J. Indian Inst. Sci.

A Multidisciplinary Reviews Journal

ISSN: 0970-4140 Coden-JIISAD

© Indian Institute of Science 2020.

ARTICLE
REVIEW
OTFS: A New Modulation Scheme for High‑Mobility
Use Cases

M. K. Ramachandran1,2, G. D. Surabhi1 and A. Chockalingam1*

Abstract | Among the several emerging use case families in 5G, high-
mobility use case family is a technologically challenging one. It is
expected that there will be a growing demand for mobile services in
vehicles, high-speed trains, and even aircraft. The degree of mobility
support required (i.e., speed) will depend upon the specific use case
(e.g., 500 km/h in bullet trains and 1000 km/h in airplanes). Mobility-on-
demand, ranging from very high mobility to low or no mobility, need to
be supported. The currently used waveforms fail to perform well in high-
mobility scenarios where the Doppler shifts witnessed are quite high
(e.g., several kHz of Doppler). Orthogonal time–frequency space (OTFS)
is a recently proposed radio access technology waveform suited very
well for high-mobility environments. It is a two-dimensional modulation
scheme in which information symbols are multiplexed in the delay–Dop‑
pler domain. We present an overview of delay–Doppler representation
of wireless channels and introduce OTFS modulation along with OTFS
basis functions. We illustrate the slow variability and sparse nature of the
delay–Doppler channel using an urban multi-lane scenario. Focusing on
MIMO-OTFS systems, we present signal detection and channel estima‑
tion schemes and their performance. MIMO-OTFS is shown to achieve
significantly better performance compared to MIMO-OFDM in high-Dop‑
pler environments operating in 4 GHz and 28 GHz frequency bands.
Keywords:  High-mobility use cases, OTFS modulation, Delay–Doppler domain, MIMO-OTFS, mmWave
communication, Signal detection, Channel estimation

1 Introduction (e.g., sensor networks, smart wearables), and high

Next-generation wireless networks (including mobility (e.g., high-speed trains, airplanes) are
5G networks) are expected to see the emergence being envisaged1. In particular, the high-mobility
of several new use cases and business models use case family is technologically challenging.
driven by the emerging needs of users and system Mobility requirements are expressed in terms of
operators. This will be enabled by the maturity of relative speed between the user and the network
current technologies and emergence of new tech- edge at which consistent user experience must be
nologies. The variety of applications and the cor- ensured. For example, in-vehicle mobile broad-
responding performance requirements in these band service requires a mobility support of up to 1
 Department of ECE,
scenarios will be vastly different. Several use case 500 km/h for bullet trains and 1000 km/h for air- Indian Institute of Science,
Bangalore 560012, India.
families such as broadband access everywhere planes. Wireless channels in such high-mobility 2
 Present Address:
(e.g., pervasive video), extreme real-time com- scenarios are doubly selective, where multipath Qualcomm India
munications (e.g., tactile internet with sub-mil- effects result in inter-symbol interference (due to Private Limited,
Bangalore 560066, India.
lisecond latency), ultra-reliable communications frequency selectivity) and Doppler shifts (due to *[email protected]
(e.g., e-health services), massive internet of things time selectivity). High Doppler shifts make the

J. Indian Inst. Sci. | VOL xxx:x | xxx–xxx 2020 | journal.iisc.ernet.in 13

 M. K. Ramachandran et al.

channel highly time selective2. The currently used the delay–Doppler domain, the doubly selec-
waveforms fail to perform well in high-mobility tive channel is sparse and slowly varying. OTFS
scenarios where the Doppler shifts encoun- modulation can be implemented using the
tered are high (e.g., several kHz of Doppler). For existing multicarrier modulation schemes (such
example, orthogonal frequency division multi- as OFDM) using additional pre- and post-pro-
plexing (OFDM)-based systems, while resilient cessing 2D transforms.
to inter-symbol interference (ISI) 3, suffer from Several recent works on OTFS have emerged
performance degradation due to inter-carrier in the literature recognizing the suitability of
interference (ICI) caused by high Doppler shifts 4. OTFS waveform for next-generation wireless
Pulse shaping is one approach to jointly mitigate systems including mmWave communication sys-
ISI and ICI in OFDM systems 5–7. This approach tems 15–36. These works have focused on vector-
uses time–frequency domain and optimal pulse ized input–output relation in OTFS, schemes
shapes for the waveform design. However, pulse for equalization/detection and channel estima-
shaping is not very effective in handling time- tion, diversity order achieved by OTFS, peak-to-
selectivity in high-mobility channels. Another average power ratio (PAPR) characteristics, pulse
approach in the literature makes use of canonical shaping, effect of oscillator phase noise, exploi-
decomposition of the received signal into delay tation of the effective channel structure for low-
and Doppler shifted versions of a basis signal, complexity linear detection, space-time coded
and proposes a delay–Doppler RAKE receiver OTFS, multiple access using OTFS, and relation
that offers joint multipath-Doppler diversity 8. between OTFS and generalized frequency divi-
Also, time-selective signaling schemes with over- sion multiplexing (GFDM). In this paper, we
lapping symbols can provide better performance present an overview of delay–Doppler represen-
despite ISI, if properly designed. Such schemes tation of wireless channels and introduce OTFS
that spread the information over time have been modulation along with OTFS basis functions. We
studied in the context of CDMA 9, 10. The above illustrate the slow variability and sparse nature of
approaches differ from OTFS modulation in the delay–Doppler channel using an urban multi-
that OTFS is designed in delay–Doppler domain lane scenario. Focusing on MIMO-OTFS systems
26
rather than in time–frequency domain. OTFS , we present signal detection and channel esti-
uses delay–Doppler domain for channel repre- mation schemes and their performance. MIMO-
sentation as well as multiplexing of information OTFS is shown to achieve significantly better
symbols. This difference between OTFS and the performance compared to MIMO-OFDM in
approach in 8 has been pointed out in 13. high-Doppler environments operating in 4 GHz
Orthogonal time–frequency space (OTFS) and 28 GHz frequency bands. For example, in a
modulation, proposed by Hadani et al. in 2 × 2 MIMO-OTFS system, a bit error rate (BER)
WCNC’2017 11, is a multiplexing scheme suited of 10−5 is achieved at a signal-to-noise ratio
well for doubly-selective channels. This two- (SNR) of 14 dB for a Doppler of 1850 Hz (500
dimensional (2D) modulation scheme uses an km/h speed at 4 GHz), whereas the performance
approach where multiplexing of information of MIMO-OFDM in the same system floors at a
symbols happens in the delay–Doppler domain BER of 0.02. A similar performance advantage of
11–14
. This is different from conventional modu- MIMO-OTFS over MIMO-OFDM is witnessed in
lation approaches (e.g., OFDM), where multi- mmWave frequency band (e.g., 28 GHz) as well.
plexing is done in the time–frequency domain. The rest of the paper is organized as follows.
OTFS basis functions (waveforms) have strong The delay–Doppler domain, its characteristics,
resilience to delay–Doppler shifts imparted by and its connection to OTFS modulation are pre-
the channel 12. OTFS has been shown to achieve sented in Sect. 2. The OTFS modulation is pre-
significant performance gains compared to sented in Sect. 3. The MIMO-OTFS system model
OFDM in high-Doppler channels with vehicular and a vectorized formulation for its input–output
speeds of 500 km/h in 4 GHz band and 40 km/h relation are presented in Sect. 4. MIMO-OTFS
in 28 GHz band 11, 15, 16. In the delay–Doppler signal detection using message passing algorithm
domain (signaling domain for OTFS), the basis and the BER performance are presented in Sect. 5.
functions for OTFS are 2D localized pulses. The The delay–Doppler channel estimation scheme
information symbols are multiplexed using and the resulting performance are presented in
these 2D pulses (basis functions). The idea is to Sect. 6. Conclusions are presented in Sect. 7.
use a transformation that relates the informa-
tion symbols and the wireless channel in the
delay–Doppler domain. Also, when viewed in

2
13 J. Indian Inst. Sci.| VOL xxx:x | xxx–xxx 2020 | journal.iisc.ernet.in
 OTFS: A New Modulation Scheme for High-Mobility Use Cases

Figure 1:  Example of a wireless channel in an urban multi-lane scenario illustrating the sparsity and slow
variability of the channel in the delay–Doppler representation.

