This article has been accepted for publication in a future issue of this journal, but has not been

fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/LCOMM.2019.2956709, IEEE
Communications Letters

IEEE COMMUNICATIONS LETTERS, VOL. XX, NO. XX MONTH 2019 1

Low-complexity Linear Equalization for OTFS Modulation

G. D. Surabhi and A. Chockalingam

Abstract—In this letter, we propose low-complexity linear equalizer followed by an initial MMSE equalization in time-
equalizers for orthogonal time frequency space (OTFS) modula- frequency domain. The performance of OTFS modulation
tion that exploit the structure of the effective channel matrix in with MMSE equalizer and decision feedback equalizer has
OTFS. The proposed approach exploits the block circulant nature
of the OTFS channel matrix to achieve significant complexity been discussed in [2]. It has been shown in [2] that the
reduction. For an N × M OTFS system, where N and M are BER performance of uncoded OTFS with MMSE equalizer
the number of Doppler and delay bins, respectively, the proposed is significantly better compared to that of OFDM for lower
approach gives exact minimum mean square error (MMSE) and modulation order (BPSK, 4-QAM, and 16-QAM) and the
zero-forcing (ZF) solutions with just O(M N log M N ) complexity, performance gains degrade for higher order QAM (64-QAM
while MMSE and ZF solutions using the traditional matrix
inversion approach require O(M 3 N 3 ) complexity. The proposed and 256-QAM). It has also been suggested in [2] that this
approach can provide low complexity initial solutions for local performance can be improved significantly by the use of a
search techniques to achieve enhanced bit error performance. non-linear equalizer following a linear equalizer. Motivated
keywords: OTFS modulation, linear equalizers, block cir- by this and the fact that the overall detection complexity can
culant matrices, computational complexity. reduce by reducing the complexity of the linear equalizer
preceding the non-linear equalizer, in this letter, we propose
I. I NTRODUCTION low-complexity linear equalizers for OTFS signal detection.
The challenge of establishing reliable high-speed communi- Our novel approach in this effort is the exploitation of the
cation in extremely dynamic environments has been addressed structure prevalent in the effective channel matrix in OTFS
recently by a new modulation technique called the ‘orthogonal to reduce the computational complexity of linear equalizers,
time frequency space (OTFS)’ modulation. It has been shown which has not been reported before.
that the OTFS modulation achieves superior bit error rate We propose low-complexity MMSE and zero-forcing (ZF)
performance (BER) compared to the conventional multicarrier equalizers for OTFS signal detection, which do not rely on the
techniques like orthogonal frequency division multiplexing traditional matrix inversion approach that conventional MMSE
(OFDM), in very high Doppler wireless environments. The and ZF equalizers employ. Instead, the proposed approach
fundamental premise of OTFS modulation is the representation achieves exact MMSE and ZF solutions at a much lower
of the channel and the information symbols in the delay- complexity compared to the matrix inversion approach. This
Doppler domain rather than the time-frequency domain as is made possible by recognizing a certain structure in the
done in conventional multicarrier modulation techniques. A effective delay-Doppler channel matrix in OTFS modulation.
key advantage of the delay-Doppler representation of wireless Specifically, the proposed approach recognizes the block cir-
channels is that the rapid fluctuation of the time varying culant nature of the OTFS channel matrix and exploits the
channel exhibits slow variations when viewed in the delay- properties of block circulant matrices to achieve significant
Doppler domain. This, along with the fact that the channel complexity reduction. For example, for an N × M OTFS
in the delay-Doppler domain has a sparse nature, simplifies system, where N and M are the number of Doppler and
channel estimation in rapidly time varying wireless channels. delay bins, respectively, the complexity of MMSE and ZF
OTFS modulation was first proposed in [1], where it was equalizers using the conventional matrix inversion approach
demonstrated to have superior error performance compared is O(M 3 N 3 ), whereas the proposed approach gives exact
to OFDM in very high Doppler environments. This was MMSE and ZF solutions using a computation complexity of
followed by several works addressing various aspects of OTFS just O(M N log M N ). This complexity reduction of linear
modulation [2]-[8]. Low complexity detection of OTFS signal equalizers can aid efficient realizations of non-linear equalizers
has been addressed using message passing based detector to achieve enhanced bit error performance at reduced overall
and Markov chain Monte Carlo based algorithm in [3] and complexity. We illustrate this by using the proposed MMSE
[4], respectively. Further, while [6] uses a minimum mean equalizer solution as the initial solution to a local search based
squared error (MMSE) detector for OTFS signal detection, [7] non-linear equalizer.
proposes an interference cancellation based low-complexity II. OTFS MODULATION
We consider OTFS modulation architected over a general
Copyright (c) 2019 IEEE. Personal use of this material is permitted.
However, permission to use this material for any other purposes must be
multicarrier modulation system as shown in Fig. 1. The
obtained from the IEEE by sending a request to [email protected]. transforms involved in OTFS modulation and demodulation
Manuscript received October 18, 2019; revised November 22; accepted are introduced in the following subsections.
November 22, 2019. The editor coordinating the review of this manuscript
and approving it for publication was Prof. Fanggang Wang.
A. OTFS transmitter
This work was supported in part by the J. C. Bose National Fellowship, The OTFS transmitter considers N M information sym-
Department of Science and Technology, Government of India, Tata Elxsi bols on an N × M delay-Doppler grid, denoted by x[k, l],
Limited, Bengaluru 560048, and the Intel India Faculty Excellence Program. k = 0, · · · , N − 1, l = 0, · · · , M − 1 from a modulation
The authors are with the Department of Electrical Communication En-
gineering, Indian Institute of Science, Bengaluru 560012, India (e-mail: alphabet A, to be transmitted in a given packet burst of
[email protected]; [email protected]). duration N T , occupying a bandwidth of B = M ∆f , where

