1

Effective Diversity of OTFS Modulation

P. Raviteja, Yi Hong, Emanuele Viterbo, and Ezio Biglieri

Abstract—Orthogonal time frequency space (OTFS) modula- Notations: Scalars, vectors, and matrices are denoted a, a,
tion, which encodes information symbols in the delay–Doppler and A, respectively. A† and Ai denote Hermitian trans-pose
domain, offers a promising solution to the problem of high
and ith power of A. We write a = vec(A) for the column-wise
Doppler sensitivity of orthogonal frequency division multiplexing
(OFDM) transmission. In this paper we study the diversity of vectorization of matrix A. Symbols a(i) and A(i, j) denote the
OTFS assuming rectangular waveforms and a delay–Doppler ith element of a and (i, j)th element of A, respectively.
channel with two paths. After introducing the concept of ef- Further, A = diag[a0, · · · , aN−1] denotes a diagonal matrix of
fective diversity (ED), which we argue to be more significant size N with {a0, · · · , aN−1} as its diagonal elements. IN is the
than “standard” diversity in the case of a large number of
transmitted symbols, we examine the conditions under which
identity matrix of size N, and FN = { √1N e2πjkl/N }kN,l−=01
OTFS achieves full ED for QAM symbols. We validate our and F†N are the N-point DFT and the IDFT matrices,
analytical results through numerical simulations, which show that respectively. Symbol ⊗ denotes Kronecker product, and CN (0,
OTFS practically achieves full ED with sufficiently large signal N0IN ) an i.i.d. Gaussian random vector with mean 0 and
constellations. covariance matrix N0IN . Notation [ · ]N denotes mod-N
Index Terms—Delay–Doppler channel, diversity, OTFS. operation, and gcd(a, b) the greatest common divisor of a and
b. Finally, Z and Z[j] denote the set of integers and the number
field whose elements have the form a+bj, with a and b
I. I NTRODUCTION 1
integers, respectively. Finally, Q(e−j2π N ) denotes the
A simple and effective way of evaluating symbol error rates cyclotomic field obtained by adjoining an N-th root of unity to
in digital communications is based on the concept of pairwise the set of rational numbers.
error probabilities (PEPs) [1], whose sum over all transmitted
signal pairs yields an upper bound to said rate. With Rayleigh II. S YSTEM M ODEL
fading channels, performance is often expressed using a single We consider OTFS modulation with one transmit and one
parameter, called diversity, representing the minimum slope of receive antenna. Let X ∈ AM ×N denote the two-dimensional
the PEP-vs.-SNR curves across all signal pairs. This turns out (QAM) information symbols in the delay–Doppler domain,
to be the slope of the actual symbol error probability for high where the QAM alphabet A ∈ {a0 , · · · , aQ−1 }. Note that A ⊂
SNR. Now, in the case of a large number of transmitted signals Z[j]. Assuming rectangular waveforms, the OTFS transmitted
and intermediate SNR, diversity may not be a meaningful signal can be written as [5]
parameter, as a large majority of PEP curves may exhibit a  
slope much steeper than the minimum. In these conditions, s = vec F†M (FM XF†N ) = (F†N ⊗ IM )x (1)
we advocate the use of an effective diversity (ED), which
accounts for the slope of the majority of PEP curves rather where x , vec(X) ∈ AM N . For efficient FFT implementa-
than the minimum one. In this paper, we focus on studying tion, we assume M and N are powers of 2.
this parameter for orthogonal time frequency space (OTFS) The received signal can be put in the form [5]
modulation, which is a recently proposed waveform for high r = Hs + w, (2)
Doppler channels [2], [3]. Some other works on OTFS may
be found in [4]–[14]. where, under the assumption that the channel admits a P -path
Recently, in [8], the authors analyzed the conventional sparse representation as described in [6], H is the M N ×M N
diversity of OTFS for ideal bi-orthogonal waveforms and channel matrix
showed that it is one. However, this analysis does not reflect P
X
the PEP values at practical SNR’s and also it is not valid for H= hi Πli ∆ki , (3)
practical waveforms. In this paper, we study the performance i=1
of OTFS using ED with practical rectangular waveforms and with Π the permutation matrix (forward cyclic shift), ∆ the
operating over a delay–Doppler channel with two paths. ED M N × M N matrix ∆ , diag z 0 , z 1 , · · · , z M N −1 , z ,