2 The Delay–Doppler Domain and OTFS velocity and the distance remain roughly constant
Modulation for at least a few milliseconds. Hence, the chan-
In this section, we discuss the delay–Doppler rep- nel in the delay–Doppler domain appears time
resentation of wireless channels and its connec- invariant for a longer observation duration as
tion to OTFS modulation. We first illustrate the compared to that in the time–frequency repre-
sparsity and slow variability of the delay–Doppler sentation 14. Besides time invariance, the delay–
representation of wireless channels and then dis- Doppler representation of the channel impulse
cuss its connections to OTFS modulation. response results in a sparse representation of the
time-varying channel, thus requiring only fewer
channel parameters to be estimated. The slow
2.1 Delay–Doppler Representation
variability and sparse nature of the delay–Dop-
of Wireless Channels
pler channel can be visualized in Figs. 1 and 2.
Different representations can be used for mode-
Figure  1 shows an example of a scenario
ling the impulse response of a linear time-varying
encountered in urban multi-lanes. In the fig-
multipath channel depending upon the param-
ure, the transmitter mounted on top of the bus
eters (independent variables) chosen. Denoting
bay transmits a signal which is intended to be
time, frequency, delay, and Doppler variables with
received by one of the moving cars (blue car in
t, f, ν , and τ , respectively, the impulse response of a
the figure). The transmitted signal gets reflected
time-varying channel can be expressed as a func-
by multiple mobile reflectors (other cars) and
tion of time–frequency H(t, f), time–delay h(t, τ ) ,
reaches the receiver. For the sake of illustration,
and delay–Doppler h(τ , ν) . In the time–frequency
we have considered four paths corresponding to
(i.e., H(t, f)) and time–delay (i.e., h(t, τ ) ) repre-
four reflectors. The paths corresponding to differ-
sentations of the channel, channel coefficients
ent reflectors are indicated with different colors.
vary with time at a rate ( ∝ 1/coherence time )
The impulse response of the wireless channel in
which depends on the mobility and the operat-
this scenario is illustrated in both delay–Doppler
ing frequency. This causes the channel to change
domain as well as time–frequency domain in the
rapidly, rendering channel estimation difficult.
same figure at time t = t0 and t = t0 +  . From
A more compact and equivalent representa-
the figure, it can be seen that the channel repre-
tion of the channel is the delay–Doppler domain
sentation in the delay–Doppler domain is very
impulse response h(τ , ν) 11, 14. The channel coef-
sparse compared to that in time–frequency rep-
ficients (taps) in this representation correspond
resentation. Also, since the channel coefficients in
to a group of reflectors with a specific delay value
the delay–Doppler representation correspond to
depending on reflectors’ relative distance and
a group of reflectors with a particular delay and
Doppler value depending on reflectors’ relative
Doppler shift, we see four non-zero channel coef-
velocity with the transmitter and receiver. The
ficients corresponding to the four reflectors in

3
J. Indian Inst. Sci. | VOL xxx:x | xxx–xxx 2020 | journal.iisc.ernet.in 13
 M. K. Ramachandran et al.

Figure 2:  Squared magnitude of the impulse response of a 300 Hz Jakes Doppler channel model with 25
uniform power delay profile taps in a time–frequency domain, b time–delay domain, and c Doppler–delay
domain.

the figure. This illustrates that the delay–Doppler reflectors. Thus, the channel-symbol coupling/
representation of the channel impulse response is interaction in this domain is given by the follow-
sparse and captures the geometry of the wireless ing double integral:
environment. Figure 1 also illustrates the channel  
at time t = t0 +  , in which one of the reflectors y(t) = h(τ , ν)x(t − τ )ej2π ν(t−τ ) dτ dν.
(green car which is circled) has decelerated. Due ν τ
(1)
to the deceleration (decrease in the speed) of one
of the reflectors, the coefficient corresponding to
this reflector now falls in a different bin on the
Doppler axis (yellow bar) in the delay–Doppler 2.2 OTFS: Modulation in the Delay–
grid, whereas other coefficients remain in their Doppler Domain
bins at time t0 . However, the time–frequency rep- Fundamentally, a signal can be represented either
resentation of the channel shows significant vari- as a superposition of delta functions (called the
ation at times t0 and t0 +  . This illustrates the temporal/time representation) or as a superpo-
slow variability of the channel in the delay–Dop- sition of complex exponential functions (called
pler representation compared to time–frequency the frequency representation). These two rep-
representation. resentations are one-dimensional (1D) and are
Figure 2 shows the squared magnitude of the interchangeable through Fourier transform.
impulse response of Jakes channel model with Apart from the time and frequency domain rep-
maximum Doppler of 300 Hz 2 with 25 delay taps resentations, a signal can also be represented as
and uniform power delay profile; (a) in time–fre- a quasi-periodic function of delay and Doppler
quency domain, (b) in time–delay domain which 12
. The delay–Doppler signal representation is
is related by Fourier transform with the time– a two-dimensional (2D) representation which
frequency domain along the delay axis, and (c) can be converted to the time and frequency rep-
in delay–Doppler domain which is related to the resentations through the Zak transforms Zt and
time–frequency domain by a transform called Zf  , respectively. The signals in the delay–Doppler
2D symplectic Fourier transform 14. As observed domain can be viewed as functions φ(τ , ν) on a
from Fig. 2, the impulse response is spread in the 2D delay–Doppler plane whose points are param-
time–frequency and time–delay representations, eterized by τ and ν . This representation is a quasi-
whereas it is peaky (localized) in a few delay– periodic representation and has an associated
Doppler bins in the delay–Doppler representa- delay period τr and a Doppler period νr , such that
tion, i.e., the delay–Doppler impulse response is τr νr = 1 . The delay–Doppler representation is
more sparse. The received signal y(t) in the delay– related to time and frequency representations by
Doppler representation is the sum of reflected Zak transforms Zt and Zf  , respectively, given by 12
copies of the transmitted signal x(t), which are
time delayed ( τ ) and frequency shifted(ν ) by the

4
13 J. Indian Inst. Sci.| VOL xxx:x | xxx–xxx 2020 | journal.iisc.ernet.in
 OTFS: A New Modulation Scheme for High-Mobility Use Cases

Figure 3:  OTFS modulation scheme.

 νr
(2) multicarrier modulation scheme such as OFDM.
Zt (φ) = ej2π tν φ(t, ν)dν, The series of transformations involved in OTFS
0
modulation and demodulation transforms a rap-
 τr idly time-varying multipath channel into a slowly
Zf (φ) = e−j2π τ f φ(τ , f )dτ . (3) varying delay–Doppler domain channel, which is
0 also sparse. The sparsity in the effective channel
Quasi-periodic signals are periodic up to a mul- in OTFS modulation can be exploited to estimate
tiplicative phase, i.e., φ(τ + nτr , ν + mνr ) = only a few channel parameters compared to that
ej2π(nντr −mτ νr ) φ(τ , ν). It is noted that the quasi- in the time–frequency domain. Also, since the
periodicity property is necessary to recover channel in the delay–Doppler domain is slowly
the 2D function φ(τ , ν) uniquely from its 1D varying, the channel can be estimated less fre-
projection Zt (φ) or Zf (φ) . Without it, a one- quently, which can reduce the overhead on chan-
dimensional signal will admit infinitely many nel estimation in a rapidly time-varying channel.
delay–Doppler representations 12. The delay– A detailed description of the OTFS modulation
Doppler signal representation is not unique scheme architected using pre- and post-process-
because τr and νr can take any value such that ing operations over a multicarrier modulation is
τr νr = 1 . When the limit τr tends to infinity and presented in Sect. 3.2.
νr tends to zero, the delay–Doppler representation
becomes the time representation. Similarly, when 3 OTFS Modulation
the limit νr tends to infinity and τr tends to zero, OTFS modulation can be implemented using pre-
the delay–Doppler representation becomes the and post-processing to existing OFDM-based
frequency representation of the signal. multicarrier modulation schemes. This is done
A fundamental feature of OTFS modulation using as a series of 2D transformations at both
that distinguishes it from other time–frequency the transmitter and receiver. Figure 3 shows the
(TF) modulation schemes is the use of delay– block diagram of the OTFS modulation scheme.
Doppler domain for multiplexing the modula- The inner box in the diagram shows the familiar
tion symbols. The symbols in the delay–Doppler multicarrier modulation in the time–frequency
domain can be converted into temporal domain domain,1 and the outer box with a pre-processor
for transmission using the Zak transform Zt . The and a post-processor block implements the OTFS
transformation that uses a single Zak transform modulation in the delay–Doppler domain. The
to convert a signal in delay–Doppler domain to a information symbols x[k, l] (e.g., QAM symbols)
signal in time domain can also be carried out in residing in the delay–Doppler domain are first
two steps. That is, the signal is first transformed transformed to the familiar time–frequency (TF)
to time–frequency domain using a 2D inverse domain signal X[n, m] using the 2D inverse sym-
symplectic finite Fourier transform, and then plectic finite Fourier transform (ISFFT) and a 2D
the resulting time–frequency signal is converted windowing is applied. The Heisenberg transform
to a temporal representation using a Heisenberg
transform. The two-step implementation used
1 
to transform the symbols in the delay–Dop- Multicarrier modulation systems such as OFDM or filtered
pler domain to time domain signal facilitates OFDM can constitute the inner core of OTFS modulation.
For example, recently, it has been shown in 25 that OTFS can
the implementation of OTFS modulation using be implemented using a GFDM framework and that OTFS
simple pre- and post-processing steps over any achieves better performance compared to GFDM.