1089-7798 (c) 2019 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/LCOMM.2019.2956709, IEEE
Communications Letters

IEEE COMMUNICATIONS LETTERS, VOL. XX, NO. XX MONTH 2019 2

 ′

l−l′
Pre-processing
Time-frequency modulation (TF domain)
Post-processing where hw k−k N T M ∆f, = hw (ν, τ )|ν= k−k′ ,τ = l−l′ and
NT M ∆f
x[k, l] ISFFT & X[n, m] Heisenberg x(t) y(t) Wigner Y [n, m] Receive y[k, l] hw (ν, τ ) is the circular convolution of the channel response
transmit channel windowing
windowing transform transform & SFFT with a windowing function w(ν, τ ) as defined in [2]. The
above equation can be vectorized as [3]
OTFS transform Inverse OTFS transform
OTFS modulation (Delay-Doppler domain) y = Hx + n, (9)
where x ∈ CM N ×1 , xk+N l = x[k, l] ∈ A, y ∈ CM N ×1 is
Fig. 1: OTFS modulation scheme.
the received vector, H ∈ CM N ×M N is the effective channel
matrix in delay-Doppler domain, and n ∈ CM N ×1 is the
∆f = 1/T . The information symbols x[k, l]s are treated as
AWGN vector whose entries are distributed as CN (0, σ 2 ).
points on the 2D N × M delay-Doppler grid and are mapped
to time-frequency plane using inverse symplectic finite Fourier III. L OW- COMPLEXITY LINEAR RECEIVERS FOR OTFS
Consider the vectorized input-output relation in OTFS given
transform (ISFFT), given by
by (9). The estimates of the transmitted symbols at the output
N −1 M −1
1 X X nk ml of a linear receiver will be of the form x̂ = f (Gy), where
X[n, m] = x[k, l]ej2π( N − M ) . (1) G is a linear transformation matrix which depends on the
MN
k=0 l=0
type of linear receiver and f (.) is a function that maps
The TF signal so obtained is converted to time domain for each entry of the vector Gy to the nearest symbol in the
transmission using Heisenberg transform, given by modulation alphabet A in terms of Euclidean distance. The
N −1 M −1
X X linear transformation matrices for MMSE and ZF are given
x(t) = X[n, m]gtx (t − nT )ej2πm∆f (t−nT ) , (2) by GMMSE = (HH H + σ 2 I)−1 HH and GZF = (HH H)−1 HH ,
n=0 m=0
respectively. Observe that the computation of both GMMSE and
where gtx (t) denotes the transmit pulse shape. The time GZF involves inversion of M N × M N matrices, which has a
domain signal x(t) is then transmitted through the time varying computational complexity of O(M 3 N 3 ). This is not attractive
wireless channel. Denoting the complex baseband response of for large values of M and N . The computation of GMMSE and
the channel in delay-Doppler domain by h(τ, ν), where τ and GZF can be achieved with significantly lower complexity if the
ν are delay and Doppler variables, respectively, the received structures of H, GMMSE , and GZF are carefully exploited.
time domain signal is given by
Z Z A. Low-complexity MMSE equalizer