is derived, and numerical simulations are used to validate our j2π
e M N , and w ∼ CN (0, N0 I).
results. After OTFS processing, the received signal in delay–
Raviteja, Hong, and Viterbo are with ECSE Department, Monash Uni- Doppler domain becomes
versity, Clayton, VIC 3800, Australia, Email: {raviteja.patchava, yi.hong,
emanuele.viterbo}@monash.edu. Biglieri is with Universitat Pompeu Fabra, y = (FN ⊗ IM )H(F†N ⊗ IM )x + (FN ⊗ IM )w
Barcelona, Email: [email protected].
The work of the first three authors was supported by the Australian Research
= Heff x + w
e (4)
Council through the Discovery Project under Grant DP160100528. The work
of Ezio Biglieri was supported by the European Research Council under the where Heff , (FN ⊗IM )H(F†N ⊗IM ) is the effective channel
H2020 Framework Programme/ERC grant agreement 694974. e = (FN ⊗ IM )w the noise vector.
matrix, and w
 2

n j2π ki([m − li]M )


e−j2π
  Ne M N , if q = [m − li]M + M[n − ki]N and m < li
ki([m − li]M )
Ξi(p, q) = e j2π M N , if q = [m − li]M + M[n − ki]N and m ≥ li (8)

0, otherwise.

In order to study the diversity of OTFS, we use the following Doppler

two observations from our earlier work [5], [6]. Observation (k) •
1. The received signal y can be written as [6] 

PX •


y= hi(FN ⊗ IM )Πli∆ki (F†N ⊗ IM )x+we =Φ(x)h+we
i=1 Delay(l)
CaseA CaseB CaseC
where h = [h1, h2, · · · , hP ] is a P × 1 vector of i.i.d. complex
Gaussian random variables, and Φ(x) is the MN×P Fig. 1. Examples for the three cases in the diversity analysis for M = N = 4
concatenated matrix
III. DIVERSITY ANALYSIS FOR P = 2
Φ(x) = [ Ξ x |...|Ξ ]
|{z}1 |{z} |{z}P x Without loss of generality, we assume that the two channel
MN×MN MN×1 MN×MN MN×1
|{z} paths are located at (l1, k1) = (0, 0) and at (l2, k2). We divide
and Ξi , (FN ⊗ IM )Πli∆ki (F†N ⊗ IM ). the diversity analysis of OTFS into the following three cases: A.
The conditional pairwise error probability for the above l2 6= 0 and k2 6= 0, B. l2 = 6 0 and k2 = 0, C. l2 = 0 and k2 6= 0.
system can be written as Fig. 1 describes the examples of the three cases for the OTFS
system of M = N = 4. Here, the dots represent the delay and
P (x → xˆ | h) = P(k y − Φ(xˆ)hk2 < ky − Φ(x)hk2 | h) Doppler positions of the two channel paths.
 k[Φ(xˆ) − 
=Q √
2NΦ(x)]hk
0
6 0 and k2 6= 0
A. l2 =
where Q is the Gaussian tail function.
Since there are no shifts across the delays or Dopplers due to
Now, using the Chernoff upperbound, the pairwise error
the first path (i.e., Ξ1 = I), the first column of Φ(δ) in (5) can
probability becomes
be written as
−kΦ(δ)hk2
 
P (x → ˆx) ≤ Eh exp 4N0 Φ1(δ) = δ = [δ0,0, δ1,0, · · · , δM−1,0, δ0,1, δ1,1, · · · , δM−1,1,
where · · · , δ0,N−1, δ1,N−1, · · · , δM−1,N−1]T (9)