5
J. Indian Inst. Sci. | VOL xxx:x | xxx–xxx 2020 | journal.iisc.ernet.in 13
 M. K. Ramachandran et al.

is then applied to the TF signal X[n, m] to trans- formed back to the TF domain using Wigner
form to the temporal domain signal x(t) for trans- transform given by
mission. The received signal y(t) at the receiver is •  Y [n, m] = Aϕrx ,y (τ , ν)|τ =nT ,ν=m�f , (6)
transformed back to a TF signal Y[n, m] through
inverse Heisenberg transform (also called Wigner where Aϕrx ,y (τ , ν) is the cross ambiguity
transform). Y[n, m] thus obtained is transformed function
to the delay–Doppler domain signal y[k, l] using 
a 2D receive windowing and the symplectic finite ∗
Aϕrx ,y (τ , ν) = ϕrx (t − τ )y(t)e−j2π ν(t−τ ) dt,
Fourier transform (SFFT), for demodulation.
In the following subsections, we describe the (7)
signal models in TF modulation and OTFS mod- and y(t) is related to x(t) by (1). The relation
ulation. Let T denote the TF modulation symbol between Y[n, m] and X[n, m] for TF modulation
time and f denote the subcarrier spacing. Let can be derived as 11
x[k, l], k = 0, . . . , N − 1 , l = 0, . . . , M − 1 denote
Y [n, m] = H[n, m]X[n, m] + V [n, m], (8)
the information symbols transmitted in a given
packet burst. Let Wtx [n, m] and Wrx [n, m] denote where V[n,  m] is the additive white Gaussian
the transmit and receive windows, respectively. noise and H[n, m] is given by
 
3.1 Time–Frequency Modulation H [n, m] = h(τ , ν)ej2π νnT e−j2π(ν+m�f )τ dνdτ .
τ ν
•  Let ϕtx (t) and ϕrx (t) denote the transmit and (9)
receive pulses, respectively. We assume ϕtx (t) ,
ϕrx (t) to be ideal pulses satisfying the bi-
orthogonality property with respect to trans- 3.2 OTFS Modulation
lations by integer multiples of time T and fre- •  When OTFS modulation is implemented as
quency f  , i.e., pre- and post-processing to TF modulation,
 the delay-period τr and Doppler period νr are
• 
e−j2π m�f (t−nT ) ϕrx
∗
(t − nT )ϕtx (t)dt = δ(m)δ(n). chosen as
 
(4) •  1 1
(τr , νr ) = , .
The bi-orthogonality property of the pulse �f T
shapes eliminates the cross-symbol interfer-
ence at the receiver. Although ideal pulses •  Let Xp [n, m] be the periodized version of
are not practically realizable, they can be X[n,  m] with period (N,  M). The SFFT of
approximated by the pulses whose support is Xp [n, m] is given by
highly concentrated in time and frequency 18, •  N −1 M−1
 
given the constraint imposed by the uncer- xp [k, l] =
nk
Xp [n, m]e−j2π( N − M ) ,
ml

tainty principle.2 The signal in the TF domain n=0 m=0

X[n, m], n = 0, . . . , N − 1 , m = 0, . . . , M − 1 (10)
is transmitted in a given packet burst. and the ISFFT is Xp [n, m] = SFFT −1
(xp [k, l]) ,
•  TF modulation/Heisenberg transform The sig- given by
nal in the time–frequency domain X[n, m] is
transformed to the time domain signal x(t) N −1 M−1
1  nk ml
using the Heisenberg transform given by Xp [n, m] = xp [k, l]ej2π( N − M ) .
MN
k=0 l=0
•  N
 −1 M−1
 (11)
j2π m�f (t−nT )
x(t) = X[n, m]ϕtx (t − nT )e .
n=0 m=0
•  Information symbols xp [k, l] ,
(5) k = 0, . . . , N − 1 , l = 0, . . . , M − 1 , are trans-
mitted in a given packet burst.
•  TF demodulation/Wigner transform At the
•  OTFS transform/pre-processing The informa-
receiver, the time domain signal is trans-
tion symbols in the delay–Doppler domain
are mapped to TF domain symbols X[n, m] as
2
 Design and performance of OTFS systems using practical
•  X[n, m] = Wtx [n, m]SFFT−1 (xp [k, l]),
pulse shapes have been discussed in 24. It has been shown in (12)
24
that the OTFS modulation with practical pulse shapes like
rectangular pulse and prolate spheroidal waveform shows bet-
where Wtx [n, m] is the square summable
ter error performance compared to the conventional OFDM. transmit windowing function.

6
13 J. Indian Inst. Sci.| VOL xxx:x | xxx–xxx 2020 | journal.iisc.ernet.in
 OTFS: A New Modulation Scheme for High-Mobility Use Cases

•  X[n, m] thus obtained is in the TF domain and support of the window function. Thus, the trans-
it is TF modulated as described in the previous mit and the receive window functions used allow
subsection, and Y[n, m] is obtained by (6). the finite 2D summation approximation of the
•  OTFS demodulation/post-processing A receive input–output relation in OTFS modulation.
window Wrx [n, m] is applied to Y[n,  m] and
periodized to obtain Yp [n, m] which has the
3.2.1  OTFS Basis Functions
period (N, M), as
The OTFS basis functions in the delay–Dop-
•  YW [n, m] =Wrx [n, m]Y [n, m], pler domain, time–frequency domain, and time
∞
 domain for different delay (l) and Doppler (k)
Yp [n, m] = YW [n − kN , m − lM]. indices are shown in Figs. 4, 5 and 6. For this illus-
k,l=−∞ tration of basis functions, we have chosen a frame
(13) size of M = 32 , N = 32 , and a subcarrier spacing
The symplectic finite Fourier transform is of f = 15 kHz. We see that, in the TF domain,
then applied to Yp [n, m] to convert it from the basis function spans the entire time–frequency
TF domain back to delay–Doppler domain plane. This resembles spread spectrum. The time
yp [k, l] , as domain waveform is a train of pulses modulated
by a tone 12. Locally, the waveform resembles TDM
yp [k, l] = SFFT(Yp [n, m]). (14) pulse (localized pulse in time domain), and glob-
ally, the shape of the train resembles a tone or
The output sequence of demodulated sym-
FDM pulse (localized pulse in frequency domain).
bols is obtained as y[k, l] = yp [k, l] for
Figure 7 shows the OTFS basis functions (car-
k = 0, 1, . . . , N − 1 and l = 0, 1, . . . , M − 1.
rier waveforms) in the time domain for Doppler
The input–output relation in OTFS modulation
and delay indices (k, l) = (0, 0) , (k, l) = (0, 15) ,
can be derived as 11
and (k, l) = (2, 0) . We see that, as the Doppler
M−1 N −1
1  index (k) changes, the frequency of the pulse
y[k, l] = x[n, m] train changes (as in FDM). Similarly, we also note
MN

m=0 n=0
 that, as the delay index (l) changes, the position
k −n l−m (15) of the pulses gets shifted in time (as in TDM).
hw , + v[k, l],
NT Mf This illustrates that the OTFS waveform is a gen-
eralization of TDM and FDM.
where v[k, l] is the additive white Gaussian noise
and
  3.3 Vectorized Formulation of the Input–
k −n l−m
hw , = hw (ν ′ , τ ′ )|ν ′ = k−n ,τ ′ = l−m , Output Relation
NT M�f NT M�f
(16) Consider a channel with P signal propaga-
tion paths (taps). Let the ith path be associated
where hw (ν ′ , τ ′ )
is the circular convolution of
with a delay value τi , a Doppler value νi , and a
the channel response with a windowing function
fade coefficient hi . The delay–Doppler channel
w(τ , ν) , given by
impulse response can be written as
 
hw (ν ′ , τ ′ ) = h(τ , ν)w(ν ′ − ν, τ ′ − τ )dτ dν, P

ν τ
h(τ , ν) = hi δ(τ − τi )δ(ν − νi ). (19)
(17)
i=1
where w(τ , ν) is given by
Assume that the transmit and receive windows
M−1
 N−1
used in modulation, Wtx [n, m] and Wrx [n, m] , are
w(τ , ν) = Wtx [n, m] αi
rectangular. Define τi = M�f and νi = NTβi
 , where
m=0 n=0
Wrx [n, m]e−j2π(νnT −τ m�f ) . (18)
αi and βi are integers denoting the indices of the
delay tap (with delay τi ) and Doppler tap (with
Note that the window w(τ , ν) in (17) is a 2D Dir-
Doppler value νi ). In practice, although the delay
ichlet kernel introduced due to the finite trans-
and Doppler values are not exactly integer multi-
mission bandwidth and finite transmission time
ples of the taps, they can be well approximated by a
interval of the signal. Any function convolved
few delay–Doppler taps in the discrete domain 40.
with a Dirichlet kernel (as in 17) is a Fourier series
With the above assumptions, the input–output
(2D in this case) approximation to the function
relation for the channel in (19) can be derived as 18
itself. The error in approximation depends on the

7
J. Indian Inst. Sci. | VOL xxx:x | xxx–xxx 2020 | journal.iisc.ernet.in 13
 M. K. Ramachandran et al.

1
2
0.5
1
0
0 0
20 20 20 20
0
0 0

-1
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
10-3

Figure 4:  OTFS basis functions in delay–Doppler domain, time–frequency domain, and time domain for
Doppler index k = 0 and delay index l = 0.

1 1

0.5 0

0
0 -1
20 20
20 20
0 0 0

-1
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
10-3

Figure 5:  OTFS basis functions in delay–Doppler domain, time–frequency domain, and time domain for
Doppler index k = 2 and delay index l = 0.