y(t) = h(τ, ν)x(t − τ )ej2πν(t−τ ) dτ dν. (3) Consider the vectorized input-output relation in (9). The
ν τ M N × M N channel matrix H in (9) is a block circulant
B. OTFS receiver matrix with M circulant blocks, each of size N × N . Let
At the receiver, the received time domain signal y(t) is BM,N denote the class of block circulant matrices with M
transformed into a TF signal using Wigner transform, which circulant blocks of size N × N . Denoting circular matrices
match filters y(t) with received pulse shape grx (t) and samples with circ(), any matrix A ∈ BM,N will be of the form
it at the lattice points t = nT and f = m∆f . The Wigner circ(A0 , A1 , · · · , AM −1 ), where each block Ai is an N × N
transform is given by circulant matrix of the form circ(ai,0 , ai,1 , · · · , ai,N −1 ) [11].
Z
′ In order to derive the low-complexity MMSE detector for
∗
Agrx ,y (t, f ) = grx (t′ − t)y(t)e−j2πf (t −t) dt′ ,
OTFS, we make use of the following three properties of BM,N .
Y [n, m] =Agrx ,y (t, f )|t=nT,f =m∆f . (4) 1) Any matrix A ∈ BM,N is diagonalized by the unitary
matrix (FM ⊗ FN ), where FM and FN denote discrete
If the transmit pulse gtx (t) and the receive pulse grx (t) satisfy Fourier transform (DFT) matrices of size M ×M and N ×
biorthogonality and robustness condition in [2], the input- N , respectively, and ⊗ denotes the Kronecker product of
output relation of the TF modulation is given by two matrices [11]. Hence, A can be written as
Y [n, m] = H[n, m]X[n, m] + V [n, m], (5) A = (FM ⊗ FN )H D(FM ⊗ FN ), (10)
where V [n, m] is the additive white Gaussian noise (AWGN) where D = diag{d1 , d2 , · · · , dM N } and di denotes the
at the output of Wigner transform, and H[n, m] is given by ith eigen value of A.
Z Z
2) For any matrix A ∈ BM,N , the entries of the matrix D
H[n, m] = h(τ, ν)ej2πνnT e−j2π(ν+m∆f )τ dνdτ. (6)
τ ν will be of the form M −1 X
The TF signal Y [n, m] so obtained is mapped back to the D= ΩiM ⊗ Di , (11)
delay-Doppler domain using symplectic finite Fourier trans- i=0

form (SFFT), given by where ΩM = diag{1, ω, · · · , ω M −1 } with ω = ej2π/M

N −1 M −1 and Di denotes the N × N diagonal matrix containing
nk ml
the eigen values of the N × N circulant block Ai .
X X
y[k, l] = Y [n, m]e−j2π( N − M ) . (7)
3) For any two matrices A, B ∈ BM,N , the matrices AT ,
n=0 m=0 P
Using (1)-(7), the input-output relation can be derived as [2] AH , AB(= BA), γ1 A + γ2 B, R−1 r
r=0 γr A , and A
−1
(if it
exists) are all in BM,N , where γ0 , · · · , γR−1 are scalars.
M −1 N −1
k − k ′ l − l′
 
1 X X Now, consider the OTFS channel matrix in (9), which is a
y[k, l] = x[k ′ , l′ ]hw , +v[k, l],
MN ′ N T M ∆f block circulant matrix with M blocks each of size N × N .
l =0 k′ =0
(8) Since H ∈ BM,N , it can be decomposed as