Φ(δ) = [Ξ1δ | . . . | ΞP δ] (5) where δm,n denotes the m + Mn element of δ. The second
column of Φ(δ) is (l2, k2) circulant shifted version of Φ1(δ)
and δ = xˆ − x is the codeword difference vector. Note that δ with some phase shifts as given in (8). Given that we exclude
⊂ Z[j]MN×1. 6 xˆ), we have at least one element δm,n 6= 0
the case δ = 0 (x =
†
Assuming h ∼ CN (0, I) and Υ(δ) = Φ (δ)Φ(δ), the upper for some m and n. Assume m ≥ l21. Let us consider two rows
bound on the average bit error probability (BER) Pe can be in Φ(δ) with δm,n as one of the element:
written as [6, Eq. (23)]
k2( m −l2)
δm,n , ej2π δδm−l (10)
M N

X X dH(x, ˆx) Y !−1 −ρ(x,xˆ) m 2,n−k2
1
Pe 6 |QMN | λi 1  k ( m)

MNlog2Q i 4N0 j2π 2M N
+l2,n+k2 , e δm,n (11)
x∈ANMxˆ6=x
(6)
The second element in (10) can be made zero by multiplying δm
where, MN log2 Q denotes the total number of bits transmit-ted −l2,n−k2 j2π 2
k ( m −l2)

in one OTFS frame, dH(x, ˆx) is the difference in number of subtracting

δ m,n e M N by the first column of Φ(δ) and
from second column. The second column entry
information bits between x and xˆ, λi’s are the non-zero in (11) becomes
eigenvalues of Υ(δ), and ρ(x, xˆ) is the number of nonzero
k 2 m k2( m −l2 δ)m−l2,n−k2
eigenvalues of Υ(δ). ej2π M Nδm,n − ej2π M N δm+l2,n+k2 (12)
Finally, the “standard” definition of diversity of OTFS is δm,n

ρ(x, xˆ) = min rank (Υ(δ)) (7) If we can show that the value in (12) is non-zero for all values
ρ, min
x6=xˆ,(x,xˆ) x6=xˆ,(x,xˆ) of δm,n ∈ Z[j], then we can say the two columns of Φ(δ) are
independent for QAM modulation and we guarantee a diversity
Observation 2. In matrix Ξi as given in (8), where 0 ≤ p ≤ order 2, which is the number P of paths and hence the
MN −1 and 0 ≤ q ≤ MN −1 [5, Theorem 1], the values of n maximum achievable value.
and m can be computed as n = b p c and m = p −nM. Note M 1Note that for m < l , the expression in (12) will be similar except a phase
2
that Ξiδ is equivalent to a circulant shift of δ with some phase n −j2π
shift of eN in the second term. Assume the value of (12) for m < l2
shifts. In next section, we derive the diversity order achieved by is zero, i.e., δm+l2,n+k2 6= 0. By repeating (12) for δm+l2,n+k2 , we can
OTFS system for two channel paths, i.e., P = 2. 6 0, for which m0 > l2.
always get a condition δm0,n0 =
 3