P
 h′i = hi e−j2π νi τi . (21)
y[k, l] = h′i x[(k − βi )N , (l − αi )M ] + v[k, l],
i=1
(20) It is assumed that the hi s are i.i.d and are distrib-
uted as CN (0, 1/P) , assuming uniform scattering
where profile. Denoting the N × M channel matrix in
the delay–Doppler grid by H   , the (k, l)th entry of

H , denoted by ĥ(k, l) , is defined as

h′i if k = βi & l = αi for some i ∈ {1, 2, . . . , P}
ĥ(k, l) =
0 otherwise.

8
13 J. Indian Inst. Sci.| VOL xxx:x | xxx–xxx 2020 | journal.iisc.ernet.in
 OTFS: A New Modulation Scheme for High-Mobility Use Cases

1
1
0.5
0
0
-1
0

20 20 20 20
0 0 0

-1
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
10-3

Figure 6:  OTFS basis functions in delay–Doppler domain, time–frequency domain, and time domain for
Doppler index k = 2 and delay index l = 2.

0.8

0.6

0.4

0.2

-0.2

-0.4

-0.6

-0.8

-1
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
10-3

Figure 7:  OTFS basis functions in time domain for (k, l) = (0, 0) , (k, l) = (0, 15) , and (k, l) = (2, 0).

Now, the input–output relation in (20) can be H[j] = [ĥ((k − 0)N , (l − 0)M )
vectorized as 18
ĥ((k − 1)N , (l − 0)M )
y = Hx + v, (22)
. . . ĥ((k − N − 1)N , (l − M − 1)M )]. (23)
where xk+Nl = x[k, l] ,
yk+Nl = y[k, l] , vk+Nl = v[k, l] , Note that, due to the modulo operations in
k = 0, . . . , N − 1, l = 0, . . . , M − 1 , and (20), each row of H in (23) has only P non-zero
H ∈ CMN ×MN  , whose jth row (  j = k + Nl , elements.
j = 0, 1, . . . , MN − 1 ), denoted by H[j] , is given
by

9
J. Indian Inst. Sci. | VOL xxx:x | xxx–xxx 2020 | journal.iisc.ernet.in 13
 M. K. Ramachandran et al.

Figure 8:  MIMO-OTFS modulation scheme.

 
4 MIMO‑OTFS Modulation H11 H12 . . . H1nt
Consider a MIMO-OTFS system 26 with nt trans-  H21 H22 . . . H2nt 
mit antennas and nr receive antennas, nt ≥ nr as HMIMO =
 .. .. .. ..  ,
. . . .
shown in Fig. 8. Information symbols are OTFS Hnr 1 Hnr 2 . . . Hnr nt
modulated and sent from each of the nt transmit T
antennas independently. Let the transmit window xMIMO = [x1 T , x2 T , · · · , xnt T ] ,
Wtx [n, m] and receive window Wrx [n, m] used in T
yMIMO = [y1 T , y2 T , · · · , ynr T ] ,
OTFS pre- and post-processing be rectangular.
T
The channel between pth transmit and qth receive vMIMO = [v1 T , v2 T , · · · , vnr T ] .
antennae is assumed to have P taps as in (19).
Therefore, the equivalent delay–Doppler channel Then (25) can be written as
model is given by yMIMO = HMIMO xMIMO + vMIMO , (26)
P
 where xMIMO ∈ Cnt NM×1 , yMIMO , vMIMO ∈ Cnr NM×1 ,
hqp (τ , ν) = hqpi δ(τ − τi )δ(ν − νi ), (24)
HMIMO ∈ Cnr NM×nt NM  . Thus due to modulo
i=1
operations, each row of the matrix HMIMO in this
p = 1, 2, . . . , nt  , q = 1, 2, . . . , nr  
. Thus, vector- representation, has only nt P non-zero elements
ized formulation in Sect. 3.3 for SISO-OTFS can and each column has only nr P non-zero elements.
be extended to MIMO-OTFS systems.
5 Signal Detection for MIMO‑OTFS
We now present an iterative message passing algo-
4.1 Vectorized Formulation of the Input– rithm for MIMO-OTFS signal detection and pre-
Output Relation for MIMO‑OTFS sent results for MIMO-OTFS and MIMO-OFDM
Let the channel matrix between the pth transmit performance comparison in highly time-selective
antenna and qth receive antenna be denoted by multipath fading channels.
Hqp . Let xp be the vector transmitted from the
pth transmit antenna and yq be the received vec-
tor at the qth receive antenna in an OTFS frame. 5.1 Signal Detection Algorithm
Both are of size NM × 1 . Then we derive a lin- for MIMO‑OTFS
ear vectorized input–output model for MIMO- Denote the set consisting of non-zero positions in
OTFS, as described below: the bth row and ath column of HMIMO by ζb and
ζa , respectively. Thus, (26) enables us to model
y1 = H11 x1 + H12 x2 + · · · + H1nt xnt + v1 , the system as a sparse factor graph having nt NM
y2 = H21 x1 + H22 x2 + · · · + H2nt xnt + v2 , variable nodes (elements of xMIMO ) and nr NM
.. .. observation nodes (elements of yMIMO ). Each
. . observation node yb in the graph is connected
ynr = Hnr 1 x1 + Hnr 2 x2 + · · · + Hnr nt xnt + vnr . to the set of variable nodes {xc , c ∈ ζb } , and each
(25) variable node xa in the graph is connected to
Define the set of observation nodes {yc , c ∈ ζa } . Thus,

10
13 J. Indian Inst. Sci.| VOL xxx:x | xxx–xxx 2020 | journal.iisc.ernet.in
 OTFS: A New Modulation Scheme for High-Mobility Use Cases

|ζb | = nt P and |ζa | = nr P . For the detection of � |A|

�
(t) (t)
xMIMO in (26), the maximum a posteriori (MAP) µba = E[Iba ] = pcb (aj )aj Hb,c ,
c∈ζb ,c� =a j=1
decision rule is 
� �|A|
(t) (t)
x̂MIMO = argmax Pr (xMIMO |yMIMO , HMIMO ), (σba )2 = Var[Iba ] =  p (aj )|aj |2 |Hb,c |2
cb
xMIMO ∈Ant NM j=1
(27) c ∈ ζb
c �= a
where A is the modulation alphabet (e.g., QAM) � �2 
� |A| �
used. The exponential complexity of (27) makes �� (t) � 
−�� pcb (aj )aj Hb,c ��  + σ 2 .
it infeasible in practice and can’t be used. Instead, � j=1 �
we can use a reduced complexity symbol by sym-
bol MAP rule for 0 ≤ a ≤ nt NM − 1 for detec-
4. Messages from xa to yb Messages passed from
tion as given below:
variable nodes xa to observation nodes yb
x̂a = argmax Pr (xa = aj |yMIMO , HMIMO ) is the pmf vector p(t+1)
ab with the elements
aj ∈A given by
1
= argmax Pr (yMIMO |xa = aj , HMIMO ) (t+1)
pab
(t)
= � pab (aj ) + (1 − �) pab
(t−1)
(aj ),
aj ∈A |A|
 (30)
≈ argmax Pr (yc |xa = aj , HMIMO ). where � ∈ (0, 1] is the damping factor for
aj ∈A c∈ζa
(28) improving convergence rate, and
(t)

The elements of xMIMO has uniform probabil- pab ∝ Pr(yc |xa = aj , HMIMO ),
ity of occurrence and the elements of yMIMO c∈ζa ,c� =b
are assumed to be nearly independent for a (31)
given xa . This assumption is true because of where
the sparsity of HMIMO . Iterative message pass-
ing based algorithm can be used to solve this Pr(yc |xa = aj , HMIMO )
 
problem as described below. The message that (t)
−|yc − µca − Hc,a aj |2
is passed from the variable node xa , for each ∝ exp 2(t)
.
a = {0, 1, . . . , nt NM − 1} 
, to the observation σc,a
node yb for b ∈ ζa , is the probability mass func-
5. Stopping criterion Repeat steps 3 & 4 till
tion (pmf) denoted by pab = {pab (aj )|aj ∈ A} (t+1)
max |pab (t)
(aj ) − pab (aj )| < ǫ (where ǫ is a
of the symbols in the constellation A . Denote a,b,aj
the element in the ath row and bth column of
small value) or the maximum number of
HMIMO by Hab . The message passing algorithm
iterations, Niter , is reached.
is described below.
6. Output Output the detected symbol as

1. Inputs yMIMO , HMIMO , Niter : max. number x̂a = argmax pa (aj ), a ∈ 0, 1, 2, . . . , nt NM − 1,

aj ∈A
of iterations. (32)
2. Initialization Iteration index t = 0 , pmf
where
p(0)
ab = 1/|A| ∀ a ∈ {0, 1, . . . , nt NM − 1} and 
b ∈ ζa. pa (aj ) = Pr(yc |xa = aj , HMIMO ).
3. Messages from yb to xa The mean (µ(t) ba ) and c∈ζa
(t) 2
variance ((σba ) ) of the interference term Iba (33)
are passed as messages from yb to xa . Iba can
be approximated as a Gaussian random vari-
able and is given by 5.2 Vectorized Formulation of the Input–
 Output Relation for MIMO‑OFDM
Iba = xc Hb,c + vb . (29) Performance comparison of MIMO-OTFS
c∈ζb ,c�=a
and MIMO-OFDM can be done by first for-
The mean and variance of Iba are given by mulating an input–output vector relation for
MIMO-OFDM similar to MIMO-OTFS. In this
section, we develop this formulation in time–fre-
quency domain, which is the signaling domain
for MIMO-OFDM. Vectorized formulation of