1089-7798 (c) 2019 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/LCOMM.2019.2956709, IEEE
Communications Letters

IEEE COMMUNICATIONS LETTERS, VOL. XX, NO. XX MONTH 2019 3

H = (FM ⊗ FN )H Λ(FM ⊗ FN ), (12) The complexity of obtaining Λ using (18) involves the
where Λ = diag{λ1 , λ2 , · · · , λM N } and λi denotes the ith computation of N M -point IDFTs. Therefore, the com-
eigen value of H. Next, consider the transformation matrix plexity of this step is O(M N log M ).
GMMSE = (HH H + σ 2 I)−1 HH . From property 3), it can be 3) Computation of eigen values of GMMSE : Next, using the
seen that if H ∈ BM,N , then GMMSE ∈ BM,N . Therefore, the eigen values of H, the eigen values of GMMSE are com-
eigen value decomposition of GMMSE will be of the form puted using (16). This steps requires the computation of
Ψ1 , Ψ2 , · · · , ΨM N using λ1 , λ2 , · · · , λM N , respectively.
GMMSE = (FM ⊗ FN )H Ψ(FM ⊗ FN ), (13)
So the order of complexity for computing Ψ is O(M N ).
where Ψ is a diagonal matrix of size M N × M N containing 4) Computation of GMMSE y: The next step is to compute
the eigen values of GMMSE . Substituting (12) in GMMSE = GMMSE y = (FM ⊗ FN )H Ψ(FM ⊗ FN )y. This equation
(HH H + σ 2 I)−1 HH and simplifying, we get can also be evaluated with low complexity using FFTs
GMMSE = (FM ⊗ FN )H (Λ∗ Λ + σ 2 I)−1 Λ∗ (FM ⊗ FN ). (14) and IFFTs. Let Y denote the N × M matrix such that
Now, comparing (13) and (14), Ψ can be written as vec(Y) = y. Then, (FM ⊗ FN )y can be written as
Ψ = (Λ∗ Λ + σ 2 I)−1 Λ∗ . (15) z = (FM ⊗ FN )y = vec(FN YFH M ). (19)
Therefore, the eigen values of GMMSE can be expressed in terms Observe that z can be obtained by computing N -point
of the eigen values of H as DFT along the columns of Y and M -point IDFT along
λ∗1 λ∗2 λ∗M N the rows of Y. Thus, the complexity of computing z
 
Ψ = diag , , · · · , . using FFTs and IFFTs is O(M N log M + M N log N ) =
|λ1 |2 + σ 2 |λ2 |2 + σ 2 |λM N |2 + σ 2
(16) O(M N log M N ). Next, Ψ(FM ⊗ FN )y can be com-
From (16), it is clear that the eigen values of GMMSE can puted by computing Ψz. Since Ψ is a diagonal matrix,
be computed from the eigen values of H. The computation of the computation of Ψz has a complexity of O(M N ).
eigen values of GMMSE using the eigen values of H is one of the Now, let q = Ψz and Q denote the N × M matrix
key steps in the proposed low-complexity MMSE equalizer for such that vec(Q) = q. With this, GMMSE y = (FM ⊗
OTFS. Using GMMSE , the transmitted symbols are estimated by FN )H Ψ(FM ⊗ FN )y can be computed as
computing GMMSE y with significantly lower complexity using GMMSE y = vec(FH
N QFM ). (20)
FFTs, IFFTs, and the properties of H and GMMSE as follows. Again, this step involves the computation of N -point
1) Computation of eigen values of each block of H: The first IDFT along the columns of Q and M -point DFT along
step in the proposed low-complexity MMSE equalizer is the rows of Q. Hence, this step has a complexity of
the computation of the eigen values of H. As discussed O(M N log M N ). Therefore, the overall complexity of
previously, H has a block circulant structure with cir- step 4 is O(2M N log M N + M N ).
culant blocks, with an eigen value decomposition of the B. Low-complexity ZF equalizer
form (12). The matrix Λ in (12) can be written as The transformation matrix for zero forcing detector is given
PM −1
Λ = i=0 ΩiM ⊗ Λi by GZF = (HH H)−1 HH . As discussed in the previous
subsection, the H is a block circulant matrix with circulant
 
