanisms available to process the received signal, by the efficient representation of
the input-output relationship leveraging a circular convolution matrix.
2
The importance of modelling the doubly-dispersive channel, especially for ISAC applications, is a highly relevant problem
currently discussed in both academia and standardization [14].
3
We remark that in the related literature, and therefore also in this article, the term Doppler shift is often used in a broad
sense, including spectral shifts of the propagated signal resulting from phenomena other than the actual Doppler effect, such as
frequency offsets and low-frequency phase noise at the local oscillators.
IEEE SIGNAL PROCESSING MAGAZINE SPECIAL ISSUE: “SIGNAL PROCESSING FOR THE ISAC REVOLUTION” 5

A. The Doubly-Dispersive Channel Model

Consider a wireless channel between a transmitter and receiver with P resolvable propagation paths,
where each p-th path, with p ∈ {1, · · · , P }, is respectively described by a corresponding complex fading
coefficient hp ∈ C, path delay, τp ∈ [0, τ max ], and Doppler shift νp ∈ [−ν max , +ν max ]. The corresponding
delay and Doppler spreads of such a doubly-dispersive channel are characterized by the maximum delay
τ max [s] and the maximum Doppler shift ±ν max [Hz], such that the channel can be described by the linear
time-variant (LTV)4 relationship between the input and the output signals, most commonly represented
as a time-variant impulse response function (TVIRF) in the time-delay domain, given by [4]
P
X
h(t, τ ) ≜ hp · ej2πνp t · δ(τ − τp ), (1)
p=1
√
where j ≜ −1 is the elementary imaginary number, t and τ denote the instantaneous time and path
delay, respectively, and δ(x) is the unit impulse function defined by δ(x) = 1 iff x = 0.
Alternatively, the TVIRF in the time-delay domain can also be represented in other domains by
leveraging appropriate linear transforms [13]. For example, the representation in the time-frequency
domain is known as time-variant transfer function (TVTF), which is obtained by a Fourier transform
(FT) on the TVIRF over the delay domain, i.e.,
Z +∞ P
X
−j2πτ f
H(t, f ) ≜ F [h(t, τ )] = h(t, τ ) · e dτ = hp · ej2πνp t · e−j2πτp f , (2)
τ→f −∞ p=1

where f is the instantaneous frequency and F[ · ] denotes the continuous FT operator.

The TVTF in the time-frequency domain readily highlights both time and frequency dispersion effects
of the channel, visible in the two fast-varying exponential terms dependent on the instantaneous time t
and instantaneous frequency f , respectively at a rate of the Doppler frequency νp and delay τp of the
corresponding p-th propagation path, as illustrated in Fig. 1a. Conversely, the Doppler-variant impulse
response function (DVIRF) in the delay-Doppler domain is obtained by an FT on the TVIRF over the
time domain, that is
Z +∞ P
X
−j2πνt
γ(ν, τ ) ≜ F [h(t, τ )] = h(t, τ ) · e dt = hp · δ(ν − νp ) · δ(τ − τp ), (3)
t→ν −∞ p=1

where the time- and frequency-selectivity characteristics are observed in the form of unique impulses in
the delay-Doppler plane corresponding to each propagation path, as illustrated in Fig. 1b.

4
The term “linear time-variant (LTV) system” is not to be confused with linear systems with time-varying delays – i.e.,
systems with delay drifts, where τp (t)) – which are also commonly described as LTV. In this article, we only consider time-
invariant delays, in compliance with the related literature on doubly-dispersive channels, e.g., [4], [8], [12].
IEEE SIGNAL PROCESSING MAGAZINE SPECIAL ISSUE: “SIGNAL PROCESSING FOR THE ISAC REVOLUTION” 6

4 1.5

3
1
jH(t; f )j

j.(t; f )j
2

0.5
1

5.905 1.5
0 0
1
0 -2
0.5 5.9 -1
1 0.5
1.5 0
2 1
2.5 Frequency [GHz] Delay [7s]
Time [ms] 3 5.895 Doppler shift [kHz] 2 0

(a) TVTF in the time-frequency domain. (b) DVIRF in the delay-Doppler domain.

Fig. 1: The doubly-dispersive channel representations with P = 3 resolvable paths, with carrier frequency
of 5.9 GHz and signal bandwidth of 10 MHz (following the IEEE 802.11p vehicular environment
specifications). The different paths are illustrated by unique colors in Fig. 1b.

Fig. 2 provides a diagram that summarizes the relationships between the various signal domains,
including the direction and the integral domain of the necessary linear transforms. Specifically, the
rhombus-shaped relationship at the center of the figure illustrates the different domain representations
of the doubly-dispersive channel as described above, which also includes the omitted Doppler-variant
transfer function (DVTF)5 in the Doppler-frequency domain.
The Wigner transform (WT) and the Heisenberg transform (HT) (illustrated in red) are generalizations
of the multiplexing (MX) and demultiplexing (DMX) operations of the classical OFDM modulator, which
transform a 2D time-frequency domain signal into the single time and frequency domains. As will be
discussed in the following section, the two transforms can be respectively implemented using the discrete
Fourier transform (DFT) and the inverse discrete Fourier transform (IDFT). Furthermore, as can be seen
in the figure, there also exist linear transforms that directly describe concatenated FTs and/or inverse
Fourier transforms (IFTs). Such linear transforms, such as the symplectic finite Fourier transform (SFFT)
and Zak transform (ZT), are leveraged in the transmitter design of next-generation waveforms such as
the OTFS [8], and are elaborated in Sec. III.
To wrap up the signal domain fundamentals, let us also address the linear canonical transform (LCT)
[13], also known as the affine Fourier transform (AFT), which is a four-parameter transform generalizing6
many of the popular transforms such as the FT, Laplace transform (LT), and Fresnel transform (fnT).
5
The Doppler-frequency domain DVTF is not addressed as often compared to the other three forms, due to its lesser intuitive
relationship with the physical phenomena. However, it is still an equally valid representation of the doubly-dispersive channel.
6 1 j
Setting specific parameters reduces the LCT to the classical transforms such as (0, 2π
, −2π, 0) for the FT, (0, 2π
, j2π, 0)
1
for the Laplace transform (LT), and (cosθ, 2π
sinθ, −2πsinθ, cosθ) to yield the θ-th order fractional Fourier transform (frFT).
IEEE SIGNAL PROCESSING MAGAZINE SPECIAL ISSUE: “SIGNAL PROCESSING FOR THE ISAC REVOLUTION” 7