11
J. Indian Inst. Sci. | VOL xxx:x | xxx–xxx 2020 | journal.iisc.ernet.in 13
 M. K. Ramachandran et al.

MIMO-OFDM can be easily obtained by extend- IDFT matrices of size M by WM×M and WM×M H  ,
ing the results of SISO-OFDM as shown below. respectively. The following notations are used:
Consider N consecutive blocks (each of size M)
to be one OFDM frame, i.e., the transmit vector •  Bcpin = (IN ⊗ TCP ) : cyclic prefix insertion
xOFDM ∈ CNM×1 (This is for a fair comparison matrix for N consecutive OFDM blocks.
with OTFS in terms of spectral efficiency). This •  Bcpre = (IN ⊗ RCP ) : cyclic prefix removal
means, iterative message passing algorithm is matrix for N consecutive OFDM blocks.
jointly applied to the NM × 1 frame. Consider •  D = (IN ⊗ W) : DFT matrix for N consecu-
the channel in (19). The time–delay representa- tive OFDM blocks.
tion h(τ , t) and the delay–Doppler representation •  DH = (IN ⊗ W H ) : IDFT matrix for N con-
h(τ , ν) are related by a Fourier transform along secutive OFDM blocks.
the time axis, and is given by •  The time–delay domain channel for a given
P OFDM frame can be written as a matrix Htd

h(τ , t) = hi ej2π νi t δ(τ − τi ). (34) of size NL × NL , using (35).
i=1
n
Thus, the linear input–output vector model for
Sample the time axis at t = nTs = Mf  . The SISO-OFDM is given by

sampled time–delay representation h(τ , n) is yOFDM = DBcpre Htd Bcpin DH xOFDM + v

given by   
HOFDM
P
 j2πνi n = HOFDM xOFDM + v,
h(τ , n) = hi e M�f δ(τ − τi ). (35) (36)
i=1 where xOFDM , yOFDM , v ∈ CNM×1 ,
In each OFDM block, let CP = P − 1 be the cyclic HOFDM ∈ CNM×NM.
prefix (CP) length used and also let L = M + CP .
Thus, the size of one OFDM frame after CP inser- 5.2.1  MIMO‑OFDM
tion to each block is NL. Let TCP = [CCP T I ]T
M In this subsection, we extend the SISO-OFDM
denote the L × M CP insertion matrix per block, results derived above to MIMO-OFDM. Denote
where CCP consists of the last CP rows of the the equivalent channel matrix between pth trans-
identity matrix IM . Also, let RCP = [0M×CP IM ] mit antenna and qth receive antenna by HOFDMqp .
be the M × L matrix which does the cyclic pre- Let xOFDMp of size NM × 1 be the transmit vector
fix removal for each block 41. Denote the DFT and from the pth transmit antenna and yOFDMq of size
NM × 1 be the received vector at the qth receive
antenna in a given OFDM frame. Define
 
HOFDM11 HOFDM12 . . . HOFDM1nt
 HOFDM21 HOFDM22 . . . HOFDM2nt 
 
HMIMO−OFDM = .. .. .. .. ,
 . . . . 
HOFDMnr 1 HOFDMnr 2 . . . HOFDMnr nt

and
T
xMIMO−OFDM = [xOFDM1 T , xOFDM2 T , · · · , xOFDMnt T ] ,
T
yMIMO−OFDM = [yOFDM1 T , yOFDM2 T , · · · , yOFDMnr T ] .

12
13 J. Indian Inst. Sci.| VOL xxx:x | xxx–xxx 2020 | journal.iisc.ernet.in
 OTFS: A New Modulation Scheme for High-Mobility Use Cases

Table 1:  Delay–Doppler profile for the channel 5.3 Performance Results and Discussion

model with P = 5 at 4 GHz carrier frequency. In this subsection, the BER performance com-
parison of MIMO-OTFS and MIMO-OFDM in
Path index (i) 1 2 3 4 5
high-mobility channel scenarios, is presented.
Delay ( τi ) ( µs) 2.08 4.164 6.246 8.328 10.41 The channel knowledge at the receiver is assumed
Doppler ( νi ) (Hz) 0 470 940 1410 1880
to be perfect. Iterative message passing-based
detection is used for both MIMO-OTFS and
Speed (km/h) 0 126.9 253.8 380.7 507.6
MIMO-OFDM. The algorithm is run for 30 itera-
tions with an ǫ value 0.01. The channel model in
(24) is used with the number of taps P = 5 . For
Table 2:  System parameters for 4 GHz system.
all the paths the fade coefficients hi are assumed
Parameter Value to be distributed as CN (0, 1) , i.e., Rayleigh fad-
ing. The channel fade coefficients are simu-
Carrier frequency (GHz) 4
lated as per (21) with the delay–Doppler profile
Frame size (M, N) (32, 32) {(τi , νi ), i = 1, . . . , P} shown in Table 1. Table 2
Subcarrier spacing (kHz) 15 gives all other simulation parameters which were
Modulation scheme BPSK used in the simulations. The carrier frequency
MIMO configuration 1 × 1 , 1 × 2 , and subcarrier spacing used are 4 GHz and 15
1 × 3 , 2 × 3 , kHz, respectively, the frame size parameters used
2 × 2 , 3 × 3 are M = N = 32 , and the modulation used is
No. of taps, P 5 BPSK.
Maximum speed (km/h) 507.6 Figure 9 shows the BER performance of OTFS
in SISO ( 1 × 1 ) and SIMO ( 1 × 2 and 1 × 3 ) set-
tings, Fig. 10 shows the performance in 2 × 2
and 2 × 3 MIMO settings, and Fig. 11 shows the
Thus, the linear vectorized input–output model performance in SISO and 2 × 2 , 3 × 3 MIMO
for MIMO-OFDM is given by settings. The maximum considered Doppler of
1880 Hz corresponds to a speed of 507.6 km/h
yMIMO−OFDM = HMIMO−OFDM xMIMO−OFDM
at 4 GHz carrier frequency. This means that, the
+ vMIMO , (37)
channel is highly time-selective. But, from the
results we see that MIMO-OTFS BER perfor-
where xMIMO−OFDM ∈ Cnt NM×1 , yMIMO−OFDM , mance is good, even in such scenarios. For exam-
vMIMO ∈ Cnr NM×1  , ple, we see in Fig. 9 that, at an SNR of about 8
HMIMO−OFDM ∈ Cnr NM×nt NM. dB, the 1 × 2 system achieves a BER of about
10−4 and the 1 × 3 system achieves a BER of

100

10-2

10-4

10-6

10-8
0 2 4 6 8 10 12

Figure 9:  BER performance of OTFS in SISO ( 1 × 1 ) and SIMO ( 1 × 2 and 1 × 3 ) settings.

13
J. Indian Inst. Sci. | VOL xxx:x | xxx–xxx 2020 | journal.iisc.ernet.in 13
 M. K. Ramachandran et al.

100

10-1

10-2

10-3

10-4

10-5

10-6
0 2 4 6 8 10 12

Figure 10:  BER performance of MIMO-OTFS in 2 × 2 and 2 × 3 MIMO systems.

100

10-1

10-2

10-3

10-4

10-5

10-6
0 2 4 6 8 10 12 14

Figure 11:  BER performance of MIMO-OTFS in 2 × 2 and 3 × 3 MIMO systems.

3 × 10−6 . Observing that the 1 × 1 SISO system of transmit–receive antennas and the resilience
achieves a BER of about 0.01 at this SNR, we see of OTFS modulation to delay–Doppler shifts in
that OTFS achieves improved BER performance doubly selective channels.
for increased number of receive antennas even in Figure 12 shows the BER performance of 2 × 2
high Doppler scenarios. This can also be observed MIMO-OTFS with (1) message passing detection,
in Fig. 10, where a BER of 3 × 10−5 is achieved at (2) minimum mean square error (MMSE) detec-
an SNR of about 12 dB in a 2 × 2 MIMO system, tion, and (3) zero forcing (ZF) detection. The sys-
whereas the same BER is achieved at an SNR of tem and channel parameters used are as given in
about 8.8 dB in a 2× 3 MIMO system. Figure. 11 Tables  1 and 2. From Fig. 12, it can be observed
shows the expected improved performance for that the performance of MIMO-OTFS with mes-
increasing nt = nr  . Thus, with the use of the sage passing detection is superior compared to
low-complexity message passing detection, we those of MMSE and ZF detectors. For example,
see that MIMO-OTFS brings in the advantages at an SNR of 12 dB, message passing detection
of linear increase in throughput with number achieves a BER of 3 × 10−5 , whereas MMSE and

14
13 J. Indian Inst. Sci.| VOL xxx:x | xxx–xxx 2020 | journal.iisc.ernet.in
 OTFS: A New Modulation Scheme for High-Mobility Use Cases

10 0

10 -1

10 -2
Bit error rate

10 -3

10 -4

10 -5
0 2 4 6 8 10 12
SNR in dB
Figure 12:  BER performance of 2 × 2 MIMO-OTFS system with (1) message passing detection, (2) mini‑
mum mean square error (MMSE) detection, and (3) zero forcing (ZF) detection.