M P−1 MP−1 MP−1
=diag Λi , e j2πi/M
Λi ,··· , e j2π(M −1)i/M
Λi ,
blocks and has an eigen value decomposition as in (12). Since
i=0 i=0 i=0

(17) GZF = (HH H)−1 HH ∈ BM,N , GZF can be written as

where Λi denotes the N × N diagonal matrix containing GZF = (FN ⊗ FM )H Ψ(FN ⊗ FM ), (21)
the eigen values of ith block in H. The computation ∗ −1 ∗
where Ψ = (Λ Λ) Λ . From the above equation, it is easy
of eigen values of H involves the following two steps: to see that the transformation matrix GZF can be computed
computing Λi s by computing the eigen values of the using FFTs, IFFTs, and diagonal matrix Ψ containing the
circulant blocks of H and computing Λ using them as in eigen values of GZF . Also, Ψ = (Λ∗ Λ)−1 Λ∗ can be obtained
(17). Since each block in H has a circulant structure, the from Λ, just by computing λ∗i /|λi |2 = 1/λi , i = 1, · · · , M N . The
eigen values of each block are computed by computing low-complexity ZF equalizer can be implemented as follows:
the DFT of the first row of each block of H. Therefore,
1) Computation of eigen values of each block of H: The
computation of Λ1 , Λ2 , · · · , ΛM requires the computa-
eigen values of M circulant blocks of H are computed
tion of eigen values of M number of N × N circulant
using DFTs. The computational complexity of this step
matrices and hence has O(M N log N ) complexity.
is O(M N log N ).
2) Computation of the eigen values of H: Next, using
2) Computation of eigen values of H: The eigen values
Λi s, the eigen values of H are computed using (17).
of H (Λ) are computed using (17) as in the case of
A closer look at (17) reveals that the entries of Λ
MMSE equalizer. This step can be implemented using
can be computed with low complexity using IFFT. Let
IDFT as discussed in Sec. III-A and has a complexity of
v1 , v2 , · · · , vM denote M column vectors of dimension
O(M N log M ).
N × 1 whose entries contain the diagonal elements of
3) Computation of eigen values of GZF : The eigen values of
T
Λ1 , Λ2 , · · · , ΛM , respectively. Let V = [v1T v2T · · · vM ]
GZF are computed using the eigen values of H as
denote the M × N matrix whose rows are the vectors
v1 , v2 , · · · , vM . Now, (17) can be computed as Ψ = diag{1/λ1 , 1/λ2 , · · · , 1/λM N }. (22)
Λ = diag{(FH T The computational complexity of this step is O(M N ).
M V) }. (18)

1089-7798 (c) 2019 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/LCOMM.2019.2956709, IEEE
Communications Letters

IEEE COMMUNICATIONS LETTERS, VOL. XX, NO. XX MONTH 2019 4

4) Computation of GZF y: The computation of GZF y is 10 0 10 0

similar to the computation of GMMSE y as in Sec. III-A.

10 -1
First, z = (FM ⊗ FN )y = vec(FN YFH M ) is computed
10 -2

using N -point DFT along the columns of Y and M -point

Bit error rate

Bit error rate

10 -2

IDFT along the rows of Y. Next, q = Ψz is computed. 10 -4

10 -3
Finally, GZF y is computed as GZF y = vec(FH N QFM ),
-6
where q = vec(Q). Overall complexity of this step is 10
10 -4

O(2M N log M N + M N ).
10 -8 10 -5
The proposed algorithm is summarized as follows. 0 5 10 15 20 0 5 10 15 20