The value in (12) becomes zero if where θ is of the form 2π Ml2N c, for any c ∈ Z. As δ =
k2 l2 (F†N ⊗ IM )† δ̂ = (FN ⊗ IM )δ̂ , from (21), the value of δ
2
δm,n = e−j2π M N δm−l2 ,n−k2 δm+l2 ,n+k2 (13) becomes
k2 l2
If l2 6= 0 and k2 6= 0 then the value e−j2π M N is ir- [δm,0 , δm,1 , · · · , δm,N −1 ]T
k2 l2
rational belonging to the cyclotomic field Q(e−j2π M N ) of h M M
iT
degree ϕ(min(M N, M N/k2 l2 )) ≥ 2 for any k2 , l2 such that = FN ejpθ δ̂(q), ejpθ ej l2 θ δ̂(q), · · · , ejpθ ej l2 (N −1)θ δ̂(q)
k2 l2 < M N/4 and the above equality cannot be satisfied by
= ejpθ δ̂(q)[0, · · · , |{z}
1 , 0, · · · , 0]T (22)
δn,m ∈ Z[j] (Here ϕ(n) denotes the Euler totient function,
(c0 +1)th entry
counting the number of relatively primes to n, and the second
argument of the min only applies for k2 l2 dividing M N .) where p = bm/l2 c, q = [m]l2 , 0 ≤ m ≤ M −1, and c0 = [c]N .
Since the typical wireless channels are underspread, i.e., Therefore, the entries of δ can be written as
l2  M, k2  N , for P = 2 paths at (0, 0) and (l2 , k2 ) with (
6 0 and k2 6= 0, we have diversity order 2.
l2 = ejpθ δq,c0 +1 , if n = c0 + 1
δm,n = (23)
Example: For M = N = 4 and k2 = l2 = 1, the condition 0, otherwise.
in (12) becomes
1 for 0 ≤ m ≤ M − 1. However, as δm,n ∈ Z[j], from (23) we
δ 2m,n = e−jπ8 δm−l2,n−k2 δm+l2,n+k2 (14)
can derive that ejpθ ∈ Z[j], which is possible only when ejθ
Since (14) cannot be satisfied for δn,m ∈ Z[j], we have takes values from the set {+1, −1, +j, −j}.
Example 1: Consider M = 2, N = 2, l2 = 1, k2 = 0 and
diversity order of 2.  Special case: For M = N = 2 and k2 =
l2 = 1, the conditions in (10) and (11) becomes 4-QAM modulation. The following are some of the δ patterns
that yield rank 1:

δ2,nm= −jδ2 [m−l2]M ,[n−k2]N (or) δ2,nm= jδ2 [m−l2]M ,[n−k2]N

[2, 2, 0, 0]T , [2, −2, 0, 0]T , [−2, 2, 0, 0]T , [−2, −2, 0, 0]T ,
[0, 0, 2, 2j]T , [0, 0, 2j, −2]T , [0, 0, −2, 2j]T , [0, 0, −2, −2j]T
However, for δn,m ∈ Z[j], neither one of the above conditions
can be satisfied, which yields full diversity.  We can clearly see that all the above patterns follow the
structure in (23). 

B. l2 6= 0 and k2 = 0 1) Upper bound on the number of pairs yielding rank 1:

From (23), for a given x, the maximum number of x̂ that yield
Let us rewrite the Φ(δ) in (7) as follows:
rank 1 can be computed as 4N (Ql2 −1), where N denotes the
all possible values of c0 , (Ql2 − 1) is the number of possible
 
  δ 0
Φ(δ) = Ξ1 Ξ2 (15)
0 δ selection of QAM symbols for a given c0 which is due to
0 ≤ q ≤ l2 − 1, and 4 is due to the set {+1, −1, +j, −j}.
Now, from (15), Υ(δ) can be written in the form
For the case of l2 not dividing M , the value of l2 in the
condition (21) is replaced by l20 = gcd(l2 , M ).
 †
δ † Λδ

δ δ
Υ(δ) = Φ† (δ)Φ(δ) = † † (16)
δ Λ δ δ† δ

where Λ , Ξ†1 Ξ2 , and we use the fact that Ξ†1 Ξ1 = Ξ†2 Ξ2 = I. C. l2 = 0 and k2 6= 0
6 0, Λ can be written as
For k2 = 0, l2 = For the case of l2 = 0 and k2 6= 0, the value of Λ in (16)
Λ= (F†N ⊗ IM ) Π † l2
(F†N ⊗ IM ) (17) becomes

Therefore, the determinant of Υ(δ) is equal to Λ = (F†N ⊗ IM )† ∆k2 (F†N ⊗ IM ) (24)