100

10-1

10-2

10-3

10-4

10-5

10-6
0 2 4 6 8 10 12 14

Figure 13:  BER performance comparison between MIMO-OTFS and MIMO-OFDM in 2 × 2 MIMO system.

ZF detectors achieve a BER of 0.003 and 0.05, in the time–frequency domain. In fact, the BER
respectively. floors at a value of about 2 × 10−2 , whereas
MIMO-OTFS vs MIMO-OFDM Figure 13 MIMO-OTFS achieves a BER of 10−5 at an SNR
shows a comparison of BER performance value close to 14 dB. This illustrates the robust
between MIMO-OTFS and MIMO-OFDM in a performance of MIMO-OTFS where the sign-
2 × 2 MIMO system with the message passing- aling is in the delay–Doppler domain and the
based detection. For the case of maximum Dop- performance superiority of MIMO-OTFS over
pler of 1880 Hz, MIMO-OFDM performance MIMO-OFDM under doubly selective channels.
degrades due to inter carrier interference (ICI)

15
J. Indian Inst. Sci. | VOL xxx:x | xxx–xxx 2020 | journal.iisc.ernet.in 13
 M. K. Ramachandran et al.

10 0
MIMO-OFDM
MIMO-OTFS
10 -1

10 -2

10 -3

10 -4

10 -5
0 2 4 6 8 10 12

Figure 14:  BER performance comparison between 2 × 2 MIMO-OTFS with rectangular pulse and 2 × 2
MIMO-OFDM.

5.4 Performance of MIMO‑OTFS compare the BER performance of MIMO-OTFS

with Rectangular Pulse and MIMO-OFDM operating in this setting.
In this subsection, we consider OTFS modulation We consider a carrier frequency of 28 GHz. The
with rectangular transmit and receive pulses 16, 24. transmitter consists of a uniform linear array
Figure  14 shows a BER performance comparison (ULA) with nt transmit antennas with an inter-
between MIMO-OFDM and MIMO-OTFS with antenna spacing of dt  . Likewise, the receiver
rectangular pulse shape in a 2 × 2 system. Both consists of a ULA with nr receive antennas with
the systems use message passing detector for signal an inter-antenna spacing of dr . As before, the
detection. The maximum Doppler considered in channel model in (24) with P taps is considered.
1880 kHz, which corresponds to a speed of 507.6 The propagation in typical mmWave frequency
km/h at a carrier frequency of 4 GHz. The delay– application scenarios is dominated by a line-of-
Doppler profile used for the simulation is given in sight (LOS) component 37. Because of this, the
Table 1. Other parameters used for the simulations magnitude of the first tap in the channel model is
are given in Table 2. From the figure, we observe considered to follow a Rician distribution and the
that MIMO-OTFS shows better BER performance magnitudes of the remaining taps are considered
compared to MIMO-OFDM even with the use of to follow Rayleigh distribution. Therefore, the
rectangular pulse. For example, at an SNR of 12 fading channel gain in the first tap between the
dB, MIMO-OTFS achieves a BER of 3 × 10−5 , pth transmit antenna and qth receive antenna can
whereas MIMO-OFDM floors at 2 × 10−2 . This be written in the form 37
SNR gain shown by OTFS modulation is because  
of the outer pre- and post-processing operations K LOS 1
hqp1 = hqp1 + hNLOS ,
involving the 2D ISFFT and SFFT that spread the K +1 K + 1 qp1
(38)
information across the entire time–frequency
plane, which enables OTFS to extract higher diver- where hLOS NLOS denote the LOS and non-
qp1 and hqp1
sity performance in the finite SNR regime. LOS channel gains, respectively, in the first tap
between the pth transmit antenna and qth receive
antenna, and K is the Rician factor, defined as
5.5 Application of MIMO‑OTFS the ratio of the powers in the LOS and non-LOS
to mmWave Communications components. The LOS gain hLOS 38
qp1 is given by
In this subsection, we consider mmWave MIMO
communications in a point-to-point setting and

16
13 J. Indian Inst. Sci.| VOL xxx:x | xxx–xxx 2020 | journal.iisc.ernet.in
 OTFS: A New Modulation Scheme for High-Mobility Use Cases

With M = N = 128 , T = 12.8 µs , the duration

of one OTFS frame is NT = 1.63 ms . Consider-
ing the relative velocity between the transmitter
and the receiver to be v = 47 km/h, the change
in the distance between the transmitter and the
receiver in one OTFS frame can be calculated as
D = vNT = 0.021 m . For illustration, consider
the distance between the fourth transmit antenna
and the second receive antenna (i.e., p = 2 ,
q = 4 ). Let d24 and d24 ′ denote the LOS path

lengths between the fourth transmit and second

receive antenna at the beginning and at the end
of an OTFS frame, respectively. Now, d24 and d24 ′
qp1 in mmWave communication.
Figure 15:  hLOS
can be calculated to be 90.00005 m and 90.02 m,
respectively. Since the difference between d24 and
′ is very small, d in (39) can be assumed to be
d24
2π
qp
−j (39)
hLOS
qp1 = e
 dqp , constant over one OTFS frame.
We also assume hNLOS qp1 to be i.i.d and dis-
where dqp denotes the LOS path length between tributed as CN (0, 1) . Also, the fading gains in
the pth transmit and qth receive antenna, and  is the other taps with no LOS component (i.e.,
the carrier wavelength. Rice factor K in those taps is zero), hNLOS qpi  ,
Figure  15 illustrates the LOS path between i = 2, . . . , P , are assumed to be i.i.d and distrib-
the pth transmit antenna and the qth receive uted as CN (0, 1) . The performance is evaluated
antenna when there is relative motion between for SISO and 2 × 2 MIMO systems. Table 3 gives
the transmitter and the receiver. Note that the the delay–Doppler profile used. Table 4 gives all
change in LOS path lengths due to the rela- other parameter values used in simulations. The
tive motion between the transmitter and the carrier frequency and subcarrier spacing used are
receiver in one OTFS frame duration is very 28 GHz and 78.125 kHz, respectively, the frame
small for low/medium mobility. This can be size parameters used are M = N = 128 , the
illustrated with an example, where we consider inter-antenna distances at the transmitter and the
nt = nr = 2 , dt = dr = 5 cm, and D = 90 m . receiver are taken as dt = dr = 6.9 cm, and the

10 0

10 -1

10 -2
Bit error rate

10 -3

10 -4

10 -5
0 2 4 6 8 10 12
SNR (dB)
Figure 16:  BER performance comparison between MIMO-OTFS and MIMO-OFDM in 2 × 2 MIMO system
at 28 GHz.

17
J. Indian Inst. Sci. | VOL xxx:x | xxx–xxx 2020 | journal.iisc.ernet.in 13
 M. K. Ramachandran et al.

Table 3:  delay–Doppler profile for the channel Table 4:  System parameters for 28 GHz system.
model with P = 5 at 28 GHz carrier frequency.
Parameter Value
Path index (i) 1 2 3 4 5
Carrier frequency (GHz) 28
Delay ( τi ) ( µs) 0.3 1 1.7 2.4 3.1 Frame size (M, N) (128,128)
Doppler ( νi ) (Hz) 0 −1220 −610 610 1220 Subcarrier spacing (kHz) 78.125
Speed (km/h) 0 47 23.5 23.5 47 Modulation scheme BPSK
MIMO configuration 2×2

modulation used is BPSK. A Rician factor of 9 K factor (dB) 9

dB as reported in 37 from 28 GHz band measure- No. of taps (P) 5
ments is used. Maximum speed (km/h) 47
In Fig. 16, we present a BER performance
comparison between OTFS and OFDM in SISO
and 2 × 2 MIMO systems at 28 GHz. Itera- can be simultaneously estimated using single
tive message passing is used for both MIMO- MIMO-OTFS frame as described below.
OTFS and MIMO-OFDM detection. From The relation between the transmitted symbols
Fig. 16, it can be observed that SISO-OFDM sys- from pth transmit antenna and the received sym-
tem achieves a BER of 10−2 at an SNR of 9 dB, bols from the qth receive antenna can be written
whereas SISO-OTFS achieves a BER of 6 × 10−5 using (15) as
for the same SNR. It is also evident from the fig-
M−1
 N−1
ure that MIMO-OTFS performs significantly bet- 1
ter compared to MIMO-OFDM at 28 GHz. For yq [k, l] = xp [n, m]
MN
m=0 n=0
example, it can be seen that 2 × 2 MIMO-OTFS  
system outperforms 2 × 2 MIMO-OFDM sys- k −n l−m (40)
hwqp , + vq [k, l].
tem by about 8 dB at a BER of 2 × 10−4 . This NT Mf
illustrates the robustness of MIMO-OTFS over
Now, if the transmitted pilot from the pth
MIMO-OFDM in high Doppler scenarios that
antenna is given by
arise due to the use of high carrier frequencies in
the mmWave band. xp [n, m] = 1 if (n, m) = (np , mp )
(41)
= 0 ∀ (n, m) � = (np , mp ),
6 Channel Estimation in delay–Doppler the received signal at the qth antenna will be
Domain  
In this section, the assumption of availability of 1 k − np l − mp
yq [k, l] = hwqp , + vq [k, l].
perfect channel state information at the receiver MN NT Mf
(CSIR) is relaxed. We present a simple scheme (42)
 