1: Inputs: y, H. SNR in dB SNR in dB

2: Step 1: Compute eigen values of M circulant blocks of H (a) (b) EVA model [12].
by computing DFTs of first row of each circulant block. Fig. 2: BER performance of the proposed low-complexity ZF
3: Step 2: Compute Λ using (18). and MMSE equalizers.
4: Step 3: Compute Ψ (Ψ = (Λ∗ Λ + σ 2 I)−1 Λ∗ in case of The channel gains (hi s) are assumed to be i.i.d and dis-
MMSE and Ψ = (Λ∗ Λ)−1 Λ∗ in case of ZF). tributed as CN (0, 1/P ). The signal-to-noise ratio (SNR) is
5: Step 4: Compute GMMSE y/GZF y by computing z = defined as E[kHxk2 ]/σ 2 , where E[.] denotes the expectation
vec(FN YFH H
M ), q = Ψz, and vec(FN QFM ).
operator and σ 2 is the noise variance. A delay Doppler grid
The complexities associated with each step of the algorithm with M = 32 and N = 32 is considered. A carrier frequency
are presented in Table I. At this stage, we make the following of 4 GHz, subcarrier spacing of 15 kHz, and BPSK modulation
remark on the proposed equalization approach. is considered. Firstly, from the figure, we observe that the
TABLE I: Order of complexity. performance of the MMSE equalizer is superior compared to
Step # Order of complexity
that of the ZF equalizer. Further, the proposed low-complexity
1: computation of eigen values of circu- O(M N log N ) ZF and MMSE equalizers show exactly the same BER perfor-
lant blocks of H mance as that of the conventional ZF and MMSE equalizers,
2: computation of Λ O(M N log M ) respectively. This demonstrates that, as expected, the proposed
3: computation of Ψ O(M N )
4: computation of Gy O(2M N log M N + M N )
low-complexity algorithms provide the exact MMSE and ZF
Overall complexity (dominated by Step 4) O(M N log M N ) solutions and hence the same BER performance as that of the
conventional MMSE and ZF equalizers. Figure 2b shows the
Remark: It should be noted that the GMMSE and GZF equal-
BER performance of OTFS with the proposed low-complexity
izers in the proposed approach perform equalization in the
MMSE and ZF equalizers considering the channel parameters
delay-Doppler domain, unlike the equalizers in [7],[8] which
according to the extended vehicular A model (EVA) in 3GPP
perform equalization in the TF domain. For the received signal
[12]. A channel with P = 9 paths with a delay profile as per
to be accessed in the TF domain, the TF equalization approach
EVA model and Jakes Doppler spectrum is considered. The
requires the two-step OTFS implementation as in Fig. 1, viz.,
Doppler shifts νi s corresponding to the ith path in different re-
time domain to TF domain (using Wigner transform) and from
alizations are generated randomly using νi =νmax cos θi , where
TF domain to delay-Doppler domain (using SFFT). On the
νmax is the maximum Doppler shift which is taken to be 1.34
other hand, the proposed direct delay-Doppler equalization
kHz, and θi s are uniformly distributed over [−π, π]. A carrier
approach allows one-step OTFS implementation using Zak
frequency of 4 GHz, subcarrier spacing of 15 kHz, and frame
transform [2], which has significant complexity advantage
size (M, N ) = (512, 12) are considered. For the considered
compared to the two-step approach with TF equalization.
system with EVA channel model parameters recommended
Specifically, the implementation of OTFS with a single Zak
by 3GPP, the proposed low-complexity MMSE/ZF equalizers
transform has a complexity of M N log N (as it involves the
require 73731 real operations, whereas the conventional matrix
use of M FFTs of length N ), whereas the two-step OTFS im-
inversion based equalizers require 9.27 × 1011 operations.
plementation has a complexity of M N log M +M N log M N .
Next, Fig. 3 shows the computational complexity of the
IV. R ESULTS AND DISCUSSIONS conventional MMSE/ZF equalizers and the proposed low-
Figure 2a shows the BER performance comparison of OTFS
complexity equalizers in terms of the number real operations
with conventional ZF and MMSE equalizers and the proposed
involved as a function of M . For simulations, N = 32 is
low-complexity ZF and MMSE equalizers. We consider a
considered and all other parameters are the same as those
channel with P paths, with the ith path having a channel gain
used in Fig. 2a. From Fig. 3, it can be observed that the
hi , delay τi , and Doppler νi . The channel response in the
complexity of the proposed MMSE equalizer is significantly
delay-Doppler
PPdomain, denoted by h(τ, ν), can be expressed
−1 less compared to that of conventional MMSE equalizer. For
as h(τ, ν)= i=0 hi δ(τ − τi )δ(ν − νi ) [3]. The simulation set
example, the number real operations required to compute
up considered in Fig. 2(a) uses the delay-Doppler value pairs
MMSE solution for a system with M = 128 and N = 32
for eight paths (P = 8) as shown in the Table given below,
is 2.75 × 1011 in case of the conventional equalizer, whereas
where the ith path has delay τi and Doppler νi .
it is 49,155 in case of the proposed MMSE equalizer.
Path, i 1 2 3 4 5 6 7 8
Performance of OFDM based OTFS: The OTFS implemen-
τi (µs) 0 1.04 2.08 3.12 4.16 5.20 6.25 7.29
νi (Hz) 0 469 469 938 1406 1406 1875 1875 tation discussed so far used the TF modulation with ideal pulse
satisfying biorthogonal and robustness property, for which