det[Υ(δ)] = |δ † δ|2 − |δ † Λδ|2 (18) Now, the determinant of Υ(δ) can be written as
= |δ̂ † δ̂|2 − |δ̂ † Πl2 δ̂|2 (19) det[Υ(δ)] = |δ̂ † δ̂|2 − |δ̂ † ∆k2 δ̂|2 (25)
where δ̂ = (F†N
⊗ IM )δ. From (19) and (7), using Cauchy- where δ̂ = (F†N ⊗ IM )δ. Therefore, using Cauchy-Schwarz
Schwarz inequality, we can see that OTFS achieves rank 1 inequality, we can see that OTFS achieves rank 1 only when
only when
δ̂ = ejθ ∆k2 δ̂ (26)
δ̂ = ejθ Πl2 δ̂ (20)
for some angle θ. For simplicity, we assume l2 divides M . for some angle θ. For simplicity, we assume k2 divides N .
Through simple algebraic computations, the condition in (20) Assuming δ̂(r) 6= 0 for some 0 ≤ r ≤ M N − 1, from (26),
can be written as we can obtain the conditions on δ̂ as

δ̂(r) = ejθ δ̂([r − l2 ]M ), for 0 ≤ r ≤ M N − 1, (21) δ̂(r0 ) = 0 for [r0 − r] M N 6= 0 (27)

k2
 4

That is, δ̂ contains at the most k2 non-zero elements separated Example 3: Consider M = 16, N = 16, l2 = 4, k2 = 0, and
by Mk2N . Now, similar to (22), the value of δ can be written 4-QAM modulation. Then the value of R(x, x̂) is ≈ 10−150 .
as 
[δm,0 , δm,1 , · · · , δm,N −1 ]T Now, consider again the union bound to Pe, as shown in (6).
h iT
= FN δ̂(m), δ̂(m + M ), · · · , δ̂(m + (N − 1)M ) (28) Each individual term in sum (6) can be viewed as a function of
SNR whose slope is essentially dictated by the value of ρ(x,
From (27), we can immediately see that the vector in (28) can ˆx). As SNR increases, the terms with higher values of ρ(x,
have at the most k2 non-zero elements for one value of m, xˆ) become increasingly less relevant to determine the value of
0 ≤ m ≤ M − 1, and all zero elements for all other m. Pe, so that for very high SNR the only significant terms in (6)
Example 2: Consider M = 2, N = 2, l2 = 0, k2 = 1 and are those with the lowest value of ρ(x, xˆ). This justifies the
4-QAM modulation. As k2 = 1, only one element in (28) can use of the standard definition (7) of diversity ρ, useful for very
be non-zero for some value of m = m0 . Therefore, the entries high SNR. For intermediate values of SNR, however, the use of
of δ can be written as standard diversity might not be appropriate whenever the
( nc number of terms in (6) with exponent higher than ρ is much
e−j2π N δm0 ,n , if m = m0 greater than the number of those with expo-nent ρ. If this is the
δm,n = (29)
0, otherwise. case, for those values of SNR the slope of the union-bound
curve might be dictated by an exponent higher than ρ. We call
for 0 ≤ n, c ≤ N −1. The following are some of the δ patterns
this exponent “effective diversity.” In our context, OTFS
that yield rank 1:
achieves full effective diversity with typical system parameters
[2, 0, 2, 0]T , [−2, 0, −2, 0]T , [−2, 0, 2, 0]T , [2, 0, −2, 0]T , as P = 2. Future work will show the extension of the above
[0, −2j, 0, 2j]T , [0, 2j, 0, 2j]T , [0, 2j, 0, −2j]T , [0, 2, 0, −2]T , analysis to P values greater than 2.
Remark 1: Note that the above diversity analysis differs
We can clearly see that all the above patterns follow the from the work in [8] in two aspects.
structure in (29).  1) The diversity analysis in [8] is proposed only for the
ideal biorthogonal waveforms that satisfy both time and
1) Upper bound on the number of pairs yielding rank 1: frequency orthogonality conditions, which is not practically
From (28), for a given x, the maximum number of x̂ yielding feasible due to Heisenberg’s uncertainty principle. Instead,
rank 1 can be computed as M kN2 (Qk2 − 1), where M denotes the analysis in our work assumes feasible rectangular
the all possible values of m, (Qk2 − 1) is the number of waveforms, which yield different input–output relations
possible selection of QAM symbols for a given m which is (extra phase shifts in (7)) compared to the ideal waveforms
due to the k2 non-zero elements in (28), and kN2 is the possible [5], [8]. Moreover, our analysis developed for rectangular
k2 non-zero sets in (28). For the case of k2 not dividing N , waveforms can be straightforwardly extended to any ar-
the value of k2 in (27) is replaced by k20 = gcd(k2 , N ) and δ̂ bitrary waveforms, for example, raised cosine waveforms
can contain at the most k20 non-zero elements. [5].
2) The work in [8] showed that the standard diversity order
D. Diversity for the three cases above of the OTFS system is one and proposed a phase rotation
based precoding scheme to achieve full diversity.
Summarizing our analysis of the three cases above, an upper However, in this work, we show that, with sufficiently
bound N (x, x̂) to the number of pairs (x, x̂) yielding rank 1 large M and N, the OTFS system achieves full effective
is diversity, thus any precoding schemes are not required.