FT

t↔f
IFT
WT (DMX) WT (DMX)
RF TV-TF t, f ↔ f
TF
t ↔ t, f
h(t) H(t, f ) H(f )
HT (MX) FT IFT HT (MX)

τ ↔f t↔ν
IFT FT
SFFT
TV-IRF DV-TF
t, f ↔ ν, τ
h(t, τ ) Γ (ν, f )
IFT ISFFT FT

t↔ν τ ↔f
IZT FT IFT ZT
DV-IRF
t ↔ ν, τ ν, τ ↔ f
γ(ν, τ )
ZT IZT

Fig. 2: Illustration of the relationship between the different signal domains and inherent transforms.
In addition to the illustrated domains, a special case of the time-frequency domain, namely the chirp
domain, also exists, which is omitted in this diagram but elaborated upon in the following sections.

In particular, the AFT L [ · ] of a time-domain signal s(t) is described by

t→u

+∞
1
Z
a 2
+ 1b ut+ 2b
d 2

 s(t) · p · e−j( 2b u t )
dt, b ̸= 0,
2π|b|

−∞
 
L s(t) ≜ (4)
t→u 1 cd 2
s(d · u) · √ · e−j 2 u

 b = 0,
a

where the four AFT parameters (a, b, c, d) are arbitrary complex scalars satisfying ad − bc = 1.
As shall be seen, the available degrees-of-freedom (DoF)s of the AFT are exploited by another
promising waveform, AFDM [12], which allows for the optimization of the AFT parameters based on
the channel statistics in order to ensure the orthogonality of the signal in doubly-dispersive systems.

B. Input-Output Relationship of Doubly-Dispersive Channel

The input-output relation of the doubly-dispersive wireless channel in time is classically described by
the linear convolution over the delay domain, leveraging the TVIRF representation [4], namely
Z +∞ P
X 
r(t) = s(t) ∗ h(t, τ ) + w(t) ≜ s(t − τ ) hp · ej2πνp t · δ(τ − τp ) dτ + w(t), (5)
−∞ p=1

where and r(t), s(t), and w(t) are respectively the received signal, input signal, and additive white
Gaussian noise (AWGN) in time, while ∗ denotes the linear convolution operator.
IEEE SIGNAL PROCESSING MAGAZINE SPECIAL ISSUE: “SIGNAL PROCESSING FOR THE ISAC REVOLUTION” 8

In turn, the discrete equivalent of eq. (5) is given by

∞ P
X
ν

j2π fp n τp
X
r[n] = s[n − ℓ] hp · e s · δ(ℓ − ∆τ ) + w[n], (6)
ℓ=0 p=1

1
where r[n] and s[n] are respectively the sampled sequences of r(t) and s(t) at a sampling rate of fs ≜ Ts

[Hz], with sampling interval Ts [s]; n ∈ {0, · · · , N − 1} is the sample index; and ℓ ≜ Tτs is the normalized
delay with delay resolution Ts .
We further introduce the normalized digital Doppler shift and normalized delay of the p-th propagation
N νp τp
path as fp ≜ fs and ℓp ≜ Ts , where the delay resolution Ts is assumed to be sufficiently high, such that
the normalized delay ℓ can be rounded to the nearest integer with negligible error, i.e., ℓp − ⌊ Tτps ⌉ ≈ 0. In
practical multicarrier wireless communications techniques, the transmit sequence in eq. (6) is prepended
with a CP to mitigate the effects of time dispersion. The prefix sequence is defined within a CP length
of Ncp samples, such that
′
s[n′ ] = s[N − n′ ] · ej2π·ϕcp (n ) , (7)

where n′ ∈ {1, · · · , Ncp }, and ϕcp (n′ ) is a function denoting the multiplicative phase term specific for
each waveform, which is set to zero if the CP does not require a phase offset, as in the OFDM.
The CP as described in eq. (7) enables the linear convolutional input-output relation of the TVIRF to
be processed as a circular convolutional response. After removing the received signal parts corresponding
to the CP, the circular convolutional input-output relationship can be described in matrix form as
P
X 
r ≜ H·s = hp ·Φp ·Wfp ·Πℓp · s + w ∈ CN ×1 , (8)
p=1

where r ∈ CN ×1 , s ∈ CN ×1 , and w ∈ CN ×1 are respectively the vectors representing the received signal, the
transmit signal, and AWGN; H ∈ CN ×N is the circular convolution effective channel matrix; Φp ∈ CN ×N
is the diagonal matrix corresponding to the p-th delayed CP phase as given in eq. (9); W ∈ CN ×N is the
diagonal matrix containing the N -th roots of unity as given in eq. (10); and Π ∈ CN ×N is the forward
cyclic shift matrix obtained by left-shifting the N × N identity matrix once, such that Πℓp corresponds
to a left-shift operation of a matrix by ℓp ∈ N0 indices.
As can be seen in eq. (8), the input-output relationship of a doubly-dispersive channel is described by

ℓp terms N −ℓp ones

 z }| { z }| { 
−j2π·ϕcp (ℓp ) −j2π·ϕcp (ℓp −1) −j2π·ϕcp (2) −j2π·ϕcp (1)
Φp ≜ diag e ,e ,··· ,e ,e , 1 , 1 , · · ·, 1 , 1 ∈ CN ×N. (9)
 
W ≜ diag 1, e−j2π/N , · · · , e−j2π(N −2)/N , e−j2π(N −1)/N ∈ CN ×N . (10)
IEEE SIGNAL PROCESSING MAGAZINE SPECIAL ISSUE: “SIGNAL PROCESSING FOR THE ISAC REVOLUTION” 9

a matrix H consisting of P off-diagonals, whose shifted positions are determined by the integer delay of
each path. In addition, the complex values along each diagonal contain the channel fading coefficient and
the phase offset information of the delayed CP and the Doppler shift of each path. As a consequence of
the circulant convolutional channel structure, the different paths are only resolvable in the delay domain
and not in the Doppler domain. Therefore, a key design objective of a double-dispersion robust waveform
must be the ability to orthogonalize the delays and Doppler shifts of the channel via means of novel
domain transforms in the modulation and demodulation of the transmit signal.