1089-7798 (c) 2019 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/LCOMM.2019.2956709, IEEE
Communications Letters

IEEE COMMUNICATIONS LETTERS, VOL. XX, NO. XX MONTH 2019 5

10 0 10 0
10 12

10 -2
10 10 10 -2

No. real operations

Bit error rate

Bit error rate

10 8 10 -4
10 -4
10 6 10 -6

10 -6
10 4 10 -8
50 100 150 200 250

10 -8 10 -10
0 5 10 15 20 0 5 10 15 20
Fig. 3: Computational complexity of conventional and pro- SNR in dB SNR in dB
posed low-complexity MMSE equalizers. (a) (b)
the input-output relation is of the form (5)† . However, for Fig. 4: Performance of proposed low-complexity ZF, MMSE
an OFDM based OTFS system [5] which uses rectangular equalizers and MMSE-LAS in OFDM based OTFS systems.
pulse (non-ideal), the input-output relation in (5) does not Importantly, though OFDM based OTFS with the proposed
hold and hence the effective channel matrix in the delay- MMSE equalizer showed some degraded performance in Fig.
Doppler domain will no longer have block circulant with 4a, we observe in Fig. 4b that the use of it with a LAS
circulant blocks structure. Hence, diagonalization of the effec- equalizer achieves almost the same improved performance
tive channel matrix using FM ⊗ FN may not yield a strictly compared to that of the use of it in OTFS with ideal pulse.
diagonal matrix. As a result, the proposed low-complexity This improved performance is achieved at a significantly low
schemes may not yield exact MMSE/ZF solutions in OFDM complexity. Figure 4b also shows the comparison of the BER
based OTFS implementations. The effect of this observation is performance of the proposed MMSE-LAS detector with the
illustrated in Fig. 4a, which shows the performance of OFDM message passing (MP) based detection proposed in [3]. From
based OTFS. From Fig. 4a, we observe that the performance this figure, we observe that MMSE-LAS and the MP detection
of the proposed low-complexity MMSE/ZF equalizers gets have almost the same performance up to 12 dB SNR, after
poor in OFDM based OTFS compared to that in OTFS with which the BER of MP detector floors, whereas the BER of
ideal pulse. This is because the proposed equalizers assume MMSE-LAS detector continues to decrease. This is because
block circulant with circulant blocks structure for the effective the Gaussian approximation used in the computation of mes-
channel matrix. However, as will be shown next, the proposed sages in MP detection involves sum of P − 1 interference
MMSE/ZF equalizer followed by a low-complexity non-linear terms. Since P = 8 in Fig. 4b, the approximation is less
equalizer achieves almost the same improved performance in accurate because of the small number of terms involved in
OTFS with ideal pulse as well as OFDM based OTFS. the approximation, and this leads to an error floor.
Proposed MMSE equalizer followed by LAS equalizer: We R EFERENCES
consider a local search based non-linear equalizer called [1] R. Hadani, S. Rakib, M. Tsatsanis, A. Monk, A. J. Goldsmith, A. F.
likelihood ascent search (LAS) equalizer [10] following the Molisch, and R. Calderbank, “Orthogonal time frequency space modula-
proposed low-complexity linear equalizer. The LAS algorithm tion,” in Proc. IEEE WCNC’2017, Mar. 2017, pp.1-7.
[2] R. Hadani, S. Rakib, S. Kons, M. Tsatsanis, A. Monk, C. Ibars, J. Delfeld,