0,
 if l2 6= 0, k2 6= 0
0
N (x, x̂) = 4N (Ql2 − 1)QM N , if l2 6= 0, k2 = 0 (30) IV. S IMULATION R ESULTS
M N (Qk20 − 1)QM N , if l = 0, k 6= 0

k0 2
2 2 In this section, we present the BER of OTFS for different
choices of system parameters. We consider P = 2, l1 =
Now, consider the ratio of N (x, x̂) to the total number of 0, k1 = 0, and 4-QAM modulation in all the simulations.
possible pairs T (x, x̂) = QM N (QM N − 1). we obtain The path coefficients are h1 , h2 ∼ CN (0, 1/2). Figs. 2a,
2b, and 3 show the BER of OTFS for different (l2 , k2 ) with

0,
 if l2 6= 0, k2 6= 0
N (x, x̂)  4N 0 , (M = 2, N = 2), (M = 4, N = 4), and (M = 16, N = 16)
R(x, x̂) = ≈ Q(M N −l2 ) if l2 6= 0, k2 = 0
T (x, x̂) respectively. We use the optimal ML detector in Figs. 2a and
MN


 , if l = 0, k =
0 2 6 0 2 2b and the message passing detector [7] in Fig. 3. Note that the
k20 Q(M N −k2 )
(31) plots of c1 (SNR)−2 and c2 (SNR)−1 are only used to identify
the slope of the curves, and do not represent an upper bound.
As l2  M, k2  N , we can see that, as M and N grow From the figures, we can observe that:
larger, R(x, x̂) quickly approaches zero. i) In all the three cases, OTFS achieves the full diversity of
M,N →∞ two with the second path located at l2 = 1, k2 = 1. This is
R(x, x̂) −−−−−−→ 0 (32) due to the absence of rank 1 pairs if k2 6= k1 and l2 =6 l1 .
 5

R EFERENCES
100 100
[1] E. Biglieri, Coding for Wireless Channels. New York: Springer, 2005.
[2] R. Hadani, S. Rakib, M. Tsatsanis, A. Monk, A. J. Goldsmith, A.
F. Molisch, and R. Calderbank, “Orthogonal time frequency space
10-2 10-2 modulation,” in Proc. IEEE WCNC, San Francisco, CA, USA, March
2017.
BER

[3] R. Hadani, S. Rakib, S. Kons, M. Tsatsanis, A. Monk, C. Ibars, J.