to estimate the channel in the delay–Doppler Now, 1 k l
can be estimated from
MN hwqp NT , Mf
domain 26. In this scheme of channel estimation,
impulse function in the delay–Doppler domain is (42), since np and mp are known at the receiver a
used as the pilot. The pilot corresponding to each priori. From this, the equivalent channel matrix
antenna is placed in the delay–Doppler grid such Ĥqp can be obtained using the vectorized formu-
they can be received without interference at the lation of Sect. 3.3.
receive antennas. Figure 17 illustrates the pilots, Observe that due to the convolutive nature of
channel response, and received signal in a 2 × 1 the input–output relation, the impulse at
MIMO system. The delay–Doppler profile and (n, m) = (np , mp ) is spread by the channel only
system parameters used for the figure is given in up to the maximum delay and Doppler spread of
Tables  1 and 2. The pilots placed in the delay– the channel. Thus, if the pilots are spaced suffi-
Doppler grid have some space reserved around ciently far apart, in the delay–Doppler domain,
it, to account for the delay and Doppler spread of they can be received without interference. Hence,
the channel. Each transmit and receive antenna the channel responses corresponding to all the
pair sees a different channel with finite and non- transmit–receive antenna pairs can be estimated
overlapping support, determined by the delay simultaneously using single MIMO-OTFS frame.
and Doppler spread of the channel 14. Hence, the Figure 17 illustrates this for a 2 × 1 MIMO-OTFS
channel for all the transmit–receive antenna pairs system with (M, N ) = (32, 32) at 4 dB SNR. The

18
13 J. Indian Inst. Sci.| VOL xxx:x | xxx–xxx 2020 | journal.iisc.ernet.in
 OTFS: A New Modulation Scheme for High-Mobility Use Cases

Figure 17:  Illustration of pilots and channel response in delay–Doppler domain in a 2 × 1 system.

pilot corresponding to the first transmit antenna channel estimation complex in doubly dispersive
is placed at (n1 , m1 ) = (0, 0) and the pilot corre- MIMO channels.
sponding to the second transmit antenna is
placed at (n2 , m2 ) = (16, 16) in the delay–Dop-
pler domain. Also, note that  the impulse 6.1 Performance Results and Discussion
responses at the receive antenna hw11 k−n 1 l−m1
NT , Mf In this subsection, the mean square error (MSE)
  of the estimated channel and the BER perfor-
k−n2 l−m2
and hw12 NT , Mf  
, corresponding to the mance of MIMO-OTFS with the estimated chan-
nel are presented. The channel estimation scheme
pilots transmitted from the first and the second described in the previous subsection is used to
transmit antennas, respectively, are non-overlap- estimate ĤMIMO , which is used for the detection
ping at the receiver. Thus, they can be simultane- of the transmitted symbols using message passing
ously estimated using a single pilot frame as algorithm. The simulations use Tables 1 and 2 for
described above. the delay–Doppler profile and the system param-
The input–output relation in OTFS is con- eters, respectively.
volutive in nature (2D convolution of symbols Figure  18 shows the normalized MSE
and the channel response in delay–Doppler ( �HMIMO − ĤMIMO �F /(MN )2 ) of the estimated
domain), whereas the input–output relation in channel, as a function of the average SNR of the
any time–frequency modulation is multiplicative. pilot frame, for a 2 × 2 MIMO-OTFS system.
This convolutive relation, along with the sparsity The system parameters used are as in Tables 1
and time-invariance of delay–Doppler impulse and 2. From the figure, it can be observed that
response, greatly simplifies the channel estima- the normalized MSE decreases with the increase
tion estimation. In Fig. 17, the pilot correspond- in the SNR of the pilot frame, as expected. Next,
ing to the transmit antenna 1 gets convolved Fig. 19 shows the BER performance of the 2 × 2
with the channel response of transmit antenna 1. MIMO-OTFS system with the estimated channel.
Similarly, the pilot corresponding to the transmit From the figure, it is observed that the achieved
antenna 2 convolves with the channel response BER performance with the estimated channel is
of transmit antenna 2. At the receiver, the pilots very close to the performance with perfect CSIR.
are spread by the channel only to the extent of For example, at an SNR of 12.5 dB, the achieved
the support of each channel in the delay–Doppler BER with the perfect knowledge of the channel
domain. Since the channel in the delay–Doppler is 2 × 10−5 , whereas the same BER is achieved
representation is sparse, the pilots can be received at an SNR of 13 dB, with the estimated channel.
without interference. In contrast, in time–fre- At the considered Doppler frequency as high as
quency domain, the received pilots overlap and 1880 Hz, channel estimation in the TF domain is
span the entire time–frequency plane, making the very difficult and results in inaccurate estimation
because of the rapid variations of the channel in

19
J. Indian Inst. Sci. | VOL xxx:x | xxx–xxx 2020 | journal.iisc.ernet.in 13
 M. K. Ramachandran et al.

10 -4
1.2

0.8

0.6

0.4

0.2

0
0 2 4 6 8 10 12 14

Figure 18:  Mean square error of the estimated channel as a function of pilot SNR in a 2 × 2 MIMO system.

100

10-1

10-2

10-3

10-4

10-5

10-6
0 2 4 6 8 10 12 14

Figure 19:  BER performance of MIMO-OTFS using the estimated channel in a 2 × 2 MIMO system.

time–frequency domain, whereas, the channel in 7 Conclusions

the delay–Doppler representation is slowly vary- We considered OTFS modulation, a 2D modu-
ing and is time-invariant over a larger observa- lation scheme, where signaling is done in the
tion time. This, along with the 2D convolutive delay–Doppler domain and can be imple-
interaction of the symbols with the channel in mented using existing multicarrier schemes
the delay–Doppler domain, enables the proposed using additional pre- and post-processing 2D
channel estimation for MIMO-OTFS to be simple transforms. It was shown to be a suitable mod-
and efficient. ulation scheme for high-mobility use cases
where the high Doppler shifts are encountered.
While conventional modulation schemes such
as OFDM fail to perform well in such high-
mobility scenarios due to Doppler-induced ICI,
OTFS achieved significantly better performance.

20
13 J. Indian Inst. Sci.| VOL xxx:x | xxx–xxx 2020 | journal.iisc.ernet.in
 OTFS: A New Modulation Scheme for High-Mobility Use Cases