III. S IGNAL M ODELS OF N EXT-G ENERATION WAVEFORMS

In this section, we provide the signal models and the transmitter structures of the various waveforms
proposed for high performance in the doubly-dispersive channel. First, the OFDM waveform is described
as a reference, followed by the next-generation waveform candidates, namely, OTFS, and AFDM, in
addition to various derivative waveforms that can be related to the latter.

A. Orthogonal Frequency Division Multiplexing (OFDM)

The well-known OFDM transmitter modulates digital symbols from the frequency domain into a time
domain signal employing an IDFT operation, as illustrated in Fig. 2. Namely, given a vector x ∈ CN ×1
consisting of N complex symbols, the OFDM transmit signal sOFDM ∈ CN ×1 is given by [15]

sOFDM = F−1
N ·x∈C
N ×1
, (11)

where FN ∈ CN ×N is the N -point DFT matrix, and hence F−1 H

N ≜ FN is the N -point IDFT matrix.

Following the above, the received signal yOFDM ∈ CN ×1 over the circular convolutional channel
described by eq. (8) is demodulated via the forward N -point DFT, i.e.,

≜ rOFDM ∈ CN ×1 ≜ GOFDM ∈ CN ×N
z }| z
{ }| {
yOFDM = FN H · sOFDM + w = (FN · H · F−1
N ) x + FN · w ∈ C
N ×1
, (12)

where, for exposition convenience, we defined the OFDM receive signal as rOFDM ∈ CN ×1 , and the
effective channel GOFDM ∈ CN ×N describing the input-output relationship of the baseband OFDM
symbols, which can be obtained by combining eqs. (8) and (12) to yield
P
X  
OFDM
G ≜ hp · FN Φp ·Wfp ·Πℓp FH
N. (13)
p=1

As illustrated in Fig. 3, the column-wise DFT and row-wise IDFT in the presence of Doppler shifts
cause the channel diagonals of the convolution matrix to be spread into a decaying band, centered at the
original diagonals, such that significant interference between each path may arise.
IEEE SIGNAL PROCESSING MAGAZINE SPECIAL ISSUE: “SIGNAL PROCESSING FOR THE ISAC REVOLUTION” 10

B. Orthogonal Time Frequency Space (OTFS)

In the OTFS modulation scheme, the information symbols are first directly placed in the delay-Doppler
domain, instead of the frequency domain as in the OFDM approach, which are then multiplexed into the
time signal. Namely, the complex baseband symbols are structured into a 2D grid of size K × L in the
delay-Doppler domain, which are first transformed into the time-frequency domain via the ISFFT, and
then into the continuous time signal via a pulse-shaping HT. As illustrated in Fig. 2, such a two-step
OTFS domain transformation process can also be achieved via a single inverse discrete Zak transform
(IDZT), which is an alternative and equivalent representation of the OTFS modulation process [8]. For
the sake of clarity, however, we will hereafter adopt the ISFFT formulation, known to be implemented
via DFTs/IDFTs.
In light of the above, the OTFS modulation process can be described mathematically as
Pulse-shaping HT ISFFT ≜x
 z }| { z }|
{  z }| {
tx −1 −1 −1
sOTFS ≜ vec P FK · FK XFL tx
= (FL ⊗ P ) · vec(X) ∈ CKL×1 , (14)

where sOTFS ∈ CKL×1 is the OTFS transmit signal vector, X ∈ CK×L is the information symbol matrix
consisting of N ≜ KL number of complex symbols7 , Ptx ∈ CK×K is the diagonal transmit pulse-
shaping filter matrix, FK ∈ CK×K and FL ∈ CL×L are the K -point and L-point DFT matrices, and
vec(·) and ⊗ denote the stacking vectorization and Kronecker product operators, respectively.
The filtered and demodulated signal yOTFS after the convolution channel H in eq. (8) is given by

≜ rOTFS ∈ CKL×1
}|z
{
yOTFS ≜ (FL ⊗ Prx ) H·sOTFS + w = GOTFS ·x + (FL ⊗ Prx )·w ∈ CKL×1 , (15)

where Prx ∈ CK×K is the diagonal matched filter matrix of Ptx , and the effective OTFS channel
GOTFS ∈ CN ×N in the delay-Doppler domain is given by
P
X
 
GOTFS ≜ hp · FL ⊗ Prx Φp ·Wfp ·Πℓp FH tx
∈ CN ×N ,


L⊗P (16)
p=1

which is the convolutional channel matrix H after a block-wise pulse-shaped8 DFT and IDFT.
It can be observed from eq. (16) that the KL = N elements in each diagonal of the convolutional
matrix of eq. (8) are spread into the OTFS effective channel via the block-wise pulse-shaping FTs, such
that the KL × KL OTFS channel matrix can be considered as a K × K grid of L × L sub-matrices
(illustrated as minor grids in Fig. 3). In light of the above, the positions of the non-zero channel coefficient

7
Without loss of generality, we assume N ≜ KL to enable direct comparison with 1D modulation schemes, i.e., OFDM.
8
The literature commonly assumes rectangular OTFS pulses [8], which reduces Ptx and Prx to K × K identity matrices.
IEEE SIGNAL PROCESSING MAGAZINE SPECIAL ISSUE: “SIGNAL PROCESSING FOR THE ISAC REVOLUTION” 11