is a low-complexity algorithm that starts with an initial solu- Y. Hebron, A. J. Goldsmith, A. F. Molisch, and R. Calderbank, “Or-
tion and searches for good solutions in the neighborhood until thogonal time frequency space modulation,” online: arXiv:1808.00519v1
a local optimum is reached. The computational complexity [cs.IT] 1 Aug 2018.
[3] P. Raviteja, K. T. Phan, Y. Hong, and E. Viterbo, “Interference can-
of the LAS algorithm depends on the computation of the cellation and iterative detection for orthogonal time frequency space
initial solution and the search operation. The complexity modulation,” IEEE Trans. Wireless Commun., vol. 17, no. 10, pp. 6501-
of obtaining an initial MMSE solution using conventional 6515, Aug. 2018.
[4] K. R. Murali and A. Chockalingam, “On OTFS modulation for high-
matrix inversion approach is O(M 3 N 3 ), which is O(M 2 N 2 ) Doppler fading channels,” in Proc. ITA’2018, Feb. 2018, pp. 1-10.
complexity per symbol. Further, it has been shown through [5] A. Farhang, A. R. Reyhani, L. E. Doyle, and B. Farhang-Boroujeny, “Low
simulations that the search operation in the LAS algorithm complexity modem structure for OFDM-based orthogonal time frequency
space modulation,” IEEE Wireless Commun. Lett., vol. 7, no. 3, pp. 344-
requires an average per-symbol complexity of O(M N ) [10]. 347, Jun. 2018.
The proposed low-complexity MMSE equalizer has a per- [6] A. Nimr, M. Chafii, M. Matthe, and G. Fettweis, “Extended GFDM frame-
symbol complexity of O(log M N ). Therefore, the use of the work: OTFS and GFDM comparison,” in Proc. IEEE GLOBECOM’2018,
Dec. 2018, pp. 1-6.
proposed low-complexity MMSE equalizer to obtain the initial [7] T. Zemen, M. Hofer, and D. Loeschenbrand, “Low-complexity equal-
solution for the LAS algorithm significantly reduces the overall ization for orthogonal time and frequency signaling (OTFS),” Online:
complexity. Figure 4b shows the BER performance of this arXiv:1710.09916v1 [cs.IT] 26 Oct 2017.
[8] T. Zemen, M. Hofer, D. Loeschenbrand, and C. Pacher, “Iterative de-
MMSE-LAS equalizer for i) OTFS with ideal pulse and ii) tection for orthogonal precoding in doubly selective channels,” in Proc.
OFDM based OTFS. From this figure, we observe that the IEEE PIMRC’2018, Sep. 2018, pp. 1-7.
performance of MMSE-LAS equalizer is significantly superior [9] T. Strohmer and S. Beaver, “Optimal OFDM design for time-frequency
dispersive channels,” IEEE Trans. Commun., vol. 51, no. 7, pp. 1111-
compared to that of the MMSE equalizer in both cases. 1122, Jul. 2003.
[10] A. Chockalingam and B. S. Rajan, Large MIMO Systems, Cambridge
† Given the constraint imposed by the uncertainty principle, ideal pulses Univ. Press, Feb. 2014.
are non-realizable in practice. However, pulses whose support is highly [11] P. J. Davis, Circulant Matrices, American Mathematical Society, 2012.
concentrated in time and frequency minimize the cross-symbol interference, [12] Evolved Universal Terrestrial Radio Access (E-UTRA); Base Station
and hence closely approximate the ideal pulses. Design of such near-ideal (BS) Radio Transmission and Reception, 3GPP TS 36.104, Ver. 14.3.0,
pulses has been addressed in [9]. Rel. 14, Apr. 2017.

1089-7798 (c) 2019 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.