Delfeld, Y. Hebron, A. J. Goldsmith, A.F. Molisch, and R. Calder-
10-4 10-4 bank, “Orthogonal time frequency space modulation,” Available online:
https://arxiv.org/pdf/1808.00519.pdf.
[4] R. Hadani and A. Monk, “OTFS: A new generation of
modulation addressing the challenges of 5G,” OTFS Physics
White Paper, Cohere Technologies, 7 Feb. 2018. Available online:
10-6 10-6
5 10 15 20 25 5 10 15 20 25 30 https://arxiv.org/pdf/1802.02623.pdf.
SNR in dB SNR in dB [5] P. Raviteja, Y. Hong, E. Viterbo, and E. Biglieri, “Practical pulse-shaping
waveforms for reduced-cyclic-prefix OTFS,” IEEE Trans. Veh. Technol.,
(a) M = 2, N = 2 (b) M = 4, N = 4 vol. 68, no. 1, pp. 957-961, Jan. 2019.
Fig. 2. BER of OTFS for different (l2 , k2 ) with P = 2 and 4-QAM. [6] E. Biglieri, P. Raviteja, and Y. Hong, “Error performance of orthogonal
time frequency space (OTFS) modulation,” in Proc. IEEE ICC Workshop
2019, Shanghai, China, May 2019.
[7] P. Raviteja, K. T. Phan, Y. Hong, and E. Viterbo, “Interference can-
10-1 cellation and iterative detection for orthogonal time frequency space
modulation,” IEEE Trans. Wireless Commun., vol. 17, no. 10, pp. 6501-
6515, Oct. 2018.
10-2 [8] G. D. Surabhi, R. M. Augustine, and A. Chockalingam, “On the diversity
of uncoded OTFS modulation in doubly-dispersive channels,” IEEE
Trans. Wireless Commun., vol. 18, no. 6, pp. 3049-3063, June 2019.
10-3 [9] K. R. Murali and A. Chockalingam, “On OTFS modulation for high-
BER

Doppler fading channels,” in Proc. ITA’2018, San Diego, Feb. 2018.

[10] L. Li, H. Wei, Y. Huang, Y. Yao, W. Ling, G. Chen, P. Li, and Y. Cai, “A
10-4 simple two-stage equalizer With simplified orthogonal time frequency
space modulation over rapidly time-varying channels,” available online:
https://arxiv.org/abs/1709.02505.
10-5 [11] T. Zemen, M. Hofer, D. Loeschenbrand, and C. Pacher, “Iterative
detection for orthogonal precoding in doubly selective channels,” in
Proc. 2018 IEEE 29th Annual International Symposium on PIMRC,
10-6 Bologna, pp. 1-7, Sept. 2018.
10 15 20 25 30 [12] A. Farhang, A. Rezazadeh, Reyhani, L. E. Doyle, and B. Farhang-
SNR in dB Boroujeny, “Low complexity modem structure for OFDM-based orthog-
onal time frequency space modulation,” IEEE Wireless Commun. Lett.,
Fig. 3. BER of OTFS for different (l2 , k2 ) with M = 16, N = 16, P = 2, vol. 7, no. 3, pp. 344-347, June 2018.
and 4-QAM. [13] A. RezazadehReyhani, A. Farhang, M. Ji, R. R. Chen, and B. Farhang-
Boroujeny, “Analysis of discrete-time MIMO OFDM-based orthogonal
time frequency space modulation,” in Proc. 2018 IEEE ICC, Kansas
City, MO, USA, pp. 1-6, May 2018.
ii) OTFS achieves rank 1 with the second paths located at [14] V. Khammammetti and S. K. Mohammed, “OTFS-based multiple-access
l2 = 1, k2 = 0 or l2 = 0, k2 = 1 for (M = 2, N = 2) and in high Doppler and delay spread wireless channels,” IEEE Wireless
(M = 4, N = 4). This is because of the upper bound in (30) Commun. Lett., vol. 8, no. 2, pp. 528-531, April 2019.
is significant for these system parameters. However, we can
see that the diversity change happens at a higher SNR with
(M = 4, N = 4) compared to the case of (M = 2, N = 2).
iii) Finally, independent of the second path location, OTFS
achieves the same performance for (M = 16, N = 16) for the
BER ranges up to 10−5 . This is justified by the small number
of rank-1 pairs as mentioned in Example 3.
From the above observations, we conclude that OTFS
system achieves full effective diversity for sufficiently large
values of M and N .

V. C ONCLUSION
We have analyzed the diversity of OTFS over two-path
channels. After deriving an upper bound on the number of
pairs that prevent the achievement of full rank, we have shown
that the number of these pairs is relatively vanishingly small
for sufficiently large values of M and N (e.g., M = N = 16).
Through analysis and simulations, we can conclude that, even
though the theoretical diversity of OTFS is one, effective
diversity, which takes value 2, is the significant parameter
expressing error performance.