elements can be deterministically obtained by the values of both delay and Doppler shift of each path,
respective to the occupied sub-matrices and the amount of left-shift of the diagonals.
For the integer Doppler shift model, each path occupies exactly K sub-matrices out of the K × K
grid in a shifted block-diagonal structure, such that the amount of right-shift is determined by the integer
value of the Doppler shift, fpint ≜ ⌊fp ⌋. For example, a path with fpint = 0 occupies the K sub-matrices
in the main block-diagonal (shift of index 0), whereas a path with fpint = 1 occupies the K sub-matrices
in the block-diagonal, which is right-shifted by an index of 1. On the other hand for negative Doppler
shifts, the block-diagonals are left-shifted by an index of |fpint |, as illustrated by path 3 in Fig. 3a.
In turn, each of the K occupied sub-matrices follows the same structure consisting of exactly L non-
zero elements in a shifted diagonal, with a left-shift relative to the main diagonal determined by the value
of the path delay ℓp . For example, for a path with delay ℓp = 0, the L non-zero elements are placed in
the main diagonal for all K sub-matrices, whereas a path with ℓp = 3 will have the L non-zero elements
in the diagonal left-shifted by three indices, for all K sub-matrices9 .
It follows that the OTFS waveform can achieve orthogonality and resolvability in the delay-Doppler
domain with integer delay and integer Doppler shifts, given ℓmax ≤ L − 1 and f max ≤ ⌊ K2 ⌋, where
 max   max 
ℓmax ≜ τ∆τ and f max ≜ N νfs are the maximum normalized delay and Doppler shift.
In contrast, in the case of fractional Doppler, the powers of the channel elements are diffused (or
“leaked”) across all K 2 sub-matrices over the main K block-diagonal sub-matrices of the integer case,
resulting in a Doppler domain interference as illustrated in Fig. 3b and Fig. 3c. The amount of such
power leakage is determined by the value of the fractional part of the Doppler shift given by fpfrac ≜
fp − fpint ∈ [0.5, +0.5), such that larger fractional parts result in more leakage.

C. Affine Frequency Division Multiplexing (AFDM)

In AFDM, a one-dimensional vector of symbols x ∈ CN ×1 is directly multiplexed into a twisted

time-frequency chirp domain using the IDAFT [16], as described by
IDAFT
z }| {
sAFDM ≜ A−1 ·x = (Λc2·FN ·Λc1 )−1 · x = (ΛH H H
c1 ·FN ·Λc2 )·x ∈ C
N ×1
, (17)

where A ≜ Λc2 FN Λc1 ∈ CN ×N is the forward N -point discrete affine Fourier transform (DAFT) matrix,
2 2
and Λci ≜ diag[e−j2πci (0) , · · · , e−j2πci (N −1) ] ∈ CN ×N is a diagonal chirp matrix with a central digital
frequency of ci .

9
A Matlab© implementation of the doubly-dispersive channel model described in this Section, as well as a convenient
visualization tool used to generate some of the figures can be found our online respository [here].
IEEE SIGNAL PROCESSING MAGAZINE SPECIAL ISSUE: “SIGNAL PROCESSING FOR THE ISAC REVOLUTION” 12

It follows that the demodulated AFDM signal over the convolution channel in eq. (8) is given by

≜rAFDM ∈ CN ×1
z }| {
yAFDM = A· (H·s AFDM
+ w) = GAFDM ·x + A·w ∈ CN ×1 , (18)

with the effective AFDM channel given by

P
AFDM
X
 
N ×N
Φp ·Wfp ·Πℓp ΛH H H


G ≜ hp · Λc2·FN ·Λc1 c1 ·FN ·Λc2 ∈ C , (19)
p=1

where the central frequencies c1 and c2 of the two diagonal chirps10 can be optimized to the channel
statistics to improve the orthogonality of the diagonals of the AFDM effective channel.
The AFDM effective channel achieves full orthogonality in the integer delay-Doppler domain when
the orthogonality condition is satisfied, which is given by

2 f max + ξ (ℓmax + 1) + ℓmax ≥ N,

(20)

where ξ ∈ N is a free parameter determining the so-called guard width of the AFDM, denoting the
number of additional guard elements around the diagonals to anticipate for Doppler-domain interference.
Assuming the latter orthogonality condition is met, the AFDM chirp frequencies are obtained by

2(f max + ξ) + 1 1
c1 = , and c2 << , (21)
2N N

where the flexibility in c2 enables fine-tuning of the waveform shape as will be discussed in Sec. IV.
In light of the above, the position of the shifted diagonal in the AFDM channel can also be described
in terms of the delay-Doppler indices of each path. Unlike the intricate block-wise structure of the OTFS
effective channel, the AFDM effective channel exhibits only a single diagonal per path, which is shifted
by a deterministic index dependent on the integer delay and integer Doppler shift. In other words, each
diagonal of the convolution channel in eq. (8) is right-shifted by an index of exactly ℓp ·2(f max +ζ)+fpint
positions, as illustrated in Fig. 3a. On the other hand, in the presence of fractional Doppler shifts, the
diagonals of the AFDM effective channels also exhibit power leakage around the main diagonal, resulting
in Doppler-domain interference as can be seen in Figs. 3b and 3c.

10
It is shown in [12] that the chirp frequencies c1 and c2 are actually correspondent to the four configurable parameters of
the AFT formulation in eq. (4).
IEEE SIGNAL PROCESSING MAGAZINE SPECIAL ISSUE: “SIGNAL PROCESSING FOR THE ISAC REVOLUTION” 13

OFDM OTFS AFDM

0 6 12 18 24 30 36 0 6 12 18 24 30 36 0 6 12 18 24 30 36
0 0 0

6 6 6

12 12 12

18 18 18

24 24 24

30 30 30

36 36 36

(a) Target parameters: {ℓ1 = 0, f1 = 0} (red), {ℓ2 = 1, f2 = −2} (blue), {ℓ3 = 3, f3 = +1} (yellow).

OFDM OTFS AFDM

0 6 12 18 24 30 36 0 6 12 18 24 30 36 0 6 12 18 24 30 36
0 0 0

6 6 6

12 12 12

18 18 18

24 24 24

30 30 30

36 36 36

(b) Target parameters: {ℓ1 = 0, f1 = 0.266} (red), {ℓ2 = 1, f2 = −2.365} (blue), {ℓ3 = 3, f3 = +1.231} (yellow).

OFDM OTFS AFDM

1 1 1

0.8 0.8 0.8

0.6 0.6 0.6

0.4 0.4 0.4

0.2 0.2 0.2

0

0 0 0
6

0 0 0
12

12

12

6 6 6
18

18

18

12 12 12
18 18 18
24

24

24

24 24 24
30

30

30

30 30 30
36

36

36

36 36 36

(c) Target parameters: {ℓ1 = 0, f1 = 0.266} (red), {ℓ2 = 1, f2 = −2.365} (blue), {ℓ3 = 3, f3 = +1.231} (yellow).