Considering a MIMO-OTFS system, we pre- 11. Hadani R, Rakib S, Tsatsanis M, Monk A, Goldsmith AJ,
sented signal detection and channel estimation Molisch AF, Calderbank R (2017) Orthogonal time fre-
schemes and their performance. MIMO-OTFS quency space modulation. In: Proc. IEEE WCNC’2017,
was shown to achieve significantly better per- pp 1–7
formance compared to MIMO-OFDM in high- 12. Hadani R, Monk A (2018) OTFS: a new generation
Doppler environments operating in 4 GHz and of modulation addressing the challenges of 5G. arXiv​
28 GHz frequency bands. The sparse and slow :1802.02623​([cs.IT] 7 Feb 2018)
time-variant nature of wireless channels in the 13. Hadani R, Rakib S, Kons S, Tsatsanis M, Monk A, Ibars C,
delay–Doppler domain enabled simple and effi- Delfeld J, Hebron Y, Goldsmith AJ, Molisch AF, Calder-
cient channel estimation. The performance and bank R (2018) Orthogonal time frequency space modula-
implementation attributes of OTFS make it a tion. arXiv​:1808.00519​v1 ([cs.IT] 1 Aug 2018)
promising modulation scheme for high-mobil- 14. Monk A, Hadani R, Tsatsanis M, Rakib S (2016) OTFS—
ity use cases. orthogonal time frequency space: a novel modulation
technique meeting 5G high mobility and massive MIMO
challenges. arXiv​:1608.02993​ ([cs.IT] 9 Aug 2016)
Publisher’s Note  15. Hadani R, Rakib S, Molisch AF, Ibars C, Monk A, Tsat-
Springer Nature remains neutral with regard to sanis M, Delfeld J, Goldsmith A, Calderbank R (Jun.
jurisdictional claims in published maps and insti- 2017) Orthogonal time frequency space (OTFS) modu-
tutional affiliations. lation for millimeter-wave communications systems. In:
Received: 20 December 2019 Accepted: 20 March 2020 Proc. IEEE MTT-S Intl. Microwave Symp., pp 681–683
16. Surabhi GD, Chockalingam A (2019) OTFS modulation
with phase noise in mmWave communications. In: Proc.
IEEE VTC’2019-Spring
17. Li L, Wei H, Huang Y, Yao Y, Ling W, Chen G, Li P, Cai
References Y (2017) A simple two-stage equalizer with simplified
1. Hattachi RE, Erfanian JJ (eds) (2015) 5G white paper.
orthogonal time frequency space modulation over rap-
NGMN Alliance, Frankfurt a. M.
idly time-varying channels. arXiv​:1709.02505​v1 ([cs.IT]
2. Jakes WC (1994) Microwave mobile communications.
8 Sep 2017)
IEEE Press, New York
18. Raviteja P, Phan KT, Hong Y, Viterbo E (2018) Interfer-
3. Goldsmith A (2005) Wireless communications. Cam-
ence cancellation and iterative detection for orthogonal
bridge University Press, Cambridge
time frequency space modulation. IEEE Trans Wireless
4. Wang T, Proakis JG, Masry E, Zeidler JR (2006) Perfor-
Commun 17(10):6501–6515
mance degradation of OFDM systems due to Doppler
19. Reyhani AR, Farhang A, Ji M, Chen R-R, Farhang-Borou-
spreading. IEEE Trans Wireless Commun 5(6):1422–1432
jeny B (2018) Analysis of discrete-time MIMO OFDM-
5. Strohmer T, Beaver S (2003) Optimal OFDM design for
based orthogonal time frequency space modulation. In:
time-frequency dispersive channels. IEEE Trans Com-
Proc. IEEE ICC’18, pp 1–6
mun 51(7):1111–1122
20. Dean T, Chowdhury M, Goldsmith A (2017) A new
6. Han F-M, Zhang X-D (2007) Hexagonal multicarrier
modulation technique for Doppler compensation in fre-
modulation: a robust transmission scheme for time-
quency-dispersive channels. In: Proc. IEEE PIMRC’2017
frequency dispersive channels. IEEE Trans Signal Process
21. Farhang A, Reyhani AR, Doyle LE, Farhang-Boroujeny
55(5):1955–1961
B (2018) Low complexity modem structure for OFDM-
7. Han F-M, Zhang X-D (2009) Wireless multicarrier digi-
based orthogonal time frequency space modulation. IEEE
tal transmission via Weyl–Heisenberg frames over time-
Wireless Commun Lett 7(3):344–347
frequency dispersive channels. IEEE Trans Commun
22. Murali KR, Chockalingam A (2018) On OTFS modu-
57(6):1721–1733
lation for high-Doppler fading channels. In: Proc.
8. Sayeed AM, Aazhang B (1999) Joint multipath-Doppler
ITA’2018, San Diego
diversity in mobile wireless communications. IEEE Trans
23. Surabhi GD, Augustine RM, Chockalingam A (2019)
Commun 47(1):123–132
On the diversity of uncoded OTFS modulation in dou-
9. Bhashyam S, Sayeed AM, Aazhang B (2000) Time-
bly-dispersive channels. IEEE Trans Wireless Commun
selective signaling and reception for communication
18(6):3049–3063
over multipath fading channels. IEEE Trans Commun
24. Raviteja P, Hong Y, Viterbo E, Biglieri E (2018) Practical
48(1):83–94
pulse-shaping waveforms for reduced-cyclic-prefix OTFS.
10. Wornell GW (1996) Spread-response precoding for com-
IEEE Trans Veh Tech 68(1):957–961
munication over fading channels. IEEE Trans Inform
25. Nimr A, Chafii M, Matthe M, Fettweis G (2018) Extended
Theory 42:488–501
GFDM framework: OTFS and GFDM comparison. In:
Proc. IEEE GLOBECOM’2018

21
J. Indian Inst. Sci. | VOL xxx:x | xxx–xxx 2020 | journal.iisc.ernet.in 13
 M. K. Ramachandran et al.

26. Ramachandran MK, Chockalingam A (2018) MIMO- 34. Augustine RM, Surabhi GD, Chockalingam A (2019)
OTFS in high-Doppler fading channels: signal detec- Space-time coded OTFS modulation in high-Doppler
tion and channel estimation. In: Proc. IEEE GLOBE- channels. In: Proc. IEEE VTC’2019-Spring
COM’2018. Abu Dhabi 35. Tiwari S, Das SS, Rangamgari V (2019) Low complexity
27. Rakib S, Hadani R (2017) Multiple access in wireless tel- LMMSE receiver for OTFS. IEEE Commun Lett. (pre-
ecommunications system for high-mobility applications, print in IEEE Xplore)
US Patent No. US9722741B1 36. Surabhi GD, Chockalingam A (2019) Low-complexity
28. Khammammetti V, Mohammed SK (2019) OTFS based linear equalization for OTFS modulation. IEEE Commun
multiple-access in high Doppler and delay spread wireless Lett. (preprint in IEEE Xplore)
channels. IEEE Trans Wireless Commun 8(2):528–531 37. Samimi MK, Maccartney GR, Sun S, Rappaport TS
29. Surabhi GD, Augustine RM, Chockalingam A (2019) (2016) 28 GHz millimeter-wave ultrawideband small-
Multiple access in the delay-Doppler domain using OTFS scale fading models in wireless channels. In: Proc. IEEE
modulation. arXiv​:1902.03415​v1 ([cs.IT] 9 Feb 2019) VTC’2016-Spring, pp 1–6
30. Augustine RM, Chockalingam A (2019) Interleaved time- 38. Liu P, Springer A (2015) Space shift keying for LOS com-
frequency multiple access using OTFS modulation. In: munication at mmWave frequencies. IEEE Wireless Com-
Proc. IEEE VTC’2019-Fall mun Lett 4(2):121–124
31. Raviteja P, Phan KT, Hong Y (2019) Embedded pilot- 39. Bohagen F, Orten P, Oien G (2007) Design of optimal
aided channel estimation for OTFS in delay-Doppler high-rank line-of-sight MIMO channels. IEEE Trans
channels. IEEE Trans Veh Tech 68(5):4906–4917 Wireless Commun 6(4):1420–1425
32. Surabhi GD, Augustine RM, Chockalingam A (2019) 40. Fish A, Gurevich S, Hadani R, Sayeed AM, Schwartz O
Peak-to-average power of OTFS modulation. IEEE Com- (2013) Delay-Doppler channel estimation in almost lin-
mun Lett 23(6):999–1002 ear complexity. IEEE Trans Inf Theory 59(11):7632–7644
33. Tiwari S, Das SS (2019) Circularly pulse shaped orthog- 41. Hlawatsch F, Matz G (2011) Wireless communications
onal time frequency space modulation. Electron Lett. over rapidly time-varying channels. Academic Press, New
arXiv​:1910.10457​ ([cs.IT] 23 Oct 2019) York

M. K. Ramachandran  received the B.Tech. from the Indian Institute of Technology, Kharagpur, India,
degree in electronics and communication in 1985, and the Ph.D. degree in electrical communication
engineering from the College of Engineer- engineering (ECE) from the Indian Institute of Science
ing Trivandrum, India, in 2015, and the (IISc), Bangalore, India, in 1993. During 1986 to 1993, he
M.Tech degree in electrical communication worked with the Transmission R&D division of the Indian
engineering from the Indian Institute of Sci- Telephone Industries Limited, Bangalore. From December
ence, Bengaluru, India, in 2018. He is currently with the 1993 to May 1996, he was a Postdoctoral Fellow and an
Wireless R&D Division, Qualcomm India Private Ltd., Ben- Assistant Project Scientist at the Department of Electrical
galuru. His interests include low-complexity signal demodu- and Computer Engineering, University of California, San
lation and channel estimation schemes for wireless Diego. From May 1996 to December 1998, he served Qual-
waveforms. comm, Inc., San Diego, CA, as a Staff Engineer/Manager in
the systems engineering group. In December 1998, he joined
G. D. Surabhi  received the B.E. degree in the faculty of the Department of ECE, IISc, Bangalore,
electronics and communication engineering India, where he is a Professor, working in the area of wireless
from Siddaganga Institute of Technology, communications and networking. Dr. Chockalingam is a
Tumkur, India, in 2014. She is currently recipient of the Swarnajayanti Fellowship from the Depart-
pursuing the Ph.D. degree with the Depart- ment of Science and Technology, Government of India. He
ment of Electrical Communication Engi- served as an Associate Editor of the IEEE TRANSACTIONS
neering, Indian Institute of Science, Bangalore, India. Her ON VEHICULAR TECHNOLOGY, and as an Editor of the
current research interests include mmWave MIMO commu- IEEE TRANSACTIONS ON WIRELESS COMMUNICA-
nication systems and orthogonal time–frequency space TIONS. He served as a Guest Editor for the IEEE JOURNAL
modulation. ON SELECTED AREAS IN COMMUNICATIONS (Special
Issue on Multiuser Detection for Advanced Communication
A. Chockalingam  (S’92-M’93-SM’98) was Systems and Networks), and for the IEEE JOURNAL OF
born in Rajapalayam, Tamil Nadu, India. He SELECTED TOPICS IN SIGNAL PROCESSING (Special
received the B.E. (Honors) degree in elec- Issue on Soft Detection on Wireless Transmission). He is a
tronics and communication engineering Fellow of the Indian National Academy of Engineering, the
from the P. S. G. College of Technology, National Academy of Sciences, India, the Indian National
Coimbatore, India, in 1984, the M.Tech. Science Academy, and the Indian Academy of Sciences.
degree in electronics and electrical communications engi-
neering (with specialization in satellite communications)

22
13 J. Indian Inst. Sci.| VOL xxx:x | xxx–xxx 2020 | journal.iisc.ernet.in