Fig. 3: Effective channel matrix structures of different waveforms in a doubly-dispersive channel with
P = 3 resolvable paths (each depicted in a different color), with corresponding normalized delays ℓp and
normalized digital Doppler shifts fp . The system size parameters are N = 36 for the OFDM and AFDM,
and K = 6, L = 6 for the OTFS. The fading colors for the fractional Doppler case correspond to the
magnitude of the elements whereas darker colors correspond to larger powers. Channel components with
a magnitude lower than 1/2N are considered negligible and not visualized in the figure.
IEEE SIGNAL PROCESSING MAGAZINE SPECIAL ISSUE: “SIGNAL PROCESSING FOR THE ISAC REVOLUTION” 14

D. Related Next-Generation Waveforms

In this section, we briefly discuss various waveforms which, as illustrated in Fig. 4 and put into a
chronological context in Fig. 5, are related to the aforementioned OFDM, OTFS, and AFDM, and can
also be potential candidates to support ISAC in B5G/6G systems. Due to space limitation, however, the
discussion is resumed to a qualitative comparison, addressed in more detail in Sec. III.
1) Intermediate Chirp Domain Waveforms: A few chirp-domain waveforms also exist, which in
commonality with OFDM and AFDM, aim at orthogonalizing delay and Doppler shift indices. Such
waveforms, which include orthogonal chirp division multiplexing (OCDM) [11] and DAFT-OFDM [17],
can be seen as special cases of AFDM, with non-ideal and simplified chirp frequencies c1 and c2 [12],
naturally exhibiting equal or worse performances depending on the doubly-dispersive channel profile.
2) Enhanced Delay-Doppler Waveforms: Various methods adopt the novel delay-Doppler signal rep-
resentation of OTFS, and have proposed enhanced delay-Doppler domain waveforms. Examples are the
transcendentally-rotated OTFS (T-OTFS) [18], which maximizes the asymptotic diversity of OTFS via a
phase-rotating precoder; orthogonal time sequency modulation (OTSM) modulation [19], which seeks to
reduce implementation complexity by leveraging a new type of domain transform; and orthogonal delay-
Doppler division multiplexing (ODDM) [20], which improves upon OTFS via an optimized pulse-shaping
filter that creates feasible pulses that are orthogonal with respect to the delay-Doppler plane resolution.
3) Filter Bank-based (Pulse-Shaping) Waveforms: Finally, various multicarrier techniques leverage
optimized pulse-shaping filter banks, such as filter bank multi-carrier (FBMC) [21] and generalized
frequency division multiplexing (GFDM) [22], which improve the out-of-band (OOB) emissions, spectral
efficiency, ISI, and ICI problems of OFDM via robust adaptation of the subcarriers and modulation pulses
with respect to the doubly-dispersive channel statistics.
Time-Frequency domain Delay-Doppler domain

Flexible subcarrier-wise pulse- Optimized pulse-shaping to

shaping given channel conditions finite delay-Doppler resolution
GFDM OFDM OTFS ODDM

Robustness to Doppler effect via

Full diversity Reduced complexity via
a fixed chirp-domain transform
via transcendental Walsh-Hadamard transform
phase rotation

FBMC DAFT- OTSM

OCDM T-OTFS
OFDM
Filter-Bank domain

Optimized subcarrier
chirp frequencies given Generalization / Optimization
AFDM channel conditions
Modified transform
Chirp domain Pulse-shaping

Fig. 4: A map of relationships between next-generation waveforms and their signal domains.
IEEE SIGNAL PROCESSING MAGAZINE SPECIAL ISSUE: “SIGNAL PROCESSING FOR THE ISAC REVOLUTION” 15

Early 1990s: 2005: 2016: 2019: 2021: 2022: 2024: 2025 – 2028 (Expected):
The OFDM is first adopted The DAFT-OFDM The OCDM 3GPP standardizes The OTSM 3GPP feasibility study ETSI Industry Spec. Introduction of ISAC as
into a commercial standard. is proposed. is proposed. mmW comm. is proposed. on ISAC use cases. Group for ISAC. native feature in 6G.

Late 1960s – Early 1970s: Late 1990s – Mid 2000s: 2009: 2016: 2019: 2021: 2022: 2023: 2024:
Fundamental concepts of OFDM The OFDM is adopted into The GFDM The OTFS The T-OTFS is The AFDM The ODDM 5GAA work 3GPP RAN study on
and FBMC are introduced. Wi-Fi and 4G LTE standards. is proposed. is proposed. proposed. is proposed. is proposed. item on ISAC. ISAC channel modelling.

Fig. 5: Timeline highlighting the invention of various waveforms, and key dates in standardization.

IV. I NTEGRATED S ENSING AND C OMMUNICATIONS (ISAC) USING

N EXT-G ENERATION WAVEFORMS

In ISAC, wireless “sensing” refers to the ability to harness the rich information about the surrounding
environment inherently embedded in radio signals affected by channel conditions. Drawing a parallel with
well-established radar technologies [23] two distinct types of sensing problems can therefore be identified,
namely: 1) the detection problem (DP), which relates to resolving the number of unique scattering points
of interest (targets) from the background clutter, and 2) the estimation problem (EP), which refers to
extracting parameters such as range, velocity, and bearing associated with the aforementioned targets.
Considering a single-antenna monostatic sensing scenario, where the transmitter and the receiver are
colocated11 , the scatterer (target) range r [m] and the relative linear velocity v [m/s] are related to the
round-trip delay τ and Doppler shift ν of the echo signal by
2r 2vfc
τ≜ , and ν≜ , (22)
c c
where c is the speed of light and fc is the carrier frequency of the transmitted signal.
For the sake of simplicity, we address here only the single-input single-output (SISO) system model, as
the multiple-input multiple-output (MIMO) extension can be trivially derived by leveraging array response
vectors and the corresponding signal processing techniques [25]. This enables the estimation of bearing,
and in the case of planar arrays, the azimuth and elevation angles of the target in 3D space.
Solutions to the DPs and EPs can be further classified into three methods based on their approaches:
a) correlation-based methods, in which echo signals are filtered with a known transmit signal in order
to yield radar parameter estimates; b) direct CSI-based methods, where radar parameters are extracted a
posteriori from the known channel matrix; and c) indirect CSI-based methods, where radar parameters
estimation and CSI acquisition are performed jointly, by leveraging the known channel structure. Each
of these approaches is discussed further in the sequel.

11
The bistatic scenario with distributed radar transmitter and receiver can also be considered for several applications and
use cases [24], where the estimated channels may be non-reciprocal. This scenario is beyond the scope of this article but is an
important setting to be addressed in future work, which can enable the full potential of ISAC.
IEEE SIGNAL PROCESSING MAGAZINE SPECIAL ISSUE: “SIGNAL PROCESSING FOR THE ISAC REVOLUTION” 16

A. Correlation-based Methods

Classical radar systems based on chirps and impulsive waveforms [23] are typical examples of the
correlation-based method, as the received echo signals are processed by a correlation (matched filter) with
a known probing signal, to directly yield the target parameters. Radar waveforms are, however, optimized
to exhibit correlation properties that can achieve high resolution in the delay and Doppler domains,
which is generally not possible to do with communication waveforms without sacrificing communication
objectives (e.g., rate, efficiency, latency, etc.).
The fundamental resolution of the correlation-based method can be analyzed through the well-known
2D ambiguity function of a waveform in delay-Doppler domain12 , which is given by
Z +∞
A(τ, ν) ≜ s(t) · s∗ (t − τ ) · ej2πνt dt, (23)
−∞

where A(τ, ν) is the 2D ambiguity function parametrized by the delay and Doppler shift values, s(t) is
the transmit signal, and (·)∗ denotes the complex conjugation operation.
By inspecting the ambiguity function behaviors of the OFDM, OTFS, and AFDM waveforms [26], the
different delay-Doppler resolution and beamlobe behaviors of the waveforms can be observed. Namely,
it is found that that OFDM shows high resolution in the delay domain, but not in Doppler, whereas
OTFS and AFDM show moderate resolution in both domains simultaneously, with AFDM exhibiting
an adjustable mainlobe width with a trade-off in the two domains by leveraging the chirp frequency
parameters c1 and c2 . Therefore naturally, correlation-based estimators of low complexity, employing
OFDM, OTFS, and AFDM [26]–[28], have been proposed, but were found to be still fundamentally
dependent on the resolution of the ambiguity function, promoting subsequent development of parameter
estimation methods based on the higher resolution of the inherent delay-Doppler modulation grid.

B. Direct CSI-based Methods

These methods operate under the assumption that channel estimation has been performed such that
the channel matrix is available, and aim to extract radar parameters a posteriori, by exploiting the
deterministic structure of the doubly-dispersive channel as described in Sec. III, which is dependent on
the target delay and Doppler shift parameters. Several sensing algorithms based on this approach have
been proposed. Assuming that CSI is obtained in the time-frequency domain as per eq. (2), the resulting
estimation problem on τp and νp is referred to as a multidimensional harmonic retrieval problem, to which
many well-known super-resolution solutions exist, such as MUltiple SIgnal Classification (MUSIC) and
estimation of signal parameters via rotational invariance techniques (ESPRIT), in addition to enhanced

12
In the case for MIMO, the ambiguity function is extended into three-dimensional (3D) including the angular domain.
IEEE SIGNAL PROCESSING MAGAZINE SPECIAL ISSUE: “SIGNAL PROCESSING FOR THE ISAC REVOLUTION” 17

tensor-based algorithms in the case of MIMO scenarios with increased dimensionality, as discussed
in [6]. The approach is known to achieve high resolution and accuracy but typically requires a large
number of continuously obtained observation samples, which together with the fundamental dependence
on the channel estimation performance, constitute a prominent challenge of such methods, since channel
estimation errors may propagate to the radar parameter estimates.
Setting these issues aside, with a sufficiently large number of observations and appropriate filtering,
the effective channel matrices of the waveforms as given in eqs. (13), (16), and (19) may be obtained
via compressive sensing and other matrix reconstruction algorithms. In such cases, the delay-Doppler
orthogonality of the OTFS and AFDM effective channels, in addition to the injective mapping between
each integer delay-Doppler pair and the position of the shifted-diagonal as discussed in Sec. III, enable
efficient extraction of the target parameters from the channel matrix element positions. However, in the
presence of fractional Doppler shifts, the resulting interference in the Doppler domain illustrated in Fig.
3b can significantly deteriorate such “position-based” methods in terms of Doppler resolution.

C. Indirect CSI-based Methods (Integrated Channel Estimation)

In cases where the channel matrix is not available but its structure is known, an optimization problem
can be formulated to jointly estimate the channel inherently incorporating the radar parameters, yielding
an integrated channel and target parameter estimation. For instance, the following minimization problem
can be formulated and solved

argmin L r − H̃(τp , νp , hp ; ∀p)·s , (24)
τp ,νp ,hp ∀p

where r ≜ Hs + w ∈ CN ×1 is the received signal, H̃(τp , νp , hp ; ∀p) ∈ CN ×N is the estimated channel

parametrized by the 3P radar parameters τp , νp , hp for p ∈ {1, · · ·, P }, and s ∈ CN ×1 is the transmit
signal, and L(·) is an arbitrary objective function, i.e., the L2 norm.
In the doubly-dispersive case, the parametrized channel H̃ is given by eq. (8), or pre-processed via
leveraging the effective channel models of specific waveforms described, e.g., by eqs. (13), (16), or (19)
for OFDM, OTFS and AFDM, respectively. Regardless, the underlying 3P -parameter estimation problem
in eq. (24) can be solved via various algorithms based on maximum likelihood, Bayesian learning,
compressive sensing, and basis expansion methods [29], [30]. Such methods are most robust and realistic
as the assumptions and the resolution of the former two types of methods may not be practical, and are
expected to be more beneficial to the communications subsystem due to the underlying joint estimation
of the channel coefficients.
IEEE SIGNAL PROCESSING MAGAZINE SPECIAL ISSUE: “SIGNAL PROCESSING FOR THE ISAC REVOLUTION” 18

V. K EY P ERFORMANCE I NDICATOR (KPI)- CENTERED C OMPARATIVE A NALYSIS

In view of all the above, we finall offer a qualitative comparison of OFDM, OTFS and AFDM –
respectively representing the classic, state-of-the-art (SotA) and most-recent alternative – ISAC-friendly
waveforms for B5G/6G systems. To this end, we consider various relevant KPIs for communication and
sensing, functions, both in terms of features and implementation aspects. The GFDM and FMCW wave-
forms are also included for comparison, for communications and radar sensing performances respectively.
The result is given in Table I, and while it is not possible to elaborate on all comparison points due to
space limitations, we briefly elaborate on a few most important of the selected KPIs. In particular, perhaps
the most important communications KPI in doubly-dispersive channels is the Doppler-shift robustness, i.e.,
the compatibility to high-mobility and EHF conditions, which is only attained by the OTFS and AFDM
waveforms due to the inherent delay-Doppler domain orthogonality. On this aspect, it is noteworthy than
OTFS achieves full diversity only in finite signal-to-noise ratio (SNR) regimes converging to first order
asymptotic diversity [18], while AFDM provides guaranteed full diversity generally [12] and is also
known to be the only Doppler-robust waveform to achieve full diversity also in MIMO scenarios. On the
other hand, due to their fundamental roots on OFDM, both AFDM and GFDM also suffer from higher
peak-to-average power ratio (PAPR), whereas OTFS enjoys a low PAPR due to the DFT-based spreading
of the symbol powers in the time-frequency domain. This advantage is closely linked to implementation
cost and hardware stability, especially in relation to the power amplifier (PA) and radio frequency (RF)
component efficiency which becomes more prominent in the massive MIMO scenarios.
Another important point to consider is the computational complexity, which is linked to various signal
processing procedures such as modulation and channel estimation. In this regard, while both OTFS and
AFDM can be interpreted as modified precoding schemes for OFDM transmitters, such that core OFDM
modulators can be reused, the one-dimensional (1D) AFDM modulator exhibits a higher efficiency than
the 2D OTFS modulator. This reduced dimension of the waveform also shows a similar advantage for
channel estimation, in terms of both computational complexity and the required piloting overhead.
In terms of target sensing performance, OTFS and AFDM exhibit a significant improvement in the
Doppler-domain ambiguity over OFDM, but such methods are restricted as they are not optimized
for correlation as with FMCW waveforms. Therefore, super-resolution methods and on-grid estimation
methods on the discrete delay-Doppler domain of the waveforms are leveraged, which can achieve
extremely high resolutions compared to those of FMCW, often used in automotive radar, given sufficient
parameterization13 such as the carrier frequency and symbol period.

13
It is expected that in B5G systems, the required parameters will be sufficiently satisfied to achieve high resolutions.
IEEE SIGNAL PROCESSING MAGAZINE SPECIAL ISSUE: “SIGNAL PROCESSING FOR THE ISAC REVOLUTION” 19

TABLE I: A comparative table of various waveforms and their ISAC KPIs, with qualitative measures:
high, medium, low. The color of each cell corresponds to the relative performance measure, ranging from
green denoting an attractive performance, yellow, to red denoting less performant.

Waveform
Key Performance Indicator
OFDM OTFS AFDM GFDM
Waveform Domain Time-frequency Delay-Doppler Chirp Filter-bank
Delay Robustness High High High High
Performance

Doppler Robustness Low High High Medium

Peak-to-Average Power Ratio High Low High Medium
Communications

Out-of-Band Emission Rate High Medium High Low

Diversity in TV Channels Low Medium High Medium
Frame Guard (CP) Overhead High Low Medium Medium
Pilot Guard Overhead Medium High Low Medium
Modulation Complexity Low High Medium High
Implementation

OFDM Compatibility High High High High

PA Strain High Low Medium Medium
MIMO Scalability High Medium High Medium
EHF Feasibility Low High High Medium
Full-Duplex Potential Low Medium High Medium

Key Performance Indicator OFDM OTFS AFDM FMCW

Delay Ambiguity Low Medium Variable Low
Doppler Ambiguity High Medium Variable Low
Performance

Peak-to-Sidelobe Ratio Low Medium Variable Low

Range Resolution High High High High
Target Sensing

Velocity Resolution Medium High High High

Max. Unambiguous Range Medium Medium Medium High
Max. Unambiguous Velocity Medium Medium Medium High
ISAC Implem.

Implementation Cost Low Medium Medium Low

Engineering Complexity Low Medium Medium Low
MIMO Array Extendibility High Medium High High
CSI Estimation Complexity Low Medium Medium High
ISAC Feasibility Medium High High Medium

The above-described properties in both communications and sensing performances must be satisfied and
be coherent, in order for a waveform to be considered a strong candidate for ISAC in B5G/6G. Clearly, as
observed from the color scaling of Table. I, OTFS and AFDM are the most promising candidates satisfying
most of the ISAC criteria, with some trade-offs between the two waveforms in terms of complexity, power
and spectral efficiency, which is to be further addressed in a future work.

VI. F UTURE W ORKS

It can be seen that while OTFS and AFDM are the most promising candidates to enable high-
performance ISAC for next-generation wireless networks, there are still many important topics to be
IEEE SIGNAL PROCESSING MAGAZINE SPECIAL ISSUE: “SIGNAL PROCESSING FOR THE ISAC REVOLUTION” 20

addressed in order to enable the incorporation of such techniques into future standards. Indeed, any given
row of Table I – such as the PAPR of AFDM or the CSI estimation complexity of OTFS – can be
a subject of optimization and development. The authors hope that this article helps the ISAC research
community with fundamental insights, techniques, and future direction to promote the development of
high-performance ISAC in doubly-dispersive environments for next-generation wireless networks.

R EFERENCES
[1] T. S. Rappaport et al., “Wireless communications and applications above 100 GHz: Opportunities and challenges for 6G
and beyond,” IEEE Access, vol. 7, pp. 78 729–78 757, 2019.
[2] N.-S. Vo, T. Q. Duong, and Z. Sheng, “The key trends in B5G technologies, services and applications,” Mobile Networks
and Applications, vol. 27, no. 4, pp. 1716–1718, 2022.
[3] H. Zhou, W. Xu, J. Chen, and W. Wang, “Evolutionary V2X technologies toward the internet of vehicles: Challenges and
opportunities,” Proceedings of the IEEE, vol. 108, no. 2, pp. 308–323, 2020.
[4] D. W. Bliss et al., Dispersive and Doubly Dispersive Channels. Cambridge University Press, 2013, pp. 341–364.
[5] T. Wang, J. Proakis, E. Masry, and J. Zeidler, “Performance degradation of OFDM systems due to Doppler spreading,”
IEEE Trans. Wireless. Commun., vol. 5, no. 6, pp. 1422–1432, 2006.
[6] J. A. Zhang, M. L. Rahman, K. Wu, X. Huang, Y. J. Guo, S. Chen, and J. Yuan, “Enabling joint communication and radar
sensing in mobile networks—a survey,” IEEE Commun. Surveys Tuts., vol. 24, no. 1, pp. 306–345, 2022.
[7] Y. Zhong et al., “Empowering the V2X network by integrated sensing and communications: Background, design, advances,
and opportunities,” IEEE Network, vol. 36, no. 4, pp. 54–60, 2022.
[8] Z. Wei, W. Yuan, S. Li, J. Yuan, G. Bharatula, R. Hadani, and L. Hanzo, “Orthogonal time-frequency space modulation:
A promising next-generation waveform,” IEEE Wireless Communications, vol. 28, no. 4, pp. 136–144, 2021.
[9] W. Anwar et al., “Performance analysis of various waveforms and coding schemes in V2X communication scenarios,” in
2020 IEEE Wireless Communications and Networking Conference (WCNC), 2020, pp. 1–8.
[10] L. Gaudio, M. Kobayashi, G. Caire, and G. Colavolpe, “On the effectiveness of OTFS for joint radar parameter estimation
and communication,” IEEE Transactions on Wireless Communications, vol. 19, no. 9, pp. 5951–5965, 2020.
[11] X. Ouyang et al., “Orthogonal chirp division multiplexing,” IEEE Trans. Commun., vol. 64, no. 9, pp. 3946–3957, 2016.
[12] A. Bemani, N. Ksairi, and M. Kountouris, “Affine frequency division multiplexing for next generation wireless communi-
cations,” IEEE Transactions on Wireless Communications, pp. 1–1, 2023.
[13] J. J. Healy et al., Linear canonical transforms: Theory and applications. Springer, 2015, vol. 198.
[14] “RP-234069: Study on channel modelling for Integrated Sensing And Communication (ISAC) for NR”. RAN Meeting
#102, Nokia Shanghai Bell, Dec. 2023.
[15] R. Prasad, OFDM for wireless communications systems. Artech House, 2004.
[16] S.-C. Pei and J.-J. Ding, “Closed-form discrete fractional and affine fourier transforms,” IEEE Trans. Sig. Proc., vol. 48,
no. 5, pp. 1338–1353, 2000.
[17] T. Erseghe, N. Laurenti, and V. Cellini, “A multicarrier architecture based upon the affine fourier transform,” IEEE
Transactions on Communications, vol. 53, no. 5, pp. 853–862, 2005.
[18] G. D. Surabhi, R. M. Augustine, and A. Chockalingam, “On the diversity of uncoded OTFS modulation in doubly-dispersive
channels,” IEEE Transactions on Wireless Communications, vol. 18, no. 6, pp. 3049–3063, 2019.
[19] T. Thaj, E. Viterbo, and Y. Hong, “Orthogonal time sequency multiplexing modulation: Analysis and low-complexity
receiver design,” IEEE Transactions on Wireless Communications, vol. 20, no. 12, pp. 7842–7855, 2021.
[20] H. Lin and J. Yuan, “Orthogonal delay-doppler division multiplexing modulation,” IEEE Transactions on Wireless
Communications, vol. 21, no. 12, pp. 11 024–11 037, 2022.
[21] B. Farhang-Boroujeny, “OFDM versus filter bank multicarrier,” IEEE Sig. Proc. Mag., vol. 28, no. 3, pp. 92–112, 2011.
[22] N. Michailow et al., “Generalized frequency division multiplexing for 5th generation cellular networks,” IEEE Transactions
on Communications, vol. 62, no. 9, pp. 3045–3061, 2014.
[23] M. A. Richards, Fundamentals of Radar Signal Processing. McGraw-Hill Professional, 2005.
[24] H. S. Rou, G. T. F. de Abreu, D. González G., and O. Gonsa, “Integrated sensing and communications for 3D object
imaging via bilinear inference,” Accepted in IEEE Transactions on Wireless Communications, 2024.
[25] J. Li and P. Stoica, MIMO radar signal processing. John Wiley & Sons, 2008.
[26] J. Zhu, Y. Tang, X. Wei, H. Yin, J. Du, Z. Wang, and Y. Liu, “A low-complexity radar system based on affine frequency
division multiplexing modulation,” arXiv preprint arXiv:2312.11125, 2023.
[27] P. Kumari, J. Choi, N. González-Prelcic, and R. W. Heath, “IEEE 802.11ad-based radar: An approach to joint vehicular
communication-radar system,” IEEE Transactions on Vehicular Technology, vol. 67, no. 4, pp. 3012–3027, 2018.
[28] P. Raviteja, K. T. Phan, Y. Hong, and E. Viterbo, “Orthogonal time frequency space (OTFS) modulation based radar
system,” in 2019 IEEE Radar Conference (RadarConf), 2019, pp. 1–6.
[29] L. Gaudio, M. Kobayashi, G. Caire, and G. Colavolpe, “Joint radar target detection and parameter estimation with
MIMO-OTFS,” in 2020 IEEE Radar Conference (RadarConf20), 2020, pp. 1–6.
[30] Y. Liu, Y. L. Guan, and D. González G., “Near-optimal BEM OTFS receiver with low pilot overhead for high-mobility
communications,” IEEE Transactions on Communications, vol. 70, no. 5, pp. 3392–3406, 2022.